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