Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2/*
   3 * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
   4 * Copyright (c) 2020, Intel Corporation. All rights reserved.
   5 */
   6
   7#include <linux/debugfs.h>
   8#include <linux/highmem.h>
   9#include <linux/module.h>
  10#include <linux/init.h>
  11#include <linux/errno.h>
  12#include <linux/pci.h>
  13#include <linux/dma-mapping.h>
  14#include <linux/slab.h>
  15#include <linux/bitmap.h>
  16#include <linux/sched.h>
  17#include <linux/sched/mm.h>
  18#include <linux/sched/task.h>
  19#include <linux/delay.h>
  20#include <rdma/ib_user_verbs.h>
  21#include <rdma/ib_addr.h>
  22#include <rdma/ib_cache.h>
  23#include <linux/mlx5/port.h>
  24#include <linux/mlx5/vport.h>
  25#include <linux/mlx5/fs.h>
  26#include <linux/mlx5/eswitch.h>
  27#include <linux/list.h>
  28#include <rdma/ib_smi.h>
  29#include <rdma/ib_umem.h>
  30#include <rdma/lag.h>
  31#include <linux/in.h>
  32#include <linux/etherdevice.h>
  33#include "mlx5_ib.h"
  34#include "ib_rep.h"
  35#include "cmd.h"
  36#include "devx.h"
  37#include "dm.h"
  38#include "fs.h"
  39#include "srq.h"
  40#include "qp.h"
  41#include "wr.h"
  42#include "restrack.h"
  43#include "counters.h"
  44#include <linux/mlx5/accel.h>
  45#include <rdma/uverbs_std_types.h>
  46#include <rdma/uverbs_ioctl.h>
  47#include <rdma/mlx5_user_ioctl_verbs.h>
  48#include <rdma/mlx5_user_ioctl_cmds.h>
  49#include <rdma/ib_umem_odp.h>
  50
  51#define UVERBS_MODULE_NAME mlx5_ib
  52#include <rdma/uverbs_named_ioctl.h>
  53
  54MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
  55MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
  56MODULE_LICENSE("Dual BSD/GPL");
  57
  58struct mlx5_ib_event_work {
  59	struct work_struct	work;
  60	union {
  61		struct mlx5_ib_dev	      *dev;
  62		struct mlx5_ib_multiport_info *mpi;
  63	};
  64	bool			is_slave;
  65	unsigned int		event;
  66	void			*param;
  67};
  68
  69enum {
  70	MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
  71};
  72
  73static struct workqueue_struct *mlx5_ib_event_wq;
  74static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
  75static LIST_HEAD(mlx5_ib_dev_list);
  76/*
  77 * This mutex should be held when accessing either of the above lists
  78 */
  79static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
  80
  81struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
  82{
  83	struct mlx5_ib_dev *dev;
  84
  85	mutex_lock(&mlx5_ib_multiport_mutex);
  86	dev = mpi->ibdev;
  87	mutex_unlock(&mlx5_ib_multiport_mutex);
  88	return dev;
  89}
  90
  91static enum rdma_link_layer
  92mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
  93{
  94	switch (port_type_cap) {
  95	case MLX5_CAP_PORT_TYPE_IB:
  96		return IB_LINK_LAYER_INFINIBAND;
  97	case MLX5_CAP_PORT_TYPE_ETH:
  98		return IB_LINK_LAYER_ETHERNET;
  99	default:
 100		return IB_LINK_LAYER_UNSPECIFIED;
 101	}
 102}
 103
 104static enum rdma_link_layer
 105mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
 106{
 107	struct mlx5_ib_dev *dev = to_mdev(device);
 108	int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
 109
 110	return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
 111}
 112
 113static int get_port_state(struct ib_device *ibdev,
 114			  u32 port_num,
 115			  enum ib_port_state *state)
 116{
 117	struct ib_port_attr attr;
 118	int ret;
 119
 120	memset(&attr, 0, sizeof(attr));
 121	ret = ibdev->ops.query_port(ibdev, port_num, &attr);
 122	if (!ret)
 123		*state = attr.state;
 124	return ret;
 125}
 126
 127static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
 128					   struct net_device *ndev,
 129					   u32 *port_num)
 130{
 131	struct net_device *rep_ndev;
 132	struct mlx5_ib_port *port;
 133	int i;
 134
 135	for (i = 0; i < dev->num_ports; i++) {
 136		port  = &dev->port[i];
 137		if (!port->rep)
 138			continue;
 139
 140		read_lock(&port->roce.netdev_lock);
 141		rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
 142						  port->rep->vport);
 143		if (rep_ndev == ndev) {
 144			read_unlock(&port->roce.netdev_lock);
 145			*port_num = i + 1;
 146			return &port->roce;
 147		}
 148		read_unlock(&port->roce.netdev_lock);
 149	}
 150
 151	return NULL;
 152}
 153
 154static int mlx5_netdev_event(struct notifier_block *this,
 155			     unsigned long event, void *ptr)
 156{
 157	struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
 158	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 159	u32 port_num = roce->native_port_num;
 160	struct mlx5_core_dev *mdev;
 161	struct mlx5_ib_dev *ibdev;
 162
 163	ibdev = roce->dev;
 164	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
 165	if (!mdev)
 166		return NOTIFY_DONE;
 167
 168	switch (event) {
 169	case NETDEV_REGISTER:
 170		/* Should already be registered during the load */
 171		if (ibdev->is_rep)
 172			break;
 173		write_lock(&roce->netdev_lock);
 174		if (ndev->dev.parent == mdev->device)
 175			roce->netdev = ndev;
 176		write_unlock(&roce->netdev_lock);
 177		break;
 178
 179	case NETDEV_UNREGISTER:
 180		/* In case of reps, ib device goes away before the netdevs */
 181		write_lock(&roce->netdev_lock);
 182		if (roce->netdev == ndev)
 183			roce->netdev = NULL;
 184		write_unlock(&roce->netdev_lock);
 185		break;
 186
 187	case NETDEV_CHANGE:
 188	case NETDEV_UP:
 189	case NETDEV_DOWN: {
 190		struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
 191		struct net_device *upper = NULL;
 192
 193		if (lag_ndev) {
 194			upper = netdev_master_upper_dev_get(lag_ndev);
 195			dev_put(lag_ndev);
 196		}
 197
 198		if (ibdev->is_rep)
 199			roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
 200		if (!roce)
 201			return NOTIFY_DONE;
 202		if ((upper == ndev || (!upper && ndev == roce->netdev))
 203		    && ibdev->ib_active) {
 204			struct ib_event ibev = { };
 205			enum ib_port_state port_state;
 206
 207			if (get_port_state(&ibdev->ib_dev, port_num,
 208					   &port_state))
 209				goto done;
 210
 211			if (roce->last_port_state == port_state)
 212				goto done;
 213
 214			roce->last_port_state = port_state;
 215			ibev.device = &ibdev->ib_dev;
 216			if (port_state == IB_PORT_DOWN)
 217				ibev.event = IB_EVENT_PORT_ERR;
 218			else if (port_state == IB_PORT_ACTIVE)
 219				ibev.event = IB_EVENT_PORT_ACTIVE;
 220			else
 221				goto done;
 222
 223			ibev.element.port_num = port_num;
 224			ib_dispatch_event(&ibev);
 225		}
 226		break;
 227	}
 228
 229	default:
 230		break;
 231	}
 232done:
 233	mlx5_ib_put_native_port_mdev(ibdev, port_num);
 234	return NOTIFY_DONE;
 235}
 236
 237static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
 238					     u32 port_num)
 239{
 240	struct mlx5_ib_dev *ibdev = to_mdev(device);
 241	struct net_device *ndev;
 242	struct mlx5_core_dev *mdev;
 243
 244	mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
 245	if (!mdev)
 246		return NULL;
 247
 248	ndev = mlx5_lag_get_roce_netdev(mdev);
 249	if (ndev)
 250		goto out;
 251
 252	/* Ensure ndev does not disappear before we invoke dev_hold()
 253	 */
 254	read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
 255	ndev = ibdev->port[port_num - 1].roce.netdev;
 256	if (ndev)
 257		dev_hold(ndev);
 258	read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
 259
 260out:
 261	mlx5_ib_put_native_port_mdev(ibdev, port_num);
 262	return ndev;
 263}
 264
 265struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
 266						   u32 ib_port_num,
 267						   u32 *native_port_num)
 268{
 269	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
 270							  ib_port_num);
 271	struct mlx5_core_dev *mdev = NULL;
 272	struct mlx5_ib_multiport_info *mpi;
 273	struct mlx5_ib_port *port;
 274
 275	if (!mlx5_core_mp_enabled(ibdev->mdev) ||
 276	    ll != IB_LINK_LAYER_ETHERNET) {
 277		if (native_port_num)
 278			*native_port_num = ib_port_num;
 279		return ibdev->mdev;
 280	}
 281
 282	if (native_port_num)
 283		*native_port_num = 1;
 284
 285	port = &ibdev->port[ib_port_num - 1];
 286	spin_lock(&port->mp.mpi_lock);
 287	mpi = ibdev->port[ib_port_num - 1].mp.mpi;
 288	if (mpi && !mpi->unaffiliate) {
 289		mdev = mpi->mdev;
 290		/* If it's the master no need to refcount, it'll exist
 291		 * as long as the ib_dev exists.
 292		 */
 293		if (!mpi->is_master)
 294			mpi->mdev_refcnt++;
 295	}
 296	spin_unlock(&port->mp.mpi_lock);
 297
 298	return mdev;
 299}
 300
 301void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
 302{
 303	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
 304							  port_num);
 305	struct mlx5_ib_multiport_info *mpi;
 306	struct mlx5_ib_port *port;
 307
 308	if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
 309		return;
 310
 311	port = &ibdev->port[port_num - 1];
 312
 313	spin_lock(&port->mp.mpi_lock);
 314	mpi = ibdev->port[port_num - 1].mp.mpi;
 315	if (mpi->is_master)
 316		goto out;
 317
 318	mpi->mdev_refcnt--;
 319	if (mpi->unaffiliate)
 320		complete(&mpi->unref_comp);
 321out:
 322	spin_unlock(&port->mp.mpi_lock);
 323}
 324
 325static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
 326					   u16 *active_speed, u8 *active_width)
 327{
 328	switch (eth_proto_oper) {
 329	case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
 330	case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
 331	case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
 332	case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
 333		*active_width = IB_WIDTH_1X;
 334		*active_speed = IB_SPEED_SDR;
 335		break;
 336	case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
 337	case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
 338	case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
 339	case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
 340	case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
 341	case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
 342	case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
 343		*active_width = IB_WIDTH_1X;
 344		*active_speed = IB_SPEED_QDR;
 345		break;
 346	case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
 347	case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
 348	case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
 349		*active_width = IB_WIDTH_1X;
 350		*active_speed = IB_SPEED_EDR;
 351		break;
 352	case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
 353	case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
 354	case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
 355	case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
 356		*active_width = IB_WIDTH_4X;
 357		*active_speed = IB_SPEED_QDR;
 358		break;
 359	case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
 360	case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
 361	case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
 362		*active_width = IB_WIDTH_1X;
 363		*active_speed = IB_SPEED_HDR;
 364		break;
 365	case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
 366		*active_width = IB_WIDTH_4X;
 367		*active_speed = IB_SPEED_FDR;
 368		break;
 369	case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
 370	case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
 371	case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
 372	case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
 373		*active_width = IB_WIDTH_4X;
 374		*active_speed = IB_SPEED_EDR;
 375		break;
 376	default:
 377		return -EINVAL;
 378	}
 379
 380	return 0;
 381}
 382
 383static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
 384					u8 *active_width)
 385{
 386	switch (eth_proto_oper) {
 387	case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
 388	case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
 389		*active_width = IB_WIDTH_1X;
 390		*active_speed = IB_SPEED_SDR;
 391		break;
 392	case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
 393		*active_width = IB_WIDTH_1X;
 394		*active_speed = IB_SPEED_DDR;
 395		break;
 396	case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
 397		*active_width = IB_WIDTH_1X;
 398		*active_speed = IB_SPEED_QDR;
 399		break;
 400	case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
 401		*active_width = IB_WIDTH_4X;
 402		*active_speed = IB_SPEED_QDR;
 403		break;
 404	case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
 405		*active_width = IB_WIDTH_1X;
 406		*active_speed = IB_SPEED_EDR;
 407		break;
 408	case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
 409		*active_width = IB_WIDTH_2X;
 410		*active_speed = IB_SPEED_EDR;
 411		break;
 412	case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
 413		*active_width = IB_WIDTH_1X;
 414		*active_speed = IB_SPEED_HDR;
 415		break;
 416	case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
 417		*active_width = IB_WIDTH_4X;
 418		*active_speed = IB_SPEED_EDR;
 419		break;
 420	case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
 421		*active_width = IB_WIDTH_2X;
 422		*active_speed = IB_SPEED_HDR;
 423		break;
 424	case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
 425		*active_width = IB_WIDTH_1X;
 426		*active_speed = IB_SPEED_NDR;
 427		break;
 428	case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
 429		*active_width = IB_WIDTH_4X;
 430		*active_speed = IB_SPEED_HDR;
 431		break;
 432	case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
 433		*active_width = IB_WIDTH_2X;
 434		*active_speed = IB_SPEED_NDR;
 435		break;
 436	case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
 437		*active_width = IB_WIDTH_4X;
 438		*active_speed = IB_SPEED_NDR;
 439		break;
 440	default:
 441		return -EINVAL;
 442	}
 443
 444	return 0;
 445}
 446
 447static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
 448				    u8 *active_width, bool ext)
 449{
 450	return ext ?
 451		translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
 452					     active_width) :
 453		translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
 454						active_width);
 455}
 456
 457static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
 458				struct ib_port_attr *props)
 459{
 460	struct mlx5_ib_dev *dev = to_mdev(device);
 461	u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
 462	struct mlx5_core_dev *mdev;
 463	struct net_device *ndev, *upper;
 464	enum ib_mtu ndev_ib_mtu;
 465	bool put_mdev = true;
 466	u32 eth_prot_oper;
 467	u32 mdev_port_num;
 468	bool ext;
 469	int err;
 470
 471	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
 472	if (!mdev) {
 473		/* This means the port isn't affiliated yet. Get the
 474		 * info for the master port instead.
 475		 */
 476		put_mdev = false;
 477		mdev = dev->mdev;
 478		mdev_port_num = 1;
 479		port_num = 1;
 480	}
 481
 482	/* Possible bad flows are checked before filling out props so in case
 483	 * of an error it will still be zeroed out.
 484	 * Use native port in case of reps
 485	 */
 486	if (dev->is_rep)
 487		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
 488					   1);
 489	else
 490		err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
 491					   mdev_port_num);
 492	if (err)
 493		goto out;
 494	ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
 495	eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
 496
 497	props->active_width     = IB_WIDTH_4X;
 498	props->active_speed     = IB_SPEED_QDR;
 499
 500	translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
 501				 &props->active_width, ext);
 502
 503	if (!dev->is_rep && dev->mdev->roce.roce_en) {
 504		u16 qkey_viol_cntr;
 505
 506		props->port_cap_flags |= IB_PORT_CM_SUP;
 507		props->ip_gids = true;
 508		props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
 509						   roce_address_table_size);
 510		mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
 511		props->qkey_viol_cntr = qkey_viol_cntr;
 512	}
 513	props->max_mtu          = IB_MTU_4096;
 514	props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
 515	props->pkey_tbl_len     = 1;
 516	props->state            = IB_PORT_DOWN;
 517	props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
 518
 519	/* If this is a stub query for an unaffiliated port stop here */
 520	if (!put_mdev)
 521		goto out;
 522
 523	ndev = mlx5_ib_get_netdev(device, port_num);
 524	if (!ndev)
 525		goto out;
 526
 527	if (dev->lag_active) {
 528		rcu_read_lock();
 529		upper = netdev_master_upper_dev_get_rcu(ndev);
 530		if (upper) {
 531			dev_put(ndev);
 532			ndev = upper;
 533			dev_hold(ndev);
 534		}
 535		rcu_read_unlock();
 536	}
 537
 538	if (netif_running(ndev) && netif_carrier_ok(ndev)) {
 539		props->state      = IB_PORT_ACTIVE;
 540		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 541	}
 542
 543	ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
 544
 545	dev_put(ndev);
 546
 547	props->active_mtu	= min(props->max_mtu, ndev_ib_mtu);
 548out:
 549	if (put_mdev)
 550		mlx5_ib_put_native_port_mdev(dev, port_num);
 551	return err;
 552}
 553
 554static int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
 555			 unsigned int index, const union ib_gid *gid,
 556			 const struct ib_gid_attr *attr)
 557{
 558	enum ib_gid_type gid_type;
 559	u16 vlan_id = 0xffff;
 560	u8 roce_version = 0;
 561	u8 roce_l3_type = 0;
 562	u8 mac[ETH_ALEN];
 563	int ret;
 564
 565	gid_type = attr->gid_type;
 566	if (gid) {
 567		ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
 568		if (ret)
 569			return ret;
 570	}
 571
 572	switch (gid_type) {
 573	case IB_GID_TYPE_ROCE:
 574		roce_version = MLX5_ROCE_VERSION_1;
 575		break;
 576	case IB_GID_TYPE_ROCE_UDP_ENCAP:
 577		roce_version = MLX5_ROCE_VERSION_2;
 578		if (gid && ipv6_addr_v4mapped((void *)gid))
 579			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
 580		else
 581			roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
 582		break;
 583
 584	default:
 585		mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
 586	}
 587
 588	return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
 589				      roce_l3_type, gid->raw, mac,
 590				      vlan_id < VLAN_CFI_MASK, vlan_id,
 591				      port_num);
 592}
 593
 594static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
 595			   __always_unused void **context)
 596{
 597	return set_roce_addr(to_mdev(attr->device), attr->port_num,
 598			     attr->index, &attr->gid, attr);
 599}
 600
 601static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
 602			   __always_unused void **context)
 603{
 604	return set_roce_addr(to_mdev(attr->device), attr->port_num,
 605			     attr->index, NULL, attr);
 606}
 607
 608__be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
 609				   const struct ib_gid_attr *attr)
 610{
 611	if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
 612		return 0;
 613
 614	return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
 615}
 616
 617static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
 618{
 619	if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
 620		return !MLX5_CAP_GEN(dev->mdev, ib_virt);
 621	return 0;
 622}
 623
 624enum {
 625	MLX5_VPORT_ACCESS_METHOD_MAD,
 626	MLX5_VPORT_ACCESS_METHOD_HCA,
 627	MLX5_VPORT_ACCESS_METHOD_NIC,
 628};
 629
 630static int mlx5_get_vport_access_method(struct ib_device *ibdev)
 631{
 632	if (mlx5_use_mad_ifc(to_mdev(ibdev)))
 633		return MLX5_VPORT_ACCESS_METHOD_MAD;
 634
 635	if (mlx5_ib_port_link_layer(ibdev, 1) ==
 636	    IB_LINK_LAYER_ETHERNET)
 637		return MLX5_VPORT_ACCESS_METHOD_NIC;
 638
 639	return MLX5_VPORT_ACCESS_METHOD_HCA;
 640}
 641
 642static void get_atomic_caps(struct mlx5_ib_dev *dev,
 643			    u8 atomic_size_qp,
 644			    struct ib_device_attr *props)
 645{
 646	u8 tmp;
 647	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
 648	u8 atomic_req_8B_endianness_mode =
 649		MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
 650
 651	/* Check if HW supports 8 bytes standard atomic operations and capable
 652	 * of host endianness respond
 653	 */
 654	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
 655	if (((atomic_operations & tmp) == tmp) &&
 656	    (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
 657	    (atomic_req_8B_endianness_mode)) {
 658		props->atomic_cap = IB_ATOMIC_HCA;
 659	} else {
 660		props->atomic_cap = IB_ATOMIC_NONE;
 661	}
 662}
 663
 664static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
 665			       struct ib_device_attr *props)
 666{
 667	u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
 668
 669	get_atomic_caps(dev, atomic_size_qp, props);
 670}
 671
 672static int mlx5_query_system_image_guid(struct ib_device *ibdev,
 673					__be64 *sys_image_guid)
 674{
 675	struct mlx5_ib_dev *dev = to_mdev(ibdev);
 676	struct mlx5_core_dev *mdev = dev->mdev;
 677	u64 tmp;
 678	int err;
 679
 680	switch (mlx5_get_vport_access_method(ibdev)) {
 681	case MLX5_VPORT_ACCESS_METHOD_MAD:
 682		return mlx5_query_mad_ifc_system_image_guid(ibdev,
 683							    sys_image_guid);
 684
 685	case MLX5_VPORT_ACCESS_METHOD_HCA:
 686		err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
 687		break;
 688
 689	case MLX5_VPORT_ACCESS_METHOD_NIC:
 690		err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
 691		break;
 692
 693	default:
 694		return -EINVAL;
 695	}
 696
 697	if (!err)
 698		*sys_image_guid = cpu_to_be64(tmp);
 699
 700	return err;
 701
 702}
 703
 704static int mlx5_query_max_pkeys(struct ib_device *ibdev,
 705				u16 *max_pkeys)
 706{
 707	struct mlx5_ib_dev *dev = to_mdev(ibdev);
 708	struct mlx5_core_dev *mdev = dev->mdev;
 709
 710	switch (mlx5_get_vport_access_method(ibdev)) {
 711	case MLX5_VPORT_ACCESS_METHOD_MAD:
 712		return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
 713
 714	case MLX5_VPORT_ACCESS_METHOD_HCA:
 715	case MLX5_VPORT_ACCESS_METHOD_NIC:
 716		*max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
 717						pkey_table_size));
 718		return 0;
 719
 720	default:
 721		return -EINVAL;
 722	}
 723}
 724
 725static int mlx5_query_vendor_id(struct ib_device *ibdev,
 726				u32 *vendor_id)
 727{
 728	struct mlx5_ib_dev *dev = to_mdev(ibdev);
 729
 730	switch (mlx5_get_vport_access_method(ibdev)) {
 731	case MLX5_VPORT_ACCESS_METHOD_MAD:
 732		return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
 733
 734	case MLX5_VPORT_ACCESS_METHOD_HCA:
 735	case MLX5_VPORT_ACCESS_METHOD_NIC:
 736		return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
 737
 738	default:
 739		return -EINVAL;
 740	}
 741}
 742
 743static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
 744				__be64 *node_guid)
 745{
 746	u64 tmp;
 747	int err;
 748
 749	switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
 750	case MLX5_VPORT_ACCESS_METHOD_MAD:
 751		return mlx5_query_mad_ifc_node_guid(dev, node_guid);
 752
 753	case MLX5_VPORT_ACCESS_METHOD_HCA:
 754		err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
 755		break;
 756
 757	case MLX5_VPORT_ACCESS_METHOD_NIC:
 758		err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
 759		break;
 760
 761	default:
 762		return -EINVAL;
 763	}
 764
 765	if (!err)
 766		*node_guid = cpu_to_be64(tmp);
 767
 768	return err;
 769}
 770
 771struct mlx5_reg_node_desc {
 772	u8	desc[IB_DEVICE_NODE_DESC_MAX];
 773};
 774
 775static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
 776{
 777	struct mlx5_reg_node_desc in;
 778
 779	if (mlx5_use_mad_ifc(dev))
 780		return mlx5_query_mad_ifc_node_desc(dev, node_desc);
 781
 782	memset(&in, 0, sizeof(in));
 783
 784	return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
 785				    sizeof(struct mlx5_reg_node_desc),
 786				    MLX5_REG_NODE_DESC, 0, 0);
 787}
 788
 789static int mlx5_ib_query_device(struct ib_device *ibdev,
 790				struct ib_device_attr *props,
 791				struct ib_udata *uhw)
 792{
 793	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
 794	struct mlx5_ib_dev *dev = to_mdev(ibdev);
 795	struct mlx5_core_dev *mdev = dev->mdev;
 796	int err = -ENOMEM;
 797	int max_sq_desc;
 798	int max_rq_sg;
 799	int max_sq_sg;
 800	u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
 801	bool raw_support = !mlx5_core_mp_enabled(mdev);
 802	struct mlx5_ib_query_device_resp resp = {};
 803	size_t resp_len;
 804	u64 max_tso;
 805
 806	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
 807	if (uhw_outlen && uhw_outlen < resp_len)
 808		return -EINVAL;
 809
 810	resp.response_length = resp_len;
 811
 812	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
 813		return -EINVAL;
 814
 815	memset(props, 0, sizeof(*props));
 816	err = mlx5_query_system_image_guid(ibdev,
 817					   &props->sys_image_guid);
 818	if (err)
 819		return err;
 820
 821	props->max_pkeys = dev->pkey_table_len;
 822
 823	err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
 824	if (err)
 825		return err;
 826
 827	props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
 828		(fw_rev_min(dev->mdev) << 16) |
 829		fw_rev_sub(dev->mdev);
 830	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
 831		IB_DEVICE_PORT_ACTIVE_EVENT		|
 832		IB_DEVICE_SYS_IMAGE_GUID		|
 833		IB_DEVICE_RC_RNR_NAK_GEN;
 834
 835	if (MLX5_CAP_GEN(mdev, pkv))
 836		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
 837	if (MLX5_CAP_GEN(mdev, qkv))
 838		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
 839	if (MLX5_CAP_GEN(mdev, apm))
 840		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
 841	if (MLX5_CAP_GEN(mdev, xrc))
 842		props->device_cap_flags |= IB_DEVICE_XRC;
 843	if (MLX5_CAP_GEN(mdev, imaicl)) {
 844		props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
 845					   IB_DEVICE_MEM_WINDOW_TYPE_2B;
 846		props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 847		/* We support 'Gappy' memory registration too */
 848		props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
 849	}
 850	/* IB_WR_REG_MR always requires changing the entity size with UMR */
 851	if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
 852		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
 853	if (MLX5_CAP_GEN(mdev, sho)) {
 854		props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
 855		/* At this stage no support for signature handover */
 856		props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
 857				      IB_PROT_T10DIF_TYPE_2 |
 858				      IB_PROT_T10DIF_TYPE_3;
 859		props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
 860				       IB_GUARD_T10DIF_CSUM;
 861	}
 862	if (MLX5_CAP_GEN(mdev, block_lb_mc))
 863		props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
 864
 865	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
 866		if (MLX5_CAP_ETH(mdev, csum_cap)) {
 867			/* Legacy bit to support old userspace libraries */
 868			props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
 869			props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
 870		}
 871
 872		if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
 873			props->raw_packet_caps |=
 874				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
 875
 876		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
 877			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
 878			if (max_tso) {
 879				resp.tso_caps.max_tso = 1 << max_tso;
 880				resp.tso_caps.supported_qpts |=
 881					1 << IB_QPT_RAW_PACKET;
 882				resp.response_length += sizeof(resp.tso_caps);
 883			}
 884		}
 885
 886		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
 887			resp.rss_caps.rx_hash_function =
 888						MLX5_RX_HASH_FUNC_TOEPLITZ;
 889			resp.rss_caps.rx_hash_fields_mask =
 890						MLX5_RX_HASH_SRC_IPV4 |
 891						MLX5_RX_HASH_DST_IPV4 |
 892						MLX5_RX_HASH_SRC_IPV6 |
 893						MLX5_RX_HASH_DST_IPV6 |
 894						MLX5_RX_HASH_SRC_PORT_TCP |
 895						MLX5_RX_HASH_DST_PORT_TCP |
 896						MLX5_RX_HASH_SRC_PORT_UDP |
 897						MLX5_RX_HASH_DST_PORT_UDP |
 898						MLX5_RX_HASH_INNER;
 899			if (mlx5_accel_ipsec_device_caps(dev->mdev) &
 900			    MLX5_ACCEL_IPSEC_CAP_DEVICE)
 901				resp.rss_caps.rx_hash_fields_mask |=
 902					MLX5_RX_HASH_IPSEC_SPI;
 903			resp.response_length += sizeof(resp.rss_caps);
 904		}
 905	} else {
 906		if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
 907			resp.response_length += sizeof(resp.tso_caps);
 908		if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
 909			resp.response_length += sizeof(resp.rss_caps);
 910	}
 911
 912	if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
 913		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 914		props->device_cap_flags |= IB_DEVICE_UD_TSO;
 915	}
 916
 917	if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
 918	    MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
 919	    raw_support)
 920		props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
 921
 922	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
 923	    MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
 924		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 925
 926	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
 927	    MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
 928	    raw_support) {
 929		/* Legacy bit to support old userspace libraries */
 930		props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
 931		props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
 932	}
 933
 934	if (MLX5_CAP_DEV_MEM(mdev, memic)) {
 935		props->max_dm_size =
 936			MLX5_CAP_DEV_MEM(mdev, max_memic_size);
 937	}
 938
 939	if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
 940		props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
 941
 942	if (MLX5_CAP_GEN(mdev, end_pad))
 943		props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
 944
 945	props->vendor_part_id	   = mdev->pdev->device;
 946	props->hw_ver		   = mdev->pdev->revision;
 947
 948	props->max_mr_size	   = ~0ull;
 949	props->page_size_cap	   = ~(min_page_size - 1);
 950	props->max_qp		   = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
 951	props->max_qp_wr	   = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
 952	max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
 953		     sizeof(struct mlx5_wqe_data_seg);
 954	max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
 955	max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
 956		     sizeof(struct mlx5_wqe_raddr_seg)) /
 957		sizeof(struct mlx5_wqe_data_seg);
 958	props->max_send_sge = max_sq_sg;
 959	props->max_recv_sge = max_rq_sg;
 960	props->max_sge_rd	   = MLX5_MAX_SGE_RD;
 961	props->max_cq		   = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
 962	props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
 963	props->max_mr		   = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 964	props->max_pd		   = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
 965	props->max_qp_rd_atom	   = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
 966	props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
 967	props->max_srq		   = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
 968	props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
 969	props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
 970	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
 971	props->max_srq_sge	   = max_rq_sg - 1;
 972	props->max_fast_reg_page_list_len =
 973		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
 974	props->max_pi_fast_reg_page_list_len =
 975		props->max_fast_reg_page_list_len / 2;
 976	props->max_sgl_rd =
 977		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
 978	get_atomic_caps_qp(dev, props);
 979	props->masked_atomic_cap   = IB_ATOMIC_NONE;
 980	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
 981	props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
 982	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 983					   props->max_mcast_grp;
 984	props->max_ah = INT_MAX;
 985	props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
 986	props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
 987
 988	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
 989		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
 990			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
 991		props->odp_caps = dev->odp_caps;
 992		if (!uhw) {
 993			/* ODP for kernel QPs is not implemented for receive
 994			 * WQEs and SRQ WQEs
 995			 */
 996			props->odp_caps.per_transport_caps.rc_odp_caps &=
 997				~(IB_ODP_SUPPORT_READ |
 998				  IB_ODP_SUPPORT_SRQ_RECV);
 999			props->odp_caps.per_transport_caps.uc_odp_caps &=
1000				~(IB_ODP_SUPPORT_READ |
1001				  IB_ODP_SUPPORT_SRQ_RECV);
1002			props->odp_caps.per_transport_caps.ud_odp_caps &=
1003				~(IB_ODP_SUPPORT_READ |
1004				  IB_ODP_SUPPORT_SRQ_RECV);
1005			props->odp_caps.per_transport_caps.xrc_odp_caps &=
1006				~(IB_ODP_SUPPORT_READ |
1007				  IB_ODP_SUPPORT_SRQ_RECV);
1008		}
1009	}
1010
1011	if (MLX5_CAP_GEN(mdev, cd))
1012		props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1013
1014	if (mlx5_core_is_vf(mdev))
1015		props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1016
1017	if (mlx5_ib_port_link_layer(ibdev, 1) ==
1018	    IB_LINK_LAYER_ETHERNET && raw_support) {
1019		props->rss_caps.max_rwq_indirection_tables =
1020			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1021		props->rss_caps.max_rwq_indirection_table_size =
1022			1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1023		props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1024		props->max_wq_type_rq =
1025			1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1026	}
1027
1028	if (MLX5_CAP_GEN(mdev, tag_matching)) {
1029		props->tm_caps.max_num_tags =
1030			(1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1031		props->tm_caps.max_ops =
1032			1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1033		props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1034	}
1035
1036	if (MLX5_CAP_GEN(mdev, tag_matching) &&
1037	    MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1038		props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1039		props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1040	}
1041
1042	if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1043		props->cq_caps.max_cq_moderation_count =
1044						MLX5_MAX_CQ_COUNT;
1045		props->cq_caps.max_cq_moderation_period =
1046						MLX5_MAX_CQ_PERIOD;
1047	}
1048
1049	if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1050		resp.response_length += sizeof(resp.cqe_comp_caps);
1051
1052		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1053			resp.cqe_comp_caps.max_num =
1054				MLX5_CAP_GEN(dev->mdev,
1055					     cqe_compression_max_num);
1056
1057			resp.cqe_comp_caps.supported_format =
1058				MLX5_IB_CQE_RES_FORMAT_HASH |
1059				MLX5_IB_CQE_RES_FORMAT_CSUM;
1060
1061			if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1062				resp.cqe_comp_caps.supported_format |=
1063					MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1064		}
1065	}
1066
1067	if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1068	    raw_support) {
1069		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1070		    MLX5_CAP_GEN(mdev, qos)) {
1071			resp.packet_pacing_caps.qp_rate_limit_max =
1072				MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1073			resp.packet_pacing_caps.qp_rate_limit_min =
1074				MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1075			resp.packet_pacing_caps.supported_qpts |=
1076				1 << IB_QPT_RAW_PACKET;
1077			if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1078			    MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1079				resp.packet_pacing_caps.cap_flags |=
1080					MLX5_IB_PP_SUPPORT_BURST;
1081		}
1082		resp.response_length += sizeof(resp.packet_pacing_caps);
1083	}
1084
1085	if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1086	    uhw_outlen) {
1087		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1088			resp.mlx5_ib_support_multi_pkt_send_wqes =
1089				MLX5_IB_ALLOW_MPW;
1090
1091		if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1092			resp.mlx5_ib_support_multi_pkt_send_wqes |=
1093				MLX5_IB_SUPPORT_EMPW;
1094
1095		resp.response_length +=
1096			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1097	}
1098
1099	if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1100		resp.response_length += sizeof(resp.flags);
1101
1102		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1103			resp.flags |=
1104				MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1105
1106		if (MLX5_CAP_GEN(mdev, cqe_128_always))
1107			resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1108		if (MLX5_CAP_GEN(mdev, qp_packet_based))
1109			resp.flags |=
1110				MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1111
1112		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1113	}
1114
1115	if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1116		resp.response_length += sizeof(resp.sw_parsing_caps);
1117		if (MLX5_CAP_ETH(mdev, swp)) {
1118			resp.sw_parsing_caps.sw_parsing_offloads |=
1119				MLX5_IB_SW_PARSING;
1120
1121			if (MLX5_CAP_ETH(mdev, swp_csum))
1122				resp.sw_parsing_caps.sw_parsing_offloads |=
1123					MLX5_IB_SW_PARSING_CSUM;
1124
1125			if (MLX5_CAP_ETH(mdev, swp_lso))
1126				resp.sw_parsing_caps.sw_parsing_offloads |=
1127					MLX5_IB_SW_PARSING_LSO;
1128
1129			if (resp.sw_parsing_caps.sw_parsing_offloads)
1130				resp.sw_parsing_caps.supported_qpts =
1131					BIT(IB_QPT_RAW_PACKET);
1132		}
1133	}
1134
1135	if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1136	    raw_support) {
1137		resp.response_length += sizeof(resp.striding_rq_caps);
1138		if (MLX5_CAP_GEN(mdev, striding_rq)) {
1139			resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1140				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1141			resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1142				MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1143			if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1144				resp.striding_rq_caps
1145					.min_single_wqe_log_num_of_strides =
1146					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1147			else
1148				resp.striding_rq_caps
1149					.min_single_wqe_log_num_of_strides =
1150					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1151			resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1152				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1153			resp.striding_rq_caps.supported_qpts =
1154				BIT(IB_QPT_RAW_PACKET);
1155		}
1156	}
1157
1158	if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1159		resp.response_length += sizeof(resp.tunnel_offloads_caps);
1160		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1161			resp.tunnel_offloads_caps |=
1162				MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1163		if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1164			resp.tunnel_offloads_caps |=
1165				MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1166		if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1167			resp.tunnel_offloads_caps |=
1168				MLX5_IB_TUNNELED_OFFLOADS_GRE;
1169		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1170			resp.tunnel_offloads_caps |=
1171				MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1172		if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1173			resp.tunnel_offloads_caps |=
1174				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1175	}
1176
1177	if (uhw_outlen) {
1178		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1179
1180		if (err)
1181			return err;
1182	}
1183
1184	return 0;
1185}
1186
1187static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1188				   u8 *ib_width)
1189{
1190	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1191
1192	if (active_width & MLX5_PTYS_WIDTH_1X)
1193		*ib_width = IB_WIDTH_1X;
1194	else if (active_width & MLX5_PTYS_WIDTH_2X)
1195		*ib_width = IB_WIDTH_2X;
1196	else if (active_width & MLX5_PTYS_WIDTH_4X)
1197		*ib_width = IB_WIDTH_4X;
1198	else if (active_width & MLX5_PTYS_WIDTH_8X)
1199		*ib_width = IB_WIDTH_8X;
1200	else if (active_width & MLX5_PTYS_WIDTH_12X)
1201		*ib_width = IB_WIDTH_12X;
1202	else {
1203		mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1204			    active_width);
1205		*ib_width = IB_WIDTH_4X;
1206	}
1207
1208	return;
1209}
1210
1211static int mlx5_mtu_to_ib_mtu(int mtu)
1212{
1213	switch (mtu) {
1214	case 256: return 1;
1215	case 512: return 2;
1216	case 1024: return 3;
1217	case 2048: return 4;
1218	case 4096: return 5;
1219	default:
1220		pr_warn("invalid mtu\n");
1221		return -1;
1222	}
1223}
1224
1225enum ib_max_vl_num {
1226	__IB_MAX_VL_0		= 1,
1227	__IB_MAX_VL_0_1		= 2,
1228	__IB_MAX_VL_0_3		= 3,
1229	__IB_MAX_VL_0_7		= 4,
1230	__IB_MAX_VL_0_14	= 5,
1231};
1232
1233enum mlx5_vl_hw_cap {
1234	MLX5_VL_HW_0	= 1,
1235	MLX5_VL_HW_0_1	= 2,
1236	MLX5_VL_HW_0_2	= 3,
1237	MLX5_VL_HW_0_3	= 4,
1238	MLX5_VL_HW_0_4	= 5,
1239	MLX5_VL_HW_0_5	= 6,
1240	MLX5_VL_HW_0_6	= 7,
1241	MLX5_VL_HW_0_7	= 8,
1242	MLX5_VL_HW_0_14	= 15
1243};
1244
1245static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1246				u8 *max_vl_num)
1247{
1248	switch (vl_hw_cap) {
1249	case MLX5_VL_HW_0:
1250		*max_vl_num = __IB_MAX_VL_0;
1251		break;
1252	case MLX5_VL_HW_0_1:
1253		*max_vl_num = __IB_MAX_VL_0_1;
1254		break;
1255	case MLX5_VL_HW_0_3:
1256		*max_vl_num = __IB_MAX_VL_0_3;
1257		break;
1258	case MLX5_VL_HW_0_7:
1259		*max_vl_num = __IB_MAX_VL_0_7;
1260		break;
1261	case MLX5_VL_HW_0_14:
1262		*max_vl_num = __IB_MAX_VL_0_14;
1263		break;
1264
1265	default:
1266		return -EINVAL;
1267	}
1268
1269	return 0;
1270}
1271
1272static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1273			       struct ib_port_attr *props)
1274{
1275	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1276	struct mlx5_core_dev *mdev = dev->mdev;
1277	struct mlx5_hca_vport_context *rep;
1278	u16 max_mtu;
1279	u16 oper_mtu;
1280	int err;
1281	u16 ib_link_width_oper;
1282	u8 vl_hw_cap;
1283
1284	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1285	if (!rep) {
1286		err = -ENOMEM;
1287		goto out;
1288	}
1289
1290	/* props being zeroed by the caller, avoid zeroing it here */
1291
1292	err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1293	if (err)
1294		goto out;
1295
1296	props->lid		= rep->lid;
1297	props->lmc		= rep->lmc;
1298	props->sm_lid		= rep->sm_lid;
1299	props->sm_sl		= rep->sm_sl;
1300	props->state		= rep->vport_state;
1301	props->phys_state	= rep->port_physical_state;
1302	props->port_cap_flags	= rep->cap_mask1;
1303	props->gid_tbl_len	= mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1304	props->max_msg_sz	= 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1305	props->pkey_tbl_len	= mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1306	props->bad_pkey_cntr	= rep->pkey_violation_counter;
1307	props->qkey_viol_cntr	= rep->qkey_violation_counter;
1308	props->subnet_timeout	= rep->subnet_timeout;
1309	props->init_type_reply	= rep->init_type_reply;
1310
1311	if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1312		props->port_cap_flags2 = rep->cap_mask2;
1313
1314	err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1315				      &props->active_speed, port);
1316	if (err)
1317		goto out;
1318
1319	translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1320
1321	mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1322
1323	props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1324
1325	mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1326
1327	props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1328
1329	err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1330	if (err)
1331		goto out;
1332
1333	err = translate_max_vl_num(ibdev, vl_hw_cap,
1334				   &props->max_vl_num);
1335out:
1336	kfree(rep);
1337	return err;
1338}
1339
1340int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1341		       struct ib_port_attr *props)
1342{
1343	unsigned int count;
1344	int ret;
1345
1346	switch (mlx5_get_vport_access_method(ibdev)) {
1347	case MLX5_VPORT_ACCESS_METHOD_MAD:
1348		ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1349		break;
1350
1351	case MLX5_VPORT_ACCESS_METHOD_HCA:
1352		ret = mlx5_query_hca_port(ibdev, port, props);
1353		break;
1354
1355	case MLX5_VPORT_ACCESS_METHOD_NIC:
1356		ret = mlx5_query_port_roce(ibdev, port, props);
1357		break;
1358
1359	default:
1360		ret = -EINVAL;
1361	}
1362
1363	if (!ret && props) {
1364		struct mlx5_ib_dev *dev = to_mdev(ibdev);
1365		struct mlx5_core_dev *mdev;
1366		bool put_mdev = true;
1367
1368		mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1369		if (!mdev) {
1370			/* If the port isn't affiliated yet query the master.
1371			 * The master and slave will have the same values.
1372			 */
1373			mdev = dev->mdev;
1374			port = 1;
1375			put_mdev = false;
1376		}
1377		count = mlx5_core_reserved_gids_count(mdev);
1378		if (put_mdev)
1379			mlx5_ib_put_native_port_mdev(dev, port);
1380		props->gid_tbl_len -= count;
1381	}
1382	return ret;
1383}
1384
1385static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1386				  struct ib_port_attr *props)
1387{
1388	return mlx5_query_port_roce(ibdev, port, props);
1389}
1390
1391static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1392				  u16 *pkey)
1393{
1394	/* Default special Pkey for representor device port as per the
1395	 * IB specification 1.3 section 10.9.1.2.
1396	 */
1397	*pkey = 0xffff;
1398	return 0;
1399}
1400
1401static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1402			     union ib_gid *gid)
1403{
1404	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1405	struct mlx5_core_dev *mdev = dev->mdev;
1406
1407	switch (mlx5_get_vport_access_method(ibdev)) {
1408	case MLX5_VPORT_ACCESS_METHOD_MAD:
1409		return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1410
1411	case MLX5_VPORT_ACCESS_METHOD_HCA:
1412		return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1413
1414	default:
1415		return -EINVAL;
1416	}
1417
1418}
1419
1420static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1421				   u16 index, u16 *pkey)
1422{
1423	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1424	struct mlx5_core_dev *mdev;
1425	bool put_mdev = true;
1426	u32 mdev_port_num;
1427	int err;
1428
1429	mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1430	if (!mdev) {
1431		/* The port isn't affiliated yet, get the PKey from the master
1432		 * port. For RoCE the PKey tables will be the same.
1433		 */
1434		put_mdev = false;
1435		mdev = dev->mdev;
1436		mdev_port_num = 1;
1437	}
1438
1439	err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1440					index, pkey);
1441	if (put_mdev)
1442		mlx5_ib_put_native_port_mdev(dev, port);
1443
1444	return err;
1445}
1446
1447static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1448			      u16 *pkey)
1449{
1450	switch (mlx5_get_vport_access_method(ibdev)) {
1451	case MLX5_VPORT_ACCESS_METHOD_MAD:
1452		return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1453
1454	case MLX5_VPORT_ACCESS_METHOD_HCA:
1455	case MLX5_VPORT_ACCESS_METHOD_NIC:
1456		return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1457	default:
1458		return -EINVAL;
1459	}
1460}
1461
1462static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1463				 struct ib_device_modify *props)
1464{
1465	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1466	struct mlx5_reg_node_desc in;
1467	struct mlx5_reg_node_desc out;
1468	int err;
1469
1470	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1471		return -EOPNOTSUPP;
1472
1473	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1474		return 0;
1475
1476	/*
1477	 * If possible, pass node desc to FW, so it can generate
1478	 * a 144 trap.  If cmd fails, just ignore.
1479	 */
1480	memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1481	err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1482				   sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1483	if (err)
1484		return err;
1485
1486	memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1487
1488	return err;
1489}
1490
1491static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1492				u32 value)
1493{
1494	struct mlx5_hca_vport_context ctx = {};
1495	struct mlx5_core_dev *mdev;
1496	u32 mdev_port_num;
1497	int err;
1498
1499	mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1500	if (!mdev)
1501		return -ENODEV;
1502
1503	err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1504	if (err)
1505		goto out;
1506
1507	if (~ctx.cap_mask1_perm & mask) {
1508		mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1509			     mask, ctx.cap_mask1_perm);
1510		err = -EINVAL;
1511		goto out;
1512	}
1513
1514	ctx.cap_mask1 = value;
1515	ctx.cap_mask1_perm = mask;
1516	err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1517						 0, &ctx);
1518
1519out:
1520	mlx5_ib_put_native_port_mdev(dev, port_num);
1521
1522	return err;
1523}
1524
1525static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1526			       struct ib_port_modify *props)
1527{
1528	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1529	struct ib_port_attr attr;
1530	u32 tmp;
1531	int err;
1532	u32 change_mask;
1533	u32 value;
1534	bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1535		      IB_LINK_LAYER_INFINIBAND);
1536
1537	/* CM layer calls ib_modify_port() regardless of the link layer. For
1538	 * Ethernet ports, qkey violation and Port capabilities are meaningless.
1539	 */
1540	if (!is_ib)
1541		return 0;
1542
1543	if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1544		change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1545		value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1546		return set_port_caps_atomic(dev, port, change_mask, value);
1547	}
1548
1549	mutex_lock(&dev->cap_mask_mutex);
1550
1551	err = ib_query_port(ibdev, port, &attr);
1552	if (err)
1553		goto out;
1554
1555	tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1556		~props->clr_port_cap_mask;
1557
1558	err = mlx5_set_port_caps(dev->mdev, port, tmp);
1559
1560out:
1561	mutex_unlock(&dev->cap_mask_mutex);
1562	return err;
1563}
1564
1565static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1566{
1567	mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1568		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1569}
1570
1571static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1572{
1573	/* Large page with non 4k uar support might limit the dynamic size */
1574	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1575		return MLX5_MIN_DYN_BFREGS;
1576
1577	return MLX5_MAX_DYN_BFREGS;
1578}
1579
1580static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1581			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
1582			     struct mlx5_bfreg_info *bfregi)
1583{
1584	int uars_per_sys_page;
1585	int bfregs_per_sys_page;
1586	int ref_bfregs = req->total_num_bfregs;
1587
1588	if (req->total_num_bfregs == 0)
1589		return -EINVAL;
1590
1591	BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1592	BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1593
1594	if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1595		return -ENOMEM;
1596
1597	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1598	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1599	/* This holds the required static allocation asked by the user */
1600	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1601	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1602		return -EINVAL;
1603
1604	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1605	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1606	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1607	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1608
1609	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, total bfregs %d, using %d sys pages\n",
1610		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1611		    lib_uar_4k ? "yes" : "no", ref_bfregs,
1612		    req->total_num_bfregs, bfregi->total_num_bfregs,
1613		    bfregi->num_sys_pages);
1614
1615	return 0;
1616}
1617
1618static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1619{
1620	struct mlx5_bfreg_info *bfregi;
1621	int err;
1622	int i;
1623
1624	bfregi = &context->bfregi;
1625	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1626		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1627		if (err)
1628			goto error;
1629
1630		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1631	}
1632
1633	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1634		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1635
1636	return 0;
1637
1638error:
1639	for (--i; i >= 0; i--)
1640		if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1641			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1642
1643	return err;
1644}
1645
1646static void deallocate_uars(struct mlx5_ib_dev *dev,
1647			    struct mlx5_ib_ucontext *context)
1648{
1649	struct mlx5_bfreg_info *bfregi;
1650	int i;
1651
1652	bfregi = &context->bfregi;
1653	for (i = 0; i < bfregi->num_sys_pages; i++)
1654		if (i < bfregi->num_static_sys_pages ||
1655		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1656			mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1657}
1658
1659int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1660{
1661	int err = 0;
1662
1663	mutex_lock(&dev->lb.mutex);
1664	if (td)
1665		dev->lb.user_td++;
1666	if (qp)
1667		dev->lb.qps++;
1668
1669	if (dev->lb.user_td == 2 ||
1670	    dev->lb.qps == 1) {
1671		if (!dev->lb.enabled) {
1672			err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1673			dev->lb.enabled = true;
1674		}
1675	}
1676
1677	mutex_unlock(&dev->lb.mutex);
1678
1679	return err;
1680}
1681
1682void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1683{
1684	mutex_lock(&dev->lb.mutex);
1685	if (td)
1686		dev->lb.user_td--;
1687	if (qp)
1688		dev->lb.qps--;
1689
1690	if (dev->lb.user_td == 1 &&
1691	    dev->lb.qps == 0) {
1692		if (dev->lb.enabled) {
1693			mlx5_nic_vport_update_local_lb(dev->mdev, false);
1694			dev->lb.enabled = false;
1695		}
1696	}
1697
1698	mutex_unlock(&dev->lb.mutex);
1699}
1700
1701static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1702					  u16 uid)
1703{
1704	int err;
1705
1706	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1707		return 0;
1708
1709	err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1710	if (err)
1711		return err;
1712
1713	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1714	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1715	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1716		return err;
1717
1718	return mlx5_ib_enable_lb(dev, true, false);
1719}
1720
1721static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1722					     u16 uid)
1723{
1724	if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1725		return;
1726
1727	mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1728
1729	if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1730	    (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1731	     !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1732		return;
1733
1734	mlx5_ib_disable_lb(dev, true, false);
1735}
1736
1737static int set_ucontext_resp(struct ib_ucontext *uctx,
1738			     struct mlx5_ib_alloc_ucontext_resp *resp)
1739{
1740	struct ib_device *ibdev = uctx->device;
1741	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1742	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1743	struct mlx5_bfreg_info *bfregi = &context->bfregi;
1744	int err;
1745
1746	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1747		err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1748					      &resp->dump_fill_mkey);
1749		if (err)
1750			return err;
1751		resp->comp_mask |=
1752			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1753	}
1754
1755	resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1756	if (dev->wc_support)
1757		resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1758						      log_bf_reg_size);
1759	resp->cache_line_size = cache_line_size();
1760	resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1761	resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1762	resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1763	resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1764	resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1765	resp->cqe_version = context->cqe_version;
1766	resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1767				MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1768	resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1769					MLX5_CAP_GEN(dev->mdev,
1770						     num_of_uars_per_page) : 1;
1771
1772	if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1773				MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1774		if (mlx5_get_flow_namespace(dev->mdev,
1775				MLX5_FLOW_NAMESPACE_EGRESS))
1776			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1777		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1778				MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1779			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1780		if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1781			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1782		if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1783				MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1784			resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1785		/* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1786	}
1787
1788	resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1789			bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1790	resp->num_ports = dev->num_ports;
1791	resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1792				      MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1793
1794	if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1795		mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1796		resp->eth_min_inline++;
1797	}
1798
1799	if (dev->mdev->clock_info)
1800		resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1801
1802	/*
1803	 * We don't want to expose information from the PCI bar that is located
1804	 * after 4096 bytes, so if the arch only supports larger pages, let's
1805	 * pretend we don't support reading the HCA's core clock. This is also
1806	 * forced by mmap function.
1807	 */
1808	if (PAGE_SIZE <= 4096) {
1809		resp->comp_mask |=
1810			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1811		resp->hca_core_clock_offset =
1812			offsetof(struct mlx5_init_seg,
1813				 internal_timer_h) % PAGE_SIZE;
1814	}
1815
1816	if (MLX5_CAP_GEN(dev->mdev, ece_support))
1817		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1818
1819	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
1820	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
1821	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
1822		resp->comp_mask |=
1823			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
1824
1825	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1826
1827	if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1828		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1829
1830	return 0;
1831}
1832
1833static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1834				  struct ib_udata *udata)
1835{
1836	struct ib_device *ibdev = uctx->device;
1837	struct mlx5_ib_dev *dev = to_mdev(ibdev);
1838	struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1839	struct mlx5_ib_alloc_ucontext_resp resp = {};
1840	struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1841	struct mlx5_bfreg_info *bfregi;
1842	int ver;
1843	int err;
1844	size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1845				     max_cqe_version);
1846	bool lib_uar_4k;
1847	bool lib_uar_dyn;
1848
1849	if (!dev->ib_active)
1850		return -EAGAIN;
1851
1852	if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1853		ver = 0;
1854	else if (udata->inlen >= min_req_v2)
1855		ver = 2;
1856	else
1857		return -EINVAL;
1858
1859	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1860	if (err)
1861		return err;
1862
1863	if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1864		return -EOPNOTSUPP;
1865
1866	if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1867		return -EOPNOTSUPP;
1868
1869	req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1870				    MLX5_NON_FP_BFREGS_PER_UAR);
1871	if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1872		return -EINVAL;
1873
1874	lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1875	lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1876	bfregi = &context->bfregi;
1877
1878	if (lib_uar_dyn) {
1879		bfregi->lib_uar_dyn = lib_uar_dyn;
1880		goto uar_done;
1881	}
1882
1883	/* updates req->total_num_bfregs */
1884	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1885	if (err)
1886		goto out_ctx;
1887
1888	mutex_init(&bfregi->lock);
1889	bfregi->lib_uar_4k = lib_uar_4k;
1890	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1891				GFP_KERNEL);
1892	if (!bfregi->count) {
1893		err = -ENOMEM;
1894		goto out_ctx;
1895	}
1896
1897	bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1898				    sizeof(*bfregi->sys_pages),
1899				    GFP_KERNEL);
1900	if (!bfregi->sys_pages) {
1901		err = -ENOMEM;
1902		goto out_count;
1903	}
1904
1905	err = allocate_uars(dev, context);
1906	if (err)
1907		goto out_sys_pages;
1908
1909uar_done:
1910	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1911		err = mlx5_ib_devx_create(dev, true);
1912		if (err < 0)
1913			goto out_uars;
1914		context->devx_uid = err;
1915	}
1916
1917	err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1918					     context->devx_uid);
1919	if (err)
1920		goto out_devx;
1921
1922	INIT_LIST_HEAD(&context->db_page_list);
1923	mutex_init(&context->db_page_mutex);
1924
1925	context->cqe_version = min_t(__u8,
1926				 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1927				 req.max_cqe_version);
1928
1929	err = set_ucontext_resp(uctx, &resp);
1930	if (err)
1931		goto out_mdev;
1932
1933	resp.response_length = min(udata->outlen, sizeof(resp));
1934	err = ib_copy_to_udata(udata, &resp, resp.response_length);
1935	if (err)
1936		goto out_mdev;
1937
1938	bfregi->ver = ver;
1939	bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1940	context->lib_caps = req.lib_caps;
1941	print_lib_caps(dev, context->lib_caps);
1942
1943	if (mlx5_ib_lag_should_assign_affinity(dev)) {
1944		u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1945
1946		atomic_set(&context->tx_port_affinity,
1947			   atomic_add_return(
1948				   1, &dev->port[port].roce.tx_port_affinity));
1949	}
1950
1951	return 0;
1952
1953out_mdev:
1954	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1955out_devx:
1956	if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1957		mlx5_ib_devx_destroy(dev, context->devx_uid);
1958
1959out_uars:
1960	deallocate_uars(dev, context);
1961
1962out_sys_pages:
1963	kfree(bfregi->sys_pages);
1964
1965out_count:
1966	kfree(bfregi->count);
1967
1968out_ctx:
1969	return err;
1970}
1971
1972static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1973				  struct uverbs_attr_bundle *attrs)
1974{
1975	struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1976	int ret;
1977
1978	ret = set_ucontext_resp(ibcontext, &uctx_resp);
1979	if (ret)
1980		return ret;
1981
1982	uctx_resp.response_length =
1983		min_t(size_t,
1984		      uverbs_attr_get_len(attrs,
1985				MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1986		      sizeof(uctx_resp));
1987
1988	ret = uverbs_copy_to_struct_or_zero(attrs,
1989					MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1990					&uctx_resp,
1991					sizeof(uctx_resp));
1992	return ret;
1993}
1994
1995static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1996{
1997	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1998	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1999	struct mlx5_bfreg_info *bfregi;
2000
2001	bfregi = &context->bfregi;
2002	mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
2003
2004	if (context->devx_uid)
2005		mlx5_ib_devx_destroy(dev, context->devx_uid);
2006
2007	deallocate_uars(dev, context);
2008	kfree(bfregi->sys_pages);
2009	kfree(bfregi->count);
2010}
2011
2012static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2013				 int uar_idx)
2014{
2015	int fw_uars_per_page;
2016
2017	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2018
2019	return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2020}
2021
2022static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2023				 int uar_idx)
2024{
2025	unsigned int fw_uars_per_page;
2026
2027	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2028				MLX5_UARS_IN_PAGE : 1;
2029
2030	return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2031}
2032
2033static int get_command(unsigned long offset)
2034{
2035	return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2036}
2037
2038static int get_arg(unsigned long offset)
2039{
2040	return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2041}
2042
2043static int get_index(unsigned long offset)
2044{
2045	return get_arg(offset);
2046}
2047
2048/* Index resides in an extra byte to enable larger values than 255 */
2049static int get_extended_index(unsigned long offset)
2050{
2051	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2052}
2053
2054
2055static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2056{
2057}
2058
2059static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2060{
2061	switch (cmd) {
2062	case MLX5_IB_MMAP_WC_PAGE:
2063		return "WC";
2064	case MLX5_IB_MMAP_REGULAR_PAGE:
2065		return "best effort WC";
2066	case MLX5_IB_MMAP_NC_PAGE:
2067		return "NC";
2068	case MLX5_IB_MMAP_DEVICE_MEM:
2069		return "Device Memory";
2070	default:
2071		return NULL;
2072	}
2073}
2074
2075static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2076					struct vm_area_struct *vma,
2077					struct mlx5_ib_ucontext *context)
2078{
2079	if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2080	    !(vma->vm_flags & VM_SHARED))
2081		return -EINVAL;
2082
2083	if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2084		return -EOPNOTSUPP;
2085
2086	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2087		return -EPERM;
2088	vma->vm_flags &= ~VM_MAYWRITE;
2089
2090	if (!dev->mdev->clock_info)
2091		return -EOPNOTSUPP;
2092
2093	return vm_insert_page(vma, vma->vm_start,
2094			      virt_to_page(dev->mdev->clock_info));
2095}
2096
2097static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2098{
2099	struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2100	struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2101	struct mlx5_var_table *var_table = &dev->var_table;
2102
2103	switch (mentry->mmap_flag) {
2104	case MLX5_IB_MMAP_TYPE_MEMIC:
2105	case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2106		mlx5_ib_dm_mmap_free(dev, mentry);
2107		break;
2108	case MLX5_IB_MMAP_TYPE_VAR:
2109		mutex_lock(&var_table->bitmap_lock);
2110		clear_bit(mentry->page_idx, var_table->bitmap);
2111		mutex_unlock(&var_table->bitmap_lock);
2112		kfree(mentry);
2113		break;
2114	case MLX5_IB_MMAP_TYPE_UAR_WC:
2115	case MLX5_IB_MMAP_TYPE_UAR_NC:
2116		mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2117		kfree(mentry);
2118		break;
2119	default:
2120		WARN_ON(true);
2121	}
2122}
2123
2124static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2125		    struct vm_area_struct *vma,
2126		    struct mlx5_ib_ucontext *context)
2127{
2128	struct mlx5_bfreg_info *bfregi = &context->bfregi;
2129	int err;
2130	unsigned long idx;
2131	phys_addr_t pfn;
2132	pgprot_t prot;
2133	u32 bfreg_dyn_idx = 0;
2134	u32 uar_index;
2135	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2136	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2137				bfregi->num_static_sys_pages;
2138
2139	if (bfregi->lib_uar_dyn)
2140		return -EINVAL;
2141
2142	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2143		return -EINVAL;
2144
2145	if (dyn_uar)
2146		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2147	else
2148		idx = get_index(vma->vm_pgoff);
2149
2150	if (idx >= max_valid_idx) {
2151		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2152			     idx, max_valid_idx);
2153		return -EINVAL;
2154	}
2155
2156	switch (cmd) {
2157	case MLX5_IB_MMAP_WC_PAGE:
2158	case MLX5_IB_MMAP_ALLOC_WC:
2159	case MLX5_IB_MMAP_REGULAR_PAGE:
2160		/* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2161		prot = pgprot_writecombine(vma->vm_page_prot);
2162		break;
2163	case MLX5_IB_MMAP_NC_PAGE:
2164		prot = pgprot_noncached(vma->vm_page_prot);
2165		break;
2166	default:
2167		return -EINVAL;
2168	}
2169
2170	if (dyn_uar) {
2171		int uars_per_page;
2172
2173		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2174		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2175		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2176			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2177				     bfreg_dyn_idx, bfregi->total_num_bfregs);
2178			return -EINVAL;
2179		}
2180
2181		mutex_lock(&bfregi->lock);
2182		/* Fail if uar already allocated, first bfreg index of each
2183		 * page holds its count.
2184		 */
2185		if (bfregi->count[bfreg_dyn_idx]) {
2186			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2187			mutex_unlock(&bfregi->lock);
2188			return -EINVAL;
2189		}
2190
2191		bfregi->count[bfreg_dyn_idx]++;
2192		mutex_unlock(&bfregi->lock);
2193
2194		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2195		if (err) {
2196			mlx5_ib_warn(dev, "UAR alloc failed\n");
2197			goto free_bfreg;
2198		}
2199	} else {
2200		uar_index = bfregi->sys_pages[idx];
2201	}
2202
2203	pfn = uar_index2pfn(dev, uar_index);
2204	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2205
2206	err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2207				prot, NULL);
2208	if (err) {
2209		mlx5_ib_err(dev,
2210			    "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2211			    err, mmap_cmd2str(cmd));
2212		goto err;
2213	}
2214
2215	if (dyn_uar)
2216		bfregi->sys_pages[idx] = uar_index;
2217	return 0;
2218
2219err:
2220	if (!dyn_uar)
2221		return err;
2222
2223	mlx5_cmd_free_uar(dev->mdev, idx);
2224
2225free_bfreg:
2226	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2227
2228	return err;
2229}
2230
2231static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2232{
2233	unsigned long idx;
2234	u8 command;
2235
2236	command = get_command(vma->vm_pgoff);
2237	idx = get_extended_index(vma->vm_pgoff);
2238
2239	return (command << 16 | idx);
2240}
2241
2242static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2243			       struct vm_area_struct *vma,
2244			       struct ib_ucontext *ucontext)
2245{
2246	struct mlx5_user_mmap_entry *mentry;
2247	struct rdma_user_mmap_entry *entry;
2248	unsigned long pgoff;
2249	pgprot_t prot;
2250	phys_addr_t pfn;
2251	int ret;
2252
2253	pgoff = mlx5_vma_to_pgoff(vma);
2254	entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2255	if (!entry)
2256		return -EINVAL;
2257
2258	mentry = to_mmmap(entry);
2259	pfn = (mentry->address >> PAGE_SHIFT);
2260	if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2261	    mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2262		prot = pgprot_noncached(vma->vm_page_prot);
2263	else
2264		prot = pgprot_writecombine(vma->vm_page_prot);
2265	ret = rdma_user_mmap_io(ucontext, vma, pfn,
2266				entry->npages * PAGE_SIZE,
2267				prot,
2268				entry);
2269	rdma_user_mmap_entry_put(&mentry->rdma_entry);
2270	return ret;
2271}
2272
2273static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2274{
2275	u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2276	u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2277
2278	return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2279		(index & 0xFF)) << PAGE_SHIFT;
2280}
2281
2282static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2283{
2284	struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2285	struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2286	unsigned long command;
2287	phys_addr_t pfn;
2288
2289	command = get_command(vma->vm_pgoff);
2290	switch (command) {
2291	case MLX5_IB_MMAP_WC_PAGE:
2292	case MLX5_IB_MMAP_ALLOC_WC:
2293		if (!dev->wc_support)
2294			return -EPERM;
2295		fallthrough;
2296	case MLX5_IB_MMAP_NC_PAGE:
2297	case MLX5_IB_MMAP_REGULAR_PAGE:
2298		return uar_mmap(dev, command, vma, context);
2299
2300	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2301		return -ENOSYS;
2302
2303	case MLX5_IB_MMAP_CORE_CLOCK:
2304		if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2305			return -EINVAL;
2306
2307		if (vma->vm_flags & VM_WRITE)
2308			return -EPERM;
2309		vma->vm_flags &= ~VM_MAYWRITE;
2310
2311		/* Don't expose to user-space information it shouldn't have */
2312		if (PAGE_SIZE > 4096)
2313			return -EOPNOTSUPP;
2314
2315		pfn = (dev->mdev->iseg_base +
2316		       offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2317			PAGE_SHIFT;
2318		return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2319					 PAGE_SIZE,
2320					 pgprot_noncached(vma->vm_page_prot),
2321					 NULL);
2322	case MLX5_IB_MMAP_CLOCK_INFO:
2323		return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2324
2325	default:
2326		return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2327	}
2328
2329	return 0;
2330}
2331
2332static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2333{
2334	struct mlx5_ib_pd *pd = to_mpd(ibpd);
2335	struct ib_device *ibdev = ibpd->device;
2336	struct mlx5_ib_alloc_pd_resp resp;
2337	int err;
2338	u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2339	u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2340	u16 uid = 0;
2341	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2342		udata, struct mlx5_ib_ucontext, ibucontext);
2343
2344	uid = context ? context->devx_uid : 0;
2345	MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2346	MLX5_SET(alloc_pd_in, in, uid, uid);
2347	err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2348	if (err)
2349		return err;
2350
2351	pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2352	pd->uid = uid;
2353	if (udata) {
2354		resp.pdn = pd->pdn;
2355		if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2356			mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2357			return -EFAULT;
2358		}
2359	}
2360
2361	return 0;
2362}
2363
2364static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2365{
2366	struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2367	struct mlx5_ib_pd *mpd = to_mpd(pd);
2368
2369	return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2370}
2371
2372static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2373{
2374	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2375	struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2376	int err;
2377	u16 uid;
2378
2379	uid = ibqp->pd ?
2380		to_mpd(ibqp->pd)->uid : 0;
2381
2382	if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2383		mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2384		return -EOPNOTSUPP;
2385	}
2386
2387	err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2388	if (err)
2389		mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2390			     ibqp->qp_num, gid->raw);
2391
2392	return err;
2393}
2394
2395static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2396{
2397	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2398	int err;
2399	u16 uid;
2400
2401	uid = ibqp->pd ?
2402		to_mpd(ibqp->pd)->uid : 0;
2403	err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2404	if (err)
2405		mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2406			     ibqp->qp_num, gid->raw);
2407
2408	return err;
2409}
2410
2411static int init_node_data(struct mlx5_ib_dev *dev)
2412{
2413	int err;
2414
2415	err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2416	if (err)
2417		return err;
2418
2419	dev->mdev->rev_id = dev->mdev->pdev->revision;
2420
2421	return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2422}
2423
2424static ssize_t fw_pages_show(struct device *device,
2425			     struct device_attribute *attr, char *buf)
2426{
2427	struct mlx5_ib_dev *dev =
2428		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2429
2430	return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2431}
2432static DEVICE_ATTR_RO(fw_pages);
2433
2434static ssize_t reg_pages_show(struct device *device,
2435			      struct device_attribute *attr, char *buf)
2436{
2437	struct mlx5_ib_dev *dev =
2438		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2439
2440	return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2441}
2442static DEVICE_ATTR_RO(reg_pages);
2443
2444static ssize_t hca_type_show(struct device *device,
2445			     struct device_attribute *attr, char *buf)
2446{
2447	struct mlx5_ib_dev *dev =
2448		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2449
2450	return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2451}
2452static DEVICE_ATTR_RO(hca_type);
2453
2454static ssize_t hw_rev_show(struct device *device,
2455			   struct device_attribute *attr, char *buf)
2456{
2457	struct mlx5_ib_dev *dev =
2458		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2459
2460	return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2461}
2462static DEVICE_ATTR_RO(hw_rev);
2463
2464static ssize_t board_id_show(struct device *device,
2465			     struct device_attribute *attr, char *buf)
2466{
2467	struct mlx5_ib_dev *dev =
2468		rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2469
2470	return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2471			  dev->mdev->board_id);
2472}
2473static DEVICE_ATTR_RO(board_id);
2474
2475static struct attribute *mlx5_class_attributes[] = {
2476	&dev_attr_hw_rev.attr,
2477	&dev_attr_hca_type.attr,
2478	&dev_attr_board_id.attr,
2479	&dev_attr_fw_pages.attr,
2480	&dev_attr_reg_pages.attr,
2481	NULL,
2482};
2483
2484static const struct attribute_group mlx5_attr_group = {
2485	.attrs = mlx5_class_attributes,
2486};
2487
2488static void pkey_change_handler(struct work_struct *work)
2489{
2490	struct mlx5_ib_port_resources *ports =
2491		container_of(work, struct mlx5_ib_port_resources,
2492			     pkey_change_work);
2493
2494	mlx5_ib_gsi_pkey_change(ports->gsi);
2495}
2496
2497static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2498{
2499	struct mlx5_ib_qp *mqp;
2500	struct mlx5_ib_cq *send_mcq, *recv_mcq;
2501	struct mlx5_core_cq *mcq;
2502	struct list_head cq_armed_list;
2503	unsigned long flags_qp;
2504	unsigned long flags_cq;
2505	unsigned long flags;
2506
2507	INIT_LIST_HEAD(&cq_armed_list);
2508
2509	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2510	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2511	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2512		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2513		if (mqp->sq.tail != mqp->sq.head) {
2514			send_mcq = to_mcq(mqp->ibqp.send_cq);
2515			spin_lock_irqsave(&send_mcq->lock, flags_cq);
2516			if (send_mcq->mcq.comp &&
2517			    mqp->ibqp.send_cq->comp_handler) {
2518				if (!send_mcq->mcq.reset_notify_added) {
2519					send_mcq->mcq.reset_notify_added = 1;
2520					list_add_tail(&send_mcq->mcq.reset_notify,
2521						      &cq_armed_list);
2522				}
2523			}
2524			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2525		}
2526		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2527		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2528		/* no handling is needed for SRQ */
2529		if (!mqp->ibqp.srq) {
2530			if (mqp->rq.tail != mqp->rq.head) {
2531				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2532				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2533				if (recv_mcq->mcq.comp &&
2534				    mqp->ibqp.recv_cq->comp_handler) {
2535					if (!recv_mcq->mcq.reset_notify_added) {
2536						recv_mcq->mcq.reset_notify_added = 1;
2537						list_add_tail(&recv_mcq->mcq.reset_notify,
2538							      &cq_armed_list);
2539					}
2540				}
2541				spin_unlock_irqrestore(&recv_mcq->lock,
2542						       flags_cq);
2543			}
2544		}
2545		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2546	}
2547	/*At that point all inflight post send were put to be executed as of we
2548	 * lock/unlock above locks Now need to arm all involved CQs.
2549	 */
2550	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2551		mcq->comp(mcq, NULL);
2552	}
2553	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2554}
2555
2556static void delay_drop_handler(struct work_struct *work)
2557{
2558	int err;
2559	struct mlx5_ib_delay_drop *delay_drop =
2560		container_of(work, struct mlx5_ib_delay_drop,
2561			     delay_drop_work);
2562
2563	atomic_inc(&delay_drop->events_cnt);
2564
2565	mutex_lock(&delay_drop->lock);
2566	err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2567	if (err) {
2568		mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2569			     delay_drop->timeout);
2570		delay_drop->activate = false;
2571	}
2572	mutex_unlock(&delay_drop->lock);
2573}
2574
2575static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2576				 struct ib_event *ibev)
2577{
2578	u32 port = (eqe->data.port.port >> 4) & 0xf;
2579
2580	switch (eqe->sub_type) {
2581	case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2582		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2583					    IB_LINK_LAYER_ETHERNET)
2584			schedule_work(&ibdev->delay_drop.delay_drop_work);
2585		break;
2586	default: /* do nothing */
2587		return;
2588	}
2589}
2590
2591static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2592			      struct ib_event *ibev)
2593{
2594	u32 port = (eqe->data.port.port >> 4) & 0xf;
2595
2596	ibev->element.port_num = port;
2597
2598	switch (eqe->sub_type) {
2599	case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2600	case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2601	case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2602		/* In RoCE, port up/down events are handled in
2603		 * mlx5_netdev_event().
2604		 */
2605		if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2606					    IB_LINK_LAYER_ETHERNET)
2607			return -EINVAL;
2608
2609		ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2610				IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2611		break;
2612
2613	case MLX5_PORT_CHANGE_SUBTYPE_LID:
2614		ibev->event = IB_EVENT_LID_CHANGE;
2615		break;
2616
2617	case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2618		ibev->event = IB_EVENT_PKEY_CHANGE;
2619		schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2620		break;
2621
2622	case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2623		ibev->event = IB_EVENT_GID_CHANGE;
2624		break;
2625
2626	case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2627		ibev->event = IB_EVENT_CLIENT_REREGISTER;
2628		break;
2629	default:
2630		return -EINVAL;
2631	}
2632
2633	return 0;
2634}
2635
2636static void mlx5_ib_handle_event(struct work_struct *_work)
2637{
2638	struct mlx5_ib_event_work *work =
2639		container_of(_work, struct mlx5_ib_event_work, work);
2640	struct mlx5_ib_dev *ibdev;
2641	struct ib_event ibev;
2642	bool fatal = false;
2643
2644	if (work->is_slave) {
2645		ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2646		if (!ibdev)
2647			goto out;
2648	} else {
2649		ibdev = work->dev;
2650	}
2651
2652	switch (work->event) {
2653	case MLX5_DEV_EVENT_SYS_ERROR:
2654		ibev.event = IB_EVENT_DEVICE_FATAL;
2655		mlx5_ib_handle_internal_error(ibdev);
2656		ibev.element.port_num  = (u8)(unsigned long)work->param;
2657		fatal = true;
2658		break;
2659	case MLX5_EVENT_TYPE_PORT_CHANGE:
2660		if (handle_port_change(ibdev, work->param, &ibev))
2661			goto out;
2662		break;
2663	case MLX5_EVENT_TYPE_GENERAL_EVENT:
2664		handle_general_event(ibdev, work->param, &ibev);
2665		fallthrough;
2666	default:
2667		goto out;
2668	}
2669
2670	ibev.device = &ibdev->ib_dev;
2671
2672	if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2673		mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2674		goto out;
2675	}
2676
2677	if (ibdev->ib_active)
2678		ib_dispatch_event(&ibev);
2679
2680	if (fatal)
2681		ibdev->ib_active = false;
2682out:
2683	kfree(work);
2684}
2685
2686static int mlx5_ib_event(struct notifier_block *nb,
2687			 unsigned long event, void *param)
2688{
2689	struct mlx5_ib_event_work *work;
2690
2691	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2692	if (!work)
2693		return NOTIFY_DONE;
2694
2695	INIT_WORK(&work->work, mlx5_ib_handle_event);
2696	work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2697	work->is_slave = false;
2698	work->param = param;
2699	work->event = event;
2700
2701	queue_work(mlx5_ib_event_wq, &work->work);
2702
2703	return NOTIFY_OK;
2704}
2705
2706static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2707				    unsigned long event, void *param)
2708{
2709	struct mlx5_ib_event_work *work;
2710
2711	work = kmalloc(sizeof(*work), GFP_ATOMIC);
2712	if (!work)
2713		return NOTIFY_DONE;
2714
2715	INIT_WORK(&work->work, mlx5_ib_handle_event);
2716	work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2717	work->is_slave = true;
2718	work->param = param;
2719	work->event = event;
2720	queue_work(mlx5_ib_event_wq, &work->work);
2721
2722	return NOTIFY_OK;
2723}
2724
2725static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2726{
2727	struct mlx5_hca_vport_context vport_ctx;
2728	int err;
2729	int port;
2730
2731	for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
2732		dev->port_caps[port - 1].has_smi = false;
2733		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2734		    MLX5_CAP_PORT_TYPE_IB) {
2735			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2736				err = mlx5_query_hca_vport_context(dev->mdev, 0,
2737								   port, 0,
2738								   &vport_ctx);
2739				if (err) {
2740					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2741						    port, err);
2742					return err;
2743				}
2744				dev->port_caps[port - 1].has_smi =
2745					vport_ctx.has_smi;
2746			} else {
2747				dev->port_caps[port - 1].has_smi = true;
2748			}
2749		}
2750	}
2751	return 0;
2752}
2753
2754static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2755{
2756	unsigned int port;
2757
2758	rdma_for_each_port (&dev->ib_dev, port)
2759		mlx5_query_ext_port_caps(dev, port);
2760}
2761
2762static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2763{
2764	switch (umr_fence_cap) {
2765	case MLX5_CAP_UMR_FENCE_NONE:
2766		return MLX5_FENCE_MODE_NONE;
2767	case MLX5_CAP_UMR_FENCE_SMALL:
2768		return MLX5_FENCE_MODE_INITIATOR_SMALL;
2769	default:
2770		return MLX5_FENCE_MODE_STRONG_ORDERING;
2771	}
2772}
2773
2774static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2775{
2776	struct mlx5_ib_resources *devr = &dev->devr;
2777	struct ib_srq_init_attr attr;
2778	struct ib_device *ibdev;
2779	struct ib_cq_init_attr cq_attr = {.cqe = 1};
2780	int port;
2781	int ret = 0;
2782
2783	ibdev = &dev->ib_dev;
2784
2785	if (!MLX5_CAP_GEN(dev->mdev, xrc))
2786		return -EOPNOTSUPP;
2787
2788	mutex_init(&devr->mutex);
2789
2790	devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
2791	if (!devr->p0)
2792		return -ENOMEM;
2793
2794	devr->p0->device  = ibdev;
2795	devr->p0->uobject = NULL;
2796	atomic_set(&devr->p0->usecnt, 0);
2797
2798	ret = mlx5_ib_alloc_pd(devr->p0, NULL);
2799	if (ret)
2800		goto error0;
2801
2802	devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
2803	if (!devr->c0) {
2804		ret = -ENOMEM;
2805		goto error1;
2806	}
2807
2808	devr->c0->device = &dev->ib_dev;
2809	atomic_set(&devr->c0->usecnt, 0);
2810
2811	ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
2812	if (ret)
2813		goto err_create_cq;
2814
2815	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2816	if (ret)
2817		goto error2;
2818
2819	ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2820	if (ret)
2821		goto error3;
2822
2823	memset(&attr, 0, sizeof(attr));
2824	attr.attr.max_sge = 1;
2825	attr.attr.max_wr = 1;
2826	attr.srq_type = IB_SRQT_XRC;
2827	attr.ext.cq = devr->c0;
2828
2829	devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2830	if (!devr->s0) {
2831		ret = -ENOMEM;
2832		goto error4;
2833	}
2834
2835	devr->s0->device	= &dev->ib_dev;
2836	devr->s0->pd		= devr->p0;
2837	devr->s0->srq_type      = IB_SRQT_XRC;
2838	devr->s0->ext.cq	= devr->c0;
2839	ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
2840	if (ret)
2841		goto err_create;
2842
2843	atomic_inc(&devr->s0->ext.cq->usecnt);
2844	atomic_inc(&devr->p0->usecnt);
2845	atomic_set(&devr->s0->usecnt, 0);
2846
2847	memset(&attr, 0, sizeof(attr));
2848	attr.attr.max_sge = 1;
2849	attr.attr.max_wr = 1;
2850	attr.srq_type = IB_SRQT_BASIC;
2851	devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2852	if (!devr->s1) {
2853		ret = -ENOMEM;
2854		goto error5;
2855	}
2856
2857	devr->s1->device	= &dev->ib_dev;
2858	devr->s1->pd		= devr->p0;
2859	devr->s1->srq_type      = IB_SRQT_BASIC;
2860	devr->s1->ext.cq	= devr->c0;
2861
2862	ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
2863	if (ret)
2864		goto error6;
2865
2866	atomic_inc(&devr->p0->usecnt);
2867	atomic_set(&devr->s1->usecnt, 0);
2868
2869	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2870		INIT_WORK(&devr->ports[port].pkey_change_work,
2871			  pkey_change_handler);
2872
2873	return 0;
2874
2875error6:
2876	kfree(devr->s1);
2877error5:
2878	mlx5_ib_destroy_srq(devr->s0, NULL);
2879err_create:
2880	kfree(devr->s0);
2881error4:
2882	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2883error3:
2884	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2885error2:
2886	mlx5_ib_destroy_cq(devr->c0, NULL);
2887err_create_cq:
2888	kfree(devr->c0);
2889error1:
2890	mlx5_ib_dealloc_pd(devr->p0, NULL);
2891error0:
2892	kfree(devr->p0);
2893	return ret;
2894}
2895
2896static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2897{
2898	struct mlx5_ib_resources *devr = &dev->devr;
2899	int port;
2900
2901	mlx5_ib_destroy_srq(devr->s1, NULL);
2902	kfree(devr->s1);
2903	mlx5_ib_destroy_srq(devr->s0, NULL);
2904	kfree(devr->s0);
2905	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2906	mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2907	mlx5_ib_destroy_cq(devr->c0, NULL);
2908	kfree(devr->c0);
2909	mlx5_ib_dealloc_pd(devr->p0, NULL);
2910	kfree(devr->p0);
2911
2912	/* Make sure no change P_Key work items are still executing */
2913	for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2914		cancel_work_sync(&devr->ports[port].pkey_change_work);
2915}
2916
2917static u32 get_core_cap_flags(struct ib_device *ibdev,
2918			      struct mlx5_hca_vport_context *rep)
2919{
2920	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2921	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2922	u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2923	u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2924	bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2925	u32 ret = 0;
2926
2927	if (rep->grh_required)
2928		ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2929
2930	if (ll == IB_LINK_LAYER_INFINIBAND)
2931		return ret | RDMA_CORE_PORT_IBA_IB;
2932
2933	if (raw_support)
2934		ret |= RDMA_CORE_PORT_RAW_PACKET;
2935
2936	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2937		return ret;
2938
2939	if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2940		return ret;
2941
2942	if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2943		ret |= RDMA_CORE_PORT_IBA_ROCE;
2944
2945	if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2946		ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2947
2948	return ret;
2949}
2950
2951static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2952			       struct ib_port_immutable *immutable)
2953{
2954	struct ib_port_attr attr;
2955	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2956	enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2957	struct mlx5_hca_vport_context rep = {0};
2958	int err;
2959
2960	err = ib_query_port(ibdev, port_num, &attr);
2961	if (err)
2962		return err;
2963
2964	if (ll == IB_LINK_LAYER_INFINIBAND) {
2965		err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2966						   &rep);
2967		if (err)
2968			return err;
2969	}
2970
2971	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2972	immutable->gid_tbl_len = attr.gid_tbl_len;
2973	immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2974	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2975
2976	return 0;
2977}
2978
2979static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2980				   struct ib_port_immutable *immutable)
2981{
2982	struct ib_port_attr attr;
2983	int err;
2984
2985	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2986
2987	err = ib_query_port(ibdev, port_num, &attr);
2988	if (err)
2989		return err;
2990
2991	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2992	immutable->gid_tbl_len = attr.gid_tbl_len;
2993	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2994
2995	return 0;
2996}
2997
2998static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2999{
3000	struct mlx5_ib_dev *dev =
3001		container_of(ibdev, struct mlx5_ib_dev, ib_dev);
3002	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
3003		 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
3004		 fw_rev_sub(dev->mdev));
3005}
3006
3007static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3008{
3009	struct mlx5_core_dev *mdev = dev->mdev;
3010	struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3011								 MLX5_FLOW_NAMESPACE_LAG);
3012	struct mlx5_flow_table *ft;
3013	int err;
3014
3015	if (!ns || !mlx5_lag_is_roce(mdev))
3016		return 0;
3017
3018	err = mlx5_cmd_create_vport_lag(mdev);
3019	if (err)
3020		return err;
3021
3022	ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3023	if (IS_ERR(ft)) {
3024		err = PTR_ERR(ft);
3025		goto err_destroy_vport_lag;
3026	}
3027
3028	dev->flow_db->lag_demux_ft = ft;
3029	dev->lag_active = true;
3030	return 0;
3031
3032err_destroy_vport_lag:
3033	mlx5_cmd_destroy_vport_lag(mdev);
3034	return err;
3035}
3036
3037static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3038{
3039	struct mlx5_core_dev *mdev = dev->mdev;
3040
3041	if (dev->lag_active) {
3042		dev->lag_active = false;
3043
3044		mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3045		dev->flow_db->lag_demux_ft = NULL;
3046
3047		mlx5_cmd_destroy_vport_lag(mdev);
3048	}
3049}
3050
3051static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3052{
3053	int err;
3054
3055	dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3056	err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3057	if (err) {
3058		dev->port[port_num].roce.nb.notifier_call = NULL;
3059		return err;
3060	}
3061
3062	return 0;
3063}
3064
3065static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3066{
3067	if (dev->port[port_num].roce.nb.notifier_call) {
3068		unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3069		dev->port[port_num].roce.nb.notifier_call = NULL;
3070	}
3071}
3072
3073static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3074{
3075	int err;
3076
3077	err = mlx5_nic_vport_enable_roce(dev->mdev);
3078	if (err)
3079		return err;
3080
3081	err = mlx5_eth_lag_init(dev);
3082	if (err)
3083		goto err_disable_roce;
3084
3085	return 0;
3086
3087err_disable_roce:
3088	mlx5_nic_vport_disable_roce(dev->mdev);
3089
3090	return err;
3091}
3092
3093static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3094{
3095	mlx5_eth_lag_cleanup(dev);
3096	mlx5_nic_vport_disable_roce(dev->mdev);
3097}
3098
3099static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3100				 enum rdma_netdev_t type,
3101				 struct rdma_netdev_alloc_params *params)
3102{
3103	if (type != RDMA_NETDEV_IPOIB)
3104		return -EOPNOTSUPP;
3105
3106	return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3107}
3108
3109static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3110				       size_t count, loff_t *pos)
3111{
3112	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3113	char lbuf[20];
3114	int len;
3115
3116	len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3117	return simple_read_from_buffer(buf, count, pos, lbuf, len);
3118}
3119
3120static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3121					size_t count, loff_t *pos)
3122{
3123	struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3124	u32 timeout;
3125	u32 var;
3126
3127	if (kstrtouint_from_user(buf, count, 0, &var))
3128		return -EFAULT;
3129
3130	timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3131			1000);
3132	if (timeout != var)
3133		mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3134			    timeout);
3135
3136	delay_drop->timeout = timeout;
3137
3138	return count;
3139}
3140
3141static const struct file_operations fops_delay_drop_timeout = {
3142	.owner	= THIS_MODULE,
3143	.open	= simple_open,
3144	.write	= delay_drop_timeout_write,
3145	.read	= delay_drop_timeout_read,
3146};
3147
3148static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3149				      struct mlx5_ib_multiport_info *mpi)
3150{
3151	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3152	struct mlx5_ib_port *port = &ibdev->port[port_num];
3153	int comps;
3154	int err;
3155	int i;
3156
3157	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3158
3159	mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3160
3161	spin_lock(&port->mp.mpi_lock);
3162	if (!mpi->ibdev) {
3163		spin_unlock(&port->mp.mpi_lock);
3164		return;
3165	}
3166
3167	mpi->ibdev = NULL;
3168
3169	spin_unlock(&port->mp.mpi_lock);
3170	if (mpi->mdev_events.notifier_call)
3171		mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3172	mpi->mdev_events.notifier_call = NULL;
3173	mlx5_remove_netdev_notifier(ibdev, port_num);
3174	spin_lock(&port->mp.mpi_lock);
3175
3176	comps = mpi->mdev_refcnt;
3177	if (comps) {
3178		mpi->unaffiliate = true;
3179		init_completion(&mpi->unref_comp);
3180		spin_unlock(&port->mp.mpi_lock);
3181
3182		for (i = 0; i < comps; i++)
3183			wait_for_completion(&mpi->unref_comp);
3184
3185		spin_lock(&port->mp.mpi_lock);
3186		mpi->unaffiliate = false;
3187	}
3188
3189	port->mp.mpi = NULL;
3190
3191	spin_unlock(&port->mp.mpi_lock);
3192
3193	err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3194
3195	mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3196	/* Log an error, still needed to cleanup the pointers and add
3197	 * it back to the list.
3198	 */
3199	if (err)
3200		mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3201			    port_num + 1);
3202
3203	ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3204}
3205
3206static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3207				    struct mlx5_ib_multiport_info *mpi)
3208{
3209	u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3210	int err;
3211
3212	lockdep_assert_held(&mlx5_ib_multiport_mutex);
3213
3214	spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3215	if (ibdev->port[port_num].mp.mpi) {
3216		mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3217			    port_num + 1);
3218		spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3219		return false;
3220	}
3221
3222	ibdev->port[port_num].mp.mpi = mpi;
3223	mpi->ibdev = ibdev;
3224	mpi->mdev_events.notifier_call = NULL;
3225	spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3226
3227	err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3228	if (err)
3229		goto unbind;
3230
3231	err = mlx5_add_netdev_notifier(ibdev, port_num);
3232	if (err) {
3233		mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3234			    port_num + 1);
3235		goto unbind;
3236	}
3237
3238	mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3239	mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3240
3241	mlx5_ib_init_cong_debugfs(ibdev, port_num);
3242
3243	return true;
3244
3245unbind:
3246	mlx5_ib_unbind_slave_port(ibdev, mpi);
3247	return false;
3248}
3249
3250static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3251{
3252	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3253	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3254							  port_num + 1);
3255	struct mlx5_ib_multiport_info *mpi;
3256	int err;
3257	u32 i;
3258
3259	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3260		return 0;
3261
3262	err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3263						     &dev->sys_image_guid);
3264	if (err)
3265		return err;
3266
3267	err = mlx5_nic_vport_enable_roce(dev->mdev);
3268	if (err)
3269		return err;
3270
3271	mutex_lock(&mlx5_ib_multiport_mutex);
3272	for (i = 0; i < dev->num_ports; i++) {
3273		bool bound = false;
3274
3275		/* build a stub multiport info struct for the native port. */
3276		if (i == port_num) {
3277			mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3278			if (!mpi) {
3279				mutex_unlock(&mlx5_ib_multiport_mutex);
3280				mlx5_nic_vport_disable_roce(dev->mdev);
3281				return -ENOMEM;
3282			}
3283
3284			mpi->is_master = true;
3285			mpi->mdev = dev->mdev;
3286			mpi->sys_image_guid = dev->sys_image_guid;
3287			dev->port[i].mp.mpi = mpi;
3288			mpi->ibdev = dev;
3289			mpi = NULL;
3290			continue;
3291		}
3292
3293		list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3294				    list) {
3295			if (dev->sys_image_guid == mpi->sys_image_guid &&
3296			    (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3297				bound = mlx5_ib_bind_slave_port(dev, mpi);
3298			}
3299
3300			if (bound) {
3301				dev_dbg(mpi->mdev->device,
3302					"removing port from unaffiliated list.\n");
3303				mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3304				list_del(&mpi->list);
3305				break;
3306			}
3307		}
3308		if (!bound)
3309			mlx5_ib_dbg(dev, "no free port found for port %d\n",
3310				    i + 1);
3311	}
3312
3313	list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3314	mutex_unlock(&mlx5_ib_multiport_mutex);
3315	return err;
3316}
3317
3318static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3319{
3320	u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3321	enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3322							  port_num + 1);
3323	u32 i;
3324
3325	if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3326		return;
3327
3328	mutex_lock(&mlx5_ib_multiport_mutex);
3329	for (i = 0; i < dev->num_ports; i++) {
3330		if (dev->port[i].mp.mpi) {
3331			/* Destroy the native port stub */
3332			if (i == port_num) {
3333				kfree(dev->port[i].mp.mpi);
3334				dev->port[i].mp.mpi = NULL;
3335			} else {
3336				mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3337					    i + 1);
3338				list_add_tail(&dev->port[i].mp.mpi->list,
3339					      &mlx5_ib_unaffiliated_port_list);
3340				mlx5_ib_unbind_slave_port(dev,
3341							  dev->port[i].mp.mpi);
3342			}
3343		}
3344	}
3345
3346	mlx5_ib_dbg(dev, "removing from devlist\n");
3347	list_del(&dev->ib_dev_list);
3348	mutex_unlock(&mlx5_ib_multiport_mutex);
3349
3350	mlx5_nic_vport_disable_roce(dev->mdev);
3351}
3352
3353static int mmap_obj_cleanup(struct ib_uobject *uobject,
3354			    enum rdma_remove_reason why,
3355			    struct uverbs_attr_bundle *attrs)
3356{
3357	struct mlx5_user_mmap_entry *obj = uobject->object;
3358
3359	rdma_user_mmap_entry_remove(&obj->rdma_entry);
3360	return 0;
3361}
3362
3363static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3364					    struct mlx5_user_mmap_entry *entry,
3365					    size_t length)
3366{
3367	return rdma_user_mmap_entry_insert_range(
3368		&c->ibucontext, &entry->rdma_entry, length,
3369		(MLX5_IB_MMAP_OFFSET_START << 16),
3370		((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3371}
3372
3373static struct mlx5_user_mmap_entry *
3374alloc_var_entry(struct mlx5_ib_ucontext *c)
3375{
3376	struct mlx5_user_mmap_entry *entry;
3377	struct mlx5_var_table *var_table;
3378	u32 page_idx;
3379	int err;
3380
3381	var_table = &to_mdev(c->ibucontext.device)->var_table;
3382	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3383	if (!entry)
3384		return ERR_PTR(-ENOMEM);
3385
3386	mutex_lock(&var_table->bitmap_lock);
3387	page_idx = find_first_zero_bit(var_table->bitmap,
3388				       var_table->num_var_hw_entries);
3389	if (page_idx >= var_table->num_var_hw_entries) {
3390		err = -ENOSPC;
3391		mutex_unlock(&var_table->bitmap_lock);
3392		goto end;
3393	}
3394
3395	set_bit(page_idx, var_table->bitmap);
3396	mutex_unlock(&var_table->bitmap_lock);
3397
3398	entry->address = var_table->hw_start_addr +
3399				(page_idx * var_table->stride_size);
3400	entry->page_idx = page_idx;
3401	entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3402
3403	err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3404					       var_table->stride_size);
3405	if (err)
3406		goto err_insert;
3407
3408	return entry;
3409
3410err_insert:
3411	mutex_lock(&var_table->bitmap_lock);
3412	clear_bit(page_idx, var_table->bitmap);
3413	mutex_unlock(&var_table->bitmap_lock);
3414end:
3415	kfree(entry);
3416	return ERR_PTR(err);
3417}
3418
3419static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3420	struct uverbs_attr_bundle *attrs)
3421{
3422	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3423		attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3424	struct mlx5_ib_ucontext *c;
3425	struct mlx5_user_mmap_entry *entry;
3426	u64 mmap_offset;
3427	u32 length;
3428	int err;
3429
3430	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3431	if (IS_ERR(c))
3432		return PTR_ERR(c);
3433
3434	entry = alloc_var_entry(c);
3435	if (IS_ERR(entry))
3436		return PTR_ERR(entry);
3437
3438	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3439	length = entry->rdma_entry.npages * PAGE_SIZE;
3440	uobj->object = entry;
3441	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3442
3443	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3444			     &mmap_offset, sizeof(mmap_offset));
3445	if (err)
3446		return err;
3447
3448	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3449			     &entry->page_idx, sizeof(entry->page_idx));
3450	if (err)
3451		return err;
3452
3453	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3454			     &length, sizeof(length));
3455	return err;
3456}
3457
3458DECLARE_UVERBS_NAMED_METHOD(
3459	MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3460	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3461			MLX5_IB_OBJECT_VAR,
3462			UVERBS_ACCESS_NEW,
3463			UA_MANDATORY),
3464	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3465			   UVERBS_ATTR_TYPE(u32),
3466			   UA_MANDATORY),
3467	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3468			   UVERBS_ATTR_TYPE(u32),
3469			   UA_MANDATORY),
3470	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3471			    UVERBS_ATTR_TYPE(u64),
3472			    UA_MANDATORY));
3473
3474DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3475	MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3476	UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3477			MLX5_IB_OBJECT_VAR,
3478			UVERBS_ACCESS_DESTROY,
3479			UA_MANDATORY));
3480
3481DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3482			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3483			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3484			    &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3485
3486static bool var_is_supported(struct ib_device *device)
3487{
3488	struct mlx5_ib_dev *dev = to_mdev(device);
3489
3490	return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3491			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3492}
3493
3494static struct mlx5_user_mmap_entry *
3495alloc_uar_entry(struct mlx5_ib_ucontext *c,
3496		enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3497{
3498	struct mlx5_user_mmap_entry *entry;
3499	struct mlx5_ib_dev *dev;
3500	u32 uar_index;
3501	int err;
3502
3503	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3504	if (!entry)
3505		return ERR_PTR(-ENOMEM);
3506
3507	dev = to_mdev(c->ibucontext.device);
3508	err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3509	if (err)
3510		goto end;
3511
3512	entry->page_idx = uar_index;
3513	entry->address = uar_index2paddress(dev, uar_index);
3514	if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3515		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3516	else
3517		entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3518
3519	err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3520	if (err)
3521		goto err_insert;
3522
3523	return entry;
3524
3525err_insert:
3526	mlx5_cmd_free_uar(dev->mdev, uar_index);
3527end:
3528	kfree(entry);
3529	return ERR_PTR(err);
3530}
3531
3532static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3533	struct uverbs_attr_bundle *attrs)
3534{
3535	struct ib_uobject *uobj = uverbs_attr_get_uobject(
3536		attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3537	enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3538	struct mlx5_ib_ucontext *c;
3539	struct mlx5_user_mmap_entry *entry;
3540	u64 mmap_offset;
3541	u32 length;
3542	int err;
3543
3544	c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3545	if (IS_ERR(c))
3546		return PTR_ERR(c);
3547
3548	err = uverbs_get_const(&alloc_type, attrs,
3549			       MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3550	if (err)
3551		return err;
3552
3553	if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3554	    alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3555		return -EOPNOTSUPP;
3556
3557	if (!to_mdev(c->ibucontext.device)->wc_support &&
3558	    alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3559		return -EOPNOTSUPP;
3560
3561	entry = alloc_uar_entry(c, alloc_type);
3562	if (IS_ERR(entry))
3563		return PTR_ERR(entry);
3564
3565	mmap_offset = mlx5_entry_to_mmap_offset(entry);
3566	length = entry->rdma_entry.npages * PAGE_SIZE;
3567	uobj->object = entry;
3568	uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3569
3570	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3571			     &mmap_offset, sizeof(mmap_offset));
3572	if (err)
3573		return err;
3574
3575	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3576			     &entry->page_idx, sizeof(entry->page_idx));
3577	if (err)
3578		return err;
3579
3580	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3581			     &length, sizeof(length));
3582	return err;
3583}
3584
3585DECLARE_UVERBS_NAMED_METHOD(
3586	MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3587	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3588			MLX5_IB_OBJECT_UAR,
3589			UVERBS_ACCESS_NEW,
3590			UA_MANDATORY),
3591	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3592			     enum mlx5_ib_uapi_uar_alloc_type,
3593			     UA_MANDATORY),
3594	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3595			   UVERBS_ATTR_TYPE(u32),
3596			   UA_MANDATORY),
3597	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3598			   UVERBS_ATTR_TYPE(u32),
3599			   UA_MANDATORY),
3600	UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3601			    UVERBS_ATTR_TYPE(u64),
3602			    UA_MANDATORY));
3603
3604DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3605	MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3606	UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3607			MLX5_IB_OBJECT_UAR,
3608			UVERBS_ACCESS_DESTROY,
3609			UA_MANDATORY));
3610
3611DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3612			    UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3613			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3614			    &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3615
3616ADD_UVERBS_ATTRIBUTES_SIMPLE(
3617	mlx5_ib_flow_action,
3618	UVERBS_OBJECT_FLOW_ACTION,
3619	UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3620	UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3621			     enum mlx5_ib_uapi_flow_action_flags));
3622
3623ADD_UVERBS_ATTRIBUTES_SIMPLE(
3624	mlx5_ib_query_context,
3625	UVERBS_OBJECT_DEVICE,
3626	UVERBS_METHOD_QUERY_CONTEXT,
3627	UVERBS_ATTR_PTR_OUT(
3628		MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3629		UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3630				   dump_fill_mkey),
3631		UA_MANDATORY));
3632
3633static const struct uapi_definition mlx5_ib_defs[] = {
3634	UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3635	UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3636	UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3637	UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3638	UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3639
3640	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3641				&mlx5_ib_flow_action),
3642	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3643	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3644				UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3645	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3646	{}
3647};
3648
3649static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3650{
3651	mlx5_ib_cleanup_multiport_master(dev);
3652	WARN_ON(!xa_empty(&dev->odp_mkeys));
3653	mutex_destroy(&dev->cap_mask_mutex);
3654	WARN_ON(!xa_empty(&dev->sig_mrs));
3655	WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3656}
3657
3658static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3659{
3660	struct mlx5_core_dev *mdev = dev->mdev;
3661	int err;
3662	int i;
3663
3664	dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3665	dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3666	dev->ib_dev.phys_port_cnt = dev->num_ports;
3667	dev->ib_dev.dev.parent = mdev->device;
3668	dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3669
3670	for (i = 0; i < dev->num_ports; i++) {
3671		spin_lock_init(&dev->port[i].mp.mpi_lock);
3672		rwlock_init(&dev->port[i].roce.netdev_lock);
3673		dev->port[i].roce.dev = dev;
3674		dev->port[i].roce.native_port_num = i + 1;
3675		dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3676	}
3677
3678	err = mlx5_ib_init_multiport_master(dev);
3679	if (err)
3680		return err;
3681
3682	err = set_has_smi_cap(dev);
3683	if (err)
3684		goto err_mp;
3685
3686	err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3687	if (err)
3688		goto err_mp;
3689
3690	if (mlx5_use_mad_ifc(dev))
3691		get_ext_port_caps(dev);
3692
3693	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3694
3695	mutex_init(&dev->cap_mask_mutex);
3696	INIT_LIST_HEAD(&dev->qp_list);
3697	spin_lock_init(&dev->reset_flow_resource_lock);
3698	xa_init(&dev->odp_mkeys);
3699	xa_init(&dev->sig_mrs);
3700	atomic_set(&dev->mkey_var, 0);
3701
3702	spin_lock_init(&dev->dm.lock);
3703	dev->dm.dev = mdev;
3704	return 0;
3705
3706err_mp:
3707	mlx5_ib_cleanup_multiport_master(dev);
3708	return err;
3709}
3710
3711static int mlx5_ib_enable_driver(struct ib_device *dev)
3712{
3713	struct mlx5_ib_dev *mdev = to_mdev(dev);
3714	int ret;
3715
3716	ret = mlx5_ib_test_wc(mdev);
3717	mlx5_ib_dbg(mdev, "Write-Combining %s",
3718		    mdev->wc_support ? "supported" : "not supported");
3719
3720	return ret;
3721}
3722
3723static const struct ib_device_ops mlx5_ib_dev_ops = {
3724	.owner = THIS_MODULE,
3725	.driver_id = RDMA_DRIVER_MLX5,
3726	.uverbs_abi_ver	= MLX5_IB_UVERBS_ABI_VERSION,
3727
3728	.add_gid = mlx5_ib_add_gid,
3729	.alloc_mr = mlx5_ib_alloc_mr,
3730	.alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3731	.alloc_pd = mlx5_ib_alloc_pd,
3732	.alloc_ucontext = mlx5_ib_alloc_ucontext,
3733	.attach_mcast = mlx5_ib_mcg_attach,
3734	.check_mr_status = mlx5_ib_check_mr_status,
3735	.create_ah = mlx5_ib_create_ah,
3736	.create_cq = mlx5_ib_create_cq,
3737	.create_qp = mlx5_ib_create_qp,
3738	.create_srq = mlx5_ib_create_srq,
3739	.create_user_ah = mlx5_ib_create_ah,
3740	.dealloc_pd = mlx5_ib_dealloc_pd,
3741	.dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3742	.del_gid = mlx5_ib_del_gid,
3743	.dereg_mr = mlx5_ib_dereg_mr,
3744	.destroy_ah = mlx5_ib_destroy_ah,
3745	.destroy_cq = mlx5_ib_destroy_cq,
3746	.destroy_qp = mlx5_ib_destroy_qp,
3747	.destroy_srq = mlx5_ib_destroy_srq,
3748	.detach_mcast = mlx5_ib_mcg_detach,
3749	.disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3750	.drain_rq = mlx5_ib_drain_rq,
3751	.drain_sq = mlx5_ib_drain_sq,
3752	.device_group = &mlx5_attr_group,
3753	.enable_driver = mlx5_ib_enable_driver,
3754	.get_dev_fw_str = get_dev_fw_str,
3755	.get_dma_mr = mlx5_ib_get_dma_mr,
3756	.get_link_layer = mlx5_ib_port_link_layer,
3757	.map_mr_sg = mlx5_ib_map_mr_sg,
3758	.map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3759	.mmap = mlx5_ib_mmap,
3760	.mmap_free = mlx5_ib_mmap_free,
3761	.modify_cq = mlx5_ib_modify_cq,
3762	.modify_device = mlx5_ib_modify_device,
3763	.modify_port = mlx5_ib_modify_port,
3764	.modify_qp = mlx5_ib_modify_qp,
3765	.modify_srq = mlx5_ib_modify_srq,
3766	.poll_cq = mlx5_ib_poll_cq,
3767	.post_recv = mlx5_ib_post_recv_nodrain,
3768	.post_send = mlx5_ib_post_send_nodrain,
3769	.post_srq_recv = mlx5_ib_post_srq_recv,
3770	.process_mad = mlx5_ib_process_mad,
3771	.query_ah = mlx5_ib_query_ah,
3772	.query_device = mlx5_ib_query_device,
3773	.query_gid = mlx5_ib_query_gid,
3774	.query_pkey = mlx5_ib_query_pkey,
3775	.query_qp = mlx5_ib_query_qp,
3776	.query_srq = mlx5_ib_query_srq,
3777	.query_ucontext = mlx5_ib_query_ucontext,
3778	.reg_user_mr = mlx5_ib_reg_user_mr,
3779	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3780	.req_notify_cq = mlx5_ib_arm_cq,
3781	.rereg_user_mr = mlx5_ib_rereg_user_mr,
3782	.resize_cq = mlx5_ib_resize_cq,
3783
3784	INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3785	INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3786	INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3787	INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3788	INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3789	INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3790};
3791
3792static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3793	.rdma_netdev_get_params = mlx5_ib_rn_get_params,
3794};
3795
3796static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3797	.get_vf_config = mlx5_ib_get_vf_config,
3798	.get_vf_guid = mlx5_ib_get_vf_guid,
3799	.get_vf_stats = mlx5_ib_get_vf_stats,
3800	.set_vf_guid = mlx5_ib_set_vf_guid,
3801	.set_vf_link_state = mlx5_ib_set_vf_link_state,
3802};
3803
3804static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3805	.alloc_mw = mlx5_ib_alloc_mw,
3806	.dealloc_mw = mlx5_ib_dealloc_mw,
3807
3808	INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3809};
3810
3811static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3812	.alloc_xrcd = mlx5_ib_alloc_xrcd,
3813	.dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3814
3815	INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3816};
3817
3818static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3819{
3820	struct mlx5_core_dev *mdev = dev->mdev;
3821	struct mlx5_var_table *var_table = &dev->var_table;
3822	u8 log_doorbell_bar_size;
3823	u8 log_doorbell_stride;
3824	u64 bar_size;
3825
3826	log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3827					log_doorbell_bar_size);
3828	log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3829					log_doorbell_stride);
3830	var_table->hw_start_addr = dev->mdev->bar_addr +
3831				MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3832					doorbell_bar_offset);
3833	bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3834	var_table->stride_size = 1ULL << log_doorbell_stride;
3835	var_table->num_var_hw_entries = div_u64(bar_size,
3836						var_table->stride_size);
3837	mutex_init(&var_table->bitmap_lock);
3838	var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3839					  GFP_KERNEL);
3840	return (var_table->bitmap) ? 0 : -ENOMEM;
3841}
3842
3843static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3844{
3845	bitmap_free(dev->var_table.bitmap);
3846}
3847
3848static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3849{
3850	struct mlx5_core_dev *mdev = dev->mdev;
3851	int err;
3852
3853	if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3854	    IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3855		ib_set_device_ops(&dev->ib_dev,
3856				  &mlx5_ib_dev_ipoib_enhanced_ops);
3857
3858	if (mlx5_core_is_pf(mdev))
3859		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3860
3861	dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3862
3863	if (MLX5_CAP_GEN(mdev, imaicl))
3864		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3865
3866	if (MLX5_CAP_GEN(mdev, xrc))
3867		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3868
3869	if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3870	    MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3871	    MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3872		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3873
3874	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3875
3876	if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3877		dev->ib_dev.driver_def = mlx5_ib_defs;
3878
3879	err = init_node_data(dev);
3880	if (err)
3881		return err;
3882
3883	if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3884	    (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3885	     MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3886		mutex_init(&dev->lb.mutex);
3887
3888	if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3889			MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3890		err = mlx5_ib_init_var_table(dev);
3891		if (err)
3892			return err;
3893	}
3894
3895	dev->ib_dev.use_cq_dim = true;
3896
3897	return 0;
3898}
3899
3900static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3901	.get_port_immutable = mlx5_port_immutable,
3902	.query_port = mlx5_ib_query_port,
3903};
3904
3905static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3906{
3907	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3908	return 0;
3909}
3910
3911static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3912	.get_port_immutable = mlx5_port_rep_immutable,
3913	.query_port = mlx5_ib_rep_query_port,
3914	.query_pkey = mlx5_ib_rep_query_pkey,
3915};
3916
3917static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3918{
3919	ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3920	return 0;
3921}
3922
3923static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3924	.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3925	.create_wq = mlx5_ib_create_wq,
3926	.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3927	.destroy_wq = mlx5_ib_destroy_wq,
3928	.get_netdev = mlx5_ib_get_netdev,
3929	.modify_wq = mlx5_ib_modify_wq,
3930
3931	INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3932			   ib_rwq_ind_tbl),
3933};
3934
3935static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3936{
3937	struct mlx5_core_dev *mdev = dev->mdev;
3938	enum rdma_link_layer ll;
3939	int port_type_cap;
3940	u32 port_num = 0;
3941	int err;
3942
3943	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3944	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3945
3946	if (ll == IB_LINK_LAYER_ETHERNET) {
3947		ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3948
3949		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3950
3951		/* Register only for native ports */
3952		err = mlx5_add_netdev_notifier(dev, port_num);
3953		if (err || dev->is_rep || !mlx5_is_roce_init_enabled(mdev))
3954			/*
3955			 * We don't enable ETH interface for
3956			 * 1. IB representors
3957			 * 2. User disabled ROCE through devlink interface
3958			 */
3959			return err;
3960
3961		err = mlx5_enable_eth(dev);
3962		if (err)
3963			goto cleanup;
3964	}
3965
3966	return 0;
3967cleanup:
3968	mlx5_remove_netdev_notifier(dev, port_num);
3969	return err;
3970}
3971
3972static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3973{
3974	struct mlx5_core_dev *mdev = dev->mdev;
3975	enum rdma_link_layer ll;
3976	int port_type_cap;
3977	u32 port_num;
3978
3979	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3980	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3981
3982	if (ll == IB_LINK_LAYER_ETHERNET) {
3983		if (!dev->is_rep)
3984			mlx5_disable_eth(dev);
3985
3986		port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3987		mlx5_remove_netdev_notifier(dev, port_num);
3988	}
3989}
3990
3991static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3992{
3993	mlx5_ib_init_cong_debugfs(dev,
3994				  mlx5_core_native_port_num(dev->mdev) - 1);
3995	return 0;
3996}
3997
3998static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
3999{
4000	mlx5_ib_cleanup_cong_debugfs(dev,
4001				     mlx5_core_native_port_num(dev->mdev) - 1);
4002}
4003
4004static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
4005{
4006	dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4007	return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4008}
4009
4010static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4011{
4012	mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4013}
4014
4015static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4016{
4017	int err;
4018
4019	err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4020	if (err)
4021		return err;
4022
4023	err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4024	if (err)
4025		mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4026
4027	return err;
4028}
4029
4030static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4031{
4032	mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4033	mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4034}
4035
4036static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4037{
4038	const char *name;
4039
4040	if (!mlx5_lag_is_roce(dev->mdev))
4041		name = "mlx5_%d";
4042	else
4043		name = "mlx5_bond_%d";
4044	return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4045}
4046
4047static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4048{
4049	int err;
4050
4051	err = mlx5_mr_cache_cleanup(dev);
4052	if (err)
4053		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4054
4055	if (dev->umrc.qp)
4056		mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4057	if (dev->umrc.cq)
4058		ib_free_cq(dev->umrc.cq);
4059	if (dev->umrc.pd)
4060		ib_dealloc_pd(dev->umrc.pd);
4061}
4062
4063static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4064{
4065	ib_unregister_device(&dev->ib_dev);
4066}
4067
4068enum {
4069	MAX_UMR_WR = 128,
4070};
4071
4072static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4073{
4074	struct ib_qp_init_attr *init_attr = NULL;
4075	struct ib_qp_attr *attr = NULL;
4076	struct ib_pd *pd;
4077	struct ib_cq *cq;
4078	struct ib_qp *qp;
4079	int ret;
4080
4081	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4082	init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4083	if (!attr || !init_attr) {
4084		ret = -ENOMEM;
4085		goto error_0;
4086	}
4087
4088	pd = ib_alloc_pd(&dev->ib_dev, 0);
4089	if (IS_ERR(pd)) {
4090		mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4091		ret = PTR_ERR(pd);
4092		goto error_0;
4093	}
4094
4095	cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4096	if (IS_ERR(cq)) {
4097		mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4098		ret = PTR_ERR(cq);
4099		goto error_2;
4100	}
4101
4102	init_attr->send_cq = cq;
4103	init_attr->recv_cq = cq;
4104	init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4105	init_attr->cap.max_send_wr = MAX_UMR_WR;
4106	init_attr->cap.max_send_sge = 1;
4107	init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4108	init_attr->port_num = 1;
4109	qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4110	if (IS_ERR(qp)) {
4111		mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4112		ret = PTR_ERR(qp);
4113		goto error_3;
4114	}
4115	qp->device     = &dev->ib_dev;
4116	qp->real_qp    = qp;
4117	qp->uobject    = NULL;
4118	qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4119	qp->send_cq    = init_attr->send_cq;
4120	qp->recv_cq    = init_attr->recv_cq;
4121
4122	attr->qp_state = IB_QPS_INIT;
4123	attr->port_num = 1;
4124	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4125				IB_QP_PORT, NULL);
4126	if (ret) {
4127		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4128		goto error_4;
4129	}
4130
4131	memset(attr, 0, sizeof(*attr));
4132	attr->qp_state = IB_QPS_RTR;
4133	attr->path_mtu = IB_MTU_256;
4134
4135	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4136	if (ret) {
4137		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4138		goto error_4;
4139	}
4140
4141	memset(attr, 0, sizeof(*attr));
4142	attr->qp_state = IB_QPS_RTS;
4143	ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4144	if (ret) {
4145		mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4146		goto error_4;
4147	}
4148
4149	dev->umrc.qp = qp;
4150	dev->umrc.cq = cq;
4151	dev->umrc.pd = pd;
4152
4153	sema_init(&dev->umrc.sem, MAX_UMR_WR);
4154	ret = mlx5_mr_cache_init(dev);
4155	if (ret) {
4156		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4157		goto error_4;
4158	}
4159
4160	kfree(attr);
4161	kfree(init_attr);
4162
4163	return 0;
4164
4165error_4:
4166	mlx5_ib_destroy_qp(qp, NULL);
4167	dev->umrc.qp = NULL;
4168
4169error_3:
4170	ib_free_cq(cq);
4171	dev->umrc.cq = NULL;
4172
4173error_2:
4174	ib_dealloc_pd(pd);
4175	dev->umrc.pd = NULL;
4176
4177error_0:
4178	kfree(attr);
4179	kfree(init_attr);
4180	return ret;
4181}
4182
4183static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4184{
4185	struct dentry *root;
4186
4187	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4188		return 0;
4189
4190	mutex_init(&dev->delay_drop.lock);
4191	dev->delay_drop.dev = dev;
4192	dev->delay_drop.activate = false;
4193	dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4194	INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4195	atomic_set(&dev->delay_drop.rqs_cnt, 0);
4196	atomic_set(&dev->delay_drop.events_cnt, 0);
4197
4198	if (!mlx5_debugfs_root)
4199		return 0;
4200
4201	root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4202	dev->delay_drop.dir_debugfs = root;
4203
4204	debugfs_create_atomic_t("num_timeout_events", 0400, root,
4205				&dev->delay_drop.events_cnt);
4206	debugfs_create_atomic_t("num_rqs", 0400, root,
4207				&dev->delay_drop.rqs_cnt);
4208	debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4209			    &fops_delay_drop_timeout);
4210	return 0;
4211}
4212
4213static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4214{
4215	if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4216		return;
4217
4218	cancel_work_sync(&dev->delay_drop.delay_drop_work);
4219	if (!dev->delay_drop.dir_debugfs)
4220		return;
4221
4222	debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4223	dev->delay_drop.dir_debugfs = NULL;
4224}
4225
4226static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4227{
4228	dev->mdev_events.notifier_call = mlx5_ib_event;
4229	mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4230	return 0;
4231}
4232
4233static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4234{
4235	mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4236}
4237
4238void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4239		      const struct mlx5_ib_profile *profile,
4240		      int stage)
4241{
4242	dev->ib_active = false;
4243
4244	/* Number of stages to cleanup */
4245	while (stage) {
4246		stage--;
4247		if (profile->stage[stage].cleanup)
4248			profile->stage[stage].cleanup(dev);
4249	}
4250
4251	kfree(dev->port);
4252	ib_dealloc_device(&dev->ib_dev);
4253}
4254
4255int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4256		  const struct mlx5_ib_profile *profile)
4257{
4258	int err;
4259	int i;
4260
4261	dev->profile = profile;
4262
4263	for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4264		if (profile->stage[i].init) {
4265			err = profile->stage[i].init(dev);
4266			if (err)
4267				goto err_out;
4268		}
4269	}
4270
4271	dev->ib_active = true;
4272	return 0;
4273
4274err_out:
4275	/* Clean up stages which were initialized */
4276	while (i) {
4277		i--;
4278		if (profile->stage[i].cleanup)
4279			profile->stage[i].cleanup(dev);
4280	}
4281	return -ENOMEM;
4282}
4283
4284static const struct mlx5_ib_profile pf_profile = {
4285	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4286		     mlx5_ib_stage_init_init,
4287		     mlx5_ib_stage_init_cleanup),
4288	STAGE_CREATE(MLX5_IB_STAGE_FS,
4289		     mlx5_ib_fs_init,
4290		     mlx5_ib_fs_cleanup),
4291	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4292		     mlx5_ib_stage_caps_init,
4293		     mlx5_ib_stage_caps_cleanup),
4294	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4295		     mlx5_ib_stage_non_default_cb,
4296		     NULL),
4297	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4298		     mlx5_ib_roce_init,
4299		     mlx5_ib_roce_cleanup),
4300	STAGE_CREATE(MLX5_IB_STAGE_QP,
4301		     mlx5_init_qp_table,
4302		     mlx5_cleanup_qp_table),
4303	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4304		     mlx5_init_srq_table,
4305		     mlx5_cleanup_srq_table),
4306	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4307		     mlx5_ib_dev_res_init,
4308		     mlx5_ib_dev_res_cleanup),
4309	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4310		     mlx5_ib_stage_dev_notifier_init,
4311		     mlx5_ib_stage_dev_notifier_cleanup),
4312	STAGE_CREATE(MLX5_IB_STAGE_ODP,
4313		     mlx5_ib_odp_init_one,
4314		     mlx5_ib_odp_cleanup_one),
4315	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4316		     mlx5_ib_counters_init,
4317		     mlx5_ib_counters_cleanup),
4318	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4319		     mlx5_ib_stage_cong_debugfs_init,
4320		     mlx5_ib_stage_cong_debugfs_cleanup),
4321	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4322		     mlx5_ib_stage_uar_init,
4323		     mlx5_ib_stage_uar_cleanup),
4324	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4325		     mlx5_ib_stage_bfrag_init,
4326		     mlx5_ib_stage_bfrag_cleanup),
4327	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4328		     NULL,
4329		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4330	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4331		     mlx5_ib_devx_init,
4332		     mlx5_ib_devx_cleanup),
4333	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4334		     mlx5_ib_stage_ib_reg_init,
4335		     mlx5_ib_stage_ib_reg_cleanup),
4336	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4337		     mlx5_ib_stage_post_ib_reg_umr_init,
4338		     NULL),
4339	STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4340		     mlx5_ib_stage_delay_drop_init,
4341		     mlx5_ib_stage_delay_drop_cleanup),
4342	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4343		     mlx5_ib_restrack_init,
4344		     NULL),
4345};
4346
4347const struct mlx5_ib_profile raw_eth_profile = {
4348	STAGE_CREATE(MLX5_IB_STAGE_INIT,
4349		     mlx5_ib_stage_init_init,
4350		     mlx5_ib_stage_init_cleanup),
4351	STAGE_CREATE(MLX5_IB_STAGE_FS,
4352		     mlx5_ib_fs_init,
4353		     mlx5_ib_fs_cleanup),
4354	STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4355		     mlx5_ib_stage_caps_init,
4356		     mlx5_ib_stage_caps_cleanup),
4357	STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4358		     mlx5_ib_stage_raw_eth_non_default_cb,
4359		     NULL),
4360	STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4361		     mlx5_ib_roce_init,
4362		     mlx5_ib_roce_cleanup),
4363	STAGE_CREATE(MLX5_IB_STAGE_QP,
4364		     mlx5_init_qp_table,
4365		     mlx5_cleanup_qp_table),
4366	STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4367		     mlx5_init_srq_table,
4368		     mlx5_cleanup_srq_table),
4369	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4370		     mlx5_ib_dev_res_init,
4371		     mlx5_ib_dev_res_cleanup),
4372	STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4373		     mlx5_ib_stage_dev_notifier_init,
4374		     mlx5_ib_stage_dev_notifier_cleanup),
4375	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4376		     mlx5_ib_counters_init,
4377		     mlx5_ib_counters_cleanup),
4378	STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4379		     mlx5_ib_stage_cong_debugfs_init,
4380		     mlx5_ib_stage_cong_debugfs_cleanup),
4381	STAGE_CREATE(MLX5_IB_STAGE_UAR,
4382		     mlx5_ib_stage_uar_init,
4383		     mlx5_ib_stage_uar_cleanup),
4384	STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4385		     mlx5_ib_stage_bfrag_init,
4386		     mlx5_ib_stage_bfrag_cleanup),
4387	STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4388		     NULL,
4389		     mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4390	STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4391		     mlx5_ib_devx_init,
4392		     mlx5_ib_devx_cleanup),
4393	STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4394		     mlx5_ib_stage_ib_reg_init,
4395		     mlx5_ib_stage_ib_reg_cleanup),
4396	STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4397		     mlx5_ib_stage_post_ib_reg_umr_init,
4398		     NULL),
4399	STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4400		     mlx5_ib_restrack_init,
4401		     NULL),
4402};
4403
4404static int mlx5r_mp_probe(struct auxiliary_device *adev,
4405			  const struct auxiliary_device_id *id)
4406{
4407	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4408	struct mlx5_core_dev *mdev = idev->mdev;
4409	struct mlx5_ib_multiport_info *mpi;
4410	struct mlx5_ib_dev *dev;
4411	bool bound = false;
4412	int err;
4413
4414	mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4415	if (!mpi)
4416		return -ENOMEM;
4417
4418	mpi->mdev = mdev;
4419	err = mlx5_query_nic_vport_system_image_guid(mdev,
4420						     &mpi->sys_image_guid);
4421	if (err) {
4422		kfree(mpi);
4423		return err;
4424	}
4425
4426	mutex_lock(&mlx5_ib_multiport_mutex);
4427	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4428		if (dev->sys_image_guid == mpi->sys_image_guid)
4429			bound = mlx5_ib_bind_slave_port(dev, mpi);
4430
4431		if (bound) {
4432			rdma_roce_rescan_device(&dev->ib_dev);
4433			mpi->ibdev->ib_active = true;
4434			break;
4435		}
4436	}
4437
4438	if (!bound) {
4439		list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4440		dev_dbg(mdev->device,
4441			"no suitable IB device found to bind to, added to unaffiliated list.\n");
4442	}
4443	mutex_unlock(&mlx5_ib_multiport_mutex);
4444
4445	dev_set_drvdata(&adev->dev, mpi);
4446	return 0;
4447}
4448
4449static void mlx5r_mp_remove(struct auxiliary_device *adev)
4450{
4451	struct mlx5_ib_multiport_info *mpi;
4452
4453	mpi = dev_get_drvdata(&adev->dev);
4454	mutex_lock(&mlx5_ib_multiport_mutex);
4455	if (mpi->ibdev)
4456		mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4457	else
4458		list_del(&mpi->list);
4459	mutex_unlock(&mlx5_ib_multiport_mutex);
4460	kfree(mpi);
4461}
4462
4463static int mlx5r_probe(struct auxiliary_device *adev,
4464		       const struct auxiliary_device_id *id)
4465{
4466	struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4467	struct mlx5_core_dev *mdev = idev->mdev;
4468	const struct mlx5_ib_profile *profile;
4469	int port_type_cap, num_ports, ret;
4470	enum rdma_link_layer ll;
4471	struct mlx5_ib_dev *dev;
4472
4473	port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4474	ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4475
4476	num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4477			MLX5_CAP_GEN(mdev, num_vhca_ports));
4478	dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4479	if (!dev)
4480		return -ENOMEM;
4481	dev->port = kcalloc(num_ports, sizeof(*dev->port),
4482			     GFP_KERNEL);
4483	if (!dev->port) {
4484		ib_dealloc_device(&dev->ib_dev);
4485		return -ENOMEM;
4486	}
4487
4488	dev->mdev = mdev;
4489	dev->num_ports = num_ports;
4490
4491	if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_init_enabled(mdev))
4492		profile = &raw_eth_profile;
4493	else
4494		profile = &pf_profile;
4495
4496	ret = __mlx5_ib_add(dev, profile);
4497	if (ret) {
4498		kfree(dev->port);
4499		ib_dealloc_device(&dev->ib_dev);
4500		return ret;
4501	}
4502
4503	dev_set_drvdata(&adev->dev, dev);
4504	return 0;
4505}
4506
4507static void mlx5r_remove(struct auxiliary_device *adev)
4508{
4509	struct mlx5_ib_dev *dev;
4510
4511	dev = dev_get_drvdata(&adev->dev);
4512	__mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4513}
4514
4515static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4516	{ .name = MLX5_ADEV_NAME ".multiport", },
4517	{},
4518};
4519
4520static const struct auxiliary_device_id mlx5r_id_table[] = {
4521	{ .name = MLX5_ADEV_NAME ".rdma", },
4522	{},
4523};
4524
4525MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4526MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4527
4528static struct auxiliary_driver mlx5r_mp_driver = {
4529	.name = "multiport",
4530	.probe = mlx5r_mp_probe,
4531	.remove = mlx5r_mp_remove,
4532	.id_table = mlx5r_mp_id_table,
4533};
4534
4535static struct auxiliary_driver mlx5r_driver = {
4536	.name = "rdma",
4537	.probe = mlx5r_probe,
4538	.remove = mlx5r_remove,
4539	.id_table = mlx5r_id_table,
4540};
4541
4542static int __init mlx5_ib_init(void)
4543{
4544	int ret;
4545
4546	xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4547	if (!xlt_emergency_page)
4548		return -ENOMEM;
4549
4550	mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4551	if (!mlx5_ib_event_wq) {
4552		free_page((unsigned long)xlt_emergency_page);
4553		return -ENOMEM;
4554	}
4555
4556	mlx5_ib_odp_init();
4557	ret = mlx5r_rep_init();
4558	if (ret)
4559		goto rep_err;
4560	ret = auxiliary_driver_register(&mlx5r_mp_driver);
4561	if (ret)
4562		goto mp_err;
4563	ret = auxiliary_driver_register(&mlx5r_driver);
4564	if (ret)
4565		goto drv_err;
4566	return 0;
4567
4568drv_err:
4569	auxiliary_driver_unregister(&mlx5r_mp_driver);
4570mp_err:
4571	mlx5r_rep_cleanup();
4572rep_err:
4573	destroy_workqueue(mlx5_ib_event_wq);
4574	free_page((unsigned long)xlt_emergency_page);
4575	return ret;
4576}
4577
4578static void __exit mlx5_ib_cleanup(void)
4579{
4580	auxiliary_driver_unregister(&mlx5r_driver);
4581	auxiliary_driver_unregister(&mlx5r_mp_driver);
4582	mlx5r_rep_cleanup();
4583
4584	destroy_workqueue(mlx5_ib_event_wq);
4585	free_page((unsigned long)xlt_emergency_page);
4586}
4587
4588module_init(mlx5_ib_init);
4589module_exit(mlx5_ib_cleanup);