drivers/nvdimm/bus.c

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