1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * net-sysfs.c - network device class and attributes 4 * 5 * Copyright (c) 2003 Stephen Hemminger <shemminger@osdl.org> 6 */ 7 8 #include <linux/capability.h> 9 #include <linux/kernel.h> 10 #include <linux/netdevice.h> 11 #include <linux/if_arp.h> 12 #include <linux/slab.h> 13 #include <linux/sched/signal.h> 14 #include <linux/sched/isolation.h> 15 #include <linux/nsproxy.h> 16 #include <net/sock.h> 17 #include <net/net_namespace.h> 18 #include <linux/rtnetlink.h> 19 #include <linux/vmalloc.h> 20 #include <linux/export.h> 21 #include <linux/jiffies.h> 22 #include <linux/pm_runtime.h> 23 #include <linux/of.h> 24 #include <linux/of_net.h> 25 #include <linux/cpu.h> 26 #include <net/netdev_lock.h> 27 #include <net/netdev_rx_queue.h> 28 #include <net/rps.h> 29 30 #include "dev.h" 31 #include "net-sysfs.h" 32 33 #ifdef CONFIG_SYSFS 34 static const char fmt_hex[] = "%#x\n"; 35 static const char fmt_dec[] = "%d\n"; 36 static const char fmt_uint[] = "%u\n"; 37 static const char fmt_ulong[] = "%lu\n"; 38 static const char fmt_u64[] = "%llu\n"; 39 40 /* There is a possible ABBA deadlock between rtnl_lock and kernfs_node->active, 41 * when unregistering a net device and accessing associated sysfs files. The 42 * potential deadlock is as follow: 43 * 44 * CPU 0 CPU 1 45 * 46 * rtnl_lock vfs_read 47 * unregister_netdevice_many kernfs_seq_start 48 * device_del / kobject_put kernfs_get_active (kn->active++) 49 * kernfs_drain sysfs_kf_seq_show 50 * wait_event( rtnl_lock 51 * kn->active == KN_DEACTIVATED_BIAS) -> waits on CPU 0 to release 52 * -> waits on CPU 1 to decrease kn->active the rtnl lock. 53 * 54 * The historical fix was to use rtnl_trylock with restart_syscall to bail out 55 * of sysfs operations when the lock couldn't be taken. This fixed the above 56 * issue as it allowed CPU 1 to bail out of the ABBA situation. 57 * 58 * But it came with performances issues, as syscalls are being restarted in 59 * loops when there was contention on the rtnl lock, with huge slow downs in 60 * specific scenarios (e.g. lots of virtual interfaces created and userspace 61 * daemons querying their attributes). 62 * 63 * The idea below is to bail out of the active kernfs_node protection 64 * (kn->active) while trying to take the rtnl lock. 65 * 66 * This replaces rtnl_lock() and still has to be used with rtnl_unlock(). The 67 * net device is guaranteed to be alive if this returns successfully. 68 */ 69 static int sysfs_rtnl_lock(struct kobject *kobj, struct attribute *attr, 70 struct net_device *ndev) 71 { 72 struct kernfs_node *kn; 73 int ret = 0; 74 75 /* First, we hold a reference to the net device as the unregistration 76 * path might run in parallel. This will ensure the net device and the 77 * associated sysfs objects won't be freed while we try to take the rtnl 78 * lock. 79 */ 80 dev_hold(ndev); 81 /* sysfs_break_active_protection was introduced to allow self-removal of 82 * devices and their associated sysfs files by bailing out of the 83 * sysfs/kernfs protection. We do this here to allow the unregistration 84 * path to complete in parallel. The following takes a reference on the 85 * kobject and the kernfs_node being accessed. 86 * 87 * This works because we hold a reference onto the net device and the 88 * unregistration path will wait for us eventually in netdev_run_todo 89 * (outside an rtnl lock section). 90 */ 91 kn = sysfs_break_active_protection(kobj, attr); 92 /* We can now try to take the rtnl lock. This can't deadlock us as the 93 * unregistration path is able to drain sysfs files (kernfs_node) thanks 94 * to the above dance. 95 */ 96 if (rtnl_lock_interruptible()) { 97 ret = -ERESTARTSYS; 98 goto unbreak; 99 } 100 /* Check dismantle on the device hasn't started, otherwise deny the 101 * operation. 102 */ 103 if (!netif_is_alive(ndev)) { 104 rtnl_unlock(); 105 ret = -ENODEV; 106 goto unbreak; 107 } 108 /* We are now sure the device dismantle hasn't started nor that it can 109 * start before we exit the locking section as we hold the rtnl lock. 110 * There's no need to keep unbreaking the sysfs protection nor to hold 111 * a net device reference from that point; that was only needed to take 112 * the rtnl lock. 113 */ 114 unbreak: 115 sysfs_unbreak_active_protection(kn); 116 dev_put(ndev); 117 118 return ret; 119 } 120 121 static int sysfs_get_link_ksettings(struct device *dev, 122 struct device_attribute *attr, 123 struct ethtool_link_ksettings *cmd) 124 { 125 struct net_device *netdev = to_net_dev(dev); 126 bool need_rtnl; 127 int ret; 128 129 /* 130 * The check is also done in netif_get_link_ksettings; this helps 131 * returning early without hitting the locking section below. 132 */ 133 if (!netdev->ethtool_ops->get_link_ksettings) 134 return -EINVAL; 135 136 need_rtnl = !netdev_need_ops_lock(netdev) || 137 (netdev->ethtool_ops->op_needs_rtnl & 138 ETHTOOL_OP_NEEDS_RTNL_LINKSETTINGS); 139 if (need_rtnl) { 140 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 141 if (ret) 142 return ret; 143 } 144 netdev_lock_ops(netdev); 145 146 if (!netif_is_alive(netdev)) { 147 ret = -ENODEV; 148 goto unlock; 149 } 150 151 ret = -EINVAL; 152 if (netif_running(netdev)) { 153 if (!netif_get_link_ksettings(netdev, cmd)) 154 ret = 0; 155 } 156 157 unlock: 158 netdev_unlock_ops(netdev); 159 if (need_rtnl) 160 rtnl_unlock(); 161 return ret; 162 } 163 164 /* use same locking rules as GIF* ioctl's */ 165 static ssize_t netdev_show(const struct device *dev, 166 struct device_attribute *attr, char *buf, 167 ssize_t (*format)(const struct net_device *, char *)) 168 { 169 struct net_device *ndev = to_net_dev(dev); 170 ssize_t ret = -EINVAL; 171 172 rcu_read_lock(); 173 if (netif_is_alive(ndev)) 174 ret = (*format)(ndev, buf); 175 rcu_read_unlock(); 176 177 return ret; 178 } 179 180 /* generate a show function for simple field */ 181 #define NETDEVICE_SHOW(field, format_string) \ 182 static ssize_t format_##field(const struct net_device *dev, char *buf) \ 183 { \ 184 return sysfs_emit(buf, format_string, READ_ONCE(dev->field)); \ 185 } \ 186 static ssize_t field##_show(struct device *dev, \ 187 struct device_attribute *attr, char *buf) \ 188 { \ 189 return netdev_show(dev, attr, buf, format_##field); \ 190 } \ 191 192 #define NETDEVICE_SHOW_RO(field, format_string) \ 193 NETDEVICE_SHOW(field, format_string); \ 194 static DEVICE_ATTR_RO(field) 195 196 #define NETDEVICE_SHOW_RW(field, format_string) \ 197 NETDEVICE_SHOW(field, format_string); \ 198 static DEVICE_ATTR_RW(field) 199 200 /* use same locking and permission rules as SIF* ioctl's */ 201 static ssize_t netdev_store(struct device *dev, struct device_attribute *attr, 202 const char *buf, size_t len, 203 int (*set)(struct net_device *, unsigned long)) 204 { 205 struct net_device *netdev = to_net_dev(dev); 206 struct net *net = dev_net(netdev); 207 unsigned long new; 208 int ret; 209 210 if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) 211 return -EPERM; 212 213 ret = kstrtoul(buf, 0, &new); 214 if (ret) 215 goto err; 216 217 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 218 if (ret) 219 goto err; 220 221 ret = (*set)(netdev, new); 222 if (ret == 0) 223 ret = len; 224 225 rtnl_unlock(); 226 err: 227 return ret; 228 } 229 230 /* Same as netdev_store() but takes netdev_lock() instead of rtnl_lock() */ 231 static ssize_t 232 netdev_lock_store(struct device *dev, struct device_attribute *attr, 233 const char *buf, size_t len, 234 int (*set)(struct net_device *, unsigned long)) 235 { 236 struct net_device *netdev = to_net_dev(dev); 237 struct net *net = dev_net(netdev); 238 unsigned long new; 239 int ret; 240 241 if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) 242 return -EPERM; 243 244 ret = kstrtoul(buf, 0, &new); 245 if (ret) 246 return ret; 247 248 netdev_lock(netdev); 249 250 if (netif_is_alive(netdev)) { 251 ret = (*set)(netdev, new); 252 if (ret == 0) 253 ret = len; 254 } 255 netdev_unlock(netdev); 256 257 return ret; 258 } 259 260 NETDEVICE_SHOW_RO(dev_id, fmt_hex); 261 NETDEVICE_SHOW_RO(dev_port, fmt_dec); 262 NETDEVICE_SHOW_RO(addr_assign_type, fmt_dec); 263 NETDEVICE_SHOW_RO(addr_len, fmt_dec); 264 NETDEVICE_SHOW_RO(ifindex, fmt_dec); 265 NETDEVICE_SHOW_RO(type, fmt_dec); 266 NETDEVICE_SHOW_RO(link_mode, fmt_dec); 267 268 static ssize_t iflink_show(struct device *dev, struct device_attribute *attr, 269 char *buf) 270 { 271 struct net_device *ndev = to_net_dev(dev); 272 273 return sysfs_emit(buf, fmt_dec, dev_get_iflink(ndev)); 274 } 275 static DEVICE_ATTR_RO(iflink); 276 277 static ssize_t format_name_assign_type(const struct net_device *dev, char *buf) 278 { 279 return sysfs_emit(buf, fmt_dec, READ_ONCE(dev->name_assign_type)); 280 } 281 282 static ssize_t name_assign_type_show(struct device *dev, 283 struct device_attribute *attr, 284 char *buf) 285 { 286 struct net_device *ndev = to_net_dev(dev); 287 ssize_t ret = -EINVAL; 288 289 if (READ_ONCE(ndev->name_assign_type) != NET_NAME_UNKNOWN) 290 ret = netdev_show(dev, attr, buf, format_name_assign_type); 291 292 return ret; 293 } 294 static DEVICE_ATTR_RO(name_assign_type); 295 296 /* use same locking rules as GIFHWADDR ioctl's (netif_get_mac_address()) */ 297 static ssize_t address_show(struct device *dev, struct device_attribute *attr, 298 char *buf) 299 { 300 struct net_device *ndev = to_net_dev(dev); 301 ssize_t ret = -EINVAL; 302 303 down_read(&dev_addr_sem); 304 305 rcu_read_lock(); 306 if (netif_is_alive(ndev)) 307 ret = sysfs_format_mac(buf, ndev->dev_addr, ndev->addr_len); 308 rcu_read_unlock(); 309 310 up_read(&dev_addr_sem); 311 return ret; 312 } 313 static DEVICE_ATTR_RO(address); 314 315 static ssize_t broadcast_show(struct device *dev, 316 struct device_attribute *attr, char *buf) 317 { 318 struct net_device *ndev = to_net_dev(dev); 319 int ret = -EINVAL; 320 321 rcu_read_lock(); 322 if (netif_is_alive(ndev)) 323 ret = sysfs_format_mac(buf, ndev->broadcast, ndev->addr_len); 324 rcu_read_unlock(); 325 return ret; 326 } 327 static DEVICE_ATTR_RO(broadcast); 328 329 static int change_carrier(struct net_device *dev, unsigned long new_carrier) 330 { 331 if (!netif_running(dev)) 332 return -EINVAL; 333 return dev_change_carrier(dev, (bool)new_carrier); 334 } 335 336 static ssize_t carrier_store(struct device *dev, struct device_attribute *attr, 337 const char *buf, size_t len) 338 { 339 struct net_device *netdev = to_net_dev(dev); 340 341 /* The check is also done in change_carrier; this helps returning early 342 * without hitting the locking section in netdev_store. 343 */ 344 if (!netdev->netdev_ops->ndo_change_carrier) 345 return -EOPNOTSUPP; 346 347 return netdev_store(dev, attr, buf, len, change_carrier); 348 } 349 350 static ssize_t carrier_show(struct device *dev, 351 struct device_attribute *attr, char *buf) 352 { 353 struct net_device *netdev = to_net_dev(dev); 354 int ret; 355 356 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 357 if (ret) 358 return ret; 359 360 ret = -EINVAL; 361 if (netif_running(netdev)) { 362 /* Synchronize carrier state with link watch, 363 * see also rtnl_getlink(). 364 */ 365 linkwatch_sync_dev(netdev); 366 367 ret = sysfs_emit(buf, fmt_dec, !!netif_carrier_ok(netdev)); 368 } 369 370 rtnl_unlock(); 371 return ret; 372 } 373 static DEVICE_ATTR_RW(carrier); 374 375 static ssize_t speed_show(struct device *dev, 376 struct device_attribute *attr, char *buf) 377 { 378 struct ethtool_link_ksettings cmd; 379 int ret; 380 381 ret = sysfs_get_link_ksettings(dev, attr, &cmd); 382 if (ret) 383 return ret; 384 385 return sysfs_emit(buf, fmt_dec, cmd.base.speed); 386 } 387 static DEVICE_ATTR_RO(speed); 388 389 static ssize_t duplex_show(struct device *dev, 390 struct device_attribute *attr, char *buf) 391 { 392 struct ethtool_link_ksettings cmd; 393 const char *duplex; 394 int ret; 395 396 ret = sysfs_get_link_ksettings(dev, attr, &cmd); 397 if (ret) 398 return ret; 399 400 switch (cmd.base.duplex) { 401 case DUPLEX_HALF: 402 duplex = "half"; 403 break; 404 case DUPLEX_FULL: 405 duplex = "full"; 406 break; 407 default: 408 duplex = "unknown"; 409 break; 410 } 411 412 return sysfs_emit(buf, "%s\n", duplex); 413 } 414 static DEVICE_ATTR_RO(duplex); 415 416 static ssize_t testing_show(struct device *dev, 417 struct device_attribute *attr, char *buf) 418 { 419 struct net_device *netdev = to_net_dev(dev); 420 421 if (netif_running(netdev)) 422 return sysfs_emit(buf, fmt_dec, !!netif_testing(netdev)); 423 424 return -EINVAL; 425 } 426 static DEVICE_ATTR_RO(testing); 427 428 static ssize_t dormant_show(struct device *dev, 429 struct device_attribute *attr, char *buf) 430 { 431 struct net_device *netdev = to_net_dev(dev); 432 433 if (netif_running(netdev)) 434 return sysfs_emit(buf, fmt_dec, !!netif_dormant(netdev)); 435 436 return -EINVAL; 437 } 438 static DEVICE_ATTR_RO(dormant); 439 440 static const char *const operstates[] = { 441 "unknown", 442 "notpresent", /* currently unused */ 443 "down", 444 "lowerlayerdown", 445 "testing", 446 "dormant", 447 "up" 448 }; 449 450 static ssize_t operstate_show(struct device *dev, 451 struct device_attribute *attr, char *buf) 452 { 453 const struct net_device *netdev = to_net_dev(dev); 454 unsigned char operstate; 455 456 operstate = READ_ONCE(netdev->operstate); 457 if (!netif_running(netdev)) 458 operstate = IF_OPER_DOWN; 459 460 if (operstate >= ARRAY_SIZE(operstates)) 461 return -EINVAL; /* should not happen */ 462 463 return sysfs_emit(buf, "%s\n", operstates[operstate]); 464 } 465 static DEVICE_ATTR_RO(operstate); 466 467 static ssize_t carrier_changes_show(struct device *dev, 468 struct device_attribute *attr, 469 char *buf) 470 { 471 struct net_device *netdev = to_net_dev(dev); 472 473 return sysfs_emit(buf, fmt_dec, 474 atomic_read(&netdev->carrier_up_count) + 475 atomic_read(&netdev->carrier_down_count)); 476 } 477 static DEVICE_ATTR_RO(carrier_changes); 478 479 static ssize_t carrier_up_count_show(struct device *dev, 480 struct device_attribute *attr, 481 char *buf) 482 { 483 struct net_device *netdev = to_net_dev(dev); 484 485 return sysfs_emit(buf, fmt_dec, atomic_read(&netdev->carrier_up_count)); 486 } 487 static DEVICE_ATTR_RO(carrier_up_count); 488 489 static ssize_t carrier_down_count_show(struct device *dev, 490 struct device_attribute *attr, 491 char *buf) 492 { 493 struct net_device *netdev = to_net_dev(dev); 494 495 return sysfs_emit(buf, fmt_dec, atomic_read(&netdev->carrier_down_count)); 496 } 497 static DEVICE_ATTR_RO(carrier_down_count); 498 499 /* read-write attributes */ 500 501 static int change_mtu(struct net_device *dev, unsigned long new_mtu) 502 { 503 return dev_set_mtu(dev, (int)new_mtu); 504 } 505 506 static ssize_t mtu_store(struct device *dev, struct device_attribute *attr, 507 const char *buf, size_t len) 508 { 509 return netdev_store(dev, attr, buf, len, change_mtu); 510 } 511 NETDEVICE_SHOW_RW(mtu, fmt_dec); 512 513 static int change_flags(struct net_device *dev, unsigned long new_flags) 514 { 515 return dev_change_flags(dev, (unsigned int)new_flags, NULL); 516 } 517 518 static ssize_t flags_store(struct device *dev, struct device_attribute *attr, 519 const char *buf, size_t len) 520 { 521 return netdev_store(dev, attr, buf, len, change_flags); 522 } 523 NETDEVICE_SHOW_RW(flags, fmt_hex); 524 525 static ssize_t tx_queue_len_store(struct device *dev, 526 struct device_attribute *attr, 527 const char *buf, size_t len) 528 { 529 if (!capable(CAP_NET_ADMIN)) 530 return -EPERM; 531 532 return netdev_store(dev, attr, buf, len, dev_change_tx_queue_len); 533 } 534 NETDEVICE_SHOW_RW(tx_queue_len, fmt_dec); 535 536 static int change_gro_flush_timeout(struct net_device *dev, unsigned long val) 537 { 538 netdev_set_gro_flush_timeout(dev, val); 539 return 0; 540 } 541 542 static ssize_t gro_flush_timeout_store(struct device *dev, 543 struct device_attribute *attr, 544 const char *buf, size_t len) 545 { 546 if (!capable(CAP_NET_ADMIN)) 547 return -EPERM; 548 549 return netdev_lock_store(dev, attr, buf, len, change_gro_flush_timeout); 550 } 551 NETDEVICE_SHOW_RW(gro_flush_timeout, fmt_ulong); 552 553 static int change_napi_defer_hard_irqs(struct net_device *dev, unsigned long val) 554 { 555 if (val > S32_MAX) 556 return -ERANGE; 557 558 netdev_set_defer_hard_irqs(dev, (u32)val); 559 return 0; 560 } 561 562 static ssize_t napi_defer_hard_irqs_store(struct device *dev, 563 struct device_attribute *attr, 564 const char *buf, size_t len) 565 { 566 if (!capable(CAP_NET_ADMIN)) 567 return -EPERM; 568 569 return netdev_lock_store(dev, attr, buf, len, 570 change_napi_defer_hard_irqs); 571 } 572 NETDEVICE_SHOW_RW(napi_defer_hard_irqs, fmt_uint); 573 574 static ssize_t ifalias_store(struct device *dev, struct device_attribute *attr, 575 const char *buf, size_t len) 576 { 577 struct net_device *netdev = to_net_dev(dev); 578 struct net *net = dev_net(netdev); 579 size_t count = len; 580 ssize_t ret; 581 582 if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) 583 return -EPERM; 584 585 /* ignore trailing newline */ 586 if (len > 0 && buf[len - 1] == '\n') 587 --count; 588 589 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 590 if (ret) 591 return ret; 592 593 ret = dev_set_alias(netdev, buf, count); 594 if (ret < 0) 595 goto err; 596 ret = len; 597 netdev_state_change(netdev); 598 err: 599 rtnl_unlock(); 600 601 return ret; 602 } 603 604 static ssize_t ifalias_show(struct device *dev, 605 struct device_attribute *attr, char *buf) 606 { 607 const struct net_device *netdev = to_net_dev(dev); 608 char tmp[IFALIASZ]; 609 ssize_t ret; 610 611 ret = dev_get_alias(netdev, tmp, sizeof(tmp)); 612 if (ret > 0) 613 ret = sysfs_emit(buf, "%s\n", tmp); 614 return ret; 615 } 616 static DEVICE_ATTR_RW(ifalias); 617 618 static int change_group(struct net_device *dev, unsigned long new_group) 619 { 620 dev_set_group(dev, (int)new_group); 621 return 0; 622 } 623 624 static ssize_t group_store(struct device *dev, struct device_attribute *attr, 625 const char *buf, size_t len) 626 { 627 return netdev_store(dev, attr, buf, len, change_group); 628 } 629 NETDEVICE_SHOW(group, fmt_dec); 630 static DEVICE_ATTR(netdev_group, 0644, group_show, group_store); 631 632 static int change_proto_down(struct net_device *dev, unsigned long proto_down) 633 { 634 return dev_change_proto_down(dev, (bool)proto_down); 635 } 636 637 static ssize_t proto_down_store(struct device *dev, 638 struct device_attribute *attr, 639 const char *buf, size_t len) 640 { 641 return netdev_store(dev, attr, buf, len, change_proto_down); 642 } 643 NETDEVICE_SHOW_RW(proto_down, fmt_dec); 644 645 static ssize_t phys_port_id_show(struct device *dev, 646 struct device_attribute *attr, char *buf) 647 { 648 struct net_device *netdev = to_net_dev(dev); 649 struct netdev_phys_item_id ppid; 650 ssize_t ret; 651 652 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 653 if (ret) 654 return ret; 655 656 ret = dev_get_phys_port_id(netdev, &ppid); 657 if (!ret) 658 ret = sysfs_emit(buf, "%*phN\n", ppid.id_len, ppid.id); 659 660 rtnl_unlock(); 661 662 return ret; 663 } 664 static DEVICE_ATTR_RO(phys_port_id); 665 666 static ssize_t phys_port_name_show(struct device *dev, 667 struct device_attribute *attr, char *buf) 668 { 669 struct net_device *netdev = to_net_dev(dev); 670 char name[IFNAMSIZ]; 671 ssize_t ret; 672 673 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 674 if (ret) 675 return ret; 676 677 ret = dev_get_phys_port_name(netdev, name, sizeof(name)); 678 if (!ret) 679 ret = sysfs_emit(buf, "%s\n", name); 680 681 rtnl_unlock(); 682 683 return ret; 684 } 685 static DEVICE_ATTR_RO(phys_port_name); 686 687 static ssize_t phys_switch_id_show(struct device *dev, 688 struct device_attribute *attr, char *buf) 689 { 690 struct net_device *netdev = to_net_dev(dev); 691 struct netdev_phys_item_id ppid = { }; 692 ssize_t ret; 693 694 ret = sysfs_rtnl_lock(&dev->kobj, &attr->attr, netdev); 695 if (ret) 696 return ret; 697 698 ret = netif_get_port_parent_id(netdev, &ppid, false); 699 if (!ret) 700 ret = sysfs_emit(buf, "%*phN\n", ppid.id_len, ppid.id); 701 702 rtnl_unlock(); 703 704 return ret; 705 } 706 static DEVICE_ATTR_RO(phys_switch_id); 707 708 static struct attribute *netdev_phys_attrs[] __ro_after_init = { 709 &dev_attr_phys_port_id.attr, 710 &dev_attr_phys_port_name.attr, 711 &dev_attr_phys_switch_id.attr, 712 NULL, 713 }; 714 715 static umode_t netdev_phys_is_visible(struct kobject *kobj, 716 struct attribute *attr, int index) 717 { 718 struct device *dev = kobj_to_dev(kobj); 719 struct net_device *netdev = to_net_dev(dev); 720 721 if (attr == &dev_attr_phys_port_id.attr) { 722 if (!netdev->netdev_ops->ndo_get_phys_port_id) 723 return 0; 724 } else if (attr == &dev_attr_phys_port_name.attr) { 725 if (!netdev->netdev_ops->ndo_get_phys_port_name && 726 !netdev->devlink_port) 727 return 0; 728 } else if (attr == &dev_attr_phys_switch_id.attr) { 729 if (!netdev->netdev_ops->ndo_get_port_parent_id && 730 !netdev->devlink_port) 731 return 0; 732 } 733 734 return attr->mode; 735 } 736 737 static const struct attribute_group netdev_phys_group = { 738 .attrs = netdev_phys_attrs, 739 .is_visible = netdev_phys_is_visible, 740 }; 741 742 static ssize_t threaded_show(struct device *dev, 743 struct device_attribute *attr, char *buf) 744 { 745 struct net_device *netdev = to_net_dev(dev); 746 ssize_t ret = -EINVAL; 747 748 rcu_read_lock(); 749 750 if (netif_is_alive(netdev)) 751 ret = sysfs_emit(buf, fmt_dec, READ_ONCE(netdev->threaded)); 752 753 rcu_read_unlock(); 754 755 return ret; 756 } 757 758 static int modify_napi_threaded(struct net_device *dev, unsigned long val) 759 { 760 struct napi_struct *napi; 761 int ret; 762 763 if (list_empty(&dev->napi_list)) 764 return -EOPNOTSUPP; 765 766 if (val != 0 && val != 1) 767 return -EOPNOTSUPP; 768 769 if (val) { 770 list_for_each_entry(napi, &dev->napi_list, dev_list) { 771 if (test_bit(NAPI_STATE_PERCPU, &napi->state)) 772 return -EOPNOTSUPP; 773 } 774 } 775 776 ret = netif_set_threaded(dev, val); 777 778 return ret; 779 } 780 781 static ssize_t threaded_store(struct device *dev, 782 struct device_attribute *attr, 783 const char *buf, size_t len) 784 { 785 return netdev_lock_store(dev, attr, buf, len, modify_napi_threaded); 786 } 787 static DEVICE_ATTR_RW(threaded); 788 789 static struct attribute *net_class_attrs[] __ro_after_init = { 790 &dev_attr_netdev_group.attr, 791 &dev_attr_type.attr, 792 &dev_attr_dev_id.attr, 793 &dev_attr_dev_port.attr, 794 &dev_attr_iflink.attr, 795 &dev_attr_ifindex.attr, 796 &dev_attr_name_assign_type.attr, 797 &dev_attr_addr_assign_type.attr, 798 &dev_attr_addr_len.attr, 799 &dev_attr_link_mode.attr, 800 &dev_attr_address.attr, 801 &dev_attr_broadcast.attr, 802 &dev_attr_speed.attr, 803 &dev_attr_duplex.attr, 804 &dev_attr_dormant.attr, 805 &dev_attr_testing.attr, 806 &dev_attr_operstate.attr, 807 &dev_attr_carrier_changes.attr, 808 &dev_attr_ifalias.attr, 809 &dev_attr_carrier.attr, 810 &dev_attr_mtu.attr, 811 &dev_attr_flags.attr, 812 &dev_attr_tx_queue_len.attr, 813 &dev_attr_gro_flush_timeout.attr, 814 &dev_attr_napi_defer_hard_irqs.attr, 815 &dev_attr_proto_down.attr, 816 &dev_attr_carrier_up_count.attr, 817 &dev_attr_carrier_down_count.attr, 818 &dev_attr_threaded.attr, 819 NULL, 820 }; 821 ATTRIBUTE_GROUPS(net_class); 822 823 /* Show a given an attribute in the statistics group */ 824 static ssize_t netstat_show(const struct device *d, 825 struct device_attribute *attr, char *buf, 826 unsigned long offset) 827 { 828 struct net_device *dev = to_net_dev(d); 829 ssize_t ret = -EINVAL; 830 831 WARN_ON(offset > sizeof(struct rtnl_link_stats64) || 832 offset % sizeof(u64) != 0); 833 834 rcu_read_lock(); 835 if (netif_is_alive(dev)) { 836 struct rtnl_link_stats64 temp; 837 const struct rtnl_link_stats64 *stats = dev_get_stats(dev, &temp); 838 839 ret = sysfs_emit(buf, fmt_u64, *(u64 *)(((u8 *)stats) + offset)); 840 } 841 rcu_read_unlock(); 842 return ret; 843 } 844 845 /* generate a read-only statistics attribute */ 846 #define NETSTAT_ENTRY(name) \ 847 static ssize_t name##_show(struct device *d, \ 848 struct device_attribute *attr, char *buf) \ 849 { \ 850 return netstat_show(d, attr, buf, \ 851 offsetof(struct rtnl_link_stats64, name)); \ 852 } \ 853 static DEVICE_ATTR_RO(name) 854 855 NETSTAT_ENTRY(rx_packets); 856 NETSTAT_ENTRY(tx_packets); 857 NETSTAT_ENTRY(rx_bytes); 858 NETSTAT_ENTRY(tx_bytes); 859 NETSTAT_ENTRY(rx_errors); 860 NETSTAT_ENTRY(tx_errors); 861 NETSTAT_ENTRY(rx_dropped); 862 NETSTAT_ENTRY(tx_dropped); 863 NETSTAT_ENTRY(multicast); 864 NETSTAT_ENTRY(collisions); 865 NETSTAT_ENTRY(rx_length_errors); 866 NETSTAT_ENTRY(rx_over_errors); 867 NETSTAT_ENTRY(rx_crc_errors); 868 NETSTAT_ENTRY(rx_frame_errors); 869 NETSTAT_ENTRY(rx_fifo_errors); 870 NETSTAT_ENTRY(rx_missed_errors); 871 NETSTAT_ENTRY(tx_aborted_errors); 872 NETSTAT_ENTRY(tx_carrier_errors); 873 NETSTAT_ENTRY(tx_fifo_errors); 874 NETSTAT_ENTRY(tx_heartbeat_errors); 875 NETSTAT_ENTRY(tx_window_errors); 876 NETSTAT_ENTRY(rx_compressed); 877 NETSTAT_ENTRY(tx_compressed); 878 NETSTAT_ENTRY(rx_nohandler); 879 880 static struct attribute *netstat_attrs[] __ro_after_init = { 881 &dev_attr_rx_packets.attr, 882 &dev_attr_tx_packets.attr, 883 &dev_attr_rx_bytes.attr, 884 &dev_attr_tx_bytes.attr, 885 &dev_attr_rx_errors.attr, 886 &dev_attr_tx_errors.attr, 887 &dev_attr_rx_dropped.attr, 888 &dev_attr_tx_dropped.attr, 889 &dev_attr_multicast.attr, 890 &dev_attr_collisions.attr, 891 &dev_attr_rx_length_errors.attr, 892 &dev_attr_rx_over_errors.attr, 893 &dev_attr_rx_crc_errors.attr, 894 &dev_attr_rx_frame_errors.attr, 895 &dev_attr_rx_fifo_errors.attr, 896 &dev_attr_rx_missed_errors.attr, 897 &dev_attr_tx_aborted_errors.attr, 898 &dev_attr_tx_carrier_errors.attr, 899 &dev_attr_tx_fifo_errors.attr, 900 &dev_attr_tx_heartbeat_errors.attr, 901 &dev_attr_tx_window_errors.attr, 902 &dev_attr_rx_compressed.attr, 903 &dev_attr_tx_compressed.attr, 904 &dev_attr_rx_nohandler.attr, 905 NULL 906 }; 907 908 static const struct attribute_group netstat_group = { 909 .name = "statistics", 910 .attrs = netstat_attrs, 911 }; 912 913 static struct attribute *wireless_attrs[] = { 914 NULL 915 }; 916 917 static const struct attribute_group wireless_group = { 918 .name = "wireless", 919 .attrs = wireless_attrs, 920 }; 921 922 static bool wireless_group_needed(struct net_device *ndev) 923 { 924 #if IS_ENABLED(CONFIG_CFG80211) 925 if (ndev->ieee80211_ptr) 926 return true; 927 #endif 928 #if IS_ENABLED(CONFIG_WIRELESS_EXT) 929 if (ndev->wireless_handlers) 930 return true; 931 #endif 932 return false; 933 } 934 935 #else /* CONFIG_SYSFS */ 936 #define net_class_groups NULL 937 #endif /* CONFIG_SYSFS */ 938 939 #ifdef CONFIG_SYSFS 940 #define to_rx_queue_attr(_attr) \ 941 container_of(_attr, struct rx_queue_attribute, attr) 942 943 #define to_rx_queue(obj) container_of(obj, struct netdev_rx_queue, kobj) 944 945 static ssize_t rx_queue_attr_show(struct kobject *kobj, struct attribute *attr, 946 char *buf) 947 { 948 const struct rx_queue_attribute *attribute = to_rx_queue_attr(attr); 949 struct netdev_rx_queue *queue = to_rx_queue(kobj); 950 951 if (!attribute->show) 952 return -EIO; 953 954 return attribute->show(queue, buf); 955 } 956 957 static ssize_t rx_queue_attr_store(struct kobject *kobj, struct attribute *attr, 958 const char *buf, size_t count) 959 { 960 const struct rx_queue_attribute *attribute = to_rx_queue_attr(attr); 961 struct netdev_rx_queue *queue = to_rx_queue(kobj); 962 963 if (!attribute->store) 964 return -EIO; 965 966 return attribute->store(queue, buf, count); 967 } 968 969 static const struct sysfs_ops rx_queue_sysfs_ops = { 970 .show = rx_queue_attr_show, 971 .store = rx_queue_attr_store, 972 }; 973 974 #ifdef CONFIG_RPS 975 static ssize_t show_rps_map(struct netdev_rx_queue *queue, char *buf) 976 { 977 struct rps_map *map; 978 cpumask_var_t mask; 979 int i, len; 980 981 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) 982 return -ENOMEM; 983 984 rcu_read_lock(); 985 map = rcu_dereference(queue->rps_map); 986 if (map) 987 for (i = 0; i < map->len; i++) 988 cpumask_set_cpu(map->cpus[i], mask); 989 990 len = sysfs_emit(buf, "%*pb\n", cpumask_pr_args(mask)); 991 rcu_read_unlock(); 992 free_cpumask_var(mask); 993 994 return len < PAGE_SIZE ? len : -EINVAL; 995 } 996 997 static int netdev_rx_queue_set_rps_mask(struct netdev_rx_queue *queue, 998 cpumask_var_t mask) 999 { 1000 static DEFINE_MUTEX(rps_map_mutex); 1001 struct rps_map *old_map, *map; 1002 int cpu, i; 1003 1004 map = kzalloc(max_t(unsigned int, 1005 RPS_MAP_SIZE(cpumask_weight(mask)), L1_CACHE_BYTES), 1006 GFP_KERNEL); 1007 if (!map) 1008 return -ENOMEM; 1009 1010 i = 0; 1011 for_each_cpu_and(cpu, mask, cpu_online_mask) 1012 map->cpus[i++] = cpu; 1013 1014 if (i) { 1015 map->len = i; 1016 } else { 1017 kfree(map); 1018 map = NULL; 1019 } 1020 1021 mutex_lock(&rps_map_mutex); 1022 old_map = rcu_dereference_protected(queue->rps_map, 1023 mutex_is_locked(&rps_map_mutex)); 1024 rcu_assign_pointer(queue->rps_map, map); 1025 1026 if (map) 1027 static_branch_inc(&rps_needed); 1028 if (old_map) 1029 static_branch_dec(&rps_needed); 1030 1031 mutex_unlock(&rps_map_mutex); 1032 1033 if (old_map) 1034 kfree_rcu(old_map, rcu); 1035 return 0; 1036 } 1037 1038 int rps_cpumask_housekeeping(struct cpumask *mask) 1039 { 1040 if (!cpumask_empty(mask)) { 1041 cpumask_and(mask, mask, housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT)); 1042 cpumask_and(mask, mask, housekeeping_cpumask(HK_TYPE_WQ)); 1043 if (cpumask_empty(mask)) 1044 return -EINVAL; 1045 } 1046 return 0; 1047 } 1048 1049 static ssize_t store_rps_map(struct netdev_rx_queue *queue, 1050 const char *buf, size_t len) 1051 { 1052 cpumask_var_t mask; 1053 int err; 1054 1055 if (!capable(CAP_NET_ADMIN)) 1056 return -EPERM; 1057 1058 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 1059 return -ENOMEM; 1060 1061 err = bitmap_parse(buf, len, cpumask_bits(mask), nr_cpumask_bits); 1062 if (err) 1063 goto out; 1064 1065 err = rps_cpumask_housekeeping(mask); 1066 if (err) 1067 goto out; 1068 1069 err = netdev_rx_queue_set_rps_mask(queue, mask); 1070 1071 out: 1072 free_cpumask_var(mask); 1073 return err ? : len; 1074 } 1075 1076 static ssize_t show_rps_dev_flow_table_cnt(struct netdev_rx_queue *queue, 1077 char *buf) 1078 { 1079 unsigned long val = 0; 1080 rps_tag_ptr tag_ptr; 1081 1082 tag_ptr = READ_ONCE(queue->rps_flow_table); 1083 if (tag_ptr) 1084 val = 1UL << rps_tag_to_log(tag_ptr); 1085 1086 return sysfs_emit(buf, "%lu\n", val); 1087 } 1088 1089 static ssize_t store_rps_dev_flow_table_cnt(struct netdev_rx_queue *queue, 1090 const char *buf, size_t len) 1091 { 1092 rps_tag_ptr otag, tag_ptr = 0UL; 1093 struct rps_dev_flow *table; 1094 unsigned long mask, count; 1095 size_t sz; 1096 int rc; 1097 1098 if (!capable(CAP_NET_ADMIN)) 1099 return -EPERM; 1100 1101 rc = kstrtoul(buf, 0, &count); 1102 if (rc < 0) 1103 return rc; 1104 1105 if (count) { 1106 mask = count - 1; 1107 /* mask = roundup_pow_of_two(count) - 1; 1108 * without overflows... 1109 */ 1110 while ((mask | (mask >> 1)) != mask) 1111 mask |= (mask >> 1); 1112 1113 /* Do not accept too large tables. */ 1114 if (mask > (INT_MAX / sizeof(*table) - 1)) 1115 return -EINVAL; 1116 1117 sz = max_t(size_t, sizeof(*table) * (mask + 1), 1118 PAGE_SIZE); 1119 if (sz <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) || 1120 is_power_of_2(sizeof(*table))) 1121 table = kvmalloc(sz, GFP_KERNEL); 1122 else 1123 table = vmalloc(sz); 1124 if (!table) 1125 return -ENOMEM; 1126 tag_ptr = (rps_tag_ptr)table; 1127 if (rps_tag_to_log(tag_ptr)) { 1128 pr_err_once("store_rps_dev_flow_table_cnt() got a non page aligned allocation.\n"); 1129 kvfree(table); 1130 return -ENOMEM; 1131 } 1132 tag_ptr |= (ilog2(mask) + 1); 1133 for (count = 0; count <= mask; count++) { 1134 table[count].cpu = RPS_NO_CPU; 1135 table[count].filter = RPS_NO_FILTER; 1136 } 1137 } 1138 1139 otag = xchg(&queue->rps_flow_table, tag_ptr); 1140 if (otag) 1141 kvfree_rcu_mightsleep(rps_tag_to_table(otag)); 1142 1143 return len; 1144 } 1145 1146 static struct rx_queue_attribute rps_cpus_attribute __ro_after_init 1147 = __ATTR(rps_cpus, 0644, show_rps_map, store_rps_map); 1148 1149 static struct rx_queue_attribute rps_dev_flow_table_cnt_attribute __ro_after_init 1150 = __ATTR(rps_flow_cnt, 0644, 1151 show_rps_dev_flow_table_cnt, store_rps_dev_flow_table_cnt); 1152 #endif /* CONFIG_RPS */ 1153 1154 static struct attribute *rx_queue_default_attrs[] __ro_after_init = { 1155 #ifdef CONFIG_RPS 1156 &rps_cpus_attribute.attr, 1157 &rps_dev_flow_table_cnt_attribute.attr, 1158 #endif 1159 NULL 1160 }; 1161 ATTRIBUTE_GROUPS(rx_queue_default); 1162 1163 static void rx_queue_release(struct kobject *kobj) 1164 { 1165 struct netdev_rx_queue *queue = to_rx_queue(kobj); 1166 #ifdef CONFIG_RPS 1167 rps_tag_ptr tag_ptr; 1168 struct rps_map *map; 1169 1170 map = rcu_dereference_protected(queue->rps_map, 1); 1171 if (map) { 1172 RCU_INIT_POINTER(queue->rps_map, NULL); 1173 kfree_rcu(map, rcu); 1174 } 1175 1176 tag_ptr = xchg(&queue->rps_flow_table, 0UL); 1177 if (tag_ptr) 1178 kvfree_rcu_mightsleep(rps_tag_to_table(tag_ptr)); 1179 #endif 1180 1181 netdev_tracker_free(queue->dev, &queue->dev_tracker); 1182 /* Pairs with the smp_mb() in rx_queue_add_kobject(). */ 1183 smp_mb(); 1184 memset(kobj, 0, sizeof(*kobj)); 1185 __dev_put(queue->dev); 1186 } 1187 1188 static const struct ns_common *rx_queue_namespace(const struct kobject *kobj) 1189 { 1190 struct netdev_rx_queue *queue = to_rx_queue(kobj); 1191 struct device *dev = &queue->dev->dev; 1192 1193 if (dev->class && dev->class->namespace) 1194 return dev->class->namespace(dev); 1195 1196 return NULL; 1197 } 1198 1199 static void rx_queue_get_ownership(const struct kobject *kobj, 1200 kuid_t *uid, kgid_t *gid) 1201 { 1202 const struct ns_common *ns = rx_queue_namespace(kobj); 1203 1204 net_ns_get_ownership(ns ? container_of(ns, struct net, ns) : NULL, 1205 uid, gid); 1206 } 1207 1208 static const struct kobj_type rx_queue_ktype = { 1209 .sysfs_ops = &rx_queue_sysfs_ops, 1210 .release = rx_queue_release, 1211 .namespace = rx_queue_namespace, 1212 .get_ownership = rx_queue_get_ownership, 1213 }; 1214 1215 static int rx_queue_default_mask(struct net_device *dev, 1216 struct netdev_rx_queue *queue) 1217 { 1218 #if IS_ENABLED(CONFIG_RPS) && IS_ENABLED(CONFIG_SYSCTL) 1219 struct cpumask *rps_default_mask; 1220 int res = 0; 1221 1222 mutex_lock(&rps_default_mask_mutex); 1223 1224 rps_default_mask = dev_net(dev)->core.rps_default_mask; 1225 if (rps_default_mask && !cpumask_empty(rps_default_mask)) 1226 res = netdev_rx_queue_set_rps_mask(queue, rps_default_mask); 1227 1228 mutex_unlock(&rps_default_mask_mutex); 1229 1230 return res; 1231 #else 1232 return 0; 1233 #endif 1234 } 1235 1236 static int rx_queue_add_kobject(struct net_device *dev, int index) 1237 { 1238 struct netdev_rx_queue *queue = dev->_rx + index; 1239 struct kobject *kobj = &queue->kobj; 1240 int error = 0; 1241 1242 /* Rx queues are cleared in rx_queue_release to allow later 1243 * re-registration. This is triggered when their kobj refcount is 1244 * dropped. 1245 * 1246 * If a queue is removed while both a read (or write) operation and a 1247 * the re-addition of the same queue are pending (waiting on rntl_lock) 1248 * it might happen that the re-addition will execute before the read, 1249 * making the initial removal to never happen (queue's kobj refcount 1250 * won't drop enough because of the pending read). In such rare case, 1251 * return to allow the removal operation to complete. 1252 */ 1253 if (unlikely(kobj->state_initialized)) { 1254 netdev_warn_once(dev, "Cannot re-add rx queues before their removal completed"); 1255 return -EAGAIN; 1256 } 1257 1258 /* Pairs with the smp_mb() in rx_queue_release(). */ 1259 smp_mb(); 1260 1261 /* Kobject_put later will trigger rx_queue_release call which 1262 * decreases dev refcount: Take that reference here 1263 */ 1264 netdev_hold(queue->dev, &queue->dev_tracker, GFP_KERNEL); 1265 1266 kobj->kset = dev->queues_kset; 1267 error = kobject_init_and_add(kobj, &rx_queue_ktype, NULL, 1268 "rx-%u", index); 1269 if (error) 1270 goto err; 1271 1272 queue->groups = rx_queue_default_groups; 1273 error = sysfs_create_groups(kobj, queue->groups); 1274 if (error) 1275 goto err; 1276 1277 if (dev->sysfs_rx_queue_group) { 1278 error = sysfs_create_group(kobj, dev->sysfs_rx_queue_group); 1279 if (error) 1280 goto err_default_groups; 1281 } 1282 1283 error = rx_queue_default_mask(dev, queue); 1284 if (error) 1285 goto err_default_groups; 1286 1287 kobject_uevent(kobj, KOBJ_ADD); 1288 1289 return error; 1290 1291 err_default_groups: 1292 sysfs_remove_groups(kobj, queue->groups); 1293 err: 1294 kobject_put(kobj); 1295 return error; 1296 } 1297 1298 static int rx_queue_change_owner(struct net_device *dev, int index, kuid_t kuid, 1299 kgid_t kgid) 1300 { 1301 struct netdev_rx_queue *queue = dev->_rx + index; 1302 struct kobject *kobj = &queue->kobj; 1303 int error; 1304 1305 error = sysfs_change_owner(kobj, kuid, kgid); 1306 if (error) 1307 return error; 1308 1309 if (dev->sysfs_rx_queue_group) 1310 error = sysfs_group_change_owner( 1311 kobj, dev->sysfs_rx_queue_group, kuid, kgid); 1312 1313 return error; 1314 } 1315 #endif /* CONFIG_SYSFS */ 1316 1317 int 1318 net_rx_queue_update_kobjects(struct net_device *dev, int old_num, int new_num) 1319 { 1320 #ifdef CONFIG_SYSFS 1321 int i; 1322 int error = 0; 1323 1324 #ifndef CONFIG_RPS 1325 if (!dev->sysfs_rx_queue_group) 1326 return 0; 1327 #endif 1328 for (i = old_num; i < new_num; i++) { 1329 error = rx_queue_add_kobject(dev, i); 1330 if (error) { 1331 new_num = old_num; 1332 break; 1333 } 1334 } 1335 1336 while (--i >= new_num) { 1337 struct netdev_rx_queue *queue = &dev->_rx[i]; 1338 struct kobject *kobj = &queue->kobj; 1339 1340 if (!check_net(dev_net(dev))) 1341 kobj->uevent_suppress = 1; 1342 if (dev->sysfs_rx_queue_group) 1343 sysfs_remove_group(kobj, dev->sysfs_rx_queue_group); 1344 sysfs_remove_groups(kobj, queue->groups); 1345 kobject_put(kobj); 1346 } 1347 1348 return error; 1349 #else 1350 return 0; 1351 #endif 1352 } 1353 1354 static int net_rx_queue_change_owner(struct net_device *dev, int num, 1355 kuid_t kuid, kgid_t kgid) 1356 { 1357 #ifdef CONFIG_SYSFS 1358 int error = 0; 1359 int i; 1360 1361 #ifndef CONFIG_RPS 1362 if (!dev->sysfs_rx_queue_group) 1363 return 0; 1364 #endif 1365 for (i = 0; i < num; i++) { 1366 error = rx_queue_change_owner(dev, i, kuid, kgid); 1367 if (error) 1368 break; 1369 } 1370 1371 return error; 1372 #else 1373 return 0; 1374 #endif 1375 } 1376 1377 #ifdef CONFIG_SYSFS 1378 /* 1379 * netdev_queue sysfs structures and functions. 1380 */ 1381 struct netdev_queue_attribute { 1382 struct attribute attr; 1383 ssize_t (*show)(struct kobject *kobj, struct attribute *attr, 1384 struct netdev_queue *queue, char *buf); 1385 ssize_t (*store)(struct kobject *kobj, struct attribute *attr, 1386 struct netdev_queue *queue, const char *buf, 1387 size_t len); 1388 }; 1389 #define to_netdev_queue_attr(_attr) \ 1390 container_of(_attr, struct netdev_queue_attribute, attr) 1391 1392 #define to_netdev_queue(obj) container_of(obj, struct netdev_queue, kobj) 1393 1394 static ssize_t netdev_queue_attr_show(struct kobject *kobj, 1395 struct attribute *attr, char *buf) 1396 { 1397 const struct netdev_queue_attribute *attribute 1398 = to_netdev_queue_attr(attr); 1399 struct netdev_queue *queue = to_netdev_queue(kobj); 1400 1401 if (!attribute->show) 1402 return -EIO; 1403 1404 return attribute->show(kobj, attr, queue, buf); 1405 } 1406 1407 static ssize_t netdev_queue_attr_store(struct kobject *kobj, 1408 struct attribute *attr, 1409 const char *buf, size_t count) 1410 { 1411 const struct netdev_queue_attribute *attribute 1412 = to_netdev_queue_attr(attr); 1413 struct netdev_queue *queue = to_netdev_queue(kobj); 1414 1415 if (!attribute->store) 1416 return -EIO; 1417 1418 return attribute->store(kobj, attr, queue, buf, count); 1419 } 1420 1421 static const struct sysfs_ops netdev_queue_sysfs_ops = { 1422 .show = netdev_queue_attr_show, 1423 .store = netdev_queue_attr_store, 1424 }; 1425 1426 static ssize_t tx_timeout_show(struct kobject *kobj, struct attribute *attr, 1427 struct netdev_queue *queue, char *buf) 1428 { 1429 unsigned long trans_timeout = atomic_long_read(&queue->trans_timeout); 1430 1431 return sysfs_emit(buf, fmt_ulong, trans_timeout); 1432 } 1433 1434 static unsigned int get_netdev_queue_index(struct netdev_queue *queue) 1435 { 1436 struct net_device *dev = queue->dev; 1437 unsigned int i; 1438 1439 i = queue - dev->_tx; 1440 BUG_ON(i >= dev->num_tx_queues); 1441 1442 return i; 1443 } 1444 1445 static ssize_t traffic_class_show(struct kobject *kobj, struct attribute *attr, 1446 struct netdev_queue *queue, char *buf) 1447 { 1448 struct net_device *dev = queue->dev; 1449 int num_tc, tc, index, ret; 1450 1451 if (!netif_is_multiqueue(dev)) 1452 return -ENOENT; 1453 1454 ret = sysfs_rtnl_lock(kobj, attr, queue->dev); 1455 if (ret) 1456 return ret; 1457 1458 index = get_netdev_queue_index(queue); 1459 1460 /* If queue belongs to subordinate dev use its TC mapping */ 1461 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 1462 1463 num_tc = READ_ONCE(dev->num_tc); 1464 tc = netdev_txq_to_tc(dev, index); 1465 1466 rtnl_unlock(); 1467 1468 if (tc < 0) 1469 return -EINVAL; 1470 1471 /* We can report the traffic class one of two ways: 1472 * Subordinate device traffic classes are reported with the traffic 1473 * class first, and then the subordinate class so for example TC0 on 1474 * subordinate device 2 will be reported as "0-2". If the queue 1475 * belongs to the root device it will be reported with just the 1476 * traffic class, so just "0" for TC 0 for example. 1477 */ 1478 return num_tc < 0 ? sysfs_emit(buf, "%d%d\n", tc, num_tc) : 1479 sysfs_emit(buf, "%d\n", tc); 1480 } 1481 1482 #ifdef CONFIG_XPS 1483 static ssize_t tx_maxrate_show(struct kobject *kobj, struct attribute *attr, 1484 struct netdev_queue *queue, char *buf) 1485 { 1486 return sysfs_emit(buf, "%lu\n", queue->tx_maxrate); 1487 } 1488 1489 static ssize_t tx_maxrate_store(struct kobject *kobj, struct attribute *attr, 1490 struct netdev_queue *queue, const char *buf, 1491 size_t len) 1492 { 1493 int err, index = get_netdev_queue_index(queue); 1494 struct net_device *dev = queue->dev; 1495 u32 rate = 0; 1496 1497 if (!capable(CAP_NET_ADMIN)) 1498 return -EPERM; 1499 1500 /* The check is also done later; this helps returning early without 1501 * hitting the locking section below. 1502 */ 1503 if (!dev->netdev_ops->ndo_set_tx_maxrate) 1504 return -EOPNOTSUPP; 1505 1506 err = kstrtou32(buf, 10, &rate); 1507 if (err < 0) 1508 return err; 1509 1510 err = sysfs_rtnl_lock(kobj, attr, dev); 1511 if (err) 1512 return err; 1513 1514 err = -EOPNOTSUPP; 1515 netdev_lock_ops(dev); 1516 if (dev->netdev_ops->ndo_set_tx_maxrate) 1517 err = dev->netdev_ops->ndo_set_tx_maxrate(dev, index, rate); 1518 netdev_unlock_ops(dev); 1519 1520 if (!err) { 1521 queue->tx_maxrate = rate; 1522 rtnl_unlock(); 1523 return len; 1524 } 1525 1526 rtnl_unlock(); 1527 return err; 1528 } 1529 1530 static struct netdev_queue_attribute queue_tx_maxrate __ro_after_init 1531 = __ATTR_RW(tx_maxrate); 1532 #endif 1533 1534 static struct netdev_queue_attribute queue_trans_timeout __ro_after_init 1535 = __ATTR_RO(tx_timeout); 1536 1537 static struct netdev_queue_attribute queue_traffic_class __ro_after_init 1538 = __ATTR_RO(traffic_class); 1539 1540 #ifdef CONFIG_BQL 1541 /* 1542 * Byte queue limits sysfs structures and functions. 1543 */ 1544 static ssize_t bql_show(char *buf, unsigned int value) 1545 { 1546 return sysfs_emit(buf, "%u\n", value); 1547 } 1548 1549 static ssize_t bql_set(const char *buf, const size_t count, 1550 unsigned int *pvalue) 1551 { 1552 unsigned int value; 1553 int err; 1554 1555 if (!strcmp(buf, "max") || !strcmp(buf, "max\n")) { 1556 value = DQL_MAX_LIMIT; 1557 } else { 1558 err = kstrtouint(buf, 10, &value); 1559 if (err < 0) 1560 return err; 1561 if (value > DQL_MAX_LIMIT) 1562 return -EINVAL; 1563 } 1564 1565 *pvalue = value; 1566 1567 return count; 1568 } 1569 1570 static ssize_t bql_show_hold_time(struct kobject *kobj, struct attribute *attr, 1571 struct netdev_queue *queue, char *buf) 1572 { 1573 struct dql *dql = &queue->dql; 1574 1575 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(dql->slack_hold_time)); 1576 } 1577 1578 static ssize_t bql_set_hold_time(struct kobject *kobj, struct attribute *attr, 1579 struct netdev_queue *queue, const char *buf, 1580 size_t len) 1581 { 1582 struct dql *dql = &queue->dql; 1583 unsigned int value; 1584 int err; 1585 1586 err = kstrtouint(buf, 10, &value); 1587 if (err < 0) 1588 return err; 1589 1590 dql->slack_hold_time = msecs_to_jiffies(value); 1591 1592 return len; 1593 } 1594 1595 static struct netdev_queue_attribute bql_hold_time_attribute __ro_after_init 1596 = __ATTR(hold_time, 0644, 1597 bql_show_hold_time, bql_set_hold_time); 1598 1599 static ssize_t bql_show_stall_thrs(struct kobject *kobj, struct attribute *attr, 1600 struct netdev_queue *queue, char *buf) 1601 { 1602 struct dql *dql = &queue->dql; 1603 1604 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(dql->stall_thrs)); 1605 } 1606 1607 static ssize_t bql_set_stall_thrs(struct kobject *kobj, struct attribute *attr, 1608 struct netdev_queue *queue, const char *buf, 1609 size_t len) 1610 { 1611 struct dql *dql = &queue->dql; 1612 unsigned int value; 1613 int err; 1614 1615 err = kstrtouint(buf, 10, &value); 1616 if (err < 0) 1617 return err; 1618 1619 value = msecs_to_jiffies(value); 1620 if (value && (value < 4 || value > 4 / 2 * BITS_PER_LONG)) 1621 return -ERANGE; 1622 1623 if (!dql->stall_thrs && value) 1624 dql->last_reap = jiffies; 1625 /* Force last_reap to be live */ 1626 smp_wmb(); 1627 dql->stall_thrs = value; 1628 1629 return len; 1630 } 1631 1632 static struct netdev_queue_attribute bql_stall_thrs_attribute __ro_after_init = 1633 __ATTR(stall_thrs, 0644, bql_show_stall_thrs, bql_set_stall_thrs); 1634 1635 static ssize_t bql_show_stall_max(struct kobject *kobj, struct attribute *attr, 1636 struct netdev_queue *queue, char *buf) 1637 { 1638 return sysfs_emit(buf, "%u\n", READ_ONCE(queue->dql.stall_max)); 1639 } 1640 1641 static ssize_t bql_set_stall_max(struct kobject *kobj, struct attribute *attr, 1642 struct netdev_queue *queue, const char *buf, 1643 size_t len) 1644 { 1645 WRITE_ONCE(queue->dql.stall_max, 0); 1646 return len; 1647 } 1648 1649 static struct netdev_queue_attribute bql_stall_max_attribute __ro_after_init = 1650 __ATTR(stall_max, 0644, bql_show_stall_max, bql_set_stall_max); 1651 1652 static ssize_t bql_show_stall_cnt(struct kobject *kobj, struct attribute *attr, 1653 struct netdev_queue *queue, char *buf) 1654 { 1655 struct dql *dql = &queue->dql; 1656 1657 return sysfs_emit(buf, "%lu\n", dql->stall_cnt); 1658 } 1659 1660 static struct netdev_queue_attribute bql_stall_cnt_attribute __ro_after_init = 1661 __ATTR(stall_cnt, 0444, bql_show_stall_cnt, NULL); 1662 1663 static ssize_t bql_show_inflight(struct kobject *kobj, struct attribute *attr, 1664 struct netdev_queue *queue, char *buf) 1665 { 1666 struct dql *dql = &queue->dql; 1667 1668 return sysfs_emit(buf, "%u\n", dql->num_queued - dql->num_completed); 1669 } 1670 1671 static struct netdev_queue_attribute bql_inflight_attribute __ro_after_init = 1672 __ATTR(inflight, 0444, bql_show_inflight, NULL); 1673 1674 #define BQL_ATTR(NAME, FIELD) \ 1675 static ssize_t bql_show_ ## NAME(struct kobject *kobj, \ 1676 struct attribute *attr, \ 1677 struct netdev_queue *queue, char *buf) \ 1678 { \ 1679 return bql_show(buf, queue->dql.FIELD); \ 1680 } \ 1681 \ 1682 static ssize_t bql_set_ ## NAME(struct kobject *kobj, \ 1683 struct attribute *attr, \ 1684 struct netdev_queue *queue, \ 1685 const char *buf, size_t len) \ 1686 { \ 1687 return bql_set(buf, len, &queue->dql.FIELD); \ 1688 } \ 1689 \ 1690 static struct netdev_queue_attribute bql_ ## NAME ## _attribute __ro_after_init \ 1691 = __ATTR(NAME, 0644, \ 1692 bql_show_ ## NAME, bql_set_ ## NAME) 1693 1694 BQL_ATTR(limit, limit); 1695 BQL_ATTR(limit_max, max_limit); 1696 BQL_ATTR(limit_min, min_limit); 1697 1698 static struct attribute *dql_attrs[] __ro_after_init = { 1699 &bql_limit_attribute.attr, 1700 &bql_limit_max_attribute.attr, 1701 &bql_limit_min_attribute.attr, 1702 &bql_hold_time_attribute.attr, 1703 &bql_inflight_attribute.attr, 1704 &bql_stall_thrs_attribute.attr, 1705 &bql_stall_cnt_attribute.attr, 1706 &bql_stall_max_attribute.attr, 1707 NULL 1708 }; 1709 1710 static const struct attribute_group dql_group = { 1711 .name = "byte_queue_limits", 1712 .attrs = dql_attrs, 1713 }; 1714 #else 1715 /* Fake declaration, all the code using it should be dead */ 1716 static const struct attribute_group dql_group = {}; 1717 #endif /* CONFIG_BQL */ 1718 1719 #ifdef CONFIG_XPS 1720 static ssize_t xps_queue_show(struct net_device *dev, unsigned int index, 1721 int tc, char *buf, enum xps_map_type type) 1722 { 1723 struct xps_dev_maps *dev_maps; 1724 unsigned long *mask; 1725 unsigned int nr_ids; 1726 int j, len; 1727 1728 rcu_read_lock(); 1729 dev_maps = rcu_dereference(dev->xps_maps[type]); 1730 1731 /* Default to nr_cpu_ids/dev->num_rx_queues and do not just return 0 1732 * when dev_maps hasn't been allocated yet, to be backward compatible. 1733 */ 1734 nr_ids = dev_maps ? dev_maps->nr_ids : 1735 (type == XPS_CPUS ? nr_cpu_ids : dev->num_rx_queues); 1736 1737 mask = bitmap_zalloc(nr_ids, GFP_NOWAIT); 1738 if (!mask) { 1739 rcu_read_unlock(); 1740 return -ENOMEM; 1741 } 1742 1743 if (!dev_maps || tc >= dev_maps->num_tc) 1744 goto out_no_maps; 1745 1746 for (j = 0; j < nr_ids; j++) { 1747 int i, tci = j * dev_maps->num_tc + tc; 1748 struct xps_map *map; 1749 1750 map = rcu_dereference(dev_maps->attr_map[tci]); 1751 if (!map) 1752 continue; 1753 1754 for (i = map->len; i--;) { 1755 if (map->queues[i] == index) { 1756 __set_bit(j, mask); 1757 break; 1758 } 1759 } 1760 } 1761 out_no_maps: 1762 rcu_read_unlock(); 1763 1764 len = sysfs_emit(buf, "%*pb\n", nr_ids, mask); 1765 bitmap_free(mask); 1766 1767 return len < PAGE_SIZE ? len : -EINVAL; 1768 } 1769 1770 static ssize_t xps_cpus_show(struct kobject *kobj, struct attribute *attr, 1771 struct netdev_queue *queue, char *buf) 1772 { 1773 struct net_device *dev = queue->dev; 1774 unsigned int index; 1775 int len, tc, ret; 1776 1777 if (!netif_is_multiqueue(dev)) 1778 return -ENOENT; 1779 1780 index = get_netdev_queue_index(queue); 1781 1782 ret = sysfs_rtnl_lock(kobj, attr, queue->dev); 1783 if (ret) 1784 return ret; 1785 1786 /* If queue belongs to subordinate dev use its map */ 1787 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 1788 1789 tc = netdev_txq_to_tc(dev, index); 1790 if (tc < 0) { 1791 rtnl_unlock(); 1792 return -EINVAL; 1793 } 1794 1795 /* Increase the net device refcnt to make sure it won't be freed while 1796 * xps_queue_show is running. 1797 */ 1798 dev_hold(dev); 1799 rtnl_unlock(); 1800 1801 len = xps_queue_show(dev, index, tc, buf, XPS_CPUS); 1802 1803 dev_put(dev); 1804 return len; 1805 } 1806 1807 static ssize_t xps_cpus_store(struct kobject *kobj, struct attribute *attr, 1808 struct netdev_queue *queue, const char *buf, 1809 size_t len) 1810 { 1811 struct net_device *dev = queue->dev; 1812 unsigned int index; 1813 cpumask_var_t mask; 1814 int err; 1815 1816 if (!netif_is_multiqueue(dev)) 1817 return -ENOENT; 1818 1819 if (!capable(CAP_NET_ADMIN)) 1820 return -EPERM; 1821 1822 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 1823 return -ENOMEM; 1824 1825 index = get_netdev_queue_index(queue); 1826 1827 err = bitmap_parse(buf, len, cpumask_bits(mask), nr_cpumask_bits); 1828 if (err) { 1829 free_cpumask_var(mask); 1830 return err; 1831 } 1832 1833 err = sysfs_rtnl_lock(kobj, attr, dev); 1834 if (err) { 1835 free_cpumask_var(mask); 1836 return err; 1837 } 1838 1839 err = netif_set_xps_queue(dev, mask, index); 1840 rtnl_unlock(); 1841 1842 free_cpumask_var(mask); 1843 1844 return err ? : len; 1845 } 1846 1847 static struct netdev_queue_attribute xps_cpus_attribute __ro_after_init 1848 = __ATTR_RW(xps_cpus); 1849 1850 static ssize_t xps_rxqs_show(struct kobject *kobj, struct attribute *attr, 1851 struct netdev_queue *queue, char *buf) 1852 { 1853 struct net_device *dev = queue->dev; 1854 unsigned int index; 1855 int tc, ret; 1856 1857 index = get_netdev_queue_index(queue); 1858 1859 ret = sysfs_rtnl_lock(kobj, attr, dev); 1860 if (ret) 1861 return ret; 1862 1863 tc = netdev_txq_to_tc(dev, index); 1864 1865 /* Increase the net device refcnt to make sure it won't be freed while 1866 * xps_queue_show is running. 1867 */ 1868 dev_hold(dev); 1869 rtnl_unlock(); 1870 1871 ret = tc >= 0 ? xps_queue_show(dev, index, tc, buf, XPS_RXQS) : -EINVAL; 1872 dev_put(dev); 1873 return ret; 1874 } 1875 1876 static ssize_t xps_rxqs_store(struct kobject *kobj, struct attribute *attr, 1877 struct netdev_queue *queue, const char *buf, 1878 size_t len) 1879 { 1880 struct net_device *dev = queue->dev; 1881 struct net *net = dev_net(dev); 1882 unsigned long *mask; 1883 unsigned int index; 1884 int err; 1885 1886 if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) 1887 return -EPERM; 1888 1889 mask = bitmap_zalloc(dev->num_rx_queues, GFP_KERNEL); 1890 if (!mask) 1891 return -ENOMEM; 1892 1893 index = get_netdev_queue_index(queue); 1894 1895 err = bitmap_parse(buf, len, mask, dev->num_rx_queues); 1896 if (err) { 1897 bitmap_free(mask); 1898 return err; 1899 } 1900 1901 err = sysfs_rtnl_lock(kobj, attr, dev); 1902 if (err) { 1903 bitmap_free(mask); 1904 return err; 1905 } 1906 1907 cpus_read_lock(); 1908 err = __netif_set_xps_queue(dev, mask, index, XPS_RXQS); 1909 cpus_read_unlock(); 1910 1911 rtnl_unlock(); 1912 1913 bitmap_free(mask); 1914 return err ? : len; 1915 } 1916 1917 static struct netdev_queue_attribute xps_rxqs_attribute __ro_after_init 1918 = __ATTR_RW(xps_rxqs); 1919 #endif /* CONFIG_XPS */ 1920 1921 static struct attribute *netdev_queue_default_attrs[] __ro_after_init = { 1922 &queue_trans_timeout.attr, 1923 &queue_traffic_class.attr, 1924 #ifdef CONFIG_XPS 1925 &xps_cpus_attribute.attr, 1926 &xps_rxqs_attribute.attr, 1927 &queue_tx_maxrate.attr, 1928 #endif 1929 NULL 1930 }; 1931 ATTRIBUTE_GROUPS(netdev_queue_default); 1932 1933 static void netdev_queue_release(struct kobject *kobj) 1934 { 1935 struct netdev_queue *queue = to_netdev_queue(kobj); 1936 1937 netdev_tracker_free(queue->dev, &queue->dev_tracker); 1938 /* Pairs with the smp_mb() in netdev_queue_add_kobject(). */ 1939 smp_mb(); 1940 memset(kobj, 0, sizeof(*kobj)); 1941 __dev_put(queue->dev); 1942 } 1943 1944 static const struct ns_common *netdev_queue_namespace(const struct kobject *kobj) 1945 { 1946 struct netdev_queue *queue = to_netdev_queue(kobj); 1947 struct device *dev = &queue->dev->dev; 1948 1949 if (dev->class && dev->class->namespace) 1950 return dev->class->namespace(dev); 1951 1952 return NULL; 1953 } 1954 1955 static void netdev_queue_get_ownership(const struct kobject *kobj, 1956 kuid_t *uid, kgid_t *gid) 1957 { 1958 const struct ns_common *ns = netdev_queue_namespace(kobj); 1959 1960 net_ns_get_ownership(ns ? container_of(ns, struct net, ns) : NULL, 1961 uid, gid); 1962 } 1963 1964 static const struct kobj_type netdev_queue_ktype = { 1965 .sysfs_ops = &netdev_queue_sysfs_ops, 1966 .release = netdev_queue_release, 1967 .namespace = netdev_queue_namespace, 1968 .get_ownership = netdev_queue_get_ownership, 1969 }; 1970 1971 static bool netdev_uses_bql(const struct net_device *dev) 1972 { 1973 if (dev->lltx || (dev->priv_flags & IFF_NO_QUEUE)) 1974 return false; 1975 1976 return IS_ENABLED(CONFIG_BQL); 1977 } 1978 1979 static int netdev_queue_add_kobject(struct net_device *dev, int index) 1980 { 1981 struct netdev_queue *queue = dev->_tx + index; 1982 struct kobject *kobj = &queue->kobj; 1983 int error = 0; 1984 1985 /* Tx queues are cleared in netdev_queue_release to allow later 1986 * re-registration. This is triggered when their kobj refcount is 1987 * dropped. 1988 * 1989 * If a queue is removed while both a read (or write) operation and a 1990 * the re-addition of the same queue are pending (waiting on rntl_lock) 1991 * it might happen that the re-addition will execute before the read, 1992 * making the initial removal to never happen (queue's kobj refcount 1993 * won't drop enough because of the pending read). In such rare case, 1994 * return to allow the removal operation to complete. 1995 */ 1996 if (unlikely(kobj->state_initialized)) { 1997 netdev_warn_once(dev, "Cannot re-add tx queues before their removal completed"); 1998 return -EAGAIN; 1999 } 2000 2001 /* Pairs with the smp_mb() in netdev_queue_release(). */ 2002 smp_mb(); 2003 2004 /* Kobject_put later will trigger netdev_queue_release call 2005 * which decreases dev refcount: Take that reference here 2006 */ 2007 netdev_hold(queue->dev, &queue->dev_tracker, GFP_KERNEL); 2008 2009 kobj->kset = dev->queues_kset; 2010 error = kobject_init_and_add(kobj, &netdev_queue_ktype, NULL, 2011 "tx-%u", index); 2012 if (error) 2013 goto err; 2014 2015 queue->groups = netdev_queue_default_groups; 2016 error = sysfs_create_groups(kobj, queue->groups); 2017 if (error) 2018 goto err; 2019 2020 if (netdev_uses_bql(dev)) { 2021 error = sysfs_create_group(kobj, &dql_group); 2022 if (error) 2023 goto err_default_groups; 2024 } 2025 2026 kobject_uevent(kobj, KOBJ_ADD); 2027 return 0; 2028 2029 err_default_groups: 2030 sysfs_remove_groups(kobj, queue->groups); 2031 err: 2032 kobject_put(kobj); 2033 return error; 2034 } 2035 2036 static int tx_queue_change_owner(struct net_device *ndev, int index, 2037 kuid_t kuid, kgid_t kgid) 2038 { 2039 struct netdev_queue *queue = ndev->_tx + index; 2040 struct kobject *kobj = &queue->kobj; 2041 int error; 2042 2043 error = sysfs_change_owner(kobj, kuid, kgid); 2044 if (error) 2045 return error; 2046 2047 if (netdev_uses_bql(ndev)) 2048 error = sysfs_group_change_owner(kobj, &dql_group, kuid, kgid); 2049 2050 return error; 2051 } 2052 #endif /* CONFIG_SYSFS */ 2053 2054 int 2055 netdev_queue_update_kobjects(struct net_device *dev, int old_num, int new_num) 2056 { 2057 #ifdef CONFIG_SYSFS 2058 int i; 2059 int error = 0; 2060 2061 /* Tx queue kobjects are allowed to be updated when a device is being 2062 * unregistered, but solely to remove queues from qdiscs. Any path 2063 * adding queues should be fixed. 2064 */ 2065 WARN(dev->reg_state == NETREG_UNREGISTERING && new_num > old_num, 2066 "New queues can't be registered after device unregistration."); 2067 2068 for (i = old_num; i < new_num; i++) { 2069 error = netdev_queue_add_kobject(dev, i); 2070 if (error) { 2071 new_num = old_num; 2072 break; 2073 } 2074 } 2075 2076 while (--i >= new_num) { 2077 struct netdev_queue *queue = dev->_tx + i; 2078 2079 if (!check_net(dev_net(dev))) 2080 queue->kobj.uevent_suppress = 1; 2081 2082 if (netdev_uses_bql(dev)) 2083 sysfs_remove_group(&queue->kobj, &dql_group); 2084 2085 sysfs_remove_groups(&queue->kobj, queue->groups); 2086 kobject_put(&queue->kobj); 2087 } 2088 2089 return error; 2090 #else 2091 return 0; 2092 #endif /* CONFIG_SYSFS */ 2093 } 2094 2095 static int net_tx_queue_change_owner(struct net_device *dev, int num, 2096 kuid_t kuid, kgid_t kgid) 2097 { 2098 #ifdef CONFIG_SYSFS 2099 int error = 0; 2100 int i; 2101 2102 for (i = 0; i < num; i++) { 2103 error = tx_queue_change_owner(dev, i, kuid, kgid); 2104 if (error) 2105 break; 2106 } 2107 2108 return error; 2109 #else 2110 return 0; 2111 #endif /* CONFIG_SYSFS */ 2112 } 2113 2114 static int register_queue_kobjects(struct net_device *dev) 2115 { 2116 int error = 0, txq = 0, rxq = 0, real_rx = 0, real_tx = 0; 2117 2118 #ifdef CONFIG_SYSFS 2119 dev->queues_kset = kset_create_and_add("queues", 2120 NULL, &dev->dev.kobj); 2121 if (!dev->queues_kset) 2122 return -ENOMEM; 2123 real_rx = dev->real_num_rx_queues; 2124 #endif 2125 real_tx = dev->real_num_tx_queues; 2126 2127 error = net_rx_queue_update_kobjects(dev, 0, real_rx); 2128 if (error) 2129 goto error; 2130 rxq = real_rx; 2131 2132 error = netdev_queue_update_kobjects(dev, 0, real_tx); 2133 if (error) 2134 goto error; 2135 txq = real_tx; 2136 2137 return 0; 2138 2139 error: 2140 netdev_queue_update_kobjects(dev, txq, 0); 2141 net_rx_queue_update_kobjects(dev, rxq, 0); 2142 #ifdef CONFIG_SYSFS 2143 kset_unregister(dev->queues_kset); 2144 #endif 2145 return error; 2146 } 2147 2148 static int queue_change_owner(struct net_device *ndev, kuid_t kuid, kgid_t kgid) 2149 { 2150 int error = 0, real_rx = 0, real_tx = 0; 2151 2152 #ifdef CONFIG_SYSFS 2153 if (ndev->queues_kset) { 2154 error = sysfs_change_owner(&ndev->queues_kset->kobj, kuid, kgid); 2155 if (error) 2156 return error; 2157 } 2158 real_rx = ndev->real_num_rx_queues; 2159 #endif 2160 real_tx = ndev->real_num_tx_queues; 2161 2162 error = net_rx_queue_change_owner(ndev, real_rx, kuid, kgid); 2163 if (error) 2164 return error; 2165 2166 error = net_tx_queue_change_owner(ndev, real_tx, kuid, kgid); 2167 if (error) 2168 return error; 2169 2170 return 0; 2171 } 2172 2173 static void remove_queue_kobjects(struct net_device *dev) 2174 { 2175 int real_rx = 0, real_tx = 0; 2176 2177 #ifdef CONFIG_SYSFS 2178 real_rx = dev->real_num_rx_queues; 2179 #endif 2180 real_tx = dev->real_num_tx_queues; 2181 2182 net_rx_queue_update_kobjects(dev, real_rx, 0); 2183 netdev_queue_update_kobjects(dev, real_tx, 0); 2184 2185 netdev_lock_ops(dev); 2186 dev->real_num_rx_queues = 0; 2187 dev->real_num_tx_queues = 0; 2188 netdev_unlock_ops(dev); 2189 #ifdef CONFIG_SYSFS 2190 kset_unregister(dev->queues_kset); 2191 #endif 2192 } 2193 2194 static bool net_current_may_mount(void) 2195 { 2196 struct net *net = current->nsproxy->net_ns; 2197 2198 return ns_capable(net->user_ns, CAP_SYS_ADMIN); 2199 } 2200 2201 static struct ns_common *net_grab_current_ns(void) 2202 { 2203 struct net *net = current->nsproxy->net_ns; 2204 #ifdef CONFIG_NET_NS 2205 if (net) 2206 refcount_inc(&net->passive); 2207 #endif 2208 return net ? to_ns_common(net) : NULL; 2209 } 2210 2211 static const struct ns_common *net_initial_ns(void) 2212 { 2213 return to_ns_common(&init_net); 2214 } 2215 2216 static const struct ns_common *net_netlink_ns(struct sock *sk) 2217 { 2218 return to_ns_common(sock_net(sk)); 2219 } 2220 2221 const struct kobj_ns_type_operations net_ns_type_operations = { 2222 .type = KOBJ_NS_TYPE_NET, 2223 .current_may_mount = net_current_may_mount, 2224 .grab_current_ns = net_grab_current_ns, 2225 .netlink_ns = net_netlink_ns, 2226 .initial_ns = net_initial_ns, 2227 .drop_ns = net_drop_ns, 2228 }; 2229 EXPORT_SYMBOL_GPL(net_ns_type_operations); 2230 2231 static int netdev_uevent(const struct device *d, struct kobj_uevent_env *env) 2232 { 2233 const struct net_device *dev = to_net_dev(d); 2234 int retval; 2235 2236 /* pass interface to uevent. */ 2237 retval = add_uevent_var(env, "INTERFACE=%s", dev->name); 2238 if (retval) 2239 goto exit; 2240 2241 /* pass ifindex to uevent. 2242 * ifindex is useful as it won't change (interface name may change) 2243 * and is what RtNetlink uses natively. 2244 */ 2245 retval = add_uevent_var(env, "IFINDEX=%d", dev->ifindex); 2246 2247 exit: 2248 return retval; 2249 } 2250 2251 /* 2252 * netdev_release -- destroy and free a dead device. 2253 * Called when last reference to device kobject is gone. 2254 */ 2255 static void netdev_release(struct device *d) 2256 { 2257 struct net_device *dev = to_net_dev(d); 2258 2259 BUG_ON(dev->reg_state != NETREG_RELEASED); 2260 2261 /* no need to wait for rcu grace period: 2262 * device is dead and about to be freed. 2263 */ 2264 kfree(rcu_access_pointer(dev->ifalias)); 2265 kvfree(dev); 2266 } 2267 2268 static const struct ns_common *net_namespace(const struct device *d) 2269 { 2270 const struct net_device *dev = to_net_dev(d); 2271 2272 return to_ns_common(dev_net(dev)); 2273 } 2274 2275 static void net_get_ownership(const struct device *d, kuid_t *uid, kgid_t *gid) 2276 { 2277 const struct net_device *dev = to_net_dev(d); 2278 const struct net *net = dev_net(dev); 2279 2280 net_ns_get_ownership(net, uid, gid); 2281 } 2282 2283 static const struct class net_class = { 2284 .name = "net", 2285 .dev_release = netdev_release, 2286 .dev_groups = net_class_groups, 2287 .dev_uevent = netdev_uevent, 2288 .ns_type = &net_ns_type_operations, 2289 .namespace = net_namespace, 2290 .get_ownership = net_get_ownership, 2291 }; 2292 2293 #ifdef CONFIG_OF 2294 static int of_dev_node_match(struct device *dev, const void *data) 2295 { 2296 for (; dev; dev = dev->parent) { 2297 if (dev->of_node == data) 2298 return 1; 2299 } 2300 2301 return 0; 2302 } 2303 2304 /* 2305 * of_find_net_device_by_node - lookup the net device for the device node 2306 * @np: OF device node 2307 * 2308 * Looks up the net_device structure corresponding with the device node. 2309 * If successful, returns a pointer to the net_device with the embedded 2310 * struct device refcount incremented by one, or NULL on failure. The 2311 * refcount must be dropped when done with the net_device. 2312 */ 2313 struct net_device *of_find_net_device_by_node(struct device_node *np) 2314 { 2315 struct device *dev; 2316 2317 dev = class_find_device(&net_class, NULL, np, of_dev_node_match); 2318 if (!dev) 2319 return NULL; 2320 2321 return to_net_dev(dev); 2322 } 2323 EXPORT_SYMBOL(of_find_net_device_by_node); 2324 #endif 2325 2326 /* Delete sysfs entries but hold kobject reference until after all 2327 * netdev references are gone. 2328 */ 2329 void netdev_unregister_kobject(struct net_device *ndev) 2330 { 2331 struct device *dev = &ndev->dev; 2332 2333 if (!check_net(dev_net(ndev))) 2334 dev_set_uevent_suppress(dev, 1); 2335 2336 kobject_get(&dev->kobj); 2337 2338 remove_queue_kobjects(ndev); 2339 2340 pm_runtime_set_memalloc_noio(dev, false); 2341 2342 device_del(dev); 2343 } 2344 2345 /* Create sysfs entries for network device. */ 2346 int netdev_register_kobject(struct net_device *ndev) 2347 { 2348 struct device *dev = &ndev->dev; 2349 const struct attribute_group **groups = ndev->sysfs_groups; 2350 int error = 0; 2351 2352 device_initialize(dev); 2353 dev->class = &net_class; 2354 dev->platform_data = ndev; 2355 dev->groups = groups; 2356 2357 dev_set_name(dev, "%s", ndev->name); 2358 2359 #ifdef CONFIG_SYSFS 2360 /* Allow for a device specific group */ 2361 if (*groups) 2362 groups++; 2363 2364 *groups++ = &netstat_group; 2365 *groups++ = &netdev_phys_group; 2366 2367 if (wireless_group_needed(ndev)) 2368 *groups++ = &wireless_group; 2369 #endif /* CONFIG_SYSFS */ 2370 2371 /* Hold back the KOBJ_ADD uevent until the device is listed. */ 2372 dev_set_uevent_suppress(dev, 1); 2373 2374 error = device_add(dev); 2375 if (error) 2376 return error; 2377 2378 error = register_queue_kobjects(ndev); 2379 if (error) { 2380 device_del(dev); 2381 return error; 2382 } 2383 2384 pm_runtime_set_memalloc_noio(dev, true); 2385 2386 return error; 2387 } 2388 2389 /* Announce a fully registered device to userspace. This pairs with the uevent 2390 * suppression from netdev_register_kobject(). 2391 */ 2392 void netdev_uevent_add(struct net_device *ndev) 2393 { 2394 struct device *dev = &ndev->dev; 2395 2396 dev_set_uevent_suppress(dev, 0); 2397 kobject_uevent(&dev->kobj, KOBJ_ADD); 2398 } 2399 2400 /* Change owner for sysfs entries when moving network devices across network 2401 * namespaces owned by different user namespaces. 2402 */ 2403 int netdev_change_owner(struct net_device *ndev, const struct net *net_old, 2404 const struct net *net_new) 2405 { 2406 kuid_t old_uid = GLOBAL_ROOT_UID, new_uid = GLOBAL_ROOT_UID; 2407 kgid_t old_gid = GLOBAL_ROOT_GID, new_gid = GLOBAL_ROOT_GID; 2408 struct device *dev = &ndev->dev; 2409 int error; 2410 2411 net_ns_get_ownership(net_old, &old_uid, &old_gid); 2412 net_ns_get_ownership(net_new, &new_uid, &new_gid); 2413 2414 /* The network namespace was changed but the owning user namespace is 2415 * identical so there's no need to change the owner of sysfs entries. 2416 */ 2417 if (uid_eq(old_uid, new_uid) && gid_eq(old_gid, new_gid)) 2418 return 0; 2419 2420 error = device_change_owner(dev, new_uid, new_gid); 2421 if (error) 2422 return error; 2423 2424 error = queue_change_owner(ndev, new_uid, new_gid); 2425 if (error) 2426 return error; 2427 2428 return 0; 2429 } 2430 2431 int netdev_class_create_file_ns(const struct class_attribute *class_attr, 2432 const struct ns_common *ns) 2433 { 2434 return class_create_file_ns(&net_class, class_attr, ns); 2435 } 2436 EXPORT_SYMBOL(netdev_class_create_file_ns); 2437 2438 void netdev_class_remove_file_ns(const struct class_attribute *class_attr, 2439 const struct ns_common *ns) 2440 { 2441 class_remove_file_ns(&net_class, class_attr, ns); 2442 } 2443 EXPORT_SYMBOL(netdev_class_remove_file_ns); 2444 2445 int __init netdev_kobject_init(void) 2446 { 2447 kobj_ns_type_register(&net_ns_type_operations); 2448 return class_register(&net_class); 2449 } 2450