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 (!rdma_is_port_valid(ib_dev, port))
 508                 return -ENOENT;
 509
 510         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 511
 512         if (ndev)
 513                 mask |= GID_ATTR_FIND_MASK_NETDEV;
 514
 515         read_lock_irqsave(&table->rwlock, flags);
 516         local_index = find_gid(table, gid, &val, false, mask, NULL);
 517         if (local_index >= 0) {
 518                 if (index)
 519                         *index = local_index;
 520                 read_unlock_irqrestore(&table->rwlock, flags);
 521                 return 0;
 522         }
 523
 524         read_unlock_irqrestore(&table->rwlock, flags);
 525         return -ENOENT;
 526 }
 527 EXPORT_SYMBOL(ib_find_cached_gid_by_port);
 528
 529 /**
 530  * ib_find_gid_by_filter - Returns the GID table index where a specified
 531  * GID value occurs
 532  * @device: The device to query.
 533  * @gid: The GID value to search for.
 534  * @port_num: The port number of the device where the GID value could be
 535  *   searched.
 536  * @filter: The filter function is executed on any matching GID in the table.
 537  *   If the filter function returns true, the corresponding index is returned,
 538  *   otherwise, we continue searching the GID table. It's guaranteed that
 539  *   while filter is executed, ndev field is valid and the structure won't
 540  *   change. filter is executed in an atomic context. filter must not be NULL.
 541  * @index: The index into the cached GID table where the GID was found.  This
 542  *   parameter may be NULL.
 543  *
 544  * ib_cache_gid_find_by_filter() searches for the specified GID value
 545  * of which the filter function returns true in the port's GID table.
 546  * This function is only supported on RoCE ports.
 547  *
 548  */
 549 static int ib_cache_gid_find_by_filter(struct ib_device *ib_dev,
 550                                        const union ib_gid *gid,
 551                                        u8 port,
 552                                        bool (*filter)(const union ib_gid *,
 553                                                       const struct ib_gid_attr *,
 554                                                       void *),
 555                                        void *context,
 556                                        u16 *index)
 557 {
 558         struct ib_gid_table *table;
 559         unsigned int i;
 560         unsigned long flags;
 561         bool found = false;
 562
 563
 564         if (!rdma_is_port_valid(ib_dev, port) ||
 565             !rdma_protocol_roce(ib_dev, port))
 566                 return -EPROTONOSUPPORT;
 567
 568         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 569
 570         read_lock_irqsave(&table->rwlock, flags);
 571         for (i = 0; i < table->sz; i++) {
 572                 struct ib_gid_attr attr;
 573
 574                 if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID)
 575                         goto next;
 576
 577                 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
 578                         goto next;
 579
 580                 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr));
 581
 582                 if (filter(gid, &attr, context))
 583                         found = true;
 584
 585 next:
 586                 if (found)
 587                         break;
 588         }
 589         read_unlock_irqrestore(&table->rwlock, flags);
 590
 591         if (!found)
 592                 return -ENOENT;
 593
 594         if (index)
 595                 *index = i;
 596         return 0;
 597 }
 598
 599 static struct ib_gid_table *alloc_gid_table(int sz)
 600 {
 601         struct ib_gid_table *table =
 602                 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
 603
 604         if (!table)
 605                 return NULL;
 606
 607         table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
 608         if (!table->data_vec)
 609                 goto err_free_table;
 610
 611         mutex_init(&table->lock);
 612
 613         table->sz = sz;
 614         rwlock_init(&table->rwlock);
 615
 616         return table;
 617
 618 err_free_table:
 619         kfree(table);
 620         return NULL;
 621 }
 622
 623 static void release_gid_table(struct ib_gid_table *table)
 624 {
 625         if (table) {
 626                 kfree(table->data_vec);
 627                 kfree(table);
 628         }
 629 }
 630
 631 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
 632                                    struct ib_gid_table *table)
 633 {
 634         int i;
 635         bool deleted = false;
 636
 637         if (!table)
 638                 return;
 639
 640         write_lock_irq(&table->rwlock);
 641         for (i = 0; i < table->sz; ++i) {
 642                 if (memcmp(&table->data_vec[i].gid, &zgid,
 643                            sizeof(table->data_vec[i].gid)))
 644                         if (!del_gid(ib_dev, port, table, i,
 645                                      table->data_vec[i].props &
 646                                      GID_ATTR_FIND_MASK_DEFAULT))
 647                                 deleted = true;
 648         }
 649         write_unlock_irq(&table->rwlock);
 650
 651         if (deleted)
 652                 dispatch_gid_change_event(ib_dev, port);
 653 }
 654
 655 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
 656                                   struct net_device *ndev,
 657                                   unsigned long gid_type_mask,
 658                                   enum ib_cache_gid_default_mode mode)
 659 {
 660         union ib_gid gid;
 661         struct ib_gid_attr gid_attr;
 662         struct ib_gid_attr zattr_type = zattr;
 663         struct ib_gid_table *table;
 664         unsigned int gid_type;
 665
 666         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 667
 668         make_default_gid(ndev, &gid);
 669         memset(&gid_attr, 0, sizeof(gid_attr));
 670         gid_attr.ndev = ndev;
 671
 672         for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
 673                 int ix;
 674                 union ib_gid current_gid;
 675                 struct ib_gid_attr current_gid_attr = {};
 676
 677                 if (1UL << gid_type & ~gid_type_mask)
 678                         continue;
 679
 680                 gid_attr.gid_type = gid_type;
 681
 682                 mutex_lock(&table->lock);
 683                 write_lock_irq(&table->rwlock);
 684                 ix = find_gid(table, NULL, &gid_attr, true,
 685                               GID_ATTR_FIND_MASK_GID_TYPE |
 686                               GID_ATTR_FIND_MASK_DEFAULT,
 687                               NULL);
 688
 689                 /* Coudn't find default GID location */
 690                 if (WARN_ON(ix < 0))
 691                         goto release;
 692
 693                 zattr_type.gid_type = gid_type;
 694
 695                 if (!__ib_cache_gid_get(ib_dev, port, ix,
 696                                         &current_gid, &current_gid_attr) &&
 697                     mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
 698                     !memcmp(&gid, &current_gid, sizeof(gid)) &&
 699                     !memcmp(&gid_attr, &current_gid_attr, sizeof(gid_attr)))
 700                         goto release;
 701
 702                 if (memcmp(&current_gid, &zgid, sizeof(current_gid)) ||
 703                     memcmp(&current_gid_attr, &zattr_type,
 704                            sizeof(current_gid_attr))) {
 705                         if (del_gid(ib_dev, port, table, ix, true)) {
 706                                 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
 707                                         ix, gid.raw);
 708                                 goto release;
 709                         } else {
 710                                 dispatch_gid_change_event(ib_dev, port);
 711                         }
 712                 }
 713
 714                 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
 715                         if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
 716                                 pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
 717                                         gid.raw);
 718                         else
 719                                 dispatch_gid_change_event(ib_dev, port);
 720                 }
 721
 722 release:
 723                 if (current_gid_attr.ndev)
 724                         dev_put(current_gid_attr.ndev);
 725                 write_unlock_irq(&table->rwlock);
 726                 mutex_unlock(&table->lock);
 727         }
 728 }
 729
 730 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
 731                                      struct ib_gid_table *table)
 732 {
 733         unsigned int i;
 734         unsigned long roce_gid_type_mask;
 735         unsigned int num_default_gids;
 736         unsigned int current_gid = 0;
 737
 738         roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 739         num_default_gids = hweight_long(roce_gid_type_mask);
 740         for (i = 0; i < num_default_gids && i < table->sz; i++) {
 741                 struct ib_gid_table_entry *entry =
 742                         &table->data_vec[i];
 743
 744                 entry->props |= GID_TABLE_ENTRY_DEFAULT;
 745                 current_gid = find_next_bit(&roce_gid_type_mask,
 746                                             BITS_PER_LONG,
 747                                             current_gid);
 748                 entry->attr.gid_type = current_gid++;
 749         }
 750
 751         return 0;
 752 }
 753
 754 static int _gid_table_setup_one(struct ib_device *ib_dev)
 755 {
 756         u8 port;
 757         struct ib_gid_table *table;
 758         int err = 0;
 759
 760         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 761                 u8 rdma_port = port + rdma_start_port(ib_dev);
 762
 763                 table =
 764                         alloc_gid_table(
 765                                 ib_dev->port_immutable[rdma_port].gid_tbl_len);
 766                 if (!table) {
 767                         err = -ENOMEM;
 768                         goto rollback_table_setup;
 769                 }
 770
 771                 err = gid_table_reserve_default(ib_dev,
 772                                                 port + rdma_start_port(ib_dev),
 773                                                 table);
 774                 if (err)
 775                         goto rollback_table_setup;
 776                 ib_dev->cache.ports[port].gid = table;
 777         }
 778
 779         return 0;
 780
 781 rollback_table_setup:
 782         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 783                 table = ib_dev->cache.ports[port].gid;
 784
 785                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 786                                        table);
 787                 release_gid_table(table);
 788         }
 789
 790         return err;
 791 }
 792
 793 static void gid_table_release_one(struct ib_device *ib_dev)
 794 {
 795         struct ib_gid_table *table;
 796         u8 port;
 797
 798         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 799                 table = ib_dev->cache.ports[port].gid;
 800                 release_gid_table(table);
 801                 ib_dev->cache.ports[port].gid = NULL;
 802         }
 803 }
 804
 805 static void gid_table_cleanup_one(struct ib_device *ib_dev)
 806 {
 807         struct ib_gid_table *table;
 808         u8 port;
 809
 810         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 811                 table = ib_dev->cache.ports[port].gid;
 812                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 813                                        table);
 814         }
 815 }
 816
 817 static int gid_table_setup_one(struct ib_device *ib_dev)
 818 {
 819         int err;
 820
 821         err = _gid_table_setup_one(ib_dev);
 822
 823         if (err)
 824                 return err;
 825
 826         err = roce_rescan_device(ib_dev);
 827
 828         if (err) {
 829                 gid_table_cleanup_one(ib_dev);
 830                 gid_table_release_one(ib_dev);
 831         }
 832
 833         return err;
 834 }
 835
 836 int ib_get_cached_gid(struct ib_device *device,
 837                       u8                port_num,
 838                       int               index,
 839                       union ib_gid     *gid,
 840                       struct ib_gid_attr *gid_attr)
 841 {
 842         int res;
 843         unsigned long flags;
 844         struct ib_gid_table *table;
 845
 846         if (!rdma_is_port_valid(device, port_num))
 847                 return -EINVAL;
 848
 849         table = device->cache.ports[port_num - rdma_start_port(device)].gid;
 850         read_lock_irqsave(&table->rwlock, flags);
 851         res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr);
 852         read_unlock_irqrestore(&table->rwlock, flags);
 853
 854         return res;
 855 }
 856 EXPORT_SYMBOL(ib_get_cached_gid);
 857
 858 int ib_find_cached_gid(struct ib_device *device,
 859                        const union ib_gid *gid,
 860                        enum ib_gid_type gid_type,
 861                        struct net_device *ndev,
 862                        u8               *port_num,
 863                        u16              *index)
 864 {
 865         return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index);
 866 }
 867 EXPORT_SYMBOL(ib_find_cached_gid);
 868
 869 int ib_find_gid_by_filter(struct ib_device *device,
 870                           const union ib_gid *gid,
 871                           u8 port_num,
 872                           bool (*filter)(const union ib_gid *gid,
 873                                          const struct ib_gid_attr *,
 874                                          void *),
 875                           void *context, u16 *index)
 876 {
 877         /* Only RoCE GID table supports filter function */
 878         if (!rdma_cap_roce_gid_table(device, port_num) && filter)
 879                 return -EPROTONOSUPPORT;
 880
 881         return ib_cache_gid_find_by_filter(device, gid,
 882                                            port_num, filter,
 883                                            context, index);
 884 }
 885 EXPORT_SYMBOL(ib_find_gid_by_filter);
 886
 887 int ib_get_cached_pkey(struct ib_device *device,
 888                        u8                port_num,
 889                        int               index,
 890                        u16              *pkey)
 891 {
 892         struct ib_pkey_cache *cache;
 893         unsigned long flags;
 894         int ret = 0;
 895
 896         if (!rdma_is_port_valid(device, port_num))
 897                 return -EINVAL;
 898
 899         read_lock_irqsave(&device->cache.lock, flags);
 900
 901         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 902
 903         if (index < 0 || index >= cache->table_len)
 904                 ret = -EINVAL;
 905         else
 906                 *pkey = cache->table[index];
 907
 908         read_unlock_irqrestore(&device->cache.lock, flags);
 909
 910         return ret;
 911 }
 912 EXPORT_SYMBOL(ib_get_cached_pkey);
 913
 914 int ib_find_cached_pkey(struct ib_device *device,
 915                         u8                port_num,
 916                         u16               pkey,
 917                         u16              *index)
 918 {
 919         struct ib_pkey_cache *cache;
 920         unsigned long flags;
 921         int i;
 922         int ret = -ENOENT;
 923         int partial_ix = -1;
 924
 925         if (!rdma_is_port_valid(device, port_num))
 926                 return -EINVAL;
 927
 928         read_lock_irqsave(&device->cache.lock, flags);
 929
 930         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 931
 932         *index = -1;
 933
 934         for (i = 0; i < cache->table_len; ++i)
 935                 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
 936                         if (cache->table[i] & 0x8000) {
 937                                 *index = i;
 938                                 ret = 0;
 939                                 break;
 940                         } else
 941                                 partial_ix = i;
 942                 }
 943
 944         if (ret && partial_ix >= 0) {
 945                 *index = partial_ix;
 946                 ret = 0;
 947         }
 948
 949         read_unlock_irqrestore(&device->cache.lock, flags);
 950
 951         return ret;
 952 }
 953 EXPORT_SYMBOL(ib_find_cached_pkey);
 954
 955 int ib_find_exact_cached_pkey(struct ib_device *device,
 956                               u8                port_num,
 957                               u16               pkey,
 958                               u16              *index)
 959 {
 960         struct ib_pkey_cache *cache;
 961         unsigned long flags;
 962         int i;
 963         int ret = -ENOENT;
 964
 965         if (!rdma_is_port_valid(device, port_num))
 966                 return -EINVAL;
 967
 968         read_lock_irqsave(&device->cache.lock, flags);
 969
 970         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 971
 972         *index = -1;
 973
 974         for (i = 0; i < cache->table_len; ++i)
 975                 if (cache->table[i] == pkey) {
 976                         *index = i;
 977                         ret = 0;
 978                         break;
 979                 }
 980
 981         read_unlock_irqrestore(&device->cache.lock, flags);
 982
 983         return ret;
 984 }
 985 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
 986
 987 int ib_get_cached_lmc(struct ib_device *device,
 988                       u8                port_num,
 989                       u8                *lmc)
 990 {
 991         unsigned long flags;
 992         int ret = 0;
 993
 994         if (!rdma_is_port_valid(device, port_num))
 995                 return -EINVAL;
 996
 997         read_lock_irqsave(&device->cache.lock, flags);
 998         *lmc = device->cache.ports[port_num - rdma_start_port(device)].lmc;
 999         read_unlock_irqrestore(&device->cache.lock, flags);
1000
1001         return ret;
1002 }
1003 EXPORT_SYMBOL(ib_get_cached_lmc);
1004
1005 int ib_get_cached_port_state(struct ib_device   *device,
1006                              u8                  port_num,
1007                              enum ib_port_state *port_state)
1008 {
1009         unsigned long flags;
1010         int ret = 0;
1011
1012         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
1013                 return -EINVAL;
1014
1015         read_lock_irqsave(&device->cache.lock, flags);
1016         *port_state = device->cache.ports[port_num
1017                 - rdma_start_port(device)].port_state;
1018         read_unlock_irqrestore(&device->cache.lock, flags);
1019
1020         return ret;
1021 }
1022 EXPORT_SYMBOL(ib_get_cached_port_state);
1023
1024 static void ib_cache_update(struct ib_device *device,
1025                             u8                port)
1026 {
1027         struct ib_port_attr       *tprops = NULL;
1028         struct ib_pkey_cache      *pkey_cache = NULL, *old_pkey_cache;
1029         struct ib_gid_cache {
1030                 int             table_len;
1031                 union ib_gid    table[0];
1032         }                         *gid_cache = NULL;
1033         int                        i;
1034         int                        ret;
1035         struct ib_gid_table       *table;
1036         bool                       use_roce_gid_table =
1037                                         rdma_cap_roce_gid_table(device, port);
1038
1039         if (!rdma_is_port_valid(device, port))
1040                 return;
1041
1042         table = device->cache.ports[port - rdma_start_port(device)].gid;
1043
1044         tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1045         if (!tprops)
1046                 return;
1047
1048         ret = ib_query_port(device, port, tprops);
1049         if (ret) {
1050                 pr_warn("ib_query_port failed (%d) for %s\n",
1051                         ret, device->name);
1052                 goto err;
1053         }
1054
1055         pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
1056                              sizeof *pkey_cache->table, GFP_KERNEL);
1057         if (!pkey_cache)
1058                 goto err;
1059
1060         pkey_cache->table_len = tprops->pkey_tbl_len;
1061
1062         if (!use_roce_gid_table) {
1063                 gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
1064                             sizeof(*gid_cache->table), GFP_KERNEL);
1065                 if (!gid_cache)
1066                         goto err;
1067
1068                 gid_cache->table_len = tprops->gid_tbl_len;
1069         }
1070
1071         for (i = 0; i < pkey_cache->table_len; ++i) {
1072                 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1073                 if (ret) {
1074                         pr_warn("ib_query_pkey failed (%d) for %s (index %d)\n",
1075                                 ret, device->name, i);
1076                         goto err;
1077                 }
1078         }
1079
1080         if (!use_roce_gid_table) {
1081                 for (i = 0;  i < gid_cache->table_len; ++i) {
1082                         ret = ib_query_gid(device, port, i,
1083                                            gid_cache->table + i, NULL);
1084                         if (ret) {
1085                                 pr_warn("ib_query_gid failed (%d) for %s (index %d)\n",
1086                                         ret, device->name, i);
1087                                 goto err;
1088                         }
1089                 }
1090         }
1091
1092         write_lock_irq(&device->cache.lock);
1093
1094         old_pkey_cache = device->cache.ports[port -
1095                 rdma_start_port(device)].pkey;
1096
1097         device->cache.ports[port - rdma_start_port(device)].pkey = pkey_cache;
1098         if (!use_roce_gid_table) {
1099                 write_lock(&table->rwlock);
1100                 for (i = 0; i < gid_cache->table_len; i++) {
1101                         modify_gid(device, port, table, i, gid_cache->table + i,
1102                                    &zattr, false);
1103                 }
1104                 write_unlock(&table->rwlock);
1105         }
1106
1107         device->cache.ports[port - rdma_start_port(device)].lmc = tprops->lmc;
1108         device->cache.ports[port - rdma_start_port(device)].port_state =
1109                 tprops->state;
1110
1111         write_unlock_irq(&device->cache.lock);
1112
1113         kfree(gid_cache);
1114         kfree(old_pkey_cache);
1115         kfree(tprops);
1116         return;
1117
1118 err:
1119         kfree(pkey_cache);
1120         kfree(gid_cache);
1121         kfree(tprops);
1122 }
1123
1124 static void ib_cache_task(struct work_struct *_work)
1125 {
1126         struct ib_update_work *work =
1127                 container_of(_work, struct ib_update_work, work);
1128
1129         ib_cache_update(work->device, work->port_num);
1130         kfree(work);
1131 }
1132
1133 static void ib_cache_event(struct ib_event_handler *handler,
1134                            struct ib_event *event)
1135 {
1136         struct ib_update_work *work;
1137
1138         if (event->event == IB_EVENT_PORT_ERR    ||
1139             event->event == IB_EVENT_PORT_ACTIVE ||
1140             event->event == IB_EVENT_LID_CHANGE  ||
1141             event->event == IB_EVENT_PKEY_CHANGE ||
1142             event->event == IB_EVENT_SM_CHANGE   ||
1143             event->event == IB_EVENT_CLIENT_REREGISTER ||
1144             event->event == IB_EVENT_GID_CHANGE) {
1145                 work = kmalloc(sizeof *work, GFP_ATOMIC);
1146                 if (work) {
1147                         INIT_WORK(&work->work, ib_cache_task);
1148                         work->device   = event->device;
1149                         work->port_num = event->element.port_num;
1150                         queue_work(ib_wq, &work->work);
1151                 }
1152         }
1153 }
1154
1155 int ib_cache_setup_one(struct ib_device *device)
1156 {
1157         int p;
1158         int err;
1159
1160         rwlock_init(&device->cache.lock);
1161
1162         device->cache.ports =
1163                 kzalloc(sizeof(*device->cache.ports) *
1164                         (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
1165         if (!device->cache.ports) {
1166                 err = -ENOMEM;
1167                 goto out;
1168         }
1169
1170         err = gid_table_setup_one(device);
1171         if (err)
1172                 goto out;
1173
1174         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1175                 ib_cache_update(device, p + rdma_start_port(device));
1176
1177         INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1178                               device, ib_cache_event);
1179         err = ib_register_event_handler(&device->cache.event_handler);
1180         if (err)
1181                 goto err;
1182
1183         return 0;
1184
1185 err:
1186         gid_table_cleanup_one(device);
1187 out:
1188         return err;
1189 }
1190
1191 void ib_cache_release_one(struct ib_device *device)
1192 {
1193         int p;
1194
1195         /*
1196          * The release function frees all the cache elements.
1197          * This function should be called as part of freeing
1198          * all the device's resources when the cache could no
1199          * longer be accessed.
1200          */
1201         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1202                 kfree(device->cache.ports[p].pkey);
1203
1204         gid_table_release_one(device);
1205         kfree(device->cache.ports);
1206 }
1207
1208 void ib_cache_cleanup_one(struct ib_device *device)
1209 {
1210         /* The cleanup function unregisters the event handler,
1211          * waits for all in-progress workqueue elements and cleans
1212          * up the GID cache. This function should be called after
1213          * the device was removed from the devices list and all
1214          * clients were removed, so the cache exists but is
1215          * non-functional and shouldn't be updated anymore.
1216          */
1217         ib_unregister_event_handler(&device->cache.event_handler);
1218         flush_workqueue(ib_wq);
1219         gid_table_cleanup_one(device);
1220 }
1221
1222 void __init ib_cache_setup(void)
1223 {
1224         roce_gid_mgmt_init();
1225 }
1226
1227 void __exit ib_cache_cleanup(void)
1228 {
1229         roce_gid_mgmt_cleanup();
1230 }