1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright(c) 2013-2015 Intel Corporation. All rights reserved. 4 */ 5 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 6 #include <linux/libnvdimm.h> 7 #include <linux/sched/mm.h> 8 #include <linux/vmalloc.h> 9 #include <linux/uaccess.h> 10 #include <linux/module.h> 11 #include <linux/blkdev.h> 12 #include <linux/fcntl.h> 13 #include <linux/async.h> 14 #include <linux/ndctl.h> 15 #include <linux/sched.h> 16 #include <linux/slab.h> 17 #include <linux/cpu.h> 18 #include <linux/fs.h> 19 #include <linux/io.h> 20 #include <linux/mm.h> 21 #include <linux/nd.h> 22 #include "nd-core.h" 23 #include "nd.h" 24 #include "pfn.h" 25 26 int nvdimm_major; 27 static int nvdimm_bus_major; 28 struct class *nd_class; 29 static DEFINE_IDA(nd_ida); 30 31 static int to_nd_device_type(struct device *dev) 32 { 33 if (is_nvdimm(dev)) 34 return ND_DEVICE_DIMM; 35 else if (is_memory(dev)) 36 return ND_DEVICE_REGION_PMEM; 37 else if (is_nd_blk(dev)) 38 return ND_DEVICE_REGION_BLK; 39 else if (is_nd_dax(dev)) 40 return ND_DEVICE_DAX_PMEM; 41 else if (is_nd_region(dev->parent)) 42 return nd_region_to_nstype(to_nd_region(dev->parent)); 43 44 return 0; 45 } 46 47 static int nvdimm_bus_uevent(struct device *dev, struct kobj_uevent_env *env) 48 { 49 return add_uevent_var(env, "MODALIAS=" ND_DEVICE_MODALIAS_FMT, 50 to_nd_device_type(dev)); 51 } 52 53 static struct module *to_bus_provider(struct device *dev) 54 { 55 /* pin bus providers while regions are enabled */ 56 if (is_nd_region(dev)) { 57 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev); 58 59 return nvdimm_bus->nd_desc->module; 60 } 61 return NULL; 62 } 63 64 static void nvdimm_bus_probe_start(struct nvdimm_bus *nvdimm_bus) 65 { 66 nvdimm_bus_lock(&nvdimm_bus->dev); 67 nvdimm_bus->probe_active++; 68 nvdimm_bus_unlock(&nvdimm_bus->dev); 69 } 70 71 static void nvdimm_bus_probe_end(struct nvdimm_bus *nvdimm_bus) 72 { 73 nvdimm_bus_lock(&nvdimm_bus->dev); 74 if (--nvdimm_bus->probe_active == 0) 75 wake_up(&nvdimm_bus->wait); 76 nvdimm_bus_unlock(&nvdimm_bus->dev); 77 } 78 79 static int nvdimm_bus_probe(struct device *dev) 80 { 81 struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver); 82 struct module *provider = to_bus_provider(dev); 83 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev); 84 int rc; 85 86 if (!try_module_get(provider)) 87 return -ENXIO; 88 89 dev_dbg(&nvdimm_bus->dev, "START: %s.probe(%s)\n", 90 dev->driver->name, dev_name(dev)); 91 92 nvdimm_bus_probe_start(nvdimm_bus); 93 debug_nvdimm_lock(dev); 94 rc = nd_drv->probe(dev); 95 debug_nvdimm_unlock(dev); 96 97 if ((rc == 0 || rc == -EOPNOTSUPP) && 98 dev->parent && is_nd_region(dev->parent)) 99 nd_region_advance_seeds(to_nd_region(dev->parent), dev); 100 nvdimm_bus_probe_end(nvdimm_bus); 101 102 dev_dbg(&nvdimm_bus->dev, "END: %s.probe(%s) = %d\n", dev->driver->name, 103 dev_name(dev), rc); 104 105 if (rc != 0) 106 module_put(provider); 107 return rc; 108 } 109 110 static void nvdimm_bus_remove(struct device *dev) 111 { 112 struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver); 113 struct module *provider = to_bus_provider(dev); 114 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev); 115 116 if (nd_drv->remove) { 117 debug_nvdimm_lock(dev); 118 nd_drv->remove(dev); 119 debug_nvdimm_unlock(dev); 120 } 121 122 dev_dbg(&nvdimm_bus->dev, "%s.remove(%s)\n", dev->driver->name, 123 dev_name(dev)); 124 module_put(provider); 125 } 126 127 static void nvdimm_bus_shutdown(struct device *dev) 128 { 129 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev); 130 struct nd_device_driver *nd_drv = NULL; 131 132 if (dev->driver) 133 nd_drv = to_nd_device_driver(dev->driver); 134 135 if (nd_drv && nd_drv->shutdown) { 136 nd_drv->shutdown(dev); 137 dev_dbg(&nvdimm_bus->dev, "%s.shutdown(%s)\n", 138 dev->driver->name, dev_name(dev)); 139 } 140 } 141 142 void nd_device_notify(struct device *dev, enum nvdimm_event event) 143 { 144 nd_device_lock(dev); 145 if (dev->driver) { 146 struct nd_device_driver *nd_drv; 147 148 nd_drv = to_nd_device_driver(dev->driver); 149 if (nd_drv->notify) 150 nd_drv->notify(dev, event); 151 } 152 nd_device_unlock(dev); 153 } 154 EXPORT_SYMBOL(nd_device_notify); 155 156 void nvdimm_region_notify(struct nd_region *nd_region, enum nvdimm_event event) 157 { 158 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(&nd_region->dev); 159 160 if (!nvdimm_bus) 161 return; 162 163 /* caller is responsible for holding a reference on the device */ 164 nd_device_notify(&nd_region->dev, event); 165 } 166 EXPORT_SYMBOL_GPL(nvdimm_region_notify); 167 168 struct clear_badblocks_context { 169 resource_size_t phys, cleared; 170 }; 171 172 static int nvdimm_clear_badblocks_region(struct device *dev, void *data) 173 { 174 struct clear_badblocks_context *ctx = data; 175 struct nd_region *nd_region; 176 resource_size_t ndr_end; 177 sector_t sector; 178 179 /* make sure device is a region */ 180 if (!is_memory(dev)) 181 return 0; 182 183 nd_region = to_nd_region(dev); 184 ndr_end = nd_region->ndr_start + nd_region->ndr_size - 1; 185 186 /* make sure we are in the region */ 187 if (ctx->phys < nd_region->ndr_start 188 || (ctx->phys + ctx->cleared) > ndr_end) 189 return 0; 190 191 sector = (ctx->phys - nd_region->ndr_start) / 512; 192 badblocks_clear(&nd_region->bb, sector, ctx->cleared / 512); 193 194 if (nd_region->bb_state) 195 sysfs_notify_dirent(nd_region->bb_state); 196 197 return 0; 198 } 199 200 static void nvdimm_clear_badblocks_regions(struct nvdimm_bus *nvdimm_bus, 201 phys_addr_t phys, u64 cleared) 202 { 203 struct clear_badblocks_context ctx = { 204 .phys = phys, 205 .cleared = cleared, 206 }; 207 208 device_for_each_child(&nvdimm_bus->dev, &ctx, 209 nvdimm_clear_badblocks_region); 210 } 211 212 static void nvdimm_account_cleared_poison(struct nvdimm_bus *nvdimm_bus, 213 phys_addr_t phys, u64 cleared) 214 { 215 if (cleared > 0) 216 badrange_forget(&nvdimm_bus->badrange, phys, cleared); 217 218 if (cleared > 0 && cleared / 512) 219 nvdimm_clear_badblocks_regions(nvdimm_bus, phys, cleared); 220 } 221 222 long nvdimm_clear_poison(struct device *dev, phys_addr_t phys, 223 unsigned int len) 224 { 225 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev); 226 struct nvdimm_bus_descriptor *nd_desc; 227 struct nd_cmd_clear_error clear_err; 228 struct nd_cmd_ars_cap ars_cap; 229 u32 clear_err_unit, mask; 230 unsigned int noio_flag; 231 int cmd_rc, rc; 232 233 if (!nvdimm_bus) 234 return -ENXIO; 235 236 nd_desc = nvdimm_bus->nd_desc; 237 /* 238 * if ndctl does not exist, it's PMEM_LEGACY and 239 * we want to just pretend everything is handled. 240 */ 241 if (!nd_desc->ndctl) 242 return len; 243 244 memset(&ars_cap, 0, sizeof(ars_cap)); 245 ars_cap.address = phys; 246 ars_cap.length = len; 247 noio_flag = memalloc_noio_save(); 248 rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_ARS_CAP, &ars_cap, 249 sizeof(ars_cap), &cmd_rc); 250 memalloc_noio_restore(noio_flag); 251 if (rc < 0) 252 return rc; 253 if (cmd_rc < 0) 254 return cmd_rc; 255 clear_err_unit = ars_cap.clear_err_unit; 256 if (!clear_err_unit || !is_power_of_2(clear_err_unit)) 257 return -ENXIO; 258 259 mask = clear_err_unit - 1; 260 if ((phys | len) & mask) 261 return -ENXIO; 262 memset(&clear_err, 0, sizeof(clear_err)); 263 clear_err.address = phys; 264 clear_err.length = len; 265 noio_flag = memalloc_noio_save(); 266 rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_CLEAR_ERROR, &clear_err, 267 sizeof(clear_err), &cmd_rc); 268 memalloc_noio_restore(noio_flag); 269 if (rc < 0) 270 return rc; 271 if (cmd_rc < 0) 272 return cmd_rc; 273 274 nvdimm_account_cleared_poison(nvdimm_bus, phys, clear_err.cleared); 275 276 return clear_err.cleared; 277 } 278 EXPORT_SYMBOL_GPL(nvdimm_clear_poison); 279 280 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv); 281 282 static struct bus_type nvdimm_bus_type = { 283 .name = "nd", 284 .uevent = nvdimm_bus_uevent, 285 .match = nvdimm_bus_match, 286 .probe = nvdimm_bus_probe, 287 .remove = nvdimm_bus_remove, 288 .shutdown = nvdimm_bus_shutdown, 289 }; 290 291 static void nvdimm_bus_release(struct device *dev) 292 { 293 struct nvdimm_bus *nvdimm_bus; 294 295 nvdimm_bus = container_of(dev, struct nvdimm_bus, dev); 296 ida_simple_remove(&nd_ida, nvdimm_bus->id); 297 kfree(nvdimm_bus); 298 } 299 300 static const struct device_type nvdimm_bus_dev_type = { 301 .release = nvdimm_bus_release, 302 .groups = nvdimm_bus_attribute_groups, 303 }; 304 305 bool is_nvdimm_bus(struct device *dev) 306 { 307 return dev->type == &nvdimm_bus_dev_type; 308 } 309 310 struct nvdimm_bus *walk_to_nvdimm_bus(struct device *nd_dev) 311 { 312 struct device *dev; 313 314 for (dev = nd_dev; dev; dev = dev->parent) 315 if (is_nvdimm_bus(dev)) 316 break; 317 dev_WARN_ONCE(nd_dev, !dev, "invalid dev, not on nd bus\n"); 318 if (dev) 319 return to_nvdimm_bus(dev); 320 return NULL; 321 } 322 323 struct nvdimm_bus *to_nvdimm_bus(struct device *dev) 324 { 325 struct nvdimm_bus *nvdimm_bus; 326 327 nvdimm_bus = container_of(dev, struct nvdimm_bus, dev); 328 WARN_ON(!is_nvdimm_bus(dev)); 329 return nvdimm_bus; 330 } 331 EXPORT_SYMBOL_GPL(to_nvdimm_bus); 332 333 struct nvdimm_bus *nvdimm_to_bus(struct nvdimm *nvdimm) 334 { 335 return to_nvdimm_bus(nvdimm->dev.parent); 336 } 337 EXPORT_SYMBOL_GPL(nvdimm_to_bus); 338 339 struct nvdimm_bus *nvdimm_bus_register(struct device *parent, 340 struct nvdimm_bus_descriptor *nd_desc) 341 { 342 struct nvdimm_bus *nvdimm_bus; 343 int rc; 344 345 nvdimm_bus = kzalloc(sizeof(*nvdimm_bus), GFP_KERNEL); 346 if (!nvdimm_bus) 347 return NULL; 348 INIT_LIST_HEAD(&nvdimm_bus->list); 349 INIT_LIST_HEAD(&nvdimm_bus->mapping_list); 350 init_waitqueue_head(&nvdimm_bus->wait); 351 nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL); 352 if (nvdimm_bus->id < 0) { 353 kfree(nvdimm_bus); 354 return NULL; 355 } 356 mutex_init(&nvdimm_bus->reconfig_mutex); 357 badrange_init(&nvdimm_bus->badrange); 358 nvdimm_bus->nd_desc = nd_desc; 359 nvdimm_bus->dev.parent = parent; 360 nvdimm_bus->dev.type = &nvdimm_bus_dev_type; 361 nvdimm_bus->dev.groups = nd_desc->attr_groups; 362 nvdimm_bus->dev.bus = &nvdimm_bus_type; 363 nvdimm_bus->dev.of_node = nd_desc->of_node; 364 device_initialize(&nvdimm_bus->dev); 365 device_set_pm_not_required(&nvdimm_bus->dev); 366 rc = dev_set_name(&nvdimm_bus->dev, "ndbus%d", nvdimm_bus->id); 367 if (rc) 368 goto err; 369 370 rc = device_add(&nvdimm_bus->dev); 371 if (rc) { 372 dev_dbg(&nvdimm_bus->dev, "registration failed: %d\n", rc); 373 goto err; 374 } 375 376 return nvdimm_bus; 377 err: 378 put_device(&nvdimm_bus->dev); 379 return NULL; 380 } 381 EXPORT_SYMBOL_GPL(nvdimm_bus_register); 382 383 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus) 384 { 385 if (!nvdimm_bus) 386 return; 387 device_unregister(&nvdimm_bus->dev); 388 } 389 EXPORT_SYMBOL_GPL(nvdimm_bus_unregister); 390 391 static int child_unregister(struct device *dev, void *data) 392 { 393 /* 394 * the singular ndctl class device per bus needs to be 395 * "device_destroy"ed, so skip it here 396 * 397 * i.e. remove classless children 398 */ 399 if (dev->class) 400 return 0; 401 402 if (is_nvdimm(dev)) 403 nvdimm_delete(to_nvdimm(dev)); 404 else 405 nd_device_unregister(dev, ND_SYNC); 406 407 return 0; 408 } 409 410 static void free_badrange_list(struct list_head *badrange_list) 411 { 412 struct badrange_entry *bre, *next; 413 414 list_for_each_entry_safe(bre, next, badrange_list, list) { 415 list_del(&bre->list); 416 kfree(bre); 417 } 418 list_del_init(badrange_list); 419 } 420 421 static void nd_bus_remove(struct device *dev) 422 { 423 struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev); 424 425 mutex_lock(&nvdimm_bus_list_mutex); 426 list_del_init(&nvdimm_bus->list); 427 mutex_unlock(&nvdimm_bus_list_mutex); 428 429 wait_event(nvdimm_bus->wait, 430 atomic_read(&nvdimm_bus->ioctl_active) == 0); 431 432 nd_synchronize(); 433 device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister); 434 435 spin_lock(&nvdimm_bus->badrange.lock); 436 free_badrange_list(&nvdimm_bus->badrange.list); 437 spin_unlock(&nvdimm_bus->badrange.lock); 438 439 nvdimm_bus_destroy_ndctl(nvdimm_bus); 440 } 441 442 static int nd_bus_probe(struct device *dev) 443 { 444 struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev); 445 int rc; 446 447 rc = nvdimm_bus_create_ndctl(nvdimm_bus); 448 if (rc) 449 return rc; 450 451 mutex_lock(&nvdimm_bus_list_mutex); 452 list_add_tail(&nvdimm_bus->list, &nvdimm_bus_list); 453 mutex_unlock(&nvdimm_bus_list_mutex); 454 455 /* enable bus provider attributes to look up their local context */ 456 dev_set_drvdata(dev, nvdimm_bus->nd_desc); 457 458 return 0; 459 } 460 461 static struct nd_device_driver nd_bus_driver = { 462 .probe = nd_bus_probe, 463 .remove = nd_bus_remove, 464 .drv = { 465 .name = "nd_bus", 466 .suppress_bind_attrs = true, 467 .bus = &nvdimm_bus_type, 468 .owner = THIS_MODULE, 469 .mod_name = KBUILD_MODNAME, 470 }, 471 }; 472 473 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv) 474 { 475 struct nd_device_driver *nd_drv = to_nd_device_driver(drv); 476 477 if (is_nvdimm_bus(dev) && nd_drv == &nd_bus_driver) 478 return true; 479 480 return !!test_bit(to_nd_device_type(dev), &nd_drv->type); 481 } 482 483 static ASYNC_DOMAIN_EXCLUSIVE(nd_async_domain); 484 485 void nd_synchronize(void) 486 { 487 async_synchronize_full_domain(&nd_async_domain); 488 } 489 EXPORT_SYMBOL_GPL(nd_synchronize); 490 491 static void nd_async_device_register(void *d, async_cookie_t cookie) 492 { 493 struct device *dev = d; 494 495 if (device_add(dev) != 0) { 496 dev_err(dev, "%s: failed\n", __func__); 497 put_device(dev); 498 } 499 put_device(dev); 500 if (dev->parent) 501 put_device(dev->parent); 502 } 503 504 static void nd_async_device_unregister(void *d, async_cookie_t cookie) 505 { 506 struct device *dev = d; 507 508 /* flush bus operations before delete */ 509 nvdimm_bus_lock(dev); 510 nvdimm_bus_unlock(dev); 511 512 device_unregister(dev); 513 put_device(dev); 514 } 515 516 void __nd_device_register(struct device *dev) 517 { 518 if (!dev) 519 return; 520 521 /* 522 * Ensure that region devices always have their NUMA node set as 523 * early as possible. This way we are able to make certain that 524 * any memory associated with the creation and the creation 525 * itself of the region is associated with the correct node. 526 */ 527 if (is_nd_region(dev)) 528 set_dev_node(dev, to_nd_region(dev)->numa_node); 529 530 dev->bus = &nvdimm_bus_type; 531 device_set_pm_not_required(dev); 532 if (dev->parent) { 533 get_device(dev->parent); 534 if (dev_to_node(dev) == NUMA_NO_NODE) 535 set_dev_node(dev, dev_to_node(dev->parent)); 536 } 537 get_device(dev); 538 539 async_schedule_dev_domain(nd_async_device_register, dev, 540 &nd_async_domain); 541 } 542 543 void nd_device_register(struct device *dev) 544 { 545 device_initialize(dev); 546 __nd_device_register(dev); 547 } 548 EXPORT_SYMBOL(nd_device_register); 549 550 void nd_device_unregister(struct device *dev, enum nd_async_mode mode) 551 { 552 bool killed; 553 554 switch (mode) { 555 case ND_ASYNC: 556 /* 557 * In the async case this is being triggered with the 558 * device lock held and the unregistration work needs to 559 * be moved out of line iff this is thread has won the 560 * race to schedule the deletion. 561 */ 562 if (!kill_device(dev)) 563 return; 564 565 get_device(dev); 566 async_schedule_domain(nd_async_device_unregister, dev, 567 &nd_async_domain); 568 break; 569 case ND_SYNC: 570 /* 571 * In the sync case the device is being unregistered due 572 * to a state change of the parent. Claim the kill state 573 * to synchronize against other unregistration requests, 574 * or otherwise let the async path handle it if the 575 * unregistration was already queued. 576 */ 577 nd_device_lock(dev); 578 killed = kill_device(dev); 579 nd_device_unlock(dev); 580 581 if (!killed) 582 return; 583 584 nd_synchronize(); 585 device_unregister(dev); 586 break; 587 } 588 } 589 EXPORT_SYMBOL(nd_device_unregister); 590 591 /** 592 * __nd_driver_register() - register a region or a namespace driver 593 * @nd_drv: driver to register 594 * @owner: automatically set by nd_driver_register() macro 595 * @mod_name: automatically set by nd_driver_register() macro 596 */ 597 int __nd_driver_register(struct nd_device_driver *nd_drv, struct module *owner, 598 const char *mod_name) 599 { 600 struct device_driver *drv = &nd_drv->drv; 601 602 if (!nd_drv->type) { 603 pr_debug("driver type bitmask not set (%ps)\n", 604 __builtin_return_address(0)); 605 return -EINVAL; 606 } 607 608 if (!nd_drv->probe) { 609 pr_debug("%s ->probe() must be specified\n", mod_name); 610 return -EINVAL; 611 } 612 613 drv->bus = &nvdimm_bus_type; 614 drv->owner = owner; 615 drv->mod_name = mod_name; 616 617 return driver_register(drv); 618 } 619 EXPORT_SYMBOL(__nd_driver_register); 620 621 void nvdimm_check_and_set_ro(struct gendisk *disk) 622 { 623 struct device *dev = disk_to_dev(disk)->parent; 624 struct nd_region *nd_region = to_nd_region(dev->parent); 625 int disk_ro = get_disk_ro(disk); 626 627 /* catch the disk up with the region ro state */ 628 if (disk_ro == nd_region->ro) 629 return; 630 631 dev_info(dev, "%s read-%s, marking %s read-%s\n", 632 dev_name(&nd_region->dev), nd_region->ro ? "only" : "write", 633 disk->disk_name, nd_region->ro ? "only" : "write"); 634 set_disk_ro(disk, nd_region->ro); 635 } 636 EXPORT_SYMBOL(nvdimm_check_and_set_ro); 637 638 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr, 639 char *buf) 640 { 641 return sprintf(buf, ND_DEVICE_MODALIAS_FMT "\n", 642 to_nd_device_type(dev)); 643 } 644 static DEVICE_ATTR_RO(modalias); 645 646 static ssize_t devtype_show(struct device *dev, struct device_attribute *attr, 647 char *buf) 648 { 649 return sprintf(buf, "%s\n", dev->type->name); 650 } 651 static DEVICE_ATTR_RO(devtype); 652 653 static struct attribute *nd_device_attributes[] = { 654 &dev_attr_modalias.attr, 655 &dev_attr_devtype.attr, 656 NULL, 657 }; 658 659 /* 660 * nd_device_attribute_group - generic attributes for all devices on an nd bus 661 */ 662 const struct attribute_group nd_device_attribute_group = { 663 .attrs = nd_device_attributes, 664 }; 665 666 static ssize_t numa_node_show(struct device *dev, 667 struct device_attribute *attr, char *buf) 668 { 669 return sprintf(buf, "%d\n", dev_to_node(dev)); 670 } 671 static DEVICE_ATTR_RO(numa_node); 672 673 static int nvdimm_dev_to_target_node(struct device *dev) 674 { 675 struct device *parent = dev->parent; 676 struct nd_region *nd_region = NULL; 677 678 if (is_nd_region(dev)) 679 nd_region = to_nd_region(dev); 680 else if (parent && is_nd_region(parent)) 681 nd_region = to_nd_region(parent); 682 683 if (!nd_region) 684 return NUMA_NO_NODE; 685 return nd_region->target_node; 686 } 687 688 static ssize_t target_node_show(struct device *dev, 689 struct device_attribute *attr, char *buf) 690 { 691 return sprintf(buf, "%d\n", nvdimm_dev_to_target_node(dev)); 692 } 693 static DEVICE_ATTR_RO(target_node); 694 695 static struct attribute *nd_numa_attributes[] = { 696 &dev_attr_numa_node.attr, 697 &dev_attr_target_node.attr, 698 NULL, 699 }; 700 701 static umode_t nd_numa_attr_visible(struct kobject *kobj, struct attribute *a, 702 int n) 703 { 704 struct device *dev = container_of(kobj, typeof(*dev), kobj); 705 706 if (!IS_ENABLED(CONFIG_NUMA)) 707 return 0; 708 709 if (a == &dev_attr_target_node.attr && 710 nvdimm_dev_to_target_node(dev) == NUMA_NO_NODE) 711 return 0; 712 713 return a->mode; 714 } 715 716 /* 717 * nd_numa_attribute_group - NUMA attributes for all devices on an nd bus 718 */ 719 const struct attribute_group nd_numa_attribute_group = { 720 .attrs = nd_numa_attributes, 721 .is_visible = nd_numa_attr_visible, 722 }; 723 724 static void ndctl_release(struct device *dev) 725 { 726 kfree(dev); 727 } 728 729 int nvdimm_bus_create_ndctl(struct nvdimm_bus *nvdimm_bus) 730 { 731 dev_t devt = MKDEV(nvdimm_bus_major, nvdimm_bus->id); 732 struct device *dev; 733 int rc; 734 735 dev = kzalloc(sizeof(*dev), GFP_KERNEL); 736 if (!dev) 737 return -ENOMEM; 738 device_initialize(dev); 739 device_set_pm_not_required(dev); 740 dev->class = nd_class; 741 dev->parent = &nvdimm_bus->dev; 742 dev->devt = devt; 743 dev->release = ndctl_release; 744 rc = dev_set_name(dev, "ndctl%d", nvdimm_bus->id); 745 if (rc) 746 goto err; 747 748 rc = device_add(dev); 749 if (rc) { 750 dev_dbg(&nvdimm_bus->dev, "failed to register ndctl%d: %d\n", 751 nvdimm_bus->id, rc); 752 goto err; 753 } 754 return 0; 755 756 err: 757 put_device(dev); 758 return rc; 759 } 760 761 void nvdimm_bus_destroy_ndctl(struct nvdimm_bus *nvdimm_bus) 762 { 763 device_destroy(nd_class, MKDEV(nvdimm_bus_major, nvdimm_bus->id)); 764 } 765 766 static const struct nd_cmd_desc __nd_cmd_dimm_descs[] = { 767 [ND_CMD_IMPLEMENTED] = { }, 768 [ND_CMD_SMART] = { 769 .out_num = 2, 770 .out_sizes = { 4, 128, }, 771 }, 772 [ND_CMD_SMART_THRESHOLD] = { 773 .out_num = 2, 774 .out_sizes = { 4, 8, }, 775 }, 776 [ND_CMD_DIMM_FLAGS] = { 777 .out_num = 2, 778 .out_sizes = { 4, 4 }, 779 }, 780 [ND_CMD_GET_CONFIG_SIZE] = { 781 .out_num = 3, 782 .out_sizes = { 4, 4, 4, }, 783 }, 784 [ND_CMD_GET_CONFIG_DATA] = { 785 .in_num = 2, 786 .in_sizes = { 4, 4, }, 787 .out_num = 2, 788 .out_sizes = { 4, UINT_MAX, }, 789 }, 790 [ND_CMD_SET_CONFIG_DATA] = { 791 .in_num = 3, 792 .in_sizes = { 4, 4, UINT_MAX, }, 793 .out_num = 1, 794 .out_sizes = { 4, }, 795 }, 796 [ND_CMD_VENDOR] = { 797 .in_num = 3, 798 .in_sizes = { 4, 4, UINT_MAX, }, 799 .out_num = 3, 800 .out_sizes = { 4, 4, UINT_MAX, }, 801 }, 802 [ND_CMD_CALL] = { 803 .in_num = 2, 804 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, }, 805 .out_num = 1, 806 .out_sizes = { UINT_MAX, }, 807 }, 808 }; 809 810 const struct nd_cmd_desc *nd_cmd_dimm_desc(int cmd) 811 { 812 if (cmd < ARRAY_SIZE(__nd_cmd_dimm_descs)) 813 return &__nd_cmd_dimm_descs[cmd]; 814 return NULL; 815 } 816 EXPORT_SYMBOL_GPL(nd_cmd_dimm_desc); 817 818 static const struct nd_cmd_desc __nd_cmd_bus_descs[] = { 819 [ND_CMD_IMPLEMENTED] = { }, 820 [ND_CMD_ARS_CAP] = { 821 .in_num = 2, 822 .in_sizes = { 8, 8, }, 823 .out_num = 4, 824 .out_sizes = { 4, 4, 4, 4, }, 825 }, 826 [ND_CMD_ARS_START] = { 827 .in_num = 5, 828 .in_sizes = { 8, 8, 2, 1, 5, }, 829 .out_num = 2, 830 .out_sizes = { 4, 4, }, 831 }, 832 [ND_CMD_ARS_STATUS] = { 833 .out_num = 3, 834 .out_sizes = { 4, 4, UINT_MAX, }, 835 }, 836 [ND_CMD_CLEAR_ERROR] = { 837 .in_num = 2, 838 .in_sizes = { 8, 8, }, 839 .out_num = 3, 840 .out_sizes = { 4, 4, 8, }, 841 }, 842 [ND_CMD_CALL] = { 843 .in_num = 2, 844 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, }, 845 .out_num = 1, 846 .out_sizes = { UINT_MAX, }, 847 }, 848 }; 849 850 const struct nd_cmd_desc *nd_cmd_bus_desc(int cmd) 851 { 852 if (cmd < ARRAY_SIZE(__nd_cmd_bus_descs)) 853 return &__nd_cmd_bus_descs[cmd]; 854 return NULL; 855 } 856 EXPORT_SYMBOL_GPL(nd_cmd_bus_desc); 857 858 u32 nd_cmd_in_size(struct nvdimm *nvdimm, int cmd, 859 const struct nd_cmd_desc *desc, int idx, void *buf) 860 { 861 if (idx >= desc->in_num) 862 return UINT_MAX; 863 864 if (desc->in_sizes[idx] < UINT_MAX) 865 return desc->in_sizes[idx]; 866 867 if (nvdimm && cmd == ND_CMD_SET_CONFIG_DATA && idx == 2) { 868 struct nd_cmd_set_config_hdr *hdr = buf; 869 870 return hdr->in_length; 871 } else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) { 872 struct nd_cmd_vendor_hdr *hdr = buf; 873 874 return hdr->in_length; 875 } else if (cmd == ND_CMD_CALL) { 876 struct nd_cmd_pkg *pkg = buf; 877 878 return pkg->nd_size_in; 879 } 880 881 return UINT_MAX; 882 } 883 EXPORT_SYMBOL_GPL(nd_cmd_in_size); 884 885 u32 nd_cmd_out_size(struct nvdimm *nvdimm, int cmd, 886 const struct nd_cmd_desc *desc, int idx, const u32 *in_field, 887 const u32 *out_field, unsigned long remainder) 888 { 889 if (idx >= desc->out_num) 890 return UINT_MAX; 891 892 if (desc->out_sizes[idx] < UINT_MAX) 893 return desc->out_sizes[idx]; 894 895 if (nvdimm && cmd == ND_CMD_GET_CONFIG_DATA && idx == 1) 896 return in_field[1]; 897 else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) 898 return out_field[1]; 899 else if (!nvdimm && cmd == ND_CMD_ARS_STATUS && idx == 2) { 900 /* 901 * Per table 9-276 ARS Data in ACPI 6.1, out_field[1] is 902 * "Size of Output Buffer in bytes, including this 903 * field." 904 */ 905 if (out_field[1] < 4) 906 return 0; 907 /* 908 * ACPI 6.1 is ambiguous if 'status' is included in the 909 * output size. If we encounter an output size that 910 * overshoots the remainder by 4 bytes, assume it was 911 * including 'status'. 912 */ 913 if (out_field[1] - 4 == remainder) 914 return remainder; 915 return out_field[1] - 8; 916 } else if (cmd == ND_CMD_CALL) { 917 struct nd_cmd_pkg *pkg = (struct nd_cmd_pkg *) in_field; 918 919 return pkg->nd_size_out; 920 } 921 922 923 return UINT_MAX; 924 } 925 EXPORT_SYMBOL_GPL(nd_cmd_out_size); 926 927 void wait_nvdimm_bus_probe_idle(struct device *dev) 928 { 929 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev); 930 931 do { 932 if (nvdimm_bus->probe_active == 0) 933 break; 934 nvdimm_bus_unlock(dev); 935 nd_device_unlock(dev); 936 wait_event(nvdimm_bus->wait, 937 nvdimm_bus->probe_active == 0); 938 nd_device_lock(dev); 939 nvdimm_bus_lock(dev); 940 } while (true); 941 } 942 943 static int nd_pmem_forget_poison_check(struct device *dev, void *data) 944 { 945 struct nd_cmd_clear_error *clear_err = 946 (struct nd_cmd_clear_error *)data; 947 struct nd_btt *nd_btt = is_nd_btt(dev) ? to_nd_btt(dev) : NULL; 948 struct nd_pfn *nd_pfn = is_nd_pfn(dev) ? to_nd_pfn(dev) : NULL; 949 struct nd_dax *nd_dax = is_nd_dax(dev) ? to_nd_dax(dev) : NULL; 950 struct nd_namespace_common *ndns = NULL; 951 struct nd_namespace_io *nsio; 952 resource_size_t offset = 0, end_trunc = 0, start, end, pstart, pend; 953 954 if (nd_dax || !dev->driver) 955 return 0; 956 957 start = clear_err->address; 958 end = clear_err->address + clear_err->cleared - 1; 959 960 if (nd_btt || nd_pfn || nd_dax) { 961 if (nd_btt) 962 ndns = nd_btt->ndns; 963 else if (nd_pfn) 964 ndns = nd_pfn->ndns; 965 else if (nd_dax) 966 ndns = nd_dax->nd_pfn.ndns; 967 968 if (!ndns) 969 return 0; 970 } else 971 ndns = to_ndns(dev); 972 973 nsio = to_nd_namespace_io(&ndns->dev); 974 pstart = nsio->res.start + offset; 975 pend = nsio->res.end - end_trunc; 976 977 if ((pstart >= start) && (pend <= end)) 978 return -EBUSY; 979 980 return 0; 981 982 } 983 984 static int nd_ns_forget_poison_check(struct device *dev, void *data) 985 { 986 return device_for_each_child(dev, data, nd_pmem_forget_poison_check); 987 } 988 989 /* set_config requires an idle interleave set */ 990 static int nd_cmd_clear_to_send(struct nvdimm_bus *nvdimm_bus, 991 struct nvdimm *nvdimm, unsigned int cmd, void *data) 992 { 993 struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc; 994 995 /* ask the bus provider if it would like to block this request */ 996 if (nd_desc->clear_to_send) { 997 int rc = nd_desc->clear_to_send(nd_desc, nvdimm, cmd, data); 998 999 if (rc) 1000 return rc; 1001 } 1002 1003 /* require clear error to go through the pmem driver */ 1004 if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR) 1005 return device_for_each_child(&nvdimm_bus->dev, data, 1006 nd_ns_forget_poison_check); 1007 1008 if (!nvdimm || cmd != ND_CMD_SET_CONFIG_DATA) 1009 return 0; 1010 1011 /* prevent label manipulation while the kernel owns label updates */ 1012 wait_nvdimm_bus_probe_idle(&nvdimm_bus->dev); 1013 if (atomic_read(&nvdimm->busy)) 1014 return -EBUSY; 1015 return 0; 1016 } 1017 1018 static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm, 1019 int read_only, unsigned int ioctl_cmd, unsigned long arg) 1020 { 1021 struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc; 1022 const struct nd_cmd_desc *desc = NULL; 1023 unsigned int cmd = _IOC_NR(ioctl_cmd); 1024 struct device *dev = &nvdimm_bus->dev; 1025 void __user *p = (void __user *) arg; 1026 char *out_env = NULL, *in_env = NULL; 1027 const char *cmd_name, *dimm_name; 1028 u32 in_len = 0, out_len = 0; 1029 unsigned int func = cmd; 1030 unsigned long cmd_mask; 1031 struct nd_cmd_pkg pkg; 1032 int rc, i, cmd_rc; 1033 void *buf = NULL; 1034 u64 buf_len = 0; 1035 1036 if (nvdimm) { 1037 desc = nd_cmd_dimm_desc(cmd); 1038 cmd_name = nvdimm_cmd_name(cmd); 1039 cmd_mask = nvdimm->cmd_mask; 1040 dimm_name = dev_name(&nvdimm->dev); 1041 } else { 1042 desc = nd_cmd_bus_desc(cmd); 1043 cmd_name = nvdimm_bus_cmd_name(cmd); 1044 cmd_mask = nd_desc->cmd_mask; 1045 dimm_name = "bus"; 1046 } 1047 1048 /* Validate command family support against bus declared support */ 1049 if (cmd == ND_CMD_CALL) { 1050 unsigned long *mask; 1051 1052 if (copy_from_user(&pkg, p, sizeof(pkg))) 1053 return -EFAULT; 1054 1055 if (nvdimm) { 1056 if (pkg.nd_family > NVDIMM_FAMILY_MAX) 1057 return -EINVAL; 1058 mask = &nd_desc->dimm_family_mask; 1059 } else { 1060 if (pkg.nd_family > NVDIMM_BUS_FAMILY_MAX) 1061 return -EINVAL; 1062 mask = &nd_desc->bus_family_mask; 1063 } 1064 1065 if (!test_bit(pkg.nd_family, mask)) 1066 return -EINVAL; 1067 } 1068 1069 if (!desc || 1070 (desc->out_num + desc->in_num == 0) || 1071 cmd > ND_CMD_CALL || 1072 !test_bit(cmd, &cmd_mask)) 1073 return -ENOTTY; 1074 1075 /* fail write commands (when read-only) */ 1076 if (read_only) 1077 switch (cmd) { 1078 case ND_CMD_VENDOR: 1079 case ND_CMD_SET_CONFIG_DATA: 1080 case ND_CMD_ARS_START: 1081 case ND_CMD_CLEAR_ERROR: 1082 case ND_CMD_CALL: 1083 dev_dbg(dev, "'%s' command while read-only.\n", 1084 nvdimm ? nvdimm_cmd_name(cmd) 1085 : nvdimm_bus_cmd_name(cmd)); 1086 return -EPERM; 1087 default: 1088 break; 1089 } 1090 1091 /* process an input envelope */ 1092 in_env = kzalloc(ND_CMD_MAX_ENVELOPE, GFP_KERNEL); 1093 if (!in_env) 1094 return -ENOMEM; 1095 for (i = 0; i < desc->in_num; i++) { 1096 u32 in_size, copy; 1097 1098 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env); 1099 if (in_size == UINT_MAX) { 1100 dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n", 1101 __func__, dimm_name, cmd_name, i); 1102 rc = -ENXIO; 1103 goto out; 1104 } 1105 if (in_len < ND_CMD_MAX_ENVELOPE) 1106 copy = min_t(u32, ND_CMD_MAX_ENVELOPE - in_len, in_size); 1107 else 1108 copy = 0; 1109 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy)) { 1110 rc = -EFAULT; 1111 goto out; 1112 } 1113 in_len += in_size; 1114 } 1115 1116 if (cmd == ND_CMD_CALL) { 1117 func = pkg.nd_command; 1118 dev_dbg(dev, "%s, idx: %llu, in: %u, out: %u, len %llu\n", 1119 dimm_name, pkg.nd_command, 1120 in_len, out_len, buf_len); 1121 } 1122 1123 /* process an output envelope */ 1124 out_env = kzalloc(ND_CMD_MAX_ENVELOPE, GFP_KERNEL); 1125 if (!out_env) { 1126 rc = -ENOMEM; 1127 goto out; 1128 } 1129 1130 for (i = 0; i < desc->out_num; i++) { 1131 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i, 1132 (u32 *) in_env, (u32 *) out_env, 0); 1133 u32 copy; 1134 1135 if (out_size == UINT_MAX) { 1136 dev_dbg(dev, "%s unknown output size cmd: %s field: %d\n", 1137 dimm_name, cmd_name, i); 1138 rc = -EFAULT; 1139 goto out; 1140 } 1141 if (out_len < ND_CMD_MAX_ENVELOPE) 1142 copy = min_t(u32, ND_CMD_MAX_ENVELOPE - out_len, out_size); 1143 else 1144 copy = 0; 1145 if (copy && copy_from_user(&out_env[out_len], 1146 p + in_len + out_len, copy)) { 1147 rc = -EFAULT; 1148 goto out; 1149 } 1150 out_len += out_size; 1151 } 1152 1153 buf_len = (u64) out_len + (u64) in_len; 1154 if (buf_len > ND_IOCTL_MAX_BUFLEN) { 1155 dev_dbg(dev, "%s cmd: %s buf_len: %llu > %d\n", dimm_name, 1156 cmd_name, buf_len, ND_IOCTL_MAX_BUFLEN); 1157 rc = -EINVAL; 1158 goto out; 1159 } 1160 1161 buf = vmalloc(buf_len); 1162 if (!buf) { 1163 rc = -ENOMEM; 1164 goto out; 1165 } 1166 1167 if (copy_from_user(buf, p, buf_len)) { 1168 rc = -EFAULT; 1169 goto out; 1170 } 1171 1172 nd_device_lock(dev); 1173 nvdimm_bus_lock(dev); 1174 rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, func, buf); 1175 if (rc) 1176 goto out_unlock; 1177 1178 rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc); 1179 if (rc < 0) 1180 goto out_unlock; 1181 1182 if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) { 1183 struct nd_cmd_clear_error *clear_err = buf; 1184 1185 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address, 1186 clear_err->cleared); 1187 } 1188 1189 if (copy_to_user(p, buf, buf_len)) 1190 rc = -EFAULT; 1191 1192 out_unlock: 1193 nvdimm_bus_unlock(dev); 1194 nd_device_unlock(dev); 1195 out: 1196 kfree(in_env); 1197 kfree(out_env); 1198 vfree(buf); 1199 return rc; 1200 } 1201 1202 enum nd_ioctl_mode { 1203 BUS_IOCTL, 1204 DIMM_IOCTL, 1205 }; 1206 1207 static int match_dimm(struct device *dev, void *data) 1208 { 1209 long id = (long) data; 1210 1211 if (is_nvdimm(dev)) { 1212 struct nvdimm *nvdimm = to_nvdimm(dev); 1213 1214 return nvdimm->id == id; 1215 } 1216 1217 return 0; 1218 } 1219 1220 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg, 1221 enum nd_ioctl_mode mode) 1222 1223 { 1224 struct nvdimm_bus *nvdimm_bus, *found = NULL; 1225 long id = (long) file->private_data; 1226 struct nvdimm *nvdimm = NULL; 1227 int rc, ro; 1228 1229 ro = ((file->f_flags & O_ACCMODE) == O_RDONLY); 1230 mutex_lock(&nvdimm_bus_list_mutex); 1231 list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) { 1232 if (mode == DIMM_IOCTL) { 1233 struct device *dev; 1234 1235 dev = device_find_child(&nvdimm_bus->dev, 1236 file->private_data, match_dimm); 1237 if (!dev) 1238 continue; 1239 nvdimm = to_nvdimm(dev); 1240 found = nvdimm_bus; 1241 } else if (nvdimm_bus->id == id) { 1242 found = nvdimm_bus; 1243 } 1244 1245 if (found) { 1246 atomic_inc(&nvdimm_bus->ioctl_active); 1247 break; 1248 } 1249 } 1250 mutex_unlock(&nvdimm_bus_list_mutex); 1251 1252 if (!found) 1253 return -ENXIO; 1254 1255 nvdimm_bus = found; 1256 rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg); 1257 1258 if (nvdimm) 1259 put_device(&nvdimm->dev); 1260 if (atomic_dec_and_test(&nvdimm_bus->ioctl_active)) 1261 wake_up(&nvdimm_bus->wait); 1262 1263 return rc; 1264 } 1265 1266 static long bus_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 1267 { 1268 return nd_ioctl(file, cmd, arg, BUS_IOCTL); 1269 } 1270 1271 static long dimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 1272 { 1273 return nd_ioctl(file, cmd, arg, DIMM_IOCTL); 1274 } 1275 1276 static int nd_open(struct inode *inode, struct file *file) 1277 { 1278 long minor = iminor(inode); 1279 1280 file->private_data = (void *) minor; 1281 return 0; 1282 } 1283 1284 static const struct file_operations nvdimm_bus_fops = { 1285 .owner = THIS_MODULE, 1286 .open = nd_open, 1287 .unlocked_ioctl = bus_ioctl, 1288 .compat_ioctl = compat_ptr_ioctl, 1289 .llseek = noop_llseek, 1290 }; 1291 1292 static const struct file_operations nvdimm_fops = { 1293 .owner = THIS_MODULE, 1294 .open = nd_open, 1295 .unlocked_ioctl = dimm_ioctl, 1296 .compat_ioctl = compat_ptr_ioctl, 1297 .llseek = noop_llseek, 1298 }; 1299 1300 int __init nvdimm_bus_init(void) 1301 { 1302 int rc; 1303 1304 rc = bus_register(&nvdimm_bus_type); 1305 if (rc) 1306 return rc; 1307 1308 rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops); 1309 if (rc < 0) 1310 goto err_bus_chrdev; 1311 nvdimm_bus_major = rc; 1312 1313 rc = register_chrdev(0, "dimmctl", &nvdimm_fops); 1314 if (rc < 0) 1315 goto err_dimm_chrdev; 1316 nvdimm_major = rc; 1317 1318 nd_class = class_create(THIS_MODULE, "nd"); 1319 if (IS_ERR(nd_class)) { 1320 rc = PTR_ERR(nd_class); 1321 goto err_class; 1322 } 1323 1324 rc = driver_register(&nd_bus_driver.drv); 1325 if (rc) 1326 goto err_nd_bus; 1327 1328 return 0; 1329 1330 err_nd_bus: 1331 class_destroy(nd_class); 1332 err_class: 1333 unregister_chrdev(nvdimm_major, "dimmctl"); 1334 err_dimm_chrdev: 1335 unregister_chrdev(nvdimm_bus_major, "ndctl"); 1336 err_bus_chrdev: 1337 bus_unregister(&nvdimm_bus_type); 1338 1339 return rc; 1340 } 1341 1342 void nvdimm_bus_exit(void) 1343 { 1344 driver_unregister(&nd_bus_driver.drv); 1345 class_destroy(nd_class); 1346 unregister_chrdev(nvdimm_bus_major, "ndctl"); 1347 unregister_chrdev(nvdimm_major, "dimmctl"); 1348 bus_unregister(&nvdimm_bus_type); 1349 ida_destroy(&nd_ida); 1350 } 1351