drivers/infiniband/core/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/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
  47 struct ib_pkey_cache {
  48         int             table_len;
  49         u16             table[0];
  50 };
  51
  52 struct ib_update_work {
  53         struct work_struct work;
  54         struct ib_device  *device;
  55         u8                 port_num;
  56 };
  57
  58 union ib_gid zgid;
  59 EXPORT_SYMBOL(zgid);
  60
  61 static const struct ib_gid_attr zattr;
  62
  63 enum gid_attr_find_mask {
  64         GID_ATTR_FIND_MASK_GID          = 1UL << 0,
  65         GID_ATTR_FIND_MASK_NETDEV       = 1UL << 1,
  66         GID_ATTR_FIND_MASK_DEFAULT      = 1UL << 2,
  67         GID_ATTR_FIND_MASK_GID_TYPE     = 1UL << 3,
  68 };
  69
  70 enum gid_table_entry_props {
  71         GID_TABLE_ENTRY_INVALID         = 1UL << 0,
  72         GID_TABLE_ENTRY_DEFAULT         = 1UL << 1,
  73 };
  74
  75 enum gid_table_write_action {
  76         GID_TABLE_WRITE_ACTION_ADD,
  77         GID_TABLE_WRITE_ACTION_DEL,
  78         /* MODIFY only updates the GID table. Currently only used by
  79          * ib_cache_update.
  80          */
  81         GID_TABLE_WRITE_ACTION_MODIFY
  82 };
  83
  84 struct ib_gid_table_entry {
  85         unsigned long       props;
  86         union ib_gid        gid;
  87         struct ib_gid_attr  attr;
  88         void               *context;
  89 };
  90
  91 struct ib_gid_table {
  92         int                  sz;
  93         /* In RoCE, adding a GID to the table requires:
  94          * (a) Find if this GID is already exists.
  95          * (b) Find a free space.
  96          * (c) Write the new GID
  97          *
  98          * Delete requires different set of operations:
  99          * (a) Find the GID
 100          * (b) Delete it.
 101          *
 102          * Add/delete should be carried out atomically.
 103          * This is done by locking this mutex from multiple
 104          * writers. We don't need this lock for IB, as the MAD
 105          * layer replaces all entries. All data_vec entries
 106          * are locked by this lock.
 107          **/
 108         struct mutex         lock;
 109         /* This lock protects the table entries from being
 110          * read and written simultaneously.
 111          */
 112         rwlock_t             rwlock;
 113         struct ib_gid_table_entry *data_vec;
 114 };
 115
 116 static void dispatch_gid_change_event(struct ib_device *ib_dev, u8 port)
 117 {
 118         if (rdma_cap_roce_gid_table(ib_dev, port)) {
 119                 struct ib_event event;
 120
 121                 event.device            = ib_dev;
 122                 event.element.port_num  = port;
 123                 event.event             = IB_EVENT_GID_CHANGE;
 124
 125                 ib_dispatch_event(&event);
 126         }
 127 }
 128
 129 static const char * const gid_type_str[] = {
 130         [IB_GID_TYPE_IB]        = "IB/RoCE v1",
 131         [IB_GID_TYPE_ROCE_UDP_ENCAP]    = "RoCE v2",
 132 };
 133
 134 const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
 135 {
 136         if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
 137                 return gid_type_str[gid_type];
 138
 139         return "Invalid GID type";
 140 }
 141 EXPORT_SYMBOL(ib_cache_gid_type_str);
 142
 143 int ib_cache_gid_parse_type_str(const char *buf)
 144 {
 145         unsigned int i;
 146         size_t len;
 147         int err = -EINVAL;
 148
 149         len = strlen(buf);
 150         if (len == 0)
 151                 return -EINVAL;
 152
 153         if (buf[len - 1] == '\n')
 154                 len--;
 155
 156         for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i)
 157                 if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) &&
 158                     len == strlen(gid_type_str[i])) {
 159                         err = i;
 160                         break;
 161                 }
 162
 163         return err;
 164 }
 165 EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
 166
 167 /* This function expects that rwlock will be write locked in all
 168  * scenarios and that lock will be locked in sleep-able (RoCE)
 169  * scenarios.
 170  */
 171 static int write_gid(struct ib_device *ib_dev, u8 port,
 172                      struct ib_gid_table *table, int ix,
 173                      const union ib_gid *gid,
 174                      const struct ib_gid_attr *attr,
 175                      enum gid_table_write_action action,
 176                      bool  default_gid)
 177         __releases(&table->rwlock) __acquires(&table->rwlock)
 178 {
 179         int ret = 0;
 180         struct net_device *old_net_dev;
 181         enum ib_gid_type old_gid_type;
 182
 183         /* in rdma_cap_roce_gid_table, this funciton should be protected by a
 184          * sleep-able lock.
 185          */
 186
 187         if (rdma_cap_roce_gid_table(ib_dev, port)) {
 188                 table->data_vec[ix].props |= GID_TABLE_ENTRY_INVALID;
 189                 write_unlock_irq(&table->rwlock);
 190                 /* GID_TABLE_WRITE_ACTION_MODIFY currently isn't supported by
 191                  * RoCE providers and thus only updates the cache.
 192                  */
 193                 if (action == GID_TABLE_WRITE_ACTION_ADD)
 194                         ret = ib_dev->add_gid(ib_dev, port, ix, gid, attr,
 195                                               &table->data_vec[ix].context);
 196                 else if (action == GID_TABLE_WRITE_ACTION_DEL)
 197                         ret = ib_dev->del_gid(ib_dev, port, ix,
 198                                               &table->data_vec[ix].context);
 199                 write_lock_irq(&table->rwlock);
 200         }
 201
 202         old_net_dev = table->data_vec[ix].attr.ndev;
 203         old_gid_type = table->data_vec[ix].attr.gid_type;
 204         if (old_net_dev && old_net_dev != attr->ndev)
 205                 dev_put(old_net_dev);
 206         /* if modify_gid failed, just delete the old gid */
 207         if (ret || action == GID_TABLE_WRITE_ACTION_DEL) {
 208                 gid = &zgid;
 209                 attr = &zattr;
 210                 table->data_vec[ix].context = NULL;
 211         }
 212
 213         memcpy(&table->data_vec[ix].gid, gid, sizeof(*gid));
 214         memcpy(&table->data_vec[ix].attr, attr, sizeof(*attr));
 215         if (default_gid) {
 216                 table->data_vec[ix].props |= GID_TABLE_ENTRY_DEFAULT;
 217                 if (action == GID_TABLE_WRITE_ACTION_DEL)
 218                         table->data_vec[ix].attr.gid_type = old_gid_type;
 219         }
 220         if (table->data_vec[ix].attr.ndev &&
 221             table->data_vec[ix].attr.ndev != old_net_dev)
 222                 dev_hold(table->data_vec[ix].attr.ndev);
 223
 224         table->data_vec[ix].props &= ~GID_TABLE_ENTRY_INVALID;
 225
 226         return ret;
 227 }
 228
 229 static int add_gid(struct ib_device *ib_dev, u8 port,
 230                    struct ib_gid_table *table, int ix,
 231                    const union ib_gid *gid,
 232                    const struct ib_gid_attr *attr,
 233                    bool  default_gid) {
 234         return write_gid(ib_dev, port, table, ix, gid, attr,
 235                          GID_TABLE_WRITE_ACTION_ADD, default_gid);
 236 }
 237
 238 static int modify_gid(struct ib_device *ib_dev, u8 port,
 239                       struct ib_gid_table *table, int ix,
 240                       const union ib_gid *gid,
 241                       const struct ib_gid_attr *attr,
 242                       bool  default_gid) {
 243         return write_gid(ib_dev, port, table, ix, gid, attr,
 244                          GID_TABLE_WRITE_ACTION_MODIFY, default_gid);
 245 }
 246
 247 static int del_gid(struct ib_device *ib_dev, u8 port,
 248                    struct ib_gid_table *table, int ix,
 249                    bool  default_gid) {
 250         return write_gid(ib_dev, port, table, ix, &zgid, &zattr,
 251                          GID_TABLE_WRITE_ACTION_DEL, default_gid);
 252 }
 253
 254 /* rwlock should be read locked */
 255 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
 256                     const struct ib_gid_attr *val, bool default_gid,
 257                     unsigned long mask, int *pempty)
 258 {
 259         int i = 0;
 260         int found = -1;
 261         int empty = pempty ? -1 : 0;
 262
 263         while (i < table->sz && (found < 0 || empty < 0)) {
 264                 struct ib_gid_table_entry *data = &table->data_vec[i];
 265                 struct ib_gid_attr *attr = &data->attr;
 266                 int curr_index = i;
 267
 268                 i++;
 269
 270                 if (data->props & GID_TABLE_ENTRY_INVALID)
 271                         continue;
 272
 273                 if (empty < 0)
 274                         if (!memcmp(&data->gid, &zgid, sizeof(*gid)) &&
 275                             !memcmp(attr, &zattr, sizeof(*attr)) &&
 276                             !data->props)
 277                                 empty = curr_index;
 278
 279                 if (found >= 0)
 280                         continue;
 281
 282                 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
 283                     attr->gid_type != val->gid_type)
 284                         continue;
 285
 286                 if (mask & GID_ATTR_FIND_MASK_GID &&
 287                     memcmp(gid, &data->gid, sizeof(*gid)))
 288                         continue;
 289
 290                 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
 291                     attr->ndev != val->ndev)
 292                         continue;
 293
 294                 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
 295                     !!(data->props & GID_TABLE_ENTRY_DEFAULT) !=
 296                     default_gid)
 297                         continue;
 298
 299                 found = curr_index;
 300         }
 301
 302         if (pempty)
 303                 *pempty = empty;
 304
 305         return found;
 306 }
 307
 308 static void make_default_gid(struct  net_device *dev, union ib_gid *gid)
 309 {
 310         gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
 311         addrconf_ifid_eui48(&gid->raw[8], dev);
 312 }
 313
 314 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
 315                      union ib_gid *gid, struct ib_gid_attr *attr)
 316 {
 317         struct ib_gid_table *table;
 318         int ix;
 319         int ret = 0;
 320         struct net_device *idev;
 321         int empty;
 322
 323         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 324
 325         if (!memcmp(gid, &zgid, sizeof(*gid)))
 326                 return -EINVAL;
 327
 328         if (ib_dev->get_netdev) {
 329                 idev = ib_dev->get_netdev(ib_dev, port);
 330                 if (idev && attr->ndev != idev) {
 331                         union ib_gid default_gid;
 332
 333                         /* Adding default GIDs in not permitted */
 334                         make_default_gid(idev, &default_gid);
 335                         if (!memcmp(gid, &default_gid, sizeof(*gid))) {
 336                                 dev_put(idev);
 337                                 return -EPERM;
 338                         }
 339                 }
 340                 if (idev)
 341                         dev_put(idev);
 342         }
 343
 344         mutex_lock(&table->lock);
 345         write_lock_irq(&table->rwlock);
 346
 347         ix = find_gid(table, gid, attr, false, GID_ATTR_FIND_MASK_GID |
 348                       GID_ATTR_FIND_MASK_GID_TYPE |
 349                       GID_ATTR_FIND_MASK_NETDEV, &empty);
 350         if (ix >= 0)
 351                 goto out_unlock;
 352
 353         if (empty < 0) {
 354                 ret = -ENOSPC;
 355                 goto out_unlock;
 356         }
 357
 358         ret = add_gid(ib_dev, port, table, empty, gid, attr, false);
 359         if (!ret)
 360                 dispatch_gid_change_event(ib_dev, port);
 361
 362 out_unlock:
 363         write_unlock_irq(&table->rwlock);
 364         mutex_unlock(&table->lock);
 365         return ret;
 366 }
 367
 368 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
 369                      union ib_gid *gid, struct ib_gid_attr *attr)
 370 {
 371         struct ib_gid_table *table;
 372         int ix;
 373
 374         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 375
 376         mutex_lock(&table->lock);
 377         write_lock_irq(&table->rwlock);
 378
 379         ix = find_gid(table, gid, attr, false,
 380                       GID_ATTR_FIND_MASK_GID      |
 381                       GID_ATTR_FIND_MASK_GID_TYPE |
 382                       GID_ATTR_FIND_MASK_NETDEV   |
 383                       GID_ATTR_FIND_MASK_DEFAULT,
 384                       NULL);
 385         if (ix < 0)
 386                 goto out_unlock;
 387
 388         if (!del_gid(ib_dev, port, table, ix, false))
 389                 dispatch_gid_change_event(ib_dev, port);
 390
 391 out_unlock:
 392         write_unlock_irq(&table->rwlock);
 393         mutex_unlock(&table->lock);
 394         return 0;
 395 }
 396
 397 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
 398                                      struct net_device *ndev)
 399 {
 400         struct ib_gid_table *table;
 401         int ix;
 402         bool deleted = false;
 403
 404         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 405
 406         mutex_lock(&table->lock);
 407         write_lock_irq(&table->rwlock);
 408
 409         for (ix = 0; ix < table->sz; ix++)
 410                 if (table->data_vec[ix].attr.ndev == ndev)
 411                         if (!del_gid(ib_dev, port, table, ix,
 412                                      !!(table->data_vec[ix].props &
 413                                         GID_TABLE_ENTRY_DEFAULT)))
 414                                 deleted = true;
 415
 416         write_unlock_irq(&table->rwlock);
 417         mutex_unlock(&table->lock);
 418
 419         if (deleted)
 420                 dispatch_gid_change_event(ib_dev, port);
 421
 422         return 0;
 423 }
 424
 425 static int __ib_cache_gid_get(struct ib_device *ib_dev, u8 port, int index,
 426                               union ib_gid *gid, struct ib_gid_attr *attr)
 427 {
 428         struct ib_gid_table *table;
 429
 430         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 431
 432         if (index < 0 || index >= table->sz)
 433                 return -EINVAL;
 434
 435         if (table->data_vec[index].props & GID_TABLE_ENTRY_INVALID)
 436                 return -EAGAIN;
 437
 438         memcpy(gid, &table->data_vec[index].gid, sizeof(*gid));
 439         if (attr) {
 440                 memcpy(attr, &table->data_vec[index].attr, sizeof(*attr));
 441                 if (attr->ndev)
 442                         dev_hold(attr->ndev);
 443         }
 444
 445         return 0;
 446 }
 447
 448 static int _ib_cache_gid_table_find(struct ib_device *ib_dev,
 449                                     const union ib_gid *gid,
 450                                     const struct ib_gid_attr *val,
 451                                     unsigned long mask,
 452                                     u8 *port, u16 *index)
 453 {
 454         struct ib_gid_table *table;
 455         u8 p;
 456         int local_index;
 457         unsigned long flags;
 458
 459         for (p = 0; p < ib_dev->phys_port_cnt; p++) {
 460                 table = ib_dev->cache.ports[p].gid;
 461                 read_lock_irqsave(&table->rwlock, flags);
 462                 local_index = find_gid(table, gid, val, false, mask, NULL);
 463                 if (local_index >= 0) {
 464                         if (index)
 465                                 *index = local_index;
 466                         if (port)
 467                                 *port = p + rdma_start_port(ib_dev);
 468                         read_unlock_irqrestore(&table->rwlock, flags);
 469                         return 0;
 470                 }
 471                 read_unlock_irqrestore(&table->rwlock, flags);
 472         }
 473
 474         return -ENOENT;
 475 }
 476
 477 static int ib_cache_gid_find(struct ib_device *ib_dev,
 478                              const union ib_gid *gid,
 479                              enum ib_gid_type gid_type,
 480                              struct net_device *ndev, u8 *port,
 481                              u16 *index)
 482 {
 483         unsigned long mask = GID_ATTR_FIND_MASK_GID |
 484                              GID_ATTR_FIND_MASK_GID_TYPE;
 485         struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
 486
 487         if (ndev)
 488                 mask |= GID_ATTR_FIND_MASK_NETDEV;
 489
 490         return _ib_cache_gid_table_find(ib_dev, gid, &gid_attr_val,
 491                                         mask, port, index);
 492 }
 493
 494 int ib_find_cached_gid_by_port(struct ib_device *ib_dev,
 495                                const union ib_gid *gid,
 496                                enum ib_gid_type gid_type,
 497                                u8 port, struct net_device *ndev,
 498                                u16 *index)
 499 {
 500         int local_index;
 501         struct ib_gid_table *table;
 502         unsigned long mask = GID_ATTR_FIND_MASK_GID |
 503                              GID_ATTR_FIND_MASK_GID_TYPE;
 504         struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
 505         unsigned long flags;
 506
 507         if (port < rdma_start_port(ib_dev) ||
 508             port > rdma_end_port(ib_dev))
 509                 return -ENOENT;
 510
 511         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 512
 513         if (ndev)
 514                 mask |= GID_ATTR_FIND_MASK_NETDEV;
 515
 516         read_lock_irqsave(&table->rwlock, flags);
 517         local_index = find_gid(table, gid, &val, false, mask, NULL);
 518         if (local_index >= 0) {
 519                 if (index)
 520                         *index = local_index;
 521                 read_unlock_irqrestore(&table->rwlock, flags);
 522                 return 0;
 523         }
 524
 525         read_unlock_irqrestore(&table->rwlock, flags);
 526         return -ENOENT;
 527 }
 528 EXPORT_SYMBOL(ib_find_cached_gid_by_port);
 529
 530 /**
 531  * ib_find_gid_by_filter - Returns the GID table index where a specified
 532  * GID value occurs
 533  * @device: The device to query.
 534  * @gid: The GID value to search for.
 535  * @port_num: The port number of the device where the GID value could be
 536  *   searched.
 537  * @filter: The filter function is executed on any matching GID in the table.
 538  *   If the filter function returns true, the corresponding index is returned,
 539  *   otherwise, we continue searching the GID table. It's guaranteed that
 540  *   while filter is executed, ndev field is valid and the structure won't
 541  *   change. filter is executed in an atomic context. filter must not be NULL.
 542  * @index: The index into the cached GID table where the GID was found.  This
 543  *   parameter may be NULL.
 544  *
 545  * ib_cache_gid_find_by_filter() searches for the specified GID value
 546  * of which the filter function returns true in the port's GID table.
 547  * This function is only supported on RoCE ports.
 548  *
 549  */
 550 static int ib_cache_gid_find_by_filter(struct ib_device *ib_dev,
 551                                        const union ib_gid *gid,
 552                                        u8 port,
 553                                        bool (*filter)(const union ib_gid *,
 554                                                       const struct ib_gid_attr *,
 555                                                       void *),
 556                                        void *context,
 557                                        u16 *index)
 558 {
 559         struct ib_gid_table *table;
 560         unsigned int i;
 561         unsigned long flags;
 562         bool found = false;
 563
 564
 565         if (port < rdma_start_port(ib_dev) ||
 566             port > rdma_end_port(ib_dev) ||
 567             !rdma_protocol_roce(ib_dev, port))
 568                 return -EPROTONOSUPPORT;
 569
 570         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 571
 572         read_lock_irqsave(&table->rwlock, flags);
 573         for (i = 0; i < table->sz; i++) {
 574                 struct ib_gid_attr attr;
 575
 576                 if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID)
 577                         goto next;
 578
 579                 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
 580                         goto next;
 581
 582                 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr));
 583
 584                 if (filter(gid, &attr, context))
 585                         found = true;
 586
 587 next:
 588                 if (found)
 589                         break;
 590         }
 591         read_unlock_irqrestore(&table->rwlock, flags);
 592
 593         if (!found)
 594                 return -ENOENT;
 595
 596         if (index)
 597                 *index = i;
 598         return 0;
 599 }
 600
 601 static struct ib_gid_table *alloc_gid_table(int sz)
 602 {
 603         struct ib_gid_table *table =
 604                 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
 605
 606         if (!table)
 607                 return NULL;
 608
 609         table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
 610         if (!table->data_vec)
 611                 goto err_free_table;
 612
 613         mutex_init(&table->lock);
 614
 615         table->sz = sz;
 616         rwlock_init(&table->rwlock);
 617
 618         return table;
 619
 620 err_free_table:
 621         kfree(table);
 622         return NULL;
 623 }
 624
 625 static void release_gid_table(struct ib_gid_table *table)
 626 {
 627         if (table) {
 628                 kfree(table->data_vec);
 629                 kfree(table);
 630         }
 631 }
 632
 633 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
 634                                    struct ib_gid_table *table)
 635 {
 636         int i;
 637         bool deleted = false;
 638
 639         if (!table)
 640                 return;
 641
 642         write_lock_irq(&table->rwlock);
 643         for (i = 0; i < table->sz; ++i) {
 644                 if (memcmp(&table->data_vec[i].gid, &zgid,
 645                            sizeof(table->data_vec[i].gid)))
 646                         if (!del_gid(ib_dev, port, table, i,
 647                                      table->data_vec[i].props &
 648                                      GID_ATTR_FIND_MASK_DEFAULT))
 649                                 deleted = true;
 650         }
 651         write_unlock_irq(&table->rwlock);
 652
 653         if (deleted)
 654                 dispatch_gid_change_event(ib_dev, port);
 655 }
 656
 657 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
 658                                   struct net_device *ndev,
 659                                   unsigned long gid_type_mask,
 660                                   enum ib_cache_gid_default_mode mode)
 661 {
 662         union ib_gid gid;
 663         struct ib_gid_attr gid_attr;
 664         struct ib_gid_attr zattr_type = zattr;
 665         struct ib_gid_table *table;
 666         unsigned int gid_type;
 667
 668         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 669
 670         make_default_gid(ndev, &gid);
 671         memset(&gid_attr, 0, sizeof(gid_attr));
 672         gid_attr.ndev = ndev;
 673
 674         for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
 675                 int ix;
 676                 union ib_gid current_gid;
 677                 struct ib_gid_attr current_gid_attr = {};
 678
 679                 if (1UL << gid_type & ~gid_type_mask)
 680                         continue;
 681
 682                 gid_attr.gid_type = gid_type;
 683
 684                 mutex_lock(&table->lock);
 685                 write_lock_irq(&table->rwlock);
 686                 ix = find_gid(table, NULL, &gid_attr, true,
 687                               GID_ATTR_FIND_MASK_GID_TYPE |
 688                               GID_ATTR_FIND_MASK_DEFAULT,
 689                               NULL);
 690
 691                 /* Coudn't find default GID location */
 692                 if (WARN_ON(ix < 0))
 693                         goto release;
 694
 695                 zattr_type.gid_type = gid_type;
 696
 697                 if (!__ib_cache_gid_get(ib_dev, port, ix,
 698                                         &current_gid, &current_gid_attr) &&
 699                     mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
 700                     !memcmp(&gid, &current_gid, sizeof(gid)) &&
 701                     !memcmp(&gid_attr, &current_gid_attr, sizeof(gid_attr)))
 702                         goto release;
 703
 704                 if (memcmp(&current_gid, &zgid, sizeof(current_gid)) ||
 705                     memcmp(&current_gid_attr, &zattr_type,
 706                            sizeof(current_gid_attr))) {
 707                         if (del_gid(ib_dev, port, table, ix, true)) {
 708                                 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
 709                                         ix, gid.raw);
 710                                 goto release;
 711                         } else {
 712                                 dispatch_gid_change_event(ib_dev, port);
 713                         }
 714                 }
 715
 716                 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
 717                         if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
 718                                 pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
 719                                         gid.raw);
 720                         else
 721                                 dispatch_gid_change_event(ib_dev, port);
 722                 }
 723
 724 release:
 725                 if (current_gid_attr.ndev)
 726                         dev_put(current_gid_attr.ndev);
 727                 write_unlock_irq(&table->rwlock);
 728                 mutex_unlock(&table->lock);
 729         }
 730 }
 731
 732 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
 733                                      struct ib_gid_table *table)
 734 {
 735         unsigned int i;
 736         unsigned long roce_gid_type_mask;
 737         unsigned int num_default_gids;
 738         unsigned int current_gid = 0;
 739
 740         roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 741         num_default_gids = hweight_long(roce_gid_type_mask);
 742         for (i = 0; i < num_default_gids && i < table->sz; i++) {
 743                 struct ib_gid_table_entry *entry =
 744                         &table->data_vec[i];
 745
 746                 entry->props |= GID_TABLE_ENTRY_DEFAULT;
 747                 current_gid = find_next_bit(&roce_gid_type_mask,
 748                                             BITS_PER_LONG,
 749                                             current_gid);
 750                 entry->attr.gid_type = current_gid++;
 751         }
 752
 753         return 0;
 754 }
 755
 756 static int _gid_table_setup_one(struct ib_device *ib_dev)
 757 {
 758         u8 port;
 759         struct ib_gid_table *table;
 760         int err = 0;
 761
 762         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 763                 u8 rdma_port = port + rdma_start_port(ib_dev);
 764
 765                 table =
 766                         alloc_gid_table(
 767                                 ib_dev->port_immutable[rdma_port].gid_tbl_len);
 768                 if (!table) {
 769                         err = -ENOMEM;
 770                         goto rollback_table_setup;
 771                 }
 772
 773                 err = gid_table_reserve_default(ib_dev,
 774                                                 port + rdma_start_port(ib_dev),
 775                                                 table);
 776                 if (err)
 777                         goto rollback_table_setup;
 778                 ib_dev->cache.ports[port].gid = table;
 779         }
 780
 781         return 0;
 782
 783 rollback_table_setup:
 784         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 785                 table = ib_dev->cache.ports[port].gid;
 786
 787                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 788                                        table);
 789                 release_gid_table(table);
 790         }
 791
 792         return err;
 793 }
 794
 795 static void gid_table_release_one(struct ib_device *ib_dev)
 796 {
 797         struct ib_gid_table *table;
 798         u8 port;
 799
 800         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 801                 table = ib_dev->cache.ports[port].gid;
 802                 release_gid_table(table);
 803                 ib_dev->cache.ports[port].gid = NULL;
 804         }
 805 }
 806
 807 static void gid_table_cleanup_one(struct ib_device *ib_dev)
 808 {
 809         struct ib_gid_table *table;
 810         u8 port;
 811
 812         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 813                 table = ib_dev->cache.ports[port].gid;
 814                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 815                                        table);
 816         }
 817 }
 818
 819 static int gid_table_setup_one(struct ib_device *ib_dev)
 820 {
 821         int err;
 822
 823         err = _gid_table_setup_one(ib_dev);
 824
 825         if (err)
 826                 return err;
 827
 828         err = roce_rescan_device(ib_dev);
 829
 830         if (err) {
 831                 gid_table_cleanup_one(ib_dev);
 832                 gid_table_release_one(ib_dev);
 833         }
 834
 835         return err;
 836 }
 837
 838 int ib_get_cached_gid(struct ib_device *device,
 839                       u8                port_num,
 840                       int               index,
 841                       union ib_gid     *gid,
 842                       struct ib_gid_attr *gid_attr)
 843 {
 844         int res;
 845         unsigned long flags;
 846         struct ib_gid_table *table;
 847
 848         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
 849                 return -EINVAL;
 850
 851         table = device->cache.ports[port_num - rdma_start_port(device)].gid;
 852         read_lock_irqsave(&table->rwlock, flags);
 853         res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr);
 854         read_unlock_irqrestore(&table->rwlock, flags);
 855
 856         return res;
 857 }
 858 EXPORT_SYMBOL(ib_get_cached_gid);
 859
 860 int ib_find_cached_gid(struct ib_device *device,
 861                        const union ib_gid *gid,
 862                        enum ib_gid_type gid_type,
 863                        struct net_device *ndev,
 864                        u8               *port_num,
 865                        u16              *index)
 866 {
 867         return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index);
 868 }
 869 EXPORT_SYMBOL(ib_find_cached_gid);
 870
 871 int ib_find_gid_by_filter(struct ib_device *device,
 872                           const union ib_gid *gid,
 873                           u8 port_num,
 874                           bool (*filter)(const union ib_gid *gid,
 875                                          const struct ib_gid_attr *,
 876                                          void *),
 877                           void *context, u16 *index)
 878 {
 879         /* Only RoCE GID table supports filter function */
 880         if (!rdma_cap_roce_gid_table(device, port_num) && filter)
 881                 return -EPROTONOSUPPORT;
 882
 883         return ib_cache_gid_find_by_filter(device, gid,
 884                                            port_num, filter,
 885                                            context, index);
 886 }
 887 EXPORT_SYMBOL(ib_find_gid_by_filter);
 888
 889 int ib_get_cached_pkey(struct ib_device *device,
 890                        u8                port_num,
 891                        int               index,
 892                        u16              *pkey)
 893 {
 894         struct ib_pkey_cache *cache;
 895         unsigned long flags;
 896         int ret = 0;
 897
 898         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
 899                 return -EINVAL;
 900
 901         read_lock_irqsave(&device->cache.lock, flags);
 902
 903         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 904
 905         if (index < 0 || index >= cache->table_len)
 906                 ret = -EINVAL;
 907         else
 908                 *pkey = cache->table[index];
 909
 910         read_unlock_irqrestore(&device->cache.lock, flags);
 911
 912         return ret;
 913 }
 914 EXPORT_SYMBOL(ib_get_cached_pkey);
 915
 916 int ib_find_cached_pkey(struct ib_device *device,
 917                         u8                port_num,
 918                         u16               pkey,
 919                         u16              *index)
 920 {
 921         struct ib_pkey_cache *cache;
 922         unsigned long flags;
 923         int i;
 924         int ret = -ENOENT;
 925         int partial_ix = -1;
 926
 927         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
 928                 return -EINVAL;
 929
 930         read_lock_irqsave(&device->cache.lock, flags);
 931
 932         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 933
 934         *index = -1;
 935
 936         for (i = 0; i < cache->table_len; ++i)
 937                 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
 938                         if (cache->table[i] & 0x8000) {
 939                                 *index = i;
 940                                 ret = 0;
 941                                 break;
 942                         } else
 943                                 partial_ix = i;
 944                 }
 945
 946         if (ret && partial_ix >= 0) {
 947                 *index = partial_ix;
 948                 ret = 0;
 949         }
 950
 951         read_unlock_irqrestore(&device->cache.lock, flags);
 952
 953         return ret;
 954 }
 955 EXPORT_SYMBOL(ib_find_cached_pkey);
 956
 957 int ib_find_exact_cached_pkey(struct ib_device *device,
 958                               u8                port_num,
 959                               u16               pkey,
 960                               u16              *index)
 961 {
 962         struct ib_pkey_cache *cache;
 963         unsigned long flags;
 964         int i;
 965         int ret = -ENOENT;
 966
 967         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
 968                 return -EINVAL;
 969
 970         read_lock_irqsave(&device->cache.lock, flags);
 971
 972         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 973
 974         *index = -1;
 975
 976         for (i = 0; i < cache->table_len; ++i)
 977                 if (cache->table[i] == pkey) {
 978                         *index = i;
 979                         ret = 0;
 980                         break;
 981                 }
 982
 983         read_unlock_irqrestore(&device->cache.lock, flags);
 984
 985         return ret;
 986 }
 987 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
 988
 989 int ib_get_cached_lmc(struct ib_device *device,
 990                       u8                port_num,
 991                       u8                *lmc)
 992 {
 993         unsigned long flags;
 994         int ret = 0;
 995
 996         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
 997                 return -EINVAL;
 998
 999         read_lock_irqsave(&device->cache.lock, flags);
1000         *lmc = device->cache.ports[port_num - rdma_start_port(device)].lmc;
1001         read_unlock_irqrestore(&device->cache.lock, flags);
1002
1003         return ret;
1004 }
1005 EXPORT_SYMBOL(ib_get_cached_lmc);
1006
1007 int ib_get_cached_port_state(struct ib_device   *device,
1008                              u8                  port_num,
1009                              enum ib_port_state *port_state)
1010 {
1011         unsigned long flags;
1012         int ret = 0;
1013
1014         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
1015                 return -EINVAL;
1016
1017         read_lock_irqsave(&device->cache.lock, flags);
1018         *port_state = device->cache.ports[port_num
1019                 - rdma_start_port(device)].port_state;
1020         read_unlock_irqrestore(&device->cache.lock, flags);
1021
1022         return ret;
1023 }
1024 EXPORT_SYMBOL(ib_get_cached_port_state);
1025
1026 static void ib_cache_update(struct ib_device *device,
1027                             u8                port)
1028 {
1029         struct ib_port_attr       *tprops = NULL;
1030         struct ib_pkey_cache      *pkey_cache = NULL, *old_pkey_cache;
1031         struct ib_gid_cache {
1032                 int             table_len;
1033                 union ib_gid    table[0];
1034         }                         *gid_cache = NULL;
1035         int                        i;
1036         int                        ret;
1037         struct ib_gid_table       *table;
1038         bool                       use_roce_gid_table =
1039                                         rdma_cap_roce_gid_table(device, port);
1040
1041         if (port < rdma_start_port(device) || port > rdma_end_port(device))
1042                 return;
1043
1044         table = device->cache.ports[port - rdma_start_port(device)].gid;
1045
1046         tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1047         if (!tprops)
1048                 return;
1049
1050         ret = ib_query_port(device, port, tprops);
1051         if (ret) {
1052                 pr_warn("ib_query_port failed (%d) for %s\n",
1053                         ret, device->name);
1054                 goto err;
1055         }
1056
1057         pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
1058                              sizeof *pkey_cache->table, GFP_KERNEL);
1059         if (!pkey_cache)
1060                 goto err;
1061
1062         pkey_cache->table_len = tprops->pkey_tbl_len;
1063
1064         if (!use_roce_gid_table) {
1065                 gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
1066                             sizeof(*gid_cache->table), GFP_KERNEL);
1067                 if (!gid_cache)
1068                         goto err;
1069
1070                 gid_cache->table_len = tprops->gid_tbl_len;
1071         }
1072
1073         for (i = 0; i < pkey_cache->table_len; ++i) {
1074                 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1075                 if (ret) {
1076                         pr_warn("ib_query_pkey failed (%d) for %s (index %d)\n",
1077                                 ret, device->name, i);
1078                         goto err;
1079                 }
1080         }
1081
1082         if (!use_roce_gid_table) {
1083                 for (i = 0;  i < gid_cache->table_len; ++i) {
1084                         ret = ib_query_gid(device, port, i,
1085                                            gid_cache->table + i, NULL);
1086                         if (ret) {
1087                                 pr_warn("ib_query_gid failed (%d) for %s (index %d)\n",
1088                                         ret, device->name, i);
1089                                 goto err;
1090                         }
1091                 }
1092         }
1093
1094         write_lock_irq(&device->cache.lock);
1095
1096         old_pkey_cache = device->cache.ports[port -
1097                 rdma_start_port(device)].pkey;
1098
1099         device->cache.ports[port - rdma_start_port(device)].pkey = pkey_cache;
1100         if (!use_roce_gid_table) {
1101                 write_lock(&table->rwlock);
1102                 for (i = 0; i < gid_cache->table_len; i++) {
1103                         modify_gid(device, port, table, i, gid_cache->table + i,
1104                                    &zattr, false);
1105                 }
1106                 write_unlock(&table->rwlock);
1107         }
1108
1109         device->cache.ports[port - rdma_start_port(device)].lmc = tprops->lmc;
1110         device->cache.ports[port - rdma_start_port(device)].port_state =
1111                 tprops->state;
1112
1113         write_unlock_irq(&device->cache.lock);
1114
1115         kfree(gid_cache);
1116         kfree(old_pkey_cache);
1117         kfree(tprops);
1118         return;
1119
1120 err:
1121         kfree(pkey_cache);
1122         kfree(gid_cache);
1123         kfree(tprops);
1124 }
1125
1126 static void ib_cache_task(struct work_struct *_work)
1127 {
1128         struct ib_update_work *work =
1129                 container_of(_work, struct ib_update_work, work);
1130
1131         ib_cache_update(work->device, work->port_num);
1132         kfree(work);
1133 }
1134
1135 static void ib_cache_event(struct ib_event_handler *handler,
1136                            struct ib_event *event)
1137 {
1138         struct ib_update_work *work;
1139
1140         if (event->event == IB_EVENT_PORT_ERR    ||
1141             event->event == IB_EVENT_PORT_ACTIVE ||
1142             event->event == IB_EVENT_LID_CHANGE  ||
1143             event->event == IB_EVENT_PKEY_CHANGE ||
1144             event->event == IB_EVENT_SM_CHANGE   ||
1145             event->event == IB_EVENT_CLIENT_REREGISTER ||
1146             event->event == IB_EVENT_GID_CHANGE) {
1147                 work = kmalloc(sizeof *work, GFP_ATOMIC);
1148                 if (work) {
1149                         INIT_WORK(&work->work, ib_cache_task);
1150                         work->device   = event->device;
1151                         work->port_num = event->element.port_num;
1152                         queue_work(ib_wq, &work->work);
1153                 }
1154         }
1155 }
1156
1157 int ib_cache_setup_one(struct ib_device *device)
1158 {
1159         int p;
1160         int err;
1161
1162         rwlock_init(&device->cache.lock);
1163
1164         device->cache.ports =
1165                 kzalloc(sizeof(*device->cache.ports) *
1166                         (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
1167         if (!device->cache.ports) {
1168                 err = -ENOMEM;
1169                 goto out;
1170         }
1171
1172         err = gid_table_setup_one(device);
1173         if (err)
1174                 goto out;
1175
1176         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1177                 ib_cache_update(device, p + rdma_start_port(device));
1178
1179         INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1180                               device, ib_cache_event);
1181         err = ib_register_event_handler(&device->cache.event_handler);
1182         if (err)
1183                 goto err;
1184
1185         return 0;
1186
1187 err:
1188         gid_table_cleanup_one(device);
1189 out:
1190         return err;
1191 }
1192
1193 void ib_cache_release_one(struct ib_device *device)
1194 {
1195         int p;
1196
1197         /*
1198          * The release function frees all the cache elements.
1199          * This function should be called as part of freeing
1200          * all the device's resources when the cache could no
1201          * longer be accessed.
1202          */
1203         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1204                 kfree(device->cache.ports[p].pkey);
1205
1206         gid_table_release_one(device);
1207         kfree(device->cache.ports);
1208 }
1209
1210 void ib_cache_cleanup_one(struct ib_device *device)
1211 {
1212         /* The cleanup function unregisters the event handler,
1213          * waits for all in-progress workqueue elements and cleans
1214          * up the GID cache. This function should be called after
1215          * the device was removed from the devices list and all
1216          * clients were removed, so the cache exists but is
1217          * non-functional and shouldn't be updated anymore.
1218          */
1219         ib_unregister_event_handler(&device->cache.event_handler);
1220         flush_workqueue(ib_wq);
1221         gid_table_cleanup_one(device);
1222 }
1223
1224 void __init ib_cache_setup(void)
1225 {
1226         roce_gid_mgmt_init();
1227 }
1228
1229 void __exit ib_cache_cleanup(void)
1230 {
1231         roce_gid_mgmt_cleanup();
1232 }