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