drivers/nvdimm/bus.c

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