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 -EAGAIN;
 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                         continue;
 577
 578                 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
 579                         continue;
 580
 581                 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr));
 582
 583                 if (filter(gid, &attr, context)) {
 584                         found = true;
 585                         if (index)
 586                                 *index = i;
 587                         break;
 588                 }
 589         }
 590         read_unlock_irqrestore(&table->rwlock, flags);
 591
 592         if (!found)
 593                 return -ENOENT;
 594         return 0;
 595 }
 596
 597 static struct ib_gid_table *alloc_gid_table(int sz)
 598 {
 599         struct ib_gid_table *table =
 600                 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
 601
 602         if (!table)
 603                 return NULL;
 604
 605         table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
 606         if (!table->data_vec)
 607                 goto err_free_table;
 608
 609         mutex_init(&table->lock);
 610
 611         table->sz = sz;
 612         rwlock_init(&table->rwlock);
 613
 614         return table;
 615
 616 err_free_table:
 617         kfree(table);
 618         return NULL;
 619 }
 620
 621 static void release_gid_table(struct ib_gid_table *table)
 622 {
 623         if (table) {
 624                 kfree(table->data_vec);
 625                 kfree(table);
 626         }
 627 }
 628
 629 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
 630                                    struct ib_gid_table *table)
 631 {
 632         int i;
 633         bool deleted = false;
 634
 635         if (!table)
 636                 return;
 637
 638         write_lock_irq(&table->rwlock);
 639         for (i = 0; i < table->sz; ++i) {
 640                 if (memcmp(&table->data_vec[i].gid, &zgid,
 641                            sizeof(table->data_vec[i].gid)))
 642                         if (!del_gid(ib_dev, port, table, i,
 643                                      table->data_vec[i].props &
 644                                      GID_ATTR_FIND_MASK_DEFAULT))
 645                                 deleted = true;
 646         }
 647         write_unlock_irq(&table->rwlock);
 648
 649         if (deleted)
 650                 dispatch_gid_change_event(ib_dev, port);
 651 }
 652
 653 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
 654                                   struct net_device *ndev,
 655                                   unsigned long gid_type_mask,
 656                                   enum ib_cache_gid_default_mode mode)
 657 {
 658         union ib_gid gid;
 659         struct ib_gid_attr gid_attr;
 660         struct ib_gid_attr zattr_type = zattr;
 661         struct ib_gid_table *table;
 662         unsigned int gid_type;
 663
 664         table = ib_dev->cache.ports[port - rdma_start_port(ib_dev)].gid;
 665
 666         make_default_gid(ndev, &gid);
 667         memset(&gid_attr, 0, sizeof(gid_attr));
 668         gid_attr.ndev = ndev;
 669
 670         for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
 671                 int ix;
 672                 union ib_gid current_gid;
 673                 struct ib_gid_attr current_gid_attr = {};
 674
 675                 if (1UL << gid_type & ~gid_type_mask)
 676                         continue;
 677
 678                 gid_attr.gid_type = gid_type;
 679
 680                 mutex_lock(&table->lock);
 681                 write_lock_irq(&table->rwlock);
 682                 ix = find_gid(table, NULL, &gid_attr, true,
 683                               GID_ATTR_FIND_MASK_GID_TYPE |
 684                               GID_ATTR_FIND_MASK_DEFAULT,
 685                               NULL);
 686
 687                 /* Coudn't find default GID location */
 688                 if (WARN_ON(ix < 0))
 689                         goto release;
 690
 691                 zattr_type.gid_type = gid_type;
 692
 693                 if (!__ib_cache_gid_get(ib_dev, port, ix,
 694                                         &current_gid, &current_gid_attr) &&
 695                     mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
 696                     !memcmp(&gid, &current_gid, sizeof(gid)) &&
 697                     !memcmp(&gid_attr, &current_gid_attr, sizeof(gid_attr)))
 698                         goto release;
 699
 700                 if (memcmp(&current_gid, &zgid, sizeof(current_gid)) ||
 701                     memcmp(&current_gid_attr, &zattr_type,
 702                            sizeof(current_gid_attr))) {
 703                         if (del_gid(ib_dev, port, table, ix, true)) {
 704                                 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
 705                                         ix, gid.raw);
 706                                 goto release;
 707                         } else {
 708                                 dispatch_gid_change_event(ib_dev, port);
 709                         }
 710                 }
 711
 712                 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
 713                         if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
 714                                 pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
 715                                         gid.raw);
 716                         else
 717                                 dispatch_gid_change_event(ib_dev, port);
 718                 }
 719
 720 release:
 721                 if (current_gid_attr.ndev)
 722                         dev_put(current_gid_attr.ndev);
 723                 write_unlock_irq(&table->rwlock);
 724                 mutex_unlock(&table->lock);
 725         }
 726 }
 727
 728 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
 729                                      struct ib_gid_table *table)
 730 {
 731         unsigned int i;
 732         unsigned long roce_gid_type_mask;
 733         unsigned int num_default_gids;
 734         unsigned int current_gid = 0;
 735
 736         roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 737         num_default_gids = hweight_long(roce_gid_type_mask);
 738         for (i = 0; i < num_default_gids && i < table->sz; i++) {
 739                 struct ib_gid_table_entry *entry =
 740                         &table->data_vec[i];
 741
 742                 entry->props |= GID_TABLE_ENTRY_DEFAULT;
 743                 current_gid = find_next_bit(&roce_gid_type_mask,
 744                                             BITS_PER_LONG,
 745                                             current_gid);
 746                 entry->attr.gid_type = current_gid++;
 747         }
 748
 749         return 0;
 750 }
 751
 752 static int _gid_table_setup_one(struct ib_device *ib_dev)
 753 {
 754         u8 port;
 755         struct ib_gid_table *table;
 756         int err = 0;
 757
 758         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 759                 u8 rdma_port = port + rdma_start_port(ib_dev);
 760
 761                 table =
 762                         alloc_gid_table(
 763                                 ib_dev->port_immutable[rdma_port].gid_tbl_len);
 764                 if (!table) {
 765                         err = -ENOMEM;
 766                         goto rollback_table_setup;
 767                 }
 768
 769                 err = gid_table_reserve_default(ib_dev,
 770                                                 port + rdma_start_port(ib_dev),
 771                                                 table);
 772                 if (err)
 773                         goto rollback_table_setup;
 774                 ib_dev->cache.ports[port].gid = table;
 775         }
 776
 777         return 0;
 778
 779 rollback_table_setup:
 780         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 781                 table = ib_dev->cache.ports[port].gid;
 782
 783                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 784                                        table);
 785                 release_gid_table(table);
 786         }
 787
 788         return err;
 789 }
 790
 791 static void gid_table_release_one(struct ib_device *ib_dev)
 792 {
 793         struct ib_gid_table *table;
 794         u8 port;
 795
 796         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 797                 table = ib_dev->cache.ports[port].gid;
 798                 release_gid_table(table);
 799                 ib_dev->cache.ports[port].gid = NULL;
 800         }
 801 }
 802
 803 static void gid_table_cleanup_one(struct ib_device *ib_dev)
 804 {
 805         struct ib_gid_table *table;
 806         u8 port;
 807
 808         for (port = 0; port < ib_dev->phys_port_cnt; port++) {
 809                 table = ib_dev->cache.ports[port].gid;
 810                 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
 811                                        table);
 812         }
 813 }
 814
 815 static int gid_table_setup_one(struct ib_device *ib_dev)
 816 {
 817         int err;
 818
 819         err = _gid_table_setup_one(ib_dev);
 820
 821         if (err)
 822                 return err;
 823
 824         rdma_roce_rescan_device(ib_dev);
 825
 826         return err;
 827 }
 828
 829 int ib_get_cached_gid(struct ib_device *device,
 830                       u8                port_num,
 831                       int               index,
 832                       union ib_gid     *gid,
 833                       struct ib_gid_attr *gid_attr)
 834 {
 835         int res;
 836         unsigned long flags;
 837         struct ib_gid_table *table;
 838
 839         if (!rdma_is_port_valid(device, port_num))
 840                 return -EINVAL;
 841
 842         table = device->cache.ports[port_num - rdma_start_port(device)].gid;
 843         read_lock_irqsave(&table->rwlock, flags);
 844         res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr);
 845         read_unlock_irqrestore(&table->rwlock, flags);
 846
 847         return res;
 848 }
 849 EXPORT_SYMBOL(ib_get_cached_gid);
 850
 851 int ib_find_cached_gid(struct ib_device *device,
 852                        const union ib_gid *gid,
 853                        enum ib_gid_type gid_type,
 854                        struct net_device *ndev,
 855                        u8               *port_num,
 856                        u16              *index)
 857 {
 858         return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index);
 859 }
 860 EXPORT_SYMBOL(ib_find_cached_gid);
 861
 862 int ib_find_gid_by_filter(struct ib_device *device,
 863                           const union ib_gid *gid,
 864                           u8 port_num,
 865                           bool (*filter)(const union ib_gid *gid,
 866                                          const struct ib_gid_attr *,
 867                                          void *),
 868                           void *context, u16 *index)
 869 {
 870         /* Only RoCE GID table supports filter function */
 871         if (!rdma_cap_roce_gid_table(device, port_num) && filter)
 872                 return -EPROTONOSUPPORT;
 873
 874         return ib_cache_gid_find_by_filter(device, gid,
 875                                            port_num, filter,
 876                                            context, index);
 877 }
 878
 879 int ib_get_cached_pkey(struct ib_device *device,
 880                        u8                port_num,
 881                        int               index,
 882                        u16              *pkey)
 883 {
 884         struct ib_pkey_cache *cache;
 885         unsigned long flags;
 886         int ret = 0;
 887
 888         if (!rdma_is_port_valid(device, port_num))
 889                 return -EINVAL;
 890
 891         read_lock_irqsave(&device->cache.lock, flags);
 892
 893         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 894
 895         if (index < 0 || index >= cache->table_len)
 896                 ret = -EINVAL;
 897         else
 898                 *pkey = cache->table[index];
 899
 900         read_unlock_irqrestore(&device->cache.lock, flags);
 901
 902         return ret;
 903 }
 904 EXPORT_SYMBOL(ib_get_cached_pkey);
 905
 906 int ib_get_cached_subnet_prefix(struct ib_device *device,
 907                                 u8                port_num,
 908                                 u64              *sn_pfx)
 909 {
 910         unsigned long flags;
 911         int p;
 912
 913         if (port_num < rdma_start_port(device) ||
 914             port_num > rdma_end_port(device))
 915                 return -EINVAL;
 916
 917         p = port_num - rdma_start_port(device);
 918         read_lock_irqsave(&device->cache.lock, flags);
 919         *sn_pfx = device->cache.ports[p].subnet_prefix;
 920         read_unlock_irqrestore(&device->cache.lock, flags);
 921
 922         return 0;
 923 }
 924 EXPORT_SYMBOL(ib_get_cached_subnet_prefix);
 925
 926 int ib_find_cached_pkey(struct ib_device *device,
 927                         u8                port_num,
 928                         u16               pkey,
 929                         u16              *index)
 930 {
 931         struct ib_pkey_cache *cache;
 932         unsigned long flags;
 933         int i;
 934         int ret = -ENOENT;
 935         int partial_ix = -1;
 936
 937         if (!rdma_is_port_valid(device, port_num))
 938                 return -EINVAL;
 939
 940         read_lock_irqsave(&device->cache.lock, flags);
 941
 942         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 943
 944         *index = -1;
 945
 946         for (i = 0; i < cache->table_len; ++i)
 947                 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
 948                         if (cache->table[i] & 0x8000) {
 949                                 *index = i;
 950                                 ret = 0;
 951                                 break;
 952                         } else
 953                                 partial_ix = i;
 954                 }
 955
 956         if (ret && partial_ix >= 0) {
 957                 *index = partial_ix;
 958                 ret = 0;
 959         }
 960
 961         read_unlock_irqrestore(&device->cache.lock, flags);
 962
 963         return ret;
 964 }
 965 EXPORT_SYMBOL(ib_find_cached_pkey);
 966
 967 int ib_find_exact_cached_pkey(struct ib_device *device,
 968                               u8                port_num,
 969                               u16               pkey,
 970                               u16              *index)
 971 {
 972         struct ib_pkey_cache *cache;
 973         unsigned long flags;
 974         int i;
 975         int ret = -ENOENT;
 976
 977         if (!rdma_is_port_valid(device, port_num))
 978                 return -EINVAL;
 979
 980         read_lock_irqsave(&device->cache.lock, flags);
 981
 982         cache = device->cache.ports[port_num - rdma_start_port(device)].pkey;
 983
 984         *index = -1;
 985
 986         for (i = 0; i < cache->table_len; ++i)
 987                 if (cache->table[i] == pkey) {
 988                         *index = i;
 989                         ret = 0;
 990                         break;
 991                 }
 992
 993         read_unlock_irqrestore(&device->cache.lock, flags);
 994
 995         return ret;
 996 }
 997 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
 998
 999 int ib_get_cached_lmc(struct ib_device *device,
1000                       u8                port_num,
1001                       u8                *lmc)
1002 {
1003         unsigned long flags;
1004         int ret = 0;
1005
1006         if (!rdma_is_port_valid(device, port_num))
1007                 return -EINVAL;
1008
1009         read_lock_irqsave(&device->cache.lock, flags);
1010         *lmc = device->cache.ports[port_num - rdma_start_port(device)].lmc;
1011         read_unlock_irqrestore(&device->cache.lock, flags);
1012
1013         return ret;
1014 }
1015 EXPORT_SYMBOL(ib_get_cached_lmc);
1016
1017 int ib_get_cached_port_state(struct ib_device   *device,
1018                              u8                  port_num,
1019                              enum ib_port_state *port_state)
1020 {
1021         unsigned long flags;
1022         int ret = 0;
1023
1024         if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
1025                 return -EINVAL;
1026
1027         read_lock_irqsave(&device->cache.lock, flags);
1028         *port_state = device->cache.ports[port_num
1029                 - rdma_start_port(device)].port_state;
1030         read_unlock_irqrestore(&device->cache.lock, flags);
1031
1032         return ret;
1033 }
1034 EXPORT_SYMBOL(ib_get_cached_port_state);
1035
1036 static void ib_cache_update(struct ib_device *device,
1037                             u8                port,
1038                             bool              enforce_security)
1039 {
1040         struct ib_port_attr       *tprops = NULL;
1041         struct ib_pkey_cache      *pkey_cache = NULL, *old_pkey_cache;
1042         struct ib_gid_cache {
1043                 int             table_len;
1044                 union ib_gid    table[0];
1045         }                         *gid_cache = NULL;
1046         int                        i;
1047         int                        ret;
1048         struct ib_gid_table       *table;
1049         bool                       use_roce_gid_table =
1050                                         rdma_cap_roce_gid_table(device, port);
1051
1052         if (!rdma_is_port_valid(device, port))
1053                 return;
1054
1055         table = device->cache.ports[port - rdma_start_port(device)].gid;
1056
1057         tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1058         if (!tprops)
1059                 return;
1060
1061         ret = ib_query_port(device, port, tprops);
1062         if (ret) {
1063                 pr_warn("ib_query_port failed (%d) for %s\n",
1064                         ret, device->name);
1065                 goto err;
1066         }
1067
1068         pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
1069                              sizeof *pkey_cache->table, GFP_KERNEL);
1070         if (!pkey_cache)
1071                 goto err;
1072
1073         pkey_cache->table_len = tprops->pkey_tbl_len;
1074
1075         if (!use_roce_gid_table) {
1076                 gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
1077                             sizeof(*gid_cache->table), GFP_KERNEL);
1078                 if (!gid_cache)
1079                         goto err;
1080
1081                 gid_cache->table_len = tprops->gid_tbl_len;
1082         }
1083
1084         for (i = 0; i < pkey_cache->table_len; ++i) {
1085                 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1086                 if (ret) {
1087                         pr_warn("ib_query_pkey failed (%d) for %s (index %d)\n",
1088                                 ret, device->name, i);
1089                         goto err;
1090                 }
1091         }
1092
1093         if (!use_roce_gid_table) {
1094                 for (i = 0;  i < gid_cache->table_len; ++i) {
1095                         ret = ib_query_gid(device, port, i,
1096                                            gid_cache->table + i, NULL);
1097                         if (ret) {
1098                                 pr_warn("ib_query_gid failed (%d) for %s (index %d)\n",
1099                                         ret, device->name, i);
1100                                 goto err;
1101                         }
1102                 }
1103         }
1104
1105         write_lock_irq(&device->cache.lock);
1106
1107         old_pkey_cache = device->cache.ports[port -
1108                 rdma_start_port(device)].pkey;
1109
1110         device->cache.ports[port - rdma_start_port(device)].pkey = pkey_cache;
1111         if (!use_roce_gid_table) {
1112                 write_lock(&table->rwlock);
1113                 for (i = 0; i < gid_cache->table_len; i++) {
1114                         modify_gid(device, port, table, i, gid_cache->table + i,
1115                                    &zattr, false);
1116                 }
1117                 write_unlock(&table->rwlock);
1118         }
1119
1120         device->cache.ports[port - rdma_start_port(device)].lmc = tprops->lmc;
1121         device->cache.ports[port - rdma_start_port(device)].port_state =
1122                 tprops->state;
1123
1124         device->cache.ports[port - rdma_start_port(device)].subnet_prefix =
1125                                                         tprops->subnet_prefix;
1126         write_unlock_irq(&device->cache.lock);
1127
1128         if (enforce_security)
1129                 ib_security_cache_change(device,
1130                                          port,
1131                                          tprops->subnet_prefix);
1132
1133         kfree(gid_cache);
1134         kfree(old_pkey_cache);
1135         kfree(tprops);
1136         return;
1137
1138 err:
1139         kfree(pkey_cache);
1140         kfree(gid_cache);
1141         kfree(tprops);
1142 }
1143
1144 static void ib_cache_task(struct work_struct *_work)
1145 {
1146         struct ib_update_work *work =
1147                 container_of(_work, struct ib_update_work, work);
1148
1149         ib_cache_update(work->device,
1150                         work->port_num,
1151                         work->enforce_security);
1152         kfree(work);
1153 }
1154
1155 static void ib_cache_event(struct ib_event_handler *handler,
1156                            struct ib_event *event)
1157 {
1158         struct ib_update_work *work;
1159
1160         if (event->event == IB_EVENT_PORT_ERR    ||
1161             event->event == IB_EVENT_PORT_ACTIVE ||
1162             event->event == IB_EVENT_LID_CHANGE  ||
1163             event->event == IB_EVENT_PKEY_CHANGE ||
1164             event->event == IB_EVENT_SM_CHANGE   ||
1165             event->event == IB_EVENT_CLIENT_REREGISTER ||
1166             event->event == IB_EVENT_GID_CHANGE) {
1167                 work = kmalloc(sizeof *work, GFP_ATOMIC);
1168                 if (work) {
1169                         INIT_WORK(&work->work, ib_cache_task);
1170                         work->device   = event->device;
1171                         work->port_num = event->element.port_num;
1172                         if (event->event == IB_EVENT_PKEY_CHANGE ||
1173                             event->event == IB_EVENT_GID_CHANGE)
1174                                 work->enforce_security = true;
1175                         else
1176                                 work->enforce_security = false;
1177
1178                         queue_work(ib_wq, &work->work);
1179                 }
1180         }
1181 }
1182
1183 int ib_cache_setup_one(struct ib_device *device)
1184 {
1185         int p;
1186         int err;
1187
1188         rwlock_init(&device->cache.lock);
1189
1190         device->cache.ports =
1191                 kzalloc(sizeof(*device->cache.ports) *
1192                         (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
1193         if (!device->cache.ports)
1194                 return -ENOMEM;
1195
1196         err = gid_table_setup_one(device);
1197         if (err) {
1198                 kfree(device->cache.ports);
1199                 device->cache.ports = NULL;
1200                 return err;
1201         }
1202
1203         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1204                 ib_cache_update(device, p + rdma_start_port(device), true);
1205
1206         INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1207                               device, ib_cache_event);
1208         ib_register_event_handler(&device->cache.event_handler);
1209         return 0;
1210 }
1211
1212 void ib_cache_release_one(struct ib_device *device)
1213 {
1214         int p;
1215
1216         /*
1217          * The release function frees all the cache elements.
1218          * This function should be called as part of freeing
1219          * all the device's resources when the cache could no
1220          * longer be accessed.
1221          */
1222         for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1223                 kfree(device->cache.ports[p].pkey);
1224
1225         gid_table_release_one(device);
1226         kfree(device->cache.ports);
1227 }
1228
1229 void ib_cache_cleanup_one(struct ib_device *device)
1230 {
1231         /* The cleanup function unregisters the event handler,
1232          * waits for all in-progress workqueue elements and cleans
1233          * up the GID cache. This function should be called after
1234          * the device was removed from the devices list and all
1235          * clients were removed, so the cache exists but is
1236          * non-functional and shouldn't be updated anymore.
1237          */
1238         ib_unregister_event_handler(&device->cache.event_handler);
1239         flush_workqueue(ib_wq);
1240         gid_table_cleanup_one(device);
1241 }
1242
1243 void __init ib_cache_setup(void)
1244 {
1245         roce_gid_mgmt_init();
1246 }
1247
1248 void __exit ib_cache_cleanup(void)
1249 {
1250         roce_gid_mgmt_cleanup();
1251 }