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