cache.c

   1/*
   2 * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3 * Copyright (c) 2005 Intel Corporation. All rights reserved.
   4 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   5 * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6 *
   7 * This software is available to you under a choice of one of two
   8 * licenses.  You may choose to be licensed under the terms of the GNU
   9 * General Public License (GPL) Version 2, available from the file
  10 * COPYING in the main directory of this source tree, or the
  11 * OpenIB.org BSD license below:
  12 *
  13 *     Redistribution and use in source and binary forms, with or
  14 *     without modification, are permitted provided that the following
  15 *     conditions are met:
  16 *
  17 *      - Redistributions of source code must retain the above
  18 *        copyright notice, this list of conditions and the following
  19 *        disclaimer.
  20 *
  21 *      - Redistributions in binary form must reproduce the above
  22 *        copyright notice, this list of conditions and the following
  23 *        disclaimer in the documentation and/or other materials
  24 *        provided with the distribution.
  25 *
  26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33 * SOFTWARE.
  34 */
  35
  36#include <linux/if_vlan.h>
  37#include <linux/errno.h>
  38#include <linux/slab.h>
  39#include <linux/workqueue.h>
  40#include <linux/netdevice.h>
  41#include <net/addrconf.h>
  42
  43#include <rdma/ib_cache.h>
  44
  45#include "core_priv.h"
  46
  47struct ib_pkey_cache {
  48	int             table_len;
  49	u16             table[] __counted_by(table_len);
  50};
  51
  52struct ib_update_work {
  53	struct work_struct work;
  54	struct ib_event event;
  55	bool enforce_security;
  56};
  57
  58union ib_gid zgid;
  59EXPORT_SYMBOL(zgid);
  60
  61enum gid_attr_find_mask {
  62	GID_ATTR_FIND_MASK_GID          = 1UL << 0,
  63	GID_ATTR_FIND_MASK_NETDEV	= 1UL << 1,
  64	GID_ATTR_FIND_MASK_DEFAULT	= 1UL << 2,
  65	GID_ATTR_FIND_MASK_GID_TYPE	= 1UL << 3,
  66};
  67
  68enum gid_table_entry_state {
  69	GID_TABLE_ENTRY_INVALID		= 1,
  70	GID_TABLE_ENTRY_VALID		= 2,
  71	/*
  72	 * Indicates that entry is pending to be removed, there may
  73	 * be active users of this GID entry.
  74	 * When last user of the GID entry releases reference to it,
  75	 * GID entry is detached from the table.
  76	 */
  77	GID_TABLE_ENTRY_PENDING_DEL	= 3,
  78};
  79
  80struct roce_gid_ndev_storage {
  81	struct rcu_head rcu_head;
  82	struct net_device *ndev;
  83};
  84
  85struct ib_gid_table_entry {
  86	struct kref			kref;
  87	struct work_struct		del_work;
  88	struct ib_gid_attr		attr;
  89	void				*context;
  90	/* Store the ndev pointer to release reference later on in
  91	 * call_rcu context because by that time gid_table_entry
  92	 * and attr might be already freed. So keep a copy of it.
  93	 * ndev_storage is freed by rcu callback.
  94	 */
  95	struct roce_gid_ndev_storage	*ndev_storage;
  96	enum gid_table_entry_state	state;
  97};
  98
  99struct ib_gid_table {
 100	int				sz;
 101	/* In RoCE, adding a GID to the table requires:
 102	 * (a) Find if this GID is already exists.
 103	 * (b) Find a free space.
 104	 * (c) Write the new GID
 105	 *
 106	 * Delete requires different set of operations:
 107	 * (a) Find the GID
 108	 * (b) Delete it.
 109	 *
 110	 **/
 111	/* Any writer to data_vec must hold this lock and the write side of
 112	 * rwlock. Readers must hold only rwlock. All writers must be in a
 113	 * sleepable context.
 114	 */
 115	struct mutex			lock;
 116	/* rwlock protects data_vec[ix]->state and entry pointer.
 117	 */
 118	rwlock_t			rwlock;
 119	struct ib_gid_table_entry	**data_vec;
 120	/* bit field, each bit indicates the index of default GID */
 121	u32				default_gid_indices;
 122};
 123
 124static void dispatch_gid_change_event(struct ib_device *ib_dev, u32 port)
 125{
 126	struct ib_event event;
 127
 128	event.device		= ib_dev;
 129	event.element.port_num	= port;
 130	event.event		= IB_EVENT_GID_CHANGE;
 131
 132	ib_dispatch_event_clients(&event);
 133}
 134
 135static const char * const gid_type_str[] = {
 136	/* IB/RoCE v1 value is set for IB_GID_TYPE_IB and IB_GID_TYPE_ROCE for
 137	 * user space compatibility reasons.
 138	 */
 139	[IB_GID_TYPE_IB]	= "IB/RoCE v1",
 140	[IB_GID_TYPE_ROCE]	= "IB/RoCE v1",
 141	[IB_GID_TYPE_ROCE_UDP_ENCAP]	= "RoCE v2",
 142};
 143
 144const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
 145{
 146	if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
 147		return gid_type_str[gid_type];
 148
 149	return "Invalid GID type";
 150}
 151EXPORT_SYMBOL(ib_cache_gid_type_str);
 152
 153/** rdma_is_zero_gid - Check if given GID is zero or not.
 154 * @gid:	GID to check
 155 * Returns true if given GID is zero, returns false otherwise.
 156 */
 157bool rdma_is_zero_gid(const union ib_gid *gid)
 158{
 159	return !memcmp(gid, &zgid, sizeof(*gid));
 160}
 161EXPORT_SYMBOL(rdma_is_zero_gid);
 162
 163/** is_gid_index_default - Check if a given index belongs to
 164 * reserved default GIDs or not.
 165 * @table:	GID table pointer
 166 * @index:	Index to check in GID table
 167 * Returns true if index is one of the reserved default GID index otherwise
 168 * returns false.
 169 */
 170static bool is_gid_index_default(const struct ib_gid_table *table,
 171				 unsigned int index)
 172{
 173	return index < 32 && (BIT(index) & table->default_gid_indices);
 174}
 175
 176int ib_cache_gid_parse_type_str(const char *buf)
 177{
 178	unsigned int i;
 179	size_t len;
 180	int err = -EINVAL;
 181
 182	len = strlen(buf);
 183	if (len == 0)
 184		return -EINVAL;
 185
 186	if (buf[len - 1] == '\n')
 187		len--;
 188
 189	for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i)
 190		if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) &&
 191		    len == strlen(gid_type_str[i])) {
 192			err = i;
 193			break;
 194		}
 195
 196	return err;
 197}
 198EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
 199
 200static struct ib_gid_table *rdma_gid_table(struct ib_device *device, u32 port)
 201{
 202	return device->port_data[port].cache.gid;
 203}
 204
 205static bool is_gid_entry_free(const struct ib_gid_table_entry *entry)
 206{
 207	return !entry;
 208}
 209
 210static bool is_gid_entry_valid(const struct ib_gid_table_entry *entry)
 211{
 212	return entry && entry->state == GID_TABLE_ENTRY_VALID;
 213}
 214
 215static void schedule_free_gid(struct kref *kref)
 216{
 217	struct ib_gid_table_entry *entry =
 218			container_of(kref, struct ib_gid_table_entry, kref);
 219
 220	queue_work(ib_wq, &entry->del_work);
 221}
 222
 223static void put_gid_ndev(struct rcu_head *head)
 224{
 225	struct roce_gid_ndev_storage *storage =
 226		container_of(head, struct roce_gid_ndev_storage, rcu_head);
 227
 228	WARN_ON(!storage->ndev);
 229	/* At this point its safe to release netdev reference,
 230	 * as all callers working on gid_attr->ndev are done
 231	 * using this netdev.
 232	 */
 233	dev_put(storage->ndev);
 234	kfree(storage);
 235}
 236
 237static void free_gid_entry_locked(struct ib_gid_table_entry *entry)
 238{
 239	struct ib_device *device = entry->attr.device;
 240	u32 port_num = entry->attr.port_num;
 241	struct ib_gid_table *table = rdma_gid_table(device, port_num);
 242
 243	dev_dbg(&device->dev, "%s port=%u index=%u gid %pI6\n", __func__,
 244		port_num, entry->attr.index, entry->attr.gid.raw);
 245
 246	write_lock_irq(&table->rwlock);
 247
 248	/*
 249	 * The only way to avoid overwriting NULL in table is
 250	 * by comparing if it is same entry in table or not!
 251	 * If new entry in table is added by the time we free here,
 252	 * don't overwrite the table entry.
 253	 */
 254	if (entry == table->data_vec[entry->attr.index])
 255		table->data_vec[entry->attr.index] = NULL;
 256	/* Now this index is ready to be allocated */
 257	write_unlock_irq(&table->rwlock);
 258
 259	if (entry->ndev_storage)
 260		call_rcu(&entry->ndev_storage->rcu_head, put_gid_ndev);
 261	kfree(entry);
 262}
 263
 264static void free_gid_entry(struct kref *kref)
 265{
 266	struct ib_gid_table_entry *entry =
 267			container_of(kref, struct ib_gid_table_entry, kref);
 268
 269	free_gid_entry_locked(entry);
 270}
 271
 272/**
 273 * free_gid_work - Release reference to the GID entry
 274 * @work: Work structure to refer to GID entry which needs to be
 275 * deleted.
 276 *
 277 * free_gid_work() frees the entry from the HCA's hardware table
 278 * if provider supports it. It releases reference to netdevice.
 279 */
 280static void free_gid_work(struct work_struct *work)
 281{
 282	struct ib_gid_table_entry *entry =
 283		container_of(work, struct ib_gid_table_entry, del_work);
 284	struct ib_device *device = entry->attr.device;
 285	u32 port_num = entry->attr.port_num;
 286	struct ib_gid_table *table = rdma_gid_table(device, port_num);
 287
 288	mutex_lock(&table->lock);
 289	free_gid_entry_locked(entry);
 290	mutex_unlock(&table->lock);
 291}
 292
 293static struct ib_gid_table_entry *
 294alloc_gid_entry(const struct ib_gid_attr *attr)
 295{
 296	struct ib_gid_table_entry *entry;
 297	struct net_device *ndev;
 298
 299	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 300	if (!entry)
 301		return NULL;
 302
 303	ndev = rcu_dereference_protected(attr->ndev, 1);
 304	if (ndev) {
 305		entry->ndev_storage = kzalloc(sizeof(*entry->ndev_storage),
 306					      GFP_KERNEL);
 307		if (!entry->ndev_storage) {
 308			kfree(entry);
 309			return NULL;
 310		}
 311		dev_hold(ndev);
 312		entry->ndev_storage->ndev = ndev;
 313	}
 314	kref_init(&entry->kref);
 315	memcpy(&entry->attr, attr, sizeof(*attr));
 316	INIT_WORK(&entry->del_work, free_gid_work);
 317	entry->state = GID_TABLE_ENTRY_INVALID;
 318	return entry;
 319}
 320
 321static void store_gid_entry(struct ib_gid_table *table,
 322			    struct ib_gid_table_entry *entry)
 323{
 324	entry->state = GID_TABLE_ENTRY_VALID;
 325
 326	dev_dbg(&entry->attr.device->dev, "%s port=%u index=%u gid %pI6\n",
 327		__func__, entry->attr.port_num, entry->attr.index,
 328		entry->attr.gid.raw);
 329
 330	lockdep_assert_held(&table->lock);
 331	write_lock_irq(&table->rwlock);
 332	table->data_vec[entry->attr.index] = entry;
 333	write_unlock_irq(&table->rwlock);
 334}
 335
 336static void get_gid_entry(struct ib_gid_table_entry *entry)
 337{
 338	kref_get(&entry->kref);
 339}
 340
 341static void put_gid_entry(struct ib_gid_table_entry *entry)
 342{
 343	kref_put(&entry->kref, schedule_free_gid);
 344}
 345
 346static void put_gid_entry_locked(struct ib_gid_table_entry *entry)
 347{
 348	kref_put(&entry->kref, free_gid_entry);
 349}
 350
 351static int add_roce_gid(struct ib_gid_table_entry *entry)
 352{
 353	const struct ib_gid_attr *attr = &entry->attr;
 354	int ret;
 355
 356	if (!attr->ndev) {
 357		dev_err(&attr->device->dev, "%s NULL netdev port=%u index=%u\n",
 358			__func__, attr->port_num, attr->index);
 359		return -EINVAL;
 360	}
 361	if (rdma_cap_roce_gid_table(attr->device, attr->port_num)) {
 362		ret = attr->device->ops.add_gid(attr, &entry->context);
 363		if (ret) {
 364			dev_err(&attr->device->dev,
 365				"%s GID add failed port=%u index=%u\n",
 366				__func__, attr->port_num, attr->index);
 367			return ret;
 368		}
 369	}
 370	return 0;
 371}
 372
 373/**
 374 * del_gid - Delete GID table entry
 375 *
 376 * @ib_dev:	IB device whose GID entry to be deleted
 377 * @port:	Port number of the IB device
 378 * @table:	GID table of the IB device for a port
 379 * @ix:		GID entry index to delete
 380 *
 381 */
 382static void del_gid(struct ib_device *ib_dev, u32 port,
 383		    struct ib_gid_table *table, int ix)
 384{
 385	struct roce_gid_ndev_storage *ndev_storage;
 386	struct ib_gid_table_entry *entry;
 387
 388	lockdep_assert_held(&table->lock);
 389
 390	dev_dbg(&ib_dev->dev, "%s port=%u index=%d gid %pI6\n", __func__, port,
 391		ix, table->data_vec[ix]->attr.gid.raw);
 392
 393	write_lock_irq(&table->rwlock);
 394	entry = table->data_vec[ix];
 395	entry->state = GID_TABLE_ENTRY_PENDING_DEL;
 396	/*
 397	 * For non RoCE protocol, GID entry slot is ready to use.
 398	 */
 399	if (!rdma_protocol_roce(ib_dev, port))
 400		table->data_vec[ix] = NULL;
 401	write_unlock_irq(&table->rwlock);
 402
 403	if (rdma_cap_roce_gid_table(ib_dev, port))
 404		ib_dev->ops.del_gid(&entry->attr, &entry->context);
 405
 406	ndev_storage = entry->ndev_storage;
 407	if (ndev_storage) {
 408		entry->ndev_storage = NULL;
 409		rcu_assign_pointer(entry->attr.ndev, NULL);
 410		call_rcu(&ndev_storage->rcu_head, put_gid_ndev);
 411	}
 412
 413	put_gid_entry_locked(entry);
 414}
 415
 416/**
 417 * add_modify_gid - Add or modify GID table entry
 418 *
 419 * @table:	GID table in which GID to be added or modified
 420 * @attr:	Attributes of the GID
 421 *
 422 * Returns 0 on success or appropriate error code. It accepts zero
 423 * GID addition for non RoCE ports for HCA's who report them as valid
 424 * GID. However such zero GIDs are not added to the cache.
 425 */
 426static int add_modify_gid(struct ib_gid_table *table,
 427			  const struct ib_gid_attr *attr)
 428{
 429	struct ib_gid_table_entry *entry;
 430	int ret = 0;
 431
 432	/*
 433	 * Invalidate any old entry in the table to make it safe to write to
 434	 * this index.
 435	 */
 436	if (is_gid_entry_valid(table->data_vec[attr->index]))
 437		del_gid(attr->device, attr->port_num, table, attr->index);
 438
 439	/*
 440	 * Some HCA's report multiple GID entries with only one valid GID, and
 441	 * leave other unused entries as the zero GID. Convert zero GIDs to
 442	 * empty table entries instead of storing them.
 443	 */
 444	if (rdma_is_zero_gid(&attr->gid))
 445		return 0;
 446
 447	entry = alloc_gid_entry(attr);
 448	if (!entry)
 449		return -ENOMEM;
 450
 451	if (rdma_protocol_roce(attr->device, attr->port_num)) {
 452		ret = add_roce_gid(entry);
 453		if (ret)
 454			goto done;
 455	}
 456
 457	store_gid_entry(table, entry);
 458	return 0;
 459
 460done:
 461	put_gid_entry(entry);
 462	return ret;
 463}
 464
 465/* rwlock should be read locked, or lock should be held */
 466static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
 467		    const struct ib_gid_attr *val, bool default_gid,
 468		    unsigned long mask, int *pempty)
 469{
 470	int i = 0;
 471	int found = -1;
 472	int empty = pempty ? -1 : 0;
 473
 474	while (i < table->sz && (found < 0 || empty < 0)) {
 475		struct ib_gid_table_entry *data = table->data_vec[i];
 476		struct ib_gid_attr *attr;
 477		int curr_index = i;
 478
 479		i++;
 480
 481		/* find_gid() is used during GID addition where it is expected
 482		 * to return a free entry slot which is not duplicate.
 483		 * Free entry slot is requested and returned if pempty is set,
 484		 * so lookup free slot only if requested.
 485		 */
 486		if (pempty && empty < 0) {
 487			if (is_gid_entry_free(data) &&
 488			    default_gid ==
 489				is_gid_index_default(table, curr_index)) {
 490				/*
 491				 * Found an invalid (free) entry; allocate it.
 492				 * If default GID is requested, then our
 493				 * found slot must be one of the DEFAULT
 494				 * reserved slots or we fail.
 495				 * This ensures that only DEFAULT reserved
 496				 * slots are used for default property GIDs.
 497				 */
 498				empty = curr_index;
 499			}
 500		}
 501
 502		/*
 503		 * Additionally find_gid() is used to find valid entry during
 504		 * lookup operation; so ignore the entries which are marked as
 505		 * pending for removal and the entries which are marked as
 506		 * invalid.
 507		 */
 508		if (!is_gid_entry_valid(data))
 509			continue;
 510
 511		if (found >= 0)
 512			continue;
 513
 514		attr = &data->attr;
 515		if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
 516		    attr->gid_type != val->gid_type)
 517			continue;
 518
 519		if (mask & GID_ATTR_FIND_MASK_GID &&
 520		    memcmp(gid, &data->attr.gid, sizeof(*gid)))
 521			continue;
 522
 523		if (mask & GID_ATTR_FIND_MASK_NETDEV &&
 524		    attr->ndev != val->ndev)
 525			continue;
 526
 527		if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
 528		    is_gid_index_default(table, curr_index) != default_gid)
 529			continue;
 530
 531		found = curr_index;
 532	}
 533
 534	if (pempty)
 535		*pempty = empty;
 536
 537	return found;
 538}
 539
 540static void make_default_gid(struct  net_device *dev, union ib_gid *gid)
 541{
 542	gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
 543	addrconf_ifid_eui48(&gid->raw[8], dev);
 544}
 545
 546static int __ib_cache_gid_add(struct ib_device *ib_dev, u32 port,
 547			      union ib_gid *gid, struct ib_gid_attr *attr,
 548			      unsigned long mask, bool default_gid)
 549{
 550	struct ib_gid_table *table;
 551	int ret = 0;
 552	int empty;
 553	int ix;
 554
 555	/* Do not allow adding zero GID in support of
 556	 * IB spec version 1.3 section 4.1.1 point (6) and
 557	 * section 12.7.10 and section 12.7.20
 558	 */
 559	if (rdma_is_zero_gid(gid))
 560		return -EINVAL;
 561
 562	table = rdma_gid_table(ib_dev, port);
 563
 564	mutex_lock(&table->lock);
 565
 566	ix = find_gid(table, gid, attr, default_gid, mask, &empty);
 567	if (ix >= 0)
 568		goto out_unlock;
 569
 570	if (empty < 0) {
 571		ret = -ENOSPC;
 572		goto out_unlock;
 573	}
 574	attr->device = ib_dev;
 575	attr->index = empty;
 576	attr->port_num = port;
 577	attr->gid = *gid;
 578	ret = add_modify_gid(table, attr);
 579	if (!ret)
 580		dispatch_gid_change_event(ib_dev, port);
 581
 582out_unlock:
 583	mutex_unlock(&table->lock);
 584	if (ret)
 585		pr_warn_ratelimited("%s: unable to add gid %pI6 error=%d\n",
 586				    __func__, gid->raw, ret);
 587	return ret;
 588}
 589
 590int ib_cache_gid_add(struct ib_device *ib_dev, u32 port,
 591		     union ib_gid *gid, struct ib_gid_attr *attr)
 592{
 593	unsigned long mask = GID_ATTR_FIND_MASK_GID |
 594			     GID_ATTR_FIND_MASK_GID_TYPE |
 595			     GID_ATTR_FIND_MASK_NETDEV;
 596
 597	return __ib_cache_gid_add(ib_dev, port, gid, attr, mask, false);
 598}
 599
 600static int
 601_ib_cache_gid_del(struct ib_device *ib_dev, u32 port,
 602		  union ib_gid *gid, struct ib_gid_attr *attr,
 603		  unsigned long mask, bool default_gid)
 604{
 605	struct ib_gid_table *table;
 606	int ret = 0;
 607	int ix;
 608
 609	table = rdma_gid_table(ib_dev, port);
 610
 611	mutex_lock(&table->lock);
 612
 613	ix = find_gid(table, gid, attr, default_gid, mask, NULL);
 614	if (ix < 0) {
 615		ret = -EINVAL;
 616		goto out_unlock;
 617	}
 618
 619	del_gid(ib_dev, port, table, ix);
 620	dispatch_gid_change_event(ib_dev, port);
 621
 622out_unlock:
 623	mutex_unlock(&table->lock);
 624	if (ret)
 625		pr_debug("%s: can't delete gid %pI6 error=%d\n",
 626			 __func__, gid->raw, ret);
 627	return ret;
 628}
 629
 630int ib_cache_gid_del(struct ib_device *ib_dev, u32 port,
 631		     union ib_gid *gid, struct ib_gid_attr *attr)
 632{
 633	unsigned long mask = GID_ATTR_FIND_MASK_GID	  |
 634			     GID_ATTR_FIND_MASK_GID_TYPE |
 635			     GID_ATTR_FIND_MASK_DEFAULT  |
 636			     GID_ATTR_FIND_MASK_NETDEV;
 637
 638	return _ib_cache_gid_del(ib_dev, port, gid, attr, mask, false);
 639}
 640
 641int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u32 port,
 642				     struct net_device *ndev)
 643{
 644	struct ib_gid_table *table;
 645	int ix;
 646	bool deleted = false;
 647
 648	table = rdma_gid_table(ib_dev, port);
 649
 650	mutex_lock(&table->lock);
 651
 652	for (ix = 0; ix < table->sz; ix++) {
 653		if (is_gid_entry_valid(table->data_vec[ix]) &&
 654		    table->data_vec[ix]->attr.ndev == ndev) {
 655			del_gid(ib_dev, port, table, ix);
 656			deleted = true;
 657		}
 658	}
 659
 660	mutex_unlock(&table->lock);
 661
 662	if (deleted)
 663		dispatch_gid_change_event(ib_dev, port);
 664
 665	return 0;
 666}
 667
 668/**
 669 * rdma_find_gid_by_port - Returns the GID entry attributes when it finds
 670 * a valid GID entry for given search parameters. It searches for the specified
 671 * GID value in the local software cache.
 672 * @ib_dev: The device to query.
 673 * @gid: The GID value to search for.
 674 * @gid_type: The GID type to search for.
 675 * @port: The port number of the device where the GID value should be searched.
 676 * @ndev: In RoCE, the net device of the device. NULL means ignore.
 677 *
 678 * Returns sgid attributes if the GID is found with valid reference or
 679 * returns ERR_PTR for the error.
 680 * The caller must invoke rdma_put_gid_attr() to release the reference.
 681 */
 682const struct ib_gid_attr *
 683rdma_find_gid_by_port(struct ib_device *ib_dev,
 684		      const union ib_gid *gid,
 685		      enum ib_gid_type gid_type,
 686		      u32 port, struct net_device *ndev)
 687{
 688	int local_index;
 689	struct ib_gid_table *table;
 690	unsigned long mask = GID_ATTR_FIND_MASK_GID |
 691			     GID_ATTR_FIND_MASK_GID_TYPE;
 692	struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
 693	const struct ib_gid_attr *attr;
 694	unsigned long flags;
 695
 696	if (!rdma_is_port_valid(ib_dev, port))
 697		return ERR_PTR(-ENOENT);
 698
 699	table = rdma_gid_table(ib_dev, port);
 700
 701	if (ndev)
 702		mask |= GID_ATTR_FIND_MASK_NETDEV;
 703
 704	read_lock_irqsave(&table->rwlock, flags);
 705	local_index = find_gid(table, gid, &val, false, mask, NULL);
 706	if (local_index >= 0) {
 707		get_gid_entry(table->data_vec[local_index]);
 708		attr = &table->data_vec[local_index]->attr;
 709		read_unlock_irqrestore(&table->rwlock, flags);
 710		return attr;
 711	}
 712
 713	read_unlock_irqrestore(&table->rwlock, flags);
 714	return ERR_PTR(-ENOENT);
 715}
 716EXPORT_SYMBOL(rdma_find_gid_by_port);
 717
 718/**
 719 * rdma_find_gid_by_filter - Returns the GID table attribute where a
 720 * specified GID value occurs
 721 * @ib_dev: The device to query.
 722 * @gid: The GID value to search for.
 723 * @port: The port number of the device where the GID value could be
 724 *   searched.
 725 * @filter: The filter function is executed on any matching GID in the table.
 726 *   If the filter function returns true, the corresponding index is returned,
 727 *   otherwise, we continue searching the GID table. It's guaranteed that
 728 *   while filter is executed, ndev field is valid and the structure won't
 729 *   change. filter is executed in an atomic context. filter must not be NULL.
 730 * @context: Private data to pass into the call-back.
 731 *
 732 * rdma_find_gid_by_filter() searches for the specified GID value
 733 * of which the filter function returns true in the port's GID table.
 734 *
 735 */
 736const struct ib_gid_attr *rdma_find_gid_by_filter(
 737	struct ib_device *ib_dev, const union ib_gid *gid, u32 port,
 738	bool (*filter)(const union ib_gid *gid, const struct ib_gid_attr *,
 739		       void *),
 740	void *context)
 741{
 742	const struct ib_gid_attr *res = ERR_PTR(-ENOENT);
 743	struct ib_gid_table *table;
 744	unsigned long flags;
 745	unsigned int i;
 746
 747	if (!rdma_is_port_valid(ib_dev, port))
 748		return ERR_PTR(-EINVAL);
 749
 750	table = rdma_gid_table(ib_dev, port);
 751
 752	read_lock_irqsave(&table->rwlock, flags);
 753	for (i = 0; i < table->sz; i++) {
 754		struct ib_gid_table_entry *entry = table->data_vec[i];
 755
 756		if (!is_gid_entry_valid(entry))
 757			continue;
 758
 759		if (memcmp(gid, &entry->attr.gid, sizeof(*gid)))
 760			continue;
 761
 762		if (filter(gid, &entry->attr, context)) {
 763			get_gid_entry(entry);
 764			res = &entry->attr;
 765			break;
 766		}
 767	}
 768	read_unlock_irqrestore(&table->rwlock, flags);
 769	return res;
 770}
 771
 772static struct ib_gid_table *alloc_gid_table(int sz)
 773{
 774	struct ib_gid_table *table = kzalloc(sizeof(*table), GFP_KERNEL);
 775
 776	if (!table)
 777		return NULL;
 778
 779	table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
 780	if (!table->data_vec)
 781		goto err_free_table;
 782
 783	mutex_init(&table->lock);
 784
 785	table->sz = sz;
 786	rwlock_init(&table->rwlock);
 787	return table;
 788
 789err_free_table:
 790	kfree(table);
 791	return NULL;
 792}
 793
 794static void release_gid_table(struct ib_device *device,
 795			      struct ib_gid_table *table)
 796{
 797	int i;
 798
 799	if (!table)
 800		return;
 801
 802	for (i = 0; i < table->sz; i++) {
 803		if (is_gid_entry_free(table->data_vec[i]))
 804			continue;
 805
 806		WARN_ONCE(true,
 807			  "GID entry ref leak for dev %s index %d ref=%u\n",
 808			  dev_name(&device->dev), i,
 809			  kref_read(&table->data_vec[i]->kref));
 810	}
 811
 812	mutex_destroy(&table->lock);
 813	kfree(table->data_vec);
 814	kfree(table);
 815}
 816
 817static void cleanup_gid_table_port(struct ib_device *ib_dev, u32 port,
 818				   struct ib_gid_table *table)
 819{
 820	int i;
 821
 822	if (!table)
 823		return;
 824
 825	mutex_lock(&table->lock);
 826	for (i = 0; i < table->sz; ++i) {
 827		if (is_gid_entry_valid(table->data_vec[i]))
 828			del_gid(ib_dev, port, table, i);
 829	}
 830	mutex_unlock(&table->lock);
 831}
 832
 833void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u32 port,
 834				  struct net_device *ndev,
 835				  unsigned long gid_type_mask,
 836				  enum ib_cache_gid_default_mode mode)
 837{
 838	union ib_gid gid = { };
 839	struct ib_gid_attr gid_attr;
 840	unsigned int gid_type;
 841	unsigned long mask;
 842
 843	mask = GID_ATTR_FIND_MASK_GID_TYPE |
 844	       GID_ATTR_FIND_MASK_DEFAULT |
 845	       GID_ATTR_FIND_MASK_NETDEV;
 846	memset(&gid_attr, 0, sizeof(gid_attr));
 847	gid_attr.ndev = ndev;
 848
 849	for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
 850		if (1UL << gid_type & ~gid_type_mask)
 851			continue;
 852
 853		gid_attr.gid_type = gid_type;
 854
 855		if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
 856			make_default_gid(ndev, &gid);
 857			__ib_cache_gid_add(ib_dev, port, &gid,
 858					   &gid_attr, mask, true);
 859		} else if (mode == IB_CACHE_GID_DEFAULT_MODE_DELETE) {
 860			_ib_cache_gid_del(ib_dev, port, &gid,
 861					  &gid_attr, mask, true);
 862		}
 863	}
 864}
 865
 866static void gid_table_reserve_default(struct ib_device *ib_dev, u32 port,
 867				      struct ib_gid_table *table)
 868{
 869	unsigned int i;
 870	unsigned long roce_gid_type_mask;
 871	unsigned int num_default_gids;
 872
 873	roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 874	num_default_gids = hweight_long(roce_gid_type_mask);
 875	/* Reserve starting indices for default GIDs */
 876	for (i = 0; i < num_default_gids && i < table->sz; i++)
 877		table->default_gid_indices |= BIT(i);
 878}
 879
 880
 881static void gid_table_release_one(struct ib_device *ib_dev)
 882{
 883	u32 p;
 884
 885	rdma_for_each_port (ib_dev, p) {
 886		release_gid_table(ib_dev, ib_dev->port_data[p].cache.gid);
 887		ib_dev->port_data[p].cache.gid = NULL;
 888	}
 889}
 890
 891static int _gid_table_setup_one(struct ib_device *ib_dev)
 892{
 893	struct ib_gid_table *table;
 894	u32 rdma_port;
 895
 896	rdma_for_each_port (ib_dev, rdma_port) {
 897		table = alloc_gid_table(
 898			ib_dev->port_data[rdma_port].immutable.gid_tbl_len);
 899		if (!table)
 900			goto rollback_table_setup;
 901
 902		gid_table_reserve_default(ib_dev, rdma_port, table);
 903		ib_dev->port_data[rdma_port].cache.gid = table;
 904	}
 905	return 0;
 906
 907rollback_table_setup:
 908	gid_table_release_one(ib_dev);
 909	return -ENOMEM;
 910}
 911
 912static void gid_table_cleanup_one(struct ib_device *ib_dev)
 913{
 914	u32 p;
 915
 916	rdma_for_each_port (ib_dev, p)
 917		cleanup_gid_table_port(ib_dev, p,
 918				       ib_dev->port_data[p].cache.gid);
 919}
 920
 921static int gid_table_setup_one(struct ib_device *ib_dev)
 922{
 923	int err;
 924
 925	err = _gid_table_setup_one(ib_dev);
 926
 927	if (err)
 928		return err;
 929
 930	/*
 931	 * Mark the device as ready for GID cache updates. This allows netdev
 932	 * event handlers to update the GID cache even before the device is
 933	 * fully registered.
 934	 */
 935	ib_device_enable_gid_updates(ib_dev);
 936
 937	rdma_roce_rescan_device(ib_dev);
 938
 939	return err;
 940}
 941
 942/**
 943 * rdma_query_gid - Read the GID content from the GID software cache
 944 * @device:		Device to query the GID
 945 * @port_num:		Port number of the device
 946 * @index:		Index of the GID table entry to read
 947 * @gid:		Pointer to GID where to store the entry's GID
 948 *
 949 * rdma_query_gid() only reads the GID entry content for requested device,
 950 * port and index. It reads for IB, RoCE and iWarp link layers.  It doesn't
 951 * hold any reference to the GID table entry in the HCA or software cache.
 952 *
 953 * Returns 0 on success or appropriate error code.
 954 *
 955 */
 956int rdma_query_gid(struct ib_device *device, u32 port_num,
 957		   int index, union ib_gid *gid)
 958{
 959	struct ib_gid_table *table;
 960	unsigned long flags;
 961	int res;
 962
 963	if (!rdma_is_port_valid(device, port_num))
 964		return -EINVAL;
 965
 966	table = rdma_gid_table(device, port_num);
 967	read_lock_irqsave(&table->rwlock, flags);
 968
 969	if (index < 0 || index >= table->sz) {
 970		res = -EINVAL;
 971		goto done;
 972	}
 973
 974	if (!is_gid_entry_valid(table->data_vec[index])) {
 975		res = -ENOENT;
 976		goto done;
 977	}
 978
 979	memcpy(gid, &table->data_vec[index]->attr.gid, sizeof(*gid));
 980	res = 0;
 981
 982done:
 983	read_unlock_irqrestore(&table->rwlock, flags);
 984	return res;
 985}
 986EXPORT_SYMBOL(rdma_query_gid);
 987
 988/**
 989 * rdma_read_gid_hw_context - Read the HW GID context from GID attribute
 990 * @attr:		Potinter to the GID attribute
 991 *
 992 * rdma_read_gid_hw_context() reads the drivers GID HW context corresponding
 993 * to the SGID attr. Callers are required to already be holding the reference
 994 * to an existing GID entry.
 995 *
 996 * Returns the HW GID context
 997 *
 998 */
 999void *rdma_read_gid_hw_context(const struct ib_gid_attr *attr)
1000{
1001	return container_of(attr, struct ib_gid_table_entry, attr)->context;
1002}
1003EXPORT_SYMBOL(rdma_read_gid_hw_context);
1004
1005/**
1006 * rdma_find_gid - Returns SGID attributes if the matching GID is found.
1007 * @device: The device to query.
1008 * @gid: The GID value to search for.
1009 * @gid_type: The GID type to search for.
1010 * @ndev: In RoCE, the net device of the device. NULL means ignore.
1011 *
1012 * rdma_find_gid() searches for the specified GID value in the software cache.
1013 *
1014 * Returns GID attributes if a valid GID is found or returns ERR_PTR for the
1015 * error. The caller must invoke rdma_put_gid_attr() to release the reference.
1016 *
1017 */
1018const struct ib_gid_attr *rdma_find_gid(struct ib_device *device,
1019					const union ib_gid *gid,
1020					enum ib_gid_type gid_type,
1021					struct net_device *ndev)
1022{
1023	unsigned long mask = GID_ATTR_FIND_MASK_GID |
1024			     GID_ATTR_FIND_MASK_GID_TYPE;
1025	struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
1026	u32 p;
1027
1028	if (ndev)
1029		mask |= GID_ATTR_FIND_MASK_NETDEV;
1030
1031	rdma_for_each_port(device, p) {
1032		struct ib_gid_table *table;
1033		unsigned long flags;
1034		int index;
1035
1036		table = device->port_data[p].cache.gid;
1037		read_lock_irqsave(&table->rwlock, flags);
1038		index = find_gid(table, gid, &gid_attr_val, false, mask, NULL);
1039		if (index >= 0) {
1040			const struct ib_gid_attr *attr;
1041
1042			get_gid_entry(table->data_vec[index]);
1043			attr = &table->data_vec[index]->attr;
1044			read_unlock_irqrestore(&table->rwlock, flags);
1045			return attr;
1046		}
1047		read_unlock_irqrestore(&table->rwlock, flags);
1048	}
1049
1050	return ERR_PTR(-ENOENT);
1051}
1052EXPORT_SYMBOL(rdma_find_gid);
1053
1054int ib_get_cached_pkey(struct ib_device *device,
1055		       u32               port_num,
1056		       int               index,
1057		       u16              *pkey)
1058{
1059	struct ib_pkey_cache *cache;
1060	unsigned long flags;
1061	int ret = 0;
1062
1063	if (!rdma_is_port_valid(device, port_num))
1064		return -EINVAL;
1065
1066	read_lock_irqsave(&device->cache_lock, flags);
1067
1068	cache = device->port_data[port_num].cache.pkey;
1069
1070	if (!cache || index < 0 || index >= cache->table_len)
1071		ret = -EINVAL;
1072	else
1073		*pkey = cache->table[index];
1074
1075	read_unlock_irqrestore(&device->cache_lock, flags);
1076
1077	return ret;
1078}
1079EXPORT_SYMBOL(ib_get_cached_pkey);
1080
1081void ib_get_cached_subnet_prefix(struct ib_device *device, u32 port_num,
1082				u64 *sn_pfx)
1083{
1084	unsigned long flags;
1085
1086	read_lock_irqsave(&device->cache_lock, flags);
1087	*sn_pfx = device->port_data[port_num].cache.subnet_prefix;
1088	read_unlock_irqrestore(&device->cache_lock, flags);
1089}
1090EXPORT_SYMBOL(ib_get_cached_subnet_prefix);
1091
1092int ib_find_cached_pkey(struct ib_device *device, u32 port_num,
1093			u16 pkey, u16 *index)
1094{
1095	struct ib_pkey_cache *cache;
1096	unsigned long flags;
1097	int i;
1098	int ret = -ENOENT;
1099	int partial_ix = -1;
1100
1101	if (!rdma_is_port_valid(device, port_num))
1102		return -EINVAL;
1103
1104	read_lock_irqsave(&device->cache_lock, flags);
1105
1106	cache = device->port_data[port_num].cache.pkey;
1107	if (!cache) {
1108		ret = -EINVAL;
1109		goto err;
1110	}
1111
1112	*index = -1;
1113
1114	for (i = 0; i < cache->table_len; ++i)
1115		if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
1116			if (cache->table[i] & 0x8000) {
1117				*index = i;
1118				ret = 0;
1119				break;
1120			} else {
1121				partial_ix = i;
1122			}
1123		}
1124
1125	if (ret && partial_ix >= 0) {
1126		*index = partial_ix;
1127		ret = 0;
1128	}
1129
1130err:
1131	read_unlock_irqrestore(&device->cache_lock, flags);
1132
1133	return ret;
1134}
1135EXPORT_SYMBOL(ib_find_cached_pkey);
1136
1137int ib_get_cached_lmc(struct ib_device *device, u32 port_num, u8 *lmc)
1138{
1139	unsigned long flags;
1140	int ret = 0;
1141
1142	if (!rdma_is_port_valid(device, port_num))
1143		return -EINVAL;
1144
1145	read_lock_irqsave(&device->cache_lock, flags);
1146	*lmc = device->port_data[port_num].cache.lmc;
1147	read_unlock_irqrestore(&device->cache_lock, flags);
1148
1149	return ret;
1150}
1151EXPORT_SYMBOL(ib_get_cached_lmc);
1152
1153int ib_get_cached_port_state(struct ib_device *device, u32 port_num,
1154			     enum ib_port_state *port_state)
1155{
1156	unsigned long flags;
1157	int ret = 0;
1158
1159	if (!rdma_is_port_valid(device, port_num))
1160		return -EINVAL;
1161
1162	read_lock_irqsave(&device->cache_lock, flags);
1163	*port_state = device->port_data[port_num].cache.port_state;
1164	read_unlock_irqrestore(&device->cache_lock, flags);
1165
1166	return ret;
1167}
1168EXPORT_SYMBOL(ib_get_cached_port_state);
1169
1170/**
1171 * rdma_get_gid_attr - Returns GID attributes for a port of a device
1172 * at a requested gid_index, if a valid GID entry exists.
1173 * @device:		The device to query.
1174 * @port_num:		The port number on the device where the GID value
1175 *			is to be queried.
1176 * @index:		Index of the GID table entry whose attributes are to
1177 *                      be queried.
1178 *
1179 * rdma_get_gid_attr() acquires reference count of gid attributes from the
1180 * cached GID table. Caller must invoke rdma_put_gid_attr() to release
1181 * reference to gid attribute regardless of link layer.
1182 *
1183 * Returns pointer to valid gid attribute or ERR_PTR for the appropriate error
1184 * code.
1185 */
1186const struct ib_gid_attr *
1187rdma_get_gid_attr(struct ib_device *device, u32 port_num, int index)
1188{
1189	const struct ib_gid_attr *attr = ERR_PTR(-ENODATA);
1190	struct ib_gid_table *table;
1191	unsigned long flags;
1192
1193	if (!rdma_is_port_valid(device, port_num))
1194		return ERR_PTR(-EINVAL);
1195
1196	table = rdma_gid_table(device, port_num);
1197	if (index < 0 || index >= table->sz)
1198		return ERR_PTR(-EINVAL);
1199
1200	read_lock_irqsave(&table->rwlock, flags);
1201	if (!is_gid_entry_valid(table->data_vec[index]))
1202		goto done;
1203
1204	get_gid_entry(table->data_vec[index]);
1205	attr = &table->data_vec[index]->attr;
1206done:
1207	read_unlock_irqrestore(&table->rwlock, flags);
1208	return attr;
1209}
1210EXPORT_SYMBOL(rdma_get_gid_attr);
1211
1212/**
1213 * rdma_query_gid_table - Reads GID table entries of all the ports of a device up to max_entries.
1214 * @device: The device to query.
1215 * @entries: Entries where GID entries are returned.
1216 * @max_entries: Maximum number of entries that can be returned.
1217 * Entries array must be allocated to hold max_entries number of entries.
1218 *
1219 * Returns number of entries on success or appropriate error code.
1220 */
1221ssize_t rdma_query_gid_table(struct ib_device *device,
1222			     struct ib_uverbs_gid_entry *entries,
1223			     size_t max_entries)
1224{
1225	const struct ib_gid_attr *gid_attr;
1226	ssize_t num_entries = 0, ret;
1227	struct ib_gid_table *table;
1228	u32 port_num, i;
1229	struct net_device *ndev;
1230	unsigned long flags;
1231
1232	rdma_for_each_port(device, port_num) {
1233		table = rdma_gid_table(device, port_num);
1234		read_lock_irqsave(&table->rwlock, flags);
1235		for (i = 0; i < table->sz; i++) {
1236			if (!is_gid_entry_valid(table->data_vec[i]))
1237				continue;
1238			if (num_entries >= max_entries) {
1239				ret = -EINVAL;
1240				goto err;
1241			}
1242
1243			gid_attr = &table->data_vec[i]->attr;
1244
1245			memcpy(&entries->gid, &gid_attr->gid,
1246			       sizeof(gid_attr->gid));
1247			entries->gid_index = gid_attr->index;
1248			entries->port_num = gid_attr->port_num;
1249			entries->gid_type = gid_attr->gid_type;
1250			ndev = rcu_dereference_protected(
1251				gid_attr->ndev,
1252				lockdep_is_held(&table->rwlock));
1253			if (ndev)
1254				entries->netdev_ifindex = ndev->ifindex;
1255
1256			num_entries++;
1257			entries++;
1258		}
1259		read_unlock_irqrestore(&table->rwlock, flags);
1260	}
1261
1262	return num_entries;
1263err:
1264	read_unlock_irqrestore(&table->rwlock, flags);
1265	return ret;
1266}
1267EXPORT_SYMBOL(rdma_query_gid_table);
1268
1269/**
1270 * rdma_put_gid_attr - Release reference to the GID attribute
1271 * @attr:		Pointer to the GID attribute whose reference
1272 *			needs to be released.
1273 *
1274 * rdma_put_gid_attr() must be used to release reference whose
1275 * reference is acquired using rdma_get_gid_attr() or any APIs
1276 * which returns a pointer to the ib_gid_attr regardless of link layer
1277 * of IB or RoCE.
1278 *
1279 */
1280void rdma_put_gid_attr(const struct ib_gid_attr *attr)
1281{
1282	struct ib_gid_table_entry *entry =
1283		container_of(attr, struct ib_gid_table_entry, attr);
1284
1285	put_gid_entry(entry);
1286}
1287EXPORT_SYMBOL(rdma_put_gid_attr);
1288
1289/**
1290 * rdma_hold_gid_attr - Get reference to existing GID attribute
1291 *
1292 * @attr:		Pointer to the GID attribute whose reference
1293 *			needs to be taken.
1294 *
1295 * Increase the reference count to a GID attribute to keep it from being
1296 * freed. Callers are required to already be holding a reference to attribute.
1297 *
1298 */
1299void rdma_hold_gid_attr(const struct ib_gid_attr *attr)
1300{
1301	struct ib_gid_table_entry *entry =
1302		container_of(attr, struct ib_gid_table_entry, attr);
1303
1304	get_gid_entry(entry);
1305}
1306EXPORT_SYMBOL(rdma_hold_gid_attr);
1307
1308/**
1309 * rdma_read_gid_attr_ndev_rcu - Read GID attribute netdevice
1310 * which must be in UP state.
1311 *
1312 * @attr:Pointer to the GID attribute
1313 *
1314 * Returns pointer to netdevice if the netdevice was attached to GID and
1315 * netdevice is in UP state. Caller must hold RCU lock as this API
1316 * reads the netdev flags which can change while netdevice migrates to
1317 * different net namespace. Returns ERR_PTR with error code otherwise.
1318 *
1319 */
1320struct net_device *rdma_read_gid_attr_ndev_rcu(const struct ib_gid_attr *attr)
1321{
1322	struct ib_gid_table_entry *entry =
1323			container_of(attr, struct ib_gid_table_entry, attr);
1324	struct ib_device *device = entry->attr.device;
1325	struct net_device *ndev = ERR_PTR(-EINVAL);
1326	u32 port_num = entry->attr.port_num;
1327	struct ib_gid_table *table;
1328	unsigned long flags;
1329	bool valid;
1330
1331	table = rdma_gid_table(device, port_num);
1332
1333	read_lock_irqsave(&table->rwlock, flags);
1334	valid = is_gid_entry_valid(table->data_vec[attr->index]);
1335	if (valid) {
1336		ndev = rcu_dereference(attr->ndev);
1337		if (!ndev)
1338			ndev = ERR_PTR(-ENODEV);
1339	}
1340	read_unlock_irqrestore(&table->rwlock, flags);
1341	return ndev;
1342}
1343EXPORT_SYMBOL(rdma_read_gid_attr_ndev_rcu);
1344
1345static int get_lower_dev_vlan(struct net_device *lower_dev,
1346			      struct netdev_nested_priv *priv)
1347{
1348	u16 *vlan_id = (u16 *)priv->data;
1349
1350	if (is_vlan_dev(lower_dev))
1351		*vlan_id = vlan_dev_vlan_id(lower_dev);
1352
1353	/* We are interested only in first level vlan device, so
1354	 * always return 1 to stop iterating over next level devices.
1355	 */
1356	return 1;
1357}
1358
1359/**
1360 * rdma_read_gid_l2_fields - Read the vlan ID and source MAC address
1361 *			     of a GID entry.
1362 *
1363 * @attr:	GID attribute pointer whose L2 fields to be read
1364 * @vlan_id:	Pointer to vlan id to fill up if the GID entry has
1365 *		vlan id. It is optional.
1366 * @smac:	Pointer to smac to fill up for a GID entry. It is optional.
1367 *
1368 * rdma_read_gid_l2_fields() returns 0 on success and returns vlan id
1369 * (if gid entry has vlan) and source MAC, or returns error.
1370 */
1371int rdma_read_gid_l2_fields(const struct ib_gid_attr *attr,
1372			    u16 *vlan_id, u8 *smac)
1373{
1374	struct netdev_nested_priv priv = {
1375		.data = (void *)vlan_id,
1376	};
1377	struct net_device *ndev;
1378
1379	rcu_read_lock();
1380	ndev = rcu_dereference(attr->ndev);
1381	if (!ndev) {
1382		rcu_read_unlock();
1383		return -ENODEV;
1384	}
1385	if (smac)
1386		ether_addr_copy(smac, ndev->dev_addr);
1387	if (vlan_id) {
1388		*vlan_id = 0xffff;
1389		if (is_vlan_dev(ndev)) {
1390			*vlan_id = vlan_dev_vlan_id(ndev);
1391		} else {
1392			/* If the netdev is upper device and if it's lower
1393			 * device is vlan device, consider vlan id of
1394			 * the lower vlan device for this gid entry.
1395			 */
1396			netdev_walk_all_lower_dev_rcu(attr->ndev,
1397					get_lower_dev_vlan, &priv);
1398		}
1399	}
1400	rcu_read_unlock();
1401	return 0;
1402}
1403EXPORT_SYMBOL(rdma_read_gid_l2_fields);
1404
1405static int config_non_roce_gid_cache(struct ib_device *device,
1406				     u32 port, struct ib_port_attr *tprops)
1407{
1408	struct ib_gid_attr gid_attr = {};
1409	struct ib_gid_table *table;
1410	int ret = 0;
1411	int i;
1412
1413	gid_attr.device = device;
1414	gid_attr.port_num = port;
1415	table = rdma_gid_table(device, port);
1416
1417	mutex_lock(&table->lock);
1418	for (i = 0; i < tprops->gid_tbl_len; ++i) {
1419		if (!device->ops.query_gid)
1420			continue;
1421		ret = device->ops.query_gid(device, port, i, &gid_attr.gid);
1422		if (ret) {
1423			dev_warn(&device->dev,
1424				 "query_gid failed (%d) for index %d\n", ret,
1425				 i);
1426			goto err;
1427		}
1428
1429		if (rdma_protocol_iwarp(device, port)) {
1430			struct net_device *ndev;
1431
1432			ndev = ib_device_get_netdev(device, port);
1433			if (!ndev)
1434				continue;
1435			RCU_INIT_POINTER(gid_attr.ndev, ndev);
1436			dev_put(ndev);
1437		}
1438
1439		gid_attr.index = i;
1440		tprops->subnet_prefix =
1441			be64_to_cpu(gid_attr.gid.global.subnet_prefix);
1442		add_modify_gid(table, &gid_attr);
1443	}
1444err:
1445	mutex_unlock(&table->lock);
1446	return ret;
1447}
1448
1449static int
1450ib_cache_update(struct ib_device *device, u32 port, bool update_gids,
1451		bool update_pkeys, bool enforce_security)
1452{
1453	struct ib_port_attr       *tprops = NULL;
1454	struct ib_pkey_cache      *pkey_cache = NULL;
1455	struct ib_pkey_cache      *old_pkey_cache = NULL;
1456	int                        i;
1457	int                        ret;
1458
1459	if (!rdma_is_port_valid(device, port))
1460		return -EINVAL;
1461
1462	tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1463	if (!tprops)
1464		return -ENOMEM;
1465
1466	ret = ib_query_port(device, port, tprops);
1467	if (ret) {
1468		dev_warn(&device->dev, "ib_query_port failed (%d)\n", ret);
1469		goto err;
1470	}
1471
1472	if (!rdma_protocol_roce(device, port) && update_gids) {
1473		ret = config_non_roce_gid_cache(device, port,
1474						tprops);
1475		if (ret)
1476			goto err;
1477	}
1478
1479	update_pkeys &= !!tprops->pkey_tbl_len;
1480
1481	if (update_pkeys) {
1482		pkey_cache = kmalloc(struct_size(pkey_cache, table,
1483						 tprops->pkey_tbl_len),
1484				     GFP_KERNEL);
1485		if (!pkey_cache) {
1486			ret = -ENOMEM;
1487			goto err;
1488		}
1489
1490		pkey_cache->table_len = tprops->pkey_tbl_len;
1491
1492		for (i = 0; i < pkey_cache->table_len; ++i) {
1493			ret = ib_query_pkey(device, port, i,
1494					    pkey_cache->table + i);
1495			if (ret) {
1496				dev_warn(&device->dev,
1497					 "ib_query_pkey failed (%d) for index %d\n",
1498					 ret, i);
1499				goto err;
1500			}
1501		}
1502	}
1503
1504	write_lock_irq(&device->cache_lock);
1505
1506	if (update_pkeys) {
1507		old_pkey_cache = device->port_data[port].cache.pkey;
1508		device->port_data[port].cache.pkey = pkey_cache;
1509	}
1510	device->port_data[port].cache.lmc = tprops->lmc;
1511
1512	if (device->port_data[port].cache.port_state != IB_PORT_NOP &&
1513	    device->port_data[port].cache.port_state != tprops->state)
1514		ibdev_info(device, "Port: %d Link %s\n", port,
1515			   ib_port_state_to_str(tprops->state));
1516
1517	device->port_data[port].cache.port_state = tprops->state;
1518
1519	device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix;
1520	write_unlock_irq(&device->cache_lock);
1521
1522	if (enforce_security)
1523		ib_security_cache_change(device,
1524					 port,
1525					 tprops->subnet_prefix);
1526
1527	kfree(old_pkey_cache);
1528	kfree(tprops);
1529	return 0;
1530
1531err:
1532	kfree(pkey_cache);
1533	kfree(tprops);
1534	return ret;
1535}
1536
1537static void ib_cache_event_task(struct work_struct *_work)
1538{
1539	struct ib_update_work *work =
1540		container_of(_work, struct ib_update_work, work);
1541	int ret;
1542
1543	/* Before distributing the cache update event, first sync
1544	 * the cache.
1545	 */
1546	ret = ib_cache_update(work->event.device, work->event.element.port_num,
1547			      work->event.event == IB_EVENT_GID_CHANGE ||
1548			      work->event.event == IB_EVENT_CLIENT_REREGISTER,
1549			      work->event.event == IB_EVENT_PKEY_CHANGE,
1550			      work->enforce_security);
1551
1552	/* GID event is notified already for individual GID entries by
1553	 * dispatch_gid_change_event(). Hence, notifiy for rest of the
1554	 * events.
1555	 */
1556	if (!ret && work->event.event != IB_EVENT_GID_CHANGE)
1557		ib_dispatch_event_clients(&work->event);
1558
1559	kfree(work);
1560}
1561
1562static void ib_generic_event_task(struct work_struct *_work)
1563{
1564	struct ib_update_work *work =
1565		container_of(_work, struct ib_update_work, work);
1566
1567	ib_dispatch_event_clients(&work->event);
1568	kfree(work);
1569}
1570
1571static bool is_cache_update_event(const struct ib_event *event)
1572{
1573	return (event->event == IB_EVENT_PORT_ERR    ||
1574		event->event == IB_EVENT_PORT_ACTIVE ||
1575		event->event == IB_EVENT_LID_CHANGE  ||
1576		event->event == IB_EVENT_PKEY_CHANGE ||
1577		event->event == IB_EVENT_CLIENT_REREGISTER ||
1578		event->event == IB_EVENT_GID_CHANGE);
1579}
1580
1581/**
1582 * ib_dispatch_event - Dispatch an asynchronous event
1583 * @event:Event to dispatch
1584 *
1585 * Low-level drivers must call ib_dispatch_event() to dispatch the
1586 * event to all registered event handlers when an asynchronous event
1587 * occurs.
1588 */
1589void ib_dispatch_event(const struct ib_event *event)
1590{
1591	struct ib_update_work *work;
1592
1593	work = kzalloc(sizeof(*work), GFP_ATOMIC);
1594	if (!work)
1595		return;
1596
1597	if (is_cache_update_event(event))
1598		INIT_WORK(&work->work, ib_cache_event_task);
1599	else
1600		INIT_WORK(&work->work, ib_generic_event_task);
1601
1602	work->event = *event;
1603	if (event->event == IB_EVENT_PKEY_CHANGE ||
1604	    event->event == IB_EVENT_GID_CHANGE)
1605		work->enforce_security = true;
1606
1607	queue_work(ib_wq, &work->work);
1608}
1609EXPORT_SYMBOL(ib_dispatch_event);
1610
1611int ib_cache_setup_one(struct ib_device *device)
1612{
1613	u32 p;
1614	int err;
1615
1616	err = gid_table_setup_one(device);
1617	if (err)
1618		return err;
1619
1620	rdma_for_each_port (device, p) {
1621		err = ib_cache_update(device, p, true, true, true);
1622		if (err) {
1623			gid_table_cleanup_one(device);
1624			return err;
1625		}
1626	}
1627
1628	return 0;
1629}
1630
1631void ib_cache_release_one(struct ib_device *device)
1632{
1633	u32 p;
1634
1635	/*
1636	 * The release function frees all the cache elements.
1637	 * This function should be called as part of freeing
1638	 * all the device's resources when the cache could no
1639	 * longer be accessed.
1640	 */
1641	rdma_for_each_port (device, p)
1642		kfree(device->port_data[p].cache.pkey);
1643
1644	gid_table_release_one(device);
1645}
1646
1647void ib_cache_cleanup_one(struct ib_device *device)
1648{
1649	/*
1650	 * Clear the GID updates mark first to prevent event handlers from
1651	 * accessing the device while it's being torn down.
1652	 */
1653	ib_device_disable_gid_updates(device);
1654
1655	/* The cleanup function waits for all in-progress workqueue
1656	 * elements and cleans up the GID cache. This function should be
1657	 * called after the device was removed from the devices list and
1658	 * all clients were removed, so the cache exists but is
1659	 * non-functional and shouldn't be updated anymore.
1660	 */
1661	flush_workqueue(ib_wq);
1662	gid_table_cleanup_one(device);
1663
1664	/*
1665	 * Flush the wq second time for any pending GID delete work.
1666	 */
1667	flush_workqueue(ib_wq);
1668}