Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
v6.8
   1/*
   2 * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4 *
   5 * This software is available to you under a choice of one of two
   6 * licenses.  You may choose to be licensed under the terms of the GNU
   7 * General Public License (GPL) Version 2, available from the file
   8 * COPYING in the main directory of this source tree, or the
   9 * OpenIB.org BSD license below:
  10 *
  11 *     Redistribution and use in source and binary forms, with or
  12 *     without modification, are permitted provided that the following
  13 *     conditions are met:
  14 *
  15 *      - Redistributions of source code must retain the above
  16 *        copyright notice, this list of conditions and the following
  17 *        disclaimer.
  18 *
  19 *      - Redistributions in binary form must reproduce the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer in the documentation and/or other materials
  22 *        provided with the distribution.
  23 *
  24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31 * SOFTWARE.
  32 */
  33
  34#include <linux/module.h>
  35#include <linux/string.h>
  36#include <linux/errno.h>
  37#include <linux/kernel.h>
  38#include <linux/slab.h>
  39#include <linux/init.h>
  40#include <linux/netdevice.h>
  41#include <net/net_namespace.h>
  42#include <linux/security.h>
  43#include <linux/notifier.h>
  44#include <linux/hashtable.h>
  45#include <rdma/rdma_netlink.h>
  46#include <rdma/ib_addr.h>
  47#include <rdma/ib_cache.h>
  48#include <rdma/rdma_counter.h>
  49
  50#include "core_priv.h"
  51#include "restrack.h"
  52
  53MODULE_AUTHOR("Roland Dreier");
  54MODULE_DESCRIPTION("core kernel InfiniBand API");
  55MODULE_LICENSE("Dual BSD/GPL");
  56
  57struct workqueue_struct *ib_comp_wq;
  58struct workqueue_struct *ib_comp_unbound_wq;
  59struct workqueue_struct *ib_wq;
  60EXPORT_SYMBOL_GPL(ib_wq);
  61static struct workqueue_struct *ib_unreg_wq;
  62
  63/*
  64 * Each of the three rwsem locks (devices, clients, client_data) protects the
  65 * xarray of the same name. Specifically it allows the caller to assert that
  66 * the MARK will/will not be changing under the lock, and for devices and
  67 * clients, that the value in the xarray is still a valid pointer. Change of
  68 * the MARK is linked to the object state, so holding the lock and testing the
  69 * MARK also asserts that the contained object is in a certain state.
  70 *
  71 * This is used to build a two stage register/unregister flow where objects
  72 * can continue to be in the xarray even though they are still in progress to
  73 * register/unregister.
  74 *
  75 * The xarray itself provides additional locking, and restartable iteration,
  76 * which is also relied on.
  77 *
  78 * Locks should not be nested, with the exception of client_data, which is
  79 * allowed to nest under the read side of the other two locks.
  80 *
  81 * The devices_rwsem also protects the device name list, any change or
  82 * assignment of device name must also hold the write side to guarantee unique
  83 * names.
  84 */
  85
  86/*
  87 * devices contains devices that have had their names assigned. The
  88 * devices may not be registered. Users that care about the registration
  89 * status need to call ib_device_try_get() on the device to ensure it is
  90 * registered, and keep it registered, for the required duration.
  91 *
  92 */
  93static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
  94static DECLARE_RWSEM(devices_rwsem);
  95#define DEVICE_REGISTERED XA_MARK_1
  96
  97static u32 highest_client_id;
  98#define CLIENT_REGISTERED XA_MARK_1
  99static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
 100static DECLARE_RWSEM(clients_rwsem);
 101
 102static void ib_client_put(struct ib_client *client)
 103{
 104	if (refcount_dec_and_test(&client->uses))
 105		complete(&client->uses_zero);
 106}
 107
 108/*
 109 * If client_data is registered then the corresponding client must also still
 110 * be registered.
 111 */
 112#define CLIENT_DATA_REGISTERED XA_MARK_1
 113
 114unsigned int rdma_dev_net_id;
 115
 116/*
 117 * A list of net namespaces is maintained in an xarray. This is necessary
 118 * because we can't get the locking right using the existing net ns list. We
 119 * would require a init_net callback after the list is updated.
 120 */
 121static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
 122/*
 123 * rwsem to protect accessing the rdma_nets xarray entries.
 124 */
 125static DECLARE_RWSEM(rdma_nets_rwsem);
 126
 127bool ib_devices_shared_netns = true;
 128module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
 129MODULE_PARM_DESC(netns_mode,
 130		 "Share device among net namespaces; default=1 (shared)");
 131/**
 132 * rdma_dev_access_netns() - Return whether an rdma device can be accessed
 133 *			     from a specified net namespace or not.
 134 * @dev:	Pointer to rdma device which needs to be checked
 135 * @net:	Pointer to net namesapce for which access to be checked
 136 *
 137 * When the rdma device is in shared mode, it ignores the net namespace.
 138 * When the rdma device is exclusive to a net namespace, rdma device net
 139 * namespace is checked against the specified one.
 
 
 
 140 */
 141bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
 142{
 143	return (ib_devices_shared_netns ||
 144		net_eq(read_pnet(&dev->coredev.rdma_net), net));
 145}
 146EXPORT_SYMBOL(rdma_dev_access_netns);
 147
 148/*
 149 * xarray has this behavior where it won't iterate over NULL values stored in
 150 * allocated arrays.  So we need our own iterator to see all values stored in
 151 * the array. This does the same thing as xa_for_each except that it also
 152 * returns NULL valued entries if the array is allocating. Simplified to only
 153 * work on simple xarrays.
 154 */
 155static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
 156			     xa_mark_t filter)
 157{
 158	XA_STATE(xas, xa, *indexp);
 159	void *entry;
 160
 161	rcu_read_lock();
 162	do {
 163		entry = xas_find_marked(&xas, ULONG_MAX, filter);
 164		if (xa_is_zero(entry))
 165			break;
 166	} while (xas_retry(&xas, entry));
 167	rcu_read_unlock();
 168
 169	if (entry) {
 170		*indexp = xas.xa_index;
 171		if (xa_is_zero(entry))
 172			return NULL;
 173		return entry;
 174	}
 175	return XA_ERROR(-ENOENT);
 176}
 177#define xan_for_each_marked(xa, index, entry, filter)                          \
 178	for (index = 0, entry = xan_find_marked(xa, &(index), filter);         \
 179	     !xa_is_err(entry);                                                \
 180	     (index)++, entry = xan_find_marked(xa, &(index), filter))
 181
 182/* RCU hash table mapping netdevice pointers to struct ib_port_data */
 183static DEFINE_SPINLOCK(ndev_hash_lock);
 184static DECLARE_HASHTABLE(ndev_hash, 5);
 185
 186static void free_netdevs(struct ib_device *ib_dev);
 187static void ib_unregister_work(struct work_struct *work);
 188static void __ib_unregister_device(struct ib_device *device);
 189static int ib_security_change(struct notifier_block *nb, unsigned long event,
 190			      void *lsm_data);
 191static void ib_policy_change_task(struct work_struct *work);
 192static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
 193
 194static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
 195			   struct va_format *vaf)
 196{
 197	if (ibdev && ibdev->dev.parent)
 198		dev_printk_emit(level[1] - '0',
 199				ibdev->dev.parent,
 200				"%s %s %s: %pV",
 201				dev_driver_string(ibdev->dev.parent),
 202				dev_name(ibdev->dev.parent),
 203				dev_name(&ibdev->dev),
 204				vaf);
 205	else if (ibdev)
 206		printk("%s%s: %pV",
 207		       level, dev_name(&ibdev->dev), vaf);
 208	else
 209		printk("%s(NULL ib_device): %pV", level, vaf);
 210}
 211
 212void ibdev_printk(const char *level, const struct ib_device *ibdev,
 213		  const char *format, ...)
 214{
 215	struct va_format vaf;
 216	va_list args;
 217
 218	va_start(args, format);
 219
 220	vaf.fmt = format;
 221	vaf.va = &args;
 222
 223	__ibdev_printk(level, ibdev, &vaf);
 224
 225	va_end(args);
 226}
 227EXPORT_SYMBOL(ibdev_printk);
 228
 229#define define_ibdev_printk_level(func, level)                  \
 230void func(const struct ib_device *ibdev, const char *fmt, ...)  \
 231{                                                               \
 232	struct va_format vaf;                                   \
 233	va_list args;                                           \
 234								\
 235	va_start(args, fmt);                                    \
 236								\
 237	vaf.fmt = fmt;                                          \
 238	vaf.va = &args;                                         \
 239								\
 240	__ibdev_printk(level, ibdev, &vaf);                     \
 241								\
 242	va_end(args);                                           \
 243}                                                               \
 244EXPORT_SYMBOL(func);
 245
 246define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
 247define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
 248define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
 249define_ibdev_printk_level(ibdev_err, KERN_ERR);
 250define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
 251define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
 252define_ibdev_printk_level(ibdev_info, KERN_INFO);
 253
 254static struct notifier_block ibdev_lsm_nb = {
 255	.notifier_call = ib_security_change,
 256};
 257
 258static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
 259				 struct net *net);
 260
 261/* Pointer to the RCU head at the start of the ib_port_data array */
 262struct ib_port_data_rcu {
 263	struct rcu_head rcu_head;
 264	struct ib_port_data pdata[];
 265};
 266
 267static void ib_device_check_mandatory(struct ib_device *device)
 268{
 269#define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
 270	static const struct {
 271		size_t offset;
 272		char  *name;
 273	} mandatory_table[] = {
 274		IB_MANDATORY_FUNC(query_device),
 275		IB_MANDATORY_FUNC(query_port),
 
 276		IB_MANDATORY_FUNC(alloc_pd),
 277		IB_MANDATORY_FUNC(dealloc_pd),
 278		IB_MANDATORY_FUNC(create_qp),
 279		IB_MANDATORY_FUNC(modify_qp),
 280		IB_MANDATORY_FUNC(destroy_qp),
 281		IB_MANDATORY_FUNC(post_send),
 282		IB_MANDATORY_FUNC(post_recv),
 283		IB_MANDATORY_FUNC(create_cq),
 284		IB_MANDATORY_FUNC(destroy_cq),
 285		IB_MANDATORY_FUNC(poll_cq),
 286		IB_MANDATORY_FUNC(req_notify_cq),
 287		IB_MANDATORY_FUNC(get_dma_mr),
 288		IB_MANDATORY_FUNC(reg_user_mr),
 289		IB_MANDATORY_FUNC(dereg_mr),
 290		IB_MANDATORY_FUNC(get_port_immutable)
 291	};
 292	int i;
 293
 294	device->kverbs_provider = true;
 295	for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
 296		if (!*(void **) ((void *) &device->ops +
 297				 mandatory_table[i].offset)) {
 298			device->kverbs_provider = false;
 299			break;
 300		}
 301	}
 302}
 303
 304/*
 305 * Caller must perform ib_device_put() to return the device reference count
 306 * when ib_device_get_by_index() returns valid device pointer.
 307 */
 308struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
 309{
 310	struct ib_device *device;
 311
 312	down_read(&devices_rwsem);
 313	device = xa_load(&devices, index);
 314	if (device) {
 315		if (!rdma_dev_access_netns(device, net)) {
 316			device = NULL;
 317			goto out;
 318		}
 319
 320		if (!ib_device_try_get(device))
 321			device = NULL;
 322	}
 323out:
 324	up_read(&devices_rwsem);
 325	return device;
 326}
 327
 328/**
 329 * ib_device_put - Release IB device reference
 330 * @device: device whose reference to be released
 331 *
 332 * ib_device_put() releases reference to the IB device to allow it to be
 333 * unregistered and eventually free.
 334 */
 335void ib_device_put(struct ib_device *device)
 336{
 337	if (refcount_dec_and_test(&device->refcount))
 338		complete(&device->unreg_completion);
 339}
 340EXPORT_SYMBOL(ib_device_put);
 341
 342static struct ib_device *__ib_device_get_by_name(const char *name)
 343{
 344	struct ib_device *device;
 345	unsigned long index;
 346
 347	xa_for_each (&devices, index, device)
 348		if (!strcmp(name, dev_name(&device->dev)))
 349			return device;
 350
 351	return NULL;
 352}
 353
 354/**
 355 * ib_device_get_by_name - Find an IB device by name
 356 * @name: The name to look for
 357 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
 358 *
 359 * Find and hold an ib_device by its name. The caller must call
 360 * ib_device_put() on the returned pointer.
 361 */
 362struct ib_device *ib_device_get_by_name(const char *name,
 363					enum rdma_driver_id driver_id)
 364{
 365	struct ib_device *device;
 366
 367	down_read(&devices_rwsem);
 368	device = __ib_device_get_by_name(name);
 369	if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
 370	    device->ops.driver_id != driver_id)
 371		device = NULL;
 372
 373	if (device) {
 374		if (!ib_device_try_get(device))
 375			device = NULL;
 376	}
 377	up_read(&devices_rwsem);
 378	return device;
 379}
 380EXPORT_SYMBOL(ib_device_get_by_name);
 381
 382static int rename_compat_devs(struct ib_device *device)
 383{
 384	struct ib_core_device *cdev;
 385	unsigned long index;
 386	int ret = 0;
 387
 388	mutex_lock(&device->compat_devs_mutex);
 389	xa_for_each (&device->compat_devs, index, cdev) {
 390		ret = device_rename(&cdev->dev, dev_name(&device->dev));
 391		if (ret) {
 392			dev_warn(&cdev->dev,
 393				 "Fail to rename compatdev to new name %s\n",
 394				 dev_name(&device->dev));
 395			break;
 396		}
 397	}
 398	mutex_unlock(&device->compat_devs_mutex);
 399	return ret;
 400}
 401
 402int ib_device_rename(struct ib_device *ibdev, const char *name)
 403{
 404	unsigned long index;
 405	void *client_data;
 406	int ret;
 407
 408	down_write(&devices_rwsem);
 409	if (!strcmp(name, dev_name(&ibdev->dev))) {
 410		up_write(&devices_rwsem);
 411		return 0;
 412	}
 413
 414	if (__ib_device_get_by_name(name)) {
 415		up_write(&devices_rwsem);
 416		return -EEXIST;
 417	}
 418
 419	ret = device_rename(&ibdev->dev, name);
 420	if (ret) {
 421		up_write(&devices_rwsem);
 422		return ret;
 423	}
 424
 425	strscpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
 426	ret = rename_compat_devs(ibdev);
 427
 428	downgrade_write(&devices_rwsem);
 429	down_read(&ibdev->client_data_rwsem);
 430	xan_for_each_marked(&ibdev->client_data, index, client_data,
 431			    CLIENT_DATA_REGISTERED) {
 432		struct ib_client *client = xa_load(&clients, index);
 433
 434		if (!client || !client->rename)
 435			continue;
 436
 437		client->rename(ibdev, client_data);
 438	}
 439	up_read(&ibdev->client_data_rwsem);
 440	up_read(&devices_rwsem);
 441	return 0;
 442}
 443
 444int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
 445{
 446	if (use_dim > 1)
 447		return -EINVAL;
 448	ibdev->use_cq_dim = use_dim;
 449
 450	return 0;
 451}
 452
 453static int alloc_name(struct ib_device *ibdev, const char *name)
 454{
 455	struct ib_device *device;
 456	unsigned long index;
 457	struct ida inuse;
 458	int rc;
 459	int i;
 460
 461	lockdep_assert_held_write(&devices_rwsem);
 462	ida_init(&inuse);
 463	xa_for_each (&devices, index, device) {
 464		char buf[IB_DEVICE_NAME_MAX];
 465
 466		if (sscanf(dev_name(&device->dev), name, &i) != 1)
 467			continue;
 468		if (i < 0 || i >= INT_MAX)
 469			continue;
 470		snprintf(buf, sizeof buf, name, i);
 471		if (strcmp(buf, dev_name(&device->dev)) != 0)
 472			continue;
 473
 474		rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
 475		if (rc < 0)
 476			goto out;
 477	}
 478
 479	rc = ida_alloc(&inuse, GFP_KERNEL);
 480	if (rc < 0)
 481		goto out;
 482
 483	rc = dev_set_name(&ibdev->dev, name, rc);
 484out:
 485	ida_destroy(&inuse);
 486	return rc;
 487}
 488
 489static void ib_device_release(struct device *device)
 490{
 491	struct ib_device *dev = container_of(device, struct ib_device, dev);
 492
 493	free_netdevs(dev);
 494	WARN_ON(refcount_read(&dev->refcount));
 495	if (dev->hw_stats_data)
 496		ib_device_release_hw_stats(dev->hw_stats_data);
 497	if (dev->port_data) {
 498		ib_cache_release_one(dev);
 499		ib_security_release_port_pkey_list(dev);
 500		rdma_counter_release(dev);
 501		kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
 502				       pdata[0]),
 503			  rcu_head);
 504	}
 505
 506	mutex_destroy(&dev->unregistration_lock);
 507	mutex_destroy(&dev->compat_devs_mutex);
 508
 509	xa_destroy(&dev->compat_devs);
 510	xa_destroy(&dev->client_data);
 511	kfree_rcu(dev, rcu_head);
 512}
 513
 514static int ib_device_uevent(const struct device *device,
 515			    struct kobj_uevent_env *env)
 516{
 517	if (add_uevent_var(env, "NAME=%s", dev_name(device)))
 518		return -ENOMEM;
 519
 520	/*
 521	 * It would be nice to pass the node GUID with the event...
 522	 */
 523
 524	return 0;
 525}
 526
 527static const void *net_namespace(const struct device *d)
 528{
 529	const struct ib_core_device *coredev =
 530			container_of(d, struct ib_core_device, dev);
 531
 532	return read_pnet(&coredev->rdma_net);
 533}
 534
 535static struct class ib_class = {
 536	.name    = "infiniband",
 537	.dev_release = ib_device_release,
 538	.dev_uevent = ib_device_uevent,
 539	.ns_type = &net_ns_type_operations,
 540	.namespace = net_namespace,
 541};
 542
 543static void rdma_init_coredev(struct ib_core_device *coredev,
 544			      struct ib_device *dev, struct net *net)
 545{
 546	/* This BUILD_BUG_ON is intended to catch layout change
 547	 * of union of ib_core_device and device.
 548	 * dev must be the first element as ib_core and providers
 549	 * driver uses it. Adding anything in ib_core_device before
 550	 * device will break this assumption.
 551	 */
 552	BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
 553		     offsetof(struct ib_device, dev));
 554
 555	coredev->dev.class = &ib_class;
 556	coredev->dev.groups = dev->groups;
 557	device_initialize(&coredev->dev);
 558	coredev->owner = dev;
 559	INIT_LIST_HEAD(&coredev->port_list);
 560	write_pnet(&coredev->rdma_net, net);
 561}
 562
 563/**
 564 * _ib_alloc_device - allocate an IB device struct
 565 * @size:size of structure to allocate
 566 *
 567 * Low-level drivers should use ib_alloc_device() to allocate &struct
 568 * ib_device.  @size is the size of the structure to be allocated,
 569 * including any private data used by the low-level driver.
 570 * ib_dealloc_device() must be used to free structures allocated with
 571 * ib_alloc_device().
 572 */
 573struct ib_device *_ib_alloc_device(size_t size)
 574{
 575	struct ib_device *device;
 576	unsigned int i;
 577
 578	if (WARN_ON(size < sizeof(struct ib_device)))
 579		return NULL;
 580
 581	device = kzalloc(size, GFP_KERNEL);
 582	if (!device)
 583		return NULL;
 584
 585	if (rdma_restrack_init(device)) {
 586		kfree(device);
 587		return NULL;
 588	}
 589
 
 590	rdma_init_coredev(&device->coredev, device, &init_net);
 591
 592	INIT_LIST_HEAD(&device->event_handler_list);
 593	spin_lock_init(&device->qp_open_list_lock);
 594	init_rwsem(&device->event_handler_rwsem);
 595	mutex_init(&device->unregistration_lock);
 596	/*
 597	 * client_data needs to be alloc because we don't want our mark to be
 598	 * destroyed if the user stores NULL in the client data.
 599	 */
 600	xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
 601	init_rwsem(&device->client_data_rwsem);
 602	xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
 603	mutex_init(&device->compat_devs_mutex);
 604	init_completion(&device->unreg_completion);
 605	INIT_WORK(&device->unregistration_work, ib_unregister_work);
 606
 607	spin_lock_init(&device->cq_pools_lock);
 608	for (i = 0; i < ARRAY_SIZE(device->cq_pools); i++)
 609		INIT_LIST_HEAD(&device->cq_pools[i]);
 610
 611	rwlock_init(&device->cache_lock);
 612
 613	device->uverbs_cmd_mask =
 614		BIT_ULL(IB_USER_VERBS_CMD_ALLOC_MW) |
 615		BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) |
 616		BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) |
 617		BIT_ULL(IB_USER_VERBS_CMD_CLOSE_XRCD) |
 618		BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) |
 619		BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
 620		BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) |
 621		BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) |
 622		BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) |
 623		BIT_ULL(IB_USER_VERBS_CMD_CREATE_XSRQ) |
 624		BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_MW) |
 625		BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) |
 626		BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) |
 627		BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) |
 628		BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) |
 629		BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) |
 630		BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) |
 631		BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) |
 632		BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) |
 633		BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) |
 634		BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) |
 635		BIT_ULL(IB_USER_VERBS_CMD_OPEN_QP) |
 636		BIT_ULL(IB_USER_VERBS_CMD_OPEN_XRCD) |
 637		BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) |
 638		BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) |
 639		BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) |
 640		BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) |
 641		BIT_ULL(IB_USER_VERBS_CMD_REG_MR) |
 642		BIT_ULL(IB_USER_VERBS_CMD_REREG_MR) |
 643		BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ);
 644	return device;
 645}
 646EXPORT_SYMBOL(_ib_alloc_device);
 647
 648/**
 649 * ib_dealloc_device - free an IB device struct
 650 * @device:structure to free
 651 *
 652 * Free a structure allocated with ib_alloc_device().
 653 */
 654void ib_dealloc_device(struct ib_device *device)
 655{
 656	if (device->ops.dealloc_driver)
 657		device->ops.dealloc_driver(device);
 658
 659	/*
 660	 * ib_unregister_driver() requires all devices to remain in the xarray
 661	 * while their ops are callable. The last op we call is dealloc_driver
 662	 * above.  This is needed to create a fence on op callbacks prior to
 663	 * allowing the driver module to unload.
 664	 */
 665	down_write(&devices_rwsem);
 666	if (xa_load(&devices, device->index) == device)
 667		xa_erase(&devices, device->index);
 668	up_write(&devices_rwsem);
 669
 670	/* Expedite releasing netdev references */
 671	free_netdevs(device);
 672
 673	WARN_ON(!xa_empty(&device->compat_devs));
 674	WARN_ON(!xa_empty(&device->client_data));
 675	WARN_ON(refcount_read(&device->refcount));
 676	rdma_restrack_clean(device);
 677	/* Balances with device_initialize */
 678	put_device(&device->dev);
 679}
 680EXPORT_SYMBOL(ib_dealloc_device);
 681
 682/*
 683 * add_client_context() and remove_client_context() must be safe against
 684 * parallel calls on the same device - registration/unregistration of both the
 685 * device and client can be occurring in parallel.
 686 *
 687 * The routines need to be a fence, any caller must not return until the add
 688 * or remove is fully completed.
 689 */
 690static int add_client_context(struct ib_device *device,
 691			      struct ib_client *client)
 692{
 693	int ret = 0;
 694
 695	if (!device->kverbs_provider && !client->no_kverbs_req)
 696		return 0;
 697
 698	down_write(&device->client_data_rwsem);
 699	/*
 700	 * So long as the client is registered hold both the client and device
 701	 * unregistration locks.
 702	 */
 703	if (!refcount_inc_not_zero(&client->uses))
 704		goto out_unlock;
 705	refcount_inc(&device->refcount);
 706
 707	/*
 708	 * Another caller to add_client_context got here first and has already
 709	 * completely initialized context.
 710	 */
 711	if (xa_get_mark(&device->client_data, client->client_id,
 712		    CLIENT_DATA_REGISTERED))
 713		goto out;
 714
 715	ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
 716			      GFP_KERNEL));
 717	if (ret)
 718		goto out;
 719	downgrade_write(&device->client_data_rwsem);
 720	if (client->add) {
 721		if (client->add(device)) {
 722			/*
 723			 * If a client fails to add then the error code is
 724			 * ignored, but we won't call any more ops on this
 725			 * client.
 726			 */
 727			xa_erase(&device->client_data, client->client_id);
 728			up_read(&device->client_data_rwsem);
 729			ib_device_put(device);
 730			ib_client_put(client);
 731			return 0;
 732		}
 733	}
 734
 735	/* Readers shall not see a client until add has been completed */
 736	xa_set_mark(&device->client_data, client->client_id,
 737		    CLIENT_DATA_REGISTERED);
 738	up_read(&device->client_data_rwsem);
 739	return 0;
 740
 741out:
 742	ib_device_put(device);
 743	ib_client_put(client);
 744out_unlock:
 745	up_write(&device->client_data_rwsem);
 746	return ret;
 747}
 748
 749static void remove_client_context(struct ib_device *device,
 750				  unsigned int client_id)
 751{
 752	struct ib_client *client;
 753	void *client_data;
 754
 755	down_write(&device->client_data_rwsem);
 756	if (!xa_get_mark(&device->client_data, client_id,
 757			 CLIENT_DATA_REGISTERED)) {
 758		up_write(&device->client_data_rwsem);
 759		return;
 760	}
 761	client_data = xa_load(&device->client_data, client_id);
 762	xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
 763	client = xa_load(&clients, client_id);
 764	up_write(&device->client_data_rwsem);
 765
 766	/*
 767	 * Notice we cannot be holding any exclusive locks when calling the
 768	 * remove callback as the remove callback can recurse back into any
 769	 * public functions in this module and thus try for any locks those
 770	 * functions take.
 771	 *
 772	 * For this reason clients and drivers should not call the
 773	 * unregistration functions will holdling any locks.
 774	 */
 775	if (client->remove)
 776		client->remove(device, client_data);
 777
 778	xa_erase(&device->client_data, client_id);
 779	ib_device_put(device);
 780	ib_client_put(client);
 781}
 782
 783static int alloc_port_data(struct ib_device *device)
 784{
 785	struct ib_port_data_rcu *pdata_rcu;
 786	u32 port;
 787
 788	if (device->port_data)
 789		return 0;
 790
 791	/* This can only be called once the physical port range is defined */
 792	if (WARN_ON(!device->phys_port_cnt))
 793		return -EINVAL;
 794
 795	/* Reserve U32_MAX so the logic to go over all the ports is sane */
 796	if (WARN_ON(device->phys_port_cnt == U32_MAX))
 797		return -EINVAL;
 798
 799	/*
 800	 * device->port_data is indexed directly by the port number to make
 801	 * access to this data as efficient as possible.
 802	 *
 803	 * Therefore port_data is declared as a 1 based array with potential
 804	 * empty slots at the beginning.
 805	 */
 806	pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
 807					size_add(rdma_end_port(device), 1)),
 808			    GFP_KERNEL);
 809	if (!pdata_rcu)
 810		return -ENOMEM;
 811	/*
 812	 * The rcu_head is put in front of the port data array and the stored
 813	 * pointer is adjusted since we never need to see that member until
 814	 * kfree_rcu.
 815	 */
 816	device->port_data = pdata_rcu->pdata;
 817
 818	rdma_for_each_port (device, port) {
 819		struct ib_port_data *pdata = &device->port_data[port];
 820
 821		pdata->ib_dev = device;
 822		spin_lock_init(&pdata->pkey_list_lock);
 823		INIT_LIST_HEAD(&pdata->pkey_list);
 824		spin_lock_init(&pdata->netdev_lock);
 825		INIT_HLIST_NODE(&pdata->ndev_hash_link);
 826	}
 827	return 0;
 828}
 829
 830static int verify_immutable(const struct ib_device *dev, u32 port)
 831{
 832	return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
 833			    rdma_max_mad_size(dev, port) != 0);
 834}
 835
 836static int setup_port_data(struct ib_device *device)
 837{
 838	u32 port;
 839	int ret;
 840
 841	ret = alloc_port_data(device);
 842	if (ret)
 843		return ret;
 844
 845	rdma_for_each_port (device, port) {
 846		struct ib_port_data *pdata = &device->port_data[port];
 847
 848		ret = device->ops.get_port_immutable(device, port,
 849						     &pdata->immutable);
 850		if (ret)
 851			return ret;
 852
 853		if (verify_immutable(device, port))
 854			return -EINVAL;
 855	}
 856	return 0;
 857}
 858
 859/**
 860 * ib_port_immutable_read() - Read rdma port's immutable data
 861 * @dev: IB device
 862 * @port: port number whose immutable data to read. It starts with index 1 and
 863 *        valid upto including rdma_end_port().
 864 */
 865const struct ib_port_immutable*
 866ib_port_immutable_read(struct ib_device *dev, unsigned int port)
 867{
 868	WARN_ON(!rdma_is_port_valid(dev, port));
 869	return &dev->port_data[port].immutable;
 870}
 871EXPORT_SYMBOL(ib_port_immutable_read);
 872
 873void ib_get_device_fw_str(struct ib_device *dev, char *str)
 874{
 875	if (dev->ops.get_dev_fw_str)
 876		dev->ops.get_dev_fw_str(dev, str);
 877	else
 878		str[0] = '\0';
 879}
 880EXPORT_SYMBOL(ib_get_device_fw_str);
 881
 882static void ib_policy_change_task(struct work_struct *work)
 883{
 884	struct ib_device *dev;
 885	unsigned long index;
 886
 887	down_read(&devices_rwsem);
 888	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
 889		unsigned int i;
 890
 891		rdma_for_each_port (dev, i) {
 892			u64 sp;
 893			ib_get_cached_subnet_prefix(dev, i, &sp);
 894			ib_security_cache_change(dev, i, sp);
 
 
 
 
 
 
 
 895		}
 896	}
 897	up_read(&devices_rwsem);
 898}
 899
 900static int ib_security_change(struct notifier_block *nb, unsigned long event,
 901			      void *lsm_data)
 902{
 903	if (event != LSM_POLICY_CHANGE)
 904		return NOTIFY_DONE;
 905
 906	schedule_work(&ib_policy_change_work);
 907	ib_mad_agent_security_change();
 908
 909	return NOTIFY_OK;
 910}
 911
 912static void compatdev_release(struct device *dev)
 913{
 914	struct ib_core_device *cdev =
 915		container_of(dev, struct ib_core_device, dev);
 916
 917	kfree(cdev);
 918}
 919
 920static int add_one_compat_dev(struct ib_device *device,
 921			      struct rdma_dev_net *rnet)
 922{
 923	struct ib_core_device *cdev;
 924	int ret;
 925
 926	lockdep_assert_held(&rdma_nets_rwsem);
 927	if (!ib_devices_shared_netns)
 928		return 0;
 929
 930	/*
 931	 * Create and add compat device in all namespaces other than where it
 932	 * is currently bound to.
 933	 */
 934	if (net_eq(read_pnet(&rnet->net),
 935		   read_pnet(&device->coredev.rdma_net)))
 936		return 0;
 937
 938	/*
 939	 * The first of init_net() or ib_register_device() to take the
 940	 * compat_devs_mutex wins and gets to add the device. Others will wait
 941	 * for completion here.
 942	 */
 943	mutex_lock(&device->compat_devs_mutex);
 944	cdev = xa_load(&device->compat_devs, rnet->id);
 945	if (cdev) {
 946		ret = 0;
 947		goto done;
 948	}
 949	ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
 950	if (ret)
 951		goto done;
 952
 953	cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
 954	if (!cdev) {
 955		ret = -ENOMEM;
 956		goto cdev_err;
 957	}
 958
 959	cdev->dev.parent = device->dev.parent;
 960	rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
 961	cdev->dev.release = compatdev_release;
 962	ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
 963	if (ret)
 964		goto add_err;
 965
 966	ret = device_add(&cdev->dev);
 967	if (ret)
 968		goto add_err;
 969	ret = ib_setup_port_attrs(cdev);
 970	if (ret)
 971		goto port_err;
 972
 973	ret = xa_err(xa_store(&device->compat_devs, rnet->id,
 974			      cdev, GFP_KERNEL));
 975	if (ret)
 976		goto insert_err;
 977
 978	mutex_unlock(&device->compat_devs_mutex);
 979	return 0;
 980
 981insert_err:
 982	ib_free_port_attrs(cdev);
 983port_err:
 984	device_del(&cdev->dev);
 985add_err:
 986	put_device(&cdev->dev);
 987cdev_err:
 988	xa_release(&device->compat_devs, rnet->id);
 989done:
 990	mutex_unlock(&device->compat_devs_mutex);
 991	return ret;
 992}
 993
 994static void remove_one_compat_dev(struct ib_device *device, u32 id)
 995{
 996	struct ib_core_device *cdev;
 997
 998	mutex_lock(&device->compat_devs_mutex);
 999	cdev = xa_erase(&device->compat_devs, id);
1000	mutex_unlock(&device->compat_devs_mutex);
1001	if (cdev) {
1002		ib_free_port_attrs(cdev);
1003		device_del(&cdev->dev);
1004		put_device(&cdev->dev);
1005	}
1006}
1007
1008static void remove_compat_devs(struct ib_device *device)
1009{
1010	struct ib_core_device *cdev;
1011	unsigned long index;
1012
1013	xa_for_each (&device->compat_devs, index, cdev)
1014		remove_one_compat_dev(device, index);
1015}
1016
1017static int add_compat_devs(struct ib_device *device)
1018{
1019	struct rdma_dev_net *rnet;
1020	unsigned long index;
1021	int ret = 0;
1022
1023	lockdep_assert_held(&devices_rwsem);
1024
1025	down_read(&rdma_nets_rwsem);
1026	xa_for_each (&rdma_nets, index, rnet) {
1027		ret = add_one_compat_dev(device, rnet);
1028		if (ret)
1029			break;
1030	}
1031	up_read(&rdma_nets_rwsem);
1032	return ret;
1033}
1034
1035static void remove_all_compat_devs(void)
1036{
1037	struct ib_compat_device *cdev;
1038	struct ib_device *dev;
1039	unsigned long index;
1040
1041	down_read(&devices_rwsem);
1042	xa_for_each (&devices, index, dev) {
1043		unsigned long c_index = 0;
1044
1045		/* Hold nets_rwsem so that any other thread modifying this
1046		 * system param can sync with this thread.
1047		 */
1048		down_read(&rdma_nets_rwsem);
1049		xa_for_each (&dev->compat_devs, c_index, cdev)
1050			remove_one_compat_dev(dev, c_index);
1051		up_read(&rdma_nets_rwsem);
1052	}
1053	up_read(&devices_rwsem);
1054}
1055
1056static int add_all_compat_devs(void)
1057{
1058	struct rdma_dev_net *rnet;
1059	struct ib_device *dev;
1060	unsigned long index;
1061	int ret = 0;
1062
1063	down_read(&devices_rwsem);
1064	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1065		unsigned long net_index = 0;
1066
1067		/* Hold nets_rwsem so that any other thread modifying this
1068		 * system param can sync with this thread.
1069		 */
1070		down_read(&rdma_nets_rwsem);
1071		xa_for_each (&rdma_nets, net_index, rnet) {
1072			ret = add_one_compat_dev(dev, rnet);
1073			if (ret)
1074				break;
1075		}
1076		up_read(&rdma_nets_rwsem);
1077	}
1078	up_read(&devices_rwsem);
1079	if (ret)
1080		remove_all_compat_devs();
1081	return ret;
1082}
1083
1084int rdma_compatdev_set(u8 enable)
1085{
1086	struct rdma_dev_net *rnet;
1087	unsigned long index;
1088	int ret = 0;
1089
1090	down_write(&rdma_nets_rwsem);
1091	if (ib_devices_shared_netns == enable) {
1092		up_write(&rdma_nets_rwsem);
1093		return 0;
1094	}
1095
1096	/* enable/disable of compat devices is not supported
1097	 * when more than default init_net exists.
1098	 */
1099	xa_for_each (&rdma_nets, index, rnet) {
1100		ret++;
1101		break;
1102	}
1103	if (!ret)
1104		ib_devices_shared_netns = enable;
1105	up_write(&rdma_nets_rwsem);
1106	if (ret)
1107		return -EBUSY;
1108
1109	if (enable)
1110		ret = add_all_compat_devs();
1111	else
1112		remove_all_compat_devs();
1113	return ret;
1114}
1115
1116static void rdma_dev_exit_net(struct net *net)
1117{
1118	struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1119	struct ib_device *dev;
1120	unsigned long index;
1121	int ret;
1122
1123	down_write(&rdma_nets_rwsem);
1124	/*
1125	 * Prevent the ID from being re-used and hide the id from xa_for_each.
1126	 */
1127	ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1128	WARN_ON(ret);
1129	up_write(&rdma_nets_rwsem);
1130
1131	down_read(&devices_rwsem);
1132	xa_for_each (&devices, index, dev) {
1133		get_device(&dev->dev);
1134		/*
1135		 * Release the devices_rwsem so that pontentially blocking
1136		 * device_del, doesn't hold the devices_rwsem for too long.
1137		 */
1138		up_read(&devices_rwsem);
1139
1140		remove_one_compat_dev(dev, rnet->id);
1141
1142		/*
1143		 * If the real device is in the NS then move it back to init.
1144		 */
1145		rdma_dev_change_netns(dev, net, &init_net);
1146
1147		put_device(&dev->dev);
1148		down_read(&devices_rwsem);
1149	}
1150	up_read(&devices_rwsem);
1151
1152	rdma_nl_net_exit(rnet);
1153	xa_erase(&rdma_nets, rnet->id);
1154}
1155
1156static __net_init int rdma_dev_init_net(struct net *net)
1157{
1158	struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1159	unsigned long index;
1160	struct ib_device *dev;
1161	int ret;
1162
1163	write_pnet(&rnet->net, net);
1164
1165	ret = rdma_nl_net_init(rnet);
1166	if (ret)
1167		return ret;
1168
1169	/* No need to create any compat devices in default init_net. */
1170	if (net_eq(net, &init_net))
1171		return 0;
1172
1173	ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1174	if (ret) {
1175		rdma_nl_net_exit(rnet);
1176		return ret;
1177	}
1178
1179	down_read(&devices_rwsem);
1180	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1181		/* Hold nets_rwsem so that netlink command cannot change
1182		 * system configuration for device sharing mode.
1183		 */
1184		down_read(&rdma_nets_rwsem);
1185		ret = add_one_compat_dev(dev, rnet);
1186		up_read(&rdma_nets_rwsem);
1187		if (ret)
1188			break;
1189	}
1190	up_read(&devices_rwsem);
1191
1192	if (ret)
1193		rdma_dev_exit_net(net);
1194
1195	return ret;
1196}
1197
1198/*
1199 * Assign the unique string device name and the unique device index. This is
1200 * undone by ib_dealloc_device.
1201 */
1202static int assign_name(struct ib_device *device, const char *name)
1203{
1204	static u32 last_id;
1205	int ret;
1206
1207	down_write(&devices_rwsem);
1208	/* Assign a unique name to the device */
1209	if (strchr(name, '%'))
1210		ret = alloc_name(device, name);
1211	else
1212		ret = dev_set_name(&device->dev, name);
1213	if (ret)
1214		goto out;
1215
1216	if (__ib_device_get_by_name(dev_name(&device->dev))) {
1217		ret = -ENFILE;
1218		goto out;
1219	}
1220	strscpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1221
1222	ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1223			&last_id, GFP_KERNEL);
1224	if (ret > 0)
1225		ret = 0;
1226
1227out:
1228	up_write(&devices_rwsem);
1229	return ret;
1230}
1231
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1232/*
1233 * setup_device() allocates memory and sets up data that requires calling the
1234 * device ops, this is the only reason these actions are not done during
1235 * ib_alloc_device. It is undone by ib_dealloc_device().
1236 */
1237static int setup_device(struct ib_device *device)
1238{
1239	struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1240	int ret;
1241
 
1242	ib_device_check_mandatory(device);
1243
1244	ret = setup_port_data(device);
1245	if (ret) {
1246		dev_warn(&device->dev, "Couldn't create per-port data\n");
1247		return ret;
1248	}
1249
1250	memset(&device->attrs, 0, sizeof(device->attrs));
1251	ret = device->ops.query_device(device, &device->attrs, &uhw);
1252	if (ret) {
1253		dev_warn(&device->dev,
1254			 "Couldn't query the device attributes\n");
1255		return ret;
1256	}
1257
1258	return 0;
1259}
1260
1261static void disable_device(struct ib_device *device)
1262{
1263	u32 cid;
1264
1265	WARN_ON(!refcount_read(&device->refcount));
1266
1267	down_write(&devices_rwsem);
1268	xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1269	up_write(&devices_rwsem);
1270
1271	/*
1272	 * Remove clients in LIFO order, see assign_client_id. This could be
1273	 * more efficient if xarray learns to reverse iterate. Since no new
1274	 * clients can be added to this ib_device past this point we only need
1275	 * the maximum possible client_id value here.
1276	 */
1277	down_read(&clients_rwsem);
1278	cid = highest_client_id;
1279	up_read(&clients_rwsem);
1280	while (cid) {
1281		cid--;
1282		remove_client_context(device, cid);
1283	}
1284
1285	ib_cq_pool_cleanup(device);
1286
1287	/* Pairs with refcount_set in enable_device */
1288	ib_device_put(device);
1289	wait_for_completion(&device->unreg_completion);
1290
1291	/*
1292	 * compat devices must be removed after device refcount drops to zero.
1293	 * Otherwise init_net() may add more compatdevs after removing compat
1294	 * devices and before device is disabled.
1295	 */
1296	remove_compat_devs(device);
1297}
1298
1299/*
1300 * An enabled device is visible to all clients and to all the public facing
1301 * APIs that return a device pointer. This always returns with a new get, even
1302 * if it fails.
1303 */
1304static int enable_device_and_get(struct ib_device *device)
1305{
1306	struct ib_client *client;
1307	unsigned long index;
1308	int ret = 0;
1309
1310	/*
1311	 * One ref belongs to the xa and the other belongs to this
1312	 * thread. This is needed to guard against parallel unregistration.
1313	 */
1314	refcount_set(&device->refcount, 2);
1315	down_write(&devices_rwsem);
1316	xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1317
1318	/*
1319	 * By using downgrade_write() we ensure that no other thread can clear
1320	 * DEVICE_REGISTERED while we are completing the client setup.
1321	 */
1322	downgrade_write(&devices_rwsem);
1323
1324	if (device->ops.enable_driver) {
1325		ret = device->ops.enable_driver(device);
1326		if (ret)
1327			goto out;
1328	}
1329
1330	down_read(&clients_rwsem);
1331	xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1332		ret = add_client_context(device, client);
1333		if (ret)
1334			break;
1335	}
1336	up_read(&clients_rwsem);
1337	if (!ret)
1338		ret = add_compat_devs(device);
1339out:
1340	up_read(&devices_rwsem);
1341	return ret;
1342}
1343
1344static void prevent_dealloc_device(struct ib_device *ib_dev)
1345{
1346}
1347
1348/**
1349 * ib_register_device - Register an IB device with IB core
1350 * @device: Device to register
1351 * @name: unique string device name. This may include a '%' which will
1352 * 	  cause a unique index to be added to the passed device name.
1353 * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
1354 *	        device will be used. In this case the caller should fully
1355 *		setup the ibdev for DMA. This usually means using dma_virt_ops.
1356 *
1357 * Low-level drivers use ib_register_device() to register their
1358 * devices with the IB core.  All registered clients will receive a
1359 * callback for each device that is added. @device must be allocated
1360 * with ib_alloc_device().
1361 *
1362 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1363 * asynchronously then the device pointer may become freed as soon as this
1364 * function returns.
1365 */
1366int ib_register_device(struct ib_device *device, const char *name,
1367		       struct device *dma_device)
1368{
1369	int ret;
1370
1371	ret = assign_name(device, name);
1372	if (ret)
1373		return ret;
1374
1375	/*
1376	 * If the caller does not provide a DMA capable device then the IB core
1377	 * will set up ib_sge and scatterlist structures that stash the kernel
1378	 * virtual address into the address field.
1379	 */
1380	WARN_ON(dma_device && !dma_device->dma_parms);
1381	device->dma_device = dma_device;
1382
1383	ret = setup_device(device);
1384	if (ret)
1385		return ret;
1386
1387	ret = ib_cache_setup_one(device);
1388	if (ret) {
1389		dev_warn(&device->dev,
1390			 "Couldn't set up InfiniBand P_Key/GID cache\n");
1391		return ret;
1392	}
1393
1394	device->groups[0] = &ib_dev_attr_group;
1395	device->groups[1] = device->ops.device_group;
1396	ret = ib_setup_device_attrs(device);
1397	if (ret)
1398		goto cache_cleanup;
1399
1400	ib_device_register_rdmacg(device);
1401
1402	rdma_counter_init(device);
1403
1404	/*
1405	 * Ensure that ADD uevent is not fired because it
1406	 * is too early amd device is not initialized yet.
1407	 */
1408	dev_set_uevent_suppress(&device->dev, true);
1409	ret = device_add(&device->dev);
1410	if (ret)
1411		goto cg_cleanup;
1412
1413	ret = ib_setup_port_attrs(&device->coredev);
1414	if (ret) {
1415		dev_warn(&device->dev,
1416			 "Couldn't register device with driver model\n");
1417		goto dev_cleanup;
1418	}
1419
1420	ret = enable_device_and_get(device);
 
 
 
1421	if (ret) {
1422		void (*dealloc_fn)(struct ib_device *);
1423
1424		/*
1425		 * If we hit this error flow then we don't want to
1426		 * automatically dealloc the device since the caller is
1427		 * expected to call ib_dealloc_device() after
1428		 * ib_register_device() fails. This is tricky due to the
1429		 * possibility for a parallel unregistration along with this
1430		 * error flow. Since we have a refcount here we know any
1431		 * parallel flow is stopped in disable_device and will see the
1432		 * special dealloc_driver pointer, causing the responsibility to
1433		 * ib_dealloc_device() to revert back to this thread.
1434		 */
1435		dealloc_fn = device->ops.dealloc_driver;
1436		device->ops.dealloc_driver = prevent_dealloc_device;
1437		ib_device_put(device);
1438		__ib_unregister_device(device);
1439		device->ops.dealloc_driver = dealloc_fn;
1440		dev_set_uevent_suppress(&device->dev, false);
1441		return ret;
1442	}
1443	dev_set_uevent_suppress(&device->dev, false);
1444	/* Mark for userspace that device is ready */
1445	kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1446	ib_device_put(device);
1447
1448	return 0;
1449
1450dev_cleanup:
1451	device_del(&device->dev);
1452cg_cleanup:
1453	dev_set_uevent_suppress(&device->dev, false);
1454	ib_device_unregister_rdmacg(device);
1455cache_cleanup:
1456	ib_cache_cleanup_one(device);
1457	return ret;
1458}
1459EXPORT_SYMBOL(ib_register_device);
1460
1461/* Callers must hold a get on the device. */
1462static void __ib_unregister_device(struct ib_device *ib_dev)
1463{
1464	/*
1465	 * We have a registration lock so that all the calls to unregister are
1466	 * fully fenced, once any unregister returns the device is truely
1467	 * unregistered even if multiple callers are unregistering it at the
1468	 * same time. This also interacts with the registration flow and
1469	 * provides sane semantics if register and unregister are racing.
1470	 */
1471	mutex_lock(&ib_dev->unregistration_lock);
1472	if (!refcount_read(&ib_dev->refcount))
1473		goto out;
1474
1475	disable_device(ib_dev);
1476
1477	/* Expedite removing unregistered pointers from the hash table */
1478	free_netdevs(ib_dev);
1479
1480	ib_free_port_attrs(&ib_dev->coredev);
1481	device_del(&ib_dev->dev);
1482	ib_device_unregister_rdmacg(ib_dev);
1483	ib_cache_cleanup_one(ib_dev);
1484
1485	/*
1486	 * Drivers using the new flow may not call ib_dealloc_device except
1487	 * in error unwind prior to registration success.
1488	 */
1489	if (ib_dev->ops.dealloc_driver &&
1490	    ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
1491		WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1492		ib_dealloc_device(ib_dev);
1493	}
1494out:
1495	mutex_unlock(&ib_dev->unregistration_lock);
1496}
1497
1498/**
1499 * ib_unregister_device - Unregister an IB device
1500 * @ib_dev: The device to unregister
1501 *
1502 * Unregister an IB device.  All clients will receive a remove callback.
1503 *
1504 * Callers should call this routine only once, and protect against races with
1505 * registration. Typically it should only be called as part of a remove
1506 * callback in an implementation of driver core's struct device_driver and
1507 * related.
1508 *
1509 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1510 * this function.
1511 */
1512void ib_unregister_device(struct ib_device *ib_dev)
1513{
1514	get_device(&ib_dev->dev);
1515	__ib_unregister_device(ib_dev);
1516	put_device(&ib_dev->dev);
1517}
1518EXPORT_SYMBOL(ib_unregister_device);
1519
1520/**
1521 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1522 * @ib_dev: The device to unregister
1523 *
1524 * This is the same as ib_unregister_device(), except it includes an internal
1525 * ib_device_put() that should match a 'get' obtained by the caller.
1526 *
1527 * It is safe to call this routine concurrently from multiple threads while
1528 * holding the 'get'. When the function returns the device is fully
1529 * unregistered.
1530 *
1531 * Drivers using this flow MUST use the driver_unregister callback to clean up
1532 * their resources associated with the device and dealloc it.
1533 */
1534void ib_unregister_device_and_put(struct ib_device *ib_dev)
1535{
1536	WARN_ON(!ib_dev->ops.dealloc_driver);
1537	get_device(&ib_dev->dev);
1538	ib_device_put(ib_dev);
1539	__ib_unregister_device(ib_dev);
1540	put_device(&ib_dev->dev);
1541}
1542EXPORT_SYMBOL(ib_unregister_device_and_put);
1543
1544/**
1545 * ib_unregister_driver - Unregister all IB devices for a driver
1546 * @driver_id: The driver to unregister
1547 *
1548 * This implements a fence for device unregistration. It only returns once all
1549 * devices associated with the driver_id have fully completed their
1550 * unregistration and returned from ib_unregister_device*().
1551 *
1552 * If device's are not yet unregistered it goes ahead and starts unregistering
1553 * them.
1554 *
1555 * This does not block creation of new devices with the given driver_id, that
1556 * is the responsibility of the caller.
1557 */
1558void ib_unregister_driver(enum rdma_driver_id driver_id)
1559{
1560	struct ib_device *ib_dev;
1561	unsigned long index;
1562
1563	down_read(&devices_rwsem);
1564	xa_for_each (&devices, index, ib_dev) {
1565		if (ib_dev->ops.driver_id != driver_id)
1566			continue;
1567
1568		get_device(&ib_dev->dev);
1569		up_read(&devices_rwsem);
1570
1571		WARN_ON(!ib_dev->ops.dealloc_driver);
1572		__ib_unregister_device(ib_dev);
1573
1574		put_device(&ib_dev->dev);
1575		down_read(&devices_rwsem);
1576	}
1577	up_read(&devices_rwsem);
1578}
1579EXPORT_SYMBOL(ib_unregister_driver);
1580
1581static void ib_unregister_work(struct work_struct *work)
1582{
1583	struct ib_device *ib_dev =
1584		container_of(work, struct ib_device, unregistration_work);
1585
1586	__ib_unregister_device(ib_dev);
1587	put_device(&ib_dev->dev);
1588}
1589
1590/**
1591 * ib_unregister_device_queued - Unregister a device using a work queue
1592 * @ib_dev: The device to unregister
1593 *
1594 * This schedules an asynchronous unregistration using a WQ for the device. A
1595 * driver should use this to avoid holding locks while doing unregistration,
1596 * such as holding the RTNL lock.
1597 *
1598 * Drivers using this API must use ib_unregister_driver before module unload
1599 * to ensure that all scheduled unregistrations have completed.
1600 */
1601void ib_unregister_device_queued(struct ib_device *ib_dev)
1602{
1603	WARN_ON(!refcount_read(&ib_dev->refcount));
1604	WARN_ON(!ib_dev->ops.dealloc_driver);
1605	get_device(&ib_dev->dev);
1606	if (!queue_work(ib_unreg_wq, &ib_dev->unregistration_work))
1607		put_device(&ib_dev->dev);
1608}
1609EXPORT_SYMBOL(ib_unregister_device_queued);
1610
1611/*
1612 * The caller must pass in a device that has the kref held and the refcount
1613 * released. If the device is in cur_net and still registered then it is moved
1614 * into net.
1615 */
1616static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1617				 struct net *net)
1618{
1619	int ret2 = -EINVAL;
1620	int ret;
1621
1622	mutex_lock(&device->unregistration_lock);
1623
1624	/*
1625	 * If a device not under ib_device_get() or if the unregistration_lock
1626	 * is not held, the namespace can be changed, or it can be unregistered.
1627	 * Check again under the lock.
1628	 */
1629	if (refcount_read(&device->refcount) == 0 ||
1630	    !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1631		ret = -ENODEV;
1632		goto out;
1633	}
1634
1635	kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1636	disable_device(device);
1637
1638	/*
1639	 * At this point no one can be using the device, so it is safe to
1640	 * change the namespace.
1641	 */
1642	write_pnet(&device->coredev.rdma_net, net);
1643
1644	down_read(&devices_rwsem);
1645	/*
1646	 * Currently rdma devices are system wide unique. So the device name
1647	 * is guaranteed free in the new namespace. Publish the new namespace
1648	 * at the sysfs level.
1649	 */
1650	ret = device_rename(&device->dev, dev_name(&device->dev));
1651	up_read(&devices_rwsem);
1652	if (ret) {
1653		dev_warn(&device->dev,
1654			 "%s: Couldn't rename device after namespace change\n",
1655			 __func__);
1656		/* Try and put things back and re-enable the device */
1657		write_pnet(&device->coredev.rdma_net, cur_net);
1658	}
1659
1660	ret2 = enable_device_and_get(device);
1661	if (ret2) {
1662		/*
1663		 * This shouldn't really happen, but if it does, let the user
1664		 * retry at later point. So don't disable the device.
1665		 */
1666		dev_warn(&device->dev,
1667			 "%s: Couldn't re-enable device after namespace change\n",
1668			 __func__);
1669	}
1670	kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1671
1672	ib_device_put(device);
1673out:
1674	mutex_unlock(&device->unregistration_lock);
1675	if (ret)
1676		return ret;
1677	return ret2;
1678}
1679
1680int ib_device_set_netns_put(struct sk_buff *skb,
1681			    struct ib_device *dev, u32 ns_fd)
1682{
1683	struct net *net;
1684	int ret;
1685
1686	net = get_net_ns_by_fd(ns_fd);
1687	if (IS_ERR(net)) {
1688		ret = PTR_ERR(net);
1689		goto net_err;
1690	}
1691
1692	if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1693		ret = -EPERM;
1694		goto ns_err;
1695	}
1696
1697	/*
1698	 * All the ib_clients, including uverbs, are reset when the namespace is
1699	 * changed and this cannot be blocked waiting for userspace to do
1700	 * something, so disassociation is mandatory.
 
1701	 */
1702	if (!dev->ops.disassociate_ucontext || ib_devices_shared_netns) {
 
1703		ret = -EOPNOTSUPP;
1704		goto ns_err;
1705	}
1706
1707	get_device(&dev->dev);
1708	ib_device_put(dev);
1709	ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1710	put_device(&dev->dev);
1711
1712	put_net(net);
1713	return ret;
1714
1715ns_err:
1716	put_net(net);
1717net_err:
1718	ib_device_put(dev);
1719	return ret;
1720}
1721
1722static struct pernet_operations rdma_dev_net_ops = {
1723	.init = rdma_dev_init_net,
1724	.exit = rdma_dev_exit_net,
1725	.id = &rdma_dev_net_id,
1726	.size = sizeof(struct rdma_dev_net),
1727};
1728
1729static int assign_client_id(struct ib_client *client)
1730{
1731	int ret;
1732
1733	down_write(&clients_rwsem);
1734	/*
1735	 * The add/remove callbacks must be called in FIFO/LIFO order. To
1736	 * achieve this we assign client_ids so they are sorted in
1737	 * registration order.
1738	 */
1739	client->client_id = highest_client_id;
1740	ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1741	if (ret)
1742		goto out;
1743
1744	highest_client_id++;
1745	xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1746
1747out:
1748	up_write(&clients_rwsem);
1749	return ret;
1750}
1751
1752static void remove_client_id(struct ib_client *client)
1753{
1754	down_write(&clients_rwsem);
1755	xa_erase(&clients, client->client_id);
1756	for (; highest_client_id; highest_client_id--)
1757		if (xa_load(&clients, highest_client_id - 1))
1758			break;
1759	up_write(&clients_rwsem);
1760}
1761
1762/**
1763 * ib_register_client - Register an IB client
1764 * @client:Client to register
1765 *
1766 * Upper level users of the IB drivers can use ib_register_client() to
1767 * register callbacks for IB device addition and removal.  When an IB
1768 * device is added, each registered client's add method will be called
1769 * (in the order the clients were registered), and when a device is
1770 * removed, each client's remove method will be called (in the reverse
1771 * order that clients were registered).  In addition, when
1772 * ib_register_client() is called, the client will receive an add
1773 * callback for all devices already registered.
1774 */
1775int ib_register_client(struct ib_client *client)
1776{
1777	struct ib_device *device;
1778	unsigned long index;
1779	int ret;
1780
1781	refcount_set(&client->uses, 1);
1782	init_completion(&client->uses_zero);
1783	ret = assign_client_id(client);
1784	if (ret)
1785		return ret;
1786
1787	down_read(&devices_rwsem);
1788	xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1789		ret = add_client_context(device, client);
1790		if (ret) {
1791			up_read(&devices_rwsem);
1792			ib_unregister_client(client);
1793			return ret;
1794		}
1795	}
1796	up_read(&devices_rwsem);
1797	return 0;
1798}
1799EXPORT_SYMBOL(ib_register_client);
1800
1801/**
1802 * ib_unregister_client - Unregister an IB client
1803 * @client:Client to unregister
1804 *
1805 * Upper level users use ib_unregister_client() to remove their client
1806 * registration.  When ib_unregister_client() is called, the client
1807 * will receive a remove callback for each IB device still registered.
1808 *
1809 * This is a full fence, once it returns no client callbacks will be called,
1810 * or are running in another thread.
1811 */
1812void ib_unregister_client(struct ib_client *client)
1813{
1814	struct ib_device *device;
1815	unsigned long index;
1816
1817	down_write(&clients_rwsem);
1818	ib_client_put(client);
1819	xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1820	up_write(&clients_rwsem);
1821
1822	/* We do not want to have locks while calling client->remove() */
1823	rcu_read_lock();
1824	xa_for_each (&devices, index, device) {
1825		if (!ib_device_try_get(device))
1826			continue;
1827		rcu_read_unlock();
1828
1829		remove_client_context(device, client->client_id);
1830
1831		ib_device_put(device);
1832		rcu_read_lock();
1833	}
1834	rcu_read_unlock();
1835
1836	/*
1837	 * remove_client_context() is not a fence, it can return even though a
1838	 * removal is ongoing. Wait until all removals are completed.
1839	 */
1840	wait_for_completion(&client->uses_zero);
1841	remove_client_id(client);
1842}
1843EXPORT_SYMBOL(ib_unregister_client);
1844
1845static int __ib_get_global_client_nl_info(const char *client_name,
1846					  struct ib_client_nl_info *res)
1847{
1848	struct ib_client *client;
1849	unsigned long index;
1850	int ret = -ENOENT;
1851
1852	down_read(&clients_rwsem);
1853	xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1854		if (strcmp(client->name, client_name) != 0)
1855			continue;
1856		if (!client->get_global_nl_info) {
1857			ret = -EOPNOTSUPP;
1858			break;
1859		}
1860		ret = client->get_global_nl_info(res);
1861		if (WARN_ON(ret == -ENOENT))
1862			ret = -EINVAL;
1863		if (!ret && res->cdev)
1864			get_device(res->cdev);
1865		break;
1866	}
1867	up_read(&clients_rwsem);
1868	return ret;
1869}
1870
1871static int __ib_get_client_nl_info(struct ib_device *ibdev,
1872				   const char *client_name,
1873				   struct ib_client_nl_info *res)
1874{
1875	unsigned long index;
1876	void *client_data;
1877	int ret = -ENOENT;
1878
1879	down_read(&ibdev->client_data_rwsem);
1880	xan_for_each_marked (&ibdev->client_data, index, client_data,
1881			     CLIENT_DATA_REGISTERED) {
1882		struct ib_client *client = xa_load(&clients, index);
1883
1884		if (!client || strcmp(client->name, client_name) != 0)
1885			continue;
1886		if (!client->get_nl_info) {
1887			ret = -EOPNOTSUPP;
1888			break;
1889		}
1890		ret = client->get_nl_info(ibdev, client_data, res);
1891		if (WARN_ON(ret == -ENOENT))
1892			ret = -EINVAL;
1893
1894		/*
1895		 * The cdev is guaranteed valid as long as we are inside the
1896		 * client_data_rwsem as remove_one can't be called. Keep it
1897		 * valid for the caller.
1898		 */
1899		if (!ret && res->cdev)
1900			get_device(res->cdev);
1901		break;
1902	}
1903	up_read(&ibdev->client_data_rwsem);
1904
1905	return ret;
1906}
1907
1908/**
1909 * ib_get_client_nl_info - Fetch the nl_info from a client
1910 * @ibdev: IB device
1911 * @client_name: Name of the client
1912 * @res: Result of the query
1913 */
1914int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1915			  struct ib_client_nl_info *res)
1916{
1917	int ret;
1918
1919	if (ibdev)
1920		ret = __ib_get_client_nl_info(ibdev, client_name, res);
1921	else
1922		ret = __ib_get_global_client_nl_info(client_name, res);
1923#ifdef CONFIG_MODULES
1924	if (ret == -ENOENT) {
1925		request_module("rdma-client-%s", client_name);
1926		if (ibdev)
1927			ret = __ib_get_client_nl_info(ibdev, client_name, res);
1928		else
1929			ret = __ib_get_global_client_nl_info(client_name, res);
1930	}
1931#endif
1932	if (ret) {
1933		if (ret == -ENOENT)
1934			return -EOPNOTSUPP;
1935		return ret;
1936	}
1937
1938	if (WARN_ON(!res->cdev))
1939		return -EINVAL;
1940	return 0;
1941}
1942
1943/**
1944 * ib_set_client_data - Set IB client context
1945 * @device:Device to set context for
1946 * @client:Client to set context for
1947 * @data:Context to set
1948 *
1949 * ib_set_client_data() sets client context data that can be retrieved with
1950 * ib_get_client_data(). This can only be called while the client is
1951 * registered to the device, once the ib_client remove() callback returns this
1952 * cannot be called.
1953 */
1954void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1955			void *data)
1956{
1957	void *rc;
1958
1959	if (WARN_ON(IS_ERR(data)))
1960		data = NULL;
1961
1962	rc = xa_store(&device->client_data, client->client_id, data,
1963		      GFP_KERNEL);
1964	WARN_ON(xa_is_err(rc));
1965}
1966EXPORT_SYMBOL(ib_set_client_data);
1967
1968/**
1969 * ib_register_event_handler - Register an IB event handler
1970 * @event_handler:Handler to register
1971 *
1972 * ib_register_event_handler() registers an event handler that will be
1973 * called back when asynchronous IB events occur (as defined in
1974 * chapter 11 of the InfiniBand Architecture Specification). This
1975 * callback occurs in workqueue context.
1976 */
1977void ib_register_event_handler(struct ib_event_handler *event_handler)
1978{
1979	down_write(&event_handler->device->event_handler_rwsem);
 
 
1980	list_add_tail(&event_handler->list,
1981		      &event_handler->device->event_handler_list);
1982	up_write(&event_handler->device->event_handler_rwsem);
1983}
1984EXPORT_SYMBOL(ib_register_event_handler);
1985
1986/**
1987 * ib_unregister_event_handler - Unregister an event handler
1988 * @event_handler:Handler to unregister
1989 *
1990 * Unregister an event handler registered with
1991 * ib_register_event_handler().
1992 */
1993void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1994{
1995	down_write(&event_handler->device->event_handler_rwsem);
 
 
1996	list_del(&event_handler->list);
1997	up_write(&event_handler->device->event_handler_rwsem);
1998}
1999EXPORT_SYMBOL(ib_unregister_event_handler);
2000
2001void ib_dispatch_event_clients(struct ib_event *event)
 
 
 
 
 
 
 
 
2002{
 
2003	struct ib_event_handler *handler;
2004
2005	down_read(&event->device->event_handler_rwsem);
2006
2007	list_for_each_entry(handler, &event->device->event_handler_list, list)
2008		handler->handler(handler, event);
2009
2010	up_read(&event->device->event_handler_rwsem);
2011}
 
2012
2013static int iw_query_port(struct ib_device *device,
2014			   u32 port_num,
2015			   struct ib_port_attr *port_attr)
2016{
2017	struct in_device *inetdev;
2018	struct net_device *netdev;
 
2019
2020	memset(port_attr, 0, sizeof(*port_attr));
2021
2022	netdev = ib_device_get_netdev(device, port_num);
2023	if (!netdev)
2024		return -ENODEV;
2025
2026	port_attr->max_mtu = IB_MTU_4096;
2027	port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
2028
2029	if (!netif_carrier_ok(netdev)) {
2030		port_attr->state = IB_PORT_DOWN;
2031		port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
2032	} else {
2033		rcu_read_lock();
2034		inetdev = __in_dev_get_rcu(netdev);
2035
2036		if (inetdev && inetdev->ifa_list) {
2037			port_attr->state = IB_PORT_ACTIVE;
2038			port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2039		} else {
2040			port_attr->state = IB_PORT_INIT;
2041			port_attr->phys_state =
2042				IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2043		}
2044
2045		rcu_read_unlock();
2046	}
2047
2048	dev_put(netdev);
2049	return device->ops.query_port(device, port_num, port_attr);
 
 
 
 
2050}
2051
2052static int __ib_query_port(struct ib_device *device,
2053			   u32 port_num,
2054			   struct ib_port_attr *port_attr)
2055{
 
2056	int err;
2057
2058	memset(port_attr, 0, sizeof(*port_attr));
2059
2060	err = device->ops.query_port(device, port_num, port_attr);
2061	if (err || port_attr->subnet_prefix)
2062		return err;
2063
2064	if (rdma_port_get_link_layer(device, port_num) !=
2065	    IB_LINK_LAYER_INFINIBAND)
2066		return 0;
2067
2068	ib_get_cached_subnet_prefix(device, port_num,
2069				    &port_attr->subnet_prefix);
 
 
 
2070	return 0;
2071}
2072
2073/**
2074 * ib_query_port - Query IB port attributes
2075 * @device:Device to query
2076 * @port_num:Port number to query
2077 * @port_attr:Port attributes
2078 *
2079 * ib_query_port() returns the attributes of a port through the
2080 * @port_attr pointer.
2081 */
2082int ib_query_port(struct ib_device *device,
2083		  u32 port_num,
2084		  struct ib_port_attr *port_attr)
2085{
2086	if (!rdma_is_port_valid(device, port_num))
2087		return -EINVAL;
2088
2089	if (rdma_protocol_iwarp(device, port_num))
2090		return iw_query_port(device, port_num, port_attr);
2091	else
2092		return __ib_query_port(device, port_num, port_attr);
2093}
2094EXPORT_SYMBOL(ib_query_port);
2095
2096static void add_ndev_hash(struct ib_port_data *pdata)
2097{
2098	unsigned long flags;
2099
2100	might_sleep();
2101
2102	spin_lock_irqsave(&ndev_hash_lock, flags);
2103	if (hash_hashed(&pdata->ndev_hash_link)) {
2104		hash_del_rcu(&pdata->ndev_hash_link);
2105		spin_unlock_irqrestore(&ndev_hash_lock, flags);
2106		/*
2107		 * We cannot do hash_add_rcu after a hash_del_rcu until the
2108		 * grace period
2109		 */
2110		synchronize_rcu();
2111		spin_lock_irqsave(&ndev_hash_lock, flags);
2112	}
2113	if (pdata->netdev)
2114		hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2115			     (uintptr_t)pdata->netdev);
2116	spin_unlock_irqrestore(&ndev_hash_lock, flags);
2117}
2118
2119/**
2120 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2121 * @ib_dev: Device to modify
2122 * @ndev: net_device to affiliate, may be NULL
2123 * @port: IB port the net_device is connected to
2124 *
2125 * Drivers should use this to link the ib_device to a netdev so the netdev
2126 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2127 * affiliated with any port.
2128 *
2129 * The caller must ensure that the given ndev is not unregistered or
2130 * unregistering, and that either the ib_device is unregistered or
2131 * ib_device_set_netdev() is called with NULL when the ndev sends a
2132 * NETDEV_UNREGISTER event.
2133 */
2134int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2135			 u32 port)
2136{
2137	struct net_device *old_ndev;
2138	struct ib_port_data *pdata;
2139	unsigned long flags;
2140	int ret;
2141
2142	/*
2143	 * Drivers wish to call this before ib_register_driver, so we have to
2144	 * setup the port data early.
2145	 */
2146	ret = alloc_port_data(ib_dev);
2147	if (ret)
2148		return ret;
2149
2150	if (!rdma_is_port_valid(ib_dev, port))
2151		return -EINVAL;
2152
2153	pdata = &ib_dev->port_data[port];
2154	spin_lock_irqsave(&pdata->netdev_lock, flags);
2155	old_ndev = rcu_dereference_protected(
2156		pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2157	if (old_ndev == ndev) {
2158		spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2159		return 0;
2160	}
2161
2162	if (old_ndev)
2163		netdev_tracker_free(ndev, &pdata->netdev_tracker);
2164	if (ndev)
2165		netdev_hold(ndev, &pdata->netdev_tracker, GFP_ATOMIC);
2166	rcu_assign_pointer(pdata->netdev, ndev);
2167	spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2168
2169	add_ndev_hash(pdata);
2170	if (old_ndev)
2171		__dev_put(old_ndev);
2172
2173	return 0;
2174}
2175EXPORT_SYMBOL(ib_device_set_netdev);
2176
2177static void free_netdevs(struct ib_device *ib_dev)
2178{
2179	unsigned long flags;
2180	u32 port;
2181
2182	if (!ib_dev->port_data)
2183		return;
2184
2185	rdma_for_each_port (ib_dev, port) {
2186		struct ib_port_data *pdata = &ib_dev->port_data[port];
2187		struct net_device *ndev;
2188
2189		spin_lock_irqsave(&pdata->netdev_lock, flags);
2190		ndev = rcu_dereference_protected(
2191			pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2192		if (ndev) {
2193			spin_lock(&ndev_hash_lock);
2194			hash_del_rcu(&pdata->ndev_hash_link);
2195			spin_unlock(&ndev_hash_lock);
2196
2197			/*
2198			 * If this is the last dev_put there is still a
2199			 * synchronize_rcu before the netdev is kfreed, so we
2200			 * can continue to rely on unlocked pointer
2201			 * comparisons after the put
2202			 */
2203			rcu_assign_pointer(pdata->netdev, NULL);
2204			netdev_put(ndev, &pdata->netdev_tracker);
2205		}
2206		spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2207	}
2208}
2209
2210struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2211					u32 port)
2212{
2213	struct ib_port_data *pdata;
2214	struct net_device *res;
2215
2216	if (!rdma_is_port_valid(ib_dev, port))
2217		return NULL;
2218
2219	pdata = &ib_dev->port_data[port];
2220
2221	/*
2222	 * New drivers should use ib_device_set_netdev() not the legacy
2223	 * get_netdev().
2224	 */
2225	if (ib_dev->ops.get_netdev)
2226		res = ib_dev->ops.get_netdev(ib_dev, port);
2227	else {
2228		spin_lock(&pdata->netdev_lock);
2229		res = rcu_dereference_protected(
2230			pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2231		if (res)
2232			dev_hold(res);
2233		spin_unlock(&pdata->netdev_lock);
2234	}
2235
2236	/*
2237	 * If we are starting to unregister expedite things by preventing
2238	 * propagation of an unregistering netdev.
2239	 */
2240	if (res && res->reg_state != NETREG_REGISTERED) {
2241		dev_put(res);
2242		return NULL;
2243	}
2244
2245	return res;
2246}
2247
2248/**
2249 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2250 * @ndev: netdev to locate
2251 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2252 *
2253 * Find and hold an ib_device that is associated with a netdev via
2254 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2255 * returned pointer.
2256 */
2257struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2258					  enum rdma_driver_id driver_id)
2259{
2260	struct ib_device *res = NULL;
2261	struct ib_port_data *cur;
2262
2263	rcu_read_lock();
2264	hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2265				    (uintptr_t)ndev) {
2266		if (rcu_access_pointer(cur->netdev) == ndev &&
2267		    (driver_id == RDMA_DRIVER_UNKNOWN ||
2268		     cur->ib_dev->ops.driver_id == driver_id) &&
2269		    ib_device_try_get(cur->ib_dev)) {
2270			res = cur->ib_dev;
2271			break;
2272		}
2273	}
2274	rcu_read_unlock();
2275
2276	return res;
2277}
2278EXPORT_SYMBOL(ib_device_get_by_netdev);
2279
2280/**
2281 * ib_enum_roce_netdev - enumerate all RoCE ports
2282 * @ib_dev : IB device we want to query
2283 * @filter: Should we call the callback?
2284 * @filter_cookie: Cookie passed to filter
2285 * @cb: Callback to call for each found RoCE ports
2286 * @cookie: Cookie passed back to the callback
2287 *
2288 * Enumerates all of the physical RoCE ports of ib_dev
2289 * which are related to netdevice and calls callback() on each
2290 * device for which filter() function returns non zero.
2291 */
2292void ib_enum_roce_netdev(struct ib_device *ib_dev,
2293			 roce_netdev_filter filter,
2294			 void *filter_cookie,
2295			 roce_netdev_callback cb,
2296			 void *cookie)
2297{
2298	u32 port;
2299
2300	rdma_for_each_port (ib_dev, port)
2301		if (rdma_protocol_roce(ib_dev, port)) {
2302			struct net_device *idev =
2303				ib_device_get_netdev(ib_dev, port);
2304
2305			if (filter(ib_dev, port, idev, filter_cookie))
2306				cb(ib_dev, port, idev, cookie);
2307
2308			if (idev)
2309				dev_put(idev);
2310		}
2311}
2312
2313/**
2314 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2315 * @filter: Should we call the callback?
2316 * @filter_cookie: Cookie passed to filter
2317 * @cb: Callback to call for each found RoCE ports
2318 * @cookie: Cookie passed back to the callback
2319 *
2320 * Enumerates all RoCE devices' physical ports which are related
2321 * to netdevices and calls callback() on each device for which
2322 * filter() function returns non zero.
2323 */
2324void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2325			      void *filter_cookie,
2326			      roce_netdev_callback cb,
2327			      void *cookie)
2328{
2329	struct ib_device *dev;
2330	unsigned long index;
2331
2332	down_read(&devices_rwsem);
2333	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2334		ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2335	up_read(&devices_rwsem);
2336}
2337
2338/*
2339 * ib_enum_all_devs - enumerate all ib_devices
2340 * @cb: Callback to call for each found ib_device
2341 *
2342 * Enumerates all ib_devices and calls callback() on each device.
2343 */
2344int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2345		     struct netlink_callback *cb)
2346{
2347	unsigned long index;
2348	struct ib_device *dev;
2349	unsigned int idx = 0;
2350	int ret = 0;
2351
2352	down_read(&devices_rwsem);
2353	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2354		if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2355			continue;
2356
2357		ret = nldev_cb(dev, skb, cb, idx);
2358		if (ret)
2359			break;
2360		idx++;
2361	}
2362	up_read(&devices_rwsem);
2363	return ret;
2364}
2365
2366/**
2367 * ib_query_pkey - Get P_Key table entry
2368 * @device:Device to query
2369 * @port_num:Port number to query
2370 * @index:P_Key table index to query
2371 * @pkey:Returned P_Key
2372 *
2373 * ib_query_pkey() fetches the specified P_Key table entry.
2374 */
2375int ib_query_pkey(struct ib_device *device,
2376		  u32 port_num, u16 index, u16 *pkey)
2377{
2378	if (!rdma_is_port_valid(device, port_num))
2379		return -EINVAL;
2380
2381	if (!device->ops.query_pkey)
2382		return -EOPNOTSUPP;
2383
2384	return device->ops.query_pkey(device, port_num, index, pkey);
2385}
2386EXPORT_SYMBOL(ib_query_pkey);
2387
2388/**
2389 * ib_modify_device - Change IB device attributes
2390 * @device:Device to modify
2391 * @device_modify_mask:Mask of attributes to change
2392 * @device_modify:New attribute values
2393 *
2394 * ib_modify_device() changes a device's attributes as specified by
2395 * the @device_modify_mask and @device_modify structure.
2396 */
2397int ib_modify_device(struct ib_device *device,
2398		     int device_modify_mask,
2399		     struct ib_device_modify *device_modify)
2400{
2401	if (!device->ops.modify_device)
2402		return -EOPNOTSUPP;
2403
2404	return device->ops.modify_device(device, device_modify_mask,
2405					 device_modify);
2406}
2407EXPORT_SYMBOL(ib_modify_device);
2408
2409/**
2410 * ib_modify_port - Modifies the attributes for the specified port.
2411 * @device: The device to modify.
2412 * @port_num: The number of the port to modify.
2413 * @port_modify_mask: Mask used to specify which attributes of the port
2414 *   to change.
2415 * @port_modify: New attribute values for the port.
2416 *
2417 * ib_modify_port() changes a port's attributes as specified by the
2418 * @port_modify_mask and @port_modify structure.
2419 */
2420int ib_modify_port(struct ib_device *device,
2421		   u32 port_num, int port_modify_mask,
2422		   struct ib_port_modify *port_modify)
2423{
2424	int rc;
2425
2426	if (!rdma_is_port_valid(device, port_num))
2427		return -EINVAL;
2428
2429	if (device->ops.modify_port)
2430		rc = device->ops.modify_port(device, port_num,
2431					     port_modify_mask,
2432					     port_modify);
2433	else if (rdma_protocol_roce(device, port_num) &&
2434		 ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 ||
2435		  (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
2436		rc = 0;
2437	else
2438		rc = -EOPNOTSUPP;
2439	return rc;
2440}
2441EXPORT_SYMBOL(ib_modify_port);
2442
2443/**
2444 * ib_find_gid - Returns the port number and GID table index where
2445 *   a specified GID value occurs. Its searches only for IB link layer.
2446 * @device: The device to query.
2447 * @gid: The GID value to search for.
2448 * @port_num: The port number of the device where the GID value was found.
2449 * @index: The index into the GID table where the GID was found.  This
2450 *   parameter may be NULL.
2451 */
2452int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2453		u32 *port_num, u16 *index)
2454{
2455	union ib_gid tmp_gid;
2456	u32 port;
2457	int ret, i;
2458
2459	rdma_for_each_port (device, port) {
2460		if (!rdma_protocol_ib(device, port))
2461			continue;
2462
2463		for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2464		     ++i) {
2465			ret = rdma_query_gid(device, port, i, &tmp_gid);
2466			if (ret)
2467				continue;
2468
2469			if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2470				*port_num = port;
2471				if (index)
2472					*index = i;
2473				return 0;
2474			}
2475		}
2476	}
2477
2478	return -ENOENT;
2479}
2480EXPORT_SYMBOL(ib_find_gid);
2481
2482/**
2483 * ib_find_pkey - Returns the PKey table index where a specified
2484 *   PKey value occurs.
2485 * @device: The device to query.
2486 * @port_num: The port number of the device to search for the PKey.
2487 * @pkey: The PKey value to search for.
2488 * @index: The index into the PKey table where the PKey was found.
2489 */
2490int ib_find_pkey(struct ib_device *device,
2491		 u32 port_num, u16 pkey, u16 *index)
2492{
2493	int ret, i;
2494	u16 tmp_pkey;
2495	int partial_ix = -1;
2496
2497	for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2498	     ++i) {
2499		ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2500		if (ret)
2501			return ret;
2502		if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2503			/* if there is full-member pkey take it.*/
2504			if (tmp_pkey & 0x8000) {
2505				*index = i;
2506				return 0;
2507			}
2508			if (partial_ix < 0)
2509				partial_ix = i;
2510		}
2511	}
2512
2513	/*no full-member, if exists take the limited*/
2514	if (partial_ix >= 0) {
2515		*index = partial_ix;
2516		return 0;
2517	}
2518	return -ENOENT;
2519}
2520EXPORT_SYMBOL(ib_find_pkey);
2521
2522/**
2523 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2524 * for a received CM request
2525 * @dev:	An RDMA device on which the request has been received.
2526 * @port:	Port number on the RDMA device.
2527 * @pkey:	The Pkey the request came on.
2528 * @gid:	A GID that the net_dev uses to communicate.
2529 * @addr:	Contains the IP address that the request specified as its
2530 *		destination.
2531 *
2532 */
2533struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2534					    u32 port,
2535					    u16 pkey,
2536					    const union ib_gid *gid,
2537					    const struct sockaddr *addr)
2538{
2539	struct net_device *net_dev = NULL;
2540	unsigned long index;
2541	void *client_data;
2542
2543	if (!rdma_protocol_ib(dev, port))
2544		return NULL;
2545
2546	/*
2547	 * Holding the read side guarantees that the client will not become
2548	 * unregistered while we are calling get_net_dev_by_params()
2549	 */
2550	down_read(&dev->client_data_rwsem);
2551	xan_for_each_marked (&dev->client_data, index, client_data,
2552			     CLIENT_DATA_REGISTERED) {
2553		struct ib_client *client = xa_load(&clients, index);
2554
2555		if (!client || !client->get_net_dev_by_params)
2556			continue;
2557
2558		net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2559							addr, client_data);
2560		if (net_dev)
2561			break;
2562	}
2563	up_read(&dev->client_data_rwsem);
2564
2565	return net_dev;
2566}
2567EXPORT_SYMBOL(ib_get_net_dev_by_params);
2568
2569void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2570{
2571	struct ib_device_ops *dev_ops = &dev->ops;
2572#define SET_DEVICE_OP(ptr, name)                                               \
2573	do {                                                                   \
2574		if (ops->name)                                                 \
2575			if (!((ptr)->name))				       \
2576				(ptr)->name = ops->name;                       \
2577	} while (0)
2578
2579#define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2580
2581	if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2582		WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2583			dev_ops->driver_id != ops->driver_id);
2584		dev_ops->driver_id = ops->driver_id;
2585	}
2586	if (ops->owner) {
2587		WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2588		dev_ops->owner = ops->owner;
2589	}
2590	if (ops->uverbs_abi_ver)
2591		dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2592
2593	dev_ops->uverbs_no_driver_id_binding |=
2594		ops->uverbs_no_driver_id_binding;
2595
2596	SET_DEVICE_OP(dev_ops, add_gid);
2597	SET_DEVICE_OP(dev_ops, advise_mr);
2598	SET_DEVICE_OP(dev_ops, alloc_dm);
2599	SET_DEVICE_OP(dev_ops, alloc_hw_device_stats);
2600	SET_DEVICE_OP(dev_ops, alloc_hw_port_stats);
2601	SET_DEVICE_OP(dev_ops, alloc_mr);
2602	SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2603	SET_DEVICE_OP(dev_ops, alloc_mw);
2604	SET_DEVICE_OP(dev_ops, alloc_pd);
2605	SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2606	SET_DEVICE_OP(dev_ops, alloc_ucontext);
2607	SET_DEVICE_OP(dev_ops, alloc_xrcd);
2608	SET_DEVICE_OP(dev_ops, attach_mcast);
2609	SET_DEVICE_OP(dev_ops, check_mr_status);
2610	SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2611	SET_DEVICE_OP(dev_ops, counter_bind_qp);
2612	SET_DEVICE_OP(dev_ops, counter_dealloc);
2613	SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2614	SET_DEVICE_OP(dev_ops, counter_update_stats);
2615	SET_DEVICE_OP(dev_ops, create_ah);
2616	SET_DEVICE_OP(dev_ops, create_counters);
2617	SET_DEVICE_OP(dev_ops, create_cq);
2618	SET_DEVICE_OP(dev_ops, create_flow);
 
2619	SET_DEVICE_OP(dev_ops, create_qp);
2620	SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2621	SET_DEVICE_OP(dev_ops, create_srq);
2622	SET_DEVICE_OP(dev_ops, create_user_ah);
2623	SET_DEVICE_OP(dev_ops, create_wq);
2624	SET_DEVICE_OP(dev_ops, dealloc_dm);
2625	SET_DEVICE_OP(dev_ops, dealloc_driver);
 
2626	SET_DEVICE_OP(dev_ops, dealloc_mw);
2627	SET_DEVICE_OP(dev_ops, dealloc_pd);
2628	SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2629	SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2630	SET_DEVICE_OP(dev_ops, del_gid);
2631	SET_DEVICE_OP(dev_ops, dereg_mr);
2632	SET_DEVICE_OP(dev_ops, destroy_ah);
2633	SET_DEVICE_OP(dev_ops, destroy_counters);
2634	SET_DEVICE_OP(dev_ops, destroy_cq);
2635	SET_DEVICE_OP(dev_ops, destroy_flow);
2636	SET_DEVICE_OP(dev_ops, destroy_flow_action);
2637	SET_DEVICE_OP(dev_ops, destroy_qp);
2638	SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2639	SET_DEVICE_OP(dev_ops, destroy_srq);
2640	SET_DEVICE_OP(dev_ops, destroy_wq);
2641	SET_DEVICE_OP(dev_ops, device_group);
2642	SET_DEVICE_OP(dev_ops, detach_mcast);
2643	SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2644	SET_DEVICE_OP(dev_ops, drain_rq);
2645	SET_DEVICE_OP(dev_ops, drain_sq);
2646	SET_DEVICE_OP(dev_ops, enable_driver);
2647	SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
2648	SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
2649	SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
2650	SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
2651	SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
2652	SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
2653	SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
2654	SET_DEVICE_OP(dev_ops, fill_res_srq_entry);
2655	SET_DEVICE_OP(dev_ops, fill_res_srq_entry_raw);
2656	SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
2657	SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2658	SET_DEVICE_OP(dev_ops, get_dma_mr);
2659	SET_DEVICE_OP(dev_ops, get_hw_stats);
2660	SET_DEVICE_OP(dev_ops, get_link_layer);
2661	SET_DEVICE_OP(dev_ops, get_netdev);
2662	SET_DEVICE_OP(dev_ops, get_numa_node);
2663	SET_DEVICE_OP(dev_ops, get_port_immutable);
2664	SET_DEVICE_OP(dev_ops, get_vector_affinity);
2665	SET_DEVICE_OP(dev_ops, get_vf_config);
2666	SET_DEVICE_OP(dev_ops, get_vf_guid);
2667	SET_DEVICE_OP(dev_ops, get_vf_stats);
 
 
2668	SET_DEVICE_OP(dev_ops, iw_accept);
2669	SET_DEVICE_OP(dev_ops, iw_add_ref);
2670	SET_DEVICE_OP(dev_ops, iw_connect);
2671	SET_DEVICE_OP(dev_ops, iw_create_listen);
2672	SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2673	SET_DEVICE_OP(dev_ops, iw_get_qp);
2674	SET_DEVICE_OP(dev_ops, iw_reject);
2675	SET_DEVICE_OP(dev_ops, iw_rem_ref);
2676	SET_DEVICE_OP(dev_ops, map_mr_sg);
2677	SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
 
2678	SET_DEVICE_OP(dev_ops, mmap);
2679	SET_DEVICE_OP(dev_ops, mmap_free);
2680	SET_DEVICE_OP(dev_ops, modify_ah);
2681	SET_DEVICE_OP(dev_ops, modify_cq);
2682	SET_DEVICE_OP(dev_ops, modify_device);
2683	SET_DEVICE_OP(dev_ops, modify_hw_stat);
2684	SET_DEVICE_OP(dev_ops, modify_port);
2685	SET_DEVICE_OP(dev_ops, modify_qp);
2686	SET_DEVICE_OP(dev_ops, modify_srq);
2687	SET_DEVICE_OP(dev_ops, modify_wq);
2688	SET_DEVICE_OP(dev_ops, peek_cq);
2689	SET_DEVICE_OP(dev_ops, poll_cq);
2690	SET_DEVICE_OP(dev_ops, port_groups);
2691	SET_DEVICE_OP(dev_ops, post_recv);
2692	SET_DEVICE_OP(dev_ops, post_send);
2693	SET_DEVICE_OP(dev_ops, post_srq_recv);
2694	SET_DEVICE_OP(dev_ops, process_mad);
2695	SET_DEVICE_OP(dev_ops, query_ah);
2696	SET_DEVICE_OP(dev_ops, query_device);
2697	SET_DEVICE_OP(dev_ops, query_gid);
2698	SET_DEVICE_OP(dev_ops, query_pkey);
2699	SET_DEVICE_OP(dev_ops, query_port);
2700	SET_DEVICE_OP(dev_ops, query_qp);
2701	SET_DEVICE_OP(dev_ops, query_srq);
2702	SET_DEVICE_OP(dev_ops, query_ucontext);
2703	SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2704	SET_DEVICE_OP(dev_ops, read_counters);
2705	SET_DEVICE_OP(dev_ops, reg_dm_mr);
2706	SET_DEVICE_OP(dev_ops, reg_user_mr);
2707	SET_DEVICE_OP(dev_ops, reg_user_mr_dmabuf);
2708	SET_DEVICE_OP(dev_ops, req_notify_cq);
2709	SET_DEVICE_OP(dev_ops, rereg_user_mr);
2710	SET_DEVICE_OP(dev_ops, resize_cq);
2711	SET_DEVICE_OP(dev_ops, set_vf_guid);
2712	SET_DEVICE_OP(dev_ops, set_vf_link_state);
 
2713
2714	SET_OBJ_SIZE(dev_ops, ib_ah);
2715	SET_OBJ_SIZE(dev_ops, ib_counters);
2716	SET_OBJ_SIZE(dev_ops, ib_cq);
2717	SET_OBJ_SIZE(dev_ops, ib_mw);
2718	SET_OBJ_SIZE(dev_ops, ib_pd);
2719	SET_OBJ_SIZE(dev_ops, ib_qp);
2720	SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
2721	SET_OBJ_SIZE(dev_ops, ib_srq);
2722	SET_OBJ_SIZE(dev_ops, ib_ucontext);
2723	SET_OBJ_SIZE(dev_ops, ib_xrcd);
2724}
2725EXPORT_SYMBOL(ib_set_device_ops);
2726
2727#ifdef CONFIG_INFINIBAND_VIRT_DMA
2728int ib_dma_virt_map_sg(struct ib_device *dev, struct scatterlist *sg, int nents)
2729{
2730	struct scatterlist *s;
2731	int i;
2732
2733	for_each_sg(sg, s, nents, i) {
2734		sg_dma_address(s) = (uintptr_t)sg_virt(s);
2735		sg_dma_len(s) = s->length;
2736	}
2737	return nents;
2738}
2739EXPORT_SYMBOL(ib_dma_virt_map_sg);
2740#endif /* CONFIG_INFINIBAND_VIRT_DMA */
2741
2742static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2743	[RDMA_NL_LS_OP_RESOLVE] = {
2744		.doit = ib_nl_handle_resolve_resp,
2745		.flags = RDMA_NL_ADMIN_PERM,
2746	},
2747	[RDMA_NL_LS_OP_SET_TIMEOUT] = {
2748		.doit = ib_nl_handle_set_timeout,
2749		.flags = RDMA_NL_ADMIN_PERM,
2750	},
2751	[RDMA_NL_LS_OP_IP_RESOLVE] = {
2752		.doit = ib_nl_handle_ip_res_resp,
2753		.flags = RDMA_NL_ADMIN_PERM,
2754	},
2755};
2756
2757static int __init ib_core_init(void)
2758{
2759	int ret = -ENOMEM;
2760
2761	ib_wq = alloc_workqueue("infiniband", 0, 0);
2762	if (!ib_wq)
2763		return -ENOMEM;
2764
2765	ib_unreg_wq = alloc_workqueue("ib-unreg-wq", WQ_UNBOUND,
2766				      WQ_UNBOUND_MAX_ACTIVE);
2767	if (!ib_unreg_wq)
2768		goto err;
2769
2770	ib_comp_wq = alloc_workqueue("ib-comp-wq",
2771			WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2772	if (!ib_comp_wq)
2773		goto err_unbound;
 
 
2774
2775	ib_comp_unbound_wq =
2776		alloc_workqueue("ib-comp-unb-wq",
2777				WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2778				WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2779	if (!ib_comp_unbound_wq)
 
2780		goto err_comp;
 
2781
2782	ret = class_register(&ib_class);
2783	if (ret) {
2784		pr_warn("Couldn't create InfiniBand device class\n");
2785		goto err_comp_unbound;
2786	}
2787
2788	rdma_nl_init();
2789
2790	ret = addr_init();
2791	if (ret) {
2792		pr_warn("Couldn't init IB address resolution\n");
2793		goto err_ibnl;
2794	}
2795
2796	ret = ib_mad_init();
2797	if (ret) {
2798		pr_warn("Couldn't init IB MAD\n");
2799		goto err_addr;
2800	}
2801
2802	ret = ib_sa_init();
2803	if (ret) {
2804		pr_warn("Couldn't init SA\n");
2805		goto err_mad;
2806	}
2807
2808	ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2809	if (ret) {
2810		pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2811		goto err_sa;
2812	}
2813
2814	ret = register_pernet_device(&rdma_dev_net_ops);
2815	if (ret) {
2816		pr_warn("Couldn't init compat dev. ret %d\n", ret);
2817		goto err_compat;
2818	}
2819
2820	nldev_init();
2821	rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2822	ret = roce_gid_mgmt_init();
2823	if (ret) {
2824		pr_warn("Couldn't init RoCE GID management\n");
2825		goto err_parent;
2826	}
2827
2828	return 0;
2829
2830err_parent:
2831	rdma_nl_unregister(RDMA_NL_LS);
2832	nldev_exit();
2833	unregister_pernet_device(&rdma_dev_net_ops);
2834err_compat:
2835	unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2836err_sa:
2837	ib_sa_cleanup();
2838err_mad:
2839	ib_mad_cleanup();
2840err_addr:
2841	addr_cleanup();
2842err_ibnl:
2843	class_unregister(&ib_class);
2844err_comp_unbound:
2845	destroy_workqueue(ib_comp_unbound_wq);
2846err_comp:
2847	destroy_workqueue(ib_comp_wq);
2848err_unbound:
2849	destroy_workqueue(ib_unreg_wq);
2850err:
2851	destroy_workqueue(ib_wq);
2852	return ret;
2853}
2854
2855static void __exit ib_core_cleanup(void)
2856{
2857	roce_gid_mgmt_cleanup();
2858	rdma_nl_unregister(RDMA_NL_LS);
2859	nldev_exit();
 
2860	unregister_pernet_device(&rdma_dev_net_ops);
2861	unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2862	ib_sa_cleanup();
2863	ib_mad_cleanup();
2864	addr_cleanup();
2865	rdma_nl_exit();
2866	class_unregister(&ib_class);
2867	destroy_workqueue(ib_comp_unbound_wq);
2868	destroy_workqueue(ib_comp_wq);
2869	/* Make sure that any pending umem accounting work is done. */
2870	destroy_workqueue(ib_wq);
2871	destroy_workqueue(ib_unreg_wq);
2872	WARN_ON(!xa_empty(&clients));
2873	WARN_ON(!xa_empty(&devices));
2874}
2875
2876MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2877
2878/* ib core relies on netdev stack to first register net_ns_type_operations
2879 * ns kobject type before ib_core initialization.
2880 */
2881fs_initcall(ib_core_init);
2882module_exit(ib_core_cleanup);
v5.4
   1/*
   2 * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4 *
   5 * This software is available to you under a choice of one of two
   6 * licenses.  You may choose to be licensed under the terms of the GNU
   7 * General Public License (GPL) Version 2, available from the file
   8 * COPYING in the main directory of this source tree, or the
   9 * OpenIB.org BSD license below:
  10 *
  11 *     Redistribution and use in source and binary forms, with or
  12 *     without modification, are permitted provided that the following
  13 *     conditions are met:
  14 *
  15 *      - Redistributions of source code must retain the above
  16 *        copyright notice, this list of conditions and the following
  17 *        disclaimer.
  18 *
  19 *      - Redistributions in binary form must reproduce the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer in the documentation and/or other materials
  22 *        provided with the distribution.
  23 *
  24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31 * SOFTWARE.
  32 */
  33
  34#include <linux/module.h>
  35#include <linux/string.h>
  36#include <linux/errno.h>
  37#include <linux/kernel.h>
  38#include <linux/slab.h>
  39#include <linux/init.h>
  40#include <linux/netdevice.h>
  41#include <net/net_namespace.h>
  42#include <linux/security.h>
  43#include <linux/notifier.h>
  44#include <linux/hashtable.h>
  45#include <rdma/rdma_netlink.h>
  46#include <rdma/ib_addr.h>
  47#include <rdma/ib_cache.h>
  48#include <rdma/rdma_counter.h>
  49
  50#include "core_priv.h"
  51#include "restrack.h"
  52
  53MODULE_AUTHOR("Roland Dreier");
  54MODULE_DESCRIPTION("core kernel InfiniBand API");
  55MODULE_LICENSE("Dual BSD/GPL");
  56
  57struct workqueue_struct *ib_comp_wq;
  58struct workqueue_struct *ib_comp_unbound_wq;
  59struct workqueue_struct *ib_wq;
  60EXPORT_SYMBOL_GPL(ib_wq);
 
  61
  62/*
  63 * Each of the three rwsem locks (devices, clients, client_data) protects the
  64 * xarray of the same name. Specifically it allows the caller to assert that
  65 * the MARK will/will not be changing under the lock, and for devices and
  66 * clients, that the value in the xarray is still a valid pointer. Change of
  67 * the MARK is linked to the object state, so holding the lock and testing the
  68 * MARK also asserts that the contained object is in a certain state.
  69 *
  70 * This is used to build a two stage register/unregister flow where objects
  71 * can continue to be in the xarray even though they are still in progress to
  72 * register/unregister.
  73 *
  74 * The xarray itself provides additional locking, and restartable iteration,
  75 * which is also relied on.
  76 *
  77 * Locks should not be nested, with the exception of client_data, which is
  78 * allowed to nest under the read side of the other two locks.
  79 *
  80 * The devices_rwsem also protects the device name list, any change or
  81 * assignment of device name must also hold the write side to guarantee unique
  82 * names.
  83 */
  84
  85/*
  86 * devices contains devices that have had their names assigned. The
  87 * devices may not be registered. Users that care about the registration
  88 * status need to call ib_device_try_get() on the device to ensure it is
  89 * registered, and keep it registered, for the required duration.
  90 *
  91 */
  92static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
  93static DECLARE_RWSEM(devices_rwsem);
  94#define DEVICE_REGISTERED XA_MARK_1
  95
  96static u32 highest_client_id;
  97#define CLIENT_REGISTERED XA_MARK_1
  98static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
  99static DECLARE_RWSEM(clients_rwsem);
 100
 101static void ib_client_put(struct ib_client *client)
 102{
 103	if (refcount_dec_and_test(&client->uses))
 104		complete(&client->uses_zero);
 105}
 106
 107/*
 108 * If client_data is registered then the corresponding client must also still
 109 * be registered.
 110 */
 111#define CLIENT_DATA_REGISTERED XA_MARK_1
 112
 113unsigned int rdma_dev_net_id;
 114
 115/*
 116 * A list of net namespaces is maintained in an xarray. This is necessary
 117 * because we can't get the locking right using the existing net ns list. We
 118 * would require a init_net callback after the list is updated.
 119 */
 120static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
 121/*
 122 * rwsem to protect accessing the rdma_nets xarray entries.
 123 */
 124static DECLARE_RWSEM(rdma_nets_rwsem);
 125
 126bool ib_devices_shared_netns = true;
 127module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
 128MODULE_PARM_DESC(netns_mode,
 129		 "Share device among net namespaces; default=1 (shared)");
 130/**
 131 * rdma_dev_access_netns() - Return whether a rdma device can be accessed
 132 *			     from a specified net namespace or not.
 133 * @device:	Pointer to rdma device which needs to be checked
 134 * @net:	Pointer to net namesapce for which access to be checked
 135 *
 136 * rdma_dev_access_netns() - Return whether a rdma device can be accessed
 137 *			     from a specified net namespace or not. When
 138 *			     rdma device is in shared mode, it ignores the
 139 *			     net namespace. When rdma device is exclusive
 140 *			     to a net namespace, rdma device net namespace is
 141 *			     checked against the specified one.
 142 */
 143bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
 144{
 145	return (ib_devices_shared_netns ||
 146		net_eq(read_pnet(&dev->coredev.rdma_net), net));
 147}
 148EXPORT_SYMBOL(rdma_dev_access_netns);
 149
 150/*
 151 * xarray has this behavior where it won't iterate over NULL values stored in
 152 * allocated arrays.  So we need our own iterator to see all values stored in
 153 * the array. This does the same thing as xa_for_each except that it also
 154 * returns NULL valued entries if the array is allocating. Simplified to only
 155 * work on simple xarrays.
 156 */
 157static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
 158			     xa_mark_t filter)
 159{
 160	XA_STATE(xas, xa, *indexp);
 161	void *entry;
 162
 163	rcu_read_lock();
 164	do {
 165		entry = xas_find_marked(&xas, ULONG_MAX, filter);
 166		if (xa_is_zero(entry))
 167			break;
 168	} while (xas_retry(&xas, entry));
 169	rcu_read_unlock();
 170
 171	if (entry) {
 172		*indexp = xas.xa_index;
 173		if (xa_is_zero(entry))
 174			return NULL;
 175		return entry;
 176	}
 177	return XA_ERROR(-ENOENT);
 178}
 179#define xan_for_each_marked(xa, index, entry, filter)                          \
 180	for (index = 0, entry = xan_find_marked(xa, &(index), filter);         \
 181	     !xa_is_err(entry);                                                \
 182	     (index)++, entry = xan_find_marked(xa, &(index), filter))
 183
 184/* RCU hash table mapping netdevice pointers to struct ib_port_data */
 185static DEFINE_SPINLOCK(ndev_hash_lock);
 186static DECLARE_HASHTABLE(ndev_hash, 5);
 187
 188static void free_netdevs(struct ib_device *ib_dev);
 189static void ib_unregister_work(struct work_struct *work);
 190static void __ib_unregister_device(struct ib_device *device);
 191static int ib_security_change(struct notifier_block *nb, unsigned long event,
 192			      void *lsm_data);
 193static void ib_policy_change_task(struct work_struct *work);
 194static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
 195
 196static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
 197			   struct va_format *vaf)
 198{
 199	if (ibdev && ibdev->dev.parent)
 200		dev_printk_emit(level[1] - '0',
 201				ibdev->dev.parent,
 202				"%s %s %s: %pV",
 203				dev_driver_string(ibdev->dev.parent),
 204				dev_name(ibdev->dev.parent),
 205				dev_name(&ibdev->dev),
 206				vaf);
 207	else if (ibdev)
 208		printk("%s%s: %pV",
 209		       level, dev_name(&ibdev->dev), vaf);
 210	else
 211		printk("%s(NULL ib_device): %pV", level, vaf);
 212}
 213
 214void ibdev_printk(const char *level, const struct ib_device *ibdev,
 215		  const char *format, ...)
 216{
 217	struct va_format vaf;
 218	va_list args;
 219
 220	va_start(args, format);
 221
 222	vaf.fmt = format;
 223	vaf.va = &args;
 224
 225	__ibdev_printk(level, ibdev, &vaf);
 226
 227	va_end(args);
 228}
 229EXPORT_SYMBOL(ibdev_printk);
 230
 231#define define_ibdev_printk_level(func, level)                  \
 232void func(const struct ib_device *ibdev, const char *fmt, ...)  \
 233{                                                               \
 234	struct va_format vaf;                                   \
 235	va_list args;                                           \
 236								\
 237	va_start(args, fmt);                                    \
 238								\
 239	vaf.fmt = fmt;                                          \
 240	vaf.va = &args;                                         \
 241								\
 242	__ibdev_printk(level, ibdev, &vaf);                     \
 243								\
 244	va_end(args);                                           \
 245}                                                               \
 246EXPORT_SYMBOL(func);
 247
 248define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
 249define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
 250define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
 251define_ibdev_printk_level(ibdev_err, KERN_ERR);
 252define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
 253define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
 254define_ibdev_printk_level(ibdev_info, KERN_INFO);
 255
 256static struct notifier_block ibdev_lsm_nb = {
 257	.notifier_call = ib_security_change,
 258};
 259
 260static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
 261				 struct net *net);
 262
 263/* Pointer to the RCU head at the start of the ib_port_data array */
 264struct ib_port_data_rcu {
 265	struct rcu_head rcu_head;
 266	struct ib_port_data pdata[];
 267};
 268
 269static void ib_device_check_mandatory(struct ib_device *device)
 270{
 271#define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
 272	static const struct {
 273		size_t offset;
 274		char  *name;
 275	} mandatory_table[] = {
 276		IB_MANDATORY_FUNC(query_device),
 277		IB_MANDATORY_FUNC(query_port),
 278		IB_MANDATORY_FUNC(query_pkey),
 279		IB_MANDATORY_FUNC(alloc_pd),
 280		IB_MANDATORY_FUNC(dealloc_pd),
 281		IB_MANDATORY_FUNC(create_qp),
 282		IB_MANDATORY_FUNC(modify_qp),
 283		IB_MANDATORY_FUNC(destroy_qp),
 284		IB_MANDATORY_FUNC(post_send),
 285		IB_MANDATORY_FUNC(post_recv),
 286		IB_MANDATORY_FUNC(create_cq),
 287		IB_MANDATORY_FUNC(destroy_cq),
 288		IB_MANDATORY_FUNC(poll_cq),
 289		IB_MANDATORY_FUNC(req_notify_cq),
 290		IB_MANDATORY_FUNC(get_dma_mr),
 
 291		IB_MANDATORY_FUNC(dereg_mr),
 292		IB_MANDATORY_FUNC(get_port_immutable)
 293	};
 294	int i;
 295
 296	device->kverbs_provider = true;
 297	for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
 298		if (!*(void **) ((void *) &device->ops +
 299				 mandatory_table[i].offset)) {
 300			device->kverbs_provider = false;
 301			break;
 302		}
 303	}
 304}
 305
 306/*
 307 * Caller must perform ib_device_put() to return the device reference count
 308 * when ib_device_get_by_index() returns valid device pointer.
 309 */
 310struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
 311{
 312	struct ib_device *device;
 313
 314	down_read(&devices_rwsem);
 315	device = xa_load(&devices, index);
 316	if (device) {
 317		if (!rdma_dev_access_netns(device, net)) {
 318			device = NULL;
 319			goto out;
 320		}
 321
 322		if (!ib_device_try_get(device))
 323			device = NULL;
 324	}
 325out:
 326	up_read(&devices_rwsem);
 327	return device;
 328}
 329
 330/**
 331 * ib_device_put - Release IB device reference
 332 * @device: device whose reference to be released
 333 *
 334 * ib_device_put() releases reference to the IB device to allow it to be
 335 * unregistered and eventually free.
 336 */
 337void ib_device_put(struct ib_device *device)
 338{
 339	if (refcount_dec_and_test(&device->refcount))
 340		complete(&device->unreg_completion);
 341}
 342EXPORT_SYMBOL(ib_device_put);
 343
 344static struct ib_device *__ib_device_get_by_name(const char *name)
 345{
 346	struct ib_device *device;
 347	unsigned long index;
 348
 349	xa_for_each (&devices, index, device)
 350		if (!strcmp(name, dev_name(&device->dev)))
 351			return device;
 352
 353	return NULL;
 354}
 355
 356/**
 357 * ib_device_get_by_name - Find an IB device by name
 358 * @name: The name to look for
 359 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
 360 *
 361 * Find and hold an ib_device by its name. The caller must call
 362 * ib_device_put() on the returned pointer.
 363 */
 364struct ib_device *ib_device_get_by_name(const char *name,
 365					enum rdma_driver_id driver_id)
 366{
 367	struct ib_device *device;
 368
 369	down_read(&devices_rwsem);
 370	device = __ib_device_get_by_name(name);
 371	if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
 372	    device->ops.driver_id != driver_id)
 373		device = NULL;
 374
 375	if (device) {
 376		if (!ib_device_try_get(device))
 377			device = NULL;
 378	}
 379	up_read(&devices_rwsem);
 380	return device;
 381}
 382EXPORT_SYMBOL(ib_device_get_by_name);
 383
 384static int rename_compat_devs(struct ib_device *device)
 385{
 386	struct ib_core_device *cdev;
 387	unsigned long index;
 388	int ret = 0;
 389
 390	mutex_lock(&device->compat_devs_mutex);
 391	xa_for_each (&device->compat_devs, index, cdev) {
 392		ret = device_rename(&cdev->dev, dev_name(&device->dev));
 393		if (ret) {
 394			dev_warn(&cdev->dev,
 395				 "Fail to rename compatdev to new name %s\n",
 396				 dev_name(&device->dev));
 397			break;
 398		}
 399	}
 400	mutex_unlock(&device->compat_devs_mutex);
 401	return ret;
 402}
 403
 404int ib_device_rename(struct ib_device *ibdev, const char *name)
 405{
 406	unsigned long index;
 407	void *client_data;
 408	int ret;
 409
 410	down_write(&devices_rwsem);
 411	if (!strcmp(name, dev_name(&ibdev->dev))) {
 412		up_write(&devices_rwsem);
 413		return 0;
 414	}
 415
 416	if (__ib_device_get_by_name(name)) {
 417		up_write(&devices_rwsem);
 418		return -EEXIST;
 419	}
 420
 421	ret = device_rename(&ibdev->dev, name);
 422	if (ret) {
 423		up_write(&devices_rwsem);
 424		return ret;
 425	}
 426
 427	strlcpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
 428	ret = rename_compat_devs(ibdev);
 429
 430	downgrade_write(&devices_rwsem);
 431	down_read(&ibdev->client_data_rwsem);
 432	xan_for_each_marked(&ibdev->client_data, index, client_data,
 433			    CLIENT_DATA_REGISTERED) {
 434		struct ib_client *client = xa_load(&clients, index);
 435
 436		if (!client || !client->rename)
 437			continue;
 438
 439		client->rename(ibdev, client_data);
 440	}
 441	up_read(&ibdev->client_data_rwsem);
 442	up_read(&devices_rwsem);
 443	return 0;
 444}
 445
 446int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
 447{
 448	if (use_dim > 1)
 449		return -EINVAL;
 450	ibdev->use_cq_dim = use_dim;
 451
 452	return 0;
 453}
 454
 455static int alloc_name(struct ib_device *ibdev, const char *name)
 456{
 457	struct ib_device *device;
 458	unsigned long index;
 459	struct ida inuse;
 460	int rc;
 461	int i;
 462
 463	lockdep_assert_held_write(&devices_rwsem);
 464	ida_init(&inuse);
 465	xa_for_each (&devices, index, device) {
 466		char buf[IB_DEVICE_NAME_MAX];
 467
 468		if (sscanf(dev_name(&device->dev), name, &i) != 1)
 469			continue;
 470		if (i < 0 || i >= INT_MAX)
 471			continue;
 472		snprintf(buf, sizeof buf, name, i);
 473		if (strcmp(buf, dev_name(&device->dev)) != 0)
 474			continue;
 475
 476		rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
 477		if (rc < 0)
 478			goto out;
 479	}
 480
 481	rc = ida_alloc(&inuse, GFP_KERNEL);
 482	if (rc < 0)
 483		goto out;
 484
 485	rc = dev_set_name(&ibdev->dev, name, rc);
 486out:
 487	ida_destroy(&inuse);
 488	return rc;
 489}
 490
 491static void ib_device_release(struct device *device)
 492{
 493	struct ib_device *dev = container_of(device, struct ib_device, dev);
 494
 495	free_netdevs(dev);
 496	WARN_ON(refcount_read(&dev->refcount));
 
 
 497	if (dev->port_data) {
 498		ib_cache_release_one(dev);
 499		ib_security_release_port_pkey_list(dev);
 500		rdma_counter_release(dev);
 501		kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
 502				       pdata[0]),
 503			  rcu_head);
 504	}
 505
 506	mutex_destroy(&dev->unregistration_lock);
 507	mutex_destroy(&dev->compat_devs_mutex);
 508
 509	xa_destroy(&dev->compat_devs);
 510	xa_destroy(&dev->client_data);
 511	kfree_rcu(dev, rcu_head);
 512}
 513
 514static int ib_device_uevent(struct device *device,
 515			    struct kobj_uevent_env *env)
 516{
 517	if (add_uevent_var(env, "NAME=%s", dev_name(device)))
 518		return -ENOMEM;
 519
 520	/*
 521	 * It would be nice to pass the node GUID with the event...
 522	 */
 523
 524	return 0;
 525}
 526
 527static const void *net_namespace(struct device *d)
 528{
 529	struct ib_core_device *coredev =
 530			container_of(d, struct ib_core_device, dev);
 531
 532	return read_pnet(&coredev->rdma_net);
 533}
 534
 535static struct class ib_class = {
 536	.name    = "infiniband",
 537	.dev_release = ib_device_release,
 538	.dev_uevent = ib_device_uevent,
 539	.ns_type = &net_ns_type_operations,
 540	.namespace = net_namespace,
 541};
 542
 543static void rdma_init_coredev(struct ib_core_device *coredev,
 544			      struct ib_device *dev, struct net *net)
 545{
 546	/* This BUILD_BUG_ON is intended to catch layout change
 547	 * of union of ib_core_device and device.
 548	 * dev must be the first element as ib_core and providers
 549	 * driver uses it. Adding anything in ib_core_device before
 550	 * device will break this assumption.
 551	 */
 552	BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
 553		     offsetof(struct ib_device, dev));
 554
 555	coredev->dev.class = &ib_class;
 556	coredev->dev.groups = dev->groups;
 557	device_initialize(&coredev->dev);
 558	coredev->owner = dev;
 559	INIT_LIST_HEAD(&coredev->port_list);
 560	write_pnet(&coredev->rdma_net, net);
 561}
 562
 563/**
 564 * _ib_alloc_device - allocate an IB device struct
 565 * @size:size of structure to allocate
 566 *
 567 * Low-level drivers should use ib_alloc_device() to allocate &struct
 568 * ib_device.  @size is the size of the structure to be allocated,
 569 * including any private data used by the low-level driver.
 570 * ib_dealloc_device() must be used to free structures allocated with
 571 * ib_alloc_device().
 572 */
 573struct ib_device *_ib_alloc_device(size_t size)
 574{
 575	struct ib_device *device;
 
 576
 577	if (WARN_ON(size < sizeof(struct ib_device)))
 578		return NULL;
 579
 580	device = kzalloc(size, GFP_KERNEL);
 581	if (!device)
 582		return NULL;
 583
 584	if (rdma_restrack_init(device)) {
 585		kfree(device);
 586		return NULL;
 587	}
 588
 589	device->groups[0] = &ib_dev_attr_group;
 590	rdma_init_coredev(&device->coredev, device, &init_net);
 591
 592	INIT_LIST_HEAD(&device->event_handler_list);
 593	spin_lock_init(&device->event_handler_lock);
 
 594	mutex_init(&device->unregistration_lock);
 595	/*
 596	 * client_data needs to be alloc because we don't want our mark to be
 597	 * destroyed if the user stores NULL in the client data.
 598	 */
 599	xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
 600	init_rwsem(&device->client_data_rwsem);
 601	xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
 602	mutex_init(&device->compat_devs_mutex);
 603	init_completion(&device->unreg_completion);
 604	INIT_WORK(&device->unregistration_work, ib_unregister_work);
 605
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 606	return device;
 607}
 608EXPORT_SYMBOL(_ib_alloc_device);
 609
 610/**
 611 * ib_dealloc_device - free an IB device struct
 612 * @device:structure to free
 613 *
 614 * Free a structure allocated with ib_alloc_device().
 615 */
 616void ib_dealloc_device(struct ib_device *device)
 617{
 618	if (device->ops.dealloc_driver)
 619		device->ops.dealloc_driver(device);
 620
 621	/*
 622	 * ib_unregister_driver() requires all devices to remain in the xarray
 623	 * while their ops are callable. The last op we call is dealloc_driver
 624	 * above.  This is needed to create a fence on op callbacks prior to
 625	 * allowing the driver module to unload.
 626	 */
 627	down_write(&devices_rwsem);
 628	if (xa_load(&devices, device->index) == device)
 629		xa_erase(&devices, device->index);
 630	up_write(&devices_rwsem);
 631
 632	/* Expedite releasing netdev references */
 633	free_netdevs(device);
 634
 635	WARN_ON(!xa_empty(&device->compat_devs));
 636	WARN_ON(!xa_empty(&device->client_data));
 637	WARN_ON(refcount_read(&device->refcount));
 638	rdma_restrack_clean(device);
 639	/* Balances with device_initialize */
 640	put_device(&device->dev);
 641}
 642EXPORT_SYMBOL(ib_dealloc_device);
 643
 644/*
 645 * add_client_context() and remove_client_context() must be safe against
 646 * parallel calls on the same device - registration/unregistration of both the
 647 * device and client can be occurring in parallel.
 648 *
 649 * The routines need to be a fence, any caller must not return until the add
 650 * or remove is fully completed.
 651 */
 652static int add_client_context(struct ib_device *device,
 653			      struct ib_client *client)
 654{
 655	int ret = 0;
 656
 657	if (!device->kverbs_provider && !client->no_kverbs_req)
 658		return 0;
 659
 660	down_write(&device->client_data_rwsem);
 661	/*
 662	 * So long as the client is registered hold both the client and device
 663	 * unregistration locks.
 664	 */
 665	if (!refcount_inc_not_zero(&client->uses))
 666		goto out_unlock;
 667	refcount_inc(&device->refcount);
 668
 669	/*
 670	 * Another caller to add_client_context got here first and has already
 671	 * completely initialized context.
 672	 */
 673	if (xa_get_mark(&device->client_data, client->client_id,
 674		    CLIENT_DATA_REGISTERED))
 675		goto out;
 676
 677	ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
 678			      GFP_KERNEL));
 679	if (ret)
 680		goto out;
 681	downgrade_write(&device->client_data_rwsem);
 682	if (client->add)
 683		client->add(device);
 
 
 
 
 
 
 
 
 
 
 
 
 684
 685	/* Readers shall not see a client until add has been completed */
 686	xa_set_mark(&device->client_data, client->client_id,
 687		    CLIENT_DATA_REGISTERED);
 688	up_read(&device->client_data_rwsem);
 689	return 0;
 690
 691out:
 692	ib_device_put(device);
 693	ib_client_put(client);
 694out_unlock:
 695	up_write(&device->client_data_rwsem);
 696	return ret;
 697}
 698
 699static void remove_client_context(struct ib_device *device,
 700				  unsigned int client_id)
 701{
 702	struct ib_client *client;
 703	void *client_data;
 704
 705	down_write(&device->client_data_rwsem);
 706	if (!xa_get_mark(&device->client_data, client_id,
 707			 CLIENT_DATA_REGISTERED)) {
 708		up_write(&device->client_data_rwsem);
 709		return;
 710	}
 711	client_data = xa_load(&device->client_data, client_id);
 712	xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
 713	client = xa_load(&clients, client_id);
 714	up_write(&device->client_data_rwsem);
 715
 716	/*
 717	 * Notice we cannot be holding any exclusive locks when calling the
 718	 * remove callback as the remove callback can recurse back into any
 719	 * public functions in this module and thus try for any locks those
 720	 * functions take.
 721	 *
 722	 * For this reason clients and drivers should not call the
 723	 * unregistration functions will holdling any locks.
 724	 */
 725	if (client->remove)
 726		client->remove(device, client_data);
 727
 728	xa_erase(&device->client_data, client_id);
 729	ib_device_put(device);
 730	ib_client_put(client);
 731}
 732
 733static int alloc_port_data(struct ib_device *device)
 734{
 735	struct ib_port_data_rcu *pdata_rcu;
 736	unsigned int port;
 737
 738	if (device->port_data)
 739		return 0;
 740
 741	/* This can only be called once the physical port range is defined */
 742	if (WARN_ON(!device->phys_port_cnt))
 743		return -EINVAL;
 744
 
 
 
 
 745	/*
 746	 * device->port_data is indexed directly by the port number to make
 747	 * access to this data as efficient as possible.
 748	 *
 749	 * Therefore port_data is declared as a 1 based array with potential
 750	 * empty slots at the beginning.
 751	 */
 752	pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
 753					rdma_end_port(device) + 1),
 754			    GFP_KERNEL);
 755	if (!pdata_rcu)
 756		return -ENOMEM;
 757	/*
 758	 * The rcu_head is put in front of the port data array and the stored
 759	 * pointer is adjusted since we never need to see that member until
 760	 * kfree_rcu.
 761	 */
 762	device->port_data = pdata_rcu->pdata;
 763
 764	rdma_for_each_port (device, port) {
 765		struct ib_port_data *pdata = &device->port_data[port];
 766
 767		pdata->ib_dev = device;
 768		spin_lock_init(&pdata->pkey_list_lock);
 769		INIT_LIST_HEAD(&pdata->pkey_list);
 770		spin_lock_init(&pdata->netdev_lock);
 771		INIT_HLIST_NODE(&pdata->ndev_hash_link);
 772	}
 773	return 0;
 774}
 775
 776static int verify_immutable(const struct ib_device *dev, u8 port)
 777{
 778	return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
 779			    rdma_max_mad_size(dev, port) != 0);
 780}
 781
 782static int setup_port_data(struct ib_device *device)
 783{
 784	unsigned int port;
 785	int ret;
 786
 787	ret = alloc_port_data(device);
 788	if (ret)
 789		return ret;
 790
 791	rdma_for_each_port (device, port) {
 792		struct ib_port_data *pdata = &device->port_data[port];
 793
 794		ret = device->ops.get_port_immutable(device, port,
 795						     &pdata->immutable);
 796		if (ret)
 797			return ret;
 798
 799		if (verify_immutable(device, port))
 800			return -EINVAL;
 801	}
 802	return 0;
 803}
 804
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 805void ib_get_device_fw_str(struct ib_device *dev, char *str)
 806{
 807	if (dev->ops.get_dev_fw_str)
 808		dev->ops.get_dev_fw_str(dev, str);
 809	else
 810		str[0] = '\0';
 811}
 812EXPORT_SYMBOL(ib_get_device_fw_str);
 813
 814static void ib_policy_change_task(struct work_struct *work)
 815{
 816	struct ib_device *dev;
 817	unsigned long index;
 818
 819	down_read(&devices_rwsem);
 820	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
 821		unsigned int i;
 822
 823		rdma_for_each_port (dev, i) {
 824			u64 sp;
 825			int ret = ib_get_cached_subnet_prefix(dev,
 826							      i,
 827							      &sp);
 828
 829			WARN_ONCE(ret,
 830				  "ib_get_cached_subnet_prefix err: %d, this should never happen here\n",
 831				  ret);
 832			if (!ret)
 833				ib_security_cache_change(dev, i, sp);
 834		}
 835	}
 836	up_read(&devices_rwsem);
 837}
 838
 839static int ib_security_change(struct notifier_block *nb, unsigned long event,
 840			      void *lsm_data)
 841{
 842	if (event != LSM_POLICY_CHANGE)
 843		return NOTIFY_DONE;
 844
 845	schedule_work(&ib_policy_change_work);
 846	ib_mad_agent_security_change();
 847
 848	return NOTIFY_OK;
 849}
 850
 851static void compatdev_release(struct device *dev)
 852{
 853	struct ib_core_device *cdev =
 854		container_of(dev, struct ib_core_device, dev);
 855
 856	kfree(cdev);
 857}
 858
 859static int add_one_compat_dev(struct ib_device *device,
 860			      struct rdma_dev_net *rnet)
 861{
 862	struct ib_core_device *cdev;
 863	int ret;
 864
 865	lockdep_assert_held(&rdma_nets_rwsem);
 866	if (!ib_devices_shared_netns)
 867		return 0;
 868
 869	/*
 870	 * Create and add compat device in all namespaces other than where it
 871	 * is currently bound to.
 872	 */
 873	if (net_eq(read_pnet(&rnet->net),
 874		   read_pnet(&device->coredev.rdma_net)))
 875		return 0;
 876
 877	/*
 878	 * The first of init_net() or ib_register_device() to take the
 879	 * compat_devs_mutex wins and gets to add the device. Others will wait
 880	 * for completion here.
 881	 */
 882	mutex_lock(&device->compat_devs_mutex);
 883	cdev = xa_load(&device->compat_devs, rnet->id);
 884	if (cdev) {
 885		ret = 0;
 886		goto done;
 887	}
 888	ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
 889	if (ret)
 890		goto done;
 891
 892	cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
 893	if (!cdev) {
 894		ret = -ENOMEM;
 895		goto cdev_err;
 896	}
 897
 898	cdev->dev.parent = device->dev.parent;
 899	rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
 900	cdev->dev.release = compatdev_release;
 901	dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
 
 
 902
 903	ret = device_add(&cdev->dev);
 904	if (ret)
 905		goto add_err;
 906	ret = ib_setup_port_attrs(cdev);
 907	if (ret)
 908		goto port_err;
 909
 910	ret = xa_err(xa_store(&device->compat_devs, rnet->id,
 911			      cdev, GFP_KERNEL));
 912	if (ret)
 913		goto insert_err;
 914
 915	mutex_unlock(&device->compat_devs_mutex);
 916	return 0;
 917
 918insert_err:
 919	ib_free_port_attrs(cdev);
 920port_err:
 921	device_del(&cdev->dev);
 922add_err:
 923	put_device(&cdev->dev);
 924cdev_err:
 925	xa_release(&device->compat_devs, rnet->id);
 926done:
 927	mutex_unlock(&device->compat_devs_mutex);
 928	return ret;
 929}
 930
 931static void remove_one_compat_dev(struct ib_device *device, u32 id)
 932{
 933	struct ib_core_device *cdev;
 934
 935	mutex_lock(&device->compat_devs_mutex);
 936	cdev = xa_erase(&device->compat_devs, id);
 937	mutex_unlock(&device->compat_devs_mutex);
 938	if (cdev) {
 939		ib_free_port_attrs(cdev);
 940		device_del(&cdev->dev);
 941		put_device(&cdev->dev);
 942	}
 943}
 944
 945static void remove_compat_devs(struct ib_device *device)
 946{
 947	struct ib_core_device *cdev;
 948	unsigned long index;
 949
 950	xa_for_each (&device->compat_devs, index, cdev)
 951		remove_one_compat_dev(device, index);
 952}
 953
 954static int add_compat_devs(struct ib_device *device)
 955{
 956	struct rdma_dev_net *rnet;
 957	unsigned long index;
 958	int ret = 0;
 959
 960	lockdep_assert_held(&devices_rwsem);
 961
 962	down_read(&rdma_nets_rwsem);
 963	xa_for_each (&rdma_nets, index, rnet) {
 964		ret = add_one_compat_dev(device, rnet);
 965		if (ret)
 966			break;
 967	}
 968	up_read(&rdma_nets_rwsem);
 969	return ret;
 970}
 971
 972static void remove_all_compat_devs(void)
 973{
 974	struct ib_compat_device *cdev;
 975	struct ib_device *dev;
 976	unsigned long index;
 977
 978	down_read(&devices_rwsem);
 979	xa_for_each (&devices, index, dev) {
 980		unsigned long c_index = 0;
 981
 982		/* Hold nets_rwsem so that any other thread modifying this
 983		 * system param can sync with this thread.
 984		 */
 985		down_read(&rdma_nets_rwsem);
 986		xa_for_each (&dev->compat_devs, c_index, cdev)
 987			remove_one_compat_dev(dev, c_index);
 988		up_read(&rdma_nets_rwsem);
 989	}
 990	up_read(&devices_rwsem);
 991}
 992
 993static int add_all_compat_devs(void)
 994{
 995	struct rdma_dev_net *rnet;
 996	struct ib_device *dev;
 997	unsigned long index;
 998	int ret = 0;
 999
1000	down_read(&devices_rwsem);
1001	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1002		unsigned long net_index = 0;
1003
1004		/* Hold nets_rwsem so that any other thread modifying this
1005		 * system param can sync with this thread.
1006		 */
1007		down_read(&rdma_nets_rwsem);
1008		xa_for_each (&rdma_nets, net_index, rnet) {
1009			ret = add_one_compat_dev(dev, rnet);
1010			if (ret)
1011				break;
1012		}
1013		up_read(&rdma_nets_rwsem);
1014	}
1015	up_read(&devices_rwsem);
1016	if (ret)
1017		remove_all_compat_devs();
1018	return ret;
1019}
1020
1021int rdma_compatdev_set(u8 enable)
1022{
1023	struct rdma_dev_net *rnet;
1024	unsigned long index;
1025	int ret = 0;
1026
1027	down_write(&rdma_nets_rwsem);
1028	if (ib_devices_shared_netns == enable) {
1029		up_write(&rdma_nets_rwsem);
1030		return 0;
1031	}
1032
1033	/* enable/disable of compat devices is not supported
1034	 * when more than default init_net exists.
1035	 */
1036	xa_for_each (&rdma_nets, index, rnet) {
1037		ret++;
1038		break;
1039	}
1040	if (!ret)
1041		ib_devices_shared_netns = enable;
1042	up_write(&rdma_nets_rwsem);
1043	if (ret)
1044		return -EBUSY;
1045
1046	if (enable)
1047		ret = add_all_compat_devs();
1048	else
1049		remove_all_compat_devs();
1050	return ret;
1051}
1052
1053static void rdma_dev_exit_net(struct net *net)
1054{
1055	struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1056	struct ib_device *dev;
1057	unsigned long index;
1058	int ret;
1059
1060	down_write(&rdma_nets_rwsem);
1061	/*
1062	 * Prevent the ID from being re-used and hide the id from xa_for_each.
1063	 */
1064	ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1065	WARN_ON(ret);
1066	up_write(&rdma_nets_rwsem);
1067
1068	down_read(&devices_rwsem);
1069	xa_for_each (&devices, index, dev) {
1070		get_device(&dev->dev);
1071		/*
1072		 * Release the devices_rwsem so that pontentially blocking
1073		 * device_del, doesn't hold the devices_rwsem for too long.
1074		 */
1075		up_read(&devices_rwsem);
1076
1077		remove_one_compat_dev(dev, rnet->id);
1078
1079		/*
1080		 * If the real device is in the NS then move it back to init.
1081		 */
1082		rdma_dev_change_netns(dev, net, &init_net);
1083
1084		put_device(&dev->dev);
1085		down_read(&devices_rwsem);
1086	}
1087	up_read(&devices_rwsem);
1088
1089	rdma_nl_net_exit(rnet);
1090	xa_erase(&rdma_nets, rnet->id);
1091}
1092
1093static __net_init int rdma_dev_init_net(struct net *net)
1094{
1095	struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1096	unsigned long index;
1097	struct ib_device *dev;
1098	int ret;
1099
1100	write_pnet(&rnet->net, net);
1101
1102	ret = rdma_nl_net_init(rnet);
1103	if (ret)
1104		return ret;
1105
1106	/* No need to create any compat devices in default init_net. */
1107	if (net_eq(net, &init_net))
1108		return 0;
1109
1110	ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1111	if (ret) {
1112		rdma_nl_net_exit(rnet);
1113		return ret;
1114	}
1115
1116	down_read(&devices_rwsem);
1117	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1118		/* Hold nets_rwsem so that netlink command cannot change
1119		 * system configuration for device sharing mode.
1120		 */
1121		down_read(&rdma_nets_rwsem);
1122		ret = add_one_compat_dev(dev, rnet);
1123		up_read(&rdma_nets_rwsem);
1124		if (ret)
1125			break;
1126	}
1127	up_read(&devices_rwsem);
1128
1129	if (ret)
1130		rdma_dev_exit_net(net);
1131
1132	return ret;
1133}
1134
1135/*
1136 * Assign the unique string device name and the unique device index. This is
1137 * undone by ib_dealloc_device.
1138 */
1139static int assign_name(struct ib_device *device, const char *name)
1140{
1141	static u32 last_id;
1142	int ret;
1143
1144	down_write(&devices_rwsem);
1145	/* Assign a unique name to the device */
1146	if (strchr(name, '%'))
1147		ret = alloc_name(device, name);
1148	else
1149		ret = dev_set_name(&device->dev, name);
1150	if (ret)
1151		goto out;
1152
1153	if (__ib_device_get_by_name(dev_name(&device->dev))) {
1154		ret = -ENFILE;
1155		goto out;
1156	}
1157	strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1158
1159	ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1160			&last_id, GFP_KERNEL);
1161	if (ret > 0)
1162		ret = 0;
1163
1164out:
1165	up_write(&devices_rwsem);
1166	return ret;
1167}
1168
1169static void setup_dma_device(struct ib_device *device)
1170{
1171	struct device *parent = device->dev.parent;
1172
1173	WARN_ON_ONCE(device->dma_device);
1174	if (device->dev.dma_ops) {
1175		/*
1176		 * The caller provided custom DMA operations. Copy the
1177		 * DMA-related fields that are used by e.g. dma_alloc_coherent()
1178		 * into device->dev.
1179		 */
1180		device->dma_device = &device->dev;
1181		if (!device->dev.dma_mask) {
1182			if (parent)
1183				device->dev.dma_mask = parent->dma_mask;
1184			else
1185				WARN_ON_ONCE(true);
1186		}
1187		if (!device->dev.coherent_dma_mask) {
1188			if (parent)
1189				device->dev.coherent_dma_mask =
1190					parent->coherent_dma_mask;
1191			else
1192				WARN_ON_ONCE(true);
1193		}
1194	} else {
1195		/*
1196		 * The caller did not provide custom DMA operations. Use the
1197		 * DMA mapping operations of the parent device.
1198		 */
1199		WARN_ON_ONCE(!parent);
1200		device->dma_device = parent;
1201	}
1202	/* Setup default max segment size for all IB devices */
1203	dma_set_max_seg_size(device->dma_device, SZ_2G);
1204
1205}
1206
1207/*
1208 * setup_device() allocates memory and sets up data that requires calling the
1209 * device ops, this is the only reason these actions are not done during
1210 * ib_alloc_device. It is undone by ib_dealloc_device().
1211 */
1212static int setup_device(struct ib_device *device)
1213{
1214	struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1215	int ret;
1216
1217	setup_dma_device(device);
1218	ib_device_check_mandatory(device);
1219
1220	ret = setup_port_data(device);
1221	if (ret) {
1222		dev_warn(&device->dev, "Couldn't create per-port data\n");
1223		return ret;
1224	}
1225
1226	memset(&device->attrs, 0, sizeof(device->attrs));
1227	ret = device->ops.query_device(device, &device->attrs, &uhw);
1228	if (ret) {
1229		dev_warn(&device->dev,
1230			 "Couldn't query the device attributes\n");
1231		return ret;
1232	}
1233
1234	return 0;
1235}
1236
1237static void disable_device(struct ib_device *device)
1238{
1239	u32 cid;
1240
1241	WARN_ON(!refcount_read(&device->refcount));
1242
1243	down_write(&devices_rwsem);
1244	xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1245	up_write(&devices_rwsem);
1246
1247	/*
1248	 * Remove clients in LIFO order, see assign_client_id. This could be
1249	 * more efficient if xarray learns to reverse iterate. Since no new
1250	 * clients can be added to this ib_device past this point we only need
1251	 * the maximum possible client_id value here.
1252	 */
1253	down_read(&clients_rwsem);
1254	cid = highest_client_id;
1255	up_read(&clients_rwsem);
1256	while (cid) {
1257		cid--;
1258		remove_client_context(device, cid);
1259	}
1260
 
 
1261	/* Pairs with refcount_set in enable_device */
1262	ib_device_put(device);
1263	wait_for_completion(&device->unreg_completion);
1264
1265	/*
1266	 * compat devices must be removed after device refcount drops to zero.
1267	 * Otherwise init_net() may add more compatdevs after removing compat
1268	 * devices and before device is disabled.
1269	 */
1270	remove_compat_devs(device);
1271}
1272
1273/*
1274 * An enabled device is visible to all clients and to all the public facing
1275 * APIs that return a device pointer. This always returns with a new get, even
1276 * if it fails.
1277 */
1278static int enable_device_and_get(struct ib_device *device)
1279{
1280	struct ib_client *client;
1281	unsigned long index;
1282	int ret = 0;
1283
1284	/*
1285	 * One ref belongs to the xa and the other belongs to this
1286	 * thread. This is needed to guard against parallel unregistration.
1287	 */
1288	refcount_set(&device->refcount, 2);
1289	down_write(&devices_rwsem);
1290	xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1291
1292	/*
1293	 * By using downgrade_write() we ensure that no other thread can clear
1294	 * DEVICE_REGISTERED while we are completing the client setup.
1295	 */
1296	downgrade_write(&devices_rwsem);
1297
1298	if (device->ops.enable_driver) {
1299		ret = device->ops.enable_driver(device);
1300		if (ret)
1301			goto out;
1302	}
1303
1304	down_read(&clients_rwsem);
1305	xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1306		ret = add_client_context(device, client);
1307		if (ret)
1308			break;
1309	}
1310	up_read(&clients_rwsem);
1311	if (!ret)
1312		ret = add_compat_devs(device);
1313out:
1314	up_read(&devices_rwsem);
1315	return ret;
1316}
1317
 
 
 
 
1318/**
1319 * ib_register_device - Register an IB device with IB core
1320 * @device:Device to register
 
 
 
 
 
1321 *
1322 * Low-level drivers use ib_register_device() to register their
1323 * devices with the IB core.  All registered clients will receive a
1324 * callback for each device that is added. @device must be allocated
1325 * with ib_alloc_device().
1326 *
1327 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1328 * asynchronously then the device pointer may become freed as soon as this
1329 * function returns.
1330 */
1331int ib_register_device(struct ib_device *device, const char *name)
 
1332{
1333	int ret;
1334
1335	ret = assign_name(device, name);
1336	if (ret)
1337		return ret;
1338
 
 
 
 
 
 
 
 
1339	ret = setup_device(device);
1340	if (ret)
1341		return ret;
1342
1343	ret = ib_cache_setup_one(device);
1344	if (ret) {
1345		dev_warn(&device->dev,
1346			 "Couldn't set up InfiniBand P_Key/GID cache\n");
1347		return ret;
1348	}
1349
 
 
 
 
 
 
1350	ib_device_register_rdmacg(device);
1351
1352	rdma_counter_init(device);
1353
1354	/*
1355	 * Ensure that ADD uevent is not fired because it
1356	 * is too early amd device is not initialized yet.
1357	 */
1358	dev_set_uevent_suppress(&device->dev, true);
1359	ret = device_add(&device->dev);
1360	if (ret)
1361		goto cg_cleanup;
1362
1363	ret = ib_device_register_sysfs(device);
1364	if (ret) {
1365		dev_warn(&device->dev,
1366			 "Couldn't register device with driver model\n");
1367		goto dev_cleanup;
1368	}
1369
1370	ret = enable_device_and_get(device);
1371	dev_set_uevent_suppress(&device->dev, false);
1372	/* Mark for userspace that device is ready */
1373	kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1374	if (ret) {
1375		void (*dealloc_fn)(struct ib_device *);
1376
1377		/*
1378		 * If we hit this error flow then we don't want to
1379		 * automatically dealloc the device since the caller is
1380		 * expected to call ib_dealloc_device() after
1381		 * ib_register_device() fails. This is tricky due to the
1382		 * possibility for a parallel unregistration along with this
1383		 * error flow. Since we have a refcount here we know any
1384		 * parallel flow is stopped in disable_device and will see the
1385		 * NULL pointers, causing the responsibility to
1386		 * ib_dealloc_device() to revert back to this thread.
1387		 */
1388		dealloc_fn = device->ops.dealloc_driver;
1389		device->ops.dealloc_driver = NULL;
1390		ib_device_put(device);
1391		__ib_unregister_device(device);
1392		device->ops.dealloc_driver = dealloc_fn;
 
1393		return ret;
1394	}
 
 
 
1395	ib_device_put(device);
1396
1397	return 0;
1398
1399dev_cleanup:
1400	device_del(&device->dev);
1401cg_cleanup:
1402	dev_set_uevent_suppress(&device->dev, false);
1403	ib_device_unregister_rdmacg(device);
 
1404	ib_cache_cleanup_one(device);
1405	return ret;
1406}
1407EXPORT_SYMBOL(ib_register_device);
1408
1409/* Callers must hold a get on the device. */
1410static void __ib_unregister_device(struct ib_device *ib_dev)
1411{
1412	/*
1413	 * We have a registration lock so that all the calls to unregister are
1414	 * fully fenced, once any unregister returns the device is truely
1415	 * unregistered even if multiple callers are unregistering it at the
1416	 * same time. This also interacts with the registration flow and
1417	 * provides sane semantics if register and unregister are racing.
1418	 */
1419	mutex_lock(&ib_dev->unregistration_lock);
1420	if (!refcount_read(&ib_dev->refcount))
1421		goto out;
1422
1423	disable_device(ib_dev);
1424
1425	/* Expedite removing unregistered pointers from the hash table */
1426	free_netdevs(ib_dev);
1427
1428	ib_device_unregister_sysfs(ib_dev);
1429	device_del(&ib_dev->dev);
1430	ib_device_unregister_rdmacg(ib_dev);
1431	ib_cache_cleanup_one(ib_dev);
1432
1433	/*
1434	 * Drivers using the new flow may not call ib_dealloc_device except
1435	 * in error unwind prior to registration success.
1436	 */
1437	if (ib_dev->ops.dealloc_driver) {
 
1438		WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1439		ib_dealloc_device(ib_dev);
1440	}
1441out:
1442	mutex_unlock(&ib_dev->unregistration_lock);
1443}
1444
1445/**
1446 * ib_unregister_device - Unregister an IB device
1447 * @device: The device to unregister
1448 *
1449 * Unregister an IB device.  All clients will receive a remove callback.
1450 *
1451 * Callers should call this routine only once, and protect against races with
1452 * registration. Typically it should only be called as part of a remove
1453 * callback in an implementation of driver core's struct device_driver and
1454 * related.
1455 *
1456 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1457 * this function.
1458 */
1459void ib_unregister_device(struct ib_device *ib_dev)
1460{
1461	get_device(&ib_dev->dev);
1462	__ib_unregister_device(ib_dev);
1463	put_device(&ib_dev->dev);
1464}
1465EXPORT_SYMBOL(ib_unregister_device);
1466
1467/**
1468 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1469 * device: The device to unregister
1470 *
1471 * This is the same as ib_unregister_device(), except it includes an internal
1472 * ib_device_put() that should match a 'get' obtained by the caller.
1473 *
1474 * It is safe to call this routine concurrently from multiple threads while
1475 * holding the 'get'. When the function returns the device is fully
1476 * unregistered.
1477 *
1478 * Drivers using this flow MUST use the driver_unregister callback to clean up
1479 * their resources associated with the device and dealloc it.
1480 */
1481void ib_unregister_device_and_put(struct ib_device *ib_dev)
1482{
1483	WARN_ON(!ib_dev->ops.dealloc_driver);
1484	get_device(&ib_dev->dev);
1485	ib_device_put(ib_dev);
1486	__ib_unregister_device(ib_dev);
1487	put_device(&ib_dev->dev);
1488}
1489EXPORT_SYMBOL(ib_unregister_device_and_put);
1490
1491/**
1492 * ib_unregister_driver - Unregister all IB devices for a driver
1493 * @driver_id: The driver to unregister
1494 *
1495 * This implements a fence for device unregistration. It only returns once all
1496 * devices associated with the driver_id have fully completed their
1497 * unregistration and returned from ib_unregister_device*().
1498 *
1499 * If device's are not yet unregistered it goes ahead and starts unregistering
1500 * them.
1501 *
1502 * This does not block creation of new devices with the given driver_id, that
1503 * is the responsibility of the caller.
1504 */
1505void ib_unregister_driver(enum rdma_driver_id driver_id)
1506{
1507	struct ib_device *ib_dev;
1508	unsigned long index;
1509
1510	down_read(&devices_rwsem);
1511	xa_for_each (&devices, index, ib_dev) {
1512		if (ib_dev->ops.driver_id != driver_id)
1513			continue;
1514
1515		get_device(&ib_dev->dev);
1516		up_read(&devices_rwsem);
1517
1518		WARN_ON(!ib_dev->ops.dealloc_driver);
1519		__ib_unregister_device(ib_dev);
1520
1521		put_device(&ib_dev->dev);
1522		down_read(&devices_rwsem);
1523	}
1524	up_read(&devices_rwsem);
1525}
1526EXPORT_SYMBOL(ib_unregister_driver);
1527
1528static void ib_unregister_work(struct work_struct *work)
1529{
1530	struct ib_device *ib_dev =
1531		container_of(work, struct ib_device, unregistration_work);
1532
1533	__ib_unregister_device(ib_dev);
1534	put_device(&ib_dev->dev);
1535}
1536
1537/**
1538 * ib_unregister_device_queued - Unregister a device using a work queue
1539 * device: The device to unregister
1540 *
1541 * This schedules an asynchronous unregistration using a WQ for the device. A
1542 * driver should use this to avoid holding locks while doing unregistration,
1543 * such as holding the RTNL lock.
1544 *
1545 * Drivers using this API must use ib_unregister_driver before module unload
1546 * to ensure that all scheduled unregistrations have completed.
1547 */
1548void ib_unregister_device_queued(struct ib_device *ib_dev)
1549{
1550	WARN_ON(!refcount_read(&ib_dev->refcount));
1551	WARN_ON(!ib_dev->ops.dealloc_driver);
1552	get_device(&ib_dev->dev);
1553	if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
1554		put_device(&ib_dev->dev);
1555}
1556EXPORT_SYMBOL(ib_unregister_device_queued);
1557
1558/*
1559 * The caller must pass in a device that has the kref held and the refcount
1560 * released. If the device is in cur_net and still registered then it is moved
1561 * into net.
1562 */
1563static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1564				 struct net *net)
1565{
1566	int ret2 = -EINVAL;
1567	int ret;
1568
1569	mutex_lock(&device->unregistration_lock);
1570
1571	/*
1572	 * If a device not under ib_device_get() or if the unregistration_lock
1573	 * is not held, the namespace can be changed, or it can be unregistered.
1574	 * Check again under the lock.
1575	 */
1576	if (refcount_read(&device->refcount) == 0 ||
1577	    !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1578		ret = -ENODEV;
1579		goto out;
1580	}
1581
1582	kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1583	disable_device(device);
1584
1585	/*
1586	 * At this point no one can be using the device, so it is safe to
1587	 * change the namespace.
1588	 */
1589	write_pnet(&device->coredev.rdma_net, net);
1590
1591	down_read(&devices_rwsem);
1592	/*
1593	 * Currently rdma devices are system wide unique. So the device name
1594	 * is guaranteed free in the new namespace. Publish the new namespace
1595	 * at the sysfs level.
1596	 */
1597	ret = device_rename(&device->dev, dev_name(&device->dev));
1598	up_read(&devices_rwsem);
1599	if (ret) {
1600		dev_warn(&device->dev,
1601			 "%s: Couldn't rename device after namespace change\n",
1602			 __func__);
1603		/* Try and put things back and re-enable the device */
1604		write_pnet(&device->coredev.rdma_net, cur_net);
1605	}
1606
1607	ret2 = enable_device_and_get(device);
1608	if (ret2) {
1609		/*
1610		 * This shouldn't really happen, but if it does, let the user
1611		 * retry at later point. So don't disable the device.
1612		 */
1613		dev_warn(&device->dev,
1614			 "%s: Couldn't re-enable device after namespace change\n",
1615			 __func__);
1616	}
1617	kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1618
1619	ib_device_put(device);
1620out:
1621	mutex_unlock(&device->unregistration_lock);
1622	if (ret)
1623		return ret;
1624	return ret2;
1625}
1626
1627int ib_device_set_netns_put(struct sk_buff *skb,
1628			    struct ib_device *dev, u32 ns_fd)
1629{
1630	struct net *net;
1631	int ret;
1632
1633	net = get_net_ns_by_fd(ns_fd);
1634	if (IS_ERR(net)) {
1635		ret = PTR_ERR(net);
1636		goto net_err;
1637	}
1638
1639	if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1640		ret = -EPERM;
1641		goto ns_err;
1642	}
1643
1644	/*
1645	 * Currently supported only for those providers which support
1646	 * disassociation and don't do port specific sysfs init. Once a
1647	 * port_cleanup infrastructure is implemented, this limitation will be
1648	 * removed.
1649	 */
1650	if (!dev->ops.disassociate_ucontext || dev->ops.init_port ||
1651	    ib_devices_shared_netns) {
1652		ret = -EOPNOTSUPP;
1653		goto ns_err;
1654	}
1655
1656	get_device(&dev->dev);
1657	ib_device_put(dev);
1658	ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1659	put_device(&dev->dev);
1660
1661	put_net(net);
1662	return ret;
1663
1664ns_err:
1665	put_net(net);
1666net_err:
1667	ib_device_put(dev);
1668	return ret;
1669}
1670
1671static struct pernet_operations rdma_dev_net_ops = {
1672	.init = rdma_dev_init_net,
1673	.exit = rdma_dev_exit_net,
1674	.id = &rdma_dev_net_id,
1675	.size = sizeof(struct rdma_dev_net),
1676};
1677
1678static int assign_client_id(struct ib_client *client)
1679{
1680	int ret;
1681
1682	down_write(&clients_rwsem);
1683	/*
1684	 * The add/remove callbacks must be called in FIFO/LIFO order. To
1685	 * achieve this we assign client_ids so they are sorted in
1686	 * registration order.
1687	 */
1688	client->client_id = highest_client_id;
1689	ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1690	if (ret)
1691		goto out;
1692
1693	highest_client_id++;
1694	xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1695
1696out:
1697	up_write(&clients_rwsem);
1698	return ret;
1699}
1700
1701static void remove_client_id(struct ib_client *client)
1702{
1703	down_write(&clients_rwsem);
1704	xa_erase(&clients, client->client_id);
1705	for (; highest_client_id; highest_client_id--)
1706		if (xa_load(&clients, highest_client_id - 1))
1707			break;
1708	up_write(&clients_rwsem);
1709}
1710
1711/**
1712 * ib_register_client - Register an IB client
1713 * @client:Client to register
1714 *
1715 * Upper level users of the IB drivers can use ib_register_client() to
1716 * register callbacks for IB device addition and removal.  When an IB
1717 * device is added, each registered client's add method will be called
1718 * (in the order the clients were registered), and when a device is
1719 * removed, each client's remove method will be called (in the reverse
1720 * order that clients were registered).  In addition, when
1721 * ib_register_client() is called, the client will receive an add
1722 * callback for all devices already registered.
1723 */
1724int ib_register_client(struct ib_client *client)
1725{
1726	struct ib_device *device;
1727	unsigned long index;
1728	int ret;
1729
1730	refcount_set(&client->uses, 1);
1731	init_completion(&client->uses_zero);
1732	ret = assign_client_id(client);
1733	if (ret)
1734		return ret;
1735
1736	down_read(&devices_rwsem);
1737	xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1738		ret = add_client_context(device, client);
1739		if (ret) {
1740			up_read(&devices_rwsem);
1741			ib_unregister_client(client);
1742			return ret;
1743		}
1744	}
1745	up_read(&devices_rwsem);
1746	return 0;
1747}
1748EXPORT_SYMBOL(ib_register_client);
1749
1750/**
1751 * ib_unregister_client - Unregister an IB client
1752 * @client:Client to unregister
1753 *
1754 * Upper level users use ib_unregister_client() to remove their client
1755 * registration.  When ib_unregister_client() is called, the client
1756 * will receive a remove callback for each IB device still registered.
1757 *
1758 * This is a full fence, once it returns no client callbacks will be called,
1759 * or are running in another thread.
1760 */
1761void ib_unregister_client(struct ib_client *client)
1762{
1763	struct ib_device *device;
1764	unsigned long index;
1765
1766	down_write(&clients_rwsem);
1767	ib_client_put(client);
1768	xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1769	up_write(&clients_rwsem);
1770
1771	/* We do not want to have locks while calling client->remove() */
1772	rcu_read_lock();
1773	xa_for_each (&devices, index, device) {
1774		if (!ib_device_try_get(device))
1775			continue;
1776		rcu_read_unlock();
1777
1778		remove_client_context(device, client->client_id);
1779
1780		ib_device_put(device);
1781		rcu_read_lock();
1782	}
1783	rcu_read_unlock();
1784
1785	/*
1786	 * remove_client_context() is not a fence, it can return even though a
1787	 * removal is ongoing. Wait until all removals are completed.
1788	 */
1789	wait_for_completion(&client->uses_zero);
1790	remove_client_id(client);
1791}
1792EXPORT_SYMBOL(ib_unregister_client);
1793
1794static int __ib_get_global_client_nl_info(const char *client_name,
1795					  struct ib_client_nl_info *res)
1796{
1797	struct ib_client *client;
1798	unsigned long index;
1799	int ret = -ENOENT;
1800
1801	down_read(&clients_rwsem);
1802	xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1803		if (strcmp(client->name, client_name) != 0)
1804			continue;
1805		if (!client->get_global_nl_info) {
1806			ret = -EOPNOTSUPP;
1807			break;
1808		}
1809		ret = client->get_global_nl_info(res);
1810		if (WARN_ON(ret == -ENOENT))
1811			ret = -EINVAL;
1812		if (!ret && res->cdev)
1813			get_device(res->cdev);
1814		break;
1815	}
1816	up_read(&clients_rwsem);
1817	return ret;
1818}
1819
1820static int __ib_get_client_nl_info(struct ib_device *ibdev,
1821				   const char *client_name,
1822				   struct ib_client_nl_info *res)
1823{
1824	unsigned long index;
1825	void *client_data;
1826	int ret = -ENOENT;
1827
1828	down_read(&ibdev->client_data_rwsem);
1829	xan_for_each_marked (&ibdev->client_data, index, client_data,
1830			     CLIENT_DATA_REGISTERED) {
1831		struct ib_client *client = xa_load(&clients, index);
1832
1833		if (!client || strcmp(client->name, client_name) != 0)
1834			continue;
1835		if (!client->get_nl_info) {
1836			ret = -EOPNOTSUPP;
1837			break;
1838		}
1839		ret = client->get_nl_info(ibdev, client_data, res);
1840		if (WARN_ON(ret == -ENOENT))
1841			ret = -EINVAL;
1842
1843		/*
1844		 * The cdev is guaranteed valid as long as we are inside the
1845		 * client_data_rwsem as remove_one can't be called. Keep it
1846		 * valid for the caller.
1847		 */
1848		if (!ret && res->cdev)
1849			get_device(res->cdev);
1850		break;
1851	}
1852	up_read(&ibdev->client_data_rwsem);
1853
1854	return ret;
1855}
1856
1857/**
1858 * ib_get_client_nl_info - Fetch the nl_info from a client
1859 * @device - IB device
1860 * @client_name - Name of the client
1861 * @res - Result of the query
1862 */
1863int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1864			  struct ib_client_nl_info *res)
1865{
1866	int ret;
1867
1868	if (ibdev)
1869		ret = __ib_get_client_nl_info(ibdev, client_name, res);
1870	else
1871		ret = __ib_get_global_client_nl_info(client_name, res);
1872#ifdef CONFIG_MODULES
1873	if (ret == -ENOENT) {
1874		request_module("rdma-client-%s", client_name);
1875		if (ibdev)
1876			ret = __ib_get_client_nl_info(ibdev, client_name, res);
1877		else
1878			ret = __ib_get_global_client_nl_info(client_name, res);
1879	}
1880#endif
1881	if (ret) {
1882		if (ret == -ENOENT)
1883			return -EOPNOTSUPP;
1884		return ret;
1885	}
1886
1887	if (WARN_ON(!res->cdev))
1888		return -EINVAL;
1889	return 0;
1890}
1891
1892/**
1893 * ib_set_client_data - Set IB client context
1894 * @device:Device to set context for
1895 * @client:Client to set context for
1896 * @data:Context to set
1897 *
1898 * ib_set_client_data() sets client context data that can be retrieved with
1899 * ib_get_client_data(). This can only be called while the client is
1900 * registered to the device, once the ib_client remove() callback returns this
1901 * cannot be called.
1902 */
1903void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1904			void *data)
1905{
1906	void *rc;
1907
1908	if (WARN_ON(IS_ERR(data)))
1909		data = NULL;
1910
1911	rc = xa_store(&device->client_data, client->client_id, data,
1912		      GFP_KERNEL);
1913	WARN_ON(xa_is_err(rc));
1914}
1915EXPORT_SYMBOL(ib_set_client_data);
1916
1917/**
1918 * ib_register_event_handler - Register an IB event handler
1919 * @event_handler:Handler to register
1920 *
1921 * ib_register_event_handler() registers an event handler that will be
1922 * called back when asynchronous IB events occur (as defined in
1923 * chapter 11 of the InfiniBand Architecture Specification).  This
1924 * callback may occur in interrupt context.
1925 */
1926void ib_register_event_handler(struct ib_event_handler *event_handler)
1927{
1928	unsigned long flags;
1929
1930	spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1931	list_add_tail(&event_handler->list,
1932		      &event_handler->device->event_handler_list);
1933	spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1934}
1935EXPORT_SYMBOL(ib_register_event_handler);
1936
1937/**
1938 * ib_unregister_event_handler - Unregister an event handler
1939 * @event_handler:Handler to unregister
1940 *
1941 * Unregister an event handler registered with
1942 * ib_register_event_handler().
1943 */
1944void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1945{
1946	unsigned long flags;
1947
1948	spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1949	list_del(&event_handler->list);
1950	spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1951}
1952EXPORT_SYMBOL(ib_unregister_event_handler);
1953
1954/**
1955 * ib_dispatch_event - Dispatch an asynchronous event
1956 * @event:Event to dispatch
1957 *
1958 * Low-level drivers must call ib_dispatch_event() to dispatch the
1959 * event to all registered event handlers when an asynchronous event
1960 * occurs.
1961 */
1962void ib_dispatch_event(struct ib_event *event)
1963{
1964	unsigned long flags;
1965	struct ib_event_handler *handler;
1966
1967	spin_lock_irqsave(&event->device->event_handler_lock, flags);
1968
1969	list_for_each_entry(handler, &event->device->event_handler_list, list)
1970		handler->handler(handler, event);
1971
1972	spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
1973}
1974EXPORT_SYMBOL(ib_dispatch_event);
1975
1976static int iw_query_port(struct ib_device *device,
1977			   u8 port_num,
1978			   struct ib_port_attr *port_attr)
1979{
1980	struct in_device *inetdev;
1981	struct net_device *netdev;
1982	int err;
1983
1984	memset(port_attr, 0, sizeof(*port_attr));
1985
1986	netdev = ib_device_get_netdev(device, port_num);
1987	if (!netdev)
1988		return -ENODEV;
1989
1990	port_attr->max_mtu = IB_MTU_4096;
1991	port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
1992
1993	if (!netif_carrier_ok(netdev)) {
1994		port_attr->state = IB_PORT_DOWN;
1995		port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
1996	} else {
1997		rcu_read_lock();
1998		inetdev = __in_dev_get_rcu(netdev);
1999
2000		if (inetdev && inetdev->ifa_list) {
2001			port_attr->state = IB_PORT_ACTIVE;
2002			port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2003		} else {
2004			port_attr->state = IB_PORT_INIT;
2005			port_attr->phys_state =
2006				IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2007		}
2008
2009		rcu_read_unlock();
2010	}
2011
2012	dev_put(netdev);
2013	err = device->ops.query_port(device, port_num, port_attr);
2014	if (err)
2015		return err;
2016
2017	return 0;
2018}
2019
2020static int __ib_query_port(struct ib_device *device,
2021			   u8 port_num,
2022			   struct ib_port_attr *port_attr)
2023{
2024	union ib_gid gid = {};
2025	int err;
2026
2027	memset(port_attr, 0, sizeof(*port_attr));
2028
2029	err = device->ops.query_port(device, port_num, port_attr);
2030	if (err || port_attr->subnet_prefix)
2031		return err;
2032
2033	if (rdma_port_get_link_layer(device, port_num) !=
2034	    IB_LINK_LAYER_INFINIBAND)
2035		return 0;
2036
2037	err = device->ops.query_gid(device, port_num, 0, &gid);
2038	if (err)
2039		return err;
2040
2041	port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
2042	return 0;
2043}
2044
2045/**
2046 * ib_query_port - Query IB port attributes
2047 * @device:Device to query
2048 * @port_num:Port number to query
2049 * @port_attr:Port attributes
2050 *
2051 * ib_query_port() returns the attributes of a port through the
2052 * @port_attr pointer.
2053 */
2054int ib_query_port(struct ib_device *device,
2055		  u8 port_num,
2056		  struct ib_port_attr *port_attr)
2057{
2058	if (!rdma_is_port_valid(device, port_num))
2059		return -EINVAL;
2060
2061	if (rdma_protocol_iwarp(device, port_num))
2062		return iw_query_port(device, port_num, port_attr);
2063	else
2064		return __ib_query_port(device, port_num, port_attr);
2065}
2066EXPORT_SYMBOL(ib_query_port);
2067
2068static void add_ndev_hash(struct ib_port_data *pdata)
2069{
2070	unsigned long flags;
2071
2072	might_sleep();
2073
2074	spin_lock_irqsave(&ndev_hash_lock, flags);
2075	if (hash_hashed(&pdata->ndev_hash_link)) {
2076		hash_del_rcu(&pdata->ndev_hash_link);
2077		spin_unlock_irqrestore(&ndev_hash_lock, flags);
2078		/*
2079		 * We cannot do hash_add_rcu after a hash_del_rcu until the
2080		 * grace period
2081		 */
2082		synchronize_rcu();
2083		spin_lock_irqsave(&ndev_hash_lock, flags);
2084	}
2085	if (pdata->netdev)
2086		hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2087			     (uintptr_t)pdata->netdev);
2088	spin_unlock_irqrestore(&ndev_hash_lock, flags);
2089}
2090
2091/**
2092 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2093 * @ib_dev: Device to modify
2094 * @ndev: net_device to affiliate, may be NULL
2095 * @port: IB port the net_device is connected to
2096 *
2097 * Drivers should use this to link the ib_device to a netdev so the netdev
2098 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2099 * affiliated with any port.
2100 *
2101 * The caller must ensure that the given ndev is not unregistered or
2102 * unregistering, and that either the ib_device is unregistered or
2103 * ib_device_set_netdev() is called with NULL when the ndev sends a
2104 * NETDEV_UNREGISTER event.
2105 */
2106int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2107			 unsigned int port)
2108{
2109	struct net_device *old_ndev;
2110	struct ib_port_data *pdata;
2111	unsigned long flags;
2112	int ret;
2113
2114	/*
2115	 * Drivers wish to call this before ib_register_driver, so we have to
2116	 * setup the port data early.
2117	 */
2118	ret = alloc_port_data(ib_dev);
2119	if (ret)
2120		return ret;
2121
2122	if (!rdma_is_port_valid(ib_dev, port))
2123		return -EINVAL;
2124
2125	pdata = &ib_dev->port_data[port];
2126	spin_lock_irqsave(&pdata->netdev_lock, flags);
2127	old_ndev = rcu_dereference_protected(
2128		pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2129	if (old_ndev == ndev) {
2130		spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2131		return 0;
2132	}
2133
 
 
2134	if (ndev)
2135		dev_hold(ndev);
2136	rcu_assign_pointer(pdata->netdev, ndev);
2137	spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2138
2139	add_ndev_hash(pdata);
2140	if (old_ndev)
2141		dev_put(old_ndev);
2142
2143	return 0;
2144}
2145EXPORT_SYMBOL(ib_device_set_netdev);
2146
2147static void free_netdevs(struct ib_device *ib_dev)
2148{
2149	unsigned long flags;
2150	unsigned int port;
2151
2152	if (!ib_dev->port_data)
2153		return;
2154
2155	rdma_for_each_port (ib_dev, port) {
2156		struct ib_port_data *pdata = &ib_dev->port_data[port];
2157		struct net_device *ndev;
2158
2159		spin_lock_irqsave(&pdata->netdev_lock, flags);
2160		ndev = rcu_dereference_protected(
2161			pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2162		if (ndev) {
2163			spin_lock(&ndev_hash_lock);
2164			hash_del_rcu(&pdata->ndev_hash_link);
2165			spin_unlock(&ndev_hash_lock);
2166
2167			/*
2168			 * If this is the last dev_put there is still a
2169			 * synchronize_rcu before the netdev is kfreed, so we
2170			 * can continue to rely on unlocked pointer
2171			 * comparisons after the put
2172			 */
2173			rcu_assign_pointer(pdata->netdev, NULL);
2174			dev_put(ndev);
2175		}
2176		spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2177	}
2178}
2179
2180struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2181					unsigned int port)
2182{
2183	struct ib_port_data *pdata;
2184	struct net_device *res;
2185
2186	if (!rdma_is_port_valid(ib_dev, port))
2187		return NULL;
2188
2189	pdata = &ib_dev->port_data[port];
2190
2191	/*
2192	 * New drivers should use ib_device_set_netdev() not the legacy
2193	 * get_netdev().
2194	 */
2195	if (ib_dev->ops.get_netdev)
2196		res = ib_dev->ops.get_netdev(ib_dev, port);
2197	else {
2198		spin_lock(&pdata->netdev_lock);
2199		res = rcu_dereference_protected(
2200			pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2201		if (res)
2202			dev_hold(res);
2203		spin_unlock(&pdata->netdev_lock);
2204	}
2205
2206	/*
2207	 * If we are starting to unregister expedite things by preventing
2208	 * propagation of an unregistering netdev.
2209	 */
2210	if (res && res->reg_state != NETREG_REGISTERED) {
2211		dev_put(res);
2212		return NULL;
2213	}
2214
2215	return res;
2216}
2217
2218/**
2219 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2220 * @ndev: netdev to locate
2221 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2222 *
2223 * Find and hold an ib_device that is associated with a netdev via
2224 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2225 * returned pointer.
2226 */
2227struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2228					  enum rdma_driver_id driver_id)
2229{
2230	struct ib_device *res = NULL;
2231	struct ib_port_data *cur;
2232
2233	rcu_read_lock();
2234	hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2235				    (uintptr_t)ndev) {
2236		if (rcu_access_pointer(cur->netdev) == ndev &&
2237		    (driver_id == RDMA_DRIVER_UNKNOWN ||
2238		     cur->ib_dev->ops.driver_id == driver_id) &&
2239		    ib_device_try_get(cur->ib_dev)) {
2240			res = cur->ib_dev;
2241			break;
2242		}
2243	}
2244	rcu_read_unlock();
2245
2246	return res;
2247}
2248EXPORT_SYMBOL(ib_device_get_by_netdev);
2249
2250/**
2251 * ib_enum_roce_netdev - enumerate all RoCE ports
2252 * @ib_dev : IB device we want to query
2253 * @filter: Should we call the callback?
2254 * @filter_cookie: Cookie passed to filter
2255 * @cb: Callback to call for each found RoCE ports
2256 * @cookie: Cookie passed back to the callback
2257 *
2258 * Enumerates all of the physical RoCE ports of ib_dev
2259 * which are related to netdevice and calls callback() on each
2260 * device for which filter() function returns non zero.
2261 */
2262void ib_enum_roce_netdev(struct ib_device *ib_dev,
2263			 roce_netdev_filter filter,
2264			 void *filter_cookie,
2265			 roce_netdev_callback cb,
2266			 void *cookie)
2267{
2268	unsigned int port;
2269
2270	rdma_for_each_port (ib_dev, port)
2271		if (rdma_protocol_roce(ib_dev, port)) {
2272			struct net_device *idev =
2273				ib_device_get_netdev(ib_dev, port);
2274
2275			if (filter(ib_dev, port, idev, filter_cookie))
2276				cb(ib_dev, port, idev, cookie);
2277
2278			if (idev)
2279				dev_put(idev);
2280		}
2281}
2282
2283/**
2284 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2285 * @filter: Should we call the callback?
2286 * @filter_cookie: Cookie passed to filter
2287 * @cb: Callback to call for each found RoCE ports
2288 * @cookie: Cookie passed back to the callback
2289 *
2290 * Enumerates all RoCE devices' physical ports which are related
2291 * to netdevices and calls callback() on each device for which
2292 * filter() function returns non zero.
2293 */
2294void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2295			      void *filter_cookie,
2296			      roce_netdev_callback cb,
2297			      void *cookie)
2298{
2299	struct ib_device *dev;
2300	unsigned long index;
2301
2302	down_read(&devices_rwsem);
2303	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2304		ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2305	up_read(&devices_rwsem);
2306}
2307
2308/**
2309 * ib_enum_all_devs - enumerate all ib_devices
2310 * @cb: Callback to call for each found ib_device
2311 *
2312 * Enumerates all ib_devices and calls callback() on each device.
2313 */
2314int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2315		     struct netlink_callback *cb)
2316{
2317	unsigned long index;
2318	struct ib_device *dev;
2319	unsigned int idx = 0;
2320	int ret = 0;
2321
2322	down_read(&devices_rwsem);
2323	xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2324		if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2325			continue;
2326
2327		ret = nldev_cb(dev, skb, cb, idx);
2328		if (ret)
2329			break;
2330		idx++;
2331	}
2332	up_read(&devices_rwsem);
2333	return ret;
2334}
2335
2336/**
2337 * ib_query_pkey - Get P_Key table entry
2338 * @device:Device to query
2339 * @port_num:Port number to query
2340 * @index:P_Key table index to query
2341 * @pkey:Returned P_Key
2342 *
2343 * ib_query_pkey() fetches the specified P_Key table entry.
2344 */
2345int ib_query_pkey(struct ib_device *device,
2346		  u8 port_num, u16 index, u16 *pkey)
2347{
2348	if (!rdma_is_port_valid(device, port_num))
2349		return -EINVAL;
2350
 
 
 
2351	return device->ops.query_pkey(device, port_num, index, pkey);
2352}
2353EXPORT_SYMBOL(ib_query_pkey);
2354
2355/**
2356 * ib_modify_device - Change IB device attributes
2357 * @device:Device to modify
2358 * @device_modify_mask:Mask of attributes to change
2359 * @device_modify:New attribute values
2360 *
2361 * ib_modify_device() changes a device's attributes as specified by
2362 * the @device_modify_mask and @device_modify structure.
2363 */
2364int ib_modify_device(struct ib_device *device,
2365		     int device_modify_mask,
2366		     struct ib_device_modify *device_modify)
2367{
2368	if (!device->ops.modify_device)
2369		return -ENOSYS;
2370
2371	return device->ops.modify_device(device, device_modify_mask,
2372					 device_modify);
2373}
2374EXPORT_SYMBOL(ib_modify_device);
2375
2376/**
2377 * ib_modify_port - Modifies the attributes for the specified port.
2378 * @device: The device to modify.
2379 * @port_num: The number of the port to modify.
2380 * @port_modify_mask: Mask used to specify which attributes of the port
2381 *   to change.
2382 * @port_modify: New attribute values for the port.
2383 *
2384 * ib_modify_port() changes a port's attributes as specified by the
2385 * @port_modify_mask and @port_modify structure.
2386 */
2387int ib_modify_port(struct ib_device *device,
2388		   u8 port_num, int port_modify_mask,
2389		   struct ib_port_modify *port_modify)
2390{
2391	int rc;
2392
2393	if (!rdma_is_port_valid(device, port_num))
2394		return -EINVAL;
2395
2396	if (device->ops.modify_port)
2397		rc = device->ops.modify_port(device, port_num,
2398					     port_modify_mask,
2399					     port_modify);
 
 
 
 
2400	else
2401		rc = rdma_protocol_roce(device, port_num) ? 0 : -ENOSYS;
2402	return rc;
2403}
2404EXPORT_SYMBOL(ib_modify_port);
2405
2406/**
2407 * ib_find_gid - Returns the port number and GID table index where
2408 *   a specified GID value occurs. Its searches only for IB link layer.
2409 * @device: The device to query.
2410 * @gid: The GID value to search for.
2411 * @port_num: The port number of the device where the GID value was found.
2412 * @index: The index into the GID table where the GID was found.  This
2413 *   parameter may be NULL.
2414 */
2415int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2416		u8 *port_num, u16 *index)
2417{
2418	union ib_gid tmp_gid;
2419	unsigned int port;
2420	int ret, i;
2421
2422	rdma_for_each_port (device, port) {
2423		if (!rdma_protocol_ib(device, port))
2424			continue;
2425
2426		for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2427		     ++i) {
2428			ret = rdma_query_gid(device, port, i, &tmp_gid);
2429			if (ret)
2430				return ret;
 
2431			if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2432				*port_num = port;
2433				if (index)
2434					*index = i;
2435				return 0;
2436			}
2437		}
2438	}
2439
2440	return -ENOENT;
2441}
2442EXPORT_SYMBOL(ib_find_gid);
2443
2444/**
2445 * ib_find_pkey - Returns the PKey table index where a specified
2446 *   PKey value occurs.
2447 * @device: The device to query.
2448 * @port_num: The port number of the device to search for the PKey.
2449 * @pkey: The PKey value to search for.
2450 * @index: The index into the PKey table where the PKey was found.
2451 */
2452int ib_find_pkey(struct ib_device *device,
2453		 u8 port_num, u16 pkey, u16 *index)
2454{
2455	int ret, i;
2456	u16 tmp_pkey;
2457	int partial_ix = -1;
2458
2459	for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2460	     ++i) {
2461		ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2462		if (ret)
2463			return ret;
2464		if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2465			/* if there is full-member pkey take it.*/
2466			if (tmp_pkey & 0x8000) {
2467				*index = i;
2468				return 0;
2469			}
2470			if (partial_ix < 0)
2471				partial_ix = i;
2472		}
2473	}
2474
2475	/*no full-member, if exists take the limited*/
2476	if (partial_ix >= 0) {
2477		*index = partial_ix;
2478		return 0;
2479	}
2480	return -ENOENT;
2481}
2482EXPORT_SYMBOL(ib_find_pkey);
2483
2484/**
2485 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2486 * for a received CM request
2487 * @dev:	An RDMA device on which the request has been received.
2488 * @port:	Port number on the RDMA device.
2489 * @pkey:	The Pkey the request came on.
2490 * @gid:	A GID that the net_dev uses to communicate.
2491 * @addr:	Contains the IP address that the request specified as its
2492 *		destination.
2493 *
2494 */
2495struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2496					    u8 port,
2497					    u16 pkey,
2498					    const union ib_gid *gid,
2499					    const struct sockaddr *addr)
2500{
2501	struct net_device *net_dev = NULL;
2502	unsigned long index;
2503	void *client_data;
2504
2505	if (!rdma_protocol_ib(dev, port))
2506		return NULL;
2507
2508	/*
2509	 * Holding the read side guarantees that the client will not become
2510	 * unregistered while we are calling get_net_dev_by_params()
2511	 */
2512	down_read(&dev->client_data_rwsem);
2513	xan_for_each_marked (&dev->client_data, index, client_data,
2514			     CLIENT_DATA_REGISTERED) {
2515		struct ib_client *client = xa_load(&clients, index);
2516
2517		if (!client || !client->get_net_dev_by_params)
2518			continue;
2519
2520		net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2521							addr, client_data);
2522		if (net_dev)
2523			break;
2524	}
2525	up_read(&dev->client_data_rwsem);
2526
2527	return net_dev;
2528}
2529EXPORT_SYMBOL(ib_get_net_dev_by_params);
2530
2531void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2532{
2533	struct ib_device_ops *dev_ops = &dev->ops;
2534#define SET_DEVICE_OP(ptr, name)                                               \
2535	do {                                                                   \
2536		if (ops->name)                                                 \
2537			if (!((ptr)->name))				       \
2538				(ptr)->name = ops->name;                       \
2539	} while (0)
2540
2541#define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2542
2543	if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2544		WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2545			dev_ops->driver_id != ops->driver_id);
2546		dev_ops->driver_id = ops->driver_id;
2547	}
2548	if (ops->owner) {
2549		WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2550		dev_ops->owner = ops->owner;
2551	}
2552	if (ops->uverbs_abi_ver)
2553		dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2554
2555	dev_ops->uverbs_no_driver_id_binding |=
2556		ops->uverbs_no_driver_id_binding;
2557
2558	SET_DEVICE_OP(dev_ops, add_gid);
2559	SET_DEVICE_OP(dev_ops, advise_mr);
2560	SET_DEVICE_OP(dev_ops, alloc_dm);
2561	SET_DEVICE_OP(dev_ops, alloc_fmr);
2562	SET_DEVICE_OP(dev_ops, alloc_hw_stats);
2563	SET_DEVICE_OP(dev_ops, alloc_mr);
2564	SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2565	SET_DEVICE_OP(dev_ops, alloc_mw);
2566	SET_DEVICE_OP(dev_ops, alloc_pd);
2567	SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2568	SET_DEVICE_OP(dev_ops, alloc_ucontext);
2569	SET_DEVICE_OP(dev_ops, alloc_xrcd);
2570	SET_DEVICE_OP(dev_ops, attach_mcast);
2571	SET_DEVICE_OP(dev_ops, check_mr_status);
2572	SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2573	SET_DEVICE_OP(dev_ops, counter_bind_qp);
2574	SET_DEVICE_OP(dev_ops, counter_dealloc);
2575	SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2576	SET_DEVICE_OP(dev_ops, counter_update_stats);
2577	SET_DEVICE_OP(dev_ops, create_ah);
2578	SET_DEVICE_OP(dev_ops, create_counters);
2579	SET_DEVICE_OP(dev_ops, create_cq);
2580	SET_DEVICE_OP(dev_ops, create_flow);
2581	SET_DEVICE_OP(dev_ops, create_flow_action_esp);
2582	SET_DEVICE_OP(dev_ops, create_qp);
2583	SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2584	SET_DEVICE_OP(dev_ops, create_srq);
 
2585	SET_DEVICE_OP(dev_ops, create_wq);
2586	SET_DEVICE_OP(dev_ops, dealloc_dm);
2587	SET_DEVICE_OP(dev_ops, dealloc_driver);
2588	SET_DEVICE_OP(dev_ops, dealloc_fmr);
2589	SET_DEVICE_OP(dev_ops, dealloc_mw);
2590	SET_DEVICE_OP(dev_ops, dealloc_pd);
2591	SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2592	SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2593	SET_DEVICE_OP(dev_ops, del_gid);
2594	SET_DEVICE_OP(dev_ops, dereg_mr);
2595	SET_DEVICE_OP(dev_ops, destroy_ah);
2596	SET_DEVICE_OP(dev_ops, destroy_counters);
2597	SET_DEVICE_OP(dev_ops, destroy_cq);
2598	SET_DEVICE_OP(dev_ops, destroy_flow);
2599	SET_DEVICE_OP(dev_ops, destroy_flow_action);
2600	SET_DEVICE_OP(dev_ops, destroy_qp);
2601	SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2602	SET_DEVICE_OP(dev_ops, destroy_srq);
2603	SET_DEVICE_OP(dev_ops, destroy_wq);
 
2604	SET_DEVICE_OP(dev_ops, detach_mcast);
2605	SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2606	SET_DEVICE_OP(dev_ops, drain_rq);
2607	SET_DEVICE_OP(dev_ops, drain_sq);
2608	SET_DEVICE_OP(dev_ops, enable_driver);
2609	SET_DEVICE_OP(dev_ops, fill_res_entry);
 
 
 
 
 
 
 
 
 
2610	SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2611	SET_DEVICE_OP(dev_ops, get_dma_mr);
2612	SET_DEVICE_OP(dev_ops, get_hw_stats);
2613	SET_DEVICE_OP(dev_ops, get_link_layer);
2614	SET_DEVICE_OP(dev_ops, get_netdev);
 
2615	SET_DEVICE_OP(dev_ops, get_port_immutable);
2616	SET_DEVICE_OP(dev_ops, get_vector_affinity);
2617	SET_DEVICE_OP(dev_ops, get_vf_config);
 
2618	SET_DEVICE_OP(dev_ops, get_vf_stats);
2619	SET_DEVICE_OP(dev_ops, init_port);
2620	SET_DEVICE_OP(dev_ops, invalidate_range);
2621	SET_DEVICE_OP(dev_ops, iw_accept);
2622	SET_DEVICE_OP(dev_ops, iw_add_ref);
2623	SET_DEVICE_OP(dev_ops, iw_connect);
2624	SET_DEVICE_OP(dev_ops, iw_create_listen);
2625	SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2626	SET_DEVICE_OP(dev_ops, iw_get_qp);
2627	SET_DEVICE_OP(dev_ops, iw_reject);
2628	SET_DEVICE_OP(dev_ops, iw_rem_ref);
2629	SET_DEVICE_OP(dev_ops, map_mr_sg);
2630	SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2631	SET_DEVICE_OP(dev_ops, map_phys_fmr);
2632	SET_DEVICE_OP(dev_ops, mmap);
 
2633	SET_DEVICE_OP(dev_ops, modify_ah);
2634	SET_DEVICE_OP(dev_ops, modify_cq);
2635	SET_DEVICE_OP(dev_ops, modify_device);
2636	SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
2637	SET_DEVICE_OP(dev_ops, modify_port);
2638	SET_DEVICE_OP(dev_ops, modify_qp);
2639	SET_DEVICE_OP(dev_ops, modify_srq);
2640	SET_DEVICE_OP(dev_ops, modify_wq);
2641	SET_DEVICE_OP(dev_ops, peek_cq);
2642	SET_DEVICE_OP(dev_ops, poll_cq);
 
2643	SET_DEVICE_OP(dev_ops, post_recv);
2644	SET_DEVICE_OP(dev_ops, post_send);
2645	SET_DEVICE_OP(dev_ops, post_srq_recv);
2646	SET_DEVICE_OP(dev_ops, process_mad);
2647	SET_DEVICE_OP(dev_ops, query_ah);
2648	SET_DEVICE_OP(dev_ops, query_device);
2649	SET_DEVICE_OP(dev_ops, query_gid);
2650	SET_DEVICE_OP(dev_ops, query_pkey);
2651	SET_DEVICE_OP(dev_ops, query_port);
2652	SET_DEVICE_OP(dev_ops, query_qp);
2653	SET_DEVICE_OP(dev_ops, query_srq);
 
2654	SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2655	SET_DEVICE_OP(dev_ops, read_counters);
2656	SET_DEVICE_OP(dev_ops, reg_dm_mr);
2657	SET_DEVICE_OP(dev_ops, reg_user_mr);
2658	SET_DEVICE_OP(dev_ops, req_ncomp_notif);
2659	SET_DEVICE_OP(dev_ops, req_notify_cq);
2660	SET_DEVICE_OP(dev_ops, rereg_user_mr);
2661	SET_DEVICE_OP(dev_ops, resize_cq);
2662	SET_DEVICE_OP(dev_ops, set_vf_guid);
2663	SET_DEVICE_OP(dev_ops, set_vf_link_state);
2664	SET_DEVICE_OP(dev_ops, unmap_fmr);
2665
2666	SET_OBJ_SIZE(dev_ops, ib_ah);
 
2667	SET_OBJ_SIZE(dev_ops, ib_cq);
 
2668	SET_OBJ_SIZE(dev_ops, ib_pd);
 
 
2669	SET_OBJ_SIZE(dev_ops, ib_srq);
2670	SET_OBJ_SIZE(dev_ops, ib_ucontext);
 
2671}
2672EXPORT_SYMBOL(ib_set_device_ops);
2673
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2674static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2675	[RDMA_NL_LS_OP_RESOLVE] = {
2676		.doit = ib_nl_handle_resolve_resp,
2677		.flags = RDMA_NL_ADMIN_PERM,
2678	},
2679	[RDMA_NL_LS_OP_SET_TIMEOUT] = {
2680		.doit = ib_nl_handle_set_timeout,
2681		.flags = RDMA_NL_ADMIN_PERM,
2682	},
2683	[RDMA_NL_LS_OP_IP_RESOLVE] = {
2684		.doit = ib_nl_handle_ip_res_resp,
2685		.flags = RDMA_NL_ADMIN_PERM,
2686	},
2687};
2688
2689static int __init ib_core_init(void)
2690{
2691	int ret;
2692
2693	ib_wq = alloc_workqueue("infiniband", 0, 0);
2694	if (!ib_wq)
2695		return -ENOMEM;
2696
 
 
 
 
 
2697	ib_comp_wq = alloc_workqueue("ib-comp-wq",
2698			WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2699	if (!ib_comp_wq) {
2700		ret = -ENOMEM;
2701		goto err;
2702	}
2703
2704	ib_comp_unbound_wq =
2705		alloc_workqueue("ib-comp-unb-wq",
2706				WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2707				WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2708	if (!ib_comp_unbound_wq) {
2709		ret = -ENOMEM;
2710		goto err_comp;
2711	}
2712
2713	ret = class_register(&ib_class);
2714	if (ret) {
2715		pr_warn("Couldn't create InfiniBand device class\n");
2716		goto err_comp_unbound;
2717	}
2718
2719	rdma_nl_init();
2720
2721	ret = addr_init();
2722	if (ret) {
2723		pr_warn("Could't init IB address resolution\n");
2724		goto err_ibnl;
2725	}
2726
2727	ret = ib_mad_init();
2728	if (ret) {
2729		pr_warn("Couldn't init IB MAD\n");
2730		goto err_addr;
2731	}
2732
2733	ret = ib_sa_init();
2734	if (ret) {
2735		pr_warn("Couldn't init SA\n");
2736		goto err_mad;
2737	}
2738
2739	ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2740	if (ret) {
2741		pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2742		goto err_sa;
2743	}
2744
2745	ret = register_pernet_device(&rdma_dev_net_ops);
2746	if (ret) {
2747		pr_warn("Couldn't init compat dev. ret %d\n", ret);
2748		goto err_compat;
2749	}
2750
2751	nldev_init();
2752	rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2753	roce_gid_mgmt_init();
 
 
 
 
2754
2755	return 0;
2756
 
 
 
 
2757err_compat:
2758	unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2759err_sa:
2760	ib_sa_cleanup();
2761err_mad:
2762	ib_mad_cleanup();
2763err_addr:
2764	addr_cleanup();
2765err_ibnl:
2766	class_unregister(&ib_class);
2767err_comp_unbound:
2768	destroy_workqueue(ib_comp_unbound_wq);
2769err_comp:
2770	destroy_workqueue(ib_comp_wq);
 
 
2771err:
2772	destroy_workqueue(ib_wq);
2773	return ret;
2774}
2775
2776static void __exit ib_core_cleanup(void)
2777{
2778	roce_gid_mgmt_cleanup();
 
2779	nldev_exit();
2780	rdma_nl_unregister(RDMA_NL_LS);
2781	unregister_pernet_device(&rdma_dev_net_ops);
2782	unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2783	ib_sa_cleanup();
2784	ib_mad_cleanup();
2785	addr_cleanup();
2786	rdma_nl_exit();
2787	class_unregister(&ib_class);
2788	destroy_workqueue(ib_comp_unbound_wq);
2789	destroy_workqueue(ib_comp_wq);
2790	/* Make sure that any pending umem accounting work is done. */
2791	destroy_workqueue(ib_wq);
2792	flush_workqueue(system_unbound_wq);
2793	WARN_ON(!xa_empty(&clients));
2794	WARN_ON(!xa_empty(&devices));
2795}
2796
2797MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2798
2799/* ib core relies on netdev stack to first register net_ns_type_operations
2800 * ns kobject type before ib_core initialization.
2801 */
2802fs_initcall(ib_core_init);
2803module_exit(ib_core_cleanup);