xref: /linux/drivers/infiniband/hw/hns/hns_roce_hw_v2.c (revision f2c6dbd220170c2396fb019ead67fbada1e23ebd)
1 /*
2  * Copyright (c) 2016-2017 Hisilicon Limited.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/acpi.h>
34 #include <linux/etherdevice.h>
35 #include <linux/interrupt.h>
36 #include <linux/iopoll.h>
37 #include <linux/kernel.h>
38 #include <linux/types.h>
39 #include <linux/workqueue.h>
40 #include <net/addrconf.h>
41 #include <rdma/ib_addr.h>
42 #include <rdma/ib_cache.h>
43 #include <rdma/ib_umem.h>
44 #include <rdma/uverbs_ioctl.h>
45 
46 #include "hnae3.h"
47 #include "hns_roce_common.h"
48 #include "hns_roce_device.h"
49 #include "hns_roce_cmd.h"
50 #include "hns_roce_hem.h"
51 #include "hns_roce_hw_v2.h"
52 
53 enum {
54 	CMD_RST_PRC_OTHERS,
55 	CMD_RST_PRC_SUCCESS,
56 	CMD_RST_PRC_EBUSY,
57 };
58 
59 enum ecc_resource_type {
60 	ECC_RESOURCE_QPC,
61 	ECC_RESOURCE_CQC,
62 	ECC_RESOURCE_MPT,
63 	ECC_RESOURCE_SRQC,
64 	ECC_RESOURCE_GMV,
65 	ECC_RESOURCE_QPC_TIMER,
66 	ECC_RESOURCE_CQC_TIMER,
67 	ECC_RESOURCE_SCCC,
68 	ECC_RESOURCE_COUNT,
69 };
70 
71 static const struct {
72 	const char *name;
73 	u8 read_bt0_op;
74 	u8 write_bt0_op;
75 } fmea_ram_res[] = {
76 	{ "ECC_RESOURCE_QPC",
77 	  HNS_ROCE_CMD_READ_QPC_BT0, HNS_ROCE_CMD_WRITE_QPC_BT0 },
78 	{ "ECC_RESOURCE_CQC",
79 	  HNS_ROCE_CMD_READ_CQC_BT0, HNS_ROCE_CMD_WRITE_CQC_BT0 },
80 	{ "ECC_RESOURCE_MPT",
81 	  HNS_ROCE_CMD_READ_MPT_BT0, HNS_ROCE_CMD_WRITE_MPT_BT0 },
82 	{ "ECC_RESOURCE_SRQC",
83 	  HNS_ROCE_CMD_READ_SRQC_BT0, HNS_ROCE_CMD_WRITE_SRQC_BT0 },
84 	/* ECC_RESOURCE_GMV is handled by cmdq, not mailbox */
85 	{ "ECC_RESOURCE_GMV",
86 	  0, 0 },
87 	{ "ECC_RESOURCE_QPC_TIMER",
88 	  HNS_ROCE_CMD_READ_QPC_TIMER_BT0, HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0 },
89 	{ "ECC_RESOURCE_CQC_TIMER",
90 	  HNS_ROCE_CMD_READ_CQC_TIMER_BT0, HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0 },
91 	{ "ECC_RESOURCE_SCCC",
92 	  HNS_ROCE_CMD_READ_SCCC_BT0, HNS_ROCE_CMD_WRITE_SCCC_BT0 },
93 };
94 
95 static inline void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
96 				   struct ib_sge *sg)
97 {
98 	dseg->lkey = cpu_to_le32(sg->lkey);
99 	dseg->addr = cpu_to_le64(sg->addr);
100 	dseg->len  = cpu_to_le32(sg->length);
101 }
102 
103 /*
104  * mapped-value = 1 + real-value
105  * The hns wr opcode real value is start from 0, In order to distinguish between
106  * initialized and uninitialized map values, we plus 1 to the actual value when
107  * defining the mapping, so that the validity can be identified by checking the
108  * mapped value is greater than 0.
109  */
110 #define HR_OPC_MAP(ib_key, hr_key) \
111 		[IB_WR_ ## ib_key] = 1 + HNS_ROCE_V2_WQE_OP_ ## hr_key
112 
113 static const u32 hns_roce_op_code[] = {
114 	HR_OPC_MAP(RDMA_WRITE,			RDMA_WRITE),
115 	HR_OPC_MAP(RDMA_WRITE_WITH_IMM,		RDMA_WRITE_WITH_IMM),
116 	HR_OPC_MAP(SEND,			SEND),
117 	HR_OPC_MAP(SEND_WITH_IMM,		SEND_WITH_IMM),
118 	HR_OPC_MAP(RDMA_READ,			RDMA_READ),
119 	HR_OPC_MAP(ATOMIC_CMP_AND_SWP,		ATOM_CMP_AND_SWAP),
120 	HR_OPC_MAP(ATOMIC_FETCH_AND_ADD,	ATOM_FETCH_AND_ADD),
121 	HR_OPC_MAP(SEND_WITH_INV,		SEND_WITH_INV),
122 	HR_OPC_MAP(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
123 	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
124 	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
125 };
126 
127 static u32 to_hr_opcode(u32 ib_opcode)
128 {
129 	if (ib_opcode >= ARRAY_SIZE(hns_roce_op_code))
130 		return HNS_ROCE_V2_WQE_OP_MASK;
131 
132 	return hns_roce_op_code[ib_opcode] ? hns_roce_op_code[ib_opcode] - 1 :
133 					     HNS_ROCE_V2_WQE_OP_MASK;
134 }
135 
136 static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
137 			 const struct ib_reg_wr *wr)
138 {
139 	struct hns_roce_wqe_frmr_seg *fseg =
140 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
141 	struct hns_roce_mr *mr = to_hr_mr(wr->mr);
142 	u64 pbl_ba;
143 
144 	/* use ib_access_flags */
145 	hr_reg_write_bool(fseg, FRMR_BIND_EN, wr->access & IB_ACCESS_MW_BIND);
146 	hr_reg_write_bool(fseg, FRMR_ATOMIC,
147 			  wr->access & IB_ACCESS_REMOTE_ATOMIC);
148 	hr_reg_write_bool(fseg, FRMR_RR, wr->access & IB_ACCESS_REMOTE_READ);
149 	hr_reg_write_bool(fseg, FRMR_RW, wr->access & IB_ACCESS_REMOTE_WRITE);
150 	hr_reg_write_bool(fseg, FRMR_LW, wr->access & IB_ACCESS_LOCAL_WRITE);
151 
152 	/* Data structure reuse may lead to confusion */
153 	pbl_ba = mr->pbl_mtr.hem_cfg.root_ba;
154 	rc_sq_wqe->msg_len = cpu_to_le32(lower_32_bits(pbl_ba));
155 	rc_sq_wqe->inv_key = cpu_to_le32(upper_32_bits(pbl_ba));
156 
157 	rc_sq_wqe->byte_16 = cpu_to_le32(wr->mr->length & 0xffffffff);
158 	rc_sq_wqe->byte_20 = cpu_to_le32(wr->mr->length >> 32);
159 	rc_sq_wqe->rkey = cpu_to_le32(wr->key);
160 	rc_sq_wqe->va = cpu_to_le64(wr->mr->iova);
161 
162 	hr_reg_write(fseg, FRMR_PBL_SIZE, mr->npages);
163 	hr_reg_write(fseg, FRMR_PBL_BUF_PG_SZ,
164 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
165 	hr_reg_clear(fseg, FRMR_BLK_MODE);
166 }
167 
168 static void set_atomic_seg(const struct ib_send_wr *wr,
169 			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
170 			   unsigned int valid_num_sge)
171 {
172 	struct hns_roce_v2_wqe_data_seg *dseg =
173 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
174 	struct hns_roce_wqe_atomic_seg *aseg =
175 		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
176 
177 	set_data_seg_v2(dseg, wr->sg_list);
178 
179 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
180 		aseg->fetchadd_swap_data = cpu_to_le64(atomic_wr(wr)->swap);
181 		aseg->cmp_data = cpu_to_le64(atomic_wr(wr)->compare_add);
182 	} else {
183 		aseg->fetchadd_swap_data =
184 			cpu_to_le64(atomic_wr(wr)->compare_add);
185 		aseg->cmp_data = 0;
186 	}
187 
188 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
189 }
190 
191 static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
192 				 const struct ib_send_wr *wr,
193 				 unsigned int *sge_idx, u32 msg_len)
194 {
195 	struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
196 	unsigned int left_len_in_pg;
197 	unsigned int idx = *sge_idx;
198 	unsigned int i = 0;
199 	unsigned int len;
200 	void *addr;
201 	void *dseg;
202 
203 	if (msg_len > qp->sq.ext_sge_cnt * HNS_ROCE_SGE_SIZE) {
204 		ibdev_err(ibdev,
205 			  "no enough extended sge space for inline data.\n");
206 		return -EINVAL;
207 	}
208 
209 	dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
210 	left_len_in_pg = hr_hw_page_align((uintptr_t)dseg) - (uintptr_t)dseg;
211 	len = wr->sg_list[0].length;
212 	addr = (void *)(unsigned long)(wr->sg_list[0].addr);
213 
214 	/* When copying data to extended sge space, the left length in page may
215 	 * not long enough for current user's sge. So the data should be
216 	 * splited into several parts, one in the first page, and the others in
217 	 * the subsequent pages.
218 	 */
219 	while (1) {
220 		if (len <= left_len_in_pg) {
221 			memcpy(dseg, addr, len);
222 
223 			idx += len / HNS_ROCE_SGE_SIZE;
224 
225 			i++;
226 			if (i >= wr->num_sge)
227 				break;
228 
229 			left_len_in_pg -= len;
230 			len = wr->sg_list[i].length;
231 			addr = (void *)(unsigned long)(wr->sg_list[i].addr);
232 			dseg += len;
233 		} else {
234 			memcpy(dseg, addr, left_len_in_pg);
235 
236 			len -= left_len_in_pg;
237 			addr += left_len_in_pg;
238 			idx += left_len_in_pg / HNS_ROCE_SGE_SIZE;
239 			dseg = hns_roce_get_extend_sge(qp,
240 						idx & (qp->sge.sge_cnt - 1));
241 			left_len_in_pg = 1 << HNS_HW_PAGE_SHIFT;
242 		}
243 	}
244 
245 	*sge_idx = idx;
246 
247 	return 0;
248 }
249 
250 static void set_extend_sge(struct hns_roce_qp *qp, struct ib_sge *sge,
251 			   unsigned int *sge_ind, unsigned int cnt)
252 {
253 	struct hns_roce_v2_wqe_data_seg *dseg;
254 	unsigned int idx = *sge_ind;
255 
256 	while (cnt > 0) {
257 		dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
258 		if (likely(sge->length)) {
259 			set_data_seg_v2(dseg, sge);
260 			idx++;
261 			cnt--;
262 		}
263 		sge++;
264 	}
265 
266 	*sge_ind = idx;
267 }
268 
269 static bool check_inl_data_len(struct hns_roce_qp *qp, unsigned int len)
270 {
271 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
272 	int mtu = ib_mtu_enum_to_int(qp->path_mtu);
273 
274 	if (mtu < 0 || len > qp->max_inline_data || len > mtu) {
275 		ibdev_err(&hr_dev->ib_dev,
276 			  "invalid length of data, data len = %u, max inline len = %u, path mtu = %d.\n",
277 			  len, qp->max_inline_data, mtu);
278 		return false;
279 	}
280 
281 	return true;
282 }
283 
284 static int set_rc_inl(struct hns_roce_qp *qp, const struct ib_send_wr *wr,
285 		      struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
286 		      unsigned int *sge_idx)
287 {
288 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
289 	u32 msg_len = le32_to_cpu(rc_sq_wqe->msg_len);
290 	struct ib_device *ibdev = &hr_dev->ib_dev;
291 	unsigned int curr_idx = *sge_idx;
292 	void *dseg = rc_sq_wqe;
293 	unsigned int i;
294 	int ret;
295 
296 	if (unlikely(wr->opcode == IB_WR_RDMA_READ)) {
297 		ibdev_err(ibdev, "invalid inline parameters!\n");
298 		return -EINVAL;
299 	}
300 
301 	if (!check_inl_data_len(qp, msg_len))
302 		return -EINVAL;
303 
304 	dseg += sizeof(struct hns_roce_v2_rc_send_wqe);
305 
306 	if (msg_len <= HNS_ROCE_V2_MAX_RC_INL_INN_SZ) {
307 		hr_reg_clear(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
308 
309 		for (i = 0; i < wr->num_sge; i++) {
310 			memcpy(dseg, ((void *)wr->sg_list[i].addr),
311 			       wr->sg_list[i].length);
312 			dseg += wr->sg_list[i].length;
313 		}
314 	} else {
315 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
316 
317 		ret = fill_ext_sge_inl_data(qp, wr, &curr_idx, msg_len);
318 		if (ret)
319 			return ret;
320 
321 		hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, curr_idx - *sge_idx);
322 	}
323 
324 	*sge_idx = curr_idx;
325 
326 	return 0;
327 }
328 
329 static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
330 			     struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
331 			     unsigned int *sge_ind,
332 			     unsigned int valid_num_sge)
333 {
334 	struct hns_roce_v2_wqe_data_seg *dseg =
335 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
336 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
337 	int j = 0;
338 	int i;
339 
340 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_MSG_START_SGE_IDX,
341 		     (*sge_ind) & (qp->sge.sge_cnt - 1));
342 
343 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_INLINE,
344 		     !!(wr->send_flags & IB_SEND_INLINE));
345 	if (wr->send_flags & IB_SEND_INLINE)
346 		return set_rc_inl(qp, wr, rc_sq_wqe, sge_ind);
347 
348 	if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
349 		for (i = 0; i < wr->num_sge; i++) {
350 			if (likely(wr->sg_list[i].length)) {
351 				set_data_seg_v2(dseg, wr->sg_list + i);
352 				dseg++;
353 			}
354 		}
355 	} else {
356 		for (i = 0; i < wr->num_sge && j < HNS_ROCE_SGE_IN_WQE; i++) {
357 			if (likely(wr->sg_list[i].length)) {
358 				set_data_seg_v2(dseg, wr->sg_list + i);
359 				dseg++;
360 				j++;
361 			}
362 		}
363 
364 		set_extend_sge(qp, wr->sg_list + i, sge_ind,
365 			       valid_num_sge - HNS_ROCE_SGE_IN_WQE);
366 	}
367 
368 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
369 
370 	return 0;
371 }
372 
373 static int check_send_valid(struct hns_roce_dev *hr_dev,
374 			    struct hns_roce_qp *hr_qp)
375 {
376 	struct ib_device *ibdev = &hr_dev->ib_dev;
377 
378 	if (unlikely(hr_qp->state == IB_QPS_RESET ||
379 		     hr_qp->state == IB_QPS_INIT ||
380 		     hr_qp->state == IB_QPS_RTR)) {
381 		ibdev_err(ibdev, "failed to post WQE, QP state %u!\n",
382 			  hr_qp->state);
383 		return -EINVAL;
384 	} else if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)) {
385 		ibdev_err(ibdev, "failed to post WQE, dev state %d!\n",
386 			  hr_dev->state);
387 		return -EIO;
388 	}
389 
390 	return 0;
391 }
392 
393 static unsigned int calc_wr_sge_num(const struct ib_send_wr *wr,
394 				    unsigned int *sge_len)
395 {
396 	unsigned int valid_num = 0;
397 	unsigned int len = 0;
398 	int i;
399 
400 	for (i = 0; i < wr->num_sge; i++) {
401 		if (likely(wr->sg_list[i].length)) {
402 			len += wr->sg_list[i].length;
403 			valid_num++;
404 		}
405 	}
406 
407 	*sge_len = len;
408 	return valid_num;
409 }
410 
411 static __le32 get_immtdata(const struct ib_send_wr *wr)
412 {
413 	switch (wr->opcode) {
414 	case IB_WR_SEND_WITH_IMM:
415 	case IB_WR_RDMA_WRITE_WITH_IMM:
416 		return cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
417 	default:
418 		return 0;
419 	}
420 }
421 
422 static int set_ud_opcode(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
423 			 const struct ib_send_wr *wr)
424 {
425 	u32 ib_op = wr->opcode;
426 
427 	if (ib_op != IB_WR_SEND && ib_op != IB_WR_SEND_WITH_IMM)
428 		return -EINVAL;
429 
430 	ud_sq_wqe->immtdata = get_immtdata(wr);
431 
432 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
433 
434 	return 0;
435 }
436 
437 static int fill_ud_av(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
438 		      struct hns_roce_ah *ah)
439 {
440 	struct ib_device *ib_dev = ah->ibah.device;
441 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
442 
443 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_UDPSPN, ah->av.udp_sport);
444 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_HOPLIMIT, ah->av.hop_limit);
445 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_TCLASS, ah->av.tclass);
446 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_FLOW_LABEL, ah->av.flowlabel);
447 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SL, ah->av.sl);
448 
449 	ud_sq_wqe->sgid_index = ah->av.gid_index;
450 
451 	memcpy(ud_sq_wqe->dmac, ah->av.mac, ETH_ALEN);
452 	memcpy(ud_sq_wqe->dgid, ah->av.dgid, GID_LEN_V2);
453 
454 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
455 		return 0;
456 
457 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN_EN, ah->av.vlan_en);
458 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN, ah->av.vlan_id);
459 
460 	return 0;
461 }
462 
463 static inline int set_ud_wqe(struct hns_roce_qp *qp,
464 			     const struct ib_send_wr *wr,
465 			     void *wqe, unsigned int *sge_idx,
466 			     unsigned int owner_bit)
467 {
468 	struct hns_roce_ah *ah = to_hr_ah(ud_wr(wr)->ah);
469 	struct hns_roce_v2_ud_send_wqe *ud_sq_wqe = wqe;
470 	unsigned int curr_idx = *sge_idx;
471 	unsigned int valid_num_sge;
472 	u32 msg_len = 0;
473 	int ret;
474 
475 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
476 
477 	ret = set_ud_opcode(ud_sq_wqe, wr);
478 	if (WARN_ON(ret))
479 		return ret;
480 
481 	ud_sq_wqe->msg_len = cpu_to_le32(msg_len);
482 
483 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_CQE,
484 		     !!(wr->send_flags & IB_SEND_SIGNALED));
485 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SE,
486 		     !!(wr->send_flags & IB_SEND_SOLICITED));
487 
488 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_PD, to_hr_pd(qp->ibqp.pd)->pdn);
489 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SGE_NUM, valid_num_sge);
490 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_MSG_START_SGE_IDX,
491 		     curr_idx & (qp->sge.sge_cnt - 1));
492 
493 	ud_sq_wqe->qkey = cpu_to_le32(ud_wr(wr)->remote_qkey & 0x80000000 ?
494 			  qp->qkey : ud_wr(wr)->remote_qkey);
495 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_DQPN, ud_wr(wr)->remote_qpn);
496 
497 	ret = fill_ud_av(ud_sq_wqe, ah);
498 	if (ret)
499 		return ret;
500 
501 	qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
502 
503 	set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
504 
505 	/*
506 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
507 	 * including new WQEs waiting for the doorbell to update the PI again.
508 	 * Therefore, the owner bit of WQE MUST be updated after all fields
509 	 * and extSGEs have been written into DDR instead of cache.
510 	 */
511 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
512 		dma_wmb();
513 
514 	*sge_idx = curr_idx;
515 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OWNER, owner_bit);
516 
517 	return 0;
518 }
519 
520 static int set_rc_opcode(struct hns_roce_dev *hr_dev,
521 			 struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
522 			 const struct ib_send_wr *wr)
523 {
524 	u32 ib_op = wr->opcode;
525 	int ret = 0;
526 
527 	rc_sq_wqe->immtdata = get_immtdata(wr);
528 
529 	switch (ib_op) {
530 	case IB_WR_RDMA_READ:
531 	case IB_WR_RDMA_WRITE:
532 	case IB_WR_RDMA_WRITE_WITH_IMM:
533 		rc_sq_wqe->rkey = cpu_to_le32(rdma_wr(wr)->rkey);
534 		rc_sq_wqe->va = cpu_to_le64(rdma_wr(wr)->remote_addr);
535 		break;
536 	case IB_WR_SEND:
537 	case IB_WR_SEND_WITH_IMM:
538 		break;
539 	case IB_WR_ATOMIC_CMP_AND_SWP:
540 	case IB_WR_ATOMIC_FETCH_AND_ADD:
541 		rc_sq_wqe->rkey = cpu_to_le32(atomic_wr(wr)->rkey);
542 		rc_sq_wqe->va = cpu_to_le64(atomic_wr(wr)->remote_addr);
543 		break;
544 	case IB_WR_REG_MR:
545 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
546 			set_frmr_seg(rc_sq_wqe, reg_wr(wr));
547 		else
548 			ret = -EOPNOTSUPP;
549 		break;
550 	case IB_WR_SEND_WITH_INV:
551 		rc_sq_wqe->inv_key = cpu_to_le32(wr->ex.invalidate_rkey);
552 		break;
553 	default:
554 		ret = -EINVAL;
555 	}
556 
557 	if (unlikely(ret))
558 		return ret;
559 
560 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
561 
562 	return ret;
563 }
564 
565 static inline int set_rc_wqe(struct hns_roce_qp *qp,
566 			     const struct ib_send_wr *wr,
567 			     void *wqe, unsigned int *sge_idx,
568 			     unsigned int owner_bit)
569 {
570 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
571 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
572 	unsigned int curr_idx = *sge_idx;
573 	unsigned int valid_num_sge;
574 	u32 msg_len = 0;
575 	int ret;
576 
577 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
578 
579 	rc_sq_wqe->msg_len = cpu_to_le32(msg_len);
580 
581 	ret = set_rc_opcode(hr_dev, rc_sq_wqe, wr);
582 	if (WARN_ON(ret))
583 		return ret;
584 
585 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_FENCE,
586 		     (wr->send_flags & IB_SEND_FENCE) ? 1 : 0);
587 
588 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SE,
589 		     (wr->send_flags & IB_SEND_SOLICITED) ? 1 : 0);
590 
591 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_CQE,
592 		     (wr->send_flags & IB_SEND_SIGNALED) ? 1 : 0);
593 
594 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
595 	    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
596 		if (msg_len != ATOMIC_WR_LEN)
597 			return -EINVAL;
598 		set_atomic_seg(wr, rc_sq_wqe, valid_num_sge);
599 	} else if (wr->opcode != IB_WR_REG_MR) {
600 		ret = set_rwqe_data_seg(&qp->ibqp, wr, rc_sq_wqe,
601 					&curr_idx, valid_num_sge);
602 		if (ret)
603 			return ret;
604 	}
605 
606 	/*
607 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
608 	 * including new WQEs waiting for the doorbell to update the PI again.
609 	 * Therefore, the owner bit of WQE MUST be updated after all fields
610 	 * and extSGEs have been written into DDR instead of cache.
611 	 */
612 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
613 		dma_wmb();
614 
615 	*sge_idx = curr_idx;
616 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OWNER, owner_bit);
617 
618 	return ret;
619 }
620 
621 static inline void update_sq_db(struct hns_roce_dev *hr_dev,
622 				struct hns_roce_qp *qp)
623 {
624 	if (unlikely(qp->state == IB_QPS_ERR)) {
625 		flush_cqe(hr_dev, qp);
626 	} else {
627 		struct hns_roce_v2_db sq_db = {};
628 
629 		hr_reg_write(&sq_db, DB_TAG, qp->qpn);
630 		hr_reg_write(&sq_db, DB_CMD, HNS_ROCE_V2_SQ_DB);
631 		hr_reg_write(&sq_db, DB_PI, qp->sq.head);
632 		hr_reg_write(&sq_db, DB_SL, qp->sl);
633 
634 		hns_roce_write64(hr_dev, (__le32 *)&sq_db, qp->sq.db_reg);
635 	}
636 }
637 
638 static inline void update_rq_db(struct hns_roce_dev *hr_dev,
639 				struct hns_roce_qp *qp)
640 {
641 	if (unlikely(qp->state == IB_QPS_ERR)) {
642 		flush_cqe(hr_dev, qp);
643 	} else {
644 		if (likely(qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)) {
645 			*qp->rdb.db_record =
646 					qp->rq.head & V2_DB_PRODUCER_IDX_M;
647 		} else {
648 			struct hns_roce_v2_db rq_db = {};
649 
650 			hr_reg_write(&rq_db, DB_TAG, qp->qpn);
651 			hr_reg_write(&rq_db, DB_CMD, HNS_ROCE_V2_RQ_DB);
652 			hr_reg_write(&rq_db, DB_PI, qp->rq.head);
653 
654 			hns_roce_write64(hr_dev, (__le32 *)&rq_db,
655 					 qp->rq.db_reg);
656 		}
657 	}
658 }
659 
660 static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
661 			      u64 __iomem *dest)
662 {
663 #define HNS_ROCE_WRITE_TIMES 8
664 	struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
665 	struct hnae3_handle *handle = priv->handle;
666 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
667 	int i;
668 
669 	if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
670 		for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
671 			writeq_relaxed(*(val + i), dest + i);
672 }
673 
674 static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
675 		       void *wqe)
676 {
677 #define HNS_ROCE_SL_SHIFT 2
678 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
679 
680 	/* All kinds of DirectWQE have the same header field layout */
681 	hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_FLAG);
682 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_L, qp->sl);
683 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_H,
684 		     qp->sl >> HNS_ROCE_SL_SHIFT);
685 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_WQE_INDEX, qp->sq.head);
686 
687 	hns_roce_write512(hr_dev, wqe, qp->sq.db_reg);
688 }
689 
690 static int hns_roce_v2_post_send(struct ib_qp *ibqp,
691 				 const struct ib_send_wr *wr,
692 				 const struct ib_send_wr **bad_wr)
693 {
694 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
695 	struct ib_device *ibdev = &hr_dev->ib_dev;
696 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
697 	unsigned long flags = 0;
698 	unsigned int owner_bit;
699 	unsigned int sge_idx;
700 	unsigned int wqe_idx;
701 	void *wqe = NULL;
702 	u32 nreq;
703 	int ret;
704 
705 	spin_lock_irqsave(&qp->sq.lock, flags);
706 
707 	ret = check_send_valid(hr_dev, qp);
708 	if (unlikely(ret)) {
709 		*bad_wr = wr;
710 		nreq = 0;
711 		goto out;
712 	}
713 
714 	sge_idx = qp->next_sge;
715 
716 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
717 		if (hns_roce_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
718 			ret = -ENOMEM;
719 			*bad_wr = wr;
720 			goto out;
721 		}
722 
723 		wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
724 
725 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
726 			ibdev_err(ibdev, "num_sge = %d > qp->sq.max_gs = %u.\n",
727 				  wr->num_sge, qp->sq.max_gs);
728 			ret = -EINVAL;
729 			*bad_wr = wr;
730 			goto out;
731 		}
732 
733 		wqe = hns_roce_get_send_wqe(qp, wqe_idx);
734 		qp->sq.wrid[wqe_idx] = wr->wr_id;
735 		owner_bit =
736 		       ~(((qp->sq.head + nreq) >> ilog2(qp->sq.wqe_cnt)) & 0x1);
737 
738 		/* Corresponding to the QP type, wqe process separately */
739 		if (ibqp->qp_type == IB_QPT_RC)
740 			ret = set_rc_wqe(qp, wr, wqe, &sge_idx, owner_bit);
741 		else
742 			ret = set_ud_wqe(qp, wr, wqe, &sge_idx, owner_bit);
743 
744 		if (unlikely(ret)) {
745 			*bad_wr = wr;
746 			goto out;
747 		}
748 	}
749 
750 out:
751 	if (likely(nreq)) {
752 		qp->sq.head += nreq;
753 		qp->next_sge = sge_idx;
754 
755 		if (nreq == 1 && !ret &&
756 		    (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
757 			write_dwqe(hr_dev, qp, wqe);
758 		else
759 			update_sq_db(hr_dev, qp);
760 	}
761 
762 	spin_unlock_irqrestore(&qp->sq.lock, flags);
763 
764 	return ret;
765 }
766 
767 static int check_recv_valid(struct hns_roce_dev *hr_dev,
768 			    struct hns_roce_qp *hr_qp)
769 {
770 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
771 		return -EIO;
772 
773 	if (hr_qp->state == IB_QPS_RESET)
774 		return -EINVAL;
775 
776 	return 0;
777 }
778 
779 static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
780 				 u32 max_sge, bool rsv)
781 {
782 	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
783 	u32 i, cnt;
784 
785 	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
786 		/* Skip zero-length sge */
787 		if (!wr->sg_list[i].length)
788 			continue;
789 		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
790 		cnt++;
791 	}
792 
793 	/* Fill a reserved sge to make hw stop reading remaining segments */
794 	if (rsv) {
795 		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
796 		dseg[cnt].addr = 0;
797 		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
798 	} else {
799 		/* Clear remaining segments to make ROCEE ignore sges */
800 		if (cnt < max_sge)
801 			memset(dseg + cnt, 0,
802 			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
803 	}
804 }
805 
806 static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
807 			u32 wqe_idx, u32 max_sge)
808 {
809 	void *wqe = NULL;
810 
811 	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
812 	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
813 }
814 
815 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
816 				 const struct ib_recv_wr *wr,
817 				 const struct ib_recv_wr **bad_wr)
818 {
819 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
820 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
821 	struct ib_device *ibdev = &hr_dev->ib_dev;
822 	u32 wqe_idx, nreq, max_sge;
823 	unsigned long flags;
824 	int ret;
825 
826 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
827 
828 	ret = check_recv_valid(hr_dev, hr_qp);
829 	if (unlikely(ret)) {
830 		*bad_wr = wr;
831 		nreq = 0;
832 		goto out;
833 	}
834 
835 	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
836 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
837 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
838 						  hr_qp->ibqp.recv_cq))) {
839 			ret = -ENOMEM;
840 			*bad_wr = wr;
841 			goto out;
842 		}
843 
844 		if (unlikely(wr->num_sge > max_sge)) {
845 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
846 				  wr->num_sge, max_sge);
847 			ret = -EINVAL;
848 			*bad_wr = wr;
849 			goto out;
850 		}
851 
852 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
853 		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
854 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
855 	}
856 
857 out:
858 	if (likely(nreq)) {
859 		hr_qp->rq.head += nreq;
860 
861 		update_rq_db(hr_dev, hr_qp);
862 	}
863 	spin_unlock_irqrestore(&hr_qp->rq.lock, flags);
864 
865 	return ret;
866 }
867 
868 static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
869 {
870 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
871 }
872 
873 static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
874 {
875 	return hns_roce_buf_offset(idx_que->mtr.kmem,
876 				   n << idx_que->entry_shift);
877 }
878 
879 static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
880 {
881 	/* always called with interrupts disabled. */
882 	spin_lock(&srq->lock);
883 
884 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
885 	srq->idx_que.tail++;
886 
887 	spin_unlock(&srq->lock);
888 }
889 
890 static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
891 {
892 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
893 
894 	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
895 }
896 
897 static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
898 				const struct ib_recv_wr *wr)
899 {
900 	struct ib_device *ib_dev = srq->ibsrq.device;
901 
902 	if (unlikely(wr->num_sge > max_sge)) {
903 		ibdev_err(ib_dev,
904 			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
905 			  wr->num_sge, max_sge);
906 		return -EINVAL;
907 	}
908 
909 	if (unlikely(hns_roce_srqwq_overflow(srq))) {
910 		ibdev_err(ib_dev,
911 			  "failed to check srqwq status, srqwq is full.\n");
912 		return -ENOMEM;
913 	}
914 
915 	return 0;
916 }
917 
918 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
919 {
920 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
921 	u32 pos;
922 
923 	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
924 	if (unlikely(pos == srq->wqe_cnt))
925 		return -ENOSPC;
926 
927 	bitmap_set(idx_que->bitmap, pos, 1);
928 	*wqe_idx = pos;
929 	return 0;
930 }
931 
932 static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
933 {
934 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
935 	unsigned int head;
936 	__le32 *buf;
937 
938 	head = idx_que->head & (srq->wqe_cnt - 1);
939 
940 	buf = get_idx_buf(idx_que, head);
941 	*buf = cpu_to_le32(wqe_idx);
942 
943 	idx_que->head++;
944 }
945 
946 static void update_srq_db(struct hns_roce_srq *srq)
947 {
948 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
949 	struct hns_roce_v2_db db;
950 
951 	hr_reg_write(&db, DB_TAG, srq->srqn);
952 	hr_reg_write(&db, DB_CMD, HNS_ROCE_V2_SRQ_DB);
953 	hr_reg_write(&db, DB_PI, srq->idx_que.head);
954 
955 	hns_roce_write64(hr_dev, (__le32 *)&db, srq->db_reg);
956 }
957 
958 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
959 				     const struct ib_recv_wr *wr,
960 				     const struct ib_recv_wr **bad_wr)
961 {
962 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
963 	unsigned long flags;
964 	int ret = 0;
965 	u32 max_sge;
966 	u32 wqe_idx;
967 	void *wqe;
968 	u32 nreq;
969 
970 	spin_lock_irqsave(&srq->lock, flags);
971 
972 	max_sge = srq->max_gs - srq->rsv_sge;
973 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
974 		ret = check_post_srq_valid(srq, max_sge, wr);
975 		if (ret) {
976 			*bad_wr = wr;
977 			break;
978 		}
979 
980 		ret = get_srq_wqe_idx(srq, &wqe_idx);
981 		if (unlikely(ret)) {
982 			*bad_wr = wr;
983 			break;
984 		}
985 
986 		wqe = get_srq_wqe_buf(srq, wqe_idx);
987 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
988 		fill_wqe_idx(srq, wqe_idx);
989 		srq->wrid[wqe_idx] = wr->wr_id;
990 	}
991 
992 	if (likely(nreq)) {
993 		if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB)
994 			*srq->rdb.db_record = srq->idx_que.head &
995 					      V2_DB_PRODUCER_IDX_M;
996 		else
997 			update_srq_db(srq);
998 	}
999 
1000 	spin_unlock_irqrestore(&srq->lock, flags);
1001 
1002 	return ret;
1003 }
1004 
1005 static u32 hns_roce_v2_cmd_hw_reseted(struct hns_roce_dev *hr_dev,
1006 				      unsigned long instance_stage,
1007 				      unsigned long reset_stage)
1008 {
1009 	/* When hardware reset has been completed once or more, we should stop
1010 	 * sending mailbox&cmq&doorbell to hardware. If now in .init_instance()
1011 	 * function, we should exit with error. If now at HNAE3_INIT_CLIENT
1012 	 * stage of soft reset process, we should exit with error, and then
1013 	 * HNAE3_INIT_CLIENT related process can rollback the operation like
1014 	 * notifing hardware to free resources, HNAE3_INIT_CLIENT related
1015 	 * process will exit with error to notify NIC driver to reschedule soft
1016 	 * reset process once again.
1017 	 */
1018 	hr_dev->is_reset = true;
1019 	hr_dev->dis_db = true;
1020 
1021 	if (reset_stage == HNS_ROCE_STATE_RST_INIT ||
1022 	    instance_stage == HNS_ROCE_STATE_INIT)
1023 		return CMD_RST_PRC_EBUSY;
1024 
1025 	return CMD_RST_PRC_SUCCESS;
1026 }
1027 
1028 static u32 hns_roce_v2_cmd_hw_resetting(struct hns_roce_dev *hr_dev,
1029 					unsigned long instance_stage,
1030 					unsigned long reset_stage)
1031 {
1032 #define HW_RESET_TIMEOUT_US 1000000
1033 #define HW_RESET_SLEEP_US 1000
1034 
1035 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1036 	struct hnae3_handle *handle = priv->handle;
1037 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1038 	unsigned long val;
1039 	int ret;
1040 
1041 	/* When hardware reset is detected, we should stop sending mailbox&cmq&
1042 	 * doorbell to hardware. If now in .init_instance() function, we should
1043 	 * exit with error. If now at HNAE3_INIT_CLIENT stage of soft reset
1044 	 * process, we should exit with error, and then HNAE3_INIT_CLIENT
1045 	 * related process can rollback the operation like notifing hardware to
1046 	 * free resources, HNAE3_INIT_CLIENT related process will exit with
1047 	 * error to notify NIC driver to reschedule soft reset process once
1048 	 * again.
1049 	 */
1050 	hr_dev->dis_db = true;
1051 
1052 	ret = read_poll_timeout(ops->ae_dev_reset_cnt, val,
1053 				val > hr_dev->reset_cnt, HW_RESET_SLEEP_US,
1054 				HW_RESET_TIMEOUT_US, false, handle);
1055 	if (!ret)
1056 		hr_dev->is_reset = true;
1057 
1058 	if (!hr_dev->is_reset || reset_stage == HNS_ROCE_STATE_RST_INIT ||
1059 	    instance_stage == HNS_ROCE_STATE_INIT)
1060 		return CMD_RST_PRC_EBUSY;
1061 
1062 	return CMD_RST_PRC_SUCCESS;
1063 }
1064 
1065 static u32 hns_roce_v2_cmd_sw_resetting(struct hns_roce_dev *hr_dev)
1066 {
1067 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1068 	struct hnae3_handle *handle = priv->handle;
1069 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1070 
1071 	/* When software reset is detected at .init_instance() function, we
1072 	 * should stop sending mailbox&cmq&doorbell to hardware, and exit
1073 	 * with error.
1074 	 */
1075 	hr_dev->dis_db = true;
1076 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt)
1077 		hr_dev->is_reset = true;
1078 
1079 	return CMD_RST_PRC_EBUSY;
1080 }
1081 
1082 static u32 check_aedev_reset_status(struct hns_roce_dev *hr_dev,
1083 				    struct hnae3_handle *handle)
1084 {
1085 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1086 	unsigned long instance_stage; /* the current instance stage */
1087 	unsigned long reset_stage; /* the current reset stage */
1088 	unsigned long reset_cnt;
1089 	bool sw_resetting;
1090 	bool hw_resetting;
1091 
1092 	/* Get information about reset from NIC driver or RoCE driver itself,
1093 	 * the meaning of the following variables from NIC driver are described
1094 	 * as below:
1095 	 * reset_cnt -- The count value of completed hardware reset.
1096 	 * hw_resetting -- Whether hardware device is resetting now.
1097 	 * sw_resetting -- Whether NIC's software reset process is running now.
1098 	 */
1099 	instance_stage = handle->rinfo.instance_state;
1100 	reset_stage = handle->rinfo.reset_state;
1101 	reset_cnt = ops->ae_dev_reset_cnt(handle);
1102 	if (reset_cnt != hr_dev->reset_cnt)
1103 		return hns_roce_v2_cmd_hw_reseted(hr_dev, instance_stage,
1104 						  reset_stage);
1105 
1106 	hw_resetting = ops->get_cmdq_stat(handle);
1107 	if (hw_resetting)
1108 		return hns_roce_v2_cmd_hw_resetting(hr_dev, instance_stage,
1109 						    reset_stage);
1110 
1111 	sw_resetting = ops->ae_dev_resetting(handle);
1112 	if (sw_resetting && instance_stage == HNS_ROCE_STATE_INIT)
1113 		return hns_roce_v2_cmd_sw_resetting(hr_dev);
1114 
1115 	return CMD_RST_PRC_OTHERS;
1116 }
1117 
1118 static bool check_device_is_in_reset(struct hns_roce_dev *hr_dev)
1119 {
1120 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1121 	struct hnae3_handle *handle = priv->handle;
1122 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1123 
1124 	if (hr_dev->reset_cnt != ops->ae_dev_reset_cnt(handle))
1125 		return true;
1126 
1127 	if (ops->get_hw_reset_stat(handle))
1128 		return true;
1129 
1130 	if (ops->ae_dev_resetting(handle))
1131 		return true;
1132 
1133 	return false;
1134 }
1135 
1136 static bool v2_chk_mbox_is_avail(struct hns_roce_dev *hr_dev, bool *busy)
1137 {
1138 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1139 	u32 status;
1140 
1141 	if (hr_dev->is_reset)
1142 		status = CMD_RST_PRC_SUCCESS;
1143 	else
1144 		status = check_aedev_reset_status(hr_dev, priv->handle);
1145 
1146 	*busy = (status == CMD_RST_PRC_EBUSY);
1147 
1148 	return status == CMD_RST_PRC_OTHERS;
1149 }
1150 
1151 static int hns_roce_alloc_cmq_desc(struct hns_roce_dev *hr_dev,
1152 				   struct hns_roce_v2_cmq_ring *ring)
1153 {
1154 	int size = ring->desc_num * sizeof(struct hns_roce_cmq_desc);
1155 
1156 	ring->desc = dma_alloc_coherent(hr_dev->dev, size,
1157 					&ring->desc_dma_addr, GFP_KERNEL);
1158 	if (!ring->desc)
1159 		return -ENOMEM;
1160 
1161 	return 0;
1162 }
1163 
1164 static void hns_roce_free_cmq_desc(struct hns_roce_dev *hr_dev,
1165 				   struct hns_roce_v2_cmq_ring *ring)
1166 {
1167 	dma_free_coherent(hr_dev->dev,
1168 			  ring->desc_num * sizeof(struct hns_roce_cmq_desc),
1169 			  ring->desc, ring->desc_dma_addr);
1170 
1171 	ring->desc_dma_addr = 0;
1172 }
1173 
1174 static int init_csq(struct hns_roce_dev *hr_dev,
1175 		    struct hns_roce_v2_cmq_ring *csq)
1176 {
1177 	dma_addr_t dma;
1178 	int ret;
1179 
1180 	csq->desc_num = CMD_CSQ_DESC_NUM;
1181 	spin_lock_init(&csq->lock);
1182 	csq->flag = TYPE_CSQ;
1183 	csq->head = 0;
1184 
1185 	ret = hns_roce_alloc_cmq_desc(hr_dev, csq);
1186 	if (ret)
1187 		return ret;
1188 
1189 	dma = csq->desc_dma_addr;
1190 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_L_REG, lower_32_bits(dma));
1191 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_H_REG, upper_32_bits(dma));
1192 	roce_write(hr_dev, ROCEE_TX_CMQ_DEPTH_REG,
1193 		   (u32)csq->desc_num >> HNS_ROCE_CMQ_DESC_NUM_S);
1194 
1195 	/* Make sure to write CI first and then PI */
1196 	roce_write(hr_dev, ROCEE_TX_CMQ_CI_REG, 0);
1197 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, 0);
1198 
1199 	return 0;
1200 }
1201 
1202 static int hns_roce_v2_cmq_init(struct hns_roce_dev *hr_dev)
1203 {
1204 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1205 	int ret;
1206 
1207 	priv->cmq.tx_timeout = HNS_ROCE_CMQ_TX_TIMEOUT;
1208 
1209 	ret = init_csq(hr_dev, &priv->cmq.csq);
1210 	if (ret)
1211 		dev_err(hr_dev->dev, "failed to init CSQ, ret = %d.\n", ret);
1212 
1213 	return ret;
1214 }
1215 
1216 static void hns_roce_v2_cmq_exit(struct hns_roce_dev *hr_dev)
1217 {
1218 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1219 
1220 	hns_roce_free_cmq_desc(hr_dev, &priv->cmq.csq);
1221 }
1222 
1223 static void hns_roce_cmq_setup_basic_desc(struct hns_roce_cmq_desc *desc,
1224 					  enum hns_roce_opcode_type opcode,
1225 					  bool is_read)
1226 {
1227 	memset((void *)desc, 0, sizeof(struct hns_roce_cmq_desc));
1228 	desc->opcode = cpu_to_le16(opcode);
1229 	desc->flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1230 	if (is_read)
1231 		desc->flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_WR);
1232 	else
1233 		desc->flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1234 }
1235 
1236 static int hns_roce_cmq_csq_done(struct hns_roce_dev *hr_dev)
1237 {
1238 	u32 tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1239 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1240 
1241 	return tail == priv->cmq.csq.head;
1242 }
1243 
1244 static void update_cmdq_status(struct hns_roce_dev *hr_dev)
1245 {
1246 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1247 	struct hnae3_handle *handle = priv->handle;
1248 
1249 	if (handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
1250 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT)
1251 		hr_dev->cmd.state = HNS_ROCE_CMDQ_STATE_FATAL_ERR;
1252 }
1253 
1254 static int hns_roce_cmd_err_convert_errno(u16 desc_ret)
1255 {
1256 	struct hns_roce_cmd_errcode errcode_table[] = {
1257 		{CMD_EXEC_SUCCESS, 0},
1258 		{CMD_NO_AUTH, -EPERM},
1259 		{CMD_NOT_EXIST, -EOPNOTSUPP},
1260 		{CMD_CRQ_FULL, -EXFULL},
1261 		{CMD_NEXT_ERR, -ENOSR},
1262 		{CMD_NOT_EXEC, -ENOTBLK},
1263 		{CMD_PARA_ERR, -EINVAL},
1264 		{CMD_RESULT_ERR, -ERANGE},
1265 		{CMD_TIMEOUT, -ETIME},
1266 		{CMD_HILINK_ERR, -ENOLINK},
1267 		{CMD_INFO_ILLEGAL, -ENXIO},
1268 		{CMD_INVALID, -EBADR},
1269 	};
1270 	u16 i;
1271 
1272 	for (i = 0; i < ARRAY_SIZE(errcode_table); i++)
1273 		if (desc_ret == errcode_table[i].return_status)
1274 			return errcode_table[i].errno;
1275 	return -EIO;
1276 }
1277 
1278 static u32 hns_roce_cmdq_tx_timeout(u16 opcode, u32 tx_timeout)
1279 {
1280 	static const struct hns_roce_cmdq_tx_timeout_map cmdq_tx_timeout[] = {
1281 		{HNS_ROCE_OPC_POST_MB, HNS_ROCE_OPC_POST_MB_TIMEOUT},
1282 	};
1283 	int i;
1284 
1285 	for (i = 0; i < ARRAY_SIZE(cmdq_tx_timeout); i++)
1286 		if (cmdq_tx_timeout[i].opcode == opcode)
1287 			return cmdq_tx_timeout[i].tx_timeout;
1288 
1289 	return tx_timeout;
1290 }
1291 
1292 static void hns_roce_wait_csq_done(struct hns_roce_dev *hr_dev, u16 opcode)
1293 {
1294 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1295 	u32 tx_timeout = hns_roce_cmdq_tx_timeout(opcode, priv->cmq.tx_timeout);
1296 	u32 timeout = 0;
1297 
1298 	do {
1299 		if (hns_roce_cmq_csq_done(hr_dev))
1300 			break;
1301 		udelay(1);
1302 	} while (++timeout < tx_timeout);
1303 }
1304 
1305 static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1306 			       struct hns_roce_cmq_desc *desc, int num)
1307 {
1308 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1309 	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
1310 	u16 desc_ret;
1311 	u32 tail;
1312 	int ret;
1313 	int i;
1314 
1315 	spin_lock_bh(&csq->lock);
1316 
1317 	tail = csq->head;
1318 
1319 	for (i = 0; i < num; i++) {
1320 		csq->desc[csq->head++] = desc[i];
1321 		if (csq->head == csq->desc_num)
1322 			csq->head = 0;
1323 	}
1324 
1325 	/* Write to hardware */
1326 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, csq->head);
1327 
1328 	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_CNT]);
1329 
1330 	hns_roce_wait_csq_done(hr_dev, le16_to_cpu(desc->opcode));
1331 	if (hns_roce_cmq_csq_done(hr_dev)) {
1332 		ret = 0;
1333 		for (i = 0; i < num; i++) {
1334 			/* check the result of hardware write back */
1335 			desc[i] = csq->desc[tail++];
1336 			if (tail == csq->desc_num)
1337 				tail = 0;
1338 
1339 			desc_ret = le16_to_cpu(desc[i].retval);
1340 			if (likely(desc_ret == CMD_EXEC_SUCCESS))
1341 				continue;
1342 
1343 			dev_err_ratelimited(hr_dev->dev,
1344 					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",
1345 					    desc->opcode, desc_ret);
1346 			ret = hns_roce_cmd_err_convert_errno(desc_ret);
1347 		}
1348 	} else {
1349 		/* FW/HW reset or incorrect number of desc */
1350 		tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1351 		dev_warn(hr_dev->dev, "CMDQ move tail from %u to %u.\n",
1352 			 csq->head, tail);
1353 		csq->head = tail;
1354 
1355 		update_cmdq_status(hr_dev);
1356 
1357 		ret = -EAGAIN;
1358 	}
1359 
1360 	spin_unlock_bh(&csq->lock);
1361 
1362 	if (ret)
1363 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_ERR_CNT]);
1364 
1365 	return ret;
1366 }
1367 
1368 static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1369 			     struct hns_roce_cmq_desc *desc, int num)
1370 {
1371 	bool busy;
1372 	int ret;
1373 
1374 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1375 		return -EIO;
1376 
1377 	if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1378 		return busy ? -EBUSY : 0;
1379 
1380 	ret = __hns_roce_cmq_send(hr_dev, desc, num);
1381 	if (ret) {
1382 		if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1383 			return busy ? -EBUSY : 0;
1384 	}
1385 
1386 	return ret;
1387 }
1388 
1389 static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev,
1390 			       dma_addr_t base_addr, u8 cmd, unsigned long tag)
1391 {
1392 	struct hns_roce_cmd_mailbox *mbox;
1393 	int ret;
1394 
1395 	mbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1396 	if (IS_ERR(mbox))
1397 		return PTR_ERR(mbox);
1398 
1399 	ret = hns_roce_cmd_mbox(hr_dev, base_addr, mbox->dma, cmd, tag);
1400 	hns_roce_free_cmd_mailbox(hr_dev, mbox);
1401 	return ret;
1402 }
1403 
1404 static int hns_roce_cmq_query_hw_info(struct hns_roce_dev *hr_dev)
1405 {
1406 	struct hns_roce_query_version *resp;
1407 	struct hns_roce_cmq_desc desc;
1408 	int ret;
1409 
1410 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_HW_VER, true);
1411 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1412 	if (ret)
1413 		return ret;
1414 
1415 	resp = (struct hns_roce_query_version *)desc.data;
1416 	hr_dev->hw_rev = le16_to_cpu(resp->rocee_hw_version);
1417 	hr_dev->vendor_id = hr_dev->pci_dev->vendor;
1418 
1419 	return 0;
1420 }
1421 
1422 static void func_clr_hw_resetting_state(struct hns_roce_dev *hr_dev,
1423 					struct hnae3_handle *handle)
1424 {
1425 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1426 	unsigned long end;
1427 
1428 	hr_dev->dis_db = true;
1429 
1430 	dev_warn(hr_dev->dev,
1431 		 "func clear is pending, device in resetting state.\n");
1432 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1433 	while (end) {
1434 		if (!ops->get_hw_reset_stat(handle)) {
1435 			hr_dev->is_reset = true;
1436 			dev_info(hr_dev->dev,
1437 				 "func clear success after reset.\n");
1438 			return;
1439 		}
1440 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1441 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1442 	}
1443 
1444 	dev_warn(hr_dev->dev, "func clear failed.\n");
1445 }
1446 
1447 static void func_clr_sw_resetting_state(struct hns_roce_dev *hr_dev,
1448 					struct hnae3_handle *handle)
1449 {
1450 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1451 	unsigned long end;
1452 
1453 	hr_dev->dis_db = true;
1454 
1455 	dev_warn(hr_dev->dev,
1456 		 "func clear is pending, device in resetting state.\n");
1457 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1458 	while (end) {
1459 		if (ops->ae_dev_reset_cnt(handle) !=
1460 		    hr_dev->reset_cnt) {
1461 			hr_dev->is_reset = true;
1462 			dev_info(hr_dev->dev,
1463 				 "func clear success after sw reset\n");
1464 			return;
1465 		}
1466 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1467 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1468 	}
1469 
1470 	dev_warn(hr_dev->dev, "func clear failed because of unfinished sw reset\n");
1471 }
1472 
1473 static void hns_roce_func_clr_rst_proc(struct hns_roce_dev *hr_dev, int retval,
1474 				       int flag)
1475 {
1476 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1477 	struct hnae3_handle *handle = priv->handle;
1478 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1479 
1480 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt) {
1481 		hr_dev->dis_db = true;
1482 		hr_dev->is_reset = true;
1483 		dev_info(hr_dev->dev, "func clear success after reset.\n");
1484 		return;
1485 	}
1486 
1487 	if (ops->get_hw_reset_stat(handle)) {
1488 		func_clr_hw_resetting_state(hr_dev, handle);
1489 		return;
1490 	}
1491 
1492 	if (ops->ae_dev_resetting(handle) &&
1493 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT) {
1494 		func_clr_sw_resetting_state(hr_dev, handle);
1495 		return;
1496 	}
1497 
1498 	if (retval && !flag)
1499 		dev_warn(hr_dev->dev,
1500 			 "func clear read failed, ret = %d.\n", retval);
1501 
1502 	dev_warn(hr_dev->dev, "func clear failed.\n");
1503 }
1504 
1505 static void __hns_roce_function_clear(struct hns_roce_dev *hr_dev, int vf_id)
1506 {
1507 	bool fclr_write_fail_flag = false;
1508 	struct hns_roce_func_clear *resp;
1509 	struct hns_roce_cmq_desc desc;
1510 	unsigned long end;
1511 	int ret = 0;
1512 
1513 	if (check_device_is_in_reset(hr_dev))
1514 		goto out;
1515 
1516 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR, false);
1517 	resp = (struct hns_roce_func_clear *)desc.data;
1518 	resp->rst_funcid_en = cpu_to_le32(vf_id);
1519 
1520 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1521 	if (ret) {
1522 		fclr_write_fail_flag = true;
1523 		dev_err(hr_dev->dev, "func clear write failed, ret = %d.\n",
1524 			 ret);
1525 		goto out;
1526 	}
1527 
1528 	msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_INTERVAL);
1529 	end = HNS_ROCE_V2_FUNC_CLEAR_TIMEOUT_MSECS;
1530 	while (end) {
1531 		if (check_device_is_in_reset(hr_dev))
1532 			goto out;
1533 		msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT);
1534 		end -= HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT;
1535 
1536 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR,
1537 					      true);
1538 
1539 		resp->rst_funcid_en = cpu_to_le32(vf_id);
1540 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1541 		if (ret)
1542 			continue;
1543 
1544 		if (hr_reg_read(resp, FUNC_CLEAR_RST_FUN_DONE)) {
1545 			if (vf_id == 0)
1546 				hr_dev->is_reset = true;
1547 			return;
1548 		}
1549 	}
1550 
1551 out:
1552 	hns_roce_func_clr_rst_proc(hr_dev, ret, fclr_write_fail_flag);
1553 }
1554 
1555 static int hns_roce_free_vf_resource(struct hns_roce_dev *hr_dev, int vf_id)
1556 {
1557 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1558 	struct hns_roce_cmq_desc desc[2];
1559 	struct hns_roce_cmq_req *req_a;
1560 
1561 	req_a = (struct hns_roce_cmq_req *)desc[0].data;
1562 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1563 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1564 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1565 	hr_reg_write(req_a, FUNC_RES_A_VF_ID, vf_id);
1566 
1567 	return hns_roce_cmq_send(hr_dev, desc, 2);
1568 }
1569 
1570 static void hns_roce_function_clear(struct hns_roce_dev *hr_dev)
1571 {
1572 	int ret;
1573 	int i;
1574 
1575 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1576 		return;
1577 
1578 	for (i = hr_dev->func_num - 1; i >= 0; i--) {
1579 		__hns_roce_function_clear(hr_dev, i);
1580 
1581 		if (i == 0)
1582 			continue;
1583 
1584 		ret = hns_roce_free_vf_resource(hr_dev, i);
1585 		if (ret)
1586 			ibdev_err(&hr_dev->ib_dev,
1587 				  "failed to free vf resource, vf_id = %d, ret = %d.\n",
1588 				  i, ret);
1589 	}
1590 }
1591 
1592 static int hns_roce_clear_extdb_list_info(struct hns_roce_dev *hr_dev)
1593 {
1594 	struct hns_roce_cmq_desc desc;
1595 	int ret;
1596 
1597 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLEAR_EXTDB_LIST_INFO,
1598 				      false);
1599 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1600 	if (ret)
1601 		ibdev_err(&hr_dev->ib_dev,
1602 			  "failed to clear extended doorbell info, ret = %d.\n",
1603 			  ret);
1604 
1605 	return ret;
1606 }
1607 
1608 static int hns_roce_query_fw_ver(struct hns_roce_dev *hr_dev)
1609 {
1610 	struct hns_roce_query_fw_info *resp;
1611 	struct hns_roce_cmq_desc desc;
1612 	int ret;
1613 
1614 	hns_roce_cmq_setup_basic_desc(&desc, HNS_QUERY_FW_VER, true);
1615 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1616 	if (ret)
1617 		return ret;
1618 
1619 	resp = (struct hns_roce_query_fw_info *)desc.data;
1620 	hr_dev->caps.fw_ver = (u64)(le32_to_cpu(resp->fw_ver));
1621 
1622 	return 0;
1623 }
1624 
1625 static int hns_roce_query_func_info(struct hns_roce_dev *hr_dev)
1626 {
1627 	struct hns_roce_cmq_desc desc;
1628 	int ret;
1629 
1630 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
1631 		hr_dev->func_num = 1;
1632 		return 0;
1633 	}
1634 
1635 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_FUNC_INFO,
1636 				      true);
1637 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1638 	if (ret) {
1639 		hr_dev->func_num = 1;
1640 		return ret;
1641 	}
1642 
1643 	hr_dev->func_num = le32_to_cpu(desc.func_info.own_func_num);
1644 	hr_dev->cong_algo_tmpl_id = le32_to_cpu(desc.func_info.own_mac_id);
1645 
1646 	return 0;
1647 }
1648 
1649 static int hns_roce_hw_v2_query_counter(struct hns_roce_dev *hr_dev,
1650 					u64 *stats, u32 port, int *num_counters)
1651 {
1652 #define CNT_PER_DESC 3
1653 	struct hns_roce_cmq_desc *desc;
1654 	int bd_idx, cnt_idx;
1655 	__le64 *cnt_data;
1656 	int desc_num;
1657 	int ret;
1658 	int i;
1659 
1660 	if (port > hr_dev->caps.num_ports)
1661 		return -EINVAL;
1662 
1663 	desc_num = DIV_ROUND_UP(HNS_ROCE_HW_CNT_TOTAL, CNT_PER_DESC);
1664 	desc = kcalloc(desc_num, sizeof(*desc), GFP_KERNEL);
1665 	if (!desc)
1666 		return -ENOMEM;
1667 
1668 	for (i = 0; i < desc_num; i++) {
1669 		hns_roce_cmq_setup_basic_desc(&desc[i],
1670 					      HNS_ROCE_OPC_QUERY_COUNTER, true);
1671 		if (i != desc_num - 1)
1672 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1673 	}
1674 
1675 	ret = hns_roce_cmq_send(hr_dev, desc, desc_num);
1676 	if (ret) {
1677 		ibdev_err(&hr_dev->ib_dev,
1678 			  "failed to get counter, ret = %d.\n", ret);
1679 		goto err_out;
1680 	}
1681 
1682 	for (i = 0; i < HNS_ROCE_HW_CNT_TOTAL && i < *num_counters; i++) {
1683 		bd_idx = i / CNT_PER_DESC;
1684 		if (!(desc[bd_idx].flag & HNS_ROCE_CMD_FLAG_NEXT) &&
1685 		    bd_idx != HNS_ROCE_HW_CNT_TOTAL / CNT_PER_DESC)
1686 			break;
1687 
1688 		cnt_data = (__le64 *)&desc[bd_idx].data[0];
1689 		cnt_idx = i % CNT_PER_DESC;
1690 		stats[i] = le64_to_cpu(cnt_data[cnt_idx]);
1691 	}
1692 	*num_counters = i;
1693 
1694 err_out:
1695 	kfree(desc);
1696 	return ret;
1697 }
1698 
1699 static int hns_roce_config_global_param(struct hns_roce_dev *hr_dev)
1700 {
1701 	struct hns_roce_cmq_desc desc;
1702 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1703 	u32 clock_cycles_of_1us;
1704 
1705 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GLOBAL_PARAM,
1706 				      false);
1707 
1708 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
1709 		clock_cycles_of_1us = HNS_ROCE_1NS_CFG;
1710 	else
1711 		clock_cycles_of_1us = HNS_ROCE_1US_CFG;
1712 
1713 	hr_reg_write(req, CFG_GLOBAL_PARAM_1US_CYCLES, clock_cycles_of_1us);
1714 	hr_reg_write(req, CFG_GLOBAL_PARAM_UDP_PORT, ROCE_V2_UDP_DPORT);
1715 
1716 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1717 }
1718 
1719 static int load_func_res_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1720 {
1721 	struct hns_roce_cmq_desc desc[2];
1722 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1723 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1724 	struct hns_roce_caps *caps = &hr_dev->caps;
1725 	enum hns_roce_opcode_type opcode;
1726 	u32 func_num;
1727 	int ret;
1728 
1729 	if (is_vf) {
1730 		opcode = HNS_ROCE_OPC_QUERY_VF_RES;
1731 		func_num = 1;
1732 	} else {
1733 		opcode = HNS_ROCE_OPC_QUERY_PF_RES;
1734 		func_num = hr_dev->func_num;
1735 	}
1736 
1737 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, true);
1738 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1739 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, true);
1740 
1741 	ret = hns_roce_cmq_send(hr_dev, desc, 2);
1742 	if (ret)
1743 		return ret;
1744 
1745 	caps->qpc_bt_num = hr_reg_read(r_a, FUNC_RES_A_QPC_BT_NUM) / func_num;
1746 	caps->srqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_SRQC_BT_NUM) / func_num;
1747 	caps->cqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_CQC_BT_NUM) / func_num;
1748 	caps->mpt_bt_num = hr_reg_read(r_a, FUNC_RES_A_MPT_BT_NUM) / func_num;
1749 	caps->eqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_EQC_BT_NUM) / func_num;
1750 	caps->smac_bt_num = hr_reg_read(r_b, FUNC_RES_B_SMAC_NUM) / func_num;
1751 	caps->sgid_bt_num = hr_reg_read(r_b, FUNC_RES_B_SGID_NUM) / func_num;
1752 	caps->sccc_bt_num = hr_reg_read(r_b, FUNC_RES_B_SCCC_BT_NUM) / func_num;
1753 
1754 	if (is_vf) {
1755 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_V_QID_NUM) / func_num;
1756 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_V_GMV_BT_NUM) /
1757 					       func_num;
1758 	} else {
1759 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_B_QID_NUM) / func_num;
1760 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_B_GMV_BT_NUM) /
1761 					       func_num;
1762 	}
1763 
1764 	return 0;
1765 }
1766 
1767 static int load_pf_timer_res_caps(struct hns_roce_dev *hr_dev)
1768 {
1769 	struct hns_roce_cmq_desc desc;
1770 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1771 	struct hns_roce_caps *caps = &hr_dev->caps;
1772 	int ret;
1773 
1774 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
1775 				      true);
1776 
1777 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1778 	if (ret)
1779 		return ret;
1780 
1781 	caps->qpc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_QPC_ITEM_NUM);
1782 	caps->cqc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_CQC_ITEM_NUM);
1783 
1784 	return 0;
1785 }
1786 
1787 static int hns_roce_query_pf_resource(struct hns_roce_dev *hr_dev)
1788 {
1789 	struct device *dev = hr_dev->dev;
1790 	int ret;
1791 
1792 	ret = load_func_res_caps(hr_dev, false);
1793 	if (ret) {
1794 		dev_err(dev, "failed to load pf res caps, ret = %d.\n", ret);
1795 		return ret;
1796 	}
1797 
1798 	ret = load_pf_timer_res_caps(hr_dev);
1799 	if (ret)
1800 		dev_err(dev, "failed to load pf timer resource, ret = %d.\n",
1801 			ret);
1802 
1803 	return ret;
1804 }
1805 
1806 static int hns_roce_query_vf_resource(struct hns_roce_dev *hr_dev)
1807 {
1808 	struct device *dev = hr_dev->dev;
1809 	int ret;
1810 
1811 	ret = load_func_res_caps(hr_dev, true);
1812 	if (ret)
1813 		dev_err(dev, "failed to load vf res caps, ret = %d.\n", ret);
1814 
1815 	return ret;
1816 }
1817 
1818 static int __hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
1819 					  u32 vf_id)
1820 {
1821 	struct hns_roce_vf_switch *swt;
1822 	struct hns_roce_cmq_desc desc;
1823 	int ret;
1824 
1825 	swt = (struct hns_roce_vf_switch *)desc.data;
1826 	hns_roce_cmq_setup_basic_desc(&desc, HNS_SWITCH_PARAMETER_CFG, true);
1827 	swt->rocee_sel |= cpu_to_le32(HNS_ICL_SWITCH_CMD_ROCEE_SEL);
1828 	hr_reg_write(swt, VF_SWITCH_VF_ID, vf_id);
1829 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1830 	if (ret)
1831 		return ret;
1832 
1833 	desc.flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1834 	desc.flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1835 	hr_reg_enable(swt, VF_SWITCH_ALW_LPBK);
1836 	hr_reg_clear(swt, VF_SWITCH_ALW_LCL_LPBK);
1837 	hr_reg_enable(swt, VF_SWITCH_ALW_DST_OVRD);
1838 
1839 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1840 }
1841 
1842 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev)
1843 {
1844 	u32 vf_id;
1845 	int ret;
1846 
1847 	for (vf_id = 0; vf_id < hr_dev->func_num; vf_id++) {
1848 		ret = __hns_roce_set_vf_switch_param(hr_dev, vf_id);
1849 		if (ret)
1850 			return ret;
1851 	}
1852 	return 0;
1853 }
1854 
1855 static int config_vf_hem_resource(struct hns_roce_dev *hr_dev, int vf_id)
1856 {
1857 	struct hns_roce_cmq_desc desc[2];
1858 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1859 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1860 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1861 	struct hns_roce_caps *caps = &hr_dev->caps;
1862 
1863 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1864 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1865 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1866 
1867 	hr_reg_write(r_a, FUNC_RES_A_VF_ID, vf_id);
1868 
1869 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_NUM, caps->qpc_bt_num);
1870 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_IDX, vf_id * caps->qpc_bt_num);
1871 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_NUM, caps->srqc_bt_num);
1872 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_IDX, vf_id * caps->srqc_bt_num);
1873 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_NUM, caps->cqc_bt_num);
1874 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_IDX, vf_id * caps->cqc_bt_num);
1875 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_NUM, caps->mpt_bt_num);
1876 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_IDX, vf_id * caps->mpt_bt_num);
1877 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_NUM, caps->eqc_bt_num);
1878 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_IDX, vf_id * caps->eqc_bt_num);
1879 	hr_reg_write(r_b, FUNC_RES_V_QID_NUM, caps->sl_num);
1880 	hr_reg_write(r_b, FUNC_RES_B_QID_IDX, vf_id * caps->sl_num);
1881 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_NUM, caps->sccc_bt_num);
1882 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_IDX, vf_id * caps->sccc_bt_num);
1883 
1884 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1885 		hr_reg_write(r_b, FUNC_RES_V_GMV_BT_NUM, caps->gmv_bt_num);
1886 		hr_reg_write(r_b, FUNC_RES_B_GMV_BT_IDX,
1887 			     vf_id * caps->gmv_bt_num);
1888 	} else {
1889 		hr_reg_write(r_b, FUNC_RES_B_SGID_NUM, caps->sgid_bt_num);
1890 		hr_reg_write(r_b, FUNC_RES_B_SGID_IDX,
1891 			     vf_id * caps->sgid_bt_num);
1892 		hr_reg_write(r_b, FUNC_RES_B_SMAC_NUM, caps->smac_bt_num);
1893 		hr_reg_write(r_b, FUNC_RES_B_SMAC_IDX,
1894 			     vf_id * caps->smac_bt_num);
1895 	}
1896 
1897 	return hns_roce_cmq_send(hr_dev, desc, 2);
1898 }
1899 
1900 static int hns_roce_alloc_vf_resource(struct hns_roce_dev *hr_dev)
1901 {
1902 	u32 func_num = max_t(u32, 1, hr_dev->func_num);
1903 	u32 vf_id;
1904 	int ret;
1905 
1906 	for (vf_id = 0; vf_id < func_num; vf_id++) {
1907 		ret = config_vf_hem_resource(hr_dev, vf_id);
1908 		if (ret) {
1909 			dev_err(hr_dev->dev,
1910 				"failed to config vf-%u hem res, ret = %d.\n",
1911 				vf_id, ret);
1912 			return ret;
1913 		}
1914 	}
1915 
1916 	return 0;
1917 }
1918 
1919 static int hns_roce_v2_set_bt(struct hns_roce_dev *hr_dev)
1920 {
1921 	struct hns_roce_cmq_desc desc;
1922 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1923 	struct hns_roce_caps *caps = &hr_dev->caps;
1924 
1925 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_BT_ATTR, false);
1926 
1927 	hr_reg_write(req, CFG_BT_ATTR_QPC_BA_PGSZ,
1928 		     caps->qpc_ba_pg_sz + PG_SHIFT_OFFSET);
1929 	hr_reg_write(req, CFG_BT_ATTR_QPC_BUF_PGSZ,
1930 		     caps->qpc_buf_pg_sz + PG_SHIFT_OFFSET);
1931 	hr_reg_write(req, CFG_BT_ATTR_QPC_HOPNUM,
1932 		     to_hr_hem_hopnum(caps->qpc_hop_num, caps->num_qps));
1933 
1934 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BA_PGSZ,
1935 		     caps->srqc_ba_pg_sz + PG_SHIFT_OFFSET);
1936 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BUF_PGSZ,
1937 		     caps->srqc_buf_pg_sz + PG_SHIFT_OFFSET);
1938 	hr_reg_write(req, CFG_BT_ATTR_SRQC_HOPNUM,
1939 		     to_hr_hem_hopnum(caps->srqc_hop_num, caps->num_srqs));
1940 
1941 	hr_reg_write(req, CFG_BT_ATTR_CQC_BA_PGSZ,
1942 		     caps->cqc_ba_pg_sz + PG_SHIFT_OFFSET);
1943 	hr_reg_write(req, CFG_BT_ATTR_CQC_BUF_PGSZ,
1944 		     caps->cqc_buf_pg_sz + PG_SHIFT_OFFSET);
1945 	hr_reg_write(req, CFG_BT_ATTR_CQC_HOPNUM,
1946 		     to_hr_hem_hopnum(caps->cqc_hop_num, caps->num_cqs));
1947 
1948 	hr_reg_write(req, CFG_BT_ATTR_MPT_BA_PGSZ,
1949 		     caps->mpt_ba_pg_sz + PG_SHIFT_OFFSET);
1950 	hr_reg_write(req, CFG_BT_ATTR_MPT_BUF_PGSZ,
1951 		     caps->mpt_buf_pg_sz + PG_SHIFT_OFFSET);
1952 	hr_reg_write(req, CFG_BT_ATTR_MPT_HOPNUM,
1953 		     to_hr_hem_hopnum(caps->mpt_hop_num, caps->num_mtpts));
1954 
1955 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BA_PGSZ,
1956 		     caps->sccc_ba_pg_sz + PG_SHIFT_OFFSET);
1957 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BUF_PGSZ,
1958 		     caps->sccc_buf_pg_sz + PG_SHIFT_OFFSET);
1959 	hr_reg_write(req, CFG_BT_ATTR_SCCC_HOPNUM,
1960 		     to_hr_hem_hopnum(caps->sccc_hop_num, caps->num_qps));
1961 
1962 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1963 }
1964 
1965 static void calc_pg_sz(u32 obj_num, u32 obj_size, u32 hop_num, u32 ctx_bt_num,
1966 		       u32 *buf_page_size, u32 *bt_page_size, u32 hem_type)
1967 {
1968 	u64 obj_per_chunk;
1969 	u64 bt_chunk_size = PAGE_SIZE;
1970 	u64 buf_chunk_size = PAGE_SIZE;
1971 	u64 obj_per_chunk_default = buf_chunk_size / obj_size;
1972 
1973 	*buf_page_size = 0;
1974 	*bt_page_size = 0;
1975 
1976 	switch (hop_num) {
1977 	case 3:
1978 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1979 				(bt_chunk_size / BA_BYTE_LEN) *
1980 				(bt_chunk_size / BA_BYTE_LEN) *
1981 				 obj_per_chunk_default;
1982 		break;
1983 	case 2:
1984 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1985 				(bt_chunk_size / BA_BYTE_LEN) *
1986 				 obj_per_chunk_default;
1987 		break;
1988 	case 1:
1989 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1990 				obj_per_chunk_default;
1991 		break;
1992 	case HNS_ROCE_HOP_NUM_0:
1993 		obj_per_chunk = ctx_bt_num * obj_per_chunk_default;
1994 		break;
1995 	default:
1996 		pr_err("table %u not support hop_num = %u!\n", hem_type,
1997 		       hop_num);
1998 		return;
1999 	}
2000 
2001 	if (hem_type >= HEM_TYPE_MTT)
2002 		*bt_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2003 	else
2004 		*buf_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2005 }
2006 
2007 static void set_hem_page_size(struct hns_roce_dev *hr_dev)
2008 {
2009 	struct hns_roce_caps *caps = &hr_dev->caps;
2010 
2011 	/* EQ */
2012 	caps->eqe_ba_pg_sz = 0;
2013 	caps->eqe_buf_pg_sz = 0;
2014 
2015 	/* Link Table */
2016 	caps->llm_buf_pg_sz = 0;
2017 
2018 	/* MR */
2019 	caps->mpt_ba_pg_sz = 0;
2020 	caps->mpt_buf_pg_sz = 0;
2021 	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
2022 	caps->pbl_buf_pg_sz = 0;
2023 	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
2024 		   caps->mpt_bt_num, &caps->mpt_buf_pg_sz, &caps->mpt_ba_pg_sz,
2025 		   HEM_TYPE_MTPT);
2026 
2027 	/* QP */
2028 	caps->qpc_ba_pg_sz = 0;
2029 	caps->qpc_buf_pg_sz = 0;
2030 	caps->qpc_timer_ba_pg_sz = 0;
2031 	caps->qpc_timer_buf_pg_sz = 0;
2032 	caps->sccc_ba_pg_sz = 0;
2033 	caps->sccc_buf_pg_sz = 0;
2034 	caps->mtt_ba_pg_sz = 0;
2035 	caps->mtt_buf_pg_sz = 0;
2036 	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
2037 		   caps->qpc_bt_num, &caps->qpc_buf_pg_sz, &caps->qpc_ba_pg_sz,
2038 		   HEM_TYPE_QPC);
2039 
2040 	if (caps->flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
2041 		calc_pg_sz(caps->num_qps, caps->sccc_sz, caps->sccc_hop_num,
2042 			   caps->sccc_bt_num, &caps->sccc_buf_pg_sz,
2043 			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
2044 
2045 	/* CQ */
2046 	caps->cqc_ba_pg_sz = 0;
2047 	caps->cqc_buf_pg_sz = 0;
2048 	caps->cqc_timer_ba_pg_sz = 0;
2049 	caps->cqc_timer_buf_pg_sz = 0;
2050 	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
2051 	caps->cqe_buf_pg_sz = 0;
2052 	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
2053 		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
2054 		   HEM_TYPE_CQC);
2055 	calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
2056 		   1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
2057 
2058 	/* SRQ */
2059 	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
2060 		caps->srqc_ba_pg_sz = 0;
2061 		caps->srqc_buf_pg_sz = 0;
2062 		caps->srqwqe_ba_pg_sz = 0;
2063 		caps->srqwqe_buf_pg_sz = 0;
2064 		caps->idx_ba_pg_sz = 0;
2065 		caps->idx_buf_pg_sz = 0;
2066 		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
2067 			   caps->srqc_hop_num, caps->srqc_bt_num,
2068 			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
2069 			   HEM_TYPE_SRQC);
2070 		calc_pg_sz(caps->num_srqwqe_segs, caps->mtt_entry_sz,
2071 			   caps->srqwqe_hop_num, 1, &caps->srqwqe_buf_pg_sz,
2072 			   &caps->srqwqe_ba_pg_sz, HEM_TYPE_SRQWQE);
2073 		calc_pg_sz(caps->num_idx_segs, caps->idx_entry_sz,
2074 			   caps->idx_hop_num, 1, &caps->idx_buf_pg_sz,
2075 			   &caps->idx_ba_pg_sz, HEM_TYPE_IDX);
2076 	}
2077 
2078 	/* GMV */
2079 	caps->gmv_ba_pg_sz = 0;
2080 	caps->gmv_buf_pg_sz = 0;
2081 }
2082 
2083 /* Apply all loaded caps before setting to hardware */
2084 static void apply_func_caps(struct hns_roce_dev *hr_dev)
2085 {
2086 #define MAX_GID_TBL_LEN 256
2087 	struct hns_roce_caps *caps = &hr_dev->caps;
2088 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2089 
2090 	/* The following configurations don't need to be got from firmware. */
2091 	caps->qpc_timer_entry_sz = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
2092 	caps->cqc_timer_entry_sz = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
2093 	caps->mtt_entry_sz = HNS_ROCE_V2_MTT_ENTRY_SZ;
2094 
2095 	caps->pbl_hop_num = HNS_ROCE_PBL_HOP_NUM;
2096 	caps->qpc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2097 	caps->cqc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2098 
2099 	caps->num_srqwqe_segs = HNS_ROCE_V2_MAX_SRQWQE_SEGS;
2100 	caps->num_idx_segs = HNS_ROCE_V2_MAX_IDX_SEGS;
2101 
2102 	if (!caps->num_comp_vectors)
2103 		caps->num_comp_vectors =
2104 			min_t(u32, caps->eqc_bt_num - HNS_ROCE_V2_AEQE_VEC_NUM,
2105 				(u32)priv->handle->rinfo.num_vectors -
2106 		(HNS_ROCE_V2_AEQE_VEC_NUM + HNS_ROCE_V2_ABNORMAL_VEC_NUM));
2107 
2108 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2109 		caps->eqe_hop_num = HNS_ROCE_V3_EQE_HOP_NUM;
2110 		caps->ceqe_size = HNS_ROCE_V3_EQE_SIZE;
2111 		caps->aeqe_size = HNS_ROCE_V3_EQE_SIZE;
2112 
2113 		/* The following configurations will be overwritten */
2114 		caps->qpc_sz = HNS_ROCE_V3_QPC_SZ;
2115 		caps->cqe_sz = HNS_ROCE_V3_CQE_SIZE;
2116 		caps->sccc_sz = HNS_ROCE_V3_SCCC_SZ;
2117 
2118 		/* The following configurations are not got from firmware */
2119 		caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;
2120 
2121 		caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0;
2122 
2123 		/* It's meaningless to support excessively large gid_table_len,
2124 		 * as the type of sgid_index in kernel struct ib_global_route
2125 		 * and userspace struct ibv_global_route are u8/uint8_t (0-255).
2126 		 */
2127 		caps->gid_table_len[0] = min_t(u32, MAX_GID_TBL_LEN,
2128 					 caps->gmv_bt_num *
2129 					 (HNS_HW_PAGE_SIZE / caps->gmv_entry_sz));
2130 
2131 		caps->gmv_entry_num = caps->gmv_bt_num * (HNS_HW_PAGE_SIZE /
2132 							  caps->gmv_entry_sz);
2133 	} else {
2134 		u32 func_num = max_t(u32, 1, hr_dev->func_num);
2135 
2136 		caps->eqe_hop_num = HNS_ROCE_V2_EQE_HOP_NUM;
2137 		caps->ceqe_size = HNS_ROCE_CEQE_SIZE;
2138 		caps->aeqe_size = HNS_ROCE_AEQE_SIZE;
2139 		caps->gid_table_len[0] /= func_num;
2140 	}
2141 
2142 	if (hr_dev->is_vf) {
2143 		caps->default_aeq_arm_st = 0x3;
2144 		caps->default_ceq_arm_st = 0x3;
2145 		caps->default_ceq_max_cnt = 0x1;
2146 		caps->default_ceq_period = 0x10;
2147 		caps->default_aeq_max_cnt = 0x1;
2148 		caps->default_aeq_period = 0x10;
2149 	}
2150 
2151 	set_hem_page_size(hr_dev);
2152 }
2153 
2154 static int hns_roce_query_caps(struct hns_roce_dev *hr_dev)
2155 {
2156 	struct hns_roce_cmq_desc desc[HNS_ROCE_QUERY_PF_CAPS_CMD_NUM];
2157 	struct hns_roce_caps *caps = &hr_dev->caps;
2158 	struct hns_roce_query_pf_caps_a *resp_a;
2159 	struct hns_roce_query_pf_caps_b *resp_b;
2160 	struct hns_roce_query_pf_caps_c *resp_c;
2161 	struct hns_roce_query_pf_caps_d *resp_d;
2162 	struct hns_roce_query_pf_caps_e *resp_e;
2163 	enum hns_roce_opcode_type cmd;
2164 	int ctx_hop_num;
2165 	int pbl_hop_num;
2166 	int ret;
2167 	int i;
2168 
2169 	cmd = hr_dev->is_vf ? HNS_ROCE_OPC_QUERY_VF_CAPS_NUM :
2170 	      HNS_ROCE_OPC_QUERY_PF_CAPS_NUM;
2171 
2172 	for (i = 0; i < HNS_ROCE_QUERY_PF_CAPS_CMD_NUM; i++) {
2173 		hns_roce_cmq_setup_basic_desc(&desc[i], cmd, true);
2174 		if (i < (HNS_ROCE_QUERY_PF_CAPS_CMD_NUM - 1))
2175 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2176 		else
2177 			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2178 	}
2179 
2180 	ret = hns_roce_cmq_send(hr_dev, desc, HNS_ROCE_QUERY_PF_CAPS_CMD_NUM);
2181 	if (ret)
2182 		return ret;
2183 
2184 	resp_a = (struct hns_roce_query_pf_caps_a *)desc[0].data;
2185 	resp_b = (struct hns_roce_query_pf_caps_b *)desc[1].data;
2186 	resp_c = (struct hns_roce_query_pf_caps_c *)desc[2].data;
2187 	resp_d = (struct hns_roce_query_pf_caps_d *)desc[3].data;
2188 	resp_e = (struct hns_roce_query_pf_caps_e *)desc[4].data;
2189 
2190 	caps->local_ca_ack_delay = resp_a->local_ca_ack_delay;
2191 	caps->max_sq_sg = le16_to_cpu(resp_a->max_sq_sg);
2192 	caps->max_sq_inline = le16_to_cpu(resp_a->max_sq_inline);
2193 	caps->max_rq_sg = le16_to_cpu(resp_a->max_rq_sg);
2194 	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
2195 	caps->max_srq_sges = le16_to_cpu(resp_a->max_srq_sges);
2196 	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
2197 	caps->num_aeq_vectors = resp_a->num_aeq_vectors;
2198 	caps->num_other_vectors = resp_a->num_other_vectors;
2199 	caps->max_sq_desc_sz = resp_a->max_sq_desc_sz;
2200 	caps->max_rq_desc_sz = resp_a->max_rq_desc_sz;
2201 
2202 	caps->mtpt_entry_sz = resp_b->mtpt_entry_sz;
2203 	caps->irrl_entry_sz = resp_b->irrl_entry_sz;
2204 	caps->trrl_entry_sz = resp_b->trrl_entry_sz;
2205 	caps->cqc_entry_sz = resp_b->cqc_entry_sz;
2206 	caps->srqc_entry_sz = resp_b->srqc_entry_sz;
2207 	caps->idx_entry_sz = resp_b->idx_entry_sz;
2208 	caps->sccc_sz = resp_b->sccc_sz;
2209 	caps->max_mtu = resp_b->max_mtu;
2210 	caps->min_cqes = resp_b->min_cqes;
2211 	caps->min_wqes = resp_b->min_wqes;
2212 	caps->page_size_cap = le32_to_cpu(resp_b->page_size_cap);
2213 	caps->pkey_table_len[0] = resp_b->pkey_table_len;
2214 	caps->phy_num_uars = resp_b->phy_num_uars;
2215 	ctx_hop_num = resp_b->ctx_hop_num;
2216 	pbl_hop_num = resp_b->pbl_hop_num;
2217 
2218 	caps->num_pds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_PDS);
2219 
2220 	caps->flags = hr_reg_read(resp_c, PF_CAPS_C_CAP_FLAGS);
2221 	caps->flags |= le16_to_cpu(resp_d->cap_flags_ex) <<
2222 		       HNS_ROCE_CAP_FLAGS_EX_SHIFT;
2223 
2224 	caps->num_cqs = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_CQS);
2225 	caps->gid_table_len[0] = hr_reg_read(resp_c, PF_CAPS_C_MAX_GID);
2226 	caps->max_cqes = 1 << hr_reg_read(resp_c, PF_CAPS_C_CQ_DEPTH);
2227 	caps->num_xrcds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_XRCDS);
2228 	caps->num_mtpts = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_MRWS);
2229 	caps->num_qps = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_QPS);
2230 	caps->max_qp_init_rdma = hr_reg_read(resp_c, PF_CAPS_C_MAX_ORD);
2231 	caps->max_qp_dest_rdma = caps->max_qp_init_rdma;
2232 	caps->max_wqes = 1 << le16_to_cpu(resp_c->sq_depth);
2233 
2234 	caps->num_srqs = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_SRQS);
2235 	caps->cong_cap = hr_reg_read(resp_d, PF_CAPS_D_CONG_CAP);
2236 	caps->max_srq_wrs = 1 << le16_to_cpu(resp_d->srq_depth);
2237 	caps->ceqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_CEQ_DEPTH);
2238 	caps->num_comp_vectors = hr_reg_read(resp_d, PF_CAPS_D_NUM_CEQS);
2239 	caps->aeqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_AEQ_DEPTH);
2240 	caps->default_cong_type = hr_reg_read(resp_d, PF_CAPS_D_DEFAULT_ALG);
2241 	caps->reserved_pds = hr_reg_read(resp_d, PF_CAPS_D_RSV_PDS);
2242 	caps->num_uars = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_UARS);
2243 	caps->reserved_qps = hr_reg_read(resp_d, PF_CAPS_D_RSV_QPS);
2244 	caps->reserved_uars = hr_reg_read(resp_d, PF_CAPS_D_RSV_UARS);
2245 
2246 	caps->reserved_mrws = hr_reg_read(resp_e, PF_CAPS_E_RSV_MRWS);
2247 	caps->chunk_sz = 1 << hr_reg_read(resp_e, PF_CAPS_E_CHUNK_SIZE_SHIFT);
2248 	caps->reserved_cqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_CQS);
2249 	caps->reserved_xrcds = hr_reg_read(resp_e, PF_CAPS_E_RSV_XRCDS);
2250 	caps->reserved_srqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_SRQS);
2251 	caps->reserved_lkey = hr_reg_read(resp_e, PF_CAPS_E_RSV_LKEYS);
2252 
2253 	caps->qpc_hop_num = ctx_hop_num;
2254 	caps->sccc_hop_num = ctx_hop_num;
2255 	caps->srqc_hop_num = ctx_hop_num;
2256 	caps->cqc_hop_num = ctx_hop_num;
2257 	caps->mpt_hop_num = ctx_hop_num;
2258 	caps->mtt_hop_num = pbl_hop_num;
2259 	caps->cqe_hop_num = pbl_hop_num;
2260 	caps->srqwqe_hop_num = pbl_hop_num;
2261 	caps->idx_hop_num = pbl_hop_num;
2262 	caps->wqe_sq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_SQWQE_HOP_NUM);
2263 	caps->wqe_sge_hop_num = hr_reg_read(resp_d, PF_CAPS_D_EX_SGE_HOP_NUM);
2264 	caps->wqe_rq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_RQWQE_HOP_NUM);
2265 
2266 	if (!(caps->page_size_cap & PAGE_SIZE))
2267 		caps->page_size_cap = HNS_ROCE_V2_PAGE_SIZE_SUPPORTED;
2268 
2269 	if (!hr_dev->is_vf) {
2270 		caps->cqe_sz = resp_a->cqe_sz;
2271 		caps->qpc_sz = le16_to_cpu(resp_b->qpc_sz);
2272 		caps->default_aeq_arm_st =
2273 				hr_reg_read(resp_d, PF_CAPS_D_AEQ_ARM_ST);
2274 		caps->default_ceq_arm_st =
2275 				hr_reg_read(resp_d, PF_CAPS_D_CEQ_ARM_ST);
2276 		caps->default_ceq_max_cnt = le16_to_cpu(resp_e->ceq_max_cnt);
2277 		caps->default_ceq_period = le16_to_cpu(resp_e->ceq_period);
2278 		caps->default_aeq_max_cnt = le16_to_cpu(resp_e->aeq_max_cnt);
2279 		caps->default_aeq_period = le16_to_cpu(resp_e->aeq_period);
2280 	}
2281 
2282 	return 0;
2283 }
2284 
2285 static int config_hem_entry_size(struct hns_roce_dev *hr_dev, u32 type, u32 val)
2286 {
2287 	struct hns_roce_cmq_desc desc;
2288 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
2289 
2290 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_ENTRY_SIZE,
2291 				      false);
2292 
2293 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_TYPE, type);
2294 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_VALUE, val);
2295 
2296 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2297 }
2298 
2299 static int hns_roce_config_entry_size(struct hns_roce_dev *hr_dev)
2300 {
2301 	struct hns_roce_caps *caps = &hr_dev->caps;
2302 	int ret;
2303 
2304 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2305 		return 0;
2306 
2307 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_QPC_SIZE,
2308 				    caps->qpc_sz);
2309 	if (ret) {
2310 		dev_err(hr_dev->dev, "failed to cfg qpc sz, ret = %d.\n", ret);
2311 		return ret;
2312 	}
2313 
2314 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_SCCC_SIZE,
2315 				    caps->sccc_sz);
2316 	if (ret)
2317 		dev_err(hr_dev->dev, "failed to cfg sccc sz, ret = %d.\n", ret);
2318 
2319 	return ret;
2320 }
2321 
2322 static int hns_roce_v2_vf_profile(struct hns_roce_dev *hr_dev)
2323 {
2324 	struct device *dev = hr_dev->dev;
2325 	int ret;
2326 
2327 	hr_dev->func_num = 1;
2328 
2329 	ret = hns_roce_query_caps(hr_dev);
2330 	if (ret) {
2331 		dev_err(dev, "failed to query VF caps, ret = %d.\n", ret);
2332 		return ret;
2333 	}
2334 
2335 	ret = hns_roce_query_vf_resource(hr_dev);
2336 	if (ret) {
2337 		dev_err(dev, "failed to query VF resource, ret = %d.\n", ret);
2338 		return ret;
2339 	}
2340 
2341 	apply_func_caps(hr_dev);
2342 
2343 	ret = hns_roce_v2_set_bt(hr_dev);
2344 	if (ret)
2345 		dev_err(dev, "failed to config VF BA table, ret = %d.\n", ret);
2346 
2347 	return ret;
2348 }
2349 
2350 static int hns_roce_v2_pf_profile(struct hns_roce_dev *hr_dev)
2351 {
2352 	struct device *dev = hr_dev->dev;
2353 	int ret;
2354 
2355 	ret = hns_roce_query_func_info(hr_dev);
2356 	if (ret) {
2357 		dev_err(dev, "failed to query func info, ret = %d.\n", ret);
2358 		return ret;
2359 	}
2360 
2361 	ret = hns_roce_config_global_param(hr_dev);
2362 	if (ret) {
2363 		dev_err(dev, "failed to config global param, ret = %d.\n", ret);
2364 		return ret;
2365 	}
2366 
2367 	ret = hns_roce_set_vf_switch_param(hr_dev);
2368 	if (ret) {
2369 		dev_err(dev, "failed to set switch param, ret = %d.\n", ret);
2370 		return ret;
2371 	}
2372 
2373 	ret = hns_roce_query_caps(hr_dev);
2374 	if (ret) {
2375 		dev_err(dev, "failed to query PF caps, ret = %d.\n", ret);
2376 		return ret;
2377 	}
2378 
2379 	ret = hns_roce_query_pf_resource(hr_dev);
2380 	if (ret) {
2381 		dev_err(dev, "failed to query pf resource, ret = %d.\n", ret);
2382 		return ret;
2383 	}
2384 
2385 	apply_func_caps(hr_dev);
2386 
2387 	ret = hns_roce_alloc_vf_resource(hr_dev);
2388 	if (ret) {
2389 		dev_err(dev, "failed to alloc vf resource, ret = %d.\n", ret);
2390 		return ret;
2391 	}
2392 
2393 	ret = hns_roce_v2_set_bt(hr_dev);
2394 	if (ret) {
2395 		dev_err(dev, "failed to config BA table, ret = %d.\n", ret);
2396 		return ret;
2397 	}
2398 
2399 	/* Configure the size of QPC, SCCC, etc. */
2400 	return hns_roce_config_entry_size(hr_dev);
2401 }
2402 
2403 static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
2404 {
2405 	struct device *dev = hr_dev->dev;
2406 	int ret;
2407 
2408 	ret = hns_roce_cmq_query_hw_info(hr_dev);
2409 	if (ret) {
2410 		dev_err(dev, "failed to query hardware info, ret = %d.\n", ret);
2411 		return ret;
2412 	}
2413 
2414 	ret = hns_roce_query_fw_ver(hr_dev);
2415 	if (ret) {
2416 		dev_err(dev, "failed to query firmware info, ret = %d.\n", ret);
2417 		return ret;
2418 	}
2419 
2420 	hr_dev->vendor_part_id = hr_dev->pci_dev->device;
2421 	hr_dev->sys_image_guid = be64_to_cpu(hr_dev->ib_dev.node_guid);
2422 
2423 	if (hr_dev->is_vf)
2424 		return hns_roce_v2_vf_profile(hr_dev);
2425 	else
2426 		return hns_roce_v2_pf_profile(hr_dev);
2427 }
2428 
2429 static void config_llm_table(struct hns_roce_buf *data_buf, void *cfg_buf)
2430 {
2431 	u32 i, next_ptr, page_num;
2432 	__le64 *entry = cfg_buf;
2433 	dma_addr_t addr;
2434 	u64 val;
2435 
2436 	page_num = data_buf->npages;
2437 	for (i = 0; i < page_num; i++) {
2438 		addr = hns_roce_buf_page(data_buf, i);
2439 		if (i == (page_num - 1))
2440 			next_ptr = 0;
2441 		else
2442 			next_ptr = i + 1;
2443 
2444 		val = HNS_ROCE_EXT_LLM_ENTRY(addr, (u64)next_ptr);
2445 		entry[i] = cpu_to_le64(val);
2446 	}
2447 }
2448 
2449 static int set_llm_cfg_to_hw(struct hns_roce_dev *hr_dev,
2450 			     struct hns_roce_link_table *table)
2451 {
2452 	struct hns_roce_cmq_desc desc[2];
2453 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
2454 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
2455 	struct hns_roce_buf *buf = table->buf;
2456 	enum hns_roce_opcode_type opcode;
2457 	dma_addr_t addr;
2458 
2459 	opcode = HNS_ROCE_OPC_CFG_EXT_LLM;
2460 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
2461 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2462 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
2463 
2464 	hr_reg_write(r_a, CFG_LLM_A_BA_L, lower_32_bits(table->table.map));
2465 	hr_reg_write(r_a, CFG_LLM_A_BA_H, upper_32_bits(table->table.map));
2466 	hr_reg_write(r_a, CFG_LLM_A_DEPTH, buf->npages);
2467 	hr_reg_write(r_a, CFG_LLM_A_PGSZ, to_hr_hw_page_shift(buf->page_shift));
2468 	hr_reg_enable(r_a, CFG_LLM_A_INIT_EN);
2469 
2470 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, 0));
2471 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_L, lower_32_bits(addr));
2472 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_H, upper_32_bits(addr));
2473 	hr_reg_write(r_a, CFG_LLM_A_HEAD_NXTPTR, 1);
2474 	hr_reg_write(r_a, CFG_LLM_A_HEAD_PTR, 0);
2475 
2476 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, buf->npages - 1));
2477 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_L, lower_32_bits(addr));
2478 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_H, upper_32_bits(addr));
2479 	hr_reg_write(r_b, CFG_LLM_B_TAIL_PTR, buf->npages - 1);
2480 
2481 	return hns_roce_cmq_send(hr_dev, desc, 2);
2482 }
2483 
2484 static struct hns_roce_link_table *
2485 alloc_link_table_buf(struct hns_roce_dev *hr_dev)
2486 {
2487 	u16 total_sl = hr_dev->caps.sl_num * hr_dev->func_num;
2488 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2489 	struct hns_roce_link_table *link_tbl;
2490 	u32 pg_shift, size, min_size;
2491 
2492 	link_tbl = &priv->ext_llm;
2493 	pg_shift = hr_dev->caps.llm_buf_pg_sz + PAGE_SHIFT;
2494 	size = hr_dev->caps.num_qps * hr_dev->func_num *
2495 	       HNS_ROCE_V2_EXT_LLM_ENTRY_SZ;
2496 	min_size = HNS_ROCE_EXT_LLM_MIN_PAGES(total_sl) << pg_shift;
2497 
2498 	/* Alloc data table */
2499 	size = max(size, min_size);
2500 	link_tbl->buf = hns_roce_buf_alloc(hr_dev, size, pg_shift, 0);
2501 	if (IS_ERR(link_tbl->buf))
2502 		return ERR_PTR(-ENOMEM);
2503 
2504 	/* Alloc config table */
2505 	size = link_tbl->buf->npages * sizeof(u64);
2506 	link_tbl->table.buf = dma_alloc_coherent(hr_dev->dev, size,
2507 						 &link_tbl->table.map,
2508 						 GFP_KERNEL);
2509 	if (!link_tbl->table.buf) {
2510 		hns_roce_buf_free(hr_dev, link_tbl->buf);
2511 		return ERR_PTR(-ENOMEM);
2512 	}
2513 
2514 	return link_tbl;
2515 }
2516 
2517 static void free_link_table_buf(struct hns_roce_dev *hr_dev,
2518 				struct hns_roce_link_table *tbl)
2519 {
2520 	if (tbl->buf) {
2521 		u32 size = tbl->buf->npages * sizeof(u64);
2522 
2523 		dma_free_coherent(hr_dev->dev, size, tbl->table.buf,
2524 				  tbl->table.map);
2525 	}
2526 
2527 	hns_roce_buf_free(hr_dev, tbl->buf);
2528 }
2529 
2530 static int hns_roce_init_link_table(struct hns_roce_dev *hr_dev)
2531 {
2532 	struct hns_roce_link_table *link_tbl;
2533 	int ret;
2534 
2535 	link_tbl = alloc_link_table_buf(hr_dev);
2536 	if (IS_ERR(link_tbl))
2537 		return -ENOMEM;
2538 
2539 	if (WARN_ON(link_tbl->buf->npages > HNS_ROCE_V2_EXT_LLM_MAX_DEPTH)) {
2540 		ret = -EINVAL;
2541 		goto err_alloc;
2542 	}
2543 
2544 	config_llm_table(link_tbl->buf, link_tbl->table.buf);
2545 	ret = set_llm_cfg_to_hw(hr_dev, link_tbl);
2546 	if (ret)
2547 		goto err_alloc;
2548 
2549 	return 0;
2550 
2551 err_alloc:
2552 	free_link_table_buf(hr_dev, link_tbl);
2553 	return ret;
2554 }
2555 
2556 static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev)
2557 {
2558 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2559 
2560 	free_link_table_buf(hr_dev, &priv->ext_llm);
2561 }
2562 
2563 static void free_dip_list(struct hns_roce_dev *hr_dev)
2564 {
2565 	struct hns_roce_dip *hr_dip;
2566 	struct hns_roce_dip *tmp;
2567 	unsigned long flags;
2568 
2569 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
2570 
2571 	list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) {
2572 		list_del(&hr_dip->node);
2573 		kfree(hr_dip);
2574 	}
2575 
2576 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
2577 }
2578 
2579 static struct ib_pd *free_mr_init_pd(struct hns_roce_dev *hr_dev)
2580 {
2581 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2582 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2583 	struct ib_device *ibdev = &hr_dev->ib_dev;
2584 	struct hns_roce_pd *hr_pd;
2585 	struct ib_pd *pd;
2586 
2587 	hr_pd = kzalloc(sizeof(*hr_pd), GFP_KERNEL);
2588 	if (ZERO_OR_NULL_PTR(hr_pd))
2589 		return NULL;
2590 	pd = &hr_pd->ibpd;
2591 	pd->device = ibdev;
2592 
2593 	if (hns_roce_alloc_pd(pd, NULL)) {
2594 		ibdev_err(ibdev, "failed to create pd for free mr.\n");
2595 		kfree(hr_pd);
2596 		return NULL;
2597 	}
2598 	free_mr->rsv_pd = to_hr_pd(pd);
2599 	free_mr->rsv_pd->ibpd.device = &hr_dev->ib_dev;
2600 	free_mr->rsv_pd->ibpd.uobject = NULL;
2601 	free_mr->rsv_pd->ibpd.__internal_mr = NULL;
2602 	atomic_set(&free_mr->rsv_pd->ibpd.usecnt, 0);
2603 
2604 	return pd;
2605 }
2606 
2607 static struct ib_cq *free_mr_init_cq(struct hns_roce_dev *hr_dev)
2608 {
2609 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2610 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2611 	struct ib_device *ibdev = &hr_dev->ib_dev;
2612 	struct ib_cq_init_attr cq_init_attr = {};
2613 	struct hns_roce_cq *hr_cq;
2614 	struct ib_cq *cq;
2615 
2616 	cq_init_attr.cqe = HNS_ROCE_FREE_MR_USED_CQE_NUM;
2617 
2618 	hr_cq = kzalloc(sizeof(*hr_cq), GFP_KERNEL);
2619 	if (ZERO_OR_NULL_PTR(hr_cq))
2620 		return NULL;
2621 
2622 	cq = &hr_cq->ib_cq;
2623 	cq->device = ibdev;
2624 
2625 	if (hns_roce_create_cq(cq, &cq_init_attr, NULL)) {
2626 		ibdev_err(ibdev, "failed to create cq for free mr.\n");
2627 		kfree(hr_cq);
2628 		return NULL;
2629 	}
2630 	free_mr->rsv_cq = to_hr_cq(cq);
2631 	free_mr->rsv_cq->ib_cq.device = &hr_dev->ib_dev;
2632 	free_mr->rsv_cq->ib_cq.uobject = NULL;
2633 	free_mr->rsv_cq->ib_cq.comp_handler = NULL;
2634 	free_mr->rsv_cq->ib_cq.event_handler = NULL;
2635 	free_mr->rsv_cq->ib_cq.cq_context = NULL;
2636 	atomic_set(&free_mr->rsv_cq->ib_cq.usecnt, 0);
2637 
2638 	return cq;
2639 }
2640 
2641 static int free_mr_init_qp(struct hns_roce_dev *hr_dev, struct ib_cq *cq,
2642 			   struct ib_qp_init_attr *init_attr, int i)
2643 {
2644 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2645 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2646 	struct ib_device *ibdev = &hr_dev->ib_dev;
2647 	struct hns_roce_qp *hr_qp;
2648 	struct ib_qp *qp;
2649 	int ret;
2650 
2651 	hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
2652 	if (ZERO_OR_NULL_PTR(hr_qp))
2653 		return -ENOMEM;
2654 
2655 	qp = &hr_qp->ibqp;
2656 	qp->device = ibdev;
2657 
2658 	ret = hns_roce_create_qp(qp, init_attr, NULL);
2659 	if (ret) {
2660 		ibdev_err(ibdev, "failed to create qp for free mr.\n");
2661 		kfree(hr_qp);
2662 		return ret;
2663 	}
2664 
2665 	free_mr->rsv_qp[i] = hr_qp;
2666 	free_mr->rsv_qp[i]->ibqp.recv_cq = cq;
2667 	free_mr->rsv_qp[i]->ibqp.send_cq = cq;
2668 
2669 	return 0;
2670 }
2671 
2672 static void free_mr_exit(struct hns_roce_dev *hr_dev)
2673 {
2674 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2675 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2676 	struct ib_qp *qp;
2677 	int i;
2678 
2679 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2680 		if (free_mr->rsv_qp[i]) {
2681 			qp = &free_mr->rsv_qp[i]->ibqp;
2682 			hns_roce_v2_destroy_qp(qp, NULL);
2683 			kfree(free_mr->rsv_qp[i]);
2684 			free_mr->rsv_qp[i] = NULL;
2685 		}
2686 	}
2687 
2688 	if (free_mr->rsv_cq) {
2689 		hns_roce_destroy_cq(&free_mr->rsv_cq->ib_cq, NULL);
2690 		kfree(free_mr->rsv_cq);
2691 		free_mr->rsv_cq = NULL;
2692 	}
2693 
2694 	if (free_mr->rsv_pd) {
2695 		hns_roce_dealloc_pd(&free_mr->rsv_pd->ibpd, NULL);
2696 		kfree(free_mr->rsv_pd);
2697 		free_mr->rsv_pd = NULL;
2698 	}
2699 
2700 	mutex_destroy(&free_mr->mutex);
2701 }
2702 
2703 static int free_mr_alloc_res(struct hns_roce_dev *hr_dev)
2704 {
2705 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2706 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2707 	struct ib_qp_init_attr qp_init_attr = {};
2708 	struct ib_pd *pd;
2709 	struct ib_cq *cq;
2710 	int ret;
2711 	int i;
2712 
2713 	pd = free_mr_init_pd(hr_dev);
2714 	if (!pd)
2715 		return -ENOMEM;
2716 
2717 	cq = free_mr_init_cq(hr_dev);
2718 	if (!cq) {
2719 		ret = -ENOMEM;
2720 		goto create_failed_cq;
2721 	}
2722 
2723 	qp_init_attr.qp_type = IB_QPT_RC;
2724 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2725 	qp_init_attr.send_cq = cq;
2726 	qp_init_attr.recv_cq = cq;
2727 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2728 		qp_init_attr.cap.max_send_wr = HNS_ROCE_FREE_MR_USED_SQWQE_NUM;
2729 		qp_init_attr.cap.max_send_sge = HNS_ROCE_FREE_MR_USED_SQSGE_NUM;
2730 		qp_init_attr.cap.max_recv_wr = HNS_ROCE_FREE_MR_USED_RQWQE_NUM;
2731 		qp_init_attr.cap.max_recv_sge = HNS_ROCE_FREE_MR_USED_RQSGE_NUM;
2732 
2733 		ret = free_mr_init_qp(hr_dev, cq, &qp_init_attr, i);
2734 		if (ret)
2735 			goto create_failed_qp;
2736 	}
2737 
2738 	return 0;
2739 
2740 create_failed_qp:
2741 	for (i--; i >= 0; i--) {
2742 		hns_roce_v2_destroy_qp(&free_mr->rsv_qp[i]->ibqp, NULL);
2743 		kfree(free_mr->rsv_qp[i]);
2744 	}
2745 	hns_roce_destroy_cq(cq, NULL);
2746 	kfree(cq);
2747 
2748 create_failed_cq:
2749 	hns_roce_dealloc_pd(pd, NULL);
2750 	kfree(pd);
2751 
2752 	return ret;
2753 }
2754 
2755 static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
2756 				 struct ib_qp_attr *attr, int sl_num)
2757 {
2758 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2759 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2760 	struct ib_device *ibdev = &hr_dev->ib_dev;
2761 	struct hns_roce_qp *hr_qp;
2762 	int loopback;
2763 	int mask;
2764 	int ret;
2765 
2766 	hr_qp = to_hr_qp(&free_mr->rsv_qp[sl_num]->ibqp);
2767 	hr_qp->free_mr_en = 1;
2768 	hr_qp->ibqp.device = ibdev;
2769 	hr_qp->ibqp.qp_type = IB_QPT_RC;
2770 
2771 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_ACCESS_FLAGS;
2772 	attr->qp_state = IB_QPS_INIT;
2773 	attr->port_num = 1;
2774 	attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
2775 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2776 				    IB_QPS_INIT, NULL);
2777 	if (ret) {
2778 		ibdev_err(ibdev, "failed to modify qp to init, ret = %d.\n",
2779 			  ret);
2780 		return ret;
2781 	}
2782 
2783 	loopback = hr_dev->loop_idc;
2784 	/* Set qpc lbi = 1 incidate loopback IO */
2785 	hr_dev->loop_idc = 1;
2786 
2787 	mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
2788 	       IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
2789 	attr->qp_state = IB_QPS_RTR;
2790 	attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2791 	attr->path_mtu = IB_MTU_256;
2792 	attr->dest_qp_num = hr_qp->qpn;
2793 	attr->rq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2794 
2795 	rdma_ah_set_sl(&attr->ah_attr, (u8)sl_num);
2796 
2797 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2798 				    IB_QPS_RTR, NULL);
2799 	hr_dev->loop_idc = loopback;
2800 	if (ret) {
2801 		ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
2802 			  ret);
2803 		return ret;
2804 	}
2805 
2806 	mask = IB_QP_STATE | IB_QP_SQ_PSN | IB_QP_RETRY_CNT | IB_QP_TIMEOUT |
2807 	       IB_QP_RNR_RETRY | IB_QP_MAX_QP_RD_ATOMIC;
2808 	attr->qp_state = IB_QPS_RTS;
2809 	attr->sq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2810 	attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
2811 	attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
2812 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_RTR,
2813 				    IB_QPS_RTS, NULL);
2814 	if (ret)
2815 		ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
2816 			  ret);
2817 
2818 	return ret;
2819 }
2820 
2821 static int free_mr_modify_qp(struct hns_roce_dev *hr_dev)
2822 {
2823 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2824 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2825 	struct ib_qp_attr attr = {};
2826 	int ret;
2827 	int i;
2828 
2829 	rdma_ah_set_grh(&attr.ah_attr, NULL, 0, 0, 1, 0);
2830 	rdma_ah_set_static_rate(&attr.ah_attr, 3);
2831 	rdma_ah_set_port_num(&attr.ah_attr, 1);
2832 
2833 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2834 		ret = free_mr_modify_rsv_qp(hr_dev, &attr, i);
2835 		if (ret)
2836 			return ret;
2837 	}
2838 
2839 	return 0;
2840 }
2841 
2842 static int free_mr_init(struct hns_roce_dev *hr_dev)
2843 {
2844 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2845 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2846 	int ret;
2847 
2848 	mutex_init(&free_mr->mutex);
2849 
2850 	ret = free_mr_alloc_res(hr_dev);
2851 	if (ret) {
2852 		mutex_destroy(&free_mr->mutex);
2853 		return ret;
2854 	}
2855 
2856 	ret = free_mr_modify_qp(hr_dev);
2857 	if (ret)
2858 		goto err_modify_qp;
2859 
2860 	return 0;
2861 
2862 err_modify_qp:
2863 	free_mr_exit(hr_dev);
2864 
2865 	return ret;
2866 }
2867 
2868 static int get_hem_table(struct hns_roce_dev *hr_dev)
2869 {
2870 	unsigned int qpc_count;
2871 	unsigned int cqc_count;
2872 	unsigned int gmv_count;
2873 	int ret;
2874 	int i;
2875 
2876 	/* Alloc memory for source address table buffer space chunk */
2877 	for (gmv_count = 0; gmv_count < hr_dev->caps.gmv_entry_num;
2878 	     gmv_count++) {
2879 		ret = hns_roce_table_get(hr_dev, &hr_dev->gmv_table, gmv_count);
2880 		if (ret)
2881 			goto err_gmv_failed;
2882 	}
2883 
2884 	if (hr_dev->is_vf)
2885 		return 0;
2886 
2887 	/* Alloc memory for QPC Timer buffer space chunk */
2888 	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
2889 	     qpc_count++) {
2890 		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
2891 					 qpc_count);
2892 		if (ret) {
2893 			dev_err(hr_dev->dev, "QPC Timer get failed\n");
2894 			goto err_qpc_timer_failed;
2895 		}
2896 	}
2897 
2898 	/* Alloc memory for CQC Timer buffer space chunk */
2899 	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
2900 	     cqc_count++) {
2901 		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
2902 					 cqc_count);
2903 		if (ret) {
2904 			dev_err(hr_dev->dev, "CQC Timer get failed\n");
2905 			goto err_cqc_timer_failed;
2906 		}
2907 	}
2908 
2909 	return 0;
2910 
2911 err_cqc_timer_failed:
2912 	for (i = 0; i < cqc_count; i++)
2913 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2914 
2915 err_qpc_timer_failed:
2916 	for (i = 0; i < qpc_count; i++)
2917 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2918 
2919 err_gmv_failed:
2920 	for (i = 0; i < gmv_count; i++)
2921 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2922 
2923 	return ret;
2924 }
2925 
2926 static void put_hem_table(struct hns_roce_dev *hr_dev)
2927 {
2928 	int i;
2929 
2930 	for (i = 0; i < hr_dev->caps.gmv_entry_num; i++)
2931 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2932 
2933 	if (hr_dev->is_vf)
2934 		return;
2935 
2936 	for (i = 0; i < hr_dev->caps.qpc_timer_bt_num; i++)
2937 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2938 
2939 	for (i = 0; i < hr_dev->caps.cqc_timer_bt_num; i++)
2940 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2941 }
2942 
2943 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
2944 {
2945 	int ret;
2946 
2947 	/* The hns ROCEE requires the extdb info to be cleared before using */
2948 	ret = hns_roce_clear_extdb_list_info(hr_dev);
2949 	if (ret)
2950 		return ret;
2951 
2952 	ret = get_hem_table(hr_dev);
2953 	if (ret)
2954 		return ret;
2955 
2956 	if (hr_dev->is_vf)
2957 		return 0;
2958 
2959 	ret = hns_roce_init_link_table(hr_dev);
2960 	if (ret) {
2961 		dev_err(hr_dev->dev, "failed to init llm, ret = %d.\n", ret);
2962 		goto err_llm_init_failed;
2963 	}
2964 
2965 	return 0;
2966 
2967 err_llm_init_failed:
2968 	put_hem_table(hr_dev);
2969 
2970 	return ret;
2971 }
2972 
2973 static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev)
2974 {
2975 	hns_roce_function_clear(hr_dev);
2976 
2977 	if (!hr_dev->is_vf)
2978 		hns_roce_free_link_table(hr_dev);
2979 
2980 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09)
2981 		free_dip_list(hr_dev);
2982 }
2983 
2984 static int hns_roce_mbox_post(struct hns_roce_dev *hr_dev,
2985 			      struct hns_roce_mbox_msg *mbox_msg)
2986 {
2987 	struct hns_roce_cmq_desc desc;
2988 	struct hns_roce_post_mbox *mb = (struct hns_roce_post_mbox *)desc.data;
2989 
2990 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_POST_MB, false);
2991 
2992 	mb->in_param_l = cpu_to_le32(mbox_msg->in_param);
2993 	mb->in_param_h = cpu_to_le32(mbox_msg->in_param >> 32);
2994 	mb->out_param_l = cpu_to_le32(mbox_msg->out_param);
2995 	mb->out_param_h = cpu_to_le32(mbox_msg->out_param >> 32);
2996 	mb->cmd_tag = cpu_to_le32(mbox_msg->tag << 8 | mbox_msg->cmd);
2997 	mb->token_event_en = cpu_to_le32(mbox_msg->event_en << 16 |
2998 					 mbox_msg->token);
2999 
3000 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3001 }
3002 
3003 static int v2_wait_mbox_complete(struct hns_roce_dev *hr_dev, u32 timeout,
3004 				 u8 *complete_status)
3005 {
3006 	struct hns_roce_mbox_status *mb_st;
3007 	struct hns_roce_cmq_desc desc;
3008 	unsigned long end;
3009 	int ret = -EBUSY;
3010 	u32 status;
3011 	bool busy;
3012 
3013 	mb_st = (struct hns_roce_mbox_status *)desc.data;
3014 	end = msecs_to_jiffies(timeout) + jiffies;
3015 	while (v2_chk_mbox_is_avail(hr_dev, &busy)) {
3016 		if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
3017 			return -EIO;
3018 
3019 		status = 0;
3020 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_MB_ST,
3021 					      true);
3022 		ret = __hns_roce_cmq_send(hr_dev, &desc, 1);
3023 		if (!ret) {
3024 			status = le32_to_cpu(mb_st->mb_status_hw_run);
3025 			/* No pending message exists in ROCEE mbox. */
3026 			if (!(status & MB_ST_HW_RUN_M))
3027 				break;
3028 		} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3029 			break;
3030 		}
3031 
3032 		if (time_after(jiffies, end)) {
3033 			dev_err_ratelimited(hr_dev->dev,
3034 					    "failed to wait mbox status 0x%x\n",
3035 					    status);
3036 			return -ETIMEDOUT;
3037 		}
3038 
3039 		cond_resched();
3040 		ret = -EBUSY;
3041 	}
3042 
3043 	if (!ret) {
3044 		*complete_status = (u8)(status & MB_ST_COMPLETE_M);
3045 	} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3046 		/* Ignore all errors if the mbox is unavailable. */
3047 		ret = 0;
3048 		*complete_status = MB_ST_COMPLETE_M;
3049 	}
3050 
3051 	return ret;
3052 }
3053 
3054 static int v2_post_mbox(struct hns_roce_dev *hr_dev,
3055 			struct hns_roce_mbox_msg *mbox_msg)
3056 {
3057 	u8 status = 0;
3058 	int ret;
3059 
3060 	/* Waiting for the mbox to be idle */
3061 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_V2_GO_BIT_TIMEOUT_MSECS,
3062 				    &status);
3063 	if (unlikely(ret)) {
3064 		dev_err_ratelimited(hr_dev->dev,
3065 				    "failed to check post mbox status = 0x%x, ret = %d.\n",
3066 				    status, ret);
3067 		return ret;
3068 	}
3069 
3070 	/* Post new message to mbox */
3071 	ret = hns_roce_mbox_post(hr_dev, mbox_msg);
3072 	if (ret)
3073 		dev_err_ratelimited(hr_dev->dev,
3074 				    "failed to post mailbox, ret = %d.\n", ret);
3075 
3076 	return ret;
3077 }
3078 
3079 static int v2_poll_mbox_done(struct hns_roce_dev *hr_dev)
3080 {
3081 	u8 status = 0;
3082 	int ret;
3083 
3084 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_CMD_TIMEOUT_MSECS,
3085 				    &status);
3086 	if (!ret) {
3087 		if (status != MB_ST_COMPLETE_SUCC)
3088 			return -EBUSY;
3089 	} else {
3090 		dev_err_ratelimited(hr_dev->dev,
3091 				    "failed to check mbox status = 0x%x, ret = %d.\n",
3092 				    status, ret);
3093 	}
3094 
3095 	return ret;
3096 }
3097 
3098 static void copy_gid(void *dest, const union ib_gid *gid)
3099 {
3100 #define GID_SIZE 4
3101 	const union ib_gid *src = gid;
3102 	__le32 (*p)[GID_SIZE] = dest;
3103 	int i;
3104 
3105 	if (!gid)
3106 		src = &zgid;
3107 
3108 	for (i = 0; i < GID_SIZE; i++)
3109 		(*p)[i] = cpu_to_le32(*(u32 *)&src->raw[i * sizeof(u32)]);
3110 }
3111 
3112 static int config_sgid_table(struct hns_roce_dev *hr_dev,
3113 			     int gid_index, const union ib_gid *gid,
3114 			     enum hns_roce_sgid_type sgid_type)
3115 {
3116 	struct hns_roce_cmq_desc desc;
3117 	struct hns_roce_cfg_sgid_tb *sgid_tb =
3118 				    (struct hns_roce_cfg_sgid_tb *)desc.data;
3119 
3120 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SGID_TB, false);
3121 
3122 	hr_reg_write(sgid_tb, CFG_SGID_TB_TABLE_IDX, gid_index);
3123 	hr_reg_write(sgid_tb, CFG_SGID_TB_VF_SGID_TYPE, sgid_type);
3124 
3125 	copy_gid(&sgid_tb->vf_sgid_l, gid);
3126 
3127 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3128 }
3129 
3130 static int config_gmv_table(struct hns_roce_dev *hr_dev,
3131 			    int gid_index, const union ib_gid *gid,
3132 			    enum hns_roce_sgid_type sgid_type,
3133 			    const struct ib_gid_attr *attr)
3134 {
3135 	struct hns_roce_cmq_desc desc[2];
3136 	struct hns_roce_cfg_gmv_tb_a *tb_a =
3137 				(struct hns_roce_cfg_gmv_tb_a *)desc[0].data;
3138 	struct hns_roce_cfg_gmv_tb_b *tb_b =
3139 				(struct hns_roce_cfg_gmv_tb_b *)desc[1].data;
3140 
3141 	u16 vlan_id = VLAN_CFI_MASK;
3142 	u8 mac[ETH_ALEN] = {};
3143 	int ret;
3144 
3145 	if (gid) {
3146 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, mac);
3147 		if (ret)
3148 			return ret;
3149 	}
3150 
3151 	hns_roce_cmq_setup_basic_desc(&desc[0], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3152 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
3153 
3154 	hns_roce_cmq_setup_basic_desc(&desc[1], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3155 
3156 	copy_gid(&tb_a->vf_sgid_l, gid);
3157 
3158 	hr_reg_write(tb_a, GMV_TB_A_VF_SGID_TYPE, sgid_type);
3159 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_EN, vlan_id < VLAN_CFI_MASK);
3160 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_ID, vlan_id);
3161 
3162 	tb_b->vf_smac_l = cpu_to_le32(*(u32 *)mac);
3163 
3164 	hr_reg_write(tb_b, GMV_TB_B_SMAC_H, *(u16 *)&mac[4]);
3165 	hr_reg_write(tb_b, GMV_TB_B_SGID_IDX, gid_index);
3166 
3167 	return hns_roce_cmq_send(hr_dev, desc, 2);
3168 }
3169 
3170 static int hns_roce_v2_set_gid(struct hns_roce_dev *hr_dev, int gid_index,
3171 			       const union ib_gid *gid,
3172 			       const struct ib_gid_attr *attr)
3173 {
3174 	enum hns_roce_sgid_type sgid_type = GID_TYPE_FLAG_ROCE_V1;
3175 	int ret;
3176 
3177 	if (gid) {
3178 		if (attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3179 			if (ipv6_addr_v4mapped((void *)gid))
3180 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV4;
3181 			else
3182 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV6;
3183 		} else if (attr->gid_type == IB_GID_TYPE_ROCE) {
3184 			sgid_type = GID_TYPE_FLAG_ROCE_V1;
3185 		}
3186 	}
3187 
3188 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
3189 		ret = config_gmv_table(hr_dev, gid_index, gid, sgid_type, attr);
3190 	else
3191 		ret = config_sgid_table(hr_dev, gid_index, gid, sgid_type);
3192 
3193 	if (ret)
3194 		ibdev_err(&hr_dev->ib_dev, "failed to set gid, ret = %d!\n",
3195 			  ret);
3196 
3197 	return ret;
3198 }
3199 
3200 static int hns_roce_v2_set_mac(struct hns_roce_dev *hr_dev, u8 phy_port,
3201 			       const u8 *addr)
3202 {
3203 	struct hns_roce_cmq_desc desc;
3204 	struct hns_roce_cfg_smac_tb *smac_tb =
3205 				    (struct hns_roce_cfg_smac_tb *)desc.data;
3206 	u16 reg_smac_h;
3207 	u32 reg_smac_l;
3208 
3209 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SMAC_TB, false);
3210 
3211 	reg_smac_l = *(u32 *)(&addr[0]);
3212 	reg_smac_h = *(u16 *)(&addr[4]);
3213 
3214 	hr_reg_write(smac_tb, CFG_SMAC_TB_IDX, phy_port);
3215 	hr_reg_write(smac_tb, CFG_SMAC_TB_VF_SMAC_H, reg_smac_h);
3216 	smac_tb->vf_smac_l = cpu_to_le32(reg_smac_l);
3217 
3218 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3219 }
3220 
3221 static int set_mtpt_pbl(struct hns_roce_dev *hr_dev,
3222 			struct hns_roce_v2_mpt_entry *mpt_entry,
3223 			struct hns_roce_mr *mr)
3224 {
3225 	u64 pages[HNS_ROCE_V2_MAX_INNER_MTPT_NUM] = { 0 };
3226 	struct ib_device *ibdev = &hr_dev->ib_dev;
3227 	dma_addr_t pbl_ba;
3228 	int ret;
3229 	int i;
3230 
3231 	ret = hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, pages,
3232 				min_t(int, ARRAY_SIZE(pages), mr->npages));
3233 	if (ret) {
3234 		ibdev_err(ibdev, "failed to find PBL mtr, ret = %d.\n", ret);
3235 		return ret;
3236 	}
3237 
3238 	/* Aligned to the hardware address access unit */
3239 	for (i = 0; i < ARRAY_SIZE(pages); i++)
3240 		pages[i] >>= MPT_PBL_BUF_ADDR_S;
3241 
3242 	pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3243 
3244 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3245 	mpt_entry->pbl_ba_l = cpu_to_le32(pbl_ba >> MPT_PBL_BA_ADDR_S);
3246 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3247 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3248 
3249 	mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
3250 	hr_reg_write(mpt_entry, MPT_PA0_H, upper_32_bits(pages[0]));
3251 
3252 	mpt_entry->pa1_l = cpu_to_le32(lower_32_bits(pages[1]));
3253 	hr_reg_write(mpt_entry, MPT_PA1_H, upper_32_bits(pages[1]));
3254 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3255 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3256 
3257 	return 0;
3258 }
3259 
3260 static int hns_roce_v2_write_mtpt(struct hns_roce_dev *hr_dev,
3261 				  void *mb_buf, struct hns_roce_mr *mr)
3262 {
3263 	struct hns_roce_v2_mpt_entry *mpt_entry;
3264 
3265 	mpt_entry = mb_buf;
3266 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3267 
3268 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3269 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3270 
3271 	hr_reg_write_bool(mpt_entry, MPT_BIND_EN,
3272 			  mr->access & IB_ACCESS_MW_BIND);
3273 	hr_reg_write_bool(mpt_entry, MPT_ATOMIC_EN,
3274 			  mr->access & IB_ACCESS_REMOTE_ATOMIC);
3275 	hr_reg_write_bool(mpt_entry, MPT_RR_EN,
3276 			  mr->access & IB_ACCESS_REMOTE_READ);
3277 	hr_reg_write_bool(mpt_entry, MPT_RW_EN,
3278 			  mr->access & IB_ACCESS_REMOTE_WRITE);
3279 	hr_reg_write_bool(mpt_entry, MPT_LW_EN,
3280 			  mr->access & IB_ACCESS_LOCAL_WRITE);
3281 
3282 	mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3283 	mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3284 	mpt_entry->lkey = cpu_to_le32(mr->key);
3285 	mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3286 	mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3287 
3288 	if (mr->type != MR_TYPE_MR)
3289 		hr_reg_enable(mpt_entry, MPT_PA);
3290 
3291 	if (mr->type == MR_TYPE_DMA)
3292 		return 0;
3293 
3294 	if (mr->pbl_hop_num != HNS_ROCE_HOP_NUM_0)
3295 		hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, mr->pbl_hop_num);
3296 
3297 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3298 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3299 	hr_reg_enable(mpt_entry, MPT_INNER_PA_VLD);
3300 
3301 	return set_mtpt_pbl(hr_dev, mpt_entry, mr);
3302 }
3303 
3304 static int hns_roce_v2_rereg_write_mtpt(struct hns_roce_dev *hr_dev,
3305 					struct hns_roce_mr *mr, int flags,
3306 					void *mb_buf)
3307 {
3308 	struct hns_roce_v2_mpt_entry *mpt_entry = mb_buf;
3309 	u32 mr_access_flags = mr->access;
3310 	int ret = 0;
3311 
3312 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3313 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3314 
3315 	if (flags & IB_MR_REREG_ACCESS) {
3316 		hr_reg_write(mpt_entry, MPT_BIND_EN,
3317 			     (mr_access_flags & IB_ACCESS_MW_BIND ? 1 : 0));
3318 		hr_reg_write(mpt_entry, MPT_ATOMIC_EN,
3319 			     mr_access_flags & IB_ACCESS_REMOTE_ATOMIC ? 1 : 0);
3320 		hr_reg_write(mpt_entry, MPT_RR_EN,
3321 			     mr_access_flags & IB_ACCESS_REMOTE_READ ? 1 : 0);
3322 		hr_reg_write(mpt_entry, MPT_RW_EN,
3323 			     mr_access_flags & IB_ACCESS_REMOTE_WRITE ? 1 : 0);
3324 		hr_reg_write(mpt_entry, MPT_LW_EN,
3325 			     mr_access_flags & IB_ACCESS_LOCAL_WRITE ? 1 : 0);
3326 	}
3327 
3328 	if (flags & IB_MR_REREG_TRANS) {
3329 		mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3330 		mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3331 		mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3332 		mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3333 
3334 		ret = set_mtpt_pbl(hr_dev, mpt_entry, mr);
3335 	}
3336 
3337 	return ret;
3338 }
3339 
3340 static int hns_roce_v2_frmr_write_mtpt(void *mb_buf, struct hns_roce_mr *mr)
3341 {
3342 	dma_addr_t pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3343 	struct hns_roce_v2_mpt_entry *mpt_entry;
3344 
3345 	mpt_entry = mb_buf;
3346 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3347 
3348 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3349 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3350 
3351 	hr_reg_enable(mpt_entry, MPT_RA_EN);
3352 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3353 
3354 	hr_reg_enable(mpt_entry, MPT_FRE);
3355 	hr_reg_clear(mpt_entry, MPT_MR_MW);
3356 	hr_reg_enable(mpt_entry, MPT_BPD);
3357 	hr_reg_clear(mpt_entry, MPT_PA);
3358 
3359 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, 1);
3360 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3361 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3362 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3363 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3364 
3365 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3366 
3367 	mpt_entry->pbl_ba_l = cpu_to_le32(lower_32_bits(pbl_ba >>
3368 							MPT_PBL_BA_ADDR_S));
3369 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3370 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3371 
3372 	return 0;
3373 }
3374 
3375 static int hns_roce_v2_mw_write_mtpt(void *mb_buf, struct hns_roce_mw *mw)
3376 {
3377 	struct hns_roce_v2_mpt_entry *mpt_entry;
3378 
3379 	mpt_entry = mb_buf;
3380 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3381 
3382 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3383 	hr_reg_write(mpt_entry, MPT_PD, mw->pdn);
3384 
3385 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3386 	hr_reg_enable(mpt_entry, MPT_LW_EN);
3387 
3388 	hr_reg_enable(mpt_entry, MPT_MR_MW);
3389 	hr_reg_enable(mpt_entry, MPT_BPD);
3390 	hr_reg_clear(mpt_entry, MPT_PA);
3391 	hr_reg_write(mpt_entry, MPT_BQP,
3392 		     mw->ibmw.type == IB_MW_TYPE_1 ? 0 : 1);
3393 
3394 	mpt_entry->lkey = cpu_to_le32(mw->rkey);
3395 
3396 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM,
3397 		     mw->pbl_hop_num == HNS_ROCE_HOP_NUM_0 ? 0 :
3398 							     mw->pbl_hop_num);
3399 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3400 		     mw->pbl_ba_pg_sz + PG_SHIFT_OFFSET);
3401 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3402 		     mw->pbl_buf_pg_sz + PG_SHIFT_OFFSET);
3403 
3404 	return 0;
3405 }
3406 
3407 static int free_mr_post_send_lp_wqe(struct hns_roce_qp *hr_qp)
3408 {
3409 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
3410 	struct ib_device *ibdev = &hr_dev->ib_dev;
3411 	const struct ib_send_wr *bad_wr;
3412 	struct ib_rdma_wr rdma_wr = {};
3413 	struct ib_send_wr *send_wr;
3414 	int ret;
3415 
3416 	send_wr = &rdma_wr.wr;
3417 	send_wr->opcode = IB_WR_RDMA_WRITE;
3418 
3419 	ret = hns_roce_v2_post_send(&hr_qp->ibqp, send_wr, &bad_wr);
3420 	if (ret) {
3421 		ibdev_err(ibdev, "failed to post wqe for free mr, ret = %d.\n",
3422 			  ret);
3423 		return ret;
3424 	}
3425 
3426 	return 0;
3427 }
3428 
3429 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3430 			       struct ib_wc *wc);
3431 
3432 static void free_mr_send_cmd_to_hw(struct hns_roce_dev *hr_dev)
3433 {
3434 	struct hns_roce_v2_priv *priv = hr_dev->priv;
3435 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
3436 	struct ib_wc wc[ARRAY_SIZE(free_mr->rsv_qp)];
3437 	struct ib_device *ibdev = &hr_dev->ib_dev;
3438 	struct hns_roce_qp *hr_qp;
3439 	unsigned long end;
3440 	int cqe_cnt = 0;
3441 	int npolled;
3442 	int ret;
3443 	int i;
3444 
3445 	/*
3446 	 * If the device initialization is not complete or in the uninstall
3447 	 * process, then there is no need to execute free mr.
3448 	 */
3449 	if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
3450 	    priv->handle->rinfo.instance_state == HNS_ROCE_STATE_INIT ||
3451 	    hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT)
3452 		return;
3453 
3454 	mutex_lock(&free_mr->mutex);
3455 
3456 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
3457 		hr_qp = free_mr->rsv_qp[i];
3458 
3459 		ret = free_mr_post_send_lp_wqe(hr_qp);
3460 		if (ret) {
3461 			ibdev_err(ibdev,
3462 				  "failed to send wqe (qp:0x%lx) for free mr, ret = %d.\n",
3463 				  hr_qp->qpn, ret);
3464 			break;
3465 		}
3466 
3467 		cqe_cnt++;
3468 	}
3469 
3470 	end = msecs_to_jiffies(HNS_ROCE_V2_FREE_MR_TIMEOUT) + jiffies;
3471 	while (cqe_cnt) {
3472 		npolled = hns_roce_v2_poll_cq(&free_mr->rsv_cq->ib_cq, cqe_cnt, wc);
3473 		if (npolled < 0) {
3474 			ibdev_err(ibdev,
3475 				  "failed to poll cqe for free mr, remain %d cqe.\n",
3476 				  cqe_cnt);
3477 			goto out;
3478 		}
3479 
3480 		if (time_after(jiffies, end)) {
3481 			ibdev_err(ibdev,
3482 				  "failed to poll cqe for free mr and timeout, remain %d cqe.\n",
3483 				  cqe_cnt);
3484 			goto out;
3485 		}
3486 		cqe_cnt -= npolled;
3487 	}
3488 
3489 out:
3490 	mutex_unlock(&free_mr->mutex);
3491 }
3492 
3493 static void hns_roce_v2_dereg_mr(struct hns_roce_dev *hr_dev)
3494 {
3495 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
3496 		free_mr_send_cmd_to_hw(hr_dev);
3497 }
3498 
3499 static void *get_cqe_v2(struct hns_roce_cq *hr_cq, int n)
3500 {
3501 	return hns_roce_buf_offset(hr_cq->mtr.kmem, n * hr_cq->cqe_size);
3502 }
3503 
3504 static void *get_sw_cqe_v2(struct hns_roce_cq *hr_cq, unsigned int n)
3505 {
3506 	struct hns_roce_v2_cqe *cqe = get_cqe_v2(hr_cq, n & hr_cq->ib_cq.cqe);
3507 
3508 	/* Get cqe when Owner bit is Conversely with the MSB of cons_idx */
3509 	return (hr_reg_read(cqe, CQE_OWNER) ^ !!(n & hr_cq->cq_depth)) ? cqe :
3510 									 NULL;
3511 }
3512 
3513 static inline void update_cq_db(struct hns_roce_dev *hr_dev,
3514 				struct hns_roce_cq *hr_cq)
3515 {
3516 	if (likely(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) {
3517 		*hr_cq->set_ci_db = hr_cq->cons_index & V2_CQ_DB_CONS_IDX_M;
3518 	} else {
3519 		struct hns_roce_v2_db cq_db = {};
3520 
3521 		hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3522 		hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB);
3523 		hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3524 		hr_reg_write(&cq_db, DB_CQ_CMD_SN, 1);
3525 
3526 		hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3527 	}
3528 }
3529 
3530 static void __hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3531 				   struct hns_roce_srq *srq)
3532 {
3533 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3534 	struct hns_roce_v2_cqe *cqe, *dest;
3535 	u32 prod_index;
3536 	int nfreed = 0;
3537 	int wqe_index;
3538 	u8 owner_bit;
3539 
3540 	for (prod_index = hr_cq->cons_index; get_sw_cqe_v2(hr_cq, prod_index);
3541 	     ++prod_index) {
3542 		if (prod_index > hr_cq->cons_index + hr_cq->ib_cq.cqe)
3543 			break;
3544 	}
3545 
3546 	/*
3547 	 * Now backwards through the CQ, removing CQ entries
3548 	 * that match our QP by overwriting them with next entries.
3549 	 */
3550 	while ((int) --prod_index - (int) hr_cq->cons_index >= 0) {
3551 		cqe = get_cqe_v2(hr_cq, prod_index & hr_cq->ib_cq.cqe);
3552 		if (hr_reg_read(cqe, CQE_LCL_QPN) == qpn) {
3553 			if (srq && hr_reg_read(cqe, CQE_S_R)) {
3554 				wqe_index = hr_reg_read(cqe, CQE_WQE_IDX);
3555 				hns_roce_free_srq_wqe(srq, wqe_index);
3556 			}
3557 			++nfreed;
3558 		} else if (nfreed) {
3559 			dest = get_cqe_v2(hr_cq, (prod_index + nfreed) &
3560 					  hr_cq->ib_cq.cqe);
3561 			owner_bit = hr_reg_read(dest, CQE_OWNER);
3562 			memcpy(dest, cqe, hr_cq->cqe_size);
3563 			hr_reg_write(dest, CQE_OWNER, owner_bit);
3564 		}
3565 	}
3566 
3567 	if (nfreed) {
3568 		hr_cq->cons_index += nfreed;
3569 		update_cq_db(hr_dev, hr_cq);
3570 	}
3571 }
3572 
3573 static void hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3574 				 struct hns_roce_srq *srq)
3575 {
3576 	spin_lock_irq(&hr_cq->lock);
3577 	__hns_roce_v2_cq_clean(hr_cq, qpn, srq);
3578 	spin_unlock_irq(&hr_cq->lock);
3579 }
3580 
3581 static void hns_roce_v2_write_cqc(struct hns_roce_dev *hr_dev,
3582 				  struct hns_roce_cq *hr_cq, void *mb_buf,
3583 				  u64 *mtts, dma_addr_t dma_handle)
3584 {
3585 	struct hns_roce_v2_cq_context *cq_context;
3586 
3587 	cq_context = mb_buf;
3588 	memset(cq_context, 0, sizeof(*cq_context));
3589 
3590 	hr_reg_write(cq_context, CQC_CQ_ST, V2_CQ_STATE_VALID);
3591 	hr_reg_write(cq_context, CQC_ARM_ST, NO_ARMED);
3592 	hr_reg_write(cq_context, CQC_SHIFT, ilog2(hr_cq->cq_depth));
3593 	hr_reg_write(cq_context, CQC_CEQN, hr_cq->vector);
3594 	hr_reg_write(cq_context, CQC_CQN, hr_cq->cqn);
3595 
3596 	if (hr_cq->cqe_size == HNS_ROCE_V3_CQE_SIZE)
3597 		hr_reg_write(cq_context, CQC_CQE_SIZE, CQE_SIZE_64B);
3598 
3599 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
3600 		hr_reg_enable(cq_context, CQC_STASH);
3601 
3602 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_L,
3603 		     to_hr_hw_page_addr(mtts[0]));
3604 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_H,
3605 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
3606 	hr_reg_write(cq_context, CQC_CQE_HOP_NUM, hr_dev->caps.cqe_hop_num ==
3607 		     HNS_ROCE_HOP_NUM_0 ? 0 : hr_dev->caps.cqe_hop_num);
3608 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_L,
3609 		     to_hr_hw_page_addr(mtts[1]));
3610 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_H,
3611 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
3612 	hr_reg_write(cq_context, CQC_CQE_BAR_PG_SZ,
3613 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.ba_pg_shift));
3614 	hr_reg_write(cq_context, CQC_CQE_BUF_PG_SZ,
3615 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.buf_pg_shift));
3616 	hr_reg_write(cq_context, CQC_CQE_BA_L, dma_handle >> CQC_CQE_BA_L_S);
3617 	hr_reg_write(cq_context, CQC_CQE_BA_H, dma_handle >> CQC_CQE_BA_H_S);
3618 	hr_reg_write_bool(cq_context, CQC_DB_RECORD_EN,
3619 			  hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB);
3620 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_L,
3621 		     ((u32)hr_cq->db.dma) >> 1);
3622 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_H,
3623 		     hr_cq->db.dma >> CQC_CQE_DB_RECORD_ADDR_H_S);
3624 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT,
3625 		     HNS_ROCE_V2_CQ_DEFAULT_BURST_NUM);
3626 	hr_reg_write(cq_context, CQC_CQ_PERIOD,
3627 		     HNS_ROCE_V2_CQ_DEFAULT_INTERVAL);
3628 }
3629 
3630 static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
3631 				     enum ib_cq_notify_flags flags)
3632 {
3633 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3634 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3635 	struct hns_roce_v2_db cq_db = {};
3636 	u32 notify_flag;
3637 
3638 	/*
3639 	 * flags = 0, then notify_flag : next
3640 	 * flags = 1, then notify flag : solocited
3641 	 */
3642 	notify_flag = (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
3643 		      V2_CQ_DB_REQ_NOT : V2_CQ_DB_REQ_NOT_SOL;
3644 
3645 	hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3646 	hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB_NOTIFY);
3647 	hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3648 	hr_reg_write(&cq_db, DB_CQ_CMD_SN, hr_cq->arm_sn);
3649 	hr_reg_write(&cq_db, DB_CQ_NOTIFY, notify_flag);
3650 
3651 	hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3652 
3653 	return 0;
3654 }
3655 
3656 static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
3657 		   int num_entries, struct ib_wc *wc)
3658 {
3659 	unsigned int left;
3660 	int npolled = 0;
3661 
3662 	left = wq->head - wq->tail;
3663 	if (left == 0)
3664 		return 0;
3665 
3666 	left = min_t(unsigned int, (unsigned int)num_entries, left);
3667 	while (npolled < left) {
3668 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3669 		wc->status = IB_WC_WR_FLUSH_ERR;
3670 		wc->vendor_err = 0;
3671 		wc->qp = &hr_qp->ibqp;
3672 
3673 		wq->tail++;
3674 		wc++;
3675 		npolled++;
3676 	}
3677 
3678 	return npolled;
3679 }
3680 
3681 static int hns_roce_v2_sw_poll_cq(struct hns_roce_cq *hr_cq, int num_entries,
3682 				  struct ib_wc *wc)
3683 {
3684 	struct hns_roce_qp *hr_qp;
3685 	int npolled = 0;
3686 
3687 	list_for_each_entry(hr_qp, &hr_cq->sq_list, sq_node) {
3688 		npolled += sw_comp(hr_qp, &hr_qp->sq,
3689 				   num_entries - npolled, wc + npolled);
3690 		if (npolled >= num_entries)
3691 			goto out;
3692 	}
3693 
3694 	list_for_each_entry(hr_qp, &hr_cq->rq_list, rq_node) {
3695 		npolled += sw_comp(hr_qp, &hr_qp->rq,
3696 				   num_entries - npolled, wc + npolled);
3697 		if (npolled >= num_entries)
3698 			goto out;
3699 	}
3700 
3701 out:
3702 	return npolled;
3703 }
3704 
3705 static void get_cqe_status(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
3706 			   struct hns_roce_cq *cq, struct hns_roce_v2_cqe *cqe,
3707 			   struct ib_wc *wc)
3708 {
3709 	static const struct {
3710 		u32 cqe_status;
3711 		enum ib_wc_status wc_status;
3712 	} map[] = {
3713 		{ HNS_ROCE_CQE_V2_SUCCESS, IB_WC_SUCCESS },
3714 		{ HNS_ROCE_CQE_V2_LOCAL_LENGTH_ERR, IB_WC_LOC_LEN_ERR },
3715 		{ HNS_ROCE_CQE_V2_LOCAL_QP_OP_ERR, IB_WC_LOC_QP_OP_ERR },
3716 		{ HNS_ROCE_CQE_V2_LOCAL_PROT_ERR, IB_WC_LOC_PROT_ERR },
3717 		{ HNS_ROCE_CQE_V2_WR_FLUSH_ERR, IB_WC_WR_FLUSH_ERR },
3718 		{ HNS_ROCE_CQE_V2_MW_BIND_ERR, IB_WC_MW_BIND_ERR },
3719 		{ HNS_ROCE_CQE_V2_BAD_RESP_ERR, IB_WC_BAD_RESP_ERR },
3720 		{ HNS_ROCE_CQE_V2_LOCAL_ACCESS_ERR, IB_WC_LOC_ACCESS_ERR },
3721 		{ HNS_ROCE_CQE_V2_REMOTE_INVAL_REQ_ERR, IB_WC_REM_INV_REQ_ERR },
3722 		{ HNS_ROCE_CQE_V2_REMOTE_ACCESS_ERR, IB_WC_REM_ACCESS_ERR },
3723 		{ HNS_ROCE_CQE_V2_REMOTE_OP_ERR, IB_WC_REM_OP_ERR },
3724 		{ HNS_ROCE_CQE_V2_TRANSPORT_RETRY_EXC_ERR,
3725 		  IB_WC_RETRY_EXC_ERR },
3726 		{ HNS_ROCE_CQE_V2_RNR_RETRY_EXC_ERR, IB_WC_RNR_RETRY_EXC_ERR },
3727 		{ HNS_ROCE_CQE_V2_REMOTE_ABORT_ERR, IB_WC_REM_ABORT_ERR },
3728 		{ HNS_ROCE_CQE_V2_GENERAL_ERR, IB_WC_GENERAL_ERR}
3729 	};
3730 
3731 	u32 cqe_status = hr_reg_read(cqe, CQE_STATUS);
3732 	int i;
3733 
3734 	wc->status = IB_WC_GENERAL_ERR;
3735 	for (i = 0; i < ARRAY_SIZE(map); i++)
3736 		if (cqe_status == map[i].cqe_status) {
3737 			wc->status = map[i].wc_status;
3738 			break;
3739 		}
3740 
3741 	if (likely(wc->status == IB_WC_SUCCESS ||
3742 		   wc->status == IB_WC_WR_FLUSH_ERR))
3743 		return;
3744 
3745 	ibdev_err_ratelimited(&hr_dev->ib_dev, "error cqe status 0x%x:\n",
3746 			      cqe_status);
3747 	print_hex_dump(KERN_DEBUG, "", DUMP_PREFIX_NONE, 16, 4, cqe,
3748 		       cq->cqe_size, false);
3749 	wc->vendor_err = hr_reg_read(cqe, CQE_SUB_STATUS);
3750 
3751 	/*
3752 	 * For hns ROCEE, GENERAL_ERR is an error type that is not defined in
3753 	 * the standard protocol, the driver must ignore it and needn't to set
3754 	 * the QP to an error state.
3755 	 */
3756 	if (cqe_status == HNS_ROCE_CQE_V2_GENERAL_ERR)
3757 		return;
3758 
3759 	flush_cqe(hr_dev, qp);
3760 }
3761 
3762 static int get_cur_qp(struct hns_roce_cq *hr_cq, struct hns_roce_v2_cqe *cqe,
3763 		      struct hns_roce_qp **cur_qp)
3764 {
3765 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3766 	struct hns_roce_qp *hr_qp = *cur_qp;
3767 	u32 qpn;
3768 
3769 	qpn = hr_reg_read(cqe, CQE_LCL_QPN);
3770 
3771 	if (!hr_qp || qpn != hr_qp->qpn) {
3772 		hr_qp = __hns_roce_qp_lookup(hr_dev, qpn);
3773 		if (unlikely(!hr_qp)) {
3774 			ibdev_err(&hr_dev->ib_dev,
3775 				  "CQ %06lx with entry for unknown QPN %06x\n",
3776 				  hr_cq->cqn, qpn);
3777 			return -EINVAL;
3778 		}
3779 		*cur_qp = hr_qp;
3780 	}
3781 
3782 	return 0;
3783 }
3784 
3785 /*
3786  * mapped-value = 1 + real-value
3787  * The ib wc opcode's real value is start from 0, In order to distinguish
3788  * between initialized and uninitialized map values, we plus 1 to the actual
3789  * value when defining the mapping, so that the validity can be identified by
3790  * checking whether the mapped value is greater than 0.
3791  */
3792 #define HR_WC_OP_MAP(hr_key, ib_key) \
3793 		[HNS_ROCE_V2_WQE_OP_ ## hr_key] = 1 + IB_WC_ ## ib_key
3794 
3795 static const u32 wc_send_op_map[] = {
3796 	HR_WC_OP_MAP(SEND,			SEND),
3797 	HR_WC_OP_MAP(SEND_WITH_INV,		SEND),
3798 	HR_WC_OP_MAP(SEND_WITH_IMM,		SEND),
3799 	HR_WC_OP_MAP(RDMA_READ,			RDMA_READ),
3800 	HR_WC_OP_MAP(RDMA_WRITE,		RDMA_WRITE),
3801 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,	RDMA_WRITE),
3802 	HR_WC_OP_MAP(ATOM_CMP_AND_SWAP,		COMP_SWAP),
3803 	HR_WC_OP_MAP(ATOM_FETCH_AND_ADD,	FETCH_ADD),
3804 	HR_WC_OP_MAP(ATOM_MSK_CMP_AND_SWAP,	MASKED_COMP_SWAP),
3805 	HR_WC_OP_MAP(ATOM_MSK_FETCH_AND_ADD,	MASKED_FETCH_ADD),
3806 	HR_WC_OP_MAP(FAST_REG_PMR,		REG_MR),
3807 	HR_WC_OP_MAP(BIND_MW,			REG_MR),
3808 };
3809 
3810 static int to_ib_wc_send_op(u32 hr_opcode)
3811 {
3812 	if (hr_opcode >= ARRAY_SIZE(wc_send_op_map))
3813 		return -EINVAL;
3814 
3815 	return wc_send_op_map[hr_opcode] ? wc_send_op_map[hr_opcode] - 1 :
3816 					   -EINVAL;
3817 }
3818 
3819 static const u32 wc_recv_op_map[] = {
3820 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,		WITH_IMM),
3821 	HR_WC_OP_MAP(SEND,				RECV),
3822 	HR_WC_OP_MAP(SEND_WITH_IMM,			WITH_IMM),
3823 	HR_WC_OP_MAP(SEND_WITH_INV,			RECV),
3824 };
3825 
3826 static int to_ib_wc_recv_op(u32 hr_opcode)
3827 {
3828 	if (hr_opcode >= ARRAY_SIZE(wc_recv_op_map))
3829 		return -EINVAL;
3830 
3831 	return wc_recv_op_map[hr_opcode] ? wc_recv_op_map[hr_opcode] - 1 :
3832 					   -EINVAL;
3833 }
3834 
3835 static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3836 {
3837 	u32 hr_opcode;
3838 	int ib_opcode;
3839 
3840 	wc->wc_flags = 0;
3841 
3842 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3843 	switch (hr_opcode) {
3844 	case HNS_ROCE_V2_WQE_OP_RDMA_READ:
3845 		wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3846 		break;
3847 	case HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM:
3848 	case HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM:
3849 		wc->wc_flags |= IB_WC_WITH_IMM;
3850 		break;
3851 	case HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP:
3852 	case HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD:
3853 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP:
3854 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD:
3855 		wc->byte_len  = 8;
3856 		break;
3857 	default:
3858 		break;
3859 	}
3860 
3861 	ib_opcode = to_ib_wc_send_op(hr_opcode);
3862 	if (ib_opcode < 0)
3863 		wc->status = IB_WC_GENERAL_ERR;
3864 	else
3865 		wc->opcode = ib_opcode;
3866 }
3867 
3868 static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3869 {
3870 	u32 hr_opcode;
3871 	int ib_opcode;
3872 
3873 	wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3874 
3875 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3876 	switch (hr_opcode) {
3877 	case HNS_ROCE_V2_OPCODE_RDMA_WRITE_IMM:
3878 	case HNS_ROCE_V2_OPCODE_SEND_WITH_IMM:
3879 		wc->wc_flags = IB_WC_WITH_IMM;
3880 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->immtdata));
3881 		break;
3882 	case HNS_ROCE_V2_OPCODE_SEND_WITH_INV:
3883 		wc->wc_flags = IB_WC_WITH_INVALIDATE;
3884 		wc->ex.invalidate_rkey = le32_to_cpu(cqe->rkey);
3885 		break;
3886 	default:
3887 		wc->wc_flags = 0;
3888 	}
3889 
3890 	ib_opcode = to_ib_wc_recv_op(hr_opcode);
3891 	if (ib_opcode < 0)
3892 		wc->status = IB_WC_GENERAL_ERR;
3893 	else
3894 		wc->opcode = ib_opcode;
3895 
3896 	wc->sl = hr_reg_read(cqe, CQE_SL);
3897 	wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
3898 	wc->slid = 0;
3899 	wc->wc_flags |= hr_reg_read(cqe, CQE_GRH) ? IB_WC_GRH : 0;
3900 	wc->port_num = hr_reg_read(cqe, CQE_PORTN);
3901 	wc->pkey_index = 0;
3902 
3903 	if (hr_reg_read(cqe, CQE_VID_VLD)) {
3904 		wc->vlan_id = hr_reg_read(cqe, CQE_VID);
3905 		wc->wc_flags |= IB_WC_WITH_VLAN;
3906 	} else {
3907 		wc->vlan_id = 0xffff;
3908 	}
3909 
3910 	wc->network_hdr_type = hr_reg_read(cqe, CQE_PORT_TYPE);
3911 
3912 	return 0;
3913 }
3914 
3915 static int hns_roce_v2_poll_one(struct hns_roce_cq *hr_cq,
3916 				struct hns_roce_qp **cur_qp, struct ib_wc *wc)
3917 {
3918 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3919 	struct hns_roce_qp *qp = *cur_qp;
3920 	struct hns_roce_srq *srq = NULL;
3921 	struct hns_roce_v2_cqe *cqe;
3922 	struct hns_roce_wq *wq;
3923 	int is_send;
3924 	u16 wqe_idx;
3925 	int ret;
3926 
3927 	cqe = get_sw_cqe_v2(hr_cq, hr_cq->cons_index);
3928 	if (!cqe)
3929 		return -EAGAIN;
3930 
3931 	++hr_cq->cons_index;
3932 	/* Memory barrier */
3933 	rmb();
3934 
3935 	ret = get_cur_qp(hr_cq, cqe, &qp);
3936 	if (ret)
3937 		return ret;
3938 
3939 	wc->qp = &qp->ibqp;
3940 	wc->vendor_err = 0;
3941 
3942 	wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
3943 
3944 	is_send = !hr_reg_read(cqe, CQE_S_R);
3945 	if (is_send) {
3946 		wq = &qp->sq;
3947 
3948 		/* If sg_signal_bit is set, tail pointer will be updated to
3949 		 * the WQE corresponding to the current CQE.
3950 		 */
3951 		if (qp->sq_signal_bits)
3952 			wq->tail += (wqe_idx - (u16)wq->tail) &
3953 				    (wq->wqe_cnt - 1);
3954 
3955 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3956 		++wq->tail;
3957 
3958 		fill_send_wc(wc, cqe);
3959 	} else {
3960 		if (qp->ibqp.srq) {
3961 			srq = to_hr_srq(qp->ibqp.srq);
3962 			wc->wr_id = srq->wrid[wqe_idx];
3963 			hns_roce_free_srq_wqe(srq, wqe_idx);
3964 		} else {
3965 			wq = &qp->rq;
3966 			wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3967 			++wq->tail;
3968 		}
3969 
3970 		ret = fill_recv_wc(wc, cqe);
3971 	}
3972 
3973 	get_cqe_status(hr_dev, qp, hr_cq, cqe, wc);
3974 	if (unlikely(wc->status != IB_WC_SUCCESS))
3975 		return 0;
3976 
3977 	return ret;
3978 }
3979 
3980 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3981 			       struct ib_wc *wc)
3982 {
3983 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3984 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3985 	struct hns_roce_qp *cur_qp = NULL;
3986 	unsigned long flags;
3987 	int npolled;
3988 
3989 	spin_lock_irqsave(&hr_cq->lock, flags);
3990 
3991 	/*
3992 	 * When the device starts to reset, the state is RST_DOWN. At this time,
3993 	 * there may still be some valid CQEs in the hardware that are not
3994 	 * polled. Therefore, it is not allowed to switch to the software mode
3995 	 * immediately. When the state changes to UNINIT, CQE no longer exists
3996 	 * in the hardware, and then switch to software mode.
3997 	 */
3998 	if (hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT) {
3999 		npolled = hns_roce_v2_sw_poll_cq(hr_cq, num_entries, wc);
4000 		goto out;
4001 	}
4002 
4003 	for (npolled = 0; npolled < num_entries; ++npolled) {
4004 		if (hns_roce_v2_poll_one(hr_cq, &cur_qp, wc + npolled))
4005 			break;
4006 	}
4007 
4008 	if (npolled)
4009 		update_cq_db(hr_dev, hr_cq);
4010 
4011 out:
4012 	spin_unlock_irqrestore(&hr_cq->lock, flags);
4013 
4014 	return npolled;
4015 }
4016 
4017 static int get_op_for_set_hem(struct hns_roce_dev *hr_dev, u32 type,
4018 			      u32 step_idx, u8 *mbox_cmd)
4019 {
4020 	u8 cmd;
4021 
4022 	switch (type) {
4023 	case HEM_TYPE_QPC:
4024 		cmd = HNS_ROCE_CMD_WRITE_QPC_BT0;
4025 		break;
4026 	case HEM_TYPE_MTPT:
4027 		cmd = HNS_ROCE_CMD_WRITE_MPT_BT0;
4028 		break;
4029 	case HEM_TYPE_CQC:
4030 		cmd = HNS_ROCE_CMD_WRITE_CQC_BT0;
4031 		break;
4032 	case HEM_TYPE_SRQC:
4033 		cmd = HNS_ROCE_CMD_WRITE_SRQC_BT0;
4034 		break;
4035 	case HEM_TYPE_SCCC:
4036 		cmd = HNS_ROCE_CMD_WRITE_SCCC_BT0;
4037 		break;
4038 	case HEM_TYPE_QPC_TIMER:
4039 		cmd = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
4040 		break;
4041 	case HEM_TYPE_CQC_TIMER:
4042 		cmd = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
4043 		break;
4044 	default:
4045 		dev_warn(hr_dev->dev, "failed to check hem type %u.\n", type);
4046 		return -EINVAL;
4047 	}
4048 
4049 	*mbox_cmd = cmd + step_idx;
4050 
4051 	return 0;
4052 }
4053 
4054 static int config_gmv_ba_to_hw(struct hns_roce_dev *hr_dev, unsigned long obj,
4055 			       dma_addr_t base_addr)
4056 {
4057 	struct hns_roce_cmq_desc desc;
4058 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
4059 	u32 idx = obj / (HNS_HW_PAGE_SIZE / hr_dev->caps.gmv_entry_sz);
4060 	u64 addr = to_hr_hw_page_addr(base_addr);
4061 
4062 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
4063 
4064 	hr_reg_write(req, CFG_GMV_BT_BA_L, lower_32_bits(addr));
4065 	hr_reg_write(req, CFG_GMV_BT_BA_H, upper_32_bits(addr));
4066 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
4067 
4068 	return hns_roce_cmq_send(hr_dev, &desc, 1);
4069 }
4070 
4071 static int set_hem_to_hw(struct hns_roce_dev *hr_dev, int obj,
4072 			 dma_addr_t base_addr, u32 hem_type, u32 step_idx)
4073 {
4074 	int ret;
4075 	u8 cmd;
4076 
4077 	if (unlikely(hem_type == HEM_TYPE_GMV))
4078 		return config_gmv_ba_to_hw(hr_dev, obj, base_addr);
4079 
4080 	if (unlikely(hem_type == HEM_TYPE_SCCC && step_idx))
4081 		return 0;
4082 
4083 	ret = get_op_for_set_hem(hr_dev, hem_type, step_idx, &cmd);
4084 	if (ret < 0)
4085 		return ret;
4086 
4087 	return config_hem_ba_to_hw(hr_dev, base_addr, cmd, obj);
4088 }
4089 
4090 static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
4091 			       struct hns_roce_hem_table *table, int obj,
4092 			       u32 step_idx)
4093 {
4094 	struct hns_roce_hem_mhop mhop;
4095 	struct hns_roce_hem *hem;
4096 	unsigned long mhop_obj = obj;
4097 	int i, j, k;
4098 	int ret = 0;
4099 	u64 hem_idx = 0;
4100 	u64 l1_idx = 0;
4101 	u64 bt_ba = 0;
4102 	u32 chunk_ba_num;
4103 	u32 hop_num;
4104 
4105 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4106 		return 0;
4107 
4108 	hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
4109 	i = mhop.l0_idx;
4110 	j = mhop.l1_idx;
4111 	k = mhop.l2_idx;
4112 	hop_num = mhop.hop_num;
4113 	chunk_ba_num = mhop.bt_chunk_size / 8;
4114 
4115 	if (hop_num == 2) {
4116 		hem_idx = i * chunk_ba_num * chunk_ba_num + j * chunk_ba_num +
4117 			  k;
4118 		l1_idx = i * chunk_ba_num + j;
4119 	} else if (hop_num == 1) {
4120 		hem_idx = i * chunk_ba_num + j;
4121 	} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
4122 		hem_idx = i;
4123 	}
4124 
4125 	if (table->type == HEM_TYPE_SCCC)
4126 		obj = mhop.l0_idx;
4127 
4128 	if (check_whether_last_step(hop_num, step_idx)) {
4129 		hem = table->hem[hem_idx];
4130 
4131 		ret = set_hem_to_hw(hr_dev, obj, hem->dma, table->type, step_idx);
4132 	} else {
4133 		if (step_idx == 0)
4134 			bt_ba = table->bt_l0_dma_addr[i];
4135 		else if (step_idx == 1 && hop_num == 2)
4136 			bt_ba = table->bt_l1_dma_addr[l1_idx];
4137 
4138 		ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type, step_idx);
4139 	}
4140 
4141 	return ret;
4142 }
4143 
4144 static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
4145 				 struct hns_roce_hem_table *table,
4146 				 int tag, u32 step_idx)
4147 {
4148 	struct hns_roce_cmd_mailbox *mailbox;
4149 	struct device *dev = hr_dev->dev;
4150 	u8 cmd = 0xff;
4151 	int ret;
4152 
4153 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4154 		return 0;
4155 
4156 	switch (table->type) {
4157 	case HEM_TYPE_QPC:
4158 		cmd = HNS_ROCE_CMD_DESTROY_QPC_BT0;
4159 		break;
4160 	case HEM_TYPE_MTPT:
4161 		cmd = HNS_ROCE_CMD_DESTROY_MPT_BT0;
4162 		break;
4163 	case HEM_TYPE_CQC:
4164 		cmd = HNS_ROCE_CMD_DESTROY_CQC_BT0;
4165 		break;
4166 	case HEM_TYPE_SRQC:
4167 		cmd = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
4168 		break;
4169 	case HEM_TYPE_SCCC:
4170 	case HEM_TYPE_QPC_TIMER:
4171 	case HEM_TYPE_CQC_TIMER:
4172 	case HEM_TYPE_GMV:
4173 		return 0;
4174 	default:
4175 		dev_warn(dev, "table %u not to be destroyed by mailbox!\n",
4176 			 table->type);
4177 		return 0;
4178 	}
4179 
4180 	cmd += step_idx;
4181 
4182 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4183 	if (IS_ERR(mailbox))
4184 		return PTR_ERR(mailbox);
4185 
4186 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, cmd, tag);
4187 
4188 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4189 	return ret;
4190 }
4191 
4192 static int hns_roce_v2_qp_modify(struct hns_roce_dev *hr_dev,
4193 				 struct hns_roce_v2_qp_context *context,
4194 				 struct hns_roce_v2_qp_context *qpc_mask,
4195 				 struct hns_roce_qp *hr_qp)
4196 {
4197 	struct hns_roce_cmd_mailbox *mailbox;
4198 	int qpc_size;
4199 	int ret;
4200 
4201 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4202 	if (IS_ERR(mailbox))
4203 		return PTR_ERR(mailbox);
4204 
4205 	/* The qpc size of HIP08 is only 256B, which is half of HIP09 */
4206 	qpc_size = hr_dev->caps.qpc_sz;
4207 	memcpy(mailbox->buf, context, qpc_size);
4208 	memcpy(mailbox->buf + qpc_size, qpc_mask, qpc_size);
4209 
4210 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
4211 				HNS_ROCE_CMD_MODIFY_QPC, hr_qp->qpn);
4212 
4213 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4214 
4215 	return ret;
4216 }
4217 
4218 static void set_access_flags(struct hns_roce_qp *hr_qp,
4219 			     struct hns_roce_v2_qp_context *context,
4220 			     struct hns_roce_v2_qp_context *qpc_mask,
4221 			     const struct ib_qp_attr *attr, int attr_mask)
4222 {
4223 	u8 dest_rd_atomic;
4224 	u32 access_flags;
4225 
4226 	dest_rd_atomic = (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) ?
4227 			 attr->max_dest_rd_atomic : hr_qp->resp_depth;
4228 
4229 	access_flags = (attr_mask & IB_QP_ACCESS_FLAGS) ?
4230 		       attr->qp_access_flags : hr_qp->atomic_rd_en;
4231 
4232 	if (!dest_rd_atomic)
4233 		access_flags &= IB_ACCESS_REMOTE_WRITE;
4234 
4235 	hr_reg_write_bool(context, QPC_RRE,
4236 			  access_flags & IB_ACCESS_REMOTE_READ);
4237 	hr_reg_clear(qpc_mask, QPC_RRE);
4238 
4239 	hr_reg_write_bool(context, QPC_RWE,
4240 			  access_flags & IB_ACCESS_REMOTE_WRITE);
4241 	hr_reg_clear(qpc_mask, QPC_RWE);
4242 
4243 	hr_reg_write_bool(context, QPC_ATE,
4244 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4245 	hr_reg_clear(qpc_mask, QPC_ATE);
4246 	hr_reg_write_bool(context, QPC_EXT_ATE,
4247 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4248 	hr_reg_clear(qpc_mask, QPC_EXT_ATE);
4249 }
4250 
4251 static void set_qpc_wqe_cnt(struct hns_roce_qp *hr_qp,
4252 			    struct hns_roce_v2_qp_context *context)
4253 {
4254 	hr_reg_write(context, QPC_SGE_SHIFT,
4255 		     to_hr_hem_entries_shift(hr_qp->sge.sge_cnt,
4256 					     hr_qp->sge.sge_shift));
4257 
4258 	hr_reg_write(context, QPC_SQ_SHIFT, ilog2(hr_qp->sq.wqe_cnt));
4259 
4260 	hr_reg_write(context, QPC_RQ_SHIFT, ilog2(hr_qp->rq.wqe_cnt));
4261 }
4262 
4263 static inline int get_cqn(struct ib_cq *ib_cq)
4264 {
4265 	return ib_cq ? to_hr_cq(ib_cq)->cqn : 0;
4266 }
4267 
4268 static inline int get_pdn(struct ib_pd *ib_pd)
4269 {
4270 	return ib_pd ? to_hr_pd(ib_pd)->pdn : 0;
4271 }
4272 
4273 static void modify_qp_reset_to_init(struct ib_qp *ibqp,
4274 				    struct hns_roce_v2_qp_context *context,
4275 				    struct hns_roce_v2_qp_context *qpc_mask)
4276 {
4277 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4278 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4279 
4280 	/*
4281 	 * In v2 engine, software pass context and context mask to hardware
4282 	 * when modifying qp. If software need modify some fields in context,
4283 	 * we should set all bits of the relevant fields in context mask to
4284 	 * 0 at the same time, else set them to 0x1.
4285 	 */
4286 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4287 
4288 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4289 
4290 	hr_reg_write(context, QPC_RQWS, ilog2(hr_qp->rq.max_gs));
4291 
4292 	set_qpc_wqe_cnt(hr_qp, context);
4293 
4294 	/* No VLAN need to set 0xFFF */
4295 	hr_reg_write(context, QPC_VLAN_ID, 0xfff);
4296 
4297 	if (ibqp->qp_type == IB_QPT_XRC_TGT) {
4298 		context->qkey_xrcd = cpu_to_le32(hr_qp->xrcdn);
4299 
4300 		hr_reg_enable(context, QPC_XRC_QP_TYPE);
4301 	}
4302 
4303 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
4304 		hr_reg_enable(context, QPC_RQ_RECORD_EN);
4305 
4306 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
4307 		hr_reg_enable(context, QPC_OWNER_MODE);
4308 
4309 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_L,
4310 		     lower_32_bits(hr_qp->rdb.dma) >> 1);
4311 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
4312 		     upper_32_bits(hr_qp->rdb.dma));
4313 
4314 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4315 
4316 	if (ibqp->srq) {
4317 		hr_reg_enable(context, QPC_SRQ_EN);
4318 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4319 	}
4320 
4321 	hr_reg_enable(context, QPC_FRE);
4322 
4323 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4324 
4325 	if (hr_dev->caps.qpc_sz < HNS_ROCE_V3_QPC_SZ)
4326 		return;
4327 
4328 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
4329 		hr_reg_enable(&context->ext, QPCEX_STASH);
4330 }
4331 
4332 static void modify_qp_init_to_init(struct ib_qp *ibqp,
4333 				   struct hns_roce_v2_qp_context *context,
4334 				   struct hns_roce_v2_qp_context *qpc_mask)
4335 {
4336 	/*
4337 	 * In v2 engine, software pass context and context mask to hardware
4338 	 * when modifying qp. If software need modify some fields in context,
4339 	 * we should set all bits of the relevant fields in context mask to
4340 	 * 0 at the same time, else set them to 0x1.
4341 	 */
4342 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4343 	hr_reg_clear(qpc_mask, QPC_TST);
4344 
4345 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4346 	hr_reg_clear(qpc_mask, QPC_PD);
4347 
4348 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4349 	hr_reg_clear(qpc_mask, QPC_RX_CQN);
4350 
4351 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4352 	hr_reg_clear(qpc_mask, QPC_TX_CQN);
4353 
4354 	if (ibqp->srq) {
4355 		hr_reg_enable(context, QPC_SRQ_EN);
4356 		hr_reg_clear(qpc_mask, QPC_SRQ_EN);
4357 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4358 		hr_reg_clear(qpc_mask, QPC_SRQN);
4359 	}
4360 }
4361 
4362 static int config_qp_rq_buf(struct hns_roce_dev *hr_dev,
4363 			    struct hns_roce_qp *hr_qp,
4364 			    struct hns_roce_v2_qp_context *context,
4365 			    struct hns_roce_v2_qp_context *qpc_mask)
4366 {
4367 	u64 mtts[MTT_MIN_COUNT] = { 0 };
4368 	u64 wqe_sge_ba;
4369 	int ret;
4370 
4371 	/* Search qp buf's mtts */
4372 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->rq.offset, mtts,
4373 				MTT_MIN_COUNT);
4374 	if (hr_qp->rq.wqe_cnt && ret) {
4375 		ibdev_err(&hr_dev->ib_dev,
4376 			  "failed to find QP(0x%lx) RQ WQE buf, ret = %d.\n",
4377 			  hr_qp->qpn, ret);
4378 		return ret;
4379 	}
4380 
4381 	wqe_sge_ba = hns_roce_get_mtr_ba(&hr_qp->mtr);
4382 
4383 	context->wqe_sge_ba = cpu_to_le32(wqe_sge_ba >> 3);
4384 	qpc_mask->wqe_sge_ba = 0;
4385 
4386 	/*
4387 	 * In v2 engine, software pass context and context mask to hardware
4388 	 * when modifying qp. If software need modify some fields in context,
4389 	 * we should set all bits of the relevant fields in context mask to
4390 	 * 0 at the same time, else set them to 0x1.
4391 	 */
4392 	hr_reg_write(context, QPC_WQE_SGE_BA_H, wqe_sge_ba >> (32 + 3));
4393 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_H);
4394 
4395 	hr_reg_write(context, QPC_SQ_HOP_NUM,
4396 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sq_hop_num,
4397 				      hr_qp->sq.wqe_cnt));
4398 	hr_reg_clear(qpc_mask, QPC_SQ_HOP_NUM);
4399 
4400 	hr_reg_write(context, QPC_SGE_HOP_NUM,
4401 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sge_hop_num,
4402 				      hr_qp->sge.sge_cnt));
4403 	hr_reg_clear(qpc_mask, QPC_SGE_HOP_NUM);
4404 
4405 	hr_reg_write(context, QPC_RQ_HOP_NUM,
4406 		     to_hr_hem_hopnum(hr_dev->caps.wqe_rq_hop_num,
4407 				      hr_qp->rq.wqe_cnt));
4408 
4409 	hr_reg_clear(qpc_mask, QPC_RQ_HOP_NUM);
4410 
4411 	hr_reg_write(context, QPC_WQE_SGE_BA_PG_SZ,
4412 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.ba_pg_shift));
4413 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_PG_SZ);
4414 
4415 	hr_reg_write(context, QPC_WQE_SGE_BUF_PG_SZ,
4416 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.buf_pg_shift));
4417 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BUF_PG_SZ);
4418 
4419 	context->rq_cur_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[0]));
4420 	qpc_mask->rq_cur_blk_addr = 0;
4421 
4422 	hr_reg_write(context, QPC_RQ_CUR_BLK_ADDR_H,
4423 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
4424 	hr_reg_clear(qpc_mask, QPC_RQ_CUR_BLK_ADDR_H);
4425 
4426 	context->rq_nxt_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[1]));
4427 	qpc_mask->rq_nxt_blk_addr = 0;
4428 
4429 	hr_reg_write(context, QPC_RQ_NXT_BLK_ADDR_H,
4430 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
4431 	hr_reg_clear(qpc_mask, QPC_RQ_NXT_BLK_ADDR_H);
4432 
4433 	return 0;
4434 }
4435 
4436 static int config_qp_sq_buf(struct hns_roce_dev *hr_dev,
4437 			    struct hns_roce_qp *hr_qp,
4438 			    struct hns_roce_v2_qp_context *context,
4439 			    struct hns_roce_v2_qp_context *qpc_mask)
4440 {
4441 	struct ib_device *ibdev = &hr_dev->ib_dev;
4442 	u64 sge_cur_blk = 0;
4443 	u64 sq_cur_blk = 0;
4444 	int ret;
4445 
4446 	/* search qp buf's mtts */
4447 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->sq.offset,
4448 				&sq_cur_blk, 1);
4449 	if (ret) {
4450 		ibdev_err(ibdev, "failed to find QP(0x%lx) SQ WQE buf, ret = %d.\n",
4451 			  hr_qp->qpn, ret);
4452 		return ret;
4453 	}
4454 	if (hr_qp->sge.sge_cnt > 0) {
4455 		ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr,
4456 					hr_qp->sge.offset, &sge_cur_blk, 1);
4457 		if (ret) {
4458 			ibdev_err(ibdev, "failed to find QP(0x%lx) SGE buf, ret = %d.\n",
4459 				  hr_qp->qpn, ret);
4460 			return ret;
4461 		}
4462 	}
4463 
4464 	/*
4465 	 * In v2 engine, software pass context and context mask to hardware
4466 	 * when modifying qp. If software need modify some fields in context,
4467 	 * we should set all bits of the relevant fields in context mask to
4468 	 * 0 at the same time, else set them to 0x1.
4469 	 */
4470 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_L,
4471 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4472 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_H,
4473 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4474 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_L);
4475 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_H);
4476 
4477 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_L,
4478 		     lower_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4479 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_H,
4480 		     upper_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4481 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_L);
4482 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_H);
4483 
4484 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_L,
4485 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4486 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_H,
4487 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4488 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_L);
4489 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_H);
4490 
4491 	return 0;
4492 }
4493 
4494 static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
4495 				  const struct ib_qp_attr *attr)
4496 {
4497 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_UD)
4498 		return IB_MTU_4096;
4499 
4500 	return attr->path_mtu;
4501 }
4502 
4503 static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
4504 				 const struct ib_qp_attr *attr, int attr_mask,
4505 				 struct hns_roce_v2_qp_context *context,
4506 				 struct hns_roce_v2_qp_context *qpc_mask,
4507 				 struct ib_udata *udata)
4508 {
4509 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
4510 					  struct hns_roce_ucontext, ibucontext);
4511 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4512 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4513 	struct ib_device *ibdev = &hr_dev->ib_dev;
4514 	dma_addr_t trrl_ba;
4515 	dma_addr_t irrl_ba;
4516 	enum ib_mtu ib_mtu;
4517 	const u8 *smac;
4518 	u8 lp_pktn_ini;
4519 	u64 *mtts;
4520 	u8 *dmac;
4521 	u32 port;
4522 	int mtu;
4523 	int ret;
4524 
4525 	ret = config_qp_rq_buf(hr_dev, hr_qp, context, qpc_mask);
4526 	if (ret) {
4527 		ibdev_err(ibdev, "failed to config rq buf, ret = %d.\n", ret);
4528 		return ret;
4529 	}
4530 
4531 	/* Search IRRL's mtts */
4532 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.irrl_table,
4533 				   hr_qp->qpn, &irrl_ba);
4534 	if (!mtts) {
4535 		ibdev_err(ibdev, "failed to find qp irrl_table.\n");
4536 		return -EINVAL;
4537 	}
4538 
4539 	/* Search TRRL's mtts */
4540 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.trrl_table,
4541 				   hr_qp->qpn, &trrl_ba);
4542 	if (!mtts) {
4543 		ibdev_err(ibdev, "failed to find qp trrl_table.\n");
4544 		return -EINVAL;
4545 	}
4546 
4547 	if (attr_mask & IB_QP_ALT_PATH) {
4548 		ibdev_err(ibdev, "INIT2RTR attr_mask (0x%x) error.\n",
4549 			  attr_mask);
4550 		return -EINVAL;
4551 	}
4552 
4553 	hr_reg_write(context, QPC_TRRL_BA_L, trrl_ba >> QPC_TRRL_BA_L_S);
4554 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_L);
4555 	context->trrl_ba = cpu_to_le32(trrl_ba >> QPC_TRRL_BA_M_S);
4556 	qpc_mask->trrl_ba = 0;
4557 	hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> QPC_TRRL_BA_H_S);
4558 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
4559 
4560 	context->irrl_ba = cpu_to_le32(irrl_ba >> QPC_IRRL_BA_L_S);
4561 	qpc_mask->irrl_ba = 0;
4562 	hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> QPC_IRRL_BA_H_S);
4563 	hr_reg_clear(qpc_mask, QPC_IRRL_BA_H);
4564 
4565 	hr_reg_enable(context, QPC_RMT_E2E);
4566 	hr_reg_clear(qpc_mask, QPC_RMT_E2E);
4567 
4568 	hr_reg_write(context, QPC_SIG_TYPE, hr_qp->sq_signal_bits);
4569 	hr_reg_clear(qpc_mask, QPC_SIG_TYPE);
4570 
4571 	port = (attr_mask & IB_QP_PORT) ? (attr->port_num - 1) : hr_qp->port;
4572 
4573 	smac = (const u8 *)hr_dev->dev_addr[port];
4574 	dmac = (u8 *)attr->ah_attr.roce.dmac;
4575 	/* when dmac equals smac or loop_idc is 1, it should loopback */
4576 	if (ether_addr_equal_unaligned(dmac, smac) ||
4577 	    hr_dev->loop_idc == 0x1) {
4578 		hr_reg_write(context, QPC_LBI, hr_dev->loop_idc);
4579 		hr_reg_clear(qpc_mask, QPC_LBI);
4580 	}
4581 
4582 	if (attr_mask & IB_QP_DEST_QPN) {
4583 		hr_reg_write(context, QPC_DQPN, attr->dest_qp_num);
4584 		hr_reg_clear(qpc_mask, QPC_DQPN);
4585 	}
4586 
4587 	memcpy(&context->dmac, dmac, sizeof(u32));
4588 	hr_reg_write(context, QPC_DMAC_H, *((u16 *)(&dmac[4])));
4589 	qpc_mask->dmac = 0;
4590 	hr_reg_clear(qpc_mask, QPC_DMAC_H);
4591 
4592 	ib_mtu = get_mtu(ibqp, attr);
4593 	hr_qp->path_mtu = ib_mtu;
4594 
4595 	mtu = ib_mtu_enum_to_int(ib_mtu);
4596 	if (WARN_ON(mtu <= 0))
4597 		return -EINVAL;
4598 #define MIN_LP_MSG_LEN 1024
4599 	/* mtu * (2 ^ lp_pktn_ini) should be in the range of 1024 to mtu */
4600 	lp_pktn_ini = ilog2(max(mtu, MIN_LP_MSG_LEN) / mtu);
4601 
4602 	if (attr_mask & IB_QP_PATH_MTU) {
4603 		hr_reg_write(context, QPC_MTU, ib_mtu);
4604 		hr_reg_clear(qpc_mask, QPC_MTU);
4605 	}
4606 
4607 	hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
4608 	hr_reg_clear(qpc_mask, QPC_LP_PKTN_INI);
4609 
4610 	/* ACK_REQ_FREQ should be larger than or equal to LP_PKTN_INI */
4611 	hr_reg_write(context, QPC_ACK_REQ_FREQ, lp_pktn_ini);
4612 	hr_reg_clear(qpc_mask, QPC_ACK_REQ_FREQ);
4613 
4614 	hr_reg_clear(qpc_mask, QPC_RX_REQ_PSN_ERR);
4615 	hr_reg_clear(qpc_mask, QPC_RX_REQ_MSN);
4616 	hr_reg_clear(qpc_mask, QPC_RX_REQ_LAST_OPTYPE);
4617 
4618 	context->rq_rnr_timer = 0;
4619 	qpc_mask->rq_rnr_timer = 0;
4620 
4621 	hr_reg_clear(qpc_mask, QPC_TRRL_HEAD_MAX);
4622 	hr_reg_clear(qpc_mask, QPC_TRRL_TAIL_MAX);
4623 
4624 #define MAX_LP_SGEN 3
4625 	/* rocee send 2^lp_sgen_ini segs every time */
4626 	hr_reg_write(context, QPC_LP_SGEN_INI, MAX_LP_SGEN);
4627 	hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
4628 
4629 	if (udata && ibqp->qp_type == IB_QPT_RC &&
4630 	    (uctx->config & HNS_ROCE_RQ_INLINE_FLAGS)) {
4631 		hr_reg_write_bool(context, QPC_RQIE,
4632 				  hr_dev->caps.flags &
4633 				  HNS_ROCE_CAP_FLAG_RQ_INLINE);
4634 		hr_reg_clear(qpc_mask, QPC_RQIE);
4635 	}
4636 
4637 	if (udata &&
4638 	    (ibqp->qp_type == IB_QPT_RC || ibqp->qp_type == IB_QPT_XRC_TGT) &&
4639 	    (uctx->config & HNS_ROCE_CQE_INLINE_FLAGS)) {
4640 		hr_reg_write_bool(context, QPC_CQEIE,
4641 				  hr_dev->caps.flags &
4642 				  HNS_ROCE_CAP_FLAG_CQE_INLINE);
4643 		hr_reg_clear(qpc_mask, QPC_CQEIE);
4644 
4645 		hr_reg_write(context, QPC_CQEIS, 0);
4646 		hr_reg_clear(qpc_mask, QPC_CQEIS);
4647 	}
4648 
4649 	return 0;
4650 }
4651 
4652 static int modify_qp_rtr_to_rts(struct ib_qp *ibqp, int attr_mask,
4653 				struct hns_roce_v2_qp_context *context,
4654 				struct hns_roce_v2_qp_context *qpc_mask)
4655 {
4656 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4657 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4658 	struct ib_device *ibdev = &hr_dev->ib_dev;
4659 	int ret;
4660 
4661 	/* Not support alternate path and path migration */
4662 	if (attr_mask & (IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE)) {
4663 		ibdev_err(ibdev, "RTR2RTS attr_mask (0x%x)error\n", attr_mask);
4664 		return -EINVAL;
4665 	}
4666 
4667 	ret = config_qp_sq_buf(hr_dev, hr_qp, context, qpc_mask);
4668 	if (ret) {
4669 		ibdev_err(ibdev, "failed to config sq buf, ret = %d.\n", ret);
4670 		return ret;
4671 	}
4672 
4673 	/*
4674 	 * Set some fields in context to zero, Because the default values
4675 	 * of all fields in context are zero, we need not set them to 0 again.
4676 	 * but we should set the relevant fields of context mask to 0.
4677 	 */
4678 	hr_reg_clear(qpc_mask, QPC_IRRL_SGE_IDX);
4679 
4680 	hr_reg_clear(qpc_mask, QPC_RX_ACK_MSN);
4681 
4682 	hr_reg_clear(qpc_mask, QPC_ACK_LAST_OPTYPE);
4683 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN_VLD);
4684 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN);
4685 
4686 	hr_reg_clear(qpc_mask, QPC_IRRL_TAIL_REAL);
4687 
4688 	hr_reg_clear(qpc_mask, QPC_RETRY_MSG_MSN);
4689 
4690 	hr_reg_clear(qpc_mask, QPC_RNR_RETRY_FLAG);
4691 
4692 	hr_reg_clear(qpc_mask, QPC_CHECK_FLG);
4693 
4694 	hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
4695 
4696 	return 0;
4697 }
4698 
4699 static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4700 			   u32 *dip_idx)
4701 {
4702 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4703 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4704 	u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx;
4705 	u32 *head =  &hr_dev->qp_table.idx_table.head;
4706 	u32 *tail =  &hr_dev->qp_table.idx_table.tail;
4707 	struct hns_roce_dip *hr_dip;
4708 	unsigned long flags;
4709 	int ret = 0;
4710 
4711 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
4712 
4713 	spare_idx[*tail] = ibqp->qp_num;
4714 	*tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1);
4715 
4716 	list_for_each_entry(hr_dip, &hr_dev->dip_list, node) {
4717 		if (!memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) {
4718 			*dip_idx = hr_dip->dip_idx;
4719 			goto out;
4720 		}
4721 	}
4722 
4723 	/* If no dgid is found, a new dip and a mapping between dgid and
4724 	 * dip_idx will be created.
4725 	 */
4726 	hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC);
4727 	if (!hr_dip) {
4728 		ret = -ENOMEM;
4729 		goto out;
4730 	}
4731 
4732 	memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4733 	hr_dip->dip_idx = *dip_idx = spare_idx[*head];
4734 	*head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1);
4735 	list_add_tail(&hr_dip->node, &hr_dev->dip_list);
4736 
4737 out:
4738 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
4739 	return ret;
4740 }
4741 
4742 enum {
4743 	CONG_DCQCN,
4744 	CONG_WINDOW,
4745 };
4746 
4747 enum {
4748 	UNSUPPORT_CONG_LEVEL,
4749 	SUPPORT_CONG_LEVEL,
4750 };
4751 
4752 enum {
4753 	CONG_LDCP,
4754 	CONG_HC3,
4755 };
4756 
4757 enum {
4758 	DIP_INVALID,
4759 	DIP_VALID,
4760 };
4761 
4762 enum {
4763 	WND_LIMIT,
4764 	WND_UNLIMIT,
4765 };
4766 
4767 static int check_cong_type(struct ib_qp *ibqp,
4768 			   struct hns_roce_congestion_algorithm *cong_alg)
4769 {
4770 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4771 
4772 	/* different congestion types match different configurations */
4773 	switch (hr_qp->cong_type) {
4774 	case CONG_TYPE_DCQCN:
4775 		cong_alg->alg_sel = CONG_DCQCN;
4776 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4777 		cong_alg->dip_vld = DIP_INVALID;
4778 		cong_alg->wnd_mode_sel = WND_LIMIT;
4779 		break;
4780 	case CONG_TYPE_LDCP:
4781 		cong_alg->alg_sel = CONG_WINDOW;
4782 		cong_alg->alg_sub_sel = CONG_LDCP;
4783 		cong_alg->dip_vld = DIP_INVALID;
4784 		cong_alg->wnd_mode_sel = WND_UNLIMIT;
4785 		break;
4786 	case CONG_TYPE_HC3:
4787 		cong_alg->alg_sel = CONG_WINDOW;
4788 		cong_alg->alg_sub_sel = CONG_HC3;
4789 		cong_alg->dip_vld = DIP_INVALID;
4790 		cong_alg->wnd_mode_sel = WND_LIMIT;
4791 		break;
4792 	case CONG_TYPE_DIP:
4793 		cong_alg->alg_sel = CONG_DCQCN;
4794 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4795 		cong_alg->dip_vld = DIP_VALID;
4796 		cong_alg->wnd_mode_sel = WND_LIMIT;
4797 		break;
4798 	default:
4799 		hr_qp->cong_type = CONG_TYPE_DCQCN;
4800 		cong_alg->alg_sel = CONG_DCQCN;
4801 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4802 		cong_alg->dip_vld = DIP_INVALID;
4803 		cong_alg->wnd_mode_sel = WND_LIMIT;
4804 		break;
4805 	}
4806 
4807 	return 0;
4808 }
4809 
4810 static int fill_cong_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4811 			   struct hns_roce_v2_qp_context *context,
4812 			   struct hns_roce_v2_qp_context *qpc_mask)
4813 {
4814 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4815 	struct hns_roce_congestion_algorithm cong_field;
4816 	struct ib_device *ibdev = ibqp->device;
4817 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
4818 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4819 	u32 dip_idx = 0;
4820 	int ret;
4821 
4822 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ||
4823 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE)
4824 		return 0;
4825 
4826 	ret = check_cong_type(ibqp, &cong_field);
4827 	if (ret)
4828 		return ret;
4829 
4830 	hr_reg_write(context, QPC_CONG_ALGO_TMPL_ID, hr_dev->cong_algo_tmpl_id +
4831 		     hr_qp->cong_type * HNS_ROCE_CONG_SIZE);
4832 	hr_reg_clear(qpc_mask, QPC_CONG_ALGO_TMPL_ID);
4833 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SEL, cong_field.alg_sel);
4834 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SEL);
4835 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SUB_SEL,
4836 		     cong_field.alg_sub_sel);
4837 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SUB_SEL);
4838 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX_VLD, cong_field.dip_vld);
4839 	hr_reg_clear(&qpc_mask->ext, QPCEX_DIP_CTX_IDX_VLD);
4840 	hr_reg_write(&context->ext, QPCEX_SQ_RQ_NOT_FORBID_EN,
4841 		     cong_field.wnd_mode_sel);
4842 	hr_reg_clear(&qpc_mask->ext, QPCEX_SQ_RQ_NOT_FORBID_EN);
4843 
4844 	/* if dip is disabled, there is no need to set dip idx */
4845 	if (cong_field.dip_vld == 0)
4846 		return 0;
4847 
4848 	ret = get_dip_ctx_idx(ibqp, attr, &dip_idx);
4849 	if (ret) {
4850 		ibdev_err(ibdev, "failed to fill cong field, ret = %d.\n", ret);
4851 		return ret;
4852 	}
4853 
4854 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX, dip_idx);
4855 	hr_reg_write(&qpc_mask->ext, QPCEX_DIP_CTX_IDX, 0);
4856 
4857 	return 0;
4858 }
4859 
4860 static int hns_roce_hw_v2_get_dscp(struct hns_roce_dev *hr_dev, u8 dscp,
4861 				   u8 *tc_mode, u8 *priority)
4862 {
4863 	struct hns_roce_v2_priv *priv = hr_dev->priv;
4864 	struct hnae3_handle *handle = priv->handle;
4865 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
4866 
4867 	if (!ops->get_dscp_prio)
4868 		return -EOPNOTSUPP;
4869 
4870 	return ops->get_dscp_prio(handle, dscp, tc_mode, priority);
4871 }
4872 
4873 bool check_sl_valid(struct hns_roce_dev *hr_dev, u8 sl)
4874 {
4875 	u32 max_sl;
4876 
4877 	max_sl = min_t(u32, MAX_SERVICE_LEVEL, hr_dev->caps.sl_num - 1);
4878 	if (unlikely(sl > max_sl)) {
4879 		ibdev_err_ratelimited(&hr_dev->ib_dev,
4880 				      "failed to set SL(%u). Shouldn't be larger than %u.\n",
4881 				      sl, max_sl);
4882 		return false;
4883 	}
4884 
4885 	return true;
4886 }
4887 
4888 static int hns_roce_set_sl(struct ib_qp *ibqp,
4889 			   const struct ib_qp_attr *attr,
4890 			   struct hns_roce_v2_qp_context *context,
4891 			   struct hns_roce_v2_qp_context *qpc_mask)
4892 {
4893 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4894 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4895 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4896 	struct ib_device *ibdev = &hr_dev->ib_dev;
4897 	int ret;
4898 
4899 	ret = hns_roce_hw_v2_get_dscp(hr_dev, get_tclass(&attr->ah_attr.grh),
4900 				      &hr_qp->tc_mode, &hr_qp->priority);
4901 	if (ret && ret != -EOPNOTSUPP &&
4902 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4903 		ibdev_err_ratelimited(ibdev,
4904 				      "failed to get dscp, ret = %d.\n", ret);
4905 		return ret;
4906 	}
4907 
4908 	if (hr_qp->tc_mode == HNAE3_TC_MAP_MODE_DSCP &&
4909 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4910 		hr_qp->sl = hr_qp->priority;
4911 	else
4912 		hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4913 
4914 	if (!check_sl_valid(hr_dev, hr_qp->sl))
4915 		return -EINVAL;
4916 
4917 	hr_reg_write(context, QPC_SL, hr_qp->sl);
4918 	hr_reg_clear(qpc_mask, QPC_SL);
4919 
4920 	return 0;
4921 }
4922 
4923 static int hns_roce_v2_set_path(struct ib_qp *ibqp,
4924 				const struct ib_qp_attr *attr,
4925 				int attr_mask,
4926 				struct hns_roce_v2_qp_context *context,
4927 				struct hns_roce_v2_qp_context *qpc_mask)
4928 {
4929 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4930 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4931 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4932 	struct ib_device *ibdev = &hr_dev->ib_dev;
4933 	const struct ib_gid_attr *gid_attr = NULL;
4934 	u8 sl = rdma_ah_get_sl(&attr->ah_attr);
4935 	int is_roce_protocol;
4936 	u16 vlan_id = 0xffff;
4937 	bool is_udp = false;
4938 	u8 ib_port;
4939 	u8 hr_port;
4940 	int ret;
4941 
4942 	/*
4943 	 * If free_mr_en of qp is set, it means that this qp comes from
4944 	 * free mr. This qp will perform the loopback operation.
4945 	 * In the loopback scenario, only sl needs to be set.
4946 	 */
4947 	if (hr_qp->free_mr_en) {
4948 		if (!check_sl_valid(hr_dev, sl))
4949 			return -EINVAL;
4950 		hr_reg_write(context, QPC_SL, sl);
4951 		hr_reg_clear(qpc_mask, QPC_SL);
4952 		hr_qp->sl = sl;
4953 		return 0;
4954 	}
4955 
4956 	ib_port = (attr_mask & IB_QP_PORT) ? attr->port_num : hr_qp->port + 1;
4957 	hr_port = ib_port - 1;
4958 	is_roce_protocol = rdma_cap_eth_ah(&hr_dev->ib_dev, ib_port) &&
4959 			   rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
4960 
4961 	if (is_roce_protocol) {
4962 		gid_attr = attr->ah_attr.grh.sgid_attr;
4963 		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
4964 		if (ret)
4965 			return ret;
4966 
4967 		is_udp = (gid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
4968 	}
4969 
4970 	/* Only HIP08 needs to set the vlan_en bits in QPC */
4971 	if (vlan_id < VLAN_N_VID &&
4972 	    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4973 		hr_reg_enable(context, QPC_RQ_VLAN_EN);
4974 		hr_reg_clear(qpc_mask, QPC_RQ_VLAN_EN);
4975 		hr_reg_enable(context, QPC_SQ_VLAN_EN);
4976 		hr_reg_clear(qpc_mask, QPC_SQ_VLAN_EN);
4977 	}
4978 
4979 	hr_reg_write(context, QPC_VLAN_ID, vlan_id);
4980 	hr_reg_clear(qpc_mask, QPC_VLAN_ID);
4981 
4982 	if (grh->sgid_index >= hr_dev->caps.gid_table_len[hr_port]) {
4983 		ibdev_err(ibdev, "sgid_index(%u) too large. max is %d\n",
4984 			  grh->sgid_index, hr_dev->caps.gid_table_len[hr_port]);
4985 		return -EINVAL;
4986 	}
4987 
4988 	if (attr->ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE) {
4989 		ibdev_err(ibdev, "ah attr is not RDMA roce type\n");
4990 		return -EINVAL;
4991 	}
4992 
4993 	hr_reg_write(context, QPC_UDPSPN,
4994 		     is_udp ? rdma_get_udp_sport(grh->flow_label, ibqp->qp_num,
4995 						 attr->dest_qp_num) :
4996 				    0);
4997 
4998 	hr_reg_clear(qpc_mask, QPC_UDPSPN);
4999 
5000 	hr_reg_write(context, QPC_GMV_IDX, grh->sgid_index);
5001 
5002 	hr_reg_clear(qpc_mask, QPC_GMV_IDX);
5003 
5004 	hr_reg_write(context, QPC_HOPLIMIT, grh->hop_limit);
5005 	hr_reg_clear(qpc_mask, QPC_HOPLIMIT);
5006 
5007 	ret = fill_cong_field(ibqp, attr, context, qpc_mask);
5008 	if (ret)
5009 		return ret;
5010 
5011 	hr_reg_write(context, QPC_TC, get_tclass(&attr->ah_attr.grh));
5012 	hr_reg_clear(qpc_mask, QPC_TC);
5013 
5014 	hr_reg_write(context, QPC_FL, grh->flow_label);
5015 	hr_reg_clear(qpc_mask, QPC_FL);
5016 	memcpy(context->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
5017 	memset(qpc_mask->dgid, 0, sizeof(grh->dgid.raw));
5018 
5019 	return  hns_roce_set_sl(ibqp, attr, context, qpc_mask);
5020 }
5021 
5022 static bool check_qp_state(enum ib_qp_state cur_state,
5023 			   enum ib_qp_state new_state)
5024 {
5025 	static const bool sm[][IB_QPS_ERR + 1] = {
5026 		[IB_QPS_RESET] = { [IB_QPS_RESET] = true,
5027 				   [IB_QPS_INIT] = true },
5028 		[IB_QPS_INIT] = { [IB_QPS_RESET] = true,
5029 				  [IB_QPS_INIT] = true,
5030 				  [IB_QPS_RTR] = true,
5031 				  [IB_QPS_ERR] = true },
5032 		[IB_QPS_RTR] = { [IB_QPS_RESET] = true,
5033 				 [IB_QPS_RTS] = true,
5034 				 [IB_QPS_ERR] = true },
5035 		[IB_QPS_RTS] = { [IB_QPS_RESET] = true,
5036 				 [IB_QPS_RTS] = true,
5037 				 [IB_QPS_ERR] = true },
5038 		[IB_QPS_SQD] = {},
5039 		[IB_QPS_SQE] = {},
5040 		[IB_QPS_ERR] = { [IB_QPS_RESET] = true,
5041 				 [IB_QPS_ERR] = true }
5042 	};
5043 
5044 	return sm[cur_state][new_state];
5045 }
5046 
5047 static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
5048 				      const struct ib_qp_attr *attr,
5049 				      int attr_mask,
5050 				      enum ib_qp_state cur_state,
5051 				      enum ib_qp_state new_state,
5052 				      struct hns_roce_v2_qp_context *context,
5053 				      struct hns_roce_v2_qp_context *qpc_mask,
5054 				      struct ib_udata *udata)
5055 {
5056 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5057 	int ret = 0;
5058 
5059 	if (!check_qp_state(cur_state, new_state)) {
5060 		ibdev_err(&hr_dev->ib_dev, "Illegal state for QP!\n");
5061 		return -EINVAL;
5062 	}
5063 
5064 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
5065 		memset(qpc_mask, 0, hr_dev->caps.qpc_sz);
5066 		modify_qp_reset_to_init(ibqp, context, qpc_mask);
5067 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
5068 		modify_qp_init_to_init(ibqp, context, qpc_mask);
5069 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
5070 		ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
5071 					    qpc_mask, udata);
5072 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
5073 		ret = modify_qp_rtr_to_rts(ibqp, attr_mask, context, qpc_mask);
5074 	}
5075 
5076 	return ret;
5077 }
5078 
5079 static bool check_qp_timeout_cfg_range(struct hns_roce_dev *hr_dev, u8 *timeout)
5080 {
5081 #define QP_ACK_TIMEOUT_MAX_HIP08 20
5082 #define QP_ACK_TIMEOUT_MAX 31
5083 
5084 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5085 		if (*timeout > QP_ACK_TIMEOUT_MAX_HIP08) {
5086 			ibdev_warn(&hr_dev->ib_dev,
5087 				   "local ACK timeout shall be 0 to 20.\n");
5088 			return false;
5089 		}
5090 		*timeout += HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5091 	} else if (hr_dev->pci_dev->revision > PCI_REVISION_ID_HIP08) {
5092 		if (*timeout > QP_ACK_TIMEOUT_MAX) {
5093 			ibdev_warn(&hr_dev->ib_dev,
5094 				   "local ACK timeout shall be 0 to 31.\n");
5095 			return false;
5096 		}
5097 	}
5098 
5099 	return true;
5100 }
5101 
5102 static int hns_roce_v2_set_opt_fields(struct ib_qp *ibqp,
5103 				      const struct ib_qp_attr *attr,
5104 				      int attr_mask,
5105 				      struct hns_roce_v2_qp_context *context,
5106 				      struct hns_roce_v2_qp_context *qpc_mask)
5107 {
5108 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5109 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5110 	int ret = 0;
5111 	u8 timeout;
5112 
5113 	if (attr_mask & IB_QP_AV) {
5114 		ret = hns_roce_v2_set_path(ibqp, attr, attr_mask, context,
5115 					   qpc_mask);
5116 		if (ret)
5117 			return ret;
5118 	}
5119 
5120 	if (attr_mask & IB_QP_TIMEOUT) {
5121 		timeout = attr->timeout;
5122 		if (check_qp_timeout_cfg_range(hr_dev, &timeout)) {
5123 			hr_reg_write(context, QPC_AT, timeout);
5124 			hr_reg_clear(qpc_mask, QPC_AT);
5125 		}
5126 	}
5127 
5128 	if (attr_mask & IB_QP_RETRY_CNT) {
5129 		hr_reg_write(context, QPC_RETRY_NUM_INIT, attr->retry_cnt);
5130 		hr_reg_clear(qpc_mask, QPC_RETRY_NUM_INIT);
5131 
5132 		hr_reg_write(context, QPC_RETRY_CNT, attr->retry_cnt);
5133 		hr_reg_clear(qpc_mask, QPC_RETRY_CNT);
5134 	}
5135 
5136 	if (attr_mask & IB_QP_RNR_RETRY) {
5137 		hr_reg_write(context, QPC_RNR_NUM_INIT, attr->rnr_retry);
5138 		hr_reg_clear(qpc_mask, QPC_RNR_NUM_INIT);
5139 
5140 		hr_reg_write(context, QPC_RNR_CNT, attr->rnr_retry);
5141 		hr_reg_clear(qpc_mask, QPC_RNR_CNT);
5142 	}
5143 
5144 	if (attr_mask & IB_QP_SQ_PSN) {
5145 		hr_reg_write(context, QPC_SQ_CUR_PSN, attr->sq_psn);
5146 		hr_reg_clear(qpc_mask, QPC_SQ_CUR_PSN);
5147 
5148 		hr_reg_write(context, QPC_SQ_MAX_PSN, attr->sq_psn);
5149 		hr_reg_clear(qpc_mask, QPC_SQ_MAX_PSN);
5150 
5151 		hr_reg_write(context, QPC_RETRY_MSG_PSN_L, attr->sq_psn);
5152 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_L);
5153 
5154 		hr_reg_write(context, QPC_RETRY_MSG_PSN_H,
5155 			     attr->sq_psn >> RETRY_MSG_PSN_SHIFT);
5156 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_H);
5157 
5158 		hr_reg_write(context, QPC_RETRY_MSG_FPKT_PSN, attr->sq_psn);
5159 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_FPKT_PSN);
5160 
5161 		hr_reg_write(context, QPC_RX_ACK_EPSN, attr->sq_psn);
5162 		hr_reg_clear(qpc_mask, QPC_RX_ACK_EPSN);
5163 	}
5164 
5165 	if ((attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) &&
5166 	     attr->max_dest_rd_atomic) {
5167 		hr_reg_write(context, QPC_RR_MAX,
5168 			     fls(attr->max_dest_rd_atomic - 1));
5169 		hr_reg_clear(qpc_mask, QPC_RR_MAX);
5170 	}
5171 
5172 	if ((attr_mask & IB_QP_MAX_QP_RD_ATOMIC) && attr->max_rd_atomic) {
5173 		hr_reg_write(context, QPC_SR_MAX, fls(attr->max_rd_atomic - 1));
5174 		hr_reg_clear(qpc_mask, QPC_SR_MAX);
5175 	}
5176 
5177 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
5178 		set_access_flags(hr_qp, context, qpc_mask, attr, attr_mask);
5179 
5180 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
5181 		hr_reg_write(context, QPC_MIN_RNR_TIME,
5182 			    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ?
5183 			    HNS_ROCE_RNR_TIMER_10NS : attr->min_rnr_timer);
5184 		hr_reg_clear(qpc_mask, QPC_MIN_RNR_TIME);
5185 	}
5186 
5187 	if (attr_mask & IB_QP_RQ_PSN) {
5188 		hr_reg_write(context, QPC_RX_REQ_EPSN, attr->rq_psn);
5189 		hr_reg_clear(qpc_mask, QPC_RX_REQ_EPSN);
5190 
5191 		hr_reg_write(context, QPC_RAQ_PSN, attr->rq_psn - 1);
5192 		hr_reg_clear(qpc_mask, QPC_RAQ_PSN);
5193 	}
5194 
5195 	if (attr_mask & IB_QP_QKEY) {
5196 		context->qkey_xrcd = cpu_to_le32(attr->qkey);
5197 		qpc_mask->qkey_xrcd = 0;
5198 		hr_qp->qkey = attr->qkey;
5199 	}
5200 
5201 	return ret;
5202 }
5203 
5204 static void hns_roce_v2_record_opt_fields(struct ib_qp *ibqp,
5205 					  const struct ib_qp_attr *attr,
5206 					  int attr_mask)
5207 {
5208 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5209 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5210 
5211 	if (attr_mask & IB_QP_ACCESS_FLAGS)
5212 		hr_qp->atomic_rd_en = attr->qp_access_flags;
5213 
5214 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
5215 		hr_qp->resp_depth = attr->max_dest_rd_atomic;
5216 	if (attr_mask & IB_QP_PORT) {
5217 		hr_qp->port = attr->port_num - 1;
5218 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
5219 	}
5220 }
5221 
5222 static void clear_qp(struct hns_roce_qp *hr_qp)
5223 {
5224 	struct ib_qp *ibqp = &hr_qp->ibqp;
5225 
5226 	if (ibqp->send_cq)
5227 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq),
5228 				     hr_qp->qpn, NULL);
5229 
5230 	if (ibqp->recv_cq  && ibqp->recv_cq != ibqp->send_cq)
5231 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->recv_cq),
5232 				     hr_qp->qpn, ibqp->srq ?
5233 				     to_hr_srq(ibqp->srq) : NULL);
5234 
5235 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
5236 		*hr_qp->rdb.db_record = 0;
5237 
5238 	hr_qp->rq.head = 0;
5239 	hr_qp->rq.tail = 0;
5240 	hr_qp->sq.head = 0;
5241 	hr_qp->sq.tail = 0;
5242 	hr_qp->next_sge = 0;
5243 }
5244 
5245 static void v2_set_flushed_fields(struct ib_qp *ibqp,
5246 				  struct hns_roce_v2_qp_context *context,
5247 				  struct hns_roce_v2_qp_context *qpc_mask)
5248 {
5249 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5250 	unsigned long sq_flag = 0;
5251 	unsigned long rq_flag = 0;
5252 
5253 	if (ibqp->qp_type == IB_QPT_XRC_TGT)
5254 		return;
5255 
5256 	spin_lock_irqsave(&hr_qp->sq.lock, sq_flag);
5257 	hr_reg_write(context, QPC_SQ_PRODUCER_IDX, hr_qp->sq.head);
5258 	hr_reg_clear(qpc_mask, QPC_SQ_PRODUCER_IDX);
5259 	hr_qp->state = IB_QPS_ERR;
5260 	spin_unlock_irqrestore(&hr_qp->sq.lock, sq_flag);
5261 
5262 	if (ibqp->srq || ibqp->qp_type == IB_QPT_XRC_INI) /* no RQ */
5263 		return;
5264 
5265 	spin_lock_irqsave(&hr_qp->rq.lock, rq_flag);
5266 	hr_reg_write(context, QPC_RQ_PRODUCER_IDX, hr_qp->rq.head);
5267 	hr_reg_clear(qpc_mask, QPC_RQ_PRODUCER_IDX);
5268 	spin_unlock_irqrestore(&hr_qp->rq.lock, rq_flag);
5269 }
5270 
5271 static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
5272 				 const struct ib_qp_attr *attr,
5273 				 int attr_mask, enum ib_qp_state cur_state,
5274 				 enum ib_qp_state new_state, struct ib_udata *udata)
5275 {
5276 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5277 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5278 	struct hns_roce_v2_qp_context ctx[2];
5279 	struct hns_roce_v2_qp_context *context = ctx;
5280 	struct hns_roce_v2_qp_context *qpc_mask = ctx + 1;
5281 	struct ib_device *ibdev = &hr_dev->ib_dev;
5282 	int ret;
5283 
5284 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
5285 		return -EOPNOTSUPP;
5286 
5287 	/*
5288 	 * In v2 engine, software pass context and context mask to hardware
5289 	 * when modifying qp. If software need modify some fields in context,
5290 	 * we should set all bits of the relevant fields in context mask to
5291 	 * 0 at the same time, else set them to 0x1.
5292 	 */
5293 	memset(context, 0, hr_dev->caps.qpc_sz);
5294 	memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
5295 
5296 	ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
5297 					 new_state, context, qpc_mask, udata);
5298 	if (ret)
5299 		goto out;
5300 
5301 	/* When QP state is err, SQ and RQ WQE should be flushed */
5302 	if (new_state == IB_QPS_ERR)
5303 		v2_set_flushed_fields(ibqp, context, qpc_mask);
5304 
5305 	/* Configure the optional fields */
5306 	ret = hns_roce_v2_set_opt_fields(ibqp, attr, attr_mask, context,
5307 					 qpc_mask);
5308 	if (ret)
5309 		goto out;
5310 
5311 	hr_reg_write_bool(context, QPC_INV_CREDIT,
5312 			  to_hr_qp_type(hr_qp->ibqp.qp_type) == SERV_TYPE_XRC ||
5313 			  ibqp->srq);
5314 	hr_reg_clear(qpc_mask, QPC_INV_CREDIT);
5315 
5316 	/* Every status migrate must change state */
5317 	hr_reg_write(context, QPC_QP_ST, new_state);
5318 	hr_reg_clear(qpc_mask, QPC_QP_ST);
5319 
5320 	/* SW pass context to HW */
5321 	ret = hns_roce_v2_qp_modify(hr_dev, context, qpc_mask, hr_qp);
5322 	if (ret) {
5323 		ibdev_err(ibdev, "failed to modify QP, ret = %d.\n", ret);
5324 		goto out;
5325 	}
5326 
5327 	hr_qp->state = new_state;
5328 
5329 	hns_roce_v2_record_opt_fields(ibqp, attr, attr_mask);
5330 
5331 	if (new_state == IB_QPS_RESET && !ibqp->uobject)
5332 		clear_qp(hr_qp);
5333 
5334 out:
5335 	return ret;
5336 }
5337 
5338 static int to_ib_qp_st(enum hns_roce_v2_qp_state state)
5339 {
5340 	static const enum ib_qp_state map[] = {
5341 		[HNS_ROCE_QP_ST_RST] = IB_QPS_RESET,
5342 		[HNS_ROCE_QP_ST_INIT] = IB_QPS_INIT,
5343 		[HNS_ROCE_QP_ST_RTR] = IB_QPS_RTR,
5344 		[HNS_ROCE_QP_ST_RTS] = IB_QPS_RTS,
5345 		[HNS_ROCE_QP_ST_SQD] = IB_QPS_SQD,
5346 		[HNS_ROCE_QP_ST_SQER] = IB_QPS_SQE,
5347 		[HNS_ROCE_QP_ST_ERR] = IB_QPS_ERR,
5348 		[HNS_ROCE_QP_ST_SQ_DRAINING] = IB_QPS_SQD
5349 	};
5350 
5351 	return (state < ARRAY_SIZE(map)) ? map[state] : -1;
5352 }
5353 
5354 static int hns_roce_v2_query_qpc(struct hns_roce_dev *hr_dev, u32 qpn,
5355 				 void *buffer)
5356 {
5357 	struct hns_roce_cmd_mailbox *mailbox;
5358 	int ret;
5359 
5360 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5361 	if (IS_ERR(mailbox))
5362 		return PTR_ERR(mailbox);
5363 
5364 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_QPC,
5365 				qpn);
5366 	if (ret)
5367 		goto out;
5368 
5369 	memcpy(buffer, mailbox->buf, hr_dev->caps.qpc_sz);
5370 
5371 out:
5372 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5373 	return ret;
5374 }
5375 
5376 static int hns_roce_v2_query_srqc(struct hns_roce_dev *hr_dev, u32 srqn,
5377 				 void *buffer)
5378 {
5379 	struct hns_roce_srq_context *context;
5380 	struct hns_roce_cmd_mailbox *mailbox;
5381 	int ret;
5382 
5383 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5384 	if (IS_ERR(mailbox))
5385 		return PTR_ERR(mailbox);
5386 
5387 	context = mailbox->buf;
5388 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SRQC,
5389 				srqn);
5390 	if (ret)
5391 		goto out;
5392 
5393 	memcpy(buffer, context, sizeof(*context));
5394 
5395 out:
5396 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5397 	return ret;
5398 }
5399 
5400 static int hns_roce_v2_query_sccc(struct hns_roce_dev *hr_dev, u32 qpn,
5401 				  void *buffer)
5402 {
5403 	struct hns_roce_v2_scc_context *context;
5404 	struct hns_roce_cmd_mailbox *mailbox;
5405 	int ret;
5406 
5407 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5408 	if (IS_ERR(mailbox))
5409 		return PTR_ERR(mailbox);
5410 
5411 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SCCC,
5412 				qpn);
5413 	if (ret)
5414 		goto out;
5415 
5416 	context = mailbox->buf;
5417 	memcpy(buffer, context, sizeof(*context));
5418 
5419 out:
5420 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5421 	return ret;
5422 }
5423 
5424 static u8 get_qp_timeout_attr(struct hns_roce_dev *hr_dev,
5425 			      struct hns_roce_v2_qp_context *context)
5426 {
5427 	u8 timeout;
5428 
5429 	timeout = (u8)hr_reg_read(context, QPC_AT);
5430 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
5431 		timeout -= HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5432 
5433 	return timeout;
5434 }
5435 
5436 static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5437 				int qp_attr_mask,
5438 				struct ib_qp_init_attr *qp_init_attr)
5439 {
5440 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5441 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5442 	struct hns_roce_v2_qp_context context = {};
5443 	struct ib_device *ibdev = &hr_dev->ib_dev;
5444 	int tmp_qp_state;
5445 	int state;
5446 	int ret;
5447 
5448 	memset(qp_attr, 0, sizeof(*qp_attr));
5449 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5450 
5451 	mutex_lock(&hr_qp->mutex);
5452 
5453 	if (hr_qp->state == IB_QPS_RESET) {
5454 		qp_attr->qp_state = IB_QPS_RESET;
5455 		ret = 0;
5456 		goto done;
5457 	}
5458 
5459 	ret = hns_roce_v2_query_qpc(hr_dev, hr_qp->qpn, &context);
5460 	if (ret) {
5461 		ibdev_err(ibdev, "failed to query QPC, ret = %d.\n", ret);
5462 		ret = -EINVAL;
5463 		goto out;
5464 	}
5465 
5466 	state = hr_reg_read(&context, QPC_QP_ST);
5467 	tmp_qp_state = to_ib_qp_st((enum hns_roce_v2_qp_state)state);
5468 	if (tmp_qp_state == -1) {
5469 		ibdev_err(ibdev, "Illegal ib_qp_state\n");
5470 		ret = -EINVAL;
5471 		goto out;
5472 	}
5473 	hr_qp->state = (u8)tmp_qp_state;
5474 	qp_attr->qp_state = (enum ib_qp_state)hr_qp->state;
5475 	qp_attr->path_mtu = (enum ib_mtu)hr_reg_read(&context, QPC_MTU);
5476 	qp_attr->path_mig_state = IB_MIG_ARMED;
5477 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
5478 	if (hr_qp->ibqp.qp_type == IB_QPT_UD)
5479 		qp_attr->qkey = le32_to_cpu(context.qkey_xrcd);
5480 
5481 	qp_attr->rq_psn = hr_reg_read(&context, QPC_RX_REQ_EPSN);
5482 	qp_attr->sq_psn = (u32)hr_reg_read(&context, QPC_SQ_CUR_PSN);
5483 	qp_attr->dest_qp_num = hr_reg_read(&context, QPC_DQPN);
5484 	qp_attr->qp_access_flags =
5485 		((hr_reg_read(&context, QPC_RRE)) << V2_QP_RRE_S) |
5486 		((hr_reg_read(&context, QPC_RWE)) << V2_QP_RWE_S) |
5487 		((hr_reg_read(&context, QPC_ATE)) << V2_QP_ATE_S);
5488 
5489 	if (hr_qp->ibqp.qp_type == IB_QPT_RC ||
5490 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5491 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
5492 		struct ib_global_route *grh =
5493 			rdma_ah_retrieve_grh(&qp_attr->ah_attr);
5494 
5495 		rdma_ah_set_sl(&qp_attr->ah_attr,
5496 			       hr_reg_read(&context, QPC_SL));
5497 		rdma_ah_set_port_num(&qp_attr->ah_attr, hr_qp->port + 1);
5498 		rdma_ah_set_ah_flags(&qp_attr->ah_attr, IB_AH_GRH);
5499 		grh->flow_label = hr_reg_read(&context, QPC_FL);
5500 		grh->sgid_index = hr_reg_read(&context, QPC_GMV_IDX);
5501 		grh->hop_limit = hr_reg_read(&context, QPC_HOPLIMIT);
5502 		grh->traffic_class = hr_reg_read(&context, QPC_TC);
5503 
5504 		memcpy(grh->dgid.raw, context.dgid, sizeof(grh->dgid.raw));
5505 	}
5506 
5507 	qp_attr->port_num = hr_qp->port + 1;
5508 	qp_attr->sq_draining = 0;
5509 	qp_attr->max_rd_atomic = 1 << hr_reg_read(&context, QPC_SR_MAX);
5510 	qp_attr->max_dest_rd_atomic = 1 << hr_reg_read(&context, QPC_RR_MAX);
5511 
5512 	qp_attr->min_rnr_timer = (u8)hr_reg_read(&context, QPC_MIN_RNR_TIME);
5513 	qp_attr->timeout = get_qp_timeout_attr(hr_dev, &context);
5514 	qp_attr->retry_cnt = hr_reg_read(&context, QPC_RETRY_NUM_INIT);
5515 	qp_attr->rnr_retry = hr_reg_read(&context, QPC_RNR_NUM_INIT);
5516 
5517 done:
5518 	qp_attr->cur_qp_state = qp_attr->qp_state;
5519 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
5520 	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
5521 	qp_attr->cap.max_inline_data = hr_qp->max_inline_data;
5522 
5523 	qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
5524 	qp_attr->cap.max_send_sge = hr_qp->sq.max_gs;
5525 
5526 	qp_init_attr->qp_context = ibqp->qp_context;
5527 	qp_init_attr->qp_type = ibqp->qp_type;
5528 	qp_init_attr->recv_cq = ibqp->recv_cq;
5529 	qp_init_attr->send_cq = ibqp->send_cq;
5530 	qp_init_attr->srq = ibqp->srq;
5531 	qp_init_attr->cap = qp_attr->cap;
5532 	qp_init_attr->sq_sig_type = hr_qp->sq_signal_bits;
5533 
5534 out:
5535 	mutex_unlock(&hr_qp->mutex);
5536 	return ret;
5537 }
5538 
5539 static inline int modify_qp_is_ok(struct hns_roce_qp *hr_qp)
5540 {
5541 	return ((hr_qp->ibqp.qp_type == IB_QPT_RC ||
5542 		 hr_qp->ibqp.qp_type == IB_QPT_UD ||
5543 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5544 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
5545 		hr_qp->state != IB_QPS_RESET);
5546 }
5547 
5548 static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
5549 					 struct hns_roce_qp *hr_qp,
5550 					 struct ib_udata *udata)
5551 {
5552 	struct ib_device *ibdev = &hr_dev->ib_dev;
5553 	struct hns_roce_cq *send_cq, *recv_cq;
5554 	unsigned long flags;
5555 	int ret = 0;
5556 
5557 	if (modify_qp_is_ok(hr_qp)) {
5558 		/* Modify qp to reset before destroying qp */
5559 		ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
5560 					    hr_qp->state, IB_QPS_RESET, udata);
5561 		if (ret)
5562 			ibdev_err(ibdev,
5563 				  "failed to modify QP to RST, ret = %d.\n",
5564 				  ret);
5565 	}
5566 
5567 	send_cq = hr_qp->ibqp.send_cq ? to_hr_cq(hr_qp->ibqp.send_cq) : NULL;
5568 	recv_cq = hr_qp->ibqp.recv_cq ? to_hr_cq(hr_qp->ibqp.recv_cq) : NULL;
5569 
5570 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
5571 	hns_roce_lock_cqs(send_cq, recv_cq);
5572 
5573 	if (!udata) {
5574 		if (recv_cq)
5575 			__hns_roce_v2_cq_clean(recv_cq, hr_qp->qpn,
5576 					       (hr_qp->ibqp.srq ?
5577 						to_hr_srq(hr_qp->ibqp.srq) :
5578 						NULL));
5579 
5580 		if (send_cq && send_cq != recv_cq)
5581 			__hns_roce_v2_cq_clean(send_cq, hr_qp->qpn, NULL);
5582 	}
5583 
5584 	hns_roce_qp_remove(hr_dev, hr_qp);
5585 
5586 	hns_roce_unlock_cqs(send_cq, recv_cq);
5587 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
5588 
5589 	return ret;
5590 }
5591 
5592 int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
5593 {
5594 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5595 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5596 	int ret;
5597 
5598 	ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata);
5599 	if (ret)
5600 		ibdev_err(&hr_dev->ib_dev,
5601 			  "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n",
5602 			  hr_qp->qpn, ret);
5603 
5604 	hns_roce_qp_destroy(hr_dev, hr_qp, udata);
5605 
5606 	return 0;
5607 }
5608 
5609 static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
5610 					    struct hns_roce_qp *hr_qp)
5611 {
5612 	struct ib_device *ibdev = &hr_dev->ib_dev;
5613 	struct hns_roce_sccc_clr_done *resp;
5614 	struct hns_roce_sccc_clr *clr;
5615 	struct hns_roce_cmq_desc desc;
5616 	int ret, i;
5617 
5618 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
5619 		return 0;
5620 
5621 	mutex_lock(&hr_dev->qp_table.scc_mutex);
5622 
5623 	/* set scc ctx clear done flag */
5624 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_RESET_SCCC, false);
5625 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5626 	if (ret) {
5627 		ibdev_err(ibdev, "failed to reset SCC ctx, ret = %d.\n", ret);
5628 		goto out;
5629 	}
5630 
5631 	/* clear scc context */
5632 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLR_SCCC, false);
5633 	clr = (struct hns_roce_sccc_clr *)desc.data;
5634 	clr->qpn = cpu_to_le32(hr_qp->qpn);
5635 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5636 	if (ret) {
5637 		ibdev_err(ibdev, "failed to clear SCC ctx, ret = %d.\n", ret);
5638 		goto out;
5639 	}
5640 
5641 	/* query scc context clear is done or not */
5642 	resp = (struct hns_roce_sccc_clr_done *)desc.data;
5643 	for (i = 0; i <= HNS_ROCE_CMQ_SCC_CLR_DONE_CNT; i++) {
5644 		hns_roce_cmq_setup_basic_desc(&desc,
5645 					      HNS_ROCE_OPC_QUERY_SCCC, true);
5646 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
5647 		if (ret) {
5648 			ibdev_err(ibdev, "failed to query clr cmq, ret = %d\n",
5649 				  ret);
5650 			goto out;
5651 		}
5652 
5653 		if (resp->clr_done)
5654 			goto out;
5655 
5656 		msleep(20);
5657 	}
5658 
5659 	ibdev_err(ibdev, "query SCC clr done flag overtime.\n");
5660 	ret = -ETIMEDOUT;
5661 
5662 out:
5663 	mutex_unlock(&hr_dev->qp_table.scc_mutex);
5664 	return ret;
5665 }
5666 
5667 #define DMA_IDX_SHIFT 3
5668 #define DMA_WQE_SHIFT 3
5669 
5670 static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
5671 					      struct hns_roce_srq_context *ctx)
5672 {
5673 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
5674 	struct ib_device *ibdev = srq->ibsrq.device;
5675 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5676 	u64 mtts_idx[MTT_MIN_COUNT] = {};
5677 	dma_addr_t dma_handle_idx;
5678 	int ret;
5679 
5680 	/* Get physical address of idx que buf */
5681 	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
5682 				ARRAY_SIZE(mtts_idx));
5683 	if (ret) {
5684 		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
5685 			  ret);
5686 		return ret;
5687 	}
5688 
5689 	dma_handle_idx = hns_roce_get_mtr_ba(&idx_que->mtr);
5690 
5691 	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
5692 		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
5693 
5694 	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> DMA_IDX_SHIFT);
5695 	hr_reg_write(ctx, SRQC_IDX_BT_BA_H,
5696 		     upper_32_bits(dma_handle_idx >> DMA_IDX_SHIFT));
5697 
5698 	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
5699 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
5700 	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
5701 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
5702 
5703 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
5704 		     to_hr_hw_page_addr(mtts_idx[0]));
5705 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
5706 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
5707 
5708 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
5709 		     to_hr_hw_page_addr(mtts_idx[1]));
5710 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
5711 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
5712 
5713 	return 0;
5714 }
5715 
5716 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
5717 {
5718 	struct ib_device *ibdev = srq->ibsrq.device;
5719 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5720 	struct hns_roce_srq_context *ctx = mb_buf;
5721 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
5722 	dma_addr_t dma_handle_wqe;
5723 	int ret;
5724 
5725 	memset(ctx, 0, sizeof(*ctx));
5726 
5727 	/* Get the physical address of srq buf */
5728 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
5729 				ARRAY_SIZE(mtts_wqe));
5730 	if (ret) {
5731 		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
5732 			  ret);
5733 		return ret;
5734 	}
5735 
5736 	dma_handle_wqe = hns_roce_get_mtr_ba(&srq->buf_mtr);
5737 
5738 	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
5739 	hr_reg_write_bool(ctx, SRQC_SRQ_TYPE,
5740 			  srq->ibsrq.srq_type == IB_SRQT_XRC);
5741 	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
5742 	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
5743 	hr_reg_write(ctx, SRQC_XRCD, srq->xrcdn);
5744 	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
5745 	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
5746 	hr_reg_write(ctx, SRQC_RQWS,
5747 		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
5748 
5749 	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
5750 		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
5751 				      srq->wqe_cnt));
5752 
5753 	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> DMA_WQE_SHIFT);
5754 	hr_reg_write(ctx, SRQC_WQE_BT_BA_H,
5755 		     upper_32_bits(dma_handle_wqe >> DMA_WQE_SHIFT));
5756 
5757 	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
5758 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
5759 	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
5760 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
5761 
5762 	if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB) {
5763 		hr_reg_enable(ctx, SRQC_DB_RECORD_EN);
5764 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_L,
5765 			     lower_32_bits(srq->rdb.dma) >> 1);
5766 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_H,
5767 			     upper_32_bits(srq->rdb.dma));
5768 	}
5769 
5770 	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
5771 }
5772 
5773 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
5774 				  struct ib_srq_attr *srq_attr,
5775 				  enum ib_srq_attr_mask srq_attr_mask,
5776 				  struct ib_udata *udata)
5777 {
5778 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5779 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5780 	struct hns_roce_srq_context *srq_context;
5781 	struct hns_roce_srq_context *srqc_mask;
5782 	struct hns_roce_cmd_mailbox *mailbox;
5783 	int ret = 0;
5784 
5785 	/* Resizing SRQs is not supported yet */
5786 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
5787 		ret = -EOPNOTSUPP;
5788 		goto out;
5789 	}
5790 
5791 	if (srq_attr_mask & IB_SRQ_LIMIT) {
5792 		if (srq_attr->srq_limit > srq->wqe_cnt) {
5793 			ret = -EINVAL;
5794 			goto out;
5795 		}
5796 
5797 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5798 		if (IS_ERR(mailbox)) {
5799 			ret = PTR_ERR(mailbox);
5800 			goto out;
5801 		}
5802 
5803 		srq_context = mailbox->buf;
5804 		srqc_mask = (struct hns_roce_srq_context *)mailbox->buf + 1;
5805 
5806 		memset(srqc_mask, 0xff, sizeof(*srqc_mask));
5807 
5808 		hr_reg_write(srq_context, SRQC_LIMIT_WL, srq_attr->srq_limit);
5809 		hr_reg_clear(srqc_mask, SRQC_LIMIT_WL);
5810 
5811 		ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5812 					HNS_ROCE_CMD_MODIFY_SRQC, srq->srqn);
5813 		hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5814 		if (ret)
5815 			ibdev_err(&hr_dev->ib_dev,
5816 				  "failed to handle cmd of modifying SRQ, ret = %d.\n",
5817 				  ret);
5818 	}
5819 
5820 out:
5821 	if (ret)
5822 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_SRQ_MODIFY_ERR_CNT]);
5823 
5824 	return ret;
5825 }
5826 
5827 static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
5828 {
5829 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5830 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5831 	struct hns_roce_srq_context *srq_context;
5832 	struct hns_roce_cmd_mailbox *mailbox;
5833 	int ret;
5834 
5835 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5836 	if (IS_ERR(mailbox))
5837 		return PTR_ERR(mailbox);
5838 
5839 	srq_context = mailbox->buf;
5840 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5841 				HNS_ROCE_CMD_QUERY_SRQC, srq->srqn);
5842 	if (ret) {
5843 		ibdev_err(&hr_dev->ib_dev,
5844 			  "failed to process cmd of querying SRQ, ret = %d.\n",
5845 			  ret);
5846 		goto out;
5847 	}
5848 
5849 	attr->srq_limit = hr_reg_read(srq_context, SRQC_LIMIT_WL);
5850 	attr->max_wr = srq->wqe_cnt;
5851 	attr->max_sge = srq->max_gs - srq->rsv_sge;
5852 
5853 out:
5854 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5855 	return ret;
5856 }
5857 
5858 static int hns_roce_v2_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
5859 {
5860 	struct hns_roce_dev *hr_dev = to_hr_dev(cq->device);
5861 	struct hns_roce_v2_cq_context *cq_context;
5862 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
5863 	struct hns_roce_v2_cq_context *cqc_mask;
5864 	struct hns_roce_cmd_mailbox *mailbox;
5865 	int ret;
5866 
5867 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5868 	ret = PTR_ERR_OR_ZERO(mailbox);
5869 	if (ret)
5870 		goto err_out;
5871 
5872 	cq_context = mailbox->buf;
5873 	cqc_mask = (struct hns_roce_v2_cq_context *)mailbox->buf + 1;
5874 
5875 	memset(cqc_mask, 0xff, sizeof(*cqc_mask));
5876 
5877 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT, cq_count);
5878 	hr_reg_clear(cqc_mask, CQC_CQ_MAX_CNT);
5879 
5880 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5881 		if (cq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
5882 			dev_info(hr_dev->dev,
5883 				 "cq_period(%u) reached the upper limit, adjusted to 65.\n",
5884 				 cq_period);
5885 			cq_period = HNS_ROCE_MAX_CQ_PERIOD_HIP08;
5886 		}
5887 		cq_period *= HNS_ROCE_CLOCK_ADJUST;
5888 	}
5889 	hr_reg_write(cq_context, CQC_CQ_PERIOD, cq_period);
5890 	hr_reg_clear(cqc_mask, CQC_CQ_PERIOD);
5891 
5892 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5893 				HNS_ROCE_CMD_MODIFY_CQC, hr_cq->cqn);
5894 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5895 	if (ret)
5896 		ibdev_err(&hr_dev->ib_dev,
5897 			  "failed to process cmd when modifying CQ, ret = %d.\n",
5898 			  ret);
5899 
5900 err_out:
5901 	if (ret)
5902 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CQ_MODIFY_ERR_CNT]);
5903 
5904 	return ret;
5905 }
5906 
5907 static int hns_roce_v2_query_cqc(struct hns_roce_dev *hr_dev, u32 cqn,
5908 				 void *buffer)
5909 {
5910 	struct hns_roce_v2_cq_context *context;
5911 	struct hns_roce_cmd_mailbox *mailbox;
5912 	int ret;
5913 
5914 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5915 	if (IS_ERR(mailbox))
5916 		return PTR_ERR(mailbox);
5917 
5918 	context = mailbox->buf;
5919 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5920 				HNS_ROCE_CMD_QUERY_CQC, cqn);
5921 	if (ret) {
5922 		ibdev_err(&hr_dev->ib_dev,
5923 			  "failed to process cmd when querying CQ, ret = %d.\n",
5924 			  ret);
5925 		goto err_mailbox;
5926 	}
5927 
5928 	memcpy(buffer, context, sizeof(*context));
5929 
5930 err_mailbox:
5931 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5932 
5933 	return ret;
5934 }
5935 
5936 static int hns_roce_v2_query_mpt(struct hns_roce_dev *hr_dev, u32 key,
5937 				 void *buffer)
5938 {
5939 	struct hns_roce_v2_mpt_entry *context;
5940 	struct hns_roce_cmd_mailbox *mailbox;
5941 	int ret;
5942 
5943 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5944 	if (IS_ERR(mailbox))
5945 		return PTR_ERR(mailbox);
5946 
5947 	context = mailbox->buf;
5948 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
5949 				key_to_hw_index(key));
5950 	if (ret) {
5951 		ibdev_err(&hr_dev->ib_dev,
5952 			  "failed to process cmd when querying MPT, ret = %d.\n",
5953 			  ret);
5954 		goto err_mailbox;
5955 	}
5956 
5957 	memcpy(buffer, context, sizeof(*context));
5958 
5959 err_mailbox:
5960 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5961 
5962 	return ret;
5963 }
5964 
5965 static void hns_roce_irq_work_handle(struct work_struct *work)
5966 {
5967 	struct hns_roce_work *irq_work =
5968 				container_of(work, struct hns_roce_work, work);
5969 	struct ib_device *ibdev = &irq_work->hr_dev->ib_dev;
5970 
5971 	switch (irq_work->event_type) {
5972 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5973 		ibdev_info(ibdev, "path migrated succeeded.\n");
5974 		break;
5975 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5976 		ibdev_warn(ibdev, "path migration failed.\n");
5977 		break;
5978 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
5979 		break;
5980 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5981 		ibdev_dbg(ibdev, "send queue drained.\n");
5982 		break;
5983 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5984 		ibdev_err(ibdev, "local work queue 0x%x catast error, sub_event type is: %d\n",
5985 			  irq_work->queue_num, irq_work->sub_type);
5986 		break;
5987 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5988 		ibdev_err(ibdev, "invalid request local work queue 0x%x error.\n",
5989 			  irq_work->queue_num);
5990 		break;
5991 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5992 		ibdev_err(ibdev, "local access violation work queue 0x%x error, sub_event type is: %d\n",
5993 			  irq_work->queue_num, irq_work->sub_type);
5994 		break;
5995 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
5996 		ibdev_dbg(ibdev, "SRQ limit reach.\n");
5997 		break;
5998 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
5999 		ibdev_dbg(ibdev, "SRQ last wqe reach.\n");
6000 		break;
6001 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6002 		ibdev_err(ibdev, "SRQ catas error.\n");
6003 		break;
6004 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6005 		ibdev_err(ibdev, "CQ 0x%x access err.\n", irq_work->queue_num);
6006 		break;
6007 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6008 		ibdev_warn(ibdev, "CQ 0x%x overflow\n", irq_work->queue_num);
6009 		break;
6010 	case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
6011 		ibdev_warn(ibdev, "DB overflow.\n");
6012 		break;
6013 	case HNS_ROCE_EVENT_TYPE_FLR:
6014 		ibdev_warn(ibdev, "function level reset.\n");
6015 		break;
6016 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6017 		ibdev_err(ibdev, "xrc domain violation error.\n");
6018 		break;
6019 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6020 		ibdev_err(ibdev, "invalid xrceth error.\n");
6021 		break;
6022 	default:
6023 		break;
6024 	}
6025 
6026 	kfree(irq_work);
6027 }
6028 
6029 static void hns_roce_v2_init_irq_work(struct hns_roce_dev *hr_dev,
6030 				      struct hns_roce_eq *eq, u32 queue_num)
6031 {
6032 	struct hns_roce_work *irq_work;
6033 
6034 	irq_work = kzalloc(sizeof(struct hns_roce_work), GFP_ATOMIC);
6035 	if (!irq_work)
6036 		return;
6037 
6038 	INIT_WORK(&irq_work->work, hns_roce_irq_work_handle);
6039 	irq_work->hr_dev = hr_dev;
6040 	irq_work->event_type = eq->event_type;
6041 	irq_work->sub_type = eq->sub_type;
6042 	irq_work->queue_num = queue_num;
6043 	queue_work(hr_dev->irq_workq, &irq_work->work);
6044 }
6045 
6046 static void update_eq_db(struct hns_roce_eq *eq)
6047 {
6048 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6049 	struct hns_roce_v2_db eq_db = {};
6050 
6051 	if (eq->type_flag == HNS_ROCE_AEQ) {
6052 		hr_reg_write(&eq_db, EQ_DB_CMD,
6053 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6054 			     HNS_ROCE_EQ_DB_CMD_AEQ :
6055 			     HNS_ROCE_EQ_DB_CMD_AEQ_ARMED);
6056 	} else {
6057 		hr_reg_write(&eq_db, EQ_DB_TAG, eq->eqn);
6058 
6059 		hr_reg_write(&eq_db, EQ_DB_CMD,
6060 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6061 			     HNS_ROCE_EQ_DB_CMD_CEQ :
6062 			     HNS_ROCE_EQ_DB_CMD_CEQ_ARMED);
6063 	}
6064 
6065 	hr_reg_write(&eq_db, EQ_DB_CI, eq->cons_index);
6066 
6067 	hns_roce_write64(hr_dev, (__le32 *)&eq_db, eq->db_reg);
6068 }
6069 
6070 static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
6071 {
6072 	struct hns_roce_aeqe *aeqe;
6073 
6074 	aeqe = hns_roce_buf_offset(eq->mtr.kmem,
6075 				   (eq->cons_index & (eq->entries - 1)) *
6076 				   eq->eqe_size);
6077 
6078 	return (hr_reg_read(aeqe, AEQE_OWNER) ^
6079 		!!(eq->cons_index & eq->entries)) ? aeqe : NULL;
6080 }
6081 
6082 static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
6083 				       struct hns_roce_eq *eq)
6084 {
6085 	struct device *dev = hr_dev->dev;
6086 	struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
6087 	irqreturn_t aeqe_found = IRQ_NONE;
6088 	int event_type;
6089 	u32 queue_num;
6090 	int sub_type;
6091 
6092 	while (aeqe) {
6093 		/* Make sure we read AEQ entry after we have checked the
6094 		 * ownership bit
6095 		 */
6096 		dma_rmb();
6097 
6098 		event_type = hr_reg_read(aeqe, AEQE_EVENT_TYPE);
6099 		sub_type = hr_reg_read(aeqe, AEQE_SUB_TYPE);
6100 		queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
6101 
6102 		switch (event_type) {
6103 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
6104 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
6105 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
6106 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
6107 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
6108 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
6109 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
6110 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
6111 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6112 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6113 			hns_roce_qp_event(hr_dev, queue_num, event_type);
6114 			break;
6115 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
6116 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6117 			hns_roce_srq_event(hr_dev, queue_num, event_type);
6118 			break;
6119 		case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6120 		case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6121 			hns_roce_cq_event(hr_dev, queue_num, event_type);
6122 			break;
6123 		case HNS_ROCE_EVENT_TYPE_MB:
6124 			hns_roce_cmd_event(hr_dev,
6125 					le16_to_cpu(aeqe->event.cmd.token),
6126 					aeqe->event.cmd.status,
6127 					le64_to_cpu(aeqe->event.cmd.out_param));
6128 			break;
6129 		case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
6130 		case HNS_ROCE_EVENT_TYPE_FLR:
6131 			break;
6132 		default:
6133 			dev_err(dev, "unhandled event %d on EQ %d at idx %u.\n",
6134 				event_type, eq->eqn, eq->cons_index);
6135 			break;
6136 		}
6137 
6138 		eq->event_type = event_type;
6139 		eq->sub_type = sub_type;
6140 		++eq->cons_index;
6141 		aeqe_found = IRQ_HANDLED;
6142 
6143 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_AEQE_CNT]);
6144 
6145 		hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
6146 
6147 		aeqe = next_aeqe_sw_v2(eq);
6148 	}
6149 
6150 	update_eq_db(eq);
6151 
6152 	return IRQ_RETVAL(aeqe_found);
6153 }
6154 
6155 static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq)
6156 {
6157 	struct hns_roce_ceqe *ceqe;
6158 
6159 	ceqe = hns_roce_buf_offset(eq->mtr.kmem,
6160 				   (eq->cons_index & (eq->entries - 1)) *
6161 				   eq->eqe_size);
6162 
6163 	return (hr_reg_read(ceqe, CEQE_OWNER) ^
6164 		!!(eq->cons_index & eq->entries)) ? ceqe : NULL;
6165 }
6166 
6167 static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_eq *eq)
6168 {
6169 	queue_work(system_bh_wq, &eq->work);
6170 
6171 	return IRQ_HANDLED;
6172 }
6173 
6174 static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
6175 {
6176 	struct hns_roce_eq *eq = eq_ptr;
6177 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6178 	irqreturn_t int_work;
6179 
6180 	if (eq->type_flag == HNS_ROCE_CEQ)
6181 		/* Completion event interrupt */
6182 		int_work = hns_roce_v2_ceq_int(eq);
6183 	else
6184 		/* Asynchronous event interrupt */
6185 		int_work = hns_roce_v2_aeq_int(hr_dev, eq);
6186 
6187 	return IRQ_RETVAL(int_work);
6188 }
6189 
6190 static irqreturn_t abnormal_interrupt_basic(struct hns_roce_dev *hr_dev,
6191 					    u32 int_st)
6192 {
6193 	struct pci_dev *pdev = hr_dev->pci_dev;
6194 	struct hnae3_ae_dev *ae_dev = pci_get_drvdata(pdev);
6195 	const struct hnae3_ae_ops *ops = ae_dev->ops;
6196 	irqreturn_t int_work = IRQ_NONE;
6197 	u32 int_en;
6198 
6199 	int_en = roce_read(hr_dev, ROCEE_VF_ABN_INT_EN_REG);
6200 
6201 	if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S)) {
6202 		dev_err(hr_dev->dev, "AEQ overflow!\n");
6203 
6204 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG,
6205 			   1 << HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S);
6206 
6207 		/* Set reset level for reset_event() */
6208 		if (ops->set_default_reset_request)
6209 			ops->set_default_reset_request(ae_dev,
6210 						       HNAE3_FUNC_RESET);
6211 		if (ops->reset_event)
6212 			ops->reset_event(pdev, NULL);
6213 
6214 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6215 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6216 
6217 		int_work = IRQ_HANDLED;
6218 	} else {
6219 		dev_err(hr_dev->dev, "there is no basic abn irq found.\n");
6220 	}
6221 
6222 	return IRQ_RETVAL(int_work);
6223 }
6224 
6225 static int fmea_ram_ecc_query(struct hns_roce_dev *hr_dev,
6226 			       struct fmea_ram_ecc *ecc_info)
6227 {
6228 	struct hns_roce_cmq_desc desc;
6229 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6230 	int ret;
6231 
6232 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_QUERY_RAM_ECC, true);
6233 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6234 	if (ret)
6235 		return ret;
6236 
6237 	ecc_info->is_ecc_err = hr_reg_read(req, QUERY_RAM_ECC_1BIT_ERR);
6238 	ecc_info->res_type = hr_reg_read(req, QUERY_RAM_ECC_RES_TYPE);
6239 	ecc_info->index = hr_reg_read(req, QUERY_RAM_ECC_TAG);
6240 
6241 	return 0;
6242 }
6243 
6244 static int fmea_recover_gmv(struct hns_roce_dev *hr_dev, u32 idx)
6245 {
6246 	struct hns_roce_cmq_desc desc;
6247 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6248 	u32 addr_upper;
6249 	u32 addr_low;
6250 	int ret;
6251 
6252 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, true);
6253 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6254 
6255 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6256 	if (ret) {
6257 		dev_err(hr_dev->dev,
6258 			"failed to execute cmd to read gmv, ret = %d.\n", ret);
6259 		return ret;
6260 	}
6261 
6262 	addr_low =  hr_reg_read(req, CFG_GMV_BT_BA_L);
6263 	addr_upper = hr_reg_read(req, CFG_GMV_BT_BA_H);
6264 
6265 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
6266 	hr_reg_write(req, CFG_GMV_BT_BA_L, addr_low);
6267 	hr_reg_write(req, CFG_GMV_BT_BA_H, addr_upper);
6268 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6269 
6270 	return hns_roce_cmq_send(hr_dev, &desc, 1);
6271 }
6272 
6273 static u64 fmea_get_ram_res_addr(u32 res_type, __le64 *data)
6274 {
6275 	if (res_type == ECC_RESOURCE_QPC_TIMER ||
6276 	    res_type == ECC_RESOURCE_CQC_TIMER ||
6277 	    res_type == ECC_RESOURCE_SCCC)
6278 		return le64_to_cpu(*data);
6279 
6280 	return le64_to_cpu(*data) << PAGE_SHIFT;
6281 }
6282 
6283 static int fmea_recover_others(struct hns_roce_dev *hr_dev, u32 res_type,
6284 			       u32 index)
6285 {
6286 	u8 write_bt0_op = fmea_ram_res[res_type].write_bt0_op;
6287 	u8 read_bt0_op = fmea_ram_res[res_type].read_bt0_op;
6288 	struct hns_roce_cmd_mailbox *mailbox;
6289 	u64 addr;
6290 	int ret;
6291 
6292 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6293 	if (IS_ERR(mailbox))
6294 		return PTR_ERR(mailbox);
6295 
6296 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, read_bt0_op, index);
6297 	if (ret) {
6298 		dev_err(hr_dev->dev,
6299 			"failed to execute cmd to read fmea ram, ret = %d.\n",
6300 			ret);
6301 		goto out;
6302 	}
6303 
6304 	addr = fmea_get_ram_res_addr(res_type, mailbox->buf);
6305 
6306 	ret = hns_roce_cmd_mbox(hr_dev, addr, 0, write_bt0_op, index);
6307 	if (ret)
6308 		dev_err(hr_dev->dev,
6309 			"failed to execute cmd to write fmea ram, ret = %d.\n",
6310 			ret);
6311 
6312 out:
6313 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6314 	return ret;
6315 }
6316 
6317 static void fmea_ram_ecc_recover(struct hns_roce_dev *hr_dev,
6318 				 struct fmea_ram_ecc *ecc_info)
6319 {
6320 	u32 res_type = ecc_info->res_type;
6321 	u32 index = ecc_info->index;
6322 	int ret;
6323 
6324 	BUILD_BUG_ON(ARRAY_SIZE(fmea_ram_res) != ECC_RESOURCE_COUNT);
6325 
6326 	if (res_type >= ECC_RESOURCE_COUNT) {
6327 		dev_err(hr_dev->dev, "unsupported fmea ram ecc type %u.\n",
6328 			res_type);
6329 		return;
6330 	}
6331 
6332 	if (res_type == ECC_RESOURCE_GMV)
6333 		ret = fmea_recover_gmv(hr_dev, index);
6334 	else
6335 		ret = fmea_recover_others(hr_dev, res_type, index);
6336 	if (ret)
6337 		dev_err(hr_dev->dev,
6338 			"failed to recover %s, index = %u, ret = %d.\n",
6339 			fmea_ram_res[res_type].name, index, ret);
6340 }
6341 
6342 static void fmea_ram_ecc_work(struct work_struct *ecc_work)
6343 {
6344 	struct hns_roce_dev *hr_dev =
6345 		container_of(ecc_work, struct hns_roce_dev, ecc_work);
6346 	struct fmea_ram_ecc ecc_info = {};
6347 
6348 	if (fmea_ram_ecc_query(hr_dev, &ecc_info)) {
6349 		dev_err(hr_dev->dev, "failed to query fmea ram ecc.\n");
6350 		return;
6351 	}
6352 
6353 	if (!ecc_info.is_ecc_err) {
6354 		dev_err(hr_dev->dev, "there is no fmea ram ecc err found.\n");
6355 		return;
6356 	}
6357 
6358 	fmea_ram_ecc_recover(hr_dev, &ecc_info);
6359 }
6360 
6361 static irqreturn_t hns_roce_v2_msix_interrupt_abn(int irq, void *dev_id)
6362 {
6363 	struct hns_roce_dev *hr_dev = dev_id;
6364 	irqreturn_t int_work = IRQ_NONE;
6365 	u32 int_st;
6366 
6367 	int_st = roce_read(hr_dev, ROCEE_VF_ABN_INT_ST_REG);
6368 
6369 	if (int_st) {
6370 		int_work = abnormal_interrupt_basic(hr_dev, int_st);
6371 	} else if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
6372 		queue_work(hr_dev->irq_workq, &hr_dev->ecc_work);
6373 		int_work = IRQ_HANDLED;
6374 	} else {
6375 		dev_err(hr_dev->dev, "there is no abnormal irq found.\n");
6376 	}
6377 
6378 	return IRQ_RETVAL(int_work);
6379 }
6380 
6381 static void hns_roce_v2_int_mask_enable(struct hns_roce_dev *hr_dev,
6382 					int eq_num, u32 enable_flag)
6383 {
6384 	int i;
6385 
6386 	for (i = 0; i < eq_num; i++)
6387 		roce_write(hr_dev, ROCEE_VF_EVENT_INT_EN_REG +
6388 			   i * EQ_REG_OFFSET, enable_flag);
6389 
6390 	roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, enable_flag);
6391 	roce_write(hr_dev, ROCEE_VF_ABN_INT_CFG_REG, enable_flag);
6392 }
6393 
6394 static void free_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6395 {
6396 	hns_roce_mtr_destroy(hr_dev, &eq->mtr);
6397 }
6398 
6399 static void hns_roce_v2_destroy_eqc(struct hns_roce_dev *hr_dev,
6400 				    struct hns_roce_eq *eq)
6401 {
6402 	struct device *dev = hr_dev->dev;
6403 	int eqn = eq->eqn;
6404 	int ret;
6405 	u8 cmd;
6406 
6407 	if (eqn < hr_dev->caps.num_comp_vectors)
6408 		cmd = HNS_ROCE_CMD_DESTROY_CEQC;
6409 	else
6410 		cmd = HNS_ROCE_CMD_DESTROY_AEQC;
6411 
6412 	ret = hns_roce_destroy_hw_ctx(hr_dev, cmd, eqn & HNS_ROCE_V2_EQN_M);
6413 	if (ret)
6414 		dev_err(dev, "[mailbox cmd] destroy eqc(%d) failed.\n", eqn);
6415 
6416 	free_eq_buf(hr_dev, eq);
6417 }
6418 
6419 static void init_eq_config(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6420 {
6421 	eq->db_reg = hr_dev->reg_base + ROCEE_VF_EQ_DB_CFG0_REG;
6422 	eq->cons_index = 0;
6423 	eq->over_ignore = HNS_ROCE_V2_EQ_OVER_IGNORE_0;
6424 	eq->coalesce = HNS_ROCE_V2_EQ_COALESCE_0;
6425 	eq->arm_st = HNS_ROCE_V2_EQ_ALWAYS_ARMED;
6426 	eq->shift = ilog2((unsigned int)eq->entries);
6427 }
6428 
6429 static int config_eqc(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq,
6430 		      void *mb_buf)
6431 {
6432 	u64 eqe_ba[MTT_MIN_COUNT] = { 0 };
6433 	struct hns_roce_eq_context *eqc;
6434 	u64 bt_ba = 0;
6435 	int ret;
6436 
6437 	eqc = mb_buf;
6438 	memset(eqc, 0, sizeof(struct hns_roce_eq_context));
6439 
6440 	init_eq_config(hr_dev, eq);
6441 
6442 	/* if not multi-hop, eqe buffer only use one trunk */
6443 	ret = hns_roce_mtr_find(hr_dev, &eq->mtr, 0, eqe_ba,
6444 				ARRAY_SIZE(eqe_ba));
6445 	if (ret) {
6446 		dev_err(hr_dev->dev, "failed to find EQE mtr, ret = %d\n", ret);
6447 		return ret;
6448 	}
6449 
6450 	bt_ba = hns_roce_get_mtr_ba(&eq->mtr);
6451 
6452 	hr_reg_write(eqc, EQC_EQ_ST, HNS_ROCE_V2_EQ_STATE_VALID);
6453 	hr_reg_write(eqc, EQC_EQE_HOP_NUM, eq->hop_num);
6454 	hr_reg_write(eqc, EQC_OVER_IGNORE, eq->over_ignore);
6455 	hr_reg_write(eqc, EQC_COALESCE, eq->coalesce);
6456 	hr_reg_write(eqc, EQC_ARM_ST, eq->arm_st);
6457 	hr_reg_write(eqc, EQC_EQN, eq->eqn);
6458 	hr_reg_write(eqc, EQC_EQE_CNT, HNS_ROCE_EQ_INIT_EQE_CNT);
6459 	hr_reg_write(eqc, EQC_EQE_BA_PG_SZ,
6460 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.ba_pg_shift));
6461 	hr_reg_write(eqc, EQC_EQE_BUF_PG_SZ,
6462 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.buf_pg_shift));
6463 	hr_reg_write(eqc, EQC_EQ_PROD_INDX, HNS_ROCE_EQ_INIT_PROD_IDX);
6464 	hr_reg_write(eqc, EQC_EQ_MAX_CNT, eq->eq_max_cnt);
6465 
6466 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6467 		if (eq->eq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
6468 			dev_info(hr_dev->dev, "eq_period(%u) reached the upper limit, adjusted to 65.\n",
6469 				 eq->eq_period);
6470 			eq->eq_period = HNS_ROCE_MAX_EQ_PERIOD;
6471 		}
6472 		eq->eq_period *= HNS_ROCE_CLOCK_ADJUST;
6473 	}
6474 
6475 	hr_reg_write(eqc, EQC_EQ_PERIOD, eq->eq_period);
6476 	hr_reg_write(eqc, EQC_EQE_REPORT_TIMER, HNS_ROCE_EQ_INIT_REPORT_TIMER);
6477 	hr_reg_write(eqc, EQC_EQE_BA_L, bt_ba >> 3);
6478 	hr_reg_write(eqc, EQC_EQE_BA_H, bt_ba >> 35);
6479 	hr_reg_write(eqc, EQC_SHIFT, eq->shift);
6480 	hr_reg_write(eqc, EQC_MSI_INDX, HNS_ROCE_EQ_INIT_MSI_IDX);
6481 	hr_reg_write(eqc, EQC_CUR_EQE_BA_L, eqe_ba[0] >> 12);
6482 	hr_reg_write(eqc, EQC_CUR_EQE_BA_M, eqe_ba[0] >> 28);
6483 	hr_reg_write(eqc, EQC_CUR_EQE_BA_H, eqe_ba[0] >> 60);
6484 	hr_reg_write(eqc, EQC_EQ_CONS_INDX, HNS_ROCE_EQ_INIT_CONS_IDX);
6485 	hr_reg_write(eqc, EQC_NEX_EQE_BA_L, eqe_ba[1] >> 12);
6486 	hr_reg_write(eqc, EQC_NEX_EQE_BA_H, eqe_ba[1] >> 44);
6487 	hr_reg_write(eqc, EQC_EQE_SIZE, eq->eqe_size == HNS_ROCE_V3_EQE_SIZE);
6488 
6489 	return 0;
6490 }
6491 
6492 static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6493 {
6494 	struct hns_roce_buf_attr buf_attr = {};
6495 	int err;
6496 
6497 	if (hr_dev->caps.eqe_hop_num == HNS_ROCE_HOP_NUM_0)
6498 		eq->hop_num = 0;
6499 	else
6500 		eq->hop_num = hr_dev->caps.eqe_hop_num;
6501 
6502 	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
6503 	buf_attr.region[0].size = eq->entries * eq->eqe_size;
6504 	buf_attr.region[0].hopnum = eq->hop_num;
6505 	buf_attr.region_count = 1;
6506 
6507 	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
6508 				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
6509 				  0);
6510 	if (err)
6511 		dev_err(hr_dev->dev, "failed to alloc EQE mtr, err %d\n", err);
6512 
6513 	return err;
6514 }
6515 
6516 static int hns_roce_v2_create_eq(struct hns_roce_dev *hr_dev,
6517 				 struct hns_roce_eq *eq, u8 eq_cmd)
6518 {
6519 	struct hns_roce_cmd_mailbox *mailbox;
6520 	int ret;
6521 
6522 	/* Allocate mailbox memory */
6523 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6524 	if (IS_ERR(mailbox))
6525 		return PTR_ERR(mailbox);
6526 
6527 	ret = alloc_eq_buf(hr_dev, eq);
6528 	if (ret)
6529 		goto free_cmd_mbox;
6530 
6531 	ret = config_eqc(hr_dev, eq, mailbox->buf);
6532 	if (ret)
6533 		goto err_cmd_mbox;
6534 
6535 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, eq_cmd, eq->eqn);
6536 	if (ret) {
6537 		dev_err(hr_dev->dev, "[mailbox cmd] create eqc failed.\n");
6538 		goto err_cmd_mbox;
6539 	}
6540 
6541 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6542 
6543 	return 0;
6544 
6545 err_cmd_mbox:
6546 	free_eq_buf(hr_dev, eq);
6547 
6548 free_cmd_mbox:
6549 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6550 
6551 	return ret;
6552 }
6553 
6554 static void hns_roce_ceq_work(struct work_struct *work)
6555 {
6556 	struct hns_roce_eq *eq = from_work(eq, work, work);
6557 	struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq);
6558 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6559 	int ceqe_num = 0;
6560 	u32 cqn;
6561 
6562 	while (ceqe && ceqe_num < hr_dev->caps.ceqe_depth) {
6563 		/* Make sure we read CEQ entry after we have checked the
6564 		 * ownership bit
6565 		 */
6566 		dma_rmb();
6567 
6568 		cqn = hr_reg_read(ceqe, CEQE_CQN);
6569 
6570 		hns_roce_cq_completion(hr_dev, cqn);
6571 
6572 		++eq->cons_index;
6573 		++ceqe_num;
6574 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CEQE_CNT]);
6575 
6576 		ceqe = next_ceqe_sw_v2(eq);
6577 	}
6578 
6579 	update_eq_db(eq);
6580 }
6581 
6582 static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
6583 				  int comp_num, int aeq_num, int other_num)
6584 {
6585 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6586 	int i, j;
6587 	int ret;
6588 
6589 	for (i = 0; i < irq_num; i++) {
6590 		hr_dev->irq_names[i] = kzalloc(HNS_ROCE_INT_NAME_LEN,
6591 					       GFP_KERNEL);
6592 		if (!hr_dev->irq_names[i]) {
6593 			ret = -ENOMEM;
6594 			goto err_kzalloc_failed;
6595 		}
6596 	}
6597 
6598 	/* irq contains: abnormal + AEQ + CEQ */
6599 	for (j = 0; j < other_num; j++)
6600 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6601 			 "hns-%s-abn-%d", pci_name(hr_dev->pci_dev), j);
6602 
6603 	for (j = other_num; j < (other_num + aeq_num); j++)
6604 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6605 			 "hns-%s-aeq-%d", pci_name(hr_dev->pci_dev), j - other_num);
6606 
6607 	for (j = (other_num + aeq_num); j < irq_num; j++)
6608 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6609 			 "hns-%s-ceq-%d", pci_name(hr_dev->pci_dev),
6610 			 j - other_num - aeq_num);
6611 
6612 	for (j = 0; j < irq_num; j++) {
6613 		if (j < other_num) {
6614 			ret = request_irq(hr_dev->irq[j],
6615 					  hns_roce_v2_msix_interrupt_abn,
6616 					  0, hr_dev->irq_names[j], hr_dev);
6617 		} else if (j < (other_num + comp_num)) {
6618 			INIT_WORK(&eq_table->eq[j - other_num].work,
6619 				  hns_roce_ceq_work);
6620 			ret = request_irq(eq_table->eq[j - other_num].irq,
6621 					  hns_roce_v2_msix_interrupt_eq,
6622 					  0, hr_dev->irq_names[j + aeq_num],
6623 					  &eq_table->eq[j - other_num]);
6624 		} else {
6625 			ret = request_irq(eq_table->eq[j - other_num].irq,
6626 					  hns_roce_v2_msix_interrupt_eq,
6627 					  0, hr_dev->irq_names[j - comp_num],
6628 					  &eq_table->eq[j - other_num]);
6629 		}
6630 
6631 		if (ret) {
6632 			dev_err(hr_dev->dev, "request irq error!\n");
6633 			goto err_request_failed;
6634 		}
6635 	}
6636 
6637 	return 0;
6638 
6639 err_request_failed:
6640 	for (j -= 1; j >= 0; j--) {
6641 		if (j < other_num) {
6642 			free_irq(hr_dev->irq[j], hr_dev);
6643 			continue;
6644 		}
6645 		free_irq(eq_table->eq[j - other_num].irq,
6646 			 &eq_table->eq[j - other_num]);
6647 		if (j < other_num + comp_num)
6648 			cancel_work_sync(&eq_table->eq[j - other_num].work);
6649 	}
6650 
6651 err_kzalloc_failed:
6652 	for (i -= 1; i >= 0; i--)
6653 		kfree(hr_dev->irq_names[i]);
6654 
6655 	return ret;
6656 }
6657 
6658 static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev)
6659 {
6660 	int irq_num;
6661 	int eq_num;
6662 	int i;
6663 
6664 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6665 	irq_num = eq_num + hr_dev->caps.num_other_vectors;
6666 
6667 	for (i = 0; i < hr_dev->caps.num_other_vectors; i++)
6668 		free_irq(hr_dev->irq[i], hr_dev);
6669 
6670 	for (i = 0; i < eq_num; i++) {
6671 		free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]);
6672 		if (i < hr_dev->caps.num_comp_vectors)
6673 			cancel_work_sync(&hr_dev->eq_table.eq[i].work);
6674 	}
6675 
6676 	for (i = 0; i < irq_num; i++)
6677 		kfree(hr_dev->irq_names[i]);
6678 }
6679 
6680 static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
6681 {
6682 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6683 	struct device *dev = hr_dev->dev;
6684 	struct hns_roce_eq *eq;
6685 	int other_num;
6686 	int comp_num;
6687 	int aeq_num;
6688 	int irq_num;
6689 	int eq_num;
6690 	u8 eq_cmd;
6691 	int ret;
6692 	int i;
6693 
6694 	other_num = hr_dev->caps.num_other_vectors;
6695 	comp_num = hr_dev->caps.num_comp_vectors;
6696 	aeq_num = hr_dev->caps.num_aeq_vectors;
6697 
6698 	eq_num = comp_num + aeq_num;
6699 	irq_num = eq_num + other_num;
6700 
6701 	eq_table->eq = kcalloc(eq_num, sizeof(*eq_table->eq), GFP_KERNEL);
6702 	if (!eq_table->eq)
6703 		return -ENOMEM;
6704 
6705 	/* create eq */
6706 	for (i = 0; i < eq_num; i++) {
6707 		eq = &eq_table->eq[i];
6708 		eq->hr_dev = hr_dev;
6709 		eq->eqn = i;
6710 		if (i < comp_num) {
6711 			/* CEQ */
6712 			eq_cmd = HNS_ROCE_CMD_CREATE_CEQC;
6713 			eq->type_flag = HNS_ROCE_CEQ;
6714 			eq->entries = hr_dev->caps.ceqe_depth;
6715 			eq->eqe_size = hr_dev->caps.ceqe_size;
6716 			eq->irq = hr_dev->irq[i + other_num + aeq_num];
6717 			eq->eq_max_cnt = HNS_ROCE_CEQ_DEFAULT_BURST_NUM;
6718 			eq->eq_period = HNS_ROCE_CEQ_DEFAULT_INTERVAL;
6719 		} else {
6720 			/* AEQ */
6721 			eq_cmd = HNS_ROCE_CMD_CREATE_AEQC;
6722 			eq->type_flag = HNS_ROCE_AEQ;
6723 			eq->entries = hr_dev->caps.aeqe_depth;
6724 			eq->eqe_size = hr_dev->caps.aeqe_size;
6725 			eq->irq = hr_dev->irq[i - comp_num + other_num];
6726 			eq->eq_max_cnt = HNS_ROCE_AEQ_DEFAULT_BURST_NUM;
6727 			eq->eq_period = HNS_ROCE_AEQ_DEFAULT_INTERVAL;
6728 		}
6729 
6730 		ret = hns_roce_v2_create_eq(hr_dev, eq, eq_cmd);
6731 		if (ret) {
6732 			dev_err(dev, "failed to create eq.\n");
6733 			goto err_create_eq_fail;
6734 		}
6735 	}
6736 
6737 	INIT_WORK(&hr_dev->ecc_work, fmea_ram_ecc_work);
6738 
6739 	hr_dev->irq_workq = alloc_ordered_workqueue("hns_roce_irq_workq", 0);
6740 	if (!hr_dev->irq_workq) {
6741 		dev_err(dev, "failed to create irq workqueue.\n");
6742 		ret = -ENOMEM;
6743 		goto err_create_eq_fail;
6744 	}
6745 
6746 	ret = __hns_roce_request_irq(hr_dev, irq_num, comp_num, aeq_num,
6747 				     other_num);
6748 	if (ret) {
6749 		dev_err(dev, "failed to request irq.\n");
6750 		goto err_request_irq_fail;
6751 	}
6752 
6753 	/* enable irq */
6754 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_ENABLE);
6755 
6756 	return 0;
6757 
6758 err_request_irq_fail:
6759 	destroy_workqueue(hr_dev->irq_workq);
6760 
6761 err_create_eq_fail:
6762 	for (i -= 1; i >= 0; i--)
6763 		hns_roce_v2_destroy_eqc(hr_dev, &eq_table->eq[i]);
6764 	kfree(eq_table->eq);
6765 
6766 	return ret;
6767 }
6768 
6769 static void hns_roce_v2_cleanup_eq_table(struct hns_roce_dev *hr_dev)
6770 {
6771 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6772 	int eq_num;
6773 	int i;
6774 
6775 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6776 
6777 	/* Disable irq */
6778 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_DISABLE);
6779 
6780 	__hns_roce_free_irq(hr_dev);
6781 	destroy_workqueue(hr_dev->irq_workq);
6782 
6783 	for (i = 0; i < eq_num; i++)
6784 		hns_roce_v2_destroy_eqc(hr_dev, &eq_table->eq[i]);
6785 
6786 	kfree(eq_table->eq);
6787 }
6788 
6789 static const struct ib_device_ops hns_roce_v2_dev_ops = {
6790 	.destroy_qp = hns_roce_v2_destroy_qp,
6791 	.modify_cq = hns_roce_v2_modify_cq,
6792 	.poll_cq = hns_roce_v2_poll_cq,
6793 	.post_recv = hns_roce_v2_post_recv,
6794 	.post_send = hns_roce_v2_post_send,
6795 	.query_qp = hns_roce_v2_query_qp,
6796 	.req_notify_cq = hns_roce_v2_req_notify_cq,
6797 };
6798 
6799 static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {
6800 	.modify_srq = hns_roce_v2_modify_srq,
6801 	.post_srq_recv = hns_roce_v2_post_srq_recv,
6802 	.query_srq = hns_roce_v2_query_srq,
6803 };
6804 
6805 static const struct hns_roce_hw hns_roce_hw_v2 = {
6806 	.cmq_init = hns_roce_v2_cmq_init,
6807 	.cmq_exit = hns_roce_v2_cmq_exit,
6808 	.hw_profile = hns_roce_v2_profile,
6809 	.hw_init = hns_roce_v2_init,
6810 	.hw_exit = hns_roce_v2_exit,
6811 	.post_mbox = v2_post_mbox,
6812 	.poll_mbox_done = v2_poll_mbox_done,
6813 	.chk_mbox_avail = v2_chk_mbox_is_avail,
6814 	.set_gid = hns_roce_v2_set_gid,
6815 	.set_mac = hns_roce_v2_set_mac,
6816 	.write_mtpt = hns_roce_v2_write_mtpt,
6817 	.rereg_write_mtpt = hns_roce_v2_rereg_write_mtpt,
6818 	.frmr_write_mtpt = hns_roce_v2_frmr_write_mtpt,
6819 	.mw_write_mtpt = hns_roce_v2_mw_write_mtpt,
6820 	.write_cqc = hns_roce_v2_write_cqc,
6821 	.set_hem = hns_roce_v2_set_hem,
6822 	.clear_hem = hns_roce_v2_clear_hem,
6823 	.modify_qp = hns_roce_v2_modify_qp,
6824 	.dereg_mr = hns_roce_v2_dereg_mr,
6825 	.qp_flow_control_init = hns_roce_v2_qp_flow_control_init,
6826 	.init_eq = hns_roce_v2_init_eq_table,
6827 	.cleanup_eq = hns_roce_v2_cleanup_eq_table,
6828 	.write_srqc = hns_roce_v2_write_srqc,
6829 	.query_cqc = hns_roce_v2_query_cqc,
6830 	.query_qpc = hns_roce_v2_query_qpc,
6831 	.query_mpt = hns_roce_v2_query_mpt,
6832 	.query_srqc = hns_roce_v2_query_srqc,
6833 	.query_sccc = hns_roce_v2_query_sccc,
6834 	.query_hw_counter = hns_roce_hw_v2_query_counter,
6835 	.get_dscp = hns_roce_hw_v2_get_dscp,
6836 	.hns_roce_dev_ops = &hns_roce_v2_dev_ops,
6837 	.hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops,
6838 };
6839 
6840 static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = {
6841 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA), 0},
6842 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA_MACSEC), 0},
6843 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA), 0},
6844 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA_MACSEC), 0},
6845 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_100G_RDMA_MACSEC), 0},
6846 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_200G_RDMA), 0},
6847 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_RDMA_DCB_PFC_VF),
6848 	 HNAE3_DEV_SUPPORT_ROCE_DCB_BITS},
6849 	/* required last entry */
6850 	{0, }
6851 };
6852 
6853 MODULE_DEVICE_TABLE(pci, hns_roce_hw_v2_pci_tbl);
6854 
6855 static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
6856 				  struct hnae3_handle *handle)
6857 {
6858 	struct hns_roce_v2_priv *priv = hr_dev->priv;
6859 	const struct pci_device_id *id;
6860 	int i;
6861 
6862 	hr_dev->pci_dev = handle->pdev;
6863 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, hr_dev->pci_dev);
6864 	hr_dev->is_vf = id->driver_data;
6865 	hr_dev->dev = &handle->pdev->dev;
6866 	hr_dev->hw = &hns_roce_hw_v2;
6867 	hr_dev->sdb_offset = ROCEE_DB_SQ_L_0_REG;
6868 	hr_dev->odb_offset = hr_dev->sdb_offset;
6869 
6870 	/* Get info from NIC driver. */
6871 	hr_dev->reg_base = handle->rinfo.roce_io_base;
6872 	hr_dev->mem_base = handle->rinfo.roce_mem_base;
6873 	hr_dev->caps.num_ports = 1;
6874 	hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
6875 	hr_dev->iboe.phy_port[0] = 0;
6876 
6877 	addrconf_addr_eui48((u8 *)&hr_dev->ib_dev.node_guid,
6878 			    hr_dev->iboe.netdevs[0]->dev_addr);
6879 
6880 	for (i = 0; i < handle->rinfo.num_vectors; i++)
6881 		hr_dev->irq[i] = pci_irq_vector(handle->pdev,
6882 						i + handle->rinfo.base_vector);
6883 
6884 	/* cmd issue mode: 0 is poll, 1 is event */
6885 	hr_dev->cmd_mod = 1;
6886 	hr_dev->loop_idc = 0;
6887 
6888 	hr_dev->reset_cnt = handle->ae_algo->ops->ae_dev_reset_cnt(handle);
6889 	priv->handle = handle;
6890 }
6891 
6892 static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6893 {
6894 	struct hns_roce_dev *hr_dev;
6895 	int ret;
6896 
6897 	hr_dev = ib_alloc_device(hns_roce_dev, ib_dev);
6898 	if (!hr_dev)
6899 		return -ENOMEM;
6900 
6901 	hr_dev->priv = kzalloc(sizeof(struct hns_roce_v2_priv), GFP_KERNEL);
6902 	if (!hr_dev->priv) {
6903 		ret = -ENOMEM;
6904 		goto error_failed_kzalloc;
6905 	}
6906 
6907 	hns_roce_hw_v2_get_cfg(hr_dev, handle);
6908 
6909 	ret = hns_roce_init(hr_dev);
6910 	if (ret) {
6911 		dev_err(hr_dev->dev, "RoCE Engine init failed!\n");
6912 		goto error_failed_roce_init;
6913 	}
6914 
6915 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6916 		ret = free_mr_init(hr_dev);
6917 		if (ret) {
6918 			dev_err(hr_dev->dev, "failed to init free mr!\n");
6919 			goto error_failed_free_mr_init;
6920 		}
6921 	}
6922 
6923 	handle->priv = hr_dev;
6924 
6925 	return 0;
6926 
6927 error_failed_free_mr_init:
6928 	hns_roce_exit(hr_dev);
6929 
6930 error_failed_roce_init:
6931 	kfree(hr_dev->priv);
6932 
6933 error_failed_kzalloc:
6934 	ib_dealloc_device(&hr_dev->ib_dev);
6935 
6936 	return ret;
6937 }
6938 
6939 static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6940 					   bool reset)
6941 {
6942 	struct hns_roce_dev *hr_dev = handle->priv;
6943 
6944 	if (!hr_dev)
6945 		return;
6946 
6947 	handle->priv = NULL;
6948 
6949 	hr_dev->state = HNS_ROCE_DEVICE_STATE_UNINIT;
6950 	hns_roce_handle_device_err(hr_dev);
6951 
6952 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
6953 		free_mr_exit(hr_dev);
6954 
6955 	hns_roce_exit(hr_dev);
6956 	kfree(hr_dev->priv);
6957 	ib_dealloc_device(&hr_dev->ib_dev);
6958 }
6959 
6960 static int hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6961 {
6962 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
6963 	const struct pci_device_id *id;
6964 	struct device *dev = &handle->pdev->dev;
6965 	int ret;
6966 
6967 	handle->rinfo.instance_state = HNS_ROCE_STATE_INIT;
6968 
6969 	if (ops->ae_dev_resetting(handle) || ops->get_hw_reset_stat(handle)) {
6970 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6971 		goto reset_chk_err;
6972 	}
6973 
6974 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, handle->pdev);
6975 	if (!id)
6976 		return 0;
6977 
6978 	if (id->driver_data && handle->pdev->revision == PCI_REVISION_ID_HIP08)
6979 		return 0;
6980 
6981 	ret = __hns_roce_hw_v2_init_instance(handle);
6982 	if (ret) {
6983 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6984 		dev_err(dev, "RoCE instance init failed! ret = %d\n", ret);
6985 		if (ops->ae_dev_resetting(handle) ||
6986 		    ops->get_hw_reset_stat(handle))
6987 			goto reset_chk_err;
6988 		else
6989 			return ret;
6990 	}
6991 
6992 	handle->rinfo.instance_state = HNS_ROCE_STATE_INITED;
6993 
6994 	return 0;
6995 
6996 reset_chk_err:
6997 	dev_err(dev, "Device is busy in resetting state.\n"
6998 		     "please retry later.\n");
6999 
7000 	return -EBUSY;
7001 }
7002 
7003 static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
7004 					   bool reset)
7005 {
7006 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED)
7007 		return;
7008 
7009 	handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;
7010 
7011 	__hns_roce_hw_v2_uninit_instance(handle, reset);
7012 
7013 	handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
7014 }
7015 static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
7016 {
7017 	struct hns_roce_dev *hr_dev;
7018 
7019 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) {
7020 		set_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
7021 		return 0;
7022 	}
7023 
7024 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_DOWN;
7025 	clear_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
7026 
7027 	hr_dev = handle->priv;
7028 	if (!hr_dev)
7029 		return 0;
7030 
7031 	hr_dev->active = false;
7032 	hr_dev->dis_db = true;
7033 	hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
7034 
7035 	return 0;
7036 }
7037 
7038 static int hns_roce_hw_v2_reset_notify_init(struct hnae3_handle *handle)
7039 {
7040 	struct device *dev = &handle->pdev->dev;
7041 	int ret;
7042 
7043 	if (test_and_clear_bit(HNS_ROCE_RST_DIRECT_RETURN,
7044 			       &handle->rinfo.state)) {
7045 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7046 		return 0;
7047 	}
7048 
7049 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INIT;
7050 
7051 	dev_info(&handle->pdev->dev, "In reset process RoCE client reinit.\n");
7052 	ret = __hns_roce_hw_v2_init_instance(handle);
7053 	if (ret) {
7054 		/* when reset notify type is HNAE3_INIT_CLIENT In reset notify
7055 		 * callback function, RoCE Engine reinitialize. If RoCE reinit
7056 		 * failed, we should inform NIC driver.
7057 		 */
7058 		handle->priv = NULL;
7059 		dev_err(dev, "In reset process RoCE reinit failed %d.\n", ret);
7060 	} else {
7061 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7062 		dev_info(dev, "reset done, RoCE client reinit finished.\n");
7063 	}
7064 
7065 	return ret;
7066 }
7067 
7068 static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle)
7069 {
7070 	if (test_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state))
7071 		return 0;
7072 
7073 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT;
7074 	dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n");
7075 	msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY);
7076 	__hns_roce_hw_v2_uninit_instance(handle, false);
7077 
7078 	return 0;
7079 }
7080 
7081 static int hns_roce_hw_v2_reset_notify(struct hnae3_handle *handle,
7082 				       enum hnae3_reset_notify_type type)
7083 {
7084 	int ret = 0;
7085 
7086 	switch (type) {
7087 	case HNAE3_DOWN_CLIENT:
7088 		ret = hns_roce_hw_v2_reset_notify_down(handle);
7089 		break;
7090 	case HNAE3_INIT_CLIENT:
7091 		ret = hns_roce_hw_v2_reset_notify_init(handle);
7092 		break;
7093 	case HNAE3_UNINIT_CLIENT:
7094 		ret = hns_roce_hw_v2_reset_notify_uninit(handle);
7095 		break;
7096 	default:
7097 		break;
7098 	}
7099 
7100 	return ret;
7101 }
7102 
7103 static const struct hnae3_client_ops hns_roce_hw_v2_ops = {
7104 	.init_instance = hns_roce_hw_v2_init_instance,
7105 	.uninit_instance = hns_roce_hw_v2_uninit_instance,
7106 	.reset_notify = hns_roce_hw_v2_reset_notify,
7107 };
7108 
7109 static struct hnae3_client hns_roce_hw_v2_client = {
7110 	.name = "hns_roce_hw_v2",
7111 	.type = HNAE3_CLIENT_ROCE,
7112 	.ops = &hns_roce_hw_v2_ops,
7113 };
7114 
7115 static int __init hns_roce_hw_v2_init(void)
7116 {
7117 	hns_roce_init_debugfs();
7118 	return hnae3_register_client(&hns_roce_hw_v2_client);
7119 }
7120 
7121 static void __exit hns_roce_hw_v2_exit(void)
7122 {
7123 	hnae3_unregister_client(&hns_roce_hw_v2_client);
7124 	hns_roce_cleanup_debugfs();
7125 }
7126 
7127 module_init(hns_roce_hw_v2_init);
7128 module_exit(hns_roce_hw_v2_exit);
7129 
7130 MODULE_LICENSE("Dual BSD/GPL");
7131 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
7132 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
7133 MODULE_AUTHOR("Shaobo Xu <xushaobo2@huawei.com>");
7134 MODULE_DESCRIPTION("Hisilicon Hip08 Family RoCE Driver");
7135