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         bool               enforce_security;
  57 };
  58
  59 union ib_gid zgid;
  60 EXPORT_SYMBOL(zgid);
  61
  62 static const struct ib_gid_attr zattr;
  63
  64 enum gid_attr_find_mask {
  65         GID_ATTR_FIND_MASK_GID          = 1UL << 0,
  66         GID_ATTR_FIND_MASK_NETDEV       = 1UL << 1,
  67         GID_ATTR_FIND_MASK_DEFAULT      = 1UL << 2,
  68         GID_ATTR_FIND_MASK_GID_TYPE     = 1UL << 3,
  69 };
  70
  71 enum gid_table_entry_props {
  72         GID_TABLE_ENTRY_INVALID         = 1UL << 0,
  73         GID_TABLE_ENTRY_DEFAULT         = 1UL << 1,
  74 };
  75
  76 enum gid_table_write_action {
  77         GID_TABLE_WRITE_ACTION_ADD,
  78         GID_TABLE_WRITE_ACTION_DEL,
  79         /* MODIFY only updates the GID table. Currently only used by
  80          * ib_cache_update.
  81          */
  82         GID_TABLE_WRITE_ACTION_MODIFY
  83 };
  84
  85 struct ib_gid_table_entry {
  86         unsigned long       props;
  87         union ib_gid        gid;
  88         struct ib_gid_attr  attr;
  89         void               *context;
  90 };
  91
  92 struct ib_gid_table {
  93         int                  sz;
  94         /* In RoCE, adding a GID to the table requires:
  95          * (a) Find if this GID is already exists.
  96          * (b) Find a free space.
  97          * (c) Write the new GID
  98          *
  99          * Delete requires different set of operations:
 100          * (a) Find the GID
 101          * (b) Delete it.
 102          *
 103          * Add/delete should be carried out atomically.
 104          * This is done by locking this mutex from multiple
 105          * writers. We don't need this lock for IB, as the MAD
 106          * layer replaces all entries. All data_vec entries
 107          * are locked by this lock.
 108          **/
 109         struct mutex         lock;
 110         /* This lock protects the table entries from being
 111          * read and written simultaneously.
 112          */
 113         rwlock_t             rwlock;
 114         struct ib_gid_table_entry *data_vec;
 115 };
 116
 117 static void dispatch_gid_change_event(struct ib_device *ib_dev, u8 port)
 118 {
 119         if (rdma_cap_roce_gid_table(ib_dev, port)) {
 120                 struct ib_event event;
 121
 122                 event.device            = ib_dev;
 123                 event.element.port_num  = port;
 124                 event.event             = IB_EVENT_GID_CHANGE;
 125
 126                 ib_dispatch_event(&event);
 127         }
 128 }
 129
 130 static const char * const gid_type_str[] = {
 131         [IB_GID_TYPE_IB]        = "IB/RoCE v1",
 132         [IB_GID_TYPE_ROCE_UDP_ENCAP]    = "RoCE v2",
 133 };
 134
 135 const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
 136 {
 137         if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
 138                 return gid_type_str[gid_type];
 139
 140         return "Invalid GID type";
 141 }
 142 EXPORT_SYMBOL(ib_cache_gid_type_str);
 143
 144 int ib_cache_gid_parse_type_str(const char *buf)
 145 {
 146         unsigned int i;
 147         size_t len;
 148         int err = -EINVAL;
 149
 150         len = strlen(buf);
 151         if (len == 0)
 152                 return -EINVAL;
 153
 154         if (buf[len - 1] == '\n')
 155                 len--;
 156
 157         for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i)
 158                 if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) &&
 159                     len == strlen(gid_type_str[i])) {
 160                         err = i;
 161                         break;
 162                 }
 163
 164         return err;
 165 }
 166 EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
 167
 168 /* This function expects that rwlock will be write locked in all
 169  * scenarios and that lock will be locked in sleep-able (RoCE)
 170  * scenarios.
 171  */
 172 static int write_gid(struct ib_device *ib_dev, u8 port,
 173                      struct ib_gid_table *table, int ix,
 174                      const union ib_gid *gid,
 175                      const struct ib_gid_attr *attr,
 176                      enum gid_table_write_action action,
 177                      bool  default_gid)
 178         __releases(&table->rwlock) __acquires(&table->rwlock)
 179 {
 180         int ret = 0;
 181         struct net_device *old_net_dev;
 182         enum ib_gid_type old_gid_type;
 183
 184         /* in rdma_cap_roce_gid_table, this funciton should be protected by a
 185          * sleep-able lock.
 186          */
 187
 188         if (rdma_cap_roce_gid_table(ib_dev, port)) {
 189                 table->data_vec[ix].props |= GID_TABLE_ENTRY_INVALID;
 190                 write_unlock_irq(&table->rwlock);
 191                 /* GID_TABLE_WRITE_ACTION_MODIFY currently isn't supported by
 192                  * RoCE providers and thus only updates the cache.
 193                  */
 194                 if (action == GID_TABLE_WRITE_ACTION_ADD)
 195                         ret = ib_dev->add_gid(ib_dev, port, ix, gid, attr,
 196                                               &table->data_vec[ix].context);
 197                 else if (action == GID_TABLE_WRITE_ACTION_DEL)
 198                         ret = ib_dev->del_gid(ib_dev, port, ix,
 199                                               &table->data_vec[ix].context);
 200                 write_lock_irq(&table->rwlock);
 201         }
 202
 203         old_net_dev = table->data_vec[ix].attr.ndev;
 204         old_gid_type = table->data_vec[ix].attr.gid_type;
 205         if (old_net_dev && old_net_dev != attr->ndev)
 206                 dev_put(old_net_dev);
 207         /* if modify_gid failed, just delete the old gid */
 208         if (ret || action == GID_TABLE_WRITE_ACTION_DEL) {
 209                 gid = &zgid;
 210                 attr = &zattr;
 211                 table->data_vec[ix].context = NULL;
 212         }
 213
 214         memcpy(&table->data_vec[ix].gid, gid, sizeof(*gid));
 215         memcpy(&table->data_vec[ix].attr, attr, sizeof(*attr));
 216         if (default_gid) {
 217                 table->data_vec[ix].props |= GID_TABLE_ENTRY_DEFAULT;
 218                 if (action == GID_TABLE_WRITE_ACTION_DEL)
 219                         table->data_vec[ix].attr.gid_type = old_gid_type;
 220         }
 221         if (table->data_vec[ix].attr.ndev &&
 222             table->data_vec[ix].attr.ndev != old_net_dev)
 223                 dev_hold(table->data_vec[ix].attr.ndev);
 224
 225         table->data_vec[ix].props &= ~GID_TABLE_ENTRY_INVALID;
 226
 227         return ret;
 228 }
 229
 230 static int add_gid(struct ib_device *ib_dev, u8 port,
 231                    struct ib_gid_table *table, int ix,
 232                    const union ib_gid *gid,
 233                    const struct ib_gid_attr *attr,
 234                    bool  default_gid) {
 235         return write_gid(ib_dev, port, table, ix, gid, attr,
 236                          GID_TABLE_WRITE_ACTION_ADD, default_gid);
 237 }
 238
 239 static int modify_gid(struct ib_device *ib_dev, u8 port,
 240                       struct ib_gid_table *table, int ix,
 241                       const union ib_gid *gid,
 242                       const struct ib_gid_attr *attr,
 243                       bool  default_gid) {
 244         return write_gid(ib_dev, port, table, ix, gid, attr,
 245                          GID_TABLE_WRITE_ACTION_MODIFY, default_gid);
 246 }
 247
 248 static int del_gid(struct ib_device *ib_dev, u8 port,
 249                    struct ib_gid_table *table, int ix,
 250                    bool  default_gid) {
 251         return write_gid(ib_dev, port, table, ix, &zgid, &zattr,
 252                          GID_TABLE_WRITE_ACTION_DEL, default_gid);
 253 }
 254
 255 /* rwlock should be read locked */
 256 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
 257                     const struct ib_gid_attr *val, bool default_gid,
 258                     unsigned long mask, int *pempty)
 259 {
 260         int i = 0;
 261         int found = -1;
 262         int empty = pempty ? -1 : 0;
 263
 264         while (i < table->sz && (found < 0 || empty < 0)) {
 265                 struct ib_gid_table_entry *data = &table->data_vec[i];
 266                 struct ib_gid_attr *attr = &data->attr;
 267                 int curr_index = i;
 268
 269                 i++;
 270
 271                 if (data->props & GID_TABLE_ENTRY_INVALID)
 272                         continue;
 273
 274                 if (empty < 0)
 275                         if (!memcmp(&data->gid, &zgid, sizeof(*gid)) &&
 276                             !memcmp(attr, &zattr, sizeof(*attr)) &&
 277                             !data->props)
 278                                 empty = curr_index;
 279
 280                 if (found >= 0)
 281                         continue;
 282
 283                 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
 284                     attr->gid_type != val->gid_type)
 285                         continue;
 286
 287                 if (mask & GID_ATTR_FIND_MASK_GID &&
 288                     memcmp(gid, &data->gid, sizeof(*gid)))
 289                         continue;
 290
 291                 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
 292                     attr->ndev != val->ndev)
 293                         continue;
 294
 295                 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
 296                     !!(data->props & GID_TABLE_ENTRY_DEFAULT) !=
 297                     default_gid)
 298                         continue;
 299
 300                 found = curr_index;
 301         }
 302
 303         if (pempty)
 304                 *pempty = empty;
 305
 306         return found;
 307 }
 308
 309 static void make_default_gid(struct  net_device *dev, union ib_gid *gid)
 310 {
 311         gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
 312         addrconf_ifid_eui48(&gid->raw[8], dev);
 313 }
 314
 315 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
 316                      union ib_gid *gid, struct ib_gid_attr *attr)
 317 {
 318         struct ib_gid_table *table;
 319         int ix;
 320         int ret = 0;
 321         struct net_device *idev;
 322         int empty;
 323
 324         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 325
 326         if (!memcmp(gid, &zgid, sizeof(*gid)))
 327                 return -EINVAL;
 328
 329         if (ib_dev->get_netdev) {
 330                 idev = ib_dev->get_netdev(ib_dev, port);
 331                 if (idev && attr->ndev != idev) {
 332                         union ib_gid default_gid;
 333
 334                         /* Adding default GIDs in not permitted */
 335                         make_default_gid(idev, &default_gid);
 336                         if (!memcmp(gid, &default_gid, sizeof(*gid))) {
 337                                 dev_put(idev);
 338                                 return -EPERM;
 339                         }
 340                 }
 341                 if (idev)
 342                         dev_put(idev);
 343         }
 344
 345         mutex_lock(&table->lock);
 346         write_lock_irq(&table->rwlock);
 347
 348         ix = find_gid(table, gid, attr, false, GID_ATTR_FIND_MASK_GID |
 349                       GID_ATTR_FIND_MASK_GID_TYPE |
 350                       GID_ATTR_FIND_MASK_NETDEV, &empty);
 351         if (ix >= 0)
 352                 goto out_unlock;
 353
 354         if (empty < 0) {
 355                 ret = -ENOSPC;
 356                 goto out_unlock;
 357         }
 358
 359         ret = add_gid(ib_dev, port, table, empty, gid, attr, false);
 360         if (!ret)
 361                 dispatch_gid_change_event(ib_dev, port);
 362
 363 out_unlock:
 364         write_unlock_irq(&table->rwlock);
 365         mutex_unlock(&table->lock);
 366         return ret;
 367 }
 368
 369 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
 370                      union ib_gid *gid, struct ib_gid_attr *attr)
 371 {
 372         struct ib_gid_table *table;
 373         int ix;
 374
 375         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 376
 377         mutex_lock(&table->lock);
 378         write_lock_irq(&table->rwlock);
 379
 380         ix = find_gid(table, gid, attr, false,
 381                       GID_ATTR_FIND_MASK_GID      |
 382                       GID_ATTR_FIND_MASK_GID_TYPE |
 383                       GID_ATTR_FIND_MASK_NETDEV   |
 384                       GID_ATTR_FIND_MASK_DEFAULT,
 385                       NULL);
 386         if (ix < 0)
 387                 goto out_unlock;
 388
 389         if (!del_gid(ib_dev, port, table, ix, false))
 390                 dispatch_gid_change_event(ib_dev, port);
 391
 392 out_unlock:
 393         write_unlock_irq(&table->rwlock);
 394         mutex_unlock(&table->lock);
 395         return 0;
 396 }
 397
 398 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
 399                                      struct net_device *ndev)
 400 {
 401         struct ib_gid_table *table;
 402         int ix;
 403         bool deleted = false;
 404
 405         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 406
 407         mutex_lock(&table->lock);
 408         write_lock_irq(&table->rwlock);
 409
 410         for (ix = 0; ix < table->sz; ix++)
 411                 if (table->data_vec[ix].attr.ndev == ndev)
 412                         if (!del_gid(ib_dev, port, table, ix,
 413                                      !!(table->data_vec[ix].props &
 414                                         GID_TABLE_ENTRY_DEFAULT)))
 415                                 deleted = true;
 416
 417         write_unlock_irq(&table->rwlock);
 418         mutex_unlock(&table->lock);
 419
 420         if (deleted)
 421                 dispatch_gid_change_event(ib_dev, port);
 422
 423         return 0;
 424 }
 425
 426 static int __ib_cache_gid_get(struct ib_device *ib_dev, u8 port, int index,
 427                               union ib_gid *gid, struct ib_gid_attr *attr)
 428 {
 429         struct ib_gid_table *table;
 430
 431         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 432
 433         if (index < 0 || index >= table->sz)
 434                 return -EINVAL;
 435
 436         if (table->data_vec[index].props & GID_TABLE_ENTRY_INVALID)
 437                 return -EINVAL;
 438
 439         memcpy(gid, &table->data_vec[index].gid, sizeof(*gid));
 440         if (attr) {
 441                 memcpy(attr, &table->data_vec[index].attr, sizeof(*attr));
 442                 if (attr->ndev)
 443                         dev_hold(attr->ndev);
 444         }
 445
 446         return 0;
 447 }
 448
 449 static int _ib_cache_gid_table_find(struct ib_device *ib_dev,
 450                                     const union ib_gid *gid,
 451                                     const struct ib_gid_attr *val,
 452                                     unsigned long mask,
 453                                     u8 *port, u16 *index)
 454 {
 455         struct ib_gid_table *table;
 456         u8 p;
 457         int local_index;
 458         unsigned long flags;
 459
 460         for (p = 0; p < ib_dev->phys_port_cnt; p++) {
 461                 table = ib_dev->cache.ports[p].gid;
 462                 read_lock_irqsave(&table->rwlock, flags);
 463                 local_index = find_gid(table, gid, val, false, mask, NULL);
 464                 if (local_index >= 0) {
 465                         if (index)
 466                                 *index = local_index;
 467                         if (port)
 468                                 *port = p + rdma_start_port(ib_dev);
 469                         read_unlock_irqrestore(&table->rwlock, flags);
 470                         return 0;
 471                 }
 472                 read_unlock_irqrestore(&table->rwlock, flags);
 473         }
 474
 475         return -ENOENT;
 476 }
 477
 478 static int ib_cache_gid_find(struct ib_device *ib_dev,
 479                              const union ib_gid *gid,
 480                              enum ib_gid_type gid_type,
 481                              struct net_device *ndev, u8 *port,
 482                              u16 *index)
 483 {
 484         unsigned long mask = GID_ATTR_FIND_MASK_GID |
 485                              GID_ATTR_FIND_MASK_GID_TYPE;
 486         struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
 487
 488         if (ndev)
 489                 mask |= GID_ATTR_FIND_MASK_NETDEV;
 490
 491         return _ib_cache_gid_table_find(ib_dev, gid, &gid_attr_val,
 492                                         mask, port, index);
 493 }
 494
 495 int ib_find_cached_gid_by_port(struct ib_device *ib_dev,
 496                                const union ib_gid *gid,
 497                                enum ib_gid_type gid_type,
 498                                u8 port, struct net_device *ndev,
 499                                u16 *index)
 500 {
 501         int local_index;
 502         struct ib_gid_table *table;
 503         unsigned long mask = GID_ATTR_FIND_MASK_GID |
 504                              GID_ATTR_FIND_MASK_GID_TYPE;
 505         struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
 506         unsigned long flags;
 507
 508         if (!rdma_is_port_valid(ib_dev, port))
 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 (!rdma_is_port_valid(ib_dev, port) ||
 566             !rdma_protocol_roce(ib_dev, port))
 567                 return -EPROTONOSUPPORT;
 568
 569         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 570
 571         read_lock_irqsave(&table->rwlock, flags);
 572         for (i = 0; i < table->sz; i++) {
 573                 struct ib_gid_attr attr;
 574
 575                 if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID)
 576                         goto next;
 577
 578                 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
 579                         goto next;
 580
 581                 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr));
 582
 583                 if (filter(gid, &attr, context))
 584                         found = true;
 585
 586 next:
 587                 if (found)
 588                         break;
 589         }
 590         read_unlock_irqrestore(&table->rwlock, flags);
 591
 592         if (!found)
 593                 return -ENOENT;
 594
 595         if (index)
 596                 *index = i;
 597         return 0;
 598 }
 599
 600 static struct ib_gid_table *alloc_gid_table(int sz)
 601 {
 602         struct ib_gid_table *table =
 603                 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
 604
 605         if (!table)
 606                 return NULL;
 607
 608         table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
 609         if (!table->data_vec)
 610                 goto err_free_table;
 611
 612         mutex_init(&table->lock);
 613
 614         table->sz = sz;
 615         rwlock_init(&table->rwlock);
 616
 617         return table;
 618
 619 err_free_table:
 620         kfree(table);
 621         return NULL;
 622 }
 623
 624 static void release_gid_table(struct ib_gid_table *table)
 625 {
 626         if (table) {
 627                 kfree(table->data_vec);
 628                 kfree(table);
 629         }
 630 }
 631
 632 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
 633                                    struct ib_gid_table *table)
 634 {
 635         int i;
 636         bool deleted = false;
 637
 638         if (!table)
 639                 return;
 640
 641         write_lock_irq(&table->rwlock);
 642         for (i = 0; i < table->sz; ++i) {
 643                 if (memcmp(&table->data_vec[i].gid, &zgid,
 644                            sizeof(table->data_vec[i].gid)))
 645                         if (!del_gid(ib_dev, port, table, i,
 646                                      table->data_vec[i].props &
 647                                      GID_ATTR_FIND_MASK_DEFAULT))
 648                                 deleted = true;
 649         }
 650         write_unlock_irq(&table->rwlock);
 651
 652         if (deleted)
 653                 dispatch_gid_change_event(ib_dev, port);
 654 }
 655
 656 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
 657                                   struct net_device *ndev,
 658                                   unsigned long gid_type_mask,
 659                                   enum ib_cache_gid_default_mode mode)
 660 {
 661         union ib_gid gid;
 662         struct ib_gid_attr gid_attr;
 663         struct ib_gid_attr zattr_type = zattr;
 664         struct ib_gid_table *table;
 665         unsigned int gid_type;
 666
 667         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 668
 669         make_default_gid(ndev, &gid);
 670         memset(&gid_attr, 0, sizeof(gid_attr));
 671         gid_attr.ndev = ndev;
 672
 673         for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
 674                 int ix;
 675                 union ib_gid current_gid;
 676                 struct ib_gid_attr current_gid_attr = {};
 677
 678                 if (1UL << gid_type & ~gid_type_mask)
 679                         continue;
 680
 681                 gid_attr.gid_type = gid_type;
 682
 683                 mutex_lock(&table->lock);
 684                 write_lock_irq(&table->rwlock);
 685                 ix = find_gid(table, NULL, &gid_attr, true,
 686                               GID_ATTR_FIND_MASK_GID_TYPE |
 687                               GID_ATTR_FIND_MASK_DEFAULT,
 688                               NULL);
 689
 690                 /* Coudn't find default GID location */
 691                 if (WARN_ON(ix < 0))
 692                         goto release;
 693
 694                 zattr_type.gid_type = gid_type;
 695
 696                 if (!__ib_cache_gid_get(ib_dev, port, ix,
 697                                         &current_gid, &current_gid_attr) &&
 698                     mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
 699                     !memcmp(&gid, &current_gid, sizeof(gid)) &&
 700                     !memcmp(&gid_attr, &current_gid_attr, sizeof(gid_attr)))
 701                         goto release;
 702
 703                 if (memcmp(&current_gid, &zgid, sizeof(current_gid)) ||
 704                     memcmp(&current_gid_attr, &zattr_type,
 705                            sizeof(current_gid_attr))) {
 706                         if (del_gid(ib_dev, port, table, ix, true)) {
 707                                 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
 708                                         ix, gid.raw);
 709                                 goto release;
 710                         } else {
 711                                 dispatch_gid_change_event(ib_dev, port);
 712                         }
 713                 }
 714
 715                 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
 716                         if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
 717                                 pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
 718                                         gid.raw);
 719                         else
 720                                 dispatch_gid_change_event(ib_dev, port);
 721                 }
 722
 723 release:
 724                 if (current_gid_attr.ndev)
 725                         dev_put(current_gid_attr.ndev);
 726                 write_unlock_irq(&table->rwlock);
 727                 mutex_unlock(&table->lock);
 728         }
 729 }
 730
 731 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
 732                                      struct ib_gid_table *table)
 733 {
 734         unsigned int i;
 735         unsigned long roce_gid_type_mask;
 736         unsigned int num_default_gids;
 737         unsigned int current_gid = 0;
 738
 739         roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 740         num_default_gids = hweight_long(roce_gid_type_mask);
 741         for (i = 0; i < num_default_gids && i < table->sz; i++) {
 742                 struct ib_gid_table_entry *entry =
 743                         &table->data_vec[i];
 744
 745                 entry->props |= GID_TABLE_ENTRY_DEFAULT;
 746                 current_gid = find_next_bit(&roce_gid_type_mask,
 747                                             BITS_PER_LONG,
 748                                             current_gid);
 749                 entry->attr.gid_type = current_gid++;
 750         }
 751
 752         return 0;
 753 }
 754
 755 static int _gid_table_setup_one(struct ib_device *ib_dev)
 756 {
 757         u8 port;
 758         struct ib_gid_table *table;
 759         int err = 0;
 760
 761         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 762                 u8 rdma_port = port + rdma_start_port(ib_dev);
 763
 764                 table =
 765                         alloc_gid_table(
 766                                 ib_dev->port_immutable[rdma_port].gid_tbl_len);
 767                 if (!table) {
 768                         err = -ENOMEM;
 769                         goto rollback_table_setup;
 770                 }
 771
 772                 err = gid_table_reserve_default(ib_dev,
 773                                                 port + rdma_start_port(ib_dev),
 774                                                 table);
 775                 if (err)
 776                         goto rollback_table_setup;
 777                 ib_dev->cache.ports[port].gid = table;
 778         }
 779
 780         return 0;
 781
 782 rollback_table_setup:
 783         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 784                 table = ib_dev->cache.ports[port].gid;
 785
 786                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 787                                        table);
 788                 release_gid_table(table);
 789         }
 790
 791         return err;
 792 }
 793
 794 static void gid_table_release_one(struct ib_device *ib_dev)
 795 {
 796         struct ib_gid_table *table;
 797         u8 port;
 798
 799         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 800                 table = ib_dev->cache.ports[port].gid;
 801                 release_gid_table(table);
 802                 ib_dev->cache.ports[port].gid = NULL;
 803         }
 804 }
 805
 806 static void gid_table_cleanup_one(struct ib_device *ib_dev)
 807 {
 808         struct ib_gid_table *table;
 809         u8 port;
 810
 811         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 812                 table = ib_dev->cache.ports[port].gid;
 813                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 814                                        table);
 815         }
 816 }
 817
 818 static int gid_table_setup_one(struct ib_device *ib_dev)
 819 {
 820         int err;
 821
 822         err = _gid_table_setup_one(ib_dev);
 823
 824         if (err)
 825                 return err;
 826
 827         err = roce_rescan_device(ib_dev);
 828
 829         if (err) {
 830                 gid_table_cleanup_one(ib_dev);
 831                 gid_table_release_one(ib_dev);
 832         }
 833
 834         return err;
 835 }
 836
 837 int ib_get_cached_gid(struct ib_device *device,
 838                       u8                port_num,
 839                       int               index,
 840                       union ib_gid     *gid,
 841                       struct ib_gid_attr *gid_attr)
 842 {
 843         int res;
 844         unsigned long flags;
 845         struct ib_gid_table *table;
 846
 847         if (!rdma_is_port_valid(device, port_num))
 848                 return -EINVAL;
 849
 850         table = device->cache.ports[port_num - rdma_start_port(device)].gid;
 851         read_lock_irqsave(&table->rwlock, flags);
 852         res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr);
 853         read_unlock_irqrestore(&table->rwlock, flags);
 854
 855         return res;
 856 }
 857 EXPORT_SYMBOL(ib_get_cached_gid);
 858
 859 int ib_find_cached_gid(struct ib_device *device,
 860                        const union ib_gid *gid,
 861                        enum ib_gid_type gid_type,
 862                        struct net_device *ndev,
 863                        u8               *port_num,
 864                        u16              *index)
 865 {
 866         return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index);
 867 }
 868 EXPORT_SYMBOL(ib_find_cached_gid);
 869
 870 int ib_find_gid_by_filter(struct ib_device *device,
 871                           const union ib_gid *gid,
 872                           u8 port_num,
 873                           bool (*filter)(const union ib_gid *gid,
 874                                          const struct ib_gid_attr *,
 875                                          void *),
 876                           void *context, u16 *index)
 877 {
 878         /* Only RoCE GID table supports filter function */
 879         if (!rdma_cap_roce_gid_table(device, port_num) && filter)
 880                 return -EPROTONOSUPPORT;
 881
 882         return ib_cache_gid_find_by_filter(device, gid,
 883                                            port_num, filter,
 884                                            context, index);
 885 }
 886 EXPORT_SYMBOL(ib_find_gid_by_filter);
 887
 888 int ib_get_cached_pkey(struct ib_device *device,
 889                        u8                port_num,
 890                        int               index,
 891                        u16              *pkey)
 892 {
 893         struct ib_pkey_cache *cache;
 894         unsigned long flags;
 895         int ret = 0;
 896
 897         if (!rdma_is_port_valid(device, port_num))
 898                 return -EINVAL;
 899
 900         read_lock_irqsave(&device->cache.lock, flags);
 901
 902         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 903
 904         if (index < 0 || index >= cache->table_len)
 905                 ret = -EINVAL;
 906         else
 907                 *pkey = cache->table[index];
 908
 909         read_unlock_irqrestore(&device->cache.lock, flags);
 910
 911         return ret;
 912 }
 913 EXPORT_SYMBOL(ib_get_cached_pkey);
 914
 915 int ib_get_cached_subnet_prefix(struct ib_device *device,
 916                                 u8                port_num,
 917                                 u64              *sn_pfx)
 918 {
 919         unsigned long flags;
 920         int p;
 921
 922         if (port_num < rdma_start_port(device) ||
 923             port_num > rdma_end_port(device))
 924                 return -EINVAL;
 925
 926         p = port_num - rdma_start_port(device);
 927         read_lock_irqsave(&device->cache.lock, flags);
 928         *sn_pfx = device->cache.ports[p].subnet_prefix;
 929         read_unlock_irqrestore(&device->cache.lock, flags);
 930
 931         return 0;
 932 }
 933 EXPORT_SYMBOL(ib_get_cached_subnet_prefix);
 934
 935 int ib_find_cached_pkey(struct ib_device *device,
 936                         u8                port_num,
 937                         u16               pkey,
 938                         u16              *index)
 939 {
 940         struct ib_pkey_cache *cache;
 941         unsigned long flags;
 942         int i;
 943         int ret = -ENOENT;
 944         int partial_ix = -1;
 945
 946         if (!rdma_is_port_valid(device, port_num))
 947                 return -EINVAL;
 948
 949         read_lock_irqsave(&device->cache.lock, flags);
 950
 951         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 952
 953         *index = -1;
 954
 955         for (i = 0; i < cache->table_len; ++i)
 956                 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
 957                         if (cache->table[i] & 0x8000) {
 958                                 *index = i;
 959                                 ret = 0;
 960                                 break;
 961                         } else
 962                                 partial_ix = i;
 963                 }
 964
 965         if (ret && partial_ix >= 0) {
 966                 *index = partial_ix;
 967                 ret = 0;
 968         }
 969
 970         read_unlock_irqrestore(&device->cache.lock, flags);
 971
 972         return ret;
 973 }
 974 EXPORT_SYMBOL(ib_find_cached_pkey);
 975
 976 int ib_find_exact_cached_pkey(struct ib_device *device,
 977                               u8                port_num,
 978                               u16               pkey,
 979                               u16              *index)
 980 {
 981         struct ib_pkey_cache *cache;
 982         unsigned long flags;
 983         int i;
 984         int ret = -ENOENT;
 985
 986         if (!rdma_is_port_valid(device, port_num))
 987                 return -EINVAL;
 988
 989         read_lock_irqsave(&device->cache.lock, flags);
 990
 991         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 992
 993         *index = -1;
 994
 995         for (i = 0; i < cache->table_len; ++i)
 996                 if (cache->table[i] == pkey) {
 997                         *index = i;
 998                         ret = 0;
 999                         break;
1000                 }
1001
1002         read_unlock_irqrestore(&device->cache.lock, flags);
1003
1004         return ret;
1005 }
1006 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
1007
1008 int ib_get_cached_lmc(struct ib_device *device,
1009                       u8                port_num,
1010                       u8                *lmc)
1011 {
1012         unsigned long flags;
1013         int ret = 0;
1014
1015         if (!rdma_is_port_valid(device, port_num))
1016                 return -EINVAL;
1017
1018         read_lock_irqsave(&device->cache.lock, flags);
1019         *lmc = device->cache.ports[port_num - rdma_start_port(device)].lmc;
1020         read_unlock_irqrestore(&device->cache.lock, flags);
1021
1022         return ret;
1023 }
1024 EXPORT_SYMBOL(ib_get_cached_lmc);
1025
1026 int ib_get_cached_port_state(struct ib_device   *device,
1027                              u8                  port_num,
1028                              enum ib_port_state *port_state)
1029 {
1030         unsigned long flags;
1031         int ret = 0;
1032
1033         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
1034                 return -EINVAL;
1035
1036         read_lock_irqsave(&device->cache.lock, flags);
1037         *port_state = device->cache.ports[port_num
1038                 - rdma_start_port(device)].port_state;
1039         read_unlock_irqrestore(&device->cache.lock, flags);
1040
1041         return ret;
1042 }
1043 EXPORT_SYMBOL(ib_get_cached_port_state);
1044
1045 static void ib_cache_update(struct ib_device *device,
1046                             u8                port,
1047                             bool              enforce_security)
1048 {
1049         struct ib_port_attr       *tprops = NULL;
1050         struct ib_pkey_cache      *pkey_cache = NULL, *old_pkey_cache;
1051         struct ib_gid_cache {
1052                 int             table_len;
1053                 union ib_gid    table[0];
1054         }                         *gid_cache = NULL;
1055         int                        i;
1056         int                        ret;
1057         struct ib_gid_table       *table;
1058         bool                       use_roce_gid_table =
1059                                         rdma_cap_roce_gid_table(device, port);
1060
1061         if (!rdma_is_port_valid(device, port))
1062                 return;
1063
1064         table = device->cache.ports[port - rdma_start_port(device)].gid;
1065
1066         tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1067         if (!tprops)
1068                 return;
1069
1070         ret = ib_query_port(device, port, tprops);
1071         if (ret) {
1072                 pr_warn("ib_query_port failed (%d) for %s\n",
1073                         ret, device->name);
1074                 goto err;
1075         }
1076
1077         pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
1078                              sizeof *pkey_cache->table, GFP_KERNEL);
1079         if (!pkey_cache)
1080                 goto err;
1081
1082         pkey_cache->table_len = tprops->pkey_tbl_len;
1083
1084         if (!use_roce_gid_table) {
1085                 gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
1086                             sizeof(*gid_cache->table), GFP_KERNEL);
1087                 if (!gid_cache)
1088                         goto err;
1089
1090                 gid_cache->table_len = tprops->gid_tbl_len;
1091         }
1092
1093         for (i = 0; i < pkey_cache->table_len; ++i) {
1094                 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1095                 if (ret) {
1096                         pr_warn("ib_query_pkey failed (%d) for %s (index %d)\n",
1097                                 ret, device->name, i);
1098                         goto err;
1099                 }
1100         }
1101
1102         if (!use_roce_gid_table) {
1103                 for (i = 0;  i < gid_cache->table_len; ++i) {
1104                         ret = ib_query_gid(device, port, i,
1105                                            gid_cache->table + i, NULL);
1106                         if (ret) {
1107                                 pr_warn("ib_query_gid failed (%d) for %s (index %d)\n",
1108                                         ret, device->name, i);
1109                                 goto err;
1110                         }
1111                 }
1112         }
1113
1114         write_lock_irq(&device->cache.lock);
1115
1116         old_pkey_cache = device->cache.ports[port -
1117                 rdma_start_port(device)].pkey;
1118
1119         device->cache.ports[port - rdma_start_port(device)].pkey = pkey_cache;
1120         if (!use_roce_gid_table) {
1121                 write_lock(&table->rwlock);
1122                 for (i = 0; i < gid_cache->table_len; i++) {
1123                         modify_gid(device, port, table, i, gid_cache->table + i,
1124                                    &zattr, false);
1125                 }
1126                 write_unlock(&table->rwlock);
1127         }
1128
1129         device->cache.ports[port - rdma_start_port(device)].lmc = tprops->lmc;
1130         device->cache.ports[port - rdma_start_port(device)].port_state =
1131                 tprops->state;
1132
1133         device->cache.ports[port - rdma_start_port(device)].subnet_prefix =
1134                                                         tprops->subnet_prefix;
1135         write_unlock_irq(&device->cache.lock);
1136
1137         if (enforce_security)
1138                 ib_security_cache_change(device,
1139                                          port,
1140                                          tprops->subnet_prefix);
1141
1142         kfree(gid_cache);
1143         kfree(old_pkey_cache);
1144         kfree(tprops);
1145         return;
1146
1147 err:
1148         kfree(pkey_cache);
1149         kfree(gid_cache);
1150         kfree(tprops);
1151 }
1152
1153 static void ib_cache_task(struct work_struct *_work)
1154 {
1155         struct ib_update_work *work =
1156                 container_of(_work, struct ib_update_work, work);
1157
1158         ib_cache_update(work->device,
1159                         work->port_num,
1160                         work->enforce_security);
1161         kfree(work);
1162 }
1163
1164 static void ib_cache_event(struct ib_event_handler *handler,
1165                            struct ib_event *event)
1166 {
1167         struct ib_update_work *work;
1168
1169         if (event->event == IB_EVENT_PORT_ERR    ||
1170             event->event == IB_EVENT_PORT_ACTIVE ||
1171             event->event == IB_EVENT_LID_CHANGE  ||
1172             event->event == IB_EVENT_PKEY_CHANGE ||
1173             event->event == IB_EVENT_SM_CHANGE   ||
1174             event->event == IB_EVENT_CLIENT_REREGISTER ||
1175             event->event == IB_EVENT_GID_CHANGE) {
1176                 work = kmalloc(sizeof *work, GFP_ATOMIC);
1177                 if (work) {
1178                         INIT_WORK(&work->work, ib_cache_task);
1179                         work->device   = event->device;
1180                         work->port_num = event->element.port_num;
1181                         if (event->event == IB_EVENT_PKEY_CHANGE ||
1182                             event->event == IB_EVENT_GID_CHANGE)
1183                                 work->enforce_security = true;
1184                         else
1185                                 work->enforce_security = false;
1186
1187                         queue_work(ib_wq, &work->work);
1188                 }
1189         }
1190 }
1191
1192 int ib_cache_setup_one(struct ib_device *device)
1193 {
1194         int p;
1195         int err;
1196
1197         rwlock_init(&device->cache.lock);
1198
1199         device->cache.ports =
1200                 kzalloc(sizeof(*device->cache.ports) *
1201                         (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
1202         if (!device->cache.ports)
1203                 return -ENOMEM;
1204
1205         err = gid_table_setup_one(device);
1206         if (err) {
1207                 kfree(device->cache.ports);
1208                 device->cache.ports = NULL;
1209                 return err;
1210         }
1211
1212         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1213                 ib_cache_update(device, p + rdma_start_port(device), true);
1214
1215         INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1216                               device, ib_cache_event);
1217         ib_register_event_handler(&device->cache.event_handler);
1218         return 0;
1219 }
1220
1221 void ib_cache_release_one(struct ib_device *device)
1222 {
1223         int p;
1224
1225         /*
1226          * The release function frees all the cache elements.
1227          * This function should be called as part of freeing
1228          * all the device's resources when the cache could no
1229          * longer be accessed.
1230          */
1231         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1232                 kfree(device->cache.ports[p].pkey);
1233
1234         gid_table_release_one(device);
1235         kfree(device->cache.ports);
1236 }
1237
1238 void ib_cache_cleanup_one(struct ib_device *device)
1239 {
1240         /* The cleanup function unregisters the event handler,
1241          * waits for all in-progress workqueue elements and cleans
1242          * up the GID cache. This function should be called after
1243          * the device was removed from the devices list and all
1244          * clients were removed, so the cache exists but is
1245          * non-functional and shouldn't be updated anymore.
1246          */
1247         ib_unregister_event_handler(&device->cache.event_handler);
1248         flush_workqueue(ib_wq);
1249         gid_table_cleanup_one(device);
1250 }
1251
1252 void __init ib_cache_setup(void)
1253 {
1254         roce_gid_mgmt_init();
1255 }
1256
1257 void __exit ib_cache_cleanup(void)
1258 {
1259         roce_gid_mgmt_cleanup();
1260 }