1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright 2018-2023 Amazon.com, Inc. or its affiliates. All rights reserved.
4  */
5 
6 #include <linux/dma-buf.h>
7 #include <linux/dma-resv.h>
8 #include <linux/vmalloc.h>
9 #include <linux/log2.h>
10 
11 #include <rdma/ib_addr.h>
12 #include <rdma/ib_umem.h>
13 #include <rdma/ib_user_verbs.h>
14 #include <rdma/ib_verbs.h>
15 #include <rdma/uverbs_ioctl.h>
16 
17 #include "efa.h"
18 #include "efa_io_defs.h"
19 
20 enum {
21 	EFA_MMAP_DMA_PAGE = 0,
22 	EFA_MMAP_IO_WC,
23 	EFA_MMAP_IO_NC,
24 };
25 
26 #define EFA_AENQ_ENABLED_GROUPS \
27 	(BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
28 	 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
29 
30 struct efa_user_mmap_entry {
31 	struct rdma_user_mmap_entry rdma_entry;
32 	u64 address;
33 	u8 mmap_flag;
34 };
35 
36 #define EFA_DEFINE_DEVICE_STATS(op) \
37 	op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
38 	op(EFA_COMPLETED_CMDS, "completed_cmds") \
39 	op(EFA_CMDS_ERR, "cmds_err") \
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_CREATE_CQ_ERR, "create_cq_err") \
45 	op(EFA_REG_MR_ERR, "reg_mr_err") \
46 	op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \
47 	op(EFA_CREATE_AH_ERR, "create_ah_err") \
48 	op(EFA_MMAP_ERR, "mmap_err")
49 
50 #define EFA_DEFINE_PORT_STATS(op) \
51 	op(EFA_TX_BYTES, "tx_bytes") \
52 	op(EFA_TX_PKTS, "tx_pkts") \
53 	op(EFA_RX_BYTES, "rx_bytes") \
54 	op(EFA_RX_PKTS, "rx_pkts") \
55 	op(EFA_RX_DROPS, "rx_drops") \
56 	op(EFA_SEND_BYTES, "send_bytes") \
57 	op(EFA_SEND_WRS, "send_wrs") \
58 	op(EFA_RECV_BYTES, "recv_bytes") \
59 	op(EFA_RECV_WRS, "recv_wrs") \
60 	op(EFA_RDMA_READ_WRS, "rdma_read_wrs") \
61 	op(EFA_RDMA_READ_BYTES, "rdma_read_bytes") \
62 	op(EFA_RDMA_READ_WR_ERR, "rdma_read_wr_err") \
63 	op(EFA_RDMA_READ_RESP_BYTES, "rdma_read_resp_bytes") \
64 	op(EFA_RDMA_WRITE_WRS, "rdma_write_wrs") \
65 	op(EFA_RDMA_WRITE_BYTES, "rdma_write_bytes") \
66 	op(EFA_RDMA_WRITE_WR_ERR, "rdma_write_wr_err") \
67 	op(EFA_RDMA_WRITE_RECV_BYTES, "rdma_write_recv_bytes") \
68 
69 #define EFA_STATS_ENUM(ename, name) ename,
70 #define EFA_STATS_STR(ename, nam) \
71 	[ename].name = nam,
72 
73 enum efa_hw_device_stats {
74 	EFA_DEFINE_DEVICE_STATS(EFA_STATS_ENUM)
75 };
76 
77 static const struct rdma_stat_desc efa_device_stats_descs[] = {
78 	EFA_DEFINE_DEVICE_STATS(EFA_STATS_STR)
79 };
80 
81 enum efa_hw_port_stats {
82 	EFA_DEFINE_PORT_STATS(EFA_STATS_ENUM)
83 };
84 
85 static const struct rdma_stat_desc efa_port_stats_descs[] = {
86 	EFA_DEFINE_PORT_STATS(EFA_STATS_STR)
87 };
88 
89 #define EFA_CHUNK_PAYLOAD_SHIFT       12
90 #define EFA_CHUNK_PAYLOAD_SIZE        BIT(EFA_CHUNK_PAYLOAD_SHIFT)
91 #define EFA_CHUNK_PAYLOAD_PTR_SIZE    8
92 
93 #define EFA_CHUNK_SHIFT               12
94 #define EFA_CHUNK_SIZE                BIT(EFA_CHUNK_SHIFT)
95 #define EFA_CHUNK_PTR_SIZE            sizeof(struct efa_com_ctrl_buff_info)
96 
97 #define EFA_PTRS_PER_CHUNK \
98 	((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
99 
100 #define EFA_CHUNK_USED_SIZE \
101 	((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
102 
103 struct pbl_chunk {
104 	dma_addr_t dma_addr;
105 	u64 *buf;
106 	u32 length;
107 };
108 
109 struct pbl_chunk_list {
110 	struct pbl_chunk *chunks;
111 	unsigned int size;
112 };
113 
114 struct pbl_context {
115 	union {
116 		struct {
117 			dma_addr_t dma_addr;
118 		} continuous;
119 		struct {
120 			u32 pbl_buf_size_in_pages;
121 			struct scatterlist *sgl;
122 			int sg_dma_cnt;
123 			struct pbl_chunk_list chunk_list;
124 		} indirect;
125 	} phys;
126 	u64 *pbl_buf;
127 	u32 pbl_buf_size_in_bytes;
128 	u8 physically_continuous;
129 };
130 
131 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
132 {
133 	return container_of(ibdev, struct efa_dev, ibdev);
134 }
135 
136 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
137 {
138 	return container_of(ibucontext, struct efa_ucontext, ibucontext);
139 }
140 
141 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
142 {
143 	return container_of(ibpd, struct efa_pd, ibpd);
144 }
145 
146 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
147 {
148 	return container_of(ibmr, struct efa_mr, ibmr);
149 }
150 
151 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
152 {
153 	return container_of(ibqp, struct efa_qp, ibqp);
154 }
155 
156 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
157 {
158 	return container_of(ibcq, struct efa_cq, ibcq);
159 }
160 
161 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
162 {
163 	return container_of(ibah, struct efa_ah, ibah);
164 }
165 
166 static inline struct efa_user_mmap_entry *
167 to_emmap(struct rdma_user_mmap_entry *rdma_entry)
168 {
169 	return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
170 }
171 
172 #define EFA_DEV_CAP(dev, cap) \
173 	((dev)->dev_attr.device_caps & \
174 	 EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_##cap##_MASK)
175 
176 #define is_reserved_cleared(reserved) \
177 	!memchr_inv(reserved, 0, sizeof(reserved))
178 
179 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
180 			       size_t size, enum dma_data_direction dir)
181 {
182 	void *addr;
183 
184 	addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
185 	if (!addr)
186 		return NULL;
187 
188 	*dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
189 	if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
190 		ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
191 		free_pages_exact(addr, size);
192 		return NULL;
193 	}
194 
195 	return addr;
196 }
197 
198 static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr,
199 			    dma_addr_t dma_addr,
200 			    size_t size, enum dma_data_direction dir)
201 {
202 	dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir);
203 	free_pages_exact(cpu_addr, size);
204 }
205 
206 int efa_query_device(struct ib_device *ibdev,
207 		     struct ib_device_attr *props,
208 		     struct ib_udata *udata)
209 {
210 	struct efa_com_get_device_attr_result *dev_attr;
211 	struct efa_ibv_ex_query_device_resp resp = {};
212 	struct efa_dev *dev = to_edev(ibdev);
213 	int err;
214 
215 	if (udata && udata->inlen &&
216 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
217 		ibdev_dbg(ibdev,
218 			  "Incompatible ABI params, udata not cleared\n");
219 		return -EINVAL;
220 	}
221 
222 	dev_attr = &dev->dev_attr;
223 
224 	memset(props, 0, sizeof(*props));
225 	props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
226 	props->page_size_cap = dev_attr->page_size_cap;
227 	props->vendor_id = dev->pdev->vendor;
228 	props->vendor_part_id = dev->pdev->device;
229 	props->hw_ver = dev->pdev->subsystem_device;
230 	props->max_qp = dev_attr->max_qp;
231 	props->max_cq = dev_attr->max_cq;
232 	props->max_pd = dev_attr->max_pd;
233 	props->max_mr = dev_attr->max_mr;
234 	props->max_ah = dev_attr->max_ah;
235 	props->max_cqe = dev_attr->max_cq_depth;
236 	props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
237 				 dev_attr->max_rq_depth);
238 	props->max_send_sge = dev_attr->max_sq_sge;
239 	props->max_recv_sge = dev_attr->max_rq_sge;
240 	props->max_sge_rd = dev_attr->max_wr_rdma_sge;
241 	props->max_pkeys = 1;
242 
243 	if (udata && udata->outlen) {
244 		resp.max_sq_sge = dev_attr->max_sq_sge;
245 		resp.max_rq_sge = dev_attr->max_rq_sge;
246 		resp.max_sq_wr = dev_attr->max_sq_depth;
247 		resp.max_rq_wr = dev_attr->max_rq_depth;
248 		resp.max_rdma_size = dev_attr->max_rdma_size;
249 
250 		resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID;
251 		if (EFA_DEV_CAP(dev, RDMA_READ))
252 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
253 
254 		if (EFA_DEV_CAP(dev, RNR_RETRY))
255 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RNR_RETRY;
256 
257 		if (EFA_DEV_CAP(dev, DATA_POLLING_128))
258 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128;
259 
260 		if (EFA_DEV_CAP(dev, RDMA_WRITE))
261 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_WRITE;
262 
263 		if (dev->neqs)
264 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS;
265 
266 		err = ib_copy_to_udata(udata, &resp,
267 				       min(sizeof(resp), udata->outlen));
268 		if (err) {
269 			ibdev_dbg(ibdev,
270 				  "Failed to copy udata for query_device\n");
271 			return err;
272 		}
273 	}
274 
275 	return 0;
276 }
277 
278 int efa_query_port(struct ib_device *ibdev, u32 port,
279 		   struct ib_port_attr *props)
280 {
281 	struct efa_dev *dev = to_edev(ibdev);
282 
283 	props->lmc = 1;
284 
285 	props->state = IB_PORT_ACTIVE;
286 	props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
287 	props->gid_tbl_len = 1;
288 	props->pkey_tbl_len = 1;
289 	props->active_speed = IB_SPEED_EDR;
290 	props->active_width = IB_WIDTH_4X;
291 	props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
292 	props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
293 	props->max_msg_sz = dev->dev_attr.mtu;
294 	props->max_vl_num = 1;
295 
296 	return 0;
297 }
298 
299 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
300 		 int qp_attr_mask,
301 		 struct ib_qp_init_attr *qp_init_attr)
302 {
303 	struct efa_dev *dev = to_edev(ibqp->device);
304 	struct efa_com_query_qp_params params = {};
305 	struct efa_com_query_qp_result result;
306 	struct efa_qp *qp = to_eqp(ibqp);
307 	int err;
308 
309 #define EFA_QUERY_QP_SUPP_MASK \
310 	(IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
311 	 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP | IB_QP_RNR_RETRY)
312 
313 	if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
314 		ibdev_dbg(&dev->ibdev,
315 			  "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
316 			  qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
317 		return -EOPNOTSUPP;
318 	}
319 
320 	memset(qp_attr, 0, sizeof(*qp_attr));
321 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
322 
323 	params.qp_handle = qp->qp_handle;
324 	err = efa_com_query_qp(&dev->edev, &params, &result);
325 	if (err)
326 		return err;
327 
328 	qp_attr->qp_state = result.qp_state;
329 	qp_attr->qkey = result.qkey;
330 	qp_attr->sq_psn = result.sq_psn;
331 	qp_attr->sq_draining = result.sq_draining;
332 	qp_attr->port_num = 1;
333 	qp_attr->rnr_retry = result.rnr_retry;
334 
335 	qp_attr->cap.max_send_wr = qp->max_send_wr;
336 	qp_attr->cap.max_recv_wr = qp->max_recv_wr;
337 	qp_attr->cap.max_send_sge = qp->max_send_sge;
338 	qp_attr->cap.max_recv_sge = qp->max_recv_sge;
339 	qp_attr->cap.max_inline_data = qp->max_inline_data;
340 
341 	qp_init_attr->qp_type = ibqp->qp_type;
342 	qp_init_attr->recv_cq = ibqp->recv_cq;
343 	qp_init_attr->send_cq = ibqp->send_cq;
344 	qp_init_attr->qp_context = ibqp->qp_context;
345 	qp_init_attr->cap = qp_attr->cap;
346 
347 	return 0;
348 }
349 
350 int efa_query_gid(struct ib_device *ibdev, u32 port, int index,
351 		  union ib_gid *gid)
352 {
353 	struct efa_dev *dev = to_edev(ibdev);
354 
355 	memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
356 
357 	return 0;
358 }
359 
360 int efa_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
361 		   u16 *pkey)
362 {
363 	if (index > 0)
364 		return -EINVAL;
365 
366 	*pkey = 0xffff;
367 	return 0;
368 }
369 
370 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
371 {
372 	struct efa_com_dealloc_pd_params params = {
373 		.pdn = pdn,
374 	};
375 
376 	return efa_com_dealloc_pd(&dev->edev, &params);
377 }
378 
379 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
380 {
381 	struct efa_dev *dev = to_edev(ibpd->device);
382 	struct efa_ibv_alloc_pd_resp resp = {};
383 	struct efa_com_alloc_pd_result result;
384 	struct efa_pd *pd = to_epd(ibpd);
385 	int err;
386 
387 	if (udata->inlen &&
388 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
389 		ibdev_dbg(&dev->ibdev,
390 			  "Incompatible ABI params, udata not cleared\n");
391 		err = -EINVAL;
392 		goto err_out;
393 	}
394 
395 	err = efa_com_alloc_pd(&dev->edev, &result);
396 	if (err)
397 		goto err_out;
398 
399 	pd->pdn = result.pdn;
400 	resp.pdn = result.pdn;
401 
402 	if (udata->outlen) {
403 		err = ib_copy_to_udata(udata, &resp,
404 				       min(sizeof(resp), udata->outlen));
405 		if (err) {
406 			ibdev_dbg(&dev->ibdev,
407 				  "Failed to copy udata for alloc_pd\n");
408 			goto err_dealloc_pd;
409 		}
410 	}
411 
412 	ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
413 
414 	return 0;
415 
416 err_dealloc_pd:
417 	efa_pd_dealloc(dev, result.pdn);
418 err_out:
419 	atomic64_inc(&dev->stats.alloc_pd_err);
420 	return err;
421 }
422 
423 int efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
424 {
425 	struct efa_dev *dev = to_edev(ibpd->device);
426 	struct efa_pd *pd = to_epd(ibpd);
427 
428 	ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
429 	efa_pd_dealloc(dev, pd->pdn);
430 	return 0;
431 }
432 
433 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
434 {
435 	struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
436 
437 	return efa_com_destroy_qp(&dev->edev, &params);
438 }
439 
440 static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp)
441 {
442 	rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
443 	rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
444 	rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
445 	rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
446 }
447 
448 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
449 {
450 	struct efa_dev *dev = to_edev(ibqp->pd->device);
451 	struct efa_qp *qp = to_eqp(ibqp);
452 	int err;
453 
454 	ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
455 
456 	err = efa_destroy_qp_handle(dev, qp->qp_handle);
457 	if (err)
458 		return err;
459 
460 	efa_qp_user_mmap_entries_remove(qp);
461 
462 	if (qp->rq_cpu_addr) {
463 		ibdev_dbg(&dev->ibdev,
464 			  "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
465 			  qp->rq_cpu_addr, qp->rq_size,
466 			  &qp->rq_dma_addr);
467 		efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
468 				qp->rq_size, DMA_TO_DEVICE);
469 	}
470 
471 	return 0;
472 }
473 
474 static struct rdma_user_mmap_entry*
475 efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
476 			   u64 address, size_t length,
477 			   u8 mmap_flag, u64 *offset)
478 {
479 	struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
480 	int err;
481 
482 	if (!entry)
483 		return NULL;
484 
485 	entry->address = address;
486 	entry->mmap_flag = mmap_flag;
487 
488 	err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
489 					  length);
490 	if (err) {
491 		kfree(entry);
492 		return NULL;
493 	}
494 	*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
495 
496 	return &entry->rdma_entry;
497 }
498 
499 static int qp_mmap_entries_setup(struct efa_qp *qp,
500 				 struct efa_dev *dev,
501 				 struct efa_ucontext *ucontext,
502 				 struct efa_com_create_qp_params *params,
503 				 struct efa_ibv_create_qp_resp *resp)
504 {
505 	size_t length;
506 	u64 address;
507 
508 	address = dev->db_bar_addr + resp->sq_db_offset;
509 	qp->sq_db_mmap_entry =
510 		efa_user_mmap_entry_insert(&ucontext->ibucontext,
511 					   address,
512 					   PAGE_SIZE, EFA_MMAP_IO_NC,
513 					   &resp->sq_db_mmap_key);
514 	if (!qp->sq_db_mmap_entry)
515 		return -ENOMEM;
516 
517 	resp->sq_db_offset &= ~PAGE_MASK;
518 
519 	address = dev->mem_bar_addr + resp->llq_desc_offset;
520 	length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
521 			    (resp->llq_desc_offset & ~PAGE_MASK));
522 
523 	qp->llq_desc_mmap_entry =
524 		efa_user_mmap_entry_insert(&ucontext->ibucontext,
525 					   address, length,
526 					   EFA_MMAP_IO_WC,
527 					   &resp->llq_desc_mmap_key);
528 	if (!qp->llq_desc_mmap_entry)
529 		goto err_remove_mmap;
530 
531 	resp->llq_desc_offset &= ~PAGE_MASK;
532 
533 	if (qp->rq_size) {
534 		address = dev->db_bar_addr + resp->rq_db_offset;
535 
536 		qp->rq_db_mmap_entry =
537 			efa_user_mmap_entry_insert(&ucontext->ibucontext,
538 						   address, PAGE_SIZE,
539 						   EFA_MMAP_IO_NC,
540 						   &resp->rq_db_mmap_key);
541 		if (!qp->rq_db_mmap_entry)
542 			goto err_remove_mmap;
543 
544 		resp->rq_db_offset &= ~PAGE_MASK;
545 
546 		address = virt_to_phys(qp->rq_cpu_addr);
547 		qp->rq_mmap_entry =
548 			efa_user_mmap_entry_insert(&ucontext->ibucontext,
549 						   address, qp->rq_size,
550 						   EFA_MMAP_DMA_PAGE,
551 						   &resp->rq_mmap_key);
552 		if (!qp->rq_mmap_entry)
553 			goto err_remove_mmap;
554 
555 		resp->rq_mmap_size = qp->rq_size;
556 	}
557 
558 	return 0;
559 
560 err_remove_mmap:
561 	efa_qp_user_mmap_entries_remove(qp);
562 
563 	return -ENOMEM;
564 }
565 
566 static int efa_qp_validate_cap(struct efa_dev *dev,
567 			       struct ib_qp_init_attr *init_attr)
568 {
569 	if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
570 		ibdev_dbg(&dev->ibdev,
571 			  "qp: requested send wr[%u] exceeds the max[%u]\n",
572 			  init_attr->cap.max_send_wr,
573 			  dev->dev_attr.max_sq_depth);
574 		return -EINVAL;
575 	}
576 	if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
577 		ibdev_dbg(&dev->ibdev,
578 			  "qp: requested receive wr[%u] exceeds the max[%u]\n",
579 			  init_attr->cap.max_recv_wr,
580 			  dev->dev_attr.max_rq_depth);
581 		return -EINVAL;
582 	}
583 	if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
584 		ibdev_dbg(&dev->ibdev,
585 			  "qp: requested sge send[%u] exceeds the max[%u]\n",
586 			  init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
587 		return -EINVAL;
588 	}
589 	if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
590 		ibdev_dbg(&dev->ibdev,
591 			  "qp: requested sge recv[%u] exceeds the max[%u]\n",
592 			  init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
593 		return -EINVAL;
594 	}
595 	if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
596 		ibdev_dbg(&dev->ibdev,
597 			  "qp: requested inline data[%u] exceeds the max[%u]\n",
598 			  init_attr->cap.max_inline_data,
599 			  dev->dev_attr.inline_buf_size);
600 		return -EINVAL;
601 	}
602 
603 	return 0;
604 }
605 
606 static int efa_qp_validate_attr(struct efa_dev *dev,
607 				struct ib_qp_init_attr *init_attr)
608 {
609 	if (init_attr->qp_type != IB_QPT_DRIVER &&
610 	    init_attr->qp_type != IB_QPT_UD) {
611 		ibdev_dbg(&dev->ibdev,
612 			  "Unsupported qp type %d\n", init_attr->qp_type);
613 		return -EOPNOTSUPP;
614 	}
615 
616 	if (init_attr->srq) {
617 		ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
618 		return -EOPNOTSUPP;
619 	}
620 
621 	if (init_attr->create_flags) {
622 		ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
623 		return -EOPNOTSUPP;
624 	}
625 
626 	return 0;
627 }
628 
629 int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
630 		  struct ib_udata *udata)
631 {
632 	struct efa_com_create_qp_params create_qp_params = {};
633 	struct efa_com_create_qp_result create_qp_resp;
634 	struct efa_dev *dev = to_edev(ibqp->device);
635 	struct efa_ibv_create_qp_resp resp = {};
636 	struct efa_ibv_create_qp cmd = {};
637 	struct efa_qp *qp = to_eqp(ibqp);
638 	struct efa_ucontext *ucontext;
639 	int err;
640 
641 	ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
642 					     ibucontext);
643 
644 	err = efa_qp_validate_cap(dev, init_attr);
645 	if (err)
646 		goto err_out;
647 
648 	err = efa_qp_validate_attr(dev, init_attr);
649 	if (err)
650 		goto err_out;
651 
652 	if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) {
653 		ibdev_dbg(&dev->ibdev,
654 			  "Incompatible ABI params, no input udata\n");
655 		err = -EINVAL;
656 		goto err_out;
657 	}
658 
659 	if (udata->inlen > sizeof(cmd) &&
660 	    !ib_is_udata_cleared(udata, sizeof(cmd),
661 				 udata->inlen - sizeof(cmd))) {
662 		ibdev_dbg(&dev->ibdev,
663 			  "Incompatible ABI params, unknown fields in udata\n");
664 		err = -EINVAL;
665 		goto err_out;
666 	}
667 
668 	err = ib_copy_from_udata(&cmd, udata,
669 				 min(sizeof(cmd), udata->inlen));
670 	if (err) {
671 		ibdev_dbg(&dev->ibdev,
672 			  "Cannot copy udata for create_qp\n");
673 		goto err_out;
674 	}
675 
676 	if (cmd.comp_mask) {
677 		ibdev_dbg(&dev->ibdev,
678 			  "Incompatible ABI params, unknown fields in udata\n");
679 		err = -EINVAL;
680 		goto err_out;
681 	}
682 
683 	create_qp_params.uarn = ucontext->uarn;
684 	create_qp_params.pd = to_epd(ibqp->pd)->pdn;
685 
686 	if (init_attr->qp_type == IB_QPT_UD) {
687 		create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
688 	} else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
689 		create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
690 	} else {
691 		ibdev_dbg(&dev->ibdev,
692 			  "Unsupported qp type %d driver qp type %d\n",
693 			  init_attr->qp_type, cmd.driver_qp_type);
694 		err = -EOPNOTSUPP;
695 		goto err_out;
696 	}
697 
698 	ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
699 		  init_attr->qp_type, cmd.driver_qp_type);
700 	create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
701 	create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
702 	create_qp_params.sq_depth = init_attr->cap.max_send_wr;
703 	create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
704 
705 	create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
706 	create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
707 	qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
708 	if (qp->rq_size) {
709 		qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
710 						    qp->rq_size, DMA_TO_DEVICE);
711 		if (!qp->rq_cpu_addr) {
712 			err = -ENOMEM;
713 			goto err_out;
714 		}
715 
716 		ibdev_dbg(&dev->ibdev,
717 			  "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
718 			  qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
719 		create_qp_params.rq_base_addr = qp->rq_dma_addr;
720 	}
721 
722 	err = efa_com_create_qp(&dev->edev, &create_qp_params,
723 				&create_qp_resp);
724 	if (err)
725 		goto err_free_mapped;
726 
727 	resp.sq_db_offset = create_qp_resp.sq_db_offset;
728 	resp.rq_db_offset = create_qp_resp.rq_db_offset;
729 	resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
730 	resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
731 	resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
732 
733 	err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
734 				    &resp);
735 	if (err)
736 		goto err_destroy_qp;
737 
738 	qp->qp_handle = create_qp_resp.qp_handle;
739 	qp->ibqp.qp_num = create_qp_resp.qp_num;
740 	qp->max_send_wr = init_attr->cap.max_send_wr;
741 	qp->max_recv_wr = init_attr->cap.max_recv_wr;
742 	qp->max_send_sge = init_attr->cap.max_send_sge;
743 	qp->max_recv_sge = init_attr->cap.max_recv_sge;
744 	qp->max_inline_data = init_attr->cap.max_inline_data;
745 
746 	if (udata->outlen) {
747 		err = ib_copy_to_udata(udata, &resp,
748 				       min(sizeof(resp), udata->outlen));
749 		if (err) {
750 			ibdev_dbg(&dev->ibdev,
751 				  "Failed to copy udata for qp[%u]\n",
752 				  create_qp_resp.qp_num);
753 			goto err_remove_mmap_entries;
754 		}
755 	}
756 
757 	ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
758 
759 	return 0;
760 
761 err_remove_mmap_entries:
762 	efa_qp_user_mmap_entries_remove(qp);
763 err_destroy_qp:
764 	efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
765 err_free_mapped:
766 	if (qp->rq_size)
767 		efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
768 				qp->rq_size, DMA_TO_DEVICE);
769 err_out:
770 	atomic64_inc(&dev->stats.create_qp_err);
771 	return err;
772 }
773 
774 static const struct {
775 	int			valid;
776 	enum ib_qp_attr_mask	req_param;
777 	enum ib_qp_attr_mask	opt_param;
778 } srd_qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
779 	[IB_QPS_RESET] = {
780 		[IB_QPS_RESET] = { .valid = 1 },
781 		[IB_QPS_INIT]  = {
782 			.valid = 1,
783 			.req_param = IB_QP_PKEY_INDEX |
784 				     IB_QP_PORT |
785 				     IB_QP_QKEY,
786 		},
787 	},
788 	[IB_QPS_INIT] = {
789 		[IB_QPS_RESET] = { .valid = 1 },
790 		[IB_QPS_ERR]   = { .valid = 1 },
791 		[IB_QPS_INIT]  = {
792 			.valid = 1,
793 			.opt_param = IB_QP_PKEY_INDEX |
794 				     IB_QP_PORT |
795 				     IB_QP_QKEY,
796 		},
797 		[IB_QPS_RTR]   = {
798 			.valid = 1,
799 			.opt_param = IB_QP_PKEY_INDEX |
800 				     IB_QP_QKEY,
801 		},
802 	},
803 	[IB_QPS_RTR] = {
804 		[IB_QPS_RESET] = { .valid = 1 },
805 		[IB_QPS_ERR]   = { .valid = 1 },
806 		[IB_QPS_RTS]   = {
807 			.valid = 1,
808 			.req_param = IB_QP_SQ_PSN,
809 			.opt_param = IB_QP_CUR_STATE |
810 				     IB_QP_QKEY |
811 				     IB_QP_RNR_RETRY,
812 
813 		}
814 	},
815 	[IB_QPS_RTS] = {
816 		[IB_QPS_RESET] = { .valid = 1 },
817 		[IB_QPS_ERR]   = { .valid = 1 },
818 		[IB_QPS_RTS]   = {
819 			.valid = 1,
820 			.opt_param = IB_QP_CUR_STATE |
821 				     IB_QP_QKEY,
822 		},
823 		[IB_QPS_SQD] = {
824 			.valid = 1,
825 			.opt_param = IB_QP_EN_SQD_ASYNC_NOTIFY,
826 		},
827 	},
828 	[IB_QPS_SQD] = {
829 		[IB_QPS_RESET] = { .valid = 1 },
830 		[IB_QPS_ERR]   = { .valid = 1 },
831 		[IB_QPS_RTS]   = {
832 			.valid = 1,
833 			.opt_param = IB_QP_CUR_STATE |
834 				     IB_QP_QKEY,
835 		},
836 		[IB_QPS_SQD] = {
837 			.valid = 1,
838 			.opt_param = IB_QP_PKEY_INDEX |
839 				     IB_QP_QKEY,
840 		}
841 	},
842 	[IB_QPS_SQE] = {
843 		[IB_QPS_RESET] = { .valid = 1 },
844 		[IB_QPS_ERR]   = { .valid = 1 },
845 		[IB_QPS_RTS]   = {
846 			.valid = 1,
847 			.opt_param = IB_QP_CUR_STATE |
848 				     IB_QP_QKEY,
849 		}
850 	},
851 	[IB_QPS_ERR] = {
852 		[IB_QPS_RESET] = { .valid = 1 },
853 		[IB_QPS_ERR]   = { .valid = 1 },
854 	}
855 };
856 
857 static bool efa_modify_srd_qp_is_ok(enum ib_qp_state cur_state,
858 				    enum ib_qp_state next_state,
859 				    enum ib_qp_attr_mask mask)
860 {
861 	enum ib_qp_attr_mask req_param, opt_param;
862 
863 	if (mask & IB_QP_CUR_STATE  &&
864 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
865 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
866 		return false;
867 
868 	if (!srd_qp_state_table[cur_state][next_state].valid)
869 		return false;
870 
871 	req_param = srd_qp_state_table[cur_state][next_state].req_param;
872 	opt_param = srd_qp_state_table[cur_state][next_state].opt_param;
873 
874 	if ((mask & req_param) != req_param)
875 		return false;
876 
877 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
878 		return false;
879 
880 	return true;
881 }
882 
883 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
884 				  struct ib_qp_attr *qp_attr, int qp_attr_mask,
885 				  enum ib_qp_state cur_state,
886 				  enum ib_qp_state new_state)
887 {
888 	int err;
889 
890 #define EFA_MODIFY_QP_SUPP_MASK \
891 	(IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
892 	 IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN | \
893 	 IB_QP_RNR_RETRY)
894 
895 	if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
896 		ibdev_dbg(&dev->ibdev,
897 			  "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
898 			  qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
899 		return -EOPNOTSUPP;
900 	}
901 
902 	if (qp->ibqp.qp_type == IB_QPT_DRIVER)
903 		err = !efa_modify_srd_qp_is_ok(cur_state, new_state,
904 					       qp_attr_mask);
905 	else
906 		err = !ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
907 					  qp_attr_mask);
908 
909 	if (err) {
910 		ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
911 		return -EINVAL;
912 	}
913 
914 	if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
915 		ibdev_dbg(&dev->ibdev, "Can't change port num\n");
916 		return -EOPNOTSUPP;
917 	}
918 
919 	if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
920 		ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
921 		return -EOPNOTSUPP;
922 	}
923 
924 	return 0;
925 }
926 
927 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
928 		  int qp_attr_mask, struct ib_udata *udata)
929 {
930 	struct efa_dev *dev = to_edev(ibqp->device);
931 	struct efa_com_modify_qp_params params = {};
932 	struct efa_qp *qp = to_eqp(ibqp);
933 	enum ib_qp_state cur_state;
934 	enum ib_qp_state new_state;
935 	int err;
936 
937 	if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
938 		return -EOPNOTSUPP;
939 
940 	if (udata->inlen &&
941 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
942 		ibdev_dbg(&dev->ibdev,
943 			  "Incompatible ABI params, udata not cleared\n");
944 		return -EINVAL;
945 	}
946 
947 	cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
948 						     qp->state;
949 	new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
950 
951 	err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
952 				     new_state);
953 	if (err)
954 		return err;
955 
956 	params.qp_handle = qp->qp_handle;
957 
958 	if (qp_attr_mask & IB_QP_STATE) {
959 		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QP_STATE,
960 			1);
961 		EFA_SET(&params.modify_mask,
962 			EFA_ADMIN_MODIFY_QP_CMD_CUR_QP_STATE, 1);
963 		params.cur_qp_state = cur_state;
964 		params.qp_state = new_state;
965 	}
966 
967 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
968 		EFA_SET(&params.modify_mask,
969 			EFA_ADMIN_MODIFY_QP_CMD_SQ_DRAINED_ASYNC_NOTIFY, 1);
970 		params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
971 	}
972 
973 	if (qp_attr_mask & IB_QP_QKEY) {
974 		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QKEY, 1);
975 		params.qkey = qp_attr->qkey;
976 	}
977 
978 	if (qp_attr_mask & IB_QP_SQ_PSN) {
979 		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_SQ_PSN, 1);
980 		params.sq_psn = qp_attr->sq_psn;
981 	}
982 
983 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
984 		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_RNR_RETRY,
985 			1);
986 		params.rnr_retry = qp_attr->rnr_retry;
987 	}
988 
989 	err = efa_com_modify_qp(&dev->edev, &params);
990 	if (err)
991 		return err;
992 
993 	qp->state = new_state;
994 
995 	return 0;
996 }
997 
998 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
999 {
1000 	struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
1001 
1002 	return efa_com_destroy_cq(&dev->edev, &params);
1003 }
1004 
1005 static void efa_cq_user_mmap_entries_remove(struct efa_cq *cq)
1006 {
1007 	rdma_user_mmap_entry_remove(cq->db_mmap_entry);
1008 	rdma_user_mmap_entry_remove(cq->mmap_entry);
1009 }
1010 
1011 int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1012 {
1013 	struct efa_dev *dev = to_edev(ibcq->device);
1014 	struct efa_cq *cq = to_ecq(ibcq);
1015 
1016 	ibdev_dbg(&dev->ibdev,
1017 		  "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
1018 		  cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
1019 
1020 	efa_destroy_cq_idx(dev, cq->cq_idx);
1021 	efa_cq_user_mmap_entries_remove(cq);
1022 	if (cq->eq) {
1023 		xa_erase(&dev->cqs_xa, cq->cq_idx);
1024 		synchronize_irq(cq->eq->irq.irqn);
1025 	}
1026 	efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1027 			DMA_FROM_DEVICE);
1028 	return 0;
1029 }
1030 
1031 static struct efa_eq *efa_vec2eq(struct efa_dev *dev, int vec)
1032 {
1033 	return &dev->eqs[vec];
1034 }
1035 
1036 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
1037 				 struct efa_ibv_create_cq_resp *resp,
1038 				 bool db_valid)
1039 {
1040 	resp->q_mmap_size = cq->size;
1041 	cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1042 						    virt_to_phys(cq->cpu_addr),
1043 						    cq->size, EFA_MMAP_DMA_PAGE,
1044 						    &resp->q_mmap_key);
1045 	if (!cq->mmap_entry)
1046 		return -ENOMEM;
1047 
1048 	if (db_valid) {
1049 		cq->db_mmap_entry =
1050 			efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1051 						   dev->db_bar_addr + resp->db_off,
1052 						   PAGE_SIZE, EFA_MMAP_IO_NC,
1053 						   &resp->db_mmap_key);
1054 		if (!cq->db_mmap_entry) {
1055 			rdma_user_mmap_entry_remove(cq->mmap_entry);
1056 			return -ENOMEM;
1057 		}
1058 
1059 		resp->db_off &= ~PAGE_MASK;
1060 		resp->comp_mask |= EFA_CREATE_CQ_RESP_DB_OFF;
1061 	}
1062 
1063 	return 0;
1064 }
1065 
1066 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1067 		  struct ib_udata *udata)
1068 {
1069 	struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
1070 		udata, struct efa_ucontext, ibucontext);
1071 	struct efa_com_create_cq_params params = {};
1072 	struct efa_ibv_create_cq_resp resp = {};
1073 	struct efa_com_create_cq_result result;
1074 	struct ib_device *ibdev = ibcq->device;
1075 	struct efa_dev *dev = to_edev(ibdev);
1076 	struct efa_ibv_create_cq cmd = {};
1077 	struct efa_cq *cq = to_ecq(ibcq);
1078 	int entries = attr->cqe;
1079 	bool set_src_addr;
1080 	int err;
1081 
1082 	ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
1083 
1084 	if (attr->flags)
1085 		return -EOPNOTSUPP;
1086 
1087 	if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
1088 		ibdev_dbg(ibdev,
1089 			  "cq: requested entries[%u] non-positive or greater than max[%u]\n",
1090 			  entries, dev->dev_attr.max_cq_depth);
1091 		err = -EINVAL;
1092 		goto err_out;
1093 	}
1094 
1095 	if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) {
1096 		ibdev_dbg(ibdev,
1097 			  "Incompatible ABI params, no input udata\n");
1098 		err = -EINVAL;
1099 		goto err_out;
1100 	}
1101 
1102 	if (udata->inlen > sizeof(cmd) &&
1103 	    !ib_is_udata_cleared(udata, sizeof(cmd),
1104 				 udata->inlen - sizeof(cmd))) {
1105 		ibdev_dbg(ibdev,
1106 			  "Incompatible ABI params, unknown fields in udata\n");
1107 		err = -EINVAL;
1108 		goto err_out;
1109 	}
1110 
1111 	err = ib_copy_from_udata(&cmd, udata,
1112 				 min(sizeof(cmd), udata->inlen));
1113 	if (err) {
1114 		ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
1115 		goto err_out;
1116 	}
1117 
1118 	if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_58)) {
1119 		ibdev_dbg(ibdev,
1120 			  "Incompatible ABI params, unknown fields in udata\n");
1121 		err = -EINVAL;
1122 		goto err_out;
1123 	}
1124 
1125 	set_src_addr = !!(cmd.flags & EFA_CREATE_CQ_WITH_SGID);
1126 	if ((cmd.cq_entry_size != sizeof(struct efa_io_rx_cdesc_ex)) &&
1127 	    (set_src_addr ||
1128 	     cmd.cq_entry_size != sizeof(struct efa_io_rx_cdesc))) {
1129 		ibdev_dbg(ibdev,
1130 			  "Invalid entry size [%u]\n", cmd.cq_entry_size);
1131 		err = -EINVAL;
1132 		goto err_out;
1133 	}
1134 
1135 	if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
1136 		ibdev_dbg(ibdev,
1137 			  "Invalid number of sub cqs[%u] expected[%u]\n",
1138 			  cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
1139 		err = -EINVAL;
1140 		goto err_out;
1141 	}
1142 
1143 	cq->ucontext = ucontext;
1144 	cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
1145 	cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
1146 					 DMA_FROM_DEVICE);
1147 	if (!cq->cpu_addr) {
1148 		err = -ENOMEM;
1149 		goto err_out;
1150 	}
1151 
1152 	params.uarn = cq->ucontext->uarn;
1153 	params.cq_depth = entries;
1154 	params.dma_addr = cq->dma_addr;
1155 	params.entry_size_in_bytes = cmd.cq_entry_size;
1156 	params.num_sub_cqs = cmd.num_sub_cqs;
1157 	params.set_src_addr = set_src_addr;
1158 	if (cmd.flags & EFA_CREATE_CQ_WITH_COMPLETION_CHANNEL) {
1159 		cq->eq = efa_vec2eq(dev, attr->comp_vector);
1160 		params.eqn = cq->eq->eeq.eqn;
1161 		params.interrupt_mode_enabled = true;
1162 	}
1163 
1164 	err = efa_com_create_cq(&dev->edev, &params, &result);
1165 	if (err)
1166 		goto err_free_mapped;
1167 
1168 	resp.db_off = result.db_off;
1169 	resp.cq_idx = result.cq_idx;
1170 	cq->cq_idx = result.cq_idx;
1171 	cq->ibcq.cqe = result.actual_depth;
1172 	WARN_ON_ONCE(entries != result.actual_depth);
1173 
1174 	err = cq_mmap_entries_setup(dev, cq, &resp, result.db_valid);
1175 	if (err) {
1176 		ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
1177 			  cq->cq_idx);
1178 		goto err_destroy_cq;
1179 	}
1180 
1181 	if (cq->eq) {
1182 		err = xa_err(xa_store(&dev->cqs_xa, cq->cq_idx, cq, GFP_KERNEL));
1183 		if (err) {
1184 			ibdev_dbg(ibdev, "Failed to store cq[%u] in xarray\n",
1185 				  cq->cq_idx);
1186 			goto err_remove_mmap;
1187 		}
1188 	}
1189 
1190 	if (udata->outlen) {
1191 		err = ib_copy_to_udata(udata, &resp,
1192 				       min(sizeof(resp), udata->outlen));
1193 		if (err) {
1194 			ibdev_dbg(ibdev,
1195 				  "Failed to copy udata for create_cq\n");
1196 			goto err_xa_erase;
1197 		}
1198 	}
1199 
1200 	ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
1201 		  cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
1202 
1203 	return 0;
1204 
1205 err_xa_erase:
1206 	if (cq->eq)
1207 		xa_erase(&dev->cqs_xa, cq->cq_idx);
1208 err_remove_mmap:
1209 	efa_cq_user_mmap_entries_remove(cq);
1210 err_destroy_cq:
1211 	efa_destroy_cq_idx(dev, cq->cq_idx);
1212 err_free_mapped:
1213 	efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1214 			DMA_FROM_DEVICE);
1215 
1216 err_out:
1217 	atomic64_inc(&dev->stats.create_cq_err);
1218 	return err;
1219 }
1220 
1221 static int umem_to_page_list(struct efa_dev *dev,
1222 			     struct ib_umem *umem,
1223 			     u64 *page_list,
1224 			     u32 hp_cnt,
1225 			     u8 hp_shift)
1226 {
1227 	u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1228 	struct ib_block_iter biter;
1229 	unsigned int hp_idx = 0;
1230 
1231 	ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1232 		  hp_cnt, pages_in_hp);
1233 
1234 	rdma_umem_for_each_dma_block(umem, &biter, BIT(hp_shift))
1235 		page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1236 
1237 	return 0;
1238 }
1239 
1240 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1241 {
1242 	struct scatterlist *sglist;
1243 	struct page *pg;
1244 	int i;
1245 
1246 	sglist = kmalloc_array(page_cnt, sizeof(*sglist), GFP_KERNEL);
1247 	if (!sglist)
1248 		return NULL;
1249 	sg_init_table(sglist, page_cnt);
1250 	for (i = 0; i < page_cnt; i++) {
1251 		pg = vmalloc_to_page(buf);
1252 		if (!pg)
1253 			goto err;
1254 		sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1255 		buf += PAGE_SIZE / sizeof(*buf);
1256 	}
1257 	return sglist;
1258 
1259 err:
1260 	kfree(sglist);
1261 	return NULL;
1262 }
1263 
1264 /*
1265  * create a chunk list of physical pages dma addresses from the supplied
1266  * scatter gather list
1267  */
1268 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1269 {
1270 	struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1271 	int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1272 	struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1273 	unsigned int chunk_list_size, chunk_idx, payload_idx;
1274 	int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1275 	struct efa_com_ctrl_buff_info *ctrl_buf;
1276 	u64 *cur_chunk_buf, *prev_chunk_buf;
1277 	struct ib_block_iter biter;
1278 	dma_addr_t dma_addr;
1279 	int i;
1280 
1281 	/* allocate a chunk list that consists of 4KB chunks */
1282 	chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1283 
1284 	chunk_list->size = chunk_list_size;
1285 	chunk_list->chunks = kcalloc(chunk_list_size,
1286 				     sizeof(*chunk_list->chunks),
1287 				     GFP_KERNEL);
1288 	if (!chunk_list->chunks)
1289 		return -ENOMEM;
1290 
1291 	ibdev_dbg(&dev->ibdev,
1292 		  "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1293 		  page_cnt);
1294 
1295 	/* allocate chunk buffers: */
1296 	for (i = 0; i < chunk_list_size; i++) {
1297 		chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1298 		if (!chunk_list->chunks[i].buf)
1299 			goto chunk_list_dealloc;
1300 
1301 		chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1302 	}
1303 	chunk_list->chunks[chunk_list_size - 1].length =
1304 		((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1305 			EFA_CHUNK_PTR_SIZE;
1306 
1307 	/* fill the dma addresses of sg list pages to chunks: */
1308 	chunk_idx = 0;
1309 	payload_idx = 0;
1310 	cur_chunk_buf = chunk_list->chunks[0].buf;
1311 	rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1312 			    EFA_CHUNK_PAYLOAD_SIZE) {
1313 		cur_chunk_buf[payload_idx++] =
1314 			rdma_block_iter_dma_address(&biter);
1315 
1316 		if (payload_idx == EFA_PTRS_PER_CHUNK) {
1317 			chunk_idx++;
1318 			cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1319 			payload_idx = 0;
1320 		}
1321 	}
1322 
1323 	/* map chunks to dma and fill chunks next ptrs */
1324 	for (i = chunk_list_size - 1; i >= 0; i--) {
1325 		dma_addr = dma_map_single(&dev->pdev->dev,
1326 					  chunk_list->chunks[i].buf,
1327 					  chunk_list->chunks[i].length,
1328 					  DMA_TO_DEVICE);
1329 		if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1330 			ibdev_err(&dev->ibdev,
1331 				  "chunk[%u] dma_map_failed\n", i);
1332 			goto chunk_list_unmap;
1333 		}
1334 
1335 		chunk_list->chunks[i].dma_addr = dma_addr;
1336 		ibdev_dbg(&dev->ibdev,
1337 			  "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1338 
1339 		if (!i)
1340 			break;
1341 
1342 		prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1343 
1344 		ctrl_buf = (struct efa_com_ctrl_buff_info *)
1345 				&prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1346 		ctrl_buf->length = chunk_list->chunks[i].length;
1347 
1348 		efa_com_set_dma_addr(dma_addr,
1349 				     &ctrl_buf->address.mem_addr_high,
1350 				     &ctrl_buf->address.mem_addr_low);
1351 	}
1352 
1353 	return 0;
1354 
1355 chunk_list_unmap:
1356 	for (; i < chunk_list_size; i++) {
1357 		dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1358 				 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1359 	}
1360 chunk_list_dealloc:
1361 	for (i = 0; i < chunk_list_size; i++)
1362 		kfree(chunk_list->chunks[i].buf);
1363 
1364 	kfree(chunk_list->chunks);
1365 	return -ENOMEM;
1366 }
1367 
1368 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1369 {
1370 	struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1371 	int i;
1372 
1373 	for (i = 0; i < chunk_list->size; i++) {
1374 		dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1375 				 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1376 		kfree(chunk_list->chunks[i].buf);
1377 	}
1378 
1379 	kfree(chunk_list->chunks);
1380 }
1381 
1382 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1383 static int pbl_continuous_initialize(struct efa_dev *dev,
1384 				     struct pbl_context *pbl)
1385 {
1386 	dma_addr_t dma_addr;
1387 
1388 	dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1389 				  pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1390 	if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1391 		ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1392 		return -ENOMEM;
1393 	}
1394 
1395 	pbl->phys.continuous.dma_addr = dma_addr;
1396 	ibdev_dbg(&dev->ibdev,
1397 		  "pbl continuous - dma_addr = %pad, size[%u]\n",
1398 		  &dma_addr, pbl->pbl_buf_size_in_bytes);
1399 
1400 	return 0;
1401 }
1402 
1403 /*
1404  * initialize pbl indirect mode:
1405  * create a chunk list out of the dma addresses of the physical pages of
1406  * pbl buffer.
1407  */
1408 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1409 {
1410 	u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, EFA_CHUNK_PAYLOAD_SIZE);
1411 	struct scatterlist *sgl;
1412 	int sg_dma_cnt, err;
1413 
1414 	BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1415 	sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1416 	if (!sgl)
1417 		return -ENOMEM;
1418 
1419 	sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1420 	if (!sg_dma_cnt) {
1421 		err = -EINVAL;
1422 		goto err_map;
1423 	}
1424 
1425 	pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1426 	pbl->phys.indirect.sgl = sgl;
1427 	pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1428 	err = pbl_chunk_list_create(dev, pbl);
1429 	if (err) {
1430 		ibdev_dbg(&dev->ibdev,
1431 			  "chunk_list creation failed[%d]\n", err);
1432 		goto err_chunk;
1433 	}
1434 
1435 	ibdev_dbg(&dev->ibdev,
1436 		  "pbl indirect - size[%u], chunks[%u]\n",
1437 		  pbl->pbl_buf_size_in_bytes,
1438 		  pbl->phys.indirect.chunk_list.size);
1439 
1440 	return 0;
1441 
1442 err_chunk:
1443 	dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1444 err_map:
1445 	kfree(sgl);
1446 	return err;
1447 }
1448 
1449 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1450 {
1451 	pbl_chunk_list_destroy(dev, pbl);
1452 	dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1453 		     pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1454 	kfree(pbl->phys.indirect.sgl);
1455 }
1456 
1457 /* create a page buffer list from a mapped user memory region */
1458 static int pbl_create(struct efa_dev *dev,
1459 		      struct pbl_context *pbl,
1460 		      struct ib_umem *umem,
1461 		      int hp_cnt,
1462 		      u8 hp_shift)
1463 {
1464 	int err;
1465 
1466 	pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1467 	pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1468 	if (!pbl->pbl_buf)
1469 		return -ENOMEM;
1470 
1471 	if (is_vmalloc_addr(pbl->pbl_buf)) {
1472 		pbl->physically_continuous = 0;
1473 		err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1474 					hp_shift);
1475 		if (err)
1476 			goto err_free;
1477 
1478 		err = pbl_indirect_initialize(dev, pbl);
1479 		if (err)
1480 			goto err_free;
1481 	} else {
1482 		pbl->physically_continuous = 1;
1483 		err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1484 					hp_shift);
1485 		if (err)
1486 			goto err_free;
1487 
1488 		err = pbl_continuous_initialize(dev, pbl);
1489 		if (err)
1490 			goto err_free;
1491 	}
1492 
1493 	ibdev_dbg(&dev->ibdev,
1494 		  "user_pbl_created: user_pages[%u], continuous[%u]\n",
1495 		  hp_cnt, pbl->physically_continuous);
1496 
1497 	return 0;
1498 
1499 err_free:
1500 	kvfree(pbl->pbl_buf);
1501 	return err;
1502 }
1503 
1504 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1505 {
1506 	if (pbl->physically_continuous)
1507 		dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1508 				 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1509 	else
1510 		pbl_indirect_terminate(dev, pbl);
1511 
1512 	kvfree(pbl->pbl_buf);
1513 }
1514 
1515 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1516 				 struct efa_com_reg_mr_params *params)
1517 {
1518 	int err;
1519 
1520 	params->inline_pbl = 1;
1521 	err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1522 				params->page_num, params->page_shift);
1523 	if (err)
1524 		return err;
1525 
1526 	ibdev_dbg(&dev->ibdev,
1527 		  "inline_pbl_array - pages[%u]\n", params->page_num);
1528 
1529 	return 0;
1530 }
1531 
1532 static int efa_create_pbl(struct efa_dev *dev,
1533 			  struct pbl_context *pbl,
1534 			  struct efa_mr *mr,
1535 			  struct efa_com_reg_mr_params *params)
1536 {
1537 	int err;
1538 
1539 	err = pbl_create(dev, pbl, mr->umem, params->page_num,
1540 			 params->page_shift);
1541 	if (err) {
1542 		ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1543 		return err;
1544 	}
1545 
1546 	params->inline_pbl = 0;
1547 	params->indirect = !pbl->physically_continuous;
1548 	if (pbl->physically_continuous) {
1549 		params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1550 
1551 		efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1552 				     &params->pbl.pbl.address.mem_addr_high,
1553 				     &params->pbl.pbl.address.mem_addr_low);
1554 	} else {
1555 		params->pbl.pbl.length =
1556 			pbl->phys.indirect.chunk_list.chunks[0].length;
1557 
1558 		efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1559 				     &params->pbl.pbl.address.mem_addr_high,
1560 				     &params->pbl.pbl.address.mem_addr_low);
1561 	}
1562 
1563 	return 0;
1564 }
1565 
1566 static struct efa_mr *efa_alloc_mr(struct ib_pd *ibpd, int access_flags,
1567 				   struct ib_udata *udata)
1568 {
1569 	struct efa_dev *dev = to_edev(ibpd->device);
1570 	int supp_access_flags;
1571 	struct efa_mr *mr;
1572 
1573 	if (udata && udata->inlen &&
1574 	    !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1575 		ibdev_dbg(&dev->ibdev,
1576 			  "Incompatible ABI params, udata not cleared\n");
1577 		return ERR_PTR(-EINVAL);
1578 	}
1579 
1580 	supp_access_flags =
1581 		IB_ACCESS_LOCAL_WRITE |
1582 		(EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0) |
1583 		(EFA_DEV_CAP(dev, RDMA_WRITE) ? IB_ACCESS_REMOTE_WRITE : 0);
1584 
1585 	access_flags &= ~IB_ACCESS_OPTIONAL;
1586 	if (access_flags & ~supp_access_flags) {
1587 		ibdev_dbg(&dev->ibdev,
1588 			  "Unsupported access flags[%#x], supported[%#x]\n",
1589 			  access_flags, supp_access_flags);
1590 		return ERR_PTR(-EOPNOTSUPP);
1591 	}
1592 
1593 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1594 	if (!mr)
1595 		return ERR_PTR(-ENOMEM);
1596 
1597 	return mr;
1598 }
1599 
1600 static int efa_register_mr(struct ib_pd *ibpd, struct efa_mr *mr, u64 start,
1601 			   u64 length, u64 virt_addr, int access_flags)
1602 {
1603 	struct efa_dev *dev = to_edev(ibpd->device);
1604 	struct efa_com_reg_mr_params params = {};
1605 	struct efa_com_reg_mr_result result = {};
1606 	struct pbl_context pbl;
1607 	unsigned int pg_sz;
1608 	int inline_size;
1609 	int err;
1610 
1611 	params.pd = to_epd(ibpd)->pdn;
1612 	params.iova = virt_addr;
1613 	params.mr_length_in_bytes = length;
1614 	params.permissions = access_flags;
1615 
1616 	pg_sz = ib_umem_find_best_pgsz(mr->umem,
1617 				       dev->dev_attr.page_size_cap,
1618 				       virt_addr);
1619 	if (!pg_sz) {
1620 		ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1621 			  dev->dev_attr.page_size_cap);
1622 		return -EOPNOTSUPP;
1623 	}
1624 
1625 	params.page_shift = order_base_2(pg_sz);
1626 	params.page_num = ib_umem_num_dma_blocks(mr->umem, pg_sz);
1627 
1628 	ibdev_dbg(&dev->ibdev,
1629 		  "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1630 		  start, length, params.page_shift, params.page_num);
1631 
1632 	inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1633 	if (params.page_num <= inline_size) {
1634 		err = efa_create_inline_pbl(dev, mr, &params);
1635 		if (err)
1636 			return err;
1637 
1638 		err = efa_com_register_mr(&dev->edev, &params, &result);
1639 		if (err)
1640 			return err;
1641 	} else {
1642 		err = efa_create_pbl(dev, &pbl, mr, &params);
1643 		if (err)
1644 			return err;
1645 
1646 		err = efa_com_register_mr(&dev->edev, &params, &result);
1647 		pbl_destroy(dev, &pbl);
1648 
1649 		if (err)
1650 			return err;
1651 	}
1652 
1653 	mr->ibmr.lkey = result.l_key;
1654 	mr->ibmr.rkey = result.r_key;
1655 	mr->ibmr.length = length;
1656 	ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1657 
1658 	return 0;
1659 }
1660 
1661 struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start,
1662 				     u64 length, u64 virt_addr,
1663 				     int fd, int access_flags,
1664 				     struct ib_udata *udata)
1665 {
1666 	struct efa_dev *dev = to_edev(ibpd->device);
1667 	struct ib_umem_dmabuf *umem_dmabuf;
1668 	struct efa_mr *mr;
1669 	int err;
1670 
1671 	mr = efa_alloc_mr(ibpd, access_flags, udata);
1672 	if (IS_ERR(mr)) {
1673 		err = PTR_ERR(mr);
1674 		goto err_out;
1675 	}
1676 
1677 	umem_dmabuf = ib_umem_dmabuf_get_pinned(ibpd->device, start, length, fd,
1678 						access_flags);
1679 	if (IS_ERR(umem_dmabuf)) {
1680 		err = PTR_ERR(umem_dmabuf);
1681 		ibdev_dbg(&dev->ibdev, "Failed to get dmabuf umem[%d]\n", err);
1682 		goto err_free;
1683 	}
1684 
1685 	mr->umem = &umem_dmabuf->umem;
1686 	err = efa_register_mr(ibpd, mr, start, length, virt_addr, access_flags);
1687 	if (err)
1688 		goto err_release;
1689 
1690 	return &mr->ibmr;
1691 
1692 err_release:
1693 	ib_umem_release(mr->umem);
1694 err_free:
1695 	kfree(mr);
1696 err_out:
1697 	atomic64_inc(&dev->stats.reg_mr_err);
1698 	return ERR_PTR(err);
1699 }
1700 
1701 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1702 			 u64 virt_addr, int access_flags,
1703 			 struct ib_udata *udata)
1704 {
1705 	struct efa_dev *dev = to_edev(ibpd->device);
1706 	struct efa_mr *mr;
1707 	int err;
1708 
1709 	mr = efa_alloc_mr(ibpd, access_flags, udata);
1710 	if (IS_ERR(mr)) {
1711 		err = PTR_ERR(mr);
1712 		goto err_out;
1713 	}
1714 
1715 	mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
1716 	if (IS_ERR(mr->umem)) {
1717 		err = PTR_ERR(mr->umem);
1718 		ibdev_dbg(&dev->ibdev,
1719 			  "Failed to pin and map user space memory[%d]\n", err);
1720 		goto err_free;
1721 	}
1722 
1723 	err = efa_register_mr(ibpd, mr, start, length, virt_addr, access_flags);
1724 	if (err)
1725 		goto err_release;
1726 
1727 	return &mr->ibmr;
1728 
1729 err_release:
1730 	ib_umem_release(mr->umem);
1731 err_free:
1732 	kfree(mr);
1733 err_out:
1734 	atomic64_inc(&dev->stats.reg_mr_err);
1735 	return ERR_PTR(err);
1736 }
1737 
1738 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1739 {
1740 	struct efa_dev *dev = to_edev(ibmr->device);
1741 	struct efa_com_dereg_mr_params params;
1742 	struct efa_mr *mr = to_emr(ibmr);
1743 	int err;
1744 
1745 	ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1746 
1747 	params.l_key = mr->ibmr.lkey;
1748 	err = efa_com_dereg_mr(&dev->edev, &params);
1749 	if (err)
1750 		return err;
1751 
1752 	ib_umem_release(mr->umem);
1753 	kfree(mr);
1754 
1755 	return 0;
1756 }
1757 
1758 int efa_get_port_immutable(struct ib_device *ibdev, u32 port_num,
1759 			   struct ib_port_immutable *immutable)
1760 {
1761 	struct ib_port_attr attr;
1762 	int err;
1763 
1764 	err = ib_query_port(ibdev, port_num, &attr);
1765 	if (err) {
1766 		ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1767 		return err;
1768 	}
1769 
1770 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
1771 	immutable->gid_tbl_len = attr.gid_tbl_len;
1772 
1773 	return 0;
1774 }
1775 
1776 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1777 {
1778 	struct efa_com_dealloc_uar_params params = {
1779 		.uarn = uarn,
1780 	};
1781 
1782 	return efa_com_dealloc_uar(&dev->edev, &params);
1783 }
1784 
1785 #define EFA_CHECK_USER_COMP(_dev, _comp_mask, _attr, _mask, _attr_str) \
1786 	(_attr_str = (!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask))) ? \
1787 		     NULL : #_attr)
1788 
1789 static int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
1790 				   const struct efa_ibv_alloc_ucontext_cmd *cmd)
1791 {
1792 	struct efa_dev *dev = to_edev(ibucontext->device);
1793 	char *attr_str;
1794 
1795 	if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, max_tx_batch,
1796 				EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH, attr_str))
1797 		goto err;
1798 
1799 	if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, min_sq_depth,
1800 				EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR,
1801 				attr_str))
1802 		goto err;
1803 
1804 	return 0;
1805 
1806 err:
1807 	ibdev_dbg(&dev->ibdev, "Userspace handshake failed for %s attribute\n",
1808 		  attr_str);
1809 	return -EOPNOTSUPP;
1810 }
1811 
1812 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1813 {
1814 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1815 	struct efa_dev *dev = to_edev(ibucontext->device);
1816 	struct efa_ibv_alloc_ucontext_resp resp = {};
1817 	struct efa_ibv_alloc_ucontext_cmd cmd = {};
1818 	struct efa_com_alloc_uar_result result;
1819 	int err;
1820 
1821 	/*
1822 	 * it's fine if the driver does not know all request fields,
1823 	 * we will ack input fields in our response.
1824 	 */
1825 
1826 	err = ib_copy_from_udata(&cmd, udata,
1827 				 min(sizeof(cmd), udata->inlen));
1828 	if (err) {
1829 		ibdev_dbg(&dev->ibdev,
1830 			  "Cannot copy udata for alloc_ucontext\n");
1831 		goto err_out;
1832 	}
1833 
1834 	err = efa_user_comp_handshake(ibucontext, &cmd);
1835 	if (err)
1836 		goto err_out;
1837 
1838 	err = efa_com_alloc_uar(&dev->edev, &result);
1839 	if (err)
1840 		goto err_out;
1841 
1842 	ucontext->uarn = result.uarn;
1843 
1844 	resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1845 	resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1846 	resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1847 	resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1848 	resp.max_llq_size = dev->dev_attr.max_llq_size;
1849 	resp.max_tx_batch = dev->dev_attr.max_tx_batch;
1850 	resp.min_sq_wr = dev->dev_attr.min_sq_depth;
1851 
1852 	err = ib_copy_to_udata(udata, &resp,
1853 			       min(sizeof(resp), udata->outlen));
1854 	if (err)
1855 		goto err_dealloc_uar;
1856 
1857 	return 0;
1858 
1859 err_dealloc_uar:
1860 	efa_dealloc_uar(dev, result.uarn);
1861 err_out:
1862 	atomic64_inc(&dev->stats.alloc_ucontext_err);
1863 	return err;
1864 }
1865 
1866 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1867 {
1868 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1869 	struct efa_dev *dev = to_edev(ibucontext->device);
1870 
1871 	efa_dealloc_uar(dev, ucontext->uarn);
1872 }
1873 
1874 void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1875 {
1876 	struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
1877 
1878 	kfree(entry);
1879 }
1880 
1881 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1882 		      struct vm_area_struct *vma)
1883 {
1884 	struct rdma_user_mmap_entry *rdma_entry;
1885 	struct efa_user_mmap_entry *entry;
1886 	unsigned long va;
1887 	int err = 0;
1888 	u64 pfn;
1889 
1890 	rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
1891 	if (!rdma_entry) {
1892 		ibdev_dbg(&dev->ibdev,
1893 			  "pgoff[%#lx] does not have valid entry\n",
1894 			  vma->vm_pgoff);
1895 		atomic64_inc(&dev->stats.mmap_err);
1896 		return -EINVAL;
1897 	}
1898 	entry = to_emmap(rdma_entry);
1899 
1900 	ibdev_dbg(&dev->ibdev,
1901 		  "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
1902 		  entry->address, rdma_entry->npages * PAGE_SIZE,
1903 		  entry->mmap_flag);
1904 
1905 	pfn = entry->address >> PAGE_SHIFT;
1906 	switch (entry->mmap_flag) {
1907 	case EFA_MMAP_IO_NC:
1908 		err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1909 					entry->rdma_entry.npages * PAGE_SIZE,
1910 					pgprot_noncached(vma->vm_page_prot),
1911 					rdma_entry);
1912 		break;
1913 	case EFA_MMAP_IO_WC:
1914 		err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1915 					entry->rdma_entry.npages * PAGE_SIZE,
1916 					pgprot_writecombine(vma->vm_page_prot),
1917 					rdma_entry);
1918 		break;
1919 	case EFA_MMAP_DMA_PAGE:
1920 		for (va = vma->vm_start; va < vma->vm_end;
1921 		     va += PAGE_SIZE, pfn++) {
1922 			err = vm_insert_page(vma, va, pfn_to_page(pfn));
1923 			if (err)
1924 				break;
1925 		}
1926 		break;
1927 	default:
1928 		err = -EINVAL;
1929 	}
1930 
1931 	if (err) {
1932 		ibdev_dbg(
1933 			&dev->ibdev,
1934 			"Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
1935 			entry->address, rdma_entry->npages * PAGE_SIZE,
1936 			entry->mmap_flag, err);
1937 		atomic64_inc(&dev->stats.mmap_err);
1938 	}
1939 
1940 	rdma_user_mmap_entry_put(rdma_entry);
1941 	return err;
1942 }
1943 
1944 int efa_mmap(struct ib_ucontext *ibucontext,
1945 	     struct vm_area_struct *vma)
1946 {
1947 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1948 	struct efa_dev *dev = to_edev(ibucontext->device);
1949 	size_t length = vma->vm_end - vma->vm_start;
1950 
1951 	ibdev_dbg(&dev->ibdev,
1952 		  "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
1953 		  vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
1954 
1955 	return __efa_mmap(dev, ucontext, vma);
1956 }
1957 
1958 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1959 {
1960 	struct efa_com_destroy_ah_params params = {
1961 		.ah = ah->ah,
1962 		.pdn = to_epd(ah->ibah.pd)->pdn,
1963 	};
1964 
1965 	return efa_com_destroy_ah(&dev->edev, &params);
1966 }
1967 
1968 int efa_create_ah(struct ib_ah *ibah,
1969 		  struct rdma_ah_init_attr *init_attr,
1970 		  struct ib_udata *udata)
1971 {
1972 	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
1973 	struct efa_dev *dev = to_edev(ibah->device);
1974 	struct efa_com_create_ah_params params = {};
1975 	struct efa_ibv_create_ah_resp resp = {};
1976 	struct efa_com_create_ah_result result;
1977 	struct efa_ah *ah = to_eah(ibah);
1978 	int err;
1979 
1980 	if (!(init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
1981 		ibdev_dbg(&dev->ibdev,
1982 			  "Create address handle is not supported in atomic context\n");
1983 		err = -EOPNOTSUPP;
1984 		goto err_out;
1985 	}
1986 
1987 	if (udata->inlen &&
1988 	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1989 		ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1990 		err = -EINVAL;
1991 		goto err_out;
1992 	}
1993 
1994 	memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1995 	       sizeof(params.dest_addr));
1996 	params.pdn = to_epd(ibah->pd)->pdn;
1997 	err = efa_com_create_ah(&dev->edev, &params, &result);
1998 	if (err)
1999 		goto err_out;
2000 
2001 	memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
2002 	ah->ah = result.ah;
2003 
2004 	resp.efa_address_handle = result.ah;
2005 
2006 	if (udata->outlen) {
2007 		err = ib_copy_to_udata(udata, &resp,
2008 				       min(sizeof(resp), udata->outlen));
2009 		if (err) {
2010 			ibdev_dbg(&dev->ibdev,
2011 				  "Failed to copy udata for create_ah response\n");
2012 			goto err_destroy_ah;
2013 		}
2014 	}
2015 	ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
2016 
2017 	return 0;
2018 
2019 err_destroy_ah:
2020 	efa_ah_destroy(dev, ah);
2021 err_out:
2022 	atomic64_inc(&dev->stats.create_ah_err);
2023 	return err;
2024 }
2025 
2026 int efa_destroy_ah(struct ib_ah *ibah, u32 flags)
2027 {
2028 	struct efa_dev *dev = to_edev(ibah->pd->device);
2029 	struct efa_ah *ah = to_eah(ibah);
2030 
2031 	ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
2032 
2033 	if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
2034 		ibdev_dbg(&dev->ibdev,
2035 			  "Destroy address handle is not supported in atomic context\n");
2036 		return -EOPNOTSUPP;
2037 	}
2038 
2039 	efa_ah_destroy(dev, ah);
2040 	return 0;
2041 }
2042 
2043 struct rdma_hw_stats *efa_alloc_hw_port_stats(struct ib_device *ibdev,
2044 					      u32 port_num)
2045 {
2046 	return rdma_alloc_hw_stats_struct(efa_port_stats_descs,
2047 					  ARRAY_SIZE(efa_port_stats_descs),
2048 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2049 }
2050 
2051 struct rdma_hw_stats *efa_alloc_hw_device_stats(struct ib_device *ibdev)
2052 {
2053 	return rdma_alloc_hw_stats_struct(efa_device_stats_descs,
2054 					  ARRAY_SIZE(efa_device_stats_descs),
2055 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2056 }
2057 
2058 static int efa_fill_device_stats(struct efa_dev *dev,
2059 				 struct rdma_hw_stats *stats)
2060 {
2061 	struct efa_com_stats_admin *as = &dev->edev.aq.stats;
2062 	struct efa_stats *s = &dev->stats;
2063 
2064 	stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
2065 	stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
2066 	stats->value[EFA_CMDS_ERR] = atomic64_read(&as->cmd_err);
2067 	stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
2068 
2069 	stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
2070 	stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->alloc_pd_err);
2071 	stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->create_qp_err);
2072 	stats->value[EFA_CREATE_CQ_ERR] = atomic64_read(&s->create_cq_err);
2073 	stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->reg_mr_err);
2074 	stats->value[EFA_ALLOC_UCONTEXT_ERR] =
2075 		atomic64_read(&s->alloc_ucontext_err);
2076 	stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->create_ah_err);
2077 	stats->value[EFA_MMAP_ERR] = atomic64_read(&s->mmap_err);
2078 
2079 	return ARRAY_SIZE(efa_device_stats_descs);
2080 }
2081 
2082 static int efa_fill_port_stats(struct efa_dev *dev, struct rdma_hw_stats *stats,
2083 			       u32 port_num)
2084 {
2085 	struct efa_com_get_stats_params params = {};
2086 	union efa_com_get_stats_result result;
2087 	struct efa_com_rdma_write_stats *rws;
2088 	struct efa_com_rdma_read_stats *rrs;
2089 	struct efa_com_messages_stats *ms;
2090 	struct efa_com_basic_stats *bs;
2091 	int err;
2092 
2093 	params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
2094 	params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
2095 
2096 	err = efa_com_get_stats(&dev->edev, &params, &result);
2097 	if (err)
2098 		return err;
2099 
2100 	bs = &result.basic_stats;
2101 	stats->value[EFA_TX_BYTES] = bs->tx_bytes;
2102 	stats->value[EFA_TX_PKTS] = bs->tx_pkts;
2103 	stats->value[EFA_RX_BYTES] = bs->rx_bytes;
2104 	stats->value[EFA_RX_PKTS] = bs->rx_pkts;
2105 	stats->value[EFA_RX_DROPS] = bs->rx_drops;
2106 
2107 	params.type = EFA_ADMIN_GET_STATS_TYPE_MESSAGES;
2108 	err = efa_com_get_stats(&dev->edev, &params, &result);
2109 	if (err)
2110 		return err;
2111 
2112 	ms = &result.messages_stats;
2113 	stats->value[EFA_SEND_BYTES] = ms->send_bytes;
2114 	stats->value[EFA_SEND_WRS] = ms->send_wrs;
2115 	stats->value[EFA_RECV_BYTES] = ms->recv_bytes;
2116 	stats->value[EFA_RECV_WRS] = ms->recv_wrs;
2117 
2118 	params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_READ;
2119 	err = efa_com_get_stats(&dev->edev, &params, &result);
2120 	if (err)
2121 		return err;
2122 
2123 	rrs = &result.rdma_read_stats;
2124 	stats->value[EFA_RDMA_READ_WRS] = rrs->read_wrs;
2125 	stats->value[EFA_RDMA_READ_BYTES] = rrs->read_bytes;
2126 	stats->value[EFA_RDMA_READ_WR_ERR] = rrs->read_wr_err;
2127 	stats->value[EFA_RDMA_READ_RESP_BYTES] = rrs->read_resp_bytes;
2128 
2129 	if (EFA_DEV_CAP(dev, RDMA_WRITE)) {
2130 		params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_WRITE;
2131 		err = efa_com_get_stats(&dev->edev, &params, &result);
2132 		if (err)
2133 			return err;
2134 
2135 		rws = &result.rdma_write_stats;
2136 		stats->value[EFA_RDMA_WRITE_WRS] = rws->write_wrs;
2137 		stats->value[EFA_RDMA_WRITE_BYTES] = rws->write_bytes;
2138 		stats->value[EFA_RDMA_WRITE_WR_ERR] = rws->write_wr_err;
2139 		stats->value[EFA_RDMA_WRITE_RECV_BYTES] = rws->write_recv_bytes;
2140 	}
2141 
2142 	return ARRAY_SIZE(efa_port_stats_descs);
2143 }
2144 
2145 int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
2146 		     u32 port_num, int index)
2147 {
2148 	if (port_num)
2149 		return efa_fill_port_stats(to_edev(ibdev), stats, port_num);
2150 	else
2151 		return efa_fill_device_stats(to_edev(ibdev), stats);
2152 }
2153 
2154 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
2155 					 u32 port_num)
2156 {
2157 	return IB_LINK_LAYER_UNSPECIFIED;
2158 }
2159 
2160