efa_verbs.c

   1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2/*
   3 * Copyright 2018-2024 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#define UVERBS_MODULE_NAME efa_ib
  17#include <rdma/uverbs_named_ioctl.h>
  18#include <rdma/ib_user_ioctl_cmds.h>
  19
  20#include "efa.h"
  21#include "efa_io_defs.h"
  22
  23enum {
  24	EFA_MMAP_DMA_PAGE = 0,
  25	EFA_MMAP_IO_WC,
  26	EFA_MMAP_IO_NC,
  27};
  28
  29struct efa_user_mmap_entry {
  30	struct rdma_user_mmap_entry rdma_entry;
  31	u64 address;
  32	u8 mmap_flag;
  33};
  34
  35#define EFA_DEFINE_DEVICE_STATS(op) \
  36	op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
  37	op(EFA_COMPLETED_CMDS, "completed_cmds") \
  38	op(EFA_CMDS_ERR, "cmds_err") \
  39	op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \
  40	op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \
  41	op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \
  42	op(EFA_CREATE_QP_ERR, "create_qp_err") \
  43	op(EFA_CREATE_CQ_ERR, "create_cq_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	op(EFA_MMAP_ERR, "mmap_err")
  48
  49#define EFA_DEFINE_PORT_STATS(op) \
  50	op(EFA_TX_BYTES, "tx_bytes") \
  51	op(EFA_TX_PKTS, "tx_pkts") \
  52	op(EFA_RX_BYTES, "rx_bytes") \
  53	op(EFA_RX_PKTS, "rx_pkts") \
  54	op(EFA_RX_DROPS, "rx_drops") \
  55	op(EFA_SEND_BYTES, "send_bytes") \
  56	op(EFA_SEND_WRS, "send_wrs") \
  57	op(EFA_RECV_BYTES, "recv_bytes") \
  58	op(EFA_RECV_WRS, "recv_wrs") \
  59	op(EFA_RDMA_READ_WRS, "rdma_read_wrs") \
  60	op(EFA_RDMA_READ_BYTES, "rdma_read_bytes") \
  61	op(EFA_RDMA_READ_WR_ERR, "rdma_read_wr_err") \
  62	op(EFA_RDMA_READ_RESP_BYTES, "rdma_read_resp_bytes") \
  63	op(EFA_RDMA_WRITE_WRS, "rdma_write_wrs") \
  64	op(EFA_RDMA_WRITE_BYTES, "rdma_write_bytes") \
  65	op(EFA_RDMA_WRITE_WR_ERR, "rdma_write_wr_err") \
  66	op(EFA_RDMA_WRITE_RECV_BYTES, "rdma_write_recv_bytes") \
  67	op(EFA_RETRANS_BYTES, "retrans_bytes") \
  68	op(EFA_RETRANS_PKTS, "retrans_pkts") \
  69	op(EFA_RETRANS_TIMEOUT_EVENS, "retrans_timeout_events") \
  70	op(EFA_UNRESPONSIVE_REMOTE_EVENTS, "unresponsive_remote_events") \
  71	op(EFA_IMPAIRED_REMOTE_CONN_EVENTS, "impaired_remote_conn_events") \
  72
  73#define EFA_STATS_ENUM(ename, name) ename,
  74#define EFA_STATS_STR(ename, nam) \
  75	[ename].name = nam,
  76
  77enum efa_hw_device_stats {
  78	EFA_DEFINE_DEVICE_STATS(EFA_STATS_ENUM)
  79};
  80
  81static const struct rdma_stat_desc efa_device_stats_descs[] = {
  82	EFA_DEFINE_DEVICE_STATS(EFA_STATS_STR)
  83};
  84
  85enum efa_hw_port_stats {
  86	EFA_DEFINE_PORT_STATS(EFA_STATS_ENUM)
  87};
  88
  89static const struct rdma_stat_desc efa_port_stats_descs[] = {
  90	EFA_DEFINE_PORT_STATS(EFA_STATS_STR)
  91};
  92
  93#define EFA_DEFAULT_LINK_SPEED_GBPS   100
  94
  95#define EFA_CHUNK_PAYLOAD_SHIFT       12
  96#define EFA_CHUNK_PAYLOAD_SIZE        BIT(EFA_CHUNK_PAYLOAD_SHIFT)
  97#define EFA_CHUNK_PAYLOAD_PTR_SIZE    8
  98
  99#define EFA_CHUNK_SHIFT               12
 100#define EFA_CHUNK_SIZE                BIT(EFA_CHUNK_SHIFT)
 101#define EFA_CHUNK_PTR_SIZE            sizeof(struct efa_com_ctrl_buff_info)
 102
 103#define EFA_PTRS_PER_CHUNK \
 104	((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
 105
 106#define EFA_CHUNK_USED_SIZE \
 107	((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
 108
 109struct pbl_chunk {
 110	dma_addr_t dma_addr;
 111	u64 *buf;
 112	u32 length;
 113};
 114
 115struct pbl_chunk_list {
 116	struct pbl_chunk *chunks;
 117	unsigned int size;
 118};
 119
 120struct pbl_context {
 121	union {
 122		struct {
 123			dma_addr_t dma_addr;
 124		} continuous;
 125		struct {
 126			u32 pbl_buf_size_in_pages;
 127			struct scatterlist *sgl;
 128			int sg_dma_cnt;
 129			struct pbl_chunk_list chunk_list;
 130		} indirect;
 131	} phys;
 132	u64 *pbl_buf;
 133	u32 pbl_buf_size_in_bytes;
 134	u8 physically_continuous;
 135};
 136
 137static inline struct efa_dev *to_edev(struct ib_device *ibdev)
 138{
 139	return container_of(ibdev, struct efa_dev, ibdev);
 140}
 141
 142static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
 143{
 144	return container_of(ibucontext, struct efa_ucontext, ibucontext);
 145}
 146
 147static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
 148{
 149	return container_of(ibpd, struct efa_pd, ibpd);
 150}
 151
 152static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
 153{
 154	return container_of(ibmr, struct efa_mr, ibmr);
 155}
 156
 157static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
 158{
 159	return container_of(ibqp, struct efa_qp, ibqp);
 160}
 161
 162static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
 163{
 164	return container_of(ibcq, struct efa_cq, ibcq);
 165}
 166
 167static inline struct efa_ah *to_eah(struct ib_ah *ibah)
 168{
 169	return container_of(ibah, struct efa_ah, ibah);
 170}
 171
 172static inline struct efa_user_mmap_entry *
 173to_emmap(struct rdma_user_mmap_entry *rdma_entry)
 174{
 175	return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
 176}
 177
 178#define EFA_DEV_CAP(dev, cap) \
 179	((dev)->dev_attr.device_caps & \
 180	 EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_##cap##_MASK)
 181
 182#define is_reserved_cleared(reserved) \
 183	!memchr_inv(reserved, 0, sizeof(reserved))
 184
 185static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
 186			       size_t size, enum dma_data_direction dir)
 187{
 188	void *addr;
 189
 190	addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
 191	if (!addr)
 192		return NULL;
 193
 194	*dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
 195	if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
 196		ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
 197		free_pages_exact(addr, size);
 198		return NULL;
 199	}
 200
 201	return addr;
 202}
 203
 204static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr,
 205			    dma_addr_t dma_addr,
 206			    size_t size, enum dma_data_direction dir)
 207{
 208	dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir);
 209	free_pages_exact(cpu_addr, size);
 210}
 211
 212int efa_query_device(struct ib_device *ibdev,
 213		     struct ib_device_attr *props,
 214		     struct ib_udata *udata)
 215{
 216	struct efa_com_get_device_attr_result *dev_attr;
 217	struct efa_ibv_ex_query_device_resp resp = {};
 218	struct efa_dev *dev = to_edev(ibdev);
 219	int err;
 220
 221	if (udata && udata->inlen &&
 222	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 223		ibdev_dbg(ibdev,
 224			  "Incompatible ABI params, udata not cleared\n");
 225		return -EINVAL;
 226	}
 227
 228	dev_attr = &dev->dev_attr;
 229
 230	memset(props, 0, sizeof(*props));
 231	props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
 232	props->page_size_cap = dev_attr->page_size_cap;
 233	props->vendor_id = dev->pdev->vendor;
 234	props->vendor_part_id = dev->pdev->device;
 235	props->hw_ver = dev->pdev->subsystem_device;
 236	props->max_qp = dev_attr->max_qp;
 237	props->max_cq = dev_attr->max_cq;
 238	props->max_pd = dev_attr->max_pd;
 239	props->max_mr = dev_attr->max_mr;
 240	props->max_ah = dev_attr->max_ah;
 241	props->max_cqe = dev_attr->max_cq_depth;
 242	props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
 243				 dev_attr->max_rq_depth);
 244	props->max_send_sge = dev_attr->max_sq_sge;
 245	props->max_recv_sge = dev_attr->max_rq_sge;
 246	props->max_sge_rd = dev_attr->max_wr_rdma_sge;
 247	props->max_pkeys = 1;
 248
 249	if (udata && udata->outlen) {
 250		resp.max_sq_sge = dev_attr->max_sq_sge;
 251		resp.max_rq_sge = dev_attr->max_rq_sge;
 252		resp.max_sq_wr = dev_attr->max_sq_depth;
 253		resp.max_rq_wr = dev_attr->max_rq_depth;
 254		resp.max_rdma_size = dev_attr->max_rdma_size;
 255
 256		resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID;
 257		resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_WITH_EXT_MEM;
 258		if (EFA_DEV_CAP(dev, RDMA_READ))
 259			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
 260
 261		if (EFA_DEV_CAP(dev, RNR_RETRY))
 262			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RNR_RETRY;
 263
 264		if (EFA_DEV_CAP(dev, DATA_POLLING_128))
 265			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128;
 266
 267		if (EFA_DEV_CAP(dev, RDMA_WRITE))
 268			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_WRITE;
 269
 270		if (EFA_DEV_CAP(dev, UNSOLICITED_WRITE_RECV))
 271			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_UNSOLICITED_WRITE_RECV;
 272
 273		if (dev->neqs)
 274			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS;
 275
 276		err = ib_copy_to_udata(udata, &resp,
 277				       min(sizeof(resp), udata->outlen));
 278		if (err) {
 279			ibdev_dbg(ibdev,
 280				  "Failed to copy udata for query_device\n");
 281			return err;
 282		}
 283	}
 284
 285	return 0;
 286}
 287
 288static void efa_link_gbps_to_speed_and_width(u16 gbps,
 289					     enum ib_port_speed *speed,
 290					     enum ib_port_width *width)
 291{
 292	if (gbps >= 400) {
 293		*width = IB_WIDTH_8X;
 294		*speed = IB_SPEED_HDR;
 295	} else if (gbps >= 200) {
 296		*width = IB_WIDTH_4X;
 297		*speed = IB_SPEED_HDR;
 298	} else if (gbps >= 120) {
 299		*width = IB_WIDTH_12X;
 300		*speed = IB_SPEED_FDR10;
 301	} else if (gbps >= 100) {
 302		*width = IB_WIDTH_4X;
 303		*speed = IB_SPEED_EDR;
 304	} else if (gbps >= 60) {
 305		*width = IB_WIDTH_12X;
 306		*speed = IB_SPEED_DDR;
 307	} else if (gbps >= 50) {
 308		*width = IB_WIDTH_1X;
 309		*speed = IB_SPEED_HDR;
 310	} else if (gbps >= 40) {
 311		*width = IB_WIDTH_4X;
 312		*speed = IB_SPEED_FDR10;
 313	} else if (gbps >= 30) {
 314		*width = IB_WIDTH_12X;
 315		*speed = IB_SPEED_SDR;
 316	} else {
 317		*width = IB_WIDTH_1X;
 318		*speed = IB_SPEED_EDR;
 319	}
 320}
 321
 322int efa_query_port(struct ib_device *ibdev, u32 port,
 323		   struct ib_port_attr *props)
 324{
 325	struct efa_dev *dev = to_edev(ibdev);
 326	enum ib_port_speed link_speed;
 327	enum ib_port_width link_width;
 328	u16 link_gbps;
 329
 330	props->lmc = 1;
 331
 332	props->state = IB_PORT_ACTIVE;
 333	props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 334	props->gid_tbl_len = 1;
 335	props->pkey_tbl_len = 1;
 336	link_gbps = dev->dev_attr.max_link_speed_gbps ?: EFA_DEFAULT_LINK_SPEED_GBPS;
 337	efa_link_gbps_to_speed_and_width(link_gbps, &link_speed, &link_width);
 338	props->active_speed = link_speed;
 339	props->active_width = link_width;
 340	props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
 341	props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
 342	props->max_msg_sz = dev->dev_attr.mtu;
 343	props->max_vl_num = 1;
 344
 345	return 0;
 346}
 347
 348int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 349		 int qp_attr_mask,
 350		 struct ib_qp_init_attr *qp_init_attr)
 351{
 352	struct efa_dev *dev = to_edev(ibqp->device);
 353	struct efa_com_query_qp_params params = {};
 354	struct efa_com_query_qp_result result;
 355	struct efa_qp *qp = to_eqp(ibqp);
 356	int err;
 357
 358#define EFA_QUERY_QP_SUPP_MASK \
 359	(IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
 360	 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP | IB_QP_RNR_RETRY)
 361
 362	if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
 363		ibdev_dbg(&dev->ibdev,
 364			  "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 365			  qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
 366		return -EOPNOTSUPP;
 367	}
 368
 369	memset(qp_attr, 0, sizeof(*qp_attr));
 370	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
 371
 372	params.qp_handle = qp->qp_handle;
 373	err = efa_com_query_qp(&dev->edev, &params, &result);
 374	if (err)
 375		return err;
 376
 377	qp_attr->qp_state = result.qp_state;
 378	qp_attr->qkey = result.qkey;
 379	qp_attr->sq_psn = result.sq_psn;
 380	qp_attr->sq_draining = result.sq_draining;
 381	qp_attr->port_num = 1;
 382	qp_attr->rnr_retry = result.rnr_retry;
 383
 384	qp_attr->cap.max_send_wr = qp->max_send_wr;
 385	qp_attr->cap.max_recv_wr = qp->max_recv_wr;
 386	qp_attr->cap.max_send_sge = qp->max_send_sge;
 387	qp_attr->cap.max_recv_sge = qp->max_recv_sge;
 388	qp_attr->cap.max_inline_data = qp->max_inline_data;
 389
 390	qp_init_attr->qp_type = ibqp->qp_type;
 391	qp_init_attr->recv_cq = ibqp->recv_cq;
 392	qp_init_attr->send_cq = ibqp->send_cq;
 393	qp_init_attr->qp_context = ibqp->qp_context;
 394	qp_init_attr->cap = qp_attr->cap;
 395
 396	return 0;
 397}
 398
 399int efa_query_gid(struct ib_device *ibdev, u32 port, int index,
 400		  union ib_gid *gid)
 401{
 402	struct efa_dev *dev = to_edev(ibdev);
 403
 404	memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
 405
 406	return 0;
 407}
 408
 409int efa_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
 410		   u16 *pkey)
 411{
 412	if (index > 0)
 413		return -EINVAL;
 414
 415	*pkey = 0xffff;
 416	return 0;
 417}
 418
 419static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
 420{
 421	struct efa_com_dealloc_pd_params params = {
 422		.pdn = pdn,
 423	};
 424
 425	return efa_com_dealloc_pd(&dev->edev, &params);
 426}
 427
 428int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 429{
 430	struct efa_dev *dev = to_edev(ibpd->device);
 431	struct efa_ibv_alloc_pd_resp resp = {};
 432	struct efa_com_alloc_pd_result result;
 433	struct efa_pd *pd = to_epd(ibpd);
 434	int err;
 435
 436	if (udata->inlen &&
 437	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 438		ibdev_dbg(&dev->ibdev,
 439			  "Incompatible ABI params, udata not cleared\n");
 440		err = -EINVAL;
 441		goto err_out;
 442	}
 443
 444	err = efa_com_alloc_pd(&dev->edev, &result);
 445	if (err)
 446		goto err_out;
 447
 448	pd->pdn = result.pdn;
 449	resp.pdn = result.pdn;
 450
 451	if (udata->outlen) {
 452		err = ib_copy_to_udata(udata, &resp,
 453				       min(sizeof(resp), udata->outlen));
 454		if (err) {
 455			ibdev_dbg(&dev->ibdev,
 456				  "Failed to copy udata for alloc_pd\n");
 457			goto err_dealloc_pd;
 458		}
 459	}
 460
 461	ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
 462
 463	return 0;
 464
 465err_dealloc_pd:
 466	efa_pd_dealloc(dev, result.pdn);
 467err_out:
 468	atomic64_inc(&dev->stats.alloc_pd_err);
 469	return err;
 470}
 471
 472int efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 473{
 474	struct efa_dev *dev = to_edev(ibpd->device);
 475	struct efa_pd *pd = to_epd(ibpd);
 476
 477	ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
 478	efa_pd_dealloc(dev, pd->pdn);
 479	return 0;
 480}
 481
 482static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
 483{
 484	struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
 485
 486	return efa_com_destroy_qp(&dev->edev, &params);
 487}
 488
 489static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp)
 490{
 491	rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
 492	rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
 493	rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
 494	rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
 495}
 496
 497int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 498{
 499	struct efa_dev *dev = to_edev(ibqp->pd->device);
 500	struct efa_qp *qp = to_eqp(ibqp);
 501	int err;
 502
 503	ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
 504
 505	err = efa_destroy_qp_handle(dev, qp->qp_handle);
 506	if (err)
 507		return err;
 508
 509	efa_qp_user_mmap_entries_remove(qp);
 510
 511	if (qp->rq_cpu_addr) {
 512		ibdev_dbg(&dev->ibdev,
 513			  "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
 514			  qp->rq_cpu_addr, qp->rq_size,
 515			  &qp->rq_dma_addr);
 516		efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
 517				qp->rq_size, DMA_TO_DEVICE);
 518	}
 519
 520	return 0;
 521}
 522
 523static struct rdma_user_mmap_entry*
 524efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
 525			   u64 address, size_t length,
 526			   u8 mmap_flag, u64 *offset)
 527{
 528	struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 529	int err;
 530
 531	if (!entry)
 532		return NULL;
 533
 534	entry->address = address;
 535	entry->mmap_flag = mmap_flag;
 536
 537	err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
 538					  length);
 539	if (err) {
 540		kfree(entry);
 541		return NULL;
 542	}
 543	*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 544
 545	return &entry->rdma_entry;
 546}
 547
 548static int qp_mmap_entries_setup(struct efa_qp *qp,
 549				 struct efa_dev *dev,
 550				 struct efa_ucontext *ucontext,
 551				 struct efa_com_create_qp_params *params,
 552				 struct efa_ibv_create_qp_resp *resp)
 553{
 554	size_t length;
 555	u64 address;
 556
 557	address = dev->db_bar_addr + resp->sq_db_offset;
 558	qp->sq_db_mmap_entry =
 559		efa_user_mmap_entry_insert(&ucontext->ibucontext,
 560					   address,
 561					   PAGE_SIZE, EFA_MMAP_IO_NC,
 562					   &resp->sq_db_mmap_key);
 563	if (!qp->sq_db_mmap_entry)
 564		return -ENOMEM;
 565
 566	resp->sq_db_offset &= ~PAGE_MASK;
 567
 568	address = dev->mem_bar_addr + resp->llq_desc_offset;
 569	length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
 570			    offset_in_page(resp->llq_desc_offset));
 571
 572	qp->llq_desc_mmap_entry =
 573		efa_user_mmap_entry_insert(&ucontext->ibucontext,
 574					   address, length,
 575					   EFA_MMAP_IO_WC,
 576					   &resp->llq_desc_mmap_key);
 577	if (!qp->llq_desc_mmap_entry)
 578		goto err_remove_mmap;
 579
 580	resp->llq_desc_offset &= ~PAGE_MASK;
 581
 582	if (qp->rq_size) {
 583		address = dev->db_bar_addr + resp->rq_db_offset;
 584
 585		qp->rq_db_mmap_entry =
 586			efa_user_mmap_entry_insert(&ucontext->ibucontext,
 587						   address, PAGE_SIZE,
 588						   EFA_MMAP_IO_NC,
 589						   &resp->rq_db_mmap_key);
 590		if (!qp->rq_db_mmap_entry)
 591			goto err_remove_mmap;
 592
 593		resp->rq_db_offset &= ~PAGE_MASK;
 594
 595		address = virt_to_phys(qp->rq_cpu_addr);
 596		qp->rq_mmap_entry =
 597			efa_user_mmap_entry_insert(&ucontext->ibucontext,
 598						   address, qp->rq_size,
 599						   EFA_MMAP_DMA_PAGE,
 600						   &resp->rq_mmap_key);
 601		if (!qp->rq_mmap_entry)
 602			goto err_remove_mmap;
 603
 604		resp->rq_mmap_size = qp->rq_size;
 605	}
 606
 607	return 0;
 608
 609err_remove_mmap:
 610	efa_qp_user_mmap_entries_remove(qp);
 611
 612	return -ENOMEM;
 613}
 614
 615static int efa_qp_validate_cap(struct efa_dev *dev,
 616			       struct ib_qp_init_attr *init_attr)
 617{
 618	if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
 619		ibdev_dbg(&dev->ibdev,
 620			  "qp: requested send wr[%u] exceeds the max[%u]\n",
 621			  init_attr->cap.max_send_wr,
 622			  dev->dev_attr.max_sq_depth);
 623		return -EINVAL;
 624	}
 625	if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
 626		ibdev_dbg(&dev->ibdev,
 627			  "qp: requested receive wr[%u] exceeds the max[%u]\n",
 628			  init_attr->cap.max_recv_wr,
 629			  dev->dev_attr.max_rq_depth);
 630		return -EINVAL;
 631	}
 632	if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
 633		ibdev_dbg(&dev->ibdev,
 634			  "qp: requested sge send[%u] exceeds the max[%u]\n",
 635			  init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
 636		return -EINVAL;
 637	}
 638	if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
 639		ibdev_dbg(&dev->ibdev,
 640			  "qp: requested sge recv[%u] exceeds the max[%u]\n",
 641			  init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
 642		return -EINVAL;
 643	}
 644	if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
 645		ibdev_dbg(&dev->ibdev,
 646			  "qp: requested inline data[%u] exceeds the max[%u]\n",
 647			  init_attr->cap.max_inline_data,
 648			  dev->dev_attr.inline_buf_size);
 649		return -EINVAL;
 650	}
 651
 652	return 0;
 653}
 654
 655static int efa_qp_validate_attr(struct efa_dev *dev,
 656				struct ib_qp_init_attr *init_attr)
 657{
 658	if (init_attr->qp_type != IB_QPT_DRIVER &&
 659	    init_attr->qp_type != IB_QPT_UD) {
 660		ibdev_dbg(&dev->ibdev,
 661			  "Unsupported qp type %d\n", init_attr->qp_type);
 662		return -EOPNOTSUPP;
 663	}
 664
 665	if (init_attr->srq) {
 666		ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
 667		return -EOPNOTSUPP;
 668	}
 669
 670	if (init_attr->create_flags) {
 671		ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
 672		return -EOPNOTSUPP;
 673	}
 674
 675	return 0;
 676}
 677
 678int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
 679		  struct ib_udata *udata)
 680{
 681	struct efa_com_create_qp_params create_qp_params = {};
 682	struct efa_com_create_qp_result create_qp_resp;
 683	struct efa_dev *dev = to_edev(ibqp->device);
 684	struct efa_ibv_create_qp_resp resp = {};
 685	struct efa_ibv_create_qp cmd = {};
 686	struct efa_qp *qp = to_eqp(ibqp);
 687	struct efa_ucontext *ucontext;
 688	u16 supported_efa_flags = 0;
 689	int err;
 690
 691	ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
 692					     ibucontext);
 693
 694	err = efa_qp_validate_cap(dev, init_attr);
 695	if (err)
 696		goto err_out;
 697
 698	err = efa_qp_validate_attr(dev, init_attr);
 699	if (err)
 700		goto err_out;
 701
 702	if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) {
 703		ibdev_dbg(&dev->ibdev,
 704			  "Incompatible ABI params, no input udata\n");
 705		err = -EINVAL;
 706		goto err_out;
 707	}
 708
 709	if (udata->inlen > sizeof(cmd) &&
 710	    !ib_is_udata_cleared(udata, sizeof(cmd),
 711				 udata->inlen - sizeof(cmd))) {
 712		ibdev_dbg(&dev->ibdev,
 713			  "Incompatible ABI params, unknown fields in udata\n");
 714		err = -EINVAL;
 715		goto err_out;
 716	}
 717
 718	err = ib_copy_from_udata(&cmd, udata,
 719				 min(sizeof(cmd), udata->inlen));
 720	if (err) {
 721		ibdev_dbg(&dev->ibdev,
 722			  "Cannot copy udata for create_qp\n");
 723		goto err_out;
 724	}
 725
 726	if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_98)) {
 727		ibdev_dbg(&dev->ibdev,
 728			  "Incompatible ABI params, unknown fields in udata\n");
 729		err = -EINVAL;
 730		goto err_out;
 731	}
 732
 733	if (EFA_DEV_CAP(dev, UNSOLICITED_WRITE_RECV))
 734		supported_efa_flags |= EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV;
 735
 736	if (cmd.flags & ~supported_efa_flags) {
 737		ibdev_dbg(&dev->ibdev, "Unsupported EFA QP create flags[%#x], supported[%#x]\n",
 738			  cmd.flags, supported_efa_flags);
 739		err = -EOPNOTSUPP;
 740		goto err_out;
 741	}
 742
 743	create_qp_params.uarn = ucontext->uarn;
 744	create_qp_params.pd = to_epd(ibqp->pd)->pdn;
 745
 746	if (init_attr->qp_type == IB_QPT_UD) {
 747		create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
 748	} else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
 749		create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
 750	} else {
 751		ibdev_dbg(&dev->ibdev,
 752			  "Unsupported qp type %d driver qp type %d\n",
 753			  init_attr->qp_type, cmd.driver_qp_type);
 754		err = -EOPNOTSUPP;
 755		goto err_out;
 756	}
 757
 758	ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
 759		  init_attr->qp_type, cmd.driver_qp_type);
 760	create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
 761	create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
 762	create_qp_params.sq_depth = init_attr->cap.max_send_wr;
 763	create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
 764
 765	create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
 766	create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
 767	qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
 768	if (qp->rq_size) {
 769		qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
 770						    qp->rq_size, DMA_TO_DEVICE);
 771		if (!qp->rq_cpu_addr) {
 772			err = -ENOMEM;
 773			goto err_out;
 774		}
 775
 776		ibdev_dbg(&dev->ibdev,
 777			  "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
 778			  qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
 779		create_qp_params.rq_base_addr = qp->rq_dma_addr;
 780	}
 781
 782	create_qp_params.sl = cmd.sl;
 783
 784	if (cmd.flags & EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV)
 785		create_qp_params.unsolicited_write_recv = true;
 786
 787	err = efa_com_create_qp(&dev->edev, &create_qp_params,
 788				&create_qp_resp);
 789	if (err)
 790		goto err_free_mapped;
 791
 792	resp.sq_db_offset = create_qp_resp.sq_db_offset;
 793	resp.rq_db_offset = create_qp_resp.rq_db_offset;
 794	resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
 795	resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
 796	resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
 797
 798	err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
 799				    &resp);
 800	if (err)
 801		goto err_destroy_qp;
 802
 803	qp->qp_handle = create_qp_resp.qp_handle;
 804	qp->ibqp.qp_num = create_qp_resp.qp_num;
 805	qp->max_send_wr = init_attr->cap.max_send_wr;
 806	qp->max_recv_wr = init_attr->cap.max_recv_wr;
 807	qp->max_send_sge = init_attr->cap.max_send_sge;
 808	qp->max_recv_sge = init_attr->cap.max_recv_sge;
 809	qp->max_inline_data = init_attr->cap.max_inline_data;
 810
 811	if (udata->outlen) {
 812		err = ib_copy_to_udata(udata, &resp,
 813				       min(sizeof(resp), udata->outlen));
 814		if (err) {
 815			ibdev_dbg(&dev->ibdev,
 816				  "Failed to copy udata for qp[%u]\n",
 817				  create_qp_resp.qp_num);
 818			goto err_remove_mmap_entries;
 819		}
 820	}
 821
 822	ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
 823
 824	return 0;
 825
 826err_remove_mmap_entries:
 827	efa_qp_user_mmap_entries_remove(qp);
 828err_destroy_qp:
 829	efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
 830err_free_mapped:
 831	if (qp->rq_size)
 832		efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
 833				qp->rq_size, DMA_TO_DEVICE);
 834err_out:
 835	atomic64_inc(&dev->stats.create_qp_err);
 836	return err;
 837}
 838
 839static const struct {
 840	int			valid;
 841	enum ib_qp_attr_mask	req_param;
 842	enum ib_qp_attr_mask	opt_param;
 843} srd_qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
 844	[IB_QPS_RESET] = {
 845		[IB_QPS_RESET] = { .valid = 1 },
 846		[IB_QPS_INIT]  = {
 847			.valid = 1,
 848			.req_param = IB_QP_PKEY_INDEX |
 849				     IB_QP_PORT |
 850				     IB_QP_QKEY,
 851		},
 852	},
 853	[IB_QPS_INIT] = {
 854		[IB_QPS_RESET] = { .valid = 1 },
 855		[IB_QPS_ERR]   = { .valid = 1 },
 856		[IB_QPS_INIT]  = {
 857			.valid = 1,
 858			.opt_param = IB_QP_PKEY_INDEX |
 859				     IB_QP_PORT |
 860				     IB_QP_QKEY,
 861		},
 862		[IB_QPS_RTR]   = {
 863			.valid = 1,
 864			.opt_param = IB_QP_PKEY_INDEX |
 865				     IB_QP_QKEY,
 866		},
 867	},
 868	[IB_QPS_RTR] = {
 869		[IB_QPS_RESET] = { .valid = 1 },
 870		[IB_QPS_ERR]   = { .valid = 1 },
 871		[IB_QPS_RTS]   = {
 872			.valid = 1,
 873			.req_param = IB_QP_SQ_PSN,
 874			.opt_param = IB_QP_CUR_STATE |
 875				     IB_QP_QKEY |
 876				     IB_QP_RNR_RETRY,
 877
 878		}
 879	},
 880	[IB_QPS_RTS] = {
 881		[IB_QPS_RESET] = { .valid = 1 },
 882		[IB_QPS_ERR]   = { .valid = 1 },
 883		[IB_QPS_RTS]   = {
 884			.valid = 1,
 885			.opt_param = IB_QP_CUR_STATE |
 886				     IB_QP_QKEY,
 887		},
 888		[IB_QPS_SQD] = {
 889			.valid = 1,
 890			.opt_param = IB_QP_EN_SQD_ASYNC_NOTIFY,
 891		},
 892	},
 893	[IB_QPS_SQD] = {
 894		[IB_QPS_RESET] = { .valid = 1 },
 895		[IB_QPS_ERR]   = { .valid = 1 },
 896		[IB_QPS_RTS]   = {
 897			.valid = 1,
 898			.opt_param = IB_QP_CUR_STATE |
 899				     IB_QP_QKEY,
 900		},
 901		[IB_QPS_SQD] = {
 902			.valid = 1,
 903			.opt_param = IB_QP_PKEY_INDEX |
 904				     IB_QP_QKEY,
 905		}
 906	},
 907	[IB_QPS_SQE] = {
 908		[IB_QPS_RESET] = { .valid = 1 },
 909		[IB_QPS_ERR]   = { .valid = 1 },
 910		[IB_QPS_RTS]   = {
 911			.valid = 1,
 912			.opt_param = IB_QP_CUR_STATE |
 913				     IB_QP_QKEY,
 914		}
 915	},
 916	[IB_QPS_ERR] = {
 917		[IB_QPS_RESET] = { .valid = 1 },
 918		[IB_QPS_ERR]   = { .valid = 1 },
 919	}
 920};
 921
 922static bool efa_modify_srd_qp_is_ok(enum ib_qp_state cur_state,
 923				    enum ib_qp_state next_state,
 924				    enum ib_qp_attr_mask mask)
 925{
 926	enum ib_qp_attr_mask req_param, opt_param;
 927
 928	if (mask & IB_QP_CUR_STATE  &&
 929	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
 930	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
 931		return false;
 932
 933	if (!srd_qp_state_table[cur_state][next_state].valid)
 934		return false;
 935
 936	req_param = srd_qp_state_table[cur_state][next_state].req_param;
 937	opt_param = srd_qp_state_table[cur_state][next_state].opt_param;
 938
 939	if ((mask & req_param) != req_param)
 940		return false;
 941
 942	if (mask & ~(req_param | opt_param | IB_QP_STATE))
 943		return false;
 944
 945	return true;
 946}
 947
 948static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
 949				  struct ib_qp_attr *qp_attr, int qp_attr_mask,
 950				  enum ib_qp_state cur_state,
 951				  enum ib_qp_state new_state)
 952{
 953	int err;
 954
 955#define EFA_MODIFY_QP_SUPP_MASK \
 956	(IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
 957	 IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN | \
 958	 IB_QP_RNR_RETRY)
 959
 960	if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
 961		ibdev_dbg(&dev->ibdev,
 962			  "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 963			  qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
 964		return -EOPNOTSUPP;
 965	}
 966
 967	if (qp->ibqp.qp_type == IB_QPT_DRIVER)
 968		err = !efa_modify_srd_qp_is_ok(cur_state, new_state,
 969					       qp_attr_mask);
 970	else
 971		err = !ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
 972					  qp_attr_mask);
 973
 974	if (err) {
 975		ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
 976		return -EINVAL;
 977	}
 978
 979	if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
 980		ibdev_dbg(&dev->ibdev, "Can't change port num\n");
 981		return -EOPNOTSUPP;
 982	}
 983
 984	if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
 985		ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
 986		return -EOPNOTSUPP;
 987	}
 988
 989	return 0;
 990}
 991
 992int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 993		  int qp_attr_mask, struct ib_udata *udata)
 994{
 995	struct efa_dev *dev = to_edev(ibqp->device);
 996	struct efa_com_modify_qp_params params = {};
 997	struct efa_qp *qp = to_eqp(ibqp);
 998	enum ib_qp_state cur_state;
 999	enum ib_qp_state new_state;
1000	int err;
1001
1002	if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1003		return -EOPNOTSUPP;
1004
1005	if (udata->inlen &&
1006	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1007		ibdev_dbg(&dev->ibdev,
1008			  "Incompatible ABI params, udata not cleared\n");
1009		return -EINVAL;
1010	}
1011
1012	cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
1013						     qp->state;
1014	new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
1015
1016	err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
1017				     new_state);
1018	if (err)
1019		return err;
1020
1021	params.qp_handle = qp->qp_handle;
1022
1023	if (qp_attr_mask & IB_QP_STATE) {
1024		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QP_STATE,
1025			1);
1026		EFA_SET(&params.modify_mask,
1027			EFA_ADMIN_MODIFY_QP_CMD_CUR_QP_STATE, 1);
1028		params.cur_qp_state = cur_state;
1029		params.qp_state = new_state;
1030	}
1031
1032	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1033		EFA_SET(&params.modify_mask,
1034			EFA_ADMIN_MODIFY_QP_CMD_SQ_DRAINED_ASYNC_NOTIFY, 1);
1035		params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
1036	}
1037
1038	if (qp_attr_mask & IB_QP_QKEY) {
1039		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QKEY, 1);
1040		params.qkey = qp_attr->qkey;
1041	}
1042
1043	if (qp_attr_mask & IB_QP_SQ_PSN) {
1044		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_SQ_PSN, 1);
1045		params.sq_psn = qp_attr->sq_psn;
1046	}
1047
1048	if (qp_attr_mask & IB_QP_RNR_RETRY) {
1049		EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_RNR_RETRY,
1050			1);
1051		params.rnr_retry = qp_attr->rnr_retry;
1052	}
1053
1054	err = efa_com_modify_qp(&dev->edev, &params);
1055	if (err)
1056		return err;
1057
1058	qp->state = new_state;
1059
1060	return 0;
1061}
1062
1063static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
1064{
1065	struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
1066
1067	return efa_com_destroy_cq(&dev->edev, &params);
1068}
1069
1070static void efa_cq_user_mmap_entries_remove(struct efa_cq *cq)
1071{
1072	rdma_user_mmap_entry_remove(cq->db_mmap_entry);
1073	rdma_user_mmap_entry_remove(cq->mmap_entry);
1074}
1075
1076int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1077{
1078	struct efa_dev *dev = to_edev(ibcq->device);
1079	struct efa_cq *cq = to_ecq(ibcq);
1080
1081	ibdev_dbg(&dev->ibdev,
1082		  "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
1083		  cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
1084
1085	efa_destroy_cq_idx(dev, cq->cq_idx);
1086	efa_cq_user_mmap_entries_remove(cq);
1087	if (cq->eq) {
1088		xa_erase(&dev->cqs_xa, cq->cq_idx);
1089		synchronize_irq(cq->eq->irq.irqn);
1090	}
1091
1092	if (cq->umem)
1093		ib_umem_release(cq->umem);
1094	else
1095		efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size, DMA_FROM_DEVICE);
1096	return 0;
1097}
1098
1099static struct efa_eq *efa_vec2eq(struct efa_dev *dev, int vec)
1100{
1101	return &dev->eqs[vec];
1102}
1103
1104static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
1105				 struct efa_ibv_create_cq_resp *resp,
1106				 bool db_valid)
1107{
1108	resp->q_mmap_size = cq->size;
1109	cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1110						    virt_to_phys(cq->cpu_addr),
1111						    cq->size, EFA_MMAP_DMA_PAGE,
1112						    &resp->q_mmap_key);
1113	if (!cq->mmap_entry)
1114		return -ENOMEM;
1115
1116	if (db_valid) {
1117		cq->db_mmap_entry =
1118			efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1119						   dev->db_bar_addr + resp->db_off,
1120						   PAGE_SIZE, EFA_MMAP_IO_NC,
1121						   &resp->db_mmap_key);
1122		if (!cq->db_mmap_entry) {
1123			rdma_user_mmap_entry_remove(cq->mmap_entry);
1124			return -ENOMEM;
1125		}
1126
1127		resp->db_off &= ~PAGE_MASK;
1128		resp->comp_mask |= EFA_CREATE_CQ_RESP_DB_OFF;
1129	}
1130
1131	return 0;
1132}
1133
1134int efa_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1135		       struct ib_umem *umem, struct uverbs_attr_bundle *attrs)
1136{
1137	struct ib_udata *udata = &attrs->driver_udata;
1138	struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
1139		udata, struct efa_ucontext, ibucontext);
1140	struct efa_com_create_cq_params params = {};
1141	struct efa_ibv_create_cq_resp resp = {};
1142	struct efa_com_create_cq_result result;
1143	struct ib_device *ibdev = ibcq->device;
1144	struct efa_dev *dev = to_edev(ibdev);
1145	struct efa_ibv_create_cq cmd = {};
1146	struct efa_cq *cq = to_ecq(ibcq);
1147	int entries = attr->cqe;
1148	bool set_src_addr;
1149	int err;
1150
1151	ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
1152
1153	if (attr->flags)
1154		return -EOPNOTSUPP;
1155
1156	if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
1157		ibdev_dbg(ibdev,
1158			  "cq: requested entries[%u] non-positive or greater than max[%u]\n",
1159			  entries, dev->dev_attr.max_cq_depth);
1160		err = -EINVAL;
1161		goto err_out;
1162	}
1163
1164	if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) {
1165		ibdev_dbg(ibdev,
1166			  "Incompatible ABI params, no input udata\n");
1167		err = -EINVAL;
1168		goto err_out;
1169	}
1170
1171	if (udata->inlen > sizeof(cmd) &&
1172	    !ib_is_udata_cleared(udata, sizeof(cmd),
1173				 udata->inlen - sizeof(cmd))) {
1174		ibdev_dbg(ibdev,
1175			  "Incompatible ABI params, unknown fields in udata\n");
1176		err = -EINVAL;
1177		goto err_out;
1178	}
1179
1180	err = ib_copy_from_udata(&cmd, udata,
1181				 min(sizeof(cmd), udata->inlen));
1182	if (err) {
1183		ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
1184		goto err_out;
1185	}
1186
1187	if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_58)) {
1188		ibdev_dbg(ibdev,
1189			  "Incompatible ABI params, unknown fields in udata\n");
1190		err = -EINVAL;
1191		goto err_out;
1192	}
1193
1194	set_src_addr = !!(cmd.flags & EFA_CREATE_CQ_WITH_SGID);
1195	if ((cmd.cq_entry_size != sizeof(struct efa_io_rx_cdesc_ex)) &&
1196	    (set_src_addr ||
1197	     cmd.cq_entry_size != sizeof(struct efa_io_rx_cdesc))) {
1198		ibdev_dbg(ibdev,
1199			  "Invalid entry size [%u]\n", cmd.cq_entry_size);
1200		err = -EINVAL;
1201		goto err_out;
1202	}
1203
1204	if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
1205		ibdev_dbg(ibdev,
1206			  "Invalid number of sub cqs[%u] expected[%u]\n",
1207			  cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
1208		err = -EINVAL;
1209		goto err_out;
1210	}
1211
1212	cq->ucontext = ucontext;
1213	cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
1214
1215	if (umem) {
1216		if (umem->length < cq->size) {
1217			ibdev_dbg(&dev->ibdev, "External memory too small\n");
1218			err = -EINVAL;
1219			goto err_out;
1220		}
1221
1222		if (!ib_umem_is_contiguous(umem)) {
1223			ibdev_dbg(&dev->ibdev, "Non contiguous CQ unsupported\n");
1224			err = -EINVAL;
1225			goto err_out;
1226		}
1227
1228		cq->cpu_addr = NULL;
1229		cq->dma_addr = ib_umem_start_dma_addr(umem);
1230		cq->umem = umem;
1231	} else {
1232		cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
1233						 DMA_FROM_DEVICE);
1234		if (!cq->cpu_addr) {
1235			err = -ENOMEM;
1236			goto err_out;
1237		}
1238	}
1239
1240	params.uarn = cq->ucontext->uarn;
1241	params.sub_cq_depth = entries;
1242	params.dma_addr = cq->dma_addr;
1243	params.entry_size_in_bytes = cmd.cq_entry_size;
1244	params.num_sub_cqs = cmd.num_sub_cqs;
1245	params.set_src_addr = set_src_addr;
1246	if (cmd.flags & EFA_CREATE_CQ_WITH_COMPLETION_CHANNEL) {
1247		cq->eq = efa_vec2eq(dev, attr->comp_vector);
1248		params.eqn = cq->eq->eeq.eqn;
1249		params.interrupt_mode_enabled = true;
1250	}
1251
1252	err = efa_com_create_cq(&dev->edev, &params, &result);
1253	if (err)
1254		goto err_free_mapped;
1255
1256	resp.db_off = result.db_off;
1257	resp.cq_idx = result.cq_idx;
1258	cq->cq_idx = result.cq_idx;
1259	cq->ibcq.cqe = result.actual_depth;
1260	WARN_ON_ONCE(entries != result.actual_depth);
1261
1262	if (!umem)
1263		err = cq_mmap_entries_setup(dev, cq, &resp, result.db_valid);
1264
1265	if (err) {
1266		ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
1267			  cq->cq_idx);
1268		goto err_destroy_cq;
1269	}
1270
1271	if (cq->eq) {
1272		err = xa_err(xa_store(&dev->cqs_xa, cq->cq_idx, cq, GFP_KERNEL));
1273		if (err) {
1274			ibdev_dbg(ibdev, "Failed to store cq[%u] in xarray\n",
1275				  cq->cq_idx);
1276			goto err_remove_mmap;
1277		}
1278	}
1279
1280	if (udata->outlen) {
1281		err = ib_copy_to_udata(udata, &resp,
1282				       min(sizeof(resp), udata->outlen));
1283		if (err) {
1284			ibdev_dbg(ibdev,
1285				  "Failed to copy udata for create_cq\n");
1286			goto err_xa_erase;
1287		}
1288	}
1289
1290	ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
1291		  cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
1292
1293	return 0;
1294
1295err_xa_erase:
1296	if (cq->eq)
1297		xa_erase(&dev->cqs_xa, cq->cq_idx);
1298err_remove_mmap:
1299	efa_cq_user_mmap_entries_remove(cq);
1300err_destroy_cq:
1301	efa_destroy_cq_idx(dev, cq->cq_idx);
1302err_free_mapped:
1303	if (!umem)
1304		efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1305				DMA_FROM_DEVICE);
1306err_out:
1307	atomic64_inc(&dev->stats.create_cq_err);
1308	return err;
1309}
1310
1311int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1312		  struct uverbs_attr_bundle *attrs)
1313{
1314	return efa_create_cq_umem(ibcq, attr, NULL, attrs);
1315}
1316
1317static int umem_to_page_list(struct efa_dev *dev,
1318			     struct ib_umem *umem,
1319			     u64 *page_list,
1320			     u32 hp_cnt,
1321			     u8 hp_shift)
1322{
1323	struct ib_block_iter biter;
1324	unsigned int hp_idx = 0;
1325
1326	rdma_umem_for_each_dma_block(umem, &biter, BIT(hp_shift))
1327		page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1328
1329	return 0;
1330}
1331
1332static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1333{
1334	struct scatterlist *sglist;
1335	struct page *pg;
1336	int i;
1337
1338	sglist = kmalloc_array(page_cnt, sizeof(*sglist), GFP_KERNEL);
1339	if (!sglist)
1340		return NULL;
1341	sg_init_table(sglist, page_cnt);
1342	for (i = 0; i < page_cnt; i++) {
1343		pg = vmalloc_to_page(buf);
1344		if (!pg)
1345			goto err;
1346		sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1347		buf += PAGE_SIZE / sizeof(*buf);
1348	}
1349	return sglist;
1350
1351err:
1352	kfree(sglist);
1353	return NULL;
1354}
1355
1356/*
1357 * create a chunk list of physical pages dma addresses from the supplied
1358 * scatter gather list
1359 */
1360static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1361{
1362	struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1363	int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1364	struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1365	unsigned int chunk_list_size, chunk_idx, payload_idx;
1366	int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1367	struct efa_com_ctrl_buff_info *ctrl_buf;
1368	u64 *cur_chunk_buf, *prev_chunk_buf;
1369	struct ib_block_iter biter;
1370	dma_addr_t dma_addr;
1371	int i;
1372
1373	/* allocate a chunk list that consists of 4KB chunks */
1374	chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1375
1376	chunk_list->size = chunk_list_size;
1377	chunk_list->chunks = kcalloc(chunk_list_size,
1378				     sizeof(*chunk_list->chunks),
1379				     GFP_KERNEL);
1380	if (!chunk_list->chunks)
1381		return -ENOMEM;
1382
1383	ibdev_dbg(&dev->ibdev,
1384		  "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1385		  page_cnt);
1386
1387	/* allocate chunk buffers: */
1388	for (i = 0; i < chunk_list_size; i++) {
1389		chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1390		if (!chunk_list->chunks[i].buf)
1391			goto chunk_list_dealloc;
1392
1393		chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1394	}
1395	chunk_list->chunks[chunk_list_size - 1].length =
1396		((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1397			EFA_CHUNK_PTR_SIZE;
1398
1399	/* fill the dma addresses of sg list pages to chunks: */
1400	chunk_idx = 0;
1401	payload_idx = 0;
1402	cur_chunk_buf = chunk_list->chunks[0].buf;
1403	rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1404			    EFA_CHUNK_PAYLOAD_SIZE) {
1405		cur_chunk_buf[payload_idx++] =
1406			rdma_block_iter_dma_address(&biter);
1407
1408		if (payload_idx == EFA_PTRS_PER_CHUNK) {
1409			chunk_idx++;
1410			cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1411			payload_idx = 0;
1412		}
1413	}
1414
1415	/* map chunks to dma and fill chunks next ptrs */
1416	for (i = chunk_list_size - 1; i >= 0; i--) {
1417		dma_addr = dma_map_single(&dev->pdev->dev,
1418					  chunk_list->chunks[i].buf,
1419					  chunk_list->chunks[i].length,
1420					  DMA_TO_DEVICE);
1421		if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1422			ibdev_err(&dev->ibdev,
1423				  "chunk[%u] dma_map_failed\n", i);
1424			goto chunk_list_unmap;
1425		}
1426
1427		chunk_list->chunks[i].dma_addr = dma_addr;
1428		ibdev_dbg(&dev->ibdev,
1429			  "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1430
1431		if (!i)
1432			break;
1433
1434		prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1435
1436		ctrl_buf = (struct efa_com_ctrl_buff_info *)
1437				&prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1438		ctrl_buf->length = chunk_list->chunks[i].length;
1439
1440		efa_com_set_dma_addr(dma_addr,
1441				     &ctrl_buf->address.mem_addr_high,
1442				     &ctrl_buf->address.mem_addr_low);
1443	}
1444
1445	return 0;
1446
1447chunk_list_unmap:
1448	for (; i < chunk_list_size; i++) {
1449		dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1450				 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1451	}
1452chunk_list_dealloc:
1453	for (i = 0; i < chunk_list_size; i++)
1454		kfree(chunk_list->chunks[i].buf);
1455
1456	kfree(chunk_list->chunks);
1457	return -ENOMEM;
1458}
1459
1460static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1461{
1462	struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1463	int i;
1464
1465	for (i = 0; i < chunk_list->size; i++) {
1466		dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1467				 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1468		kfree(chunk_list->chunks[i].buf);
1469	}
1470
1471	kfree(chunk_list->chunks);
1472}
1473
1474/* initialize pbl continuous mode: map pbl buffer to a dma address. */
1475static int pbl_continuous_initialize(struct efa_dev *dev,
1476				     struct pbl_context *pbl)
1477{
1478	dma_addr_t dma_addr;
1479
1480	dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1481				  pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1482	if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1483		ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1484		return -ENOMEM;
1485	}
1486
1487	pbl->phys.continuous.dma_addr = dma_addr;
1488	ibdev_dbg(&dev->ibdev,
1489		  "pbl continuous - dma_addr = %pad, size[%u]\n",
1490		  &dma_addr, pbl->pbl_buf_size_in_bytes);
1491
1492	return 0;
1493}
1494
1495/*
1496 * initialize pbl indirect mode:
1497 * create a chunk list out of the dma addresses of the physical pages of
1498 * pbl buffer.
1499 */
1500static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1501{
1502	u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, EFA_CHUNK_PAYLOAD_SIZE);
1503	struct scatterlist *sgl;
1504	int sg_dma_cnt, err;
1505
1506	BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1507	sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1508	if (!sgl)
1509		return -ENOMEM;
1510
1511	sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1512	if (!sg_dma_cnt) {
1513		err = -EINVAL;
1514		goto err_map;
1515	}
1516
1517	pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1518	pbl->phys.indirect.sgl = sgl;
1519	pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1520	err = pbl_chunk_list_create(dev, pbl);
1521	if (err) {
1522		ibdev_dbg(&dev->ibdev,
1523			  "chunk_list creation failed[%d]\n", err);
1524		goto err_chunk;
1525	}
1526
1527	ibdev_dbg(&dev->ibdev,
1528		  "pbl indirect - size[%u], chunks[%u]\n",
1529		  pbl->pbl_buf_size_in_bytes,
1530		  pbl->phys.indirect.chunk_list.size);
1531
1532	return 0;
1533
1534err_chunk:
1535	dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1536err_map:
1537	kfree(sgl);
1538	return err;
1539}
1540
1541static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1542{
1543	pbl_chunk_list_destroy(dev, pbl);
1544	dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1545		     pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1546	kfree(pbl->phys.indirect.sgl);
1547}
1548
1549/* create a page buffer list from a mapped user memory region */
1550static int pbl_create(struct efa_dev *dev,
1551		      struct pbl_context *pbl,
1552		      struct ib_umem *umem,
1553		      int hp_cnt,
1554		      u8 hp_shift)
1555{
1556	int err;
1557
1558	pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1559	pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1560	if (!pbl->pbl_buf)
1561		return -ENOMEM;
1562
1563	if (is_vmalloc_addr(pbl->pbl_buf)) {
1564		pbl->physically_continuous = 0;
1565		err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1566					hp_shift);
1567		if (err)
1568			goto err_free;
1569
1570		err = pbl_indirect_initialize(dev, pbl);
1571		if (err)
1572			goto err_free;
1573	} else {
1574		pbl->physically_continuous = 1;
1575		err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1576					hp_shift);
1577		if (err)
1578			goto err_free;
1579
1580		err = pbl_continuous_initialize(dev, pbl);
1581		if (err)
1582			goto err_free;
1583	}
1584
1585	ibdev_dbg(&dev->ibdev,
1586		  "user_pbl_created: user_pages[%u], continuous[%u]\n",
1587		  hp_cnt, pbl->physically_continuous);
1588
1589	return 0;
1590
1591err_free:
1592	kvfree(pbl->pbl_buf);
1593	return err;
1594}
1595
1596static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1597{
1598	if (pbl->physically_continuous)
1599		dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1600				 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1601	else
1602		pbl_indirect_terminate(dev, pbl);
1603
1604	kvfree(pbl->pbl_buf);
1605}
1606
1607static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1608				 struct efa_com_reg_mr_params *params)
1609{
1610	int err;
1611
1612	params->inline_pbl = 1;
1613	err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1614				params->page_num, params->page_shift);
1615	if (err)
1616		return err;
1617
1618	ibdev_dbg(&dev->ibdev,
1619		  "inline_pbl_array - pages[%u]\n", params->page_num);
1620
1621	return 0;
1622}
1623
1624static int efa_create_pbl(struct efa_dev *dev,
1625			  struct pbl_context *pbl,
1626			  struct efa_mr *mr,
1627			  struct efa_com_reg_mr_params *params)
1628{
1629	int err;
1630
1631	err = pbl_create(dev, pbl, mr->umem, params->page_num,
1632			 params->page_shift);
1633	if (err) {
1634		ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1635		return err;
1636	}
1637
1638	params->inline_pbl = 0;
1639	params->indirect = !pbl->physically_continuous;
1640	if (pbl->physically_continuous) {
1641		params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1642
1643		efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1644				     &params->pbl.pbl.address.mem_addr_high,
1645				     &params->pbl.pbl.address.mem_addr_low);
1646	} else {
1647		params->pbl.pbl.length =
1648			pbl->phys.indirect.chunk_list.chunks[0].length;
1649
1650		efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1651				     &params->pbl.pbl.address.mem_addr_high,
1652				     &params->pbl.pbl.address.mem_addr_low);
1653	}
1654
1655	return 0;
1656}
1657
1658static struct efa_mr *efa_alloc_mr(struct ib_pd *ibpd, int access_flags,
1659				   struct ib_udata *udata)
1660{
1661	struct efa_dev *dev = to_edev(ibpd->device);
1662	int supp_access_flags;
1663	struct efa_mr *mr;
1664
1665	if (udata && udata->inlen &&
1666	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1667		ibdev_dbg(&dev->ibdev,
1668			  "Incompatible ABI params, udata not cleared\n");
1669		return ERR_PTR(-EINVAL);
1670	}
1671
1672	supp_access_flags =
1673		IB_ACCESS_LOCAL_WRITE |
1674		(EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0) |
1675		(EFA_DEV_CAP(dev, RDMA_WRITE) ? IB_ACCESS_REMOTE_WRITE : 0);
1676
1677	access_flags &= ~IB_ACCESS_OPTIONAL;
1678	if (access_flags & ~supp_access_flags) {
1679		ibdev_dbg(&dev->ibdev,
1680			  "Unsupported access flags[%#x], supported[%#x]\n",
1681			  access_flags, supp_access_flags);
1682		return ERR_PTR(-EOPNOTSUPP);
1683	}
1684
1685	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1686	if (!mr)
1687		return ERR_PTR(-ENOMEM);
1688
1689	return mr;
1690}
1691
1692static int efa_register_mr(struct ib_pd *ibpd, struct efa_mr *mr, u64 start,
1693			   u64 length, u64 virt_addr, int access_flags)
1694{
1695	struct efa_dev *dev = to_edev(ibpd->device);
1696	struct efa_com_reg_mr_params params = {};
1697	struct efa_com_reg_mr_result result = {};
1698	struct pbl_context pbl;
1699	unsigned int pg_sz;
1700	int inline_size;
1701	int err;
1702
1703	params.pd = to_epd(ibpd)->pdn;
1704	params.iova = virt_addr;
1705	params.mr_length_in_bytes = length;
1706	params.permissions = access_flags;
1707
1708	pg_sz = ib_umem_find_best_pgsz(mr->umem,
1709				       dev->dev_attr.page_size_cap,
1710				       virt_addr);
1711	if (!pg_sz) {
1712		ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1713			  dev->dev_attr.page_size_cap);
1714		return -EOPNOTSUPP;
1715	}
1716
1717	params.page_shift = order_base_2(pg_sz);
1718	params.page_num = ib_umem_num_dma_blocks(mr->umem, pg_sz);
1719
1720	ibdev_dbg(&dev->ibdev,
1721		  "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1722		  start, length, params.page_shift, params.page_num);
1723
1724	inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1725	if (params.page_num <= inline_size) {
1726		err = efa_create_inline_pbl(dev, mr, &params);
1727		if (err)
1728			return err;
1729
1730		err = efa_com_register_mr(&dev->edev, &params, &result);
1731		if (err)
1732			return err;
1733	} else {
1734		err = efa_create_pbl(dev, &pbl, mr, &params);
1735		if (err)
1736			return err;
1737
1738		err = efa_com_register_mr(&dev->edev, &params, &result);
1739		pbl_destroy(dev, &pbl);
1740
1741		if (err)
1742			return err;
1743	}
1744
1745	mr->ibmr.lkey = result.l_key;
1746	mr->ibmr.rkey = result.r_key;
1747	mr->ibmr.length = length;
1748	mr->ic_info.recv_ic_id = result.ic_info.recv_ic_id;
1749	mr->ic_info.rdma_read_ic_id = result.ic_info.rdma_read_ic_id;
1750	mr->ic_info.rdma_recv_ic_id = result.ic_info.rdma_recv_ic_id;
1751	mr->ic_info.recv_ic_id_valid = result.ic_info.recv_ic_id_valid;
1752	mr->ic_info.rdma_read_ic_id_valid = result.ic_info.rdma_read_ic_id_valid;
1753	mr->ic_info.rdma_recv_ic_id_valid = result.ic_info.rdma_recv_ic_id_valid;
1754	ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1755
1756	return 0;
1757}
1758
1759struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start,
1760				     u64 length, u64 virt_addr,
1761				     int fd, int access_flags,
1762				     struct ib_dmah *dmah,
1763				     struct uverbs_attr_bundle *attrs)
1764{
1765	struct efa_dev *dev = to_edev(ibpd->device);
1766	struct ib_umem_dmabuf *umem_dmabuf;
1767	struct efa_mr *mr;
1768	int err;
1769
1770	if (dmah) {
1771		err = -EOPNOTSUPP;
1772		goto err_out;
1773	}
1774
1775	mr = efa_alloc_mr(ibpd, access_flags, &attrs->driver_udata);
1776	if (IS_ERR(mr)) {
1777		err = PTR_ERR(mr);
1778		goto err_out;
1779	}
1780
1781	umem_dmabuf = ib_umem_dmabuf_get_pinned(ibpd->device, start, length, fd,
1782						access_flags);
1783	if (IS_ERR(umem_dmabuf)) {
1784		err = PTR_ERR(umem_dmabuf);
1785		ibdev_dbg(&dev->ibdev, "Failed to get dmabuf umem[%pe]\n",
1786			  umem_dmabuf);
1787		goto err_free;
1788	}
1789
1790	mr->umem = &umem_dmabuf->umem;
1791	err = efa_register_mr(ibpd, mr, start, length, virt_addr, access_flags);
1792	if (err)
1793		goto err_release;
1794
1795	return &mr->ibmr;
1796
1797err_release:
1798	ib_umem_release(mr->umem);
1799err_free:
1800	kfree(mr);
1801err_out:
1802	atomic64_inc(&dev->stats.reg_mr_err);
1803	return ERR_PTR(err);
1804}
1805
1806struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1807			 u64 virt_addr, int access_flags,
1808			 struct ib_dmah *dmah,
1809			 struct ib_udata *udata)
1810{
1811	struct efa_dev *dev = to_edev(ibpd->device);
1812	struct efa_mr *mr;
1813	int err;
1814
1815	if (dmah) {
1816		err = -EOPNOTSUPP;
1817		goto err_out;
1818	}
1819
1820	mr = efa_alloc_mr(ibpd, access_flags, udata);
1821	if (IS_ERR(mr)) {
1822		err = PTR_ERR(mr);
1823		goto err_out;
1824	}
1825
1826	mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
1827	if (IS_ERR(mr->umem)) {
1828		err = PTR_ERR(mr->umem);
1829		ibdev_dbg(&dev->ibdev,
1830			  "Failed to pin and map user space memory[%pe]\n",
1831			  mr->umem);
1832		goto err_free;
1833	}
1834
1835	err = efa_register_mr(ibpd, mr, start, length, virt_addr, access_flags);
1836	if (err)
1837		goto err_release;
1838
1839	return &mr->ibmr;
1840
1841err_release:
1842	ib_umem_release(mr->umem);
1843err_free:
1844	kfree(mr);
1845err_out:
1846	atomic64_inc(&dev->stats.reg_mr_err);
1847	return ERR_PTR(err);
1848}
1849
1850static int UVERBS_HANDLER(EFA_IB_METHOD_MR_QUERY)(struct uverbs_attr_bundle *attrs)
1851{
1852	struct ib_mr *ibmr = uverbs_attr_get_obj(attrs, EFA_IB_ATTR_QUERY_MR_HANDLE);
1853	struct efa_mr *mr = to_emr(ibmr);
1854	u16 ic_id_validity = 0;
1855	int ret;
1856
1857	ret = uverbs_copy_to(attrs, EFA_IB_ATTR_QUERY_MR_RESP_RECV_IC_ID,
1858			     &mr->ic_info.recv_ic_id, sizeof(mr->ic_info.recv_ic_id));
1859	if (ret)
1860		return ret;
1861
1862	ret = uverbs_copy_to(attrs, EFA_IB_ATTR_QUERY_MR_RESP_RDMA_READ_IC_ID,
1863			     &mr->ic_info.rdma_read_ic_id, sizeof(mr->ic_info.rdma_read_ic_id));
1864	if (ret)
1865		return ret;
1866
1867	ret = uverbs_copy_to(attrs, EFA_IB_ATTR_QUERY_MR_RESP_RDMA_RECV_IC_ID,
1868			     &mr->ic_info.rdma_recv_ic_id, sizeof(mr->ic_info.rdma_recv_ic_id));
1869	if (ret)
1870		return ret;
1871
1872	if (mr->ic_info.recv_ic_id_valid)
1873		ic_id_validity |= EFA_QUERY_MR_VALIDITY_RECV_IC_ID;
1874	if (mr->ic_info.rdma_read_ic_id_valid)
1875		ic_id_validity |= EFA_QUERY_MR_VALIDITY_RDMA_READ_IC_ID;
1876	if (mr->ic_info.rdma_recv_ic_id_valid)
1877		ic_id_validity |= EFA_QUERY_MR_VALIDITY_RDMA_RECV_IC_ID;
1878
1879	return uverbs_copy_to(attrs, EFA_IB_ATTR_QUERY_MR_RESP_IC_ID_VALIDITY,
1880			      &ic_id_validity, sizeof(ic_id_validity));
1881}
1882
1883int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1884{
1885	struct efa_dev *dev = to_edev(ibmr->device);
1886	struct efa_com_dereg_mr_params params;
1887	struct efa_mr *mr = to_emr(ibmr);
1888	int err;
1889
1890	ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1891
1892	params.l_key = mr->ibmr.lkey;
1893	err = efa_com_dereg_mr(&dev->edev, &params);
1894	if (err)
1895		return err;
1896
1897	ib_umem_release(mr->umem);
1898	kfree(mr);
1899
1900	return 0;
1901}
1902
1903int efa_get_port_immutable(struct ib_device *ibdev, u32 port_num,
1904			   struct ib_port_immutable *immutable)
1905{
1906	struct ib_port_attr attr;
1907	int err;
1908
1909	err = ib_query_port(ibdev, port_num, &attr);
1910	if (err) {
1911		ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1912		return err;
1913	}
1914
1915	immutable->pkey_tbl_len = attr.pkey_tbl_len;
1916	immutable->gid_tbl_len = attr.gid_tbl_len;
1917
1918	return 0;
1919}
1920
1921static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1922{
1923	struct efa_com_dealloc_uar_params params = {
1924		.uarn = uarn,
1925	};
1926
1927	return efa_com_dealloc_uar(&dev->edev, &params);
1928}
1929
1930#define EFA_CHECK_USER_COMP(_dev, _comp_mask, _attr, _mask, _attr_str) \
1931	(_attr_str = (!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask))) ? \
1932		     NULL : #_attr)
1933
1934static int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
1935				   const struct efa_ibv_alloc_ucontext_cmd *cmd)
1936{
1937	struct efa_dev *dev = to_edev(ibucontext->device);
1938	char *attr_str;
1939
1940	if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, max_tx_batch,
1941				EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH, attr_str))
1942		goto err;
1943
1944	if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, min_sq_depth,
1945				EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR,
1946				attr_str))
1947		goto err;
1948
1949	return 0;
1950
1951err:
1952	ibdev_dbg(&dev->ibdev, "Userspace handshake failed for %s attribute\n",
1953		  attr_str);
1954	return -EOPNOTSUPP;
1955}
1956
1957int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1958{
1959	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1960	struct efa_dev *dev = to_edev(ibucontext->device);
1961	struct efa_ibv_alloc_ucontext_resp resp = {};
1962	struct efa_ibv_alloc_ucontext_cmd cmd = {};
1963	struct efa_com_alloc_uar_result result;
1964	int err;
1965
1966	/*
1967	 * it's fine if the driver does not know all request fields,
1968	 * we will ack input fields in our response.
1969	 */
1970
1971	err = ib_copy_from_udata(&cmd, udata,
1972				 min(sizeof(cmd), udata->inlen));
1973	if (err) {
1974		ibdev_dbg(&dev->ibdev,
1975			  "Cannot copy udata for alloc_ucontext\n");
1976		goto err_out;
1977	}
1978
1979	err = efa_user_comp_handshake(ibucontext, &cmd);
1980	if (err)
1981		goto err_out;
1982
1983	err = efa_com_alloc_uar(&dev->edev, &result);
1984	if (err)
1985		goto err_out;
1986
1987	ucontext->uarn = result.uarn;
1988
1989	resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1990	resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1991	resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1992	resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1993	resp.max_llq_size = dev->dev_attr.max_llq_size;
1994	resp.max_tx_batch = dev->dev_attr.max_tx_batch;
1995	resp.min_sq_wr = dev->dev_attr.min_sq_depth;
1996
1997	err = ib_copy_to_udata(udata, &resp,
1998			       min(sizeof(resp), udata->outlen));
1999	if (err)
2000		goto err_dealloc_uar;
2001
2002	return 0;
2003
2004err_dealloc_uar:
2005	efa_dealloc_uar(dev, result.uarn);
2006err_out:
2007	atomic64_inc(&dev->stats.alloc_ucontext_err);
2008	return err;
2009}
2010
2011void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
2012{
2013	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
2014	struct efa_dev *dev = to_edev(ibucontext->device);
2015
2016	efa_dealloc_uar(dev, ucontext->uarn);
2017}
2018
2019void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
2020{
2021	struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
2022
2023	kfree(entry);
2024}
2025
2026static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
2027		      struct vm_area_struct *vma)
2028{
2029	struct rdma_user_mmap_entry *rdma_entry;
2030	struct efa_user_mmap_entry *entry;
2031	unsigned long va;
2032	int err = 0;
2033	u64 pfn;
2034
2035	rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
2036	if (!rdma_entry) {
2037		ibdev_dbg(&dev->ibdev,
2038			  "pgoff[%#lx] does not have valid entry\n",
2039			  vma->vm_pgoff);
2040		atomic64_inc(&dev->stats.mmap_err);
2041		return -EINVAL;
2042	}
2043	entry = to_emmap(rdma_entry);
2044
2045	ibdev_dbg(&dev->ibdev,
2046		  "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
2047		  entry->address, rdma_entry->npages * PAGE_SIZE,
2048		  entry->mmap_flag);
2049
2050	pfn = entry->address >> PAGE_SHIFT;
2051	switch (entry->mmap_flag) {
2052	case EFA_MMAP_IO_NC:
2053		err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
2054					entry->rdma_entry.npages * PAGE_SIZE,
2055					pgprot_noncached(vma->vm_page_prot),
2056					rdma_entry);
2057		break;
2058	case EFA_MMAP_IO_WC:
2059		err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
2060					entry->rdma_entry.npages * PAGE_SIZE,
2061					pgprot_writecombine(vma->vm_page_prot),
2062					rdma_entry);
2063		break;
2064	case EFA_MMAP_DMA_PAGE:
2065		for (va = vma->vm_start; va < vma->vm_end;
2066		     va += PAGE_SIZE, pfn++) {
2067			err = vm_insert_page(vma, va, pfn_to_page(pfn));
2068			if (err)
2069				break;
2070		}
2071		break;
2072	default:
2073		err = -EINVAL;
2074	}
2075
2076	if (err) {
2077		ibdev_dbg(
2078			&dev->ibdev,
2079			"Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
2080			entry->address, rdma_entry->npages * PAGE_SIZE,
2081			entry->mmap_flag, err);
2082		atomic64_inc(&dev->stats.mmap_err);
2083	}
2084
2085	rdma_user_mmap_entry_put(rdma_entry);
2086	return err;
2087}
2088
2089int efa_mmap(struct ib_ucontext *ibucontext,
2090	     struct vm_area_struct *vma)
2091{
2092	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
2093	struct efa_dev *dev = to_edev(ibucontext->device);
2094	size_t length = vma->vm_end - vma->vm_start;
2095
2096	ibdev_dbg(&dev->ibdev,
2097		  "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
2098		  vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
2099
2100	return __efa_mmap(dev, ucontext, vma);
2101}
2102
2103static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
2104{
2105	struct efa_com_destroy_ah_params params = {
2106		.ah = ah->ah,
2107		.pdn = to_epd(ah->ibah.pd)->pdn,
2108	};
2109
2110	return efa_com_destroy_ah(&dev->edev, &params);
2111}
2112
2113int efa_create_ah(struct ib_ah *ibah,
2114		  struct rdma_ah_init_attr *init_attr,
2115		  struct ib_udata *udata)
2116{
2117	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
2118	struct efa_dev *dev = to_edev(ibah->device);
2119	struct efa_com_create_ah_params params = {};
2120	struct efa_ibv_create_ah_resp resp = {};
2121	struct efa_com_create_ah_result result;
2122	struct efa_ah *ah = to_eah(ibah);
2123	int err;
2124
2125	if (!(init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
2126		ibdev_dbg(&dev->ibdev,
2127			  "Create address handle is not supported in atomic context\n");
2128		err = -EOPNOTSUPP;
2129		goto err_out;
2130	}
2131
2132	if (udata->inlen &&
2133	    !ib_is_udata_cleared(udata, 0, udata->inlen)) {
2134		ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
2135		err = -EINVAL;
2136		goto err_out;
2137	}
2138
2139	memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
2140	       sizeof(params.dest_addr));
2141	params.pdn = to_epd(ibah->pd)->pdn;
2142	err = efa_com_create_ah(&dev->edev, &params, &result);
2143	if (err)
2144		goto err_out;
2145
2146	memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
2147	ah->ah = result.ah;
2148
2149	resp.efa_address_handle = result.ah;
2150
2151	if (udata->outlen) {
2152		err = ib_copy_to_udata(udata, &resp,
2153				       min(sizeof(resp), udata->outlen));
2154		if (err) {
2155			ibdev_dbg(&dev->ibdev,
2156				  "Failed to copy udata for create_ah response\n");
2157			goto err_destroy_ah;
2158		}
2159	}
2160	ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
2161
2162	return 0;
2163
2164err_destroy_ah:
2165	efa_ah_destroy(dev, ah);
2166err_out:
2167	atomic64_inc(&dev->stats.create_ah_err);
2168	return err;
2169}
2170
2171int efa_destroy_ah(struct ib_ah *ibah, u32 flags)
2172{
2173	struct efa_dev *dev = to_edev(ibah->pd->device);
2174	struct efa_ah *ah = to_eah(ibah);
2175
2176	ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
2177
2178	if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
2179		ibdev_dbg(&dev->ibdev,
2180			  "Destroy address handle is not supported in atomic context\n");
2181		return -EOPNOTSUPP;
2182	}
2183
2184	efa_ah_destroy(dev, ah);
2185	return 0;
2186}
2187
2188struct rdma_hw_stats *efa_alloc_hw_port_stats(struct ib_device *ibdev,
2189					      u32 port_num)
2190{
2191	return rdma_alloc_hw_stats_struct(efa_port_stats_descs,
2192					  ARRAY_SIZE(efa_port_stats_descs),
2193					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2194}
2195
2196struct rdma_hw_stats *efa_alloc_hw_device_stats(struct ib_device *ibdev)
2197{
2198	return rdma_alloc_hw_stats_struct(efa_device_stats_descs,
2199					  ARRAY_SIZE(efa_device_stats_descs),
2200					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
2201}
2202
2203static int efa_fill_device_stats(struct efa_dev *dev,
2204				 struct rdma_hw_stats *stats)
2205{
2206	struct efa_com_stats_admin *as = &dev->edev.aq.stats;
2207	struct efa_stats *s = &dev->stats;
2208
2209	stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
2210	stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
2211	stats->value[EFA_CMDS_ERR] = atomic64_read(&as->cmd_err);
2212	stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
2213
2214	stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
2215	stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->alloc_pd_err);
2216	stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->create_qp_err);
2217	stats->value[EFA_CREATE_CQ_ERR] = atomic64_read(&s->create_cq_err);
2218	stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->reg_mr_err);
2219	stats->value[EFA_ALLOC_UCONTEXT_ERR] =
2220		atomic64_read(&s->alloc_ucontext_err);
2221	stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->create_ah_err);
2222	stats->value[EFA_MMAP_ERR] = atomic64_read(&s->mmap_err);
2223
2224	return ARRAY_SIZE(efa_device_stats_descs);
2225}
2226
2227static int efa_fill_port_stats(struct efa_dev *dev, struct rdma_hw_stats *stats,
2228			       u32 port_num)
2229{
2230	struct efa_com_get_stats_params params = {};
2231	union efa_com_get_stats_result result;
2232	struct efa_com_rdma_write_stats *rws;
2233	struct efa_com_rdma_read_stats *rrs;
2234	struct efa_com_messages_stats *ms;
2235	struct efa_com_network_stats *ns;
2236	struct efa_com_basic_stats *bs;
2237	int err;
2238
2239	params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
2240	params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
2241
2242	err = efa_com_get_stats(&dev->edev, &params, &result);
2243	if (err)
2244		return err;
2245
2246	bs = &result.basic_stats;
2247	stats->value[EFA_TX_BYTES] = bs->tx_bytes;
2248	stats->value[EFA_TX_PKTS] = bs->tx_pkts;
2249	stats->value[EFA_RX_BYTES] = bs->rx_bytes;
2250	stats->value[EFA_RX_PKTS] = bs->rx_pkts;
2251	stats->value[EFA_RX_DROPS] = bs->rx_drops;
2252
2253	params.type = EFA_ADMIN_GET_STATS_TYPE_MESSAGES;
2254	err = efa_com_get_stats(&dev->edev, &params, &result);
2255	if (err)
2256		return err;
2257
2258	ms = &result.messages_stats;
2259	stats->value[EFA_SEND_BYTES] = ms->send_bytes;
2260	stats->value[EFA_SEND_WRS] = ms->send_wrs;
2261	stats->value[EFA_RECV_BYTES] = ms->recv_bytes;
2262	stats->value[EFA_RECV_WRS] = ms->recv_wrs;
2263
2264	params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_READ;
2265	err = efa_com_get_stats(&dev->edev, &params, &result);
2266	if (err)
2267		return err;
2268
2269	rrs = &result.rdma_read_stats;
2270	stats->value[EFA_RDMA_READ_WRS] = rrs->read_wrs;
2271	stats->value[EFA_RDMA_READ_BYTES] = rrs->read_bytes;
2272	stats->value[EFA_RDMA_READ_WR_ERR] = rrs->read_wr_err;
2273	stats->value[EFA_RDMA_READ_RESP_BYTES] = rrs->read_resp_bytes;
2274
2275	if (EFA_DEV_CAP(dev, RDMA_WRITE)) {
2276		params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_WRITE;
2277		err = efa_com_get_stats(&dev->edev, &params, &result);
2278		if (err)
2279			return err;
2280
2281		rws = &result.rdma_write_stats;
2282		stats->value[EFA_RDMA_WRITE_WRS] = rws->write_wrs;
2283		stats->value[EFA_RDMA_WRITE_BYTES] = rws->write_bytes;
2284		stats->value[EFA_RDMA_WRITE_WR_ERR] = rws->write_wr_err;
2285		stats->value[EFA_RDMA_WRITE_RECV_BYTES] = rws->write_recv_bytes;
2286	}
2287
2288	params.type = EFA_ADMIN_GET_STATS_TYPE_NETWORK;
2289	err = efa_com_get_stats(&dev->edev, &params, &result);
2290	if (err)
2291		return err;
2292
2293	ns = &result.network_stats;
2294	stats->value[EFA_RETRANS_BYTES] = ns->retrans_bytes;
2295	stats->value[EFA_RETRANS_PKTS] = ns->retrans_pkts;
2296	stats->value[EFA_RETRANS_TIMEOUT_EVENS] = ns->retrans_timeout_events;
2297	stats->value[EFA_UNRESPONSIVE_REMOTE_EVENTS] = ns->unresponsive_remote_events;
2298	stats->value[EFA_IMPAIRED_REMOTE_CONN_EVENTS] = ns->impaired_remote_conn_events;
2299
2300	return ARRAY_SIZE(efa_port_stats_descs);
2301}
2302
2303int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
2304		     u32 port_num, int index)
2305{
2306	if (port_num)
2307		return efa_fill_port_stats(to_edev(ibdev), stats, port_num);
2308	else
2309		return efa_fill_device_stats(to_edev(ibdev), stats);
2310}
2311
2312enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
2313					 u32 port_num)
2314{
2315	return IB_LINK_LAYER_UNSPECIFIED;
2316}
2317
2318DECLARE_UVERBS_NAMED_METHOD(EFA_IB_METHOD_MR_QUERY,
2319			    UVERBS_ATTR_IDR(EFA_IB_ATTR_QUERY_MR_HANDLE,
2320					    UVERBS_OBJECT_MR,
2321					    UVERBS_ACCESS_READ,
2322					    UA_MANDATORY),
2323			    UVERBS_ATTR_PTR_OUT(EFA_IB_ATTR_QUERY_MR_RESP_IC_ID_VALIDITY,
2324						UVERBS_ATTR_TYPE(u16),
2325						UA_MANDATORY),
2326			    UVERBS_ATTR_PTR_OUT(EFA_IB_ATTR_QUERY_MR_RESP_RECV_IC_ID,
2327						UVERBS_ATTR_TYPE(u16),
2328						UA_MANDATORY),
2329			    UVERBS_ATTR_PTR_OUT(EFA_IB_ATTR_QUERY_MR_RESP_RDMA_READ_IC_ID,
2330						UVERBS_ATTR_TYPE(u16),
2331						UA_MANDATORY),
2332			    UVERBS_ATTR_PTR_OUT(EFA_IB_ATTR_QUERY_MR_RESP_RDMA_RECV_IC_ID,
2333						UVERBS_ATTR_TYPE(u16),
2334						UA_MANDATORY));
2335
2336ADD_UVERBS_METHODS(efa_mr,
2337		   UVERBS_OBJECT_MR,
2338		   &UVERBS_METHOD(EFA_IB_METHOD_MR_QUERY));
2339
2340const struct uapi_definition efa_uapi_defs[] = {
2341	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_MR,
2342				&efa_mr),
2343	{},
2344};