1 /*
2  * Copyright (c) 2016-2017 Hisilicon Limited.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/acpi.h>
34 #include <linux/etherdevice.h>
35 #include <linux/interrupt.h>
36 #include <linux/iopoll.h>
37 #include <linux/kernel.h>
38 #include <linux/types.h>
39 #include <net/addrconf.h>
40 #include <rdma/ib_addr.h>
41 #include <rdma/ib_cache.h>
42 #include <rdma/ib_umem.h>
43 #include <rdma/uverbs_ioctl.h>
44 
45 #include "hnae3.h"
46 #include "hns_roce_common.h"
47 #include "hns_roce_device.h"
48 #include "hns_roce_cmd.h"
49 #include "hns_roce_hem.h"
50 #include "hns_roce_hw_v2.h"
51 
52 enum {
53 	CMD_RST_PRC_OTHERS,
54 	CMD_RST_PRC_SUCCESS,
55 	CMD_RST_PRC_EBUSY,
56 };
57 
58 static inline void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
59 				   struct ib_sge *sg)
60 {
61 	dseg->lkey = cpu_to_le32(sg->lkey);
62 	dseg->addr = cpu_to_le64(sg->addr);
63 	dseg->len  = cpu_to_le32(sg->length);
64 }
65 
66 /*
67  * mapped-value = 1 + real-value
68  * The hns wr opcode real value is start from 0, In order to distinguish between
69  * initialized and uninitialized map values, we plus 1 to the actual value when
70  * defining the mapping, so that the validity can be identified by checking the
71  * mapped value is greater than 0.
72  */
73 #define HR_OPC_MAP(ib_key, hr_key) \
74 		[IB_WR_ ## ib_key] = 1 + HNS_ROCE_V2_WQE_OP_ ## hr_key
75 
76 static const u32 hns_roce_op_code[] = {
77 	HR_OPC_MAP(RDMA_WRITE,			RDMA_WRITE),
78 	HR_OPC_MAP(RDMA_WRITE_WITH_IMM,		RDMA_WRITE_WITH_IMM),
79 	HR_OPC_MAP(SEND,			SEND),
80 	HR_OPC_MAP(SEND_WITH_IMM,		SEND_WITH_IMM),
81 	HR_OPC_MAP(RDMA_READ,			RDMA_READ),
82 	HR_OPC_MAP(ATOMIC_CMP_AND_SWP,		ATOM_CMP_AND_SWAP),
83 	HR_OPC_MAP(ATOMIC_FETCH_AND_ADD,	ATOM_FETCH_AND_ADD),
84 	HR_OPC_MAP(SEND_WITH_INV,		SEND_WITH_INV),
85 	HR_OPC_MAP(LOCAL_INV,			LOCAL_INV),
86 	HR_OPC_MAP(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
87 	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
88 	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
89 };
90 
91 static u32 to_hr_opcode(u32 ib_opcode)
92 {
93 	if (ib_opcode >= ARRAY_SIZE(hns_roce_op_code))
94 		return HNS_ROCE_V2_WQE_OP_MASK;
95 
96 	return hns_roce_op_code[ib_opcode] ? hns_roce_op_code[ib_opcode] - 1 :
97 					     HNS_ROCE_V2_WQE_OP_MASK;
98 }
99 
100 static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
101 			 const struct ib_reg_wr *wr)
102 {
103 	struct hns_roce_wqe_frmr_seg *fseg =
104 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
105 	struct hns_roce_mr *mr = to_hr_mr(wr->mr);
106 	u64 pbl_ba;
107 
108 	/* use ib_access_flags */
109 	hr_reg_write_bool(fseg, FRMR_BIND_EN, wr->access & IB_ACCESS_MW_BIND);
110 	hr_reg_write_bool(fseg, FRMR_ATOMIC,
111 			  wr->access & IB_ACCESS_REMOTE_ATOMIC);
112 	hr_reg_write_bool(fseg, FRMR_RR, wr->access & IB_ACCESS_REMOTE_READ);
113 	hr_reg_write_bool(fseg, FRMR_RW, wr->access & IB_ACCESS_REMOTE_WRITE);
114 	hr_reg_write_bool(fseg, FRMR_LW, wr->access & IB_ACCESS_LOCAL_WRITE);
115 
116 	/* Data structure reuse may lead to confusion */
117 	pbl_ba = mr->pbl_mtr.hem_cfg.root_ba;
118 	rc_sq_wqe->msg_len = cpu_to_le32(lower_32_bits(pbl_ba));
119 	rc_sq_wqe->inv_key = cpu_to_le32(upper_32_bits(pbl_ba));
120 
121 	rc_sq_wqe->byte_16 = cpu_to_le32(wr->mr->length & 0xffffffff);
122 	rc_sq_wqe->byte_20 = cpu_to_le32(wr->mr->length >> 32);
123 	rc_sq_wqe->rkey = cpu_to_le32(wr->key);
124 	rc_sq_wqe->va = cpu_to_le64(wr->mr->iova);
125 
126 	hr_reg_write(fseg, FRMR_PBL_SIZE, mr->npages);
127 	hr_reg_write(fseg, FRMR_PBL_BUF_PG_SZ,
128 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
129 	hr_reg_clear(fseg, FRMR_BLK_MODE);
130 }
131 
132 static void set_atomic_seg(const struct ib_send_wr *wr,
133 			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
134 			   unsigned int valid_num_sge)
135 {
136 	struct hns_roce_v2_wqe_data_seg *dseg =
137 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
138 	struct hns_roce_wqe_atomic_seg *aseg =
139 		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
140 
141 	set_data_seg_v2(dseg, wr->sg_list);
142 
143 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
144 		aseg->fetchadd_swap_data = cpu_to_le64(atomic_wr(wr)->swap);
145 		aseg->cmp_data = cpu_to_le64(atomic_wr(wr)->compare_add);
146 	} else {
147 		aseg->fetchadd_swap_data =
148 			cpu_to_le64(atomic_wr(wr)->compare_add);
149 		aseg->cmp_data = 0;
150 	}
151 
152 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
153 }
154 
155 static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
156 				 const struct ib_send_wr *wr,
157 				 unsigned int *sge_idx, u32 msg_len)
158 {
159 	struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
160 	unsigned int dseg_len = sizeof(struct hns_roce_v2_wqe_data_seg);
161 	unsigned int ext_sge_sz = qp->sq.max_gs * dseg_len;
162 	unsigned int left_len_in_pg;
163 	unsigned int idx = *sge_idx;
164 	unsigned int i = 0;
165 	unsigned int len;
166 	void *addr;
167 	void *dseg;
168 
169 	if (msg_len > ext_sge_sz) {
170 		ibdev_err(ibdev,
171 			  "no enough extended sge space for inline data.\n");
172 		return -EINVAL;
173 	}
174 
175 	dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
176 	left_len_in_pg = hr_hw_page_align((uintptr_t)dseg) - (uintptr_t)dseg;
177 	len = wr->sg_list[0].length;
178 	addr = (void *)(unsigned long)(wr->sg_list[0].addr);
179 
180 	/* When copying data to extended sge space, the left length in page may
181 	 * not long enough for current user's sge. So the data should be
182 	 * splited into several parts, one in the first page, and the others in
183 	 * the subsequent pages.
184 	 */
185 	while (1) {
186 		if (len <= left_len_in_pg) {
187 			memcpy(dseg, addr, len);
188 
189 			idx += len / dseg_len;
190 
191 			i++;
192 			if (i >= wr->num_sge)
193 				break;
194 
195 			left_len_in_pg -= len;
196 			len = wr->sg_list[i].length;
197 			addr = (void *)(unsigned long)(wr->sg_list[i].addr);
198 			dseg += len;
199 		} else {
200 			memcpy(dseg, addr, left_len_in_pg);
201 
202 			len -= left_len_in_pg;
203 			addr += left_len_in_pg;
204 			idx += left_len_in_pg / dseg_len;
205 			dseg = hns_roce_get_extend_sge(qp,
206 						idx & (qp->sge.sge_cnt - 1));
207 			left_len_in_pg = 1 << HNS_HW_PAGE_SHIFT;
208 		}
209 	}
210 
211 	*sge_idx = idx;
212 
213 	return 0;
214 }
215 
216 static void set_extend_sge(struct hns_roce_qp *qp, struct ib_sge *sge,
217 			   unsigned int *sge_ind, unsigned int cnt)
218 {
219 	struct hns_roce_v2_wqe_data_seg *dseg;
220 	unsigned int idx = *sge_ind;
221 
222 	while (cnt > 0) {
223 		dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
224 		if (likely(sge->length)) {
225 			set_data_seg_v2(dseg, sge);
226 			idx++;
227 			cnt--;
228 		}
229 		sge++;
230 	}
231 
232 	*sge_ind = idx;
233 }
234 
235 static bool check_inl_data_len(struct hns_roce_qp *qp, unsigned int len)
236 {
237 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
238 	int mtu = ib_mtu_enum_to_int(qp->path_mtu);
239 
240 	if (len > qp->max_inline_data || len > mtu) {
241 		ibdev_err(&hr_dev->ib_dev,
242 			  "invalid length of data, data len = %u, max inline len = %u, path mtu = %d.\n",
243 			  len, qp->max_inline_data, mtu);
244 		return false;
245 	}
246 
247 	return true;
248 }
249 
250 static int set_rc_inl(struct hns_roce_qp *qp, const struct ib_send_wr *wr,
251 		      struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
252 		      unsigned int *sge_idx)
253 {
254 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
255 	u32 msg_len = le32_to_cpu(rc_sq_wqe->msg_len);
256 	struct ib_device *ibdev = &hr_dev->ib_dev;
257 	unsigned int curr_idx = *sge_idx;
258 	void *dseg = rc_sq_wqe;
259 	unsigned int i;
260 	int ret;
261 
262 	if (unlikely(wr->opcode == IB_WR_RDMA_READ)) {
263 		ibdev_err(ibdev, "invalid inline parameters!\n");
264 		return -EINVAL;
265 	}
266 
267 	if (!check_inl_data_len(qp, msg_len))
268 		return -EINVAL;
269 
270 	dseg += sizeof(struct hns_roce_v2_rc_send_wqe);
271 
272 	if (msg_len <= HNS_ROCE_V2_MAX_RC_INL_INN_SZ) {
273 		hr_reg_clear(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
274 
275 		for (i = 0; i < wr->num_sge; i++) {
276 			memcpy(dseg, ((void *)wr->sg_list[i].addr),
277 			       wr->sg_list[i].length);
278 			dseg += wr->sg_list[i].length;
279 		}
280 	} else {
281 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
282 
283 		ret = fill_ext_sge_inl_data(qp, wr, &curr_idx, msg_len);
284 		if (ret)
285 			return ret;
286 
287 		hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, curr_idx - *sge_idx);
288 	}
289 
290 	*sge_idx = curr_idx;
291 
292 	return 0;
293 }
294 
295 static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
296 			     struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
297 			     unsigned int *sge_ind,
298 			     unsigned int valid_num_sge)
299 {
300 	struct hns_roce_v2_wqe_data_seg *dseg =
301 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
302 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
303 	int j = 0;
304 	int i;
305 
306 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_MSG_START_SGE_IDX,
307 		     (*sge_ind) & (qp->sge.sge_cnt - 1));
308 
309 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_INLINE,
310 		     !!(wr->send_flags & IB_SEND_INLINE));
311 	if (wr->send_flags & IB_SEND_INLINE)
312 		return set_rc_inl(qp, wr, rc_sq_wqe, sge_ind);
313 
314 	if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
315 		for (i = 0; i < wr->num_sge; i++) {
316 			if (likely(wr->sg_list[i].length)) {
317 				set_data_seg_v2(dseg, wr->sg_list + i);
318 				dseg++;
319 			}
320 		}
321 	} else {
322 		for (i = 0; i < wr->num_sge && j < HNS_ROCE_SGE_IN_WQE; i++) {
323 			if (likely(wr->sg_list[i].length)) {
324 				set_data_seg_v2(dseg, wr->sg_list + i);
325 				dseg++;
326 				j++;
327 			}
328 		}
329 
330 		set_extend_sge(qp, wr->sg_list + i, sge_ind,
331 			       valid_num_sge - HNS_ROCE_SGE_IN_WQE);
332 	}
333 
334 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
335 
336 	return 0;
337 }
338 
339 static int check_send_valid(struct hns_roce_dev *hr_dev,
340 			    struct hns_roce_qp *hr_qp)
341 {
342 	struct ib_device *ibdev = &hr_dev->ib_dev;
343 	struct ib_qp *ibqp = &hr_qp->ibqp;
344 
345 	if (unlikely(ibqp->qp_type != IB_QPT_RC &&
346 		     ibqp->qp_type != IB_QPT_GSI &&
347 		     ibqp->qp_type != IB_QPT_UD)) {
348 		ibdev_err(ibdev, "Not supported QP(0x%x)type!\n",
349 			  ibqp->qp_type);
350 		return -EOPNOTSUPP;
351 	} else if (unlikely(hr_qp->state == IB_QPS_RESET ||
352 		   hr_qp->state == IB_QPS_INIT ||
353 		   hr_qp->state == IB_QPS_RTR)) {
354 		ibdev_err(ibdev, "failed to post WQE, QP state %u!\n",
355 			  hr_qp->state);
356 		return -EINVAL;
357 	} else if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)) {
358 		ibdev_err(ibdev, "failed to post WQE, dev state %d!\n",
359 			  hr_dev->state);
360 		return -EIO;
361 	}
362 
363 	return 0;
364 }
365 
366 static unsigned int calc_wr_sge_num(const struct ib_send_wr *wr,
367 				    unsigned int *sge_len)
368 {
369 	unsigned int valid_num = 0;
370 	unsigned int len = 0;
371 	int i;
372 
373 	for (i = 0; i < wr->num_sge; i++) {
374 		if (likely(wr->sg_list[i].length)) {
375 			len += wr->sg_list[i].length;
376 			valid_num++;
377 		}
378 	}
379 
380 	*sge_len = len;
381 	return valid_num;
382 }
383 
384 static __le32 get_immtdata(const struct ib_send_wr *wr)
385 {
386 	switch (wr->opcode) {
387 	case IB_WR_SEND_WITH_IMM:
388 	case IB_WR_RDMA_WRITE_WITH_IMM:
389 		return cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
390 	default:
391 		return 0;
392 	}
393 }
394 
395 static int set_ud_opcode(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
396 			 const struct ib_send_wr *wr)
397 {
398 	u32 ib_op = wr->opcode;
399 
400 	if (ib_op != IB_WR_SEND && ib_op != IB_WR_SEND_WITH_IMM)
401 		return -EINVAL;
402 
403 	ud_sq_wqe->immtdata = get_immtdata(wr);
404 
405 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
406 
407 	return 0;
408 }
409 
410 static int fill_ud_av(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
411 		      struct hns_roce_ah *ah)
412 {
413 	struct ib_device *ib_dev = ah->ibah.device;
414 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
415 
416 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_UDPSPN, ah->av.udp_sport);
417 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_HOPLIMIT, ah->av.hop_limit);
418 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_TCLASS, ah->av.tclass);
419 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_FLOW_LABEL, ah->av.flowlabel);
420 
421 	if (WARN_ON(ah->av.sl > MAX_SERVICE_LEVEL))
422 		return -EINVAL;
423 
424 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SL, ah->av.sl);
425 
426 	ud_sq_wqe->sgid_index = ah->av.gid_index;
427 
428 	memcpy(ud_sq_wqe->dmac, ah->av.mac, ETH_ALEN);
429 	memcpy(ud_sq_wqe->dgid, ah->av.dgid, GID_LEN_V2);
430 
431 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
432 		return 0;
433 
434 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN_EN, ah->av.vlan_en);
435 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN, ah->av.vlan_id);
436 
437 	return 0;
438 }
439 
440 static inline int set_ud_wqe(struct hns_roce_qp *qp,
441 			     const struct ib_send_wr *wr,
442 			     void *wqe, unsigned int *sge_idx,
443 			     unsigned int owner_bit)
444 {
445 	struct hns_roce_ah *ah = to_hr_ah(ud_wr(wr)->ah);
446 	struct hns_roce_v2_ud_send_wqe *ud_sq_wqe = wqe;
447 	unsigned int curr_idx = *sge_idx;
448 	unsigned int valid_num_sge;
449 	u32 msg_len = 0;
450 	int ret;
451 
452 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
453 
454 	ret = set_ud_opcode(ud_sq_wqe, wr);
455 	if (WARN_ON(ret))
456 		return ret;
457 
458 	ud_sq_wqe->msg_len = cpu_to_le32(msg_len);
459 
460 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_CQE,
461 		     !!(wr->send_flags & IB_SEND_SIGNALED));
462 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SE,
463 		     !!(wr->send_flags & IB_SEND_SOLICITED));
464 
465 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_PD, to_hr_pd(qp->ibqp.pd)->pdn);
466 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SGE_NUM, valid_num_sge);
467 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_MSG_START_SGE_IDX,
468 		     curr_idx & (qp->sge.sge_cnt - 1));
469 
470 	ud_sq_wqe->qkey = cpu_to_le32(ud_wr(wr)->remote_qkey & 0x80000000 ?
471 			  qp->qkey : ud_wr(wr)->remote_qkey);
472 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_DQPN, ud_wr(wr)->remote_qpn);
473 
474 	ret = fill_ud_av(ud_sq_wqe, ah);
475 	if (ret)
476 		return ret;
477 
478 	qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
479 
480 	set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
481 
482 	/*
483 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
484 	 * including new WQEs waiting for the doorbell to update the PI again.
485 	 * Therefore, the owner bit of WQE MUST be updated after all fields
486 	 * and extSGEs have been written into DDR instead of cache.
487 	 */
488 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
489 		dma_wmb();
490 
491 	*sge_idx = curr_idx;
492 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OWNER, owner_bit);
493 
494 	return 0;
495 }
496 
497 static int set_rc_opcode(struct hns_roce_dev *hr_dev,
498 			 struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
499 			 const struct ib_send_wr *wr)
500 {
501 	u32 ib_op = wr->opcode;
502 	int ret = 0;
503 
504 	rc_sq_wqe->immtdata = get_immtdata(wr);
505 
506 	switch (ib_op) {
507 	case IB_WR_RDMA_READ:
508 	case IB_WR_RDMA_WRITE:
509 	case IB_WR_RDMA_WRITE_WITH_IMM:
510 		rc_sq_wqe->rkey = cpu_to_le32(rdma_wr(wr)->rkey);
511 		rc_sq_wqe->va = cpu_to_le64(rdma_wr(wr)->remote_addr);
512 		break;
513 	case IB_WR_SEND:
514 	case IB_WR_SEND_WITH_IMM:
515 		break;
516 	case IB_WR_ATOMIC_CMP_AND_SWP:
517 	case IB_WR_ATOMIC_FETCH_AND_ADD:
518 		rc_sq_wqe->rkey = cpu_to_le32(atomic_wr(wr)->rkey);
519 		rc_sq_wqe->va = cpu_to_le64(atomic_wr(wr)->remote_addr);
520 		break;
521 	case IB_WR_REG_MR:
522 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
523 			set_frmr_seg(rc_sq_wqe, reg_wr(wr));
524 		else
525 			ret = -EOPNOTSUPP;
526 		break;
527 	case IB_WR_LOCAL_INV:
528 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_SO);
529 		fallthrough;
530 	case IB_WR_SEND_WITH_INV:
531 		rc_sq_wqe->inv_key = cpu_to_le32(wr->ex.invalidate_rkey);
532 		break;
533 	default:
534 		ret = -EINVAL;
535 	}
536 
537 	if (unlikely(ret))
538 		return ret;
539 
540 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
541 
542 	return ret;
543 }
544 
545 static inline int set_rc_wqe(struct hns_roce_qp *qp,
546 			     const struct ib_send_wr *wr,
547 			     void *wqe, unsigned int *sge_idx,
548 			     unsigned int owner_bit)
549 {
550 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
551 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
552 	unsigned int curr_idx = *sge_idx;
553 	unsigned int valid_num_sge;
554 	u32 msg_len = 0;
555 	int ret;
556 
557 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
558 
559 	rc_sq_wqe->msg_len = cpu_to_le32(msg_len);
560 
561 	ret = set_rc_opcode(hr_dev, rc_sq_wqe, wr);
562 	if (WARN_ON(ret))
563 		return ret;
564 
565 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_FENCE,
566 		     (wr->send_flags & IB_SEND_FENCE) ? 1 : 0);
567 
568 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SE,
569 		     (wr->send_flags & IB_SEND_SOLICITED) ? 1 : 0);
570 
571 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_CQE,
572 		     (wr->send_flags & IB_SEND_SIGNALED) ? 1 : 0);
573 
574 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
575 	    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
576 		set_atomic_seg(wr, rc_sq_wqe, valid_num_sge);
577 	else if (wr->opcode != IB_WR_REG_MR)
578 		ret = set_rwqe_data_seg(&qp->ibqp, wr, rc_sq_wqe,
579 					&curr_idx, valid_num_sge);
580 
581 	/*
582 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
583 	 * including new WQEs waiting for the doorbell to update the PI again.
584 	 * Therefore, the owner bit of WQE MUST be updated after all fields
585 	 * and extSGEs have been written into DDR instead of cache.
586 	 */
587 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
588 		dma_wmb();
589 
590 	*sge_idx = curr_idx;
591 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OWNER, owner_bit);
592 
593 	return ret;
594 }
595 
596 static inline void update_sq_db(struct hns_roce_dev *hr_dev,
597 				struct hns_roce_qp *qp)
598 {
599 	if (unlikely(qp->state == IB_QPS_ERR)) {
600 		flush_cqe(hr_dev, qp);
601 	} else {
602 		struct hns_roce_v2_db sq_db = {};
603 
604 		hr_reg_write(&sq_db, DB_TAG, qp->doorbell_qpn);
605 		hr_reg_write(&sq_db, DB_CMD, HNS_ROCE_V2_SQ_DB);
606 		hr_reg_write(&sq_db, DB_PI, qp->sq.head);
607 		hr_reg_write(&sq_db, DB_SL, qp->sl);
608 
609 		hns_roce_write64(hr_dev, (__le32 *)&sq_db, qp->sq.db_reg);
610 	}
611 }
612 
613 static inline void update_rq_db(struct hns_roce_dev *hr_dev,
614 				struct hns_roce_qp *qp)
615 {
616 	if (unlikely(qp->state == IB_QPS_ERR)) {
617 		flush_cqe(hr_dev, qp);
618 	} else {
619 		if (likely(qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)) {
620 			*qp->rdb.db_record =
621 					qp->rq.head & V2_DB_PRODUCER_IDX_M;
622 		} else {
623 			struct hns_roce_v2_db rq_db = {};
624 
625 			hr_reg_write(&rq_db, DB_TAG, qp->qpn);
626 			hr_reg_write(&rq_db, DB_CMD, HNS_ROCE_V2_RQ_DB);
627 			hr_reg_write(&rq_db, DB_PI, qp->rq.head);
628 
629 			hns_roce_write64(hr_dev, (__le32 *)&rq_db,
630 					 qp->rq.db_reg);
631 		}
632 	}
633 }
634 
635 static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
636 			      u64 __iomem *dest)
637 {
638 #define HNS_ROCE_WRITE_TIMES 8
639 	struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
640 	struct hnae3_handle *handle = priv->handle;
641 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
642 	int i;
643 
644 	if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
645 		for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
646 			writeq_relaxed(*(val + i), dest + i);
647 }
648 
649 static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
650 		       void *wqe)
651 {
652 #define HNS_ROCE_SL_SHIFT 2
653 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
654 
655 	/* All kinds of DirectWQE have the same header field layout */
656 	hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_FLAG);
657 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_L, qp->sl);
658 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_H,
659 		     qp->sl >> HNS_ROCE_SL_SHIFT);
660 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_WQE_INDEX, qp->sq.head);
661 
662 	hns_roce_write512(hr_dev, wqe, qp->sq.db_reg);
663 }
664 
665 static int hns_roce_v2_post_send(struct ib_qp *ibqp,
666 				 const struct ib_send_wr *wr,
667 				 const struct ib_send_wr **bad_wr)
668 {
669 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
670 	struct ib_device *ibdev = &hr_dev->ib_dev;
671 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
672 	unsigned long flags = 0;
673 	unsigned int owner_bit;
674 	unsigned int sge_idx;
675 	unsigned int wqe_idx;
676 	void *wqe = NULL;
677 	u32 nreq;
678 	int ret;
679 
680 	spin_lock_irqsave(&qp->sq.lock, flags);
681 
682 	ret = check_send_valid(hr_dev, qp);
683 	if (unlikely(ret)) {
684 		*bad_wr = wr;
685 		nreq = 0;
686 		goto out;
687 	}
688 
689 	sge_idx = qp->next_sge;
690 
691 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
692 		if (hns_roce_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
693 			ret = -ENOMEM;
694 			*bad_wr = wr;
695 			goto out;
696 		}
697 
698 		wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
699 
700 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
701 			ibdev_err(ibdev, "num_sge = %d > qp->sq.max_gs = %u.\n",
702 				  wr->num_sge, qp->sq.max_gs);
703 			ret = -EINVAL;
704 			*bad_wr = wr;
705 			goto out;
706 		}
707 
708 		wqe = hns_roce_get_send_wqe(qp, wqe_idx);
709 		qp->sq.wrid[wqe_idx] = wr->wr_id;
710 		owner_bit =
711 		       ~(((qp->sq.head + nreq) >> ilog2(qp->sq.wqe_cnt)) & 0x1);
712 
713 		/* Corresponding to the QP type, wqe process separately */
714 		if (ibqp->qp_type == IB_QPT_RC)
715 			ret = set_rc_wqe(qp, wr, wqe, &sge_idx, owner_bit);
716 		else
717 			ret = set_ud_wqe(qp, wr, wqe, &sge_idx, owner_bit);
718 
719 		if (unlikely(ret)) {
720 			*bad_wr = wr;
721 			goto out;
722 		}
723 	}
724 
725 out:
726 	if (likely(nreq)) {
727 		qp->sq.head += nreq;
728 		qp->next_sge = sge_idx;
729 
730 		if (nreq == 1 && (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
731 			write_dwqe(hr_dev, qp, wqe);
732 		else
733 			update_sq_db(hr_dev, qp);
734 	}
735 
736 	spin_unlock_irqrestore(&qp->sq.lock, flags);
737 
738 	return ret;
739 }
740 
741 static int check_recv_valid(struct hns_roce_dev *hr_dev,
742 			    struct hns_roce_qp *hr_qp)
743 {
744 	struct ib_device *ibdev = &hr_dev->ib_dev;
745 	struct ib_qp *ibqp = &hr_qp->ibqp;
746 
747 	if (unlikely(ibqp->qp_type != IB_QPT_RC &&
748 		     ibqp->qp_type != IB_QPT_GSI &&
749 		     ibqp->qp_type != IB_QPT_UD)) {
750 		ibdev_err(ibdev, "unsupported qp type, qp_type = %d.\n",
751 			  ibqp->qp_type);
752 		return -EOPNOTSUPP;
753 	}
754 
755 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
756 		return -EIO;
757 
758 	if (hr_qp->state == IB_QPS_RESET)
759 		return -EINVAL;
760 
761 	return 0;
762 }
763 
764 static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
765 				 u32 max_sge, bool rsv)
766 {
767 	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
768 	u32 i, cnt;
769 
770 	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
771 		/* Skip zero-length sge */
772 		if (!wr->sg_list[i].length)
773 			continue;
774 		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
775 		cnt++;
776 	}
777 
778 	/* Fill a reserved sge to make hw stop reading remaining segments */
779 	if (rsv) {
780 		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
781 		dseg[cnt].addr = 0;
782 		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
783 	} else {
784 		/* Clear remaining segments to make ROCEE ignore sges */
785 		if (cnt < max_sge)
786 			memset(dseg + cnt, 0,
787 			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
788 	}
789 }
790 
791 static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
792 			u32 wqe_idx, u32 max_sge)
793 {
794 	struct hns_roce_rinl_sge *sge_list;
795 	void *wqe = NULL;
796 	u32 i;
797 
798 	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
799 	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
800 
801 	/* rq support inline data */
802 	if (hr_qp->rq_inl_buf.wqe_cnt) {
803 		sge_list = hr_qp->rq_inl_buf.wqe_list[wqe_idx].sg_list;
804 		hr_qp->rq_inl_buf.wqe_list[wqe_idx].sge_cnt = (u32)wr->num_sge;
805 		for (i = 0; i < wr->num_sge; i++) {
806 			sge_list[i].addr = (void *)(u64)wr->sg_list[i].addr;
807 			sge_list[i].len = wr->sg_list[i].length;
808 		}
809 	}
810 }
811 
812 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
813 				 const struct ib_recv_wr *wr,
814 				 const struct ib_recv_wr **bad_wr)
815 {
816 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
817 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
818 	struct ib_device *ibdev = &hr_dev->ib_dev;
819 	u32 wqe_idx, nreq, max_sge;
820 	unsigned long flags;
821 	int ret;
822 
823 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
824 
825 	ret = check_recv_valid(hr_dev, hr_qp);
826 	if (unlikely(ret)) {
827 		*bad_wr = wr;
828 		nreq = 0;
829 		goto out;
830 	}
831 
832 	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
833 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
834 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
835 						  hr_qp->ibqp.recv_cq))) {
836 			ret = -ENOMEM;
837 			*bad_wr = wr;
838 			goto out;
839 		}
840 
841 		if (unlikely(wr->num_sge > max_sge)) {
842 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
843 				  wr->num_sge, max_sge);
844 			ret = -EINVAL;
845 			*bad_wr = wr;
846 			goto out;
847 		}
848 
849 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
850 		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
851 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
852 	}
853 
854 out:
855 	if (likely(nreq)) {
856 		hr_qp->rq.head += nreq;
857 
858 		update_rq_db(hr_dev, hr_qp);
859 	}
860 	spin_unlock_irqrestore(&hr_qp->rq.lock, flags);
861 
862 	return ret;
863 }
864 
865 static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
866 {
867 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
868 }
869 
870 static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
871 {
872 	return hns_roce_buf_offset(idx_que->mtr.kmem,
873 				   n << idx_que->entry_shift);
874 }
875 
876 static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
877 {
878 	/* always called with interrupts disabled. */
879 	spin_lock(&srq->lock);
880 
881 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
882 	srq->idx_que.tail++;
883 
884 	spin_unlock(&srq->lock);
885 }
886 
887 static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
888 {
889 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
890 
891 	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
892 }
893 
894 static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
895 				const struct ib_recv_wr *wr)
896 {
897 	struct ib_device *ib_dev = srq->ibsrq.device;
898 
899 	if (unlikely(wr->num_sge > max_sge)) {
900 		ibdev_err(ib_dev,
901 			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
902 			  wr->num_sge, max_sge);
903 		return -EINVAL;
904 	}
905 
906 	if (unlikely(hns_roce_srqwq_overflow(srq))) {
907 		ibdev_err(ib_dev,
908 			  "failed to check srqwq status, srqwq is full.\n");
909 		return -ENOMEM;
910 	}
911 
912 	return 0;
913 }
914 
915 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
916 {
917 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
918 	u32 pos;
919 
920 	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
921 	if (unlikely(pos == srq->wqe_cnt))
922 		return -ENOSPC;
923 
924 	bitmap_set(idx_que->bitmap, pos, 1);
925 	*wqe_idx = pos;
926 	return 0;
927 }
928 
929 static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
930 {
931 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
932 	unsigned int head;
933 	__le32 *buf;
934 
935 	head = idx_que->head & (srq->wqe_cnt - 1);
936 
937 	buf = get_idx_buf(idx_que, head);
938 	*buf = cpu_to_le32(wqe_idx);
939 
940 	idx_que->head++;
941 }
942 
943 static void update_srq_db(struct hns_roce_v2_db *db, struct hns_roce_srq *srq)
944 {
945 	hr_reg_write(db, DB_TAG, srq->srqn);
946 	hr_reg_write(db, DB_CMD, HNS_ROCE_V2_SRQ_DB);
947 	hr_reg_write(db, DB_PI, srq->idx_que.head);
948 }
949 
950 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
951 				     const struct ib_recv_wr *wr,
952 				     const struct ib_recv_wr **bad_wr)
953 {
954 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
955 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
956 	struct hns_roce_v2_db srq_db;
957 	unsigned long flags;
958 	int ret = 0;
959 	u32 max_sge;
960 	u32 wqe_idx;
961 	void *wqe;
962 	u32 nreq;
963 
964 	spin_lock_irqsave(&srq->lock, flags);
965 
966 	max_sge = srq->max_gs - srq->rsv_sge;
967 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
968 		ret = check_post_srq_valid(srq, max_sge, wr);
969 		if (ret) {
970 			*bad_wr = wr;
971 			break;
972 		}
973 
974 		ret = get_srq_wqe_idx(srq, &wqe_idx);
975 		if (unlikely(ret)) {
976 			*bad_wr = wr;
977 			break;
978 		}
979 
980 		wqe = get_srq_wqe_buf(srq, wqe_idx);
981 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
982 		fill_wqe_idx(srq, wqe_idx);
983 		srq->wrid[wqe_idx] = wr->wr_id;
984 	}
985 
986 	if (likely(nreq)) {
987 		update_srq_db(&srq_db, srq);
988 
989 		hns_roce_write64(hr_dev, (__le32 *)&srq_db, srq->db_reg);
990 	}
991 
992 	spin_unlock_irqrestore(&srq->lock, flags);
993 
994 	return ret;
995 }
996 
997 static u32 hns_roce_v2_cmd_hw_reseted(struct hns_roce_dev *hr_dev,
998 				      unsigned long instance_stage,
999 				      unsigned long reset_stage)
1000 {
1001 	/* When hardware reset has been completed once or more, we should stop
1002 	 * sending mailbox&cmq&doorbell to hardware. If now in .init_instance()
1003 	 * function, we should exit with error. If now at HNAE3_INIT_CLIENT
1004 	 * stage of soft reset process, we should exit with error, and then
1005 	 * HNAE3_INIT_CLIENT related process can rollback the operation like
1006 	 * notifing hardware to free resources, HNAE3_INIT_CLIENT related
1007 	 * process will exit with error to notify NIC driver to reschedule soft
1008 	 * reset process once again.
1009 	 */
1010 	hr_dev->is_reset = true;
1011 	hr_dev->dis_db = true;
1012 
1013 	if (reset_stage == HNS_ROCE_STATE_RST_INIT ||
1014 	    instance_stage == HNS_ROCE_STATE_INIT)
1015 		return CMD_RST_PRC_EBUSY;
1016 
1017 	return CMD_RST_PRC_SUCCESS;
1018 }
1019 
1020 static u32 hns_roce_v2_cmd_hw_resetting(struct hns_roce_dev *hr_dev,
1021 					unsigned long instance_stage,
1022 					unsigned long reset_stage)
1023 {
1024 #define HW_RESET_TIMEOUT_US 1000000
1025 #define HW_RESET_SLEEP_US 1000
1026 
1027 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1028 	struct hnae3_handle *handle = priv->handle;
1029 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1030 	unsigned long val;
1031 	int ret;
1032 
1033 	/* When hardware reset is detected, we should stop sending mailbox&cmq&
1034 	 * doorbell to hardware. If now in .init_instance() function, we should
1035 	 * exit with error. If now at HNAE3_INIT_CLIENT stage of soft reset
1036 	 * process, we should exit with error, and then HNAE3_INIT_CLIENT
1037 	 * related process can rollback the operation like notifing hardware to
1038 	 * free resources, HNAE3_INIT_CLIENT related process will exit with
1039 	 * error to notify NIC driver to reschedule soft reset process once
1040 	 * again.
1041 	 */
1042 	hr_dev->dis_db = true;
1043 
1044 	ret = read_poll_timeout(ops->ae_dev_reset_cnt, val,
1045 				val > hr_dev->reset_cnt, HW_RESET_SLEEP_US,
1046 				HW_RESET_TIMEOUT_US, false, handle);
1047 	if (!ret)
1048 		hr_dev->is_reset = true;
1049 
1050 	if (!hr_dev->is_reset || reset_stage == HNS_ROCE_STATE_RST_INIT ||
1051 	    instance_stage == HNS_ROCE_STATE_INIT)
1052 		return CMD_RST_PRC_EBUSY;
1053 
1054 	return CMD_RST_PRC_SUCCESS;
1055 }
1056 
1057 static u32 hns_roce_v2_cmd_sw_resetting(struct hns_roce_dev *hr_dev)
1058 {
1059 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1060 	struct hnae3_handle *handle = priv->handle;
1061 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1062 
1063 	/* When software reset is detected at .init_instance() function, we
1064 	 * should stop sending mailbox&cmq&doorbell to hardware, and exit
1065 	 * with error.
1066 	 */
1067 	hr_dev->dis_db = true;
1068 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt)
1069 		hr_dev->is_reset = true;
1070 
1071 	return CMD_RST_PRC_EBUSY;
1072 }
1073 
1074 static u32 check_aedev_reset_status(struct hns_roce_dev *hr_dev,
1075 				    struct hnae3_handle *handle)
1076 {
1077 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1078 	unsigned long instance_stage; /* the current instance stage */
1079 	unsigned long reset_stage; /* the current reset stage */
1080 	unsigned long reset_cnt;
1081 	bool sw_resetting;
1082 	bool hw_resetting;
1083 
1084 	/* Get information about reset from NIC driver or RoCE driver itself,
1085 	 * the meaning of the following variables from NIC driver are described
1086 	 * as below:
1087 	 * reset_cnt -- The count value of completed hardware reset.
1088 	 * hw_resetting -- Whether hardware device is resetting now.
1089 	 * sw_resetting -- Whether NIC's software reset process is running now.
1090 	 */
1091 	instance_stage = handle->rinfo.instance_state;
1092 	reset_stage = handle->rinfo.reset_state;
1093 	reset_cnt = ops->ae_dev_reset_cnt(handle);
1094 	if (reset_cnt != hr_dev->reset_cnt)
1095 		return hns_roce_v2_cmd_hw_reseted(hr_dev, instance_stage,
1096 						  reset_stage);
1097 
1098 	hw_resetting = ops->get_cmdq_stat(handle);
1099 	if (hw_resetting)
1100 		return hns_roce_v2_cmd_hw_resetting(hr_dev, instance_stage,
1101 						    reset_stage);
1102 
1103 	sw_resetting = ops->ae_dev_resetting(handle);
1104 	if (sw_resetting && instance_stage == HNS_ROCE_STATE_INIT)
1105 		return hns_roce_v2_cmd_sw_resetting(hr_dev);
1106 
1107 	return CMD_RST_PRC_OTHERS;
1108 }
1109 
1110 static bool check_device_is_in_reset(struct hns_roce_dev *hr_dev)
1111 {
1112 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1113 	struct hnae3_handle *handle = priv->handle;
1114 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1115 
1116 	if (hr_dev->reset_cnt != ops->ae_dev_reset_cnt(handle))
1117 		return true;
1118 
1119 	if (ops->get_hw_reset_stat(handle))
1120 		return true;
1121 
1122 	if (ops->ae_dev_resetting(handle))
1123 		return true;
1124 
1125 	return false;
1126 }
1127 
1128 static bool v2_chk_mbox_is_avail(struct hns_roce_dev *hr_dev, bool *busy)
1129 {
1130 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1131 	u32 status;
1132 
1133 	if (hr_dev->is_reset)
1134 		status = CMD_RST_PRC_SUCCESS;
1135 	else
1136 		status = check_aedev_reset_status(hr_dev, priv->handle);
1137 
1138 	*busy = (status == CMD_RST_PRC_EBUSY);
1139 
1140 	return status == CMD_RST_PRC_OTHERS;
1141 }
1142 
1143 static int hns_roce_alloc_cmq_desc(struct hns_roce_dev *hr_dev,
1144 				   struct hns_roce_v2_cmq_ring *ring)
1145 {
1146 	int size = ring->desc_num * sizeof(struct hns_roce_cmq_desc);
1147 
1148 	ring->desc = dma_alloc_coherent(hr_dev->dev, size,
1149 					&ring->desc_dma_addr, GFP_KERNEL);
1150 	if (!ring->desc)
1151 		return -ENOMEM;
1152 
1153 	return 0;
1154 }
1155 
1156 static void hns_roce_free_cmq_desc(struct hns_roce_dev *hr_dev,
1157 				   struct hns_roce_v2_cmq_ring *ring)
1158 {
1159 	dma_free_coherent(hr_dev->dev,
1160 			  ring->desc_num * sizeof(struct hns_roce_cmq_desc),
1161 			  ring->desc, ring->desc_dma_addr);
1162 
1163 	ring->desc_dma_addr = 0;
1164 }
1165 
1166 static int init_csq(struct hns_roce_dev *hr_dev,
1167 		    struct hns_roce_v2_cmq_ring *csq)
1168 {
1169 	dma_addr_t dma;
1170 	int ret;
1171 
1172 	csq->desc_num = CMD_CSQ_DESC_NUM;
1173 	spin_lock_init(&csq->lock);
1174 	csq->flag = TYPE_CSQ;
1175 	csq->head = 0;
1176 
1177 	ret = hns_roce_alloc_cmq_desc(hr_dev, csq);
1178 	if (ret)
1179 		return ret;
1180 
1181 	dma = csq->desc_dma_addr;
1182 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_L_REG, lower_32_bits(dma));
1183 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_H_REG, upper_32_bits(dma));
1184 	roce_write(hr_dev, ROCEE_TX_CMQ_DEPTH_REG,
1185 		   (u32)csq->desc_num >> HNS_ROCE_CMQ_DESC_NUM_S);
1186 
1187 	/* Make sure to write CI first and then PI */
1188 	roce_write(hr_dev, ROCEE_TX_CMQ_CI_REG, 0);
1189 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, 0);
1190 
1191 	return 0;
1192 }
1193 
1194 static int hns_roce_v2_cmq_init(struct hns_roce_dev *hr_dev)
1195 {
1196 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1197 	int ret;
1198 
1199 	priv->cmq.tx_timeout = HNS_ROCE_CMQ_TX_TIMEOUT;
1200 
1201 	ret = init_csq(hr_dev, &priv->cmq.csq);
1202 	if (ret)
1203 		dev_err(hr_dev->dev, "failed to init CSQ, ret = %d.\n", ret);
1204 
1205 	return ret;
1206 }
1207 
1208 static void hns_roce_v2_cmq_exit(struct hns_roce_dev *hr_dev)
1209 {
1210 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1211 
1212 	hns_roce_free_cmq_desc(hr_dev, &priv->cmq.csq);
1213 }
1214 
1215 static void hns_roce_cmq_setup_basic_desc(struct hns_roce_cmq_desc *desc,
1216 					  enum hns_roce_opcode_type opcode,
1217 					  bool is_read)
1218 {
1219 	memset((void *)desc, 0, sizeof(struct hns_roce_cmq_desc));
1220 	desc->opcode = cpu_to_le16(opcode);
1221 	desc->flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1222 	if (is_read)
1223 		desc->flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_WR);
1224 	else
1225 		desc->flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1226 }
1227 
1228 static int hns_roce_cmq_csq_done(struct hns_roce_dev *hr_dev)
1229 {
1230 	u32 tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1231 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1232 
1233 	return tail == priv->cmq.csq.head;
1234 }
1235 
1236 static void update_cmdq_status(struct hns_roce_dev *hr_dev)
1237 {
1238 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1239 	struct hnae3_handle *handle = priv->handle;
1240 
1241 	if (handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
1242 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT)
1243 		hr_dev->cmd.state = HNS_ROCE_CMDQ_STATE_FATAL_ERR;
1244 }
1245 
1246 static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1247 			       struct hns_roce_cmq_desc *desc, int num)
1248 {
1249 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1250 	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
1251 	u32 timeout = 0;
1252 	u16 desc_ret;
1253 	u32 tail;
1254 	int ret;
1255 	int i;
1256 
1257 	spin_lock_bh(&csq->lock);
1258 
1259 	tail = csq->head;
1260 
1261 	for (i = 0; i < num; i++) {
1262 		csq->desc[csq->head++] = desc[i];
1263 		if (csq->head == csq->desc_num)
1264 			csq->head = 0;
1265 	}
1266 
1267 	/* Write to hardware */
1268 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, csq->head);
1269 
1270 	do {
1271 		if (hns_roce_cmq_csq_done(hr_dev))
1272 			break;
1273 		udelay(1);
1274 	} while (++timeout < priv->cmq.tx_timeout);
1275 
1276 	if (hns_roce_cmq_csq_done(hr_dev)) {
1277 		ret = 0;
1278 		for (i = 0; i < num; i++) {
1279 			/* check the result of hardware write back */
1280 			desc[i] = csq->desc[tail++];
1281 			if (tail == csq->desc_num)
1282 				tail = 0;
1283 
1284 			desc_ret = le16_to_cpu(desc[i].retval);
1285 			if (likely(desc_ret == CMD_EXEC_SUCCESS))
1286 				continue;
1287 
1288 			dev_err_ratelimited(hr_dev->dev,
1289 					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",
1290 					    desc->opcode, desc_ret);
1291 			ret = -EIO;
1292 		}
1293 	} else {
1294 		/* FW/HW reset or incorrect number of desc */
1295 		tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1296 		dev_warn(hr_dev->dev, "CMDQ move tail from %u to %u.\n",
1297 			 csq->head, tail);
1298 		csq->head = tail;
1299 
1300 		update_cmdq_status(hr_dev);
1301 
1302 		ret = -EAGAIN;
1303 	}
1304 
1305 	spin_unlock_bh(&csq->lock);
1306 
1307 	return ret;
1308 }
1309 
1310 static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1311 			     struct hns_roce_cmq_desc *desc, int num)
1312 {
1313 	bool busy;
1314 	int ret;
1315 
1316 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1317 		return -EIO;
1318 
1319 	if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1320 		return busy ? -EBUSY : 0;
1321 
1322 	ret = __hns_roce_cmq_send(hr_dev, desc, num);
1323 	if (ret) {
1324 		if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1325 			return busy ? -EBUSY : 0;
1326 	}
1327 
1328 	return ret;
1329 }
1330 
1331 static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev,
1332 			       dma_addr_t base_addr, u8 cmd, unsigned long tag)
1333 {
1334 	struct hns_roce_cmd_mailbox *mbox;
1335 	int ret;
1336 
1337 	mbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1338 	if (IS_ERR(mbox))
1339 		return PTR_ERR(mbox);
1340 
1341 	ret = hns_roce_cmd_mbox(hr_dev, base_addr, mbox->dma, cmd, tag);
1342 	hns_roce_free_cmd_mailbox(hr_dev, mbox);
1343 	return ret;
1344 }
1345 
1346 static int hns_roce_cmq_query_hw_info(struct hns_roce_dev *hr_dev)
1347 {
1348 	struct hns_roce_query_version *resp;
1349 	struct hns_roce_cmq_desc desc;
1350 	int ret;
1351 
1352 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_HW_VER, true);
1353 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1354 	if (ret)
1355 		return ret;
1356 
1357 	resp = (struct hns_roce_query_version *)desc.data;
1358 	hr_dev->hw_rev = le16_to_cpu(resp->rocee_hw_version);
1359 	hr_dev->vendor_id = hr_dev->pci_dev->vendor;
1360 
1361 	return 0;
1362 }
1363 
1364 static void func_clr_hw_resetting_state(struct hns_roce_dev *hr_dev,
1365 					struct hnae3_handle *handle)
1366 {
1367 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1368 	unsigned long end;
1369 
1370 	hr_dev->dis_db = true;
1371 
1372 	dev_warn(hr_dev->dev,
1373 		 "Func clear is pending, device in resetting state.\n");
1374 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1375 	while (end) {
1376 		if (!ops->get_hw_reset_stat(handle)) {
1377 			hr_dev->is_reset = true;
1378 			dev_info(hr_dev->dev,
1379 				 "Func clear success after reset.\n");
1380 			return;
1381 		}
1382 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1383 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1384 	}
1385 
1386 	dev_warn(hr_dev->dev, "Func clear failed.\n");
1387 }
1388 
1389 static void func_clr_sw_resetting_state(struct hns_roce_dev *hr_dev,
1390 					struct hnae3_handle *handle)
1391 {
1392 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1393 	unsigned long end;
1394 
1395 	hr_dev->dis_db = true;
1396 
1397 	dev_warn(hr_dev->dev,
1398 		 "Func clear is pending, device in resetting state.\n");
1399 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1400 	while (end) {
1401 		if (ops->ae_dev_reset_cnt(handle) !=
1402 		    hr_dev->reset_cnt) {
1403 			hr_dev->is_reset = true;
1404 			dev_info(hr_dev->dev,
1405 				 "Func clear success after sw reset\n");
1406 			return;
1407 		}
1408 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1409 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1410 	}
1411 
1412 	dev_warn(hr_dev->dev, "Func clear failed because of unfinished sw reset\n");
1413 }
1414 
1415 static void hns_roce_func_clr_rst_proc(struct hns_roce_dev *hr_dev, int retval,
1416 				       int flag)
1417 {
1418 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1419 	struct hnae3_handle *handle = priv->handle;
1420 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1421 
1422 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt) {
1423 		hr_dev->dis_db = true;
1424 		hr_dev->is_reset = true;
1425 		dev_info(hr_dev->dev, "Func clear success after reset.\n");
1426 		return;
1427 	}
1428 
1429 	if (ops->get_hw_reset_stat(handle)) {
1430 		func_clr_hw_resetting_state(hr_dev, handle);
1431 		return;
1432 	}
1433 
1434 	if (ops->ae_dev_resetting(handle) &&
1435 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT) {
1436 		func_clr_sw_resetting_state(hr_dev, handle);
1437 		return;
1438 	}
1439 
1440 	if (retval && !flag)
1441 		dev_warn(hr_dev->dev,
1442 			 "Func clear read failed, ret = %d.\n", retval);
1443 
1444 	dev_warn(hr_dev->dev, "Func clear failed.\n");
1445 }
1446 
1447 static void __hns_roce_function_clear(struct hns_roce_dev *hr_dev, int vf_id)
1448 {
1449 	bool fclr_write_fail_flag = false;
1450 	struct hns_roce_func_clear *resp;
1451 	struct hns_roce_cmq_desc desc;
1452 	unsigned long end;
1453 	int ret = 0;
1454 
1455 	if (check_device_is_in_reset(hr_dev))
1456 		goto out;
1457 
1458 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR, false);
1459 	resp = (struct hns_roce_func_clear *)desc.data;
1460 	resp->rst_funcid_en = cpu_to_le32(vf_id);
1461 
1462 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1463 	if (ret) {
1464 		fclr_write_fail_flag = true;
1465 		dev_err(hr_dev->dev, "Func clear write failed, ret = %d.\n",
1466 			 ret);
1467 		goto out;
1468 	}
1469 
1470 	msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_INTERVAL);
1471 	end = HNS_ROCE_V2_FUNC_CLEAR_TIMEOUT_MSECS;
1472 	while (end) {
1473 		if (check_device_is_in_reset(hr_dev))
1474 			goto out;
1475 		msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT);
1476 		end -= HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT;
1477 
1478 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR,
1479 					      true);
1480 
1481 		resp->rst_funcid_en = cpu_to_le32(vf_id);
1482 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1483 		if (ret)
1484 			continue;
1485 
1486 		if (hr_reg_read(resp, FUNC_CLEAR_RST_FUN_DONE)) {
1487 			if (vf_id == 0)
1488 				hr_dev->is_reset = true;
1489 			return;
1490 		}
1491 	}
1492 
1493 out:
1494 	hns_roce_func_clr_rst_proc(hr_dev, ret, fclr_write_fail_flag);
1495 }
1496 
1497 static int hns_roce_free_vf_resource(struct hns_roce_dev *hr_dev, int vf_id)
1498 {
1499 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1500 	struct hns_roce_cmq_desc desc[2];
1501 	struct hns_roce_cmq_req *req_a;
1502 
1503 	req_a = (struct hns_roce_cmq_req *)desc[0].data;
1504 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1505 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1506 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1507 	hr_reg_write(req_a, FUNC_RES_A_VF_ID, vf_id);
1508 
1509 	return hns_roce_cmq_send(hr_dev, desc, 2);
1510 }
1511 
1512 static void hns_roce_function_clear(struct hns_roce_dev *hr_dev)
1513 {
1514 	int ret;
1515 	int i;
1516 
1517 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1518 		return;
1519 
1520 	for (i = hr_dev->func_num - 1; i >= 0; i--) {
1521 		__hns_roce_function_clear(hr_dev, i);
1522 
1523 		if (i == 0)
1524 			continue;
1525 
1526 		ret = hns_roce_free_vf_resource(hr_dev, i);
1527 		if (ret)
1528 			ibdev_err(&hr_dev->ib_dev,
1529 				  "failed to free vf resource, vf_id = %d, ret = %d.\n",
1530 				  i, ret);
1531 	}
1532 }
1533 
1534 static int hns_roce_clear_extdb_list_info(struct hns_roce_dev *hr_dev)
1535 {
1536 	struct hns_roce_cmq_desc desc;
1537 	int ret;
1538 
1539 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLEAR_EXTDB_LIST_INFO,
1540 				      false);
1541 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1542 	if (ret)
1543 		ibdev_err(&hr_dev->ib_dev,
1544 			  "failed to clear extended doorbell info, ret = %d.\n",
1545 			  ret);
1546 
1547 	return ret;
1548 }
1549 
1550 static int hns_roce_query_fw_ver(struct hns_roce_dev *hr_dev)
1551 {
1552 	struct hns_roce_query_fw_info *resp;
1553 	struct hns_roce_cmq_desc desc;
1554 	int ret;
1555 
1556 	hns_roce_cmq_setup_basic_desc(&desc, HNS_QUERY_FW_VER, true);
1557 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1558 	if (ret)
1559 		return ret;
1560 
1561 	resp = (struct hns_roce_query_fw_info *)desc.data;
1562 	hr_dev->caps.fw_ver = (u64)(le32_to_cpu(resp->fw_ver));
1563 
1564 	return 0;
1565 }
1566 
1567 static int hns_roce_query_func_info(struct hns_roce_dev *hr_dev)
1568 {
1569 	struct hns_roce_cmq_desc desc;
1570 	int ret;
1571 
1572 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
1573 		hr_dev->func_num = 1;
1574 		return 0;
1575 	}
1576 
1577 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_FUNC_INFO,
1578 				      true);
1579 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1580 	if (ret) {
1581 		hr_dev->func_num = 1;
1582 		return ret;
1583 	}
1584 
1585 	hr_dev->func_num = le32_to_cpu(desc.func_info.own_func_num);
1586 	hr_dev->cong_algo_tmpl_id = le32_to_cpu(desc.func_info.own_mac_id);
1587 
1588 	return 0;
1589 }
1590 
1591 static int hns_roce_config_global_param(struct hns_roce_dev *hr_dev)
1592 {
1593 	struct hns_roce_cmq_desc desc;
1594 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1595 	u32 clock_cycles_of_1us;
1596 
1597 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GLOBAL_PARAM,
1598 				      false);
1599 
1600 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
1601 		clock_cycles_of_1us = HNS_ROCE_1NS_CFG;
1602 	else
1603 		clock_cycles_of_1us = HNS_ROCE_1US_CFG;
1604 
1605 	hr_reg_write(req, CFG_GLOBAL_PARAM_1US_CYCLES, clock_cycles_of_1us);
1606 	hr_reg_write(req, CFG_GLOBAL_PARAM_UDP_PORT, ROCE_V2_UDP_DPORT);
1607 
1608 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1609 }
1610 
1611 static int load_func_res_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1612 {
1613 	struct hns_roce_cmq_desc desc[2];
1614 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1615 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1616 	struct hns_roce_caps *caps = &hr_dev->caps;
1617 	enum hns_roce_opcode_type opcode;
1618 	u32 func_num;
1619 	int ret;
1620 
1621 	if (is_vf) {
1622 		opcode = HNS_ROCE_OPC_QUERY_VF_RES;
1623 		func_num = 1;
1624 	} else {
1625 		opcode = HNS_ROCE_OPC_QUERY_PF_RES;
1626 		func_num = hr_dev->func_num;
1627 	}
1628 
1629 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, true);
1630 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1631 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, true);
1632 
1633 	ret = hns_roce_cmq_send(hr_dev, desc, 2);
1634 	if (ret)
1635 		return ret;
1636 
1637 	caps->qpc_bt_num = hr_reg_read(r_a, FUNC_RES_A_QPC_BT_NUM) / func_num;
1638 	caps->srqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_SRQC_BT_NUM) / func_num;
1639 	caps->cqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_CQC_BT_NUM) / func_num;
1640 	caps->mpt_bt_num = hr_reg_read(r_a, FUNC_RES_A_MPT_BT_NUM) / func_num;
1641 	caps->eqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_EQC_BT_NUM) / func_num;
1642 	caps->smac_bt_num = hr_reg_read(r_b, FUNC_RES_B_SMAC_NUM) / func_num;
1643 	caps->sgid_bt_num = hr_reg_read(r_b, FUNC_RES_B_SGID_NUM) / func_num;
1644 	caps->sccc_bt_num = hr_reg_read(r_b, FUNC_RES_B_SCCC_BT_NUM) / func_num;
1645 
1646 	if (is_vf) {
1647 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_V_QID_NUM) / func_num;
1648 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_V_GMV_BT_NUM) /
1649 					       func_num;
1650 	} else {
1651 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_B_QID_NUM) / func_num;
1652 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_B_GMV_BT_NUM) /
1653 					       func_num;
1654 	}
1655 
1656 	return 0;
1657 }
1658 
1659 static int load_ext_cfg_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1660 {
1661 	struct hns_roce_cmq_desc desc;
1662 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1663 	struct hns_roce_caps *caps = &hr_dev->caps;
1664 	u32 func_num, qp_num;
1665 	int ret;
1666 
1667 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_EXT_CFG, true);
1668 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1669 	if (ret)
1670 		return ret;
1671 
1672 	func_num = is_vf ? 1 : max_t(u32, 1, hr_dev->func_num);
1673 	qp_num = hr_reg_read(req, EXT_CFG_QP_PI_NUM) / func_num;
1674 	caps->num_pi_qps = round_down(qp_num, HNS_ROCE_QP_BANK_NUM);
1675 
1676 	qp_num = hr_reg_read(req, EXT_CFG_QP_NUM) / func_num;
1677 	caps->num_qps = round_down(qp_num, HNS_ROCE_QP_BANK_NUM);
1678 
1679 	return 0;
1680 }
1681 
1682 static int load_pf_timer_res_caps(struct hns_roce_dev *hr_dev)
1683 {
1684 	struct hns_roce_cmq_desc desc;
1685 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1686 	struct hns_roce_caps *caps = &hr_dev->caps;
1687 	int ret;
1688 
1689 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
1690 				      true);
1691 
1692 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1693 	if (ret)
1694 		return ret;
1695 
1696 	caps->qpc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_QPC_ITEM_NUM);
1697 	caps->cqc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_CQC_ITEM_NUM);
1698 
1699 	return 0;
1700 }
1701 
1702 static int query_func_resource_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1703 {
1704 	struct device *dev = hr_dev->dev;
1705 	int ret;
1706 
1707 	ret = load_func_res_caps(hr_dev, is_vf);
1708 	if (ret) {
1709 		dev_err(dev, "failed to load res caps, ret = %d (%s).\n", ret,
1710 			is_vf ? "vf" : "pf");
1711 		return ret;
1712 	}
1713 
1714 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1715 		ret = load_ext_cfg_caps(hr_dev, is_vf);
1716 		if (ret)
1717 			dev_err(dev, "failed to load ext cfg, ret = %d (%s).\n",
1718 				ret, is_vf ? "vf" : "pf");
1719 	}
1720 
1721 	return ret;
1722 }
1723 
1724 static int hns_roce_query_pf_resource(struct hns_roce_dev *hr_dev)
1725 {
1726 	struct device *dev = hr_dev->dev;
1727 	int ret;
1728 
1729 	ret = query_func_resource_caps(hr_dev, false);
1730 	if (ret)
1731 		return ret;
1732 
1733 	ret = load_pf_timer_res_caps(hr_dev);
1734 	if (ret)
1735 		dev_err(dev, "failed to load pf timer resource, ret = %d.\n",
1736 			ret);
1737 
1738 	return ret;
1739 }
1740 
1741 static int hns_roce_query_vf_resource(struct hns_roce_dev *hr_dev)
1742 {
1743 	return query_func_resource_caps(hr_dev, true);
1744 }
1745 
1746 static int __hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
1747 					  u32 vf_id)
1748 {
1749 	struct hns_roce_vf_switch *swt;
1750 	struct hns_roce_cmq_desc desc;
1751 	int ret;
1752 
1753 	swt = (struct hns_roce_vf_switch *)desc.data;
1754 	hns_roce_cmq_setup_basic_desc(&desc, HNS_SWITCH_PARAMETER_CFG, true);
1755 	swt->rocee_sel |= cpu_to_le32(HNS_ICL_SWITCH_CMD_ROCEE_SEL);
1756 	hr_reg_write(swt, VF_SWITCH_VF_ID, vf_id);
1757 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1758 	if (ret)
1759 		return ret;
1760 
1761 	desc.flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1762 	desc.flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1763 	hr_reg_enable(swt, VF_SWITCH_ALW_LPBK);
1764 	hr_reg_clear(swt, VF_SWITCH_ALW_LCL_LPBK);
1765 	hr_reg_enable(swt, VF_SWITCH_ALW_DST_OVRD);
1766 
1767 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1768 }
1769 
1770 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev)
1771 {
1772 	u32 vf_id;
1773 	int ret;
1774 
1775 	for (vf_id = 0; vf_id < hr_dev->func_num; vf_id++) {
1776 		ret = __hns_roce_set_vf_switch_param(hr_dev, vf_id);
1777 		if (ret)
1778 			return ret;
1779 	}
1780 	return 0;
1781 }
1782 
1783 static int config_vf_hem_resource(struct hns_roce_dev *hr_dev, int vf_id)
1784 {
1785 	struct hns_roce_cmq_desc desc[2];
1786 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1787 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1788 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1789 	struct hns_roce_caps *caps = &hr_dev->caps;
1790 
1791 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1792 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1793 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1794 
1795 	hr_reg_write(r_a, FUNC_RES_A_VF_ID, vf_id);
1796 
1797 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_NUM, caps->qpc_bt_num);
1798 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_IDX, vf_id * caps->qpc_bt_num);
1799 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_NUM, caps->srqc_bt_num);
1800 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_IDX, vf_id * caps->srqc_bt_num);
1801 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_NUM, caps->cqc_bt_num);
1802 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_IDX, vf_id * caps->cqc_bt_num);
1803 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_NUM, caps->mpt_bt_num);
1804 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_IDX, vf_id * caps->mpt_bt_num);
1805 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_NUM, caps->eqc_bt_num);
1806 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_IDX, vf_id * caps->eqc_bt_num);
1807 	hr_reg_write(r_b, FUNC_RES_V_QID_NUM, caps->sl_num);
1808 	hr_reg_write(r_b, FUNC_RES_B_QID_IDX, vf_id * caps->sl_num);
1809 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_NUM, caps->sccc_bt_num);
1810 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_IDX, vf_id * caps->sccc_bt_num);
1811 
1812 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1813 		hr_reg_write(r_b, FUNC_RES_V_GMV_BT_NUM, caps->gmv_bt_num);
1814 		hr_reg_write(r_b, FUNC_RES_B_GMV_BT_IDX,
1815 			     vf_id * caps->gmv_bt_num);
1816 	} else {
1817 		hr_reg_write(r_b, FUNC_RES_B_SGID_NUM, caps->sgid_bt_num);
1818 		hr_reg_write(r_b, FUNC_RES_B_SGID_IDX,
1819 			     vf_id * caps->sgid_bt_num);
1820 		hr_reg_write(r_b, FUNC_RES_B_SMAC_NUM, caps->smac_bt_num);
1821 		hr_reg_write(r_b, FUNC_RES_B_SMAC_IDX,
1822 			     vf_id * caps->smac_bt_num);
1823 	}
1824 
1825 	return hns_roce_cmq_send(hr_dev, desc, 2);
1826 }
1827 
1828 static int config_vf_ext_resource(struct hns_roce_dev *hr_dev, u32 vf_id)
1829 {
1830 	struct hns_roce_cmq_desc desc;
1831 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1832 	struct hns_roce_caps *caps = &hr_dev->caps;
1833 
1834 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_EXT_CFG, false);
1835 
1836 	hr_reg_write(req, EXT_CFG_VF_ID, vf_id);
1837 
1838 	hr_reg_write(req, EXT_CFG_QP_PI_NUM, caps->num_pi_qps);
1839 	hr_reg_write(req, EXT_CFG_QP_PI_IDX, vf_id * caps->num_pi_qps);
1840 	hr_reg_write(req, EXT_CFG_QP_NUM, caps->num_qps);
1841 	hr_reg_write(req, EXT_CFG_QP_IDX, vf_id * caps->num_qps);
1842 
1843 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1844 }
1845 
1846 static int hns_roce_alloc_vf_resource(struct hns_roce_dev *hr_dev)
1847 {
1848 	u32 func_num = max_t(u32, 1, hr_dev->func_num);
1849 	u32 vf_id;
1850 	int ret;
1851 
1852 	for (vf_id = 0; vf_id < func_num; vf_id++) {
1853 		ret = config_vf_hem_resource(hr_dev, vf_id);
1854 		if (ret) {
1855 			dev_err(hr_dev->dev,
1856 				"failed to config vf-%u hem res, ret = %d.\n",
1857 				vf_id, ret);
1858 			return ret;
1859 		}
1860 
1861 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1862 			ret = config_vf_ext_resource(hr_dev, vf_id);
1863 			if (ret) {
1864 				dev_err(hr_dev->dev,
1865 					"failed to config vf-%u ext res, ret = %d.\n",
1866 					vf_id, ret);
1867 				return ret;
1868 			}
1869 		}
1870 	}
1871 
1872 	return 0;
1873 }
1874 
1875 static int hns_roce_v2_set_bt(struct hns_roce_dev *hr_dev)
1876 {
1877 	struct hns_roce_cmq_desc desc;
1878 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1879 	struct hns_roce_caps *caps = &hr_dev->caps;
1880 
1881 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_BT_ATTR, false);
1882 
1883 	hr_reg_write(req, CFG_BT_ATTR_QPC_BA_PGSZ,
1884 		     caps->qpc_ba_pg_sz + PG_SHIFT_OFFSET);
1885 	hr_reg_write(req, CFG_BT_ATTR_QPC_BUF_PGSZ,
1886 		     caps->qpc_buf_pg_sz + PG_SHIFT_OFFSET);
1887 	hr_reg_write(req, CFG_BT_ATTR_QPC_HOPNUM,
1888 		     to_hr_hem_hopnum(caps->qpc_hop_num, caps->num_qps));
1889 
1890 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BA_PGSZ,
1891 		     caps->srqc_ba_pg_sz + PG_SHIFT_OFFSET);
1892 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BUF_PGSZ,
1893 		     caps->srqc_buf_pg_sz + PG_SHIFT_OFFSET);
1894 	hr_reg_write(req, CFG_BT_ATTR_SRQC_HOPNUM,
1895 		     to_hr_hem_hopnum(caps->srqc_hop_num, caps->num_srqs));
1896 
1897 	hr_reg_write(req, CFG_BT_ATTR_CQC_BA_PGSZ,
1898 		     caps->cqc_ba_pg_sz + PG_SHIFT_OFFSET);
1899 	hr_reg_write(req, CFG_BT_ATTR_CQC_BUF_PGSZ,
1900 		     caps->cqc_buf_pg_sz + PG_SHIFT_OFFSET);
1901 	hr_reg_write(req, CFG_BT_ATTR_CQC_HOPNUM,
1902 		     to_hr_hem_hopnum(caps->cqc_hop_num, caps->num_cqs));
1903 
1904 	hr_reg_write(req, CFG_BT_ATTR_MPT_BA_PGSZ,
1905 		     caps->mpt_ba_pg_sz + PG_SHIFT_OFFSET);
1906 	hr_reg_write(req, CFG_BT_ATTR_MPT_BUF_PGSZ,
1907 		     caps->mpt_buf_pg_sz + PG_SHIFT_OFFSET);
1908 	hr_reg_write(req, CFG_BT_ATTR_MPT_HOPNUM,
1909 		     to_hr_hem_hopnum(caps->mpt_hop_num, caps->num_mtpts));
1910 
1911 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BA_PGSZ,
1912 		     caps->sccc_ba_pg_sz + PG_SHIFT_OFFSET);
1913 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BUF_PGSZ,
1914 		     caps->sccc_buf_pg_sz + PG_SHIFT_OFFSET);
1915 	hr_reg_write(req, CFG_BT_ATTR_SCCC_HOPNUM,
1916 		     to_hr_hem_hopnum(caps->sccc_hop_num, caps->num_qps));
1917 
1918 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1919 }
1920 
1921 /* Use default caps when hns_roce_query_pf_caps() failed or init VF profile */
1922 static void set_default_caps(struct hns_roce_dev *hr_dev)
1923 {
1924 	struct hns_roce_caps *caps = &hr_dev->caps;
1925 
1926 	caps->num_qps		= HNS_ROCE_V2_MAX_QP_NUM;
1927 	caps->max_wqes		= HNS_ROCE_V2_MAX_WQE_NUM;
1928 	caps->num_cqs		= HNS_ROCE_V2_MAX_CQ_NUM;
1929 	caps->num_srqs		= HNS_ROCE_V2_MAX_SRQ_NUM;
1930 	caps->min_cqes		= HNS_ROCE_MIN_CQE_NUM;
1931 	caps->max_cqes		= HNS_ROCE_V2_MAX_CQE_NUM;
1932 	caps->max_sq_sg		= HNS_ROCE_V2_MAX_SQ_SGE_NUM;
1933 	caps->max_extend_sg	= HNS_ROCE_V2_MAX_EXTEND_SGE_NUM;
1934 	caps->max_rq_sg		= HNS_ROCE_V2_MAX_RQ_SGE_NUM;
1935 
1936 	caps->num_uars		= HNS_ROCE_V2_UAR_NUM;
1937 	caps->phy_num_uars	= HNS_ROCE_V2_PHY_UAR_NUM;
1938 	caps->num_aeq_vectors	= HNS_ROCE_V2_AEQE_VEC_NUM;
1939 	caps->num_other_vectors = HNS_ROCE_V2_ABNORMAL_VEC_NUM;
1940 	caps->num_comp_vectors	= 0;
1941 
1942 	caps->num_mtpts		= HNS_ROCE_V2_MAX_MTPT_NUM;
1943 	caps->num_pds		= HNS_ROCE_V2_MAX_PD_NUM;
1944 	caps->num_qpc_timer	= HNS_ROCE_V2_MAX_QPC_TIMER_NUM;
1945 	caps->cqc_timer_bt_num	= HNS_ROCE_V2_MAX_CQC_TIMER_BT_NUM;
1946 
1947 	caps->max_qp_init_rdma	= HNS_ROCE_V2_MAX_QP_INIT_RDMA;
1948 	caps->max_qp_dest_rdma	= HNS_ROCE_V2_MAX_QP_DEST_RDMA;
1949 	caps->max_sq_desc_sz	= HNS_ROCE_V2_MAX_SQ_DESC_SZ;
1950 	caps->max_rq_desc_sz	= HNS_ROCE_V2_MAX_RQ_DESC_SZ;
1951 	caps->max_srq_desc_sz	= HNS_ROCE_V2_MAX_SRQ_DESC_SZ;
1952 	caps->irrl_entry_sz	= HNS_ROCE_V2_IRRL_ENTRY_SZ;
1953 	caps->trrl_entry_sz	= HNS_ROCE_V2_EXT_ATOMIC_TRRL_ENTRY_SZ;
1954 	caps->cqc_entry_sz	= HNS_ROCE_V2_CQC_ENTRY_SZ;
1955 	caps->srqc_entry_sz	= HNS_ROCE_V2_SRQC_ENTRY_SZ;
1956 	caps->mtpt_entry_sz	= HNS_ROCE_V2_MTPT_ENTRY_SZ;
1957 	caps->idx_entry_sz	= HNS_ROCE_V2_IDX_ENTRY_SZ;
1958 	caps->page_size_cap	= HNS_ROCE_V2_PAGE_SIZE_SUPPORTED;
1959 	caps->reserved_lkey	= 0;
1960 	caps->reserved_pds	= 0;
1961 	caps->reserved_mrws	= 1;
1962 	caps->reserved_uars	= 0;
1963 	caps->reserved_cqs	= 0;
1964 	caps->reserved_srqs	= 0;
1965 	caps->reserved_qps	= HNS_ROCE_V2_RSV_QPS;
1966 
1967 	caps->qpc_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
1968 	caps->srqc_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
1969 	caps->cqc_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
1970 	caps->mpt_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
1971 	caps->sccc_hop_num	= HNS_ROCE_SCCC_HOP_NUM;
1972 
1973 	caps->mtt_hop_num	= HNS_ROCE_MTT_HOP_NUM;
1974 	caps->wqe_sq_hop_num	= HNS_ROCE_SQWQE_HOP_NUM;
1975 	caps->wqe_sge_hop_num	= HNS_ROCE_EXT_SGE_HOP_NUM;
1976 	caps->wqe_rq_hop_num	= HNS_ROCE_RQWQE_HOP_NUM;
1977 	caps->cqe_hop_num	= HNS_ROCE_CQE_HOP_NUM;
1978 	caps->srqwqe_hop_num	= HNS_ROCE_SRQWQE_HOP_NUM;
1979 	caps->idx_hop_num	= HNS_ROCE_IDX_HOP_NUM;
1980 	caps->chunk_sz          = HNS_ROCE_V2_TABLE_CHUNK_SIZE;
1981 
1982 	caps->flags		= HNS_ROCE_CAP_FLAG_REREG_MR |
1983 				  HNS_ROCE_CAP_FLAG_ROCE_V1_V2 |
1984 				  HNS_ROCE_CAP_FLAG_CQ_RECORD_DB |
1985 				  HNS_ROCE_CAP_FLAG_QP_RECORD_DB;
1986 
1987 	caps->pkey_table_len[0] = 1;
1988 	caps->ceqe_depth	= HNS_ROCE_V2_COMP_EQE_NUM;
1989 	caps->aeqe_depth	= HNS_ROCE_V2_ASYNC_EQE_NUM;
1990 	caps->local_ca_ack_delay = 0;
1991 	caps->max_mtu = IB_MTU_4096;
1992 
1993 	caps->max_srq_wrs	= HNS_ROCE_V2_MAX_SRQ_WR;
1994 	caps->max_srq_sges	= HNS_ROCE_V2_MAX_SRQ_SGE;
1995 
1996 	caps->flags |= HNS_ROCE_CAP_FLAG_ATOMIC | HNS_ROCE_CAP_FLAG_MW |
1997 		       HNS_ROCE_CAP_FLAG_SRQ | HNS_ROCE_CAP_FLAG_FRMR |
1998 		       HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL | HNS_ROCE_CAP_FLAG_XRC;
1999 
2000 	caps->gid_table_len[0] = HNS_ROCE_V2_GID_INDEX_NUM;
2001 
2002 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2003 		caps->flags |= HNS_ROCE_CAP_FLAG_STASH |
2004 			       HNS_ROCE_CAP_FLAG_DIRECT_WQE;
2005 		caps->max_sq_inline = HNS_ROCE_V3_MAX_SQ_INLINE;
2006 	} else {
2007 		caps->max_sq_inline = HNS_ROCE_V2_MAX_SQ_INLINE;
2008 
2009 		/* The following configuration are only valid for HIP08 */
2010 		caps->qpc_sz = HNS_ROCE_V2_QPC_SZ;
2011 		caps->sccc_sz = HNS_ROCE_V2_SCCC_SZ;
2012 		caps->cqe_sz = HNS_ROCE_V2_CQE_SIZE;
2013 	}
2014 }
2015 
2016 static void calc_pg_sz(u32 obj_num, u32 obj_size, u32 hop_num, u32 ctx_bt_num,
2017 		       u32 *buf_page_size, u32 *bt_page_size, u32 hem_type)
2018 {
2019 	u64 obj_per_chunk;
2020 	u64 bt_chunk_size = PAGE_SIZE;
2021 	u64 buf_chunk_size = PAGE_SIZE;
2022 	u64 obj_per_chunk_default = buf_chunk_size / obj_size;
2023 
2024 	*buf_page_size = 0;
2025 	*bt_page_size = 0;
2026 
2027 	switch (hop_num) {
2028 	case 3:
2029 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
2030 				(bt_chunk_size / BA_BYTE_LEN) *
2031 				(bt_chunk_size / BA_BYTE_LEN) *
2032 				 obj_per_chunk_default;
2033 		break;
2034 	case 2:
2035 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
2036 				(bt_chunk_size / BA_BYTE_LEN) *
2037 				 obj_per_chunk_default;
2038 		break;
2039 	case 1:
2040 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
2041 				obj_per_chunk_default;
2042 		break;
2043 	case HNS_ROCE_HOP_NUM_0:
2044 		obj_per_chunk = ctx_bt_num * obj_per_chunk_default;
2045 		break;
2046 	default:
2047 		pr_err("table %u not support hop_num = %u!\n", hem_type,
2048 		       hop_num);
2049 		return;
2050 	}
2051 
2052 	if (hem_type >= HEM_TYPE_MTT)
2053 		*bt_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2054 	else
2055 		*buf_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2056 }
2057 
2058 static void set_hem_page_size(struct hns_roce_dev *hr_dev)
2059 {
2060 	struct hns_roce_caps *caps = &hr_dev->caps;
2061 
2062 	/* EQ */
2063 	caps->eqe_ba_pg_sz = 0;
2064 	caps->eqe_buf_pg_sz = 0;
2065 
2066 	/* Link Table */
2067 	caps->llm_buf_pg_sz = 0;
2068 
2069 	/* MR */
2070 	caps->mpt_ba_pg_sz = 0;
2071 	caps->mpt_buf_pg_sz = 0;
2072 	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
2073 	caps->pbl_buf_pg_sz = 0;
2074 	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
2075 		   caps->mpt_bt_num, &caps->mpt_buf_pg_sz, &caps->mpt_ba_pg_sz,
2076 		   HEM_TYPE_MTPT);
2077 
2078 	/* QP */
2079 	caps->qpc_ba_pg_sz = 0;
2080 	caps->qpc_buf_pg_sz = 0;
2081 	caps->qpc_timer_ba_pg_sz = 0;
2082 	caps->qpc_timer_buf_pg_sz = 0;
2083 	caps->sccc_ba_pg_sz = 0;
2084 	caps->sccc_buf_pg_sz = 0;
2085 	caps->mtt_ba_pg_sz = 0;
2086 	caps->mtt_buf_pg_sz = 0;
2087 	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
2088 		   caps->qpc_bt_num, &caps->qpc_buf_pg_sz, &caps->qpc_ba_pg_sz,
2089 		   HEM_TYPE_QPC);
2090 
2091 	if (caps->flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
2092 		calc_pg_sz(caps->num_qps, caps->sccc_sz, caps->sccc_hop_num,
2093 			   caps->sccc_bt_num, &caps->sccc_buf_pg_sz,
2094 			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
2095 
2096 	/* CQ */
2097 	caps->cqc_ba_pg_sz = 0;
2098 	caps->cqc_buf_pg_sz = 0;
2099 	caps->cqc_timer_ba_pg_sz = 0;
2100 	caps->cqc_timer_buf_pg_sz = 0;
2101 	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
2102 	caps->cqe_buf_pg_sz = 0;
2103 	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
2104 		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
2105 		   HEM_TYPE_CQC);
2106 	calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
2107 		   1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
2108 
2109 	/* SRQ */
2110 	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
2111 		caps->srqc_ba_pg_sz = 0;
2112 		caps->srqc_buf_pg_sz = 0;
2113 		caps->srqwqe_ba_pg_sz = 0;
2114 		caps->srqwqe_buf_pg_sz = 0;
2115 		caps->idx_ba_pg_sz = 0;
2116 		caps->idx_buf_pg_sz = 0;
2117 		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
2118 			   caps->srqc_hop_num, caps->srqc_bt_num,
2119 			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
2120 			   HEM_TYPE_SRQC);
2121 		calc_pg_sz(caps->num_srqwqe_segs, caps->mtt_entry_sz,
2122 			   caps->srqwqe_hop_num, 1, &caps->srqwqe_buf_pg_sz,
2123 			   &caps->srqwqe_ba_pg_sz, HEM_TYPE_SRQWQE);
2124 		calc_pg_sz(caps->num_idx_segs, caps->idx_entry_sz,
2125 			   caps->idx_hop_num, 1, &caps->idx_buf_pg_sz,
2126 			   &caps->idx_ba_pg_sz, HEM_TYPE_IDX);
2127 	}
2128 
2129 	/* GMV */
2130 	caps->gmv_ba_pg_sz = 0;
2131 	caps->gmv_buf_pg_sz = 0;
2132 }
2133 
2134 /* Apply all loaded caps before setting to hardware */
2135 static void apply_func_caps(struct hns_roce_dev *hr_dev)
2136 {
2137 	struct hns_roce_caps *caps = &hr_dev->caps;
2138 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2139 
2140 	/* The following configurations don't need to be got from firmware. */
2141 	caps->qpc_timer_entry_sz = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
2142 	caps->cqc_timer_entry_sz = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
2143 	caps->mtt_entry_sz = HNS_ROCE_V2_MTT_ENTRY_SZ;
2144 
2145 	caps->pbl_hop_num = HNS_ROCE_PBL_HOP_NUM;
2146 	caps->qpc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2147 	caps->cqc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2148 
2149 	caps->num_xrcds = HNS_ROCE_V2_MAX_XRCD_NUM;
2150 	caps->reserved_xrcds = HNS_ROCE_V2_RSV_XRCD_NUM;
2151 
2152 	caps->num_mtt_segs = HNS_ROCE_V2_MAX_MTT_SEGS;
2153 	caps->num_srqwqe_segs = HNS_ROCE_V2_MAX_SRQWQE_SEGS;
2154 	caps->num_idx_segs = HNS_ROCE_V2_MAX_IDX_SEGS;
2155 
2156 	if (!caps->num_comp_vectors)
2157 		caps->num_comp_vectors = min_t(u32, caps->eqc_bt_num - 1,
2158 				  (u32)priv->handle->rinfo.num_vectors - 2);
2159 
2160 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2161 		caps->eqe_hop_num = HNS_ROCE_V3_EQE_HOP_NUM;
2162 		caps->ceqe_size = HNS_ROCE_V3_EQE_SIZE;
2163 		caps->aeqe_size = HNS_ROCE_V3_EQE_SIZE;
2164 
2165 		/* The following configurations will be overwritten */
2166 		caps->qpc_sz = HNS_ROCE_V3_QPC_SZ;
2167 		caps->cqe_sz = HNS_ROCE_V3_CQE_SIZE;
2168 		caps->sccc_sz = HNS_ROCE_V3_SCCC_SZ;
2169 
2170 		/* The following configurations are not got from firmware */
2171 		caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;
2172 
2173 		caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0;
2174 		caps->gid_table_len[0] = caps->gmv_bt_num *
2175 					(HNS_HW_PAGE_SIZE / caps->gmv_entry_sz);
2176 
2177 		caps->gmv_entry_num = caps->gmv_bt_num * (PAGE_SIZE /
2178 							  caps->gmv_entry_sz);
2179 	} else {
2180 		u32 func_num = max_t(u32, 1, hr_dev->func_num);
2181 
2182 		caps->eqe_hop_num = HNS_ROCE_V2_EQE_HOP_NUM;
2183 		caps->ceqe_size = HNS_ROCE_CEQE_SIZE;
2184 		caps->aeqe_size = HNS_ROCE_AEQE_SIZE;
2185 		caps->gid_table_len[0] /= func_num;
2186 	}
2187 
2188 	if (hr_dev->is_vf) {
2189 		caps->default_aeq_arm_st = 0x3;
2190 		caps->default_ceq_arm_st = 0x3;
2191 		caps->default_ceq_max_cnt = 0x1;
2192 		caps->default_ceq_period = 0x10;
2193 		caps->default_aeq_max_cnt = 0x1;
2194 		caps->default_aeq_period = 0x10;
2195 	}
2196 
2197 	set_hem_page_size(hr_dev);
2198 }
2199 
2200 static int hns_roce_query_pf_caps(struct hns_roce_dev *hr_dev)
2201 {
2202 	struct hns_roce_cmq_desc desc[HNS_ROCE_QUERY_PF_CAPS_CMD_NUM];
2203 	struct hns_roce_caps *caps = &hr_dev->caps;
2204 	struct hns_roce_query_pf_caps_a *resp_a;
2205 	struct hns_roce_query_pf_caps_b *resp_b;
2206 	struct hns_roce_query_pf_caps_c *resp_c;
2207 	struct hns_roce_query_pf_caps_d *resp_d;
2208 	struct hns_roce_query_pf_caps_e *resp_e;
2209 	int ctx_hop_num;
2210 	int pbl_hop_num;
2211 	int ret;
2212 	int i;
2213 
2214 	for (i = 0; i < HNS_ROCE_QUERY_PF_CAPS_CMD_NUM; i++) {
2215 		hns_roce_cmq_setup_basic_desc(&desc[i],
2216 					      HNS_ROCE_OPC_QUERY_PF_CAPS_NUM,
2217 					      true);
2218 		if (i < (HNS_ROCE_QUERY_PF_CAPS_CMD_NUM - 1))
2219 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2220 		else
2221 			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2222 	}
2223 
2224 	ret = hns_roce_cmq_send(hr_dev, desc, HNS_ROCE_QUERY_PF_CAPS_CMD_NUM);
2225 	if (ret)
2226 		return ret;
2227 
2228 	resp_a = (struct hns_roce_query_pf_caps_a *)desc[0].data;
2229 	resp_b = (struct hns_roce_query_pf_caps_b *)desc[1].data;
2230 	resp_c = (struct hns_roce_query_pf_caps_c *)desc[2].data;
2231 	resp_d = (struct hns_roce_query_pf_caps_d *)desc[3].data;
2232 	resp_e = (struct hns_roce_query_pf_caps_e *)desc[4].data;
2233 
2234 	caps->local_ca_ack_delay     = resp_a->local_ca_ack_delay;
2235 	caps->max_sq_sg		     = le16_to_cpu(resp_a->max_sq_sg);
2236 	caps->max_sq_inline	     = le16_to_cpu(resp_a->max_sq_inline);
2237 	caps->max_rq_sg		     = le16_to_cpu(resp_a->max_rq_sg);
2238 	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
2239 	caps->max_extend_sg	     = le32_to_cpu(resp_a->max_extend_sg);
2240 	caps->num_qpc_timer	     = le16_to_cpu(resp_a->num_qpc_timer);
2241 	caps->max_srq_sges	     = le16_to_cpu(resp_a->max_srq_sges);
2242 	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
2243 	caps->num_aeq_vectors	     = resp_a->num_aeq_vectors;
2244 	caps->num_other_vectors	     = resp_a->num_other_vectors;
2245 	caps->max_sq_desc_sz	     = resp_a->max_sq_desc_sz;
2246 	caps->max_rq_desc_sz	     = resp_a->max_rq_desc_sz;
2247 	caps->max_srq_desc_sz	     = resp_a->max_srq_desc_sz;
2248 	caps->cqe_sz		     = resp_a->cqe_sz;
2249 
2250 	caps->mtpt_entry_sz	     = resp_b->mtpt_entry_sz;
2251 	caps->irrl_entry_sz	     = resp_b->irrl_entry_sz;
2252 	caps->trrl_entry_sz	     = resp_b->trrl_entry_sz;
2253 	caps->cqc_entry_sz	     = resp_b->cqc_entry_sz;
2254 	caps->srqc_entry_sz	     = resp_b->srqc_entry_sz;
2255 	caps->idx_entry_sz	     = resp_b->idx_entry_sz;
2256 	caps->sccc_sz		     = resp_b->sccc_sz;
2257 	caps->max_mtu		     = resp_b->max_mtu;
2258 	caps->qpc_sz		     = le16_to_cpu(resp_b->qpc_sz);
2259 	caps->min_cqes		     = resp_b->min_cqes;
2260 	caps->min_wqes		     = resp_b->min_wqes;
2261 	caps->page_size_cap	     = le32_to_cpu(resp_b->page_size_cap);
2262 	caps->pkey_table_len[0]	     = resp_b->pkey_table_len;
2263 	caps->phy_num_uars	     = resp_b->phy_num_uars;
2264 	ctx_hop_num		     = resp_b->ctx_hop_num;
2265 	pbl_hop_num		     = resp_b->pbl_hop_num;
2266 
2267 	caps->num_pds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_PDS);
2268 
2269 	caps->flags = hr_reg_read(resp_c, PF_CAPS_C_CAP_FLAGS);
2270 	caps->flags |= le16_to_cpu(resp_d->cap_flags_ex) <<
2271 		       HNS_ROCE_CAP_FLAGS_EX_SHIFT;
2272 
2273 	caps->num_cqs = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_CQS);
2274 	caps->gid_table_len[0] = hr_reg_read(resp_c, PF_CAPS_C_MAX_GID);
2275 	caps->max_cqes = 1 << hr_reg_read(resp_c, PF_CAPS_C_CQ_DEPTH);
2276 	caps->num_mtpts = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_MRWS);
2277 	caps->num_qps = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_QPS);
2278 	caps->max_qp_init_rdma = hr_reg_read(resp_c, PF_CAPS_C_MAX_ORD);
2279 	caps->max_qp_dest_rdma = caps->max_qp_init_rdma;
2280 	caps->max_wqes = 1 << le16_to_cpu(resp_c->sq_depth);
2281 
2282 	caps->num_srqs = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_SRQS);
2283 	caps->cong_type = hr_reg_read(resp_d, PF_CAPS_D_CONG_TYPE);
2284 	caps->max_srq_wrs = 1 << le16_to_cpu(resp_d->srq_depth);
2285 	caps->ceqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_CEQ_DEPTH);
2286 	caps->num_comp_vectors = hr_reg_read(resp_d, PF_CAPS_D_NUM_CEQS);
2287 	caps->aeqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_AEQ_DEPTH);
2288 	caps->default_aeq_arm_st = hr_reg_read(resp_d, PF_CAPS_D_AEQ_ARM_ST);
2289 	caps->default_ceq_arm_st = hr_reg_read(resp_d, PF_CAPS_D_CEQ_ARM_ST);
2290 	caps->reserved_pds = hr_reg_read(resp_d, PF_CAPS_D_RSV_PDS);
2291 	caps->num_uars = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_UARS);
2292 	caps->reserved_qps = hr_reg_read(resp_d, PF_CAPS_D_RSV_QPS);
2293 	caps->reserved_uars = hr_reg_read(resp_d, PF_CAPS_D_RSV_UARS);
2294 
2295 	caps->reserved_mrws = hr_reg_read(resp_e, PF_CAPS_E_RSV_MRWS);
2296 	caps->chunk_sz = 1 << hr_reg_read(resp_e, PF_CAPS_E_CHUNK_SIZE_SHIFT);
2297 	caps->reserved_cqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_CQS);
2298 	caps->reserved_srqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_SRQS);
2299 	caps->reserved_lkey = hr_reg_read(resp_e, PF_CAPS_E_RSV_LKEYS);
2300 	caps->default_ceq_max_cnt = le16_to_cpu(resp_e->ceq_max_cnt);
2301 	caps->default_ceq_period = le16_to_cpu(resp_e->ceq_period);
2302 	caps->default_aeq_max_cnt = le16_to_cpu(resp_e->aeq_max_cnt);
2303 	caps->default_aeq_period = le16_to_cpu(resp_e->aeq_period);
2304 
2305 	caps->qpc_hop_num = ctx_hop_num;
2306 	caps->sccc_hop_num = ctx_hop_num;
2307 	caps->srqc_hop_num = ctx_hop_num;
2308 	caps->cqc_hop_num = ctx_hop_num;
2309 	caps->mpt_hop_num = ctx_hop_num;
2310 	caps->mtt_hop_num = pbl_hop_num;
2311 	caps->cqe_hop_num = pbl_hop_num;
2312 	caps->srqwqe_hop_num = pbl_hop_num;
2313 	caps->idx_hop_num = pbl_hop_num;
2314 	caps->wqe_sq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_SQWQE_HOP_NUM);
2315 	caps->wqe_sge_hop_num = hr_reg_read(resp_d, PF_CAPS_D_EX_SGE_HOP_NUM);
2316 	caps->wqe_rq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_RQWQE_HOP_NUM);
2317 
2318 	return 0;
2319 }
2320 
2321 static int config_hem_entry_size(struct hns_roce_dev *hr_dev, u32 type, u32 val)
2322 {
2323 	struct hns_roce_cmq_desc desc;
2324 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
2325 
2326 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_ENTRY_SIZE,
2327 				      false);
2328 
2329 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_TYPE, type);
2330 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_VALUE, val);
2331 
2332 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2333 }
2334 
2335 static int hns_roce_config_entry_size(struct hns_roce_dev *hr_dev)
2336 {
2337 	struct hns_roce_caps *caps = &hr_dev->caps;
2338 	int ret;
2339 
2340 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2341 		return 0;
2342 
2343 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_QPC_SIZE,
2344 				    caps->qpc_sz);
2345 	if (ret) {
2346 		dev_err(hr_dev->dev, "failed to cfg qpc sz, ret = %d.\n", ret);
2347 		return ret;
2348 	}
2349 
2350 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_SCCC_SIZE,
2351 				    caps->sccc_sz);
2352 	if (ret)
2353 		dev_err(hr_dev->dev, "failed to cfg sccc sz, ret = %d.\n", ret);
2354 
2355 	return ret;
2356 }
2357 
2358 static int hns_roce_v2_vf_profile(struct hns_roce_dev *hr_dev)
2359 {
2360 	struct device *dev = hr_dev->dev;
2361 	int ret;
2362 
2363 	hr_dev->func_num = 1;
2364 
2365 	set_default_caps(hr_dev);
2366 
2367 	ret = hns_roce_query_vf_resource(hr_dev);
2368 	if (ret) {
2369 		dev_err(dev, "failed to query VF resource, ret = %d.\n", ret);
2370 		return ret;
2371 	}
2372 
2373 	apply_func_caps(hr_dev);
2374 
2375 	ret = hns_roce_v2_set_bt(hr_dev);
2376 	if (ret)
2377 		dev_err(dev, "failed to config VF BA table, ret = %d.\n", ret);
2378 
2379 	return ret;
2380 }
2381 
2382 static int hns_roce_v2_pf_profile(struct hns_roce_dev *hr_dev)
2383 {
2384 	struct device *dev = hr_dev->dev;
2385 	int ret;
2386 
2387 	ret = hns_roce_query_func_info(hr_dev);
2388 	if (ret) {
2389 		dev_err(dev, "failed to query func info, ret = %d.\n", ret);
2390 		return ret;
2391 	}
2392 
2393 	ret = hns_roce_config_global_param(hr_dev);
2394 	if (ret) {
2395 		dev_err(dev, "failed to config global param, ret = %d.\n", ret);
2396 		return ret;
2397 	}
2398 
2399 	ret = hns_roce_set_vf_switch_param(hr_dev);
2400 	if (ret) {
2401 		dev_err(dev, "failed to set switch param, ret = %d.\n", ret);
2402 		return ret;
2403 	}
2404 
2405 	ret = hns_roce_query_pf_caps(hr_dev);
2406 	if (ret)
2407 		set_default_caps(hr_dev);
2408 
2409 	ret = hns_roce_query_pf_resource(hr_dev);
2410 	if (ret) {
2411 		dev_err(dev, "failed to query pf resource, ret = %d.\n", ret);
2412 		return ret;
2413 	}
2414 
2415 	apply_func_caps(hr_dev);
2416 
2417 	ret = hns_roce_alloc_vf_resource(hr_dev);
2418 	if (ret) {
2419 		dev_err(dev, "failed to alloc vf resource, ret = %d.\n", ret);
2420 		return ret;
2421 	}
2422 
2423 	ret = hns_roce_v2_set_bt(hr_dev);
2424 	if (ret) {
2425 		dev_err(dev, "failed to config BA table, ret = %d.\n", ret);
2426 		return ret;
2427 	}
2428 
2429 	/* Configure the size of QPC, SCCC, etc. */
2430 	return hns_roce_config_entry_size(hr_dev);
2431 }
2432 
2433 static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
2434 {
2435 	struct device *dev = hr_dev->dev;
2436 	int ret;
2437 
2438 	ret = hns_roce_cmq_query_hw_info(hr_dev);
2439 	if (ret) {
2440 		dev_err(dev, "failed to query hardware info, ret = %d.\n", ret);
2441 		return ret;
2442 	}
2443 
2444 	ret = hns_roce_query_fw_ver(hr_dev);
2445 	if (ret) {
2446 		dev_err(dev, "failed to query firmware info, ret = %d.\n", ret);
2447 		return ret;
2448 	}
2449 
2450 	hr_dev->vendor_part_id = hr_dev->pci_dev->device;
2451 	hr_dev->sys_image_guid = be64_to_cpu(hr_dev->ib_dev.node_guid);
2452 
2453 	if (hr_dev->is_vf)
2454 		return hns_roce_v2_vf_profile(hr_dev);
2455 	else
2456 		return hns_roce_v2_pf_profile(hr_dev);
2457 }
2458 
2459 static void config_llm_table(struct hns_roce_buf *data_buf, void *cfg_buf)
2460 {
2461 	u32 i, next_ptr, page_num;
2462 	__le64 *entry = cfg_buf;
2463 	dma_addr_t addr;
2464 	u64 val;
2465 
2466 	page_num = data_buf->npages;
2467 	for (i = 0; i < page_num; i++) {
2468 		addr = hns_roce_buf_page(data_buf, i);
2469 		if (i == (page_num - 1))
2470 			next_ptr = 0;
2471 		else
2472 			next_ptr = i + 1;
2473 
2474 		val = HNS_ROCE_EXT_LLM_ENTRY(addr, (u64)next_ptr);
2475 		entry[i] = cpu_to_le64(val);
2476 	}
2477 }
2478 
2479 static int set_llm_cfg_to_hw(struct hns_roce_dev *hr_dev,
2480 			     struct hns_roce_link_table *table)
2481 {
2482 	struct hns_roce_cmq_desc desc[2];
2483 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
2484 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
2485 	struct hns_roce_buf *buf = table->buf;
2486 	enum hns_roce_opcode_type opcode;
2487 	dma_addr_t addr;
2488 
2489 	opcode = HNS_ROCE_OPC_CFG_EXT_LLM;
2490 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
2491 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2492 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
2493 
2494 	hr_reg_write(r_a, CFG_LLM_A_BA_L, lower_32_bits(table->table.map));
2495 	hr_reg_write(r_a, CFG_LLM_A_BA_H, upper_32_bits(table->table.map));
2496 	hr_reg_write(r_a, CFG_LLM_A_DEPTH, buf->npages);
2497 	hr_reg_write(r_a, CFG_LLM_A_PGSZ, to_hr_hw_page_shift(buf->page_shift));
2498 	hr_reg_enable(r_a, CFG_LLM_A_INIT_EN);
2499 
2500 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, 0));
2501 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_L, lower_32_bits(addr));
2502 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_H, upper_32_bits(addr));
2503 	hr_reg_write(r_a, CFG_LLM_A_HEAD_NXTPTR, 1);
2504 	hr_reg_write(r_a, CFG_LLM_A_HEAD_PTR, 0);
2505 
2506 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, buf->npages - 1));
2507 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_L, lower_32_bits(addr));
2508 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_H, upper_32_bits(addr));
2509 	hr_reg_write(r_b, CFG_LLM_B_TAIL_PTR, buf->npages - 1);
2510 
2511 	return hns_roce_cmq_send(hr_dev, desc, 2);
2512 }
2513 
2514 static struct hns_roce_link_table *
2515 alloc_link_table_buf(struct hns_roce_dev *hr_dev)
2516 {
2517 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2518 	struct hns_roce_link_table *link_tbl;
2519 	u32 pg_shift, size, min_size;
2520 
2521 	link_tbl = &priv->ext_llm;
2522 	pg_shift = hr_dev->caps.llm_buf_pg_sz + PAGE_SHIFT;
2523 	size = hr_dev->caps.num_qps * HNS_ROCE_V2_EXT_LLM_ENTRY_SZ;
2524 	min_size = HNS_ROCE_EXT_LLM_MIN_PAGES(hr_dev->caps.sl_num) << pg_shift;
2525 
2526 	/* Alloc data table */
2527 	size = max(size, min_size);
2528 	link_tbl->buf = hns_roce_buf_alloc(hr_dev, size, pg_shift, 0);
2529 	if (IS_ERR(link_tbl->buf))
2530 		return ERR_PTR(-ENOMEM);
2531 
2532 	/* Alloc config table */
2533 	size = link_tbl->buf->npages * sizeof(u64);
2534 	link_tbl->table.buf = dma_alloc_coherent(hr_dev->dev, size,
2535 						 &link_tbl->table.map,
2536 						 GFP_KERNEL);
2537 	if (!link_tbl->table.buf) {
2538 		hns_roce_buf_free(hr_dev, link_tbl->buf);
2539 		return ERR_PTR(-ENOMEM);
2540 	}
2541 
2542 	return link_tbl;
2543 }
2544 
2545 static void free_link_table_buf(struct hns_roce_dev *hr_dev,
2546 				struct hns_roce_link_table *tbl)
2547 {
2548 	if (tbl->buf) {
2549 		u32 size = tbl->buf->npages * sizeof(u64);
2550 
2551 		dma_free_coherent(hr_dev->dev, size, tbl->table.buf,
2552 				  tbl->table.map);
2553 	}
2554 
2555 	hns_roce_buf_free(hr_dev, tbl->buf);
2556 }
2557 
2558 static int hns_roce_init_link_table(struct hns_roce_dev *hr_dev)
2559 {
2560 	struct hns_roce_link_table *link_tbl;
2561 	int ret;
2562 
2563 	link_tbl = alloc_link_table_buf(hr_dev);
2564 	if (IS_ERR(link_tbl))
2565 		return -ENOMEM;
2566 
2567 	if (WARN_ON(link_tbl->buf->npages > HNS_ROCE_V2_EXT_LLM_MAX_DEPTH)) {
2568 		ret = -EINVAL;
2569 		goto err_alloc;
2570 	}
2571 
2572 	config_llm_table(link_tbl->buf, link_tbl->table.buf);
2573 	ret = set_llm_cfg_to_hw(hr_dev, link_tbl);
2574 	if (ret)
2575 		goto err_alloc;
2576 
2577 	return 0;
2578 
2579 err_alloc:
2580 	free_link_table_buf(hr_dev, link_tbl);
2581 	return ret;
2582 }
2583 
2584 static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev)
2585 {
2586 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2587 
2588 	free_link_table_buf(hr_dev, &priv->ext_llm);
2589 }
2590 
2591 static void free_dip_list(struct hns_roce_dev *hr_dev)
2592 {
2593 	struct hns_roce_dip *hr_dip;
2594 	struct hns_roce_dip *tmp;
2595 	unsigned long flags;
2596 
2597 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
2598 
2599 	list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) {
2600 		list_del(&hr_dip->node);
2601 		kfree(hr_dip);
2602 	}
2603 
2604 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
2605 }
2606 
2607 static void free_mr_exit(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 	int ret;
2612 	int i;
2613 
2614 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2615 		if (free_mr->rsv_qp[i]) {
2616 			ret = ib_destroy_qp(free_mr->rsv_qp[i]);
2617 			if (ret)
2618 				ibdev_err(&hr_dev->ib_dev,
2619 					  "failed to destroy qp in free mr.\n");
2620 
2621 			free_mr->rsv_qp[i] = NULL;
2622 		}
2623 	}
2624 
2625 	if (free_mr->rsv_cq) {
2626 		ib_destroy_cq(free_mr->rsv_cq);
2627 		free_mr->rsv_cq = NULL;
2628 	}
2629 
2630 	if (free_mr->rsv_pd) {
2631 		ib_dealloc_pd(free_mr->rsv_pd);
2632 		free_mr->rsv_pd = NULL;
2633 	}
2634 }
2635 
2636 static int free_mr_alloc_res(struct hns_roce_dev *hr_dev)
2637 {
2638 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2639 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2640 	struct ib_device *ibdev = &hr_dev->ib_dev;
2641 	struct ib_cq_init_attr cq_init_attr = {};
2642 	struct ib_qp_init_attr qp_init_attr = {};
2643 	struct ib_pd *pd;
2644 	struct ib_cq *cq;
2645 	struct ib_qp *qp;
2646 	int ret;
2647 	int i;
2648 
2649 	pd = ib_alloc_pd(ibdev, 0);
2650 	if (IS_ERR(pd)) {
2651 		ibdev_err(ibdev, "failed to create pd for free mr.\n");
2652 		return PTR_ERR(pd);
2653 	}
2654 	free_mr->rsv_pd = pd;
2655 
2656 	cq_init_attr.cqe = HNS_ROCE_FREE_MR_USED_CQE_NUM;
2657 	cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_init_attr);
2658 	if (IS_ERR(cq)) {
2659 		ibdev_err(ibdev, "failed to create cq for free mr.\n");
2660 		ret = PTR_ERR(cq);
2661 		goto create_failed;
2662 	}
2663 	free_mr->rsv_cq = cq;
2664 
2665 	qp_init_attr.qp_type = IB_QPT_RC;
2666 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2667 	qp_init_attr.send_cq = free_mr->rsv_cq;
2668 	qp_init_attr.recv_cq = free_mr->rsv_cq;
2669 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2670 		qp_init_attr.cap.max_send_wr = HNS_ROCE_FREE_MR_USED_SQWQE_NUM;
2671 		qp_init_attr.cap.max_send_sge = HNS_ROCE_FREE_MR_USED_SQSGE_NUM;
2672 		qp_init_attr.cap.max_recv_wr = HNS_ROCE_FREE_MR_USED_RQWQE_NUM;
2673 		qp_init_attr.cap.max_recv_sge = HNS_ROCE_FREE_MR_USED_RQSGE_NUM;
2674 
2675 		qp = ib_create_qp(free_mr->rsv_pd, &qp_init_attr);
2676 		if (IS_ERR(qp)) {
2677 			ibdev_err(ibdev, "failed to create qp for free mr.\n");
2678 			ret = PTR_ERR(qp);
2679 			goto create_failed;
2680 		}
2681 
2682 		free_mr->rsv_qp[i] = qp;
2683 	}
2684 
2685 	return 0;
2686 
2687 create_failed:
2688 	free_mr_exit(hr_dev);
2689 
2690 	return ret;
2691 }
2692 
2693 static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
2694 				 struct ib_qp_attr *attr, int sl_num)
2695 {
2696 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2697 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2698 	struct ib_device *ibdev = &hr_dev->ib_dev;
2699 	struct hns_roce_qp *hr_qp;
2700 	int loopback;
2701 	int mask;
2702 	int ret;
2703 
2704 	hr_qp = to_hr_qp(free_mr->rsv_qp[sl_num]);
2705 	hr_qp->free_mr_en = 1;
2706 
2707 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_ACCESS_FLAGS;
2708 	attr->qp_state = IB_QPS_INIT;
2709 	attr->port_num = 1;
2710 	attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
2711 	ret = ib_modify_qp(&hr_qp->ibqp, attr, mask);
2712 	if (ret) {
2713 		ibdev_err(ibdev, "failed to modify qp to init, ret = %d.\n",
2714 			  ret);
2715 		return ret;
2716 	}
2717 
2718 	loopback = hr_dev->loop_idc;
2719 	/* Set qpc lbi = 1 incidate loopback IO */
2720 	hr_dev->loop_idc = 1;
2721 
2722 	mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
2723 	       IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
2724 	attr->qp_state = IB_QPS_RTR;
2725 	attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2726 	attr->path_mtu = IB_MTU_256;
2727 	attr->dest_qp_num = hr_qp->qpn;
2728 	attr->rq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2729 
2730 	rdma_ah_set_sl(&attr->ah_attr, (u8)sl_num);
2731 
2732 	ret = ib_modify_qp(&hr_qp->ibqp, attr, mask);
2733 	hr_dev->loop_idc = loopback;
2734 	if (ret) {
2735 		ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
2736 			  ret);
2737 		return ret;
2738 	}
2739 
2740 	mask = IB_QP_STATE | IB_QP_SQ_PSN | IB_QP_RETRY_CNT | IB_QP_TIMEOUT |
2741 	       IB_QP_RNR_RETRY | IB_QP_MAX_QP_RD_ATOMIC;
2742 	attr->qp_state = IB_QPS_RTS;
2743 	attr->sq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2744 	attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
2745 	attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
2746 	ret = ib_modify_qp(&hr_qp->ibqp, attr, mask);
2747 	if (ret)
2748 		ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
2749 			  ret);
2750 
2751 	return ret;
2752 }
2753 
2754 static int free_mr_modify_qp(struct hns_roce_dev *hr_dev)
2755 {
2756 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2757 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2758 	struct ib_qp_attr attr = {};
2759 	int ret;
2760 	int i;
2761 
2762 	rdma_ah_set_grh(&attr.ah_attr, NULL, 0, 0, 1, 0);
2763 	rdma_ah_set_static_rate(&attr.ah_attr, 3);
2764 	rdma_ah_set_port_num(&attr.ah_attr, 1);
2765 
2766 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2767 		ret = free_mr_modify_rsv_qp(hr_dev, &attr, i);
2768 		if (ret)
2769 			return ret;
2770 	}
2771 
2772 	return 0;
2773 }
2774 
2775 static int free_mr_init(struct hns_roce_dev *hr_dev)
2776 {
2777 	int ret;
2778 
2779 	ret = free_mr_alloc_res(hr_dev);
2780 	if (ret)
2781 		return ret;
2782 
2783 	ret = free_mr_modify_qp(hr_dev);
2784 	if (ret)
2785 		goto err_modify_qp;
2786 
2787 	return 0;
2788 
2789 err_modify_qp:
2790 	free_mr_exit(hr_dev);
2791 
2792 	return ret;
2793 }
2794 
2795 static int get_hem_table(struct hns_roce_dev *hr_dev)
2796 {
2797 	unsigned int qpc_count;
2798 	unsigned int cqc_count;
2799 	unsigned int gmv_count;
2800 	int ret;
2801 	int i;
2802 
2803 	/* Alloc memory for source address table buffer space chunk */
2804 	for (gmv_count = 0; gmv_count < hr_dev->caps.gmv_entry_num;
2805 	     gmv_count++) {
2806 		ret = hns_roce_table_get(hr_dev, &hr_dev->gmv_table, gmv_count);
2807 		if (ret)
2808 			goto err_gmv_failed;
2809 	}
2810 
2811 	if (hr_dev->is_vf)
2812 		return 0;
2813 
2814 	/* Alloc memory for QPC Timer buffer space chunk */
2815 	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
2816 	     qpc_count++) {
2817 		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
2818 					 qpc_count);
2819 		if (ret) {
2820 			dev_err(hr_dev->dev, "QPC Timer get failed\n");
2821 			goto err_qpc_timer_failed;
2822 		}
2823 	}
2824 
2825 	/* Alloc memory for CQC Timer buffer space chunk */
2826 	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
2827 	     cqc_count++) {
2828 		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
2829 					 cqc_count);
2830 		if (ret) {
2831 			dev_err(hr_dev->dev, "CQC Timer get failed\n");
2832 			goto err_cqc_timer_failed;
2833 		}
2834 	}
2835 
2836 	return 0;
2837 
2838 err_cqc_timer_failed:
2839 	for (i = 0; i < cqc_count; i++)
2840 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2841 
2842 err_qpc_timer_failed:
2843 	for (i = 0; i < qpc_count; i++)
2844 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2845 
2846 err_gmv_failed:
2847 	for (i = 0; i < gmv_count; i++)
2848 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2849 
2850 	return ret;
2851 }
2852 
2853 static void put_hem_table(struct hns_roce_dev *hr_dev)
2854 {
2855 	int i;
2856 
2857 	for (i = 0; i < hr_dev->caps.gmv_entry_num; i++)
2858 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2859 
2860 	if (hr_dev->is_vf)
2861 		return;
2862 
2863 	for (i = 0; i < hr_dev->caps.qpc_timer_bt_num; i++)
2864 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2865 
2866 	for (i = 0; i < hr_dev->caps.cqc_timer_bt_num; i++)
2867 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2868 }
2869 
2870 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
2871 {
2872 	int ret;
2873 
2874 	/* The hns ROCEE requires the extdb info to be cleared before using */
2875 	ret = hns_roce_clear_extdb_list_info(hr_dev);
2876 	if (ret)
2877 		return ret;
2878 
2879 	ret = get_hem_table(hr_dev);
2880 	if (ret)
2881 		return ret;
2882 
2883 	if (hr_dev->is_vf)
2884 		return 0;
2885 
2886 	ret = hns_roce_init_link_table(hr_dev);
2887 	if (ret) {
2888 		dev_err(hr_dev->dev, "failed to init llm, ret = %d.\n", ret);
2889 		goto err_llm_init_failed;
2890 	}
2891 
2892 	return 0;
2893 
2894 err_llm_init_failed:
2895 	put_hem_table(hr_dev);
2896 
2897 	return ret;
2898 }
2899 
2900 static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev)
2901 {
2902 	hns_roce_function_clear(hr_dev);
2903 
2904 	if (!hr_dev->is_vf)
2905 		hns_roce_free_link_table(hr_dev);
2906 
2907 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09)
2908 		free_dip_list(hr_dev);
2909 }
2910 
2911 static int hns_roce_mbox_post(struct hns_roce_dev *hr_dev,
2912 			      struct hns_roce_mbox_msg *mbox_msg)
2913 {
2914 	struct hns_roce_cmq_desc desc;
2915 	struct hns_roce_post_mbox *mb = (struct hns_roce_post_mbox *)desc.data;
2916 
2917 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_POST_MB, false);
2918 
2919 	mb->in_param_l = cpu_to_le32(mbox_msg->in_param);
2920 	mb->in_param_h = cpu_to_le32(mbox_msg->in_param >> 32);
2921 	mb->out_param_l = cpu_to_le32(mbox_msg->out_param);
2922 	mb->out_param_h = cpu_to_le32(mbox_msg->out_param >> 32);
2923 	mb->cmd_tag = cpu_to_le32(mbox_msg->tag << 8 | mbox_msg->cmd);
2924 	mb->token_event_en = cpu_to_le32(mbox_msg->event_en << 16 |
2925 					 mbox_msg->token);
2926 
2927 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2928 }
2929 
2930 static int v2_wait_mbox_complete(struct hns_roce_dev *hr_dev, u32 timeout,
2931 				 u8 *complete_status)
2932 {
2933 	struct hns_roce_mbox_status *mb_st;
2934 	struct hns_roce_cmq_desc desc;
2935 	unsigned long end;
2936 	int ret = -EBUSY;
2937 	u32 status;
2938 	bool busy;
2939 
2940 	mb_st = (struct hns_roce_mbox_status *)desc.data;
2941 	end = msecs_to_jiffies(timeout) + jiffies;
2942 	while (v2_chk_mbox_is_avail(hr_dev, &busy)) {
2943 		if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
2944 			return -EIO;
2945 
2946 		status = 0;
2947 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_MB_ST,
2948 					      true);
2949 		ret = __hns_roce_cmq_send(hr_dev, &desc, 1);
2950 		if (!ret) {
2951 			status = le32_to_cpu(mb_st->mb_status_hw_run);
2952 			/* No pending message exists in ROCEE mbox. */
2953 			if (!(status & MB_ST_HW_RUN_M))
2954 				break;
2955 		} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
2956 			break;
2957 		}
2958 
2959 		if (time_after(jiffies, end)) {
2960 			dev_err_ratelimited(hr_dev->dev,
2961 					    "failed to wait mbox status 0x%x\n",
2962 					    status);
2963 			return -ETIMEDOUT;
2964 		}
2965 
2966 		cond_resched();
2967 		ret = -EBUSY;
2968 	}
2969 
2970 	if (!ret) {
2971 		*complete_status = (u8)(status & MB_ST_COMPLETE_M);
2972 	} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
2973 		/* Ignore all errors if the mbox is unavailable. */
2974 		ret = 0;
2975 		*complete_status = MB_ST_COMPLETE_M;
2976 	}
2977 
2978 	return ret;
2979 }
2980 
2981 static int v2_post_mbox(struct hns_roce_dev *hr_dev,
2982 			struct hns_roce_mbox_msg *mbox_msg)
2983 {
2984 	u8 status = 0;
2985 	int ret;
2986 
2987 	/* Waiting for the mbox to be idle */
2988 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_V2_GO_BIT_TIMEOUT_MSECS,
2989 				    &status);
2990 	if (unlikely(ret)) {
2991 		dev_err_ratelimited(hr_dev->dev,
2992 				    "failed to check post mbox status = 0x%x, ret = %d.\n",
2993 				    status, ret);
2994 		return ret;
2995 	}
2996 
2997 	/* Post new message to mbox */
2998 	ret = hns_roce_mbox_post(hr_dev, mbox_msg);
2999 	if (ret)
3000 		dev_err_ratelimited(hr_dev->dev,
3001 				    "failed to post mailbox, ret = %d.\n", ret);
3002 
3003 	return ret;
3004 }
3005 
3006 static int v2_poll_mbox_done(struct hns_roce_dev *hr_dev)
3007 {
3008 	u8 status = 0;
3009 	int ret;
3010 
3011 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_CMD_TIMEOUT_MSECS,
3012 				    &status);
3013 	if (!ret) {
3014 		if (status != MB_ST_COMPLETE_SUCC)
3015 			return -EBUSY;
3016 	} else {
3017 		dev_err_ratelimited(hr_dev->dev,
3018 				    "failed to check mbox status = 0x%x, ret = %d.\n",
3019 				    status, ret);
3020 	}
3021 
3022 	return ret;
3023 }
3024 
3025 static void copy_gid(void *dest, const union ib_gid *gid)
3026 {
3027 #define GID_SIZE 4
3028 	const union ib_gid *src = gid;
3029 	__le32 (*p)[GID_SIZE] = dest;
3030 	int i;
3031 
3032 	if (!gid)
3033 		src = &zgid;
3034 
3035 	for (i = 0; i < GID_SIZE; i++)
3036 		(*p)[i] = cpu_to_le32(*(u32 *)&src->raw[i * sizeof(u32)]);
3037 }
3038 
3039 static int config_sgid_table(struct hns_roce_dev *hr_dev,
3040 			     int gid_index, const union ib_gid *gid,
3041 			     enum hns_roce_sgid_type sgid_type)
3042 {
3043 	struct hns_roce_cmq_desc desc;
3044 	struct hns_roce_cfg_sgid_tb *sgid_tb =
3045 				    (struct hns_roce_cfg_sgid_tb *)desc.data;
3046 
3047 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SGID_TB, false);
3048 
3049 	hr_reg_write(sgid_tb, CFG_SGID_TB_TABLE_IDX, gid_index);
3050 	hr_reg_write(sgid_tb, CFG_SGID_TB_VF_SGID_TYPE, sgid_type);
3051 
3052 	copy_gid(&sgid_tb->vf_sgid_l, gid);
3053 
3054 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3055 }
3056 
3057 static int config_gmv_table(struct hns_roce_dev *hr_dev,
3058 			    int gid_index, const union ib_gid *gid,
3059 			    enum hns_roce_sgid_type sgid_type,
3060 			    const struct ib_gid_attr *attr)
3061 {
3062 	struct hns_roce_cmq_desc desc[2];
3063 	struct hns_roce_cfg_gmv_tb_a *tb_a =
3064 				(struct hns_roce_cfg_gmv_tb_a *)desc[0].data;
3065 	struct hns_roce_cfg_gmv_tb_b *tb_b =
3066 				(struct hns_roce_cfg_gmv_tb_b *)desc[1].data;
3067 
3068 	u16 vlan_id = VLAN_CFI_MASK;
3069 	u8 mac[ETH_ALEN] = {};
3070 	int ret;
3071 
3072 	if (gid) {
3073 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, mac);
3074 		if (ret)
3075 			return ret;
3076 	}
3077 
3078 	hns_roce_cmq_setup_basic_desc(&desc[0], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3079 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
3080 
3081 	hns_roce_cmq_setup_basic_desc(&desc[1], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3082 
3083 	copy_gid(&tb_a->vf_sgid_l, gid);
3084 
3085 	hr_reg_write(tb_a, GMV_TB_A_VF_SGID_TYPE, sgid_type);
3086 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_EN, vlan_id < VLAN_CFI_MASK);
3087 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_ID, vlan_id);
3088 
3089 	tb_b->vf_smac_l = cpu_to_le32(*(u32 *)mac);
3090 
3091 	hr_reg_write(tb_b, GMV_TB_B_SMAC_H, *(u16 *)&mac[4]);
3092 	hr_reg_write(tb_b, GMV_TB_B_SGID_IDX, gid_index);
3093 
3094 	return hns_roce_cmq_send(hr_dev, desc, 2);
3095 }
3096 
3097 static int hns_roce_v2_set_gid(struct hns_roce_dev *hr_dev, int gid_index,
3098 			       const union ib_gid *gid,
3099 			       const struct ib_gid_attr *attr)
3100 {
3101 	enum hns_roce_sgid_type sgid_type = GID_TYPE_FLAG_ROCE_V1;
3102 	int ret;
3103 
3104 	if (gid) {
3105 		if (attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3106 			if (ipv6_addr_v4mapped((void *)gid))
3107 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV4;
3108 			else
3109 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV6;
3110 		} else if (attr->gid_type == IB_GID_TYPE_ROCE) {
3111 			sgid_type = GID_TYPE_FLAG_ROCE_V1;
3112 		}
3113 	}
3114 
3115 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
3116 		ret = config_gmv_table(hr_dev, gid_index, gid, sgid_type, attr);
3117 	else
3118 		ret = config_sgid_table(hr_dev, gid_index, gid, sgid_type);
3119 
3120 	if (ret)
3121 		ibdev_err(&hr_dev->ib_dev, "failed to set gid, ret = %d!\n",
3122 			  ret);
3123 
3124 	return ret;
3125 }
3126 
3127 static int hns_roce_v2_set_mac(struct hns_roce_dev *hr_dev, u8 phy_port,
3128 			       const u8 *addr)
3129 {
3130 	struct hns_roce_cmq_desc desc;
3131 	struct hns_roce_cfg_smac_tb *smac_tb =
3132 				    (struct hns_roce_cfg_smac_tb *)desc.data;
3133 	u16 reg_smac_h;
3134 	u32 reg_smac_l;
3135 
3136 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SMAC_TB, false);
3137 
3138 	reg_smac_l = *(u32 *)(&addr[0]);
3139 	reg_smac_h = *(u16 *)(&addr[4]);
3140 
3141 	hr_reg_write(smac_tb, CFG_SMAC_TB_IDX, phy_port);
3142 	hr_reg_write(smac_tb, CFG_SMAC_TB_VF_SMAC_H, reg_smac_h);
3143 	smac_tb->vf_smac_l = cpu_to_le32(reg_smac_l);
3144 
3145 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3146 }
3147 
3148 static int set_mtpt_pbl(struct hns_roce_dev *hr_dev,
3149 			struct hns_roce_v2_mpt_entry *mpt_entry,
3150 			struct hns_roce_mr *mr)
3151 {
3152 	u64 pages[HNS_ROCE_V2_MAX_INNER_MTPT_NUM] = { 0 };
3153 	struct ib_device *ibdev = &hr_dev->ib_dev;
3154 	dma_addr_t pbl_ba;
3155 	int i, count;
3156 
3157 	count = hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, pages,
3158 				  ARRAY_SIZE(pages), &pbl_ba);
3159 	if (count < 1) {
3160 		ibdev_err(ibdev, "failed to find PBL mtr, count = %d.\n",
3161 			  count);
3162 		return -ENOBUFS;
3163 	}
3164 
3165 	/* Aligned to the hardware address access unit */
3166 	for (i = 0; i < count; i++)
3167 		pages[i] >>= 6;
3168 
3169 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3170 	mpt_entry->pbl_ba_l = cpu_to_le32(pbl_ba >> 3);
3171 	hr_reg_write(mpt_entry, MPT_PBL_BA_H, upper_32_bits(pbl_ba >> 3));
3172 
3173 	mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
3174 	hr_reg_write(mpt_entry, MPT_PA0_H, upper_32_bits(pages[0]));
3175 
3176 	mpt_entry->pa1_l = cpu_to_le32(lower_32_bits(pages[1]));
3177 	hr_reg_write(mpt_entry, MPT_PA1_H, upper_32_bits(pages[1]));
3178 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3179 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3180 
3181 	return 0;
3182 }
3183 
3184 static int hns_roce_v2_write_mtpt(struct hns_roce_dev *hr_dev,
3185 				  void *mb_buf, struct hns_roce_mr *mr)
3186 {
3187 	struct hns_roce_v2_mpt_entry *mpt_entry;
3188 
3189 	mpt_entry = mb_buf;
3190 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3191 
3192 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3193 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3194 	hr_reg_enable(mpt_entry, MPT_L_INV_EN);
3195 
3196 	hr_reg_write_bool(mpt_entry, MPT_BIND_EN,
3197 			  mr->access & IB_ACCESS_MW_BIND);
3198 	hr_reg_write_bool(mpt_entry, MPT_ATOMIC_EN,
3199 			  mr->access & IB_ACCESS_REMOTE_ATOMIC);
3200 	hr_reg_write_bool(mpt_entry, MPT_RR_EN,
3201 			  mr->access & IB_ACCESS_REMOTE_READ);
3202 	hr_reg_write_bool(mpt_entry, MPT_RW_EN,
3203 			  mr->access & IB_ACCESS_REMOTE_WRITE);
3204 	hr_reg_write_bool(mpt_entry, MPT_LW_EN,
3205 			  mr->access & IB_ACCESS_LOCAL_WRITE);
3206 
3207 	mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3208 	mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3209 	mpt_entry->lkey = cpu_to_le32(mr->key);
3210 	mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3211 	mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3212 
3213 	if (mr->type != MR_TYPE_MR)
3214 		hr_reg_enable(mpt_entry, MPT_PA);
3215 
3216 	if (mr->type == MR_TYPE_DMA)
3217 		return 0;
3218 
3219 	if (mr->pbl_hop_num != HNS_ROCE_HOP_NUM_0)
3220 		hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, mr->pbl_hop_num);
3221 
3222 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3223 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3224 	hr_reg_enable(mpt_entry, MPT_INNER_PA_VLD);
3225 
3226 	return set_mtpt_pbl(hr_dev, mpt_entry, mr);
3227 }
3228 
3229 static int hns_roce_v2_rereg_write_mtpt(struct hns_roce_dev *hr_dev,
3230 					struct hns_roce_mr *mr, int flags,
3231 					void *mb_buf)
3232 {
3233 	struct hns_roce_v2_mpt_entry *mpt_entry = mb_buf;
3234 	u32 mr_access_flags = mr->access;
3235 	int ret = 0;
3236 
3237 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3238 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3239 
3240 	if (flags & IB_MR_REREG_ACCESS) {
3241 		hr_reg_write(mpt_entry, MPT_BIND_EN,
3242 			     (mr_access_flags & IB_ACCESS_MW_BIND ? 1 : 0));
3243 		hr_reg_write(mpt_entry, MPT_ATOMIC_EN,
3244 			     mr_access_flags & IB_ACCESS_REMOTE_ATOMIC ? 1 : 0);
3245 		hr_reg_write(mpt_entry, MPT_RR_EN,
3246 			     mr_access_flags & IB_ACCESS_REMOTE_READ ? 1 : 0);
3247 		hr_reg_write(mpt_entry, MPT_RW_EN,
3248 			     mr_access_flags & IB_ACCESS_REMOTE_WRITE ? 1 : 0);
3249 		hr_reg_write(mpt_entry, MPT_LW_EN,
3250 			     mr_access_flags & IB_ACCESS_LOCAL_WRITE ? 1 : 0);
3251 	}
3252 
3253 	if (flags & IB_MR_REREG_TRANS) {
3254 		mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3255 		mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3256 		mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3257 		mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3258 
3259 		ret = set_mtpt_pbl(hr_dev, mpt_entry, mr);
3260 	}
3261 
3262 	return ret;
3263 }
3264 
3265 static int hns_roce_v2_frmr_write_mtpt(struct hns_roce_dev *hr_dev,
3266 				       void *mb_buf, struct hns_roce_mr *mr)
3267 {
3268 	struct ib_device *ibdev = &hr_dev->ib_dev;
3269 	struct hns_roce_v2_mpt_entry *mpt_entry;
3270 	dma_addr_t pbl_ba = 0;
3271 
3272 	mpt_entry = mb_buf;
3273 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3274 
3275 	if (hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, NULL, 0, &pbl_ba) < 0) {
3276 		ibdev_err(ibdev, "failed to find frmr mtr.\n");
3277 		return -ENOBUFS;
3278 	}
3279 
3280 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3281 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3282 
3283 	hr_reg_enable(mpt_entry, MPT_RA_EN);
3284 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3285 	hr_reg_enable(mpt_entry, MPT_L_INV_EN);
3286 
3287 	hr_reg_enable(mpt_entry, MPT_FRE);
3288 	hr_reg_clear(mpt_entry, MPT_MR_MW);
3289 	hr_reg_enable(mpt_entry, MPT_BPD);
3290 	hr_reg_clear(mpt_entry, MPT_PA);
3291 
3292 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, 1);
3293 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3294 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3295 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3296 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3297 
3298 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3299 
3300 	mpt_entry->pbl_ba_l = cpu_to_le32(lower_32_bits(pbl_ba >> 3));
3301 	hr_reg_write(mpt_entry, MPT_PBL_BA_H, upper_32_bits(pbl_ba >> 3));
3302 
3303 	return 0;
3304 }
3305 
3306 static int hns_roce_v2_mw_write_mtpt(void *mb_buf, struct hns_roce_mw *mw)
3307 {
3308 	struct hns_roce_v2_mpt_entry *mpt_entry;
3309 
3310 	mpt_entry = mb_buf;
3311 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3312 
3313 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3314 	hr_reg_write(mpt_entry, MPT_PD, mw->pdn);
3315 
3316 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3317 	hr_reg_enable(mpt_entry, MPT_L_INV_EN);
3318 	hr_reg_enable(mpt_entry, MPT_LW_EN);
3319 
3320 	hr_reg_enable(mpt_entry, MPT_MR_MW);
3321 	hr_reg_enable(mpt_entry, MPT_BPD);
3322 	hr_reg_clear(mpt_entry, MPT_PA);
3323 	hr_reg_write(mpt_entry, MPT_BQP,
3324 		     mw->ibmw.type == IB_MW_TYPE_1 ? 0 : 1);
3325 
3326 	mpt_entry->lkey = cpu_to_le32(mw->rkey);
3327 
3328 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM,
3329 		     mw->pbl_hop_num == HNS_ROCE_HOP_NUM_0 ? 0 :
3330 							     mw->pbl_hop_num);
3331 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3332 		     mw->pbl_ba_pg_sz + PG_SHIFT_OFFSET);
3333 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3334 		     mw->pbl_buf_pg_sz + PG_SHIFT_OFFSET);
3335 
3336 	return 0;
3337 }
3338 
3339 static int free_mr_post_send_lp_wqe(struct hns_roce_qp *hr_qp)
3340 {
3341 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
3342 	struct ib_device *ibdev = &hr_dev->ib_dev;
3343 	const struct ib_send_wr *bad_wr;
3344 	struct ib_rdma_wr rdma_wr = {};
3345 	struct ib_send_wr *send_wr;
3346 	int ret;
3347 
3348 	send_wr = &rdma_wr.wr;
3349 	send_wr->opcode = IB_WR_RDMA_WRITE;
3350 
3351 	ret = hns_roce_v2_post_send(&hr_qp->ibqp, send_wr, &bad_wr);
3352 	if (ret) {
3353 		ibdev_err(ibdev, "failed to post wqe for free mr, ret = %d.\n",
3354 			  ret);
3355 		return ret;
3356 	}
3357 
3358 	return 0;
3359 }
3360 
3361 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3362 			       struct ib_wc *wc);
3363 
3364 static void free_mr_send_cmd_to_hw(struct hns_roce_dev *hr_dev)
3365 {
3366 	struct hns_roce_v2_priv *priv = hr_dev->priv;
3367 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
3368 	struct ib_wc wc[ARRAY_SIZE(free_mr->rsv_qp)];
3369 	struct ib_device *ibdev = &hr_dev->ib_dev;
3370 	struct hns_roce_qp *hr_qp;
3371 	unsigned long end;
3372 	int cqe_cnt = 0;
3373 	int npolled;
3374 	int ret;
3375 	int i;
3376 
3377 	/*
3378 	 * If the device initialization is not complete or in the uninstall
3379 	 * process, then there is no need to execute free mr.
3380 	 */
3381 	if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
3382 	    priv->handle->rinfo.instance_state == HNS_ROCE_STATE_INIT ||
3383 	    hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT)
3384 		return;
3385 
3386 	mutex_lock(&free_mr->mutex);
3387 
3388 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
3389 		hr_qp = to_hr_qp(free_mr->rsv_qp[i]);
3390 
3391 		ret = free_mr_post_send_lp_wqe(hr_qp);
3392 		if (ret) {
3393 			ibdev_err(ibdev,
3394 				  "failed to send wqe (qp:0x%lx) for free mr, ret = %d.\n",
3395 				  hr_qp->qpn, ret);
3396 			break;
3397 		}
3398 
3399 		cqe_cnt++;
3400 	}
3401 
3402 	end = msecs_to_jiffies(HNS_ROCE_V2_FREE_MR_TIMEOUT) + jiffies;
3403 	while (cqe_cnt) {
3404 		npolled = hns_roce_v2_poll_cq(free_mr->rsv_cq, cqe_cnt, wc);
3405 		if (npolled < 0) {
3406 			ibdev_err(ibdev,
3407 				  "failed to poll cqe for free mr, remain %d cqe.\n",
3408 				  cqe_cnt);
3409 			goto out;
3410 		}
3411 
3412 		if (time_after(jiffies, end)) {
3413 			ibdev_err(ibdev,
3414 				  "failed to poll cqe for free mr and timeout, remain %d cqe.\n",
3415 				  cqe_cnt);
3416 			goto out;
3417 		}
3418 		cqe_cnt -= npolled;
3419 	}
3420 
3421 out:
3422 	mutex_unlock(&free_mr->mutex);
3423 }
3424 
3425 static void hns_roce_v2_dereg_mr(struct hns_roce_dev *hr_dev)
3426 {
3427 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
3428 		free_mr_send_cmd_to_hw(hr_dev);
3429 }
3430 
3431 static void *get_cqe_v2(struct hns_roce_cq *hr_cq, int n)
3432 {
3433 	return hns_roce_buf_offset(hr_cq->mtr.kmem, n * hr_cq->cqe_size);
3434 }
3435 
3436 static void *get_sw_cqe_v2(struct hns_roce_cq *hr_cq, unsigned int n)
3437 {
3438 	struct hns_roce_v2_cqe *cqe = get_cqe_v2(hr_cq, n & hr_cq->ib_cq.cqe);
3439 
3440 	/* Get cqe when Owner bit is Conversely with the MSB of cons_idx */
3441 	return (hr_reg_read(cqe, CQE_OWNER) ^ !!(n & hr_cq->cq_depth)) ? cqe :
3442 									 NULL;
3443 }
3444 
3445 static inline void update_cq_db(struct hns_roce_dev *hr_dev,
3446 				struct hns_roce_cq *hr_cq)
3447 {
3448 	if (likely(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) {
3449 		*hr_cq->set_ci_db = hr_cq->cons_index & V2_CQ_DB_CONS_IDX_M;
3450 	} else {
3451 		struct hns_roce_v2_db cq_db = {};
3452 
3453 		hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3454 		hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB);
3455 		hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3456 		hr_reg_write(&cq_db, DB_CQ_CMD_SN, 1);
3457 
3458 		hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3459 	}
3460 }
3461 
3462 static void __hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3463 				   struct hns_roce_srq *srq)
3464 {
3465 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3466 	struct hns_roce_v2_cqe *cqe, *dest;
3467 	u32 prod_index;
3468 	int nfreed = 0;
3469 	int wqe_index;
3470 	u8 owner_bit;
3471 
3472 	for (prod_index = hr_cq->cons_index; get_sw_cqe_v2(hr_cq, prod_index);
3473 	     ++prod_index) {
3474 		if (prod_index > hr_cq->cons_index + hr_cq->ib_cq.cqe)
3475 			break;
3476 	}
3477 
3478 	/*
3479 	 * Now backwards through the CQ, removing CQ entries
3480 	 * that match our QP by overwriting them with next entries.
3481 	 */
3482 	while ((int) --prod_index - (int) hr_cq->cons_index >= 0) {
3483 		cqe = get_cqe_v2(hr_cq, prod_index & hr_cq->ib_cq.cqe);
3484 		if (hr_reg_read(cqe, CQE_LCL_QPN) == qpn) {
3485 			if (srq && hr_reg_read(cqe, CQE_S_R)) {
3486 				wqe_index = hr_reg_read(cqe, CQE_WQE_IDX);
3487 				hns_roce_free_srq_wqe(srq, wqe_index);
3488 			}
3489 			++nfreed;
3490 		} else if (nfreed) {
3491 			dest = get_cqe_v2(hr_cq, (prod_index + nfreed) &
3492 					  hr_cq->ib_cq.cqe);
3493 			owner_bit = hr_reg_read(dest, CQE_OWNER);
3494 			memcpy(dest, cqe, hr_cq->cqe_size);
3495 			hr_reg_write(dest, CQE_OWNER, owner_bit);
3496 		}
3497 	}
3498 
3499 	if (nfreed) {
3500 		hr_cq->cons_index += nfreed;
3501 		update_cq_db(hr_dev, hr_cq);
3502 	}
3503 }
3504 
3505 static void hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3506 				 struct hns_roce_srq *srq)
3507 {
3508 	spin_lock_irq(&hr_cq->lock);
3509 	__hns_roce_v2_cq_clean(hr_cq, qpn, srq);
3510 	spin_unlock_irq(&hr_cq->lock);
3511 }
3512 
3513 static void hns_roce_v2_write_cqc(struct hns_roce_dev *hr_dev,
3514 				  struct hns_roce_cq *hr_cq, void *mb_buf,
3515 				  u64 *mtts, dma_addr_t dma_handle)
3516 {
3517 	struct hns_roce_v2_cq_context *cq_context;
3518 
3519 	cq_context = mb_buf;
3520 	memset(cq_context, 0, sizeof(*cq_context));
3521 
3522 	hr_reg_write(cq_context, CQC_CQ_ST, V2_CQ_STATE_VALID);
3523 	hr_reg_write(cq_context, CQC_ARM_ST, NO_ARMED);
3524 	hr_reg_write(cq_context, CQC_SHIFT, ilog2(hr_cq->cq_depth));
3525 	hr_reg_write(cq_context, CQC_CEQN, hr_cq->vector);
3526 	hr_reg_write(cq_context, CQC_CQN, hr_cq->cqn);
3527 
3528 	if (hr_cq->cqe_size == HNS_ROCE_V3_CQE_SIZE)
3529 		hr_reg_write(cq_context, CQC_CQE_SIZE, CQE_SIZE_64B);
3530 
3531 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
3532 		hr_reg_enable(cq_context, CQC_STASH);
3533 
3534 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_L,
3535 		     to_hr_hw_page_addr(mtts[0]));
3536 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_H,
3537 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
3538 	hr_reg_write(cq_context, CQC_CQE_HOP_NUM, hr_dev->caps.cqe_hop_num ==
3539 		     HNS_ROCE_HOP_NUM_0 ? 0 : hr_dev->caps.cqe_hop_num);
3540 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_L,
3541 		     to_hr_hw_page_addr(mtts[1]));
3542 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_H,
3543 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
3544 	hr_reg_write(cq_context, CQC_CQE_BAR_PG_SZ,
3545 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.ba_pg_shift));
3546 	hr_reg_write(cq_context, CQC_CQE_BUF_PG_SZ,
3547 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.buf_pg_shift));
3548 	hr_reg_write(cq_context, CQC_CQE_BA_L, dma_handle >> 3);
3549 	hr_reg_write(cq_context, CQC_CQE_BA_H, (dma_handle >> (32 + 3)));
3550 	hr_reg_write_bool(cq_context, CQC_DB_RECORD_EN,
3551 			  hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB);
3552 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_L,
3553 		     ((u32)hr_cq->db.dma) >> 1);
3554 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_H,
3555 		     hr_cq->db.dma >> 32);
3556 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT,
3557 		     HNS_ROCE_V2_CQ_DEFAULT_BURST_NUM);
3558 	hr_reg_write(cq_context, CQC_CQ_PERIOD,
3559 		     HNS_ROCE_V2_CQ_DEFAULT_INTERVAL);
3560 }
3561 
3562 static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
3563 				     enum ib_cq_notify_flags flags)
3564 {
3565 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3566 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3567 	struct hns_roce_v2_db cq_db = {};
3568 	u32 notify_flag;
3569 
3570 	/*
3571 	 * flags = 0, then notify_flag : next
3572 	 * flags = 1, then notify flag : solocited
3573 	 */
3574 	notify_flag = (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
3575 		      V2_CQ_DB_REQ_NOT : V2_CQ_DB_REQ_NOT_SOL;
3576 
3577 	hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3578 	hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB_NOTIFY);
3579 	hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3580 	hr_reg_write(&cq_db, DB_CQ_CMD_SN, hr_cq->arm_sn);
3581 	hr_reg_write(&cq_db, DB_CQ_NOTIFY, notify_flag);
3582 
3583 	hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3584 
3585 	return 0;
3586 }
3587 
3588 static int hns_roce_handle_recv_inl_wqe(struct hns_roce_v2_cqe *cqe,
3589 					struct hns_roce_qp *qp,
3590 					struct ib_wc *wc)
3591 {
3592 	struct hns_roce_rinl_sge *sge_list;
3593 	u32 wr_num, wr_cnt, sge_num;
3594 	u32 sge_cnt, data_len, size;
3595 	void *wqe_buf;
3596 
3597 	wr_num = hr_reg_read(cqe, CQE_WQE_IDX);
3598 	wr_cnt = wr_num & (qp->rq.wqe_cnt - 1);
3599 
3600 	sge_list = qp->rq_inl_buf.wqe_list[wr_cnt].sg_list;
3601 	sge_num = qp->rq_inl_buf.wqe_list[wr_cnt].sge_cnt;
3602 	wqe_buf = hns_roce_get_recv_wqe(qp, wr_cnt);
3603 	data_len = wc->byte_len;
3604 
3605 	for (sge_cnt = 0; (sge_cnt < sge_num) && (data_len); sge_cnt++) {
3606 		size = min(sge_list[sge_cnt].len, data_len);
3607 		memcpy((void *)sge_list[sge_cnt].addr, wqe_buf, size);
3608 
3609 		data_len -= size;
3610 		wqe_buf += size;
3611 	}
3612 
3613 	if (unlikely(data_len)) {
3614 		wc->status = IB_WC_LOC_LEN_ERR;
3615 		return -EAGAIN;
3616 	}
3617 
3618 	return 0;
3619 }
3620 
3621 static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
3622 		   int num_entries, struct ib_wc *wc)
3623 {
3624 	unsigned int left;
3625 	int npolled = 0;
3626 
3627 	left = wq->head - wq->tail;
3628 	if (left == 0)
3629 		return 0;
3630 
3631 	left = min_t(unsigned int, (unsigned int)num_entries, left);
3632 	while (npolled < left) {
3633 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3634 		wc->status = IB_WC_WR_FLUSH_ERR;
3635 		wc->vendor_err = 0;
3636 		wc->qp = &hr_qp->ibqp;
3637 
3638 		wq->tail++;
3639 		wc++;
3640 		npolled++;
3641 	}
3642 
3643 	return npolled;
3644 }
3645 
3646 static int hns_roce_v2_sw_poll_cq(struct hns_roce_cq *hr_cq, int num_entries,
3647 				  struct ib_wc *wc)
3648 {
3649 	struct hns_roce_qp *hr_qp;
3650 	int npolled = 0;
3651 
3652 	list_for_each_entry(hr_qp, &hr_cq->sq_list, sq_node) {
3653 		npolled += sw_comp(hr_qp, &hr_qp->sq,
3654 				   num_entries - npolled, wc + npolled);
3655 		if (npolled >= num_entries)
3656 			goto out;
3657 	}
3658 
3659 	list_for_each_entry(hr_qp, &hr_cq->rq_list, rq_node) {
3660 		npolled += sw_comp(hr_qp, &hr_qp->rq,
3661 				   num_entries - npolled, wc + npolled);
3662 		if (npolled >= num_entries)
3663 			goto out;
3664 	}
3665 
3666 out:
3667 	return npolled;
3668 }
3669 
3670 static void get_cqe_status(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
3671 			   struct hns_roce_cq *cq, struct hns_roce_v2_cqe *cqe,
3672 			   struct ib_wc *wc)
3673 {
3674 	static const struct {
3675 		u32 cqe_status;
3676 		enum ib_wc_status wc_status;
3677 	} map[] = {
3678 		{ HNS_ROCE_CQE_V2_SUCCESS, IB_WC_SUCCESS },
3679 		{ HNS_ROCE_CQE_V2_LOCAL_LENGTH_ERR, IB_WC_LOC_LEN_ERR },
3680 		{ HNS_ROCE_CQE_V2_LOCAL_QP_OP_ERR, IB_WC_LOC_QP_OP_ERR },
3681 		{ HNS_ROCE_CQE_V2_LOCAL_PROT_ERR, IB_WC_LOC_PROT_ERR },
3682 		{ HNS_ROCE_CQE_V2_WR_FLUSH_ERR, IB_WC_WR_FLUSH_ERR },
3683 		{ HNS_ROCE_CQE_V2_MW_BIND_ERR, IB_WC_MW_BIND_ERR },
3684 		{ HNS_ROCE_CQE_V2_BAD_RESP_ERR, IB_WC_BAD_RESP_ERR },
3685 		{ HNS_ROCE_CQE_V2_LOCAL_ACCESS_ERR, IB_WC_LOC_ACCESS_ERR },
3686 		{ HNS_ROCE_CQE_V2_REMOTE_INVAL_REQ_ERR, IB_WC_REM_INV_REQ_ERR },
3687 		{ HNS_ROCE_CQE_V2_REMOTE_ACCESS_ERR, IB_WC_REM_ACCESS_ERR },
3688 		{ HNS_ROCE_CQE_V2_REMOTE_OP_ERR, IB_WC_REM_OP_ERR },
3689 		{ HNS_ROCE_CQE_V2_TRANSPORT_RETRY_EXC_ERR,
3690 		  IB_WC_RETRY_EXC_ERR },
3691 		{ HNS_ROCE_CQE_V2_RNR_RETRY_EXC_ERR, IB_WC_RNR_RETRY_EXC_ERR },
3692 		{ HNS_ROCE_CQE_V2_REMOTE_ABORT_ERR, IB_WC_REM_ABORT_ERR },
3693 		{ HNS_ROCE_CQE_V2_GENERAL_ERR, IB_WC_GENERAL_ERR}
3694 	};
3695 
3696 	u32 cqe_status = hr_reg_read(cqe, CQE_STATUS);
3697 	int i;
3698 
3699 	wc->status = IB_WC_GENERAL_ERR;
3700 	for (i = 0; i < ARRAY_SIZE(map); i++)
3701 		if (cqe_status == map[i].cqe_status) {
3702 			wc->status = map[i].wc_status;
3703 			break;
3704 		}
3705 
3706 	if (likely(wc->status == IB_WC_SUCCESS ||
3707 		   wc->status == IB_WC_WR_FLUSH_ERR))
3708 		return;
3709 
3710 	ibdev_err(&hr_dev->ib_dev, "error cqe status 0x%x:\n", cqe_status);
3711 	print_hex_dump(KERN_ERR, "", DUMP_PREFIX_NONE, 16, 4, cqe,
3712 		       cq->cqe_size, false);
3713 	wc->vendor_err = hr_reg_read(cqe, CQE_SUB_STATUS);
3714 
3715 	/*
3716 	 * For hns ROCEE, GENERAL_ERR is an error type that is not defined in
3717 	 * the standard protocol, the driver must ignore it and needn't to set
3718 	 * the QP to an error state.
3719 	 */
3720 	if (cqe_status == HNS_ROCE_CQE_V2_GENERAL_ERR)
3721 		return;
3722 
3723 	flush_cqe(hr_dev, qp);
3724 }
3725 
3726 static int get_cur_qp(struct hns_roce_cq *hr_cq, struct hns_roce_v2_cqe *cqe,
3727 		      struct hns_roce_qp **cur_qp)
3728 {
3729 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3730 	struct hns_roce_qp *hr_qp = *cur_qp;
3731 	u32 qpn;
3732 
3733 	qpn = hr_reg_read(cqe, CQE_LCL_QPN);
3734 
3735 	if (!hr_qp || qpn != hr_qp->qpn) {
3736 		hr_qp = __hns_roce_qp_lookup(hr_dev, qpn);
3737 		if (unlikely(!hr_qp)) {
3738 			ibdev_err(&hr_dev->ib_dev,
3739 				  "CQ %06lx with entry for unknown QPN %06x\n",
3740 				  hr_cq->cqn, qpn);
3741 			return -EINVAL;
3742 		}
3743 		*cur_qp = hr_qp;
3744 	}
3745 
3746 	return 0;
3747 }
3748 
3749 /*
3750  * mapped-value = 1 + real-value
3751  * The ib wc opcode's real value is start from 0, In order to distinguish
3752  * between initialized and uninitialized map values, we plus 1 to the actual
3753  * value when defining the mapping, so that the validity can be identified by
3754  * checking whether the mapped value is greater than 0.
3755  */
3756 #define HR_WC_OP_MAP(hr_key, ib_key) \
3757 		[HNS_ROCE_V2_WQE_OP_ ## hr_key] = 1 + IB_WC_ ## ib_key
3758 
3759 static const u32 wc_send_op_map[] = {
3760 	HR_WC_OP_MAP(SEND,			SEND),
3761 	HR_WC_OP_MAP(SEND_WITH_INV,		SEND),
3762 	HR_WC_OP_MAP(SEND_WITH_IMM,		SEND),
3763 	HR_WC_OP_MAP(RDMA_READ,			RDMA_READ),
3764 	HR_WC_OP_MAP(RDMA_WRITE,		RDMA_WRITE),
3765 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,	RDMA_WRITE),
3766 	HR_WC_OP_MAP(LOCAL_INV,			LOCAL_INV),
3767 	HR_WC_OP_MAP(ATOM_CMP_AND_SWAP,		COMP_SWAP),
3768 	HR_WC_OP_MAP(ATOM_FETCH_AND_ADD,	FETCH_ADD),
3769 	HR_WC_OP_MAP(ATOM_MSK_CMP_AND_SWAP,	MASKED_COMP_SWAP),
3770 	HR_WC_OP_MAP(ATOM_MSK_FETCH_AND_ADD,	MASKED_FETCH_ADD),
3771 	HR_WC_OP_MAP(FAST_REG_PMR,		REG_MR),
3772 	HR_WC_OP_MAP(BIND_MW,			REG_MR),
3773 };
3774 
3775 static int to_ib_wc_send_op(u32 hr_opcode)
3776 {
3777 	if (hr_opcode >= ARRAY_SIZE(wc_send_op_map))
3778 		return -EINVAL;
3779 
3780 	return wc_send_op_map[hr_opcode] ? wc_send_op_map[hr_opcode] - 1 :
3781 					   -EINVAL;
3782 }
3783 
3784 static const u32 wc_recv_op_map[] = {
3785 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,		WITH_IMM),
3786 	HR_WC_OP_MAP(SEND,				RECV),
3787 	HR_WC_OP_MAP(SEND_WITH_IMM,			WITH_IMM),
3788 	HR_WC_OP_MAP(SEND_WITH_INV,			RECV),
3789 };
3790 
3791 static int to_ib_wc_recv_op(u32 hr_opcode)
3792 {
3793 	if (hr_opcode >= ARRAY_SIZE(wc_recv_op_map))
3794 		return -EINVAL;
3795 
3796 	return wc_recv_op_map[hr_opcode] ? wc_recv_op_map[hr_opcode] - 1 :
3797 					   -EINVAL;
3798 }
3799 
3800 static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3801 {
3802 	u32 hr_opcode;
3803 	int ib_opcode;
3804 
3805 	wc->wc_flags = 0;
3806 
3807 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3808 	switch (hr_opcode) {
3809 	case HNS_ROCE_V2_WQE_OP_RDMA_READ:
3810 		wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3811 		break;
3812 	case HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM:
3813 	case HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM:
3814 		wc->wc_flags |= IB_WC_WITH_IMM;
3815 		break;
3816 	case HNS_ROCE_V2_WQE_OP_LOCAL_INV:
3817 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3818 		break;
3819 	case HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP:
3820 	case HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD:
3821 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP:
3822 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD:
3823 		wc->byte_len  = 8;
3824 		break;
3825 	default:
3826 		break;
3827 	}
3828 
3829 	ib_opcode = to_ib_wc_send_op(hr_opcode);
3830 	if (ib_opcode < 0)
3831 		wc->status = IB_WC_GENERAL_ERR;
3832 	else
3833 		wc->opcode = ib_opcode;
3834 }
3835 
3836 static inline bool is_rq_inl_enabled(struct ib_wc *wc, u32 hr_opcode,
3837 				     struct hns_roce_v2_cqe *cqe)
3838 {
3839 	return wc->qp->qp_type != IB_QPT_UD && wc->qp->qp_type != IB_QPT_GSI &&
3840 	       (hr_opcode == HNS_ROCE_V2_OPCODE_SEND ||
3841 		hr_opcode == HNS_ROCE_V2_OPCODE_SEND_WITH_IMM ||
3842 		hr_opcode == HNS_ROCE_V2_OPCODE_SEND_WITH_INV) &&
3843 	       hr_reg_read(cqe, CQE_RQ_INLINE);
3844 }
3845 
3846 static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3847 {
3848 	struct hns_roce_qp *qp = to_hr_qp(wc->qp);
3849 	u32 hr_opcode;
3850 	int ib_opcode;
3851 	int ret;
3852 
3853 	wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3854 
3855 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3856 	switch (hr_opcode) {
3857 	case HNS_ROCE_V2_OPCODE_RDMA_WRITE_IMM:
3858 	case HNS_ROCE_V2_OPCODE_SEND_WITH_IMM:
3859 		wc->wc_flags = IB_WC_WITH_IMM;
3860 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->immtdata));
3861 		break;
3862 	case HNS_ROCE_V2_OPCODE_SEND_WITH_INV:
3863 		wc->wc_flags = IB_WC_WITH_INVALIDATE;
3864 		wc->ex.invalidate_rkey = le32_to_cpu(cqe->rkey);
3865 		break;
3866 	default:
3867 		wc->wc_flags = 0;
3868 	}
3869 
3870 	ib_opcode = to_ib_wc_recv_op(hr_opcode);
3871 	if (ib_opcode < 0)
3872 		wc->status = IB_WC_GENERAL_ERR;
3873 	else
3874 		wc->opcode = ib_opcode;
3875 
3876 	if (is_rq_inl_enabled(wc, hr_opcode, cqe)) {
3877 		ret = hns_roce_handle_recv_inl_wqe(cqe, qp, wc);
3878 		if (unlikely(ret))
3879 			return ret;
3880 	}
3881 
3882 	wc->sl = hr_reg_read(cqe, CQE_SL);
3883 	wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
3884 	wc->slid = 0;
3885 	wc->wc_flags |= hr_reg_read(cqe, CQE_GRH) ? IB_WC_GRH : 0;
3886 	wc->port_num = hr_reg_read(cqe, CQE_PORTN);
3887 	wc->pkey_index = 0;
3888 
3889 	if (hr_reg_read(cqe, CQE_VID_VLD)) {
3890 		wc->vlan_id = hr_reg_read(cqe, CQE_VID);
3891 		wc->wc_flags |= IB_WC_WITH_VLAN;
3892 	} else {
3893 		wc->vlan_id = 0xffff;
3894 	}
3895 
3896 	wc->network_hdr_type = hr_reg_read(cqe, CQE_PORT_TYPE);
3897 
3898 	return 0;
3899 }
3900 
3901 static int hns_roce_v2_poll_one(struct hns_roce_cq *hr_cq,
3902 				struct hns_roce_qp **cur_qp, struct ib_wc *wc)
3903 {
3904 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3905 	struct hns_roce_qp *qp = *cur_qp;
3906 	struct hns_roce_srq *srq = NULL;
3907 	struct hns_roce_v2_cqe *cqe;
3908 	struct hns_roce_wq *wq;
3909 	int is_send;
3910 	u16 wqe_idx;
3911 	int ret;
3912 
3913 	cqe = get_sw_cqe_v2(hr_cq, hr_cq->cons_index);
3914 	if (!cqe)
3915 		return -EAGAIN;
3916 
3917 	++hr_cq->cons_index;
3918 	/* Memory barrier */
3919 	rmb();
3920 
3921 	ret = get_cur_qp(hr_cq, cqe, &qp);
3922 	if (ret)
3923 		return ret;
3924 
3925 	wc->qp = &qp->ibqp;
3926 	wc->vendor_err = 0;
3927 
3928 	wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
3929 
3930 	is_send = !hr_reg_read(cqe, CQE_S_R);
3931 	if (is_send) {
3932 		wq = &qp->sq;
3933 
3934 		/* If sg_signal_bit is set, tail pointer will be updated to
3935 		 * the WQE corresponding to the current CQE.
3936 		 */
3937 		if (qp->sq_signal_bits)
3938 			wq->tail += (wqe_idx - (u16)wq->tail) &
3939 				    (wq->wqe_cnt - 1);
3940 
3941 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3942 		++wq->tail;
3943 
3944 		fill_send_wc(wc, cqe);
3945 	} else {
3946 		if (qp->ibqp.srq) {
3947 			srq = to_hr_srq(qp->ibqp.srq);
3948 			wc->wr_id = srq->wrid[wqe_idx];
3949 			hns_roce_free_srq_wqe(srq, wqe_idx);
3950 		} else {
3951 			wq = &qp->rq;
3952 			wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3953 			++wq->tail;
3954 		}
3955 
3956 		ret = fill_recv_wc(wc, cqe);
3957 	}
3958 
3959 	get_cqe_status(hr_dev, qp, hr_cq, cqe, wc);
3960 	if (unlikely(wc->status != IB_WC_SUCCESS))
3961 		return 0;
3962 
3963 	return ret;
3964 }
3965 
3966 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3967 			       struct ib_wc *wc)
3968 {
3969 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3970 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3971 	struct hns_roce_qp *cur_qp = NULL;
3972 	unsigned long flags;
3973 	int npolled;
3974 
3975 	spin_lock_irqsave(&hr_cq->lock, flags);
3976 
3977 	/*
3978 	 * When the device starts to reset, the state is RST_DOWN. At this time,
3979 	 * there may still be some valid CQEs in the hardware that are not
3980 	 * polled. Therefore, it is not allowed to switch to the software mode
3981 	 * immediately. When the state changes to UNINIT, CQE no longer exists
3982 	 * in the hardware, and then switch to software mode.
3983 	 */
3984 	if (hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT) {
3985 		npolled = hns_roce_v2_sw_poll_cq(hr_cq, num_entries, wc);
3986 		goto out;
3987 	}
3988 
3989 	for (npolled = 0; npolled < num_entries; ++npolled) {
3990 		if (hns_roce_v2_poll_one(hr_cq, &cur_qp, wc + npolled))
3991 			break;
3992 	}
3993 
3994 	if (npolled)
3995 		update_cq_db(hr_dev, hr_cq);
3996 
3997 out:
3998 	spin_unlock_irqrestore(&hr_cq->lock, flags);
3999 
4000 	return npolled;
4001 }
4002 
4003 static int get_op_for_set_hem(struct hns_roce_dev *hr_dev, u32 type,
4004 			      u32 step_idx, u8 *mbox_cmd)
4005 {
4006 	u8 cmd;
4007 
4008 	switch (type) {
4009 	case HEM_TYPE_QPC:
4010 		cmd = HNS_ROCE_CMD_WRITE_QPC_BT0;
4011 		break;
4012 	case HEM_TYPE_MTPT:
4013 		cmd = HNS_ROCE_CMD_WRITE_MPT_BT0;
4014 		break;
4015 	case HEM_TYPE_CQC:
4016 		cmd = HNS_ROCE_CMD_WRITE_CQC_BT0;
4017 		break;
4018 	case HEM_TYPE_SRQC:
4019 		cmd = HNS_ROCE_CMD_WRITE_SRQC_BT0;
4020 		break;
4021 	case HEM_TYPE_SCCC:
4022 		cmd = HNS_ROCE_CMD_WRITE_SCCC_BT0;
4023 		break;
4024 	case HEM_TYPE_QPC_TIMER:
4025 		cmd = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
4026 		break;
4027 	case HEM_TYPE_CQC_TIMER:
4028 		cmd = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
4029 		break;
4030 	default:
4031 		dev_warn(hr_dev->dev, "failed to check hem type %u.\n", type);
4032 		return -EINVAL;
4033 	}
4034 
4035 	*mbox_cmd = cmd + step_idx;
4036 
4037 	return 0;
4038 }
4039 
4040 static int config_gmv_ba_to_hw(struct hns_roce_dev *hr_dev, unsigned long obj,
4041 			       dma_addr_t base_addr)
4042 {
4043 	struct hns_roce_cmq_desc desc;
4044 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
4045 	u32 idx = obj / (HNS_HW_PAGE_SIZE / hr_dev->caps.gmv_entry_sz);
4046 	u64 addr = to_hr_hw_page_addr(base_addr);
4047 
4048 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
4049 
4050 	hr_reg_write(req, CFG_GMV_BT_BA_L, lower_32_bits(addr));
4051 	hr_reg_write(req, CFG_GMV_BT_BA_H, upper_32_bits(addr));
4052 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
4053 
4054 	return hns_roce_cmq_send(hr_dev, &desc, 1);
4055 }
4056 
4057 static int set_hem_to_hw(struct hns_roce_dev *hr_dev, int obj,
4058 			 dma_addr_t base_addr, u32 hem_type, u32 step_idx)
4059 {
4060 	int ret;
4061 	u8 cmd;
4062 
4063 	if (unlikely(hem_type == HEM_TYPE_GMV))
4064 		return config_gmv_ba_to_hw(hr_dev, obj, base_addr);
4065 
4066 	if (unlikely(hem_type == HEM_TYPE_SCCC && step_idx))
4067 		return 0;
4068 
4069 	ret = get_op_for_set_hem(hr_dev, hem_type, step_idx, &cmd);
4070 	if (ret < 0)
4071 		return ret;
4072 
4073 	return config_hem_ba_to_hw(hr_dev, base_addr, cmd, obj);
4074 }
4075 
4076 static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
4077 			       struct hns_roce_hem_table *table, int obj,
4078 			       u32 step_idx)
4079 {
4080 	struct hns_roce_hem_iter iter;
4081 	struct hns_roce_hem_mhop mhop;
4082 	struct hns_roce_hem *hem;
4083 	unsigned long mhop_obj = obj;
4084 	int i, j, k;
4085 	int ret = 0;
4086 	u64 hem_idx = 0;
4087 	u64 l1_idx = 0;
4088 	u64 bt_ba = 0;
4089 	u32 chunk_ba_num;
4090 	u32 hop_num;
4091 
4092 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4093 		return 0;
4094 
4095 	hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
4096 	i = mhop.l0_idx;
4097 	j = mhop.l1_idx;
4098 	k = mhop.l2_idx;
4099 	hop_num = mhop.hop_num;
4100 	chunk_ba_num = mhop.bt_chunk_size / 8;
4101 
4102 	if (hop_num == 2) {
4103 		hem_idx = i * chunk_ba_num * chunk_ba_num + j * chunk_ba_num +
4104 			  k;
4105 		l1_idx = i * chunk_ba_num + j;
4106 	} else if (hop_num == 1) {
4107 		hem_idx = i * chunk_ba_num + j;
4108 	} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
4109 		hem_idx = i;
4110 	}
4111 
4112 	if (table->type == HEM_TYPE_SCCC)
4113 		obj = mhop.l0_idx;
4114 
4115 	if (check_whether_last_step(hop_num, step_idx)) {
4116 		hem = table->hem[hem_idx];
4117 		for (hns_roce_hem_first(hem, &iter);
4118 		     !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
4119 			bt_ba = hns_roce_hem_addr(&iter);
4120 			ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type,
4121 					    step_idx);
4122 		}
4123 	} else {
4124 		if (step_idx == 0)
4125 			bt_ba = table->bt_l0_dma_addr[i];
4126 		else if (step_idx == 1 && hop_num == 2)
4127 			bt_ba = table->bt_l1_dma_addr[l1_idx];
4128 
4129 		ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type, step_idx);
4130 	}
4131 
4132 	return ret;
4133 }
4134 
4135 static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
4136 				 struct hns_roce_hem_table *table,
4137 				 int tag, u32 step_idx)
4138 {
4139 	struct hns_roce_cmd_mailbox *mailbox;
4140 	struct device *dev = hr_dev->dev;
4141 	u8 cmd = 0xff;
4142 	int ret;
4143 
4144 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4145 		return 0;
4146 
4147 	switch (table->type) {
4148 	case HEM_TYPE_QPC:
4149 		cmd = HNS_ROCE_CMD_DESTROY_QPC_BT0;
4150 		break;
4151 	case HEM_TYPE_MTPT:
4152 		cmd = HNS_ROCE_CMD_DESTROY_MPT_BT0;
4153 		break;
4154 	case HEM_TYPE_CQC:
4155 		cmd = HNS_ROCE_CMD_DESTROY_CQC_BT0;
4156 		break;
4157 	case HEM_TYPE_SRQC:
4158 		cmd = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
4159 		break;
4160 	case HEM_TYPE_SCCC:
4161 	case HEM_TYPE_QPC_TIMER:
4162 	case HEM_TYPE_CQC_TIMER:
4163 	case HEM_TYPE_GMV:
4164 		return 0;
4165 	default:
4166 		dev_warn(dev, "table %u not to be destroyed by mailbox!\n",
4167 			 table->type);
4168 		return 0;
4169 	}
4170 
4171 	cmd += step_idx;
4172 
4173 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4174 	if (IS_ERR(mailbox))
4175 		return PTR_ERR(mailbox);
4176 
4177 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, cmd, tag);
4178 
4179 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4180 	return ret;
4181 }
4182 
4183 static int hns_roce_v2_qp_modify(struct hns_roce_dev *hr_dev,
4184 				 struct hns_roce_v2_qp_context *context,
4185 				 struct hns_roce_v2_qp_context *qpc_mask,
4186 				 struct hns_roce_qp *hr_qp)
4187 {
4188 	struct hns_roce_cmd_mailbox *mailbox;
4189 	int qpc_size;
4190 	int ret;
4191 
4192 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4193 	if (IS_ERR(mailbox))
4194 		return PTR_ERR(mailbox);
4195 
4196 	/* The qpc size of HIP08 is only 256B, which is half of HIP09 */
4197 	qpc_size = hr_dev->caps.qpc_sz;
4198 	memcpy(mailbox->buf, context, qpc_size);
4199 	memcpy(mailbox->buf + qpc_size, qpc_mask, qpc_size);
4200 
4201 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
4202 				HNS_ROCE_CMD_MODIFY_QPC, hr_qp->qpn);
4203 
4204 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4205 
4206 	return ret;
4207 }
4208 
4209 static void set_access_flags(struct hns_roce_qp *hr_qp,
4210 			     struct hns_roce_v2_qp_context *context,
4211 			     struct hns_roce_v2_qp_context *qpc_mask,
4212 			     const struct ib_qp_attr *attr, int attr_mask)
4213 {
4214 	u8 dest_rd_atomic;
4215 	u32 access_flags;
4216 
4217 	dest_rd_atomic = (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) ?
4218 			 attr->max_dest_rd_atomic : hr_qp->resp_depth;
4219 
4220 	access_flags = (attr_mask & IB_QP_ACCESS_FLAGS) ?
4221 		       attr->qp_access_flags : hr_qp->atomic_rd_en;
4222 
4223 	if (!dest_rd_atomic)
4224 		access_flags &= IB_ACCESS_REMOTE_WRITE;
4225 
4226 	hr_reg_write_bool(context, QPC_RRE,
4227 			  access_flags & IB_ACCESS_REMOTE_READ);
4228 	hr_reg_clear(qpc_mask, QPC_RRE);
4229 
4230 	hr_reg_write_bool(context, QPC_RWE,
4231 			  access_flags & IB_ACCESS_REMOTE_WRITE);
4232 	hr_reg_clear(qpc_mask, QPC_RWE);
4233 
4234 	hr_reg_write_bool(context, QPC_ATE,
4235 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4236 	hr_reg_clear(qpc_mask, QPC_ATE);
4237 	hr_reg_write_bool(context, QPC_EXT_ATE,
4238 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4239 	hr_reg_clear(qpc_mask, QPC_EXT_ATE);
4240 }
4241 
4242 static void set_qpc_wqe_cnt(struct hns_roce_qp *hr_qp,
4243 			    struct hns_roce_v2_qp_context *context,
4244 			    struct hns_roce_v2_qp_context *qpc_mask)
4245 {
4246 	hr_reg_write(context, QPC_SGE_SHIFT,
4247 		     to_hr_hem_entries_shift(hr_qp->sge.sge_cnt,
4248 					     hr_qp->sge.sge_shift));
4249 
4250 	hr_reg_write(context, QPC_SQ_SHIFT, ilog2(hr_qp->sq.wqe_cnt));
4251 
4252 	hr_reg_write(context, QPC_RQ_SHIFT, ilog2(hr_qp->rq.wqe_cnt));
4253 }
4254 
4255 static inline int get_cqn(struct ib_cq *ib_cq)
4256 {
4257 	return ib_cq ? to_hr_cq(ib_cq)->cqn : 0;
4258 }
4259 
4260 static inline int get_pdn(struct ib_pd *ib_pd)
4261 {
4262 	return ib_pd ? to_hr_pd(ib_pd)->pdn : 0;
4263 }
4264 
4265 static void modify_qp_reset_to_init(struct ib_qp *ibqp,
4266 				    const struct ib_qp_attr *attr,
4267 				    int attr_mask,
4268 				    struct hns_roce_v2_qp_context *context,
4269 				    struct hns_roce_v2_qp_context *qpc_mask)
4270 {
4271 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4272 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4273 
4274 	/*
4275 	 * In v2 engine, software pass context and context mask to hardware
4276 	 * when modifying qp. If software need modify some fields in context,
4277 	 * we should set all bits of the relevant fields in context mask to
4278 	 * 0 at the same time, else set them to 0x1.
4279 	 */
4280 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4281 
4282 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4283 
4284 	hr_reg_write(context, QPC_RQWS, ilog2(hr_qp->rq.max_gs));
4285 
4286 	set_qpc_wqe_cnt(hr_qp, context, qpc_mask);
4287 
4288 	/* No VLAN need to set 0xFFF */
4289 	hr_reg_write(context, QPC_VLAN_ID, 0xfff);
4290 
4291 	if (ibqp->qp_type == IB_QPT_XRC_TGT) {
4292 		context->qkey_xrcd = cpu_to_le32(hr_qp->xrcdn);
4293 
4294 		hr_reg_enable(context, QPC_XRC_QP_TYPE);
4295 	}
4296 
4297 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
4298 		hr_reg_enable(context, QPC_RQ_RECORD_EN);
4299 
4300 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
4301 		hr_reg_enable(context, QPC_OWNER_MODE);
4302 
4303 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_L,
4304 		     lower_32_bits(hr_qp->rdb.dma) >> 1);
4305 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
4306 		     upper_32_bits(hr_qp->rdb.dma));
4307 
4308 	if (ibqp->qp_type != IB_QPT_UD && ibqp->qp_type != IB_QPT_GSI)
4309 		hr_reg_write_bool(context, QPC_RQIE,
4310 			     hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE);
4311 
4312 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4313 
4314 	if (ibqp->srq) {
4315 		hr_reg_enable(context, QPC_SRQ_EN);
4316 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4317 	}
4318 
4319 	hr_reg_enable(context, QPC_FRE);
4320 
4321 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4322 
4323 	if (hr_dev->caps.qpc_sz < HNS_ROCE_V3_QPC_SZ)
4324 		return;
4325 
4326 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
4327 		hr_reg_enable(&context->ext, QPCEX_STASH);
4328 }
4329 
4330 static void modify_qp_init_to_init(struct ib_qp *ibqp,
4331 				   const struct ib_qp_attr *attr, int attr_mask,
4332 				   struct hns_roce_v2_qp_context *context,
4333 				   struct hns_roce_v2_qp_context *qpc_mask)
4334 {
4335 	/*
4336 	 * In v2 engine, software pass context and context mask to hardware
4337 	 * when modifying qp. If software need modify some fields in context,
4338 	 * we should set all bits of the relevant fields in context mask to
4339 	 * 0 at the same time, else set them to 0x1.
4340 	 */
4341 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4342 	hr_reg_clear(qpc_mask, QPC_TST);
4343 
4344 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4345 	hr_reg_clear(qpc_mask, QPC_PD);
4346 
4347 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4348 	hr_reg_clear(qpc_mask, QPC_RX_CQN);
4349 
4350 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4351 	hr_reg_clear(qpc_mask, QPC_TX_CQN);
4352 
4353 	if (ibqp->srq) {
4354 		hr_reg_enable(context, QPC_SRQ_EN);
4355 		hr_reg_clear(qpc_mask, QPC_SRQ_EN);
4356 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4357 		hr_reg_clear(qpc_mask, QPC_SRQN);
4358 	}
4359 }
4360 
4361 static int config_qp_rq_buf(struct hns_roce_dev *hr_dev,
4362 			    struct hns_roce_qp *hr_qp,
4363 			    struct hns_roce_v2_qp_context *context,
4364 			    struct hns_roce_v2_qp_context *qpc_mask)
4365 {
4366 	u64 mtts[MTT_MIN_COUNT] = { 0 };
4367 	u64 wqe_sge_ba;
4368 	int count;
4369 
4370 	/* Search qp buf's mtts */
4371 	count = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->rq.offset, mtts,
4372 				  MTT_MIN_COUNT, &wqe_sge_ba);
4373 	if (hr_qp->rq.wqe_cnt && count < 1) {
4374 		ibdev_err(&hr_dev->ib_dev,
4375 			  "failed to find RQ WQE, QPN = 0x%lx.\n", hr_qp->qpn);
4376 		return -EINVAL;
4377 	}
4378 
4379 	context->wqe_sge_ba = cpu_to_le32(wqe_sge_ba >> 3);
4380 	qpc_mask->wqe_sge_ba = 0;
4381 
4382 	/*
4383 	 * In v2 engine, software pass context and context mask to hardware
4384 	 * when modifying qp. If software need modify some fields in context,
4385 	 * we should set all bits of the relevant fields in context mask to
4386 	 * 0 at the same time, else set them to 0x1.
4387 	 */
4388 	hr_reg_write(context, QPC_WQE_SGE_BA_H, wqe_sge_ba >> (32 + 3));
4389 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_H);
4390 
4391 	hr_reg_write(context, QPC_SQ_HOP_NUM,
4392 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sq_hop_num,
4393 				      hr_qp->sq.wqe_cnt));
4394 	hr_reg_clear(qpc_mask, QPC_SQ_HOP_NUM);
4395 
4396 	hr_reg_write(context, QPC_SGE_HOP_NUM,
4397 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sge_hop_num,
4398 				      hr_qp->sge.sge_cnt));
4399 	hr_reg_clear(qpc_mask, QPC_SGE_HOP_NUM);
4400 
4401 	hr_reg_write(context, QPC_RQ_HOP_NUM,
4402 		     to_hr_hem_hopnum(hr_dev->caps.wqe_rq_hop_num,
4403 				      hr_qp->rq.wqe_cnt));
4404 
4405 	hr_reg_clear(qpc_mask, QPC_RQ_HOP_NUM);
4406 
4407 	hr_reg_write(context, QPC_WQE_SGE_BA_PG_SZ,
4408 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.ba_pg_shift));
4409 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_PG_SZ);
4410 
4411 	hr_reg_write(context, QPC_WQE_SGE_BUF_PG_SZ,
4412 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.buf_pg_shift));
4413 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BUF_PG_SZ);
4414 
4415 	context->rq_cur_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[0]));
4416 	qpc_mask->rq_cur_blk_addr = 0;
4417 
4418 	hr_reg_write(context, QPC_RQ_CUR_BLK_ADDR_H,
4419 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
4420 	hr_reg_clear(qpc_mask, QPC_RQ_CUR_BLK_ADDR_H);
4421 
4422 	context->rq_nxt_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[1]));
4423 	qpc_mask->rq_nxt_blk_addr = 0;
4424 
4425 	hr_reg_write(context, QPC_RQ_NXT_BLK_ADDR_H,
4426 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
4427 	hr_reg_clear(qpc_mask, QPC_RQ_NXT_BLK_ADDR_H);
4428 
4429 	return 0;
4430 }
4431 
4432 static int config_qp_sq_buf(struct hns_roce_dev *hr_dev,
4433 			    struct hns_roce_qp *hr_qp,
4434 			    struct hns_roce_v2_qp_context *context,
4435 			    struct hns_roce_v2_qp_context *qpc_mask)
4436 {
4437 	struct ib_device *ibdev = &hr_dev->ib_dev;
4438 	u64 sge_cur_blk = 0;
4439 	u64 sq_cur_blk = 0;
4440 	int count;
4441 
4442 	/* search qp buf's mtts */
4443 	count = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, 0, &sq_cur_blk, 1, NULL);
4444 	if (count < 1) {
4445 		ibdev_err(ibdev, "failed to find QP(0x%lx) SQ buf.\n",
4446 			  hr_qp->qpn);
4447 		return -EINVAL;
4448 	}
4449 	if (hr_qp->sge.sge_cnt > 0) {
4450 		count = hns_roce_mtr_find(hr_dev, &hr_qp->mtr,
4451 					  hr_qp->sge.offset,
4452 					  &sge_cur_blk, 1, NULL);
4453 		if (count < 1) {
4454 			ibdev_err(ibdev, "failed to find QP(0x%lx) SGE buf.\n",
4455 				  hr_qp->qpn);
4456 			return -EINVAL;
4457 		}
4458 	}
4459 
4460 	/*
4461 	 * In v2 engine, software pass context and context mask to hardware
4462 	 * when modifying qp. If software need modify some fields in context,
4463 	 * we should set all bits of the relevant fields in context mask to
4464 	 * 0 at the same time, else set them to 0x1.
4465 	 */
4466 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_L,
4467 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4468 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_H,
4469 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4470 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_L);
4471 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_H);
4472 
4473 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_L,
4474 		     lower_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4475 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_H,
4476 		     upper_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4477 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_L);
4478 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_H);
4479 
4480 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_L,
4481 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4482 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_H,
4483 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4484 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_L);
4485 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_H);
4486 
4487 	return 0;
4488 }
4489 
4490 static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
4491 				  const struct ib_qp_attr *attr)
4492 {
4493 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_UD)
4494 		return IB_MTU_4096;
4495 
4496 	return attr->path_mtu;
4497 }
4498 
4499 static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
4500 				 const struct ib_qp_attr *attr, int attr_mask,
4501 				 struct hns_roce_v2_qp_context *context,
4502 				 struct hns_roce_v2_qp_context *qpc_mask)
4503 {
4504 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4505 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4506 	struct ib_device *ibdev = &hr_dev->ib_dev;
4507 	dma_addr_t trrl_ba;
4508 	dma_addr_t irrl_ba;
4509 	enum ib_mtu ib_mtu;
4510 	const u8 *smac;
4511 	u8 lp_pktn_ini;
4512 	u64 *mtts;
4513 	u8 *dmac;
4514 	u32 port;
4515 	int mtu;
4516 	int ret;
4517 
4518 	ret = config_qp_rq_buf(hr_dev, hr_qp, context, qpc_mask);
4519 	if (ret) {
4520 		ibdev_err(ibdev, "failed to config rq buf, ret = %d.\n", ret);
4521 		return ret;
4522 	}
4523 
4524 	/* Search IRRL's mtts */
4525 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.irrl_table,
4526 				   hr_qp->qpn, &irrl_ba);
4527 	if (!mtts) {
4528 		ibdev_err(ibdev, "failed to find qp irrl_table.\n");
4529 		return -EINVAL;
4530 	}
4531 
4532 	/* Search TRRL's mtts */
4533 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.trrl_table,
4534 				   hr_qp->qpn, &trrl_ba);
4535 	if (!mtts) {
4536 		ibdev_err(ibdev, "failed to find qp trrl_table.\n");
4537 		return -EINVAL;
4538 	}
4539 
4540 	if (attr_mask & IB_QP_ALT_PATH) {
4541 		ibdev_err(ibdev, "INIT2RTR attr_mask (0x%x) error.\n",
4542 			  attr_mask);
4543 		return -EINVAL;
4544 	}
4545 
4546 	hr_reg_write(context, QPC_TRRL_BA_L, trrl_ba >> 4);
4547 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_L);
4548 	context->trrl_ba = cpu_to_le32(trrl_ba >> (16 + 4));
4549 	qpc_mask->trrl_ba = 0;
4550 	hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> (32 + 16 + 4));
4551 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
4552 
4553 	context->irrl_ba = cpu_to_le32(irrl_ba >> 6);
4554 	qpc_mask->irrl_ba = 0;
4555 	hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> (32 + 6));
4556 	hr_reg_clear(qpc_mask, QPC_IRRL_BA_H);
4557 
4558 	hr_reg_enable(context, QPC_RMT_E2E);
4559 	hr_reg_clear(qpc_mask, QPC_RMT_E2E);
4560 
4561 	hr_reg_write(context, QPC_SIG_TYPE, hr_qp->sq_signal_bits);
4562 	hr_reg_clear(qpc_mask, QPC_SIG_TYPE);
4563 
4564 	port = (attr_mask & IB_QP_PORT) ? (attr->port_num - 1) : hr_qp->port;
4565 
4566 	smac = (const u8 *)hr_dev->dev_addr[port];
4567 	dmac = (u8 *)attr->ah_attr.roce.dmac;
4568 	/* when dmac equals smac or loop_idc is 1, it should loopback */
4569 	if (ether_addr_equal_unaligned(dmac, smac) ||
4570 	    hr_dev->loop_idc == 0x1) {
4571 		hr_reg_write(context, QPC_LBI, hr_dev->loop_idc);
4572 		hr_reg_clear(qpc_mask, QPC_LBI);
4573 	}
4574 
4575 	if (attr_mask & IB_QP_DEST_QPN) {
4576 		hr_reg_write(context, QPC_DQPN, attr->dest_qp_num);
4577 		hr_reg_clear(qpc_mask, QPC_DQPN);
4578 	}
4579 
4580 	memcpy(&(context->dmac), dmac, sizeof(u32));
4581 	hr_reg_write(context, QPC_DMAC_H, *((u16 *)(&dmac[4])));
4582 	qpc_mask->dmac = 0;
4583 	hr_reg_clear(qpc_mask, QPC_DMAC_H);
4584 
4585 	ib_mtu = get_mtu(ibqp, attr);
4586 	hr_qp->path_mtu = ib_mtu;
4587 
4588 	mtu = ib_mtu_enum_to_int(ib_mtu);
4589 	if (WARN_ON(mtu <= 0))
4590 		return -EINVAL;
4591 #define MAX_LP_MSG_LEN 16384
4592 	/* MTU * (2 ^ LP_PKTN_INI) shouldn't be bigger than 16KB */
4593 	lp_pktn_ini = ilog2(MAX_LP_MSG_LEN / mtu);
4594 	if (WARN_ON(lp_pktn_ini >= 0xF))
4595 		return -EINVAL;
4596 
4597 	if (attr_mask & IB_QP_PATH_MTU) {
4598 		hr_reg_write(context, QPC_MTU, ib_mtu);
4599 		hr_reg_clear(qpc_mask, QPC_MTU);
4600 	}
4601 
4602 	hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
4603 	hr_reg_clear(qpc_mask, QPC_LP_PKTN_INI);
4604 
4605 	/* ACK_REQ_FREQ should be larger than or equal to LP_PKTN_INI */
4606 	hr_reg_write(context, QPC_ACK_REQ_FREQ, lp_pktn_ini);
4607 	hr_reg_clear(qpc_mask, QPC_ACK_REQ_FREQ);
4608 
4609 	hr_reg_clear(qpc_mask, QPC_RX_REQ_PSN_ERR);
4610 	hr_reg_clear(qpc_mask, QPC_RX_REQ_MSN);
4611 	hr_reg_clear(qpc_mask, QPC_RX_REQ_LAST_OPTYPE);
4612 
4613 	context->rq_rnr_timer = 0;
4614 	qpc_mask->rq_rnr_timer = 0;
4615 
4616 	hr_reg_clear(qpc_mask, QPC_TRRL_HEAD_MAX);
4617 	hr_reg_clear(qpc_mask, QPC_TRRL_TAIL_MAX);
4618 
4619 	/* rocee send 2^lp_sgen_ini segs every time */
4620 	hr_reg_write(context, QPC_LP_SGEN_INI, 3);
4621 	hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
4622 
4623 	return 0;
4624 }
4625 
4626 static int modify_qp_rtr_to_rts(struct ib_qp *ibqp,
4627 				const struct ib_qp_attr *attr, int attr_mask,
4628 				struct hns_roce_v2_qp_context *context,
4629 				struct hns_roce_v2_qp_context *qpc_mask)
4630 {
4631 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4632 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4633 	struct ib_device *ibdev = &hr_dev->ib_dev;
4634 	int ret;
4635 
4636 	/* Not support alternate path and path migration */
4637 	if (attr_mask & (IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE)) {
4638 		ibdev_err(ibdev, "RTR2RTS attr_mask (0x%x)error\n", attr_mask);
4639 		return -EINVAL;
4640 	}
4641 
4642 	ret = config_qp_sq_buf(hr_dev, hr_qp, context, qpc_mask);
4643 	if (ret) {
4644 		ibdev_err(ibdev, "failed to config sq buf, ret = %d.\n", ret);
4645 		return ret;
4646 	}
4647 
4648 	/*
4649 	 * Set some fields in context to zero, Because the default values
4650 	 * of all fields in context are zero, we need not set them to 0 again.
4651 	 * but we should set the relevant fields of context mask to 0.
4652 	 */
4653 	hr_reg_clear(qpc_mask, QPC_IRRL_SGE_IDX);
4654 
4655 	hr_reg_clear(qpc_mask, QPC_RX_ACK_MSN);
4656 
4657 	hr_reg_clear(qpc_mask, QPC_ACK_LAST_OPTYPE);
4658 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN_VLD);
4659 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN);
4660 
4661 	hr_reg_clear(qpc_mask, QPC_IRRL_TAIL_REAL);
4662 
4663 	hr_reg_clear(qpc_mask, QPC_RETRY_MSG_MSN);
4664 
4665 	hr_reg_clear(qpc_mask, QPC_RNR_RETRY_FLAG);
4666 
4667 	hr_reg_clear(qpc_mask, QPC_CHECK_FLG);
4668 
4669 	hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
4670 
4671 	return 0;
4672 }
4673 
4674 static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4675 			   u32 *dip_idx)
4676 {
4677 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4678 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4679 	u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx;
4680 	u32 *head =  &hr_dev->qp_table.idx_table.head;
4681 	u32 *tail =  &hr_dev->qp_table.idx_table.tail;
4682 	struct hns_roce_dip *hr_dip;
4683 	unsigned long flags;
4684 	int ret = 0;
4685 
4686 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
4687 
4688 	spare_idx[*tail] = ibqp->qp_num;
4689 	*tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1);
4690 
4691 	list_for_each_entry(hr_dip, &hr_dev->dip_list, node) {
4692 		if (!memcmp(grh->dgid.raw, hr_dip->dgid, 16)) {
4693 			*dip_idx = hr_dip->dip_idx;
4694 			goto out;
4695 		}
4696 	}
4697 
4698 	/* If no dgid is found, a new dip and a mapping between dgid and
4699 	 * dip_idx will be created.
4700 	 */
4701 	hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC);
4702 	if (!hr_dip) {
4703 		ret = -ENOMEM;
4704 		goto out;
4705 	}
4706 
4707 	memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4708 	hr_dip->dip_idx = *dip_idx = spare_idx[*head];
4709 	*head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1);
4710 	list_add_tail(&hr_dip->node, &hr_dev->dip_list);
4711 
4712 out:
4713 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
4714 	return ret;
4715 }
4716 
4717 enum {
4718 	CONG_DCQCN,
4719 	CONG_WINDOW,
4720 };
4721 
4722 enum {
4723 	UNSUPPORT_CONG_LEVEL,
4724 	SUPPORT_CONG_LEVEL,
4725 };
4726 
4727 enum {
4728 	CONG_LDCP,
4729 	CONG_HC3,
4730 };
4731 
4732 enum {
4733 	DIP_INVALID,
4734 	DIP_VALID,
4735 };
4736 
4737 enum {
4738 	WND_LIMIT,
4739 	WND_UNLIMIT,
4740 };
4741 
4742 static int check_cong_type(struct ib_qp *ibqp,
4743 			   struct hns_roce_congestion_algorithm *cong_alg)
4744 {
4745 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4746 
4747 	/* different congestion types match different configurations */
4748 	switch (hr_dev->caps.cong_type) {
4749 	case CONG_TYPE_DCQCN:
4750 		cong_alg->alg_sel = CONG_DCQCN;
4751 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4752 		cong_alg->dip_vld = DIP_INVALID;
4753 		cong_alg->wnd_mode_sel = WND_LIMIT;
4754 		break;
4755 	case CONG_TYPE_LDCP:
4756 		cong_alg->alg_sel = CONG_WINDOW;
4757 		cong_alg->alg_sub_sel = CONG_LDCP;
4758 		cong_alg->dip_vld = DIP_INVALID;
4759 		cong_alg->wnd_mode_sel = WND_UNLIMIT;
4760 		break;
4761 	case CONG_TYPE_HC3:
4762 		cong_alg->alg_sel = CONG_WINDOW;
4763 		cong_alg->alg_sub_sel = CONG_HC3;
4764 		cong_alg->dip_vld = DIP_INVALID;
4765 		cong_alg->wnd_mode_sel = WND_LIMIT;
4766 		break;
4767 	case CONG_TYPE_DIP:
4768 		cong_alg->alg_sel = CONG_DCQCN;
4769 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4770 		cong_alg->dip_vld = DIP_VALID;
4771 		cong_alg->wnd_mode_sel = WND_LIMIT;
4772 		break;
4773 	default:
4774 		ibdev_err(&hr_dev->ib_dev,
4775 			  "error type(%u) for congestion selection.\n",
4776 			  hr_dev->caps.cong_type);
4777 		return -EINVAL;
4778 	}
4779 
4780 	return 0;
4781 }
4782 
4783 static int fill_cong_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4784 			   struct hns_roce_v2_qp_context *context,
4785 			   struct hns_roce_v2_qp_context *qpc_mask)
4786 {
4787 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4788 	struct hns_roce_congestion_algorithm cong_field;
4789 	struct ib_device *ibdev = ibqp->device;
4790 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
4791 	u32 dip_idx = 0;
4792 	int ret;
4793 
4794 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ||
4795 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE)
4796 		return 0;
4797 
4798 	ret = check_cong_type(ibqp, &cong_field);
4799 	if (ret)
4800 		return ret;
4801 
4802 	hr_reg_write(context, QPC_CONG_ALGO_TMPL_ID, hr_dev->cong_algo_tmpl_id +
4803 		     hr_dev->caps.cong_type * HNS_ROCE_CONG_SIZE);
4804 	hr_reg_clear(qpc_mask, QPC_CONG_ALGO_TMPL_ID);
4805 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SEL, cong_field.alg_sel);
4806 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SEL);
4807 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SUB_SEL,
4808 		     cong_field.alg_sub_sel);
4809 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SUB_SEL);
4810 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX_VLD, cong_field.dip_vld);
4811 	hr_reg_clear(&qpc_mask->ext, QPCEX_DIP_CTX_IDX_VLD);
4812 	hr_reg_write(&context->ext, QPCEX_SQ_RQ_NOT_FORBID_EN,
4813 		     cong_field.wnd_mode_sel);
4814 	hr_reg_clear(&qpc_mask->ext, QPCEX_SQ_RQ_NOT_FORBID_EN);
4815 
4816 	/* if dip is disabled, there is no need to set dip idx */
4817 	if (cong_field.dip_vld == 0)
4818 		return 0;
4819 
4820 	ret = get_dip_ctx_idx(ibqp, attr, &dip_idx);
4821 	if (ret) {
4822 		ibdev_err(ibdev, "failed to fill cong field, ret = %d.\n", ret);
4823 		return ret;
4824 	}
4825 
4826 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX, dip_idx);
4827 	hr_reg_write(&qpc_mask->ext, QPCEX_DIP_CTX_IDX, 0);
4828 
4829 	return 0;
4830 }
4831 
4832 static int hns_roce_v2_set_path(struct ib_qp *ibqp,
4833 				const struct ib_qp_attr *attr,
4834 				int attr_mask,
4835 				struct hns_roce_v2_qp_context *context,
4836 				struct hns_roce_v2_qp_context *qpc_mask)
4837 {
4838 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4839 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4840 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4841 	struct ib_device *ibdev = &hr_dev->ib_dev;
4842 	const struct ib_gid_attr *gid_attr = NULL;
4843 	int is_roce_protocol;
4844 	u16 vlan_id = 0xffff;
4845 	bool is_udp = false;
4846 	u8 ib_port;
4847 	u8 hr_port;
4848 	int ret;
4849 
4850 	/*
4851 	 * If free_mr_en of qp is set, it means that this qp comes from
4852 	 * free mr. This qp will perform the loopback operation.
4853 	 * In the loopback scenario, only sl needs to be set.
4854 	 */
4855 	if (hr_qp->free_mr_en) {
4856 		hr_reg_write(context, QPC_SL, rdma_ah_get_sl(&attr->ah_attr));
4857 		hr_reg_clear(qpc_mask, QPC_SL);
4858 		hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4859 		return 0;
4860 	}
4861 
4862 	ib_port = (attr_mask & IB_QP_PORT) ? attr->port_num : hr_qp->port + 1;
4863 	hr_port = ib_port - 1;
4864 	is_roce_protocol = rdma_cap_eth_ah(&hr_dev->ib_dev, ib_port) &&
4865 			   rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
4866 
4867 	if (is_roce_protocol) {
4868 		gid_attr = attr->ah_attr.grh.sgid_attr;
4869 		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
4870 		if (ret)
4871 			return ret;
4872 
4873 		is_udp = (gid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
4874 	}
4875 
4876 	/* Only HIP08 needs to set the vlan_en bits in QPC */
4877 	if (vlan_id < VLAN_N_VID &&
4878 	    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4879 		hr_reg_enable(context, QPC_RQ_VLAN_EN);
4880 		hr_reg_clear(qpc_mask, QPC_RQ_VLAN_EN);
4881 		hr_reg_enable(context, QPC_SQ_VLAN_EN);
4882 		hr_reg_clear(qpc_mask, QPC_SQ_VLAN_EN);
4883 	}
4884 
4885 	hr_reg_write(context, QPC_VLAN_ID, vlan_id);
4886 	hr_reg_clear(qpc_mask, QPC_VLAN_ID);
4887 
4888 	if (grh->sgid_index >= hr_dev->caps.gid_table_len[hr_port]) {
4889 		ibdev_err(ibdev, "sgid_index(%u) too large. max is %d\n",
4890 			  grh->sgid_index, hr_dev->caps.gid_table_len[hr_port]);
4891 		return -EINVAL;
4892 	}
4893 
4894 	if (attr->ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE) {
4895 		ibdev_err(ibdev, "ah attr is not RDMA roce type\n");
4896 		return -EINVAL;
4897 	}
4898 
4899 	hr_reg_write(context, QPC_UDPSPN,
4900 		     is_udp ? rdma_get_udp_sport(grh->flow_label, ibqp->qp_num,
4901 						 attr->dest_qp_num) :
4902 				    0);
4903 
4904 	hr_reg_clear(qpc_mask, QPC_UDPSPN);
4905 
4906 	hr_reg_write(context, QPC_GMV_IDX, grh->sgid_index);
4907 
4908 	hr_reg_clear(qpc_mask, QPC_GMV_IDX);
4909 
4910 	hr_reg_write(context, QPC_HOPLIMIT, grh->hop_limit);
4911 	hr_reg_clear(qpc_mask, QPC_HOPLIMIT);
4912 
4913 	ret = fill_cong_field(ibqp, attr, context, qpc_mask);
4914 	if (ret)
4915 		return ret;
4916 
4917 	hr_reg_write(context, QPC_TC, get_tclass(&attr->ah_attr.grh));
4918 	hr_reg_clear(qpc_mask, QPC_TC);
4919 
4920 	hr_reg_write(context, QPC_FL, grh->flow_label);
4921 	hr_reg_clear(qpc_mask, QPC_FL);
4922 	memcpy(context->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4923 	memset(qpc_mask->dgid, 0, sizeof(grh->dgid.raw));
4924 
4925 	hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4926 	if (unlikely(hr_qp->sl > MAX_SERVICE_LEVEL)) {
4927 		ibdev_err(ibdev,
4928 			  "failed to fill QPC, sl (%u) shouldn't be larger than %d.\n",
4929 			  hr_qp->sl, MAX_SERVICE_LEVEL);
4930 		return -EINVAL;
4931 	}
4932 
4933 	hr_reg_write(context, QPC_SL, hr_qp->sl);
4934 	hr_reg_clear(qpc_mask, QPC_SL);
4935 
4936 	return 0;
4937 }
4938 
4939 static bool check_qp_state(enum ib_qp_state cur_state,
4940 			   enum ib_qp_state new_state)
4941 {
4942 	static const bool sm[][IB_QPS_ERR + 1] = {
4943 		[IB_QPS_RESET] = { [IB_QPS_RESET] = true,
4944 				   [IB_QPS_INIT] = true },
4945 		[IB_QPS_INIT] = { [IB_QPS_RESET] = true,
4946 				  [IB_QPS_INIT] = true,
4947 				  [IB_QPS_RTR] = true,
4948 				  [IB_QPS_ERR] = true },
4949 		[IB_QPS_RTR] = { [IB_QPS_RESET] = true,
4950 				 [IB_QPS_RTS] = true,
4951 				 [IB_QPS_ERR] = true },
4952 		[IB_QPS_RTS] = { [IB_QPS_RESET] = true,
4953 				 [IB_QPS_RTS] = true,
4954 				 [IB_QPS_ERR] = true },
4955 		[IB_QPS_SQD] = {},
4956 		[IB_QPS_SQE] = {},
4957 		[IB_QPS_ERR] = { [IB_QPS_RESET] = true,
4958 				 [IB_QPS_ERR] = true }
4959 	};
4960 
4961 	return sm[cur_state][new_state];
4962 }
4963 
4964 static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
4965 				      const struct ib_qp_attr *attr,
4966 				      int attr_mask,
4967 				      enum ib_qp_state cur_state,
4968 				      enum ib_qp_state new_state,
4969 				      struct hns_roce_v2_qp_context *context,
4970 				      struct hns_roce_v2_qp_context *qpc_mask)
4971 {
4972 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4973 	int ret = 0;
4974 
4975 	if (!check_qp_state(cur_state, new_state)) {
4976 		ibdev_err(&hr_dev->ib_dev, "Illegal state for QP!\n");
4977 		return -EINVAL;
4978 	}
4979 
4980 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4981 		memset(qpc_mask, 0, hr_dev->caps.qpc_sz);
4982 		modify_qp_reset_to_init(ibqp, attr, attr_mask, context,
4983 					qpc_mask);
4984 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
4985 		modify_qp_init_to_init(ibqp, attr, attr_mask, context,
4986 				       qpc_mask);
4987 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4988 		ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
4989 					    qpc_mask);
4990 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
4991 		ret = modify_qp_rtr_to_rts(ibqp, attr, attr_mask, context,
4992 					   qpc_mask);
4993 	}
4994 
4995 	return ret;
4996 }
4997 
4998 static bool check_qp_timeout_cfg_range(struct hns_roce_dev *hr_dev, u8 *timeout)
4999 {
5000 #define QP_ACK_TIMEOUT_MAX_HIP08 20
5001 #define QP_ACK_TIMEOUT_OFFSET 10
5002 #define QP_ACK_TIMEOUT_MAX 31
5003 
5004 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5005 		if (*timeout > QP_ACK_TIMEOUT_MAX_HIP08) {
5006 			ibdev_warn(&hr_dev->ib_dev,
5007 				   "Local ACK timeout shall be 0 to 20.\n");
5008 			return false;
5009 		}
5010 		*timeout += QP_ACK_TIMEOUT_OFFSET;
5011 	} else if (hr_dev->pci_dev->revision > PCI_REVISION_ID_HIP08) {
5012 		if (*timeout > QP_ACK_TIMEOUT_MAX) {
5013 			ibdev_warn(&hr_dev->ib_dev,
5014 				   "Local ACK timeout shall be 0 to 31.\n");
5015 			return false;
5016 		}
5017 	}
5018 
5019 	return true;
5020 }
5021 
5022 static int hns_roce_v2_set_opt_fields(struct ib_qp *ibqp,
5023 				      const struct ib_qp_attr *attr,
5024 				      int attr_mask,
5025 				      struct hns_roce_v2_qp_context *context,
5026 				      struct hns_roce_v2_qp_context *qpc_mask)
5027 {
5028 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5029 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5030 	int ret = 0;
5031 	u8 timeout;
5032 
5033 	if (attr_mask & IB_QP_AV) {
5034 		ret = hns_roce_v2_set_path(ibqp, attr, attr_mask, context,
5035 					   qpc_mask);
5036 		if (ret)
5037 			return ret;
5038 	}
5039 
5040 	if (attr_mask & IB_QP_TIMEOUT) {
5041 		timeout = attr->timeout;
5042 		if (check_qp_timeout_cfg_range(hr_dev, &timeout)) {
5043 			hr_reg_write(context, QPC_AT, timeout);
5044 			hr_reg_clear(qpc_mask, QPC_AT);
5045 		}
5046 	}
5047 
5048 	if (attr_mask & IB_QP_RETRY_CNT) {
5049 		hr_reg_write(context, QPC_RETRY_NUM_INIT, attr->retry_cnt);
5050 		hr_reg_clear(qpc_mask, QPC_RETRY_NUM_INIT);
5051 
5052 		hr_reg_write(context, QPC_RETRY_CNT, attr->retry_cnt);
5053 		hr_reg_clear(qpc_mask, QPC_RETRY_CNT);
5054 	}
5055 
5056 	if (attr_mask & IB_QP_RNR_RETRY) {
5057 		hr_reg_write(context, QPC_RNR_NUM_INIT, attr->rnr_retry);
5058 		hr_reg_clear(qpc_mask, QPC_RNR_NUM_INIT);
5059 
5060 		hr_reg_write(context, QPC_RNR_CNT, attr->rnr_retry);
5061 		hr_reg_clear(qpc_mask, QPC_RNR_CNT);
5062 	}
5063 
5064 	if (attr_mask & IB_QP_SQ_PSN) {
5065 		hr_reg_write(context, QPC_SQ_CUR_PSN, attr->sq_psn);
5066 		hr_reg_clear(qpc_mask, QPC_SQ_CUR_PSN);
5067 
5068 		hr_reg_write(context, QPC_SQ_MAX_PSN, attr->sq_psn);
5069 		hr_reg_clear(qpc_mask, QPC_SQ_MAX_PSN);
5070 
5071 		hr_reg_write(context, QPC_RETRY_MSG_PSN_L, attr->sq_psn);
5072 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_L);
5073 
5074 		hr_reg_write(context, QPC_RETRY_MSG_PSN_H,
5075 			     attr->sq_psn >> RETRY_MSG_PSN_SHIFT);
5076 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_H);
5077 
5078 		hr_reg_write(context, QPC_RETRY_MSG_FPKT_PSN, attr->sq_psn);
5079 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_FPKT_PSN);
5080 
5081 		hr_reg_write(context, QPC_RX_ACK_EPSN, attr->sq_psn);
5082 		hr_reg_clear(qpc_mask, QPC_RX_ACK_EPSN);
5083 	}
5084 
5085 	if ((attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) &&
5086 	     attr->max_dest_rd_atomic) {
5087 		hr_reg_write(context, QPC_RR_MAX,
5088 			     fls(attr->max_dest_rd_atomic - 1));
5089 		hr_reg_clear(qpc_mask, QPC_RR_MAX);
5090 	}
5091 
5092 	if ((attr_mask & IB_QP_MAX_QP_RD_ATOMIC) && attr->max_rd_atomic) {
5093 		hr_reg_write(context, QPC_SR_MAX, fls(attr->max_rd_atomic - 1));
5094 		hr_reg_clear(qpc_mask, QPC_SR_MAX);
5095 	}
5096 
5097 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
5098 		set_access_flags(hr_qp, context, qpc_mask, attr, attr_mask);
5099 
5100 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
5101 		hr_reg_write(context, QPC_MIN_RNR_TIME,
5102 			    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ?
5103 			    HNS_ROCE_RNR_TIMER_10NS : attr->min_rnr_timer);
5104 		hr_reg_clear(qpc_mask, QPC_MIN_RNR_TIME);
5105 	}
5106 
5107 	if (attr_mask & IB_QP_RQ_PSN) {
5108 		hr_reg_write(context, QPC_RX_REQ_EPSN, attr->rq_psn);
5109 		hr_reg_clear(qpc_mask, QPC_RX_REQ_EPSN);
5110 
5111 		hr_reg_write(context, QPC_RAQ_PSN, attr->rq_psn - 1);
5112 		hr_reg_clear(qpc_mask, QPC_RAQ_PSN);
5113 	}
5114 
5115 	if (attr_mask & IB_QP_QKEY) {
5116 		context->qkey_xrcd = cpu_to_le32(attr->qkey);
5117 		qpc_mask->qkey_xrcd = 0;
5118 		hr_qp->qkey = attr->qkey;
5119 	}
5120 
5121 	return ret;
5122 }
5123 
5124 static void hns_roce_v2_record_opt_fields(struct ib_qp *ibqp,
5125 					  const struct ib_qp_attr *attr,
5126 					  int attr_mask)
5127 {
5128 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5129 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5130 
5131 	if (attr_mask & IB_QP_ACCESS_FLAGS)
5132 		hr_qp->atomic_rd_en = attr->qp_access_flags;
5133 
5134 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
5135 		hr_qp->resp_depth = attr->max_dest_rd_atomic;
5136 	if (attr_mask & IB_QP_PORT) {
5137 		hr_qp->port = attr->port_num - 1;
5138 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
5139 	}
5140 }
5141 
5142 static void clear_qp(struct hns_roce_qp *hr_qp)
5143 {
5144 	struct ib_qp *ibqp = &hr_qp->ibqp;
5145 
5146 	if (ibqp->send_cq)
5147 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq),
5148 				     hr_qp->qpn, NULL);
5149 
5150 	if (ibqp->recv_cq  && ibqp->recv_cq != ibqp->send_cq)
5151 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->recv_cq),
5152 				     hr_qp->qpn, ibqp->srq ?
5153 				     to_hr_srq(ibqp->srq) : NULL);
5154 
5155 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
5156 		*hr_qp->rdb.db_record = 0;
5157 
5158 	hr_qp->rq.head = 0;
5159 	hr_qp->rq.tail = 0;
5160 	hr_qp->sq.head = 0;
5161 	hr_qp->sq.tail = 0;
5162 	hr_qp->next_sge = 0;
5163 }
5164 
5165 static void v2_set_flushed_fields(struct ib_qp *ibqp,
5166 				  struct hns_roce_v2_qp_context *context,
5167 				  struct hns_roce_v2_qp_context *qpc_mask)
5168 {
5169 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5170 	unsigned long sq_flag = 0;
5171 	unsigned long rq_flag = 0;
5172 
5173 	if (ibqp->qp_type == IB_QPT_XRC_TGT)
5174 		return;
5175 
5176 	spin_lock_irqsave(&hr_qp->sq.lock, sq_flag);
5177 	hr_reg_write(context, QPC_SQ_PRODUCER_IDX, hr_qp->sq.head);
5178 	hr_reg_clear(qpc_mask, QPC_SQ_PRODUCER_IDX);
5179 	hr_qp->state = IB_QPS_ERR;
5180 	spin_unlock_irqrestore(&hr_qp->sq.lock, sq_flag);
5181 
5182 	if (ibqp->srq || ibqp->qp_type == IB_QPT_XRC_INI) /* no RQ */
5183 		return;
5184 
5185 	spin_lock_irqsave(&hr_qp->rq.lock, rq_flag);
5186 	hr_reg_write(context, QPC_RQ_PRODUCER_IDX, hr_qp->rq.head);
5187 	hr_reg_clear(qpc_mask, QPC_RQ_PRODUCER_IDX);
5188 	spin_unlock_irqrestore(&hr_qp->rq.lock, rq_flag);
5189 }
5190 
5191 static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
5192 				 const struct ib_qp_attr *attr,
5193 				 int attr_mask, enum ib_qp_state cur_state,
5194 				 enum ib_qp_state new_state)
5195 {
5196 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5197 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5198 	struct hns_roce_v2_qp_context ctx[2];
5199 	struct hns_roce_v2_qp_context *context = ctx;
5200 	struct hns_roce_v2_qp_context *qpc_mask = ctx + 1;
5201 	struct ib_device *ibdev = &hr_dev->ib_dev;
5202 	int ret;
5203 
5204 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
5205 		return -EOPNOTSUPP;
5206 
5207 	/*
5208 	 * In v2 engine, software pass context and context mask to hardware
5209 	 * when modifying qp. If software need modify some fields in context,
5210 	 * we should set all bits of the relevant fields in context mask to
5211 	 * 0 at the same time, else set them to 0x1.
5212 	 */
5213 	memset(context, 0, hr_dev->caps.qpc_sz);
5214 	memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
5215 
5216 	ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
5217 					 new_state, context, qpc_mask);
5218 	if (ret)
5219 		goto out;
5220 
5221 	/* When QP state is err, SQ and RQ WQE should be flushed */
5222 	if (new_state == IB_QPS_ERR)
5223 		v2_set_flushed_fields(ibqp, context, qpc_mask);
5224 
5225 	/* Configure the optional fields */
5226 	ret = hns_roce_v2_set_opt_fields(ibqp, attr, attr_mask, context,
5227 					 qpc_mask);
5228 	if (ret)
5229 		goto out;
5230 
5231 	hr_reg_write_bool(context, QPC_INV_CREDIT,
5232 			  to_hr_qp_type(hr_qp->ibqp.qp_type) == SERV_TYPE_XRC ||
5233 			  ibqp->srq);
5234 	hr_reg_clear(qpc_mask, QPC_INV_CREDIT);
5235 
5236 	/* Every status migrate must change state */
5237 	hr_reg_write(context, QPC_QP_ST, new_state);
5238 	hr_reg_clear(qpc_mask, QPC_QP_ST);
5239 
5240 	/* SW pass context to HW */
5241 	ret = hns_roce_v2_qp_modify(hr_dev, context, qpc_mask, hr_qp);
5242 	if (ret) {
5243 		ibdev_err(ibdev, "failed to modify QP, ret = %d.\n", ret);
5244 		goto out;
5245 	}
5246 
5247 	hr_qp->state = new_state;
5248 
5249 	hns_roce_v2_record_opt_fields(ibqp, attr, attr_mask);
5250 
5251 	if (new_state == IB_QPS_RESET && !ibqp->uobject)
5252 		clear_qp(hr_qp);
5253 
5254 out:
5255 	return ret;
5256 }
5257 
5258 static int to_ib_qp_st(enum hns_roce_v2_qp_state state)
5259 {
5260 	static const enum ib_qp_state map[] = {
5261 		[HNS_ROCE_QP_ST_RST] = IB_QPS_RESET,
5262 		[HNS_ROCE_QP_ST_INIT] = IB_QPS_INIT,
5263 		[HNS_ROCE_QP_ST_RTR] = IB_QPS_RTR,
5264 		[HNS_ROCE_QP_ST_RTS] = IB_QPS_RTS,
5265 		[HNS_ROCE_QP_ST_SQD] = IB_QPS_SQD,
5266 		[HNS_ROCE_QP_ST_SQER] = IB_QPS_SQE,
5267 		[HNS_ROCE_QP_ST_ERR] = IB_QPS_ERR,
5268 		[HNS_ROCE_QP_ST_SQ_DRAINING] = IB_QPS_SQD
5269 	};
5270 
5271 	return (state < ARRAY_SIZE(map)) ? map[state] : -1;
5272 }
5273 
5274 static int hns_roce_v2_query_qpc(struct hns_roce_dev *hr_dev,
5275 				 struct hns_roce_qp *hr_qp,
5276 				 struct hns_roce_v2_qp_context *hr_context)
5277 {
5278 	struct hns_roce_cmd_mailbox *mailbox;
5279 	int ret;
5280 
5281 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5282 	if (IS_ERR(mailbox))
5283 		return PTR_ERR(mailbox);
5284 
5285 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_QPC,
5286 				hr_qp->qpn);
5287 	if (ret)
5288 		goto out;
5289 
5290 	memcpy(hr_context, mailbox->buf, hr_dev->caps.qpc_sz);
5291 
5292 out:
5293 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5294 	return ret;
5295 }
5296 
5297 static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5298 				int qp_attr_mask,
5299 				struct ib_qp_init_attr *qp_init_attr)
5300 {
5301 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5302 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5303 	struct hns_roce_v2_qp_context context = {};
5304 	struct ib_device *ibdev = &hr_dev->ib_dev;
5305 	int tmp_qp_state;
5306 	int state;
5307 	int ret;
5308 
5309 	memset(qp_attr, 0, sizeof(*qp_attr));
5310 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5311 
5312 	mutex_lock(&hr_qp->mutex);
5313 
5314 	if (hr_qp->state == IB_QPS_RESET) {
5315 		qp_attr->qp_state = IB_QPS_RESET;
5316 		ret = 0;
5317 		goto done;
5318 	}
5319 
5320 	ret = hns_roce_v2_query_qpc(hr_dev, hr_qp, &context);
5321 	if (ret) {
5322 		ibdev_err(ibdev, "failed to query QPC, ret = %d.\n", ret);
5323 		ret = -EINVAL;
5324 		goto out;
5325 	}
5326 
5327 	state = hr_reg_read(&context, QPC_QP_ST);
5328 	tmp_qp_state = to_ib_qp_st((enum hns_roce_v2_qp_state)state);
5329 	if (tmp_qp_state == -1) {
5330 		ibdev_err(ibdev, "Illegal ib_qp_state\n");
5331 		ret = -EINVAL;
5332 		goto out;
5333 	}
5334 	hr_qp->state = (u8)tmp_qp_state;
5335 	qp_attr->qp_state = (enum ib_qp_state)hr_qp->state;
5336 	qp_attr->path_mtu = (enum ib_mtu)hr_reg_read(&context, QPC_MTU);
5337 	qp_attr->path_mig_state = IB_MIG_ARMED;
5338 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
5339 	if (hr_qp->ibqp.qp_type == IB_QPT_UD)
5340 		qp_attr->qkey = le32_to_cpu(context.qkey_xrcd);
5341 
5342 	qp_attr->rq_psn = hr_reg_read(&context, QPC_RX_REQ_EPSN);
5343 	qp_attr->sq_psn = (u32)hr_reg_read(&context, QPC_SQ_CUR_PSN);
5344 	qp_attr->dest_qp_num = hr_reg_read(&context, QPC_DQPN);
5345 	qp_attr->qp_access_flags =
5346 		((hr_reg_read(&context, QPC_RRE)) << V2_QP_RRE_S) |
5347 		((hr_reg_read(&context, QPC_RWE)) << V2_QP_RWE_S) |
5348 		((hr_reg_read(&context, QPC_ATE)) << V2_QP_ATE_S);
5349 
5350 	if (hr_qp->ibqp.qp_type == IB_QPT_RC ||
5351 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5352 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
5353 		struct ib_global_route *grh =
5354 			rdma_ah_retrieve_grh(&qp_attr->ah_attr);
5355 
5356 		rdma_ah_set_sl(&qp_attr->ah_attr,
5357 			       hr_reg_read(&context, QPC_SL));
5358 		grh->flow_label = hr_reg_read(&context, QPC_FL);
5359 		grh->sgid_index = hr_reg_read(&context, QPC_GMV_IDX);
5360 		grh->hop_limit = hr_reg_read(&context, QPC_HOPLIMIT);
5361 		grh->traffic_class = hr_reg_read(&context, QPC_TC);
5362 
5363 		memcpy(grh->dgid.raw, context.dgid, sizeof(grh->dgid.raw));
5364 	}
5365 
5366 	qp_attr->port_num = hr_qp->port + 1;
5367 	qp_attr->sq_draining = 0;
5368 	qp_attr->max_rd_atomic = 1 << hr_reg_read(&context, QPC_SR_MAX);
5369 	qp_attr->max_dest_rd_atomic = 1 << hr_reg_read(&context, QPC_RR_MAX);
5370 
5371 	qp_attr->min_rnr_timer = (u8)hr_reg_read(&context, QPC_MIN_RNR_TIME);
5372 	qp_attr->timeout = (u8)hr_reg_read(&context, QPC_AT);
5373 	qp_attr->retry_cnt = hr_reg_read(&context, QPC_RETRY_NUM_INIT);
5374 	qp_attr->rnr_retry = hr_reg_read(&context, QPC_RNR_NUM_INIT);
5375 
5376 done:
5377 	qp_attr->cur_qp_state = qp_attr->qp_state;
5378 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
5379 	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
5380 	qp_attr->cap.max_inline_data = hr_qp->max_inline_data;
5381 
5382 	qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
5383 	qp_attr->cap.max_send_sge = hr_qp->sq.max_gs;
5384 
5385 	qp_init_attr->qp_context = ibqp->qp_context;
5386 	qp_init_attr->qp_type = ibqp->qp_type;
5387 	qp_init_attr->recv_cq = ibqp->recv_cq;
5388 	qp_init_attr->send_cq = ibqp->send_cq;
5389 	qp_init_attr->srq = ibqp->srq;
5390 	qp_init_attr->cap = qp_attr->cap;
5391 	qp_init_attr->sq_sig_type = hr_qp->sq_signal_bits;
5392 
5393 out:
5394 	mutex_unlock(&hr_qp->mutex);
5395 	return ret;
5396 }
5397 
5398 static inline int modify_qp_is_ok(struct hns_roce_qp *hr_qp)
5399 {
5400 	return ((hr_qp->ibqp.qp_type == IB_QPT_RC ||
5401 		 hr_qp->ibqp.qp_type == IB_QPT_UD ||
5402 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5403 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
5404 		hr_qp->state != IB_QPS_RESET);
5405 }
5406 
5407 static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
5408 					 struct hns_roce_qp *hr_qp,
5409 					 struct ib_udata *udata)
5410 {
5411 	struct ib_device *ibdev = &hr_dev->ib_dev;
5412 	struct hns_roce_cq *send_cq, *recv_cq;
5413 	unsigned long flags;
5414 	int ret = 0;
5415 
5416 	if (modify_qp_is_ok(hr_qp)) {
5417 		/* Modify qp to reset before destroying qp */
5418 		ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
5419 					    hr_qp->state, IB_QPS_RESET);
5420 		if (ret)
5421 			ibdev_err(ibdev,
5422 				  "failed to modify QP to RST, ret = %d.\n",
5423 				  ret);
5424 	}
5425 
5426 	send_cq = hr_qp->ibqp.send_cq ? to_hr_cq(hr_qp->ibqp.send_cq) : NULL;
5427 	recv_cq = hr_qp->ibqp.recv_cq ? to_hr_cq(hr_qp->ibqp.recv_cq) : NULL;
5428 
5429 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
5430 	hns_roce_lock_cqs(send_cq, recv_cq);
5431 
5432 	if (!udata) {
5433 		if (recv_cq)
5434 			__hns_roce_v2_cq_clean(recv_cq, hr_qp->qpn,
5435 					       (hr_qp->ibqp.srq ?
5436 						to_hr_srq(hr_qp->ibqp.srq) :
5437 						NULL));
5438 
5439 		if (send_cq && send_cq != recv_cq)
5440 			__hns_roce_v2_cq_clean(send_cq, hr_qp->qpn, NULL);
5441 	}
5442 
5443 	hns_roce_qp_remove(hr_dev, hr_qp);
5444 
5445 	hns_roce_unlock_cqs(send_cq, recv_cq);
5446 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
5447 
5448 	return ret;
5449 }
5450 
5451 static int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
5452 {
5453 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5454 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5455 	int ret;
5456 
5457 	ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata);
5458 	if (ret)
5459 		ibdev_err(&hr_dev->ib_dev,
5460 			  "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n",
5461 			  hr_qp->qpn, ret);
5462 
5463 	hns_roce_qp_destroy(hr_dev, hr_qp, udata);
5464 
5465 	return 0;
5466 }
5467 
5468 static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
5469 					    struct hns_roce_qp *hr_qp)
5470 {
5471 	struct ib_device *ibdev = &hr_dev->ib_dev;
5472 	struct hns_roce_sccc_clr_done *resp;
5473 	struct hns_roce_sccc_clr *clr;
5474 	struct hns_roce_cmq_desc desc;
5475 	int ret, i;
5476 
5477 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
5478 		return 0;
5479 
5480 	mutex_lock(&hr_dev->qp_table.scc_mutex);
5481 
5482 	/* set scc ctx clear done flag */
5483 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_RESET_SCCC, false);
5484 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5485 	if (ret) {
5486 		ibdev_err(ibdev, "failed to reset SCC ctx, ret = %d.\n", ret);
5487 		goto out;
5488 	}
5489 
5490 	/* clear scc context */
5491 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLR_SCCC, false);
5492 	clr = (struct hns_roce_sccc_clr *)desc.data;
5493 	clr->qpn = cpu_to_le32(hr_qp->qpn);
5494 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5495 	if (ret) {
5496 		ibdev_err(ibdev, "failed to clear SCC ctx, ret = %d.\n", ret);
5497 		goto out;
5498 	}
5499 
5500 	/* query scc context clear is done or not */
5501 	resp = (struct hns_roce_sccc_clr_done *)desc.data;
5502 	for (i = 0; i <= HNS_ROCE_CMQ_SCC_CLR_DONE_CNT; i++) {
5503 		hns_roce_cmq_setup_basic_desc(&desc,
5504 					      HNS_ROCE_OPC_QUERY_SCCC, true);
5505 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
5506 		if (ret) {
5507 			ibdev_err(ibdev, "failed to query clr cmq, ret = %d\n",
5508 				  ret);
5509 			goto out;
5510 		}
5511 
5512 		if (resp->clr_done)
5513 			goto out;
5514 
5515 		msleep(20);
5516 	}
5517 
5518 	ibdev_err(ibdev, "Query SCC clr done flag overtime.\n");
5519 	ret = -ETIMEDOUT;
5520 
5521 out:
5522 	mutex_unlock(&hr_dev->qp_table.scc_mutex);
5523 	return ret;
5524 }
5525 
5526 #define DMA_IDX_SHIFT 3
5527 #define DMA_WQE_SHIFT 3
5528 
5529 static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
5530 					      struct hns_roce_srq_context *ctx)
5531 {
5532 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
5533 	struct ib_device *ibdev = srq->ibsrq.device;
5534 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5535 	u64 mtts_idx[MTT_MIN_COUNT] = {};
5536 	dma_addr_t dma_handle_idx = 0;
5537 	int ret;
5538 
5539 	/* Get physical address of idx que buf */
5540 	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
5541 				ARRAY_SIZE(mtts_idx), &dma_handle_idx);
5542 	if (ret < 1) {
5543 		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
5544 			  ret);
5545 		return -ENOBUFS;
5546 	}
5547 
5548 	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
5549 		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
5550 
5551 	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> DMA_IDX_SHIFT);
5552 	hr_reg_write(ctx, SRQC_IDX_BT_BA_H,
5553 		     upper_32_bits(dma_handle_idx >> DMA_IDX_SHIFT));
5554 
5555 	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
5556 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
5557 	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
5558 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
5559 
5560 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
5561 		     to_hr_hw_page_addr(mtts_idx[0]));
5562 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
5563 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
5564 
5565 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
5566 		     to_hr_hw_page_addr(mtts_idx[1]));
5567 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
5568 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
5569 
5570 	return 0;
5571 }
5572 
5573 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
5574 {
5575 	struct ib_device *ibdev = srq->ibsrq.device;
5576 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5577 	struct hns_roce_srq_context *ctx = mb_buf;
5578 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
5579 	dma_addr_t dma_handle_wqe = 0;
5580 	int ret;
5581 
5582 	memset(ctx, 0, sizeof(*ctx));
5583 
5584 	/* Get the physical address of srq buf */
5585 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
5586 				ARRAY_SIZE(mtts_wqe), &dma_handle_wqe);
5587 	if (ret < 1) {
5588 		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
5589 			  ret);
5590 		return -ENOBUFS;
5591 	}
5592 
5593 	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
5594 	hr_reg_write_bool(ctx, SRQC_SRQ_TYPE,
5595 			  srq->ibsrq.srq_type == IB_SRQT_XRC);
5596 	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
5597 	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
5598 	hr_reg_write(ctx, SRQC_XRCD, srq->xrcdn);
5599 	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
5600 	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
5601 	hr_reg_write(ctx, SRQC_RQWS,
5602 		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
5603 
5604 	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
5605 		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
5606 				      srq->wqe_cnt));
5607 
5608 	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> DMA_WQE_SHIFT);
5609 	hr_reg_write(ctx, SRQC_WQE_BT_BA_H,
5610 		     upper_32_bits(dma_handle_wqe >> DMA_WQE_SHIFT));
5611 
5612 	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
5613 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
5614 	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
5615 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
5616 
5617 	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
5618 }
5619 
5620 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
5621 				  struct ib_srq_attr *srq_attr,
5622 				  enum ib_srq_attr_mask srq_attr_mask,
5623 				  struct ib_udata *udata)
5624 {
5625 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5626 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5627 	struct hns_roce_srq_context *srq_context;
5628 	struct hns_roce_srq_context *srqc_mask;
5629 	struct hns_roce_cmd_mailbox *mailbox;
5630 	int ret;
5631 
5632 	/* Resizing SRQs is not supported yet */
5633 	if (srq_attr_mask & IB_SRQ_MAX_WR)
5634 		return -EINVAL;
5635 
5636 	if (srq_attr_mask & IB_SRQ_LIMIT) {
5637 		if (srq_attr->srq_limit > srq->wqe_cnt)
5638 			return -EINVAL;
5639 
5640 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5641 		if (IS_ERR(mailbox))
5642 			return PTR_ERR(mailbox);
5643 
5644 		srq_context = mailbox->buf;
5645 		srqc_mask = (struct hns_roce_srq_context *)mailbox->buf + 1;
5646 
5647 		memset(srqc_mask, 0xff, sizeof(*srqc_mask));
5648 
5649 		hr_reg_write(srq_context, SRQC_LIMIT_WL, srq_attr->srq_limit);
5650 		hr_reg_clear(srqc_mask, SRQC_LIMIT_WL);
5651 
5652 		ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5653 					HNS_ROCE_CMD_MODIFY_SRQC, srq->srqn);
5654 		hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5655 		if (ret) {
5656 			ibdev_err(&hr_dev->ib_dev,
5657 				  "failed to handle cmd of modifying SRQ, ret = %d.\n",
5658 				  ret);
5659 			return ret;
5660 		}
5661 	}
5662 
5663 	return 0;
5664 }
5665 
5666 static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
5667 {
5668 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5669 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5670 	struct hns_roce_srq_context *srq_context;
5671 	struct hns_roce_cmd_mailbox *mailbox;
5672 	int ret;
5673 
5674 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5675 	if (IS_ERR(mailbox))
5676 		return PTR_ERR(mailbox);
5677 
5678 	srq_context = mailbox->buf;
5679 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5680 				HNS_ROCE_CMD_QUERY_SRQC, srq->srqn);
5681 	if (ret) {
5682 		ibdev_err(&hr_dev->ib_dev,
5683 			  "failed to process cmd of querying SRQ, ret = %d.\n",
5684 			  ret);
5685 		goto out;
5686 	}
5687 
5688 	attr->srq_limit = hr_reg_read(srq_context, SRQC_LIMIT_WL);
5689 	attr->max_wr = srq->wqe_cnt;
5690 	attr->max_sge = srq->max_gs - srq->rsv_sge;
5691 
5692 out:
5693 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5694 	return ret;
5695 }
5696 
5697 static int hns_roce_v2_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
5698 {
5699 	struct hns_roce_dev *hr_dev = to_hr_dev(cq->device);
5700 	struct hns_roce_v2_cq_context *cq_context;
5701 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
5702 	struct hns_roce_v2_cq_context *cqc_mask;
5703 	struct hns_roce_cmd_mailbox *mailbox;
5704 	int ret;
5705 
5706 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5707 	if (IS_ERR(mailbox))
5708 		return PTR_ERR(mailbox);
5709 
5710 	cq_context = mailbox->buf;
5711 	cqc_mask = (struct hns_roce_v2_cq_context *)mailbox->buf + 1;
5712 
5713 	memset(cqc_mask, 0xff, sizeof(*cqc_mask));
5714 
5715 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT, cq_count);
5716 	hr_reg_clear(cqc_mask, CQC_CQ_MAX_CNT);
5717 
5718 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5719 		if (cq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
5720 			dev_info(hr_dev->dev,
5721 				 "cq_period(%u) reached the upper limit, adjusted to 65.\n",
5722 				 cq_period);
5723 			cq_period = HNS_ROCE_MAX_CQ_PERIOD;
5724 		}
5725 		cq_period *= HNS_ROCE_CLOCK_ADJUST;
5726 	}
5727 	hr_reg_write(cq_context, CQC_CQ_PERIOD, cq_period);
5728 	hr_reg_clear(cqc_mask, CQC_CQ_PERIOD);
5729 
5730 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5731 				HNS_ROCE_CMD_MODIFY_CQC, hr_cq->cqn);
5732 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5733 	if (ret)
5734 		ibdev_err(&hr_dev->ib_dev,
5735 			  "failed to process cmd when modifying CQ, ret = %d.\n",
5736 			  ret);
5737 
5738 	return ret;
5739 }
5740 
5741 static void hns_roce_irq_work_handle(struct work_struct *work)
5742 {
5743 	struct hns_roce_work *irq_work =
5744 				container_of(work, struct hns_roce_work, work);
5745 	struct ib_device *ibdev = &irq_work->hr_dev->ib_dev;
5746 
5747 	switch (irq_work->event_type) {
5748 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5749 		ibdev_info(ibdev, "Path migrated succeeded.\n");
5750 		break;
5751 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5752 		ibdev_warn(ibdev, "Path migration failed.\n");
5753 		break;
5754 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
5755 		break;
5756 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5757 		ibdev_warn(ibdev, "Send queue drained.\n");
5758 		break;
5759 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5760 		ibdev_err(ibdev, "Local work queue 0x%x catast error, sub_event type is: %d\n",
5761 			  irq_work->queue_num, irq_work->sub_type);
5762 		break;
5763 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5764 		ibdev_err(ibdev, "Invalid request local work queue 0x%x error.\n",
5765 			  irq_work->queue_num);
5766 		break;
5767 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5768 		ibdev_err(ibdev, "Local access violation work queue 0x%x error, sub_event type is: %d\n",
5769 			  irq_work->queue_num, irq_work->sub_type);
5770 		break;
5771 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
5772 		ibdev_warn(ibdev, "SRQ limit reach.\n");
5773 		break;
5774 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
5775 		ibdev_warn(ibdev, "SRQ last wqe reach.\n");
5776 		break;
5777 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
5778 		ibdev_err(ibdev, "SRQ catas error.\n");
5779 		break;
5780 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
5781 		ibdev_err(ibdev, "CQ 0x%x access err.\n", irq_work->queue_num);
5782 		break;
5783 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
5784 		ibdev_warn(ibdev, "CQ 0x%x overflow\n", irq_work->queue_num);
5785 		break;
5786 	case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
5787 		ibdev_warn(ibdev, "DB overflow.\n");
5788 		break;
5789 	case HNS_ROCE_EVENT_TYPE_FLR:
5790 		ibdev_warn(ibdev, "Function level reset.\n");
5791 		break;
5792 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
5793 		ibdev_err(ibdev, "xrc domain violation error.\n");
5794 		break;
5795 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
5796 		ibdev_err(ibdev, "invalid xrceth error.\n");
5797 		break;
5798 	default:
5799 		break;
5800 	}
5801 
5802 	kfree(irq_work);
5803 }
5804 
5805 static void hns_roce_v2_init_irq_work(struct hns_roce_dev *hr_dev,
5806 				      struct hns_roce_eq *eq, u32 queue_num)
5807 {
5808 	struct hns_roce_work *irq_work;
5809 
5810 	irq_work = kzalloc(sizeof(struct hns_roce_work), GFP_ATOMIC);
5811 	if (!irq_work)
5812 		return;
5813 
5814 	INIT_WORK(&(irq_work->work), hns_roce_irq_work_handle);
5815 	irq_work->hr_dev = hr_dev;
5816 	irq_work->event_type = eq->event_type;
5817 	irq_work->sub_type = eq->sub_type;
5818 	irq_work->queue_num = queue_num;
5819 	queue_work(hr_dev->irq_workq, &(irq_work->work));
5820 }
5821 
5822 static void update_eq_db(struct hns_roce_eq *eq)
5823 {
5824 	struct hns_roce_dev *hr_dev = eq->hr_dev;
5825 	struct hns_roce_v2_db eq_db = {};
5826 
5827 	if (eq->type_flag == HNS_ROCE_AEQ) {
5828 		hr_reg_write(&eq_db, EQ_DB_CMD,
5829 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
5830 			     HNS_ROCE_EQ_DB_CMD_AEQ :
5831 			     HNS_ROCE_EQ_DB_CMD_AEQ_ARMED);
5832 	} else {
5833 		hr_reg_write(&eq_db, EQ_DB_TAG, eq->eqn);
5834 
5835 		hr_reg_write(&eq_db, EQ_DB_CMD,
5836 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
5837 			     HNS_ROCE_EQ_DB_CMD_CEQ :
5838 			     HNS_ROCE_EQ_DB_CMD_CEQ_ARMED);
5839 	}
5840 
5841 	hr_reg_write(&eq_db, EQ_DB_CI, eq->cons_index);
5842 
5843 	hns_roce_write64(hr_dev, (__le32 *)&eq_db, eq->db_reg);
5844 }
5845 
5846 static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
5847 {
5848 	struct hns_roce_aeqe *aeqe;
5849 
5850 	aeqe = hns_roce_buf_offset(eq->mtr.kmem,
5851 				   (eq->cons_index & (eq->entries - 1)) *
5852 				   eq->eqe_size);
5853 
5854 	return (hr_reg_read(aeqe, AEQE_OWNER) ^
5855 		!!(eq->cons_index & eq->entries)) ? aeqe : NULL;
5856 }
5857 
5858 static int hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
5859 			       struct hns_roce_eq *eq)
5860 {
5861 	struct device *dev = hr_dev->dev;
5862 	struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
5863 	int aeqe_found = 0;
5864 	int event_type;
5865 	u32 queue_num;
5866 	int sub_type;
5867 
5868 	while (aeqe) {
5869 		/* Make sure we read AEQ entry after we have checked the
5870 		 * ownership bit
5871 		 */
5872 		dma_rmb();
5873 
5874 		event_type = hr_reg_read(aeqe, AEQE_EVENT_TYPE);
5875 		sub_type = hr_reg_read(aeqe, AEQE_SUB_TYPE);
5876 		queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
5877 
5878 		switch (event_type) {
5879 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5880 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5881 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
5882 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5883 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5884 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
5885 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5886 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5887 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
5888 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
5889 			hns_roce_qp_event(hr_dev, queue_num, event_type);
5890 			break;
5891 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
5892 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
5893 			hns_roce_srq_event(hr_dev, queue_num, event_type);
5894 			break;
5895 		case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
5896 		case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
5897 			hns_roce_cq_event(hr_dev, queue_num, event_type);
5898 			break;
5899 		case HNS_ROCE_EVENT_TYPE_MB:
5900 			hns_roce_cmd_event(hr_dev,
5901 					le16_to_cpu(aeqe->event.cmd.token),
5902 					aeqe->event.cmd.status,
5903 					le64_to_cpu(aeqe->event.cmd.out_param));
5904 			break;
5905 		case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
5906 		case HNS_ROCE_EVENT_TYPE_FLR:
5907 			break;
5908 		default:
5909 			dev_err(dev, "Unhandled event %d on EQ %d at idx %u.\n",
5910 				event_type, eq->eqn, eq->cons_index);
5911 			break;
5912 		}
5913 
5914 		eq->event_type = event_type;
5915 		eq->sub_type = sub_type;
5916 		++eq->cons_index;
5917 		aeqe_found = 1;
5918 
5919 		hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
5920 
5921 		aeqe = next_aeqe_sw_v2(eq);
5922 	}
5923 
5924 	update_eq_db(eq);
5925 	return aeqe_found;
5926 }
5927 
5928 static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq)
5929 {
5930 	struct hns_roce_ceqe *ceqe;
5931 
5932 	ceqe = hns_roce_buf_offset(eq->mtr.kmem,
5933 				   (eq->cons_index & (eq->entries - 1)) *
5934 				   eq->eqe_size);
5935 
5936 	return (hr_reg_read(ceqe, CEQE_OWNER) ^
5937 		!!(eq->cons_index & eq->entries)) ? ceqe : NULL;
5938 }
5939 
5940 static int hns_roce_v2_ceq_int(struct hns_roce_dev *hr_dev,
5941 			       struct hns_roce_eq *eq)
5942 {
5943 	struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq);
5944 	int ceqe_found = 0;
5945 	u32 cqn;
5946 
5947 	while (ceqe) {
5948 		/* Make sure we read CEQ entry after we have checked the
5949 		 * ownership bit
5950 		 */
5951 		dma_rmb();
5952 
5953 		cqn = hr_reg_read(ceqe, CEQE_CQN);
5954 
5955 		hns_roce_cq_completion(hr_dev, cqn);
5956 
5957 		++eq->cons_index;
5958 		ceqe_found = 1;
5959 
5960 		ceqe = next_ceqe_sw_v2(eq);
5961 	}
5962 
5963 	update_eq_db(eq);
5964 
5965 	return ceqe_found;
5966 }
5967 
5968 static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
5969 {
5970 	struct hns_roce_eq *eq = eq_ptr;
5971 	struct hns_roce_dev *hr_dev = eq->hr_dev;
5972 	int int_work;
5973 
5974 	if (eq->type_flag == HNS_ROCE_CEQ)
5975 		/* Completion event interrupt */
5976 		int_work = hns_roce_v2_ceq_int(hr_dev, eq);
5977 	else
5978 		/* Asychronous event interrupt */
5979 		int_work = hns_roce_v2_aeq_int(hr_dev, eq);
5980 
5981 	return IRQ_RETVAL(int_work);
5982 }
5983 
5984 static irqreturn_t hns_roce_v2_msix_interrupt_abn(int irq, void *dev_id)
5985 {
5986 	struct hns_roce_dev *hr_dev = dev_id;
5987 	struct device *dev = hr_dev->dev;
5988 	int int_work = 0;
5989 	u32 int_st;
5990 	u32 int_en;
5991 
5992 	/* Abnormal interrupt */
5993 	int_st = roce_read(hr_dev, ROCEE_VF_ABN_INT_ST_REG);
5994 	int_en = roce_read(hr_dev, ROCEE_VF_ABN_INT_EN_REG);
5995 
5996 	if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S)) {
5997 		struct pci_dev *pdev = hr_dev->pci_dev;
5998 		struct hnae3_ae_dev *ae_dev = pci_get_drvdata(pdev);
5999 		const struct hnae3_ae_ops *ops = ae_dev->ops;
6000 
6001 		dev_err(dev, "AEQ overflow!\n");
6002 
6003 		int_st |= 1 << HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S;
6004 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG, int_st);
6005 
6006 		/* Set reset level for reset_event() */
6007 		if (ops->set_default_reset_request)
6008 			ops->set_default_reset_request(ae_dev,
6009 						       HNAE3_FUNC_RESET);
6010 		if (ops->reset_event)
6011 			ops->reset_event(pdev, NULL);
6012 
6013 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6014 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6015 
6016 		int_work = 1;
6017 	} else if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_RAS_INT_S)) {
6018 		dev_err(dev, "RAS interrupt!\n");
6019 
6020 		int_st |= 1 << HNS_ROCE_V2_VF_INT_ST_RAS_INT_S;
6021 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG, int_st);
6022 
6023 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6024 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6025 
6026 		int_work = 1;
6027 	} else {
6028 		dev_err(dev, "There is no abnormal irq found!\n");
6029 	}
6030 
6031 	return IRQ_RETVAL(int_work);
6032 }
6033 
6034 static void hns_roce_v2_int_mask_enable(struct hns_roce_dev *hr_dev,
6035 					int eq_num, u32 enable_flag)
6036 {
6037 	int i;
6038 
6039 	for (i = 0; i < eq_num; i++)
6040 		roce_write(hr_dev, ROCEE_VF_EVENT_INT_EN_REG +
6041 			   i * EQ_REG_OFFSET, enable_flag);
6042 
6043 	roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, enable_flag);
6044 	roce_write(hr_dev, ROCEE_VF_ABN_INT_CFG_REG, enable_flag);
6045 }
6046 
6047 static void hns_roce_v2_destroy_eqc(struct hns_roce_dev *hr_dev, u32 eqn)
6048 {
6049 	struct device *dev = hr_dev->dev;
6050 	int ret;
6051 	u8 cmd;
6052 
6053 	if (eqn < hr_dev->caps.num_comp_vectors)
6054 		cmd = HNS_ROCE_CMD_DESTROY_CEQC;
6055 	else
6056 		cmd = HNS_ROCE_CMD_DESTROY_AEQC;
6057 
6058 	ret = hns_roce_destroy_hw_ctx(hr_dev, cmd, eqn & HNS_ROCE_V2_EQN_M);
6059 	if (ret)
6060 		dev_err(dev, "[mailbox cmd] destroy eqc(%u) failed.\n", eqn);
6061 }
6062 
6063 static void free_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6064 {
6065 	hns_roce_mtr_destroy(hr_dev, &eq->mtr);
6066 }
6067 
6068 static void init_eq_config(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6069 {
6070 	eq->db_reg = hr_dev->reg_base + ROCEE_VF_EQ_DB_CFG0_REG;
6071 	eq->cons_index = 0;
6072 	eq->over_ignore = HNS_ROCE_V2_EQ_OVER_IGNORE_0;
6073 	eq->coalesce = HNS_ROCE_V2_EQ_COALESCE_0;
6074 	eq->arm_st = HNS_ROCE_V2_EQ_ALWAYS_ARMED;
6075 	eq->shift = ilog2((unsigned int)eq->entries);
6076 }
6077 
6078 static int config_eqc(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq,
6079 		      void *mb_buf)
6080 {
6081 	u64 eqe_ba[MTT_MIN_COUNT] = { 0 };
6082 	struct hns_roce_eq_context *eqc;
6083 	u64 bt_ba = 0;
6084 	int count;
6085 
6086 	eqc = mb_buf;
6087 	memset(eqc, 0, sizeof(struct hns_roce_eq_context));
6088 
6089 	init_eq_config(hr_dev, eq);
6090 
6091 	/* if not multi-hop, eqe buffer only use one trunk */
6092 	count = hns_roce_mtr_find(hr_dev, &eq->mtr, 0, eqe_ba, MTT_MIN_COUNT,
6093 				  &bt_ba);
6094 	if (count < 1) {
6095 		dev_err(hr_dev->dev, "failed to find EQE mtr\n");
6096 		return -ENOBUFS;
6097 	}
6098 
6099 	hr_reg_write(eqc, EQC_EQ_ST, HNS_ROCE_V2_EQ_STATE_VALID);
6100 	hr_reg_write(eqc, EQC_EQE_HOP_NUM, eq->hop_num);
6101 	hr_reg_write(eqc, EQC_OVER_IGNORE, eq->over_ignore);
6102 	hr_reg_write(eqc, EQC_COALESCE, eq->coalesce);
6103 	hr_reg_write(eqc, EQC_ARM_ST, eq->arm_st);
6104 	hr_reg_write(eqc, EQC_EQN, eq->eqn);
6105 	hr_reg_write(eqc, EQC_EQE_CNT, HNS_ROCE_EQ_INIT_EQE_CNT);
6106 	hr_reg_write(eqc, EQC_EQE_BA_PG_SZ,
6107 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.ba_pg_shift));
6108 	hr_reg_write(eqc, EQC_EQE_BUF_PG_SZ,
6109 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.buf_pg_shift));
6110 	hr_reg_write(eqc, EQC_EQ_PROD_INDX, HNS_ROCE_EQ_INIT_PROD_IDX);
6111 	hr_reg_write(eqc, EQC_EQ_MAX_CNT, eq->eq_max_cnt);
6112 
6113 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6114 		if (eq->eq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
6115 			dev_info(hr_dev->dev, "eq_period(%u) reached the upper limit, adjusted to 65.\n",
6116 				 eq->eq_period);
6117 			eq->eq_period = HNS_ROCE_MAX_EQ_PERIOD;
6118 		}
6119 		eq->eq_period *= HNS_ROCE_CLOCK_ADJUST;
6120 	}
6121 
6122 	hr_reg_write(eqc, EQC_EQ_PERIOD, eq->eq_period);
6123 	hr_reg_write(eqc, EQC_EQE_REPORT_TIMER, HNS_ROCE_EQ_INIT_REPORT_TIMER);
6124 	hr_reg_write(eqc, EQC_EQE_BA_L, bt_ba >> 3);
6125 	hr_reg_write(eqc, EQC_EQE_BA_H, bt_ba >> 35);
6126 	hr_reg_write(eqc, EQC_SHIFT, eq->shift);
6127 	hr_reg_write(eqc, EQC_MSI_INDX, HNS_ROCE_EQ_INIT_MSI_IDX);
6128 	hr_reg_write(eqc, EQC_CUR_EQE_BA_L, eqe_ba[0] >> 12);
6129 	hr_reg_write(eqc, EQC_CUR_EQE_BA_M, eqe_ba[0] >> 28);
6130 	hr_reg_write(eqc, EQC_CUR_EQE_BA_H, eqe_ba[0] >> 60);
6131 	hr_reg_write(eqc, EQC_EQ_CONS_INDX, HNS_ROCE_EQ_INIT_CONS_IDX);
6132 	hr_reg_write(eqc, EQC_NEX_EQE_BA_L, eqe_ba[1] >> 12);
6133 	hr_reg_write(eqc, EQC_NEX_EQE_BA_H, eqe_ba[1] >> 44);
6134 	hr_reg_write(eqc, EQC_EQE_SIZE, eq->eqe_size == HNS_ROCE_V3_EQE_SIZE);
6135 
6136 	return 0;
6137 }
6138 
6139 static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6140 {
6141 	struct hns_roce_buf_attr buf_attr = {};
6142 	int err;
6143 
6144 	if (hr_dev->caps.eqe_hop_num == HNS_ROCE_HOP_NUM_0)
6145 		eq->hop_num = 0;
6146 	else
6147 		eq->hop_num = hr_dev->caps.eqe_hop_num;
6148 
6149 	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
6150 	buf_attr.region[0].size = eq->entries * eq->eqe_size;
6151 	buf_attr.region[0].hopnum = eq->hop_num;
6152 	buf_attr.region_count = 1;
6153 
6154 	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
6155 				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
6156 				  0);
6157 	if (err)
6158 		dev_err(hr_dev->dev, "Failed to alloc EQE mtr, err %d\n", err);
6159 
6160 	return err;
6161 }
6162 
6163 static int hns_roce_v2_create_eq(struct hns_roce_dev *hr_dev,
6164 				 struct hns_roce_eq *eq, u8 eq_cmd)
6165 {
6166 	struct hns_roce_cmd_mailbox *mailbox;
6167 	int ret;
6168 
6169 	/* Allocate mailbox memory */
6170 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6171 	if (IS_ERR(mailbox))
6172 		return PTR_ERR(mailbox);
6173 
6174 	ret = alloc_eq_buf(hr_dev, eq);
6175 	if (ret)
6176 		goto free_cmd_mbox;
6177 
6178 	ret = config_eqc(hr_dev, eq, mailbox->buf);
6179 	if (ret)
6180 		goto err_cmd_mbox;
6181 
6182 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, eq_cmd, eq->eqn);
6183 	if (ret) {
6184 		dev_err(hr_dev->dev, "[mailbox cmd] create eqc failed.\n");
6185 		goto err_cmd_mbox;
6186 	}
6187 
6188 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6189 
6190 	return 0;
6191 
6192 err_cmd_mbox:
6193 	free_eq_buf(hr_dev, eq);
6194 
6195 free_cmd_mbox:
6196 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6197 
6198 	return ret;
6199 }
6200 
6201 static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
6202 				  int comp_num, int aeq_num, int other_num)
6203 {
6204 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6205 	int i, j;
6206 	int ret;
6207 
6208 	for (i = 0; i < irq_num; i++) {
6209 		hr_dev->irq_names[i] = kzalloc(HNS_ROCE_INT_NAME_LEN,
6210 					       GFP_KERNEL);
6211 		if (!hr_dev->irq_names[i]) {
6212 			ret = -ENOMEM;
6213 			goto err_kzalloc_failed;
6214 		}
6215 	}
6216 
6217 	/* irq contains: abnormal + AEQ + CEQ */
6218 	for (j = 0; j < other_num; j++)
6219 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6220 			 "hns-abn-%d", j);
6221 
6222 	for (j = other_num; j < (other_num + aeq_num); j++)
6223 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6224 			 "hns-aeq-%d", j - other_num);
6225 
6226 	for (j = (other_num + aeq_num); j < irq_num; j++)
6227 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6228 			 "hns-ceq-%d", j - other_num - aeq_num);
6229 
6230 	for (j = 0; j < irq_num; j++) {
6231 		if (j < other_num)
6232 			ret = request_irq(hr_dev->irq[j],
6233 					  hns_roce_v2_msix_interrupt_abn,
6234 					  0, hr_dev->irq_names[j], hr_dev);
6235 
6236 		else if (j < (other_num + comp_num))
6237 			ret = request_irq(eq_table->eq[j - other_num].irq,
6238 					  hns_roce_v2_msix_interrupt_eq,
6239 					  0, hr_dev->irq_names[j + aeq_num],
6240 					  &eq_table->eq[j - other_num]);
6241 		else
6242 			ret = request_irq(eq_table->eq[j - other_num].irq,
6243 					  hns_roce_v2_msix_interrupt_eq,
6244 					  0, hr_dev->irq_names[j - comp_num],
6245 					  &eq_table->eq[j - other_num]);
6246 		if (ret) {
6247 			dev_err(hr_dev->dev, "Request irq error!\n");
6248 			goto err_request_failed;
6249 		}
6250 	}
6251 
6252 	return 0;
6253 
6254 err_request_failed:
6255 	for (j -= 1; j >= 0; j--)
6256 		if (j < other_num)
6257 			free_irq(hr_dev->irq[j], hr_dev);
6258 		else
6259 			free_irq(eq_table->eq[j - other_num].irq,
6260 				 &eq_table->eq[j - other_num]);
6261 
6262 err_kzalloc_failed:
6263 	for (i -= 1; i >= 0; i--)
6264 		kfree(hr_dev->irq_names[i]);
6265 
6266 	return ret;
6267 }
6268 
6269 static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev)
6270 {
6271 	int irq_num;
6272 	int eq_num;
6273 	int i;
6274 
6275 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6276 	irq_num = eq_num + hr_dev->caps.num_other_vectors;
6277 
6278 	for (i = 0; i < hr_dev->caps.num_other_vectors; i++)
6279 		free_irq(hr_dev->irq[i], hr_dev);
6280 
6281 	for (i = 0; i < eq_num; i++)
6282 		free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]);
6283 
6284 	for (i = 0; i < irq_num; i++)
6285 		kfree(hr_dev->irq_names[i]);
6286 }
6287 
6288 static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
6289 {
6290 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6291 	struct device *dev = hr_dev->dev;
6292 	struct hns_roce_eq *eq;
6293 	int other_num;
6294 	int comp_num;
6295 	int aeq_num;
6296 	int irq_num;
6297 	int eq_num;
6298 	u8 eq_cmd;
6299 	int ret;
6300 	int i;
6301 
6302 	other_num = hr_dev->caps.num_other_vectors;
6303 	comp_num = hr_dev->caps.num_comp_vectors;
6304 	aeq_num = hr_dev->caps.num_aeq_vectors;
6305 
6306 	eq_num = comp_num + aeq_num;
6307 	irq_num = eq_num + other_num;
6308 
6309 	eq_table->eq = kcalloc(eq_num, sizeof(*eq_table->eq), GFP_KERNEL);
6310 	if (!eq_table->eq)
6311 		return -ENOMEM;
6312 
6313 	/* create eq */
6314 	for (i = 0; i < eq_num; i++) {
6315 		eq = &eq_table->eq[i];
6316 		eq->hr_dev = hr_dev;
6317 		eq->eqn = i;
6318 		if (i < comp_num) {
6319 			/* CEQ */
6320 			eq_cmd = HNS_ROCE_CMD_CREATE_CEQC;
6321 			eq->type_flag = HNS_ROCE_CEQ;
6322 			eq->entries = hr_dev->caps.ceqe_depth;
6323 			eq->eqe_size = hr_dev->caps.ceqe_size;
6324 			eq->irq = hr_dev->irq[i + other_num + aeq_num];
6325 			eq->eq_max_cnt = HNS_ROCE_CEQ_DEFAULT_BURST_NUM;
6326 			eq->eq_period = HNS_ROCE_CEQ_DEFAULT_INTERVAL;
6327 		} else {
6328 			/* AEQ */
6329 			eq_cmd = HNS_ROCE_CMD_CREATE_AEQC;
6330 			eq->type_flag = HNS_ROCE_AEQ;
6331 			eq->entries = hr_dev->caps.aeqe_depth;
6332 			eq->eqe_size = hr_dev->caps.aeqe_size;
6333 			eq->irq = hr_dev->irq[i - comp_num + other_num];
6334 			eq->eq_max_cnt = HNS_ROCE_AEQ_DEFAULT_BURST_NUM;
6335 			eq->eq_period = HNS_ROCE_AEQ_DEFAULT_INTERVAL;
6336 		}
6337 
6338 		ret = hns_roce_v2_create_eq(hr_dev, eq, eq_cmd);
6339 		if (ret) {
6340 			dev_err(dev, "failed to create eq.\n");
6341 			goto err_create_eq_fail;
6342 		}
6343 	}
6344 
6345 	hr_dev->irq_workq = alloc_ordered_workqueue("hns_roce_irq_workq", 0);
6346 	if (!hr_dev->irq_workq) {
6347 		dev_err(dev, "failed to create irq workqueue.\n");
6348 		ret = -ENOMEM;
6349 		goto err_create_eq_fail;
6350 	}
6351 
6352 	ret = __hns_roce_request_irq(hr_dev, irq_num, comp_num, aeq_num,
6353 				     other_num);
6354 	if (ret) {
6355 		dev_err(dev, "failed to request irq.\n");
6356 		goto err_request_irq_fail;
6357 	}
6358 
6359 	/* enable irq */
6360 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_ENABLE);
6361 
6362 	return 0;
6363 
6364 err_request_irq_fail:
6365 	destroy_workqueue(hr_dev->irq_workq);
6366 
6367 err_create_eq_fail:
6368 	for (i -= 1; i >= 0; i--)
6369 		free_eq_buf(hr_dev, &eq_table->eq[i]);
6370 	kfree(eq_table->eq);
6371 
6372 	return ret;
6373 }
6374 
6375 static void hns_roce_v2_cleanup_eq_table(struct hns_roce_dev *hr_dev)
6376 {
6377 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6378 	int eq_num;
6379 	int i;
6380 
6381 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6382 
6383 	/* Disable irq */
6384 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_DISABLE);
6385 
6386 	__hns_roce_free_irq(hr_dev);
6387 	destroy_workqueue(hr_dev->irq_workq);
6388 
6389 	for (i = 0; i < eq_num; i++) {
6390 		hns_roce_v2_destroy_eqc(hr_dev, i);
6391 
6392 		free_eq_buf(hr_dev, &eq_table->eq[i]);
6393 	}
6394 
6395 	kfree(eq_table->eq);
6396 }
6397 
6398 static const struct hns_roce_dfx_hw hns_roce_dfx_hw_v2 = {
6399 	.query_cqc_info = hns_roce_v2_query_cqc_info,
6400 };
6401 
6402 static const struct ib_device_ops hns_roce_v2_dev_ops = {
6403 	.destroy_qp = hns_roce_v2_destroy_qp,
6404 	.modify_cq = hns_roce_v2_modify_cq,
6405 	.poll_cq = hns_roce_v2_poll_cq,
6406 	.post_recv = hns_roce_v2_post_recv,
6407 	.post_send = hns_roce_v2_post_send,
6408 	.query_qp = hns_roce_v2_query_qp,
6409 	.req_notify_cq = hns_roce_v2_req_notify_cq,
6410 };
6411 
6412 static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {
6413 	.modify_srq = hns_roce_v2_modify_srq,
6414 	.post_srq_recv = hns_roce_v2_post_srq_recv,
6415 	.query_srq = hns_roce_v2_query_srq,
6416 };
6417 
6418 static const struct hns_roce_hw hns_roce_hw_v2 = {
6419 	.cmq_init = hns_roce_v2_cmq_init,
6420 	.cmq_exit = hns_roce_v2_cmq_exit,
6421 	.hw_profile = hns_roce_v2_profile,
6422 	.hw_init = hns_roce_v2_init,
6423 	.hw_exit = hns_roce_v2_exit,
6424 	.post_mbox = v2_post_mbox,
6425 	.poll_mbox_done = v2_poll_mbox_done,
6426 	.chk_mbox_avail = v2_chk_mbox_is_avail,
6427 	.set_gid = hns_roce_v2_set_gid,
6428 	.set_mac = hns_roce_v2_set_mac,
6429 	.write_mtpt = hns_roce_v2_write_mtpt,
6430 	.rereg_write_mtpt = hns_roce_v2_rereg_write_mtpt,
6431 	.frmr_write_mtpt = hns_roce_v2_frmr_write_mtpt,
6432 	.mw_write_mtpt = hns_roce_v2_mw_write_mtpt,
6433 	.write_cqc = hns_roce_v2_write_cqc,
6434 	.set_hem = hns_roce_v2_set_hem,
6435 	.clear_hem = hns_roce_v2_clear_hem,
6436 	.modify_qp = hns_roce_v2_modify_qp,
6437 	.dereg_mr = hns_roce_v2_dereg_mr,
6438 	.qp_flow_control_init = hns_roce_v2_qp_flow_control_init,
6439 	.init_eq = hns_roce_v2_init_eq_table,
6440 	.cleanup_eq = hns_roce_v2_cleanup_eq_table,
6441 	.write_srqc = hns_roce_v2_write_srqc,
6442 	.hns_roce_dev_ops = &hns_roce_v2_dev_ops,
6443 	.hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops,
6444 };
6445 
6446 static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = {
6447 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA), 0},
6448 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA_MACSEC), 0},
6449 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA), 0},
6450 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA_MACSEC), 0},
6451 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_100G_RDMA_MACSEC), 0},
6452 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_200G_RDMA), 0},
6453 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_RDMA_DCB_PFC_VF),
6454 	 HNAE3_DEV_SUPPORT_ROCE_DCB_BITS},
6455 	/* required last entry */
6456 	{0, }
6457 };
6458 
6459 MODULE_DEVICE_TABLE(pci, hns_roce_hw_v2_pci_tbl);
6460 
6461 static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
6462 				  struct hnae3_handle *handle)
6463 {
6464 	struct hns_roce_v2_priv *priv = hr_dev->priv;
6465 	const struct pci_device_id *id;
6466 	int i;
6467 
6468 	hr_dev->pci_dev = handle->pdev;
6469 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, hr_dev->pci_dev);
6470 	hr_dev->is_vf = id->driver_data;
6471 	hr_dev->dev = &handle->pdev->dev;
6472 	hr_dev->hw = &hns_roce_hw_v2;
6473 	hr_dev->dfx = &hns_roce_dfx_hw_v2;
6474 	hr_dev->sdb_offset = ROCEE_DB_SQ_L_0_REG;
6475 	hr_dev->odb_offset = hr_dev->sdb_offset;
6476 
6477 	/* Get info from NIC driver. */
6478 	hr_dev->reg_base = handle->rinfo.roce_io_base;
6479 	hr_dev->mem_base = handle->rinfo.roce_mem_base;
6480 	hr_dev->caps.num_ports = 1;
6481 	hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
6482 	hr_dev->iboe.phy_port[0] = 0;
6483 
6484 	addrconf_addr_eui48((u8 *)&hr_dev->ib_dev.node_guid,
6485 			    hr_dev->iboe.netdevs[0]->dev_addr);
6486 
6487 	for (i = 0; i < handle->rinfo.num_vectors; i++)
6488 		hr_dev->irq[i] = pci_irq_vector(handle->pdev,
6489 						i + handle->rinfo.base_vector);
6490 
6491 	/* cmd issue mode: 0 is poll, 1 is event */
6492 	hr_dev->cmd_mod = 1;
6493 	hr_dev->loop_idc = 0;
6494 
6495 	hr_dev->reset_cnt = handle->ae_algo->ops->ae_dev_reset_cnt(handle);
6496 	priv->handle = handle;
6497 }
6498 
6499 static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6500 {
6501 	struct hns_roce_dev *hr_dev;
6502 	int ret;
6503 
6504 	hr_dev = ib_alloc_device(hns_roce_dev, ib_dev);
6505 	if (!hr_dev)
6506 		return -ENOMEM;
6507 
6508 	hr_dev->priv = kzalloc(sizeof(struct hns_roce_v2_priv), GFP_KERNEL);
6509 	if (!hr_dev->priv) {
6510 		ret = -ENOMEM;
6511 		goto error_failed_kzalloc;
6512 	}
6513 
6514 	hns_roce_hw_v2_get_cfg(hr_dev, handle);
6515 
6516 	ret = hns_roce_init(hr_dev);
6517 	if (ret) {
6518 		dev_err(hr_dev->dev, "RoCE Engine init failed!\n");
6519 		goto error_failed_cfg;
6520 	}
6521 
6522 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6523 		ret = free_mr_init(hr_dev);
6524 		if (ret) {
6525 			dev_err(hr_dev->dev, "failed to init free mr!\n");
6526 			goto error_failed_roce_init;
6527 		}
6528 	}
6529 
6530 	handle->priv = hr_dev;
6531 
6532 	return 0;
6533 
6534 error_failed_roce_init:
6535 	hns_roce_exit(hr_dev);
6536 
6537 error_failed_cfg:
6538 	kfree(hr_dev->priv);
6539 
6540 error_failed_kzalloc:
6541 	ib_dealloc_device(&hr_dev->ib_dev);
6542 
6543 	return ret;
6544 }
6545 
6546 static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6547 					   bool reset)
6548 {
6549 	struct hns_roce_dev *hr_dev = handle->priv;
6550 
6551 	if (!hr_dev)
6552 		return;
6553 
6554 	handle->priv = NULL;
6555 
6556 	hr_dev->state = HNS_ROCE_DEVICE_STATE_UNINIT;
6557 	hns_roce_handle_device_err(hr_dev);
6558 
6559 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
6560 		free_mr_exit(hr_dev);
6561 
6562 	hns_roce_exit(hr_dev);
6563 	kfree(hr_dev->priv);
6564 	ib_dealloc_device(&hr_dev->ib_dev);
6565 }
6566 
6567 static int hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6568 {
6569 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
6570 	const struct pci_device_id *id;
6571 	struct device *dev = &handle->pdev->dev;
6572 	int ret;
6573 
6574 	handle->rinfo.instance_state = HNS_ROCE_STATE_INIT;
6575 
6576 	if (ops->ae_dev_resetting(handle) || ops->get_hw_reset_stat(handle)) {
6577 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6578 		goto reset_chk_err;
6579 	}
6580 
6581 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, handle->pdev);
6582 	if (!id)
6583 		return 0;
6584 
6585 	if (id->driver_data && handle->pdev->revision == PCI_REVISION_ID_HIP08)
6586 		return 0;
6587 
6588 	ret = __hns_roce_hw_v2_init_instance(handle);
6589 	if (ret) {
6590 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6591 		dev_err(dev, "RoCE instance init failed! ret = %d\n", ret);
6592 		if (ops->ae_dev_resetting(handle) ||
6593 		    ops->get_hw_reset_stat(handle))
6594 			goto reset_chk_err;
6595 		else
6596 			return ret;
6597 	}
6598 
6599 	handle->rinfo.instance_state = HNS_ROCE_STATE_INITED;
6600 
6601 	return 0;
6602 
6603 reset_chk_err:
6604 	dev_err(dev, "Device is busy in resetting state.\n"
6605 		     "please retry later.\n");
6606 
6607 	return -EBUSY;
6608 }
6609 
6610 static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6611 					   bool reset)
6612 {
6613 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED)
6614 		return;
6615 
6616 	handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;
6617 
6618 	__hns_roce_hw_v2_uninit_instance(handle, reset);
6619 
6620 	handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6621 }
6622 static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
6623 {
6624 	struct hns_roce_dev *hr_dev;
6625 
6626 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) {
6627 		set_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
6628 		return 0;
6629 	}
6630 
6631 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_DOWN;
6632 	clear_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
6633 
6634 	hr_dev = handle->priv;
6635 	if (!hr_dev)
6636 		return 0;
6637 
6638 	hr_dev->active = false;
6639 	hr_dev->dis_db = true;
6640 	hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
6641 
6642 	return 0;
6643 }
6644 
6645 static int hns_roce_hw_v2_reset_notify_init(struct hnae3_handle *handle)
6646 {
6647 	struct device *dev = &handle->pdev->dev;
6648 	int ret;
6649 
6650 	if (test_and_clear_bit(HNS_ROCE_RST_DIRECT_RETURN,
6651 			       &handle->rinfo.state)) {
6652 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
6653 		return 0;
6654 	}
6655 
6656 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INIT;
6657 
6658 	dev_info(&handle->pdev->dev, "In reset process RoCE client reinit.\n");
6659 	ret = __hns_roce_hw_v2_init_instance(handle);
6660 	if (ret) {
6661 		/* when reset notify type is HNAE3_INIT_CLIENT In reset notify
6662 		 * callback function, RoCE Engine reinitialize. If RoCE reinit
6663 		 * failed, we should inform NIC driver.
6664 		 */
6665 		handle->priv = NULL;
6666 		dev_err(dev, "In reset process RoCE reinit failed %d.\n", ret);
6667 	} else {
6668 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
6669 		dev_info(dev, "Reset done, RoCE client reinit finished.\n");
6670 	}
6671 
6672 	return ret;
6673 }
6674 
6675 static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle)
6676 {
6677 	if (test_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state))
6678 		return 0;
6679 
6680 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT;
6681 	dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n");
6682 	msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY);
6683 	__hns_roce_hw_v2_uninit_instance(handle, false);
6684 
6685 	return 0;
6686 }
6687 
6688 static int hns_roce_hw_v2_reset_notify(struct hnae3_handle *handle,
6689 				       enum hnae3_reset_notify_type type)
6690 {
6691 	int ret = 0;
6692 
6693 	switch (type) {
6694 	case HNAE3_DOWN_CLIENT:
6695 		ret = hns_roce_hw_v2_reset_notify_down(handle);
6696 		break;
6697 	case HNAE3_INIT_CLIENT:
6698 		ret = hns_roce_hw_v2_reset_notify_init(handle);
6699 		break;
6700 	case HNAE3_UNINIT_CLIENT:
6701 		ret = hns_roce_hw_v2_reset_notify_uninit(handle);
6702 		break;
6703 	default:
6704 		break;
6705 	}
6706 
6707 	return ret;
6708 }
6709 
6710 static const struct hnae3_client_ops hns_roce_hw_v2_ops = {
6711 	.init_instance = hns_roce_hw_v2_init_instance,
6712 	.uninit_instance = hns_roce_hw_v2_uninit_instance,
6713 	.reset_notify = hns_roce_hw_v2_reset_notify,
6714 };
6715 
6716 static struct hnae3_client hns_roce_hw_v2_client = {
6717 	.name = "hns_roce_hw_v2",
6718 	.type = HNAE3_CLIENT_ROCE,
6719 	.ops = &hns_roce_hw_v2_ops,
6720 };
6721 
6722 static int __init hns_roce_hw_v2_init(void)
6723 {
6724 	return hnae3_register_client(&hns_roce_hw_v2_client);
6725 }
6726 
6727 static void __exit hns_roce_hw_v2_exit(void)
6728 {
6729 	hnae3_unregister_client(&hns_roce_hw_v2_client);
6730 }
6731 
6732 module_init(hns_roce_hw_v2_init);
6733 module_exit(hns_roce_hw_v2_exit);
6734 
6735 MODULE_LICENSE("Dual BSD/GPL");
6736 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
6737 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
6738 MODULE_AUTHOR("Shaobo Xu <xushaobo2@huawei.com>");
6739 MODULE_DESCRIPTION("Hisilicon Hip08 Family RoCE Driver");
6740