xref: /linux/drivers/infiniband/hw/hns/hns_roce_hw_v2.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 <linux/workqueue.h>
40 #include <net/addrconf.h>
41 #include <rdma/ib_addr.h>
42 #include <rdma/ib_cache.h>
43 #include <rdma/ib_umem.h>
44 #include <rdma/uverbs_ioctl.h>
45 
46 #include "hnae3.h"
47 #include "hns_roce_common.h"
48 #include "hns_roce_device.h"
49 #include "hns_roce_cmd.h"
50 #include "hns_roce_hem.h"
51 #include "hns_roce_hw_v2.h"
52 
53 enum {
54 	CMD_RST_PRC_OTHERS,
55 	CMD_RST_PRC_SUCCESS,
56 	CMD_RST_PRC_EBUSY,
57 };
58 
59 enum ecc_resource_type {
60 	ECC_RESOURCE_QPC,
61 	ECC_RESOURCE_CQC,
62 	ECC_RESOURCE_MPT,
63 	ECC_RESOURCE_SRQC,
64 	ECC_RESOURCE_GMV,
65 	ECC_RESOURCE_QPC_TIMER,
66 	ECC_RESOURCE_CQC_TIMER,
67 	ECC_RESOURCE_SCCC,
68 	ECC_RESOURCE_COUNT,
69 };
70 
71 static const struct {
72 	const char *name;
73 	u8 read_bt0_op;
74 	u8 write_bt0_op;
75 } fmea_ram_res[] = {
76 	{ "ECC_RESOURCE_QPC",
77 	  HNS_ROCE_CMD_READ_QPC_BT0, HNS_ROCE_CMD_WRITE_QPC_BT0 },
78 	{ "ECC_RESOURCE_CQC",
79 	  HNS_ROCE_CMD_READ_CQC_BT0, HNS_ROCE_CMD_WRITE_CQC_BT0 },
80 	{ "ECC_RESOURCE_MPT",
81 	  HNS_ROCE_CMD_READ_MPT_BT0, HNS_ROCE_CMD_WRITE_MPT_BT0 },
82 	{ "ECC_RESOURCE_SRQC",
83 	  HNS_ROCE_CMD_READ_SRQC_BT0, HNS_ROCE_CMD_WRITE_SRQC_BT0 },
84 	/* ECC_RESOURCE_GMV is handled by cmdq, not mailbox */
85 	{ "ECC_RESOURCE_GMV",
86 	  0, 0 },
87 	{ "ECC_RESOURCE_QPC_TIMER",
88 	  HNS_ROCE_CMD_READ_QPC_TIMER_BT0, HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0 },
89 	{ "ECC_RESOURCE_CQC_TIMER",
90 	  HNS_ROCE_CMD_READ_CQC_TIMER_BT0, HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0 },
91 	{ "ECC_RESOURCE_SCCC",
92 	  HNS_ROCE_CMD_READ_SCCC_BT0, HNS_ROCE_CMD_WRITE_SCCC_BT0 },
93 };
94 
95 static inline void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
96 				   struct ib_sge *sg)
97 {
98 	dseg->lkey = cpu_to_le32(sg->lkey);
99 	dseg->addr = cpu_to_le64(sg->addr);
100 	dseg->len  = cpu_to_le32(sg->length);
101 }
102 
103 /*
104  * mapped-value = 1 + real-value
105  * The hns wr opcode real value is start from 0, In order to distinguish between
106  * initialized and uninitialized map values, we plus 1 to the actual value when
107  * defining the mapping, so that the validity can be identified by checking the
108  * mapped value is greater than 0.
109  */
110 #define HR_OPC_MAP(ib_key, hr_key) \
111 		[IB_WR_ ## ib_key] = 1 + HNS_ROCE_V2_WQE_OP_ ## hr_key
112 
113 static const u32 hns_roce_op_code[] = {
114 	HR_OPC_MAP(RDMA_WRITE,			RDMA_WRITE),
115 	HR_OPC_MAP(RDMA_WRITE_WITH_IMM,		RDMA_WRITE_WITH_IMM),
116 	HR_OPC_MAP(SEND,			SEND),
117 	HR_OPC_MAP(SEND_WITH_IMM,		SEND_WITH_IMM),
118 	HR_OPC_MAP(RDMA_READ,			RDMA_READ),
119 	HR_OPC_MAP(ATOMIC_CMP_AND_SWP,		ATOM_CMP_AND_SWAP),
120 	HR_OPC_MAP(ATOMIC_FETCH_AND_ADD,	ATOM_FETCH_AND_ADD),
121 	HR_OPC_MAP(SEND_WITH_INV,		SEND_WITH_INV),
122 	HR_OPC_MAP(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
123 	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
124 	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
125 };
126 
127 static u32 to_hr_opcode(u32 ib_opcode)
128 {
129 	if (ib_opcode >= ARRAY_SIZE(hns_roce_op_code))
130 		return HNS_ROCE_V2_WQE_OP_MASK;
131 
132 	return hns_roce_op_code[ib_opcode] ? hns_roce_op_code[ib_opcode] - 1 :
133 					     HNS_ROCE_V2_WQE_OP_MASK;
134 }
135 
136 static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
137 			 const struct ib_reg_wr *wr)
138 {
139 	struct hns_roce_wqe_frmr_seg *fseg =
140 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
141 	struct hns_roce_mr *mr = to_hr_mr(wr->mr);
142 	u64 pbl_ba;
143 
144 	/* use ib_access_flags */
145 	hr_reg_write_bool(fseg, FRMR_BIND_EN, wr->access & IB_ACCESS_MW_BIND);
146 	hr_reg_write_bool(fseg, FRMR_ATOMIC,
147 			  wr->access & IB_ACCESS_REMOTE_ATOMIC);
148 	hr_reg_write_bool(fseg, FRMR_RR, wr->access & IB_ACCESS_REMOTE_READ);
149 	hr_reg_write_bool(fseg, FRMR_RW, wr->access & IB_ACCESS_REMOTE_WRITE);
150 	hr_reg_write_bool(fseg, FRMR_LW, wr->access & IB_ACCESS_LOCAL_WRITE);
151 
152 	/* Data structure reuse may lead to confusion */
153 	pbl_ba = mr->pbl_mtr.hem_cfg.root_ba;
154 	rc_sq_wqe->msg_len = cpu_to_le32(lower_32_bits(pbl_ba));
155 	rc_sq_wqe->inv_key = cpu_to_le32(upper_32_bits(pbl_ba));
156 
157 	rc_sq_wqe->byte_16 = cpu_to_le32(wr->mr->length & 0xffffffff);
158 	rc_sq_wqe->byte_20 = cpu_to_le32(wr->mr->length >> 32);
159 	rc_sq_wqe->rkey = cpu_to_le32(wr->key);
160 	rc_sq_wqe->va = cpu_to_le64(wr->mr->iova);
161 
162 	hr_reg_write(fseg, FRMR_PBL_SIZE, mr->npages);
163 	hr_reg_write(fseg, FRMR_PBL_BUF_PG_SZ,
164 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
165 	hr_reg_clear(fseg, FRMR_BLK_MODE);
166 }
167 
168 static void set_atomic_seg(const struct ib_send_wr *wr,
169 			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
170 			   unsigned int valid_num_sge)
171 {
172 	struct hns_roce_v2_wqe_data_seg *dseg =
173 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
174 	struct hns_roce_wqe_atomic_seg *aseg =
175 		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
176 
177 	set_data_seg_v2(dseg, wr->sg_list);
178 
179 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
180 		aseg->fetchadd_swap_data = cpu_to_le64(atomic_wr(wr)->swap);
181 		aseg->cmp_data = cpu_to_le64(atomic_wr(wr)->compare_add);
182 	} else {
183 		aseg->fetchadd_swap_data =
184 			cpu_to_le64(atomic_wr(wr)->compare_add);
185 		aseg->cmp_data = 0;
186 	}
187 
188 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
189 }
190 
191 static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
192 				 const struct ib_send_wr *wr,
193 				 unsigned int *sge_idx, u32 msg_len)
194 {
195 	struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
196 	unsigned int left_len_in_pg;
197 	unsigned int idx = *sge_idx;
198 	unsigned int i = 0;
199 	unsigned int len;
200 	void *addr;
201 	void *dseg;
202 
203 	if (msg_len > qp->sq.ext_sge_cnt * HNS_ROCE_SGE_SIZE) {
204 		ibdev_err(ibdev,
205 			  "no enough extended sge space for inline data.\n");
206 		return -EINVAL;
207 	}
208 
209 	dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
210 	left_len_in_pg = hr_hw_page_align((uintptr_t)dseg) - (uintptr_t)dseg;
211 	len = wr->sg_list[0].length;
212 	addr = (void *)(unsigned long)(wr->sg_list[0].addr);
213 
214 	/* When copying data to extended sge space, the left length in page may
215 	 * not long enough for current user's sge. So the data should be
216 	 * splited into several parts, one in the first page, and the others in
217 	 * the subsequent pages.
218 	 */
219 	while (1) {
220 		if (len <= left_len_in_pg) {
221 			memcpy(dseg, addr, len);
222 
223 			idx += len / HNS_ROCE_SGE_SIZE;
224 
225 			i++;
226 			if (i >= wr->num_sge)
227 				break;
228 
229 			left_len_in_pg -= len;
230 			len = wr->sg_list[i].length;
231 			addr = (void *)(unsigned long)(wr->sg_list[i].addr);
232 			dseg += len;
233 		} else {
234 			memcpy(dseg, addr, left_len_in_pg);
235 
236 			len -= left_len_in_pg;
237 			addr += left_len_in_pg;
238 			idx += left_len_in_pg / HNS_ROCE_SGE_SIZE;
239 			dseg = hns_roce_get_extend_sge(qp,
240 						idx & (qp->sge.sge_cnt - 1));
241 			left_len_in_pg = 1 << HNS_HW_PAGE_SHIFT;
242 		}
243 	}
244 
245 	*sge_idx = idx;
246 
247 	return 0;
248 }
249 
250 static void set_extend_sge(struct hns_roce_qp *qp, struct ib_sge *sge,
251 			   unsigned int *sge_ind, unsigned int cnt)
252 {
253 	struct hns_roce_v2_wqe_data_seg *dseg;
254 	unsigned int idx = *sge_ind;
255 
256 	while (cnt > 0) {
257 		dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
258 		if (likely(sge->length)) {
259 			set_data_seg_v2(dseg, sge);
260 			idx++;
261 			cnt--;
262 		}
263 		sge++;
264 	}
265 
266 	*sge_ind = idx;
267 }
268 
269 static bool check_inl_data_len(struct hns_roce_qp *qp, unsigned int len)
270 {
271 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
272 	int mtu = ib_mtu_enum_to_int(qp->path_mtu);
273 
274 	if (mtu < 0 || len > qp->max_inline_data || len > mtu) {
275 		ibdev_err(&hr_dev->ib_dev,
276 			  "invalid length of data, data len = %u, max inline len = %u, path mtu = %d.\n",
277 			  len, qp->max_inline_data, mtu);
278 		return false;
279 	}
280 
281 	return true;
282 }
283 
284 static int set_rc_inl(struct hns_roce_qp *qp, const struct ib_send_wr *wr,
285 		      struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
286 		      unsigned int *sge_idx)
287 {
288 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
289 	u32 msg_len = le32_to_cpu(rc_sq_wqe->msg_len);
290 	struct ib_device *ibdev = &hr_dev->ib_dev;
291 	unsigned int curr_idx = *sge_idx;
292 	void *dseg = rc_sq_wqe;
293 	unsigned int i;
294 	int ret;
295 
296 	if (unlikely(wr->opcode == IB_WR_RDMA_READ)) {
297 		ibdev_err(ibdev, "invalid inline parameters!\n");
298 		return -EINVAL;
299 	}
300 
301 	if (!check_inl_data_len(qp, msg_len))
302 		return -EINVAL;
303 
304 	dseg += sizeof(struct hns_roce_v2_rc_send_wqe);
305 
306 	if (msg_len <= HNS_ROCE_V2_MAX_RC_INL_INN_SZ) {
307 		hr_reg_clear(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
308 
309 		for (i = 0; i < wr->num_sge; i++) {
310 			memcpy(dseg, ((void *)wr->sg_list[i].addr),
311 			       wr->sg_list[i].length);
312 			dseg += wr->sg_list[i].length;
313 		}
314 	} else {
315 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
316 
317 		ret = fill_ext_sge_inl_data(qp, wr, &curr_idx, msg_len);
318 		if (ret)
319 			return ret;
320 
321 		hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, curr_idx - *sge_idx);
322 	}
323 
324 	*sge_idx = curr_idx;
325 
326 	return 0;
327 }
328 
329 static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
330 			     struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
331 			     unsigned int *sge_ind,
332 			     unsigned int valid_num_sge)
333 {
334 	struct hns_roce_v2_wqe_data_seg *dseg =
335 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
336 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
337 	int j = 0;
338 	int i;
339 
340 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_MSG_START_SGE_IDX,
341 		     (*sge_ind) & (qp->sge.sge_cnt - 1));
342 
343 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_INLINE,
344 		     !!(wr->send_flags & IB_SEND_INLINE));
345 	if (wr->send_flags & IB_SEND_INLINE)
346 		return set_rc_inl(qp, wr, rc_sq_wqe, sge_ind);
347 
348 	if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
349 		for (i = 0; i < wr->num_sge; i++) {
350 			if (likely(wr->sg_list[i].length)) {
351 				set_data_seg_v2(dseg, wr->sg_list + i);
352 				dseg++;
353 			}
354 		}
355 	} else {
356 		for (i = 0; i < wr->num_sge && j < HNS_ROCE_SGE_IN_WQE; i++) {
357 			if (likely(wr->sg_list[i].length)) {
358 				set_data_seg_v2(dseg, wr->sg_list + i);
359 				dseg++;
360 				j++;
361 			}
362 		}
363 
364 		set_extend_sge(qp, wr->sg_list + i, sge_ind,
365 			       valid_num_sge - HNS_ROCE_SGE_IN_WQE);
366 	}
367 
368 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
369 
370 	return 0;
371 }
372 
373 static int check_send_valid(struct hns_roce_dev *hr_dev,
374 			    struct hns_roce_qp *hr_qp)
375 {
376 	struct ib_device *ibdev = &hr_dev->ib_dev;
377 
378 	if (unlikely(hr_qp->state == IB_QPS_RESET ||
379 		     hr_qp->state == IB_QPS_INIT ||
380 		     hr_qp->state == IB_QPS_RTR)) {
381 		ibdev_err(ibdev, "failed to post WQE, QP state %u!\n",
382 			  hr_qp->state);
383 		return -EINVAL;
384 	} else if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)) {
385 		ibdev_err(ibdev, "failed to post WQE, dev state %d!\n",
386 			  hr_dev->state);
387 		return -EIO;
388 	}
389 
390 	return 0;
391 }
392 
393 static unsigned int calc_wr_sge_num(const struct ib_send_wr *wr,
394 				    unsigned int *sge_len)
395 {
396 	unsigned int valid_num = 0;
397 	unsigned int len = 0;
398 	int i;
399 
400 	for (i = 0; i < wr->num_sge; i++) {
401 		if (likely(wr->sg_list[i].length)) {
402 			len += wr->sg_list[i].length;
403 			valid_num++;
404 		}
405 	}
406 
407 	*sge_len = len;
408 	return valid_num;
409 }
410 
411 static __le32 get_immtdata(const struct ib_send_wr *wr)
412 {
413 	switch (wr->opcode) {
414 	case IB_WR_SEND_WITH_IMM:
415 	case IB_WR_RDMA_WRITE_WITH_IMM:
416 		return cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
417 	default:
418 		return 0;
419 	}
420 }
421 
422 static int set_ud_opcode(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
423 			 const struct ib_send_wr *wr)
424 {
425 	u32 ib_op = wr->opcode;
426 
427 	if (ib_op != IB_WR_SEND && ib_op != IB_WR_SEND_WITH_IMM)
428 		return -EINVAL;
429 
430 	ud_sq_wqe->immtdata = get_immtdata(wr);
431 
432 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
433 
434 	return 0;
435 }
436 
437 static int fill_ud_av(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
438 		      struct hns_roce_ah *ah)
439 {
440 	struct ib_device *ib_dev = ah->ibah.device;
441 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
442 
443 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_UDPSPN, ah->av.udp_sport);
444 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_HOPLIMIT, ah->av.hop_limit);
445 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_TCLASS, ah->av.tclass);
446 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_FLOW_LABEL, ah->av.flowlabel);
447 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SL, ah->av.sl);
448 
449 	ud_sq_wqe->sgid_index = ah->av.gid_index;
450 
451 	memcpy(ud_sq_wqe->dmac, ah->av.mac, ETH_ALEN);
452 	memcpy(ud_sq_wqe->dgid, ah->av.dgid, GID_LEN_V2);
453 
454 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
455 		return 0;
456 
457 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN_EN, ah->av.vlan_en);
458 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN, ah->av.vlan_id);
459 
460 	return 0;
461 }
462 
463 static inline int set_ud_wqe(struct hns_roce_qp *qp,
464 			     const struct ib_send_wr *wr,
465 			     void *wqe, unsigned int *sge_idx,
466 			     unsigned int owner_bit)
467 {
468 	struct hns_roce_ah *ah = to_hr_ah(ud_wr(wr)->ah);
469 	struct hns_roce_v2_ud_send_wqe *ud_sq_wqe = wqe;
470 	unsigned int curr_idx = *sge_idx;
471 	unsigned int valid_num_sge;
472 	u32 msg_len = 0;
473 	int ret;
474 
475 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
476 
477 	ret = set_ud_opcode(ud_sq_wqe, wr);
478 	if (WARN_ON(ret))
479 		return ret;
480 
481 	ud_sq_wqe->msg_len = cpu_to_le32(msg_len);
482 
483 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_CQE,
484 		     !!(wr->send_flags & IB_SEND_SIGNALED));
485 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SE,
486 		     !!(wr->send_flags & IB_SEND_SOLICITED));
487 
488 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_PD, to_hr_pd(qp->ibqp.pd)->pdn);
489 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SGE_NUM, valid_num_sge);
490 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_MSG_START_SGE_IDX,
491 		     curr_idx & (qp->sge.sge_cnt - 1));
492 
493 	ud_sq_wqe->qkey = cpu_to_le32(ud_wr(wr)->remote_qkey & 0x80000000 ?
494 			  qp->qkey : ud_wr(wr)->remote_qkey);
495 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_DQPN, ud_wr(wr)->remote_qpn);
496 
497 	ret = fill_ud_av(ud_sq_wqe, ah);
498 	if (ret)
499 		return ret;
500 
501 	qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
502 
503 	set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
504 
505 	/*
506 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
507 	 * including new WQEs waiting for the doorbell to update the PI again.
508 	 * Therefore, the owner bit of WQE MUST be updated after all fields
509 	 * and extSGEs have been written into DDR instead of cache.
510 	 */
511 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
512 		dma_wmb();
513 
514 	*sge_idx = curr_idx;
515 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OWNER, owner_bit);
516 
517 	return 0;
518 }
519 
520 static int set_rc_opcode(struct hns_roce_dev *hr_dev,
521 			 struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
522 			 const struct ib_send_wr *wr)
523 {
524 	u32 ib_op = wr->opcode;
525 	int ret = 0;
526 
527 	rc_sq_wqe->immtdata = get_immtdata(wr);
528 
529 	switch (ib_op) {
530 	case IB_WR_RDMA_READ:
531 	case IB_WR_RDMA_WRITE:
532 	case IB_WR_RDMA_WRITE_WITH_IMM:
533 		rc_sq_wqe->rkey = cpu_to_le32(rdma_wr(wr)->rkey);
534 		rc_sq_wqe->va = cpu_to_le64(rdma_wr(wr)->remote_addr);
535 		break;
536 	case IB_WR_SEND:
537 	case IB_WR_SEND_WITH_IMM:
538 		break;
539 	case IB_WR_ATOMIC_CMP_AND_SWP:
540 	case IB_WR_ATOMIC_FETCH_AND_ADD:
541 		rc_sq_wqe->rkey = cpu_to_le32(atomic_wr(wr)->rkey);
542 		rc_sq_wqe->va = cpu_to_le64(atomic_wr(wr)->remote_addr);
543 		break;
544 	case IB_WR_REG_MR:
545 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
546 			set_frmr_seg(rc_sq_wqe, reg_wr(wr));
547 		else
548 			ret = -EOPNOTSUPP;
549 		break;
550 	case IB_WR_SEND_WITH_INV:
551 		rc_sq_wqe->inv_key = cpu_to_le32(wr->ex.invalidate_rkey);
552 		break;
553 	default:
554 		ret = -EINVAL;
555 	}
556 
557 	if (unlikely(ret))
558 		return ret;
559 
560 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
561 
562 	return ret;
563 }
564 
565 static inline int set_rc_wqe(struct hns_roce_qp *qp,
566 			     const struct ib_send_wr *wr,
567 			     void *wqe, unsigned int *sge_idx,
568 			     unsigned int owner_bit)
569 {
570 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
571 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
572 	unsigned int curr_idx = *sge_idx;
573 	unsigned int valid_num_sge;
574 	u32 msg_len = 0;
575 	int ret;
576 
577 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
578 
579 	rc_sq_wqe->msg_len = cpu_to_le32(msg_len);
580 
581 	ret = set_rc_opcode(hr_dev, rc_sq_wqe, wr);
582 	if (WARN_ON(ret))
583 		return ret;
584 
585 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_FENCE,
586 		     (wr->send_flags & IB_SEND_FENCE) ? 1 : 0);
587 
588 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SE,
589 		     (wr->send_flags & IB_SEND_SOLICITED) ? 1 : 0);
590 
591 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_CQE,
592 		     (wr->send_flags & IB_SEND_SIGNALED) ? 1 : 0);
593 
594 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
595 	    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
596 		if (msg_len != ATOMIC_WR_LEN)
597 			return -EINVAL;
598 		set_atomic_seg(wr, rc_sq_wqe, valid_num_sge);
599 	} else if (wr->opcode != IB_WR_REG_MR) {
600 		ret = set_rwqe_data_seg(&qp->ibqp, wr, rc_sq_wqe,
601 					&curr_idx, valid_num_sge);
602 		if (ret)
603 			return ret;
604 	}
605 
606 	/*
607 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
608 	 * including new WQEs waiting for the doorbell to update the PI again.
609 	 * Therefore, the owner bit of WQE MUST be updated after all fields
610 	 * and extSGEs have been written into DDR instead of cache.
611 	 */
612 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
613 		dma_wmb();
614 
615 	*sge_idx = curr_idx;
616 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OWNER, owner_bit);
617 
618 	return ret;
619 }
620 
621 static inline void update_sq_db(struct hns_roce_dev *hr_dev,
622 				struct hns_roce_qp *qp)
623 {
624 	if (unlikely(qp->state == IB_QPS_ERR)) {
625 		flush_cqe(hr_dev, qp);
626 	} else {
627 		struct hns_roce_v2_db sq_db = {};
628 
629 		hr_reg_write(&sq_db, DB_TAG, qp->qpn);
630 		hr_reg_write(&sq_db, DB_CMD, HNS_ROCE_V2_SQ_DB);
631 		hr_reg_write(&sq_db, DB_PI, qp->sq.head);
632 		hr_reg_write(&sq_db, DB_SL, qp->sl);
633 
634 		hns_roce_write64(hr_dev, (__le32 *)&sq_db, qp->sq.db_reg);
635 	}
636 }
637 
638 static inline void update_rq_db(struct hns_roce_dev *hr_dev,
639 				struct hns_roce_qp *qp)
640 {
641 	if (unlikely(qp->state == IB_QPS_ERR)) {
642 		flush_cqe(hr_dev, qp);
643 	} else {
644 		if (likely(qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)) {
645 			*qp->rdb.db_record =
646 					qp->rq.head & V2_DB_PRODUCER_IDX_M;
647 		} else {
648 			struct hns_roce_v2_db rq_db = {};
649 
650 			hr_reg_write(&rq_db, DB_TAG, qp->qpn);
651 			hr_reg_write(&rq_db, DB_CMD, HNS_ROCE_V2_RQ_DB);
652 			hr_reg_write(&rq_db, DB_PI, qp->rq.head);
653 
654 			hns_roce_write64(hr_dev, (__le32 *)&rq_db,
655 					 qp->rq.db_reg);
656 		}
657 	}
658 }
659 
660 static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
661 			      u64 __iomem *dest)
662 {
663 #define HNS_ROCE_WRITE_TIMES 8
664 	struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
665 	struct hnae3_handle *handle = priv->handle;
666 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
667 	int i;
668 
669 	if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
670 		for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
671 			writeq_relaxed(*(val + i), dest + i);
672 }
673 
674 static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
675 		       void *wqe)
676 {
677 #define HNS_ROCE_SL_SHIFT 2
678 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
679 
680 	/* All kinds of DirectWQE have the same header field layout */
681 	hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_FLAG);
682 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_L, qp->sl);
683 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_H,
684 		     qp->sl >> HNS_ROCE_SL_SHIFT);
685 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_WQE_INDEX, qp->sq.head);
686 
687 	hns_roce_write512(hr_dev, wqe, qp->sq.db_reg);
688 }
689 
690 static int hns_roce_v2_post_send(struct ib_qp *ibqp,
691 				 const struct ib_send_wr *wr,
692 				 const struct ib_send_wr **bad_wr)
693 {
694 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
695 	struct ib_device *ibdev = &hr_dev->ib_dev;
696 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
697 	unsigned long flags = 0;
698 	unsigned int owner_bit;
699 	unsigned int sge_idx;
700 	unsigned int wqe_idx;
701 	void *wqe = NULL;
702 	u32 nreq;
703 	int ret;
704 
705 	spin_lock_irqsave(&qp->sq.lock, flags);
706 
707 	ret = check_send_valid(hr_dev, qp);
708 	if (unlikely(ret)) {
709 		*bad_wr = wr;
710 		nreq = 0;
711 		goto out;
712 	}
713 
714 	sge_idx = qp->next_sge;
715 
716 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
717 		if (hns_roce_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
718 			ret = -ENOMEM;
719 			*bad_wr = wr;
720 			goto out;
721 		}
722 
723 		wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
724 
725 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
726 			ibdev_err(ibdev, "num_sge = %d > qp->sq.max_gs = %u.\n",
727 				  wr->num_sge, qp->sq.max_gs);
728 			ret = -EINVAL;
729 			*bad_wr = wr;
730 			goto out;
731 		}
732 
733 		wqe = hns_roce_get_send_wqe(qp, wqe_idx);
734 		qp->sq.wrid[wqe_idx] = wr->wr_id;
735 		owner_bit =
736 		       ~(((qp->sq.head + nreq) >> ilog2(qp->sq.wqe_cnt)) & 0x1);
737 
738 		/* Corresponding to the QP type, wqe process separately */
739 		if (ibqp->qp_type == IB_QPT_RC)
740 			ret = set_rc_wqe(qp, wr, wqe, &sge_idx, owner_bit);
741 		else
742 			ret = set_ud_wqe(qp, wr, wqe, &sge_idx, owner_bit);
743 
744 		if (unlikely(ret)) {
745 			*bad_wr = wr;
746 			goto out;
747 		}
748 	}
749 
750 out:
751 	if (likely(nreq)) {
752 		qp->sq.head += nreq;
753 		qp->next_sge = sge_idx;
754 
755 		if (nreq == 1 && !ret &&
756 		    (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
757 			write_dwqe(hr_dev, qp, wqe);
758 		else
759 			update_sq_db(hr_dev, qp);
760 	}
761 
762 	spin_unlock_irqrestore(&qp->sq.lock, flags);
763 
764 	return ret;
765 }
766 
767 static int check_recv_valid(struct hns_roce_dev *hr_dev,
768 			    struct hns_roce_qp *hr_qp)
769 {
770 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
771 		return -EIO;
772 
773 	if (hr_qp->state == IB_QPS_RESET)
774 		return -EINVAL;
775 
776 	return 0;
777 }
778 
779 static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
780 				 u32 max_sge, bool rsv)
781 {
782 	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
783 	u32 i, cnt;
784 
785 	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
786 		/* Skip zero-length sge */
787 		if (!wr->sg_list[i].length)
788 			continue;
789 		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
790 		cnt++;
791 	}
792 
793 	/* Fill a reserved sge to make hw stop reading remaining segments */
794 	if (rsv) {
795 		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
796 		dseg[cnt].addr = 0;
797 		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
798 	} else {
799 		/* Clear remaining segments to make ROCEE ignore sges */
800 		if (cnt < max_sge)
801 			memset(dseg + cnt, 0,
802 			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
803 	}
804 }
805 
806 static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
807 			u32 wqe_idx, u32 max_sge)
808 {
809 	void *wqe = NULL;
810 
811 	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
812 	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
813 }
814 
815 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
816 				 const struct ib_recv_wr *wr,
817 				 const struct ib_recv_wr **bad_wr)
818 {
819 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
820 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
821 	struct ib_device *ibdev = &hr_dev->ib_dev;
822 	u32 wqe_idx, nreq, max_sge;
823 	unsigned long flags;
824 	int ret;
825 
826 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
827 
828 	ret = check_recv_valid(hr_dev, hr_qp);
829 	if (unlikely(ret)) {
830 		*bad_wr = wr;
831 		nreq = 0;
832 		goto out;
833 	}
834 
835 	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
836 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
837 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
838 						  hr_qp->ibqp.recv_cq))) {
839 			ret = -ENOMEM;
840 			*bad_wr = wr;
841 			goto out;
842 		}
843 
844 		if (unlikely(wr->num_sge > max_sge)) {
845 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
846 				  wr->num_sge, max_sge);
847 			ret = -EINVAL;
848 			*bad_wr = wr;
849 			goto out;
850 		}
851 
852 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
853 		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
854 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
855 	}
856 
857 out:
858 	if (likely(nreq)) {
859 		hr_qp->rq.head += nreq;
860 
861 		update_rq_db(hr_dev, hr_qp);
862 	}
863 	spin_unlock_irqrestore(&hr_qp->rq.lock, flags);
864 
865 	return ret;
866 }
867 
868 static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
869 {
870 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
871 }
872 
873 static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
874 {
875 	return hns_roce_buf_offset(idx_que->mtr.kmem,
876 				   n << idx_que->entry_shift);
877 }
878 
879 static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
880 {
881 	/* always called with interrupts disabled. */
882 	spin_lock(&srq->lock);
883 
884 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
885 	srq->idx_que.tail++;
886 
887 	spin_unlock(&srq->lock);
888 }
889 
890 static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
891 {
892 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
893 
894 	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
895 }
896 
897 static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
898 				const struct ib_recv_wr *wr)
899 {
900 	struct ib_device *ib_dev = srq->ibsrq.device;
901 
902 	if (unlikely(wr->num_sge > max_sge)) {
903 		ibdev_err(ib_dev,
904 			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
905 			  wr->num_sge, max_sge);
906 		return -EINVAL;
907 	}
908 
909 	if (unlikely(hns_roce_srqwq_overflow(srq))) {
910 		ibdev_err(ib_dev,
911 			  "failed to check srqwq status, srqwq is full.\n");
912 		return -ENOMEM;
913 	}
914 
915 	return 0;
916 }
917 
918 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
919 {
920 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
921 	u32 pos;
922 
923 	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
924 	if (unlikely(pos == srq->wqe_cnt))
925 		return -ENOSPC;
926 
927 	bitmap_set(idx_que->bitmap, pos, 1);
928 	*wqe_idx = pos;
929 	return 0;
930 }
931 
932 static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
933 {
934 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
935 	unsigned int head;
936 	__le32 *buf;
937 
938 	head = idx_que->head & (srq->wqe_cnt - 1);
939 
940 	buf = get_idx_buf(idx_que, head);
941 	*buf = cpu_to_le32(wqe_idx);
942 
943 	idx_que->head++;
944 }
945 
946 static void update_srq_db(struct hns_roce_srq *srq)
947 {
948 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
949 	struct hns_roce_v2_db db;
950 
951 	hr_reg_write(&db, DB_TAG, srq->srqn);
952 	hr_reg_write(&db, DB_CMD, HNS_ROCE_V2_SRQ_DB);
953 	hr_reg_write(&db, DB_PI, srq->idx_que.head);
954 
955 	hns_roce_write64(hr_dev, (__le32 *)&db, srq->db_reg);
956 }
957 
958 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
959 				     const struct ib_recv_wr *wr,
960 				     const struct ib_recv_wr **bad_wr)
961 {
962 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
963 	unsigned long flags;
964 	int ret = 0;
965 	u32 max_sge;
966 	u32 wqe_idx;
967 	void *wqe;
968 	u32 nreq;
969 
970 	spin_lock_irqsave(&srq->lock, flags);
971 
972 	max_sge = srq->max_gs - srq->rsv_sge;
973 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
974 		ret = check_post_srq_valid(srq, max_sge, wr);
975 		if (ret) {
976 			*bad_wr = wr;
977 			break;
978 		}
979 
980 		ret = get_srq_wqe_idx(srq, &wqe_idx);
981 		if (unlikely(ret)) {
982 			*bad_wr = wr;
983 			break;
984 		}
985 
986 		wqe = get_srq_wqe_buf(srq, wqe_idx);
987 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
988 		fill_wqe_idx(srq, wqe_idx);
989 		srq->wrid[wqe_idx] = wr->wr_id;
990 	}
991 
992 	if (likely(nreq)) {
993 		if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB)
994 			*srq->rdb.db_record = srq->idx_que.head &
995 					      V2_DB_PRODUCER_IDX_M;
996 		else
997 			update_srq_db(srq);
998 	}
999 
1000 	spin_unlock_irqrestore(&srq->lock, flags);
1001 
1002 	return ret;
1003 }
1004 
1005 static u32 hns_roce_v2_cmd_hw_reseted(struct hns_roce_dev *hr_dev,
1006 				      unsigned long instance_stage,
1007 				      unsigned long reset_stage)
1008 {
1009 	/* When hardware reset has been completed once or more, we should stop
1010 	 * sending mailbox&cmq&doorbell to hardware. If now in .init_instance()
1011 	 * function, we should exit with error. If now at HNAE3_INIT_CLIENT
1012 	 * stage of soft reset process, we should exit with error, and then
1013 	 * HNAE3_INIT_CLIENT related process can rollback the operation like
1014 	 * notifing hardware to free resources, HNAE3_INIT_CLIENT related
1015 	 * process will exit with error to notify NIC driver to reschedule soft
1016 	 * reset process once again.
1017 	 */
1018 	hr_dev->is_reset = true;
1019 	hr_dev->dis_db = true;
1020 
1021 	if (reset_stage == HNS_ROCE_STATE_RST_INIT ||
1022 	    instance_stage == HNS_ROCE_STATE_INIT)
1023 		return CMD_RST_PRC_EBUSY;
1024 
1025 	return CMD_RST_PRC_SUCCESS;
1026 }
1027 
1028 static u32 hns_roce_v2_cmd_hw_resetting(struct hns_roce_dev *hr_dev,
1029 					unsigned long instance_stage,
1030 					unsigned long reset_stage)
1031 {
1032 #define HW_RESET_TIMEOUT_US 1000000
1033 #define HW_RESET_SLEEP_US 1000
1034 
1035 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1036 	struct hnae3_handle *handle = priv->handle;
1037 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1038 	unsigned long val;
1039 	int ret;
1040 
1041 	/* When hardware reset is detected, we should stop sending mailbox&cmq&
1042 	 * doorbell to hardware. If now in .init_instance() function, we should
1043 	 * exit with error. If now at HNAE3_INIT_CLIENT stage of soft reset
1044 	 * process, we should exit with error, and then HNAE3_INIT_CLIENT
1045 	 * related process can rollback the operation like notifing hardware to
1046 	 * free resources, HNAE3_INIT_CLIENT related process will exit with
1047 	 * error to notify NIC driver to reschedule soft reset process once
1048 	 * again.
1049 	 */
1050 	hr_dev->dis_db = true;
1051 
1052 	ret = read_poll_timeout(ops->ae_dev_reset_cnt, val,
1053 				val > hr_dev->reset_cnt, HW_RESET_SLEEP_US,
1054 				HW_RESET_TIMEOUT_US, false, handle);
1055 	if (!ret)
1056 		hr_dev->is_reset = true;
1057 
1058 	if (!hr_dev->is_reset || reset_stage == HNS_ROCE_STATE_RST_INIT ||
1059 	    instance_stage == HNS_ROCE_STATE_INIT)
1060 		return CMD_RST_PRC_EBUSY;
1061 
1062 	return CMD_RST_PRC_SUCCESS;
1063 }
1064 
1065 static u32 hns_roce_v2_cmd_sw_resetting(struct hns_roce_dev *hr_dev)
1066 {
1067 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1068 	struct hnae3_handle *handle = priv->handle;
1069 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1070 
1071 	/* When software reset is detected at .init_instance() function, we
1072 	 * should stop sending mailbox&cmq&doorbell to hardware, and exit
1073 	 * with error.
1074 	 */
1075 	hr_dev->dis_db = true;
1076 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt)
1077 		hr_dev->is_reset = true;
1078 
1079 	return CMD_RST_PRC_EBUSY;
1080 }
1081 
1082 static u32 check_aedev_reset_status(struct hns_roce_dev *hr_dev,
1083 				    struct hnae3_handle *handle)
1084 {
1085 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1086 	unsigned long instance_stage; /* the current instance stage */
1087 	unsigned long reset_stage; /* the current reset stage */
1088 	unsigned long reset_cnt;
1089 	bool sw_resetting;
1090 	bool hw_resetting;
1091 
1092 	/* Get information about reset from NIC driver or RoCE driver itself,
1093 	 * the meaning of the following variables from NIC driver are described
1094 	 * as below:
1095 	 * reset_cnt -- The count value of completed hardware reset.
1096 	 * hw_resetting -- Whether hardware device is resetting now.
1097 	 * sw_resetting -- Whether NIC's software reset process is running now.
1098 	 */
1099 	instance_stage = handle->rinfo.instance_state;
1100 	reset_stage = handle->rinfo.reset_state;
1101 	reset_cnt = ops->ae_dev_reset_cnt(handle);
1102 	if (reset_cnt != hr_dev->reset_cnt)
1103 		return hns_roce_v2_cmd_hw_reseted(hr_dev, instance_stage,
1104 						  reset_stage);
1105 
1106 	hw_resetting = ops->get_cmdq_stat(handle);
1107 	if (hw_resetting)
1108 		return hns_roce_v2_cmd_hw_resetting(hr_dev, instance_stage,
1109 						    reset_stage);
1110 
1111 	sw_resetting = ops->ae_dev_resetting(handle);
1112 	if (sw_resetting && instance_stage == HNS_ROCE_STATE_INIT)
1113 		return hns_roce_v2_cmd_sw_resetting(hr_dev);
1114 
1115 	return CMD_RST_PRC_OTHERS;
1116 }
1117 
1118 static bool check_device_is_in_reset(struct hns_roce_dev *hr_dev)
1119 {
1120 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1121 	struct hnae3_handle *handle = priv->handle;
1122 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1123 
1124 	if (hr_dev->reset_cnt != ops->ae_dev_reset_cnt(handle))
1125 		return true;
1126 
1127 	if (ops->get_hw_reset_stat(handle))
1128 		return true;
1129 
1130 	if (ops->ae_dev_resetting(handle))
1131 		return true;
1132 
1133 	return false;
1134 }
1135 
1136 static bool v2_chk_mbox_is_avail(struct hns_roce_dev *hr_dev, bool *busy)
1137 {
1138 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1139 	u32 status;
1140 
1141 	if (hr_dev->is_reset)
1142 		status = CMD_RST_PRC_SUCCESS;
1143 	else
1144 		status = check_aedev_reset_status(hr_dev, priv->handle);
1145 
1146 	*busy = (status == CMD_RST_PRC_EBUSY);
1147 
1148 	return status == CMD_RST_PRC_OTHERS;
1149 }
1150 
1151 static int hns_roce_alloc_cmq_desc(struct hns_roce_dev *hr_dev,
1152 				   struct hns_roce_v2_cmq_ring *ring)
1153 {
1154 	int size = ring->desc_num * sizeof(struct hns_roce_cmq_desc);
1155 
1156 	ring->desc = dma_alloc_coherent(hr_dev->dev, size,
1157 					&ring->desc_dma_addr, GFP_KERNEL);
1158 	if (!ring->desc)
1159 		return -ENOMEM;
1160 
1161 	return 0;
1162 }
1163 
1164 static void hns_roce_free_cmq_desc(struct hns_roce_dev *hr_dev,
1165 				   struct hns_roce_v2_cmq_ring *ring)
1166 {
1167 	dma_free_coherent(hr_dev->dev,
1168 			  ring->desc_num * sizeof(struct hns_roce_cmq_desc),
1169 			  ring->desc, ring->desc_dma_addr);
1170 
1171 	ring->desc_dma_addr = 0;
1172 }
1173 
1174 static int init_csq(struct hns_roce_dev *hr_dev,
1175 		    struct hns_roce_v2_cmq_ring *csq)
1176 {
1177 	dma_addr_t dma;
1178 	int ret;
1179 
1180 	csq->desc_num = CMD_CSQ_DESC_NUM;
1181 	spin_lock_init(&csq->lock);
1182 	csq->flag = TYPE_CSQ;
1183 	csq->head = 0;
1184 
1185 	ret = hns_roce_alloc_cmq_desc(hr_dev, csq);
1186 	if (ret)
1187 		return ret;
1188 
1189 	dma = csq->desc_dma_addr;
1190 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_L_REG, lower_32_bits(dma));
1191 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_H_REG, upper_32_bits(dma));
1192 	roce_write(hr_dev, ROCEE_TX_CMQ_DEPTH_REG,
1193 		   (u32)csq->desc_num >> HNS_ROCE_CMQ_DESC_NUM_S);
1194 
1195 	/* Make sure to write CI first and then PI */
1196 	roce_write(hr_dev, ROCEE_TX_CMQ_CI_REG, 0);
1197 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, 0);
1198 
1199 	return 0;
1200 }
1201 
1202 static int hns_roce_v2_cmq_init(struct hns_roce_dev *hr_dev)
1203 {
1204 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1205 	int ret;
1206 
1207 	priv->cmq.tx_timeout = HNS_ROCE_CMQ_TX_TIMEOUT;
1208 
1209 	ret = init_csq(hr_dev, &priv->cmq.csq);
1210 	if (ret)
1211 		dev_err(hr_dev->dev, "failed to init CSQ, ret = %d.\n", ret);
1212 
1213 	return ret;
1214 }
1215 
1216 static void hns_roce_v2_cmq_exit(struct hns_roce_dev *hr_dev)
1217 {
1218 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1219 
1220 	hns_roce_free_cmq_desc(hr_dev, &priv->cmq.csq);
1221 }
1222 
1223 static void hns_roce_cmq_setup_basic_desc(struct hns_roce_cmq_desc *desc,
1224 					  enum hns_roce_opcode_type opcode,
1225 					  bool is_read)
1226 {
1227 	memset((void *)desc, 0, sizeof(struct hns_roce_cmq_desc));
1228 	desc->opcode = cpu_to_le16(opcode);
1229 	desc->flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1230 	if (is_read)
1231 		desc->flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_WR);
1232 	else
1233 		desc->flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1234 }
1235 
1236 static int hns_roce_cmq_csq_done(struct hns_roce_dev *hr_dev)
1237 {
1238 	u32 tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1239 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1240 
1241 	return tail == priv->cmq.csq.head;
1242 }
1243 
1244 static void update_cmdq_status(struct hns_roce_dev *hr_dev)
1245 {
1246 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1247 	struct hnae3_handle *handle = priv->handle;
1248 
1249 	if (handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
1250 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT)
1251 		hr_dev->cmd.state = HNS_ROCE_CMDQ_STATE_FATAL_ERR;
1252 }
1253 
1254 static int hns_roce_cmd_err_convert_errno(u16 desc_ret)
1255 {
1256 	struct hns_roce_cmd_errcode errcode_table[] = {
1257 		{CMD_EXEC_SUCCESS, 0},
1258 		{CMD_NO_AUTH, -EPERM},
1259 		{CMD_NOT_EXIST, -EOPNOTSUPP},
1260 		{CMD_CRQ_FULL, -EXFULL},
1261 		{CMD_NEXT_ERR, -ENOSR},
1262 		{CMD_NOT_EXEC, -ENOTBLK},
1263 		{CMD_PARA_ERR, -EINVAL},
1264 		{CMD_RESULT_ERR, -ERANGE},
1265 		{CMD_TIMEOUT, -ETIME},
1266 		{CMD_HILINK_ERR, -ENOLINK},
1267 		{CMD_INFO_ILLEGAL, -ENXIO},
1268 		{CMD_INVALID, -EBADR},
1269 	};
1270 	u16 i;
1271 
1272 	for (i = 0; i < ARRAY_SIZE(errcode_table); i++)
1273 		if (desc_ret == errcode_table[i].return_status)
1274 			return errcode_table[i].errno;
1275 	return -EIO;
1276 }
1277 
1278 static u32 hns_roce_cmdq_tx_timeout(u16 opcode, u32 tx_timeout)
1279 {
1280 	static const struct hns_roce_cmdq_tx_timeout_map cmdq_tx_timeout[] = {
1281 		{HNS_ROCE_OPC_POST_MB, HNS_ROCE_OPC_POST_MB_TIMEOUT},
1282 	};
1283 	int i;
1284 
1285 	for (i = 0; i < ARRAY_SIZE(cmdq_tx_timeout); i++)
1286 		if (cmdq_tx_timeout[i].opcode == opcode)
1287 			return cmdq_tx_timeout[i].tx_timeout;
1288 
1289 	return tx_timeout;
1290 }
1291 
1292 static void hns_roce_wait_csq_done(struct hns_roce_dev *hr_dev, u16 opcode)
1293 {
1294 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1295 	u32 tx_timeout = hns_roce_cmdq_tx_timeout(opcode, priv->cmq.tx_timeout);
1296 	u32 timeout = 0;
1297 
1298 	do {
1299 		if (hns_roce_cmq_csq_done(hr_dev))
1300 			break;
1301 		udelay(1);
1302 	} while (++timeout < tx_timeout);
1303 }
1304 
1305 static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1306 			       struct hns_roce_cmq_desc *desc, int num)
1307 {
1308 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1309 	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
1310 	u16 desc_ret;
1311 	u32 tail;
1312 	int ret;
1313 	int i;
1314 
1315 	spin_lock_bh(&csq->lock);
1316 
1317 	tail = csq->head;
1318 
1319 	for (i = 0; i < num; i++) {
1320 		csq->desc[csq->head++] = desc[i];
1321 		if (csq->head == csq->desc_num)
1322 			csq->head = 0;
1323 	}
1324 
1325 	/* Write to hardware */
1326 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, csq->head);
1327 
1328 	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_CNT]);
1329 
1330 	hns_roce_wait_csq_done(hr_dev, le16_to_cpu(desc->opcode));
1331 	if (hns_roce_cmq_csq_done(hr_dev)) {
1332 		ret = 0;
1333 		for (i = 0; i < num; i++) {
1334 			/* check the result of hardware write back */
1335 			desc[i] = csq->desc[tail++];
1336 			if (tail == csq->desc_num)
1337 				tail = 0;
1338 
1339 			desc_ret = le16_to_cpu(desc[i].retval);
1340 			if (likely(desc_ret == CMD_EXEC_SUCCESS))
1341 				continue;
1342 
1343 			dev_err_ratelimited(hr_dev->dev,
1344 					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",
1345 					    desc->opcode, desc_ret);
1346 			ret = hns_roce_cmd_err_convert_errno(desc_ret);
1347 		}
1348 	} else {
1349 		/* FW/HW reset or incorrect number of desc */
1350 		tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1351 		dev_warn(hr_dev->dev, "CMDQ move tail from %u to %u.\n",
1352 			 csq->head, tail);
1353 		csq->head = tail;
1354 
1355 		update_cmdq_status(hr_dev);
1356 
1357 		ret = -EAGAIN;
1358 	}
1359 
1360 	spin_unlock_bh(&csq->lock);
1361 
1362 	if (ret)
1363 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_ERR_CNT]);
1364 
1365 	return ret;
1366 }
1367 
1368 static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1369 			     struct hns_roce_cmq_desc *desc, int num)
1370 {
1371 	bool busy;
1372 	int ret;
1373 
1374 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1375 		return -EIO;
1376 
1377 	if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1378 		return busy ? -EBUSY : 0;
1379 
1380 	ret = __hns_roce_cmq_send(hr_dev, desc, num);
1381 	if (ret) {
1382 		if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1383 			return busy ? -EBUSY : 0;
1384 	}
1385 
1386 	return ret;
1387 }
1388 
1389 static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev,
1390 			       dma_addr_t base_addr, u8 cmd, unsigned long tag)
1391 {
1392 	struct hns_roce_cmd_mailbox *mbox;
1393 	int ret;
1394 
1395 	mbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1396 	if (IS_ERR(mbox))
1397 		return PTR_ERR(mbox);
1398 
1399 	ret = hns_roce_cmd_mbox(hr_dev, base_addr, mbox->dma, cmd, tag);
1400 	hns_roce_free_cmd_mailbox(hr_dev, mbox);
1401 	return ret;
1402 }
1403 
1404 static int hns_roce_cmq_query_hw_info(struct hns_roce_dev *hr_dev)
1405 {
1406 	struct hns_roce_query_version *resp;
1407 	struct hns_roce_cmq_desc desc;
1408 	int ret;
1409 
1410 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_HW_VER, true);
1411 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1412 	if (ret)
1413 		return ret;
1414 
1415 	resp = (struct hns_roce_query_version *)desc.data;
1416 	hr_dev->hw_rev = le16_to_cpu(resp->rocee_hw_version);
1417 	hr_dev->vendor_id = hr_dev->pci_dev->vendor;
1418 
1419 	return 0;
1420 }
1421 
1422 static void func_clr_hw_resetting_state(struct hns_roce_dev *hr_dev,
1423 					struct hnae3_handle *handle)
1424 {
1425 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1426 	unsigned long end;
1427 
1428 	hr_dev->dis_db = true;
1429 
1430 	dev_warn(hr_dev->dev,
1431 		 "func clear is pending, device in resetting state.\n");
1432 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1433 	while (end) {
1434 		if (!ops->get_hw_reset_stat(handle)) {
1435 			hr_dev->is_reset = true;
1436 			dev_info(hr_dev->dev,
1437 				 "func clear success after reset.\n");
1438 			return;
1439 		}
1440 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1441 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1442 	}
1443 
1444 	dev_warn(hr_dev->dev, "func clear failed.\n");
1445 }
1446 
1447 static void func_clr_sw_resetting_state(struct hns_roce_dev *hr_dev,
1448 					struct hnae3_handle *handle)
1449 {
1450 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1451 	unsigned long end;
1452 
1453 	hr_dev->dis_db = true;
1454 
1455 	dev_warn(hr_dev->dev,
1456 		 "func clear is pending, device in resetting state.\n");
1457 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1458 	while (end) {
1459 		if (ops->ae_dev_reset_cnt(handle) !=
1460 		    hr_dev->reset_cnt) {
1461 			hr_dev->is_reset = true;
1462 			dev_info(hr_dev->dev,
1463 				 "func clear success after sw reset\n");
1464 			return;
1465 		}
1466 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1467 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1468 	}
1469 
1470 	dev_warn(hr_dev->dev, "func clear failed because of unfinished sw reset\n");
1471 }
1472 
1473 static void hns_roce_func_clr_rst_proc(struct hns_roce_dev *hr_dev, int retval,
1474 				       int flag)
1475 {
1476 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1477 	struct hnae3_handle *handle = priv->handle;
1478 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1479 
1480 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt) {
1481 		hr_dev->dis_db = true;
1482 		hr_dev->is_reset = true;
1483 		dev_info(hr_dev->dev, "func clear success after reset.\n");
1484 		return;
1485 	}
1486 
1487 	if (ops->get_hw_reset_stat(handle)) {
1488 		func_clr_hw_resetting_state(hr_dev, handle);
1489 		return;
1490 	}
1491 
1492 	if (ops->ae_dev_resetting(handle) &&
1493 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT) {
1494 		func_clr_sw_resetting_state(hr_dev, handle);
1495 		return;
1496 	}
1497 
1498 	if (retval && !flag)
1499 		dev_warn(hr_dev->dev,
1500 			 "func clear read failed, ret = %d.\n", retval);
1501 
1502 	dev_warn(hr_dev->dev, "func clear failed.\n");
1503 }
1504 
1505 static void __hns_roce_function_clear(struct hns_roce_dev *hr_dev, int vf_id)
1506 {
1507 	bool fclr_write_fail_flag = false;
1508 	struct hns_roce_func_clear *resp;
1509 	struct hns_roce_cmq_desc desc;
1510 	unsigned long end;
1511 	int ret = 0;
1512 
1513 	if (check_device_is_in_reset(hr_dev))
1514 		goto out;
1515 
1516 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR, false);
1517 	resp = (struct hns_roce_func_clear *)desc.data;
1518 	resp->rst_funcid_en = cpu_to_le32(vf_id);
1519 
1520 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1521 	if (ret) {
1522 		fclr_write_fail_flag = true;
1523 		dev_err(hr_dev->dev, "func clear write failed, ret = %d.\n",
1524 			 ret);
1525 		goto out;
1526 	}
1527 
1528 	msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_INTERVAL);
1529 	end = HNS_ROCE_V2_FUNC_CLEAR_TIMEOUT_MSECS;
1530 	while (end) {
1531 		if (check_device_is_in_reset(hr_dev))
1532 			goto out;
1533 		msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT);
1534 		end -= HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT;
1535 
1536 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR,
1537 					      true);
1538 
1539 		resp->rst_funcid_en = cpu_to_le32(vf_id);
1540 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1541 		if (ret)
1542 			continue;
1543 
1544 		if (hr_reg_read(resp, FUNC_CLEAR_RST_FUN_DONE)) {
1545 			if (vf_id == 0)
1546 				hr_dev->is_reset = true;
1547 			return;
1548 		}
1549 	}
1550 
1551 out:
1552 	hns_roce_func_clr_rst_proc(hr_dev, ret, fclr_write_fail_flag);
1553 }
1554 
1555 static int hns_roce_free_vf_resource(struct hns_roce_dev *hr_dev, int vf_id)
1556 {
1557 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1558 	struct hns_roce_cmq_desc desc[2];
1559 	struct hns_roce_cmq_req *req_a;
1560 
1561 	req_a = (struct hns_roce_cmq_req *)desc[0].data;
1562 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1563 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1564 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1565 	hr_reg_write(req_a, FUNC_RES_A_VF_ID, vf_id);
1566 
1567 	return hns_roce_cmq_send(hr_dev, desc, 2);
1568 }
1569 
1570 static void hns_roce_function_clear(struct hns_roce_dev *hr_dev)
1571 {
1572 	int ret;
1573 	int i;
1574 
1575 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1576 		return;
1577 
1578 	for (i = hr_dev->func_num - 1; i >= 0; i--) {
1579 		__hns_roce_function_clear(hr_dev, i);
1580 
1581 		if (i == 0)
1582 			continue;
1583 
1584 		ret = hns_roce_free_vf_resource(hr_dev, i);
1585 		if (ret)
1586 			ibdev_err(&hr_dev->ib_dev,
1587 				  "failed to free vf resource, vf_id = %d, ret = %d.\n",
1588 				  i, ret);
1589 	}
1590 }
1591 
1592 static int hns_roce_clear_extdb_list_info(struct hns_roce_dev *hr_dev)
1593 {
1594 	struct hns_roce_cmq_desc desc;
1595 	int ret;
1596 
1597 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLEAR_EXTDB_LIST_INFO,
1598 				      false);
1599 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1600 	if (ret)
1601 		ibdev_err(&hr_dev->ib_dev,
1602 			  "failed to clear extended doorbell info, ret = %d.\n",
1603 			  ret);
1604 
1605 	return ret;
1606 }
1607 
1608 static int hns_roce_query_fw_ver(struct hns_roce_dev *hr_dev)
1609 {
1610 	struct hns_roce_query_fw_info *resp;
1611 	struct hns_roce_cmq_desc desc;
1612 	int ret;
1613 
1614 	hns_roce_cmq_setup_basic_desc(&desc, HNS_QUERY_FW_VER, true);
1615 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1616 	if (ret)
1617 		return ret;
1618 
1619 	resp = (struct hns_roce_query_fw_info *)desc.data;
1620 	hr_dev->caps.fw_ver = (u64)(le32_to_cpu(resp->fw_ver));
1621 
1622 	return 0;
1623 }
1624 
1625 static int hns_roce_query_func_info(struct hns_roce_dev *hr_dev)
1626 {
1627 	struct hns_roce_cmq_desc desc;
1628 	int ret;
1629 
1630 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
1631 		hr_dev->func_num = 1;
1632 		return 0;
1633 	}
1634 
1635 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_FUNC_INFO,
1636 				      true);
1637 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1638 	if (ret) {
1639 		hr_dev->func_num = 1;
1640 		return ret;
1641 	}
1642 
1643 	hr_dev->func_num = le32_to_cpu(desc.func_info.own_func_num);
1644 	hr_dev->cong_algo_tmpl_id = le32_to_cpu(desc.func_info.own_mac_id);
1645 
1646 	return 0;
1647 }
1648 
1649 static int hns_roce_hw_v2_query_counter(struct hns_roce_dev *hr_dev,
1650 					u64 *stats, u32 port, int *num_counters)
1651 {
1652 #define CNT_PER_DESC 3
1653 	struct hns_roce_cmq_desc *desc;
1654 	int bd_idx, cnt_idx;
1655 	__le64 *cnt_data;
1656 	int desc_num;
1657 	int ret;
1658 	int i;
1659 
1660 	if (port > hr_dev->caps.num_ports)
1661 		return -EINVAL;
1662 
1663 	desc_num = DIV_ROUND_UP(HNS_ROCE_HW_CNT_TOTAL, CNT_PER_DESC);
1664 	desc = kcalloc(desc_num, sizeof(*desc), GFP_KERNEL);
1665 	if (!desc)
1666 		return -ENOMEM;
1667 
1668 	for (i = 0; i < desc_num; i++) {
1669 		hns_roce_cmq_setup_basic_desc(&desc[i],
1670 					      HNS_ROCE_OPC_QUERY_COUNTER, true);
1671 		if (i != desc_num - 1)
1672 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1673 	}
1674 
1675 	ret = hns_roce_cmq_send(hr_dev, desc, desc_num);
1676 	if (ret) {
1677 		ibdev_err(&hr_dev->ib_dev,
1678 			  "failed to get counter, ret = %d.\n", ret);
1679 		goto err_out;
1680 	}
1681 
1682 	for (i = 0; i < HNS_ROCE_HW_CNT_TOTAL && i < *num_counters; i++) {
1683 		bd_idx = i / CNT_PER_DESC;
1684 		if (bd_idx != HNS_ROCE_HW_CNT_TOTAL / CNT_PER_DESC &&
1685 		    !(desc[bd_idx].flag & cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT)))
1686 			break;
1687 
1688 		cnt_data = (__le64 *)&desc[bd_idx].data[0];
1689 		cnt_idx = i % CNT_PER_DESC;
1690 		stats[i] = le64_to_cpu(cnt_data[cnt_idx]);
1691 	}
1692 	*num_counters = i;
1693 
1694 err_out:
1695 	kfree(desc);
1696 	return ret;
1697 }
1698 
1699 static int hns_roce_config_global_param(struct hns_roce_dev *hr_dev)
1700 {
1701 	struct hns_roce_cmq_desc desc;
1702 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1703 	u32 clock_cycles_of_1us;
1704 
1705 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GLOBAL_PARAM,
1706 				      false);
1707 
1708 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
1709 		clock_cycles_of_1us = HNS_ROCE_1NS_CFG;
1710 	else
1711 		clock_cycles_of_1us = HNS_ROCE_1US_CFG;
1712 
1713 	hr_reg_write(req, CFG_GLOBAL_PARAM_1US_CYCLES, clock_cycles_of_1us);
1714 	hr_reg_write(req, CFG_GLOBAL_PARAM_UDP_PORT, ROCE_V2_UDP_DPORT);
1715 
1716 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1717 }
1718 
1719 static int load_func_res_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1720 {
1721 	struct hns_roce_cmq_desc desc[2];
1722 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1723 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1724 	struct hns_roce_caps *caps = &hr_dev->caps;
1725 	enum hns_roce_opcode_type opcode;
1726 	u32 func_num;
1727 	int ret;
1728 
1729 	if (is_vf) {
1730 		opcode = HNS_ROCE_OPC_QUERY_VF_RES;
1731 		func_num = 1;
1732 	} else {
1733 		opcode = HNS_ROCE_OPC_QUERY_PF_RES;
1734 		func_num = hr_dev->func_num;
1735 	}
1736 
1737 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, true);
1738 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1739 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, true);
1740 
1741 	ret = hns_roce_cmq_send(hr_dev, desc, 2);
1742 	if (ret)
1743 		return ret;
1744 
1745 	caps->qpc_bt_num = hr_reg_read(r_a, FUNC_RES_A_QPC_BT_NUM) / func_num;
1746 	caps->srqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_SRQC_BT_NUM) / func_num;
1747 	caps->cqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_CQC_BT_NUM) / func_num;
1748 	caps->mpt_bt_num = hr_reg_read(r_a, FUNC_RES_A_MPT_BT_NUM) / func_num;
1749 	caps->eqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_EQC_BT_NUM) / func_num;
1750 	caps->smac_bt_num = hr_reg_read(r_b, FUNC_RES_B_SMAC_NUM) / func_num;
1751 	caps->sgid_bt_num = hr_reg_read(r_b, FUNC_RES_B_SGID_NUM) / func_num;
1752 	caps->sccc_bt_num = hr_reg_read(r_b, FUNC_RES_B_SCCC_BT_NUM) / func_num;
1753 
1754 	if (is_vf) {
1755 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_V_QID_NUM) / func_num;
1756 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_V_GMV_BT_NUM) /
1757 					       func_num;
1758 	} else {
1759 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_B_QID_NUM) / func_num;
1760 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_B_GMV_BT_NUM) /
1761 					       func_num;
1762 	}
1763 
1764 	return 0;
1765 }
1766 
1767 static int load_pf_timer_res_caps(struct hns_roce_dev *hr_dev)
1768 {
1769 	struct hns_roce_cmq_desc desc;
1770 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1771 	struct hns_roce_caps *caps = &hr_dev->caps;
1772 	int ret;
1773 
1774 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
1775 				      true);
1776 
1777 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1778 	if (ret)
1779 		return ret;
1780 
1781 	caps->qpc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_QPC_ITEM_NUM);
1782 	caps->cqc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_CQC_ITEM_NUM);
1783 
1784 	return 0;
1785 }
1786 
1787 static int hns_roce_query_pf_resource(struct hns_roce_dev *hr_dev)
1788 {
1789 	struct device *dev = hr_dev->dev;
1790 	int ret;
1791 
1792 	ret = load_func_res_caps(hr_dev, false);
1793 	if (ret) {
1794 		dev_err(dev, "failed to load pf res caps, ret = %d.\n", ret);
1795 		return ret;
1796 	}
1797 
1798 	ret = load_pf_timer_res_caps(hr_dev);
1799 	if (ret)
1800 		dev_err(dev, "failed to load pf timer resource, ret = %d.\n",
1801 			ret);
1802 
1803 	return ret;
1804 }
1805 
1806 static int hns_roce_query_vf_resource(struct hns_roce_dev *hr_dev)
1807 {
1808 	struct device *dev = hr_dev->dev;
1809 	int ret;
1810 
1811 	ret = load_func_res_caps(hr_dev, true);
1812 	if (ret)
1813 		dev_err(dev, "failed to load vf res caps, ret = %d.\n", ret);
1814 
1815 	return ret;
1816 }
1817 
1818 static int __hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
1819 					  u32 vf_id)
1820 {
1821 	struct hns_roce_vf_switch *swt;
1822 	struct hns_roce_cmq_desc desc;
1823 	int ret;
1824 
1825 	swt = (struct hns_roce_vf_switch *)desc.data;
1826 	hns_roce_cmq_setup_basic_desc(&desc, HNS_SWITCH_PARAMETER_CFG, true);
1827 	swt->rocee_sel |= cpu_to_le32(HNS_ICL_SWITCH_CMD_ROCEE_SEL);
1828 	hr_reg_write(swt, VF_SWITCH_VF_ID, vf_id);
1829 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1830 	if (ret)
1831 		return ret;
1832 
1833 	desc.flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1834 	desc.flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1835 	hr_reg_enable(swt, VF_SWITCH_ALW_LPBK);
1836 	hr_reg_clear(swt, VF_SWITCH_ALW_LCL_LPBK);
1837 	hr_reg_enable(swt, VF_SWITCH_ALW_DST_OVRD);
1838 
1839 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1840 }
1841 
1842 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev)
1843 {
1844 	u32 vf_id;
1845 	int ret;
1846 
1847 	for (vf_id = 0; vf_id < hr_dev->func_num; vf_id++) {
1848 		ret = __hns_roce_set_vf_switch_param(hr_dev, vf_id);
1849 		if (ret)
1850 			return ret;
1851 	}
1852 	return 0;
1853 }
1854 
1855 static int config_vf_hem_resource(struct hns_roce_dev *hr_dev, int vf_id)
1856 {
1857 	struct hns_roce_cmq_desc desc[2];
1858 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1859 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1860 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1861 	struct hns_roce_caps *caps = &hr_dev->caps;
1862 
1863 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1864 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1865 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1866 
1867 	hr_reg_write(r_a, FUNC_RES_A_VF_ID, vf_id);
1868 
1869 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_NUM, caps->qpc_bt_num);
1870 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_IDX, vf_id * caps->qpc_bt_num);
1871 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_NUM, caps->srqc_bt_num);
1872 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_IDX, vf_id * caps->srqc_bt_num);
1873 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_NUM, caps->cqc_bt_num);
1874 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_IDX, vf_id * caps->cqc_bt_num);
1875 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_NUM, caps->mpt_bt_num);
1876 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_IDX, vf_id * caps->mpt_bt_num);
1877 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_NUM, caps->eqc_bt_num);
1878 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_IDX, vf_id * caps->eqc_bt_num);
1879 	hr_reg_write(r_b, FUNC_RES_V_QID_NUM, caps->sl_num);
1880 	hr_reg_write(r_b, FUNC_RES_B_QID_IDX, vf_id * caps->sl_num);
1881 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_NUM, caps->sccc_bt_num);
1882 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_IDX, vf_id * caps->sccc_bt_num);
1883 
1884 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1885 		hr_reg_write(r_b, FUNC_RES_V_GMV_BT_NUM, caps->gmv_bt_num);
1886 		hr_reg_write(r_b, FUNC_RES_B_GMV_BT_IDX,
1887 			     vf_id * caps->gmv_bt_num);
1888 	} else {
1889 		hr_reg_write(r_b, FUNC_RES_B_SGID_NUM, caps->sgid_bt_num);
1890 		hr_reg_write(r_b, FUNC_RES_B_SGID_IDX,
1891 			     vf_id * caps->sgid_bt_num);
1892 		hr_reg_write(r_b, FUNC_RES_B_SMAC_NUM, caps->smac_bt_num);
1893 		hr_reg_write(r_b, FUNC_RES_B_SMAC_IDX,
1894 			     vf_id * caps->smac_bt_num);
1895 	}
1896 
1897 	return hns_roce_cmq_send(hr_dev, desc, 2);
1898 }
1899 
1900 static int hns_roce_alloc_vf_resource(struct hns_roce_dev *hr_dev)
1901 {
1902 	u32 func_num = max_t(u32, 1, hr_dev->func_num);
1903 	u32 vf_id;
1904 	int ret;
1905 
1906 	for (vf_id = 0; vf_id < func_num; vf_id++) {
1907 		ret = config_vf_hem_resource(hr_dev, vf_id);
1908 		if (ret) {
1909 			dev_err(hr_dev->dev,
1910 				"failed to config vf-%u hem res, ret = %d.\n",
1911 				vf_id, ret);
1912 			return ret;
1913 		}
1914 	}
1915 
1916 	return 0;
1917 }
1918 
1919 static int hns_roce_v2_set_bt(struct hns_roce_dev *hr_dev)
1920 {
1921 	struct hns_roce_cmq_desc desc;
1922 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1923 	struct hns_roce_caps *caps = &hr_dev->caps;
1924 
1925 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_BT_ATTR, false);
1926 
1927 	hr_reg_write(req, CFG_BT_ATTR_QPC_BA_PGSZ,
1928 		     caps->qpc_ba_pg_sz + PG_SHIFT_OFFSET);
1929 	hr_reg_write(req, CFG_BT_ATTR_QPC_BUF_PGSZ,
1930 		     caps->qpc_buf_pg_sz + PG_SHIFT_OFFSET);
1931 	hr_reg_write(req, CFG_BT_ATTR_QPC_HOPNUM,
1932 		     to_hr_hem_hopnum(caps->qpc_hop_num, caps->num_qps));
1933 
1934 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BA_PGSZ,
1935 		     caps->srqc_ba_pg_sz + PG_SHIFT_OFFSET);
1936 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BUF_PGSZ,
1937 		     caps->srqc_buf_pg_sz + PG_SHIFT_OFFSET);
1938 	hr_reg_write(req, CFG_BT_ATTR_SRQC_HOPNUM,
1939 		     to_hr_hem_hopnum(caps->srqc_hop_num, caps->num_srqs));
1940 
1941 	hr_reg_write(req, CFG_BT_ATTR_CQC_BA_PGSZ,
1942 		     caps->cqc_ba_pg_sz + PG_SHIFT_OFFSET);
1943 	hr_reg_write(req, CFG_BT_ATTR_CQC_BUF_PGSZ,
1944 		     caps->cqc_buf_pg_sz + PG_SHIFT_OFFSET);
1945 	hr_reg_write(req, CFG_BT_ATTR_CQC_HOPNUM,
1946 		     to_hr_hem_hopnum(caps->cqc_hop_num, caps->num_cqs));
1947 
1948 	hr_reg_write(req, CFG_BT_ATTR_MPT_BA_PGSZ,
1949 		     caps->mpt_ba_pg_sz + PG_SHIFT_OFFSET);
1950 	hr_reg_write(req, CFG_BT_ATTR_MPT_BUF_PGSZ,
1951 		     caps->mpt_buf_pg_sz + PG_SHIFT_OFFSET);
1952 	hr_reg_write(req, CFG_BT_ATTR_MPT_HOPNUM,
1953 		     to_hr_hem_hopnum(caps->mpt_hop_num, caps->num_mtpts));
1954 
1955 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BA_PGSZ,
1956 		     caps->sccc_ba_pg_sz + PG_SHIFT_OFFSET);
1957 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BUF_PGSZ,
1958 		     caps->sccc_buf_pg_sz + PG_SHIFT_OFFSET);
1959 	hr_reg_write(req, CFG_BT_ATTR_SCCC_HOPNUM,
1960 		     to_hr_hem_hopnum(caps->sccc_hop_num, caps->num_qps));
1961 
1962 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1963 }
1964 
1965 static void calc_pg_sz(u32 obj_num, u32 obj_size, u32 hop_num, u32 ctx_bt_num,
1966 		       u32 *buf_page_size, u32 *bt_page_size, u32 hem_type)
1967 {
1968 	u64 obj_per_chunk;
1969 	u64 bt_chunk_size = PAGE_SIZE;
1970 	u64 buf_chunk_size = PAGE_SIZE;
1971 	u64 obj_per_chunk_default = buf_chunk_size / obj_size;
1972 
1973 	*buf_page_size = 0;
1974 	*bt_page_size = 0;
1975 
1976 	switch (hop_num) {
1977 	case 3:
1978 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1979 				(bt_chunk_size / BA_BYTE_LEN) *
1980 				(bt_chunk_size / BA_BYTE_LEN) *
1981 				 obj_per_chunk_default;
1982 		break;
1983 	case 2:
1984 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1985 				(bt_chunk_size / BA_BYTE_LEN) *
1986 				 obj_per_chunk_default;
1987 		break;
1988 	case 1:
1989 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1990 				obj_per_chunk_default;
1991 		break;
1992 	case HNS_ROCE_HOP_NUM_0:
1993 		obj_per_chunk = ctx_bt_num * obj_per_chunk_default;
1994 		break;
1995 	default:
1996 		pr_err("table %u not support hop_num = %u!\n", hem_type,
1997 		       hop_num);
1998 		return;
1999 	}
2000 
2001 	if (hem_type >= HEM_TYPE_MTT)
2002 		*bt_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2003 	else
2004 		*buf_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2005 }
2006 
2007 static void set_hem_page_size(struct hns_roce_dev *hr_dev)
2008 {
2009 	struct hns_roce_caps *caps = &hr_dev->caps;
2010 
2011 	/* EQ */
2012 	caps->eqe_ba_pg_sz = 0;
2013 	caps->eqe_buf_pg_sz = 0;
2014 
2015 	/* Link Table */
2016 	caps->llm_buf_pg_sz = 0;
2017 
2018 	/* MR */
2019 	caps->mpt_ba_pg_sz = 0;
2020 	caps->mpt_buf_pg_sz = 0;
2021 	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
2022 	caps->pbl_buf_pg_sz = 0;
2023 	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
2024 		   caps->mpt_bt_num, &caps->mpt_buf_pg_sz, &caps->mpt_ba_pg_sz,
2025 		   HEM_TYPE_MTPT);
2026 
2027 	/* QP */
2028 	caps->qpc_ba_pg_sz = 0;
2029 	caps->qpc_buf_pg_sz = 0;
2030 	caps->qpc_timer_ba_pg_sz = 0;
2031 	caps->qpc_timer_buf_pg_sz = 0;
2032 	caps->sccc_ba_pg_sz = 0;
2033 	caps->sccc_buf_pg_sz = 0;
2034 	caps->mtt_ba_pg_sz = 0;
2035 	caps->mtt_buf_pg_sz = 0;
2036 	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
2037 		   caps->qpc_bt_num, &caps->qpc_buf_pg_sz, &caps->qpc_ba_pg_sz,
2038 		   HEM_TYPE_QPC);
2039 
2040 	if (caps->flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
2041 		calc_pg_sz(caps->num_qps, caps->sccc_sz, caps->sccc_hop_num,
2042 			   caps->sccc_bt_num, &caps->sccc_buf_pg_sz,
2043 			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
2044 
2045 	/* CQ */
2046 	caps->cqc_ba_pg_sz = 0;
2047 	caps->cqc_buf_pg_sz = 0;
2048 	caps->cqc_timer_ba_pg_sz = 0;
2049 	caps->cqc_timer_buf_pg_sz = 0;
2050 	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
2051 	caps->cqe_buf_pg_sz = 0;
2052 	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
2053 		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
2054 		   HEM_TYPE_CQC);
2055 	calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
2056 		   1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
2057 
2058 	/* SRQ */
2059 	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
2060 		caps->srqc_ba_pg_sz = 0;
2061 		caps->srqc_buf_pg_sz = 0;
2062 		caps->srqwqe_ba_pg_sz = 0;
2063 		caps->srqwqe_buf_pg_sz = 0;
2064 		caps->idx_ba_pg_sz = 0;
2065 		caps->idx_buf_pg_sz = 0;
2066 		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
2067 			   caps->srqc_hop_num, caps->srqc_bt_num,
2068 			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
2069 			   HEM_TYPE_SRQC);
2070 		calc_pg_sz(caps->num_srqwqe_segs, caps->mtt_entry_sz,
2071 			   caps->srqwqe_hop_num, 1, &caps->srqwqe_buf_pg_sz,
2072 			   &caps->srqwqe_ba_pg_sz, HEM_TYPE_SRQWQE);
2073 		calc_pg_sz(caps->num_idx_segs, caps->idx_entry_sz,
2074 			   caps->idx_hop_num, 1, &caps->idx_buf_pg_sz,
2075 			   &caps->idx_ba_pg_sz, HEM_TYPE_IDX);
2076 	}
2077 
2078 	/* GMV */
2079 	caps->gmv_ba_pg_sz = 0;
2080 	caps->gmv_buf_pg_sz = 0;
2081 }
2082 
2083 /* Apply all loaded caps before setting to hardware */
2084 static void apply_func_caps(struct hns_roce_dev *hr_dev)
2085 {
2086 #define MAX_GID_TBL_LEN 256
2087 	struct hns_roce_caps *caps = &hr_dev->caps;
2088 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2089 
2090 	/* The following configurations don't need to be got from firmware. */
2091 	caps->qpc_timer_entry_sz = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
2092 	caps->cqc_timer_entry_sz = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
2093 	caps->mtt_entry_sz = HNS_ROCE_V2_MTT_ENTRY_SZ;
2094 
2095 	caps->pbl_hop_num = HNS_ROCE_PBL_HOP_NUM;
2096 	caps->qpc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2097 	caps->cqc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2098 
2099 	caps->num_srqwqe_segs = HNS_ROCE_V2_MAX_SRQWQE_SEGS;
2100 	caps->num_idx_segs = HNS_ROCE_V2_MAX_IDX_SEGS;
2101 
2102 	if (!caps->num_comp_vectors)
2103 		caps->num_comp_vectors =
2104 			min_t(u32, caps->eqc_bt_num - HNS_ROCE_V2_AEQE_VEC_NUM,
2105 				(u32)priv->handle->rinfo.num_vectors -
2106 		(HNS_ROCE_V2_AEQE_VEC_NUM + HNS_ROCE_V2_ABNORMAL_VEC_NUM));
2107 
2108 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2109 		caps->eqe_hop_num = HNS_ROCE_V3_EQE_HOP_NUM;
2110 		caps->ceqe_size = HNS_ROCE_V3_EQE_SIZE;
2111 		caps->aeqe_size = HNS_ROCE_V3_EQE_SIZE;
2112 
2113 		/* The following configurations will be overwritten */
2114 		caps->qpc_sz = HNS_ROCE_V3_QPC_SZ;
2115 		caps->cqe_sz = HNS_ROCE_V3_CQE_SIZE;
2116 		caps->sccc_sz = HNS_ROCE_V3_SCCC_SZ;
2117 
2118 		/* The following configurations are not got from firmware */
2119 		caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;
2120 
2121 		caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0;
2122 
2123 		/* It's meaningless to support excessively large gid_table_len,
2124 		 * as the type of sgid_index in kernel struct ib_global_route
2125 		 * and userspace struct ibv_global_route are u8/uint8_t (0-255).
2126 		 */
2127 		caps->gid_table_len[0] = min_t(u32, MAX_GID_TBL_LEN,
2128 					 caps->gmv_bt_num *
2129 					 (HNS_HW_PAGE_SIZE / caps->gmv_entry_sz));
2130 
2131 		caps->gmv_entry_num = caps->gmv_bt_num * (HNS_HW_PAGE_SIZE /
2132 							  caps->gmv_entry_sz);
2133 	} else {
2134 		u32 func_num = max_t(u32, 1, hr_dev->func_num);
2135 
2136 		caps->eqe_hop_num = HNS_ROCE_V2_EQE_HOP_NUM;
2137 		caps->ceqe_size = HNS_ROCE_CEQE_SIZE;
2138 		caps->aeqe_size = HNS_ROCE_AEQE_SIZE;
2139 		caps->gid_table_len[0] /= func_num;
2140 	}
2141 
2142 	if (hr_dev->is_vf) {
2143 		caps->default_aeq_arm_st = 0x3;
2144 		caps->default_ceq_arm_st = 0x3;
2145 		caps->default_ceq_max_cnt = 0x1;
2146 		caps->default_ceq_period = 0x10;
2147 		caps->default_aeq_max_cnt = 0x1;
2148 		caps->default_aeq_period = 0x10;
2149 	}
2150 
2151 	set_hem_page_size(hr_dev);
2152 }
2153 
2154 static int hns_roce_query_caps(struct hns_roce_dev *hr_dev)
2155 {
2156 	struct hns_roce_cmq_desc desc[HNS_ROCE_QUERY_PF_CAPS_CMD_NUM];
2157 	struct hns_roce_caps *caps = &hr_dev->caps;
2158 	struct hns_roce_query_pf_caps_a *resp_a;
2159 	struct hns_roce_query_pf_caps_b *resp_b;
2160 	struct hns_roce_query_pf_caps_c *resp_c;
2161 	struct hns_roce_query_pf_caps_d *resp_d;
2162 	struct hns_roce_query_pf_caps_e *resp_e;
2163 	enum hns_roce_opcode_type cmd;
2164 	int ctx_hop_num;
2165 	int pbl_hop_num;
2166 	int ret;
2167 	int i;
2168 
2169 	cmd = hr_dev->is_vf ? HNS_ROCE_OPC_QUERY_VF_CAPS_NUM :
2170 	      HNS_ROCE_OPC_QUERY_PF_CAPS_NUM;
2171 
2172 	for (i = 0; i < HNS_ROCE_QUERY_PF_CAPS_CMD_NUM; i++) {
2173 		hns_roce_cmq_setup_basic_desc(&desc[i], cmd, true);
2174 		if (i < (HNS_ROCE_QUERY_PF_CAPS_CMD_NUM - 1))
2175 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2176 		else
2177 			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2178 	}
2179 
2180 	ret = hns_roce_cmq_send(hr_dev, desc, HNS_ROCE_QUERY_PF_CAPS_CMD_NUM);
2181 	if (ret)
2182 		return ret;
2183 
2184 	resp_a = (struct hns_roce_query_pf_caps_a *)desc[0].data;
2185 	resp_b = (struct hns_roce_query_pf_caps_b *)desc[1].data;
2186 	resp_c = (struct hns_roce_query_pf_caps_c *)desc[2].data;
2187 	resp_d = (struct hns_roce_query_pf_caps_d *)desc[3].data;
2188 	resp_e = (struct hns_roce_query_pf_caps_e *)desc[4].data;
2189 
2190 	caps->local_ca_ack_delay = resp_a->local_ca_ack_delay;
2191 	caps->max_sq_sg = le16_to_cpu(resp_a->max_sq_sg);
2192 	caps->max_sq_inline = le16_to_cpu(resp_a->max_sq_inline);
2193 	caps->max_rq_sg = le16_to_cpu(resp_a->max_rq_sg);
2194 	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
2195 	caps->max_srq_sges = le16_to_cpu(resp_a->max_srq_sges);
2196 	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
2197 	caps->num_aeq_vectors = resp_a->num_aeq_vectors;
2198 	caps->num_other_vectors = resp_a->num_other_vectors;
2199 	caps->max_sq_desc_sz = resp_a->max_sq_desc_sz;
2200 	caps->max_rq_desc_sz = resp_a->max_rq_desc_sz;
2201 
2202 	caps->mtpt_entry_sz = resp_b->mtpt_entry_sz;
2203 	caps->irrl_entry_sz = resp_b->irrl_entry_sz;
2204 	caps->trrl_entry_sz = resp_b->trrl_entry_sz;
2205 	caps->cqc_entry_sz = resp_b->cqc_entry_sz;
2206 	caps->srqc_entry_sz = resp_b->srqc_entry_sz;
2207 	caps->idx_entry_sz = resp_b->idx_entry_sz;
2208 	caps->sccc_sz = resp_b->sccc_sz;
2209 	caps->max_mtu = resp_b->max_mtu;
2210 	caps->min_cqes = resp_b->min_cqes;
2211 	caps->min_wqes = resp_b->min_wqes;
2212 	caps->page_size_cap = le32_to_cpu(resp_b->page_size_cap);
2213 	caps->pkey_table_len[0] = resp_b->pkey_table_len;
2214 	caps->phy_num_uars = resp_b->phy_num_uars;
2215 	ctx_hop_num = resp_b->ctx_hop_num;
2216 	pbl_hop_num = resp_b->pbl_hop_num;
2217 
2218 	caps->num_pds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_PDS);
2219 
2220 	caps->flags = hr_reg_read(resp_c, PF_CAPS_C_CAP_FLAGS);
2221 	caps->flags |= le16_to_cpu(resp_d->cap_flags_ex) <<
2222 		       HNS_ROCE_CAP_FLAGS_EX_SHIFT;
2223 
2224 	caps->num_cqs = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_CQS);
2225 	caps->gid_table_len[0] = hr_reg_read(resp_c, PF_CAPS_C_MAX_GID);
2226 	caps->max_cqes = 1 << hr_reg_read(resp_c, PF_CAPS_C_CQ_DEPTH);
2227 	caps->num_xrcds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_XRCDS);
2228 	caps->num_mtpts = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_MRWS);
2229 	caps->num_qps = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_QPS);
2230 	caps->max_qp_init_rdma = hr_reg_read(resp_c, PF_CAPS_C_MAX_ORD);
2231 	caps->max_qp_dest_rdma = caps->max_qp_init_rdma;
2232 	caps->max_wqes = 1 << le16_to_cpu(resp_c->sq_depth);
2233 
2234 	caps->num_srqs = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_SRQS);
2235 	caps->cong_cap = hr_reg_read(resp_d, PF_CAPS_D_CONG_CAP);
2236 	caps->max_srq_wrs = 1 << le16_to_cpu(resp_d->srq_depth);
2237 	caps->ceqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_CEQ_DEPTH);
2238 	caps->num_comp_vectors = hr_reg_read(resp_d, PF_CAPS_D_NUM_CEQS);
2239 	caps->aeqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_AEQ_DEPTH);
2240 	caps->default_cong_type = hr_reg_read(resp_d, PF_CAPS_D_DEFAULT_ALG);
2241 	caps->reserved_pds = hr_reg_read(resp_d, PF_CAPS_D_RSV_PDS);
2242 	caps->num_uars = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_UARS);
2243 	caps->reserved_qps = hr_reg_read(resp_d, PF_CAPS_D_RSV_QPS);
2244 	caps->reserved_uars = hr_reg_read(resp_d, PF_CAPS_D_RSV_UARS);
2245 
2246 	caps->reserved_mrws = hr_reg_read(resp_e, PF_CAPS_E_RSV_MRWS);
2247 	caps->chunk_sz = 1 << hr_reg_read(resp_e, PF_CAPS_E_CHUNK_SIZE_SHIFT);
2248 	caps->reserved_cqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_CQS);
2249 	caps->reserved_xrcds = hr_reg_read(resp_e, PF_CAPS_E_RSV_XRCDS);
2250 	caps->reserved_srqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_SRQS);
2251 	caps->reserved_lkey = hr_reg_read(resp_e, PF_CAPS_E_RSV_LKEYS);
2252 
2253 	caps->qpc_hop_num = ctx_hop_num;
2254 	caps->sccc_hop_num = ctx_hop_num;
2255 	caps->srqc_hop_num = ctx_hop_num;
2256 	caps->cqc_hop_num = ctx_hop_num;
2257 	caps->mpt_hop_num = ctx_hop_num;
2258 	caps->mtt_hop_num = pbl_hop_num;
2259 	caps->cqe_hop_num = pbl_hop_num;
2260 	caps->srqwqe_hop_num = pbl_hop_num;
2261 	caps->idx_hop_num = pbl_hop_num;
2262 	caps->wqe_sq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_SQWQE_HOP_NUM);
2263 	caps->wqe_sge_hop_num = hr_reg_read(resp_d, PF_CAPS_D_EX_SGE_HOP_NUM);
2264 	caps->wqe_rq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_RQWQE_HOP_NUM);
2265 
2266 	if (!(caps->page_size_cap & PAGE_SIZE))
2267 		caps->page_size_cap = HNS_ROCE_V2_PAGE_SIZE_SUPPORTED;
2268 
2269 	if (!hr_dev->is_vf) {
2270 		caps->cqe_sz = resp_a->cqe_sz;
2271 		caps->qpc_sz = le16_to_cpu(resp_b->qpc_sz);
2272 		caps->default_aeq_arm_st =
2273 				hr_reg_read(resp_d, PF_CAPS_D_AEQ_ARM_ST);
2274 		caps->default_ceq_arm_st =
2275 				hr_reg_read(resp_d, PF_CAPS_D_CEQ_ARM_ST);
2276 		caps->default_ceq_max_cnt = le16_to_cpu(resp_e->ceq_max_cnt);
2277 		caps->default_ceq_period = le16_to_cpu(resp_e->ceq_period);
2278 		caps->default_aeq_max_cnt = le16_to_cpu(resp_e->aeq_max_cnt);
2279 		caps->default_aeq_period = le16_to_cpu(resp_e->aeq_period);
2280 	}
2281 
2282 	return 0;
2283 }
2284 
2285 static int config_hem_entry_size(struct hns_roce_dev *hr_dev, u32 type, u32 val)
2286 {
2287 	struct hns_roce_cmq_desc desc;
2288 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
2289 
2290 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_ENTRY_SIZE,
2291 				      false);
2292 
2293 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_TYPE, type);
2294 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_VALUE, val);
2295 
2296 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2297 }
2298 
2299 static int hns_roce_config_entry_size(struct hns_roce_dev *hr_dev)
2300 {
2301 	struct hns_roce_caps *caps = &hr_dev->caps;
2302 	int ret;
2303 
2304 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2305 		return 0;
2306 
2307 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_QPC_SIZE,
2308 				    caps->qpc_sz);
2309 	if (ret) {
2310 		dev_err(hr_dev->dev, "failed to cfg qpc sz, ret = %d.\n", ret);
2311 		return ret;
2312 	}
2313 
2314 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_SCCC_SIZE,
2315 				    caps->sccc_sz);
2316 	if (ret)
2317 		dev_err(hr_dev->dev, "failed to cfg sccc sz, ret = %d.\n", ret);
2318 
2319 	return ret;
2320 }
2321 
2322 static int hns_roce_v2_vf_profile(struct hns_roce_dev *hr_dev)
2323 {
2324 	struct device *dev = hr_dev->dev;
2325 	int ret;
2326 
2327 	hr_dev->func_num = 1;
2328 
2329 	ret = hns_roce_query_caps(hr_dev);
2330 	if (ret) {
2331 		dev_err(dev, "failed to query VF caps, ret = %d.\n", ret);
2332 		return ret;
2333 	}
2334 
2335 	ret = hns_roce_query_vf_resource(hr_dev);
2336 	if (ret) {
2337 		dev_err(dev, "failed to query VF resource, ret = %d.\n", ret);
2338 		return ret;
2339 	}
2340 
2341 	apply_func_caps(hr_dev);
2342 
2343 	ret = hns_roce_v2_set_bt(hr_dev);
2344 	if (ret)
2345 		dev_err(dev, "failed to config VF BA table, ret = %d.\n", ret);
2346 
2347 	return ret;
2348 }
2349 
2350 static int hns_roce_v2_pf_profile(struct hns_roce_dev *hr_dev)
2351 {
2352 	struct device *dev = hr_dev->dev;
2353 	int ret;
2354 
2355 	ret = hns_roce_query_func_info(hr_dev);
2356 	if (ret) {
2357 		dev_err(dev, "failed to query func info, ret = %d.\n", ret);
2358 		return ret;
2359 	}
2360 
2361 	ret = hns_roce_config_global_param(hr_dev);
2362 	if (ret) {
2363 		dev_err(dev, "failed to config global param, ret = %d.\n", ret);
2364 		return ret;
2365 	}
2366 
2367 	ret = hns_roce_set_vf_switch_param(hr_dev);
2368 	if (ret) {
2369 		dev_err(dev, "failed to set switch param, ret = %d.\n", ret);
2370 		return ret;
2371 	}
2372 
2373 	ret = hns_roce_query_caps(hr_dev);
2374 	if (ret) {
2375 		dev_err(dev, "failed to query PF caps, ret = %d.\n", ret);
2376 		return ret;
2377 	}
2378 
2379 	ret = hns_roce_query_pf_resource(hr_dev);
2380 	if (ret) {
2381 		dev_err(dev, "failed to query pf resource, ret = %d.\n", ret);
2382 		return ret;
2383 	}
2384 
2385 	apply_func_caps(hr_dev);
2386 
2387 	ret = hns_roce_alloc_vf_resource(hr_dev);
2388 	if (ret) {
2389 		dev_err(dev, "failed to alloc vf resource, ret = %d.\n", ret);
2390 		return ret;
2391 	}
2392 
2393 	ret = hns_roce_v2_set_bt(hr_dev);
2394 	if (ret) {
2395 		dev_err(dev, "failed to config BA table, ret = %d.\n", ret);
2396 		return ret;
2397 	}
2398 
2399 	/* Configure the size of QPC, SCCC, etc. */
2400 	return hns_roce_config_entry_size(hr_dev);
2401 }
2402 
2403 static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
2404 {
2405 	struct device *dev = hr_dev->dev;
2406 	int ret;
2407 
2408 	ret = hns_roce_cmq_query_hw_info(hr_dev);
2409 	if (ret) {
2410 		dev_err(dev, "failed to query hardware info, ret = %d.\n", ret);
2411 		return ret;
2412 	}
2413 
2414 	ret = hns_roce_query_fw_ver(hr_dev);
2415 	if (ret) {
2416 		dev_err(dev, "failed to query firmware info, ret = %d.\n", ret);
2417 		return ret;
2418 	}
2419 
2420 	hr_dev->vendor_part_id = hr_dev->pci_dev->device;
2421 	hr_dev->sys_image_guid = be64_to_cpu(hr_dev->ib_dev.node_guid);
2422 
2423 	if (hr_dev->is_vf)
2424 		return hns_roce_v2_vf_profile(hr_dev);
2425 	else
2426 		return hns_roce_v2_pf_profile(hr_dev);
2427 }
2428 
2429 static void config_llm_table(struct hns_roce_buf *data_buf, void *cfg_buf)
2430 {
2431 	u32 i, next_ptr, page_num;
2432 	__le64 *entry = cfg_buf;
2433 	dma_addr_t addr;
2434 	u64 val;
2435 
2436 	page_num = data_buf->npages;
2437 	for (i = 0; i < page_num; i++) {
2438 		addr = hns_roce_buf_page(data_buf, i);
2439 		if (i == (page_num - 1))
2440 			next_ptr = 0;
2441 		else
2442 			next_ptr = i + 1;
2443 
2444 		val = HNS_ROCE_EXT_LLM_ENTRY(addr, (u64)next_ptr);
2445 		entry[i] = cpu_to_le64(val);
2446 	}
2447 }
2448 
2449 static int set_llm_cfg_to_hw(struct hns_roce_dev *hr_dev,
2450 			     struct hns_roce_link_table *table)
2451 {
2452 	struct hns_roce_cmq_desc desc[2];
2453 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
2454 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
2455 	struct hns_roce_buf *buf = table->buf;
2456 	enum hns_roce_opcode_type opcode;
2457 	dma_addr_t addr;
2458 
2459 	opcode = HNS_ROCE_OPC_CFG_EXT_LLM;
2460 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
2461 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2462 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
2463 
2464 	hr_reg_write(r_a, CFG_LLM_A_BA_L, lower_32_bits(table->table.map));
2465 	hr_reg_write(r_a, CFG_LLM_A_BA_H, upper_32_bits(table->table.map));
2466 	hr_reg_write(r_a, CFG_LLM_A_DEPTH, buf->npages);
2467 	hr_reg_write(r_a, CFG_LLM_A_PGSZ, to_hr_hw_page_shift(buf->page_shift));
2468 	hr_reg_enable(r_a, CFG_LLM_A_INIT_EN);
2469 
2470 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, 0));
2471 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_L, lower_32_bits(addr));
2472 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_H, upper_32_bits(addr));
2473 	hr_reg_write(r_a, CFG_LLM_A_HEAD_NXTPTR, 1);
2474 	hr_reg_write(r_a, CFG_LLM_A_HEAD_PTR, 0);
2475 
2476 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, buf->npages - 1));
2477 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_L, lower_32_bits(addr));
2478 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_H, upper_32_bits(addr));
2479 	hr_reg_write(r_b, CFG_LLM_B_TAIL_PTR, buf->npages - 1);
2480 
2481 	return hns_roce_cmq_send(hr_dev, desc, 2);
2482 }
2483 
2484 static struct hns_roce_link_table *
2485 alloc_link_table_buf(struct hns_roce_dev *hr_dev)
2486 {
2487 	u16 total_sl = hr_dev->caps.sl_num * hr_dev->func_num;
2488 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2489 	struct hns_roce_link_table *link_tbl;
2490 	u32 pg_shift, size, min_size;
2491 
2492 	link_tbl = &priv->ext_llm;
2493 	pg_shift = hr_dev->caps.llm_buf_pg_sz + PAGE_SHIFT;
2494 	size = hr_dev->caps.num_qps * hr_dev->func_num *
2495 	       HNS_ROCE_V2_EXT_LLM_ENTRY_SZ;
2496 	min_size = HNS_ROCE_EXT_LLM_MIN_PAGES(total_sl) << pg_shift;
2497 
2498 	/* Alloc data table */
2499 	size = max(size, min_size);
2500 	link_tbl->buf = hns_roce_buf_alloc(hr_dev, size, pg_shift, 0);
2501 	if (IS_ERR(link_tbl->buf))
2502 		return ERR_PTR(-ENOMEM);
2503 
2504 	/* Alloc config table */
2505 	size = link_tbl->buf->npages * sizeof(u64);
2506 	link_tbl->table.buf = dma_alloc_coherent(hr_dev->dev, size,
2507 						 &link_tbl->table.map,
2508 						 GFP_KERNEL);
2509 	if (!link_tbl->table.buf) {
2510 		hns_roce_buf_free(hr_dev, link_tbl->buf);
2511 		return ERR_PTR(-ENOMEM);
2512 	}
2513 
2514 	return link_tbl;
2515 }
2516 
2517 static void free_link_table_buf(struct hns_roce_dev *hr_dev,
2518 				struct hns_roce_link_table *tbl)
2519 {
2520 	if (tbl->buf) {
2521 		u32 size = tbl->buf->npages * sizeof(u64);
2522 
2523 		dma_free_coherent(hr_dev->dev, size, tbl->table.buf,
2524 				  tbl->table.map);
2525 	}
2526 
2527 	hns_roce_buf_free(hr_dev, tbl->buf);
2528 }
2529 
2530 static int hns_roce_init_link_table(struct hns_roce_dev *hr_dev)
2531 {
2532 	struct hns_roce_link_table *link_tbl;
2533 	int ret;
2534 
2535 	link_tbl = alloc_link_table_buf(hr_dev);
2536 	if (IS_ERR(link_tbl))
2537 		return -ENOMEM;
2538 
2539 	if (WARN_ON(link_tbl->buf->npages > HNS_ROCE_V2_EXT_LLM_MAX_DEPTH)) {
2540 		ret = -EINVAL;
2541 		goto err_alloc;
2542 	}
2543 
2544 	config_llm_table(link_tbl->buf, link_tbl->table.buf);
2545 	ret = set_llm_cfg_to_hw(hr_dev, link_tbl);
2546 	if (ret)
2547 		goto err_alloc;
2548 
2549 	return 0;
2550 
2551 err_alloc:
2552 	free_link_table_buf(hr_dev, link_tbl);
2553 	return ret;
2554 }
2555 
2556 static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev)
2557 {
2558 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2559 
2560 	free_link_table_buf(hr_dev, &priv->ext_llm);
2561 }
2562 
2563 static void free_dip_list(struct hns_roce_dev *hr_dev)
2564 {
2565 	struct hns_roce_dip *hr_dip;
2566 	struct hns_roce_dip *tmp;
2567 	unsigned long flags;
2568 
2569 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
2570 
2571 	list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) {
2572 		list_del(&hr_dip->node);
2573 		kfree(hr_dip);
2574 	}
2575 
2576 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
2577 }
2578 
2579 static struct ib_pd *free_mr_init_pd(struct hns_roce_dev *hr_dev)
2580 {
2581 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2582 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2583 	struct ib_device *ibdev = &hr_dev->ib_dev;
2584 	struct hns_roce_pd *hr_pd;
2585 	struct ib_pd *pd;
2586 
2587 	hr_pd = kzalloc(sizeof(*hr_pd), GFP_KERNEL);
2588 	if (ZERO_OR_NULL_PTR(hr_pd))
2589 		return NULL;
2590 	pd = &hr_pd->ibpd;
2591 	pd->device = ibdev;
2592 
2593 	if (hns_roce_alloc_pd(pd, NULL)) {
2594 		ibdev_err(ibdev, "failed to create pd for free mr.\n");
2595 		kfree(hr_pd);
2596 		return NULL;
2597 	}
2598 	free_mr->rsv_pd = to_hr_pd(pd);
2599 	free_mr->rsv_pd->ibpd.device = &hr_dev->ib_dev;
2600 	free_mr->rsv_pd->ibpd.uobject = NULL;
2601 	free_mr->rsv_pd->ibpd.__internal_mr = NULL;
2602 	atomic_set(&free_mr->rsv_pd->ibpd.usecnt, 0);
2603 
2604 	return pd;
2605 }
2606 
2607 static struct ib_cq *free_mr_init_cq(struct hns_roce_dev *hr_dev)
2608 {
2609 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2610 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2611 	struct ib_device *ibdev = &hr_dev->ib_dev;
2612 	struct ib_cq_init_attr cq_init_attr = {};
2613 	struct hns_roce_cq *hr_cq;
2614 	struct ib_cq *cq;
2615 
2616 	cq_init_attr.cqe = HNS_ROCE_FREE_MR_USED_CQE_NUM;
2617 
2618 	hr_cq = kzalloc(sizeof(*hr_cq), GFP_KERNEL);
2619 	if (ZERO_OR_NULL_PTR(hr_cq))
2620 		return NULL;
2621 
2622 	cq = &hr_cq->ib_cq;
2623 	cq->device = ibdev;
2624 
2625 	if (hns_roce_create_cq(cq, &cq_init_attr, NULL)) {
2626 		ibdev_err(ibdev, "failed to create cq for free mr.\n");
2627 		kfree(hr_cq);
2628 		return NULL;
2629 	}
2630 	free_mr->rsv_cq = to_hr_cq(cq);
2631 	free_mr->rsv_cq->ib_cq.device = &hr_dev->ib_dev;
2632 	free_mr->rsv_cq->ib_cq.uobject = NULL;
2633 	free_mr->rsv_cq->ib_cq.comp_handler = NULL;
2634 	free_mr->rsv_cq->ib_cq.event_handler = NULL;
2635 	free_mr->rsv_cq->ib_cq.cq_context = NULL;
2636 	atomic_set(&free_mr->rsv_cq->ib_cq.usecnt, 0);
2637 
2638 	return cq;
2639 }
2640 
2641 static int free_mr_init_qp(struct hns_roce_dev *hr_dev, struct ib_cq *cq,
2642 			   struct ib_qp_init_attr *init_attr, int i)
2643 {
2644 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2645 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2646 	struct ib_device *ibdev = &hr_dev->ib_dev;
2647 	struct hns_roce_qp *hr_qp;
2648 	struct ib_qp *qp;
2649 	int ret;
2650 
2651 	hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
2652 	if (ZERO_OR_NULL_PTR(hr_qp))
2653 		return -ENOMEM;
2654 
2655 	qp = &hr_qp->ibqp;
2656 	qp->device = ibdev;
2657 
2658 	ret = hns_roce_create_qp(qp, init_attr, NULL);
2659 	if (ret) {
2660 		ibdev_err(ibdev, "failed to create qp for free mr.\n");
2661 		kfree(hr_qp);
2662 		return ret;
2663 	}
2664 
2665 	free_mr->rsv_qp[i] = hr_qp;
2666 	free_mr->rsv_qp[i]->ibqp.recv_cq = cq;
2667 	free_mr->rsv_qp[i]->ibqp.send_cq = cq;
2668 
2669 	return 0;
2670 }
2671 
2672 static void free_mr_exit(struct hns_roce_dev *hr_dev)
2673 {
2674 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2675 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2676 	struct ib_qp *qp;
2677 	int i;
2678 
2679 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2680 		if (free_mr->rsv_qp[i]) {
2681 			qp = &free_mr->rsv_qp[i]->ibqp;
2682 			hns_roce_v2_destroy_qp(qp, NULL);
2683 			kfree(free_mr->rsv_qp[i]);
2684 			free_mr->rsv_qp[i] = NULL;
2685 		}
2686 	}
2687 
2688 	if (free_mr->rsv_cq) {
2689 		hns_roce_destroy_cq(&free_mr->rsv_cq->ib_cq, NULL);
2690 		kfree(free_mr->rsv_cq);
2691 		free_mr->rsv_cq = NULL;
2692 	}
2693 
2694 	if (free_mr->rsv_pd) {
2695 		hns_roce_dealloc_pd(&free_mr->rsv_pd->ibpd, NULL);
2696 		kfree(free_mr->rsv_pd);
2697 		free_mr->rsv_pd = NULL;
2698 	}
2699 
2700 	mutex_destroy(&free_mr->mutex);
2701 }
2702 
2703 static int free_mr_alloc_res(struct hns_roce_dev *hr_dev)
2704 {
2705 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2706 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2707 	struct ib_qp_init_attr qp_init_attr = {};
2708 	struct ib_pd *pd;
2709 	struct ib_cq *cq;
2710 	int ret;
2711 	int i;
2712 
2713 	pd = free_mr_init_pd(hr_dev);
2714 	if (!pd)
2715 		return -ENOMEM;
2716 
2717 	cq = free_mr_init_cq(hr_dev);
2718 	if (!cq) {
2719 		ret = -ENOMEM;
2720 		goto create_failed_cq;
2721 	}
2722 
2723 	qp_init_attr.qp_type = IB_QPT_RC;
2724 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2725 	qp_init_attr.send_cq = cq;
2726 	qp_init_attr.recv_cq = cq;
2727 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2728 		qp_init_attr.cap.max_send_wr = HNS_ROCE_FREE_MR_USED_SQWQE_NUM;
2729 		qp_init_attr.cap.max_send_sge = HNS_ROCE_FREE_MR_USED_SQSGE_NUM;
2730 		qp_init_attr.cap.max_recv_wr = HNS_ROCE_FREE_MR_USED_RQWQE_NUM;
2731 		qp_init_attr.cap.max_recv_sge = HNS_ROCE_FREE_MR_USED_RQSGE_NUM;
2732 
2733 		ret = free_mr_init_qp(hr_dev, cq, &qp_init_attr, i);
2734 		if (ret)
2735 			goto create_failed_qp;
2736 	}
2737 
2738 	return 0;
2739 
2740 create_failed_qp:
2741 	for (i--; i >= 0; i--) {
2742 		hns_roce_v2_destroy_qp(&free_mr->rsv_qp[i]->ibqp, NULL);
2743 		kfree(free_mr->rsv_qp[i]);
2744 	}
2745 	hns_roce_destroy_cq(cq, NULL);
2746 	kfree(cq);
2747 
2748 create_failed_cq:
2749 	hns_roce_dealloc_pd(pd, NULL);
2750 	kfree(pd);
2751 
2752 	return ret;
2753 }
2754 
2755 static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
2756 				 struct ib_qp_attr *attr, int sl_num)
2757 {
2758 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2759 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2760 	struct ib_device *ibdev = &hr_dev->ib_dev;
2761 	struct hns_roce_qp *hr_qp;
2762 	int loopback;
2763 	int mask;
2764 	int ret;
2765 
2766 	hr_qp = to_hr_qp(&free_mr->rsv_qp[sl_num]->ibqp);
2767 	hr_qp->free_mr_en = 1;
2768 	hr_qp->ibqp.device = ibdev;
2769 	hr_qp->ibqp.qp_type = IB_QPT_RC;
2770 
2771 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_ACCESS_FLAGS;
2772 	attr->qp_state = IB_QPS_INIT;
2773 	attr->port_num = 1;
2774 	attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
2775 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2776 				    IB_QPS_INIT, NULL);
2777 	if (ret) {
2778 		ibdev_err(ibdev, "failed to modify qp to init, ret = %d.\n",
2779 			  ret);
2780 		return ret;
2781 	}
2782 
2783 	loopback = hr_dev->loop_idc;
2784 	/* Set qpc lbi = 1 incidate loopback IO */
2785 	hr_dev->loop_idc = 1;
2786 
2787 	mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
2788 	       IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
2789 	attr->qp_state = IB_QPS_RTR;
2790 	attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2791 	attr->path_mtu = IB_MTU_256;
2792 	attr->dest_qp_num = hr_qp->qpn;
2793 	attr->rq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2794 
2795 	rdma_ah_set_sl(&attr->ah_attr, (u8)sl_num);
2796 
2797 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2798 				    IB_QPS_RTR, NULL);
2799 	hr_dev->loop_idc = loopback;
2800 	if (ret) {
2801 		ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
2802 			  ret);
2803 		return ret;
2804 	}
2805 
2806 	mask = IB_QP_STATE | IB_QP_SQ_PSN | IB_QP_RETRY_CNT | IB_QP_TIMEOUT |
2807 	       IB_QP_RNR_RETRY | IB_QP_MAX_QP_RD_ATOMIC;
2808 	attr->qp_state = IB_QPS_RTS;
2809 	attr->sq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2810 	attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
2811 	attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
2812 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_RTR,
2813 				    IB_QPS_RTS, NULL);
2814 	if (ret)
2815 		ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
2816 			  ret);
2817 
2818 	return ret;
2819 }
2820 
2821 static int free_mr_modify_qp(struct hns_roce_dev *hr_dev)
2822 {
2823 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2824 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2825 	struct ib_qp_attr attr = {};
2826 	int ret;
2827 	int i;
2828 
2829 	rdma_ah_set_grh(&attr.ah_attr, NULL, 0, 0, 1, 0);
2830 	rdma_ah_set_static_rate(&attr.ah_attr, 3);
2831 	rdma_ah_set_port_num(&attr.ah_attr, 1);
2832 
2833 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2834 		ret = free_mr_modify_rsv_qp(hr_dev, &attr, i);
2835 		if (ret)
2836 			return ret;
2837 	}
2838 
2839 	return 0;
2840 }
2841 
2842 static int free_mr_init(struct hns_roce_dev *hr_dev)
2843 {
2844 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2845 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2846 	int ret;
2847 
2848 	mutex_init(&free_mr->mutex);
2849 
2850 	ret = free_mr_alloc_res(hr_dev);
2851 	if (ret) {
2852 		mutex_destroy(&free_mr->mutex);
2853 		return ret;
2854 	}
2855 
2856 	ret = free_mr_modify_qp(hr_dev);
2857 	if (ret)
2858 		goto err_modify_qp;
2859 
2860 	return 0;
2861 
2862 err_modify_qp:
2863 	free_mr_exit(hr_dev);
2864 
2865 	return ret;
2866 }
2867 
2868 static int get_hem_table(struct hns_roce_dev *hr_dev)
2869 {
2870 	unsigned int qpc_count;
2871 	unsigned int cqc_count;
2872 	unsigned int gmv_count;
2873 	int ret;
2874 	int i;
2875 
2876 	/* Alloc memory for source address table buffer space chunk */
2877 	for (gmv_count = 0; gmv_count < hr_dev->caps.gmv_entry_num;
2878 	     gmv_count++) {
2879 		ret = hns_roce_table_get(hr_dev, &hr_dev->gmv_table, gmv_count);
2880 		if (ret)
2881 			goto err_gmv_failed;
2882 	}
2883 
2884 	if (hr_dev->is_vf)
2885 		return 0;
2886 
2887 	/* Alloc memory for QPC Timer buffer space chunk */
2888 	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
2889 	     qpc_count++) {
2890 		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
2891 					 qpc_count);
2892 		if (ret) {
2893 			dev_err(hr_dev->dev, "QPC Timer get failed\n");
2894 			goto err_qpc_timer_failed;
2895 		}
2896 	}
2897 
2898 	/* Alloc memory for CQC Timer buffer space chunk */
2899 	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
2900 	     cqc_count++) {
2901 		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
2902 					 cqc_count);
2903 		if (ret) {
2904 			dev_err(hr_dev->dev, "CQC Timer get failed\n");
2905 			goto err_cqc_timer_failed;
2906 		}
2907 	}
2908 
2909 	return 0;
2910 
2911 err_cqc_timer_failed:
2912 	for (i = 0; i < cqc_count; i++)
2913 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2914 
2915 err_qpc_timer_failed:
2916 	for (i = 0; i < qpc_count; i++)
2917 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2918 
2919 err_gmv_failed:
2920 	for (i = 0; i < gmv_count; i++)
2921 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2922 
2923 	return ret;
2924 }
2925 
2926 static void put_hem_table(struct hns_roce_dev *hr_dev)
2927 {
2928 	int i;
2929 
2930 	for (i = 0; i < hr_dev->caps.gmv_entry_num; i++)
2931 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2932 
2933 	if (hr_dev->is_vf)
2934 		return;
2935 
2936 	for (i = 0; i < hr_dev->caps.qpc_timer_bt_num; i++)
2937 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2938 
2939 	for (i = 0; i < hr_dev->caps.cqc_timer_bt_num; i++)
2940 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2941 }
2942 
2943 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
2944 {
2945 	int ret;
2946 
2947 	/* The hns ROCEE requires the extdb info to be cleared before using */
2948 	ret = hns_roce_clear_extdb_list_info(hr_dev);
2949 	if (ret)
2950 		return ret;
2951 
2952 	ret = get_hem_table(hr_dev);
2953 	if (ret)
2954 		return ret;
2955 
2956 	if (hr_dev->is_vf)
2957 		return 0;
2958 
2959 	ret = hns_roce_init_link_table(hr_dev);
2960 	if (ret) {
2961 		dev_err(hr_dev->dev, "failed to init llm, ret = %d.\n", ret);
2962 		goto err_llm_init_failed;
2963 	}
2964 
2965 	return 0;
2966 
2967 err_llm_init_failed:
2968 	put_hem_table(hr_dev);
2969 
2970 	return ret;
2971 }
2972 
2973 static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev)
2974 {
2975 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2976 		free_mr_exit(hr_dev);
2977 
2978 	hns_roce_function_clear(hr_dev);
2979 
2980 	if (!hr_dev->is_vf)
2981 		hns_roce_free_link_table(hr_dev);
2982 
2983 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09)
2984 		free_dip_list(hr_dev);
2985 }
2986 
2987 static int hns_roce_mbox_post(struct hns_roce_dev *hr_dev,
2988 			      struct hns_roce_mbox_msg *mbox_msg)
2989 {
2990 	struct hns_roce_cmq_desc desc;
2991 	struct hns_roce_post_mbox *mb = (struct hns_roce_post_mbox *)desc.data;
2992 
2993 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_POST_MB, false);
2994 
2995 	mb->in_param_l = cpu_to_le32(mbox_msg->in_param);
2996 	mb->in_param_h = cpu_to_le32(mbox_msg->in_param >> 32);
2997 	mb->out_param_l = cpu_to_le32(mbox_msg->out_param);
2998 	mb->out_param_h = cpu_to_le32(mbox_msg->out_param >> 32);
2999 	mb->cmd_tag = cpu_to_le32(mbox_msg->tag << 8 | mbox_msg->cmd);
3000 	mb->token_event_en = cpu_to_le32(mbox_msg->event_en << 16 |
3001 					 mbox_msg->token);
3002 
3003 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3004 }
3005 
3006 static int v2_wait_mbox_complete(struct hns_roce_dev *hr_dev, u32 timeout,
3007 				 u8 *complete_status)
3008 {
3009 	struct hns_roce_mbox_status *mb_st;
3010 	struct hns_roce_cmq_desc desc;
3011 	unsigned long end;
3012 	int ret = -EBUSY;
3013 	u32 status;
3014 	bool busy;
3015 
3016 	mb_st = (struct hns_roce_mbox_status *)desc.data;
3017 	end = msecs_to_jiffies(timeout) + jiffies;
3018 	while (v2_chk_mbox_is_avail(hr_dev, &busy)) {
3019 		if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
3020 			return -EIO;
3021 
3022 		status = 0;
3023 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_MB_ST,
3024 					      true);
3025 		ret = __hns_roce_cmq_send(hr_dev, &desc, 1);
3026 		if (!ret) {
3027 			status = le32_to_cpu(mb_st->mb_status_hw_run);
3028 			/* No pending message exists in ROCEE mbox. */
3029 			if (!(status & MB_ST_HW_RUN_M))
3030 				break;
3031 		} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3032 			break;
3033 		}
3034 
3035 		if (time_after(jiffies, end)) {
3036 			dev_err_ratelimited(hr_dev->dev,
3037 					    "failed to wait mbox status 0x%x\n",
3038 					    status);
3039 			return -ETIMEDOUT;
3040 		}
3041 
3042 		cond_resched();
3043 		ret = -EBUSY;
3044 	}
3045 
3046 	if (!ret) {
3047 		*complete_status = (u8)(status & MB_ST_COMPLETE_M);
3048 	} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3049 		/* Ignore all errors if the mbox is unavailable. */
3050 		ret = 0;
3051 		*complete_status = MB_ST_COMPLETE_M;
3052 	}
3053 
3054 	return ret;
3055 }
3056 
3057 static int v2_post_mbox(struct hns_roce_dev *hr_dev,
3058 			struct hns_roce_mbox_msg *mbox_msg)
3059 {
3060 	u8 status = 0;
3061 	int ret;
3062 
3063 	/* Waiting for the mbox to be idle */
3064 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_V2_GO_BIT_TIMEOUT_MSECS,
3065 				    &status);
3066 	if (unlikely(ret)) {
3067 		dev_err_ratelimited(hr_dev->dev,
3068 				    "failed to check post mbox status = 0x%x, ret = %d.\n",
3069 				    status, ret);
3070 		return ret;
3071 	}
3072 
3073 	/* Post new message to mbox */
3074 	ret = hns_roce_mbox_post(hr_dev, mbox_msg);
3075 	if (ret)
3076 		dev_err_ratelimited(hr_dev->dev,
3077 				    "failed to post mailbox, ret = %d.\n", ret);
3078 
3079 	return ret;
3080 }
3081 
3082 static int v2_poll_mbox_done(struct hns_roce_dev *hr_dev)
3083 {
3084 	u8 status = 0;
3085 	int ret;
3086 
3087 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_CMD_TIMEOUT_MSECS,
3088 				    &status);
3089 	if (!ret) {
3090 		if (status != MB_ST_COMPLETE_SUCC)
3091 			return -EBUSY;
3092 	} else {
3093 		dev_err_ratelimited(hr_dev->dev,
3094 				    "failed to check mbox status = 0x%x, ret = %d.\n",
3095 				    status, ret);
3096 	}
3097 
3098 	return ret;
3099 }
3100 
3101 static void copy_gid(void *dest, const union ib_gid *gid)
3102 {
3103 #define GID_SIZE 4
3104 	const union ib_gid *src = gid;
3105 	__le32 (*p)[GID_SIZE] = dest;
3106 	int i;
3107 
3108 	if (!gid)
3109 		src = &zgid;
3110 
3111 	for (i = 0; i < GID_SIZE; i++)
3112 		(*p)[i] = cpu_to_le32(*(u32 *)&src->raw[i * sizeof(u32)]);
3113 }
3114 
3115 static int config_sgid_table(struct hns_roce_dev *hr_dev,
3116 			     int gid_index, const union ib_gid *gid,
3117 			     enum hns_roce_sgid_type sgid_type)
3118 {
3119 	struct hns_roce_cmq_desc desc;
3120 	struct hns_roce_cfg_sgid_tb *sgid_tb =
3121 				    (struct hns_roce_cfg_sgid_tb *)desc.data;
3122 
3123 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SGID_TB, false);
3124 
3125 	hr_reg_write(sgid_tb, CFG_SGID_TB_TABLE_IDX, gid_index);
3126 	hr_reg_write(sgid_tb, CFG_SGID_TB_VF_SGID_TYPE, sgid_type);
3127 
3128 	copy_gid(&sgid_tb->vf_sgid_l, gid);
3129 
3130 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3131 }
3132 
3133 static int config_gmv_table(struct hns_roce_dev *hr_dev,
3134 			    int gid_index, const union ib_gid *gid,
3135 			    enum hns_roce_sgid_type sgid_type,
3136 			    const struct ib_gid_attr *attr)
3137 {
3138 	struct hns_roce_cmq_desc desc[2];
3139 	struct hns_roce_cfg_gmv_tb_a *tb_a =
3140 				(struct hns_roce_cfg_gmv_tb_a *)desc[0].data;
3141 	struct hns_roce_cfg_gmv_tb_b *tb_b =
3142 				(struct hns_roce_cfg_gmv_tb_b *)desc[1].data;
3143 
3144 	u16 vlan_id = VLAN_CFI_MASK;
3145 	u8 mac[ETH_ALEN] = {};
3146 	int ret;
3147 
3148 	if (gid) {
3149 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, mac);
3150 		if (ret)
3151 			return ret;
3152 	}
3153 
3154 	hns_roce_cmq_setup_basic_desc(&desc[0], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3155 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
3156 
3157 	hns_roce_cmq_setup_basic_desc(&desc[1], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3158 
3159 	copy_gid(&tb_a->vf_sgid_l, gid);
3160 
3161 	hr_reg_write(tb_a, GMV_TB_A_VF_SGID_TYPE, sgid_type);
3162 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_EN, vlan_id < VLAN_CFI_MASK);
3163 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_ID, vlan_id);
3164 
3165 	tb_b->vf_smac_l = cpu_to_le32(*(u32 *)mac);
3166 
3167 	hr_reg_write(tb_b, GMV_TB_B_SMAC_H, *(u16 *)&mac[4]);
3168 	hr_reg_write(tb_b, GMV_TB_B_SGID_IDX, gid_index);
3169 
3170 	return hns_roce_cmq_send(hr_dev, desc, 2);
3171 }
3172 
3173 static int hns_roce_v2_set_gid(struct hns_roce_dev *hr_dev, int gid_index,
3174 			       const union ib_gid *gid,
3175 			       const struct ib_gid_attr *attr)
3176 {
3177 	enum hns_roce_sgid_type sgid_type = GID_TYPE_FLAG_ROCE_V1;
3178 	int ret;
3179 
3180 	if (gid) {
3181 		if (attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3182 			if (ipv6_addr_v4mapped((void *)gid))
3183 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV4;
3184 			else
3185 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV6;
3186 		} else if (attr->gid_type == IB_GID_TYPE_ROCE) {
3187 			sgid_type = GID_TYPE_FLAG_ROCE_V1;
3188 		}
3189 	}
3190 
3191 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
3192 		ret = config_gmv_table(hr_dev, gid_index, gid, sgid_type, attr);
3193 	else
3194 		ret = config_sgid_table(hr_dev, gid_index, gid, sgid_type);
3195 
3196 	if (ret)
3197 		ibdev_err(&hr_dev->ib_dev, "failed to set gid, ret = %d!\n",
3198 			  ret);
3199 
3200 	return ret;
3201 }
3202 
3203 static int hns_roce_v2_set_mac(struct hns_roce_dev *hr_dev, u8 phy_port,
3204 			       const u8 *addr)
3205 {
3206 	struct hns_roce_cmq_desc desc;
3207 	struct hns_roce_cfg_smac_tb *smac_tb =
3208 				    (struct hns_roce_cfg_smac_tb *)desc.data;
3209 	u16 reg_smac_h;
3210 	u32 reg_smac_l;
3211 
3212 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SMAC_TB, false);
3213 
3214 	reg_smac_l = *(u32 *)(&addr[0]);
3215 	reg_smac_h = *(u16 *)(&addr[4]);
3216 
3217 	hr_reg_write(smac_tb, CFG_SMAC_TB_IDX, phy_port);
3218 	hr_reg_write(smac_tb, CFG_SMAC_TB_VF_SMAC_H, reg_smac_h);
3219 	smac_tb->vf_smac_l = cpu_to_le32(reg_smac_l);
3220 
3221 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3222 }
3223 
3224 static int set_mtpt_pbl(struct hns_roce_dev *hr_dev,
3225 			struct hns_roce_v2_mpt_entry *mpt_entry,
3226 			struct hns_roce_mr *mr)
3227 {
3228 	u64 pages[HNS_ROCE_V2_MAX_INNER_MTPT_NUM] = { 0 };
3229 	struct ib_device *ibdev = &hr_dev->ib_dev;
3230 	dma_addr_t pbl_ba;
3231 	int ret;
3232 	int i;
3233 
3234 	ret = hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, pages,
3235 				min_t(int, ARRAY_SIZE(pages), mr->npages));
3236 	if (ret) {
3237 		ibdev_err(ibdev, "failed to find PBL mtr, ret = %d.\n", ret);
3238 		return ret;
3239 	}
3240 
3241 	/* Aligned to the hardware address access unit */
3242 	for (i = 0; i < ARRAY_SIZE(pages); i++)
3243 		pages[i] >>= MPT_PBL_BUF_ADDR_S;
3244 
3245 	pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3246 
3247 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3248 	mpt_entry->pbl_ba_l = cpu_to_le32(pbl_ba >> MPT_PBL_BA_ADDR_S);
3249 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3250 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3251 
3252 	mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
3253 	hr_reg_write(mpt_entry, MPT_PA0_H, upper_32_bits(pages[0]));
3254 
3255 	mpt_entry->pa1_l = cpu_to_le32(lower_32_bits(pages[1]));
3256 	hr_reg_write(mpt_entry, MPT_PA1_H, upper_32_bits(pages[1]));
3257 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3258 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3259 
3260 	return 0;
3261 }
3262 
3263 static int hns_roce_v2_write_mtpt(struct hns_roce_dev *hr_dev,
3264 				  void *mb_buf, struct hns_roce_mr *mr)
3265 {
3266 	struct hns_roce_v2_mpt_entry *mpt_entry;
3267 
3268 	mpt_entry = mb_buf;
3269 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3270 
3271 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3272 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3273 
3274 	hr_reg_write_bool(mpt_entry, MPT_BIND_EN,
3275 			  mr->access & IB_ACCESS_MW_BIND);
3276 	hr_reg_write_bool(mpt_entry, MPT_ATOMIC_EN,
3277 			  mr->access & IB_ACCESS_REMOTE_ATOMIC);
3278 	hr_reg_write_bool(mpt_entry, MPT_RR_EN,
3279 			  mr->access & IB_ACCESS_REMOTE_READ);
3280 	hr_reg_write_bool(mpt_entry, MPT_RW_EN,
3281 			  mr->access & IB_ACCESS_REMOTE_WRITE);
3282 	hr_reg_write_bool(mpt_entry, MPT_LW_EN,
3283 			  mr->access & IB_ACCESS_LOCAL_WRITE);
3284 
3285 	mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3286 	mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3287 	mpt_entry->lkey = cpu_to_le32(mr->key);
3288 	mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3289 	mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3290 
3291 	if (mr->type != MR_TYPE_MR)
3292 		hr_reg_enable(mpt_entry, MPT_PA);
3293 
3294 	if (mr->type == MR_TYPE_DMA)
3295 		return 0;
3296 
3297 	if (mr->pbl_hop_num != HNS_ROCE_HOP_NUM_0)
3298 		hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, mr->pbl_hop_num);
3299 
3300 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3301 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3302 	hr_reg_enable(mpt_entry, MPT_INNER_PA_VLD);
3303 
3304 	return set_mtpt_pbl(hr_dev, mpt_entry, mr);
3305 }
3306 
3307 static int hns_roce_v2_rereg_write_mtpt(struct hns_roce_dev *hr_dev,
3308 					struct hns_roce_mr *mr, int flags,
3309 					void *mb_buf)
3310 {
3311 	struct hns_roce_v2_mpt_entry *mpt_entry = mb_buf;
3312 	u32 mr_access_flags = mr->access;
3313 	int ret = 0;
3314 
3315 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3316 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3317 
3318 	if (flags & IB_MR_REREG_ACCESS) {
3319 		hr_reg_write(mpt_entry, MPT_BIND_EN,
3320 			     (mr_access_flags & IB_ACCESS_MW_BIND ? 1 : 0));
3321 		hr_reg_write(mpt_entry, MPT_ATOMIC_EN,
3322 			     mr_access_flags & IB_ACCESS_REMOTE_ATOMIC ? 1 : 0);
3323 		hr_reg_write(mpt_entry, MPT_RR_EN,
3324 			     mr_access_flags & IB_ACCESS_REMOTE_READ ? 1 : 0);
3325 		hr_reg_write(mpt_entry, MPT_RW_EN,
3326 			     mr_access_flags & IB_ACCESS_REMOTE_WRITE ? 1 : 0);
3327 		hr_reg_write(mpt_entry, MPT_LW_EN,
3328 			     mr_access_flags & IB_ACCESS_LOCAL_WRITE ? 1 : 0);
3329 	}
3330 
3331 	if (flags & IB_MR_REREG_TRANS) {
3332 		mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3333 		mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3334 		mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3335 		mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3336 
3337 		ret = set_mtpt_pbl(hr_dev, mpt_entry, mr);
3338 	}
3339 
3340 	return ret;
3341 }
3342 
3343 static int hns_roce_v2_frmr_write_mtpt(void *mb_buf, struct hns_roce_mr *mr)
3344 {
3345 	dma_addr_t pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3346 	struct hns_roce_v2_mpt_entry *mpt_entry;
3347 
3348 	mpt_entry = mb_buf;
3349 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3350 
3351 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3352 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3353 
3354 	hr_reg_enable(mpt_entry, MPT_RA_EN);
3355 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3356 
3357 	hr_reg_enable(mpt_entry, MPT_FRE);
3358 	hr_reg_clear(mpt_entry, MPT_MR_MW);
3359 	hr_reg_enable(mpt_entry, MPT_BPD);
3360 	hr_reg_clear(mpt_entry, MPT_PA);
3361 
3362 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, 1);
3363 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3364 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3365 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3366 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3367 
3368 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3369 
3370 	mpt_entry->pbl_ba_l = cpu_to_le32(lower_32_bits(pbl_ba >>
3371 							MPT_PBL_BA_ADDR_S));
3372 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3373 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3374 
3375 	return 0;
3376 }
3377 
3378 static int hns_roce_v2_mw_write_mtpt(void *mb_buf, struct hns_roce_mw *mw)
3379 {
3380 	struct hns_roce_v2_mpt_entry *mpt_entry;
3381 
3382 	mpt_entry = mb_buf;
3383 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3384 
3385 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3386 	hr_reg_write(mpt_entry, MPT_PD, mw->pdn);
3387 
3388 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3389 	hr_reg_enable(mpt_entry, MPT_LW_EN);
3390 
3391 	hr_reg_enable(mpt_entry, MPT_MR_MW);
3392 	hr_reg_enable(mpt_entry, MPT_BPD);
3393 	hr_reg_clear(mpt_entry, MPT_PA);
3394 	hr_reg_write(mpt_entry, MPT_BQP,
3395 		     mw->ibmw.type == IB_MW_TYPE_1 ? 0 : 1);
3396 
3397 	mpt_entry->lkey = cpu_to_le32(mw->rkey);
3398 
3399 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM,
3400 		     mw->pbl_hop_num == HNS_ROCE_HOP_NUM_0 ? 0 :
3401 							     mw->pbl_hop_num);
3402 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3403 		     mw->pbl_ba_pg_sz + PG_SHIFT_OFFSET);
3404 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3405 		     mw->pbl_buf_pg_sz + PG_SHIFT_OFFSET);
3406 
3407 	return 0;
3408 }
3409 
3410 static int free_mr_post_send_lp_wqe(struct hns_roce_qp *hr_qp)
3411 {
3412 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
3413 	struct ib_device *ibdev = &hr_dev->ib_dev;
3414 	const struct ib_send_wr *bad_wr;
3415 	struct ib_rdma_wr rdma_wr = {};
3416 	struct ib_send_wr *send_wr;
3417 	int ret;
3418 
3419 	send_wr = &rdma_wr.wr;
3420 	send_wr->opcode = IB_WR_RDMA_WRITE;
3421 
3422 	ret = hns_roce_v2_post_send(&hr_qp->ibqp, send_wr, &bad_wr);
3423 	if (ret) {
3424 		ibdev_err(ibdev, "failed to post wqe for free mr, ret = %d.\n",
3425 			  ret);
3426 		return ret;
3427 	}
3428 
3429 	return 0;
3430 }
3431 
3432 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3433 			       struct ib_wc *wc);
3434 
3435 static void free_mr_send_cmd_to_hw(struct hns_roce_dev *hr_dev)
3436 {
3437 	struct hns_roce_v2_priv *priv = hr_dev->priv;
3438 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
3439 	struct ib_wc wc[ARRAY_SIZE(free_mr->rsv_qp)];
3440 	struct ib_device *ibdev = &hr_dev->ib_dev;
3441 	struct hns_roce_qp *hr_qp;
3442 	unsigned long end;
3443 	int cqe_cnt = 0;
3444 	int npolled;
3445 	int ret;
3446 	int i;
3447 
3448 	/*
3449 	 * If the device initialization is not complete or in the uninstall
3450 	 * process, then there is no need to execute free mr.
3451 	 */
3452 	if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
3453 	    priv->handle->rinfo.instance_state == HNS_ROCE_STATE_INIT ||
3454 	    hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT)
3455 		return;
3456 
3457 	mutex_lock(&free_mr->mutex);
3458 
3459 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
3460 		hr_qp = free_mr->rsv_qp[i];
3461 
3462 		ret = free_mr_post_send_lp_wqe(hr_qp);
3463 		if (ret) {
3464 			ibdev_err(ibdev,
3465 				  "failed to send wqe (qp:0x%lx) for free mr, ret = %d.\n",
3466 				  hr_qp->qpn, ret);
3467 			break;
3468 		}
3469 
3470 		cqe_cnt++;
3471 	}
3472 
3473 	end = msecs_to_jiffies(HNS_ROCE_V2_FREE_MR_TIMEOUT) + jiffies;
3474 	while (cqe_cnt) {
3475 		npolled = hns_roce_v2_poll_cq(&free_mr->rsv_cq->ib_cq, cqe_cnt, wc);
3476 		if (npolled < 0) {
3477 			ibdev_err(ibdev,
3478 				  "failed to poll cqe for free mr, remain %d cqe.\n",
3479 				  cqe_cnt);
3480 			goto out;
3481 		}
3482 
3483 		if (time_after(jiffies, end)) {
3484 			ibdev_err(ibdev,
3485 				  "failed to poll cqe for free mr and timeout, remain %d cqe.\n",
3486 				  cqe_cnt);
3487 			goto out;
3488 		}
3489 		cqe_cnt -= npolled;
3490 	}
3491 
3492 out:
3493 	mutex_unlock(&free_mr->mutex);
3494 }
3495 
3496 static void hns_roce_v2_dereg_mr(struct hns_roce_dev *hr_dev)
3497 {
3498 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
3499 		free_mr_send_cmd_to_hw(hr_dev);
3500 }
3501 
3502 static void *get_cqe_v2(struct hns_roce_cq *hr_cq, int n)
3503 {
3504 	return hns_roce_buf_offset(hr_cq->mtr.kmem, n * hr_cq->cqe_size);
3505 }
3506 
3507 static void *get_sw_cqe_v2(struct hns_roce_cq *hr_cq, unsigned int n)
3508 {
3509 	struct hns_roce_v2_cqe *cqe = get_cqe_v2(hr_cq, n & hr_cq->ib_cq.cqe);
3510 
3511 	/* Get cqe when Owner bit is Conversely with the MSB of cons_idx */
3512 	return (hr_reg_read(cqe, CQE_OWNER) ^ !!(n & hr_cq->cq_depth)) ? cqe :
3513 									 NULL;
3514 }
3515 
3516 static inline void update_cq_db(struct hns_roce_dev *hr_dev,
3517 				struct hns_roce_cq *hr_cq)
3518 {
3519 	if (likely(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) {
3520 		*hr_cq->set_ci_db = hr_cq->cons_index & V2_CQ_DB_CONS_IDX_M;
3521 	} else {
3522 		struct hns_roce_v2_db cq_db = {};
3523 
3524 		hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3525 		hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB);
3526 		hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3527 		hr_reg_write(&cq_db, DB_CQ_CMD_SN, 1);
3528 
3529 		hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3530 	}
3531 }
3532 
3533 static void __hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3534 				   struct hns_roce_srq *srq)
3535 {
3536 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3537 	struct hns_roce_v2_cqe *cqe, *dest;
3538 	u32 prod_index;
3539 	int nfreed = 0;
3540 	int wqe_index;
3541 	u8 owner_bit;
3542 
3543 	for (prod_index = hr_cq->cons_index; get_sw_cqe_v2(hr_cq, prod_index);
3544 	     ++prod_index) {
3545 		if (prod_index > hr_cq->cons_index + hr_cq->ib_cq.cqe)
3546 			break;
3547 	}
3548 
3549 	/*
3550 	 * Now backwards through the CQ, removing CQ entries
3551 	 * that match our QP by overwriting them with next entries.
3552 	 */
3553 	while ((int) --prod_index - (int) hr_cq->cons_index >= 0) {
3554 		cqe = get_cqe_v2(hr_cq, prod_index & hr_cq->ib_cq.cqe);
3555 		if (hr_reg_read(cqe, CQE_LCL_QPN) == qpn) {
3556 			if (srq && hr_reg_read(cqe, CQE_S_R)) {
3557 				wqe_index = hr_reg_read(cqe, CQE_WQE_IDX);
3558 				hns_roce_free_srq_wqe(srq, wqe_index);
3559 			}
3560 			++nfreed;
3561 		} else if (nfreed) {
3562 			dest = get_cqe_v2(hr_cq, (prod_index + nfreed) &
3563 					  hr_cq->ib_cq.cqe);
3564 			owner_bit = hr_reg_read(dest, CQE_OWNER);
3565 			memcpy(dest, cqe, hr_cq->cqe_size);
3566 			hr_reg_write(dest, CQE_OWNER, owner_bit);
3567 		}
3568 	}
3569 
3570 	if (nfreed) {
3571 		hr_cq->cons_index += nfreed;
3572 		update_cq_db(hr_dev, hr_cq);
3573 	}
3574 }
3575 
3576 static void hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3577 				 struct hns_roce_srq *srq)
3578 {
3579 	spin_lock_irq(&hr_cq->lock);
3580 	__hns_roce_v2_cq_clean(hr_cq, qpn, srq);
3581 	spin_unlock_irq(&hr_cq->lock);
3582 }
3583 
3584 static void hns_roce_v2_write_cqc(struct hns_roce_dev *hr_dev,
3585 				  struct hns_roce_cq *hr_cq, void *mb_buf,
3586 				  u64 *mtts, dma_addr_t dma_handle)
3587 {
3588 	struct hns_roce_v2_cq_context *cq_context;
3589 
3590 	cq_context = mb_buf;
3591 	memset(cq_context, 0, sizeof(*cq_context));
3592 
3593 	hr_reg_write(cq_context, CQC_CQ_ST, V2_CQ_STATE_VALID);
3594 	hr_reg_write(cq_context, CQC_ARM_ST, NO_ARMED);
3595 	hr_reg_write(cq_context, CQC_SHIFT, ilog2(hr_cq->cq_depth));
3596 	hr_reg_write(cq_context, CQC_CEQN, hr_cq->vector);
3597 	hr_reg_write(cq_context, CQC_CQN, hr_cq->cqn);
3598 
3599 	if (hr_cq->cqe_size == HNS_ROCE_V3_CQE_SIZE)
3600 		hr_reg_write(cq_context, CQC_CQE_SIZE, CQE_SIZE_64B);
3601 
3602 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
3603 		hr_reg_enable(cq_context, CQC_STASH);
3604 
3605 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_L,
3606 		     to_hr_hw_page_addr(mtts[0]));
3607 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_H,
3608 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
3609 	hr_reg_write(cq_context, CQC_CQE_HOP_NUM, hr_dev->caps.cqe_hop_num ==
3610 		     HNS_ROCE_HOP_NUM_0 ? 0 : hr_dev->caps.cqe_hop_num);
3611 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_L,
3612 		     to_hr_hw_page_addr(mtts[1]));
3613 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_H,
3614 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
3615 	hr_reg_write(cq_context, CQC_CQE_BAR_PG_SZ,
3616 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.ba_pg_shift));
3617 	hr_reg_write(cq_context, CQC_CQE_BUF_PG_SZ,
3618 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.buf_pg_shift));
3619 	hr_reg_write(cq_context, CQC_CQE_BA_L, dma_handle >> CQC_CQE_BA_L_S);
3620 	hr_reg_write(cq_context, CQC_CQE_BA_H, dma_handle >> CQC_CQE_BA_H_S);
3621 	hr_reg_write_bool(cq_context, CQC_DB_RECORD_EN,
3622 			  hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB);
3623 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_L,
3624 		     ((u32)hr_cq->db.dma) >> 1);
3625 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_H,
3626 		     hr_cq->db.dma >> CQC_CQE_DB_RECORD_ADDR_H_S);
3627 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT,
3628 		     HNS_ROCE_V2_CQ_DEFAULT_BURST_NUM);
3629 	hr_reg_write(cq_context, CQC_CQ_PERIOD,
3630 		     HNS_ROCE_V2_CQ_DEFAULT_INTERVAL);
3631 }
3632 
3633 static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
3634 				     enum ib_cq_notify_flags flags)
3635 {
3636 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3637 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3638 	struct hns_roce_v2_db cq_db = {};
3639 	u32 notify_flag;
3640 
3641 	/*
3642 	 * flags = 0, then notify_flag : next
3643 	 * flags = 1, then notify flag : solocited
3644 	 */
3645 	notify_flag = (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
3646 		      V2_CQ_DB_REQ_NOT : V2_CQ_DB_REQ_NOT_SOL;
3647 
3648 	hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3649 	hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB_NOTIFY);
3650 	hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3651 	hr_reg_write(&cq_db, DB_CQ_CMD_SN, hr_cq->arm_sn);
3652 	hr_reg_write(&cq_db, DB_CQ_NOTIFY, notify_flag);
3653 
3654 	hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3655 
3656 	return 0;
3657 }
3658 
3659 static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
3660 		   int num_entries, struct ib_wc *wc)
3661 {
3662 	unsigned int left;
3663 	int npolled = 0;
3664 
3665 	left = wq->head - wq->tail;
3666 	if (left == 0)
3667 		return 0;
3668 
3669 	left = min_t(unsigned int, (unsigned int)num_entries, left);
3670 	while (npolled < left) {
3671 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3672 		wc->status = IB_WC_WR_FLUSH_ERR;
3673 		wc->vendor_err = 0;
3674 		wc->qp = &hr_qp->ibqp;
3675 
3676 		wq->tail++;
3677 		wc++;
3678 		npolled++;
3679 	}
3680 
3681 	return npolled;
3682 }
3683 
3684 static int hns_roce_v2_sw_poll_cq(struct hns_roce_cq *hr_cq, int num_entries,
3685 				  struct ib_wc *wc)
3686 {
3687 	struct hns_roce_qp *hr_qp;
3688 	int npolled = 0;
3689 
3690 	list_for_each_entry(hr_qp, &hr_cq->sq_list, sq_node) {
3691 		npolled += sw_comp(hr_qp, &hr_qp->sq,
3692 				   num_entries - npolled, wc + npolled);
3693 		if (npolled >= num_entries)
3694 			goto out;
3695 	}
3696 
3697 	list_for_each_entry(hr_qp, &hr_cq->rq_list, rq_node) {
3698 		npolled += sw_comp(hr_qp, &hr_qp->rq,
3699 				   num_entries - npolled, wc + npolled);
3700 		if (npolled >= num_entries)
3701 			goto out;
3702 	}
3703 
3704 out:
3705 	return npolled;
3706 }
3707 
3708 static void get_cqe_status(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
3709 			   struct hns_roce_cq *cq, struct hns_roce_v2_cqe *cqe,
3710 			   struct ib_wc *wc)
3711 {
3712 	static const struct {
3713 		u32 cqe_status;
3714 		enum ib_wc_status wc_status;
3715 	} map[] = {
3716 		{ HNS_ROCE_CQE_V2_SUCCESS, IB_WC_SUCCESS },
3717 		{ HNS_ROCE_CQE_V2_LOCAL_LENGTH_ERR, IB_WC_LOC_LEN_ERR },
3718 		{ HNS_ROCE_CQE_V2_LOCAL_QP_OP_ERR, IB_WC_LOC_QP_OP_ERR },
3719 		{ HNS_ROCE_CQE_V2_LOCAL_PROT_ERR, IB_WC_LOC_PROT_ERR },
3720 		{ HNS_ROCE_CQE_V2_WR_FLUSH_ERR, IB_WC_WR_FLUSH_ERR },
3721 		{ HNS_ROCE_CQE_V2_MW_BIND_ERR, IB_WC_MW_BIND_ERR },
3722 		{ HNS_ROCE_CQE_V2_BAD_RESP_ERR, IB_WC_BAD_RESP_ERR },
3723 		{ HNS_ROCE_CQE_V2_LOCAL_ACCESS_ERR, IB_WC_LOC_ACCESS_ERR },
3724 		{ HNS_ROCE_CQE_V2_REMOTE_INVAL_REQ_ERR, IB_WC_REM_INV_REQ_ERR },
3725 		{ HNS_ROCE_CQE_V2_REMOTE_ACCESS_ERR, IB_WC_REM_ACCESS_ERR },
3726 		{ HNS_ROCE_CQE_V2_REMOTE_OP_ERR, IB_WC_REM_OP_ERR },
3727 		{ HNS_ROCE_CQE_V2_TRANSPORT_RETRY_EXC_ERR,
3728 		  IB_WC_RETRY_EXC_ERR },
3729 		{ HNS_ROCE_CQE_V2_RNR_RETRY_EXC_ERR, IB_WC_RNR_RETRY_EXC_ERR },
3730 		{ HNS_ROCE_CQE_V2_REMOTE_ABORT_ERR, IB_WC_REM_ABORT_ERR },
3731 		{ HNS_ROCE_CQE_V2_GENERAL_ERR, IB_WC_GENERAL_ERR}
3732 	};
3733 
3734 	u32 cqe_status = hr_reg_read(cqe, CQE_STATUS);
3735 	int i;
3736 
3737 	wc->status = IB_WC_GENERAL_ERR;
3738 	for (i = 0; i < ARRAY_SIZE(map); i++)
3739 		if (cqe_status == map[i].cqe_status) {
3740 			wc->status = map[i].wc_status;
3741 			break;
3742 		}
3743 
3744 	if (likely(wc->status == IB_WC_SUCCESS ||
3745 		   wc->status == IB_WC_WR_FLUSH_ERR))
3746 		return;
3747 
3748 	ibdev_err_ratelimited(&hr_dev->ib_dev, "error cqe status 0x%x:\n",
3749 			      cqe_status);
3750 	print_hex_dump(KERN_DEBUG, "", DUMP_PREFIX_NONE, 16, 4, cqe,
3751 		       cq->cqe_size, false);
3752 	wc->vendor_err = hr_reg_read(cqe, CQE_SUB_STATUS);
3753 
3754 	/*
3755 	 * For hns ROCEE, GENERAL_ERR is an error type that is not defined in
3756 	 * the standard protocol, the driver must ignore it and needn't to set
3757 	 * the QP to an error state.
3758 	 */
3759 	if (cqe_status == HNS_ROCE_CQE_V2_GENERAL_ERR)
3760 		return;
3761 
3762 	flush_cqe(hr_dev, qp);
3763 }
3764 
3765 static int get_cur_qp(struct hns_roce_cq *hr_cq, struct hns_roce_v2_cqe *cqe,
3766 		      struct hns_roce_qp **cur_qp)
3767 {
3768 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3769 	struct hns_roce_qp *hr_qp = *cur_qp;
3770 	u32 qpn;
3771 
3772 	qpn = hr_reg_read(cqe, CQE_LCL_QPN);
3773 
3774 	if (!hr_qp || qpn != hr_qp->qpn) {
3775 		hr_qp = __hns_roce_qp_lookup(hr_dev, qpn);
3776 		if (unlikely(!hr_qp)) {
3777 			ibdev_err(&hr_dev->ib_dev,
3778 				  "CQ %06lx with entry for unknown QPN %06x\n",
3779 				  hr_cq->cqn, qpn);
3780 			return -EINVAL;
3781 		}
3782 		*cur_qp = hr_qp;
3783 	}
3784 
3785 	return 0;
3786 }
3787 
3788 /*
3789  * mapped-value = 1 + real-value
3790  * The ib wc opcode's real value is start from 0, In order to distinguish
3791  * between initialized and uninitialized map values, we plus 1 to the actual
3792  * value when defining the mapping, so that the validity can be identified by
3793  * checking whether the mapped value is greater than 0.
3794  */
3795 #define HR_WC_OP_MAP(hr_key, ib_key) \
3796 		[HNS_ROCE_V2_WQE_OP_ ## hr_key] = 1 + IB_WC_ ## ib_key
3797 
3798 static const u32 wc_send_op_map[] = {
3799 	HR_WC_OP_MAP(SEND,			SEND),
3800 	HR_WC_OP_MAP(SEND_WITH_INV,		SEND),
3801 	HR_WC_OP_MAP(SEND_WITH_IMM,		SEND),
3802 	HR_WC_OP_MAP(RDMA_READ,			RDMA_READ),
3803 	HR_WC_OP_MAP(RDMA_WRITE,		RDMA_WRITE),
3804 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,	RDMA_WRITE),
3805 	HR_WC_OP_MAP(ATOM_CMP_AND_SWAP,		COMP_SWAP),
3806 	HR_WC_OP_MAP(ATOM_FETCH_AND_ADD,	FETCH_ADD),
3807 	HR_WC_OP_MAP(ATOM_MSK_CMP_AND_SWAP,	MASKED_COMP_SWAP),
3808 	HR_WC_OP_MAP(ATOM_MSK_FETCH_AND_ADD,	MASKED_FETCH_ADD),
3809 	HR_WC_OP_MAP(FAST_REG_PMR,		REG_MR),
3810 	HR_WC_OP_MAP(BIND_MW,			REG_MR),
3811 };
3812 
3813 static int to_ib_wc_send_op(u32 hr_opcode)
3814 {
3815 	if (hr_opcode >= ARRAY_SIZE(wc_send_op_map))
3816 		return -EINVAL;
3817 
3818 	return wc_send_op_map[hr_opcode] ? wc_send_op_map[hr_opcode] - 1 :
3819 					   -EINVAL;
3820 }
3821 
3822 static const u32 wc_recv_op_map[] = {
3823 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,		WITH_IMM),
3824 	HR_WC_OP_MAP(SEND,				RECV),
3825 	HR_WC_OP_MAP(SEND_WITH_IMM,			WITH_IMM),
3826 	HR_WC_OP_MAP(SEND_WITH_INV,			RECV),
3827 };
3828 
3829 static int to_ib_wc_recv_op(u32 hr_opcode)
3830 {
3831 	if (hr_opcode >= ARRAY_SIZE(wc_recv_op_map))
3832 		return -EINVAL;
3833 
3834 	return wc_recv_op_map[hr_opcode] ? wc_recv_op_map[hr_opcode] - 1 :
3835 					   -EINVAL;
3836 }
3837 
3838 static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3839 {
3840 	u32 hr_opcode;
3841 	int ib_opcode;
3842 
3843 	wc->wc_flags = 0;
3844 
3845 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3846 	switch (hr_opcode) {
3847 	case HNS_ROCE_V2_WQE_OP_RDMA_READ:
3848 		wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3849 		break;
3850 	case HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM:
3851 	case HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM:
3852 		wc->wc_flags |= IB_WC_WITH_IMM;
3853 		break;
3854 	case HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP:
3855 	case HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD:
3856 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP:
3857 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD:
3858 		wc->byte_len  = 8;
3859 		break;
3860 	default:
3861 		break;
3862 	}
3863 
3864 	ib_opcode = to_ib_wc_send_op(hr_opcode);
3865 	if (ib_opcode < 0)
3866 		wc->status = IB_WC_GENERAL_ERR;
3867 	else
3868 		wc->opcode = ib_opcode;
3869 }
3870 
3871 static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3872 {
3873 	u32 hr_opcode;
3874 	int ib_opcode;
3875 
3876 	wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3877 
3878 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3879 	switch (hr_opcode) {
3880 	case HNS_ROCE_V2_OPCODE_RDMA_WRITE_IMM:
3881 	case HNS_ROCE_V2_OPCODE_SEND_WITH_IMM:
3882 		wc->wc_flags = IB_WC_WITH_IMM;
3883 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->immtdata));
3884 		break;
3885 	case HNS_ROCE_V2_OPCODE_SEND_WITH_INV:
3886 		wc->wc_flags = IB_WC_WITH_INVALIDATE;
3887 		wc->ex.invalidate_rkey = le32_to_cpu(cqe->rkey);
3888 		break;
3889 	default:
3890 		wc->wc_flags = 0;
3891 	}
3892 
3893 	ib_opcode = to_ib_wc_recv_op(hr_opcode);
3894 	if (ib_opcode < 0)
3895 		wc->status = IB_WC_GENERAL_ERR;
3896 	else
3897 		wc->opcode = ib_opcode;
3898 
3899 	wc->sl = hr_reg_read(cqe, CQE_SL);
3900 	wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
3901 	wc->slid = 0;
3902 	wc->wc_flags |= hr_reg_read(cqe, CQE_GRH) ? IB_WC_GRH : 0;
3903 	wc->port_num = hr_reg_read(cqe, CQE_PORTN);
3904 	wc->pkey_index = 0;
3905 
3906 	if (hr_reg_read(cqe, CQE_VID_VLD)) {
3907 		wc->vlan_id = hr_reg_read(cqe, CQE_VID);
3908 		wc->wc_flags |= IB_WC_WITH_VLAN;
3909 	} else {
3910 		wc->vlan_id = 0xffff;
3911 	}
3912 
3913 	wc->network_hdr_type = hr_reg_read(cqe, CQE_PORT_TYPE);
3914 
3915 	return 0;
3916 }
3917 
3918 static int hns_roce_v2_poll_one(struct hns_roce_cq *hr_cq,
3919 				struct hns_roce_qp **cur_qp, struct ib_wc *wc)
3920 {
3921 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3922 	struct hns_roce_qp *qp = *cur_qp;
3923 	struct hns_roce_srq *srq = NULL;
3924 	struct hns_roce_v2_cqe *cqe;
3925 	struct hns_roce_wq *wq;
3926 	int is_send;
3927 	u16 wqe_idx;
3928 	int ret;
3929 
3930 	cqe = get_sw_cqe_v2(hr_cq, hr_cq->cons_index);
3931 	if (!cqe)
3932 		return -EAGAIN;
3933 
3934 	++hr_cq->cons_index;
3935 	/* Memory barrier */
3936 	rmb();
3937 
3938 	ret = get_cur_qp(hr_cq, cqe, &qp);
3939 	if (ret)
3940 		return ret;
3941 
3942 	wc->qp = &qp->ibqp;
3943 	wc->vendor_err = 0;
3944 
3945 	wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
3946 
3947 	is_send = !hr_reg_read(cqe, CQE_S_R);
3948 	if (is_send) {
3949 		wq = &qp->sq;
3950 
3951 		/* If sg_signal_bit is set, tail pointer will be updated to
3952 		 * the WQE corresponding to the current CQE.
3953 		 */
3954 		if (qp->sq_signal_bits)
3955 			wq->tail += (wqe_idx - (u16)wq->tail) &
3956 				    (wq->wqe_cnt - 1);
3957 
3958 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3959 		++wq->tail;
3960 
3961 		fill_send_wc(wc, cqe);
3962 	} else {
3963 		if (qp->ibqp.srq) {
3964 			srq = to_hr_srq(qp->ibqp.srq);
3965 			wc->wr_id = srq->wrid[wqe_idx];
3966 			hns_roce_free_srq_wqe(srq, wqe_idx);
3967 		} else {
3968 			wq = &qp->rq;
3969 			wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3970 			++wq->tail;
3971 		}
3972 
3973 		ret = fill_recv_wc(wc, cqe);
3974 	}
3975 
3976 	get_cqe_status(hr_dev, qp, hr_cq, cqe, wc);
3977 	if (unlikely(wc->status != IB_WC_SUCCESS))
3978 		return 0;
3979 
3980 	return ret;
3981 }
3982 
3983 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3984 			       struct ib_wc *wc)
3985 {
3986 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3987 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3988 	struct hns_roce_qp *cur_qp = NULL;
3989 	unsigned long flags;
3990 	int npolled;
3991 
3992 	spin_lock_irqsave(&hr_cq->lock, flags);
3993 
3994 	/*
3995 	 * When the device starts to reset, the state is RST_DOWN. At this time,
3996 	 * there may still be some valid CQEs in the hardware that are not
3997 	 * polled. Therefore, it is not allowed to switch to the software mode
3998 	 * immediately. When the state changes to UNINIT, CQE no longer exists
3999 	 * in the hardware, and then switch to software mode.
4000 	 */
4001 	if (hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT) {
4002 		npolled = hns_roce_v2_sw_poll_cq(hr_cq, num_entries, wc);
4003 		goto out;
4004 	}
4005 
4006 	for (npolled = 0; npolled < num_entries; ++npolled) {
4007 		if (hns_roce_v2_poll_one(hr_cq, &cur_qp, wc + npolled))
4008 			break;
4009 	}
4010 
4011 	if (npolled)
4012 		update_cq_db(hr_dev, hr_cq);
4013 
4014 out:
4015 	spin_unlock_irqrestore(&hr_cq->lock, flags);
4016 
4017 	return npolled;
4018 }
4019 
4020 static int get_op_for_set_hem(struct hns_roce_dev *hr_dev, u32 type,
4021 			      u32 step_idx, u8 *mbox_cmd)
4022 {
4023 	u8 cmd;
4024 
4025 	switch (type) {
4026 	case HEM_TYPE_QPC:
4027 		cmd = HNS_ROCE_CMD_WRITE_QPC_BT0;
4028 		break;
4029 	case HEM_TYPE_MTPT:
4030 		cmd = HNS_ROCE_CMD_WRITE_MPT_BT0;
4031 		break;
4032 	case HEM_TYPE_CQC:
4033 		cmd = HNS_ROCE_CMD_WRITE_CQC_BT0;
4034 		break;
4035 	case HEM_TYPE_SRQC:
4036 		cmd = HNS_ROCE_CMD_WRITE_SRQC_BT0;
4037 		break;
4038 	case HEM_TYPE_SCCC:
4039 		cmd = HNS_ROCE_CMD_WRITE_SCCC_BT0;
4040 		break;
4041 	case HEM_TYPE_QPC_TIMER:
4042 		cmd = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
4043 		break;
4044 	case HEM_TYPE_CQC_TIMER:
4045 		cmd = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
4046 		break;
4047 	default:
4048 		dev_warn(hr_dev->dev, "failed to check hem type %u.\n", type);
4049 		return -EINVAL;
4050 	}
4051 
4052 	*mbox_cmd = cmd + step_idx;
4053 
4054 	return 0;
4055 }
4056 
4057 static int config_gmv_ba_to_hw(struct hns_roce_dev *hr_dev, unsigned long obj,
4058 			       dma_addr_t base_addr)
4059 {
4060 	struct hns_roce_cmq_desc desc;
4061 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
4062 	u32 idx = obj / (HNS_HW_PAGE_SIZE / hr_dev->caps.gmv_entry_sz);
4063 	u64 addr = to_hr_hw_page_addr(base_addr);
4064 
4065 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
4066 
4067 	hr_reg_write(req, CFG_GMV_BT_BA_L, lower_32_bits(addr));
4068 	hr_reg_write(req, CFG_GMV_BT_BA_H, upper_32_bits(addr));
4069 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
4070 
4071 	return hns_roce_cmq_send(hr_dev, &desc, 1);
4072 }
4073 
4074 static int set_hem_to_hw(struct hns_roce_dev *hr_dev, int obj,
4075 			 dma_addr_t base_addr, u32 hem_type, u32 step_idx)
4076 {
4077 	int ret;
4078 	u8 cmd;
4079 
4080 	if (unlikely(hem_type == HEM_TYPE_GMV))
4081 		return config_gmv_ba_to_hw(hr_dev, obj, base_addr);
4082 
4083 	if (unlikely(hem_type == HEM_TYPE_SCCC && step_idx))
4084 		return 0;
4085 
4086 	ret = get_op_for_set_hem(hr_dev, hem_type, step_idx, &cmd);
4087 	if (ret < 0)
4088 		return ret;
4089 
4090 	return config_hem_ba_to_hw(hr_dev, base_addr, cmd, obj);
4091 }
4092 
4093 static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
4094 			       struct hns_roce_hem_table *table, int obj,
4095 			       u32 step_idx)
4096 {
4097 	struct hns_roce_hem_mhop mhop;
4098 	struct hns_roce_hem *hem;
4099 	unsigned long mhop_obj = obj;
4100 	int i, j, k;
4101 	int ret = 0;
4102 	u64 hem_idx = 0;
4103 	u64 l1_idx = 0;
4104 	u64 bt_ba = 0;
4105 	u32 chunk_ba_num;
4106 	u32 hop_num;
4107 
4108 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4109 		return 0;
4110 
4111 	hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
4112 	i = mhop.l0_idx;
4113 	j = mhop.l1_idx;
4114 	k = mhop.l2_idx;
4115 	hop_num = mhop.hop_num;
4116 	chunk_ba_num = mhop.bt_chunk_size / 8;
4117 
4118 	if (hop_num == 2) {
4119 		hem_idx = i * chunk_ba_num * chunk_ba_num + j * chunk_ba_num +
4120 			  k;
4121 		l1_idx = i * chunk_ba_num + j;
4122 	} else if (hop_num == 1) {
4123 		hem_idx = i * chunk_ba_num + j;
4124 	} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
4125 		hem_idx = i;
4126 	}
4127 
4128 	if (table->type == HEM_TYPE_SCCC)
4129 		obj = mhop.l0_idx;
4130 
4131 	if (check_whether_last_step(hop_num, step_idx)) {
4132 		hem = table->hem[hem_idx];
4133 
4134 		ret = set_hem_to_hw(hr_dev, obj, hem->dma, table->type, step_idx);
4135 	} else {
4136 		if (step_idx == 0)
4137 			bt_ba = table->bt_l0_dma_addr[i];
4138 		else if (step_idx == 1 && hop_num == 2)
4139 			bt_ba = table->bt_l1_dma_addr[l1_idx];
4140 
4141 		ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type, step_idx);
4142 	}
4143 
4144 	return ret;
4145 }
4146 
4147 static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
4148 				 struct hns_roce_hem_table *table,
4149 				 int tag, u32 step_idx)
4150 {
4151 	struct hns_roce_cmd_mailbox *mailbox;
4152 	struct device *dev = hr_dev->dev;
4153 	u8 cmd = 0xff;
4154 	int ret;
4155 
4156 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4157 		return 0;
4158 
4159 	switch (table->type) {
4160 	case HEM_TYPE_QPC:
4161 		cmd = HNS_ROCE_CMD_DESTROY_QPC_BT0;
4162 		break;
4163 	case HEM_TYPE_MTPT:
4164 		cmd = HNS_ROCE_CMD_DESTROY_MPT_BT0;
4165 		break;
4166 	case HEM_TYPE_CQC:
4167 		cmd = HNS_ROCE_CMD_DESTROY_CQC_BT0;
4168 		break;
4169 	case HEM_TYPE_SRQC:
4170 		cmd = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
4171 		break;
4172 	case HEM_TYPE_SCCC:
4173 	case HEM_TYPE_QPC_TIMER:
4174 	case HEM_TYPE_CQC_TIMER:
4175 	case HEM_TYPE_GMV:
4176 		return 0;
4177 	default:
4178 		dev_warn(dev, "table %u not to be destroyed by mailbox!\n",
4179 			 table->type);
4180 		return 0;
4181 	}
4182 
4183 	cmd += step_idx;
4184 
4185 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4186 	if (IS_ERR(mailbox))
4187 		return PTR_ERR(mailbox);
4188 
4189 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, cmd, tag);
4190 
4191 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4192 	return ret;
4193 }
4194 
4195 static int hns_roce_v2_qp_modify(struct hns_roce_dev *hr_dev,
4196 				 struct hns_roce_v2_qp_context *context,
4197 				 struct hns_roce_v2_qp_context *qpc_mask,
4198 				 struct hns_roce_qp *hr_qp)
4199 {
4200 	struct hns_roce_cmd_mailbox *mailbox;
4201 	int qpc_size;
4202 	int ret;
4203 
4204 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4205 	if (IS_ERR(mailbox))
4206 		return PTR_ERR(mailbox);
4207 
4208 	/* The qpc size of HIP08 is only 256B, which is half of HIP09 */
4209 	qpc_size = hr_dev->caps.qpc_sz;
4210 	memcpy(mailbox->buf, context, qpc_size);
4211 	memcpy(mailbox->buf + qpc_size, qpc_mask, qpc_size);
4212 
4213 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
4214 				HNS_ROCE_CMD_MODIFY_QPC, hr_qp->qpn);
4215 
4216 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4217 
4218 	return ret;
4219 }
4220 
4221 static void set_access_flags(struct hns_roce_qp *hr_qp,
4222 			     struct hns_roce_v2_qp_context *context,
4223 			     struct hns_roce_v2_qp_context *qpc_mask,
4224 			     const struct ib_qp_attr *attr, int attr_mask)
4225 {
4226 	u8 dest_rd_atomic;
4227 	u32 access_flags;
4228 
4229 	dest_rd_atomic = (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) ?
4230 			 attr->max_dest_rd_atomic : hr_qp->resp_depth;
4231 
4232 	access_flags = (attr_mask & IB_QP_ACCESS_FLAGS) ?
4233 		       attr->qp_access_flags : hr_qp->atomic_rd_en;
4234 
4235 	if (!dest_rd_atomic)
4236 		access_flags &= IB_ACCESS_REMOTE_WRITE;
4237 
4238 	hr_reg_write_bool(context, QPC_RRE,
4239 			  access_flags & IB_ACCESS_REMOTE_READ);
4240 	hr_reg_clear(qpc_mask, QPC_RRE);
4241 
4242 	hr_reg_write_bool(context, QPC_RWE,
4243 			  access_flags & IB_ACCESS_REMOTE_WRITE);
4244 	hr_reg_clear(qpc_mask, QPC_RWE);
4245 
4246 	hr_reg_write_bool(context, QPC_ATE,
4247 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4248 	hr_reg_clear(qpc_mask, QPC_ATE);
4249 	hr_reg_write_bool(context, QPC_EXT_ATE,
4250 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4251 	hr_reg_clear(qpc_mask, QPC_EXT_ATE);
4252 }
4253 
4254 static void set_qpc_wqe_cnt(struct hns_roce_qp *hr_qp,
4255 			    struct hns_roce_v2_qp_context *context)
4256 {
4257 	hr_reg_write(context, QPC_SGE_SHIFT,
4258 		     to_hr_hem_entries_shift(hr_qp->sge.sge_cnt,
4259 					     hr_qp->sge.sge_shift));
4260 
4261 	hr_reg_write(context, QPC_SQ_SHIFT, ilog2(hr_qp->sq.wqe_cnt));
4262 
4263 	hr_reg_write(context, QPC_RQ_SHIFT, ilog2(hr_qp->rq.wqe_cnt));
4264 }
4265 
4266 static inline int get_cqn(struct ib_cq *ib_cq)
4267 {
4268 	return ib_cq ? to_hr_cq(ib_cq)->cqn : 0;
4269 }
4270 
4271 static inline int get_pdn(struct ib_pd *ib_pd)
4272 {
4273 	return ib_pd ? to_hr_pd(ib_pd)->pdn : 0;
4274 }
4275 
4276 static void modify_qp_reset_to_init(struct ib_qp *ibqp,
4277 				    struct hns_roce_v2_qp_context *context,
4278 				    struct hns_roce_v2_qp_context *qpc_mask)
4279 {
4280 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4281 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4282 
4283 	/*
4284 	 * In v2 engine, software pass context and context mask to hardware
4285 	 * when modifying qp. If software need modify some fields in context,
4286 	 * we should set all bits of the relevant fields in context mask to
4287 	 * 0 at the same time, else set them to 0x1.
4288 	 */
4289 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4290 
4291 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4292 
4293 	hr_reg_write(context, QPC_RQWS, ilog2(hr_qp->rq.max_gs));
4294 
4295 	set_qpc_wqe_cnt(hr_qp, context);
4296 
4297 	/* No VLAN need to set 0xFFF */
4298 	hr_reg_write(context, QPC_VLAN_ID, 0xfff);
4299 
4300 	if (ibqp->qp_type == IB_QPT_XRC_TGT) {
4301 		context->qkey_xrcd = cpu_to_le32(hr_qp->xrcdn);
4302 
4303 		hr_reg_enable(context, QPC_XRC_QP_TYPE);
4304 	}
4305 
4306 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
4307 		hr_reg_enable(context, QPC_RQ_RECORD_EN);
4308 
4309 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
4310 		hr_reg_enable(context, QPC_OWNER_MODE);
4311 
4312 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_L,
4313 		     lower_32_bits(hr_qp->rdb.dma) >> 1);
4314 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
4315 		     upper_32_bits(hr_qp->rdb.dma));
4316 
4317 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4318 
4319 	if (ibqp->srq) {
4320 		hr_reg_enable(context, QPC_SRQ_EN);
4321 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4322 	}
4323 
4324 	hr_reg_enable(context, QPC_FRE);
4325 
4326 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4327 
4328 	if (hr_dev->caps.qpc_sz < HNS_ROCE_V3_QPC_SZ)
4329 		return;
4330 
4331 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
4332 		hr_reg_enable(&context->ext, QPCEX_STASH);
4333 }
4334 
4335 static void modify_qp_init_to_init(struct ib_qp *ibqp,
4336 				   struct hns_roce_v2_qp_context *context,
4337 				   struct hns_roce_v2_qp_context *qpc_mask)
4338 {
4339 	/*
4340 	 * In v2 engine, software pass context and context mask to hardware
4341 	 * when modifying qp. If software need modify some fields in context,
4342 	 * we should set all bits of the relevant fields in context mask to
4343 	 * 0 at the same time, else set them to 0x1.
4344 	 */
4345 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4346 	hr_reg_clear(qpc_mask, QPC_TST);
4347 
4348 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4349 	hr_reg_clear(qpc_mask, QPC_PD);
4350 
4351 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4352 	hr_reg_clear(qpc_mask, QPC_RX_CQN);
4353 
4354 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4355 	hr_reg_clear(qpc_mask, QPC_TX_CQN);
4356 
4357 	if (ibqp->srq) {
4358 		hr_reg_enable(context, QPC_SRQ_EN);
4359 		hr_reg_clear(qpc_mask, QPC_SRQ_EN);
4360 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4361 		hr_reg_clear(qpc_mask, QPC_SRQN);
4362 	}
4363 }
4364 
4365 static int config_qp_rq_buf(struct hns_roce_dev *hr_dev,
4366 			    struct hns_roce_qp *hr_qp,
4367 			    struct hns_roce_v2_qp_context *context,
4368 			    struct hns_roce_v2_qp_context *qpc_mask)
4369 {
4370 	u64 mtts[MTT_MIN_COUNT] = { 0 };
4371 	u64 wqe_sge_ba;
4372 	int ret;
4373 
4374 	/* Search qp buf's mtts */
4375 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->rq.offset, mtts,
4376 				MTT_MIN_COUNT);
4377 	if (hr_qp->rq.wqe_cnt && ret) {
4378 		ibdev_err(&hr_dev->ib_dev,
4379 			  "failed to find QP(0x%lx) RQ WQE buf, ret = %d.\n",
4380 			  hr_qp->qpn, ret);
4381 		return ret;
4382 	}
4383 
4384 	wqe_sge_ba = hns_roce_get_mtr_ba(&hr_qp->mtr);
4385 
4386 	context->wqe_sge_ba = cpu_to_le32(wqe_sge_ba >> 3);
4387 	qpc_mask->wqe_sge_ba = 0;
4388 
4389 	/*
4390 	 * In v2 engine, software pass context and context mask to hardware
4391 	 * when modifying qp. If software need modify some fields in context,
4392 	 * we should set all bits of the relevant fields in context mask to
4393 	 * 0 at the same time, else set them to 0x1.
4394 	 */
4395 	hr_reg_write(context, QPC_WQE_SGE_BA_H, wqe_sge_ba >> (32 + 3));
4396 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_H);
4397 
4398 	hr_reg_write(context, QPC_SQ_HOP_NUM,
4399 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sq_hop_num,
4400 				      hr_qp->sq.wqe_cnt));
4401 	hr_reg_clear(qpc_mask, QPC_SQ_HOP_NUM);
4402 
4403 	hr_reg_write(context, QPC_SGE_HOP_NUM,
4404 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sge_hop_num,
4405 				      hr_qp->sge.sge_cnt));
4406 	hr_reg_clear(qpc_mask, QPC_SGE_HOP_NUM);
4407 
4408 	hr_reg_write(context, QPC_RQ_HOP_NUM,
4409 		     to_hr_hem_hopnum(hr_dev->caps.wqe_rq_hop_num,
4410 				      hr_qp->rq.wqe_cnt));
4411 
4412 	hr_reg_clear(qpc_mask, QPC_RQ_HOP_NUM);
4413 
4414 	hr_reg_write(context, QPC_WQE_SGE_BA_PG_SZ,
4415 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.ba_pg_shift));
4416 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_PG_SZ);
4417 
4418 	hr_reg_write(context, QPC_WQE_SGE_BUF_PG_SZ,
4419 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.buf_pg_shift));
4420 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BUF_PG_SZ);
4421 
4422 	context->rq_cur_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[0]));
4423 	qpc_mask->rq_cur_blk_addr = 0;
4424 
4425 	hr_reg_write(context, QPC_RQ_CUR_BLK_ADDR_H,
4426 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
4427 	hr_reg_clear(qpc_mask, QPC_RQ_CUR_BLK_ADDR_H);
4428 
4429 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4430 		context->rq_nxt_blk_addr =
4431 				cpu_to_le32(to_hr_hw_page_addr(mtts[1]));
4432 		qpc_mask->rq_nxt_blk_addr = 0;
4433 		hr_reg_write(context, QPC_RQ_NXT_BLK_ADDR_H,
4434 			     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
4435 		hr_reg_clear(qpc_mask, QPC_RQ_NXT_BLK_ADDR_H);
4436 	}
4437 
4438 	return 0;
4439 }
4440 
4441 static int config_qp_sq_buf(struct hns_roce_dev *hr_dev,
4442 			    struct hns_roce_qp *hr_qp,
4443 			    struct hns_roce_v2_qp_context *context,
4444 			    struct hns_roce_v2_qp_context *qpc_mask)
4445 {
4446 	struct ib_device *ibdev = &hr_dev->ib_dev;
4447 	u64 sge_cur_blk = 0;
4448 	u64 sq_cur_blk = 0;
4449 	int ret;
4450 
4451 	/* search qp buf's mtts */
4452 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->sq.offset,
4453 				&sq_cur_blk, 1);
4454 	if (ret) {
4455 		ibdev_err(ibdev, "failed to find QP(0x%lx) SQ WQE buf, ret = %d.\n",
4456 			  hr_qp->qpn, ret);
4457 		return ret;
4458 	}
4459 	if (hr_qp->sge.sge_cnt > 0) {
4460 		ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr,
4461 					hr_qp->sge.offset, &sge_cur_blk, 1);
4462 		if (ret) {
4463 			ibdev_err(ibdev, "failed to find QP(0x%lx) SGE buf, ret = %d.\n",
4464 				  hr_qp->qpn, ret);
4465 			return ret;
4466 		}
4467 	}
4468 
4469 	/*
4470 	 * In v2 engine, software pass context and context mask to hardware
4471 	 * when modifying qp. If software need modify some fields in context,
4472 	 * we should set all bits of the relevant fields in context mask to
4473 	 * 0 at the same time, else set them to 0x1.
4474 	 */
4475 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_L,
4476 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4477 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_H,
4478 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4479 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_L);
4480 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_H);
4481 
4482 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_L,
4483 		     lower_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4484 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_H,
4485 		     upper_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4486 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_L);
4487 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_H);
4488 
4489 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_L,
4490 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4491 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_H,
4492 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4493 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_L);
4494 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_H);
4495 
4496 	return 0;
4497 }
4498 
4499 static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
4500 				  const struct ib_qp_attr *attr)
4501 {
4502 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_UD)
4503 		return IB_MTU_4096;
4504 
4505 	return attr->path_mtu;
4506 }
4507 
4508 static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
4509 				 const struct ib_qp_attr *attr, int attr_mask,
4510 				 struct hns_roce_v2_qp_context *context,
4511 				 struct hns_roce_v2_qp_context *qpc_mask,
4512 				 struct ib_udata *udata)
4513 {
4514 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
4515 					  struct hns_roce_ucontext, ibucontext);
4516 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4517 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4518 	struct ib_device *ibdev = &hr_dev->ib_dev;
4519 	dma_addr_t trrl_ba;
4520 	dma_addr_t irrl_ba;
4521 	enum ib_mtu ib_mtu;
4522 	const u8 *smac;
4523 	u8 lp_pktn_ini;
4524 	u64 *mtts;
4525 	u8 *dmac;
4526 	u32 port;
4527 	int mtu;
4528 	int ret;
4529 
4530 	ret = config_qp_rq_buf(hr_dev, hr_qp, context, qpc_mask);
4531 	if (ret) {
4532 		ibdev_err(ibdev, "failed to config rq buf, ret = %d.\n", ret);
4533 		return ret;
4534 	}
4535 
4536 	/* Search IRRL's mtts */
4537 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.irrl_table,
4538 				   hr_qp->qpn, &irrl_ba);
4539 	if (!mtts) {
4540 		ibdev_err(ibdev, "failed to find qp irrl_table.\n");
4541 		return -EINVAL;
4542 	}
4543 
4544 	/* Search TRRL's mtts */
4545 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.trrl_table,
4546 				   hr_qp->qpn, &trrl_ba);
4547 	if (!mtts) {
4548 		ibdev_err(ibdev, "failed to find qp trrl_table.\n");
4549 		return -EINVAL;
4550 	}
4551 
4552 	if (attr_mask & IB_QP_ALT_PATH) {
4553 		ibdev_err(ibdev, "INIT2RTR attr_mask (0x%x) error.\n",
4554 			  attr_mask);
4555 		return -EINVAL;
4556 	}
4557 
4558 	hr_reg_write(context, QPC_TRRL_BA_L, trrl_ba >> QPC_TRRL_BA_L_S);
4559 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_L);
4560 	context->trrl_ba = cpu_to_le32(trrl_ba >> QPC_TRRL_BA_M_S);
4561 	qpc_mask->trrl_ba = 0;
4562 	hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> QPC_TRRL_BA_H_S);
4563 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
4564 
4565 	context->irrl_ba = cpu_to_le32(irrl_ba >> QPC_IRRL_BA_L_S);
4566 	qpc_mask->irrl_ba = 0;
4567 	hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> QPC_IRRL_BA_H_S);
4568 	hr_reg_clear(qpc_mask, QPC_IRRL_BA_H);
4569 
4570 	hr_reg_enable(context, QPC_RMT_E2E);
4571 	hr_reg_clear(qpc_mask, QPC_RMT_E2E);
4572 
4573 	hr_reg_write(context, QPC_SIG_TYPE, hr_qp->sq_signal_bits);
4574 	hr_reg_clear(qpc_mask, QPC_SIG_TYPE);
4575 
4576 	port = (attr_mask & IB_QP_PORT) ? (attr->port_num - 1) : hr_qp->port;
4577 
4578 	smac = (const u8 *)hr_dev->dev_addr[port];
4579 	dmac = (u8 *)attr->ah_attr.roce.dmac;
4580 	/* when dmac equals smac or loop_idc is 1, it should loopback */
4581 	if (ether_addr_equal_unaligned(dmac, smac) ||
4582 	    hr_dev->loop_idc == 0x1) {
4583 		hr_reg_write(context, QPC_LBI, hr_dev->loop_idc);
4584 		hr_reg_clear(qpc_mask, QPC_LBI);
4585 	}
4586 
4587 	if (attr_mask & IB_QP_DEST_QPN) {
4588 		hr_reg_write(context, QPC_DQPN, attr->dest_qp_num);
4589 		hr_reg_clear(qpc_mask, QPC_DQPN);
4590 	}
4591 
4592 	memcpy(&context->dmac, dmac, sizeof(u32));
4593 	hr_reg_write(context, QPC_DMAC_H, *((u16 *)(&dmac[4])));
4594 	qpc_mask->dmac = 0;
4595 	hr_reg_clear(qpc_mask, QPC_DMAC_H);
4596 
4597 	ib_mtu = get_mtu(ibqp, attr);
4598 	hr_qp->path_mtu = ib_mtu;
4599 
4600 	mtu = ib_mtu_enum_to_int(ib_mtu);
4601 	if (WARN_ON(mtu <= 0))
4602 		return -EINVAL;
4603 #define MIN_LP_MSG_LEN 1024
4604 	/* mtu * (2 ^ lp_pktn_ini) should be in the range of 1024 to mtu */
4605 	lp_pktn_ini = ilog2(max(mtu, MIN_LP_MSG_LEN) / mtu);
4606 
4607 	if (attr_mask & IB_QP_PATH_MTU) {
4608 		hr_reg_write(context, QPC_MTU, ib_mtu);
4609 		hr_reg_clear(qpc_mask, QPC_MTU);
4610 	}
4611 
4612 	hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
4613 	hr_reg_clear(qpc_mask, QPC_LP_PKTN_INI);
4614 
4615 	/* ACK_REQ_FREQ should be larger than or equal to LP_PKTN_INI */
4616 	hr_reg_write(context, QPC_ACK_REQ_FREQ, lp_pktn_ini);
4617 	hr_reg_clear(qpc_mask, QPC_ACK_REQ_FREQ);
4618 
4619 	hr_reg_clear(qpc_mask, QPC_RX_REQ_PSN_ERR);
4620 	hr_reg_clear(qpc_mask, QPC_RX_REQ_MSN);
4621 	hr_reg_clear(qpc_mask, QPC_RX_REQ_LAST_OPTYPE);
4622 
4623 	context->rq_rnr_timer = 0;
4624 	qpc_mask->rq_rnr_timer = 0;
4625 
4626 	hr_reg_clear(qpc_mask, QPC_TRRL_HEAD_MAX);
4627 	hr_reg_clear(qpc_mask, QPC_TRRL_TAIL_MAX);
4628 
4629 #define MAX_LP_SGEN 3
4630 	/* rocee send 2^lp_sgen_ini segs every time */
4631 	hr_reg_write(context, QPC_LP_SGEN_INI, MAX_LP_SGEN);
4632 	hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
4633 
4634 	if (udata && ibqp->qp_type == IB_QPT_RC &&
4635 	    (uctx->config & HNS_ROCE_RQ_INLINE_FLAGS)) {
4636 		hr_reg_write_bool(context, QPC_RQIE,
4637 				  hr_dev->caps.flags &
4638 				  HNS_ROCE_CAP_FLAG_RQ_INLINE);
4639 		hr_reg_clear(qpc_mask, QPC_RQIE);
4640 	}
4641 
4642 	if (udata &&
4643 	    (ibqp->qp_type == IB_QPT_RC || ibqp->qp_type == IB_QPT_XRC_TGT) &&
4644 	    (uctx->config & HNS_ROCE_CQE_INLINE_FLAGS)) {
4645 		hr_reg_write_bool(context, QPC_CQEIE,
4646 				  hr_dev->caps.flags &
4647 				  HNS_ROCE_CAP_FLAG_CQE_INLINE);
4648 		hr_reg_clear(qpc_mask, QPC_CQEIE);
4649 
4650 		hr_reg_write(context, QPC_CQEIS, 0);
4651 		hr_reg_clear(qpc_mask, QPC_CQEIS);
4652 	}
4653 
4654 	return 0;
4655 }
4656 
4657 static int modify_qp_rtr_to_rts(struct ib_qp *ibqp, int attr_mask,
4658 				struct hns_roce_v2_qp_context *context,
4659 				struct hns_roce_v2_qp_context *qpc_mask)
4660 {
4661 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4662 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4663 	struct ib_device *ibdev = &hr_dev->ib_dev;
4664 	int ret;
4665 
4666 	/* Not support alternate path and path migration */
4667 	if (attr_mask & (IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE)) {
4668 		ibdev_err(ibdev, "RTR2RTS attr_mask (0x%x)error\n", attr_mask);
4669 		return -EINVAL;
4670 	}
4671 
4672 	ret = config_qp_sq_buf(hr_dev, hr_qp, context, qpc_mask);
4673 	if (ret) {
4674 		ibdev_err(ibdev, "failed to config sq buf, ret = %d.\n", ret);
4675 		return ret;
4676 	}
4677 
4678 	/*
4679 	 * Set some fields in context to zero, Because the default values
4680 	 * of all fields in context are zero, we need not set them to 0 again.
4681 	 * but we should set the relevant fields of context mask to 0.
4682 	 */
4683 	hr_reg_clear(qpc_mask, QPC_IRRL_SGE_IDX);
4684 
4685 	hr_reg_clear(qpc_mask, QPC_RX_ACK_MSN);
4686 
4687 	hr_reg_clear(qpc_mask, QPC_ACK_LAST_OPTYPE);
4688 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN_VLD);
4689 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN);
4690 
4691 	hr_reg_clear(qpc_mask, QPC_IRRL_TAIL_REAL);
4692 
4693 	hr_reg_clear(qpc_mask, QPC_RETRY_MSG_MSN);
4694 
4695 	hr_reg_clear(qpc_mask, QPC_RNR_RETRY_FLAG);
4696 
4697 	hr_reg_clear(qpc_mask, QPC_CHECK_FLG);
4698 
4699 	hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
4700 
4701 	return 0;
4702 }
4703 
4704 static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4705 			   u32 *dip_idx)
4706 {
4707 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4708 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4709 	u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx;
4710 	u32 *head =  &hr_dev->qp_table.idx_table.head;
4711 	u32 *tail =  &hr_dev->qp_table.idx_table.tail;
4712 	struct hns_roce_dip *hr_dip;
4713 	unsigned long flags;
4714 	int ret = 0;
4715 
4716 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
4717 
4718 	spare_idx[*tail] = ibqp->qp_num;
4719 	*tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1);
4720 
4721 	list_for_each_entry(hr_dip, &hr_dev->dip_list, node) {
4722 		if (!memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) {
4723 			*dip_idx = hr_dip->dip_idx;
4724 			goto out;
4725 		}
4726 	}
4727 
4728 	/* If no dgid is found, a new dip and a mapping between dgid and
4729 	 * dip_idx will be created.
4730 	 */
4731 	hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC);
4732 	if (!hr_dip) {
4733 		ret = -ENOMEM;
4734 		goto out;
4735 	}
4736 
4737 	memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4738 	hr_dip->dip_idx = *dip_idx = spare_idx[*head];
4739 	*head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1);
4740 	list_add_tail(&hr_dip->node, &hr_dev->dip_list);
4741 
4742 out:
4743 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
4744 	return ret;
4745 }
4746 
4747 enum {
4748 	CONG_DCQCN,
4749 	CONG_WINDOW,
4750 };
4751 
4752 enum {
4753 	UNSUPPORT_CONG_LEVEL,
4754 	SUPPORT_CONG_LEVEL,
4755 };
4756 
4757 enum {
4758 	CONG_LDCP,
4759 	CONG_HC3,
4760 };
4761 
4762 enum {
4763 	DIP_INVALID,
4764 	DIP_VALID,
4765 };
4766 
4767 enum {
4768 	WND_LIMIT,
4769 	WND_UNLIMIT,
4770 };
4771 
4772 static int check_cong_type(struct ib_qp *ibqp,
4773 			   struct hns_roce_congestion_algorithm *cong_alg)
4774 {
4775 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4776 
4777 	/* different congestion types match different configurations */
4778 	switch (hr_qp->cong_type) {
4779 	case CONG_TYPE_DCQCN:
4780 		cong_alg->alg_sel = CONG_DCQCN;
4781 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4782 		cong_alg->dip_vld = DIP_INVALID;
4783 		cong_alg->wnd_mode_sel = WND_LIMIT;
4784 		break;
4785 	case CONG_TYPE_LDCP:
4786 		cong_alg->alg_sel = CONG_WINDOW;
4787 		cong_alg->alg_sub_sel = CONG_LDCP;
4788 		cong_alg->dip_vld = DIP_INVALID;
4789 		cong_alg->wnd_mode_sel = WND_UNLIMIT;
4790 		break;
4791 	case CONG_TYPE_HC3:
4792 		cong_alg->alg_sel = CONG_WINDOW;
4793 		cong_alg->alg_sub_sel = CONG_HC3;
4794 		cong_alg->dip_vld = DIP_INVALID;
4795 		cong_alg->wnd_mode_sel = WND_LIMIT;
4796 		break;
4797 	case CONG_TYPE_DIP:
4798 		cong_alg->alg_sel = CONG_DCQCN;
4799 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4800 		cong_alg->dip_vld = DIP_VALID;
4801 		cong_alg->wnd_mode_sel = WND_LIMIT;
4802 		break;
4803 	default:
4804 		hr_qp->cong_type = CONG_TYPE_DCQCN;
4805 		cong_alg->alg_sel = CONG_DCQCN;
4806 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4807 		cong_alg->dip_vld = DIP_INVALID;
4808 		cong_alg->wnd_mode_sel = WND_LIMIT;
4809 		break;
4810 	}
4811 
4812 	return 0;
4813 }
4814 
4815 static int fill_cong_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4816 			   struct hns_roce_v2_qp_context *context,
4817 			   struct hns_roce_v2_qp_context *qpc_mask)
4818 {
4819 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4820 	struct hns_roce_congestion_algorithm cong_field;
4821 	struct ib_device *ibdev = ibqp->device;
4822 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
4823 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4824 	u32 dip_idx = 0;
4825 	int ret;
4826 
4827 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ||
4828 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE)
4829 		return 0;
4830 
4831 	ret = check_cong_type(ibqp, &cong_field);
4832 	if (ret)
4833 		return ret;
4834 
4835 	hr_reg_write(context, QPC_CONG_ALGO_TMPL_ID, hr_dev->cong_algo_tmpl_id +
4836 		     hr_qp->cong_type * HNS_ROCE_CONG_SIZE);
4837 	hr_reg_clear(qpc_mask, QPC_CONG_ALGO_TMPL_ID);
4838 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SEL, cong_field.alg_sel);
4839 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SEL);
4840 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SUB_SEL,
4841 		     cong_field.alg_sub_sel);
4842 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SUB_SEL);
4843 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX_VLD, cong_field.dip_vld);
4844 	hr_reg_clear(&qpc_mask->ext, QPCEX_DIP_CTX_IDX_VLD);
4845 	hr_reg_write(&context->ext, QPCEX_SQ_RQ_NOT_FORBID_EN,
4846 		     cong_field.wnd_mode_sel);
4847 	hr_reg_clear(&qpc_mask->ext, QPCEX_SQ_RQ_NOT_FORBID_EN);
4848 
4849 	/* if dip is disabled, there is no need to set dip idx */
4850 	if (cong_field.dip_vld == 0)
4851 		return 0;
4852 
4853 	ret = get_dip_ctx_idx(ibqp, attr, &dip_idx);
4854 	if (ret) {
4855 		ibdev_err(ibdev, "failed to fill cong field, ret = %d.\n", ret);
4856 		return ret;
4857 	}
4858 
4859 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX, dip_idx);
4860 	hr_reg_write(&qpc_mask->ext, QPCEX_DIP_CTX_IDX, 0);
4861 
4862 	return 0;
4863 }
4864 
4865 static int hns_roce_hw_v2_get_dscp(struct hns_roce_dev *hr_dev, u8 dscp,
4866 				   u8 *tc_mode, u8 *priority)
4867 {
4868 	struct hns_roce_v2_priv *priv = hr_dev->priv;
4869 	struct hnae3_handle *handle = priv->handle;
4870 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
4871 
4872 	if (!ops->get_dscp_prio)
4873 		return -EOPNOTSUPP;
4874 
4875 	return ops->get_dscp_prio(handle, dscp, tc_mode, priority);
4876 }
4877 
4878 bool check_sl_valid(struct hns_roce_dev *hr_dev, u8 sl)
4879 {
4880 	u32 max_sl;
4881 
4882 	max_sl = min_t(u32, MAX_SERVICE_LEVEL, hr_dev->caps.sl_num - 1);
4883 	if (unlikely(sl > max_sl)) {
4884 		ibdev_err_ratelimited(&hr_dev->ib_dev,
4885 				      "failed to set SL(%u). Shouldn't be larger than %u.\n",
4886 				      sl, max_sl);
4887 		return false;
4888 	}
4889 
4890 	return true;
4891 }
4892 
4893 static int hns_roce_set_sl(struct ib_qp *ibqp,
4894 			   const struct ib_qp_attr *attr,
4895 			   struct hns_roce_v2_qp_context *context,
4896 			   struct hns_roce_v2_qp_context *qpc_mask)
4897 {
4898 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4899 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4900 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4901 	struct ib_device *ibdev = &hr_dev->ib_dev;
4902 	int ret;
4903 
4904 	ret = hns_roce_hw_v2_get_dscp(hr_dev, get_tclass(&attr->ah_attr.grh),
4905 				      &hr_qp->tc_mode, &hr_qp->priority);
4906 	if (ret && ret != -EOPNOTSUPP &&
4907 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4908 		ibdev_err_ratelimited(ibdev,
4909 				      "failed to get dscp, ret = %d.\n", ret);
4910 		return ret;
4911 	}
4912 
4913 	if (hr_qp->tc_mode == HNAE3_TC_MAP_MODE_DSCP &&
4914 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4915 		hr_qp->sl = hr_qp->priority;
4916 	else
4917 		hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4918 
4919 	if (!check_sl_valid(hr_dev, hr_qp->sl))
4920 		return -EINVAL;
4921 
4922 	hr_reg_write(context, QPC_SL, hr_qp->sl);
4923 	hr_reg_clear(qpc_mask, QPC_SL);
4924 
4925 	return 0;
4926 }
4927 
4928 static int hns_roce_v2_set_path(struct ib_qp *ibqp,
4929 				const struct ib_qp_attr *attr,
4930 				int attr_mask,
4931 				struct hns_roce_v2_qp_context *context,
4932 				struct hns_roce_v2_qp_context *qpc_mask)
4933 {
4934 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4935 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4936 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4937 	struct ib_device *ibdev = &hr_dev->ib_dev;
4938 	const struct ib_gid_attr *gid_attr = NULL;
4939 	u8 sl = rdma_ah_get_sl(&attr->ah_attr);
4940 	int is_roce_protocol;
4941 	u16 vlan_id = 0xffff;
4942 	bool is_udp = false;
4943 	u8 ib_port;
4944 	u8 hr_port;
4945 	int ret;
4946 
4947 	/*
4948 	 * If free_mr_en of qp is set, it means that this qp comes from
4949 	 * free mr. This qp will perform the loopback operation.
4950 	 * In the loopback scenario, only sl needs to be set.
4951 	 */
4952 	if (hr_qp->free_mr_en) {
4953 		if (!check_sl_valid(hr_dev, sl))
4954 			return -EINVAL;
4955 		hr_reg_write(context, QPC_SL, sl);
4956 		hr_reg_clear(qpc_mask, QPC_SL);
4957 		hr_qp->sl = sl;
4958 		return 0;
4959 	}
4960 
4961 	ib_port = (attr_mask & IB_QP_PORT) ? attr->port_num : hr_qp->port + 1;
4962 	hr_port = ib_port - 1;
4963 	is_roce_protocol = rdma_cap_eth_ah(&hr_dev->ib_dev, ib_port) &&
4964 			   rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
4965 
4966 	if (is_roce_protocol) {
4967 		gid_attr = attr->ah_attr.grh.sgid_attr;
4968 		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
4969 		if (ret)
4970 			return ret;
4971 
4972 		is_udp = (gid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
4973 	}
4974 
4975 	/* Only HIP08 needs to set the vlan_en bits in QPC */
4976 	if (vlan_id < VLAN_N_VID &&
4977 	    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4978 		hr_reg_enable(context, QPC_RQ_VLAN_EN);
4979 		hr_reg_clear(qpc_mask, QPC_RQ_VLAN_EN);
4980 		hr_reg_enable(context, QPC_SQ_VLAN_EN);
4981 		hr_reg_clear(qpc_mask, QPC_SQ_VLAN_EN);
4982 	}
4983 
4984 	hr_reg_write(context, QPC_VLAN_ID, vlan_id);
4985 	hr_reg_clear(qpc_mask, QPC_VLAN_ID);
4986 
4987 	if (grh->sgid_index >= hr_dev->caps.gid_table_len[hr_port]) {
4988 		ibdev_err(ibdev, "sgid_index(%u) too large. max is %d\n",
4989 			  grh->sgid_index, hr_dev->caps.gid_table_len[hr_port]);
4990 		return -EINVAL;
4991 	}
4992 
4993 	if (attr->ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE) {
4994 		ibdev_err(ibdev, "ah attr is not RDMA roce type\n");
4995 		return -EINVAL;
4996 	}
4997 
4998 	hr_reg_write(context, QPC_UDPSPN,
4999 		     is_udp ? rdma_get_udp_sport(grh->flow_label, ibqp->qp_num,
5000 						 attr->dest_qp_num) :
5001 				    0);
5002 
5003 	hr_reg_clear(qpc_mask, QPC_UDPSPN);
5004 
5005 	hr_reg_write(context, QPC_GMV_IDX, grh->sgid_index);
5006 
5007 	hr_reg_clear(qpc_mask, QPC_GMV_IDX);
5008 
5009 	hr_reg_write(context, QPC_HOPLIMIT, grh->hop_limit);
5010 	hr_reg_clear(qpc_mask, QPC_HOPLIMIT);
5011 
5012 	ret = fill_cong_field(ibqp, attr, context, qpc_mask);
5013 	if (ret)
5014 		return ret;
5015 
5016 	hr_reg_write(context, QPC_TC, get_tclass(&attr->ah_attr.grh));
5017 	hr_reg_clear(qpc_mask, QPC_TC);
5018 
5019 	hr_reg_write(context, QPC_FL, grh->flow_label);
5020 	hr_reg_clear(qpc_mask, QPC_FL);
5021 	memcpy(context->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
5022 	memset(qpc_mask->dgid, 0, sizeof(grh->dgid.raw));
5023 
5024 	return  hns_roce_set_sl(ibqp, attr, context, qpc_mask);
5025 }
5026 
5027 static bool check_qp_state(enum ib_qp_state cur_state,
5028 			   enum ib_qp_state new_state)
5029 {
5030 	static const bool sm[][IB_QPS_ERR + 1] = {
5031 		[IB_QPS_RESET] = { [IB_QPS_RESET] = true,
5032 				   [IB_QPS_INIT] = true },
5033 		[IB_QPS_INIT] = { [IB_QPS_RESET] = true,
5034 				  [IB_QPS_INIT] = true,
5035 				  [IB_QPS_RTR] = true,
5036 				  [IB_QPS_ERR] = true },
5037 		[IB_QPS_RTR] = { [IB_QPS_RESET] = true,
5038 				 [IB_QPS_RTS] = true,
5039 				 [IB_QPS_ERR] = true },
5040 		[IB_QPS_RTS] = { [IB_QPS_RESET] = true,
5041 				 [IB_QPS_RTS] = true,
5042 				 [IB_QPS_ERR] = true },
5043 		[IB_QPS_SQD] = {},
5044 		[IB_QPS_SQE] = {},
5045 		[IB_QPS_ERR] = { [IB_QPS_RESET] = true,
5046 				 [IB_QPS_ERR] = true }
5047 	};
5048 
5049 	return sm[cur_state][new_state];
5050 }
5051 
5052 static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
5053 				      const struct ib_qp_attr *attr,
5054 				      int attr_mask,
5055 				      enum ib_qp_state cur_state,
5056 				      enum ib_qp_state new_state,
5057 				      struct hns_roce_v2_qp_context *context,
5058 				      struct hns_roce_v2_qp_context *qpc_mask,
5059 				      struct ib_udata *udata)
5060 {
5061 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5062 	int ret = 0;
5063 
5064 	if (!check_qp_state(cur_state, new_state)) {
5065 		ibdev_err(&hr_dev->ib_dev, "Illegal state for QP!\n");
5066 		return -EINVAL;
5067 	}
5068 
5069 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
5070 		memset(qpc_mask, 0, hr_dev->caps.qpc_sz);
5071 		modify_qp_reset_to_init(ibqp, context, qpc_mask);
5072 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
5073 		modify_qp_init_to_init(ibqp, context, qpc_mask);
5074 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
5075 		ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
5076 					    qpc_mask, udata);
5077 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
5078 		ret = modify_qp_rtr_to_rts(ibqp, attr_mask, context, qpc_mask);
5079 	}
5080 
5081 	return ret;
5082 }
5083 
5084 static bool check_qp_timeout_cfg_range(struct hns_roce_dev *hr_dev, u8 *timeout)
5085 {
5086 #define QP_ACK_TIMEOUT_MAX_HIP08 20
5087 #define QP_ACK_TIMEOUT_MAX 31
5088 
5089 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5090 		if (*timeout > QP_ACK_TIMEOUT_MAX_HIP08) {
5091 			ibdev_warn(&hr_dev->ib_dev,
5092 				   "local ACK timeout shall be 0 to 20.\n");
5093 			return false;
5094 		}
5095 		*timeout += HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5096 	} else if (hr_dev->pci_dev->revision > PCI_REVISION_ID_HIP08) {
5097 		if (*timeout > QP_ACK_TIMEOUT_MAX) {
5098 			ibdev_warn(&hr_dev->ib_dev,
5099 				   "local ACK timeout shall be 0 to 31.\n");
5100 			return false;
5101 		}
5102 	}
5103 
5104 	return true;
5105 }
5106 
5107 static int hns_roce_v2_set_opt_fields(struct ib_qp *ibqp,
5108 				      const struct ib_qp_attr *attr,
5109 				      int attr_mask,
5110 				      struct hns_roce_v2_qp_context *context,
5111 				      struct hns_roce_v2_qp_context *qpc_mask)
5112 {
5113 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5114 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5115 	int ret = 0;
5116 	u8 timeout;
5117 
5118 	if (attr_mask & IB_QP_AV) {
5119 		ret = hns_roce_v2_set_path(ibqp, attr, attr_mask, context,
5120 					   qpc_mask);
5121 		if (ret)
5122 			return ret;
5123 	}
5124 
5125 	if (attr_mask & IB_QP_TIMEOUT) {
5126 		timeout = attr->timeout;
5127 		if (check_qp_timeout_cfg_range(hr_dev, &timeout)) {
5128 			hr_reg_write(context, QPC_AT, timeout);
5129 			hr_reg_clear(qpc_mask, QPC_AT);
5130 		}
5131 	}
5132 
5133 	if (attr_mask & IB_QP_RETRY_CNT) {
5134 		hr_reg_write(context, QPC_RETRY_NUM_INIT, attr->retry_cnt);
5135 		hr_reg_clear(qpc_mask, QPC_RETRY_NUM_INIT);
5136 
5137 		hr_reg_write(context, QPC_RETRY_CNT, attr->retry_cnt);
5138 		hr_reg_clear(qpc_mask, QPC_RETRY_CNT);
5139 	}
5140 
5141 	if (attr_mask & IB_QP_RNR_RETRY) {
5142 		hr_reg_write(context, QPC_RNR_NUM_INIT, attr->rnr_retry);
5143 		hr_reg_clear(qpc_mask, QPC_RNR_NUM_INIT);
5144 
5145 		hr_reg_write(context, QPC_RNR_CNT, attr->rnr_retry);
5146 		hr_reg_clear(qpc_mask, QPC_RNR_CNT);
5147 	}
5148 
5149 	if (attr_mask & IB_QP_SQ_PSN) {
5150 		hr_reg_write(context, QPC_SQ_CUR_PSN, attr->sq_psn);
5151 		hr_reg_clear(qpc_mask, QPC_SQ_CUR_PSN);
5152 
5153 		hr_reg_write(context, QPC_SQ_MAX_PSN, attr->sq_psn);
5154 		hr_reg_clear(qpc_mask, QPC_SQ_MAX_PSN);
5155 
5156 		hr_reg_write(context, QPC_RETRY_MSG_PSN_L, attr->sq_psn);
5157 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_L);
5158 
5159 		hr_reg_write(context, QPC_RETRY_MSG_PSN_H,
5160 			     attr->sq_psn >> RETRY_MSG_PSN_SHIFT);
5161 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_H);
5162 
5163 		hr_reg_write(context, QPC_RETRY_MSG_FPKT_PSN, attr->sq_psn);
5164 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_FPKT_PSN);
5165 
5166 		hr_reg_write(context, QPC_RX_ACK_EPSN, attr->sq_psn);
5167 		hr_reg_clear(qpc_mask, QPC_RX_ACK_EPSN);
5168 	}
5169 
5170 	if ((attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) &&
5171 	     attr->max_dest_rd_atomic) {
5172 		hr_reg_write(context, QPC_RR_MAX,
5173 			     fls(attr->max_dest_rd_atomic - 1));
5174 		hr_reg_clear(qpc_mask, QPC_RR_MAX);
5175 	}
5176 
5177 	if ((attr_mask & IB_QP_MAX_QP_RD_ATOMIC) && attr->max_rd_atomic) {
5178 		hr_reg_write(context, QPC_SR_MAX, fls(attr->max_rd_atomic - 1));
5179 		hr_reg_clear(qpc_mask, QPC_SR_MAX);
5180 	}
5181 
5182 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
5183 		set_access_flags(hr_qp, context, qpc_mask, attr, attr_mask);
5184 
5185 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
5186 		hr_reg_write(context, QPC_MIN_RNR_TIME,
5187 			    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ?
5188 			    HNS_ROCE_RNR_TIMER_10NS : attr->min_rnr_timer);
5189 		hr_reg_clear(qpc_mask, QPC_MIN_RNR_TIME);
5190 	}
5191 
5192 	if (attr_mask & IB_QP_RQ_PSN) {
5193 		hr_reg_write(context, QPC_RX_REQ_EPSN, attr->rq_psn);
5194 		hr_reg_clear(qpc_mask, QPC_RX_REQ_EPSN);
5195 
5196 		hr_reg_write(context, QPC_RAQ_PSN, attr->rq_psn - 1);
5197 		hr_reg_clear(qpc_mask, QPC_RAQ_PSN);
5198 	}
5199 
5200 	if (attr_mask & IB_QP_QKEY) {
5201 		context->qkey_xrcd = cpu_to_le32(attr->qkey);
5202 		qpc_mask->qkey_xrcd = 0;
5203 		hr_qp->qkey = attr->qkey;
5204 	}
5205 
5206 	return ret;
5207 }
5208 
5209 static void hns_roce_v2_record_opt_fields(struct ib_qp *ibqp,
5210 					  const struct ib_qp_attr *attr,
5211 					  int attr_mask)
5212 {
5213 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5214 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5215 
5216 	if (attr_mask & IB_QP_ACCESS_FLAGS)
5217 		hr_qp->atomic_rd_en = attr->qp_access_flags;
5218 
5219 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
5220 		hr_qp->resp_depth = attr->max_dest_rd_atomic;
5221 	if (attr_mask & IB_QP_PORT) {
5222 		hr_qp->port = attr->port_num - 1;
5223 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
5224 	}
5225 }
5226 
5227 static void clear_qp(struct hns_roce_qp *hr_qp)
5228 {
5229 	struct ib_qp *ibqp = &hr_qp->ibqp;
5230 
5231 	if (ibqp->send_cq)
5232 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq),
5233 				     hr_qp->qpn, NULL);
5234 
5235 	if (ibqp->recv_cq  && ibqp->recv_cq != ibqp->send_cq)
5236 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->recv_cq),
5237 				     hr_qp->qpn, ibqp->srq ?
5238 				     to_hr_srq(ibqp->srq) : NULL);
5239 
5240 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
5241 		*hr_qp->rdb.db_record = 0;
5242 
5243 	hr_qp->rq.head = 0;
5244 	hr_qp->rq.tail = 0;
5245 	hr_qp->sq.head = 0;
5246 	hr_qp->sq.tail = 0;
5247 	hr_qp->next_sge = 0;
5248 }
5249 
5250 static void v2_set_flushed_fields(struct ib_qp *ibqp,
5251 				  struct hns_roce_v2_qp_context *context,
5252 				  struct hns_roce_v2_qp_context *qpc_mask)
5253 {
5254 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5255 	unsigned long sq_flag = 0;
5256 	unsigned long rq_flag = 0;
5257 
5258 	if (ibqp->qp_type == IB_QPT_XRC_TGT)
5259 		return;
5260 
5261 	spin_lock_irqsave(&hr_qp->sq.lock, sq_flag);
5262 	hr_reg_write(context, QPC_SQ_PRODUCER_IDX, hr_qp->sq.head);
5263 	hr_reg_clear(qpc_mask, QPC_SQ_PRODUCER_IDX);
5264 	hr_qp->state = IB_QPS_ERR;
5265 	spin_unlock_irqrestore(&hr_qp->sq.lock, sq_flag);
5266 
5267 	if (ibqp->srq || ibqp->qp_type == IB_QPT_XRC_INI) /* no RQ */
5268 		return;
5269 
5270 	spin_lock_irqsave(&hr_qp->rq.lock, rq_flag);
5271 	hr_reg_write(context, QPC_RQ_PRODUCER_IDX, hr_qp->rq.head);
5272 	hr_reg_clear(qpc_mask, QPC_RQ_PRODUCER_IDX);
5273 	spin_unlock_irqrestore(&hr_qp->rq.lock, rq_flag);
5274 }
5275 
5276 static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
5277 				 const struct ib_qp_attr *attr,
5278 				 int attr_mask, enum ib_qp_state cur_state,
5279 				 enum ib_qp_state new_state, struct ib_udata *udata)
5280 {
5281 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5282 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5283 	struct hns_roce_v2_qp_context ctx[2];
5284 	struct hns_roce_v2_qp_context *context = ctx;
5285 	struct hns_roce_v2_qp_context *qpc_mask = ctx + 1;
5286 	struct ib_device *ibdev = &hr_dev->ib_dev;
5287 	int ret;
5288 
5289 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
5290 		return -EOPNOTSUPP;
5291 
5292 	/*
5293 	 * In v2 engine, software pass context and context mask to hardware
5294 	 * when modifying qp. If software need modify some fields in context,
5295 	 * we should set all bits of the relevant fields in context mask to
5296 	 * 0 at the same time, else set them to 0x1.
5297 	 */
5298 	memset(context, 0, hr_dev->caps.qpc_sz);
5299 	memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
5300 
5301 	ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
5302 					 new_state, context, qpc_mask, udata);
5303 	if (ret)
5304 		goto out;
5305 
5306 	/* When QP state is err, SQ and RQ WQE should be flushed */
5307 	if (new_state == IB_QPS_ERR)
5308 		v2_set_flushed_fields(ibqp, context, qpc_mask);
5309 
5310 	/* Configure the optional fields */
5311 	ret = hns_roce_v2_set_opt_fields(ibqp, attr, attr_mask, context,
5312 					 qpc_mask);
5313 	if (ret)
5314 		goto out;
5315 
5316 	hr_reg_write_bool(context, QPC_INV_CREDIT,
5317 			  to_hr_qp_type(hr_qp->ibqp.qp_type) == SERV_TYPE_XRC ||
5318 			  ibqp->srq);
5319 	hr_reg_clear(qpc_mask, QPC_INV_CREDIT);
5320 
5321 	/* Every status migrate must change state */
5322 	hr_reg_write(context, QPC_QP_ST, new_state);
5323 	hr_reg_clear(qpc_mask, QPC_QP_ST);
5324 
5325 	/* SW pass context to HW */
5326 	ret = hns_roce_v2_qp_modify(hr_dev, context, qpc_mask, hr_qp);
5327 	if (ret) {
5328 		ibdev_err(ibdev, "failed to modify QP, ret = %d.\n", ret);
5329 		goto out;
5330 	}
5331 
5332 	hr_qp->state = new_state;
5333 
5334 	hns_roce_v2_record_opt_fields(ibqp, attr, attr_mask);
5335 
5336 	if (new_state == IB_QPS_RESET && !ibqp->uobject)
5337 		clear_qp(hr_qp);
5338 
5339 out:
5340 	return ret;
5341 }
5342 
5343 static int to_ib_qp_st(enum hns_roce_v2_qp_state state)
5344 {
5345 	static const enum ib_qp_state map[] = {
5346 		[HNS_ROCE_QP_ST_RST] = IB_QPS_RESET,
5347 		[HNS_ROCE_QP_ST_INIT] = IB_QPS_INIT,
5348 		[HNS_ROCE_QP_ST_RTR] = IB_QPS_RTR,
5349 		[HNS_ROCE_QP_ST_RTS] = IB_QPS_RTS,
5350 		[HNS_ROCE_QP_ST_SQD] = IB_QPS_SQD,
5351 		[HNS_ROCE_QP_ST_SQER] = IB_QPS_SQE,
5352 		[HNS_ROCE_QP_ST_ERR] = IB_QPS_ERR,
5353 		[HNS_ROCE_QP_ST_SQ_DRAINING] = IB_QPS_SQD
5354 	};
5355 
5356 	return (state < ARRAY_SIZE(map)) ? map[state] : -1;
5357 }
5358 
5359 static int hns_roce_v2_query_qpc(struct hns_roce_dev *hr_dev, u32 qpn,
5360 				 void *buffer)
5361 {
5362 	struct hns_roce_cmd_mailbox *mailbox;
5363 	int ret;
5364 
5365 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5366 	if (IS_ERR(mailbox))
5367 		return PTR_ERR(mailbox);
5368 
5369 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_QPC,
5370 				qpn);
5371 	if (ret)
5372 		goto out;
5373 
5374 	memcpy(buffer, mailbox->buf, hr_dev->caps.qpc_sz);
5375 
5376 out:
5377 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5378 	return ret;
5379 }
5380 
5381 static int hns_roce_v2_query_srqc(struct hns_roce_dev *hr_dev, u32 srqn,
5382 				 void *buffer)
5383 {
5384 	struct hns_roce_srq_context *context;
5385 	struct hns_roce_cmd_mailbox *mailbox;
5386 	int ret;
5387 
5388 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5389 	if (IS_ERR(mailbox))
5390 		return PTR_ERR(mailbox);
5391 
5392 	context = mailbox->buf;
5393 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SRQC,
5394 				srqn);
5395 	if (ret)
5396 		goto out;
5397 
5398 	memcpy(buffer, context, sizeof(*context));
5399 
5400 out:
5401 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5402 	return ret;
5403 }
5404 
5405 static int hns_roce_v2_query_sccc(struct hns_roce_dev *hr_dev, u32 qpn,
5406 				  void *buffer)
5407 {
5408 	struct hns_roce_v2_scc_context *context;
5409 	struct hns_roce_cmd_mailbox *mailbox;
5410 	int ret;
5411 
5412 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5413 	if (IS_ERR(mailbox))
5414 		return PTR_ERR(mailbox);
5415 
5416 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SCCC,
5417 				qpn);
5418 	if (ret)
5419 		goto out;
5420 
5421 	context = mailbox->buf;
5422 	memcpy(buffer, context, sizeof(*context));
5423 
5424 out:
5425 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5426 	return ret;
5427 }
5428 
5429 static u8 get_qp_timeout_attr(struct hns_roce_dev *hr_dev,
5430 			      struct hns_roce_v2_qp_context *context)
5431 {
5432 	u8 timeout;
5433 
5434 	timeout = (u8)hr_reg_read(context, QPC_AT);
5435 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
5436 		timeout -= HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5437 
5438 	return timeout;
5439 }
5440 
5441 static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5442 				int qp_attr_mask,
5443 				struct ib_qp_init_attr *qp_init_attr)
5444 {
5445 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5446 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5447 	struct hns_roce_v2_qp_context context = {};
5448 	struct ib_device *ibdev = &hr_dev->ib_dev;
5449 	int tmp_qp_state;
5450 	int state;
5451 	int ret;
5452 
5453 	memset(qp_attr, 0, sizeof(*qp_attr));
5454 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5455 
5456 	mutex_lock(&hr_qp->mutex);
5457 
5458 	if (hr_qp->state == IB_QPS_RESET) {
5459 		qp_attr->qp_state = IB_QPS_RESET;
5460 		ret = 0;
5461 		goto done;
5462 	}
5463 
5464 	ret = hns_roce_v2_query_qpc(hr_dev, hr_qp->qpn, &context);
5465 	if (ret) {
5466 		ibdev_err(ibdev, "failed to query QPC, ret = %d.\n", ret);
5467 		ret = -EINVAL;
5468 		goto out;
5469 	}
5470 
5471 	state = hr_reg_read(&context, QPC_QP_ST);
5472 	tmp_qp_state = to_ib_qp_st((enum hns_roce_v2_qp_state)state);
5473 	if (tmp_qp_state == -1) {
5474 		ibdev_err(ibdev, "Illegal ib_qp_state\n");
5475 		ret = -EINVAL;
5476 		goto out;
5477 	}
5478 	hr_qp->state = (u8)tmp_qp_state;
5479 	qp_attr->qp_state = (enum ib_qp_state)hr_qp->state;
5480 	qp_attr->path_mtu = (enum ib_mtu)hr_reg_read(&context, QPC_MTU);
5481 	qp_attr->path_mig_state = IB_MIG_ARMED;
5482 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
5483 	if (hr_qp->ibqp.qp_type == IB_QPT_UD)
5484 		qp_attr->qkey = le32_to_cpu(context.qkey_xrcd);
5485 
5486 	qp_attr->rq_psn = hr_reg_read(&context, QPC_RX_REQ_EPSN);
5487 	qp_attr->sq_psn = (u32)hr_reg_read(&context, QPC_SQ_CUR_PSN);
5488 	qp_attr->dest_qp_num = hr_reg_read(&context, QPC_DQPN);
5489 	qp_attr->qp_access_flags =
5490 		((hr_reg_read(&context, QPC_RRE)) << V2_QP_RRE_S) |
5491 		((hr_reg_read(&context, QPC_RWE)) << V2_QP_RWE_S) |
5492 		((hr_reg_read(&context, QPC_ATE)) << V2_QP_ATE_S);
5493 
5494 	if (hr_qp->ibqp.qp_type == IB_QPT_RC ||
5495 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5496 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
5497 		struct ib_global_route *grh =
5498 			rdma_ah_retrieve_grh(&qp_attr->ah_attr);
5499 
5500 		rdma_ah_set_sl(&qp_attr->ah_attr,
5501 			       hr_reg_read(&context, QPC_SL));
5502 		rdma_ah_set_port_num(&qp_attr->ah_attr, hr_qp->port + 1);
5503 		rdma_ah_set_ah_flags(&qp_attr->ah_attr, IB_AH_GRH);
5504 		grh->flow_label = hr_reg_read(&context, QPC_FL);
5505 		grh->sgid_index = hr_reg_read(&context, QPC_GMV_IDX);
5506 		grh->hop_limit = hr_reg_read(&context, QPC_HOPLIMIT);
5507 		grh->traffic_class = hr_reg_read(&context, QPC_TC);
5508 
5509 		memcpy(grh->dgid.raw, context.dgid, sizeof(grh->dgid.raw));
5510 	}
5511 
5512 	qp_attr->port_num = hr_qp->port + 1;
5513 	qp_attr->sq_draining = 0;
5514 	qp_attr->max_rd_atomic = 1 << hr_reg_read(&context, QPC_SR_MAX);
5515 	qp_attr->max_dest_rd_atomic = 1 << hr_reg_read(&context, QPC_RR_MAX);
5516 
5517 	qp_attr->min_rnr_timer = (u8)hr_reg_read(&context, QPC_MIN_RNR_TIME);
5518 	qp_attr->timeout = get_qp_timeout_attr(hr_dev, &context);
5519 	qp_attr->retry_cnt = hr_reg_read(&context, QPC_RETRY_NUM_INIT);
5520 	qp_attr->rnr_retry = hr_reg_read(&context, QPC_RNR_NUM_INIT);
5521 
5522 done:
5523 	qp_attr->cur_qp_state = qp_attr->qp_state;
5524 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
5525 	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
5526 	qp_attr->cap.max_inline_data = hr_qp->max_inline_data;
5527 
5528 	qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
5529 	qp_attr->cap.max_send_sge = hr_qp->sq.max_gs;
5530 
5531 	qp_init_attr->qp_context = ibqp->qp_context;
5532 	qp_init_attr->qp_type = ibqp->qp_type;
5533 	qp_init_attr->recv_cq = ibqp->recv_cq;
5534 	qp_init_attr->send_cq = ibqp->send_cq;
5535 	qp_init_attr->srq = ibqp->srq;
5536 	qp_init_attr->cap = qp_attr->cap;
5537 	qp_init_attr->sq_sig_type = hr_qp->sq_signal_bits;
5538 
5539 out:
5540 	mutex_unlock(&hr_qp->mutex);
5541 	return ret;
5542 }
5543 
5544 static inline int modify_qp_is_ok(struct hns_roce_qp *hr_qp)
5545 {
5546 	return ((hr_qp->ibqp.qp_type == IB_QPT_RC ||
5547 		 hr_qp->ibqp.qp_type == IB_QPT_UD ||
5548 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5549 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
5550 		hr_qp->state != IB_QPS_RESET);
5551 }
5552 
5553 static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
5554 					 struct hns_roce_qp *hr_qp,
5555 					 struct ib_udata *udata)
5556 {
5557 	struct ib_device *ibdev = &hr_dev->ib_dev;
5558 	struct hns_roce_cq *send_cq, *recv_cq;
5559 	unsigned long flags;
5560 	int ret = 0;
5561 
5562 	if (modify_qp_is_ok(hr_qp)) {
5563 		/* Modify qp to reset before destroying qp */
5564 		ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
5565 					    hr_qp->state, IB_QPS_RESET, udata);
5566 		if (ret)
5567 			ibdev_err(ibdev,
5568 				  "failed to modify QP to RST, ret = %d.\n",
5569 				  ret);
5570 	}
5571 
5572 	send_cq = hr_qp->ibqp.send_cq ? to_hr_cq(hr_qp->ibqp.send_cq) : NULL;
5573 	recv_cq = hr_qp->ibqp.recv_cq ? to_hr_cq(hr_qp->ibqp.recv_cq) : NULL;
5574 
5575 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
5576 	hns_roce_lock_cqs(send_cq, recv_cq);
5577 
5578 	if (!udata) {
5579 		if (recv_cq)
5580 			__hns_roce_v2_cq_clean(recv_cq, hr_qp->qpn,
5581 					       (hr_qp->ibqp.srq ?
5582 						to_hr_srq(hr_qp->ibqp.srq) :
5583 						NULL));
5584 
5585 		if (send_cq && send_cq != recv_cq)
5586 			__hns_roce_v2_cq_clean(send_cq, hr_qp->qpn, NULL);
5587 	}
5588 
5589 	hns_roce_qp_remove(hr_dev, hr_qp);
5590 
5591 	hns_roce_unlock_cqs(send_cq, recv_cq);
5592 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
5593 
5594 	return ret;
5595 }
5596 
5597 int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
5598 {
5599 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5600 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5601 	int ret;
5602 
5603 	ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata);
5604 	if (ret)
5605 		ibdev_err(&hr_dev->ib_dev,
5606 			  "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n",
5607 			  hr_qp->qpn, ret);
5608 
5609 	hns_roce_qp_destroy(hr_dev, hr_qp, udata);
5610 
5611 	return 0;
5612 }
5613 
5614 static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
5615 					    struct hns_roce_qp *hr_qp)
5616 {
5617 	struct ib_device *ibdev = &hr_dev->ib_dev;
5618 	struct hns_roce_sccc_clr_done *resp;
5619 	struct hns_roce_sccc_clr *clr;
5620 	struct hns_roce_cmq_desc desc;
5621 	int ret, i;
5622 
5623 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
5624 		return 0;
5625 
5626 	mutex_lock(&hr_dev->qp_table.scc_mutex);
5627 
5628 	/* set scc ctx clear done flag */
5629 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_RESET_SCCC, false);
5630 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5631 	if (ret) {
5632 		ibdev_err(ibdev, "failed to reset SCC ctx, ret = %d.\n", ret);
5633 		goto out;
5634 	}
5635 
5636 	/* clear scc context */
5637 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLR_SCCC, false);
5638 	clr = (struct hns_roce_sccc_clr *)desc.data;
5639 	clr->qpn = cpu_to_le32(hr_qp->qpn);
5640 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5641 	if (ret) {
5642 		ibdev_err(ibdev, "failed to clear SCC ctx, ret = %d.\n", ret);
5643 		goto out;
5644 	}
5645 
5646 	/* query scc context clear is done or not */
5647 	resp = (struct hns_roce_sccc_clr_done *)desc.data;
5648 	for (i = 0; i <= HNS_ROCE_CMQ_SCC_CLR_DONE_CNT; i++) {
5649 		hns_roce_cmq_setup_basic_desc(&desc,
5650 					      HNS_ROCE_OPC_QUERY_SCCC, true);
5651 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
5652 		if (ret) {
5653 			ibdev_err(ibdev, "failed to query clr cmq, ret = %d\n",
5654 				  ret);
5655 			goto out;
5656 		}
5657 
5658 		if (resp->clr_done)
5659 			goto out;
5660 
5661 		msleep(20);
5662 	}
5663 
5664 	ibdev_err(ibdev, "query SCC clr done flag overtime.\n");
5665 	ret = -ETIMEDOUT;
5666 
5667 out:
5668 	mutex_unlock(&hr_dev->qp_table.scc_mutex);
5669 	return ret;
5670 }
5671 
5672 #define DMA_IDX_SHIFT 3
5673 #define DMA_WQE_SHIFT 3
5674 
5675 static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
5676 					      struct hns_roce_srq_context *ctx)
5677 {
5678 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
5679 	struct ib_device *ibdev = srq->ibsrq.device;
5680 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5681 	u64 mtts_idx[MTT_MIN_COUNT] = {};
5682 	dma_addr_t dma_handle_idx;
5683 	int ret;
5684 
5685 	/* Get physical address of idx que buf */
5686 	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
5687 				ARRAY_SIZE(mtts_idx));
5688 	if (ret) {
5689 		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
5690 			  ret);
5691 		return ret;
5692 	}
5693 
5694 	dma_handle_idx = hns_roce_get_mtr_ba(&idx_que->mtr);
5695 
5696 	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
5697 		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
5698 
5699 	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> DMA_IDX_SHIFT);
5700 	hr_reg_write(ctx, SRQC_IDX_BT_BA_H,
5701 		     upper_32_bits(dma_handle_idx >> DMA_IDX_SHIFT));
5702 
5703 	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
5704 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
5705 	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
5706 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
5707 
5708 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
5709 		     to_hr_hw_page_addr(mtts_idx[0]));
5710 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
5711 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
5712 
5713 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
5714 		     to_hr_hw_page_addr(mtts_idx[1]));
5715 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
5716 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
5717 
5718 	return 0;
5719 }
5720 
5721 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
5722 {
5723 	struct ib_device *ibdev = srq->ibsrq.device;
5724 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5725 	struct hns_roce_srq_context *ctx = mb_buf;
5726 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
5727 	dma_addr_t dma_handle_wqe;
5728 	int ret;
5729 
5730 	memset(ctx, 0, sizeof(*ctx));
5731 
5732 	/* Get the physical address of srq buf */
5733 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
5734 				ARRAY_SIZE(mtts_wqe));
5735 	if (ret) {
5736 		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
5737 			  ret);
5738 		return ret;
5739 	}
5740 
5741 	dma_handle_wqe = hns_roce_get_mtr_ba(&srq->buf_mtr);
5742 
5743 	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
5744 	hr_reg_write_bool(ctx, SRQC_SRQ_TYPE,
5745 			  srq->ibsrq.srq_type == IB_SRQT_XRC);
5746 	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
5747 	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
5748 	hr_reg_write(ctx, SRQC_XRCD, srq->xrcdn);
5749 	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
5750 	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
5751 	hr_reg_write(ctx, SRQC_RQWS,
5752 		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
5753 
5754 	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
5755 		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
5756 				      srq->wqe_cnt));
5757 
5758 	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> DMA_WQE_SHIFT);
5759 	hr_reg_write(ctx, SRQC_WQE_BT_BA_H,
5760 		     upper_32_bits(dma_handle_wqe >> DMA_WQE_SHIFT));
5761 
5762 	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
5763 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
5764 	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
5765 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
5766 
5767 	if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB) {
5768 		hr_reg_enable(ctx, SRQC_DB_RECORD_EN);
5769 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_L,
5770 			     lower_32_bits(srq->rdb.dma) >> 1);
5771 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_H,
5772 			     upper_32_bits(srq->rdb.dma));
5773 	}
5774 
5775 	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
5776 }
5777 
5778 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
5779 				  struct ib_srq_attr *srq_attr,
5780 				  enum ib_srq_attr_mask srq_attr_mask,
5781 				  struct ib_udata *udata)
5782 {
5783 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5784 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5785 	struct hns_roce_srq_context *srq_context;
5786 	struct hns_roce_srq_context *srqc_mask;
5787 	struct hns_roce_cmd_mailbox *mailbox;
5788 	int ret = 0;
5789 
5790 	/* Resizing SRQs is not supported yet */
5791 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
5792 		ret = -EOPNOTSUPP;
5793 		goto out;
5794 	}
5795 
5796 	if (srq_attr_mask & IB_SRQ_LIMIT) {
5797 		if (srq_attr->srq_limit > srq->wqe_cnt) {
5798 			ret = -EINVAL;
5799 			goto out;
5800 		}
5801 
5802 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5803 		if (IS_ERR(mailbox)) {
5804 			ret = PTR_ERR(mailbox);
5805 			goto out;
5806 		}
5807 
5808 		srq_context = mailbox->buf;
5809 		srqc_mask = (struct hns_roce_srq_context *)mailbox->buf + 1;
5810 
5811 		memset(srqc_mask, 0xff, sizeof(*srqc_mask));
5812 
5813 		hr_reg_write(srq_context, SRQC_LIMIT_WL, srq_attr->srq_limit);
5814 		hr_reg_clear(srqc_mask, SRQC_LIMIT_WL);
5815 
5816 		ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5817 					HNS_ROCE_CMD_MODIFY_SRQC, srq->srqn);
5818 		hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5819 		if (ret)
5820 			ibdev_err(&hr_dev->ib_dev,
5821 				  "failed to handle cmd of modifying SRQ, ret = %d.\n",
5822 				  ret);
5823 	}
5824 
5825 out:
5826 	if (ret)
5827 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_SRQ_MODIFY_ERR_CNT]);
5828 
5829 	return ret;
5830 }
5831 
5832 static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
5833 {
5834 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5835 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5836 	struct hns_roce_srq_context *srq_context;
5837 	struct hns_roce_cmd_mailbox *mailbox;
5838 	int ret;
5839 
5840 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5841 	if (IS_ERR(mailbox))
5842 		return PTR_ERR(mailbox);
5843 
5844 	srq_context = mailbox->buf;
5845 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5846 				HNS_ROCE_CMD_QUERY_SRQC, srq->srqn);
5847 	if (ret) {
5848 		ibdev_err(&hr_dev->ib_dev,
5849 			  "failed to process cmd of querying SRQ, ret = %d.\n",
5850 			  ret);
5851 		goto out;
5852 	}
5853 
5854 	attr->srq_limit = hr_reg_read(srq_context, SRQC_LIMIT_WL);
5855 	attr->max_wr = srq->wqe_cnt;
5856 	attr->max_sge = srq->max_gs - srq->rsv_sge;
5857 
5858 out:
5859 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5860 	return ret;
5861 }
5862 
5863 static int hns_roce_v2_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
5864 {
5865 	struct hns_roce_dev *hr_dev = to_hr_dev(cq->device);
5866 	struct hns_roce_v2_cq_context *cq_context;
5867 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
5868 	struct hns_roce_v2_cq_context *cqc_mask;
5869 	struct hns_roce_cmd_mailbox *mailbox;
5870 	int ret;
5871 
5872 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5873 	ret = PTR_ERR_OR_ZERO(mailbox);
5874 	if (ret)
5875 		goto err_out;
5876 
5877 	cq_context = mailbox->buf;
5878 	cqc_mask = (struct hns_roce_v2_cq_context *)mailbox->buf + 1;
5879 
5880 	memset(cqc_mask, 0xff, sizeof(*cqc_mask));
5881 
5882 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT, cq_count);
5883 	hr_reg_clear(cqc_mask, CQC_CQ_MAX_CNT);
5884 
5885 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5886 		if (cq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
5887 			dev_info(hr_dev->dev,
5888 				 "cq_period(%u) reached the upper limit, adjusted to 65.\n",
5889 				 cq_period);
5890 			cq_period = HNS_ROCE_MAX_CQ_PERIOD_HIP08;
5891 		}
5892 		cq_period *= HNS_ROCE_CLOCK_ADJUST;
5893 	}
5894 	hr_reg_write(cq_context, CQC_CQ_PERIOD, cq_period);
5895 	hr_reg_clear(cqc_mask, CQC_CQ_PERIOD);
5896 
5897 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5898 				HNS_ROCE_CMD_MODIFY_CQC, hr_cq->cqn);
5899 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5900 	if (ret)
5901 		ibdev_err(&hr_dev->ib_dev,
5902 			  "failed to process cmd when modifying CQ, ret = %d.\n",
5903 			  ret);
5904 
5905 err_out:
5906 	if (ret)
5907 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CQ_MODIFY_ERR_CNT]);
5908 
5909 	return ret;
5910 }
5911 
5912 static int hns_roce_v2_query_cqc(struct hns_roce_dev *hr_dev, u32 cqn,
5913 				 void *buffer)
5914 {
5915 	struct hns_roce_v2_cq_context *context;
5916 	struct hns_roce_cmd_mailbox *mailbox;
5917 	int ret;
5918 
5919 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5920 	if (IS_ERR(mailbox))
5921 		return PTR_ERR(mailbox);
5922 
5923 	context = mailbox->buf;
5924 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5925 				HNS_ROCE_CMD_QUERY_CQC, cqn);
5926 	if (ret) {
5927 		ibdev_err(&hr_dev->ib_dev,
5928 			  "failed to process cmd when querying CQ, ret = %d.\n",
5929 			  ret);
5930 		goto err_mailbox;
5931 	}
5932 
5933 	memcpy(buffer, context, sizeof(*context));
5934 
5935 err_mailbox:
5936 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5937 
5938 	return ret;
5939 }
5940 
5941 static int hns_roce_v2_query_mpt(struct hns_roce_dev *hr_dev, u32 key,
5942 				 void *buffer)
5943 {
5944 	struct hns_roce_v2_mpt_entry *context;
5945 	struct hns_roce_cmd_mailbox *mailbox;
5946 	int ret;
5947 
5948 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5949 	if (IS_ERR(mailbox))
5950 		return PTR_ERR(mailbox);
5951 
5952 	context = mailbox->buf;
5953 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
5954 				key_to_hw_index(key));
5955 	if (ret) {
5956 		ibdev_err(&hr_dev->ib_dev,
5957 			  "failed to process cmd when querying MPT, ret = %d.\n",
5958 			  ret);
5959 		goto err_mailbox;
5960 	}
5961 
5962 	memcpy(buffer, context, sizeof(*context));
5963 
5964 err_mailbox:
5965 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5966 
5967 	return ret;
5968 }
5969 
5970 static void hns_roce_irq_work_handle(struct work_struct *work)
5971 {
5972 	struct hns_roce_work *irq_work =
5973 				container_of(work, struct hns_roce_work, work);
5974 	struct ib_device *ibdev = &irq_work->hr_dev->ib_dev;
5975 
5976 	switch (irq_work->event_type) {
5977 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5978 		ibdev_info(ibdev, "path migrated succeeded.\n");
5979 		break;
5980 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5981 		ibdev_warn(ibdev, "path migration failed.\n");
5982 		break;
5983 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
5984 		break;
5985 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5986 		ibdev_dbg(ibdev, "send queue drained.\n");
5987 		break;
5988 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5989 		ibdev_err(ibdev, "local work queue 0x%x catast error, sub_event type is: %d\n",
5990 			  irq_work->queue_num, irq_work->sub_type);
5991 		break;
5992 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5993 		ibdev_err(ibdev, "invalid request local work queue 0x%x error.\n",
5994 			  irq_work->queue_num);
5995 		break;
5996 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5997 		ibdev_err(ibdev, "local access violation work queue 0x%x error, sub_event type is: %d\n",
5998 			  irq_work->queue_num, irq_work->sub_type);
5999 		break;
6000 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
6001 		ibdev_dbg(ibdev, "SRQ limit reach.\n");
6002 		break;
6003 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
6004 		ibdev_dbg(ibdev, "SRQ last wqe reach.\n");
6005 		break;
6006 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6007 		ibdev_err(ibdev, "SRQ catas error.\n");
6008 		break;
6009 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6010 		ibdev_err(ibdev, "CQ 0x%x access err.\n", irq_work->queue_num);
6011 		break;
6012 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6013 		ibdev_warn(ibdev, "CQ 0x%x overflow\n", irq_work->queue_num);
6014 		break;
6015 	case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
6016 		ibdev_warn(ibdev, "DB overflow.\n");
6017 		break;
6018 	case HNS_ROCE_EVENT_TYPE_FLR:
6019 		ibdev_warn(ibdev, "function level reset.\n");
6020 		break;
6021 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6022 		ibdev_err(ibdev, "xrc domain violation error.\n");
6023 		break;
6024 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6025 		ibdev_err(ibdev, "invalid xrceth error.\n");
6026 		break;
6027 	default:
6028 		break;
6029 	}
6030 
6031 	kfree(irq_work);
6032 }
6033 
6034 static void hns_roce_v2_init_irq_work(struct hns_roce_dev *hr_dev,
6035 				      struct hns_roce_eq *eq, u32 queue_num)
6036 {
6037 	struct hns_roce_work *irq_work;
6038 
6039 	irq_work = kzalloc(sizeof(struct hns_roce_work), GFP_ATOMIC);
6040 	if (!irq_work)
6041 		return;
6042 
6043 	INIT_WORK(&irq_work->work, hns_roce_irq_work_handle);
6044 	irq_work->hr_dev = hr_dev;
6045 	irq_work->event_type = eq->event_type;
6046 	irq_work->sub_type = eq->sub_type;
6047 	irq_work->queue_num = queue_num;
6048 	queue_work(hr_dev->irq_workq, &irq_work->work);
6049 }
6050 
6051 static void update_eq_db(struct hns_roce_eq *eq)
6052 {
6053 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6054 	struct hns_roce_v2_db eq_db = {};
6055 
6056 	if (eq->type_flag == HNS_ROCE_AEQ) {
6057 		hr_reg_write(&eq_db, EQ_DB_CMD,
6058 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6059 			     HNS_ROCE_EQ_DB_CMD_AEQ :
6060 			     HNS_ROCE_EQ_DB_CMD_AEQ_ARMED);
6061 	} else {
6062 		hr_reg_write(&eq_db, EQ_DB_TAG, eq->eqn);
6063 
6064 		hr_reg_write(&eq_db, EQ_DB_CMD,
6065 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6066 			     HNS_ROCE_EQ_DB_CMD_CEQ :
6067 			     HNS_ROCE_EQ_DB_CMD_CEQ_ARMED);
6068 	}
6069 
6070 	hr_reg_write(&eq_db, EQ_DB_CI, eq->cons_index);
6071 
6072 	hns_roce_write64(hr_dev, (__le32 *)&eq_db, eq->db_reg);
6073 }
6074 
6075 static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
6076 {
6077 	struct hns_roce_aeqe *aeqe;
6078 
6079 	aeqe = hns_roce_buf_offset(eq->mtr.kmem,
6080 				   (eq->cons_index & (eq->entries - 1)) *
6081 				   eq->eqe_size);
6082 
6083 	return (hr_reg_read(aeqe, AEQE_OWNER) ^
6084 		!!(eq->cons_index & eq->entries)) ? aeqe : NULL;
6085 }
6086 
6087 static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
6088 				       struct hns_roce_eq *eq)
6089 {
6090 	struct device *dev = hr_dev->dev;
6091 	struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
6092 	irqreturn_t aeqe_found = IRQ_NONE;
6093 	int event_type;
6094 	u32 queue_num;
6095 	int sub_type;
6096 
6097 	while (aeqe) {
6098 		/* Make sure we read AEQ entry after we have checked the
6099 		 * ownership bit
6100 		 */
6101 		dma_rmb();
6102 
6103 		event_type = hr_reg_read(aeqe, AEQE_EVENT_TYPE);
6104 		sub_type = hr_reg_read(aeqe, AEQE_SUB_TYPE);
6105 		queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
6106 
6107 		switch (event_type) {
6108 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
6109 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
6110 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
6111 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
6112 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
6113 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
6114 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
6115 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
6116 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6117 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6118 			hns_roce_qp_event(hr_dev, queue_num, event_type);
6119 			break;
6120 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
6121 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6122 			hns_roce_srq_event(hr_dev, queue_num, event_type);
6123 			break;
6124 		case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6125 		case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6126 			hns_roce_cq_event(hr_dev, queue_num, event_type);
6127 			break;
6128 		case HNS_ROCE_EVENT_TYPE_MB:
6129 			hns_roce_cmd_event(hr_dev,
6130 					le16_to_cpu(aeqe->event.cmd.token),
6131 					aeqe->event.cmd.status,
6132 					le64_to_cpu(aeqe->event.cmd.out_param));
6133 			break;
6134 		case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
6135 		case HNS_ROCE_EVENT_TYPE_FLR:
6136 			break;
6137 		default:
6138 			dev_err(dev, "unhandled event %d on EQ %d at idx %u.\n",
6139 				event_type, eq->eqn, eq->cons_index);
6140 			break;
6141 		}
6142 
6143 		eq->event_type = event_type;
6144 		eq->sub_type = sub_type;
6145 		++eq->cons_index;
6146 		aeqe_found = IRQ_HANDLED;
6147 
6148 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_AEQE_CNT]);
6149 
6150 		hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
6151 
6152 		aeqe = next_aeqe_sw_v2(eq);
6153 	}
6154 
6155 	update_eq_db(eq);
6156 
6157 	return IRQ_RETVAL(aeqe_found);
6158 }
6159 
6160 static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq)
6161 {
6162 	struct hns_roce_ceqe *ceqe;
6163 
6164 	ceqe = hns_roce_buf_offset(eq->mtr.kmem,
6165 				   (eq->cons_index & (eq->entries - 1)) *
6166 				   eq->eqe_size);
6167 
6168 	return (hr_reg_read(ceqe, CEQE_OWNER) ^
6169 		!!(eq->cons_index & eq->entries)) ? ceqe : NULL;
6170 }
6171 
6172 static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_eq *eq)
6173 {
6174 	queue_work(system_bh_wq, &eq->work);
6175 
6176 	return IRQ_HANDLED;
6177 }
6178 
6179 static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
6180 {
6181 	struct hns_roce_eq *eq = eq_ptr;
6182 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6183 	irqreturn_t int_work;
6184 
6185 	if (eq->type_flag == HNS_ROCE_CEQ)
6186 		/* Completion event interrupt */
6187 		int_work = hns_roce_v2_ceq_int(eq);
6188 	else
6189 		/* Asynchronous event interrupt */
6190 		int_work = hns_roce_v2_aeq_int(hr_dev, eq);
6191 
6192 	return IRQ_RETVAL(int_work);
6193 }
6194 
6195 static irqreturn_t abnormal_interrupt_basic(struct hns_roce_dev *hr_dev,
6196 					    u32 int_st)
6197 {
6198 	struct pci_dev *pdev = hr_dev->pci_dev;
6199 	struct hnae3_ae_dev *ae_dev = pci_get_drvdata(pdev);
6200 	const struct hnae3_ae_ops *ops = ae_dev->ops;
6201 	enum hnae3_reset_type reset_type;
6202 	irqreturn_t int_work = IRQ_NONE;
6203 	u32 int_en;
6204 
6205 	int_en = roce_read(hr_dev, ROCEE_VF_ABN_INT_EN_REG);
6206 
6207 	if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S)) {
6208 		dev_err(hr_dev->dev, "AEQ overflow!\n");
6209 
6210 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG,
6211 			   1 << HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S);
6212 
6213 		reset_type = hr_dev->is_vf ?
6214 			     HNAE3_VF_FUNC_RESET : HNAE3_FUNC_RESET;
6215 
6216 		/* Set reset level for reset_event() */
6217 		if (ops->set_default_reset_request)
6218 			ops->set_default_reset_request(ae_dev, reset_type);
6219 		if (ops->reset_event)
6220 			ops->reset_event(pdev, NULL);
6221 
6222 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6223 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6224 
6225 		int_work = IRQ_HANDLED;
6226 	} else {
6227 		dev_err(hr_dev->dev, "there is no basic abn irq found.\n");
6228 	}
6229 
6230 	return IRQ_RETVAL(int_work);
6231 }
6232 
6233 static int fmea_ram_ecc_query(struct hns_roce_dev *hr_dev,
6234 			       struct fmea_ram_ecc *ecc_info)
6235 {
6236 	struct hns_roce_cmq_desc desc;
6237 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6238 	int ret;
6239 
6240 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_QUERY_RAM_ECC, true);
6241 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6242 	if (ret)
6243 		return ret;
6244 
6245 	ecc_info->is_ecc_err = hr_reg_read(req, QUERY_RAM_ECC_1BIT_ERR);
6246 	ecc_info->res_type = hr_reg_read(req, QUERY_RAM_ECC_RES_TYPE);
6247 	ecc_info->index = hr_reg_read(req, QUERY_RAM_ECC_TAG);
6248 
6249 	return 0;
6250 }
6251 
6252 static int fmea_recover_gmv(struct hns_roce_dev *hr_dev, u32 idx)
6253 {
6254 	struct hns_roce_cmq_desc desc;
6255 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6256 	u32 addr_upper;
6257 	u32 addr_low;
6258 	int ret;
6259 
6260 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, true);
6261 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6262 
6263 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6264 	if (ret) {
6265 		dev_err(hr_dev->dev,
6266 			"failed to execute cmd to read gmv, ret = %d.\n", ret);
6267 		return ret;
6268 	}
6269 
6270 	addr_low =  hr_reg_read(req, CFG_GMV_BT_BA_L);
6271 	addr_upper = hr_reg_read(req, CFG_GMV_BT_BA_H);
6272 
6273 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
6274 	hr_reg_write(req, CFG_GMV_BT_BA_L, addr_low);
6275 	hr_reg_write(req, CFG_GMV_BT_BA_H, addr_upper);
6276 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6277 
6278 	return hns_roce_cmq_send(hr_dev, &desc, 1);
6279 }
6280 
6281 static u64 fmea_get_ram_res_addr(u32 res_type, __le64 *data)
6282 {
6283 	if (res_type == ECC_RESOURCE_QPC_TIMER ||
6284 	    res_type == ECC_RESOURCE_CQC_TIMER ||
6285 	    res_type == ECC_RESOURCE_SCCC)
6286 		return le64_to_cpu(*data);
6287 
6288 	return le64_to_cpu(*data) << HNS_HW_PAGE_SHIFT;
6289 }
6290 
6291 static int fmea_recover_others(struct hns_roce_dev *hr_dev, u32 res_type,
6292 			       u32 index)
6293 {
6294 	u8 write_bt0_op = fmea_ram_res[res_type].write_bt0_op;
6295 	u8 read_bt0_op = fmea_ram_res[res_type].read_bt0_op;
6296 	struct hns_roce_cmd_mailbox *mailbox;
6297 	u64 addr;
6298 	int ret;
6299 
6300 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6301 	if (IS_ERR(mailbox))
6302 		return PTR_ERR(mailbox);
6303 
6304 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, read_bt0_op, index);
6305 	if (ret) {
6306 		dev_err(hr_dev->dev,
6307 			"failed to execute cmd to read fmea ram, ret = %d.\n",
6308 			ret);
6309 		goto out;
6310 	}
6311 
6312 	addr = fmea_get_ram_res_addr(res_type, mailbox->buf);
6313 
6314 	ret = hns_roce_cmd_mbox(hr_dev, addr, 0, write_bt0_op, index);
6315 	if (ret)
6316 		dev_err(hr_dev->dev,
6317 			"failed to execute cmd to write fmea ram, ret = %d.\n",
6318 			ret);
6319 
6320 out:
6321 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6322 	return ret;
6323 }
6324 
6325 static void fmea_ram_ecc_recover(struct hns_roce_dev *hr_dev,
6326 				 struct fmea_ram_ecc *ecc_info)
6327 {
6328 	u32 res_type = ecc_info->res_type;
6329 	u32 index = ecc_info->index;
6330 	int ret;
6331 
6332 	BUILD_BUG_ON(ARRAY_SIZE(fmea_ram_res) != ECC_RESOURCE_COUNT);
6333 
6334 	if (res_type >= ECC_RESOURCE_COUNT) {
6335 		dev_err(hr_dev->dev, "unsupported fmea ram ecc type %u.\n",
6336 			res_type);
6337 		return;
6338 	}
6339 
6340 	if (res_type == ECC_RESOURCE_GMV)
6341 		ret = fmea_recover_gmv(hr_dev, index);
6342 	else
6343 		ret = fmea_recover_others(hr_dev, res_type, index);
6344 	if (ret)
6345 		dev_err(hr_dev->dev,
6346 			"failed to recover %s, index = %u, ret = %d.\n",
6347 			fmea_ram_res[res_type].name, index, ret);
6348 }
6349 
6350 static void fmea_ram_ecc_work(struct work_struct *ecc_work)
6351 {
6352 	struct hns_roce_dev *hr_dev =
6353 		container_of(ecc_work, struct hns_roce_dev, ecc_work);
6354 	struct fmea_ram_ecc ecc_info = {};
6355 
6356 	if (fmea_ram_ecc_query(hr_dev, &ecc_info)) {
6357 		dev_err(hr_dev->dev, "failed to query fmea ram ecc.\n");
6358 		return;
6359 	}
6360 
6361 	if (!ecc_info.is_ecc_err) {
6362 		dev_err(hr_dev->dev, "there is no fmea ram ecc err found.\n");
6363 		return;
6364 	}
6365 
6366 	fmea_ram_ecc_recover(hr_dev, &ecc_info);
6367 }
6368 
6369 static irqreturn_t hns_roce_v2_msix_interrupt_abn(int irq, void *dev_id)
6370 {
6371 	struct hns_roce_dev *hr_dev = dev_id;
6372 	irqreturn_t int_work = IRQ_NONE;
6373 	u32 int_st;
6374 
6375 	int_st = roce_read(hr_dev, ROCEE_VF_ABN_INT_ST_REG);
6376 
6377 	if (int_st) {
6378 		int_work = abnormal_interrupt_basic(hr_dev, int_st);
6379 	} else if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
6380 		queue_work(hr_dev->irq_workq, &hr_dev->ecc_work);
6381 		int_work = IRQ_HANDLED;
6382 	} else {
6383 		dev_err(hr_dev->dev, "there is no abnormal irq found.\n");
6384 	}
6385 
6386 	return IRQ_RETVAL(int_work);
6387 }
6388 
6389 static void hns_roce_v2_int_mask_enable(struct hns_roce_dev *hr_dev,
6390 					int eq_num, u32 enable_flag)
6391 {
6392 	int i;
6393 
6394 	for (i = 0; i < eq_num; i++)
6395 		roce_write(hr_dev, ROCEE_VF_EVENT_INT_EN_REG +
6396 			   i * EQ_REG_OFFSET, enable_flag);
6397 
6398 	roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, enable_flag);
6399 	roce_write(hr_dev, ROCEE_VF_ABN_INT_CFG_REG, enable_flag);
6400 }
6401 
6402 static void free_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6403 {
6404 	hns_roce_mtr_destroy(hr_dev, &eq->mtr);
6405 }
6406 
6407 static void hns_roce_v2_destroy_eqc(struct hns_roce_dev *hr_dev,
6408 				    struct hns_roce_eq *eq)
6409 {
6410 	struct device *dev = hr_dev->dev;
6411 	int eqn = eq->eqn;
6412 	int ret;
6413 	u8 cmd;
6414 
6415 	if (eqn < hr_dev->caps.num_comp_vectors)
6416 		cmd = HNS_ROCE_CMD_DESTROY_CEQC;
6417 	else
6418 		cmd = HNS_ROCE_CMD_DESTROY_AEQC;
6419 
6420 	ret = hns_roce_destroy_hw_ctx(hr_dev, cmd, eqn & HNS_ROCE_V2_EQN_M);
6421 	if (ret)
6422 		dev_err(dev, "[mailbox cmd] destroy eqc(%d) failed.\n", eqn);
6423 
6424 	free_eq_buf(hr_dev, eq);
6425 }
6426 
6427 static void init_eq_config(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6428 {
6429 	eq->db_reg = hr_dev->reg_base + ROCEE_VF_EQ_DB_CFG0_REG;
6430 	eq->cons_index = 0;
6431 	eq->over_ignore = HNS_ROCE_V2_EQ_OVER_IGNORE_0;
6432 	eq->coalesce = HNS_ROCE_V2_EQ_COALESCE_0;
6433 	eq->arm_st = HNS_ROCE_V2_EQ_ALWAYS_ARMED;
6434 	eq->shift = ilog2((unsigned int)eq->entries);
6435 }
6436 
6437 static int config_eqc(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq,
6438 		      void *mb_buf)
6439 {
6440 	u64 eqe_ba[MTT_MIN_COUNT] = { 0 };
6441 	struct hns_roce_eq_context *eqc;
6442 	u64 bt_ba = 0;
6443 	int ret;
6444 
6445 	eqc = mb_buf;
6446 	memset(eqc, 0, sizeof(struct hns_roce_eq_context));
6447 
6448 	init_eq_config(hr_dev, eq);
6449 
6450 	/* if not multi-hop, eqe buffer only use one trunk */
6451 	ret = hns_roce_mtr_find(hr_dev, &eq->mtr, 0, eqe_ba,
6452 				ARRAY_SIZE(eqe_ba));
6453 	if (ret) {
6454 		dev_err(hr_dev->dev, "failed to find EQE mtr, ret = %d\n", ret);
6455 		return ret;
6456 	}
6457 
6458 	bt_ba = hns_roce_get_mtr_ba(&eq->mtr);
6459 
6460 	hr_reg_write(eqc, EQC_EQ_ST, HNS_ROCE_V2_EQ_STATE_VALID);
6461 	hr_reg_write(eqc, EQC_EQE_HOP_NUM, eq->hop_num);
6462 	hr_reg_write(eqc, EQC_OVER_IGNORE, eq->over_ignore);
6463 	hr_reg_write(eqc, EQC_COALESCE, eq->coalesce);
6464 	hr_reg_write(eqc, EQC_ARM_ST, eq->arm_st);
6465 	hr_reg_write(eqc, EQC_EQN, eq->eqn);
6466 	hr_reg_write(eqc, EQC_EQE_CNT, HNS_ROCE_EQ_INIT_EQE_CNT);
6467 	hr_reg_write(eqc, EQC_EQE_BA_PG_SZ,
6468 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.ba_pg_shift));
6469 	hr_reg_write(eqc, EQC_EQE_BUF_PG_SZ,
6470 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.buf_pg_shift));
6471 	hr_reg_write(eqc, EQC_EQ_PROD_INDX, HNS_ROCE_EQ_INIT_PROD_IDX);
6472 	hr_reg_write(eqc, EQC_EQ_MAX_CNT, eq->eq_max_cnt);
6473 
6474 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6475 		if (eq->eq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
6476 			dev_info(hr_dev->dev, "eq_period(%u) reached the upper limit, adjusted to 65.\n",
6477 				 eq->eq_period);
6478 			eq->eq_period = HNS_ROCE_MAX_EQ_PERIOD;
6479 		}
6480 		eq->eq_period *= HNS_ROCE_CLOCK_ADJUST;
6481 	}
6482 
6483 	hr_reg_write(eqc, EQC_EQ_PERIOD, eq->eq_period);
6484 	hr_reg_write(eqc, EQC_EQE_REPORT_TIMER, HNS_ROCE_EQ_INIT_REPORT_TIMER);
6485 	hr_reg_write(eqc, EQC_EQE_BA_L, bt_ba >> 3);
6486 	hr_reg_write(eqc, EQC_EQE_BA_H, bt_ba >> 35);
6487 	hr_reg_write(eqc, EQC_SHIFT, eq->shift);
6488 	hr_reg_write(eqc, EQC_MSI_INDX, HNS_ROCE_EQ_INIT_MSI_IDX);
6489 	hr_reg_write(eqc, EQC_CUR_EQE_BA_L, eqe_ba[0] >> 12);
6490 	hr_reg_write(eqc, EQC_CUR_EQE_BA_M, eqe_ba[0] >> 28);
6491 	hr_reg_write(eqc, EQC_CUR_EQE_BA_H, eqe_ba[0] >> 60);
6492 	hr_reg_write(eqc, EQC_EQ_CONS_INDX, HNS_ROCE_EQ_INIT_CONS_IDX);
6493 	hr_reg_write(eqc, EQC_NEX_EQE_BA_L, eqe_ba[1] >> 12);
6494 	hr_reg_write(eqc, EQC_NEX_EQE_BA_H, eqe_ba[1] >> 44);
6495 	hr_reg_write(eqc, EQC_EQE_SIZE, eq->eqe_size == HNS_ROCE_V3_EQE_SIZE);
6496 
6497 	return 0;
6498 }
6499 
6500 static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6501 {
6502 	struct hns_roce_buf_attr buf_attr = {};
6503 	int err;
6504 
6505 	if (hr_dev->caps.eqe_hop_num == HNS_ROCE_HOP_NUM_0)
6506 		eq->hop_num = 0;
6507 	else
6508 		eq->hop_num = hr_dev->caps.eqe_hop_num;
6509 
6510 	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
6511 	buf_attr.region[0].size = eq->entries * eq->eqe_size;
6512 	buf_attr.region[0].hopnum = eq->hop_num;
6513 	buf_attr.region_count = 1;
6514 
6515 	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
6516 				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
6517 				  0);
6518 	if (err)
6519 		dev_err(hr_dev->dev, "failed to alloc EQE mtr, err %d\n", err);
6520 
6521 	return err;
6522 }
6523 
6524 static int hns_roce_v2_create_eq(struct hns_roce_dev *hr_dev,
6525 				 struct hns_roce_eq *eq, u8 eq_cmd)
6526 {
6527 	struct hns_roce_cmd_mailbox *mailbox;
6528 	int ret;
6529 
6530 	/* Allocate mailbox memory */
6531 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6532 	if (IS_ERR(mailbox))
6533 		return PTR_ERR(mailbox);
6534 
6535 	ret = alloc_eq_buf(hr_dev, eq);
6536 	if (ret)
6537 		goto free_cmd_mbox;
6538 
6539 	ret = config_eqc(hr_dev, eq, mailbox->buf);
6540 	if (ret)
6541 		goto err_cmd_mbox;
6542 
6543 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, eq_cmd, eq->eqn);
6544 	if (ret) {
6545 		dev_err(hr_dev->dev, "[mailbox cmd] create eqc failed.\n");
6546 		goto err_cmd_mbox;
6547 	}
6548 
6549 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6550 
6551 	return 0;
6552 
6553 err_cmd_mbox:
6554 	free_eq_buf(hr_dev, eq);
6555 
6556 free_cmd_mbox:
6557 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6558 
6559 	return ret;
6560 }
6561 
6562 static void hns_roce_ceq_work(struct work_struct *work)
6563 {
6564 	struct hns_roce_eq *eq = from_work(eq, work, work);
6565 	struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq);
6566 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6567 	int ceqe_num = 0;
6568 	u32 cqn;
6569 
6570 	while (ceqe && ceqe_num < hr_dev->caps.ceqe_depth) {
6571 		/* Make sure we read CEQ entry after we have checked the
6572 		 * ownership bit
6573 		 */
6574 		dma_rmb();
6575 
6576 		cqn = hr_reg_read(ceqe, CEQE_CQN);
6577 
6578 		hns_roce_cq_completion(hr_dev, cqn);
6579 
6580 		++eq->cons_index;
6581 		++ceqe_num;
6582 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CEQE_CNT]);
6583 
6584 		ceqe = next_ceqe_sw_v2(eq);
6585 	}
6586 
6587 	update_eq_db(eq);
6588 }
6589 
6590 static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
6591 				  int comp_num, int aeq_num, int other_num)
6592 {
6593 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6594 	int i, j;
6595 	int ret;
6596 
6597 	for (i = 0; i < irq_num; i++) {
6598 		hr_dev->irq_names[i] = kzalloc(HNS_ROCE_INT_NAME_LEN,
6599 					       GFP_KERNEL);
6600 		if (!hr_dev->irq_names[i]) {
6601 			ret = -ENOMEM;
6602 			goto err_kzalloc_failed;
6603 		}
6604 	}
6605 
6606 	/* irq contains: abnormal + AEQ + CEQ */
6607 	for (j = 0; j < other_num; j++)
6608 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6609 			 "hns-%s-abn-%d", pci_name(hr_dev->pci_dev), j);
6610 
6611 	for (j = other_num; j < (other_num + aeq_num); j++)
6612 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6613 			 "hns-%s-aeq-%d", pci_name(hr_dev->pci_dev), j - other_num);
6614 
6615 	for (j = (other_num + aeq_num); j < irq_num; j++)
6616 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6617 			 "hns-%s-ceq-%d", pci_name(hr_dev->pci_dev),
6618 			 j - other_num - aeq_num);
6619 
6620 	for (j = 0; j < irq_num; j++) {
6621 		if (j < other_num) {
6622 			ret = request_irq(hr_dev->irq[j],
6623 					  hns_roce_v2_msix_interrupt_abn,
6624 					  0, hr_dev->irq_names[j], hr_dev);
6625 		} else if (j < (other_num + comp_num)) {
6626 			INIT_WORK(&eq_table->eq[j - other_num].work,
6627 				  hns_roce_ceq_work);
6628 			ret = request_irq(eq_table->eq[j - other_num].irq,
6629 					  hns_roce_v2_msix_interrupt_eq,
6630 					  0, hr_dev->irq_names[j + aeq_num],
6631 					  &eq_table->eq[j - other_num]);
6632 		} else {
6633 			ret = request_irq(eq_table->eq[j - other_num].irq,
6634 					  hns_roce_v2_msix_interrupt_eq,
6635 					  0, hr_dev->irq_names[j - comp_num],
6636 					  &eq_table->eq[j - other_num]);
6637 		}
6638 
6639 		if (ret) {
6640 			dev_err(hr_dev->dev, "request irq error!\n");
6641 			goto err_request_failed;
6642 		}
6643 	}
6644 
6645 	return 0;
6646 
6647 err_request_failed:
6648 	for (j -= 1; j >= 0; j--) {
6649 		if (j < other_num) {
6650 			free_irq(hr_dev->irq[j], hr_dev);
6651 			continue;
6652 		}
6653 		free_irq(eq_table->eq[j - other_num].irq,
6654 			 &eq_table->eq[j - other_num]);
6655 		if (j < other_num + comp_num)
6656 			cancel_work_sync(&eq_table->eq[j - other_num].work);
6657 	}
6658 
6659 err_kzalloc_failed:
6660 	for (i -= 1; i >= 0; i--)
6661 		kfree(hr_dev->irq_names[i]);
6662 
6663 	return ret;
6664 }
6665 
6666 static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev)
6667 {
6668 	int irq_num;
6669 	int eq_num;
6670 	int i;
6671 
6672 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6673 	irq_num = eq_num + hr_dev->caps.num_other_vectors;
6674 
6675 	for (i = 0; i < hr_dev->caps.num_other_vectors; i++)
6676 		free_irq(hr_dev->irq[i], hr_dev);
6677 
6678 	for (i = 0; i < eq_num; i++) {
6679 		free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]);
6680 		if (i < hr_dev->caps.num_comp_vectors)
6681 			cancel_work_sync(&hr_dev->eq_table.eq[i].work);
6682 	}
6683 
6684 	for (i = 0; i < irq_num; i++)
6685 		kfree(hr_dev->irq_names[i]);
6686 }
6687 
6688 static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
6689 {
6690 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6691 	struct device *dev = hr_dev->dev;
6692 	struct hns_roce_eq *eq;
6693 	int other_num;
6694 	int comp_num;
6695 	int aeq_num;
6696 	int irq_num;
6697 	int eq_num;
6698 	u8 eq_cmd;
6699 	int ret;
6700 	int i;
6701 
6702 	other_num = hr_dev->caps.num_other_vectors;
6703 	comp_num = hr_dev->caps.num_comp_vectors;
6704 	aeq_num = hr_dev->caps.num_aeq_vectors;
6705 
6706 	eq_num = comp_num + aeq_num;
6707 	irq_num = eq_num + other_num;
6708 
6709 	eq_table->eq = kcalloc(eq_num, sizeof(*eq_table->eq), GFP_KERNEL);
6710 	if (!eq_table->eq)
6711 		return -ENOMEM;
6712 
6713 	/* create eq */
6714 	for (i = 0; i < eq_num; i++) {
6715 		eq = &eq_table->eq[i];
6716 		eq->hr_dev = hr_dev;
6717 		eq->eqn = i;
6718 		if (i < comp_num) {
6719 			/* CEQ */
6720 			eq_cmd = HNS_ROCE_CMD_CREATE_CEQC;
6721 			eq->type_flag = HNS_ROCE_CEQ;
6722 			eq->entries = hr_dev->caps.ceqe_depth;
6723 			eq->eqe_size = hr_dev->caps.ceqe_size;
6724 			eq->irq = hr_dev->irq[i + other_num + aeq_num];
6725 			eq->eq_max_cnt = HNS_ROCE_CEQ_DEFAULT_BURST_NUM;
6726 			eq->eq_period = HNS_ROCE_CEQ_DEFAULT_INTERVAL;
6727 		} else {
6728 			/* AEQ */
6729 			eq_cmd = HNS_ROCE_CMD_CREATE_AEQC;
6730 			eq->type_flag = HNS_ROCE_AEQ;
6731 			eq->entries = hr_dev->caps.aeqe_depth;
6732 			eq->eqe_size = hr_dev->caps.aeqe_size;
6733 			eq->irq = hr_dev->irq[i - comp_num + other_num];
6734 			eq->eq_max_cnt = HNS_ROCE_AEQ_DEFAULT_BURST_NUM;
6735 			eq->eq_period = HNS_ROCE_AEQ_DEFAULT_INTERVAL;
6736 		}
6737 
6738 		ret = hns_roce_v2_create_eq(hr_dev, eq, eq_cmd);
6739 		if (ret) {
6740 			dev_err(dev, "failed to create eq.\n");
6741 			goto err_create_eq_fail;
6742 		}
6743 	}
6744 
6745 	INIT_WORK(&hr_dev->ecc_work, fmea_ram_ecc_work);
6746 
6747 	hr_dev->irq_workq = alloc_ordered_workqueue("hns_roce_irq_workq", 0);
6748 	if (!hr_dev->irq_workq) {
6749 		dev_err(dev, "failed to create irq workqueue.\n");
6750 		ret = -ENOMEM;
6751 		goto err_create_eq_fail;
6752 	}
6753 
6754 	ret = __hns_roce_request_irq(hr_dev, irq_num, comp_num, aeq_num,
6755 				     other_num);
6756 	if (ret) {
6757 		dev_err(dev, "failed to request irq.\n");
6758 		goto err_request_irq_fail;
6759 	}
6760 
6761 	/* enable irq */
6762 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_ENABLE);
6763 
6764 	return 0;
6765 
6766 err_request_irq_fail:
6767 	destroy_workqueue(hr_dev->irq_workq);
6768 
6769 err_create_eq_fail:
6770 	for (i -= 1; i >= 0; i--)
6771 		hns_roce_v2_destroy_eqc(hr_dev, &eq_table->eq[i]);
6772 	kfree(eq_table->eq);
6773 
6774 	return ret;
6775 }
6776 
6777 static void hns_roce_v2_cleanup_eq_table(struct hns_roce_dev *hr_dev)
6778 {
6779 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6780 	int eq_num;
6781 	int i;
6782 
6783 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6784 
6785 	/* Disable irq */
6786 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_DISABLE);
6787 
6788 	__hns_roce_free_irq(hr_dev);
6789 	destroy_workqueue(hr_dev->irq_workq);
6790 
6791 	for (i = 0; i < eq_num; i++)
6792 		hns_roce_v2_destroy_eqc(hr_dev, &eq_table->eq[i]);
6793 
6794 	kfree(eq_table->eq);
6795 }
6796 
6797 static const struct ib_device_ops hns_roce_v2_dev_ops = {
6798 	.destroy_qp = hns_roce_v2_destroy_qp,
6799 	.modify_cq = hns_roce_v2_modify_cq,
6800 	.poll_cq = hns_roce_v2_poll_cq,
6801 	.post_recv = hns_roce_v2_post_recv,
6802 	.post_send = hns_roce_v2_post_send,
6803 	.query_qp = hns_roce_v2_query_qp,
6804 	.req_notify_cq = hns_roce_v2_req_notify_cq,
6805 };
6806 
6807 static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {
6808 	.modify_srq = hns_roce_v2_modify_srq,
6809 	.post_srq_recv = hns_roce_v2_post_srq_recv,
6810 	.query_srq = hns_roce_v2_query_srq,
6811 };
6812 
6813 static const struct hns_roce_hw hns_roce_hw_v2 = {
6814 	.cmq_init = hns_roce_v2_cmq_init,
6815 	.cmq_exit = hns_roce_v2_cmq_exit,
6816 	.hw_profile = hns_roce_v2_profile,
6817 	.hw_init = hns_roce_v2_init,
6818 	.hw_exit = hns_roce_v2_exit,
6819 	.post_mbox = v2_post_mbox,
6820 	.poll_mbox_done = v2_poll_mbox_done,
6821 	.chk_mbox_avail = v2_chk_mbox_is_avail,
6822 	.set_gid = hns_roce_v2_set_gid,
6823 	.set_mac = hns_roce_v2_set_mac,
6824 	.write_mtpt = hns_roce_v2_write_mtpt,
6825 	.rereg_write_mtpt = hns_roce_v2_rereg_write_mtpt,
6826 	.frmr_write_mtpt = hns_roce_v2_frmr_write_mtpt,
6827 	.mw_write_mtpt = hns_roce_v2_mw_write_mtpt,
6828 	.write_cqc = hns_roce_v2_write_cqc,
6829 	.set_hem = hns_roce_v2_set_hem,
6830 	.clear_hem = hns_roce_v2_clear_hem,
6831 	.modify_qp = hns_roce_v2_modify_qp,
6832 	.dereg_mr = hns_roce_v2_dereg_mr,
6833 	.qp_flow_control_init = hns_roce_v2_qp_flow_control_init,
6834 	.init_eq = hns_roce_v2_init_eq_table,
6835 	.cleanup_eq = hns_roce_v2_cleanup_eq_table,
6836 	.write_srqc = hns_roce_v2_write_srqc,
6837 	.query_cqc = hns_roce_v2_query_cqc,
6838 	.query_qpc = hns_roce_v2_query_qpc,
6839 	.query_mpt = hns_roce_v2_query_mpt,
6840 	.query_srqc = hns_roce_v2_query_srqc,
6841 	.query_sccc = hns_roce_v2_query_sccc,
6842 	.query_hw_counter = hns_roce_hw_v2_query_counter,
6843 	.get_dscp = hns_roce_hw_v2_get_dscp,
6844 	.hns_roce_dev_ops = &hns_roce_v2_dev_ops,
6845 	.hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops,
6846 };
6847 
6848 static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = {
6849 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA), 0},
6850 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA_MACSEC), 0},
6851 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA), 0},
6852 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA_MACSEC), 0},
6853 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_100G_RDMA_MACSEC), 0},
6854 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_200G_RDMA), 0},
6855 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_RDMA_DCB_PFC_VF),
6856 	 HNAE3_DEV_SUPPORT_ROCE_DCB_BITS},
6857 	/* required last entry */
6858 	{0, }
6859 };
6860 
6861 MODULE_DEVICE_TABLE(pci, hns_roce_hw_v2_pci_tbl);
6862 
6863 static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
6864 				  struct hnae3_handle *handle)
6865 {
6866 	struct hns_roce_v2_priv *priv = hr_dev->priv;
6867 	const struct pci_device_id *id;
6868 	int i;
6869 
6870 	hr_dev->pci_dev = handle->pdev;
6871 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, hr_dev->pci_dev);
6872 	hr_dev->is_vf = id->driver_data;
6873 	hr_dev->dev = &handle->pdev->dev;
6874 	hr_dev->hw = &hns_roce_hw_v2;
6875 	hr_dev->sdb_offset = ROCEE_DB_SQ_L_0_REG;
6876 	hr_dev->odb_offset = hr_dev->sdb_offset;
6877 
6878 	/* Get info from NIC driver. */
6879 	hr_dev->reg_base = handle->rinfo.roce_io_base;
6880 	hr_dev->mem_base = handle->rinfo.roce_mem_base;
6881 	hr_dev->caps.num_ports = 1;
6882 	hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
6883 	hr_dev->iboe.phy_port[0] = 0;
6884 
6885 	addrconf_addr_eui48((u8 *)&hr_dev->ib_dev.node_guid,
6886 			    hr_dev->iboe.netdevs[0]->dev_addr);
6887 
6888 	for (i = 0; i < handle->rinfo.num_vectors; i++)
6889 		hr_dev->irq[i] = pci_irq_vector(handle->pdev,
6890 						i + handle->rinfo.base_vector);
6891 
6892 	/* cmd issue mode: 0 is poll, 1 is event */
6893 	hr_dev->cmd_mod = 1;
6894 	hr_dev->loop_idc = 0;
6895 
6896 	hr_dev->reset_cnt = handle->ae_algo->ops->ae_dev_reset_cnt(handle);
6897 	priv->handle = handle;
6898 }
6899 
6900 static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6901 {
6902 	struct hns_roce_dev *hr_dev;
6903 	int ret;
6904 
6905 	hr_dev = ib_alloc_device(hns_roce_dev, ib_dev);
6906 	if (!hr_dev)
6907 		return -ENOMEM;
6908 
6909 	hr_dev->priv = kzalloc(sizeof(struct hns_roce_v2_priv), GFP_KERNEL);
6910 	if (!hr_dev->priv) {
6911 		ret = -ENOMEM;
6912 		goto error_failed_kzalloc;
6913 	}
6914 
6915 	hns_roce_hw_v2_get_cfg(hr_dev, handle);
6916 
6917 	ret = hns_roce_init(hr_dev);
6918 	if (ret) {
6919 		dev_err(hr_dev->dev, "RoCE Engine init failed!\n");
6920 		goto error_failed_roce_init;
6921 	}
6922 
6923 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6924 		ret = free_mr_init(hr_dev);
6925 		if (ret) {
6926 			dev_err(hr_dev->dev, "failed to init free mr!\n");
6927 			goto error_failed_free_mr_init;
6928 		}
6929 	}
6930 
6931 	handle->priv = hr_dev;
6932 
6933 	return 0;
6934 
6935 error_failed_free_mr_init:
6936 	hns_roce_exit(hr_dev);
6937 
6938 error_failed_roce_init:
6939 	kfree(hr_dev->priv);
6940 
6941 error_failed_kzalloc:
6942 	ib_dealloc_device(&hr_dev->ib_dev);
6943 
6944 	return ret;
6945 }
6946 
6947 static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6948 					   bool reset)
6949 {
6950 	struct hns_roce_dev *hr_dev = handle->priv;
6951 
6952 	if (!hr_dev)
6953 		return;
6954 
6955 	handle->priv = NULL;
6956 
6957 	hr_dev->state = HNS_ROCE_DEVICE_STATE_UNINIT;
6958 	hns_roce_handle_device_err(hr_dev);
6959 
6960 	hns_roce_exit(hr_dev);
6961 	kfree(hr_dev->priv);
6962 	ib_dealloc_device(&hr_dev->ib_dev);
6963 }
6964 
6965 static int hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6966 {
6967 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
6968 	const struct pci_device_id *id;
6969 	struct device *dev = &handle->pdev->dev;
6970 	int ret;
6971 
6972 	handle->rinfo.instance_state = HNS_ROCE_STATE_INIT;
6973 
6974 	if (ops->ae_dev_resetting(handle) || ops->get_hw_reset_stat(handle)) {
6975 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6976 		goto reset_chk_err;
6977 	}
6978 
6979 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, handle->pdev);
6980 	if (!id)
6981 		return 0;
6982 
6983 	if (id->driver_data && handle->pdev->revision == PCI_REVISION_ID_HIP08)
6984 		return 0;
6985 
6986 	ret = __hns_roce_hw_v2_init_instance(handle);
6987 	if (ret) {
6988 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6989 		dev_err(dev, "RoCE instance init failed! ret = %d\n", ret);
6990 		if (ops->ae_dev_resetting(handle) ||
6991 		    ops->get_hw_reset_stat(handle))
6992 			goto reset_chk_err;
6993 		else
6994 			return ret;
6995 	}
6996 
6997 	handle->rinfo.instance_state = HNS_ROCE_STATE_INITED;
6998 
6999 	return 0;
7000 
7001 reset_chk_err:
7002 	dev_err(dev, "Device is busy in resetting state.\n"
7003 		     "please retry later.\n");
7004 
7005 	return -EBUSY;
7006 }
7007 
7008 static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
7009 					   bool reset)
7010 {
7011 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED)
7012 		return;
7013 
7014 	handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;
7015 
7016 	__hns_roce_hw_v2_uninit_instance(handle, reset);
7017 
7018 	handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
7019 }
7020 static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
7021 {
7022 	struct hns_roce_dev *hr_dev;
7023 
7024 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) {
7025 		set_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
7026 		return 0;
7027 	}
7028 
7029 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_DOWN;
7030 	clear_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
7031 
7032 	hr_dev = handle->priv;
7033 	if (!hr_dev)
7034 		return 0;
7035 
7036 	hr_dev->active = false;
7037 	hr_dev->dis_db = true;
7038 	hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
7039 
7040 	return 0;
7041 }
7042 
7043 static int hns_roce_hw_v2_reset_notify_init(struct hnae3_handle *handle)
7044 {
7045 	struct device *dev = &handle->pdev->dev;
7046 	int ret;
7047 
7048 	if (test_and_clear_bit(HNS_ROCE_RST_DIRECT_RETURN,
7049 			       &handle->rinfo.state)) {
7050 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7051 		return 0;
7052 	}
7053 
7054 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INIT;
7055 
7056 	dev_info(&handle->pdev->dev, "In reset process RoCE client reinit.\n");
7057 	ret = __hns_roce_hw_v2_init_instance(handle);
7058 	if (ret) {
7059 		/* when reset notify type is HNAE3_INIT_CLIENT In reset notify
7060 		 * callback function, RoCE Engine reinitialize. If RoCE reinit
7061 		 * failed, we should inform NIC driver.
7062 		 */
7063 		handle->priv = NULL;
7064 		dev_err(dev, "In reset process RoCE reinit failed %d.\n", ret);
7065 	} else {
7066 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7067 		dev_info(dev, "reset done, RoCE client reinit finished.\n");
7068 	}
7069 
7070 	return ret;
7071 }
7072 
7073 static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle)
7074 {
7075 	if (test_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state))
7076 		return 0;
7077 
7078 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT;
7079 	dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n");
7080 	msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY);
7081 	__hns_roce_hw_v2_uninit_instance(handle, false);
7082 
7083 	return 0;
7084 }
7085 
7086 static int hns_roce_hw_v2_reset_notify(struct hnae3_handle *handle,
7087 				       enum hnae3_reset_notify_type type)
7088 {
7089 	int ret = 0;
7090 
7091 	switch (type) {
7092 	case HNAE3_DOWN_CLIENT:
7093 		ret = hns_roce_hw_v2_reset_notify_down(handle);
7094 		break;
7095 	case HNAE3_INIT_CLIENT:
7096 		ret = hns_roce_hw_v2_reset_notify_init(handle);
7097 		break;
7098 	case HNAE3_UNINIT_CLIENT:
7099 		ret = hns_roce_hw_v2_reset_notify_uninit(handle);
7100 		break;
7101 	default:
7102 		break;
7103 	}
7104 
7105 	return ret;
7106 }
7107 
7108 static const struct hnae3_client_ops hns_roce_hw_v2_ops = {
7109 	.init_instance = hns_roce_hw_v2_init_instance,
7110 	.uninit_instance = hns_roce_hw_v2_uninit_instance,
7111 	.reset_notify = hns_roce_hw_v2_reset_notify,
7112 };
7113 
7114 static struct hnae3_client hns_roce_hw_v2_client = {
7115 	.name = "hns_roce_hw_v2",
7116 	.type = HNAE3_CLIENT_ROCE,
7117 	.ops = &hns_roce_hw_v2_ops,
7118 };
7119 
7120 static int __init hns_roce_hw_v2_init(void)
7121 {
7122 	hns_roce_init_debugfs();
7123 	return hnae3_register_client(&hns_roce_hw_v2_client);
7124 }
7125 
7126 static void __exit hns_roce_hw_v2_exit(void)
7127 {
7128 	hnae3_unregister_client(&hns_roce_hw_v2_client);
7129 	hns_roce_cleanup_debugfs();
7130 }
7131 
7132 module_init(hns_roce_hw_v2_init);
7133 module_exit(hns_roce_hw_v2_exit);
7134 
7135 MODULE_LICENSE("Dual BSD/GPL");
7136 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
7137 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
7138 MODULE_AUTHOR("Shaobo Xu <xushaobo2@huawei.com>");
7139 MODULE_DESCRIPTION("Hisilicon Hip08 Family RoCE Driver");
7140