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