1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright 2018-2019 Amazon.com, Inc. or its affiliates. All rights reserved.
4  */
5 
6 #include <linux/vmalloc.h>
7 
8 #include <rdma/ib_addr.h>
9 #include <rdma/ib_umem.h>
10 #include <rdma/ib_user_verbs.h>
11 #include <rdma/ib_verbs.h>
12 #include <rdma/uverbs_ioctl.h>
13 
14 #include "efa.h"
15 
16 enum {
17 	EFA_MMAP_DMA_PAGE = 0,
18 	EFA_MMAP_IO_WC,
19 	EFA_MMAP_IO_NC,
20 };
21 
22 #define EFA_AENQ_ENABLED_GROUPS \
23 	(BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
24 	 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
25 
26 struct efa_user_mmap_entry {
27 	struct rdma_user_mmap_entry rdma_entry;
28 	u64 address;
29 	u8 mmap_flag;
30 };
31 
32 #define EFA_DEFINE_STATS(op) \
33 	op(EFA_TX_BYTES, "tx_bytes") \
34 	op(EFA_TX_PKTS, "tx_pkts") \
35 	op(EFA_RX_BYTES, "rx_bytes") \
36 	op(EFA_RX_PKTS, "rx_pkts") \
37 	op(EFA_RX_DROPS, "rx_drops") \
38 	op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
39 	op(EFA_COMPLETED_CMDS, "completed_cmds") \
40 	op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \
41 	op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \
42 	op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \
43 	op(EFA_CREATE_QP_ERR, "create_qp_err") \
44 	op(EFA_REG_MR_ERR, "reg_mr_err") \
45 	op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \
46 	op(EFA_CREATE_AH_ERR, "create_ah_err")
47 
48 #define EFA_STATS_ENUM(ename, name) ename,
49 #define EFA_STATS_STR(ename, name) [ename] = name,
50 
51 enum efa_hw_stats {
52 	EFA_DEFINE_STATS(EFA_STATS_ENUM)
53 };
54 
55 static const char *const efa_stats_names[] = {
56 	EFA_DEFINE_STATS(EFA_STATS_STR)
57 };
58 
59 #define EFA_CHUNK_PAYLOAD_SHIFT       12
60 #define EFA_CHUNK_PAYLOAD_SIZE        BIT(EFA_CHUNK_PAYLOAD_SHIFT)
61 #define EFA_CHUNK_PAYLOAD_PTR_SIZE    8
62 
63 #define EFA_CHUNK_SHIFT               12
64 #define EFA_CHUNK_SIZE                BIT(EFA_CHUNK_SHIFT)
65 #define EFA_CHUNK_PTR_SIZE            sizeof(struct efa_com_ctrl_buff_info)
66 
67 #define EFA_PTRS_PER_CHUNK \
68 	((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
69 
70 #define EFA_CHUNK_USED_SIZE \
71 	((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
72 
73 struct pbl_chunk {
74 	dma_addr_t dma_addr;
75 	u64 *buf;
76 	u32 length;
77 };
78 
79 struct pbl_chunk_list {
80 	struct pbl_chunk *chunks;
81 	unsigned int size;
82 };
83 
84 struct pbl_context {
85 	union {
86 		struct {
87 			dma_addr_t dma_addr;
88 		} continuous;
89 		struct {
90 			u32 pbl_buf_size_in_pages;
91 			struct scatterlist *sgl;
92 			int sg_dma_cnt;
93 			struct pbl_chunk_list chunk_list;
94 		} indirect;
95 	} phys;
96 	u64 *pbl_buf;
97 	u32 pbl_buf_size_in_bytes;
98 	u8 physically_continuous;
99 };
100 
101 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
102 {
103 	return container_of(ibdev, struct efa_dev, ibdev);
104 }
105 
106 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
107 {
108 	return container_of(ibucontext, struct efa_ucontext, ibucontext);
109 }
110 
111 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
112 {
113 	return container_of(ibpd, struct efa_pd, ibpd);
114 }
115 
116 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
117 {
118 	return container_of(ibmr, struct efa_mr, ibmr);
119 }
120 
121 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
122 {
123 	return container_of(ibqp, struct efa_qp, ibqp);
124 }
125 
126 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
127 {
128 	return container_of(ibcq, struct efa_cq, ibcq);
129 }
130 
131 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
132 {
133 	return container_of(ibah, struct efa_ah, ibah);
134 }
135 
136 static inline struct efa_user_mmap_entry *
137 to_emmap(struct rdma_user_mmap_entry *rdma_entry)
138 {
139 	return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
140 }
141 
142 static inline bool is_rdma_read_cap(struct efa_dev *dev)
143 {
144 	return dev->dev_attr.device_caps & EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RDMA_READ_MASK;
145 }
146 
147 #define field_avail(x, fld, sz) (offsetof(typeof(x), fld) + \
148 				 sizeof_field(typeof(x), fld) <= (sz))
149 
150 #define is_reserved_cleared(reserved) \
151 	!memchr_inv(reserved, 0, sizeof(reserved))
152 
153 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
154 			       size_t size, enum dma_data_direction dir)
155 {
156 	void *addr;
157 
158 	addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
159 	if (!addr)
160 		return NULL;
161 
162 	*dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
163 	if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
164 		ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
165 		free_pages_exact(addr, size);
166 		return NULL;
167 	}
168 
169 	return addr;
170 }
171 
172 int efa_query_device(struct ib_device *ibdev,
173 		     struct ib_device_attr *props,
174 		     struct ib_udata *udata)
175 {
176 	struct efa_com_get_device_attr_result *dev_attr;
177 	struct efa_ibv_ex_query_device_resp resp = {};
178 	struct efa_dev *dev = to_edev(ibdev);
179 	int err;
180 
181 	if (udata && udata->inlen &&
182 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
183 		ibdev_dbg(ibdev,
184 			  "Incompatible ABI params, udata not cleared\n");
185 		return -EINVAL;
186 	}
187 
188 	dev_attr = &dev->dev_attr;
189 
190 	memset(props, 0, sizeof(*props));
191 	props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
192 	props->page_size_cap = dev_attr->page_size_cap;
193 	props->vendor_id = dev->pdev->vendor;
194 	props->vendor_part_id = dev->pdev->device;
195 	props->hw_ver = dev->pdev->subsystem_device;
196 	props->max_qp = dev_attr->max_qp;
197 	props->max_cq = dev_attr->max_cq;
198 	props->max_pd = dev_attr->max_pd;
199 	props->max_mr = dev_attr->max_mr;
200 	props->max_ah = dev_attr->max_ah;
201 	props->max_cqe = dev_attr->max_cq_depth;
202 	props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
203 				 dev_attr->max_rq_depth);
204 	props->max_send_sge = dev_attr->max_sq_sge;
205 	props->max_recv_sge = dev_attr->max_rq_sge;
206 	props->max_sge_rd = dev_attr->max_wr_rdma_sge;
207 
208 	if (udata && udata->outlen) {
209 		resp.max_sq_sge = dev_attr->max_sq_sge;
210 		resp.max_rq_sge = dev_attr->max_rq_sge;
211 		resp.max_sq_wr = dev_attr->max_sq_depth;
212 		resp.max_rq_wr = dev_attr->max_rq_depth;
213 		resp.max_rdma_size = dev_attr->max_rdma_size;
214 
215 		if (is_rdma_read_cap(dev))
216 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
217 
218 		err = ib_copy_to_udata(udata, &resp,
219 				       min(sizeof(resp), udata->outlen));
220 		if (err) {
221 			ibdev_dbg(ibdev,
222 				  "Failed to copy udata for query_device\n");
223 			return err;
224 		}
225 	}
226 
227 	return 0;
228 }
229 
230 int efa_query_port(struct ib_device *ibdev, u8 port,
231 		   struct ib_port_attr *props)
232 {
233 	struct efa_dev *dev = to_edev(ibdev);
234 
235 	props->lmc = 1;
236 
237 	props->state = IB_PORT_ACTIVE;
238 	props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
239 	props->gid_tbl_len = 1;
240 	props->pkey_tbl_len = 1;
241 	props->active_speed = IB_SPEED_EDR;
242 	props->active_width = IB_WIDTH_4X;
243 	props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
244 	props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
245 	props->max_msg_sz = dev->dev_attr.mtu;
246 	props->max_vl_num = 1;
247 
248 	return 0;
249 }
250 
251 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
252 		 int qp_attr_mask,
253 		 struct ib_qp_init_attr *qp_init_attr)
254 {
255 	struct efa_dev *dev = to_edev(ibqp->device);
256 	struct efa_com_query_qp_params params = {};
257 	struct efa_com_query_qp_result result;
258 	struct efa_qp *qp = to_eqp(ibqp);
259 	int err;
260 
261 #define EFA_QUERY_QP_SUPP_MASK \
262 	(IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
263 	 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP)
264 
265 	if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
266 		ibdev_dbg(&dev->ibdev,
267 			  "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
268 			  qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
269 		return -EOPNOTSUPP;
270 	}
271 
272 	memset(qp_attr, 0, sizeof(*qp_attr));
273 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
274 
275 	params.qp_handle = qp->qp_handle;
276 	err = efa_com_query_qp(&dev->edev, &params, &result);
277 	if (err)
278 		return err;
279 
280 	qp_attr->qp_state = result.qp_state;
281 	qp_attr->qkey = result.qkey;
282 	qp_attr->sq_psn = result.sq_psn;
283 	qp_attr->sq_draining = result.sq_draining;
284 	qp_attr->port_num = 1;
285 
286 	qp_attr->cap.max_send_wr = qp->max_send_wr;
287 	qp_attr->cap.max_recv_wr = qp->max_recv_wr;
288 	qp_attr->cap.max_send_sge = qp->max_send_sge;
289 	qp_attr->cap.max_recv_sge = qp->max_recv_sge;
290 	qp_attr->cap.max_inline_data = qp->max_inline_data;
291 
292 	qp_init_attr->qp_type = ibqp->qp_type;
293 	qp_init_attr->recv_cq = ibqp->recv_cq;
294 	qp_init_attr->send_cq = ibqp->send_cq;
295 	qp_init_attr->qp_context = ibqp->qp_context;
296 	qp_init_attr->cap = qp_attr->cap;
297 
298 	return 0;
299 }
300 
301 int efa_query_gid(struct ib_device *ibdev, u8 port, int index,
302 		  union ib_gid *gid)
303 {
304 	struct efa_dev *dev = to_edev(ibdev);
305 
306 	memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
307 
308 	return 0;
309 }
310 
311 int efa_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
312 		   u16 *pkey)
313 {
314 	if (index > 0)
315 		return -EINVAL;
316 
317 	*pkey = 0xffff;
318 	return 0;
319 }
320 
321 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
322 {
323 	struct efa_com_dealloc_pd_params params = {
324 		.pdn = pdn,
325 	};
326 
327 	return efa_com_dealloc_pd(&dev->edev, &params);
328 }
329 
330 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
331 {
332 	struct efa_dev *dev = to_edev(ibpd->device);
333 	struct efa_ibv_alloc_pd_resp resp = {};
334 	struct efa_com_alloc_pd_result result;
335 	struct efa_pd *pd = to_epd(ibpd);
336 	int err;
337 
338 	if (udata->inlen &&
339 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
340 		ibdev_dbg(&dev->ibdev,
341 			  "Incompatible ABI params, udata not cleared\n");
342 		err = -EINVAL;
343 		goto err_out;
344 	}
345 
346 	err = efa_com_alloc_pd(&dev->edev, &result);
347 	if (err)
348 		goto err_out;
349 
350 	pd->pdn = result.pdn;
351 	resp.pdn = result.pdn;
352 
353 	if (udata->outlen) {
354 		err = ib_copy_to_udata(udata, &resp,
355 				       min(sizeof(resp), udata->outlen));
356 		if (err) {
357 			ibdev_dbg(&dev->ibdev,
358 				  "Failed to copy udata for alloc_pd\n");
359 			goto err_dealloc_pd;
360 		}
361 	}
362 
363 	ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
364 
365 	return 0;
366 
367 err_dealloc_pd:
368 	efa_pd_dealloc(dev, result.pdn);
369 err_out:
370 	atomic64_inc(&dev->stats.sw_stats.alloc_pd_err);
371 	return err;
372 }
373 
374 void efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
375 {
376 	struct efa_dev *dev = to_edev(ibpd->device);
377 	struct efa_pd *pd = to_epd(ibpd);
378 
379 	ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
380 	efa_pd_dealloc(dev, pd->pdn);
381 }
382 
383 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
384 {
385 	struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
386 
387 	return efa_com_destroy_qp(&dev->edev, &params);
388 }
389 
390 static void efa_qp_user_mmap_entries_remove(struct efa_ucontext *uctx,
391 					    struct efa_qp *qp)
392 {
393 	rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
394 	rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
395 	rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
396 	rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
397 }
398 
399 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
400 {
401 	struct efa_ucontext *ucontext = rdma_udata_to_drv_context(udata,
402 		struct efa_ucontext, ibucontext);
403 	struct efa_dev *dev = to_edev(ibqp->pd->device);
404 	struct efa_qp *qp = to_eqp(ibqp);
405 	int err;
406 
407 	ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
408 	err = efa_destroy_qp_handle(dev, qp->qp_handle);
409 	if (err)
410 		return err;
411 
412 	if (qp->rq_cpu_addr) {
413 		ibdev_dbg(&dev->ibdev,
414 			  "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
415 			  qp->rq_cpu_addr, qp->rq_size,
416 			  &qp->rq_dma_addr);
417 		dma_unmap_single(&dev->pdev->dev, qp->rq_dma_addr, qp->rq_size,
418 				 DMA_TO_DEVICE);
419 	}
420 
421 	efa_qp_user_mmap_entries_remove(ucontext, qp);
422 	kfree(qp);
423 	return 0;
424 }
425 
426 static struct rdma_user_mmap_entry*
427 efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
428 			   u64 address, size_t length,
429 			   u8 mmap_flag, u64 *offset)
430 {
431 	struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
432 	int err;
433 
434 	if (!entry)
435 		return NULL;
436 
437 	entry->address = address;
438 	entry->mmap_flag = mmap_flag;
439 
440 	err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
441 					  length);
442 	if (err) {
443 		kfree(entry);
444 		return NULL;
445 	}
446 	*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
447 
448 	return &entry->rdma_entry;
449 }
450 
451 static int qp_mmap_entries_setup(struct efa_qp *qp,
452 				 struct efa_dev *dev,
453 				 struct efa_ucontext *ucontext,
454 				 struct efa_com_create_qp_params *params,
455 				 struct efa_ibv_create_qp_resp *resp)
456 {
457 	size_t length;
458 	u64 address;
459 
460 	address = dev->db_bar_addr + resp->sq_db_offset;
461 	qp->sq_db_mmap_entry =
462 		efa_user_mmap_entry_insert(&ucontext->ibucontext,
463 					   address,
464 					   PAGE_SIZE, EFA_MMAP_IO_NC,
465 					   &resp->sq_db_mmap_key);
466 	if (!qp->sq_db_mmap_entry)
467 		return -ENOMEM;
468 
469 	resp->sq_db_offset &= ~PAGE_MASK;
470 
471 	address = dev->mem_bar_addr + resp->llq_desc_offset;
472 	length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
473 			    (resp->llq_desc_offset & ~PAGE_MASK));
474 
475 	qp->llq_desc_mmap_entry =
476 		efa_user_mmap_entry_insert(&ucontext->ibucontext,
477 					   address, length,
478 					   EFA_MMAP_IO_WC,
479 					   &resp->llq_desc_mmap_key);
480 	if (!qp->llq_desc_mmap_entry)
481 		goto err_remove_mmap;
482 
483 	resp->llq_desc_offset &= ~PAGE_MASK;
484 
485 	if (qp->rq_size) {
486 		address = dev->db_bar_addr + resp->rq_db_offset;
487 
488 		qp->rq_db_mmap_entry =
489 			efa_user_mmap_entry_insert(&ucontext->ibucontext,
490 						   address, PAGE_SIZE,
491 						   EFA_MMAP_IO_NC,
492 						   &resp->rq_db_mmap_key);
493 		if (!qp->rq_db_mmap_entry)
494 			goto err_remove_mmap;
495 
496 		resp->rq_db_offset &= ~PAGE_MASK;
497 
498 		address = virt_to_phys(qp->rq_cpu_addr);
499 		qp->rq_mmap_entry =
500 			efa_user_mmap_entry_insert(&ucontext->ibucontext,
501 						   address, qp->rq_size,
502 						   EFA_MMAP_DMA_PAGE,
503 						   &resp->rq_mmap_key);
504 		if (!qp->rq_mmap_entry)
505 			goto err_remove_mmap;
506 
507 		resp->rq_mmap_size = qp->rq_size;
508 	}
509 
510 	return 0;
511 
512 err_remove_mmap:
513 	efa_qp_user_mmap_entries_remove(ucontext, qp);
514 
515 	return -ENOMEM;
516 }
517 
518 static int efa_qp_validate_cap(struct efa_dev *dev,
519 			       struct ib_qp_init_attr *init_attr)
520 {
521 	if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
522 		ibdev_dbg(&dev->ibdev,
523 			  "qp: requested send wr[%u] exceeds the max[%u]\n",
524 			  init_attr->cap.max_send_wr,
525 			  dev->dev_attr.max_sq_depth);
526 		return -EINVAL;
527 	}
528 	if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
529 		ibdev_dbg(&dev->ibdev,
530 			  "qp: requested receive wr[%u] exceeds the max[%u]\n",
531 			  init_attr->cap.max_recv_wr,
532 			  dev->dev_attr.max_rq_depth);
533 		return -EINVAL;
534 	}
535 	if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
536 		ibdev_dbg(&dev->ibdev,
537 			  "qp: requested sge send[%u] exceeds the max[%u]\n",
538 			  init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
539 		return -EINVAL;
540 	}
541 	if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
542 		ibdev_dbg(&dev->ibdev,
543 			  "qp: requested sge recv[%u] exceeds the max[%u]\n",
544 			  init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
545 		return -EINVAL;
546 	}
547 	if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
548 		ibdev_dbg(&dev->ibdev,
549 			  "qp: requested inline data[%u] exceeds the max[%u]\n",
550 			  init_attr->cap.max_inline_data,
551 			  dev->dev_attr.inline_buf_size);
552 		return -EINVAL;
553 	}
554 
555 	return 0;
556 }
557 
558 static int efa_qp_validate_attr(struct efa_dev *dev,
559 				struct ib_qp_init_attr *init_attr)
560 {
561 	if (init_attr->qp_type != IB_QPT_DRIVER &&
562 	    init_attr->qp_type != IB_QPT_UD) {
563 		ibdev_dbg(&dev->ibdev,
564 			  "Unsupported qp type %d\n", init_attr->qp_type);
565 		return -EOPNOTSUPP;
566 	}
567 
568 	if (init_attr->srq) {
569 		ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
570 		return -EOPNOTSUPP;
571 	}
572 
573 	if (init_attr->create_flags) {
574 		ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
575 		return -EOPNOTSUPP;
576 	}
577 
578 	return 0;
579 }
580 
581 struct ib_qp *efa_create_qp(struct ib_pd *ibpd,
582 			    struct ib_qp_init_attr *init_attr,
583 			    struct ib_udata *udata)
584 {
585 	struct efa_com_create_qp_params create_qp_params = {};
586 	struct efa_com_create_qp_result create_qp_resp;
587 	struct efa_dev *dev = to_edev(ibpd->device);
588 	struct efa_ibv_create_qp_resp resp = {};
589 	struct efa_ibv_create_qp cmd = {};
590 	struct efa_ucontext *ucontext;
591 	struct efa_qp *qp;
592 	int err;
593 
594 	ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
595 					     ibucontext);
596 
597 	err = efa_qp_validate_cap(dev, init_attr);
598 	if (err)
599 		goto err_out;
600 
601 	err = efa_qp_validate_attr(dev, init_attr);
602 	if (err)
603 		goto err_out;
604 
605 	if (!field_avail(cmd, driver_qp_type, udata->inlen)) {
606 		ibdev_dbg(&dev->ibdev,
607 			  "Incompatible ABI params, no input udata\n");
608 		err = -EINVAL;
609 		goto err_out;
610 	}
611 
612 	if (udata->inlen > sizeof(cmd) &&
613 	    !ib_is_udata_cleared(udata, sizeof(cmd),
614 				 udata->inlen - sizeof(cmd))) {
615 		ibdev_dbg(&dev->ibdev,
616 			  "Incompatible ABI params, unknown fields in udata\n");
617 		err = -EINVAL;
618 		goto err_out;
619 	}
620 
621 	err = ib_copy_from_udata(&cmd, udata,
622 				 min(sizeof(cmd), udata->inlen));
623 	if (err) {
624 		ibdev_dbg(&dev->ibdev,
625 			  "Cannot copy udata for create_qp\n");
626 		goto err_out;
627 	}
628 
629 	if (cmd.comp_mask) {
630 		ibdev_dbg(&dev->ibdev,
631 			  "Incompatible ABI params, unknown fields in udata\n");
632 		err = -EINVAL;
633 		goto err_out;
634 	}
635 
636 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
637 	if (!qp) {
638 		err = -ENOMEM;
639 		goto err_out;
640 	}
641 
642 	create_qp_params.uarn = ucontext->uarn;
643 	create_qp_params.pd = to_epd(ibpd)->pdn;
644 
645 	if (init_attr->qp_type == IB_QPT_UD) {
646 		create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
647 	} else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
648 		create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
649 	} else {
650 		ibdev_dbg(&dev->ibdev,
651 			  "Unsupported qp type %d driver qp type %d\n",
652 			  init_attr->qp_type, cmd.driver_qp_type);
653 		err = -EOPNOTSUPP;
654 		goto err_free_qp;
655 	}
656 
657 	ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
658 		  init_attr->qp_type, cmd.driver_qp_type);
659 	create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
660 	create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
661 	create_qp_params.sq_depth = init_attr->cap.max_send_wr;
662 	create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
663 
664 	create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
665 	create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
666 	qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
667 	if (qp->rq_size) {
668 		qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
669 						    qp->rq_size, DMA_TO_DEVICE);
670 		if (!qp->rq_cpu_addr) {
671 			err = -ENOMEM;
672 			goto err_free_qp;
673 		}
674 
675 		ibdev_dbg(&dev->ibdev,
676 			  "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
677 			  qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
678 		create_qp_params.rq_base_addr = qp->rq_dma_addr;
679 	}
680 
681 	err = efa_com_create_qp(&dev->edev, &create_qp_params,
682 				&create_qp_resp);
683 	if (err)
684 		goto err_free_mapped;
685 
686 	resp.sq_db_offset = create_qp_resp.sq_db_offset;
687 	resp.rq_db_offset = create_qp_resp.rq_db_offset;
688 	resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
689 	resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
690 	resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
691 
692 	err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
693 				    &resp);
694 	if (err)
695 		goto err_destroy_qp;
696 
697 	qp->qp_handle = create_qp_resp.qp_handle;
698 	qp->ibqp.qp_num = create_qp_resp.qp_num;
699 	qp->ibqp.qp_type = init_attr->qp_type;
700 	qp->max_send_wr = init_attr->cap.max_send_wr;
701 	qp->max_recv_wr = init_attr->cap.max_recv_wr;
702 	qp->max_send_sge = init_attr->cap.max_send_sge;
703 	qp->max_recv_sge = init_attr->cap.max_recv_sge;
704 	qp->max_inline_data = init_attr->cap.max_inline_data;
705 
706 	if (udata->outlen) {
707 		err = ib_copy_to_udata(udata, &resp,
708 				       min(sizeof(resp), udata->outlen));
709 		if (err) {
710 			ibdev_dbg(&dev->ibdev,
711 				  "Failed to copy udata for qp[%u]\n",
712 				  create_qp_resp.qp_num);
713 			goto err_remove_mmap_entries;
714 		}
715 	}
716 
717 	ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
718 
719 	return &qp->ibqp;
720 
721 err_remove_mmap_entries:
722 	efa_qp_user_mmap_entries_remove(ucontext, qp);
723 err_destroy_qp:
724 	efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
725 err_free_mapped:
726 	if (qp->rq_size) {
727 		dma_unmap_single(&dev->pdev->dev, qp->rq_dma_addr, qp->rq_size,
728 				 DMA_TO_DEVICE);
729 
730 		if (!qp->rq_mmap_entry)
731 			free_pages_exact(qp->rq_cpu_addr, qp->rq_size);
732 	}
733 err_free_qp:
734 	kfree(qp);
735 err_out:
736 	atomic64_inc(&dev->stats.sw_stats.create_qp_err);
737 	return ERR_PTR(err);
738 }
739 
740 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
741 				  struct ib_qp_attr *qp_attr, int qp_attr_mask,
742 				  enum ib_qp_state cur_state,
743 				  enum ib_qp_state new_state)
744 {
745 #define EFA_MODIFY_QP_SUPP_MASK \
746 	(IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
747 	 IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN)
748 
749 	if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
750 		ibdev_dbg(&dev->ibdev,
751 			  "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
752 			  qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
753 		return -EOPNOTSUPP;
754 	}
755 
756 	if (!ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
757 				qp_attr_mask)) {
758 		ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
759 		return -EINVAL;
760 	}
761 
762 	if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
763 		ibdev_dbg(&dev->ibdev, "Can't change port num\n");
764 		return -EOPNOTSUPP;
765 	}
766 
767 	if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
768 		ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
769 		return -EOPNOTSUPP;
770 	}
771 
772 	return 0;
773 }
774 
775 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
776 		  int qp_attr_mask, struct ib_udata *udata)
777 {
778 	struct efa_dev *dev = to_edev(ibqp->device);
779 	struct efa_com_modify_qp_params params = {};
780 	struct efa_qp *qp = to_eqp(ibqp);
781 	enum ib_qp_state cur_state;
782 	enum ib_qp_state new_state;
783 	int err;
784 
785 	if (udata->inlen &&
786 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
787 		ibdev_dbg(&dev->ibdev,
788 			  "Incompatible ABI params, udata not cleared\n");
789 		return -EINVAL;
790 	}
791 
792 	cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
793 						     qp->state;
794 	new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
795 
796 	err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
797 				     new_state);
798 	if (err)
799 		return err;
800 
801 	params.qp_handle = qp->qp_handle;
802 
803 	if (qp_attr_mask & IB_QP_STATE) {
804 		params.modify_mask |= BIT(EFA_ADMIN_QP_STATE_BIT) |
805 				      BIT(EFA_ADMIN_CUR_QP_STATE_BIT);
806 		params.cur_qp_state = qp_attr->cur_qp_state;
807 		params.qp_state = qp_attr->qp_state;
808 	}
809 
810 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
811 		params.modify_mask |=
812 			BIT(EFA_ADMIN_SQ_DRAINED_ASYNC_NOTIFY_BIT);
813 		params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
814 	}
815 
816 	if (qp_attr_mask & IB_QP_QKEY) {
817 		params.modify_mask |= BIT(EFA_ADMIN_QKEY_BIT);
818 		params.qkey = qp_attr->qkey;
819 	}
820 
821 	if (qp_attr_mask & IB_QP_SQ_PSN) {
822 		params.modify_mask |= BIT(EFA_ADMIN_SQ_PSN_BIT);
823 		params.sq_psn = qp_attr->sq_psn;
824 	}
825 
826 	err = efa_com_modify_qp(&dev->edev, &params);
827 	if (err)
828 		return err;
829 
830 	qp->state = new_state;
831 
832 	return 0;
833 }
834 
835 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
836 {
837 	struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
838 
839 	return efa_com_destroy_cq(&dev->edev, &params);
840 }
841 
842 void efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
843 {
844 	struct efa_dev *dev = to_edev(ibcq->device);
845 	struct efa_cq *cq = to_ecq(ibcq);
846 
847 	ibdev_dbg(&dev->ibdev,
848 		  "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
849 		  cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
850 
851 	efa_destroy_cq_idx(dev, cq->cq_idx);
852 	dma_unmap_single(&dev->pdev->dev, cq->dma_addr, cq->size,
853 			 DMA_FROM_DEVICE);
854 	rdma_user_mmap_entry_remove(cq->mmap_entry);
855 }
856 
857 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
858 				 struct efa_ibv_create_cq_resp *resp)
859 {
860 	resp->q_mmap_size = cq->size;
861 	cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
862 						    virt_to_phys(cq->cpu_addr),
863 						    cq->size, EFA_MMAP_DMA_PAGE,
864 						    &resp->q_mmap_key);
865 	if (!cq->mmap_entry)
866 		return -ENOMEM;
867 
868 	return 0;
869 }
870 
871 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
872 		  struct ib_udata *udata)
873 {
874 	struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
875 		udata, struct efa_ucontext, ibucontext);
876 	struct efa_ibv_create_cq_resp resp = {};
877 	struct efa_com_create_cq_params params;
878 	struct efa_com_create_cq_result result;
879 	struct ib_device *ibdev = ibcq->device;
880 	struct efa_dev *dev = to_edev(ibdev);
881 	struct efa_ibv_create_cq cmd = {};
882 	struct efa_cq *cq = to_ecq(ibcq);
883 	int entries = attr->cqe;
884 	int err;
885 
886 	ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
887 
888 	if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
889 		ibdev_dbg(ibdev,
890 			  "cq: requested entries[%u] non-positive or greater than max[%u]\n",
891 			  entries, dev->dev_attr.max_cq_depth);
892 		err = -EINVAL;
893 		goto err_out;
894 	}
895 
896 	if (!field_avail(cmd, num_sub_cqs, udata->inlen)) {
897 		ibdev_dbg(ibdev,
898 			  "Incompatible ABI params, no input udata\n");
899 		err = -EINVAL;
900 		goto err_out;
901 	}
902 
903 	if (udata->inlen > sizeof(cmd) &&
904 	    !ib_is_udata_cleared(udata, sizeof(cmd),
905 				 udata->inlen - sizeof(cmd))) {
906 		ibdev_dbg(ibdev,
907 			  "Incompatible ABI params, unknown fields in udata\n");
908 		err = -EINVAL;
909 		goto err_out;
910 	}
911 
912 	err = ib_copy_from_udata(&cmd, udata,
913 				 min(sizeof(cmd), udata->inlen));
914 	if (err) {
915 		ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
916 		goto err_out;
917 	}
918 
919 	if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) {
920 		ibdev_dbg(ibdev,
921 			  "Incompatible ABI params, unknown fields in udata\n");
922 		err = -EINVAL;
923 		goto err_out;
924 	}
925 
926 	if (!cmd.cq_entry_size) {
927 		ibdev_dbg(ibdev,
928 			  "Invalid entry size [%u]\n", cmd.cq_entry_size);
929 		err = -EINVAL;
930 		goto err_out;
931 	}
932 
933 	if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
934 		ibdev_dbg(ibdev,
935 			  "Invalid number of sub cqs[%u] expected[%u]\n",
936 			  cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
937 		err = -EINVAL;
938 		goto err_out;
939 	}
940 
941 	cq->ucontext = ucontext;
942 	cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
943 	cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
944 					 DMA_FROM_DEVICE);
945 	if (!cq->cpu_addr) {
946 		err = -ENOMEM;
947 		goto err_out;
948 	}
949 
950 	params.uarn = cq->ucontext->uarn;
951 	params.cq_depth = entries;
952 	params.dma_addr = cq->dma_addr;
953 	params.entry_size_in_bytes = cmd.cq_entry_size;
954 	params.num_sub_cqs = cmd.num_sub_cqs;
955 	err = efa_com_create_cq(&dev->edev, &params, &result);
956 	if (err)
957 		goto err_free_mapped;
958 
959 	resp.cq_idx = result.cq_idx;
960 	cq->cq_idx = result.cq_idx;
961 	cq->ibcq.cqe = result.actual_depth;
962 	WARN_ON_ONCE(entries != result.actual_depth);
963 
964 	err = cq_mmap_entries_setup(dev, cq, &resp);
965 	if (err) {
966 		ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
967 			  cq->cq_idx);
968 		goto err_destroy_cq;
969 	}
970 
971 	if (udata->outlen) {
972 		err = ib_copy_to_udata(udata, &resp,
973 				       min(sizeof(resp), udata->outlen));
974 		if (err) {
975 			ibdev_dbg(ibdev,
976 				  "Failed to copy udata for create_cq\n");
977 			goto err_remove_mmap;
978 		}
979 	}
980 
981 	ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
982 		  cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
983 
984 	return 0;
985 
986 err_remove_mmap:
987 	rdma_user_mmap_entry_remove(cq->mmap_entry);
988 err_destroy_cq:
989 	efa_destroy_cq_idx(dev, cq->cq_idx);
990 err_free_mapped:
991 	dma_unmap_single(&dev->pdev->dev, cq->dma_addr, cq->size,
992 			 DMA_FROM_DEVICE);
993 	if (!cq->mmap_entry)
994 		free_pages_exact(cq->cpu_addr, cq->size);
995 
996 err_out:
997 	atomic64_inc(&dev->stats.sw_stats.create_cq_err);
998 	return err;
999 }
1000 
1001 static int umem_to_page_list(struct efa_dev *dev,
1002 			     struct ib_umem *umem,
1003 			     u64 *page_list,
1004 			     u32 hp_cnt,
1005 			     u8 hp_shift)
1006 {
1007 	u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1008 	struct ib_block_iter biter;
1009 	unsigned int hp_idx = 0;
1010 
1011 	ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1012 		  hp_cnt, pages_in_hp);
1013 
1014 	rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap,
1015 			    BIT(hp_shift))
1016 		page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1017 
1018 	return 0;
1019 }
1020 
1021 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1022 {
1023 	struct scatterlist *sglist;
1024 	struct page *pg;
1025 	int i;
1026 
1027 	sglist = kcalloc(page_cnt, sizeof(*sglist), GFP_KERNEL);
1028 	if (!sglist)
1029 		return NULL;
1030 	sg_init_table(sglist, page_cnt);
1031 	for (i = 0; i < page_cnt; i++) {
1032 		pg = vmalloc_to_page(buf);
1033 		if (!pg)
1034 			goto err;
1035 		sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1036 		buf += PAGE_SIZE / sizeof(*buf);
1037 	}
1038 	return sglist;
1039 
1040 err:
1041 	kfree(sglist);
1042 	return NULL;
1043 }
1044 
1045 /*
1046  * create a chunk list of physical pages dma addresses from the supplied
1047  * scatter gather list
1048  */
1049 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1050 {
1051 	struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1052 	int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1053 	struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1054 	unsigned int chunk_list_size, chunk_idx, payload_idx;
1055 	int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1056 	struct efa_com_ctrl_buff_info *ctrl_buf;
1057 	u64 *cur_chunk_buf, *prev_chunk_buf;
1058 	struct ib_block_iter biter;
1059 	dma_addr_t dma_addr;
1060 	int i;
1061 
1062 	/* allocate a chunk list that consists of 4KB chunks */
1063 	chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1064 
1065 	chunk_list->size = chunk_list_size;
1066 	chunk_list->chunks = kcalloc(chunk_list_size,
1067 				     sizeof(*chunk_list->chunks),
1068 				     GFP_KERNEL);
1069 	if (!chunk_list->chunks)
1070 		return -ENOMEM;
1071 
1072 	ibdev_dbg(&dev->ibdev,
1073 		  "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1074 		  page_cnt);
1075 
1076 	/* allocate chunk buffers: */
1077 	for (i = 0; i < chunk_list_size; i++) {
1078 		chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1079 		if (!chunk_list->chunks[i].buf)
1080 			goto chunk_list_dealloc;
1081 
1082 		chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1083 	}
1084 	chunk_list->chunks[chunk_list_size - 1].length =
1085 		((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1086 			EFA_CHUNK_PTR_SIZE;
1087 
1088 	/* fill the dma addresses of sg list pages to chunks: */
1089 	chunk_idx = 0;
1090 	payload_idx = 0;
1091 	cur_chunk_buf = chunk_list->chunks[0].buf;
1092 	rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1093 			    EFA_CHUNK_PAYLOAD_SIZE) {
1094 		cur_chunk_buf[payload_idx++] =
1095 			rdma_block_iter_dma_address(&biter);
1096 
1097 		if (payload_idx == EFA_PTRS_PER_CHUNK) {
1098 			chunk_idx++;
1099 			cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1100 			payload_idx = 0;
1101 		}
1102 	}
1103 
1104 	/* map chunks to dma and fill chunks next ptrs */
1105 	for (i = chunk_list_size - 1; i >= 0; i--) {
1106 		dma_addr = dma_map_single(&dev->pdev->dev,
1107 					  chunk_list->chunks[i].buf,
1108 					  chunk_list->chunks[i].length,
1109 					  DMA_TO_DEVICE);
1110 		if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1111 			ibdev_err(&dev->ibdev,
1112 				  "chunk[%u] dma_map_failed\n", i);
1113 			goto chunk_list_unmap;
1114 		}
1115 
1116 		chunk_list->chunks[i].dma_addr = dma_addr;
1117 		ibdev_dbg(&dev->ibdev,
1118 			  "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1119 
1120 		if (!i)
1121 			break;
1122 
1123 		prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1124 
1125 		ctrl_buf = (struct efa_com_ctrl_buff_info *)
1126 				&prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1127 		ctrl_buf->length = chunk_list->chunks[i].length;
1128 
1129 		efa_com_set_dma_addr(dma_addr,
1130 				     &ctrl_buf->address.mem_addr_high,
1131 				     &ctrl_buf->address.mem_addr_low);
1132 	}
1133 
1134 	return 0;
1135 
1136 chunk_list_unmap:
1137 	for (; i < chunk_list_size; i++) {
1138 		dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1139 				 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1140 	}
1141 chunk_list_dealloc:
1142 	for (i = 0; i < chunk_list_size; i++)
1143 		kfree(chunk_list->chunks[i].buf);
1144 
1145 	kfree(chunk_list->chunks);
1146 	return -ENOMEM;
1147 }
1148 
1149 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1150 {
1151 	struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1152 	int i;
1153 
1154 	for (i = 0; i < chunk_list->size; i++) {
1155 		dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1156 				 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1157 		kfree(chunk_list->chunks[i].buf);
1158 	}
1159 
1160 	kfree(chunk_list->chunks);
1161 }
1162 
1163 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1164 static int pbl_continuous_initialize(struct efa_dev *dev,
1165 				     struct pbl_context *pbl)
1166 {
1167 	dma_addr_t dma_addr;
1168 
1169 	dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1170 				  pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1171 	if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1172 		ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1173 		return -ENOMEM;
1174 	}
1175 
1176 	pbl->phys.continuous.dma_addr = dma_addr;
1177 	ibdev_dbg(&dev->ibdev,
1178 		  "pbl continuous - dma_addr = %pad, size[%u]\n",
1179 		  &dma_addr, pbl->pbl_buf_size_in_bytes);
1180 
1181 	return 0;
1182 }
1183 
1184 /*
1185  * initialize pbl indirect mode:
1186  * create a chunk list out of the dma addresses of the physical pages of
1187  * pbl buffer.
1188  */
1189 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1190 {
1191 	u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1192 	struct scatterlist *sgl;
1193 	int sg_dma_cnt, err;
1194 
1195 	BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1196 	sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1197 	if (!sgl)
1198 		return -ENOMEM;
1199 
1200 	sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1201 	if (!sg_dma_cnt) {
1202 		err = -EINVAL;
1203 		goto err_map;
1204 	}
1205 
1206 	pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1207 	pbl->phys.indirect.sgl = sgl;
1208 	pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1209 	err = pbl_chunk_list_create(dev, pbl);
1210 	if (err) {
1211 		ibdev_dbg(&dev->ibdev,
1212 			  "chunk_list creation failed[%d]\n", err);
1213 		goto err_chunk;
1214 	}
1215 
1216 	ibdev_dbg(&dev->ibdev,
1217 		  "pbl indirect - size[%u], chunks[%u]\n",
1218 		  pbl->pbl_buf_size_in_bytes,
1219 		  pbl->phys.indirect.chunk_list.size);
1220 
1221 	return 0;
1222 
1223 err_chunk:
1224 	dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1225 err_map:
1226 	kfree(sgl);
1227 	return err;
1228 }
1229 
1230 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1231 {
1232 	pbl_chunk_list_destroy(dev, pbl);
1233 	dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1234 		     pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1235 	kfree(pbl->phys.indirect.sgl);
1236 }
1237 
1238 /* create a page buffer list from a mapped user memory region */
1239 static int pbl_create(struct efa_dev *dev,
1240 		      struct pbl_context *pbl,
1241 		      struct ib_umem *umem,
1242 		      int hp_cnt,
1243 		      u8 hp_shift)
1244 {
1245 	int err;
1246 
1247 	pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1248 	pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1249 	if (!pbl->pbl_buf)
1250 		return -ENOMEM;
1251 
1252 	if (is_vmalloc_addr(pbl->pbl_buf)) {
1253 		pbl->physically_continuous = 0;
1254 		err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1255 					hp_shift);
1256 		if (err)
1257 			goto err_free;
1258 
1259 		err = pbl_indirect_initialize(dev, pbl);
1260 		if (err)
1261 			goto err_free;
1262 	} else {
1263 		pbl->physically_continuous = 1;
1264 		err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1265 					hp_shift);
1266 		if (err)
1267 			goto err_free;
1268 
1269 		err = pbl_continuous_initialize(dev, pbl);
1270 		if (err)
1271 			goto err_free;
1272 	}
1273 
1274 	ibdev_dbg(&dev->ibdev,
1275 		  "user_pbl_created: user_pages[%u], continuous[%u]\n",
1276 		  hp_cnt, pbl->physically_continuous);
1277 
1278 	return 0;
1279 
1280 err_free:
1281 	kvfree(pbl->pbl_buf);
1282 	return err;
1283 }
1284 
1285 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1286 {
1287 	if (pbl->physically_continuous)
1288 		dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1289 				 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1290 	else
1291 		pbl_indirect_terminate(dev, pbl);
1292 
1293 	kvfree(pbl->pbl_buf);
1294 }
1295 
1296 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1297 				 struct efa_com_reg_mr_params *params)
1298 {
1299 	int err;
1300 
1301 	params->inline_pbl = 1;
1302 	err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1303 				params->page_num, params->page_shift);
1304 	if (err)
1305 		return err;
1306 
1307 	ibdev_dbg(&dev->ibdev,
1308 		  "inline_pbl_array - pages[%u]\n", params->page_num);
1309 
1310 	return 0;
1311 }
1312 
1313 static int efa_create_pbl(struct efa_dev *dev,
1314 			  struct pbl_context *pbl,
1315 			  struct efa_mr *mr,
1316 			  struct efa_com_reg_mr_params *params)
1317 {
1318 	int err;
1319 
1320 	err = pbl_create(dev, pbl, mr->umem, params->page_num,
1321 			 params->page_shift);
1322 	if (err) {
1323 		ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1324 		return err;
1325 	}
1326 
1327 	params->inline_pbl = 0;
1328 	params->indirect = !pbl->physically_continuous;
1329 	if (pbl->physically_continuous) {
1330 		params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1331 
1332 		efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1333 				     &params->pbl.pbl.address.mem_addr_high,
1334 				     &params->pbl.pbl.address.mem_addr_low);
1335 	} else {
1336 		params->pbl.pbl.length =
1337 			pbl->phys.indirect.chunk_list.chunks[0].length;
1338 
1339 		efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1340 				     &params->pbl.pbl.address.mem_addr_high,
1341 				     &params->pbl.pbl.address.mem_addr_low);
1342 	}
1343 
1344 	return 0;
1345 }
1346 
1347 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1348 			 u64 virt_addr, int access_flags,
1349 			 struct ib_udata *udata)
1350 {
1351 	struct efa_dev *dev = to_edev(ibpd->device);
1352 	struct efa_com_reg_mr_params params = {};
1353 	struct efa_com_reg_mr_result result = {};
1354 	struct pbl_context pbl;
1355 	int supp_access_flags;
1356 	unsigned int pg_sz;
1357 	struct efa_mr *mr;
1358 	int inline_size;
1359 	int err;
1360 
1361 	if (udata->inlen &&
1362 	    !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1363 		ibdev_dbg(&dev->ibdev,
1364 			  "Incompatible ABI params, udata not cleared\n");
1365 		err = -EINVAL;
1366 		goto err_out;
1367 	}
1368 
1369 	supp_access_flags =
1370 		IB_ACCESS_LOCAL_WRITE |
1371 		(is_rdma_read_cap(dev) ? IB_ACCESS_REMOTE_READ : 0);
1372 
1373 	if (access_flags & ~supp_access_flags) {
1374 		ibdev_dbg(&dev->ibdev,
1375 			  "Unsupported access flags[%#x], supported[%#x]\n",
1376 			  access_flags, supp_access_flags);
1377 		err = -EOPNOTSUPP;
1378 		goto err_out;
1379 	}
1380 
1381 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1382 	if (!mr) {
1383 		err = -ENOMEM;
1384 		goto err_out;
1385 	}
1386 
1387 	mr->umem = ib_umem_get(udata, start, length, access_flags);
1388 	if (IS_ERR(mr->umem)) {
1389 		err = PTR_ERR(mr->umem);
1390 		ibdev_dbg(&dev->ibdev,
1391 			  "Failed to pin and map user space memory[%d]\n", err);
1392 		goto err_free;
1393 	}
1394 
1395 	params.pd = to_epd(ibpd)->pdn;
1396 	params.iova = virt_addr;
1397 	params.mr_length_in_bytes = length;
1398 	params.permissions = access_flags;
1399 
1400 	pg_sz = ib_umem_find_best_pgsz(mr->umem,
1401 				       dev->dev_attr.page_size_cap,
1402 				       virt_addr);
1403 	if (!pg_sz) {
1404 		err = -EOPNOTSUPP;
1405 		ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1406 			  dev->dev_attr.page_size_cap);
1407 		goto err_unmap;
1408 	}
1409 
1410 	params.page_shift = __ffs(pg_sz);
1411 	params.page_num = DIV_ROUND_UP(length + (start & (pg_sz - 1)),
1412 				       pg_sz);
1413 
1414 	ibdev_dbg(&dev->ibdev,
1415 		  "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1416 		  start, length, params.page_shift, params.page_num);
1417 
1418 	inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1419 	if (params.page_num <= inline_size) {
1420 		err = efa_create_inline_pbl(dev, mr, &params);
1421 		if (err)
1422 			goto err_unmap;
1423 
1424 		err = efa_com_register_mr(&dev->edev, &params, &result);
1425 		if (err)
1426 			goto err_unmap;
1427 	} else {
1428 		err = efa_create_pbl(dev, &pbl, mr, &params);
1429 		if (err)
1430 			goto err_unmap;
1431 
1432 		err = efa_com_register_mr(&dev->edev, &params, &result);
1433 		pbl_destroy(dev, &pbl);
1434 
1435 		if (err)
1436 			goto err_unmap;
1437 	}
1438 
1439 	mr->ibmr.lkey = result.l_key;
1440 	mr->ibmr.rkey = result.r_key;
1441 	mr->ibmr.length = length;
1442 	ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1443 
1444 	return &mr->ibmr;
1445 
1446 err_unmap:
1447 	ib_umem_release(mr->umem);
1448 err_free:
1449 	kfree(mr);
1450 err_out:
1451 	atomic64_inc(&dev->stats.sw_stats.reg_mr_err);
1452 	return ERR_PTR(err);
1453 }
1454 
1455 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1456 {
1457 	struct efa_dev *dev = to_edev(ibmr->device);
1458 	struct efa_com_dereg_mr_params params;
1459 	struct efa_mr *mr = to_emr(ibmr);
1460 	int err;
1461 
1462 	ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1463 
1464 	params.l_key = mr->ibmr.lkey;
1465 	err = efa_com_dereg_mr(&dev->edev, &params);
1466 	if (err)
1467 		return err;
1468 
1469 	ib_umem_release(mr->umem);
1470 	kfree(mr);
1471 
1472 	return 0;
1473 }
1474 
1475 int efa_get_port_immutable(struct ib_device *ibdev, u8 port_num,
1476 			   struct ib_port_immutable *immutable)
1477 {
1478 	struct ib_port_attr attr;
1479 	int err;
1480 
1481 	err = ib_query_port(ibdev, port_num, &attr);
1482 	if (err) {
1483 		ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1484 		return err;
1485 	}
1486 
1487 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
1488 	immutable->gid_tbl_len = attr.gid_tbl_len;
1489 
1490 	return 0;
1491 }
1492 
1493 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1494 {
1495 	struct efa_com_dealloc_uar_params params = {
1496 		.uarn = uarn,
1497 	};
1498 
1499 	return efa_com_dealloc_uar(&dev->edev, &params);
1500 }
1501 
1502 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1503 {
1504 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1505 	struct efa_dev *dev = to_edev(ibucontext->device);
1506 	struct efa_ibv_alloc_ucontext_resp resp = {};
1507 	struct efa_com_alloc_uar_result result;
1508 	int err;
1509 
1510 	/*
1511 	 * it's fine if the driver does not know all request fields,
1512 	 * we will ack input fields in our response.
1513 	 */
1514 
1515 	err = efa_com_alloc_uar(&dev->edev, &result);
1516 	if (err)
1517 		goto err_out;
1518 
1519 	ucontext->uarn = result.uarn;
1520 
1521 	resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1522 	resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1523 	resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1524 	resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1525 	resp.max_llq_size = dev->dev_attr.max_llq_size;
1526 
1527 	if (udata && udata->outlen) {
1528 		err = ib_copy_to_udata(udata, &resp,
1529 				       min(sizeof(resp), udata->outlen));
1530 		if (err)
1531 			goto err_dealloc_uar;
1532 	}
1533 
1534 	return 0;
1535 
1536 err_dealloc_uar:
1537 	efa_dealloc_uar(dev, result.uarn);
1538 err_out:
1539 	atomic64_inc(&dev->stats.sw_stats.alloc_ucontext_err);
1540 	return err;
1541 }
1542 
1543 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1544 {
1545 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1546 	struct efa_dev *dev = to_edev(ibucontext->device);
1547 
1548 	efa_dealloc_uar(dev, ucontext->uarn);
1549 }
1550 
1551 void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1552 {
1553 	struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
1554 
1555 	/* DMA mapping is already gone, now free the pages */
1556 	if (entry->mmap_flag == EFA_MMAP_DMA_PAGE)
1557 		free_pages_exact(phys_to_virt(entry->address),
1558 				 entry->rdma_entry.npages * PAGE_SIZE);
1559 	kfree(entry);
1560 }
1561 
1562 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1563 		      struct vm_area_struct *vma)
1564 {
1565 	struct rdma_user_mmap_entry *rdma_entry;
1566 	struct efa_user_mmap_entry *entry;
1567 	unsigned long va;
1568 	int err = 0;
1569 	u64 pfn;
1570 
1571 	rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
1572 	if (!rdma_entry) {
1573 		ibdev_dbg(&dev->ibdev,
1574 			  "pgoff[%#lx] does not have valid entry\n",
1575 			  vma->vm_pgoff);
1576 		return -EINVAL;
1577 	}
1578 	entry = to_emmap(rdma_entry);
1579 
1580 	ibdev_dbg(&dev->ibdev,
1581 		  "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
1582 		  entry->address, rdma_entry->npages * PAGE_SIZE,
1583 		  entry->mmap_flag);
1584 
1585 	pfn = entry->address >> PAGE_SHIFT;
1586 	switch (entry->mmap_flag) {
1587 	case EFA_MMAP_IO_NC:
1588 		err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1589 					entry->rdma_entry.npages * PAGE_SIZE,
1590 					pgprot_noncached(vma->vm_page_prot),
1591 					rdma_entry);
1592 		break;
1593 	case EFA_MMAP_IO_WC:
1594 		err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1595 					entry->rdma_entry.npages * PAGE_SIZE,
1596 					pgprot_writecombine(vma->vm_page_prot),
1597 					rdma_entry);
1598 		break;
1599 	case EFA_MMAP_DMA_PAGE:
1600 		for (va = vma->vm_start; va < vma->vm_end;
1601 		     va += PAGE_SIZE, pfn++) {
1602 			err = vm_insert_page(vma, va, pfn_to_page(pfn));
1603 			if (err)
1604 				break;
1605 		}
1606 		break;
1607 	default:
1608 		err = -EINVAL;
1609 	}
1610 
1611 	if (err) {
1612 		ibdev_dbg(
1613 			&dev->ibdev,
1614 			"Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
1615 			entry->address, rdma_entry->npages * PAGE_SIZE,
1616 			entry->mmap_flag, err);
1617 	}
1618 
1619 	rdma_user_mmap_entry_put(rdma_entry);
1620 	return err;
1621 }
1622 
1623 int efa_mmap(struct ib_ucontext *ibucontext,
1624 	     struct vm_area_struct *vma)
1625 {
1626 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1627 	struct efa_dev *dev = to_edev(ibucontext->device);
1628 	size_t length = vma->vm_end - vma->vm_start;
1629 
1630 	ibdev_dbg(&dev->ibdev,
1631 		  "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
1632 		  vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
1633 
1634 	return __efa_mmap(dev, ucontext, vma);
1635 }
1636 
1637 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1638 {
1639 	struct efa_com_destroy_ah_params params = {
1640 		.ah = ah->ah,
1641 		.pdn = to_epd(ah->ibah.pd)->pdn,
1642 	};
1643 
1644 	return efa_com_destroy_ah(&dev->edev, &params);
1645 }
1646 
1647 int efa_create_ah(struct ib_ah *ibah,
1648 		  struct rdma_ah_attr *ah_attr,
1649 		  u32 flags,
1650 		  struct ib_udata *udata)
1651 {
1652 	struct efa_dev *dev = to_edev(ibah->device);
1653 	struct efa_com_create_ah_params params = {};
1654 	struct efa_ibv_create_ah_resp resp = {};
1655 	struct efa_com_create_ah_result result;
1656 	struct efa_ah *ah = to_eah(ibah);
1657 	int err;
1658 
1659 	if (!(flags & RDMA_CREATE_AH_SLEEPABLE)) {
1660 		ibdev_dbg(&dev->ibdev,
1661 			  "Create address handle is not supported in atomic context\n");
1662 		err = -EOPNOTSUPP;
1663 		goto err_out;
1664 	}
1665 
1666 	if (udata->inlen &&
1667 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1668 		ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1669 		err = -EINVAL;
1670 		goto err_out;
1671 	}
1672 
1673 	memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1674 	       sizeof(params.dest_addr));
1675 	params.pdn = to_epd(ibah->pd)->pdn;
1676 	err = efa_com_create_ah(&dev->edev, &params, &result);
1677 	if (err)
1678 		goto err_out;
1679 
1680 	memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1681 	ah->ah = result.ah;
1682 
1683 	resp.efa_address_handle = result.ah;
1684 
1685 	if (udata->outlen) {
1686 		err = ib_copy_to_udata(udata, &resp,
1687 				       min(sizeof(resp), udata->outlen));
1688 		if (err) {
1689 			ibdev_dbg(&dev->ibdev,
1690 				  "Failed to copy udata for create_ah response\n");
1691 			goto err_destroy_ah;
1692 		}
1693 	}
1694 	ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
1695 
1696 	return 0;
1697 
1698 err_destroy_ah:
1699 	efa_ah_destroy(dev, ah);
1700 err_out:
1701 	atomic64_inc(&dev->stats.sw_stats.create_ah_err);
1702 	return err;
1703 }
1704 
1705 void efa_destroy_ah(struct ib_ah *ibah, u32 flags)
1706 {
1707 	struct efa_dev *dev = to_edev(ibah->pd->device);
1708 	struct efa_ah *ah = to_eah(ibah);
1709 
1710 	ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
1711 
1712 	if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
1713 		ibdev_dbg(&dev->ibdev,
1714 			  "Destroy address handle is not supported in atomic context\n");
1715 		return;
1716 	}
1717 
1718 	efa_ah_destroy(dev, ah);
1719 }
1720 
1721 struct rdma_hw_stats *efa_alloc_hw_stats(struct ib_device *ibdev, u8 port_num)
1722 {
1723 	return rdma_alloc_hw_stats_struct(efa_stats_names,
1724 					  ARRAY_SIZE(efa_stats_names),
1725 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
1726 }
1727 
1728 int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1729 		     u8 port_num, int index)
1730 {
1731 	struct efa_com_get_stats_params params = {};
1732 	union efa_com_get_stats_result result;
1733 	struct efa_dev *dev = to_edev(ibdev);
1734 	struct efa_com_basic_stats *bs;
1735 	struct efa_com_stats_admin *as;
1736 	struct efa_stats *s;
1737 	int err;
1738 
1739 	params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
1740 	params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
1741 
1742 	err = efa_com_get_stats(&dev->edev, &params, &result);
1743 	if (err)
1744 		return err;
1745 
1746 	bs = &result.basic_stats;
1747 	stats->value[EFA_TX_BYTES] = bs->tx_bytes;
1748 	stats->value[EFA_TX_PKTS] = bs->tx_pkts;
1749 	stats->value[EFA_RX_BYTES] = bs->rx_bytes;
1750 	stats->value[EFA_RX_PKTS] = bs->rx_pkts;
1751 	stats->value[EFA_RX_DROPS] = bs->rx_drops;
1752 
1753 	as = &dev->edev.aq.stats;
1754 	stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
1755 	stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
1756 	stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
1757 
1758 	s = &dev->stats;
1759 	stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
1760 	stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->sw_stats.alloc_pd_err);
1761 	stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->sw_stats.create_qp_err);
1762 	stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->sw_stats.reg_mr_err);
1763 	stats->value[EFA_ALLOC_UCONTEXT_ERR] = atomic64_read(&s->sw_stats.alloc_ucontext_err);
1764 	stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->sw_stats.create_ah_err);
1765 
1766 	return ARRAY_SIZE(efa_stats_names);
1767 }
1768 
1769 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
1770 					 u8 port_num)
1771 {
1772 	return IB_LINK_LAYER_UNSPECIFIED;
1773 }
1774 
1775