xref: /linux/drivers/infiniband/hw/hns/hns_roce_hw_v2.c (revision 7f4f3b14e8079ecde096bd734af10e30d40c27b7)
1 /*
2  * Copyright (c) 2016-2017 Hisilicon Limited.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/acpi.h>
34 #include <linux/etherdevice.h>
35 #include <linux/interrupt.h>
36 #include <linux/iopoll.h>
37 #include <linux/kernel.h>
38 #include <linux/types.h>
39 #include <linux/workqueue.h>
40 #include <net/addrconf.h>
41 #include <rdma/ib_addr.h>
42 #include <rdma/ib_cache.h>
43 #include <rdma/ib_umem.h>
44 #include <rdma/uverbs_ioctl.h>
45 
46 #include "hnae3.h"
47 #include "hns_roce_common.h"
48 #include "hns_roce_device.h"
49 #include "hns_roce_cmd.h"
50 #include "hns_roce_hem.h"
51 #include "hns_roce_hw_v2.h"
52 
53 enum {
54 	CMD_RST_PRC_OTHERS,
55 	CMD_RST_PRC_SUCCESS,
56 	CMD_RST_PRC_EBUSY,
57 };
58 
59 enum ecc_resource_type {
60 	ECC_RESOURCE_QPC,
61 	ECC_RESOURCE_CQC,
62 	ECC_RESOURCE_MPT,
63 	ECC_RESOURCE_SRQC,
64 	ECC_RESOURCE_GMV,
65 	ECC_RESOURCE_QPC_TIMER,
66 	ECC_RESOURCE_CQC_TIMER,
67 	ECC_RESOURCE_SCCC,
68 	ECC_RESOURCE_COUNT,
69 };
70 
71 static const struct {
72 	const char *name;
73 	u8 read_bt0_op;
74 	u8 write_bt0_op;
75 } fmea_ram_res[] = {
76 	{ "ECC_RESOURCE_QPC",
77 	  HNS_ROCE_CMD_READ_QPC_BT0, HNS_ROCE_CMD_WRITE_QPC_BT0 },
78 	{ "ECC_RESOURCE_CQC",
79 	  HNS_ROCE_CMD_READ_CQC_BT0, HNS_ROCE_CMD_WRITE_CQC_BT0 },
80 	{ "ECC_RESOURCE_MPT",
81 	  HNS_ROCE_CMD_READ_MPT_BT0, HNS_ROCE_CMD_WRITE_MPT_BT0 },
82 	{ "ECC_RESOURCE_SRQC",
83 	  HNS_ROCE_CMD_READ_SRQC_BT0, HNS_ROCE_CMD_WRITE_SRQC_BT0 },
84 	/* ECC_RESOURCE_GMV is handled by cmdq, not mailbox */
85 	{ "ECC_RESOURCE_GMV",
86 	  0, 0 },
87 	{ "ECC_RESOURCE_QPC_TIMER",
88 	  HNS_ROCE_CMD_READ_QPC_TIMER_BT0, HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0 },
89 	{ "ECC_RESOURCE_CQC_TIMER",
90 	  HNS_ROCE_CMD_READ_CQC_TIMER_BT0, HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0 },
91 	{ "ECC_RESOURCE_SCCC",
92 	  HNS_ROCE_CMD_READ_SCCC_BT0, HNS_ROCE_CMD_WRITE_SCCC_BT0 },
93 };
94 
95 static inline void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
96 				   struct ib_sge *sg)
97 {
98 	dseg->lkey = cpu_to_le32(sg->lkey);
99 	dseg->addr = cpu_to_le64(sg->addr);
100 	dseg->len  = cpu_to_le32(sg->length);
101 }
102 
103 /*
104  * mapped-value = 1 + real-value
105  * The hns wr opcode real value is start from 0, In order to distinguish between
106  * initialized and uninitialized map values, we plus 1 to the actual value when
107  * defining the mapping, so that the validity can be identified by checking the
108  * mapped value is greater than 0.
109  */
110 #define HR_OPC_MAP(ib_key, hr_key) \
111 		[IB_WR_ ## ib_key] = 1 + HNS_ROCE_V2_WQE_OP_ ## hr_key
112 
113 static const u32 hns_roce_op_code[] = {
114 	HR_OPC_MAP(RDMA_WRITE,			RDMA_WRITE),
115 	HR_OPC_MAP(RDMA_WRITE_WITH_IMM,		RDMA_WRITE_WITH_IMM),
116 	HR_OPC_MAP(SEND,			SEND),
117 	HR_OPC_MAP(SEND_WITH_IMM,		SEND_WITH_IMM),
118 	HR_OPC_MAP(RDMA_READ,			RDMA_READ),
119 	HR_OPC_MAP(ATOMIC_CMP_AND_SWP,		ATOM_CMP_AND_SWAP),
120 	HR_OPC_MAP(ATOMIC_FETCH_AND_ADD,	ATOM_FETCH_AND_ADD),
121 	HR_OPC_MAP(SEND_WITH_INV,		SEND_WITH_INV),
122 	HR_OPC_MAP(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
123 	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
124 	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
125 };
126 
127 static u32 to_hr_opcode(u32 ib_opcode)
128 {
129 	if (ib_opcode >= ARRAY_SIZE(hns_roce_op_code))
130 		return HNS_ROCE_V2_WQE_OP_MASK;
131 
132 	return hns_roce_op_code[ib_opcode] ? hns_roce_op_code[ib_opcode] - 1 :
133 					     HNS_ROCE_V2_WQE_OP_MASK;
134 }
135 
136 static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
137 			 const struct ib_reg_wr *wr)
138 {
139 	struct hns_roce_wqe_frmr_seg *fseg =
140 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
141 	struct hns_roce_mr *mr = to_hr_mr(wr->mr);
142 	u64 pbl_ba;
143 
144 	/* use ib_access_flags */
145 	hr_reg_write_bool(fseg, FRMR_BIND_EN, wr->access & IB_ACCESS_MW_BIND);
146 	hr_reg_write_bool(fseg, FRMR_ATOMIC,
147 			  wr->access & IB_ACCESS_REMOTE_ATOMIC);
148 	hr_reg_write_bool(fseg, FRMR_RR, wr->access & IB_ACCESS_REMOTE_READ);
149 	hr_reg_write_bool(fseg, FRMR_RW, wr->access & IB_ACCESS_REMOTE_WRITE);
150 	hr_reg_write_bool(fseg, FRMR_LW, wr->access & IB_ACCESS_LOCAL_WRITE);
151 
152 	/* Data structure reuse may lead to confusion */
153 	pbl_ba = mr->pbl_mtr.hem_cfg.root_ba;
154 	rc_sq_wqe->msg_len = cpu_to_le32(lower_32_bits(pbl_ba));
155 	rc_sq_wqe->inv_key = cpu_to_le32(upper_32_bits(pbl_ba));
156 
157 	rc_sq_wqe->byte_16 = cpu_to_le32(wr->mr->length & 0xffffffff);
158 	rc_sq_wqe->byte_20 = cpu_to_le32(wr->mr->length >> 32);
159 	rc_sq_wqe->rkey = cpu_to_le32(wr->key);
160 	rc_sq_wqe->va = cpu_to_le64(wr->mr->iova);
161 
162 	hr_reg_write(fseg, FRMR_PBL_SIZE, mr->npages);
163 	hr_reg_write(fseg, FRMR_PBL_BUF_PG_SZ,
164 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
165 	hr_reg_clear(fseg, FRMR_BLK_MODE);
166 }
167 
168 static void set_atomic_seg(const struct ib_send_wr *wr,
169 			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
170 			   unsigned int valid_num_sge)
171 {
172 	struct hns_roce_v2_wqe_data_seg *dseg =
173 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
174 	struct hns_roce_wqe_atomic_seg *aseg =
175 		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
176 
177 	set_data_seg_v2(dseg, wr->sg_list);
178 
179 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
180 		aseg->fetchadd_swap_data = cpu_to_le64(atomic_wr(wr)->swap);
181 		aseg->cmp_data = cpu_to_le64(atomic_wr(wr)->compare_add);
182 	} else {
183 		aseg->fetchadd_swap_data =
184 			cpu_to_le64(atomic_wr(wr)->compare_add);
185 		aseg->cmp_data = 0;
186 	}
187 
188 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
189 }
190 
191 static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
192 				 const struct ib_send_wr *wr,
193 				 unsigned int *sge_idx, u32 msg_len)
194 {
195 	struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
196 	unsigned int left_len_in_pg;
197 	unsigned int idx = *sge_idx;
198 	unsigned int i = 0;
199 	unsigned int len;
200 	void *addr;
201 	void *dseg;
202 
203 	if (msg_len > qp->sq.ext_sge_cnt * HNS_ROCE_SGE_SIZE) {
204 		ibdev_err(ibdev,
205 			  "no enough extended sge space for inline data.\n");
206 		return -EINVAL;
207 	}
208 
209 	dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
210 	left_len_in_pg = hr_hw_page_align((uintptr_t)dseg) - (uintptr_t)dseg;
211 	len = wr->sg_list[0].length;
212 	addr = (void *)(unsigned long)(wr->sg_list[0].addr);
213 
214 	/* When copying data to extended sge space, the left length in page may
215 	 * not long enough for current user's sge. So the data should be
216 	 * splited into several parts, one in the first page, and the others in
217 	 * the subsequent pages.
218 	 */
219 	while (1) {
220 		if (len <= left_len_in_pg) {
221 			memcpy(dseg, addr, len);
222 
223 			idx += len / HNS_ROCE_SGE_SIZE;
224 
225 			i++;
226 			if (i >= wr->num_sge)
227 				break;
228 
229 			left_len_in_pg -= len;
230 			len = wr->sg_list[i].length;
231 			addr = (void *)(unsigned long)(wr->sg_list[i].addr);
232 			dseg += len;
233 		} else {
234 			memcpy(dseg, addr, left_len_in_pg);
235 
236 			len -= left_len_in_pg;
237 			addr += left_len_in_pg;
238 			idx += left_len_in_pg / HNS_ROCE_SGE_SIZE;
239 			dseg = hns_roce_get_extend_sge(qp,
240 						idx & (qp->sge.sge_cnt - 1));
241 			left_len_in_pg = 1 << HNS_HW_PAGE_SHIFT;
242 		}
243 	}
244 
245 	*sge_idx = idx;
246 
247 	return 0;
248 }
249 
250 static void set_extend_sge(struct hns_roce_qp *qp, struct ib_sge *sge,
251 			   unsigned int *sge_ind, unsigned int cnt)
252 {
253 	struct hns_roce_v2_wqe_data_seg *dseg;
254 	unsigned int idx = *sge_ind;
255 
256 	while (cnt > 0) {
257 		dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
258 		if (likely(sge->length)) {
259 			set_data_seg_v2(dseg, sge);
260 			idx++;
261 			cnt--;
262 		}
263 		sge++;
264 	}
265 
266 	*sge_ind = idx;
267 }
268 
269 static bool check_inl_data_len(struct hns_roce_qp *qp, unsigned int len)
270 {
271 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
272 	int mtu = ib_mtu_enum_to_int(qp->path_mtu);
273 
274 	if (mtu < 0 || len > qp->max_inline_data || len > mtu) {
275 		ibdev_err(&hr_dev->ib_dev,
276 			  "invalid length of data, data len = %u, max inline len = %u, path mtu = %d.\n",
277 			  len, qp->max_inline_data, mtu);
278 		return false;
279 	}
280 
281 	return true;
282 }
283 
284 static int set_rc_inl(struct hns_roce_qp *qp, const struct ib_send_wr *wr,
285 		      struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
286 		      unsigned int *sge_idx)
287 {
288 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
289 	u32 msg_len = le32_to_cpu(rc_sq_wqe->msg_len);
290 	struct ib_device *ibdev = &hr_dev->ib_dev;
291 	unsigned int curr_idx = *sge_idx;
292 	void *dseg = rc_sq_wqe;
293 	unsigned int i;
294 	int ret;
295 
296 	if (unlikely(wr->opcode == IB_WR_RDMA_READ)) {
297 		ibdev_err(ibdev, "invalid inline parameters!\n");
298 		return -EINVAL;
299 	}
300 
301 	if (!check_inl_data_len(qp, msg_len))
302 		return -EINVAL;
303 
304 	dseg += sizeof(struct hns_roce_v2_rc_send_wqe);
305 
306 	if (msg_len <= HNS_ROCE_V2_MAX_RC_INL_INN_SZ) {
307 		hr_reg_clear(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
308 
309 		for (i = 0; i < wr->num_sge; i++) {
310 			memcpy(dseg, ((void *)wr->sg_list[i].addr),
311 			       wr->sg_list[i].length);
312 			dseg += wr->sg_list[i].length;
313 		}
314 	} else {
315 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
316 
317 		ret = fill_ext_sge_inl_data(qp, wr, &curr_idx, msg_len);
318 		if (ret)
319 			return ret;
320 
321 		hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, curr_idx - *sge_idx);
322 	}
323 
324 	*sge_idx = curr_idx;
325 
326 	return 0;
327 }
328 
329 static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
330 			     struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
331 			     unsigned int *sge_ind,
332 			     unsigned int valid_num_sge)
333 {
334 	struct hns_roce_v2_wqe_data_seg *dseg =
335 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
336 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
337 	int j = 0;
338 	int i;
339 
340 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_MSG_START_SGE_IDX,
341 		     (*sge_ind) & (qp->sge.sge_cnt - 1));
342 
343 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_INLINE,
344 		     !!(wr->send_flags & IB_SEND_INLINE));
345 	if (wr->send_flags & IB_SEND_INLINE)
346 		return set_rc_inl(qp, wr, rc_sq_wqe, sge_ind);
347 
348 	if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
349 		for (i = 0; i < wr->num_sge; i++) {
350 			if (likely(wr->sg_list[i].length)) {
351 				set_data_seg_v2(dseg, wr->sg_list + i);
352 				dseg++;
353 			}
354 		}
355 	} else {
356 		for (i = 0; i < wr->num_sge && j < HNS_ROCE_SGE_IN_WQE; i++) {
357 			if (likely(wr->sg_list[i].length)) {
358 				set_data_seg_v2(dseg, wr->sg_list + i);
359 				dseg++;
360 				j++;
361 			}
362 		}
363 
364 		set_extend_sge(qp, wr->sg_list + i, sge_ind,
365 			       valid_num_sge - HNS_ROCE_SGE_IN_WQE);
366 	}
367 
368 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
369 
370 	return 0;
371 }
372 
373 static int check_send_valid(struct hns_roce_dev *hr_dev,
374 			    struct hns_roce_qp *hr_qp)
375 {
376 	if (unlikely(hr_qp->state == IB_QPS_RESET ||
377 		     hr_qp->state == IB_QPS_INIT ||
378 		     hr_qp->state == IB_QPS_RTR))
379 		return -EINVAL;
380 	else if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
381 		return -EIO;
382 
383 	return 0;
384 }
385 
386 static unsigned int calc_wr_sge_num(const struct ib_send_wr *wr,
387 				    unsigned int *sge_len)
388 {
389 	unsigned int valid_num = 0;
390 	unsigned int len = 0;
391 	int i;
392 
393 	for (i = 0; i < wr->num_sge; i++) {
394 		if (likely(wr->sg_list[i].length)) {
395 			len += wr->sg_list[i].length;
396 			valid_num++;
397 		}
398 	}
399 
400 	*sge_len = len;
401 	return valid_num;
402 }
403 
404 static __le32 get_immtdata(const struct ib_send_wr *wr)
405 {
406 	switch (wr->opcode) {
407 	case IB_WR_SEND_WITH_IMM:
408 	case IB_WR_RDMA_WRITE_WITH_IMM:
409 		return cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
410 	default:
411 		return 0;
412 	}
413 }
414 
415 static int set_ud_opcode(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
416 			 const struct ib_send_wr *wr)
417 {
418 	u32 ib_op = wr->opcode;
419 
420 	if (ib_op != IB_WR_SEND && ib_op != IB_WR_SEND_WITH_IMM)
421 		return -EINVAL;
422 
423 	ud_sq_wqe->immtdata = get_immtdata(wr);
424 
425 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
426 
427 	return 0;
428 }
429 
430 static int fill_ud_av(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
431 		      struct hns_roce_ah *ah)
432 {
433 	struct ib_device *ib_dev = ah->ibah.device;
434 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
435 
436 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_UDPSPN, ah->av.udp_sport);
437 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_HOPLIMIT, ah->av.hop_limit);
438 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_TCLASS, ah->av.tclass);
439 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_FLOW_LABEL, ah->av.flowlabel);
440 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SL, ah->av.sl);
441 
442 	ud_sq_wqe->sgid_index = ah->av.gid_index;
443 
444 	memcpy(ud_sq_wqe->dmac, ah->av.mac, ETH_ALEN);
445 	memcpy(ud_sq_wqe->dgid, ah->av.dgid, GID_LEN_V2);
446 
447 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
448 		return 0;
449 
450 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN_EN, ah->av.vlan_en);
451 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN, ah->av.vlan_id);
452 
453 	return 0;
454 }
455 
456 static inline int set_ud_wqe(struct hns_roce_qp *qp,
457 			     const struct ib_send_wr *wr,
458 			     void *wqe, unsigned int *sge_idx,
459 			     unsigned int owner_bit)
460 {
461 	struct hns_roce_ah *ah = to_hr_ah(ud_wr(wr)->ah);
462 	struct hns_roce_v2_ud_send_wqe *ud_sq_wqe = wqe;
463 	unsigned int curr_idx = *sge_idx;
464 	unsigned int valid_num_sge;
465 	u32 msg_len = 0;
466 	int ret;
467 
468 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
469 
470 	ret = set_ud_opcode(ud_sq_wqe, wr);
471 	if (WARN_ON(ret))
472 		return ret;
473 
474 	ud_sq_wqe->msg_len = cpu_to_le32(msg_len);
475 
476 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_CQE,
477 		     !!(wr->send_flags & IB_SEND_SIGNALED));
478 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SE,
479 		     !!(wr->send_flags & IB_SEND_SOLICITED));
480 
481 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_PD, to_hr_pd(qp->ibqp.pd)->pdn);
482 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SGE_NUM, valid_num_sge);
483 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_MSG_START_SGE_IDX,
484 		     curr_idx & (qp->sge.sge_cnt - 1));
485 
486 	ud_sq_wqe->qkey = cpu_to_le32(ud_wr(wr)->remote_qkey & 0x80000000 ?
487 			  qp->qkey : ud_wr(wr)->remote_qkey);
488 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_DQPN, ud_wr(wr)->remote_qpn);
489 
490 	ret = fill_ud_av(ud_sq_wqe, ah);
491 	if (ret)
492 		return ret;
493 
494 	qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
495 
496 	set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
497 
498 	/*
499 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
500 	 * including new WQEs waiting for the doorbell to update the PI again.
501 	 * Therefore, the owner bit of WQE MUST be updated after all fields
502 	 * and extSGEs have been written into DDR instead of cache.
503 	 */
504 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
505 		dma_wmb();
506 
507 	*sge_idx = curr_idx;
508 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OWNER, owner_bit);
509 
510 	return 0;
511 }
512 
513 static int set_rc_opcode(struct hns_roce_dev *hr_dev,
514 			 struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
515 			 const struct ib_send_wr *wr)
516 {
517 	u32 ib_op = wr->opcode;
518 	int ret = 0;
519 
520 	rc_sq_wqe->immtdata = get_immtdata(wr);
521 
522 	switch (ib_op) {
523 	case IB_WR_RDMA_READ:
524 	case IB_WR_RDMA_WRITE:
525 	case IB_WR_RDMA_WRITE_WITH_IMM:
526 		rc_sq_wqe->rkey = cpu_to_le32(rdma_wr(wr)->rkey);
527 		rc_sq_wqe->va = cpu_to_le64(rdma_wr(wr)->remote_addr);
528 		break;
529 	case IB_WR_SEND:
530 	case IB_WR_SEND_WITH_IMM:
531 		break;
532 	case IB_WR_ATOMIC_CMP_AND_SWP:
533 	case IB_WR_ATOMIC_FETCH_AND_ADD:
534 		rc_sq_wqe->rkey = cpu_to_le32(atomic_wr(wr)->rkey);
535 		rc_sq_wqe->va = cpu_to_le64(atomic_wr(wr)->remote_addr);
536 		break;
537 	case IB_WR_REG_MR:
538 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
539 			set_frmr_seg(rc_sq_wqe, reg_wr(wr));
540 		else
541 			ret = -EOPNOTSUPP;
542 		break;
543 	case IB_WR_SEND_WITH_INV:
544 		rc_sq_wqe->inv_key = cpu_to_le32(wr->ex.invalidate_rkey);
545 		break;
546 	default:
547 		ret = -EINVAL;
548 	}
549 
550 	if (unlikely(ret))
551 		return ret;
552 
553 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
554 
555 	return ret;
556 }
557 
558 static inline int set_rc_wqe(struct hns_roce_qp *qp,
559 			     const struct ib_send_wr *wr,
560 			     void *wqe, unsigned int *sge_idx,
561 			     unsigned int owner_bit)
562 {
563 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
564 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
565 	unsigned int curr_idx = *sge_idx;
566 	unsigned int valid_num_sge;
567 	u32 msg_len = 0;
568 	int ret;
569 
570 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
571 
572 	rc_sq_wqe->msg_len = cpu_to_le32(msg_len);
573 
574 	ret = set_rc_opcode(hr_dev, rc_sq_wqe, wr);
575 	if (WARN_ON(ret))
576 		return ret;
577 
578 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SO,
579 		     (wr->send_flags & IB_SEND_FENCE) ? 1 : 0);
580 
581 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SE,
582 		     (wr->send_flags & IB_SEND_SOLICITED) ? 1 : 0);
583 
584 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_CQE,
585 		     (wr->send_flags & IB_SEND_SIGNALED) ? 1 : 0);
586 
587 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
588 	    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
589 		if (msg_len != ATOMIC_WR_LEN)
590 			return -EINVAL;
591 		set_atomic_seg(wr, rc_sq_wqe, valid_num_sge);
592 	} else if (wr->opcode != IB_WR_REG_MR) {
593 		ret = set_rwqe_data_seg(&qp->ibqp, wr, rc_sq_wqe,
594 					&curr_idx, valid_num_sge);
595 		if (ret)
596 			return ret;
597 	}
598 
599 	/*
600 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
601 	 * including new WQEs waiting for the doorbell to update the PI again.
602 	 * Therefore, the owner bit of WQE MUST be updated after all fields
603 	 * and extSGEs have been written into DDR instead of cache.
604 	 */
605 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
606 		dma_wmb();
607 
608 	*sge_idx = curr_idx;
609 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OWNER, owner_bit);
610 
611 	return ret;
612 }
613 
614 static inline void update_sq_db(struct hns_roce_dev *hr_dev,
615 				struct hns_roce_qp *qp)
616 {
617 	if (unlikely(qp->state == IB_QPS_ERR)) {
618 		flush_cqe(hr_dev, qp);
619 	} else {
620 		struct hns_roce_v2_db sq_db = {};
621 
622 		hr_reg_write(&sq_db, DB_TAG, qp->qpn);
623 		hr_reg_write(&sq_db, DB_CMD, HNS_ROCE_V2_SQ_DB);
624 		hr_reg_write(&sq_db, DB_PI, qp->sq.head);
625 		hr_reg_write(&sq_db, DB_SL, qp->sl);
626 
627 		hns_roce_write64(hr_dev, (__le32 *)&sq_db, qp->sq.db_reg);
628 	}
629 }
630 
631 static inline void update_rq_db(struct hns_roce_dev *hr_dev,
632 				struct hns_roce_qp *qp)
633 {
634 	if (unlikely(qp->state == IB_QPS_ERR)) {
635 		flush_cqe(hr_dev, qp);
636 	} else {
637 		if (likely(qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)) {
638 			*qp->rdb.db_record =
639 					qp->rq.head & V2_DB_PRODUCER_IDX_M;
640 		} else {
641 			struct hns_roce_v2_db rq_db = {};
642 
643 			hr_reg_write(&rq_db, DB_TAG, qp->qpn);
644 			hr_reg_write(&rq_db, DB_CMD, HNS_ROCE_V2_RQ_DB);
645 			hr_reg_write(&rq_db, DB_PI, qp->rq.head);
646 
647 			hns_roce_write64(hr_dev, (__le32 *)&rq_db,
648 					 qp->rq.db_reg);
649 		}
650 	}
651 }
652 
653 static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
654 			      u64 __iomem *dest)
655 {
656 #define HNS_ROCE_WRITE_TIMES 8
657 	struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
658 	struct hnae3_handle *handle = priv->handle;
659 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
660 	int i;
661 
662 	if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
663 		for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
664 			writeq_relaxed(*(val + i), dest + i);
665 }
666 
667 static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
668 		       void *wqe)
669 {
670 #define HNS_ROCE_SL_SHIFT 2
671 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
672 
673 	/* All kinds of DirectWQE have the same header field layout */
674 	hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_FLAG);
675 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_L, qp->sl);
676 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_H,
677 		     qp->sl >> HNS_ROCE_SL_SHIFT);
678 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_WQE_INDEX, qp->sq.head);
679 
680 	hns_roce_write512(hr_dev, wqe, qp->sq.db_reg);
681 }
682 
683 static int hns_roce_v2_post_send(struct ib_qp *ibqp,
684 				 const struct ib_send_wr *wr,
685 				 const struct ib_send_wr **bad_wr)
686 {
687 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
688 	struct ib_device *ibdev = &hr_dev->ib_dev;
689 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
690 	unsigned long flags = 0;
691 	unsigned int owner_bit;
692 	unsigned int sge_idx;
693 	unsigned int wqe_idx;
694 	void *wqe = NULL;
695 	u32 nreq;
696 	int ret;
697 
698 	spin_lock_irqsave(&qp->sq.lock, flags);
699 
700 	ret = check_send_valid(hr_dev, qp);
701 	if (unlikely(ret)) {
702 		*bad_wr = wr;
703 		nreq = 0;
704 		goto out;
705 	}
706 
707 	sge_idx = qp->next_sge;
708 
709 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
710 		if (hns_roce_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
711 			ret = -ENOMEM;
712 			*bad_wr = wr;
713 			goto out;
714 		}
715 
716 		wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
717 
718 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
719 			ibdev_err(ibdev, "num_sge = %d > qp->sq.max_gs = %u.\n",
720 				  wr->num_sge, qp->sq.max_gs);
721 			ret = -EINVAL;
722 			*bad_wr = wr;
723 			goto out;
724 		}
725 
726 		wqe = hns_roce_get_send_wqe(qp, wqe_idx);
727 		qp->sq.wrid[wqe_idx] = wr->wr_id;
728 		owner_bit =
729 		       ~(((qp->sq.head + nreq) >> ilog2(qp->sq.wqe_cnt)) & 0x1);
730 
731 		/* Corresponding to the QP type, wqe process separately */
732 		if (ibqp->qp_type == IB_QPT_RC)
733 			ret = set_rc_wqe(qp, wr, wqe, &sge_idx, owner_bit);
734 		else
735 			ret = set_ud_wqe(qp, wr, wqe, &sge_idx, owner_bit);
736 
737 		if (unlikely(ret)) {
738 			*bad_wr = wr;
739 			goto out;
740 		}
741 	}
742 
743 out:
744 	if (likely(nreq)) {
745 		qp->sq.head += nreq;
746 		qp->next_sge = sge_idx;
747 
748 		if (nreq == 1 && !ret &&
749 		    (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
750 			write_dwqe(hr_dev, qp, wqe);
751 		else
752 			update_sq_db(hr_dev, qp);
753 	}
754 
755 	spin_unlock_irqrestore(&qp->sq.lock, flags);
756 
757 	return ret;
758 }
759 
760 static int check_recv_valid(struct hns_roce_dev *hr_dev,
761 			    struct hns_roce_qp *hr_qp)
762 {
763 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
764 		return -EIO;
765 
766 	if (hr_qp->state == IB_QPS_RESET)
767 		return -EINVAL;
768 
769 	return 0;
770 }
771 
772 static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
773 				 u32 max_sge, bool rsv)
774 {
775 	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
776 	u32 i, cnt;
777 
778 	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
779 		/* Skip zero-length sge */
780 		if (!wr->sg_list[i].length)
781 			continue;
782 		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
783 		cnt++;
784 	}
785 
786 	/* Fill a reserved sge to make hw stop reading remaining segments */
787 	if (rsv) {
788 		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
789 		dseg[cnt].addr = 0;
790 		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
791 	} else {
792 		/* Clear remaining segments to make ROCEE ignore sges */
793 		if (cnt < max_sge)
794 			memset(dseg + cnt, 0,
795 			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
796 	}
797 }
798 
799 static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
800 			u32 wqe_idx, u32 max_sge)
801 {
802 	void *wqe = NULL;
803 
804 	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
805 	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
806 }
807 
808 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
809 				 const struct ib_recv_wr *wr,
810 				 const struct ib_recv_wr **bad_wr)
811 {
812 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
813 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
814 	struct ib_device *ibdev = &hr_dev->ib_dev;
815 	u32 wqe_idx, nreq, max_sge;
816 	unsigned long flags;
817 	int ret;
818 
819 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
820 
821 	ret = check_recv_valid(hr_dev, hr_qp);
822 	if (unlikely(ret)) {
823 		*bad_wr = wr;
824 		nreq = 0;
825 		goto out;
826 	}
827 
828 	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
829 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
830 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
831 						  hr_qp->ibqp.recv_cq))) {
832 			ret = -ENOMEM;
833 			*bad_wr = wr;
834 			goto out;
835 		}
836 
837 		if (unlikely(wr->num_sge > max_sge)) {
838 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
839 				  wr->num_sge, max_sge);
840 			ret = -EINVAL;
841 			*bad_wr = wr;
842 			goto out;
843 		}
844 
845 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
846 		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
847 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
848 	}
849 
850 out:
851 	if (likely(nreq)) {
852 		hr_qp->rq.head += nreq;
853 
854 		update_rq_db(hr_dev, hr_qp);
855 	}
856 	spin_unlock_irqrestore(&hr_qp->rq.lock, flags);
857 
858 	return ret;
859 }
860 
861 static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
862 {
863 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
864 }
865 
866 static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
867 {
868 	return hns_roce_buf_offset(idx_que->mtr.kmem,
869 				   n << idx_que->entry_shift);
870 }
871 
872 static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
873 {
874 	/* always called with interrupts disabled. */
875 	spin_lock(&srq->lock);
876 
877 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
878 	srq->idx_que.tail++;
879 
880 	spin_unlock(&srq->lock);
881 }
882 
883 static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
884 {
885 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
886 
887 	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
888 }
889 
890 static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
891 				const struct ib_recv_wr *wr)
892 {
893 	struct ib_device *ib_dev = srq->ibsrq.device;
894 
895 	if (unlikely(wr->num_sge > max_sge)) {
896 		ibdev_err(ib_dev,
897 			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
898 			  wr->num_sge, max_sge);
899 		return -EINVAL;
900 	}
901 
902 	if (unlikely(hns_roce_srqwq_overflow(srq))) {
903 		ibdev_err(ib_dev,
904 			  "failed to check srqwq status, srqwq is full.\n");
905 		return -ENOMEM;
906 	}
907 
908 	return 0;
909 }
910 
911 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
912 {
913 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
914 	u32 pos;
915 
916 	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
917 	if (unlikely(pos == srq->wqe_cnt))
918 		return -ENOSPC;
919 
920 	bitmap_set(idx_que->bitmap, pos, 1);
921 	*wqe_idx = pos;
922 	return 0;
923 }
924 
925 static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
926 {
927 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
928 	unsigned int head;
929 	__le32 *buf;
930 
931 	head = idx_que->head & (srq->wqe_cnt - 1);
932 
933 	buf = get_idx_buf(idx_que, head);
934 	*buf = cpu_to_le32(wqe_idx);
935 
936 	idx_que->head++;
937 }
938 
939 static void update_srq_db(struct hns_roce_srq *srq)
940 {
941 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
942 	struct hns_roce_v2_db db;
943 
944 	hr_reg_write(&db, DB_TAG, srq->srqn);
945 	hr_reg_write(&db, DB_CMD, HNS_ROCE_V2_SRQ_DB);
946 	hr_reg_write(&db, DB_PI, srq->idx_que.head);
947 
948 	hns_roce_write64(hr_dev, (__le32 *)&db, srq->db_reg);
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_srq *srq = to_hr_srq(ibsrq);
956 	unsigned long flags;
957 	int ret = 0;
958 	u32 max_sge;
959 	u32 wqe_idx;
960 	void *wqe;
961 	u32 nreq;
962 
963 	spin_lock_irqsave(&srq->lock, flags);
964 
965 	max_sge = srq->max_gs - srq->rsv_sge;
966 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
967 		ret = check_post_srq_valid(srq, max_sge, wr);
968 		if (ret) {
969 			*bad_wr = wr;
970 			break;
971 		}
972 
973 		ret = get_srq_wqe_idx(srq, &wqe_idx);
974 		if (unlikely(ret)) {
975 			*bad_wr = wr;
976 			break;
977 		}
978 
979 		wqe = get_srq_wqe_buf(srq, wqe_idx);
980 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
981 		fill_wqe_idx(srq, wqe_idx);
982 		srq->wrid[wqe_idx] = wr->wr_id;
983 	}
984 
985 	if (likely(nreq)) {
986 		if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB)
987 			*srq->rdb.db_record = srq->idx_que.head &
988 					      V2_DB_PRODUCER_IDX_M;
989 		else
990 			update_srq_db(srq);
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 u32 hns_roce_cmdq_tx_timeout(u16 opcode, u32 tx_timeout)
1272 {
1273 	static const struct hns_roce_cmdq_tx_timeout_map cmdq_tx_timeout[] = {
1274 		{HNS_ROCE_OPC_POST_MB, HNS_ROCE_OPC_POST_MB_TIMEOUT},
1275 	};
1276 	int i;
1277 
1278 	for (i = 0; i < ARRAY_SIZE(cmdq_tx_timeout); i++)
1279 		if (cmdq_tx_timeout[i].opcode == opcode)
1280 			return cmdq_tx_timeout[i].tx_timeout;
1281 
1282 	return tx_timeout;
1283 }
1284 
1285 static void hns_roce_wait_csq_done(struct hns_roce_dev *hr_dev, u16 opcode)
1286 {
1287 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1288 	u32 tx_timeout = hns_roce_cmdq_tx_timeout(opcode, priv->cmq.tx_timeout);
1289 	u32 timeout = 0;
1290 
1291 	do {
1292 		if (hns_roce_cmq_csq_done(hr_dev))
1293 			break;
1294 		udelay(1);
1295 	} while (++timeout < tx_timeout);
1296 }
1297 
1298 static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1299 			       struct hns_roce_cmq_desc *desc, int num)
1300 {
1301 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1302 	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
1303 	u16 desc_ret;
1304 	u32 tail;
1305 	int ret;
1306 	int i;
1307 
1308 	spin_lock_bh(&csq->lock);
1309 
1310 	tail = csq->head;
1311 
1312 	for (i = 0; i < num; i++) {
1313 		csq->desc[csq->head++] = desc[i];
1314 		if (csq->head == csq->desc_num)
1315 			csq->head = 0;
1316 	}
1317 
1318 	/* Write to hardware */
1319 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, csq->head);
1320 
1321 	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_CNT]);
1322 
1323 	hns_roce_wait_csq_done(hr_dev, le16_to_cpu(desc->opcode));
1324 	if (hns_roce_cmq_csq_done(hr_dev)) {
1325 		ret = 0;
1326 		for (i = 0; i < num; i++) {
1327 			/* check the result of hardware write back */
1328 			desc[i] = csq->desc[tail++];
1329 			if (tail == csq->desc_num)
1330 				tail = 0;
1331 
1332 			desc_ret = le16_to_cpu(desc[i].retval);
1333 			if (likely(desc_ret == CMD_EXEC_SUCCESS))
1334 				continue;
1335 
1336 			dev_err_ratelimited(hr_dev->dev,
1337 					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",
1338 					    desc->opcode, desc_ret);
1339 			ret = hns_roce_cmd_err_convert_errno(desc_ret);
1340 		}
1341 	} else {
1342 		/* FW/HW reset or incorrect number of desc */
1343 		tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1344 		dev_warn(hr_dev->dev, "CMDQ move tail from %u to %u.\n",
1345 			 csq->head, tail);
1346 		csq->head = tail;
1347 
1348 		update_cmdq_status(hr_dev);
1349 
1350 		ret = -EAGAIN;
1351 	}
1352 
1353 	spin_unlock_bh(&csq->lock);
1354 
1355 	if (ret)
1356 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_ERR_CNT]);
1357 
1358 	return ret;
1359 }
1360 
1361 static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1362 			     struct hns_roce_cmq_desc *desc, int num)
1363 {
1364 	bool busy;
1365 	int ret;
1366 
1367 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1368 		return -EIO;
1369 
1370 	if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1371 		return busy ? -EBUSY : 0;
1372 
1373 	ret = __hns_roce_cmq_send(hr_dev, desc, num);
1374 	if (ret) {
1375 		if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1376 			return busy ? -EBUSY : 0;
1377 	}
1378 
1379 	return ret;
1380 }
1381 
1382 static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev,
1383 			       dma_addr_t base_addr, u8 cmd, unsigned long tag)
1384 {
1385 	struct hns_roce_cmd_mailbox *mbox;
1386 	int ret;
1387 
1388 	mbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1389 	if (IS_ERR(mbox))
1390 		return PTR_ERR(mbox);
1391 
1392 	ret = hns_roce_cmd_mbox(hr_dev, base_addr, mbox->dma, cmd, tag);
1393 	hns_roce_free_cmd_mailbox(hr_dev, mbox);
1394 	return ret;
1395 }
1396 
1397 static int hns_roce_cmq_query_hw_info(struct hns_roce_dev *hr_dev)
1398 {
1399 	struct hns_roce_query_version *resp;
1400 	struct hns_roce_cmq_desc desc;
1401 	int ret;
1402 
1403 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_HW_VER, true);
1404 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1405 	if (ret)
1406 		return ret;
1407 
1408 	resp = (struct hns_roce_query_version *)desc.data;
1409 	hr_dev->hw_rev = le16_to_cpu(resp->rocee_hw_version);
1410 	hr_dev->vendor_id = hr_dev->pci_dev->vendor;
1411 
1412 	return 0;
1413 }
1414 
1415 static void func_clr_hw_resetting_state(struct hns_roce_dev *hr_dev,
1416 					struct hnae3_handle *handle)
1417 {
1418 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1419 	unsigned long end;
1420 
1421 	hr_dev->dis_db = true;
1422 
1423 	dev_warn(hr_dev->dev,
1424 		 "func clear is pending, device in resetting state.\n");
1425 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1426 	while (end) {
1427 		if (!ops->get_hw_reset_stat(handle)) {
1428 			hr_dev->is_reset = true;
1429 			dev_info(hr_dev->dev,
1430 				 "func clear success after 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.\n");
1438 }
1439 
1440 static void func_clr_sw_resetting_state(struct hns_roce_dev *hr_dev,
1441 					struct hnae3_handle *handle)
1442 {
1443 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1444 	unsigned long end;
1445 
1446 	hr_dev->dis_db = true;
1447 
1448 	dev_warn(hr_dev->dev,
1449 		 "func clear is pending, device in resetting state.\n");
1450 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1451 	while (end) {
1452 		if (ops->ae_dev_reset_cnt(handle) !=
1453 		    hr_dev->reset_cnt) {
1454 			hr_dev->is_reset = true;
1455 			dev_info(hr_dev->dev,
1456 				 "func clear success after sw reset\n");
1457 			return;
1458 		}
1459 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1460 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1461 	}
1462 
1463 	dev_warn(hr_dev->dev, "func clear failed because of unfinished sw reset\n");
1464 }
1465 
1466 static void hns_roce_func_clr_rst_proc(struct hns_roce_dev *hr_dev, int retval,
1467 				       int flag)
1468 {
1469 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1470 	struct hnae3_handle *handle = priv->handle;
1471 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1472 
1473 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt) {
1474 		hr_dev->dis_db = true;
1475 		hr_dev->is_reset = true;
1476 		dev_info(hr_dev->dev, "func clear success after reset.\n");
1477 		return;
1478 	}
1479 
1480 	if (ops->get_hw_reset_stat(handle)) {
1481 		func_clr_hw_resetting_state(hr_dev, handle);
1482 		return;
1483 	}
1484 
1485 	if (ops->ae_dev_resetting(handle) &&
1486 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT) {
1487 		func_clr_sw_resetting_state(hr_dev, handle);
1488 		return;
1489 	}
1490 
1491 	if (retval && !flag)
1492 		dev_warn(hr_dev->dev,
1493 			 "func clear read failed, ret = %d.\n", retval);
1494 
1495 	dev_warn(hr_dev->dev, "func clear failed.\n");
1496 }
1497 
1498 static void __hns_roce_function_clear(struct hns_roce_dev *hr_dev, int vf_id)
1499 {
1500 	bool fclr_write_fail_flag = false;
1501 	struct hns_roce_func_clear *resp;
1502 	struct hns_roce_cmq_desc desc;
1503 	unsigned long end;
1504 	int ret = 0;
1505 
1506 	if (check_device_is_in_reset(hr_dev))
1507 		goto out;
1508 
1509 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR, false);
1510 	resp = (struct hns_roce_func_clear *)desc.data;
1511 	resp->rst_funcid_en = cpu_to_le32(vf_id);
1512 
1513 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1514 	if (ret) {
1515 		fclr_write_fail_flag = true;
1516 		dev_err(hr_dev->dev, "func clear write failed, ret = %d.\n",
1517 			 ret);
1518 		goto out;
1519 	}
1520 
1521 	msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_INTERVAL);
1522 	end = HNS_ROCE_V2_FUNC_CLEAR_TIMEOUT_MSECS;
1523 	while (end) {
1524 		if (check_device_is_in_reset(hr_dev))
1525 			goto out;
1526 		msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT);
1527 		end -= HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT;
1528 
1529 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR,
1530 					      true);
1531 
1532 		resp->rst_funcid_en = cpu_to_le32(vf_id);
1533 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1534 		if (ret)
1535 			continue;
1536 
1537 		if (hr_reg_read(resp, FUNC_CLEAR_RST_FUN_DONE)) {
1538 			if (vf_id == 0)
1539 				hr_dev->is_reset = true;
1540 			return;
1541 		}
1542 	}
1543 
1544 out:
1545 	hns_roce_func_clr_rst_proc(hr_dev, ret, fclr_write_fail_flag);
1546 }
1547 
1548 static int hns_roce_free_vf_resource(struct hns_roce_dev *hr_dev, int vf_id)
1549 {
1550 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1551 	struct hns_roce_cmq_desc desc[2];
1552 	struct hns_roce_cmq_req *req_a;
1553 
1554 	req_a = (struct hns_roce_cmq_req *)desc[0].data;
1555 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1556 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1557 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1558 	hr_reg_write(req_a, FUNC_RES_A_VF_ID, vf_id);
1559 
1560 	return hns_roce_cmq_send(hr_dev, desc, 2);
1561 }
1562 
1563 static void hns_roce_function_clear(struct hns_roce_dev *hr_dev)
1564 {
1565 	int ret;
1566 	int i;
1567 
1568 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1569 		return;
1570 
1571 	for (i = hr_dev->func_num - 1; i >= 0; i--) {
1572 		__hns_roce_function_clear(hr_dev, i);
1573 
1574 		if (i == 0)
1575 			continue;
1576 
1577 		ret = hns_roce_free_vf_resource(hr_dev, i);
1578 		if (ret)
1579 			ibdev_err(&hr_dev->ib_dev,
1580 				  "failed to free vf resource, vf_id = %d, ret = %d.\n",
1581 				  i, ret);
1582 	}
1583 }
1584 
1585 static int hns_roce_clear_extdb_list_info(struct hns_roce_dev *hr_dev)
1586 {
1587 	struct hns_roce_cmq_desc desc;
1588 	int ret;
1589 
1590 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLEAR_EXTDB_LIST_INFO,
1591 				      false);
1592 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1593 	if (ret)
1594 		ibdev_err(&hr_dev->ib_dev,
1595 			  "failed to clear extended doorbell info, ret = %d.\n",
1596 			  ret);
1597 
1598 	return ret;
1599 }
1600 
1601 static int hns_roce_query_fw_ver(struct hns_roce_dev *hr_dev)
1602 {
1603 	struct hns_roce_query_fw_info *resp;
1604 	struct hns_roce_cmq_desc desc;
1605 	int ret;
1606 
1607 	hns_roce_cmq_setup_basic_desc(&desc, HNS_QUERY_FW_VER, true);
1608 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1609 	if (ret)
1610 		return ret;
1611 
1612 	resp = (struct hns_roce_query_fw_info *)desc.data;
1613 	hr_dev->caps.fw_ver = (u64)(le32_to_cpu(resp->fw_ver));
1614 
1615 	return 0;
1616 }
1617 
1618 static int hns_roce_query_func_info(struct hns_roce_dev *hr_dev)
1619 {
1620 	struct hns_roce_cmq_desc desc;
1621 	int ret;
1622 
1623 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
1624 		hr_dev->func_num = 1;
1625 		return 0;
1626 	}
1627 
1628 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_FUNC_INFO,
1629 				      true);
1630 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1631 	if (ret) {
1632 		hr_dev->func_num = 1;
1633 		return ret;
1634 	}
1635 
1636 	hr_dev->func_num = le32_to_cpu(desc.func_info.own_func_num);
1637 	hr_dev->cong_algo_tmpl_id = le32_to_cpu(desc.func_info.own_mac_id);
1638 
1639 	return 0;
1640 }
1641 
1642 static int hns_roce_hw_v2_query_counter(struct hns_roce_dev *hr_dev,
1643 					u64 *stats, u32 port, int *num_counters)
1644 {
1645 #define CNT_PER_DESC 3
1646 	struct hns_roce_cmq_desc *desc;
1647 	int bd_idx, cnt_idx;
1648 	__le64 *cnt_data;
1649 	int desc_num;
1650 	int ret;
1651 	int i;
1652 
1653 	if (port > hr_dev->caps.num_ports)
1654 		return -EINVAL;
1655 
1656 	desc_num = DIV_ROUND_UP(HNS_ROCE_HW_CNT_TOTAL, CNT_PER_DESC);
1657 	desc = kcalloc(desc_num, sizeof(*desc), GFP_KERNEL);
1658 	if (!desc)
1659 		return -ENOMEM;
1660 
1661 	for (i = 0; i < desc_num; i++) {
1662 		hns_roce_cmq_setup_basic_desc(&desc[i],
1663 					      HNS_ROCE_OPC_QUERY_COUNTER, true);
1664 		if (i != desc_num - 1)
1665 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1666 	}
1667 
1668 	ret = hns_roce_cmq_send(hr_dev, desc, desc_num);
1669 	if (ret) {
1670 		ibdev_err(&hr_dev->ib_dev,
1671 			  "failed to get counter, ret = %d.\n", ret);
1672 		goto err_out;
1673 	}
1674 
1675 	for (i = 0; i < HNS_ROCE_HW_CNT_TOTAL && i < *num_counters; i++) {
1676 		bd_idx = i / CNT_PER_DESC;
1677 		if (bd_idx != HNS_ROCE_HW_CNT_TOTAL / CNT_PER_DESC &&
1678 		    !(desc[bd_idx].flag & cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT)))
1679 			break;
1680 
1681 		cnt_data = (__le64 *)&desc[bd_idx].data[0];
1682 		cnt_idx = i % CNT_PER_DESC;
1683 		stats[i] = le64_to_cpu(cnt_data[cnt_idx]);
1684 	}
1685 	*num_counters = i;
1686 
1687 err_out:
1688 	kfree(desc);
1689 	return ret;
1690 }
1691 
1692 static int hns_roce_config_global_param(struct hns_roce_dev *hr_dev)
1693 {
1694 	struct hns_roce_cmq_desc desc;
1695 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1696 	u32 clock_cycles_of_1us;
1697 
1698 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GLOBAL_PARAM,
1699 				      false);
1700 
1701 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
1702 		clock_cycles_of_1us = HNS_ROCE_1NS_CFG;
1703 	else
1704 		clock_cycles_of_1us = HNS_ROCE_1US_CFG;
1705 
1706 	hr_reg_write(req, CFG_GLOBAL_PARAM_1US_CYCLES, clock_cycles_of_1us);
1707 	hr_reg_write(req, CFG_GLOBAL_PARAM_UDP_PORT, ROCE_V2_UDP_DPORT);
1708 
1709 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1710 }
1711 
1712 static int load_func_res_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1713 {
1714 	struct hns_roce_cmq_desc desc[2];
1715 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1716 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1717 	struct hns_roce_caps *caps = &hr_dev->caps;
1718 	enum hns_roce_opcode_type opcode;
1719 	u32 func_num;
1720 	int ret;
1721 
1722 	if (is_vf) {
1723 		opcode = HNS_ROCE_OPC_QUERY_VF_RES;
1724 		func_num = 1;
1725 	} else {
1726 		opcode = HNS_ROCE_OPC_QUERY_PF_RES;
1727 		func_num = hr_dev->func_num;
1728 	}
1729 
1730 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, true);
1731 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1732 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, true);
1733 
1734 	ret = hns_roce_cmq_send(hr_dev, desc, 2);
1735 	if (ret)
1736 		return ret;
1737 
1738 	caps->qpc_bt_num = hr_reg_read(r_a, FUNC_RES_A_QPC_BT_NUM) / func_num;
1739 	caps->srqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_SRQC_BT_NUM) / func_num;
1740 	caps->cqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_CQC_BT_NUM) / func_num;
1741 	caps->mpt_bt_num = hr_reg_read(r_a, FUNC_RES_A_MPT_BT_NUM) / func_num;
1742 	caps->eqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_EQC_BT_NUM) / func_num;
1743 	caps->smac_bt_num = hr_reg_read(r_b, FUNC_RES_B_SMAC_NUM) / func_num;
1744 	caps->sgid_bt_num = hr_reg_read(r_b, FUNC_RES_B_SGID_NUM) / func_num;
1745 	caps->sccc_bt_num = hr_reg_read(r_b, FUNC_RES_B_SCCC_BT_NUM) / func_num;
1746 
1747 	if (is_vf) {
1748 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_V_QID_NUM) / func_num;
1749 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_V_GMV_BT_NUM) /
1750 					       func_num;
1751 	} else {
1752 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_B_QID_NUM) / func_num;
1753 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_B_GMV_BT_NUM) /
1754 					       func_num;
1755 	}
1756 
1757 	return 0;
1758 }
1759 
1760 static int load_pf_timer_res_caps(struct hns_roce_dev *hr_dev)
1761 {
1762 	struct hns_roce_cmq_desc desc;
1763 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1764 	struct hns_roce_caps *caps = &hr_dev->caps;
1765 	int ret;
1766 
1767 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
1768 				      true);
1769 
1770 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1771 	if (ret)
1772 		return ret;
1773 
1774 	caps->qpc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_QPC_ITEM_NUM);
1775 	caps->cqc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_CQC_ITEM_NUM);
1776 
1777 	return 0;
1778 }
1779 
1780 static int hns_roce_query_pf_resource(struct hns_roce_dev *hr_dev)
1781 {
1782 	struct device *dev = hr_dev->dev;
1783 	int ret;
1784 
1785 	ret = load_func_res_caps(hr_dev, false);
1786 	if (ret) {
1787 		dev_err(dev, "failed to load pf res caps, ret = %d.\n", ret);
1788 		return ret;
1789 	}
1790 
1791 	ret = load_pf_timer_res_caps(hr_dev);
1792 	if (ret)
1793 		dev_err(dev, "failed to load pf timer resource, ret = %d.\n",
1794 			ret);
1795 
1796 	return ret;
1797 }
1798 
1799 static int hns_roce_query_vf_resource(struct hns_roce_dev *hr_dev)
1800 {
1801 	struct device *dev = hr_dev->dev;
1802 	int ret;
1803 
1804 	ret = load_func_res_caps(hr_dev, true);
1805 	if (ret)
1806 		dev_err(dev, "failed to load vf res caps, ret = %d.\n", ret);
1807 
1808 	return ret;
1809 }
1810 
1811 static int __hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
1812 					  u32 vf_id)
1813 {
1814 	struct hns_roce_vf_switch *swt;
1815 	struct hns_roce_cmq_desc desc;
1816 	int ret;
1817 
1818 	swt = (struct hns_roce_vf_switch *)desc.data;
1819 	hns_roce_cmq_setup_basic_desc(&desc, HNS_SWITCH_PARAMETER_CFG, true);
1820 	swt->rocee_sel |= cpu_to_le32(HNS_ICL_SWITCH_CMD_ROCEE_SEL);
1821 	hr_reg_write(swt, VF_SWITCH_VF_ID, vf_id);
1822 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1823 	if (ret)
1824 		return ret;
1825 
1826 	desc.flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1827 	desc.flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1828 	hr_reg_enable(swt, VF_SWITCH_ALW_LPBK);
1829 	hr_reg_clear(swt, VF_SWITCH_ALW_LCL_LPBK);
1830 	hr_reg_enable(swt, VF_SWITCH_ALW_DST_OVRD);
1831 
1832 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1833 }
1834 
1835 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev)
1836 {
1837 	u32 vf_id;
1838 	int ret;
1839 
1840 	for (vf_id = 0; vf_id < hr_dev->func_num; vf_id++) {
1841 		ret = __hns_roce_set_vf_switch_param(hr_dev, vf_id);
1842 		if (ret)
1843 			return ret;
1844 	}
1845 	return 0;
1846 }
1847 
1848 static int config_vf_hem_resource(struct hns_roce_dev *hr_dev, int vf_id)
1849 {
1850 	struct hns_roce_cmq_desc desc[2];
1851 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1852 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1853 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1854 	struct hns_roce_caps *caps = &hr_dev->caps;
1855 
1856 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1857 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1858 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1859 
1860 	hr_reg_write(r_a, FUNC_RES_A_VF_ID, vf_id);
1861 
1862 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_NUM, caps->qpc_bt_num);
1863 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_IDX, vf_id * caps->qpc_bt_num);
1864 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_NUM, caps->srqc_bt_num);
1865 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_IDX, vf_id * caps->srqc_bt_num);
1866 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_NUM, caps->cqc_bt_num);
1867 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_IDX, vf_id * caps->cqc_bt_num);
1868 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_NUM, caps->mpt_bt_num);
1869 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_IDX, vf_id * caps->mpt_bt_num);
1870 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_NUM, caps->eqc_bt_num);
1871 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_IDX, vf_id * caps->eqc_bt_num);
1872 	hr_reg_write(r_b, FUNC_RES_V_QID_NUM, caps->sl_num);
1873 	hr_reg_write(r_b, FUNC_RES_B_QID_IDX, vf_id * caps->sl_num);
1874 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_NUM, caps->sccc_bt_num);
1875 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_IDX, vf_id * caps->sccc_bt_num);
1876 
1877 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1878 		hr_reg_write(r_b, FUNC_RES_V_GMV_BT_NUM, caps->gmv_bt_num);
1879 		hr_reg_write(r_b, FUNC_RES_B_GMV_BT_IDX,
1880 			     vf_id * caps->gmv_bt_num);
1881 	} else {
1882 		hr_reg_write(r_b, FUNC_RES_B_SGID_NUM, caps->sgid_bt_num);
1883 		hr_reg_write(r_b, FUNC_RES_B_SGID_IDX,
1884 			     vf_id * caps->sgid_bt_num);
1885 		hr_reg_write(r_b, FUNC_RES_B_SMAC_NUM, caps->smac_bt_num);
1886 		hr_reg_write(r_b, FUNC_RES_B_SMAC_IDX,
1887 			     vf_id * caps->smac_bt_num);
1888 	}
1889 
1890 	return hns_roce_cmq_send(hr_dev, desc, 2);
1891 }
1892 
1893 static int hns_roce_alloc_vf_resource(struct hns_roce_dev *hr_dev)
1894 {
1895 	u32 func_num = max_t(u32, 1, hr_dev->func_num);
1896 	u32 vf_id;
1897 	int ret;
1898 
1899 	for (vf_id = 0; vf_id < func_num; vf_id++) {
1900 		ret = config_vf_hem_resource(hr_dev, vf_id);
1901 		if (ret) {
1902 			dev_err(hr_dev->dev,
1903 				"failed to config vf-%u hem res, ret = %d.\n",
1904 				vf_id, ret);
1905 			return ret;
1906 		}
1907 	}
1908 
1909 	return 0;
1910 }
1911 
1912 static int hns_roce_v2_set_bt(struct hns_roce_dev *hr_dev)
1913 {
1914 	struct hns_roce_cmq_desc desc;
1915 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1916 	struct hns_roce_caps *caps = &hr_dev->caps;
1917 
1918 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_BT_ATTR, false);
1919 
1920 	hr_reg_write(req, CFG_BT_ATTR_QPC_BA_PGSZ,
1921 		     caps->qpc_ba_pg_sz + PG_SHIFT_OFFSET);
1922 	hr_reg_write(req, CFG_BT_ATTR_QPC_BUF_PGSZ,
1923 		     caps->qpc_buf_pg_sz + PG_SHIFT_OFFSET);
1924 	hr_reg_write(req, CFG_BT_ATTR_QPC_HOPNUM,
1925 		     to_hr_hem_hopnum(caps->qpc_hop_num, caps->num_qps));
1926 
1927 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BA_PGSZ,
1928 		     caps->srqc_ba_pg_sz + PG_SHIFT_OFFSET);
1929 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BUF_PGSZ,
1930 		     caps->srqc_buf_pg_sz + PG_SHIFT_OFFSET);
1931 	hr_reg_write(req, CFG_BT_ATTR_SRQC_HOPNUM,
1932 		     to_hr_hem_hopnum(caps->srqc_hop_num, caps->num_srqs));
1933 
1934 	hr_reg_write(req, CFG_BT_ATTR_CQC_BA_PGSZ,
1935 		     caps->cqc_ba_pg_sz + PG_SHIFT_OFFSET);
1936 	hr_reg_write(req, CFG_BT_ATTR_CQC_BUF_PGSZ,
1937 		     caps->cqc_buf_pg_sz + PG_SHIFT_OFFSET);
1938 	hr_reg_write(req, CFG_BT_ATTR_CQC_HOPNUM,
1939 		     to_hr_hem_hopnum(caps->cqc_hop_num, caps->num_cqs));
1940 
1941 	hr_reg_write(req, CFG_BT_ATTR_MPT_BA_PGSZ,
1942 		     caps->mpt_ba_pg_sz + PG_SHIFT_OFFSET);
1943 	hr_reg_write(req, CFG_BT_ATTR_MPT_BUF_PGSZ,
1944 		     caps->mpt_buf_pg_sz + PG_SHIFT_OFFSET);
1945 	hr_reg_write(req, CFG_BT_ATTR_MPT_HOPNUM,
1946 		     to_hr_hem_hopnum(caps->mpt_hop_num, caps->num_mtpts));
1947 
1948 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BA_PGSZ,
1949 		     caps->sccc_ba_pg_sz + PG_SHIFT_OFFSET);
1950 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BUF_PGSZ,
1951 		     caps->sccc_buf_pg_sz + PG_SHIFT_OFFSET);
1952 	hr_reg_write(req, CFG_BT_ATTR_SCCC_HOPNUM,
1953 		     to_hr_hem_hopnum(caps->sccc_hop_num, caps->num_qps));
1954 
1955 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1956 }
1957 
1958 static void calc_pg_sz(u32 obj_num, u32 obj_size, u32 hop_num, u32 ctx_bt_num,
1959 		       u32 *buf_page_size, u32 *bt_page_size, u32 hem_type)
1960 {
1961 	u64 obj_per_chunk;
1962 	u64 bt_chunk_size = PAGE_SIZE;
1963 	u64 buf_chunk_size = PAGE_SIZE;
1964 	u64 obj_per_chunk_default = buf_chunk_size / obj_size;
1965 
1966 	*buf_page_size = 0;
1967 	*bt_page_size = 0;
1968 
1969 	switch (hop_num) {
1970 	case 3:
1971 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1972 				(bt_chunk_size / BA_BYTE_LEN) *
1973 				(bt_chunk_size / BA_BYTE_LEN) *
1974 				 obj_per_chunk_default;
1975 		break;
1976 	case 2:
1977 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1978 				(bt_chunk_size / BA_BYTE_LEN) *
1979 				 obj_per_chunk_default;
1980 		break;
1981 	case 1:
1982 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1983 				obj_per_chunk_default;
1984 		break;
1985 	case HNS_ROCE_HOP_NUM_0:
1986 		obj_per_chunk = ctx_bt_num * obj_per_chunk_default;
1987 		break;
1988 	default:
1989 		pr_err("table %u not support hop_num = %u!\n", hem_type,
1990 		       hop_num);
1991 		return;
1992 	}
1993 
1994 	if (hem_type >= HEM_TYPE_MTT)
1995 		*bt_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
1996 	else
1997 		*buf_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
1998 }
1999 
2000 static void set_hem_page_size(struct hns_roce_dev *hr_dev)
2001 {
2002 	struct hns_roce_caps *caps = &hr_dev->caps;
2003 
2004 	/* EQ */
2005 	caps->eqe_ba_pg_sz = 0;
2006 	caps->eqe_buf_pg_sz = 0;
2007 
2008 	/* Link Table */
2009 	caps->llm_buf_pg_sz = 0;
2010 
2011 	/* MR */
2012 	caps->mpt_ba_pg_sz = 0;
2013 	caps->mpt_buf_pg_sz = 0;
2014 	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
2015 	caps->pbl_buf_pg_sz = 0;
2016 	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
2017 		   caps->mpt_bt_num, &caps->mpt_buf_pg_sz, &caps->mpt_ba_pg_sz,
2018 		   HEM_TYPE_MTPT);
2019 
2020 	/* QP */
2021 	caps->qpc_ba_pg_sz = 0;
2022 	caps->qpc_buf_pg_sz = 0;
2023 	caps->qpc_timer_ba_pg_sz = 0;
2024 	caps->qpc_timer_buf_pg_sz = 0;
2025 	caps->sccc_ba_pg_sz = 0;
2026 	caps->sccc_buf_pg_sz = 0;
2027 	caps->mtt_ba_pg_sz = 0;
2028 	caps->mtt_buf_pg_sz = 0;
2029 	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
2030 		   caps->qpc_bt_num, &caps->qpc_buf_pg_sz, &caps->qpc_ba_pg_sz,
2031 		   HEM_TYPE_QPC);
2032 
2033 	if (caps->flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
2034 		calc_pg_sz(caps->num_qps, caps->sccc_sz, caps->sccc_hop_num,
2035 			   caps->sccc_bt_num, &caps->sccc_buf_pg_sz,
2036 			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
2037 
2038 	/* CQ */
2039 	caps->cqc_ba_pg_sz = 0;
2040 	caps->cqc_buf_pg_sz = 0;
2041 	caps->cqc_timer_ba_pg_sz = 0;
2042 	caps->cqc_timer_buf_pg_sz = 0;
2043 	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
2044 	caps->cqe_buf_pg_sz = 0;
2045 	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
2046 		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
2047 		   HEM_TYPE_CQC);
2048 	calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
2049 		   1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
2050 
2051 	/* SRQ */
2052 	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
2053 		caps->srqc_ba_pg_sz = 0;
2054 		caps->srqc_buf_pg_sz = 0;
2055 		caps->srqwqe_ba_pg_sz = 0;
2056 		caps->srqwqe_buf_pg_sz = 0;
2057 		caps->idx_ba_pg_sz = 0;
2058 		caps->idx_buf_pg_sz = 0;
2059 		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
2060 			   caps->srqc_hop_num, caps->srqc_bt_num,
2061 			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
2062 			   HEM_TYPE_SRQC);
2063 		calc_pg_sz(caps->num_srqwqe_segs, caps->mtt_entry_sz,
2064 			   caps->srqwqe_hop_num, 1, &caps->srqwqe_buf_pg_sz,
2065 			   &caps->srqwqe_ba_pg_sz, HEM_TYPE_SRQWQE);
2066 		calc_pg_sz(caps->num_idx_segs, caps->idx_entry_sz,
2067 			   caps->idx_hop_num, 1, &caps->idx_buf_pg_sz,
2068 			   &caps->idx_ba_pg_sz, HEM_TYPE_IDX);
2069 	}
2070 
2071 	/* GMV */
2072 	caps->gmv_ba_pg_sz = 0;
2073 	caps->gmv_buf_pg_sz = 0;
2074 }
2075 
2076 /* Apply all loaded caps before setting to hardware */
2077 static void apply_func_caps(struct hns_roce_dev *hr_dev)
2078 {
2079 #define MAX_GID_TBL_LEN 256
2080 	struct hns_roce_caps *caps = &hr_dev->caps;
2081 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2082 
2083 	/* The following configurations don't need to be got from firmware. */
2084 	caps->qpc_timer_entry_sz = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
2085 	caps->cqc_timer_entry_sz = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
2086 	caps->mtt_entry_sz = HNS_ROCE_V2_MTT_ENTRY_SZ;
2087 
2088 	caps->pbl_hop_num = HNS_ROCE_PBL_HOP_NUM;
2089 	caps->qpc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2090 	caps->cqc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2091 
2092 	caps->num_srqwqe_segs = HNS_ROCE_V2_MAX_SRQWQE_SEGS;
2093 	caps->num_idx_segs = HNS_ROCE_V2_MAX_IDX_SEGS;
2094 
2095 	if (!caps->num_comp_vectors)
2096 		caps->num_comp_vectors =
2097 			min_t(u32, caps->eqc_bt_num - HNS_ROCE_V2_AEQE_VEC_NUM,
2098 				(u32)priv->handle->rinfo.num_vectors -
2099 		(HNS_ROCE_V2_AEQE_VEC_NUM + HNS_ROCE_V2_ABNORMAL_VEC_NUM));
2100 
2101 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2102 		caps->eqe_hop_num = HNS_ROCE_V3_EQE_HOP_NUM;
2103 		caps->ceqe_size = HNS_ROCE_V3_EQE_SIZE;
2104 		caps->aeqe_size = HNS_ROCE_V3_EQE_SIZE;
2105 
2106 		/* The following configurations will be overwritten */
2107 		caps->qpc_sz = HNS_ROCE_V3_QPC_SZ;
2108 		caps->cqe_sz = HNS_ROCE_V3_CQE_SIZE;
2109 		caps->sccc_sz = HNS_ROCE_V3_SCCC_SZ;
2110 
2111 		/* The following configurations are not got from firmware */
2112 		caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;
2113 
2114 		caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0;
2115 
2116 		/* It's meaningless to support excessively large gid_table_len,
2117 		 * as the type of sgid_index in kernel struct ib_global_route
2118 		 * and userspace struct ibv_global_route are u8/uint8_t (0-255).
2119 		 */
2120 		caps->gid_table_len[0] = min_t(u32, MAX_GID_TBL_LEN,
2121 					 caps->gmv_bt_num *
2122 					 (HNS_HW_PAGE_SIZE / caps->gmv_entry_sz));
2123 
2124 		caps->gmv_entry_num = caps->gmv_bt_num * (HNS_HW_PAGE_SIZE /
2125 							  caps->gmv_entry_sz);
2126 	} else {
2127 		u32 func_num = max_t(u32, 1, hr_dev->func_num);
2128 
2129 		caps->eqe_hop_num = HNS_ROCE_V2_EQE_HOP_NUM;
2130 		caps->ceqe_size = HNS_ROCE_CEQE_SIZE;
2131 		caps->aeqe_size = HNS_ROCE_AEQE_SIZE;
2132 		caps->gid_table_len[0] /= func_num;
2133 	}
2134 
2135 	if (hr_dev->is_vf) {
2136 		caps->default_aeq_arm_st = 0x3;
2137 		caps->default_ceq_arm_st = 0x3;
2138 		caps->default_ceq_max_cnt = 0x1;
2139 		caps->default_ceq_period = 0x10;
2140 		caps->default_aeq_max_cnt = 0x1;
2141 		caps->default_aeq_period = 0x10;
2142 	}
2143 
2144 	set_hem_page_size(hr_dev);
2145 }
2146 
2147 static int hns_roce_query_caps(struct hns_roce_dev *hr_dev)
2148 {
2149 	struct hns_roce_cmq_desc desc[HNS_ROCE_QUERY_PF_CAPS_CMD_NUM];
2150 	struct hns_roce_caps *caps = &hr_dev->caps;
2151 	struct hns_roce_query_pf_caps_a *resp_a;
2152 	struct hns_roce_query_pf_caps_b *resp_b;
2153 	struct hns_roce_query_pf_caps_c *resp_c;
2154 	struct hns_roce_query_pf_caps_d *resp_d;
2155 	struct hns_roce_query_pf_caps_e *resp_e;
2156 	enum hns_roce_opcode_type cmd;
2157 	int ctx_hop_num;
2158 	int pbl_hop_num;
2159 	int ret;
2160 	int i;
2161 
2162 	cmd = hr_dev->is_vf ? HNS_ROCE_OPC_QUERY_VF_CAPS_NUM :
2163 	      HNS_ROCE_OPC_QUERY_PF_CAPS_NUM;
2164 
2165 	for (i = 0; i < HNS_ROCE_QUERY_PF_CAPS_CMD_NUM; i++) {
2166 		hns_roce_cmq_setup_basic_desc(&desc[i], cmd, true);
2167 		if (i < (HNS_ROCE_QUERY_PF_CAPS_CMD_NUM - 1))
2168 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2169 		else
2170 			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2171 	}
2172 
2173 	ret = hns_roce_cmq_send(hr_dev, desc, HNS_ROCE_QUERY_PF_CAPS_CMD_NUM);
2174 	if (ret)
2175 		return ret;
2176 
2177 	resp_a = (struct hns_roce_query_pf_caps_a *)desc[0].data;
2178 	resp_b = (struct hns_roce_query_pf_caps_b *)desc[1].data;
2179 	resp_c = (struct hns_roce_query_pf_caps_c *)desc[2].data;
2180 	resp_d = (struct hns_roce_query_pf_caps_d *)desc[3].data;
2181 	resp_e = (struct hns_roce_query_pf_caps_e *)desc[4].data;
2182 
2183 	caps->local_ca_ack_delay = resp_a->local_ca_ack_delay;
2184 	caps->max_sq_sg = le16_to_cpu(resp_a->max_sq_sg);
2185 	caps->max_sq_inline = le16_to_cpu(resp_a->max_sq_inline);
2186 	caps->max_rq_sg = le16_to_cpu(resp_a->max_rq_sg);
2187 	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
2188 	caps->max_srq_sges = le16_to_cpu(resp_a->max_srq_sges);
2189 	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
2190 	caps->num_aeq_vectors = resp_a->num_aeq_vectors;
2191 	caps->num_other_vectors = resp_a->num_other_vectors;
2192 	caps->max_sq_desc_sz = resp_a->max_sq_desc_sz;
2193 	caps->max_rq_desc_sz = resp_a->max_rq_desc_sz;
2194 
2195 	caps->mtpt_entry_sz = resp_b->mtpt_entry_sz;
2196 	caps->irrl_entry_sz = resp_b->irrl_entry_sz;
2197 	caps->trrl_entry_sz = resp_b->trrl_entry_sz;
2198 	caps->cqc_entry_sz = resp_b->cqc_entry_sz;
2199 	caps->srqc_entry_sz = resp_b->srqc_entry_sz;
2200 	caps->idx_entry_sz = resp_b->idx_entry_sz;
2201 	caps->sccc_sz = resp_b->sccc_sz;
2202 	caps->max_mtu = resp_b->max_mtu;
2203 	caps->min_cqes = resp_b->min_cqes;
2204 	caps->min_wqes = resp_b->min_wqes;
2205 	caps->page_size_cap = le32_to_cpu(resp_b->page_size_cap);
2206 	caps->pkey_table_len[0] = resp_b->pkey_table_len;
2207 	caps->phy_num_uars = resp_b->phy_num_uars;
2208 	ctx_hop_num = resp_b->ctx_hop_num;
2209 	pbl_hop_num = resp_b->pbl_hop_num;
2210 
2211 	caps->num_pds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_PDS);
2212 
2213 	caps->flags = hr_reg_read(resp_c, PF_CAPS_C_CAP_FLAGS);
2214 	caps->flags |= le16_to_cpu(resp_d->cap_flags_ex) <<
2215 		       HNS_ROCE_CAP_FLAGS_EX_SHIFT;
2216 
2217 	caps->num_cqs = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_CQS);
2218 	caps->gid_table_len[0] = hr_reg_read(resp_c, PF_CAPS_C_MAX_GID);
2219 	caps->max_cqes = 1 << hr_reg_read(resp_c, PF_CAPS_C_CQ_DEPTH);
2220 	caps->num_xrcds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_XRCDS);
2221 	caps->num_mtpts = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_MRWS);
2222 	caps->num_qps = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_QPS);
2223 	caps->max_qp_init_rdma = hr_reg_read(resp_c, PF_CAPS_C_MAX_ORD);
2224 	caps->max_qp_dest_rdma = caps->max_qp_init_rdma;
2225 	caps->max_wqes = 1 << le16_to_cpu(resp_c->sq_depth);
2226 
2227 	caps->num_srqs = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_SRQS);
2228 	caps->cong_cap = hr_reg_read(resp_d, PF_CAPS_D_CONG_CAP);
2229 	caps->max_srq_wrs = 1 << le16_to_cpu(resp_d->srq_depth);
2230 	caps->ceqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_CEQ_DEPTH);
2231 	caps->num_comp_vectors = hr_reg_read(resp_d, PF_CAPS_D_NUM_CEQS);
2232 	caps->aeqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_AEQ_DEPTH);
2233 	caps->default_cong_type = hr_reg_read(resp_d, PF_CAPS_D_DEFAULT_ALG);
2234 	caps->reserved_pds = hr_reg_read(resp_d, PF_CAPS_D_RSV_PDS);
2235 	caps->num_uars = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_UARS);
2236 	caps->reserved_qps = hr_reg_read(resp_d, PF_CAPS_D_RSV_QPS);
2237 	caps->reserved_uars = hr_reg_read(resp_d, PF_CAPS_D_RSV_UARS);
2238 
2239 	caps->reserved_mrws = hr_reg_read(resp_e, PF_CAPS_E_RSV_MRWS);
2240 	caps->chunk_sz = 1 << hr_reg_read(resp_e, PF_CAPS_E_CHUNK_SIZE_SHIFT);
2241 	caps->reserved_cqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_CQS);
2242 	caps->reserved_xrcds = hr_reg_read(resp_e, PF_CAPS_E_RSV_XRCDS);
2243 	caps->reserved_srqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_SRQS);
2244 	caps->reserved_lkey = hr_reg_read(resp_e, PF_CAPS_E_RSV_LKEYS);
2245 
2246 	caps->qpc_hop_num = ctx_hop_num;
2247 	caps->sccc_hop_num = ctx_hop_num;
2248 	caps->srqc_hop_num = ctx_hop_num;
2249 	caps->cqc_hop_num = ctx_hop_num;
2250 	caps->mpt_hop_num = ctx_hop_num;
2251 	caps->mtt_hop_num = pbl_hop_num;
2252 	caps->cqe_hop_num = pbl_hop_num;
2253 	caps->srqwqe_hop_num = pbl_hop_num;
2254 	caps->idx_hop_num = pbl_hop_num;
2255 	caps->wqe_sq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_SQWQE_HOP_NUM);
2256 	caps->wqe_sge_hop_num = hr_reg_read(resp_d, PF_CAPS_D_EX_SGE_HOP_NUM);
2257 	caps->wqe_rq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_RQWQE_HOP_NUM);
2258 
2259 	if (!(caps->page_size_cap & PAGE_SIZE))
2260 		caps->page_size_cap = HNS_ROCE_V2_PAGE_SIZE_SUPPORTED;
2261 
2262 	if (!hr_dev->is_vf) {
2263 		caps->cqe_sz = resp_a->cqe_sz;
2264 		caps->qpc_sz = le16_to_cpu(resp_b->qpc_sz);
2265 		caps->default_aeq_arm_st =
2266 				hr_reg_read(resp_d, PF_CAPS_D_AEQ_ARM_ST);
2267 		caps->default_ceq_arm_st =
2268 				hr_reg_read(resp_d, PF_CAPS_D_CEQ_ARM_ST);
2269 		caps->default_ceq_max_cnt = le16_to_cpu(resp_e->ceq_max_cnt);
2270 		caps->default_ceq_period = le16_to_cpu(resp_e->ceq_period);
2271 		caps->default_aeq_max_cnt = le16_to_cpu(resp_e->aeq_max_cnt);
2272 		caps->default_aeq_period = le16_to_cpu(resp_e->aeq_period);
2273 	}
2274 
2275 	return 0;
2276 }
2277 
2278 static int config_hem_entry_size(struct hns_roce_dev *hr_dev, u32 type, u32 val)
2279 {
2280 	struct hns_roce_cmq_desc desc;
2281 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
2282 
2283 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_ENTRY_SIZE,
2284 				      false);
2285 
2286 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_TYPE, type);
2287 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_VALUE, val);
2288 
2289 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2290 }
2291 
2292 static int hns_roce_config_entry_size(struct hns_roce_dev *hr_dev)
2293 {
2294 	struct hns_roce_caps *caps = &hr_dev->caps;
2295 	int ret;
2296 
2297 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2298 		return 0;
2299 
2300 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_QPC_SIZE,
2301 				    caps->qpc_sz);
2302 	if (ret) {
2303 		dev_err(hr_dev->dev, "failed to cfg qpc sz, ret = %d.\n", ret);
2304 		return ret;
2305 	}
2306 
2307 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_SCCC_SIZE,
2308 				    caps->sccc_sz);
2309 	if (ret)
2310 		dev_err(hr_dev->dev, "failed to cfg sccc sz, ret = %d.\n", ret);
2311 
2312 	return ret;
2313 }
2314 
2315 static int hns_roce_v2_vf_profile(struct hns_roce_dev *hr_dev)
2316 {
2317 	struct device *dev = hr_dev->dev;
2318 	int ret;
2319 
2320 	hr_dev->func_num = 1;
2321 
2322 	ret = hns_roce_query_caps(hr_dev);
2323 	if (ret) {
2324 		dev_err(dev, "failed to query VF caps, ret = %d.\n", ret);
2325 		return ret;
2326 	}
2327 
2328 	ret = hns_roce_query_vf_resource(hr_dev);
2329 	if (ret) {
2330 		dev_err(dev, "failed to query VF resource, ret = %d.\n", ret);
2331 		return ret;
2332 	}
2333 
2334 	apply_func_caps(hr_dev);
2335 
2336 	ret = hns_roce_v2_set_bt(hr_dev);
2337 	if (ret)
2338 		dev_err(dev, "failed to config VF BA table, ret = %d.\n", ret);
2339 
2340 	return ret;
2341 }
2342 
2343 static int hns_roce_v2_pf_profile(struct hns_roce_dev *hr_dev)
2344 {
2345 	struct device *dev = hr_dev->dev;
2346 	int ret;
2347 
2348 	ret = hns_roce_query_func_info(hr_dev);
2349 	if (ret) {
2350 		dev_err(dev, "failed to query func info, ret = %d.\n", ret);
2351 		return ret;
2352 	}
2353 
2354 	ret = hns_roce_config_global_param(hr_dev);
2355 	if (ret) {
2356 		dev_err(dev, "failed to config global param, ret = %d.\n", ret);
2357 		return ret;
2358 	}
2359 
2360 	ret = hns_roce_set_vf_switch_param(hr_dev);
2361 	if (ret) {
2362 		dev_err(dev, "failed to set switch param, ret = %d.\n", ret);
2363 		return ret;
2364 	}
2365 
2366 	ret = hns_roce_query_caps(hr_dev);
2367 	if (ret) {
2368 		dev_err(dev, "failed to query PF caps, ret = %d.\n", ret);
2369 		return ret;
2370 	}
2371 
2372 	ret = hns_roce_query_pf_resource(hr_dev);
2373 	if (ret) {
2374 		dev_err(dev, "failed to query pf resource, ret = %d.\n", ret);
2375 		return ret;
2376 	}
2377 
2378 	apply_func_caps(hr_dev);
2379 
2380 	ret = hns_roce_alloc_vf_resource(hr_dev);
2381 	if (ret) {
2382 		dev_err(dev, "failed to alloc vf resource, ret = %d.\n", ret);
2383 		return ret;
2384 	}
2385 
2386 	ret = hns_roce_v2_set_bt(hr_dev);
2387 	if (ret) {
2388 		dev_err(dev, "failed to config BA table, ret = %d.\n", ret);
2389 		return ret;
2390 	}
2391 
2392 	/* Configure the size of QPC, SCCC, etc. */
2393 	return hns_roce_config_entry_size(hr_dev);
2394 }
2395 
2396 static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
2397 {
2398 	struct device *dev = hr_dev->dev;
2399 	int ret;
2400 
2401 	ret = hns_roce_cmq_query_hw_info(hr_dev);
2402 	if (ret) {
2403 		dev_err(dev, "failed to query hardware info, ret = %d.\n", ret);
2404 		return ret;
2405 	}
2406 
2407 	ret = hns_roce_query_fw_ver(hr_dev);
2408 	if (ret) {
2409 		dev_err(dev, "failed to query firmware info, ret = %d.\n", ret);
2410 		return ret;
2411 	}
2412 
2413 	hr_dev->vendor_part_id = hr_dev->pci_dev->device;
2414 	hr_dev->sys_image_guid = be64_to_cpu(hr_dev->ib_dev.node_guid);
2415 
2416 	if (hr_dev->is_vf)
2417 		return hns_roce_v2_vf_profile(hr_dev);
2418 	else
2419 		return hns_roce_v2_pf_profile(hr_dev);
2420 }
2421 
2422 static void config_llm_table(struct hns_roce_buf *data_buf, void *cfg_buf)
2423 {
2424 	u32 i, next_ptr, page_num;
2425 	__le64 *entry = cfg_buf;
2426 	dma_addr_t addr;
2427 	u64 val;
2428 
2429 	page_num = data_buf->npages;
2430 	for (i = 0; i < page_num; i++) {
2431 		addr = hns_roce_buf_page(data_buf, i);
2432 		if (i == (page_num - 1))
2433 			next_ptr = 0;
2434 		else
2435 			next_ptr = i + 1;
2436 
2437 		val = HNS_ROCE_EXT_LLM_ENTRY(addr, (u64)next_ptr);
2438 		entry[i] = cpu_to_le64(val);
2439 	}
2440 }
2441 
2442 static int set_llm_cfg_to_hw(struct hns_roce_dev *hr_dev,
2443 			     struct hns_roce_link_table *table)
2444 {
2445 	struct hns_roce_cmq_desc desc[2];
2446 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
2447 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
2448 	struct hns_roce_buf *buf = table->buf;
2449 	enum hns_roce_opcode_type opcode;
2450 	dma_addr_t addr;
2451 
2452 	opcode = HNS_ROCE_OPC_CFG_EXT_LLM;
2453 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
2454 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2455 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
2456 
2457 	hr_reg_write(r_a, CFG_LLM_A_BA_L, lower_32_bits(table->table.map));
2458 	hr_reg_write(r_a, CFG_LLM_A_BA_H, upper_32_bits(table->table.map));
2459 	hr_reg_write(r_a, CFG_LLM_A_DEPTH, buf->npages);
2460 	hr_reg_write(r_a, CFG_LLM_A_PGSZ, to_hr_hw_page_shift(buf->page_shift));
2461 	hr_reg_enable(r_a, CFG_LLM_A_INIT_EN);
2462 
2463 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, 0));
2464 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_L, lower_32_bits(addr));
2465 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_H, upper_32_bits(addr));
2466 	hr_reg_write(r_a, CFG_LLM_A_HEAD_NXTPTR, 1);
2467 	hr_reg_write(r_a, CFG_LLM_A_HEAD_PTR, 0);
2468 
2469 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, buf->npages - 1));
2470 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_L, lower_32_bits(addr));
2471 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_H, upper_32_bits(addr));
2472 	hr_reg_write(r_b, CFG_LLM_B_TAIL_PTR, buf->npages - 1);
2473 
2474 	return hns_roce_cmq_send(hr_dev, desc, 2);
2475 }
2476 
2477 static struct hns_roce_link_table *
2478 alloc_link_table_buf(struct hns_roce_dev *hr_dev)
2479 {
2480 	u16 total_sl = hr_dev->caps.sl_num * hr_dev->func_num;
2481 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2482 	struct hns_roce_link_table *link_tbl;
2483 	u32 pg_shift, size, min_size;
2484 
2485 	link_tbl = &priv->ext_llm;
2486 	pg_shift = hr_dev->caps.llm_buf_pg_sz + PAGE_SHIFT;
2487 	size = hr_dev->caps.num_qps * hr_dev->func_num *
2488 	       HNS_ROCE_V2_EXT_LLM_ENTRY_SZ;
2489 	min_size = HNS_ROCE_EXT_LLM_MIN_PAGES(total_sl) << pg_shift;
2490 
2491 	/* Alloc data table */
2492 	size = max(size, min_size);
2493 	link_tbl->buf = hns_roce_buf_alloc(hr_dev, size, pg_shift, 0);
2494 	if (IS_ERR(link_tbl->buf))
2495 		return ERR_PTR(-ENOMEM);
2496 
2497 	/* Alloc config table */
2498 	size = link_tbl->buf->npages * sizeof(u64);
2499 	link_tbl->table.buf = dma_alloc_coherent(hr_dev->dev, size,
2500 						 &link_tbl->table.map,
2501 						 GFP_KERNEL);
2502 	if (!link_tbl->table.buf) {
2503 		hns_roce_buf_free(hr_dev, link_tbl->buf);
2504 		return ERR_PTR(-ENOMEM);
2505 	}
2506 
2507 	return link_tbl;
2508 }
2509 
2510 static void free_link_table_buf(struct hns_roce_dev *hr_dev,
2511 				struct hns_roce_link_table *tbl)
2512 {
2513 	if (tbl->buf) {
2514 		u32 size = tbl->buf->npages * sizeof(u64);
2515 
2516 		dma_free_coherent(hr_dev->dev, size, tbl->table.buf,
2517 				  tbl->table.map);
2518 	}
2519 
2520 	hns_roce_buf_free(hr_dev, tbl->buf);
2521 }
2522 
2523 static int hns_roce_init_link_table(struct hns_roce_dev *hr_dev)
2524 {
2525 	struct hns_roce_link_table *link_tbl;
2526 	int ret;
2527 
2528 	link_tbl = alloc_link_table_buf(hr_dev);
2529 	if (IS_ERR(link_tbl))
2530 		return -ENOMEM;
2531 
2532 	if (WARN_ON(link_tbl->buf->npages > HNS_ROCE_V2_EXT_LLM_MAX_DEPTH)) {
2533 		ret = -EINVAL;
2534 		goto err_alloc;
2535 	}
2536 
2537 	config_llm_table(link_tbl->buf, link_tbl->table.buf);
2538 	ret = set_llm_cfg_to_hw(hr_dev, link_tbl);
2539 	if (ret)
2540 		goto err_alloc;
2541 
2542 	return 0;
2543 
2544 err_alloc:
2545 	free_link_table_buf(hr_dev, link_tbl);
2546 	return ret;
2547 }
2548 
2549 static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev)
2550 {
2551 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2552 
2553 	free_link_table_buf(hr_dev, &priv->ext_llm);
2554 }
2555 
2556 static void free_dip_entry(struct hns_roce_dev *hr_dev)
2557 {
2558 	struct hns_roce_dip *hr_dip;
2559 	unsigned long idx;
2560 
2561 	xa_lock(&hr_dev->qp_table.dip_xa);
2562 
2563 	xa_for_each(&hr_dev->qp_table.dip_xa, idx, hr_dip) {
2564 		__xa_erase(&hr_dev->qp_table.dip_xa, hr_dip->dip_idx);
2565 		kfree(hr_dip);
2566 	}
2567 
2568 	xa_unlock(&hr_dev->qp_table.dip_xa);
2569 }
2570 
2571 static struct ib_pd *free_mr_init_pd(struct hns_roce_dev *hr_dev)
2572 {
2573 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2574 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2575 	struct ib_device *ibdev = &hr_dev->ib_dev;
2576 	struct hns_roce_pd *hr_pd;
2577 	struct ib_pd *pd;
2578 
2579 	hr_pd = kzalloc(sizeof(*hr_pd), GFP_KERNEL);
2580 	if (ZERO_OR_NULL_PTR(hr_pd))
2581 		return NULL;
2582 	pd = &hr_pd->ibpd;
2583 	pd->device = ibdev;
2584 
2585 	if (hns_roce_alloc_pd(pd, NULL)) {
2586 		ibdev_err(ibdev, "failed to create pd for free mr.\n");
2587 		kfree(hr_pd);
2588 		return NULL;
2589 	}
2590 	free_mr->rsv_pd = to_hr_pd(pd);
2591 	free_mr->rsv_pd->ibpd.device = &hr_dev->ib_dev;
2592 	free_mr->rsv_pd->ibpd.uobject = NULL;
2593 	free_mr->rsv_pd->ibpd.__internal_mr = NULL;
2594 	atomic_set(&free_mr->rsv_pd->ibpd.usecnt, 0);
2595 
2596 	return pd;
2597 }
2598 
2599 static struct ib_cq *free_mr_init_cq(struct hns_roce_dev *hr_dev)
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 ib_cq_init_attr cq_init_attr = {};
2605 	struct hns_roce_cq *hr_cq;
2606 	struct ib_cq *cq;
2607 
2608 	cq_init_attr.cqe = HNS_ROCE_FREE_MR_USED_CQE_NUM;
2609 
2610 	hr_cq = kzalloc(sizeof(*hr_cq), GFP_KERNEL);
2611 	if (ZERO_OR_NULL_PTR(hr_cq))
2612 		return NULL;
2613 
2614 	cq = &hr_cq->ib_cq;
2615 	cq->device = ibdev;
2616 
2617 	if (hns_roce_create_cq(cq, &cq_init_attr, NULL)) {
2618 		ibdev_err(ibdev, "failed to create cq for free mr.\n");
2619 		kfree(hr_cq);
2620 		return NULL;
2621 	}
2622 	free_mr->rsv_cq = to_hr_cq(cq);
2623 	free_mr->rsv_cq->ib_cq.device = &hr_dev->ib_dev;
2624 	free_mr->rsv_cq->ib_cq.uobject = NULL;
2625 	free_mr->rsv_cq->ib_cq.comp_handler = NULL;
2626 	free_mr->rsv_cq->ib_cq.event_handler = NULL;
2627 	free_mr->rsv_cq->ib_cq.cq_context = NULL;
2628 	atomic_set(&free_mr->rsv_cq->ib_cq.usecnt, 0);
2629 
2630 	return cq;
2631 }
2632 
2633 static int free_mr_init_qp(struct hns_roce_dev *hr_dev, struct ib_cq *cq,
2634 			   struct ib_qp_init_attr *init_attr, int i)
2635 {
2636 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2637 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2638 	struct ib_device *ibdev = &hr_dev->ib_dev;
2639 	struct hns_roce_qp *hr_qp;
2640 	struct ib_qp *qp;
2641 	int ret;
2642 
2643 	hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
2644 	if (ZERO_OR_NULL_PTR(hr_qp))
2645 		return -ENOMEM;
2646 
2647 	qp = &hr_qp->ibqp;
2648 	qp->device = ibdev;
2649 
2650 	ret = hns_roce_create_qp(qp, init_attr, NULL);
2651 	if (ret) {
2652 		ibdev_err(ibdev, "failed to create qp for free mr.\n");
2653 		kfree(hr_qp);
2654 		return ret;
2655 	}
2656 
2657 	free_mr->rsv_qp[i] = hr_qp;
2658 	free_mr->rsv_qp[i]->ibqp.recv_cq = cq;
2659 	free_mr->rsv_qp[i]->ibqp.send_cq = cq;
2660 
2661 	return 0;
2662 }
2663 
2664 static void free_mr_exit(struct hns_roce_dev *hr_dev)
2665 {
2666 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2667 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2668 	struct ib_qp *qp;
2669 	int i;
2670 
2671 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2672 		if (free_mr->rsv_qp[i]) {
2673 			qp = &free_mr->rsv_qp[i]->ibqp;
2674 			hns_roce_v2_destroy_qp(qp, NULL);
2675 			kfree(free_mr->rsv_qp[i]);
2676 			free_mr->rsv_qp[i] = NULL;
2677 		}
2678 	}
2679 
2680 	if (free_mr->rsv_cq) {
2681 		hns_roce_destroy_cq(&free_mr->rsv_cq->ib_cq, NULL);
2682 		kfree(free_mr->rsv_cq);
2683 		free_mr->rsv_cq = NULL;
2684 	}
2685 
2686 	if (free_mr->rsv_pd) {
2687 		hns_roce_dealloc_pd(&free_mr->rsv_pd->ibpd, NULL);
2688 		kfree(free_mr->rsv_pd);
2689 		free_mr->rsv_pd = NULL;
2690 	}
2691 
2692 	mutex_destroy(&free_mr->mutex);
2693 }
2694 
2695 static int free_mr_alloc_res(struct hns_roce_dev *hr_dev)
2696 {
2697 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2698 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2699 	struct ib_qp_init_attr qp_init_attr = {};
2700 	struct ib_pd *pd;
2701 	struct ib_cq *cq;
2702 	int ret;
2703 	int i;
2704 
2705 	pd = free_mr_init_pd(hr_dev);
2706 	if (!pd)
2707 		return -ENOMEM;
2708 
2709 	cq = free_mr_init_cq(hr_dev);
2710 	if (!cq) {
2711 		ret = -ENOMEM;
2712 		goto create_failed_cq;
2713 	}
2714 
2715 	qp_init_attr.qp_type = IB_QPT_RC;
2716 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2717 	qp_init_attr.send_cq = cq;
2718 	qp_init_attr.recv_cq = cq;
2719 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2720 		qp_init_attr.cap.max_send_wr = HNS_ROCE_FREE_MR_USED_SQWQE_NUM;
2721 		qp_init_attr.cap.max_send_sge = HNS_ROCE_FREE_MR_USED_SQSGE_NUM;
2722 		qp_init_attr.cap.max_recv_wr = HNS_ROCE_FREE_MR_USED_RQWQE_NUM;
2723 		qp_init_attr.cap.max_recv_sge = HNS_ROCE_FREE_MR_USED_RQSGE_NUM;
2724 
2725 		ret = free_mr_init_qp(hr_dev, cq, &qp_init_attr, i);
2726 		if (ret)
2727 			goto create_failed_qp;
2728 	}
2729 
2730 	return 0;
2731 
2732 create_failed_qp:
2733 	for (i--; i >= 0; i--) {
2734 		hns_roce_v2_destroy_qp(&free_mr->rsv_qp[i]->ibqp, NULL);
2735 		kfree(free_mr->rsv_qp[i]);
2736 	}
2737 	hns_roce_destroy_cq(cq, NULL);
2738 	kfree(cq);
2739 
2740 create_failed_cq:
2741 	hns_roce_dealloc_pd(pd, NULL);
2742 	kfree(pd);
2743 
2744 	return ret;
2745 }
2746 
2747 static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
2748 				 struct ib_qp_attr *attr, int sl_num)
2749 {
2750 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2751 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2752 	struct ib_device *ibdev = &hr_dev->ib_dev;
2753 	struct hns_roce_qp *hr_qp;
2754 	int loopback;
2755 	int mask;
2756 	int ret;
2757 
2758 	hr_qp = to_hr_qp(&free_mr->rsv_qp[sl_num]->ibqp);
2759 	hr_qp->free_mr_en = 1;
2760 	hr_qp->ibqp.device = ibdev;
2761 	hr_qp->ibqp.qp_type = IB_QPT_RC;
2762 
2763 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_ACCESS_FLAGS;
2764 	attr->qp_state = IB_QPS_INIT;
2765 	attr->port_num = 1;
2766 	attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
2767 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2768 				    IB_QPS_INIT, NULL);
2769 	if (ret) {
2770 		ibdev_err_ratelimited(ibdev, "failed to modify qp to init, ret = %d.\n",
2771 				      ret);
2772 		return ret;
2773 	}
2774 
2775 	loopback = hr_dev->loop_idc;
2776 	/* Set qpc lbi = 1 incidate loopback IO */
2777 	hr_dev->loop_idc = 1;
2778 
2779 	mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
2780 	       IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
2781 	attr->qp_state = IB_QPS_RTR;
2782 	attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2783 	attr->path_mtu = IB_MTU_256;
2784 	attr->dest_qp_num = hr_qp->qpn;
2785 	attr->rq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2786 
2787 	rdma_ah_set_sl(&attr->ah_attr, (u8)sl_num);
2788 
2789 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2790 				    IB_QPS_RTR, NULL);
2791 	hr_dev->loop_idc = loopback;
2792 	if (ret) {
2793 		ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
2794 			  ret);
2795 		return ret;
2796 	}
2797 
2798 	mask = IB_QP_STATE | IB_QP_SQ_PSN | IB_QP_RETRY_CNT | IB_QP_TIMEOUT |
2799 	       IB_QP_RNR_RETRY | IB_QP_MAX_QP_RD_ATOMIC;
2800 	attr->qp_state = IB_QPS_RTS;
2801 	attr->sq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2802 	attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
2803 	attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
2804 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_RTR,
2805 				    IB_QPS_RTS, NULL);
2806 	if (ret)
2807 		ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
2808 			  ret);
2809 
2810 	return ret;
2811 }
2812 
2813 static int free_mr_modify_qp(struct hns_roce_dev *hr_dev)
2814 {
2815 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2816 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2817 	struct ib_qp_attr attr = {};
2818 	int ret;
2819 	int i;
2820 
2821 	rdma_ah_set_grh(&attr.ah_attr, NULL, 0, 0, 1, 0);
2822 	rdma_ah_set_static_rate(&attr.ah_attr, 3);
2823 	rdma_ah_set_port_num(&attr.ah_attr, 1);
2824 
2825 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2826 		ret = free_mr_modify_rsv_qp(hr_dev, &attr, i);
2827 		if (ret)
2828 			return ret;
2829 	}
2830 
2831 	return 0;
2832 }
2833 
2834 static int free_mr_init(struct hns_roce_dev *hr_dev)
2835 {
2836 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2837 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2838 	int ret;
2839 
2840 	mutex_init(&free_mr->mutex);
2841 
2842 	ret = free_mr_alloc_res(hr_dev);
2843 	if (ret) {
2844 		mutex_destroy(&free_mr->mutex);
2845 		return ret;
2846 	}
2847 
2848 	ret = free_mr_modify_qp(hr_dev);
2849 	if (ret)
2850 		goto err_modify_qp;
2851 
2852 	return 0;
2853 
2854 err_modify_qp:
2855 	free_mr_exit(hr_dev);
2856 
2857 	return ret;
2858 }
2859 
2860 static int get_hem_table(struct hns_roce_dev *hr_dev)
2861 {
2862 	unsigned int qpc_count;
2863 	unsigned int cqc_count;
2864 	unsigned int gmv_count;
2865 	int ret;
2866 	int i;
2867 
2868 	/* Alloc memory for source address table buffer space chunk */
2869 	for (gmv_count = 0; gmv_count < hr_dev->caps.gmv_entry_num;
2870 	     gmv_count++) {
2871 		ret = hns_roce_table_get(hr_dev, &hr_dev->gmv_table, gmv_count);
2872 		if (ret)
2873 			goto err_gmv_failed;
2874 	}
2875 
2876 	if (hr_dev->is_vf)
2877 		return 0;
2878 
2879 	/* Alloc memory for QPC Timer buffer space chunk */
2880 	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
2881 	     qpc_count++) {
2882 		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
2883 					 qpc_count);
2884 		if (ret) {
2885 			dev_err(hr_dev->dev, "QPC Timer get failed\n");
2886 			goto err_qpc_timer_failed;
2887 		}
2888 	}
2889 
2890 	/* Alloc memory for CQC Timer buffer space chunk */
2891 	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
2892 	     cqc_count++) {
2893 		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
2894 					 cqc_count);
2895 		if (ret) {
2896 			dev_err(hr_dev->dev, "CQC Timer get failed\n");
2897 			goto err_cqc_timer_failed;
2898 		}
2899 	}
2900 
2901 	return 0;
2902 
2903 err_cqc_timer_failed:
2904 	for (i = 0; i < cqc_count; i++)
2905 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2906 
2907 err_qpc_timer_failed:
2908 	for (i = 0; i < qpc_count; i++)
2909 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2910 
2911 err_gmv_failed:
2912 	for (i = 0; i < gmv_count; i++)
2913 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2914 
2915 	return ret;
2916 }
2917 
2918 static void put_hem_table(struct hns_roce_dev *hr_dev)
2919 {
2920 	int i;
2921 
2922 	for (i = 0; i < hr_dev->caps.gmv_entry_num; i++)
2923 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2924 
2925 	if (hr_dev->is_vf)
2926 		return;
2927 
2928 	for (i = 0; i < hr_dev->caps.qpc_timer_bt_num; i++)
2929 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2930 
2931 	for (i = 0; i < hr_dev->caps.cqc_timer_bt_num; i++)
2932 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2933 }
2934 
2935 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
2936 {
2937 	int ret;
2938 
2939 	/* The hns ROCEE requires the extdb info to be cleared before using */
2940 	ret = hns_roce_clear_extdb_list_info(hr_dev);
2941 	if (ret)
2942 		return ret;
2943 
2944 	ret = get_hem_table(hr_dev);
2945 	if (ret)
2946 		return ret;
2947 
2948 	if (hr_dev->is_vf)
2949 		return 0;
2950 
2951 	ret = hns_roce_init_link_table(hr_dev);
2952 	if (ret) {
2953 		dev_err(hr_dev->dev, "failed to init llm, ret = %d.\n", ret);
2954 		goto err_llm_init_failed;
2955 	}
2956 
2957 	return 0;
2958 
2959 err_llm_init_failed:
2960 	put_hem_table(hr_dev);
2961 
2962 	return ret;
2963 }
2964 
2965 static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev)
2966 {
2967 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2968 		free_mr_exit(hr_dev);
2969 
2970 	hns_roce_function_clear(hr_dev);
2971 
2972 	if (!hr_dev->is_vf)
2973 		hns_roce_free_link_table(hr_dev);
2974 
2975 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09)
2976 		free_dip_entry(hr_dev);
2977 }
2978 
2979 static int hns_roce_mbox_post(struct hns_roce_dev *hr_dev,
2980 			      struct hns_roce_mbox_msg *mbox_msg)
2981 {
2982 	struct hns_roce_cmq_desc desc;
2983 	struct hns_roce_post_mbox *mb = (struct hns_roce_post_mbox *)desc.data;
2984 
2985 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_POST_MB, false);
2986 
2987 	mb->in_param_l = cpu_to_le32(mbox_msg->in_param);
2988 	mb->in_param_h = cpu_to_le32(mbox_msg->in_param >> 32);
2989 	mb->out_param_l = cpu_to_le32(mbox_msg->out_param);
2990 	mb->out_param_h = cpu_to_le32(mbox_msg->out_param >> 32);
2991 	mb->cmd_tag = cpu_to_le32(mbox_msg->tag << 8 | mbox_msg->cmd);
2992 	mb->token_event_en = cpu_to_le32(mbox_msg->event_en << 16 |
2993 					 mbox_msg->token);
2994 
2995 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2996 }
2997 
2998 static int v2_wait_mbox_complete(struct hns_roce_dev *hr_dev, u32 timeout,
2999 				 u8 *complete_status)
3000 {
3001 	struct hns_roce_mbox_status *mb_st;
3002 	struct hns_roce_cmq_desc desc;
3003 	unsigned long end;
3004 	int ret = -EBUSY;
3005 	u32 status;
3006 	bool busy;
3007 
3008 	mb_st = (struct hns_roce_mbox_status *)desc.data;
3009 	end = msecs_to_jiffies(timeout) + jiffies;
3010 	while (v2_chk_mbox_is_avail(hr_dev, &busy)) {
3011 		if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
3012 			return -EIO;
3013 
3014 		status = 0;
3015 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_MB_ST,
3016 					      true);
3017 		ret = __hns_roce_cmq_send(hr_dev, &desc, 1);
3018 		if (!ret) {
3019 			status = le32_to_cpu(mb_st->mb_status_hw_run);
3020 			/* No pending message exists in ROCEE mbox. */
3021 			if (!(status & MB_ST_HW_RUN_M))
3022 				break;
3023 		} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3024 			break;
3025 		}
3026 
3027 		if (time_after(jiffies, end)) {
3028 			dev_err_ratelimited(hr_dev->dev,
3029 					    "failed to wait mbox status 0x%x\n",
3030 					    status);
3031 			return -ETIMEDOUT;
3032 		}
3033 
3034 		cond_resched();
3035 		ret = -EBUSY;
3036 	}
3037 
3038 	if (!ret) {
3039 		*complete_status = (u8)(status & MB_ST_COMPLETE_M);
3040 	} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3041 		/* Ignore all errors if the mbox is unavailable. */
3042 		ret = 0;
3043 		*complete_status = MB_ST_COMPLETE_M;
3044 	}
3045 
3046 	return ret;
3047 }
3048 
3049 static int v2_post_mbox(struct hns_roce_dev *hr_dev,
3050 			struct hns_roce_mbox_msg *mbox_msg)
3051 {
3052 	u8 status = 0;
3053 	int ret;
3054 
3055 	/* Waiting for the mbox to be idle */
3056 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_V2_GO_BIT_TIMEOUT_MSECS,
3057 				    &status);
3058 	if (unlikely(ret)) {
3059 		dev_err_ratelimited(hr_dev->dev,
3060 				    "failed to check post mbox status = 0x%x, ret = %d.\n",
3061 				    status, ret);
3062 		return ret;
3063 	}
3064 
3065 	/* Post new message to mbox */
3066 	ret = hns_roce_mbox_post(hr_dev, mbox_msg);
3067 	if (ret)
3068 		dev_err_ratelimited(hr_dev->dev,
3069 				    "failed to post mailbox, ret = %d.\n", ret);
3070 
3071 	return ret;
3072 }
3073 
3074 static int v2_poll_mbox_done(struct hns_roce_dev *hr_dev)
3075 {
3076 	u8 status = 0;
3077 	int ret;
3078 
3079 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_CMD_TIMEOUT_MSECS,
3080 				    &status);
3081 	if (!ret) {
3082 		if (status != MB_ST_COMPLETE_SUCC)
3083 			return -EBUSY;
3084 	} else {
3085 		dev_err_ratelimited(hr_dev->dev,
3086 				    "failed to check mbox status = 0x%x, ret = %d.\n",
3087 				    status, ret);
3088 	}
3089 
3090 	return ret;
3091 }
3092 
3093 static void copy_gid(void *dest, const union ib_gid *gid)
3094 {
3095 #define GID_SIZE 4
3096 	const union ib_gid *src = gid;
3097 	__le32 (*p)[GID_SIZE] = dest;
3098 	int i;
3099 
3100 	if (!gid)
3101 		src = &zgid;
3102 
3103 	for (i = 0; i < GID_SIZE; i++)
3104 		(*p)[i] = cpu_to_le32(*(u32 *)&src->raw[i * sizeof(u32)]);
3105 }
3106 
3107 static int config_sgid_table(struct hns_roce_dev *hr_dev,
3108 			     int gid_index, const union ib_gid *gid,
3109 			     enum hns_roce_sgid_type sgid_type)
3110 {
3111 	struct hns_roce_cmq_desc desc;
3112 	struct hns_roce_cfg_sgid_tb *sgid_tb =
3113 				    (struct hns_roce_cfg_sgid_tb *)desc.data;
3114 
3115 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SGID_TB, false);
3116 
3117 	hr_reg_write(sgid_tb, CFG_SGID_TB_TABLE_IDX, gid_index);
3118 	hr_reg_write(sgid_tb, CFG_SGID_TB_VF_SGID_TYPE, sgid_type);
3119 
3120 	copy_gid(&sgid_tb->vf_sgid_l, gid);
3121 
3122 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3123 }
3124 
3125 static int config_gmv_table(struct hns_roce_dev *hr_dev,
3126 			    int gid_index, const union ib_gid *gid,
3127 			    enum hns_roce_sgid_type sgid_type,
3128 			    const struct ib_gid_attr *attr)
3129 {
3130 	struct hns_roce_cmq_desc desc[2];
3131 	struct hns_roce_cfg_gmv_tb_a *tb_a =
3132 				(struct hns_roce_cfg_gmv_tb_a *)desc[0].data;
3133 	struct hns_roce_cfg_gmv_tb_b *tb_b =
3134 				(struct hns_roce_cfg_gmv_tb_b *)desc[1].data;
3135 
3136 	u16 vlan_id = VLAN_CFI_MASK;
3137 	u8 mac[ETH_ALEN] = {};
3138 	int ret;
3139 
3140 	if (gid) {
3141 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, mac);
3142 		if (ret)
3143 			return ret;
3144 	}
3145 
3146 	hns_roce_cmq_setup_basic_desc(&desc[0], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3147 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
3148 
3149 	hns_roce_cmq_setup_basic_desc(&desc[1], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3150 
3151 	copy_gid(&tb_a->vf_sgid_l, gid);
3152 
3153 	hr_reg_write(tb_a, GMV_TB_A_VF_SGID_TYPE, sgid_type);
3154 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_EN, vlan_id < VLAN_CFI_MASK);
3155 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_ID, vlan_id);
3156 
3157 	tb_b->vf_smac_l = cpu_to_le32(*(u32 *)mac);
3158 
3159 	hr_reg_write(tb_b, GMV_TB_B_SMAC_H, *(u16 *)&mac[4]);
3160 	hr_reg_write(tb_b, GMV_TB_B_SGID_IDX, gid_index);
3161 
3162 	return hns_roce_cmq_send(hr_dev, desc, 2);
3163 }
3164 
3165 static int hns_roce_v2_set_gid(struct hns_roce_dev *hr_dev, int gid_index,
3166 			       const union ib_gid *gid,
3167 			       const struct ib_gid_attr *attr)
3168 {
3169 	enum hns_roce_sgid_type sgid_type = GID_TYPE_FLAG_ROCE_V1;
3170 	int ret;
3171 
3172 	if (gid) {
3173 		if (attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3174 			if (ipv6_addr_v4mapped((void *)gid))
3175 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV4;
3176 			else
3177 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV6;
3178 		} else if (attr->gid_type == IB_GID_TYPE_ROCE) {
3179 			sgid_type = GID_TYPE_FLAG_ROCE_V1;
3180 		}
3181 	}
3182 
3183 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
3184 		ret = config_gmv_table(hr_dev, gid_index, gid, sgid_type, attr);
3185 	else
3186 		ret = config_sgid_table(hr_dev, gid_index, gid, sgid_type);
3187 
3188 	if (ret)
3189 		ibdev_err(&hr_dev->ib_dev, "failed to set gid, ret = %d!\n",
3190 			  ret);
3191 
3192 	return ret;
3193 }
3194 
3195 static int hns_roce_v2_set_mac(struct hns_roce_dev *hr_dev, u8 phy_port,
3196 			       const u8 *addr)
3197 {
3198 	struct hns_roce_cmq_desc desc;
3199 	struct hns_roce_cfg_smac_tb *smac_tb =
3200 				    (struct hns_roce_cfg_smac_tb *)desc.data;
3201 	u16 reg_smac_h;
3202 	u32 reg_smac_l;
3203 
3204 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SMAC_TB, false);
3205 
3206 	reg_smac_l = *(u32 *)(&addr[0]);
3207 	reg_smac_h = *(u16 *)(&addr[4]);
3208 
3209 	hr_reg_write(smac_tb, CFG_SMAC_TB_IDX, phy_port);
3210 	hr_reg_write(smac_tb, CFG_SMAC_TB_VF_SMAC_H, reg_smac_h);
3211 	smac_tb->vf_smac_l = cpu_to_le32(reg_smac_l);
3212 
3213 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3214 }
3215 
3216 static int set_mtpt_pbl(struct hns_roce_dev *hr_dev,
3217 			struct hns_roce_v2_mpt_entry *mpt_entry,
3218 			struct hns_roce_mr *mr)
3219 {
3220 	u64 pages[HNS_ROCE_V2_MAX_INNER_MTPT_NUM] = { 0 };
3221 	struct ib_device *ibdev = &hr_dev->ib_dev;
3222 	dma_addr_t pbl_ba;
3223 	int ret;
3224 	int i;
3225 
3226 	ret = hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, pages,
3227 				min_t(int, ARRAY_SIZE(pages), mr->npages));
3228 	if (ret) {
3229 		ibdev_err(ibdev, "failed to find PBL mtr, ret = %d.\n", ret);
3230 		return ret;
3231 	}
3232 
3233 	/* Aligned to the hardware address access unit */
3234 	for (i = 0; i < ARRAY_SIZE(pages); i++)
3235 		pages[i] >>= MPT_PBL_BUF_ADDR_S;
3236 
3237 	pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3238 
3239 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3240 	mpt_entry->pbl_ba_l = cpu_to_le32(pbl_ba >> MPT_PBL_BA_ADDR_S);
3241 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3242 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3243 
3244 	mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
3245 	hr_reg_write(mpt_entry, MPT_PA0_H, upper_32_bits(pages[0]));
3246 
3247 	mpt_entry->pa1_l = cpu_to_le32(lower_32_bits(pages[1]));
3248 	hr_reg_write(mpt_entry, MPT_PA1_H, upper_32_bits(pages[1]));
3249 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3250 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3251 
3252 	return 0;
3253 }
3254 
3255 static int hns_roce_v2_write_mtpt(struct hns_roce_dev *hr_dev,
3256 				  void *mb_buf, struct hns_roce_mr *mr)
3257 {
3258 	struct hns_roce_v2_mpt_entry *mpt_entry;
3259 
3260 	mpt_entry = mb_buf;
3261 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3262 
3263 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3264 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3265 
3266 	hr_reg_write_bool(mpt_entry, MPT_BIND_EN,
3267 			  mr->access & IB_ACCESS_MW_BIND);
3268 	hr_reg_write_bool(mpt_entry, MPT_ATOMIC_EN,
3269 			  mr->access & IB_ACCESS_REMOTE_ATOMIC);
3270 	hr_reg_write_bool(mpt_entry, MPT_RR_EN,
3271 			  mr->access & IB_ACCESS_REMOTE_READ);
3272 	hr_reg_write_bool(mpt_entry, MPT_RW_EN,
3273 			  mr->access & IB_ACCESS_REMOTE_WRITE);
3274 	hr_reg_write_bool(mpt_entry, MPT_LW_EN,
3275 			  mr->access & IB_ACCESS_LOCAL_WRITE);
3276 
3277 	mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3278 	mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3279 	mpt_entry->lkey = cpu_to_le32(mr->key);
3280 	mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3281 	mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3282 
3283 	if (mr->type != MR_TYPE_MR)
3284 		hr_reg_enable(mpt_entry, MPT_PA);
3285 
3286 	if (mr->type == MR_TYPE_DMA)
3287 		return 0;
3288 
3289 	if (mr->pbl_hop_num != HNS_ROCE_HOP_NUM_0)
3290 		hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, mr->pbl_hop_num);
3291 
3292 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3293 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3294 	hr_reg_enable(mpt_entry, MPT_INNER_PA_VLD);
3295 
3296 	return set_mtpt_pbl(hr_dev, mpt_entry, mr);
3297 }
3298 
3299 static int hns_roce_v2_rereg_write_mtpt(struct hns_roce_dev *hr_dev,
3300 					struct hns_roce_mr *mr, int flags,
3301 					void *mb_buf)
3302 {
3303 	struct hns_roce_v2_mpt_entry *mpt_entry = mb_buf;
3304 	u32 mr_access_flags = mr->access;
3305 	int ret = 0;
3306 
3307 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3308 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3309 
3310 	if (flags & IB_MR_REREG_ACCESS) {
3311 		hr_reg_write(mpt_entry, MPT_BIND_EN,
3312 			     (mr_access_flags & IB_ACCESS_MW_BIND ? 1 : 0));
3313 		hr_reg_write(mpt_entry, MPT_ATOMIC_EN,
3314 			     mr_access_flags & IB_ACCESS_REMOTE_ATOMIC ? 1 : 0);
3315 		hr_reg_write(mpt_entry, MPT_RR_EN,
3316 			     mr_access_flags & IB_ACCESS_REMOTE_READ ? 1 : 0);
3317 		hr_reg_write(mpt_entry, MPT_RW_EN,
3318 			     mr_access_flags & IB_ACCESS_REMOTE_WRITE ? 1 : 0);
3319 		hr_reg_write(mpt_entry, MPT_LW_EN,
3320 			     mr_access_flags & IB_ACCESS_LOCAL_WRITE ? 1 : 0);
3321 	}
3322 
3323 	if (flags & IB_MR_REREG_TRANS) {
3324 		mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3325 		mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3326 		mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3327 		mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3328 
3329 		ret = set_mtpt_pbl(hr_dev, mpt_entry, mr);
3330 	}
3331 
3332 	return ret;
3333 }
3334 
3335 static int hns_roce_v2_frmr_write_mtpt(void *mb_buf, struct hns_roce_mr *mr)
3336 {
3337 	dma_addr_t pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3338 	struct hns_roce_v2_mpt_entry *mpt_entry;
3339 
3340 	mpt_entry = mb_buf;
3341 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3342 
3343 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3344 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3345 
3346 	hr_reg_enable(mpt_entry, MPT_RA_EN);
3347 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3348 
3349 	hr_reg_enable(mpt_entry, MPT_FRE);
3350 	hr_reg_clear(mpt_entry, MPT_MR_MW);
3351 	hr_reg_enable(mpt_entry, MPT_BPD);
3352 	hr_reg_clear(mpt_entry, MPT_PA);
3353 
3354 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, 1);
3355 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3356 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3357 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3358 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3359 
3360 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3361 
3362 	mpt_entry->pbl_ba_l = cpu_to_le32(lower_32_bits(pbl_ba >>
3363 							MPT_PBL_BA_ADDR_S));
3364 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3365 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3366 
3367 	return 0;
3368 }
3369 
3370 static int hns_roce_v2_mw_write_mtpt(void *mb_buf, struct hns_roce_mw *mw)
3371 {
3372 	struct hns_roce_v2_mpt_entry *mpt_entry;
3373 
3374 	mpt_entry = mb_buf;
3375 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3376 
3377 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3378 	hr_reg_write(mpt_entry, MPT_PD, mw->pdn);
3379 
3380 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3381 	hr_reg_enable(mpt_entry, MPT_LW_EN);
3382 
3383 	hr_reg_enable(mpt_entry, MPT_MR_MW);
3384 	hr_reg_enable(mpt_entry, MPT_BPD);
3385 	hr_reg_clear(mpt_entry, MPT_PA);
3386 	hr_reg_write(mpt_entry, MPT_BQP,
3387 		     mw->ibmw.type == IB_MW_TYPE_1 ? 0 : 1);
3388 
3389 	mpt_entry->lkey = cpu_to_le32(mw->rkey);
3390 
3391 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM,
3392 		     mw->pbl_hop_num == HNS_ROCE_HOP_NUM_0 ? 0 :
3393 							     mw->pbl_hop_num);
3394 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3395 		     mw->pbl_ba_pg_sz + PG_SHIFT_OFFSET);
3396 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3397 		     mw->pbl_buf_pg_sz + PG_SHIFT_OFFSET);
3398 
3399 	return 0;
3400 }
3401 
3402 static int free_mr_post_send_lp_wqe(struct hns_roce_qp *hr_qp)
3403 {
3404 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
3405 	struct ib_device *ibdev = &hr_dev->ib_dev;
3406 	const struct ib_send_wr *bad_wr;
3407 	struct ib_rdma_wr rdma_wr = {};
3408 	struct ib_send_wr *send_wr;
3409 	int ret;
3410 
3411 	send_wr = &rdma_wr.wr;
3412 	send_wr->opcode = IB_WR_RDMA_WRITE;
3413 
3414 	ret = hns_roce_v2_post_send(&hr_qp->ibqp, send_wr, &bad_wr);
3415 	if (ret) {
3416 		ibdev_err_ratelimited(ibdev, "failed to post wqe for free mr, ret = %d.\n",
3417 				      ret);
3418 		return ret;
3419 	}
3420 
3421 	return 0;
3422 }
3423 
3424 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3425 			       struct ib_wc *wc);
3426 
3427 static void free_mr_send_cmd_to_hw(struct hns_roce_dev *hr_dev)
3428 {
3429 	struct hns_roce_v2_priv *priv = hr_dev->priv;
3430 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
3431 	struct ib_wc wc[ARRAY_SIZE(free_mr->rsv_qp)];
3432 	struct ib_device *ibdev = &hr_dev->ib_dev;
3433 	struct hns_roce_qp *hr_qp;
3434 	unsigned long end;
3435 	int cqe_cnt = 0;
3436 	int npolled;
3437 	int ret;
3438 	int i;
3439 
3440 	/*
3441 	 * If the device initialization is not complete or in the uninstall
3442 	 * process, then there is no need to execute free mr.
3443 	 */
3444 	if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
3445 	    priv->handle->rinfo.instance_state == HNS_ROCE_STATE_INIT ||
3446 	    hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT)
3447 		return;
3448 
3449 	mutex_lock(&free_mr->mutex);
3450 
3451 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
3452 		hr_qp = free_mr->rsv_qp[i];
3453 
3454 		ret = free_mr_post_send_lp_wqe(hr_qp);
3455 		if (ret) {
3456 			ibdev_err_ratelimited(ibdev,
3457 					      "failed to send wqe (qp:0x%lx) for free mr, ret = %d.\n",
3458 					      hr_qp->qpn, ret);
3459 			break;
3460 		}
3461 
3462 		cqe_cnt++;
3463 	}
3464 
3465 	end = msecs_to_jiffies(HNS_ROCE_V2_FREE_MR_TIMEOUT) + jiffies;
3466 	while (cqe_cnt) {
3467 		npolled = hns_roce_v2_poll_cq(&free_mr->rsv_cq->ib_cq, cqe_cnt, wc);
3468 		if (npolled < 0) {
3469 			ibdev_err_ratelimited(ibdev,
3470 					      "failed to poll cqe for free mr, remain %d cqe.\n",
3471 					      cqe_cnt);
3472 			goto out;
3473 		}
3474 
3475 		if (time_after(jiffies, end)) {
3476 			ibdev_err_ratelimited(ibdev,
3477 					      "failed to poll cqe for free mr and timeout, remain %d cqe.\n",
3478 					      cqe_cnt);
3479 			goto out;
3480 		}
3481 		cqe_cnt -= npolled;
3482 	}
3483 
3484 out:
3485 	mutex_unlock(&free_mr->mutex);
3486 }
3487 
3488 static void hns_roce_v2_dereg_mr(struct hns_roce_dev *hr_dev)
3489 {
3490 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
3491 		free_mr_send_cmd_to_hw(hr_dev);
3492 }
3493 
3494 static void *get_cqe_v2(struct hns_roce_cq *hr_cq, int n)
3495 {
3496 	return hns_roce_buf_offset(hr_cq->mtr.kmem, n * hr_cq->cqe_size);
3497 }
3498 
3499 static void *get_sw_cqe_v2(struct hns_roce_cq *hr_cq, unsigned int n)
3500 {
3501 	struct hns_roce_v2_cqe *cqe = get_cqe_v2(hr_cq, n & hr_cq->ib_cq.cqe);
3502 
3503 	/* Get cqe when Owner bit is Conversely with the MSB of cons_idx */
3504 	return (hr_reg_read(cqe, CQE_OWNER) ^ !!(n & hr_cq->cq_depth)) ? cqe :
3505 									 NULL;
3506 }
3507 
3508 static inline void update_cq_db(struct hns_roce_dev *hr_dev,
3509 				struct hns_roce_cq *hr_cq)
3510 {
3511 	if (likely(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) {
3512 		*hr_cq->set_ci_db = hr_cq->cons_index & V2_CQ_DB_CONS_IDX_M;
3513 	} else {
3514 		struct hns_roce_v2_db cq_db = {};
3515 
3516 		hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3517 		hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB);
3518 		hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3519 		hr_reg_write(&cq_db, DB_CQ_CMD_SN, 1);
3520 
3521 		hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
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 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3529 	struct hns_roce_v2_cqe *cqe, *dest;
3530 	u32 prod_index;
3531 	int nfreed = 0;
3532 	int wqe_index;
3533 	u8 owner_bit;
3534 
3535 	for (prod_index = hr_cq->cons_index; get_sw_cqe_v2(hr_cq, prod_index);
3536 	     ++prod_index) {
3537 		if (prod_index > hr_cq->cons_index + hr_cq->ib_cq.cqe)
3538 			break;
3539 	}
3540 
3541 	/*
3542 	 * Now backwards through the CQ, removing CQ entries
3543 	 * that match our QP by overwriting them with next entries.
3544 	 */
3545 	while ((int) --prod_index - (int) hr_cq->cons_index >= 0) {
3546 		cqe = get_cqe_v2(hr_cq, prod_index & hr_cq->ib_cq.cqe);
3547 		if (hr_reg_read(cqe, CQE_LCL_QPN) == qpn) {
3548 			if (srq && hr_reg_read(cqe, CQE_S_R)) {
3549 				wqe_index = hr_reg_read(cqe, CQE_WQE_IDX);
3550 				hns_roce_free_srq_wqe(srq, wqe_index);
3551 			}
3552 			++nfreed;
3553 		} else if (nfreed) {
3554 			dest = get_cqe_v2(hr_cq, (prod_index + nfreed) &
3555 					  hr_cq->ib_cq.cqe);
3556 			owner_bit = hr_reg_read(dest, CQE_OWNER);
3557 			memcpy(dest, cqe, hr_cq->cqe_size);
3558 			hr_reg_write(dest, CQE_OWNER, owner_bit);
3559 		}
3560 	}
3561 
3562 	if (nfreed) {
3563 		hr_cq->cons_index += nfreed;
3564 		update_cq_db(hr_dev, hr_cq);
3565 	}
3566 }
3567 
3568 static void hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3569 				 struct hns_roce_srq *srq)
3570 {
3571 	spin_lock_irq(&hr_cq->lock);
3572 	__hns_roce_v2_cq_clean(hr_cq, qpn, srq);
3573 	spin_unlock_irq(&hr_cq->lock);
3574 }
3575 
3576 static void hns_roce_v2_write_cqc(struct hns_roce_dev *hr_dev,
3577 				  struct hns_roce_cq *hr_cq, void *mb_buf,
3578 				  u64 *mtts, dma_addr_t dma_handle)
3579 {
3580 	struct hns_roce_v2_cq_context *cq_context;
3581 
3582 	cq_context = mb_buf;
3583 	memset(cq_context, 0, sizeof(*cq_context));
3584 
3585 	hr_reg_write(cq_context, CQC_CQ_ST, V2_CQ_STATE_VALID);
3586 	hr_reg_write(cq_context, CQC_ARM_ST, NO_ARMED);
3587 	hr_reg_write(cq_context, CQC_SHIFT, ilog2(hr_cq->cq_depth));
3588 	hr_reg_write(cq_context, CQC_CEQN, hr_cq->vector);
3589 	hr_reg_write(cq_context, CQC_CQN, hr_cq->cqn);
3590 
3591 	if (hr_cq->cqe_size == HNS_ROCE_V3_CQE_SIZE)
3592 		hr_reg_write(cq_context, CQC_CQE_SIZE, CQE_SIZE_64B);
3593 
3594 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
3595 		hr_reg_enable(cq_context, CQC_STASH);
3596 
3597 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_L,
3598 		     to_hr_hw_page_addr(mtts[0]));
3599 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_H,
3600 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
3601 	hr_reg_write(cq_context, CQC_CQE_HOP_NUM, hr_dev->caps.cqe_hop_num ==
3602 		     HNS_ROCE_HOP_NUM_0 ? 0 : hr_dev->caps.cqe_hop_num);
3603 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_L,
3604 		     to_hr_hw_page_addr(mtts[1]));
3605 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_H,
3606 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
3607 	hr_reg_write(cq_context, CQC_CQE_BAR_PG_SZ,
3608 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.ba_pg_shift));
3609 	hr_reg_write(cq_context, CQC_CQE_BUF_PG_SZ,
3610 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.buf_pg_shift));
3611 	hr_reg_write(cq_context, CQC_CQE_BA_L, dma_handle >> CQC_CQE_BA_L_S);
3612 	hr_reg_write(cq_context, CQC_CQE_BA_H, dma_handle >> CQC_CQE_BA_H_S);
3613 	hr_reg_write_bool(cq_context, CQC_DB_RECORD_EN,
3614 			  hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB);
3615 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_L,
3616 		     ((u32)hr_cq->db.dma) >> 1);
3617 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_H,
3618 		     hr_cq->db.dma >> CQC_CQE_DB_RECORD_ADDR_H_S);
3619 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT,
3620 		     HNS_ROCE_V2_CQ_DEFAULT_BURST_NUM);
3621 	hr_reg_write(cq_context, CQC_CQ_PERIOD,
3622 		     HNS_ROCE_V2_CQ_DEFAULT_INTERVAL);
3623 }
3624 
3625 static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
3626 				     enum ib_cq_notify_flags flags)
3627 {
3628 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3629 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3630 	struct hns_roce_v2_db cq_db = {};
3631 	u32 notify_flag;
3632 
3633 	/*
3634 	 * flags = 0, then notify_flag : next
3635 	 * flags = 1, then notify flag : solocited
3636 	 */
3637 	notify_flag = (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
3638 		      V2_CQ_DB_REQ_NOT : V2_CQ_DB_REQ_NOT_SOL;
3639 
3640 	hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3641 	hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB_NOTIFY);
3642 	hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3643 	hr_reg_write(&cq_db, DB_CQ_CMD_SN, hr_cq->arm_sn);
3644 	hr_reg_write(&cq_db, DB_CQ_NOTIFY, notify_flag);
3645 
3646 	hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3647 
3648 	return 0;
3649 }
3650 
3651 static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
3652 		   int num_entries, struct ib_wc *wc)
3653 {
3654 	unsigned int left;
3655 	int npolled = 0;
3656 
3657 	left = wq->head - wq->tail;
3658 	if (left == 0)
3659 		return 0;
3660 
3661 	left = min_t(unsigned int, (unsigned int)num_entries, left);
3662 	while (npolled < left) {
3663 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3664 		wc->status = IB_WC_WR_FLUSH_ERR;
3665 		wc->vendor_err = 0;
3666 		wc->qp = &hr_qp->ibqp;
3667 
3668 		wq->tail++;
3669 		wc++;
3670 		npolled++;
3671 	}
3672 
3673 	return npolled;
3674 }
3675 
3676 static int hns_roce_v2_sw_poll_cq(struct hns_roce_cq *hr_cq, int num_entries,
3677 				  struct ib_wc *wc)
3678 {
3679 	struct hns_roce_qp *hr_qp;
3680 	int npolled = 0;
3681 
3682 	list_for_each_entry(hr_qp, &hr_cq->sq_list, sq_node) {
3683 		npolled += sw_comp(hr_qp, &hr_qp->sq,
3684 				   num_entries - npolled, wc + npolled);
3685 		if (npolled >= num_entries)
3686 			goto out;
3687 	}
3688 
3689 	list_for_each_entry(hr_qp, &hr_cq->rq_list, rq_node) {
3690 		npolled += sw_comp(hr_qp, &hr_qp->rq,
3691 				   num_entries - npolled, wc + npolled);
3692 		if (npolled >= num_entries)
3693 			goto out;
3694 	}
3695 
3696 out:
3697 	return npolled;
3698 }
3699 
3700 static void get_cqe_status(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
3701 			   struct hns_roce_cq *cq, struct hns_roce_v2_cqe *cqe,
3702 			   struct ib_wc *wc)
3703 {
3704 	static const struct {
3705 		u32 cqe_status;
3706 		enum ib_wc_status wc_status;
3707 	} map[] = {
3708 		{ HNS_ROCE_CQE_V2_SUCCESS, IB_WC_SUCCESS },
3709 		{ HNS_ROCE_CQE_V2_LOCAL_LENGTH_ERR, IB_WC_LOC_LEN_ERR },
3710 		{ HNS_ROCE_CQE_V2_LOCAL_QP_OP_ERR, IB_WC_LOC_QP_OP_ERR },
3711 		{ HNS_ROCE_CQE_V2_LOCAL_PROT_ERR, IB_WC_LOC_PROT_ERR },
3712 		{ HNS_ROCE_CQE_V2_WR_FLUSH_ERR, IB_WC_WR_FLUSH_ERR },
3713 		{ HNS_ROCE_CQE_V2_MW_BIND_ERR, IB_WC_MW_BIND_ERR },
3714 		{ HNS_ROCE_CQE_V2_BAD_RESP_ERR, IB_WC_BAD_RESP_ERR },
3715 		{ HNS_ROCE_CQE_V2_LOCAL_ACCESS_ERR, IB_WC_LOC_ACCESS_ERR },
3716 		{ HNS_ROCE_CQE_V2_REMOTE_INVAL_REQ_ERR, IB_WC_REM_INV_REQ_ERR },
3717 		{ HNS_ROCE_CQE_V2_REMOTE_ACCESS_ERR, IB_WC_REM_ACCESS_ERR },
3718 		{ HNS_ROCE_CQE_V2_REMOTE_OP_ERR, IB_WC_REM_OP_ERR },
3719 		{ HNS_ROCE_CQE_V2_TRANSPORT_RETRY_EXC_ERR,
3720 		  IB_WC_RETRY_EXC_ERR },
3721 		{ HNS_ROCE_CQE_V2_RNR_RETRY_EXC_ERR, IB_WC_RNR_RETRY_EXC_ERR },
3722 		{ HNS_ROCE_CQE_V2_REMOTE_ABORT_ERR, IB_WC_REM_ABORT_ERR },
3723 		{ HNS_ROCE_CQE_V2_GENERAL_ERR, IB_WC_GENERAL_ERR}
3724 	};
3725 
3726 	u32 cqe_status = hr_reg_read(cqe, CQE_STATUS);
3727 	int i;
3728 
3729 	wc->status = IB_WC_GENERAL_ERR;
3730 	for (i = 0; i < ARRAY_SIZE(map); i++)
3731 		if (cqe_status == map[i].cqe_status) {
3732 			wc->status = map[i].wc_status;
3733 			break;
3734 		}
3735 
3736 	if (likely(wc->status == IB_WC_SUCCESS ||
3737 		   wc->status == IB_WC_WR_FLUSH_ERR))
3738 		return;
3739 
3740 	ibdev_err_ratelimited(&hr_dev->ib_dev, "error cqe status 0x%x:\n",
3741 			      cqe_status);
3742 	print_hex_dump(KERN_DEBUG, "", DUMP_PREFIX_NONE, 16, 4, cqe,
3743 		       cq->cqe_size, false);
3744 	wc->vendor_err = hr_reg_read(cqe, CQE_SUB_STATUS);
3745 
3746 	/*
3747 	 * For hns ROCEE, GENERAL_ERR is an error type that is not defined in
3748 	 * the standard protocol, the driver must ignore it and needn't to set
3749 	 * the QP to an error state.
3750 	 */
3751 	if (cqe_status == HNS_ROCE_CQE_V2_GENERAL_ERR)
3752 		return;
3753 
3754 	flush_cqe(hr_dev, qp);
3755 }
3756 
3757 static int get_cur_qp(struct hns_roce_cq *hr_cq, struct hns_roce_v2_cqe *cqe,
3758 		      struct hns_roce_qp **cur_qp)
3759 {
3760 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3761 	struct hns_roce_qp *hr_qp = *cur_qp;
3762 	u32 qpn;
3763 
3764 	qpn = hr_reg_read(cqe, CQE_LCL_QPN);
3765 
3766 	if (!hr_qp || qpn != hr_qp->qpn) {
3767 		hr_qp = __hns_roce_qp_lookup(hr_dev, qpn);
3768 		if (unlikely(!hr_qp)) {
3769 			ibdev_err(&hr_dev->ib_dev,
3770 				  "CQ %06lx with entry for unknown QPN %06x\n",
3771 				  hr_cq->cqn, qpn);
3772 			return -EINVAL;
3773 		}
3774 		*cur_qp = hr_qp;
3775 	}
3776 
3777 	return 0;
3778 }
3779 
3780 /*
3781  * mapped-value = 1 + real-value
3782  * The ib wc opcode's real value is start from 0, In order to distinguish
3783  * between initialized and uninitialized map values, we plus 1 to the actual
3784  * value when defining the mapping, so that the validity can be identified by
3785  * checking whether the mapped value is greater than 0.
3786  */
3787 #define HR_WC_OP_MAP(hr_key, ib_key) \
3788 		[HNS_ROCE_V2_WQE_OP_ ## hr_key] = 1 + IB_WC_ ## ib_key
3789 
3790 static const u32 wc_send_op_map[] = {
3791 	HR_WC_OP_MAP(SEND,			SEND),
3792 	HR_WC_OP_MAP(SEND_WITH_INV,		SEND),
3793 	HR_WC_OP_MAP(SEND_WITH_IMM,		SEND),
3794 	HR_WC_OP_MAP(RDMA_READ,			RDMA_READ),
3795 	HR_WC_OP_MAP(RDMA_WRITE,		RDMA_WRITE),
3796 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,	RDMA_WRITE),
3797 	HR_WC_OP_MAP(ATOM_CMP_AND_SWAP,		COMP_SWAP),
3798 	HR_WC_OP_MAP(ATOM_FETCH_AND_ADD,	FETCH_ADD),
3799 	HR_WC_OP_MAP(ATOM_MSK_CMP_AND_SWAP,	MASKED_COMP_SWAP),
3800 	HR_WC_OP_MAP(ATOM_MSK_FETCH_AND_ADD,	MASKED_FETCH_ADD),
3801 	HR_WC_OP_MAP(FAST_REG_PMR,		REG_MR),
3802 	HR_WC_OP_MAP(BIND_MW,			REG_MR),
3803 };
3804 
3805 static int to_ib_wc_send_op(u32 hr_opcode)
3806 {
3807 	if (hr_opcode >= ARRAY_SIZE(wc_send_op_map))
3808 		return -EINVAL;
3809 
3810 	return wc_send_op_map[hr_opcode] ? wc_send_op_map[hr_opcode] - 1 :
3811 					   -EINVAL;
3812 }
3813 
3814 static const u32 wc_recv_op_map[] = {
3815 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,		WITH_IMM),
3816 	HR_WC_OP_MAP(SEND,				RECV),
3817 	HR_WC_OP_MAP(SEND_WITH_IMM,			WITH_IMM),
3818 	HR_WC_OP_MAP(SEND_WITH_INV,			RECV),
3819 };
3820 
3821 static int to_ib_wc_recv_op(u32 hr_opcode)
3822 {
3823 	if (hr_opcode >= ARRAY_SIZE(wc_recv_op_map))
3824 		return -EINVAL;
3825 
3826 	return wc_recv_op_map[hr_opcode] ? wc_recv_op_map[hr_opcode] - 1 :
3827 					   -EINVAL;
3828 }
3829 
3830 static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3831 {
3832 	u32 hr_opcode;
3833 	int ib_opcode;
3834 
3835 	wc->wc_flags = 0;
3836 
3837 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3838 	switch (hr_opcode) {
3839 	case HNS_ROCE_V2_WQE_OP_RDMA_READ:
3840 		wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3841 		break;
3842 	case HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM:
3843 	case HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM:
3844 		wc->wc_flags |= IB_WC_WITH_IMM;
3845 		break;
3846 	case HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP:
3847 	case HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD:
3848 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP:
3849 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD:
3850 		wc->byte_len  = 8;
3851 		break;
3852 	default:
3853 		break;
3854 	}
3855 
3856 	ib_opcode = to_ib_wc_send_op(hr_opcode);
3857 	if (ib_opcode < 0)
3858 		wc->status = IB_WC_GENERAL_ERR;
3859 	else
3860 		wc->opcode = ib_opcode;
3861 }
3862 
3863 static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3864 {
3865 	u32 hr_opcode;
3866 	int ib_opcode;
3867 
3868 	wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3869 
3870 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3871 	switch (hr_opcode) {
3872 	case HNS_ROCE_V2_OPCODE_RDMA_WRITE_IMM:
3873 	case HNS_ROCE_V2_OPCODE_SEND_WITH_IMM:
3874 		wc->wc_flags = IB_WC_WITH_IMM;
3875 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->immtdata));
3876 		break;
3877 	case HNS_ROCE_V2_OPCODE_SEND_WITH_INV:
3878 		wc->wc_flags = IB_WC_WITH_INVALIDATE;
3879 		wc->ex.invalidate_rkey = le32_to_cpu(cqe->rkey);
3880 		break;
3881 	default:
3882 		wc->wc_flags = 0;
3883 	}
3884 
3885 	ib_opcode = to_ib_wc_recv_op(hr_opcode);
3886 	if (ib_opcode < 0)
3887 		wc->status = IB_WC_GENERAL_ERR;
3888 	else
3889 		wc->opcode = ib_opcode;
3890 
3891 	wc->sl = hr_reg_read(cqe, CQE_SL);
3892 	wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
3893 	wc->slid = 0;
3894 	wc->wc_flags |= hr_reg_read(cqe, CQE_GRH) ? IB_WC_GRH : 0;
3895 	wc->port_num = hr_reg_read(cqe, CQE_PORTN);
3896 	wc->pkey_index = 0;
3897 
3898 	if (hr_reg_read(cqe, CQE_VID_VLD)) {
3899 		wc->vlan_id = hr_reg_read(cqe, CQE_VID);
3900 		wc->wc_flags |= IB_WC_WITH_VLAN;
3901 	} else {
3902 		wc->vlan_id = 0xffff;
3903 	}
3904 
3905 	wc->network_hdr_type = hr_reg_read(cqe, CQE_PORT_TYPE);
3906 
3907 	return 0;
3908 }
3909 
3910 static int hns_roce_v2_poll_one(struct hns_roce_cq *hr_cq,
3911 				struct hns_roce_qp **cur_qp, struct ib_wc *wc)
3912 {
3913 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3914 	struct hns_roce_qp *qp = *cur_qp;
3915 	struct hns_roce_srq *srq = NULL;
3916 	struct hns_roce_v2_cqe *cqe;
3917 	struct hns_roce_wq *wq;
3918 	int is_send;
3919 	u16 wqe_idx;
3920 	int ret;
3921 
3922 	cqe = get_sw_cqe_v2(hr_cq, hr_cq->cons_index);
3923 	if (!cqe)
3924 		return -EAGAIN;
3925 
3926 	++hr_cq->cons_index;
3927 	/* Memory barrier */
3928 	rmb();
3929 
3930 	ret = get_cur_qp(hr_cq, cqe, &qp);
3931 	if (ret)
3932 		return ret;
3933 
3934 	wc->qp = &qp->ibqp;
3935 	wc->vendor_err = 0;
3936 
3937 	wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
3938 
3939 	is_send = !hr_reg_read(cqe, CQE_S_R);
3940 	if (is_send) {
3941 		wq = &qp->sq;
3942 
3943 		/* If sg_signal_bit is set, tail pointer will be updated to
3944 		 * the WQE corresponding to the current CQE.
3945 		 */
3946 		if (qp->sq_signal_bits)
3947 			wq->tail += (wqe_idx - (u16)wq->tail) &
3948 				    (wq->wqe_cnt - 1);
3949 
3950 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3951 		++wq->tail;
3952 
3953 		fill_send_wc(wc, cqe);
3954 	} else {
3955 		if (qp->ibqp.srq) {
3956 			srq = to_hr_srq(qp->ibqp.srq);
3957 			wc->wr_id = srq->wrid[wqe_idx];
3958 			hns_roce_free_srq_wqe(srq, wqe_idx);
3959 		} else {
3960 			wq = &qp->rq;
3961 			wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3962 			++wq->tail;
3963 		}
3964 
3965 		ret = fill_recv_wc(wc, cqe);
3966 	}
3967 
3968 	get_cqe_status(hr_dev, qp, hr_cq, cqe, wc);
3969 	if (unlikely(wc->status != IB_WC_SUCCESS))
3970 		return 0;
3971 
3972 	return ret;
3973 }
3974 
3975 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3976 			       struct ib_wc *wc)
3977 {
3978 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3979 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3980 	struct hns_roce_qp *cur_qp = NULL;
3981 	unsigned long flags;
3982 	int npolled;
3983 
3984 	spin_lock_irqsave(&hr_cq->lock, flags);
3985 
3986 	/*
3987 	 * When the device starts to reset, the state is RST_DOWN. At this time,
3988 	 * there may still be some valid CQEs in the hardware that are not
3989 	 * polled. Therefore, it is not allowed to switch to the software mode
3990 	 * immediately. When the state changes to UNINIT, CQE no longer exists
3991 	 * in the hardware, and then switch to software mode.
3992 	 */
3993 	if (hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT) {
3994 		npolled = hns_roce_v2_sw_poll_cq(hr_cq, num_entries, wc);
3995 		goto out;
3996 	}
3997 
3998 	for (npolled = 0; npolled < num_entries; ++npolled) {
3999 		if (hns_roce_v2_poll_one(hr_cq, &cur_qp, wc + npolled))
4000 			break;
4001 	}
4002 
4003 	if (npolled)
4004 		update_cq_db(hr_dev, hr_cq);
4005 
4006 out:
4007 	spin_unlock_irqrestore(&hr_cq->lock, flags);
4008 
4009 	return npolled;
4010 }
4011 
4012 static int get_op_for_set_hem(struct hns_roce_dev *hr_dev, u32 type,
4013 			      u32 step_idx, u8 *mbox_cmd)
4014 {
4015 	u8 cmd;
4016 
4017 	switch (type) {
4018 	case HEM_TYPE_QPC:
4019 		cmd = HNS_ROCE_CMD_WRITE_QPC_BT0;
4020 		break;
4021 	case HEM_TYPE_MTPT:
4022 		cmd = HNS_ROCE_CMD_WRITE_MPT_BT0;
4023 		break;
4024 	case HEM_TYPE_CQC:
4025 		cmd = HNS_ROCE_CMD_WRITE_CQC_BT0;
4026 		break;
4027 	case HEM_TYPE_SRQC:
4028 		cmd = HNS_ROCE_CMD_WRITE_SRQC_BT0;
4029 		break;
4030 	case HEM_TYPE_SCCC:
4031 		cmd = HNS_ROCE_CMD_WRITE_SCCC_BT0;
4032 		break;
4033 	case HEM_TYPE_QPC_TIMER:
4034 		cmd = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
4035 		break;
4036 	case HEM_TYPE_CQC_TIMER:
4037 		cmd = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
4038 		break;
4039 	default:
4040 		dev_warn(hr_dev->dev, "failed to check hem type %u.\n", type);
4041 		return -EINVAL;
4042 	}
4043 
4044 	*mbox_cmd = cmd + step_idx;
4045 
4046 	return 0;
4047 }
4048 
4049 static int config_gmv_ba_to_hw(struct hns_roce_dev *hr_dev, unsigned long obj,
4050 			       dma_addr_t base_addr)
4051 {
4052 	struct hns_roce_cmq_desc desc;
4053 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
4054 	u32 idx = obj / (HNS_HW_PAGE_SIZE / hr_dev->caps.gmv_entry_sz);
4055 	u64 addr = to_hr_hw_page_addr(base_addr);
4056 
4057 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
4058 
4059 	hr_reg_write(req, CFG_GMV_BT_BA_L, lower_32_bits(addr));
4060 	hr_reg_write(req, CFG_GMV_BT_BA_H, upper_32_bits(addr));
4061 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
4062 
4063 	return hns_roce_cmq_send(hr_dev, &desc, 1);
4064 }
4065 
4066 static int set_hem_to_hw(struct hns_roce_dev *hr_dev, int obj,
4067 			 dma_addr_t base_addr, u32 hem_type, u32 step_idx)
4068 {
4069 	int ret;
4070 	u8 cmd;
4071 
4072 	if (unlikely(hem_type == HEM_TYPE_GMV))
4073 		return config_gmv_ba_to_hw(hr_dev, obj, base_addr);
4074 
4075 	if (unlikely(hem_type == HEM_TYPE_SCCC && step_idx))
4076 		return 0;
4077 
4078 	ret = get_op_for_set_hem(hr_dev, hem_type, step_idx, &cmd);
4079 	if (ret < 0)
4080 		return ret;
4081 
4082 	return config_hem_ba_to_hw(hr_dev, base_addr, cmd, obj);
4083 }
4084 
4085 static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
4086 			       struct hns_roce_hem_table *table, int obj,
4087 			       u32 step_idx)
4088 {
4089 	struct hns_roce_hem_mhop mhop;
4090 	struct hns_roce_hem *hem;
4091 	unsigned long mhop_obj = obj;
4092 	int i, j, k;
4093 	int ret = 0;
4094 	u64 hem_idx = 0;
4095 	u64 l1_idx = 0;
4096 	u64 bt_ba = 0;
4097 	u32 chunk_ba_num;
4098 	u32 hop_num;
4099 
4100 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4101 		return 0;
4102 
4103 	hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
4104 	i = mhop.l0_idx;
4105 	j = mhop.l1_idx;
4106 	k = mhop.l2_idx;
4107 	hop_num = mhop.hop_num;
4108 	chunk_ba_num = mhop.bt_chunk_size / 8;
4109 
4110 	if (hop_num == 2) {
4111 		hem_idx = i * chunk_ba_num * chunk_ba_num + j * chunk_ba_num +
4112 			  k;
4113 		l1_idx = i * chunk_ba_num + j;
4114 	} else if (hop_num == 1) {
4115 		hem_idx = i * chunk_ba_num + j;
4116 	} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
4117 		hem_idx = i;
4118 	}
4119 
4120 	if (table->type == HEM_TYPE_SCCC)
4121 		obj = mhop.l0_idx;
4122 
4123 	if (check_whether_last_step(hop_num, step_idx)) {
4124 		hem = table->hem[hem_idx];
4125 
4126 		ret = set_hem_to_hw(hr_dev, obj, hem->dma, table->type, step_idx);
4127 	} else {
4128 		if (step_idx == 0)
4129 			bt_ba = table->bt_l0_dma_addr[i];
4130 		else if (step_idx == 1 && hop_num == 2)
4131 			bt_ba = table->bt_l1_dma_addr[l1_idx];
4132 
4133 		ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type, step_idx);
4134 	}
4135 
4136 	return ret;
4137 }
4138 
4139 static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
4140 				 struct hns_roce_hem_table *table,
4141 				 int tag, u32 step_idx)
4142 {
4143 	struct hns_roce_cmd_mailbox *mailbox;
4144 	struct device *dev = hr_dev->dev;
4145 	u8 cmd = 0xff;
4146 	int ret;
4147 
4148 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4149 		return 0;
4150 
4151 	switch (table->type) {
4152 	case HEM_TYPE_QPC:
4153 		cmd = HNS_ROCE_CMD_DESTROY_QPC_BT0;
4154 		break;
4155 	case HEM_TYPE_MTPT:
4156 		cmd = HNS_ROCE_CMD_DESTROY_MPT_BT0;
4157 		break;
4158 	case HEM_TYPE_CQC:
4159 		cmd = HNS_ROCE_CMD_DESTROY_CQC_BT0;
4160 		break;
4161 	case HEM_TYPE_SRQC:
4162 		cmd = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
4163 		break;
4164 	case HEM_TYPE_SCCC:
4165 	case HEM_TYPE_QPC_TIMER:
4166 	case HEM_TYPE_CQC_TIMER:
4167 	case HEM_TYPE_GMV:
4168 		return 0;
4169 	default:
4170 		dev_warn(dev, "table %u not to be destroyed by mailbox!\n",
4171 			 table->type);
4172 		return 0;
4173 	}
4174 
4175 	cmd += step_idx;
4176 
4177 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4178 	if (IS_ERR(mailbox))
4179 		return PTR_ERR(mailbox);
4180 
4181 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, cmd, tag);
4182 
4183 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4184 	return ret;
4185 }
4186 
4187 static int hns_roce_v2_qp_modify(struct hns_roce_dev *hr_dev,
4188 				 struct hns_roce_v2_qp_context *context,
4189 				 struct hns_roce_v2_qp_context *qpc_mask,
4190 				 struct hns_roce_qp *hr_qp)
4191 {
4192 	struct hns_roce_cmd_mailbox *mailbox;
4193 	int qpc_size;
4194 	int ret;
4195 
4196 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4197 	if (IS_ERR(mailbox))
4198 		return PTR_ERR(mailbox);
4199 
4200 	/* The qpc size of HIP08 is only 256B, which is half of HIP09 */
4201 	qpc_size = hr_dev->caps.qpc_sz;
4202 	memcpy(mailbox->buf, context, qpc_size);
4203 	memcpy(mailbox->buf + qpc_size, qpc_mask, qpc_size);
4204 
4205 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
4206 				HNS_ROCE_CMD_MODIFY_QPC, hr_qp->qpn);
4207 
4208 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4209 
4210 	return ret;
4211 }
4212 
4213 static void set_access_flags(struct hns_roce_qp *hr_qp,
4214 			     struct hns_roce_v2_qp_context *context,
4215 			     struct hns_roce_v2_qp_context *qpc_mask,
4216 			     const struct ib_qp_attr *attr, int attr_mask)
4217 {
4218 	u8 dest_rd_atomic;
4219 	u32 access_flags;
4220 
4221 	dest_rd_atomic = (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) ?
4222 			 attr->max_dest_rd_atomic : hr_qp->resp_depth;
4223 
4224 	access_flags = (attr_mask & IB_QP_ACCESS_FLAGS) ?
4225 		       attr->qp_access_flags : hr_qp->atomic_rd_en;
4226 
4227 	if (!dest_rd_atomic)
4228 		access_flags &= IB_ACCESS_REMOTE_WRITE;
4229 
4230 	hr_reg_write_bool(context, QPC_RRE,
4231 			  access_flags & IB_ACCESS_REMOTE_READ);
4232 	hr_reg_clear(qpc_mask, QPC_RRE);
4233 
4234 	hr_reg_write_bool(context, QPC_RWE,
4235 			  access_flags & IB_ACCESS_REMOTE_WRITE);
4236 	hr_reg_clear(qpc_mask, QPC_RWE);
4237 
4238 	hr_reg_write_bool(context, QPC_ATE,
4239 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4240 	hr_reg_clear(qpc_mask, QPC_ATE);
4241 	hr_reg_write_bool(context, QPC_EXT_ATE,
4242 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4243 	hr_reg_clear(qpc_mask, QPC_EXT_ATE);
4244 }
4245 
4246 static void set_qpc_wqe_cnt(struct hns_roce_qp *hr_qp,
4247 			    struct hns_roce_v2_qp_context *context)
4248 {
4249 	hr_reg_write(context, QPC_SGE_SHIFT,
4250 		     to_hr_hem_entries_shift(hr_qp->sge.sge_cnt,
4251 					     hr_qp->sge.sge_shift));
4252 
4253 	hr_reg_write(context, QPC_SQ_SHIFT, ilog2(hr_qp->sq.wqe_cnt));
4254 
4255 	hr_reg_write(context, QPC_RQ_SHIFT, ilog2(hr_qp->rq.wqe_cnt));
4256 }
4257 
4258 static inline int get_cqn(struct ib_cq *ib_cq)
4259 {
4260 	return ib_cq ? to_hr_cq(ib_cq)->cqn : 0;
4261 }
4262 
4263 static inline int get_pdn(struct ib_pd *ib_pd)
4264 {
4265 	return ib_pd ? to_hr_pd(ib_pd)->pdn : 0;
4266 }
4267 
4268 static void modify_qp_reset_to_init(struct ib_qp *ibqp,
4269 				    struct hns_roce_v2_qp_context *context,
4270 				    struct hns_roce_v2_qp_context *qpc_mask)
4271 {
4272 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4273 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4274 
4275 	/*
4276 	 * In v2 engine, software pass context and context mask to hardware
4277 	 * when modifying qp. If software need modify some fields in context,
4278 	 * we should set all bits of the relevant fields in context mask to
4279 	 * 0 at the same time, else set them to 0x1.
4280 	 */
4281 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4282 
4283 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4284 
4285 	hr_reg_write(context, QPC_RQWS, ilog2(hr_qp->rq.max_gs));
4286 
4287 	set_qpc_wqe_cnt(hr_qp, context);
4288 
4289 	/* No VLAN need to set 0xFFF */
4290 	hr_reg_write(context, QPC_VLAN_ID, 0xfff);
4291 
4292 	if (ibqp->qp_type == IB_QPT_XRC_TGT) {
4293 		context->qkey_xrcd = cpu_to_le32(hr_qp->xrcdn);
4294 
4295 		hr_reg_enable(context, QPC_XRC_QP_TYPE);
4296 	}
4297 
4298 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
4299 		hr_reg_enable(context, QPC_RQ_RECORD_EN);
4300 
4301 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
4302 		hr_reg_enable(context, QPC_OWNER_MODE);
4303 
4304 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_L,
4305 		     lower_32_bits(hr_qp->rdb.dma) >> 1);
4306 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
4307 		     upper_32_bits(hr_qp->rdb.dma));
4308 
4309 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4310 
4311 	if (ibqp->srq) {
4312 		hr_reg_enable(context, QPC_SRQ_EN);
4313 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4314 	}
4315 
4316 	hr_reg_enable(context, QPC_FRE);
4317 
4318 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4319 
4320 	if (hr_dev->caps.qpc_sz < HNS_ROCE_V3_QPC_SZ)
4321 		return;
4322 
4323 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
4324 		hr_reg_enable(&context->ext, QPCEX_STASH);
4325 }
4326 
4327 static void modify_qp_init_to_init(struct ib_qp *ibqp,
4328 				   struct hns_roce_v2_qp_context *context,
4329 				   struct hns_roce_v2_qp_context *qpc_mask)
4330 {
4331 	/*
4332 	 * In v2 engine, software pass context and context mask to hardware
4333 	 * when modifying qp. If software need modify some fields in context,
4334 	 * we should set all bits of the relevant fields in context mask to
4335 	 * 0 at the same time, else set them to 0x1.
4336 	 */
4337 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4338 	hr_reg_clear(qpc_mask, QPC_TST);
4339 
4340 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4341 	hr_reg_clear(qpc_mask, QPC_PD);
4342 
4343 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4344 	hr_reg_clear(qpc_mask, QPC_RX_CQN);
4345 
4346 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4347 	hr_reg_clear(qpc_mask, QPC_TX_CQN);
4348 
4349 	if (ibqp->srq) {
4350 		hr_reg_enable(context, QPC_SRQ_EN);
4351 		hr_reg_clear(qpc_mask, QPC_SRQ_EN);
4352 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4353 		hr_reg_clear(qpc_mask, QPC_SRQN);
4354 	}
4355 }
4356 
4357 static int config_qp_rq_buf(struct hns_roce_dev *hr_dev,
4358 			    struct hns_roce_qp *hr_qp,
4359 			    struct hns_roce_v2_qp_context *context,
4360 			    struct hns_roce_v2_qp_context *qpc_mask)
4361 {
4362 	u64 mtts[MTT_MIN_COUNT] = { 0 };
4363 	u64 wqe_sge_ba;
4364 	int ret;
4365 
4366 	/* Search qp buf's mtts */
4367 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->rq.offset, mtts,
4368 				MTT_MIN_COUNT);
4369 	if (hr_qp->rq.wqe_cnt && ret) {
4370 		ibdev_err(&hr_dev->ib_dev,
4371 			  "failed to find QP(0x%lx) RQ WQE buf, ret = %d.\n",
4372 			  hr_qp->qpn, ret);
4373 		return ret;
4374 	}
4375 
4376 	wqe_sge_ba = hns_roce_get_mtr_ba(&hr_qp->mtr);
4377 
4378 	context->wqe_sge_ba = cpu_to_le32(wqe_sge_ba >> 3);
4379 	qpc_mask->wqe_sge_ba = 0;
4380 
4381 	/*
4382 	 * In v2 engine, software pass context and context mask to hardware
4383 	 * when modifying qp. If software need modify some fields in context,
4384 	 * we should set all bits of the relevant fields in context mask to
4385 	 * 0 at the same time, else set them to 0x1.
4386 	 */
4387 	hr_reg_write(context, QPC_WQE_SGE_BA_H, wqe_sge_ba >> (32 + 3));
4388 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_H);
4389 
4390 	hr_reg_write(context, QPC_SQ_HOP_NUM,
4391 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sq_hop_num,
4392 				      hr_qp->sq.wqe_cnt));
4393 	hr_reg_clear(qpc_mask, QPC_SQ_HOP_NUM);
4394 
4395 	hr_reg_write(context, QPC_SGE_HOP_NUM,
4396 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sge_hop_num,
4397 				      hr_qp->sge.sge_cnt));
4398 	hr_reg_clear(qpc_mask, QPC_SGE_HOP_NUM);
4399 
4400 	hr_reg_write(context, QPC_RQ_HOP_NUM,
4401 		     to_hr_hem_hopnum(hr_dev->caps.wqe_rq_hop_num,
4402 				      hr_qp->rq.wqe_cnt));
4403 
4404 	hr_reg_clear(qpc_mask, QPC_RQ_HOP_NUM);
4405 
4406 	hr_reg_write(context, QPC_WQE_SGE_BA_PG_SZ,
4407 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.ba_pg_shift));
4408 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_PG_SZ);
4409 
4410 	hr_reg_write(context, QPC_WQE_SGE_BUF_PG_SZ,
4411 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.buf_pg_shift));
4412 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BUF_PG_SZ);
4413 
4414 	context->rq_cur_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[0]));
4415 	qpc_mask->rq_cur_blk_addr = 0;
4416 
4417 	hr_reg_write(context, QPC_RQ_CUR_BLK_ADDR_H,
4418 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
4419 	hr_reg_clear(qpc_mask, QPC_RQ_CUR_BLK_ADDR_H);
4420 
4421 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4422 		context->rq_nxt_blk_addr =
4423 				cpu_to_le32(to_hr_hw_page_addr(mtts[1]));
4424 		qpc_mask->rq_nxt_blk_addr = 0;
4425 		hr_reg_write(context, QPC_RQ_NXT_BLK_ADDR_H,
4426 			     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
4427 		hr_reg_clear(qpc_mask, QPC_RQ_NXT_BLK_ADDR_H);
4428 	}
4429 
4430 	return 0;
4431 }
4432 
4433 static int config_qp_sq_buf(struct hns_roce_dev *hr_dev,
4434 			    struct hns_roce_qp *hr_qp,
4435 			    struct hns_roce_v2_qp_context *context,
4436 			    struct hns_roce_v2_qp_context *qpc_mask)
4437 {
4438 	struct ib_device *ibdev = &hr_dev->ib_dev;
4439 	u64 sge_cur_blk = 0;
4440 	u64 sq_cur_blk = 0;
4441 	int ret;
4442 
4443 	/* search qp buf's mtts */
4444 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->sq.offset,
4445 				&sq_cur_blk, 1);
4446 	if (ret) {
4447 		ibdev_err(ibdev, "failed to find QP(0x%lx) SQ WQE buf, ret = %d.\n",
4448 			  hr_qp->qpn, ret);
4449 		return ret;
4450 	}
4451 	if (hr_qp->sge.sge_cnt > 0) {
4452 		ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr,
4453 					hr_qp->sge.offset, &sge_cur_blk, 1);
4454 		if (ret) {
4455 			ibdev_err(ibdev, "failed to find QP(0x%lx) SGE buf, ret = %d.\n",
4456 				  hr_qp->qpn, ret);
4457 			return ret;
4458 		}
4459 	}
4460 
4461 	/*
4462 	 * In v2 engine, software pass context and context mask to hardware
4463 	 * when modifying qp. If software need modify some fields in context,
4464 	 * we should set all bits of the relevant fields in context mask to
4465 	 * 0 at the same time, else set them to 0x1.
4466 	 */
4467 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_L,
4468 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4469 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_H,
4470 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4471 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_L);
4472 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_H);
4473 
4474 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_L,
4475 		     lower_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4476 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_H,
4477 		     upper_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4478 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_L);
4479 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_H);
4480 
4481 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_L,
4482 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4483 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_H,
4484 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4485 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_L);
4486 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_H);
4487 
4488 	return 0;
4489 }
4490 
4491 static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
4492 				  const struct ib_qp_attr *attr)
4493 {
4494 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_UD)
4495 		return IB_MTU_4096;
4496 
4497 	return attr->path_mtu;
4498 }
4499 
4500 static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
4501 				 const struct ib_qp_attr *attr, int attr_mask,
4502 				 struct hns_roce_v2_qp_context *context,
4503 				 struct hns_roce_v2_qp_context *qpc_mask,
4504 				 struct ib_udata *udata)
4505 {
4506 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
4507 					  struct hns_roce_ucontext, ibucontext);
4508 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4509 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4510 	struct ib_device *ibdev = &hr_dev->ib_dev;
4511 	dma_addr_t trrl_ba;
4512 	dma_addr_t irrl_ba;
4513 	enum ib_mtu ib_mtu;
4514 	const u8 *smac;
4515 	u8 lp_pktn_ini;
4516 	u64 *mtts;
4517 	u8 *dmac;
4518 	u32 port;
4519 	int mtu;
4520 	int ret;
4521 
4522 	ret = config_qp_rq_buf(hr_dev, hr_qp, context, qpc_mask);
4523 	if (ret) {
4524 		ibdev_err(ibdev, "failed to config rq buf, ret = %d.\n", ret);
4525 		return ret;
4526 	}
4527 
4528 	/* Search IRRL's mtts */
4529 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.irrl_table,
4530 				   hr_qp->qpn, &irrl_ba);
4531 	if (!mtts) {
4532 		ibdev_err(ibdev, "failed to find qp irrl_table.\n");
4533 		return -EINVAL;
4534 	}
4535 
4536 	/* Search TRRL's mtts */
4537 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.trrl_table,
4538 				   hr_qp->qpn, &trrl_ba);
4539 	if (!mtts) {
4540 		ibdev_err(ibdev, "failed to find qp trrl_table.\n");
4541 		return -EINVAL;
4542 	}
4543 
4544 	if (attr_mask & IB_QP_ALT_PATH) {
4545 		ibdev_err(ibdev, "INIT2RTR attr_mask (0x%x) error.\n",
4546 			  attr_mask);
4547 		return -EINVAL;
4548 	}
4549 
4550 	hr_reg_write(context, QPC_TRRL_BA_L, trrl_ba >> QPC_TRRL_BA_L_S);
4551 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_L);
4552 	context->trrl_ba = cpu_to_le32(trrl_ba >> QPC_TRRL_BA_M_S);
4553 	qpc_mask->trrl_ba = 0;
4554 	hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> QPC_TRRL_BA_H_S);
4555 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
4556 
4557 	context->irrl_ba = cpu_to_le32(irrl_ba >> QPC_IRRL_BA_L_S);
4558 	qpc_mask->irrl_ba = 0;
4559 	hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> QPC_IRRL_BA_H_S);
4560 	hr_reg_clear(qpc_mask, QPC_IRRL_BA_H);
4561 
4562 	hr_reg_enable(context, QPC_RMT_E2E);
4563 	hr_reg_clear(qpc_mask, QPC_RMT_E2E);
4564 
4565 	hr_reg_write(context, QPC_SIG_TYPE, hr_qp->sq_signal_bits);
4566 	hr_reg_clear(qpc_mask, QPC_SIG_TYPE);
4567 
4568 	port = (attr_mask & IB_QP_PORT) ? (attr->port_num - 1) : hr_qp->port;
4569 
4570 	smac = (const u8 *)hr_dev->dev_addr[port];
4571 	dmac = (u8 *)attr->ah_attr.roce.dmac;
4572 	/* when dmac equals smac or loop_idc is 1, it should loopback */
4573 	if (ether_addr_equal_unaligned(dmac, smac) ||
4574 	    hr_dev->loop_idc == 0x1) {
4575 		hr_reg_write(context, QPC_LBI, hr_dev->loop_idc);
4576 		hr_reg_clear(qpc_mask, QPC_LBI);
4577 	}
4578 
4579 	if (attr_mask & IB_QP_DEST_QPN) {
4580 		hr_reg_write(context, QPC_DQPN, attr->dest_qp_num);
4581 		hr_reg_clear(qpc_mask, QPC_DQPN);
4582 	}
4583 
4584 	memcpy(&context->dmac, dmac, sizeof(u32));
4585 	hr_reg_write(context, QPC_DMAC_H, *((u16 *)(&dmac[4])));
4586 	qpc_mask->dmac = 0;
4587 	hr_reg_clear(qpc_mask, QPC_DMAC_H);
4588 
4589 	ib_mtu = get_mtu(ibqp, attr);
4590 	hr_qp->path_mtu = ib_mtu;
4591 
4592 	mtu = ib_mtu_enum_to_int(ib_mtu);
4593 	if (WARN_ON(mtu <= 0))
4594 		return -EINVAL;
4595 #define MIN_LP_MSG_LEN 1024
4596 	/* mtu * (2 ^ lp_pktn_ini) should be in the range of 1024 to mtu */
4597 	lp_pktn_ini = ilog2(max(mtu, MIN_LP_MSG_LEN) / mtu);
4598 
4599 	if (attr_mask & IB_QP_PATH_MTU) {
4600 		hr_reg_write(context, QPC_MTU, ib_mtu);
4601 		hr_reg_clear(qpc_mask, QPC_MTU);
4602 	}
4603 
4604 	hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
4605 	hr_reg_clear(qpc_mask, QPC_LP_PKTN_INI);
4606 
4607 	/* ACK_REQ_FREQ should be larger than or equal to LP_PKTN_INI */
4608 	hr_reg_write(context, QPC_ACK_REQ_FREQ, lp_pktn_ini);
4609 	hr_reg_clear(qpc_mask, QPC_ACK_REQ_FREQ);
4610 
4611 	hr_reg_clear(qpc_mask, QPC_RX_REQ_PSN_ERR);
4612 	hr_reg_clear(qpc_mask, QPC_RX_REQ_MSN);
4613 	hr_reg_clear(qpc_mask, QPC_RX_REQ_LAST_OPTYPE);
4614 
4615 	context->rq_rnr_timer = 0;
4616 	qpc_mask->rq_rnr_timer = 0;
4617 
4618 	hr_reg_clear(qpc_mask, QPC_TRRL_HEAD_MAX);
4619 	hr_reg_clear(qpc_mask, QPC_TRRL_TAIL_MAX);
4620 
4621 #define MAX_LP_SGEN 3
4622 	/* rocee send 2^lp_sgen_ini segs every time */
4623 	hr_reg_write(context, QPC_LP_SGEN_INI, MAX_LP_SGEN);
4624 	hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
4625 
4626 	if (udata && ibqp->qp_type == IB_QPT_RC &&
4627 	    (uctx->config & HNS_ROCE_RQ_INLINE_FLAGS)) {
4628 		hr_reg_write_bool(context, QPC_RQIE,
4629 				  hr_dev->caps.flags &
4630 				  HNS_ROCE_CAP_FLAG_RQ_INLINE);
4631 		hr_reg_clear(qpc_mask, QPC_RQIE);
4632 	}
4633 
4634 	if (udata &&
4635 	    (ibqp->qp_type == IB_QPT_RC || ibqp->qp_type == IB_QPT_XRC_TGT) &&
4636 	    (uctx->config & HNS_ROCE_CQE_INLINE_FLAGS)) {
4637 		hr_reg_write_bool(context, QPC_CQEIE,
4638 				  hr_dev->caps.flags &
4639 				  HNS_ROCE_CAP_FLAG_CQE_INLINE);
4640 		hr_reg_clear(qpc_mask, QPC_CQEIE);
4641 
4642 		hr_reg_write(context, QPC_CQEIS, 0);
4643 		hr_reg_clear(qpc_mask, QPC_CQEIS);
4644 	}
4645 
4646 	return 0;
4647 }
4648 
4649 static int modify_qp_rtr_to_rts(struct ib_qp *ibqp, int attr_mask,
4650 				struct hns_roce_v2_qp_context *context,
4651 				struct hns_roce_v2_qp_context *qpc_mask)
4652 {
4653 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4654 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4655 	struct ib_device *ibdev = &hr_dev->ib_dev;
4656 	int ret;
4657 
4658 	/* Not support alternate path and path migration */
4659 	if (attr_mask & (IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE)) {
4660 		ibdev_err(ibdev, "RTR2RTS attr_mask (0x%x)error\n", attr_mask);
4661 		return -EINVAL;
4662 	}
4663 
4664 	ret = config_qp_sq_buf(hr_dev, hr_qp, context, qpc_mask);
4665 	if (ret) {
4666 		ibdev_err(ibdev, "failed to config sq buf, ret = %d.\n", ret);
4667 		return ret;
4668 	}
4669 
4670 	/*
4671 	 * Set some fields in context to zero, Because the default values
4672 	 * of all fields in context are zero, we need not set them to 0 again.
4673 	 * but we should set the relevant fields of context mask to 0.
4674 	 */
4675 	hr_reg_clear(qpc_mask, QPC_IRRL_SGE_IDX);
4676 
4677 	hr_reg_clear(qpc_mask, QPC_RX_ACK_MSN);
4678 
4679 	hr_reg_clear(qpc_mask, QPC_ACK_LAST_OPTYPE);
4680 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN_VLD);
4681 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN);
4682 
4683 	hr_reg_clear(qpc_mask, QPC_IRRL_TAIL_REAL);
4684 
4685 	hr_reg_clear(qpc_mask, QPC_RETRY_MSG_MSN);
4686 
4687 	hr_reg_clear(qpc_mask, QPC_RNR_RETRY_FLAG);
4688 
4689 	hr_reg_clear(qpc_mask, QPC_CHECK_FLG);
4690 
4691 	hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
4692 
4693 	return 0;
4694 }
4695 
4696 static int alloc_dip_entry(struct xarray *dip_xa, u32 qpn)
4697 {
4698 	struct hns_roce_dip *hr_dip;
4699 	int ret;
4700 
4701 	hr_dip = xa_load(dip_xa, qpn);
4702 	if (hr_dip)
4703 		return 0;
4704 
4705 	hr_dip = kzalloc(sizeof(*hr_dip), GFP_KERNEL);
4706 	if (!hr_dip)
4707 		return -ENOMEM;
4708 
4709 	ret = xa_err(xa_store(dip_xa, qpn, hr_dip, GFP_KERNEL));
4710 	if (ret)
4711 		kfree(hr_dip);
4712 
4713 	return ret;
4714 }
4715 
4716 static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4717 			   u32 *dip_idx)
4718 {
4719 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4720 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4721 	struct xarray *dip_xa = &hr_dev->qp_table.dip_xa;
4722 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4723 	struct hns_roce_dip *hr_dip;
4724 	unsigned long idx;
4725 	int ret = 0;
4726 
4727 	ret = alloc_dip_entry(dip_xa, ibqp->qp_num);
4728 	if (ret)
4729 		return ret;
4730 
4731 	xa_lock(dip_xa);
4732 
4733 	xa_for_each(dip_xa, idx, hr_dip) {
4734 		if (hr_dip->qp_cnt &&
4735 		    !memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) {
4736 			*dip_idx = hr_dip->dip_idx;
4737 			hr_dip->qp_cnt++;
4738 			hr_qp->dip = hr_dip;
4739 			goto out;
4740 		}
4741 	}
4742 
4743 	/* If no dgid is found, a new dip and a mapping between dgid and
4744 	 * dip_idx will be created.
4745 	 */
4746 	xa_for_each(dip_xa, idx, hr_dip) {
4747 		if (hr_dip->qp_cnt)
4748 			continue;
4749 
4750 		*dip_idx = idx;
4751 		memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4752 		hr_dip->dip_idx = idx;
4753 		hr_dip->qp_cnt++;
4754 		hr_qp->dip = hr_dip;
4755 		break;
4756 	}
4757 
4758 	/* This should never happen. */
4759 	if (WARN_ON_ONCE(!hr_qp->dip))
4760 		ret = -ENOSPC;
4761 
4762 out:
4763 	xa_unlock(dip_xa);
4764 	return ret;
4765 }
4766 
4767 enum {
4768 	CONG_DCQCN,
4769 	CONG_WINDOW,
4770 };
4771 
4772 enum {
4773 	UNSUPPORT_CONG_LEVEL,
4774 	SUPPORT_CONG_LEVEL,
4775 };
4776 
4777 enum {
4778 	CONG_LDCP,
4779 	CONG_HC3,
4780 };
4781 
4782 enum {
4783 	DIP_INVALID,
4784 	DIP_VALID,
4785 };
4786 
4787 enum {
4788 	WND_LIMIT,
4789 	WND_UNLIMIT,
4790 };
4791 
4792 static int check_cong_type(struct ib_qp *ibqp,
4793 			   struct hns_roce_congestion_algorithm *cong_alg)
4794 {
4795 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4796 
4797 	/* different congestion types match different configurations */
4798 	switch (hr_qp->cong_type) {
4799 	case CONG_TYPE_DCQCN:
4800 		cong_alg->alg_sel = CONG_DCQCN;
4801 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4802 		cong_alg->dip_vld = DIP_INVALID;
4803 		cong_alg->wnd_mode_sel = WND_LIMIT;
4804 		break;
4805 	case CONG_TYPE_LDCP:
4806 		cong_alg->alg_sel = CONG_WINDOW;
4807 		cong_alg->alg_sub_sel = CONG_LDCP;
4808 		cong_alg->dip_vld = DIP_INVALID;
4809 		cong_alg->wnd_mode_sel = WND_UNLIMIT;
4810 		break;
4811 	case CONG_TYPE_HC3:
4812 		cong_alg->alg_sel = CONG_WINDOW;
4813 		cong_alg->alg_sub_sel = CONG_HC3;
4814 		cong_alg->dip_vld = DIP_INVALID;
4815 		cong_alg->wnd_mode_sel = WND_LIMIT;
4816 		break;
4817 	case CONG_TYPE_DIP:
4818 		cong_alg->alg_sel = CONG_DCQCN;
4819 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4820 		cong_alg->dip_vld = DIP_VALID;
4821 		cong_alg->wnd_mode_sel = WND_LIMIT;
4822 		break;
4823 	default:
4824 		hr_qp->cong_type = CONG_TYPE_DCQCN;
4825 		cong_alg->alg_sel = CONG_DCQCN;
4826 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4827 		cong_alg->dip_vld = DIP_INVALID;
4828 		cong_alg->wnd_mode_sel = WND_LIMIT;
4829 		break;
4830 	}
4831 
4832 	return 0;
4833 }
4834 
4835 static int fill_cong_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4836 			   struct hns_roce_v2_qp_context *context,
4837 			   struct hns_roce_v2_qp_context *qpc_mask)
4838 {
4839 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4840 	struct hns_roce_congestion_algorithm cong_field;
4841 	struct ib_device *ibdev = ibqp->device;
4842 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
4843 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4844 	u32 dip_idx = 0;
4845 	int ret;
4846 
4847 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ||
4848 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE)
4849 		return 0;
4850 
4851 	ret = check_cong_type(ibqp, &cong_field);
4852 	if (ret)
4853 		return ret;
4854 
4855 	hr_reg_write(context, QPC_CONG_ALGO_TMPL_ID, hr_dev->cong_algo_tmpl_id +
4856 		     hr_qp->cong_type * HNS_ROCE_CONG_SIZE);
4857 	hr_reg_clear(qpc_mask, QPC_CONG_ALGO_TMPL_ID);
4858 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SEL, cong_field.alg_sel);
4859 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SEL);
4860 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SUB_SEL,
4861 		     cong_field.alg_sub_sel);
4862 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SUB_SEL);
4863 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX_VLD, cong_field.dip_vld);
4864 	hr_reg_clear(&qpc_mask->ext, QPCEX_DIP_CTX_IDX_VLD);
4865 	hr_reg_write(&context->ext, QPCEX_SQ_RQ_NOT_FORBID_EN,
4866 		     cong_field.wnd_mode_sel);
4867 	hr_reg_clear(&qpc_mask->ext, QPCEX_SQ_RQ_NOT_FORBID_EN);
4868 
4869 	/* if dip is disabled, there is no need to set dip idx */
4870 	if (cong_field.dip_vld == 0)
4871 		return 0;
4872 
4873 	ret = get_dip_ctx_idx(ibqp, attr, &dip_idx);
4874 	if (ret) {
4875 		ibdev_err(ibdev, "failed to fill cong field, ret = %d.\n", ret);
4876 		return ret;
4877 	}
4878 
4879 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX, dip_idx);
4880 	hr_reg_write(&qpc_mask->ext, QPCEX_DIP_CTX_IDX, 0);
4881 
4882 	return 0;
4883 }
4884 
4885 static int hns_roce_hw_v2_get_dscp(struct hns_roce_dev *hr_dev, u8 dscp,
4886 				   u8 *tc_mode, u8 *priority)
4887 {
4888 	struct hns_roce_v2_priv *priv = hr_dev->priv;
4889 	struct hnae3_handle *handle = priv->handle;
4890 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
4891 
4892 	if (!ops->get_dscp_prio)
4893 		return -EOPNOTSUPP;
4894 
4895 	return ops->get_dscp_prio(handle, dscp, tc_mode, priority);
4896 }
4897 
4898 bool check_sl_valid(struct hns_roce_dev *hr_dev, u8 sl)
4899 {
4900 	u32 max_sl;
4901 
4902 	max_sl = min_t(u32, MAX_SERVICE_LEVEL, hr_dev->caps.sl_num - 1);
4903 	if (unlikely(sl > max_sl)) {
4904 		ibdev_err_ratelimited(&hr_dev->ib_dev,
4905 				      "failed to set SL(%u). Shouldn't be larger than %u.\n",
4906 				      sl, max_sl);
4907 		return false;
4908 	}
4909 
4910 	return true;
4911 }
4912 
4913 static int hns_roce_set_sl(struct ib_qp *ibqp,
4914 			   const struct ib_qp_attr *attr,
4915 			   struct hns_roce_v2_qp_context *context,
4916 			   struct hns_roce_v2_qp_context *qpc_mask)
4917 {
4918 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4919 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4920 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4921 	struct ib_device *ibdev = &hr_dev->ib_dev;
4922 	int ret;
4923 
4924 	ret = hns_roce_hw_v2_get_dscp(hr_dev, get_tclass(&attr->ah_attr.grh),
4925 				      &hr_qp->tc_mode, &hr_qp->priority);
4926 	if (ret && ret != -EOPNOTSUPP &&
4927 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4928 		ibdev_err_ratelimited(ibdev,
4929 				      "failed to get dscp, ret = %d.\n", ret);
4930 		return ret;
4931 	}
4932 
4933 	if (hr_qp->tc_mode == HNAE3_TC_MAP_MODE_DSCP &&
4934 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4935 		hr_qp->sl = hr_qp->priority;
4936 	else
4937 		hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4938 
4939 	if (!check_sl_valid(hr_dev, hr_qp->sl))
4940 		return -EINVAL;
4941 
4942 	hr_reg_write(context, QPC_SL, hr_qp->sl);
4943 	hr_reg_clear(qpc_mask, QPC_SL);
4944 
4945 	return 0;
4946 }
4947 
4948 static int hns_roce_v2_set_path(struct ib_qp *ibqp,
4949 				const struct ib_qp_attr *attr,
4950 				int attr_mask,
4951 				struct hns_roce_v2_qp_context *context,
4952 				struct hns_roce_v2_qp_context *qpc_mask)
4953 {
4954 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4955 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4956 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4957 	struct ib_device *ibdev = &hr_dev->ib_dev;
4958 	const struct ib_gid_attr *gid_attr = NULL;
4959 	u8 sl = rdma_ah_get_sl(&attr->ah_attr);
4960 	int is_roce_protocol;
4961 	u16 vlan_id = 0xffff;
4962 	bool is_udp = false;
4963 	u8 ib_port;
4964 	u8 hr_port;
4965 	int ret;
4966 
4967 	/*
4968 	 * If free_mr_en of qp is set, it means that this qp comes from
4969 	 * free mr. This qp will perform the loopback operation.
4970 	 * In the loopback scenario, only sl needs to be set.
4971 	 */
4972 	if (hr_qp->free_mr_en) {
4973 		if (!check_sl_valid(hr_dev, sl))
4974 			return -EINVAL;
4975 		hr_reg_write(context, QPC_SL, sl);
4976 		hr_reg_clear(qpc_mask, QPC_SL);
4977 		hr_qp->sl = sl;
4978 		return 0;
4979 	}
4980 
4981 	ib_port = (attr_mask & IB_QP_PORT) ? attr->port_num : hr_qp->port + 1;
4982 	hr_port = ib_port - 1;
4983 	is_roce_protocol = rdma_cap_eth_ah(&hr_dev->ib_dev, ib_port) &&
4984 			   rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
4985 
4986 	if (is_roce_protocol) {
4987 		gid_attr = attr->ah_attr.grh.sgid_attr;
4988 		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
4989 		if (ret)
4990 			return ret;
4991 
4992 		is_udp = (gid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
4993 	}
4994 
4995 	/* Only HIP08 needs to set the vlan_en bits in QPC */
4996 	if (vlan_id < VLAN_N_VID &&
4997 	    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4998 		hr_reg_enable(context, QPC_RQ_VLAN_EN);
4999 		hr_reg_clear(qpc_mask, QPC_RQ_VLAN_EN);
5000 		hr_reg_enable(context, QPC_SQ_VLAN_EN);
5001 		hr_reg_clear(qpc_mask, QPC_SQ_VLAN_EN);
5002 	}
5003 
5004 	hr_reg_write(context, QPC_VLAN_ID, vlan_id);
5005 	hr_reg_clear(qpc_mask, QPC_VLAN_ID);
5006 
5007 	if (grh->sgid_index >= hr_dev->caps.gid_table_len[hr_port]) {
5008 		ibdev_err(ibdev, "sgid_index(%u) too large. max is %d\n",
5009 			  grh->sgid_index, hr_dev->caps.gid_table_len[hr_port]);
5010 		return -EINVAL;
5011 	}
5012 
5013 	if (attr->ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE) {
5014 		ibdev_err(ibdev, "ah attr is not RDMA roce type\n");
5015 		return -EINVAL;
5016 	}
5017 
5018 	hr_reg_write(context, QPC_UDPSPN,
5019 		     is_udp ? rdma_get_udp_sport(grh->flow_label, ibqp->qp_num,
5020 						 attr->dest_qp_num) :
5021 				    0);
5022 
5023 	hr_reg_clear(qpc_mask, QPC_UDPSPN);
5024 
5025 	hr_reg_write(context, QPC_GMV_IDX, grh->sgid_index);
5026 
5027 	hr_reg_clear(qpc_mask, QPC_GMV_IDX);
5028 
5029 	hr_reg_write(context, QPC_HOPLIMIT, grh->hop_limit);
5030 	hr_reg_clear(qpc_mask, QPC_HOPLIMIT);
5031 
5032 	ret = fill_cong_field(ibqp, attr, context, qpc_mask);
5033 	if (ret)
5034 		return ret;
5035 
5036 	hr_reg_write(context, QPC_TC, get_tclass(&attr->ah_attr.grh));
5037 	hr_reg_clear(qpc_mask, QPC_TC);
5038 
5039 	hr_reg_write(context, QPC_FL, grh->flow_label);
5040 	hr_reg_clear(qpc_mask, QPC_FL);
5041 	memcpy(context->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
5042 	memset(qpc_mask->dgid, 0, sizeof(grh->dgid.raw));
5043 
5044 	return  hns_roce_set_sl(ibqp, attr, context, qpc_mask);
5045 }
5046 
5047 static bool check_qp_state(enum ib_qp_state cur_state,
5048 			   enum ib_qp_state new_state)
5049 {
5050 	static const bool sm[][IB_QPS_ERR + 1] = {
5051 		[IB_QPS_RESET] = { [IB_QPS_RESET] = true,
5052 				   [IB_QPS_INIT] = true },
5053 		[IB_QPS_INIT] = { [IB_QPS_RESET] = true,
5054 				  [IB_QPS_INIT] = true,
5055 				  [IB_QPS_RTR] = true,
5056 				  [IB_QPS_ERR] = true },
5057 		[IB_QPS_RTR] = { [IB_QPS_RESET] = true,
5058 				 [IB_QPS_RTS] = true,
5059 				 [IB_QPS_ERR] = true },
5060 		[IB_QPS_RTS] = { [IB_QPS_RESET] = true,
5061 				 [IB_QPS_RTS] = true,
5062 				 [IB_QPS_ERR] = true },
5063 		[IB_QPS_SQD] = {},
5064 		[IB_QPS_SQE] = {},
5065 		[IB_QPS_ERR] = { [IB_QPS_RESET] = true,
5066 				 [IB_QPS_ERR] = true }
5067 	};
5068 
5069 	return sm[cur_state][new_state];
5070 }
5071 
5072 static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
5073 				      const struct ib_qp_attr *attr,
5074 				      int attr_mask,
5075 				      enum ib_qp_state cur_state,
5076 				      enum ib_qp_state new_state,
5077 				      struct hns_roce_v2_qp_context *context,
5078 				      struct hns_roce_v2_qp_context *qpc_mask,
5079 				      struct ib_udata *udata)
5080 {
5081 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5082 	int ret = 0;
5083 
5084 	if (!check_qp_state(cur_state, new_state))
5085 		return -EINVAL;
5086 
5087 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
5088 		memset(qpc_mask, 0, hr_dev->caps.qpc_sz);
5089 		modify_qp_reset_to_init(ibqp, context, qpc_mask);
5090 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
5091 		modify_qp_init_to_init(ibqp, context, qpc_mask);
5092 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
5093 		ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
5094 					    qpc_mask, udata);
5095 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
5096 		ret = modify_qp_rtr_to_rts(ibqp, attr_mask, context, qpc_mask);
5097 	}
5098 
5099 	return ret;
5100 }
5101 
5102 static bool check_qp_timeout_cfg_range(struct hns_roce_dev *hr_dev, u8 *timeout)
5103 {
5104 #define QP_ACK_TIMEOUT_MAX_HIP08 20
5105 #define QP_ACK_TIMEOUT_MAX 31
5106 
5107 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5108 		if (*timeout > QP_ACK_TIMEOUT_MAX_HIP08) {
5109 			ibdev_warn(&hr_dev->ib_dev,
5110 				   "local ACK timeout shall be 0 to 20.\n");
5111 			return false;
5112 		}
5113 		*timeout += HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5114 	} else if (hr_dev->pci_dev->revision > PCI_REVISION_ID_HIP08) {
5115 		if (*timeout > QP_ACK_TIMEOUT_MAX) {
5116 			ibdev_warn(&hr_dev->ib_dev,
5117 				   "local ACK timeout shall be 0 to 31.\n");
5118 			return false;
5119 		}
5120 	}
5121 
5122 	return true;
5123 }
5124 
5125 static int hns_roce_v2_set_opt_fields(struct ib_qp *ibqp,
5126 				      const struct ib_qp_attr *attr,
5127 				      int attr_mask,
5128 				      struct hns_roce_v2_qp_context *context,
5129 				      struct hns_roce_v2_qp_context *qpc_mask)
5130 {
5131 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5132 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5133 	int ret = 0;
5134 	u8 timeout;
5135 
5136 	if (attr_mask & IB_QP_AV) {
5137 		ret = hns_roce_v2_set_path(ibqp, attr, attr_mask, context,
5138 					   qpc_mask);
5139 		if (ret)
5140 			return ret;
5141 	}
5142 
5143 	if (attr_mask & IB_QP_TIMEOUT) {
5144 		timeout = attr->timeout;
5145 		if (check_qp_timeout_cfg_range(hr_dev, &timeout)) {
5146 			hr_reg_write(context, QPC_AT, timeout);
5147 			hr_reg_clear(qpc_mask, QPC_AT);
5148 		}
5149 	}
5150 
5151 	if (attr_mask & IB_QP_RETRY_CNT) {
5152 		hr_reg_write(context, QPC_RETRY_NUM_INIT, attr->retry_cnt);
5153 		hr_reg_clear(qpc_mask, QPC_RETRY_NUM_INIT);
5154 
5155 		hr_reg_write(context, QPC_RETRY_CNT, attr->retry_cnt);
5156 		hr_reg_clear(qpc_mask, QPC_RETRY_CNT);
5157 	}
5158 
5159 	if (attr_mask & IB_QP_RNR_RETRY) {
5160 		hr_reg_write(context, QPC_RNR_NUM_INIT, attr->rnr_retry);
5161 		hr_reg_clear(qpc_mask, QPC_RNR_NUM_INIT);
5162 
5163 		hr_reg_write(context, QPC_RNR_CNT, attr->rnr_retry);
5164 		hr_reg_clear(qpc_mask, QPC_RNR_CNT);
5165 	}
5166 
5167 	if (attr_mask & IB_QP_SQ_PSN) {
5168 		hr_reg_write(context, QPC_SQ_CUR_PSN, attr->sq_psn);
5169 		hr_reg_clear(qpc_mask, QPC_SQ_CUR_PSN);
5170 
5171 		hr_reg_write(context, QPC_SQ_MAX_PSN, attr->sq_psn);
5172 		hr_reg_clear(qpc_mask, QPC_SQ_MAX_PSN);
5173 
5174 		hr_reg_write(context, QPC_RETRY_MSG_PSN_L, attr->sq_psn);
5175 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_L);
5176 
5177 		hr_reg_write(context, QPC_RETRY_MSG_PSN_H,
5178 			     attr->sq_psn >> RETRY_MSG_PSN_SHIFT);
5179 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_H);
5180 
5181 		hr_reg_write(context, QPC_RETRY_MSG_FPKT_PSN, attr->sq_psn);
5182 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_FPKT_PSN);
5183 
5184 		hr_reg_write(context, QPC_RX_ACK_EPSN, attr->sq_psn);
5185 		hr_reg_clear(qpc_mask, QPC_RX_ACK_EPSN);
5186 	}
5187 
5188 	if ((attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) &&
5189 	     attr->max_dest_rd_atomic) {
5190 		hr_reg_write(context, QPC_RR_MAX,
5191 			     fls(attr->max_dest_rd_atomic - 1));
5192 		hr_reg_clear(qpc_mask, QPC_RR_MAX);
5193 	}
5194 
5195 	if ((attr_mask & IB_QP_MAX_QP_RD_ATOMIC) && attr->max_rd_atomic) {
5196 		hr_reg_write(context, QPC_SR_MAX, fls(attr->max_rd_atomic - 1));
5197 		hr_reg_clear(qpc_mask, QPC_SR_MAX);
5198 	}
5199 
5200 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
5201 		set_access_flags(hr_qp, context, qpc_mask, attr, attr_mask);
5202 
5203 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
5204 		hr_reg_write(context, QPC_MIN_RNR_TIME,
5205 			    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ?
5206 			    HNS_ROCE_RNR_TIMER_10NS : attr->min_rnr_timer);
5207 		hr_reg_clear(qpc_mask, QPC_MIN_RNR_TIME);
5208 	}
5209 
5210 	if (attr_mask & IB_QP_RQ_PSN) {
5211 		hr_reg_write(context, QPC_RX_REQ_EPSN, attr->rq_psn);
5212 		hr_reg_clear(qpc_mask, QPC_RX_REQ_EPSN);
5213 
5214 		hr_reg_write(context, QPC_RAQ_PSN, attr->rq_psn - 1);
5215 		hr_reg_clear(qpc_mask, QPC_RAQ_PSN);
5216 	}
5217 
5218 	if (attr_mask & IB_QP_QKEY) {
5219 		context->qkey_xrcd = cpu_to_le32(attr->qkey);
5220 		qpc_mask->qkey_xrcd = 0;
5221 		hr_qp->qkey = attr->qkey;
5222 	}
5223 
5224 	return ret;
5225 }
5226 
5227 static void hns_roce_v2_record_opt_fields(struct ib_qp *ibqp,
5228 					  const struct ib_qp_attr *attr,
5229 					  int attr_mask)
5230 {
5231 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5232 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5233 
5234 	if (attr_mask & IB_QP_ACCESS_FLAGS)
5235 		hr_qp->atomic_rd_en = attr->qp_access_flags;
5236 
5237 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
5238 		hr_qp->resp_depth = attr->max_dest_rd_atomic;
5239 	if (attr_mask & IB_QP_PORT) {
5240 		hr_qp->port = attr->port_num - 1;
5241 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
5242 	}
5243 }
5244 
5245 static void clear_qp(struct hns_roce_qp *hr_qp)
5246 {
5247 	struct ib_qp *ibqp = &hr_qp->ibqp;
5248 
5249 	if (ibqp->send_cq)
5250 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq),
5251 				     hr_qp->qpn, NULL);
5252 
5253 	if (ibqp->recv_cq  && ibqp->recv_cq != ibqp->send_cq)
5254 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->recv_cq),
5255 				     hr_qp->qpn, ibqp->srq ?
5256 				     to_hr_srq(ibqp->srq) : NULL);
5257 
5258 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
5259 		*hr_qp->rdb.db_record = 0;
5260 
5261 	hr_qp->rq.head = 0;
5262 	hr_qp->rq.tail = 0;
5263 	hr_qp->sq.head = 0;
5264 	hr_qp->sq.tail = 0;
5265 	hr_qp->next_sge = 0;
5266 }
5267 
5268 static void v2_set_flushed_fields(struct ib_qp *ibqp,
5269 				  struct hns_roce_v2_qp_context *context,
5270 				  struct hns_roce_v2_qp_context *qpc_mask)
5271 {
5272 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5273 	unsigned long sq_flag = 0;
5274 	unsigned long rq_flag = 0;
5275 
5276 	if (ibqp->qp_type == IB_QPT_XRC_TGT)
5277 		return;
5278 
5279 	spin_lock_irqsave(&hr_qp->sq.lock, sq_flag);
5280 	hr_reg_write(context, QPC_SQ_PRODUCER_IDX, hr_qp->sq.head);
5281 	hr_reg_clear(qpc_mask, QPC_SQ_PRODUCER_IDX);
5282 	hr_qp->state = IB_QPS_ERR;
5283 	spin_unlock_irqrestore(&hr_qp->sq.lock, sq_flag);
5284 
5285 	if (ibqp->srq || ibqp->qp_type == IB_QPT_XRC_INI) /* no RQ */
5286 		return;
5287 
5288 	spin_lock_irqsave(&hr_qp->rq.lock, rq_flag);
5289 	hr_reg_write(context, QPC_RQ_PRODUCER_IDX, hr_qp->rq.head);
5290 	hr_reg_clear(qpc_mask, QPC_RQ_PRODUCER_IDX);
5291 	spin_unlock_irqrestore(&hr_qp->rq.lock, rq_flag);
5292 }
5293 
5294 static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
5295 				 const struct ib_qp_attr *attr,
5296 				 int attr_mask, enum ib_qp_state cur_state,
5297 				 enum ib_qp_state new_state, struct ib_udata *udata)
5298 {
5299 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5300 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5301 	struct hns_roce_v2_qp_context ctx[2];
5302 	struct hns_roce_v2_qp_context *context = ctx;
5303 	struct hns_roce_v2_qp_context *qpc_mask = ctx + 1;
5304 	struct ib_device *ibdev = &hr_dev->ib_dev;
5305 	int ret;
5306 
5307 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
5308 		return -EOPNOTSUPP;
5309 
5310 	/*
5311 	 * In v2 engine, software pass context and context mask to hardware
5312 	 * when modifying qp. If software need modify some fields in context,
5313 	 * we should set all bits of the relevant fields in context mask to
5314 	 * 0 at the same time, else set them to 0x1.
5315 	 */
5316 	memset(context, 0, hr_dev->caps.qpc_sz);
5317 	memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
5318 
5319 	ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
5320 					 new_state, context, qpc_mask, udata);
5321 	if (ret)
5322 		goto out;
5323 
5324 	/* When QP state is err, SQ and RQ WQE should be flushed */
5325 	if (new_state == IB_QPS_ERR)
5326 		v2_set_flushed_fields(ibqp, context, qpc_mask);
5327 
5328 	/* Configure the optional fields */
5329 	ret = hns_roce_v2_set_opt_fields(ibqp, attr, attr_mask, context,
5330 					 qpc_mask);
5331 	if (ret)
5332 		goto out;
5333 
5334 	hr_reg_write_bool(context, QPC_INV_CREDIT,
5335 			  to_hr_qp_type(hr_qp->ibqp.qp_type) == SERV_TYPE_XRC ||
5336 			  ibqp->srq);
5337 	hr_reg_clear(qpc_mask, QPC_INV_CREDIT);
5338 
5339 	/* Every status migrate must change state */
5340 	hr_reg_write(context, QPC_QP_ST, new_state);
5341 	hr_reg_clear(qpc_mask, QPC_QP_ST);
5342 
5343 	/* SW pass context to HW */
5344 	ret = hns_roce_v2_qp_modify(hr_dev, context, qpc_mask, hr_qp);
5345 	if (ret) {
5346 		ibdev_err_ratelimited(ibdev, "failed to modify QP, ret = %d.\n", ret);
5347 		goto out;
5348 	}
5349 
5350 	hr_qp->state = new_state;
5351 
5352 	hns_roce_v2_record_opt_fields(ibqp, attr, attr_mask);
5353 
5354 	if (new_state == IB_QPS_RESET && !ibqp->uobject)
5355 		clear_qp(hr_qp);
5356 
5357 out:
5358 	return ret;
5359 }
5360 
5361 static int to_ib_qp_st(enum hns_roce_v2_qp_state state)
5362 {
5363 	static const enum ib_qp_state map[] = {
5364 		[HNS_ROCE_QP_ST_RST] = IB_QPS_RESET,
5365 		[HNS_ROCE_QP_ST_INIT] = IB_QPS_INIT,
5366 		[HNS_ROCE_QP_ST_RTR] = IB_QPS_RTR,
5367 		[HNS_ROCE_QP_ST_RTS] = IB_QPS_RTS,
5368 		[HNS_ROCE_QP_ST_SQD] = IB_QPS_SQD,
5369 		[HNS_ROCE_QP_ST_SQER] = IB_QPS_SQE,
5370 		[HNS_ROCE_QP_ST_ERR] = IB_QPS_ERR,
5371 		[HNS_ROCE_QP_ST_SQ_DRAINING] = IB_QPS_SQD
5372 	};
5373 
5374 	return (state < ARRAY_SIZE(map)) ? map[state] : -1;
5375 }
5376 
5377 static int hns_roce_v2_query_qpc(struct hns_roce_dev *hr_dev, u32 qpn,
5378 				 void *buffer)
5379 {
5380 	struct hns_roce_cmd_mailbox *mailbox;
5381 	int ret;
5382 
5383 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5384 	if (IS_ERR(mailbox))
5385 		return PTR_ERR(mailbox);
5386 
5387 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_QPC,
5388 				qpn);
5389 	if (ret)
5390 		goto out;
5391 
5392 	memcpy(buffer, mailbox->buf, hr_dev->caps.qpc_sz);
5393 
5394 out:
5395 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5396 	return ret;
5397 }
5398 
5399 static int hns_roce_v2_query_srqc(struct hns_roce_dev *hr_dev, u32 srqn,
5400 				 void *buffer)
5401 {
5402 	struct hns_roce_srq_context *context;
5403 	struct hns_roce_cmd_mailbox *mailbox;
5404 	int ret;
5405 
5406 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5407 	if (IS_ERR(mailbox))
5408 		return PTR_ERR(mailbox);
5409 
5410 	context = mailbox->buf;
5411 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SRQC,
5412 				srqn);
5413 	if (ret)
5414 		goto out;
5415 
5416 	memcpy(buffer, context, sizeof(*context));
5417 
5418 out:
5419 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5420 	return ret;
5421 }
5422 
5423 static int hns_roce_v2_query_sccc(struct hns_roce_dev *hr_dev, u32 qpn,
5424 				  void *buffer)
5425 {
5426 	struct hns_roce_v2_scc_context *context;
5427 	struct hns_roce_cmd_mailbox *mailbox;
5428 	int ret;
5429 
5430 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5431 	if (IS_ERR(mailbox))
5432 		return PTR_ERR(mailbox);
5433 
5434 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SCCC,
5435 				qpn);
5436 	if (ret)
5437 		goto out;
5438 
5439 	context = mailbox->buf;
5440 	memcpy(buffer, context, sizeof(*context));
5441 
5442 out:
5443 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5444 	return ret;
5445 }
5446 
5447 static u8 get_qp_timeout_attr(struct hns_roce_dev *hr_dev,
5448 			      struct hns_roce_v2_qp_context *context)
5449 {
5450 	u8 timeout;
5451 
5452 	timeout = (u8)hr_reg_read(context, QPC_AT);
5453 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
5454 		timeout -= HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5455 
5456 	return timeout;
5457 }
5458 
5459 static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5460 				int qp_attr_mask,
5461 				struct ib_qp_init_attr *qp_init_attr)
5462 {
5463 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5464 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5465 	struct hns_roce_v2_qp_context context = {};
5466 	struct ib_device *ibdev = &hr_dev->ib_dev;
5467 	int tmp_qp_state;
5468 	int state;
5469 	int ret;
5470 
5471 	memset(qp_attr, 0, sizeof(*qp_attr));
5472 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5473 
5474 	mutex_lock(&hr_qp->mutex);
5475 
5476 	if (hr_qp->state == IB_QPS_RESET) {
5477 		qp_attr->qp_state = IB_QPS_RESET;
5478 		ret = 0;
5479 		goto done;
5480 	}
5481 
5482 	ret = hns_roce_v2_query_qpc(hr_dev, hr_qp->qpn, &context);
5483 	if (ret) {
5484 		ibdev_err_ratelimited(ibdev,
5485 				      "failed to query QPC, ret = %d.\n",
5486 				      ret);
5487 		ret = -EINVAL;
5488 		goto out;
5489 	}
5490 
5491 	state = hr_reg_read(&context, QPC_QP_ST);
5492 	tmp_qp_state = to_ib_qp_st((enum hns_roce_v2_qp_state)state);
5493 	if (tmp_qp_state == -1) {
5494 		ibdev_err_ratelimited(ibdev, "Illegal ib_qp_state\n");
5495 		ret = -EINVAL;
5496 		goto out;
5497 	}
5498 	hr_qp->state = (u8)tmp_qp_state;
5499 	qp_attr->qp_state = (enum ib_qp_state)hr_qp->state;
5500 	qp_attr->path_mtu = (enum ib_mtu)hr_reg_read(&context, QPC_MTU);
5501 	qp_attr->path_mig_state = IB_MIG_ARMED;
5502 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
5503 	if (hr_qp->ibqp.qp_type == IB_QPT_UD)
5504 		qp_attr->qkey = le32_to_cpu(context.qkey_xrcd);
5505 
5506 	qp_attr->rq_psn = hr_reg_read(&context, QPC_RX_REQ_EPSN);
5507 	qp_attr->sq_psn = (u32)hr_reg_read(&context, QPC_SQ_CUR_PSN);
5508 	qp_attr->dest_qp_num = hr_reg_read(&context, QPC_DQPN);
5509 	qp_attr->qp_access_flags =
5510 		((hr_reg_read(&context, QPC_RRE)) << V2_QP_RRE_S) |
5511 		((hr_reg_read(&context, QPC_RWE)) << V2_QP_RWE_S) |
5512 		((hr_reg_read(&context, QPC_ATE)) << V2_QP_ATE_S);
5513 
5514 	if (hr_qp->ibqp.qp_type == IB_QPT_RC ||
5515 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5516 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
5517 		struct ib_global_route *grh =
5518 			rdma_ah_retrieve_grh(&qp_attr->ah_attr);
5519 
5520 		rdma_ah_set_sl(&qp_attr->ah_attr,
5521 			       hr_reg_read(&context, QPC_SL));
5522 		rdma_ah_set_port_num(&qp_attr->ah_attr, hr_qp->port + 1);
5523 		rdma_ah_set_ah_flags(&qp_attr->ah_attr, IB_AH_GRH);
5524 		grh->flow_label = hr_reg_read(&context, QPC_FL);
5525 		grh->sgid_index = hr_reg_read(&context, QPC_GMV_IDX);
5526 		grh->hop_limit = hr_reg_read(&context, QPC_HOPLIMIT);
5527 		grh->traffic_class = hr_reg_read(&context, QPC_TC);
5528 
5529 		memcpy(grh->dgid.raw, context.dgid, sizeof(grh->dgid.raw));
5530 	}
5531 
5532 	qp_attr->port_num = hr_qp->port + 1;
5533 	qp_attr->sq_draining = 0;
5534 	qp_attr->max_rd_atomic = 1 << hr_reg_read(&context, QPC_SR_MAX);
5535 	qp_attr->max_dest_rd_atomic = 1 << hr_reg_read(&context, QPC_RR_MAX);
5536 
5537 	qp_attr->min_rnr_timer = (u8)hr_reg_read(&context, QPC_MIN_RNR_TIME);
5538 	qp_attr->timeout = get_qp_timeout_attr(hr_dev, &context);
5539 	qp_attr->retry_cnt = hr_reg_read(&context, QPC_RETRY_NUM_INIT);
5540 	qp_attr->rnr_retry = hr_reg_read(&context, QPC_RNR_NUM_INIT);
5541 
5542 done:
5543 	qp_attr->cur_qp_state = qp_attr->qp_state;
5544 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
5545 	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
5546 	qp_attr->cap.max_inline_data = hr_qp->max_inline_data;
5547 
5548 	qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
5549 	qp_attr->cap.max_send_sge = hr_qp->sq.max_gs;
5550 
5551 	qp_init_attr->qp_context = ibqp->qp_context;
5552 	qp_init_attr->qp_type = ibqp->qp_type;
5553 	qp_init_attr->recv_cq = ibqp->recv_cq;
5554 	qp_init_attr->send_cq = ibqp->send_cq;
5555 	qp_init_attr->srq = ibqp->srq;
5556 	qp_init_attr->cap = qp_attr->cap;
5557 	qp_init_attr->sq_sig_type = hr_qp->sq_signal_bits;
5558 
5559 out:
5560 	mutex_unlock(&hr_qp->mutex);
5561 	return ret;
5562 }
5563 
5564 static inline int modify_qp_is_ok(struct hns_roce_qp *hr_qp)
5565 {
5566 	return ((hr_qp->ibqp.qp_type == IB_QPT_RC ||
5567 		 hr_qp->ibqp.qp_type == IB_QPT_UD ||
5568 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5569 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
5570 		hr_qp->state != IB_QPS_RESET);
5571 }
5572 
5573 static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
5574 					 struct hns_roce_qp *hr_qp,
5575 					 struct ib_udata *udata)
5576 {
5577 	struct ib_device *ibdev = &hr_dev->ib_dev;
5578 	struct hns_roce_cq *send_cq, *recv_cq;
5579 	unsigned long flags;
5580 	int ret = 0;
5581 
5582 	if (modify_qp_is_ok(hr_qp)) {
5583 		/* Modify qp to reset before destroying qp */
5584 		ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
5585 					    hr_qp->state, IB_QPS_RESET, udata);
5586 		if (ret)
5587 			ibdev_err_ratelimited(ibdev,
5588 					      "failed to modify QP to RST, ret = %d.\n",
5589 					      ret);
5590 	}
5591 
5592 	send_cq = hr_qp->ibqp.send_cq ? to_hr_cq(hr_qp->ibqp.send_cq) : NULL;
5593 	recv_cq = hr_qp->ibqp.recv_cq ? to_hr_cq(hr_qp->ibqp.recv_cq) : NULL;
5594 
5595 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
5596 	hns_roce_lock_cqs(send_cq, recv_cq);
5597 
5598 	if (!udata) {
5599 		if (recv_cq)
5600 			__hns_roce_v2_cq_clean(recv_cq, hr_qp->qpn,
5601 					       (hr_qp->ibqp.srq ?
5602 						to_hr_srq(hr_qp->ibqp.srq) :
5603 						NULL));
5604 
5605 		if (send_cq && send_cq != recv_cq)
5606 			__hns_roce_v2_cq_clean(send_cq, hr_qp->qpn, NULL);
5607 	}
5608 
5609 	hns_roce_qp_remove(hr_dev, hr_qp);
5610 
5611 	hns_roce_unlock_cqs(send_cq, recv_cq);
5612 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
5613 
5614 	return ret;
5615 }
5616 
5617 static void put_dip_ctx_idx(struct hns_roce_dev *hr_dev,
5618 			    struct hns_roce_qp *hr_qp)
5619 {
5620 	struct hns_roce_dip *hr_dip = hr_qp->dip;
5621 
5622 	xa_lock(&hr_dev->qp_table.dip_xa);
5623 
5624 	hr_dip->qp_cnt--;
5625 	if (!hr_dip->qp_cnt)
5626 		memset(hr_dip->dgid, 0, GID_LEN_V2);
5627 
5628 	xa_unlock(&hr_dev->qp_table.dip_xa);
5629 }
5630 
5631 int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
5632 {
5633 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5634 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5635 	unsigned long flags;
5636 	int ret;
5637 
5638 	/* Make sure flush_cqe() is completed */
5639 	spin_lock_irqsave(&hr_qp->flush_lock, flags);
5640 	set_bit(HNS_ROCE_STOP_FLUSH_FLAG, &hr_qp->flush_flag);
5641 	spin_unlock_irqrestore(&hr_qp->flush_lock, flags);
5642 	flush_work(&hr_qp->flush_work.work);
5643 
5644 	if (hr_qp->cong_type == CONG_TYPE_DIP)
5645 		put_dip_ctx_idx(hr_dev, hr_qp);
5646 
5647 	ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata);
5648 	if (ret)
5649 		ibdev_err_ratelimited(&hr_dev->ib_dev,
5650 				      "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n",
5651 				      hr_qp->qpn, ret);
5652 
5653 	hns_roce_qp_destroy(hr_dev, hr_qp, udata);
5654 
5655 	return 0;
5656 }
5657 
5658 static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
5659 					    struct hns_roce_qp *hr_qp)
5660 {
5661 	struct ib_device *ibdev = &hr_dev->ib_dev;
5662 	struct hns_roce_sccc_clr_done *resp;
5663 	struct hns_roce_sccc_clr *clr;
5664 	struct hns_roce_cmq_desc desc;
5665 	int ret, i;
5666 
5667 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
5668 		return 0;
5669 
5670 	mutex_lock(&hr_dev->qp_table.scc_mutex);
5671 
5672 	/* set scc ctx clear done flag */
5673 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_RESET_SCCC, false);
5674 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5675 	if (ret) {
5676 		ibdev_err(ibdev, "failed to reset SCC ctx, ret = %d.\n", ret);
5677 		goto out;
5678 	}
5679 
5680 	/* clear scc context */
5681 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLR_SCCC, false);
5682 	clr = (struct hns_roce_sccc_clr *)desc.data;
5683 	clr->qpn = cpu_to_le32(hr_qp->qpn);
5684 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5685 	if (ret) {
5686 		ibdev_err(ibdev, "failed to clear SCC ctx, ret = %d.\n", ret);
5687 		goto out;
5688 	}
5689 
5690 	/* query scc context clear is done or not */
5691 	resp = (struct hns_roce_sccc_clr_done *)desc.data;
5692 	for (i = 0; i <= HNS_ROCE_CMQ_SCC_CLR_DONE_CNT; i++) {
5693 		hns_roce_cmq_setup_basic_desc(&desc,
5694 					      HNS_ROCE_OPC_QUERY_SCCC, true);
5695 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
5696 		if (ret) {
5697 			ibdev_err(ibdev, "failed to query clr cmq, ret = %d\n",
5698 				  ret);
5699 			goto out;
5700 		}
5701 
5702 		if (resp->clr_done)
5703 			goto out;
5704 
5705 		msleep(20);
5706 	}
5707 
5708 	ibdev_err(ibdev, "query SCC clr done flag overtime.\n");
5709 	ret = -ETIMEDOUT;
5710 
5711 out:
5712 	mutex_unlock(&hr_dev->qp_table.scc_mutex);
5713 	return ret;
5714 }
5715 
5716 #define DMA_IDX_SHIFT 3
5717 #define DMA_WQE_SHIFT 3
5718 
5719 static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
5720 					      struct hns_roce_srq_context *ctx)
5721 {
5722 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
5723 	struct ib_device *ibdev = srq->ibsrq.device;
5724 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5725 	u64 mtts_idx[MTT_MIN_COUNT] = {};
5726 	dma_addr_t dma_handle_idx;
5727 	int ret;
5728 
5729 	/* Get physical address of idx que buf */
5730 	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
5731 				ARRAY_SIZE(mtts_idx));
5732 	if (ret) {
5733 		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
5734 			  ret);
5735 		return ret;
5736 	}
5737 
5738 	dma_handle_idx = hns_roce_get_mtr_ba(&idx_que->mtr);
5739 
5740 	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
5741 		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
5742 
5743 	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> DMA_IDX_SHIFT);
5744 	hr_reg_write(ctx, SRQC_IDX_BT_BA_H,
5745 		     upper_32_bits(dma_handle_idx >> DMA_IDX_SHIFT));
5746 
5747 	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
5748 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
5749 	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
5750 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
5751 
5752 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
5753 		     to_hr_hw_page_addr(mtts_idx[0]));
5754 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
5755 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
5756 
5757 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
5758 		     to_hr_hw_page_addr(mtts_idx[1]));
5759 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
5760 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
5761 
5762 	return 0;
5763 }
5764 
5765 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
5766 {
5767 	struct ib_device *ibdev = srq->ibsrq.device;
5768 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5769 	struct hns_roce_srq_context *ctx = mb_buf;
5770 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
5771 	dma_addr_t dma_handle_wqe;
5772 	int ret;
5773 
5774 	memset(ctx, 0, sizeof(*ctx));
5775 
5776 	/* Get the physical address of srq buf */
5777 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
5778 				ARRAY_SIZE(mtts_wqe));
5779 	if (ret) {
5780 		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
5781 			  ret);
5782 		return ret;
5783 	}
5784 
5785 	dma_handle_wqe = hns_roce_get_mtr_ba(&srq->buf_mtr);
5786 
5787 	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
5788 	hr_reg_write_bool(ctx, SRQC_SRQ_TYPE,
5789 			  srq->ibsrq.srq_type == IB_SRQT_XRC);
5790 	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
5791 	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
5792 	hr_reg_write(ctx, SRQC_XRCD, srq->xrcdn);
5793 	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
5794 	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
5795 	hr_reg_write(ctx, SRQC_RQWS,
5796 		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
5797 
5798 	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
5799 		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
5800 				      srq->wqe_cnt));
5801 
5802 	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> DMA_WQE_SHIFT);
5803 	hr_reg_write(ctx, SRQC_WQE_BT_BA_H,
5804 		     upper_32_bits(dma_handle_wqe >> DMA_WQE_SHIFT));
5805 
5806 	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
5807 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
5808 	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
5809 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
5810 
5811 	if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB) {
5812 		hr_reg_enable(ctx, SRQC_DB_RECORD_EN);
5813 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_L,
5814 			     lower_32_bits(srq->rdb.dma) >> 1);
5815 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_H,
5816 			     upper_32_bits(srq->rdb.dma));
5817 	}
5818 
5819 	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
5820 }
5821 
5822 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
5823 				  struct ib_srq_attr *srq_attr,
5824 				  enum ib_srq_attr_mask srq_attr_mask,
5825 				  struct ib_udata *udata)
5826 {
5827 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5828 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5829 	struct hns_roce_srq_context *srq_context;
5830 	struct hns_roce_srq_context *srqc_mask;
5831 	struct hns_roce_cmd_mailbox *mailbox;
5832 	int ret = 0;
5833 
5834 	/* Resizing SRQs is not supported yet */
5835 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
5836 		ret = -EOPNOTSUPP;
5837 		goto out;
5838 	}
5839 
5840 	if (srq_attr_mask & IB_SRQ_LIMIT) {
5841 		if (srq_attr->srq_limit > srq->wqe_cnt) {
5842 			ret = -EINVAL;
5843 			goto out;
5844 		}
5845 
5846 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5847 		if (IS_ERR(mailbox)) {
5848 			ret = PTR_ERR(mailbox);
5849 			goto out;
5850 		}
5851 
5852 		srq_context = mailbox->buf;
5853 		srqc_mask = (struct hns_roce_srq_context *)mailbox->buf + 1;
5854 
5855 		memset(srqc_mask, 0xff, sizeof(*srqc_mask));
5856 
5857 		hr_reg_write(srq_context, SRQC_LIMIT_WL, srq_attr->srq_limit);
5858 		hr_reg_clear(srqc_mask, SRQC_LIMIT_WL);
5859 
5860 		ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5861 					HNS_ROCE_CMD_MODIFY_SRQC, srq->srqn);
5862 		hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5863 		if (ret)
5864 			ibdev_err(&hr_dev->ib_dev,
5865 				  "failed to handle cmd of modifying SRQ, ret = %d.\n",
5866 				  ret);
5867 	}
5868 
5869 out:
5870 	if (ret)
5871 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_SRQ_MODIFY_ERR_CNT]);
5872 
5873 	return ret;
5874 }
5875 
5876 static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
5877 {
5878 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5879 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5880 	struct hns_roce_srq_context *srq_context;
5881 	struct hns_roce_cmd_mailbox *mailbox;
5882 	int ret;
5883 
5884 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5885 	if (IS_ERR(mailbox))
5886 		return PTR_ERR(mailbox);
5887 
5888 	srq_context = mailbox->buf;
5889 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5890 				HNS_ROCE_CMD_QUERY_SRQC, srq->srqn);
5891 	if (ret) {
5892 		ibdev_err(&hr_dev->ib_dev,
5893 			  "failed to process cmd of querying SRQ, ret = %d.\n",
5894 			  ret);
5895 		goto out;
5896 	}
5897 
5898 	attr->srq_limit = hr_reg_read(srq_context, SRQC_LIMIT_WL);
5899 	attr->max_wr = srq->wqe_cnt;
5900 	attr->max_sge = srq->max_gs - srq->rsv_sge;
5901 
5902 out:
5903 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5904 	return ret;
5905 }
5906 
5907 static int hns_roce_v2_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
5908 {
5909 	struct hns_roce_dev *hr_dev = to_hr_dev(cq->device);
5910 	struct hns_roce_v2_cq_context *cq_context;
5911 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
5912 	struct hns_roce_v2_cq_context *cqc_mask;
5913 	struct hns_roce_cmd_mailbox *mailbox;
5914 	int ret;
5915 
5916 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5917 	ret = PTR_ERR_OR_ZERO(mailbox);
5918 	if (ret)
5919 		goto err_out;
5920 
5921 	cq_context = mailbox->buf;
5922 	cqc_mask = (struct hns_roce_v2_cq_context *)mailbox->buf + 1;
5923 
5924 	memset(cqc_mask, 0xff, sizeof(*cqc_mask));
5925 
5926 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT, cq_count);
5927 	hr_reg_clear(cqc_mask, CQC_CQ_MAX_CNT);
5928 
5929 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5930 		if (cq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
5931 			dev_info(hr_dev->dev,
5932 				 "cq_period(%u) reached the upper limit, adjusted to 65.\n",
5933 				 cq_period);
5934 			cq_period = HNS_ROCE_MAX_CQ_PERIOD_HIP08;
5935 		}
5936 		cq_period *= HNS_ROCE_CLOCK_ADJUST;
5937 	}
5938 	hr_reg_write(cq_context, CQC_CQ_PERIOD, cq_period);
5939 	hr_reg_clear(cqc_mask, CQC_CQ_PERIOD);
5940 
5941 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5942 				HNS_ROCE_CMD_MODIFY_CQC, hr_cq->cqn);
5943 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5944 	if (ret)
5945 		ibdev_err_ratelimited(&hr_dev->ib_dev,
5946 				      "failed to process cmd when modifying CQ, ret = %d.\n",
5947 				      ret);
5948 
5949 err_out:
5950 	if (ret)
5951 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CQ_MODIFY_ERR_CNT]);
5952 
5953 	return ret;
5954 }
5955 
5956 static int hns_roce_v2_query_cqc(struct hns_roce_dev *hr_dev, u32 cqn,
5957 				 void *buffer)
5958 {
5959 	struct hns_roce_v2_cq_context *context;
5960 	struct hns_roce_cmd_mailbox *mailbox;
5961 	int ret;
5962 
5963 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5964 	if (IS_ERR(mailbox))
5965 		return PTR_ERR(mailbox);
5966 
5967 	context = mailbox->buf;
5968 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5969 				HNS_ROCE_CMD_QUERY_CQC, cqn);
5970 	if (ret) {
5971 		ibdev_err_ratelimited(&hr_dev->ib_dev,
5972 				      "failed to process cmd when querying CQ, ret = %d.\n",
5973 				      ret);
5974 		goto err_mailbox;
5975 	}
5976 
5977 	memcpy(buffer, context, sizeof(*context));
5978 
5979 err_mailbox:
5980 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5981 
5982 	return ret;
5983 }
5984 
5985 static int hns_roce_v2_query_mpt(struct hns_roce_dev *hr_dev, u32 key,
5986 				 void *buffer)
5987 {
5988 	struct hns_roce_v2_mpt_entry *context;
5989 	struct hns_roce_cmd_mailbox *mailbox;
5990 	int ret;
5991 
5992 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5993 	if (IS_ERR(mailbox))
5994 		return PTR_ERR(mailbox);
5995 
5996 	context = mailbox->buf;
5997 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
5998 				key_to_hw_index(key));
5999 	if (ret) {
6000 		ibdev_err(&hr_dev->ib_dev,
6001 			  "failed to process cmd when querying MPT, ret = %d.\n",
6002 			  ret);
6003 		goto err_mailbox;
6004 	}
6005 
6006 	memcpy(buffer, context, sizeof(*context));
6007 
6008 err_mailbox:
6009 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6010 
6011 	return ret;
6012 }
6013 
6014 static void dump_aeqe_log(struct hns_roce_work *irq_work)
6015 {
6016 	struct hns_roce_dev *hr_dev = irq_work->hr_dev;
6017 	struct ib_device *ibdev = &hr_dev->ib_dev;
6018 
6019 	switch (irq_work->event_type) {
6020 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
6021 		ibdev_info(ibdev, "path migrated succeeded.\n");
6022 		break;
6023 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
6024 		ibdev_warn(ibdev, "path migration failed.\n");
6025 		break;
6026 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
6027 		break;
6028 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
6029 		ibdev_dbg(ibdev, "send queue drained.\n");
6030 		break;
6031 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
6032 		ibdev_err(ibdev, "local work queue 0x%x catast error, sub_event type is: %d\n",
6033 			  irq_work->queue_num, irq_work->sub_type);
6034 		break;
6035 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
6036 		ibdev_err(ibdev, "invalid request local work queue 0x%x error.\n",
6037 			  irq_work->queue_num);
6038 		break;
6039 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
6040 		ibdev_err(ibdev, "local access violation work queue 0x%x error, sub_event type is: %d\n",
6041 			  irq_work->queue_num, irq_work->sub_type);
6042 		break;
6043 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
6044 		ibdev_dbg(ibdev, "SRQ limit reach.\n");
6045 		break;
6046 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
6047 		ibdev_dbg(ibdev, "SRQ last wqe reach.\n");
6048 		break;
6049 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6050 		ibdev_err(ibdev, "SRQ catas error.\n");
6051 		break;
6052 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6053 		ibdev_err(ibdev, "CQ 0x%x access err.\n", irq_work->queue_num);
6054 		break;
6055 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6056 		ibdev_warn(ibdev, "CQ 0x%x overflow\n", irq_work->queue_num);
6057 		break;
6058 	case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
6059 		ibdev_warn(ibdev, "DB overflow.\n");
6060 		break;
6061 	case HNS_ROCE_EVENT_TYPE_MB:
6062 		break;
6063 	case HNS_ROCE_EVENT_TYPE_FLR:
6064 		ibdev_warn(ibdev, "function level reset.\n");
6065 		break;
6066 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6067 		ibdev_err(ibdev, "xrc domain violation error.\n");
6068 		break;
6069 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6070 		ibdev_err(ibdev, "invalid xrceth error.\n");
6071 		break;
6072 	default:
6073 		ibdev_info(ibdev, "Undefined event %d.\n",
6074 			   irq_work->event_type);
6075 		break;
6076 	}
6077 }
6078 
6079 static void hns_roce_irq_work_handle(struct work_struct *work)
6080 {
6081 	struct hns_roce_work *irq_work =
6082 				container_of(work, struct hns_roce_work, work);
6083 	struct hns_roce_dev *hr_dev = irq_work->hr_dev;
6084 	int event_type = irq_work->event_type;
6085 	u32 queue_num = irq_work->queue_num;
6086 
6087 	switch (event_type) {
6088 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
6089 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
6090 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
6091 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
6092 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
6093 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
6094 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
6095 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
6096 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6097 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6098 		hns_roce_qp_event(hr_dev, queue_num, event_type);
6099 		break;
6100 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
6101 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6102 		hns_roce_srq_event(hr_dev, queue_num, event_type);
6103 		break;
6104 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6105 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6106 		hns_roce_cq_event(hr_dev, queue_num, event_type);
6107 		break;
6108 	default:
6109 		break;
6110 	}
6111 
6112 	dump_aeqe_log(irq_work);
6113 
6114 	kfree(irq_work);
6115 }
6116 
6117 static void hns_roce_v2_init_irq_work(struct hns_roce_dev *hr_dev,
6118 				      struct hns_roce_eq *eq, u32 queue_num)
6119 {
6120 	struct hns_roce_work *irq_work;
6121 
6122 	irq_work = kzalloc(sizeof(struct hns_roce_work), GFP_ATOMIC);
6123 	if (!irq_work)
6124 		return;
6125 
6126 	INIT_WORK(&irq_work->work, hns_roce_irq_work_handle);
6127 	irq_work->hr_dev = hr_dev;
6128 	irq_work->event_type = eq->event_type;
6129 	irq_work->sub_type = eq->sub_type;
6130 	irq_work->queue_num = queue_num;
6131 	queue_work(hr_dev->irq_workq, &irq_work->work);
6132 }
6133 
6134 static void update_eq_db(struct hns_roce_eq *eq)
6135 {
6136 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6137 	struct hns_roce_v2_db eq_db = {};
6138 
6139 	if (eq->type_flag == HNS_ROCE_AEQ) {
6140 		hr_reg_write(&eq_db, EQ_DB_CMD,
6141 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6142 			     HNS_ROCE_EQ_DB_CMD_AEQ :
6143 			     HNS_ROCE_EQ_DB_CMD_AEQ_ARMED);
6144 	} else {
6145 		hr_reg_write(&eq_db, EQ_DB_TAG, eq->eqn);
6146 
6147 		hr_reg_write(&eq_db, EQ_DB_CMD,
6148 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6149 			     HNS_ROCE_EQ_DB_CMD_CEQ :
6150 			     HNS_ROCE_EQ_DB_CMD_CEQ_ARMED);
6151 	}
6152 
6153 	hr_reg_write(&eq_db, EQ_DB_CI, eq->cons_index);
6154 
6155 	hns_roce_write64(hr_dev, (__le32 *)&eq_db, eq->db_reg);
6156 }
6157 
6158 static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
6159 {
6160 	struct hns_roce_aeqe *aeqe;
6161 
6162 	aeqe = hns_roce_buf_offset(eq->mtr.kmem,
6163 				   (eq->cons_index & (eq->entries - 1)) *
6164 				   eq->eqe_size);
6165 
6166 	return (hr_reg_read(aeqe, AEQE_OWNER) ^
6167 		!!(eq->cons_index & eq->entries)) ? aeqe : NULL;
6168 }
6169 
6170 static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
6171 				       struct hns_roce_eq *eq)
6172 {
6173 	struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
6174 	irqreturn_t aeqe_found = IRQ_NONE;
6175 	int num_aeqes = 0;
6176 	int event_type;
6177 	u32 queue_num;
6178 	int sub_type;
6179 
6180 	while (aeqe && num_aeqes < HNS_AEQ_POLLING_BUDGET) {
6181 		/* Make sure we read AEQ entry after we have checked the
6182 		 * ownership bit
6183 		 */
6184 		dma_rmb();
6185 
6186 		event_type = hr_reg_read(aeqe, AEQE_EVENT_TYPE);
6187 		sub_type = hr_reg_read(aeqe, AEQE_SUB_TYPE);
6188 		queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
6189 
6190 		switch (event_type) {
6191 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
6192 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
6193 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
6194 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6195 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6196 			hns_roce_flush_cqe(hr_dev, queue_num);
6197 			break;
6198 		case HNS_ROCE_EVENT_TYPE_MB:
6199 			hns_roce_cmd_event(hr_dev,
6200 					le16_to_cpu(aeqe->event.cmd.token),
6201 					aeqe->event.cmd.status,
6202 					le64_to_cpu(aeqe->event.cmd.out_param));
6203 			break;
6204 		default:
6205 			break;
6206 		}
6207 
6208 		eq->event_type = event_type;
6209 		eq->sub_type = sub_type;
6210 		++eq->cons_index;
6211 		aeqe_found = IRQ_HANDLED;
6212 
6213 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_AEQE_CNT]);
6214 
6215 		hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
6216 
6217 		aeqe = next_aeqe_sw_v2(eq);
6218 		++num_aeqes;
6219 	}
6220 
6221 	update_eq_db(eq);
6222 
6223 	return IRQ_RETVAL(aeqe_found);
6224 }
6225 
6226 static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq)
6227 {
6228 	struct hns_roce_ceqe *ceqe;
6229 
6230 	ceqe = hns_roce_buf_offset(eq->mtr.kmem,
6231 				   (eq->cons_index & (eq->entries - 1)) *
6232 				   eq->eqe_size);
6233 
6234 	return (hr_reg_read(ceqe, CEQE_OWNER) ^
6235 		!!(eq->cons_index & eq->entries)) ? ceqe : NULL;
6236 }
6237 
6238 static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_eq *eq)
6239 {
6240 	queue_work(system_bh_wq, &eq->work);
6241 
6242 	return IRQ_HANDLED;
6243 }
6244 
6245 static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
6246 {
6247 	struct hns_roce_eq *eq = eq_ptr;
6248 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6249 	irqreturn_t int_work;
6250 
6251 	if (eq->type_flag == HNS_ROCE_CEQ)
6252 		/* Completion event interrupt */
6253 		int_work = hns_roce_v2_ceq_int(eq);
6254 	else
6255 		/* Asynchronous event interrupt */
6256 		int_work = hns_roce_v2_aeq_int(hr_dev, eq);
6257 
6258 	return IRQ_RETVAL(int_work);
6259 }
6260 
6261 static irqreturn_t abnormal_interrupt_basic(struct hns_roce_dev *hr_dev,
6262 					    u32 int_st)
6263 {
6264 	struct pci_dev *pdev = hr_dev->pci_dev;
6265 	struct hnae3_ae_dev *ae_dev = pci_get_drvdata(pdev);
6266 	const struct hnae3_ae_ops *ops = ae_dev->ops;
6267 	enum hnae3_reset_type reset_type;
6268 	irqreturn_t int_work = IRQ_NONE;
6269 	u32 int_en;
6270 
6271 	int_en = roce_read(hr_dev, ROCEE_VF_ABN_INT_EN_REG);
6272 
6273 	if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S)) {
6274 		dev_err(hr_dev->dev, "AEQ overflow!\n");
6275 
6276 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG,
6277 			   1 << HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S);
6278 
6279 		reset_type = hr_dev->is_vf ?
6280 			     HNAE3_VF_FUNC_RESET : HNAE3_FUNC_RESET;
6281 
6282 		/* Set reset level for reset_event() */
6283 		if (ops->set_default_reset_request)
6284 			ops->set_default_reset_request(ae_dev, reset_type);
6285 		if (ops->reset_event)
6286 			ops->reset_event(pdev, NULL);
6287 
6288 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6289 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6290 
6291 		int_work = IRQ_HANDLED;
6292 	} else {
6293 		dev_err(hr_dev->dev, "there is no basic abn irq found.\n");
6294 	}
6295 
6296 	return IRQ_RETVAL(int_work);
6297 }
6298 
6299 static int fmea_ram_ecc_query(struct hns_roce_dev *hr_dev,
6300 			       struct fmea_ram_ecc *ecc_info)
6301 {
6302 	struct hns_roce_cmq_desc desc;
6303 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6304 	int ret;
6305 
6306 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_QUERY_RAM_ECC, true);
6307 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6308 	if (ret)
6309 		return ret;
6310 
6311 	ecc_info->is_ecc_err = hr_reg_read(req, QUERY_RAM_ECC_1BIT_ERR);
6312 	ecc_info->res_type = hr_reg_read(req, QUERY_RAM_ECC_RES_TYPE);
6313 	ecc_info->index = hr_reg_read(req, QUERY_RAM_ECC_TAG);
6314 
6315 	return 0;
6316 }
6317 
6318 static int fmea_recover_gmv(struct hns_roce_dev *hr_dev, u32 idx)
6319 {
6320 	struct hns_roce_cmq_desc desc;
6321 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6322 	u32 addr_upper;
6323 	u32 addr_low;
6324 	int ret;
6325 
6326 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, true);
6327 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6328 
6329 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6330 	if (ret) {
6331 		dev_err(hr_dev->dev,
6332 			"failed to execute cmd to read gmv, ret = %d.\n", ret);
6333 		return ret;
6334 	}
6335 
6336 	addr_low =  hr_reg_read(req, CFG_GMV_BT_BA_L);
6337 	addr_upper = hr_reg_read(req, CFG_GMV_BT_BA_H);
6338 
6339 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
6340 	hr_reg_write(req, CFG_GMV_BT_BA_L, addr_low);
6341 	hr_reg_write(req, CFG_GMV_BT_BA_H, addr_upper);
6342 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6343 
6344 	return hns_roce_cmq_send(hr_dev, &desc, 1);
6345 }
6346 
6347 static u64 fmea_get_ram_res_addr(u32 res_type, __le64 *data)
6348 {
6349 	if (res_type == ECC_RESOURCE_QPC_TIMER ||
6350 	    res_type == ECC_RESOURCE_CQC_TIMER ||
6351 	    res_type == ECC_RESOURCE_SCCC)
6352 		return le64_to_cpu(*data);
6353 
6354 	return le64_to_cpu(*data) << HNS_HW_PAGE_SHIFT;
6355 }
6356 
6357 static int fmea_recover_others(struct hns_roce_dev *hr_dev, u32 res_type,
6358 			       u32 index)
6359 {
6360 	u8 write_bt0_op = fmea_ram_res[res_type].write_bt0_op;
6361 	u8 read_bt0_op = fmea_ram_res[res_type].read_bt0_op;
6362 	struct hns_roce_cmd_mailbox *mailbox;
6363 	u64 addr;
6364 	int ret;
6365 
6366 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6367 	if (IS_ERR(mailbox))
6368 		return PTR_ERR(mailbox);
6369 
6370 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, read_bt0_op, index);
6371 	if (ret) {
6372 		dev_err(hr_dev->dev,
6373 			"failed to execute cmd to read fmea ram, ret = %d.\n",
6374 			ret);
6375 		goto out;
6376 	}
6377 
6378 	addr = fmea_get_ram_res_addr(res_type, mailbox->buf);
6379 
6380 	ret = hns_roce_cmd_mbox(hr_dev, addr, 0, write_bt0_op, index);
6381 	if (ret)
6382 		dev_err(hr_dev->dev,
6383 			"failed to execute cmd to write fmea ram, ret = %d.\n",
6384 			ret);
6385 
6386 out:
6387 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6388 	return ret;
6389 }
6390 
6391 static void fmea_ram_ecc_recover(struct hns_roce_dev *hr_dev,
6392 				 struct fmea_ram_ecc *ecc_info)
6393 {
6394 	u32 res_type = ecc_info->res_type;
6395 	u32 index = ecc_info->index;
6396 	int ret;
6397 
6398 	BUILD_BUG_ON(ARRAY_SIZE(fmea_ram_res) != ECC_RESOURCE_COUNT);
6399 
6400 	if (res_type >= ECC_RESOURCE_COUNT) {
6401 		dev_err(hr_dev->dev, "unsupported fmea ram ecc type %u.\n",
6402 			res_type);
6403 		return;
6404 	}
6405 
6406 	if (res_type == ECC_RESOURCE_GMV)
6407 		ret = fmea_recover_gmv(hr_dev, index);
6408 	else
6409 		ret = fmea_recover_others(hr_dev, res_type, index);
6410 	if (ret)
6411 		dev_err(hr_dev->dev,
6412 			"failed to recover %s, index = %u, ret = %d.\n",
6413 			fmea_ram_res[res_type].name, index, ret);
6414 }
6415 
6416 static void fmea_ram_ecc_work(struct work_struct *ecc_work)
6417 {
6418 	struct hns_roce_dev *hr_dev =
6419 		container_of(ecc_work, struct hns_roce_dev, ecc_work);
6420 	struct fmea_ram_ecc ecc_info = {};
6421 
6422 	if (fmea_ram_ecc_query(hr_dev, &ecc_info)) {
6423 		dev_err(hr_dev->dev, "failed to query fmea ram ecc.\n");
6424 		return;
6425 	}
6426 
6427 	if (!ecc_info.is_ecc_err) {
6428 		dev_err(hr_dev->dev, "there is no fmea ram ecc err found.\n");
6429 		return;
6430 	}
6431 
6432 	fmea_ram_ecc_recover(hr_dev, &ecc_info);
6433 }
6434 
6435 static irqreturn_t hns_roce_v2_msix_interrupt_abn(int irq, void *dev_id)
6436 {
6437 	struct hns_roce_dev *hr_dev = dev_id;
6438 	irqreturn_t int_work = IRQ_NONE;
6439 	u32 int_st;
6440 
6441 	int_st = roce_read(hr_dev, ROCEE_VF_ABN_INT_ST_REG);
6442 
6443 	if (int_st) {
6444 		int_work = abnormal_interrupt_basic(hr_dev, int_st);
6445 	} else if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
6446 		queue_work(hr_dev->irq_workq, &hr_dev->ecc_work);
6447 		int_work = IRQ_HANDLED;
6448 	} else {
6449 		dev_err(hr_dev->dev, "there is no abnormal irq found.\n");
6450 	}
6451 
6452 	return IRQ_RETVAL(int_work);
6453 }
6454 
6455 static void hns_roce_v2_int_mask_enable(struct hns_roce_dev *hr_dev,
6456 					int eq_num, u32 enable_flag)
6457 {
6458 	int i;
6459 
6460 	for (i = 0; i < eq_num; i++)
6461 		roce_write(hr_dev, ROCEE_VF_EVENT_INT_EN_REG +
6462 			   i * EQ_REG_OFFSET, enable_flag);
6463 
6464 	roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, enable_flag);
6465 	roce_write(hr_dev, ROCEE_VF_ABN_INT_CFG_REG, enable_flag);
6466 }
6467 
6468 static void free_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6469 {
6470 	hns_roce_mtr_destroy(hr_dev, &eq->mtr);
6471 }
6472 
6473 static void hns_roce_v2_destroy_eqc(struct hns_roce_dev *hr_dev,
6474 				    struct hns_roce_eq *eq)
6475 {
6476 	struct device *dev = hr_dev->dev;
6477 	int eqn = eq->eqn;
6478 	int ret;
6479 	u8 cmd;
6480 
6481 	if (eqn < hr_dev->caps.num_comp_vectors)
6482 		cmd = HNS_ROCE_CMD_DESTROY_CEQC;
6483 	else
6484 		cmd = HNS_ROCE_CMD_DESTROY_AEQC;
6485 
6486 	ret = hns_roce_destroy_hw_ctx(hr_dev, cmd, eqn & HNS_ROCE_V2_EQN_M);
6487 	if (ret)
6488 		dev_err(dev, "[mailbox cmd] destroy eqc(%d) failed.\n", eqn);
6489 
6490 	free_eq_buf(hr_dev, eq);
6491 }
6492 
6493 static void init_eq_config(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6494 {
6495 	eq->db_reg = hr_dev->reg_base + ROCEE_VF_EQ_DB_CFG0_REG;
6496 	eq->cons_index = 0;
6497 	eq->over_ignore = HNS_ROCE_V2_EQ_OVER_IGNORE_0;
6498 	eq->coalesce = HNS_ROCE_V2_EQ_COALESCE_0;
6499 	eq->arm_st = HNS_ROCE_V2_EQ_ALWAYS_ARMED;
6500 	eq->shift = ilog2((unsigned int)eq->entries);
6501 }
6502 
6503 static int config_eqc(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq,
6504 		      void *mb_buf)
6505 {
6506 	u64 eqe_ba[MTT_MIN_COUNT] = { 0 };
6507 	struct hns_roce_eq_context *eqc;
6508 	u64 bt_ba = 0;
6509 	int ret;
6510 
6511 	eqc = mb_buf;
6512 	memset(eqc, 0, sizeof(struct hns_roce_eq_context));
6513 
6514 	init_eq_config(hr_dev, eq);
6515 
6516 	/* if not multi-hop, eqe buffer only use one trunk */
6517 	ret = hns_roce_mtr_find(hr_dev, &eq->mtr, 0, eqe_ba,
6518 				ARRAY_SIZE(eqe_ba));
6519 	if (ret) {
6520 		dev_err(hr_dev->dev, "failed to find EQE mtr, ret = %d\n", ret);
6521 		return ret;
6522 	}
6523 
6524 	bt_ba = hns_roce_get_mtr_ba(&eq->mtr);
6525 
6526 	hr_reg_write(eqc, EQC_EQ_ST, HNS_ROCE_V2_EQ_STATE_VALID);
6527 	hr_reg_write(eqc, EQC_EQE_HOP_NUM, eq->hop_num);
6528 	hr_reg_write(eqc, EQC_OVER_IGNORE, eq->over_ignore);
6529 	hr_reg_write(eqc, EQC_COALESCE, eq->coalesce);
6530 	hr_reg_write(eqc, EQC_ARM_ST, eq->arm_st);
6531 	hr_reg_write(eqc, EQC_EQN, eq->eqn);
6532 	hr_reg_write(eqc, EQC_EQE_CNT, HNS_ROCE_EQ_INIT_EQE_CNT);
6533 	hr_reg_write(eqc, EQC_EQE_BA_PG_SZ,
6534 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.ba_pg_shift));
6535 	hr_reg_write(eqc, EQC_EQE_BUF_PG_SZ,
6536 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.buf_pg_shift));
6537 	hr_reg_write(eqc, EQC_EQ_PROD_INDX, HNS_ROCE_EQ_INIT_PROD_IDX);
6538 	hr_reg_write(eqc, EQC_EQ_MAX_CNT, eq->eq_max_cnt);
6539 
6540 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6541 		if (eq->eq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
6542 			dev_info(hr_dev->dev, "eq_period(%u) reached the upper limit, adjusted to 65.\n",
6543 				 eq->eq_period);
6544 			eq->eq_period = HNS_ROCE_MAX_EQ_PERIOD;
6545 		}
6546 		eq->eq_period *= HNS_ROCE_CLOCK_ADJUST;
6547 	}
6548 
6549 	hr_reg_write(eqc, EQC_EQ_PERIOD, eq->eq_period);
6550 	hr_reg_write(eqc, EQC_EQE_REPORT_TIMER, HNS_ROCE_EQ_INIT_REPORT_TIMER);
6551 	hr_reg_write(eqc, EQC_EQE_BA_L, bt_ba >> 3);
6552 	hr_reg_write(eqc, EQC_EQE_BA_H, bt_ba >> 35);
6553 	hr_reg_write(eqc, EQC_SHIFT, eq->shift);
6554 	hr_reg_write(eqc, EQC_MSI_INDX, HNS_ROCE_EQ_INIT_MSI_IDX);
6555 	hr_reg_write(eqc, EQC_CUR_EQE_BA_L, eqe_ba[0] >> 12);
6556 	hr_reg_write(eqc, EQC_CUR_EQE_BA_M, eqe_ba[0] >> 28);
6557 	hr_reg_write(eqc, EQC_CUR_EQE_BA_H, eqe_ba[0] >> 60);
6558 	hr_reg_write(eqc, EQC_EQ_CONS_INDX, HNS_ROCE_EQ_INIT_CONS_IDX);
6559 	hr_reg_write(eqc, EQC_NEX_EQE_BA_L, eqe_ba[1] >> 12);
6560 	hr_reg_write(eqc, EQC_NEX_EQE_BA_H, eqe_ba[1] >> 44);
6561 	hr_reg_write(eqc, EQC_EQE_SIZE, eq->eqe_size == HNS_ROCE_V3_EQE_SIZE);
6562 
6563 	return 0;
6564 }
6565 
6566 static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6567 {
6568 	struct hns_roce_buf_attr buf_attr = {};
6569 	int err;
6570 
6571 	if (hr_dev->caps.eqe_hop_num == HNS_ROCE_HOP_NUM_0)
6572 		eq->hop_num = 0;
6573 	else
6574 		eq->hop_num = hr_dev->caps.eqe_hop_num;
6575 
6576 	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
6577 	buf_attr.region[0].size = eq->entries * eq->eqe_size;
6578 	buf_attr.region[0].hopnum = eq->hop_num;
6579 	buf_attr.region_count = 1;
6580 
6581 	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
6582 				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
6583 				  0);
6584 	if (err)
6585 		dev_err(hr_dev->dev, "failed to alloc EQE mtr, err %d\n", err);
6586 
6587 	return err;
6588 }
6589 
6590 static int hns_roce_v2_create_eq(struct hns_roce_dev *hr_dev,
6591 				 struct hns_roce_eq *eq, u8 eq_cmd)
6592 {
6593 	struct hns_roce_cmd_mailbox *mailbox;
6594 	int ret;
6595 
6596 	/* Allocate mailbox memory */
6597 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6598 	if (IS_ERR(mailbox))
6599 		return PTR_ERR(mailbox);
6600 
6601 	ret = alloc_eq_buf(hr_dev, eq);
6602 	if (ret)
6603 		goto free_cmd_mbox;
6604 
6605 	ret = config_eqc(hr_dev, eq, mailbox->buf);
6606 	if (ret)
6607 		goto err_cmd_mbox;
6608 
6609 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, eq_cmd, eq->eqn);
6610 	if (ret) {
6611 		dev_err(hr_dev->dev, "[mailbox cmd] create eqc failed.\n");
6612 		goto err_cmd_mbox;
6613 	}
6614 
6615 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6616 
6617 	return 0;
6618 
6619 err_cmd_mbox:
6620 	free_eq_buf(hr_dev, eq);
6621 
6622 free_cmd_mbox:
6623 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6624 
6625 	return ret;
6626 }
6627 
6628 static void hns_roce_ceq_work(struct work_struct *work)
6629 {
6630 	struct hns_roce_eq *eq = from_work(eq, work, work);
6631 	struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq);
6632 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6633 	int ceqe_num = 0;
6634 	u32 cqn;
6635 
6636 	while (ceqe && ceqe_num < hr_dev->caps.ceqe_depth) {
6637 		/* Make sure we read CEQ entry after we have checked the
6638 		 * ownership bit
6639 		 */
6640 		dma_rmb();
6641 
6642 		cqn = hr_reg_read(ceqe, CEQE_CQN);
6643 
6644 		hns_roce_cq_completion(hr_dev, cqn);
6645 
6646 		++eq->cons_index;
6647 		++ceqe_num;
6648 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CEQE_CNT]);
6649 
6650 		ceqe = next_ceqe_sw_v2(eq);
6651 	}
6652 
6653 	update_eq_db(eq);
6654 }
6655 
6656 static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
6657 				  int comp_num, int aeq_num, int other_num)
6658 {
6659 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6660 	int i, j;
6661 	int ret;
6662 
6663 	for (i = 0; i < irq_num; i++) {
6664 		hr_dev->irq_names[i] = kzalloc(HNS_ROCE_INT_NAME_LEN,
6665 					       GFP_KERNEL);
6666 		if (!hr_dev->irq_names[i]) {
6667 			ret = -ENOMEM;
6668 			goto err_kzalloc_failed;
6669 		}
6670 	}
6671 
6672 	/* irq contains: abnormal + AEQ + CEQ */
6673 	for (j = 0; j < other_num; j++)
6674 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6675 			 "hns-%s-abn-%d", pci_name(hr_dev->pci_dev), j);
6676 
6677 	for (j = other_num; j < (other_num + aeq_num); j++)
6678 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6679 			 "hns-%s-aeq-%d", pci_name(hr_dev->pci_dev), j - other_num);
6680 
6681 	for (j = (other_num + aeq_num); j < irq_num; j++)
6682 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6683 			 "hns-%s-ceq-%d", pci_name(hr_dev->pci_dev),
6684 			 j - other_num - aeq_num);
6685 
6686 	for (j = 0; j < irq_num; j++) {
6687 		if (j < other_num) {
6688 			ret = request_irq(hr_dev->irq[j],
6689 					  hns_roce_v2_msix_interrupt_abn,
6690 					  0, hr_dev->irq_names[j], hr_dev);
6691 		} else if (j < (other_num + comp_num)) {
6692 			INIT_WORK(&eq_table->eq[j - other_num].work,
6693 				  hns_roce_ceq_work);
6694 			ret = request_irq(eq_table->eq[j - other_num].irq,
6695 					  hns_roce_v2_msix_interrupt_eq,
6696 					  0, hr_dev->irq_names[j + aeq_num],
6697 					  &eq_table->eq[j - other_num]);
6698 		} else {
6699 			ret = request_irq(eq_table->eq[j - other_num].irq,
6700 					  hns_roce_v2_msix_interrupt_eq,
6701 					  0, hr_dev->irq_names[j - comp_num],
6702 					  &eq_table->eq[j - other_num]);
6703 		}
6704 
6705 		if (ret) {
6706 			dev_err(hr_dev->dev, "request irq error!\n");
6707 			goto err_request_failed;
6708 		}
6709 	}
6710 
6711 	return 0;
6712 
6713 err_request_failed:
6714 	for (j -= 1; j >= 0; j--) {
6715 		if (j < other_num) {
6716 			free_irq(hr_dev->irq[j], hr_dev);
6717 			continue;
6718 		}
6719 		free_irq(eq_table->eq[j - other_num].irq,
6720 			 &eq_table->eq[j - other_num]);
6721 		if (j < other_num + comp_num)
6722 			cancel_work_sync(&eq_table->eq[j - other_num].work);
6723 	}
6724 
6725 err_kzalloc_failed:
6726 	for (i -= 1; i >= 0; i--)
6727 		kfree(hr_dev->irq_names[i]);
6728 
6729 	return ret;
6730 }
6731 
6732 static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev)
6733 {
6734 	int irq_num;
6735 	int eq_num;
6736 	int i;
6737 
6738 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6739 	irq_num = eq_num + hr_dev->caps.num_other_vectors;
6740 
6741 	for (i = 0; i < hr_dev->caps.num_other_vectors; i++)
6742 		free_irq(hr_dev->irq[i], hr_dev);
6743 
6744 	for (i = 0; i < eq_num; i++) {
6745 		free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]);
6746 		if (i < hr_dev->caps.num_comp_vectors)
6747 			cancel_work_sync(&hr_dev->eq_table.eq[i].work);
6748 	}
6749 
6750 	for (i = 0; i < irq_num; i++)
6751 		kfree(hr_dev->irq_names[i]);
6752 }
6753 
6754 static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
6755 {
6756 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6757 	struct device *dev = hr_dev->dev;
6758 	struct hns_roce_eq *eq;
6759 	int other_num;
6760 	int comp_num;
6761 	int aeq_num;
6762 	int irq_num;
6763 	int eq_num;
6764 	u8 eq_cmd;
6765 	int ret;
6766 	int i;
6767 
6768 	if (hr_dev->caps.aeqe_depth < HNS_AEQ_POLLING_BUDGET)
6769 		return -EINVAL;
6770 
6771 	other_num = hr_dev->caps.num_other_vectors;
6772 	comp_num = hr_dev->caps.num_comp_vectors;
6773 	aeq_num = hr_dev->caps.num_aeq_vectors;
6774 
6775 	eq_num = comp_num + aeq_num;
6776 	irq_num = eq_num + other_num;
6777 
6778 	eq_table->eq = kcalloc(eq_num, sizeof(*eq_table->eq), GFP_KERNEL);
6779 	if (!eq_table->eq)
6780 		return -ENOMEM;
6781 
6782 	/* create eq */
6783 	for (i = 0; i < eq_num; i++) {
6784 		eq = &eq_table->eq[i];
6785 		eq->hr_dev = hr_dev;
6786 		eq->eqn = i;
6787 		if (i < comp_num) {
6788 			/* CEQ */
6789 			eq_cmd = HNS_ROCE_CMD_CREATE_CEQC;
6790 			eq->type_flag = HNS_ROCE_CEQ;
6791 			eq->entries = hr_dev->caps.ceqe_depth;
6792 			eq->eqe_size = hr_dev->caps.ceqe_size;
6793 			eq->irq = hr_dev->irq[i + other_num + aeq_num];
6794 			eq->eq_max_cnt = HNS_ROCE_CEQ_DEFAULT_BURST_NUM;
6795 			eq->eq_period = HNS_ROCE_CEQ_DEFAULT_INTERVAL;
6796 		} else {
6797 			/* AEQ */
6798 			eq_cmd = HNS_ROCE_CMD_CREATE_AEQC;
6799 			eq->type_flag = HNS_ROCE_AEQ;
6800 			eq->entries = hr_dev->caps.aeqe_depth;
6801 			eq->eqe_size = hr_dev->caps.aeqe_size;
6802 			eq->irq = hr_dev->irq[i - comp_num + other_num];
6803 			eq->eq_max_cnt = HNS_ROCE_AEQ_DEFAULT_BURST_NUM;
6804 			eq->eq_period = HNS_ROCE_AEQ_DEFAULT_INTERVAL;
6805 		}
6806 
6807 		ret = hns_roce_v2_create_eq(hr_dev, eq, eq_cmd);
6808 		if (ret) {
6809 			dev_err(dev, "failed to create eq.\n");
6810 			goto err_create_eq_fail;
6811 		}
6812 	}
6813 
6814 	INIT_WORK(&hr_dev->ecc_work, fmea_ram_ecc_work);
6815 
6816 	hr_dev->irq_workq = alloc_ordered_workqueue("hns_roce_irq_workq", 0);
6817 	if (!hr_dev->irq_workq) {
6818 		dev_err(dev, "failed to create irq workqueue.\n");
6819 		ret = -ENOMEM;
6820 		goto err_create_eq_fail;
6821 	}
6822 
6823 	ret = __hns_roce_request_irq(hr_dev, irq_num, comp_num, aeq_num,
6824 				     other_num);
6825 	if (ret) {
6826 		dev_err(dev, "failed to request irq.\n");
6827 		goto err_request_irq_fail;
6828 	}
6829 
6830 	/* enable irq */
6831 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_ENABLE);
6832 
6833 	return 0;
6834 
6835 err_request_irq_fail:
6836 	destroy_workqueue(hr_dev->irq_workq);
6837 
6838 err_create_eq_fail:
6839 	for (i -= 1; i >= 0; i--)
6840 		hns_roce_v2_destroy_eqc(hr_dev, &eq_table->eq[i]);
6841 	kfree(eq_table->eq);
6842 
6843 	return ret;
6844 }
6845 
6846 static void hns_roce_v2_cleanup_eq_table(struct hns_roce_dev *hr_dev)
6847 {
6848 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6849 	int eq_num;
6850 	int i;
6851 
6852 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6853 
6854 	/* Disable irq */
6855 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_DISABLE);
6856 
6857 	__hns_roce_free_irq(hr_dev);
6858 	destroy_workqueue(hr_dev->irq_workq);
6859 
6860 	for (i = 0; i < eq_num; i++)
6861 		hns_roce_v2_destroy_eqc(hr_dev, &eq_table->eq[i]);
6862 
6863 	kfree(eq_table->eq);
6864 }
6865 
6866 static const struct ib_device_ops hns_roce_v2_dev_ops = {
6867 	.destroy_qp = hns_roce_v2_destroy_qp,
6868 	.modify_cq = hns_roce_v2_modify_cq,
6869 	.poll_cq = hns_roce_v2_poll_cq,
6870 	.post_recv = hns_roce_v2_post_recv,
6871 	.post_send = hns_roce_v2_post_send,
6872 	.query_qp = hns_roce_v2_query_qp,
6873 	.req_notify_cq = hns_roce_v2_req_notify_cq,
6874 };
6875 
6876 static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {
6877 	.modify_srq = hns_roce_v2_modify_srq,
6878 	.post_srq_recv = hns_roce_v2_post_srq_recv,
6879 	.query_srq = hns_roce_v2_query_srq,
6880 };
6881 
6882 static const struct hns_roce_hw hns_roce_hw_v2 = {
6883 	.cmq_init = hns_roce_v2_cmq_init,
6884 	.cmq_exit = hns_roce_v2_cmq_exit,
6885 	.hw_profile = hns_roce_v2_profile,
6886 	.hw_init = hns_roce_v2_init,
6887 	.hw_exit = hns_roce_v2_exit,
6888 	.post_mbox = v2_post_mbox,
6889 	.poll_mbox_done = v2_poll_mbox_done,
6890 	.chk_mbox_avail = v2_chk_mbox_is_avail,
6891 	.set_gid = hns_roce_v2_set_gid,
6892 	.set_mac = hns_roce_v2_set_mac,
6893 	.write_mtpt = hns_roce_v2_write_mtpt,
6894 	.rereg_write_mtpt = hns_roce_v2_rereg_write_mtpt,
6895 	.frmr_write_mtpt = hns_roce_v2_frmr_write_mtpt,
6896 	.mw_write_mtpt = hns_roce_v2_mw_write_mtpt,
6897 	.write_cqc = hns_roce_v2_write_cqc,
6898 	.set_hem = hns_roce_v2_set_hem,
6899 	.clear_hem = hns_roce_v2_clear_hem,
6900 	.modify_qp = hns_roce_v2_modify_qp,
6901 	.dereg_mr = hns_roce_v2_dereg_mr,
6902 	.qp_flow_control_init = hns_roce_v2_qp_flow_control_init,
6903 	.init_eq = hns_roce_v2_init_eq_table,
6904 	.cleanup_eq = hns_roce_v2_cleanup_eq_table,
6905 	.write_srqc = hns_roce_v2_write_srqc,
6906 	.query_cqc = hns_roce_v2_query_cqc,
6907 	.query_qpc = hns_roce_v2_query_qpc,
6908 	.query_mpt = hns_roce_v2_query_mpt,
6909 	.query_srqc = hns_roce_v2_query_srqc,
6910 	.query_sccc = hns_roce_v2_query_sccc,
6911 	.query_hw_counter = hns_roce_hw_v2_query_counter,
6912 	.get_dscp = hns_roce_hw_v2_get_dscp,
6913 	.hns_roce_dev_ops = &hns_roce_v2_dev_ops,
6914 	.hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops,
6915 };
6916 
6917 static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = {
6918 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA), 0},
6919 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA_MACSEC), 0},
6920 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA), 0},
6921 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA_MACSEC), 0},
6922 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_100G_RDMA_MACSEC), 0},
6923 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_200G_RDMA), 0},
6924 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_RDMA_DCB_PFC_VF),
6925 	 HNAE3_DEV_SUPPORT_ROCE_DCB_BITS},
6926 	/* required last entry */
6927 	{0, }
6928 };
6929 
6930 MODULE_DEVICE_TABLE(pci, hns_roce_hw_v2_pci_tbl);
6931 
6932 static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
6933 				  struct hnae3_handle *handle)
6934 {
6935 	struct hns_roce_v2_priv *priv = hr_dev->priv;
6936 	const struct pci_device_id *id;
6937 	int i;
6938 
6939 	hr_dev->pci_dev = handle->pdev;
6940 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, hr_dev->pci_dev);
6941 	hr_dev->is_vf = id->driver_data;
6942 	hr_dev->dev = &handle->pdev->dev;
6943 	hr_dev->hw = &hns_roce_hw_v2;
6944 	hr_dev->sdb_offset = ROCEE_DB_SQ_L_0_REG;
6945 	hr_dev->odb_offset = hr_dev->sdb_offset;
6946 
6947 	/* Get info from NIC driver. */
6948 	hr_dev->reg_base = handle->rinfo.roce_io_base;
6949 	hr_dev->mem_base = handle->rinfo.roce_mem_base;
6950 	hr_dev->caps.num_ports = 1;
6951 	hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
6952 	hr_dev->iboe.phy_port[0] = 0;
6953 
6954 	addrconf_addr_eui48((u8 *)&hr_dev->ib_dev.node_guid,
6955 			    hr_dev->iboe.netdevs[0]->dev_addr);
6956 
6957 	for (i = 0; i < handle->rinfo.num_vectors; i++)
6958 		hr_dev->irq[i] = pci_irq_vector(handle->pdev,
6959 						i + handle->rinfo.base_vector);
6960 
6961 	/* cmd issue mode: 0 is poll, 1 is event */
6962 	hr_dev->cmd_mod = 1;
6963 	hr_dev->loop_idc = 0;
6964 
6965 	hr_dev->reset_cnt = handle->ae_algo->ops->ae_dev_reset_cnt(handle);
6966 	priv->handle = handle;
6967 }
6968 
6969 static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6970 {
6971 	struct hns_roce_dev *hr_dev;
6972 	int ret;
6973 
6974 	hr_dev = ib_alloc_device(hns_roce_dev, ib_dev);
6975 	if (!hr_dev)
6976 		return -ENOMEM;
6977 
6978 	hr_dev->priv = kzalloc(sizeof(struct hns_roce_v2_priv), GFP_KERNEL);
6979 	if (!hr_dev->priv) {
6980 		ret = -ENOMEM;
6981 		goto error_failed_kzalloc;
6982 	}
6983 
6984 	hns_roce_hw_v2_get_cfg(hr_dev, handle);
6985 
6986 	ret = hns_roce_init(hr_dev);
6987 	if (ret) {
6988 		dev_err(hr_dev->dev, "RoCE Engine init failed!\n");
6989 		goto error_failed_roce_init;
6990 	}
6991 
6992 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6993 		ret = free_mr_init(hr_dev);
6994 		if (ret) {
6995 			dev_err(hr_dev->dev, "failed to init free mr!\n");
6996 			goto error_failed_free_mr_init;
6997 		}
6998 	}
6999 
7000 	handle->priv = hr_dev;
7001 
7002 	return 0;
7003 
7004 error_failed_free_mr_init:
7005 	hns_roce_exit(hr_dev);
7006 
7007 error_failed_roce_init:
7008 	kfree(hr_dev->priv);
7009 
7010 error_failed_kzalloc:
7011 	ib_dealloc_device(&hr_dev->ib_dev);
7012 
7013 	return ret;
7014 }
7015 
7016 static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
7017 					   bool reset)
7018 {
7019 	struct hns_roce_dev *hr_dev = handle->priv;
7020 
7021 	if (!hr_dev)
7022 		return;
7023 
7024 	handle->priv = NULL;
7025 
7026 	hr_dev->state = HNS_ROCE_DEVICE_STATE_UNINIT;
7027 	hns_roce_handle_device_err(hr_dev);
7028 
7029 	hns_roce_exit(hr_dev);
7030 	kfree(hr_dev->priv);
7031 	ib_dealloc_device(&hr_dev->ib_dev);
7032 }
7033 
7034 static int hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
7035 {
7036 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
7037 	const struct pci_device_id *id;
7038 	struct device *dev = &handle->pdev->dev;
7039 	int ret;
7040 
7041 	handle->rinfo.instance_state = HNS_ROCE_STATE_INIT;
7042 
7043 	if (ops->ae_dev_resetting(handle) || ops->get_hw_reset_stat(handle)) {
7044 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
7045 		goto reset_chk_err;
7046 	}
7047 
7048 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, handle->pdev);
7049 	if (!id)
7050 		return 0;
7051 
7052 	if (id->driver_data && handle->pdev->revision == PCI_REVISION_ID_HIP08)
7053 		return 0;
7054 
7055 	ret = __hns_roce_hw_v2_init_instance(handle);
7056 	if (ret) {
7057 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
7058 		dev_err(dev, "RoCE instance init failed! ret = %d\n", ret);
7059 		if (ops->ae_dev_resetting(handle) ||
7060 		    ops->get_hw_reset_stat(handle))
7061 			goto reset_chk_err;
7062 		else
7063 			return ret;
7064 	}
7065 
7066 	handle->rinfo.instance_state = HNS_ROCE_STATE_INITED;
7067 
7068 	return 0;
7069 
7070 reset_chk_err:
7071 	dev_err(dev, "Device is busy in resetting state.\n"
7072 		     "please retry later.\n");
7073 
7074 	return -EBUSY;
7075 }
7076 
7077 static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
7078 					   bool reset)
7079 {
7080 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED)
7081 		return;
7082 
7083 	handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;
7084 
7085 	__hns_roce_hw_v2_uninit_instance(handle, reset);
7086 
7087 	handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
7088 }
7089 
7090 static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
7091 {
7092 	struct hns_roce_dev *hr_dev;
7093 
7094 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) {
7095 		set_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
7096 		return 0;
7097 	}
7098 
7099 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_DOWN;
7100 	clear_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
7101 
7102 	hr_dev = handle->priv;
7103 	if (!hr_dev)
7104 		return 0;
7105 
7106 	hr_dev->active = false;
7107 	hr_dev->dis_db = true;
7108 
7109 	rdma_user_mmap_disassociate(&hr_dev->ib_dev);
7110 
7111 	hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
7112 
7113 	return 0;
7114 }
7115 
7116 static int hns_roce_hw_v2_reset_notify_init(struct hnae3_handle *handle)
7117 {
7118 	struct device *dev = &handle->pdev->dev;
7119 	int ret;
7120 
7121 	if (test_and_clear_bit(HNS_ROCE_RST_DIRECT_RETURN,
7122 			       &handle->rinfo.state)) {
7123 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7124 		return 0;
7125 	}
7126 
7127 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INIT;
7128 
7129 	dev_info(&handle->pdev->dev, "In reset process RoCE client reinit.\n");
7130 	ret = __hns_roce_hw_v2_init_instance(handle);
7131 	if (ret) {
7132 		/* when reset notify type is HNAE3_INIT_CLIENT In reset notify
7133 		 * callback function, RoCE Engine reinitialize. If RoCE reinit
7134 		 * failed, we should inform NIC driver.
7135 		 */
7136 		handle->priv = NULL;
7137 		dev_err(dev, "In reset process RoCE reinit failed %d.\n", ret);
7138 	} else {
7139 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7140 		dev_info(dev, "reset done, RoCE client reinit finished.\n");
7141 	}
7142 
7143 	return ret;
7144 }
7145 
7146 static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle)
7147 {
7148 	if (test_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state))
7149 		return 0;
7150 
7151 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT;
7152 	dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n");
7153 	msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY);
7154 	__hns_roce_hw_v2_uninit_instance(handle, false);
7155 
7156 	return 0;
7157 }
7158 
7159 static int hns_roce_hw_v2_reset_notify(struct hnae3_handle *handle,
7160 				       enum hnae3_reset_notify_type type)
7161 {
7162 	int ret = 0;
7163 
7164 	switch (type) {
7165 	case HNAE3_DOWN_CLIENT:
7166 		ret = hns_roce_hw_v2_reset_notify_down(handle);
7167 		break;
7168 	case HNAE3_INIT_CLIENT:
7169 		ret = hns_roce_hw_v2_reset_notify_init(handle);
7170 		break;
7171 	case HNAE3_UNINIT_CLIENT:
7172 		ret = hns_roce_hw_v2_reset_notify_uninit(handle);
7173 		break;
7174 	default:
7175 		break;
7176 	}
7177 
7178 	return ret;
7179 }
7180 
7181 static const struct hnae3_client_ops hns_roce_hw_v2_ops = {
7182 	.init_instance = hns_roce_hw_v2_init_instance,
7183 	.uninit_instance = hns_roce_hw_v2_uninit_instance,
7184 	.reset_notify = hns_roce_hw_v2_reset_notify,
7185 };
7186 
7187 static struct hnae3_client hns_roce_hw_v2_client = {
7188 	.name = "hns_roce_hw_v2",
7189 	.type = HNAE3_CLIENT_ROCE,
7190 	.ops = &hns_roce_hw_v2_ops,
7191 };
7192 
7193 static int __init hns_roce_hw_v2_init(void)
7194 {
7195 	hns_roce_init_debugfs();
7196 	return hnae3_register_client(&hns_roce_hw_v2_client);
7197 }
7198 
7199 static void __exit hns_roce_hw_v2_exit(void)
7200 {
7201 	hnae3_unregister_client(&hns_roce_hw_v2_client);
7202 	hns_roce_cleanup_debugfs();
7203 }
7204 
7205 module_init(hns_roce_hw_v2_init);
7206 module_exit(hns_roce_hw_v2_exit);
7207 
7208 MODULE_LICENSE("Dual BSD/GPL");
7209 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
7210 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
7211 MODULE_AUTHOR("Shaobo Xu <xushaobo2@huawei.com>");
7212 MODULE_DESCRIPTION("Hisilicon Hip08 Family RoCE Driver");
7213