1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2004 Topspin Communications. All rights reserved. 5 * Copyright (c) 2005 Intel Corporation. All rights reserved. 6 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 7 * Copyright (c) 2005 Voltaire, Inc. All rights reserved. 8 * 9 * This software is available to you under a choice of one of two 10 * licenses. You may choose to be licensed under the terms of the GNU 11 * General Public License (GPL) Version 2, available from the file 12 * COPYING in the main directory of this source tree, or the 13 * OpenIB.org BSD license below: 14 * 15 * Redistribution and use in source and binary forms, with or 16 * without modification, are permitted provided that the following 17 * conditions are met: 18 * 19 * - Redistributions of source code must retain the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer. 22 * 23 * - Redistributions in binary form must reproduce the above 24 * copyright notice, this list of conditions and the following 25 * disclaimer in the documentation and/or other materials 26 * provided with the distribution. 27 * 28 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 29 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 30 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 31 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 32 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 33 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 34 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 35 * SOFTWARE. 36 */ 37 38 #include <sys/cdefs.h> 39 __FBSDID("$FreeBSD$"); 40 41 #include <linux/module.h> 42 #include <linux/errno.h> 43 #include <linux/slab.h> 44 #include <linux/workqueue.h> 45 #include <linux/netdevice.h> 46 #include <linux/in6.h> 47 48 #include <rdma/ib_cache.h> 49 50 #include "core_priv.h" 51 52 struct ib_pkey_cache { 53 int table_len; 54 u16 table[0]; 55 }; 56 57 struct ib_update_work { 58 struct work_struct work; 59 struct ib_device *device; 60 u8 port_num; 61 }; 62 63 union ib_gid zgid; 64 EXPORT_SYMBOL(zgid); 65 66 static const struct ib_gid_attr zattr; 67 68 enum gid_attr_find_mask { 69 GID_ATTR_FIND_MASK_GID = 1UL << 0, 70 GID_ATTR_FIND_MASK_NETDEV = 1UL << 1, 71 GID_ATTR_FIND_MASK_DEFAULT = 1UL << 2, 72 GID_ATTR_FIND_MASK_GID_TYPE = 1UL << 3, 73 }; 74 75 enum gid_table_entry_props { 76 GID_TABLE_ENTRY_INVALID = 1UL << 0, 77 GID_TABLE_ENTRY_DEFAULT = 1UL << 1, 78 }; 79 80 enum gid_table_write_action { 81 GID_TABLE_WRITE_ACTION_ADD, 82 GID_TABLE_WRITE_ACTION_DEL, 83 /* MODIFY only updates the GID table. Currently only used by 84 * ib_cache_update. 85 */ 86 GID_TABLE_WRITE_ACTION_MODIFY 87 }; 88 89 struct ib_gid_table_entry { 90 unsigned long props; 91 union ib_gid gid; 92 struct ib_gid_attr attr; 93 void *context; 94 }; 95 96 struct ib_gid_table { 97 int sz; 98 /* In RoCE, adding a GID to the table requires: 99 * (a) Find if this GID is already exists. 100 * (b) Find a free space. 101 * (c) Write the new GID 102 * 103 * Delete requires different set of operations: 104 * (a) Find the GID 105 * (b) Delete it. 106 * 107 * Add/delete should be carried out atomically. 108 * This is done by locking this mutex from multiple 109 * writers. We don't need this lock for IB, as the MAD 110 * layer replaces all entries. All data_vec entries 111 * are locked by this lock. 112 **/ 113 struct mutex lock; 114 /* This lock protects the table entries from being 115 * read and written simultaneously. 116 */ 117 rwlock_t rwlock; 118 struct ib_gid_table_entry *data_vec; 119 }; 120 121 static void dispatch_gid_change_event(struct ib_device *ib_dev, u8 port) 122 { 123 if (rdma_cap_roce_gid_table(ib_dev, port)) { 124 struct ib_event event; 125 126 event.device = ib_dev; 127 event.element.port_num = port; 128 event.event = IB_EVENT_GID_CHANGE; 129 130 ib_dispatch_event(&event); 131 } 132 } 133 134 static const char * const gid_type_str[] = { 135 [IB_GID_TYPE_IB] = "IB/RoCE v1", 136 [IB_GID_TYPE_ROCE_UDP_ENCAP] = "RoCE v2", 137 }; 138 139 const char *ib_cache_gid_type_str(enum ib_gid_type gid_type) 140 { 141 if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type]) 142 return gid_type_str[gid_type]; 143 144 return "Invalid GID type"; 145 } 146 EXPORT_SYMBOL(ib_cache_gid_type_str); 147 148 int ib_cache_gid_parse_type_str(const char *buf) 149 { 150 unsigned int i; 151 size_t len; 152 int err = -EINVAL; 153 154 len = strlen(buf); 155 if (len == 0) 156 return -EINVAL; 157 158 if (buf[len - 1] == '\n') 159 len--; 160 161 for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i) 162 if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) && 163 len == strlen(gid_type_str[i])) { 164 err = i; 165 break; 166 } 167 168 return err; 169 } 170 EXPORT_SYMBOL(ib_cache_gid_parse_type_str); 171 172 /* This function expects that rwlock will be write locked in all 173 * scenarios and that lock will be locked in sleep-able (RoCE) 174 * scenarios. 175 */ 176 static int write_gid(struct ib_device *ib_dev, u8 port, 177 struct ib_gid_table *table, int ix, 178 const union ib_gid *gid, 179 const struct ib_gid_attr *attr, 180 enum gid_table_write_action action, 181 bool default_gid) 182 __releases(&table->rwlock) __acquires(&table->rwlock) 183 { 184 int ret = 0; 185 struct net_device *old_net_dev; 186 enum ib_gid_type old_gid_type; 187 188 /* in rdma_cap_roce_gid_table, this funciton should be protected by a 189 * sleep-able lock. 190 */ 191 192 if (rdma_cap_roce_gid_table(ib_dev, port)) { 193 table->data_vec[ix].props |= GID_TABLE_ENTRY_INVALID; 194 write_unlock_irq(&table->rwlock); 195 /* GID_TABLE_WRITE_ACTION_MODIFY currently isn't supported by 196 * RoCE providers and thus only updates the cache. 197 */ 198 if (action == GID_TABLE_WRITE_ACTION_ADD) 199 ret = ib_dev->add_gid(ib_dev, port, ix, gid, attr, 200 &table->data_vec[ix].context); 201 else if (action == GID_TABLE_WRITE_ACTION_DEL) 202 ret = ib_dev->del_gid(ib_dev, port, ix, 203 &table->data_vec[ix].context); 204 write_lock_irq(&table->rwlock); 205 } 206 207 old_net_dev = table->data_vec[ix].attr.ndev; 208 old_gid_type = table->data_vec[ix].attr.gid_type; 209 if (old_net_dev && old_net_dev != attr->ndev) 210 dev_put(old_net_dev); 211 /* if modify_gid failed, just delete the old gid */ 212 if (ret || action == GID_TABLE_WRITE_ACTION_DEL) { 213 gid = &zgid; 214 attr = &zattr; 215 table->data_vec[ix].context = NULL; 216 } 217 218 memcpy(&table->data_vec[ix].gid, gid, sizeof(*gid)); 219 memcpy(&table->data_vec[ix].attr, attr, sizeof(*attr)); 220 if (default_gid) { 221 table->data_vec[ix].props |= GID_TABLE_ENTRY_DEFAULT; 222 if (action == GID_TABLE_WRITE_ACTION_DEL) 223 table->data_vec[ix].attr.gid_type = old_gid_type; 224 } 225 if (table->data_vec[ix].attr.ndev && 226 table->data_vec[ix].attr.ndev != old_net_dev) 227 dev_hold(table->data_vec[ix].attr.ndev); 228 229 table->data_vec[ix].props &= ~GID_TABLE_ENTRY_INVALID; 230 231 return ret; 232 } 233 234 static int add_gid(struct ib_device *ib_dev, u8 port, 235 struct ib_gid_table *table, int ix, 236 const union ib_gid *gid, 237 const struct ib_gid_attr *attr, 238 bool default_gid) { 239 return write_gid(ib_dev, port, table, ix, gid, attr, 240 GID_TABLE_WRITE_ACTION_ADD, default_gid); 241 } 242 243 static int modify_gid(struct ib_device *ib_dev, u8 port, 244 struct ib_gid_table *table, int ix, 245 const union ib_gid *gid, 246 const struct ib_gid_attr *attr, 247 bool default_gid) { 248 return write_gid(ib_dev, port, table, ix, gid, attr, 249 GID_TABLE_WRITE_ACTION_MODIFY, default_gid); 250 } 251 252 static int del_gid(struct ib_device *ib_dev, u8 port, 253 struct ib_gid_table *table, int ix, 254 bool default_gid) { 255 return write_gid(ib_dev, port, table, ix, &zgid, &zattr, 256 GID_TABLE_WRITE_ACTION_DEL, default_gid); 257 } 258 259 /* rwlock should be read locked */ 260 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid, 261 const struct ib_gid_attr *val, bool default_gid, 262 unsigned long mask, int *pempty) 263 { 264 int i = 0; 265 int found = -1; 266 int empty = pempty ? -1 : 0; 267 268 while (i < table->sz && (found < 0 || empty < 0)) { 269 struct ib_gid_table_entry *data = &table->data_vec[i]; 270 struct ib_gid_attr *attr = &data->attr; 271 int curr_index = i; 272 273 i++; 274 275 if (data->props & GID_TABLE_ENTRY_INVALID) 276 continue; 277 278 if (empty < 0) 279 if (!memcmp(&data->gid, &zgid, sizeof(*gid)) && 280 !memcmp(attr, &zattr, sizeof(*attr)) && 281 !data->props) 282 empty = curr_index; 283 284 if (found >= 0) 285 continue; 286 287 if (mask & GID_ATTR_FIND_MASK_GID_TYPE && 288 attr->gid_type != val->gid_type) 289 continue; 290 291 if (mask & GID_ATTR_FIND_MASK_GID && 292 memcmp(gid, &data->gid, sizeof(*gid))) 293 continue; 294 295 if (mask & GID_ATTR_FIND_MASK_NETDEV && 296 attr->ndev != val->ndev) 297 continue; 298 299 if (mask & GID_ATTR_FIND_MASK_DEFAULT && 300 !!(data->props & GID_TABLE_ENTRY_DEFAULT) != 301 default_gid) 302 continue; 303 304 found = curr_index; 305 } 306 307 if (pempty) 308 *pempty = empty; 309 310 return found; 311 } 312 313 static void addrconf_ifid_eui48(u8 *eui, struct net_device *dev) 314 { 315 if (dev->if_addrlen != ETH_ALEN) 316 return; 317 memcpy(eui, IF_LLADDR(dev), 3); 318 memcpy(eui + 5, IF_LLADDR(dev) + 3, 3); 319 320 /* NOTE: The scope ID is added by the GID to IP conversion */ 321 322 eui[3] = 0xFF; 323 eui[4] = 0xFE; 324 eui[0] ^= 2; 325 } 326 327 static void make_default_gid(struct net_device *dev, union ib_gid *gid) 328 { 329 gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL); 330 addrconf_ifid_eui48(&gid->raw[8], dev); 331 } 332 333 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port, 334 union ib_gid *gid, struct ib_gid_attr *attr) 335 { 336 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 337 struct ib_gid_table *table; 338 int ix; 339 int ret = 0; 340 int empty; 341 342 table = ports_table[port - rdma_start_port(ib_dev)]; 343 344 if (!memcmp(gid, &zgid, sizeof(*gid))) 345 return -EINVAL; 346 347 mutex_lock(&table->lock); 348 write_lock_irq(&table->rwlock); 349 350 ix = find_gid(table, gid, attr, false, GID_ATTR_FIND_MASK_GID | 351 GID_ATTR_FIND_MASK_GID_TYPE | 352 GID_ATTR_FIND_MASK_NETDEV, &empty); 353 if (ix >= 0) 354 goto out_unlock; 355 356 if (empty < 0) { 357 ret = -ENOSPC; 358 goto out_unlock; 359 } 360 361 ret = add_gid(ib_dev, port, table, empty, gid, attr, false); 362 if (!ret) 363 dispatch_gid_change_event(ib_dev, port); 364 365 out_unlock: 366 write_unlock_irq(&table->rwlock); 367 mutex_unlock(&table->lock); 368 return ret; 369 } 370 371 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port, 372 union ib_gid *gid, struct ib_gid_attr *attr) 373 { 374 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 375 struct ib_gid_table *table; 376 int ix; 377 378 table = ports_table[port - rdma_start_port(ib_dev)]; 379 380 mutex_lock(&table->lock); 381 write_lock_irq(&table->rwlock); 382 383 ix = find_gid(table, gid, attr, false, 384 GID_ATTR_FIND_MASK_GID | 385 GID_ATTR_FIND_MASK_GID_TYPE | 386 GID_ATTR_FIND_MASK_NETDEV | 387 GID_ATTR_FIND_MASK_DEFAULT, 388 NULL); 389 if (ix < 0) 390 goto out_unlock; 391 392 if (!del_gid(ib_dev, port, table, ix, false)) 393 dispatch_gid_change_event(ib_dev, port); 394 395 out_unlock: 396 write_unlock_irq(&table->rwlock); 397 mutex_unlock(&table->lock); 398 return 0; 399 } 400 401 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port, 402 struct net_device *ndev) 403 { 404 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 405 struct ib_gid_table *table; 406 int ix; 407 bool deleted = false; 408 409 table = ports_table[port - rdma_start_port(ib_dev)]; 410 411 mutex_lock(&table->lock); 412 write_lock_irq(&table->rwlock); 413 414 for (ix = 0; ix < table->sz; ix++) 415 if (table->data_vec[ix].attr.ndev == ndev) 416 if (!del_gid(ib_dev, port, table, ix, 417 !!(table->data_vec[ix].props & 418 GID_TABLE_ENTRY_DEFAULT))) 419 deleted = true; 420 421 write_unlock_irq(&table->rwlock); 422 mutex_unlock(&table->lock); 423 424 if (deleted) 425 dispatch_gid_change_event(ib_dev, port); 426 427 return 0; 428 } 429 430 static int __ib_cache_gid_get(struct ib_device *ib_dev, u8 port, int index, 431 union ib_gid *gid, struct ib_gid_attr *attr) 432 { 433 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 434 struct ib_gid_table *table; 435 436 table = ports_table[port - rdma_start_port(ib_dev)]; 437 438 if (index < 0 || index >= table->sz) 439 return -EINVAL; 440 441 if (table->data_vec[index].props & GID_TABLE_ENTRY_INVALID) 442 return -EAGAIN; 443 444 memcpy(gid, &table->data_vec[index].gid, sizeof(*gid)); 445 if (attr) { 446 memcpy(attr, &table->data_vec[index].attr, sizeof(*attr)); 447 /* make sure network device is valid and attached */ 448 if (attr->ndev != NULL && 449 (attr->ndev->if_flags & IFF_DYING) == 0 && 450 attr->ndev->if_addr != NULL) 451 dev_hold(attr->ndev); 452 else 453 attr->ndev = NULL; 454 } 455 456 return 0; 457 } 458 459 static int _ib_cache_gid_table_find(struct ib_device *ib_dev, 460 const union ib_gid *gid, 461 const struct ib_gid_attr *val, 462 unsigned long mask, 463 u8 *port, u16 *index) 464 { 465 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 466 struct ib_gid_table *table; 467 u8 p; 468 int local_index; 469 unsigned long flags; 470 471 for (p = 0; p < ib_dev->phys_port_cnt; p++) { 472 table = ports_table[p]; 473 read_lock_irqsave(&table->rwlock, flags); 474 local_index = find_gid(table, gid, val, false, mask, NULL); 475 if (local_index >= 0) { 476 if (index) 477 *index = local_index; 478 if (port) 479 *port = p + rdma_start_port(ib_dev); 480 read_unlock_irqrestore(&table->rwlock, flags); 481 return 0; 482 } 483 read_unlock_irqrestore(&table->rwlock, flags); 484 } 485 486 return -ENOENT; 487 } 488 489 static int ib_cache_gid_find(struct ib_device *ib_dev, 490 const union ib_gid *gid, 491 enum ib_gid_type gid_type, 492 struct net_device *ndev, u8 *port, 493 u16 *index) 494 { 495 unsigned long mask = GID_ATTR_FIND_MASK_GID | 496 GID_ATTR_FIND_MASK_GID_TYPE; 497 struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type}; 498 499 if (ndev) 500 mask |= GID_ATTR_FIND_MASK_NETDEV; 501 502 return _ib_cache_gid_table_find(ib_dev, gid, &gid_attr_val, 503 mask, port, index); 504 } 505 506 int ib_find_cached_gid_by_port(struct ib_device *ib_dev, 507 const union ib_gid *gid, 508 enum ib_gid_type gid_type, 509 u8 port, struct net_device *ndev, 510 u16 *index) 511 { 512 int local_index; 513 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 514 struct ib_gid_table *table; 515 unsigned long mask = GID_ATTR_FIND_MASK_GID | 516 GID_ATTR_FIND_MASK_GID_TYPE; 517 struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type}; 518 unsigned long flags; 519 520 if (port < rdma_start_port(ib_dev) || 521 port > rdma_end_port(ib_dev)) 522 return -ENOENT; 523 524 table = ports_table[port - rdma_start_port(ib_dev)]; 525 526 if (ndev) 527 mask |= GID_ATTR_FIND_MASK_NETDEV; 528 529 read_lock_irqsave(&table->rwlock, flags); 530 local_index = find_gid(table, gid, &val, false, mask, NULL); 531 if (local_index >= 0) { 532 if (index) 533 *index = local_index; 534 read_unlock_irqrestore(&table->rwlock, flags); 535 return 0; 536 } 537 538 read_unlock_irqrestore(&table->rwlock, flags); 539 return -ENOENT; 540 } 541 EXPORT_SYMBOL(ib_find_cached_gid_by_port); 542 543 /** 544 * ib_find_gid_by_filter - Returns the GID table index where a specified 545 * GID value occurs 546 * @device: The device to query. 547 * @gid: The GID value to search for. 548 * @port_num: The port number of the device where the GID value could be 549 * searched. 550 * @filter: The filter function is executed on any matching GID in the table. 551 * If the filter function returns true, the corresponding index is returned, 552 * otherwise, we continue searching the GID table. It's guaranteed that 553 * while filter is executed, ndev field is valid and the structure won't 554 * change. filter is executed in an atomic context. filter must not be NULL. 555 * @index: The index into the cached GID table where the GID was found. This 556 * parameter may be NULL. 557 * 558 * ib_cache_gid_find_by_filter() searches for the specified GID value 559 * of which the filter function returns true in the port's GID table. 560 * This function is only supported on RoCE ports. 561 * 562 */ 563 static int ib_cache_gid_find_by_filter(struct ib_device *ib_dev, 564 const union ib_gid *gid, 565 u8 port, 566 bool (*filter)(const union ib_gid *, 567 const struct ib_gid_attr *, 568 void *), 569 void *context, 570 u16 *index) 571 { 572 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 573 struct ib_gid_table *table; 574 unsigned int i; 575 unsigned long flags; 576 bool found = false; 577 578 if (!ports_table) 579 return -EOPNOTSUPP; 580 581 if (port < rdma_start_port(ib_dev) || 582 port > rdma_end_port(ib_dev) || 583 !rdma_protocol_roce(ib_dev, port)) 584 return -EPROTONOSUPPORT; 585 586 table = ports_table[port - rdma_start_port(ib_dev)]; 587 588 read_lock_irqsave(&table->rwlock, flags); 589 for (i = 0; i < table->sz; i++) { 590 struct ib_gid_attr attr; 591 592 if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID) 593 goto next; 594 595 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid))) 596 goto next; 597 598 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr)); 599 600 if (filter(gid, &attr, context)) 601 found = true; 602 603 next: 604 if (found) 605 break; 606 } 607 read_unlock_irqrestore(&table->rwlock, flags); 608 609 if (!found) 610 return -ENOENT; 611 612 if (index) 613 *index = i; 614 return 0; 615 } 616 617 static struct ib_gid_table *alloc_gid_table(int sz) 618 { 619 struct ib_gid_table *table = 620 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL); 621 622 if (!table) 623 return NULL; 624 625 table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL); 626 if (!table->data_vec) 627 goto err_free_table; 628 629 mutex_init(&table->lock); 630 631 table->sz = sz; 632 rwlock_init(&table->rwlock); 633 634 return table; 635 636 err_free_table: 637 kfree(table); 638 return NULL; 639 } 640 641 static void release_gid_table(struct ib_gid_table *table) 642 { 643 if (table) { 644 kfree(table->data_vec); 645 kfree(table); 646 } 647 } 648 649 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port, 650 struct ib_gid_table *table) 651 { 652 int i; 653 bool deleted = false; 654 655 if (!table) 656 return; 657 658 write_lock_irq(&table->rwlock); 659 for (i = 0; i < table->sz; ++i) { 660 if (memcmp(&table->data_vec[i].gid, &zgid, 661 sizeof(table->data_vec[i].gid))) 662 if (!del_gid(ib_dev, port, table, i, 663 table->data_vec[i].props & 664 GID_ATTR_FIND_MASK_DEFAULT)) 665 deleted = true; 666 } 667 write_unlock_irq(&table->rwlock); 668 669 if (deleted) 670 dispatch_gid_change_event(ib_dev, port); 671 } 672 673 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port, 674 struct net_device *ndev, 675 unsigned long gid_type_mask, 676 enum ib_cache_gid_default_mode mode) 677 { 678 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache; 679 union ib_gid gid; 680 struct ib_gid_attr gid_attr; 681 struct ib_gid_attr zattr_type = zattr; 682 struct ib_gid_table *table; 683 unsigned int gid_type; 684 685 table = ports_table[port - rdma_start_port(ib_dev)]; 686 687 make_default_gid(ndev, &gid); 688 memset(&gid_attr, 0, sizeof(gid_attr)); 689 gid_attr.ndev = ndev; 690 691 /* Default GID is created using unique GUID and local subnet prefix, 692 * as described in section 4.1.1 and 3.5.10 in IB spec 1.3. 693 * Therefore don't create RoCEv2 default GID based on it that 694 * resembles as IPv6 GID based on link local address when IPv6 is 695 * disabled in kernel. 696 */ 697 #ifndef INET6 698 gid_type_mask &= ~BIT(IB_GID_TYPE_ROCE_UDP_ENCAP); 699 #endif 700 701 for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) { 702 int ix; 703 union ib_gid current_gid; 704 struct ib_gid_attr current_gid_attr = {}; 705 706 if (1UL << gid_type & ~gid_type_mask) 707 continue; 708 709 gid_attr.gid_type = gid_type; 710 711 mutex_lock(&table->lock); 712 write_lock_irq(&table->rwlock); 713 ix = find_gid(table, NULL, &gid_attr, true, 714 GID_ATTR_FIND_MASK_GID_TYPE | 715 GID_ATTR_FIND_MASK_DEFAULT, 716 NULL); 717 718 /* Coudn't find default GID location */ 719 if (WARN_ON(ix < 0)) 720 goto release; 721 722 zattr_type.gid_type = gid_type; 723 724 if (!__ib_cache_gid_get(ib_dev, port, ix, 725 ¤t_gid, ¤t_gid_attr) && 726 mode == IB_CACHE_GID_DEFAULT_MODE_SET && 727 !memcmp(&gid, ¤t_gid, sizeof(gid)) && 728 !memcmp(&gid_attr, ¤t_gid_attr, sizeof(gid_attr))) 729 goto release; 730 731 if (memcmp(¤t_gid, &zgid, sizeof(current_gid)) || 732 memcmp(¤t_gid_attr, &zattr_type, 733 sizeof(current_gid_attr))) { 734 if (del_gid(ib_dev, port, table, ix, true)) { 735 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n", 736 ix, gid.raw); 737 goto release; 738 } else { 739 dispatch_gid_change_event(ib_dev, port); 740 } 741 } 742 743 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) { 744 if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true)) 745 pr_warn("ib_cache_gid: unable to add default gid %pI6\n", 746 gid.raw); 747 else 748 dispatch_gid_change_event(ib_dev, port); 749 } 750 751 release: 752 if (current_gid_attr.ndev) 753 dev_put(current_gid_attr.ndev); 754 write_unlock_irq(&table->rwlock); 755 mutex_unlock(&table->lock); 756 } 757 } 758 759 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port, 760 struct ib_gid_table *table) 761 { 762 unsigned int i; 763 unsigned long roce_gid_type_mask; 764 unsigned int num_default_gids; 765 unsigned int current_gid = 0; 766 767 roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port); 768 num_default_gids = hweight_long(roce_gid_type_mask); 769 for (i = 0; i < num_default_gids && i < table->sz; i++) { 770 struct ib_gid_table_entry *entry = 771 &table->data_vec[i]; 772 773 entry->props |= GID_TABLE_ENTRY_DEFAULT; 774 current_gid = find_next_bit(&roce_gid_type_mask, 775 BITS_PER_LONG, 776 current_gid); 777 entry->attr.gid_type = current_gid++; 778 } 779 780 return 0; 781 } 782 783 static int _gid_table_setup_one(struct ib_device *ib_dev) 784 { 785 u8 port; 786 struct ib_gid_table **table; 787 int err = 0; 788 789 table = kcalloc(ib_dev->phys_port_cnt, sizeof(*table), GFP_KERNEL); 790 791 if (!table) { 792 pr_warn("failed to allocate ib gid cache for %s\n", 793 ib_dev->name); 794 return -ENOMEM; 795 } 796 797 for (port = 0; port < ib_dev->phys_port_cnt; port++) { 798 u8 rdma_port = port + rdma_start_port(ib_dev); 799 800 table[port] = 801 alloc_gid_table( 802 ib_dev->port_immutable[rdma_port].gid_tbl_len); 803 if (!table[port]) { 804 err = -ENOMEM; 805 goto rollback_table_setup; 806 } 807 808 err = gid_table_reserve_default(ib_dev, 809 port + rdma_start_port(ib_dev), 810 table[port]); 811 if (err) 812 goto rollback_table_setup; 813 } 814 815 ib_dev->cache.gid_cache = table; 816 return 0; 817 818 rollback_table_setup: 819 for (port = 0; port < ib_dev->phys_port_cnt; port++) { 820 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev), 821 table[port]); 822 release_gid_table(table[port]); 823 } 824 825 kfree(table); 826 return err; 827 } 828 829 static void gid_table_release_one(struct ib_device *ib_dev) 830 { 831 struct ib_gid_table **table = ib_dev->cache.gid_cache; 832 u8 port; 833 834 if (!table) 835 return; 836 837 for (port = 0; port < ib_dev->phys_port_cnt; port++) 838 release_gid_table(table[port]); 839 840 kfree(table); 841 ib_dev->cache.gid_cache = NULL; 842 } 843 844 static void gid_table_cleanup_one(struct ib_device *ib_dev) 845 { 846 struct ib_gid_table **table = ib_dev->cache.gid_cache; 847 u8 port; 848 849 if (!table) 850 return; 851 852 for (port = 0; port < ib_dev->phys_port_cnt; port++) 853 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev), 854 table[port]); 855 } 856 857 static int gid_table_setup_one(struct ib_device *ib_dev) 858 { 859 int err; 860 861 err = _gid_table_setup_one(ib_dev); 862 863 if (err) 864 return err; 865 866 err = roce_rescan_device(ib_dev); 867 868 if (err) { 869 gid_table_cleanup_one(ib_dev); 870 gid_table_release_one(ib_dev); 871 } 872 873 return err; 874 } 875 876 int ib_get_cached_gid(struct ib_device *device, 877 u8 port_num, 878 int index, 879 union ib_gid *gid, 880 struct ib_gid_attr *gid_attr) 881 { 882 int res; 883 unsigned long flags; 884 struct ib_gid_table **ports_table = device->cache.gid_cache; 885 struct ib_gid_table *table = ports_table[port_num - rdma_start_port(device)]; 886 887 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device)) 888 return -EINVAL; 889 890 read_lock_irqsave(&table->rwlock, flags); 891 res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr); 892 read_unlock_irqrestore(&table->rwlock, flags); 893 894 return res; 895 } 896 EXPORT_SYMBOL(ib_get_cached_gid); 897 898 int ib_find_cached_gid(struct ib_device *device, 899 const union ib_gid *gid, 900 enum ib_gid_type gid_type, 901 struct net_device *ndev, 902 u8 *port_num, 903 u16 *index) 904 { 905 return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index); 906 } 907 EXPORT_SYMBOL(ib_find_cached_gid); 908 909 int ib_find_gid_by_filter(struct ib_device *device, 910 const union ib_gid *gid, 911 u8 port_num, 912 bool (*filter)(const union ib_gid *gid, 913 const struct ib_gid_attr *, 914 void *), 915 void *context, u16 *index) 916 { 917 /* Only RoCE GID table supports filter function */ 918 if (!rdma_cap_roce_gid_table(device, port_num) && filter) 919 return -EPROTONOSUPPORT; 920 921 return ib_cache_gid_find_by_filter(device, gid, 922 port_num, filter, 923 context, index); 924 } 925 EXPORT_SYMBOL(ib_find_gid_by_filter); 926 927 int ib_get_cached_pkey(struct ib_device *device, 928 u8 port_num, 929 int index, 930 u16 *pkey) 931 { 932 struct ib_pkey_cache *cache; 933 unsigned long flags; 934 int ret = 0; 935 936 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device)) 937 return -EINVAL; 938 939 read_lock_irqsave(&device->cache.lock, flags); 940 941 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)]; 942 943 if (index < 0 || index >= cache->table_len) 944 ret = -EINVAL; 945 else 946 *pkey = cache->table[index]; 947 948 read_unlock_irqrestore(&device->cache.lock, flags); 949 950 return ret; 951 } 952 EXPORT_SYMBOL(ib_get_cached_pkey); 953 954 int ib_find_cached_pkey(struct ib_device *device, 955 u8 port_num, 956 u16 pkey, 957 u16 *index) 958 { 959 struct ib_pkey_cache *cache; 960 unsigned long flags; 961 int i; 962 int ret = -ENOENT; 963 int partial_ix = -1; 964 965 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device)) 966 return -EINVAL; 967 968 read_lock_irqsave(&device->cache.lock, flags); 969 970 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)]; 971 972 *index = -1; 973 974 for (i = 0; i < cache->table_len; ++i) 975 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) { 976 if (cache->table[i] & 0x8000) { 977 *index = i; 978 ret = 0; 979 break; 980 } else 981 partial_ix = i; 982 } 983 984 if (ret && partial_ix >= 0) { 985 *index = partial_ix; 986 ret = 0; 987 } 988 989 read_unlock_irqrestore(&device->cache.lock, flags); 990 991 return ret; 992 } 993 EXPORT_SYMBOL(ib_find_cached_pkey); 994 995 int ib_find_exact_cached_pkey(struct ib_device *device, 996 u8 port_num, 997 u16 pkey, 998 u16 *index) 999 { 1000 struct ib_pkey_cache *cache; 1001 unsigned long flags; 1002 int i; 1003 int ret = -ENOENT; 1004 1005 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device)) 1006 return -EINVAL; 1007 1008 read_lock_irqsave(&device->cache.lock, flags); 1009 1010 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)]; 1011 1012 *index = -1; 1013 1014 for (i = 0; i < cache->table_len; ++i) 1015 if (cache->table[i] == pkey) { 1016 *index = i; 1017 ret = 0; 1018 break; 1019 } 1020 1021 read_unlock_irqrestore(&device->cache.lock, flags); 1022 1023 return ret; 1024 } 1025 EXPORT_SYMBOL(ib_find_exact_cached_pkey); 1026 1027 int ib_get_cached_lmc(struct ib_device *device, 1028 u8 port_num, 1029 u8 *lmc) 1030 { 1031 unsigned long flags; 1032 int ret = 0; 1033 1034 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device)) 1035 return -EINVAL; 1036 1037 read_lock_irqsave(&device->cache.lock, flags); 1038 *lmc = device->cache.lmc_cache[port_num - rdma_start_port(device)]; 1039 read_unlock_irqrestore(&device->cache.lock, flags); 1040 1041 return ret; 1042 } 1043 EXPORT_SYMBOL(ib_get_cached_lmc); 1044 1045 static void ib_cache_update(struct ib_device *device, 1046 u8 port) 1047 { 1048 struct ib_port_attr *tprops = NULL; 1049 struct ib_pkey_cache *pkey_cache = NULL, *old_pkey_cache; 1050 struct ib_gid_cache { 1051 int table_len; 1052 union ib_gid table[0]; 1053 } *gid_cache = NULL; 1054 int i; 1055 int ret; 1056 struct ib_gid_table *table; 1057 struct ib_gid_table **ports_table = device->cache.gid_cache; 1058 bool use_roce_gid_table = 1059 rdma_cap_roce_gid_table(device, port); 1060 1061 if (port < rdma_start_port(device) || port > rdma_end_port(device)) 1062 return; 1063 1064 table = ports_table[port - rdma_start_port(device)]; 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.pkey_cache[port - rdma_start_port(device)]; 1117 1118 device->cache.pkey_cache[port - rdma_start_port(device)] = pkey_cache; 1119 if (!use_roce_gid_table) { 1120 write_lock(&table->rwlock); 1121 for (i = 0; i < gid_cache->table_len; i++) { 1122 modify_gid(device, port, table, i, gid_cache->table + i, 1123 &zattr, false); 1124 } 1125 write_unlock(&table->rwlock); 1126 } 1127 1128 device->cache.lmc_cache[port - rdma_start_port(device)] = tprops->lmc; 1129 1130 write_unlock_irq(&device->cache.lock); 1131 1132 kfree(gid_cache); 1133 kfree(old_pkey_cache); 1134 kfree(tprops); 1135 return; 1136 1137 err: 1138 kfree(pkey_cache); 1139 kfree(gid_cache); 1140 kfree(tprops); 1141 } 1142 1143 static void ib_cache_task(struct work_struct *_work) 1144 { 1145 struct ib_update_work *work = 1146 container_of(_work, struct ib_update_work, work); 1147 1148 ib_cache_update(work->device, work->port_num); 1149 kfree(work); 1150 } 1151 1152 static void ib_cache_event(struct ib_event_handler *handler, 1153 struct ib_event *event) 1154 { 1155 struct ib_update_work *work; 1156 1157 if (event->event == IB_EVENT_PORT_ERR || 1158 event->event == IB_EVENT_PORT_ACTIVE || 1159 event->event == IB_EVENT_LID_CHANGE || 1160 event->event == IB_EVENT_PKEY_CHANGE || 1161 event->event == IB_EVENT_SM_CHANGE || 1162 event->event == IB_EVENT_CLIENT_REREGISTER || 1163 event->event == IB_EVENT_GID_CHANGE) { 1164 work = kmalloc(sizeof *work, GFP_ATOMIC); 1165 if (work) { 1166 INIT_WORK(&work->work, ib_cache_task); 1167 work->device = event->device; 1168 work->port_num = event->element.port_num; 1169 queue_work(ib_wq, &work->work); 1170 } 1171 } 1172 } 1173 1174 int ib_cache_setup_one(struct ib_device *device) 1175 { 1176 int p; 1177 int err; 1178 1179 rwlock_init(&device->cache.lock); 1180 1181 device->cache.pkey_cache = 1182 kzalloc(sizeof *device->cache.pkey_cache * 1183 (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL); 1184 device->cache.lmc_cache = kmalloc(sizeof *device->cache.lmc_cache * 1185 (rdma_end_port(device) - 1186 rdma_start_port(device) + 1), 1187 GFP_KERNEL); 1188 if (!device->cache.pkey_cache || 1189 !device->cache.lmc_cache) { 1190 pr_warn("Couldn't allocate cache for %s\n", device->name); 1191 return -ENOMEM; 1192 } 1193 1194 err = gid_table_setup_one(device); 1195 if (err) 1196 /* Allocated memory will be cleaned in the release function */ 1197 return err; 1198 1199 for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p) 1200 ib_cache_update(device, p + rdma_start_port(device)); 1201 1202 INIT_IB_EVENT_HANDLER(&device->cache.event_handler, 1203 device, ib_cache_event); 1204 err = ib_register_event_handler(&device->cache.event_handler); 1205 if (err) 1206 goto err; 1207 1208 return 0; 1209 1210 err: 1211 gid_table_cleanup_one(device); 1212 return err; 1213 } 1214 1215 void ib_cache_release_one(struct ib_device *device) 1216 { 1217 int p; 1218 1219 /* 1220 * The release function frees all the cache elements. 1221 * This function should be called as part of freeing 1222 * all the device's resources when the cache could no 1223 * longer be accessed. 1224 */ 1225 if (device->cache.pkey_cache) 1226 for (p = 0; 1227 p <= rdma_end_port(device) - rdma_start_port(device); ++p) 1228 kfree(device->cache.pkey_cache[p]); 1229 1230 gid_table_release_one(device); 1231 kfree(device->cache.pkey_cache); 1232 kfree(device->cache.lmc_cache); 1233 } 1234 1235 void ib_cache_cleanup_one(struct ib_device *device) 1236 { 1237 /* The cleanup function unregisters the event handler, 1238 * waits for all in-progress workqueue elements and cleans 1239 * up the GID cache. This function should be called after 1240 * the device was removed from the devices list and all 1241 * clients were removed, so the cache exists but is 1242 * non-functional and shouldn't be updated anymore. 1243 */ 1244 ib_unregister_event_handler(&device->cache.event_handler); 1245 flush_workqueue(ib_wq); 1246 gid_table_cleanup_one(device); 1247 } 1248 1249 void __init ib_cache_setup(void) 1250 { 1251 roce_gid_mgmt_init(); 1252 } 1253 1254 void __exit ib_cache_cleanup(void) 1255 { 1256 roce_gid_mgmt_cleanup(); 1257 } 1258