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/vmalloc.h>
  15 #include <linux/uaccess.h>
  16 #include <linux/module.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/fcntl.h>
  19 #include <linux/async.h>
  20 #include <linux/genhd.h>
  21 #include <linux/ndctl.h>
  22 #include <linux/sched.h>
  23 #include <linux/slab.h>
  24 #include <linux/fs.h>
  25 #include <linux/io.h>
  26 #include <linux/mm.h>
  27 #include <linux/nd.h>
  28 #include "nd-core.h"
  29 #include "nd.h"
  30 #include "pfn.h"
  31
  32 int nvdimm_major;
  33 static int nvdimm_bus_major;
  34 static struct class *nd_class;
  35 static DEFINE_IDA(nd_ida);
  36
  37 static int to_nd_device_type(struct device *dev)
  38 {
  39         if (is_nvdimm(dev))
  40                 return ND_DEVICE_DIMM;
  41         else if (is_memory(dev))
  42                 return ND_DEVICE_REGION_PMEM;
  43         else if (is_nd_blk(dev))
  44                 return ND_DEVICE_REGION_BLK;
  45         else if (is_nd_dax(dev))
  46                 return ND_DEVICE_DAX_PMEM;
  47         else if (is_nd_region(dev->parent))
  48                 return nd_region_to_nstype(to_nd_region(dev->parent));
  49
  50         return 0;
  51 }
  52
  53 static int nvdimm_bus_uevent(struct device *dev, struct kobj_uevent_env *env)
  54 {
  55         /*
  56          * Ensure that region devices always have their numa node set as
  57          * early as possible.
  58          */
  59         if (is_nd_region(dev))
  60                 set_dev_node(dev, to_nd_region(dev)->numa_node);
  61         return add_uevent_var(env, "MODALIAS=" ND_DEVICE_MODALIAS_FMT,
  62                         to_nd_device_type(dev));
  63 }
  64
  65 static struct module *to_bus_provider(struct device *dev)
  66 {
  67         /* pin bus providers while regions are enabled */
  68         if (is_nd_region(dev)) {
  69                 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
  70
  71                 return nvdimm_bus->nd_desc->module;
  72         }
  73         return NULL;
  74 }
  75
  76 static void nvdimm_bus_probe_start(struct nvdimm_bus *nvdimm_bus)
  77 {
  78         nvdimm_bus_lock(&nvdimm_bus->dev);
  79         nvdimm_bus->probe_active++;
  80         nvdimm_bus_unlock(&nvdimm_bus->dev);
  81 }
  82
  83 static void nvdimm_bus_probe_end(struct nvdimm_bus *nvdimm_bus)
  84 {
  85         nvdimm_bus_lock(&nvdimm_bus->dev);
  86         if (--nvdimm_bus->probe_active == 0)
  87                 wake_up(&nvdimm_bus->probe_wait);
  88         nvdimm_bus_unlock(&nvdimm_bus->dev);
  89 }
  90
  91 static int nvdimm_bus_probe(struct device *dev)
  92 {
  93         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
  94         struct module *provider = to_bus_provider(dev);
  95         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
  96         int rc;
  97
  98         if (!try_module_get(provider))
  99                 return -ENXIO;
 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, "%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                 nvdimm_forget_poison(nvdimm_bus, 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         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         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_ARS_CAP, &ars_cap,
 254                         sizeof(ars_cap), &cmd_rc);
 255         if (rc < 0)
 256                 return rc;
 257         if (cmd_rc < 0)
 258                 return cmd_rc;
 259         clear_err_unit = ars_cap.clear_err_unit;
 260         if (!clear_err_unit || !is_power_of_2(clear_err_unit))
 261                 return -ENXIO;
 262
 263         mask = clear_err_unit - 1;
 264         if ((phys | len) & mask)
 265                 return -ENXIO;
 266         memset(&clear_err, 0, sizeof(clear_err));
 267         clear_err.address = phys;
 268         clear_err.length = len;
 269         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_CLEAR_ERROR, &clear_err,
 270                         sizeof(clear_err), &cmd_rc);
 271         if (rc < 0)
 272                 return rc;
 273         if (cmd_rc < 0)
 274                 return cmd_rc;
 275
 276         nvdimm_account_cleared_poison(nvdimm_bus, phys, clear_err.cleared);
 277
 278         return clear_err.cleared;
 279 }
 280 EXPORT_SYMBOL_GPL(nvdimm_clear_poison);
 281
 282 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv);
 283
 284 static struct bus_type nvdimm_bus_type = {
 285         .name = "nd",
 286         .uevent = nvdimm_bus_uevent,
 287         .match = nvdimm_bus_match,
 288         .probe = nvdimm_bus_probe,
 289         .remove = nvdimm_bus_remove,
 290         .shutdown = nvdimm_bus_shutdown,
 291 };
 292
 293 static void nvdimm_bus_release(struct device *dev)
 294 {
 295         struct nvdimm_bus *nvdimm_bus;
 296
 297         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 298         ida_simple_remove(&nd_ida, nvdimm_bus->id);
 299         kfree(nvdimm_bus);
 300 }
 301
 302 static bool is_nvdimm_bus(struct device *dev)
 303 {
 304         return dev->release == nvdimm_bus_release;
 305 }
 306
 307 struct nvdimm_bus *walk_to_nvdimm_bus(struct device *nd_dev)
 308 {
 309         struct device *dev;
 310
 311         for (dev = nd_dev; dev; dev = dev->parent)
 312                 if (is_nvdimm_bus(dev))
 313                         break;
 314         dev_WARN_ONCE(nd_dev, !dev, "invalid dev, not on nd bus\n");
 315         if (dev)
 316                 return to_nvdimm_bus(dev);
 317         return NULL;
 318 }
 319
 320 struct nvdimm_bus *to_nvdimm_bus(struct device *dev)
 321 {
 322         struct nvdimm_bus *nvdimm_bus;
 323
 324         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 325         WARN_ON(!is_nvdimm_bus(dev));
 326         return nvdimm_bus;
 327 }
 328 EXPORT_SYMBOL_GPL(to_nvdimm_bus);
 329
 330 struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
 331                 struct nvdimm_bus_descriptor *nd_desc)
 332 {
 333         struct nvdimm_bus *nvdimm_bus;
 334         int rc;
 335
 336         nvdimm_bus = kzalloc(sizeof(*nvdimm_bus), GFP_KERNEL);
 337         if (!nvdimm_bus)
 338                 return NULL;
 339         INIT_LIST_HEAD(&nvdimm_bus->list);
 340         INIT_LIST_HEAD(&nvdimm_bus->mapping_list);
 341         INIT_LIST_HEAD(&nvdimm_bus->poison_list);
 342         init_waitqueue_head(&nvdimm_bus->probe_wait);
 343         nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL);
 344         mutex_init(&nvdimm_bus->reconfig_mutex);
 345         spin_lock_init(&nvdimm_bus->poison_lock);
 346         if (nvdimm_bus->id < 0) {
 347                 kfree(nvdimm_bus);
 348                 return NULL;
 349         }
 350         nvdimm_bus->nd_desc = nd_desc;
 351         nvdimm_bus->dev.parent = parent;
 352         nvdimm_bus->dev.release = nvdimm_bus_release;
 353         nvdimm_bus->dev.groups = nd_desc->attr_groups;
 354         nvdimm_bus->dev.bus = &nvdimm_bus_type;
 355         dev_set_name(&nvdimm_bus->dev, "ndbus%d", nvdimm_bus->id);
 356         rc = device_register(&nvdimm_bus->dev);
 357         if (rc) {
 358                 dev_dbg(&nvdimm_bus->dev, "registration failed: %d\n", rc);
 359                 goto err;
 360         }
 361
 362         return nvdimm_bus;
 363  err:
 364         put_device(&nvdimm_bus->dev);
 365         return NULL;
 366 }
 367 EXPORT_SYMBOL_GPL(nvdimm_bus_register);
 368
 369 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus)
 370 {
 371         if (!nvdimm_bus)
 372                 return;
 373         device_unregister(&nvdimm_bus->dev);
 374 }
 375 EXPORT_SYMBOL_GPL(nvdimm_bus_unregister);
 376
 377 static int child_unregister(struct device *dev, void *data)
 378 {
 379         /*
 380          * the singular ndctl class device per bus needs to be
 381          * "device_destroy"ed, so skip it here
 382          *
 383          * i.e. remove classless children
 384          */
 385         if (dev->class)
 386                 /* pass */;
 387         else
 388                 nd_device_unregister(dev, ND_SYNC);
 389         return 0;
 390 }
 391
 392 static void free_poison_list(struct list_head *poison_list)
 393 {
 394         struct nd_poison *pl, *next;
 395
 396         list_for_each_entry_safe(pl, next, poison_list, list) {
 397                 list_del(&pl->list);
 398                 kfree(pl);
 399         }
 400         list_del_init(poison_list);
 401 }
 402
 403 static int nd_bus_remove(struct device *dev)
 404 {
 405         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 406
 407         mutex_lock(&nvdimm_bus_list_mutex);
 408         list_del_init(&nvdimm_bus->list);
 409         mutex_unlock(&nvdimm_bus_list_mutex);
 410
 411         nd_synchronize();
 412         device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister);
 413
 414         spin_lock(&nvdimm_bus->poison_lock);
 415         free_poison_list(&nvdimm_bus->poison_list);
 416         spin_unlock(&nvdimm_bus->poison_lock);
 417
 418         nvdimm_bus_destroy_ndctl(nvdimm_bus);
 419
 420         return 0;
 421 }
 422
 423 static int nd_bus_probe(struct device *dev)
 424 {
 425         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 426         int rc;
 427
 428         rc = nvdimm_bus_create_ndctl(nvdimm_bus);
 429         if (rc)
 430                 return rc;
 431
 432         mutex_lock(&nvdimm_bus_list_mutex);
 433         list_add_tail(&nvdimm_bus->list, &nvdimm_bus_list);
 434         mutex_unlock(&nvdimm_bus_list_mutex);
 435
 436         /* enable bus provider attributes to look up their local context */
 437         dev_set_drvdata(dev, nvdimm_bus->nd_desc);
 438
 439         return 0;
 440 }
 441
 442 static struct nd_device_driver nd_bus_driver = {
 443         .probe = nd_bus_probe,
 444         .remove = nd_bus_remove,
 445         .drv = {
 446                 .name = "nd_bus",
 447                 .suppress_bind_attrs = true,
 448                 .bus = &nvdimm_bus_type,
 449                 .owner = THIS_MODULE,
 450                 .mod_name = KBUILD_MODNAME,
 451         },
 452 };
 453
 454 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv)
 455 {
 456         struct nd_device_driver *nd_drv = to_nd_device_driver(drv);
 457
 458         if (is_nvdimm_bus(dev) && nd_drv == &nd_bus_driver)
 459                 return true;
 460
 461         return !!test_bit(to_nd_device_type(dev), &nd_drv->type);
 462 }
 463
 464 static ASYNC_DOMAIN_EXCLUSIVE(nd_async_domain);
 465
 466 void nd_synchronize(void)
 467 {
 468         async_synchronize_full_domain(&nd_async_domain);
 469 }
 470 EXPORT_SYMBOL_GPL(nd_synchronize);
 471
 472 static void nd_async_device_register(void *d, async_cookie_t cookie)
 473 {
 474         struct device *dev = d;
 475
 476         if (device_add(dev) != 0) {
 477                 dev_err(dev, "%s: failed\n", __func__);
 478                 put_device(dev);
 479         }
 480         put_device(dev);
 481 }
 482
 483 static void nd_async_device_unregister(void *d, async_cookie_t cookie)
 484 {
 485         struct device *dev = d;
 486
 487         /* flush bus operations before delete */
 488         nvdimm_bus_lock(dev);
 489         nvdimm_bus_unlock(dev);
 490
 491         device_unregister(dev);
 492         put_device(dev);
 493 }
 494
 495 void __nd_device_register(struct device *dev)
 496 {
 497         if (!dev)
 498                 return;
 499         dev->bus = &nvdimm_bus_type;
 500         get_device(dev);
 501         async_schedule_domain(nd_async_device_register, dev,
 502                         &nd_async_domain);
 503 }
 504
 505 void nd_device_register(struct device *dev)
 506 {
 507         device_initialize(dev);
 508         __nd_device_register(dev);
 509 }
 510 EXPORT_SYMBOL(nd_device_register);
 511
 512 void nd_device_unregister(struct device *dev, enum nd_async_mode mode)
 513 {
 514         switch (mode) {
 515         case ND_ASYNC:
 516                 get_device(dev);
 517                 async_schedule_domain(nd_async_device_unregister, dev,
 518                                 &nd_async_domain);
 519                 break;
 520         case ND_SYNC:
 521                 nd_synchronize();
 522                 device_unregister(dev);
 523                 break;
 524         }
 525 }
 526 EXPORT_SYMBOL(nd_device_unregister);
 527
 528 /**
 529  * __nd_driver_register() - register a region or a namespace driver
 530  * @nd_drv: driver to register
 531  * @owner: automatically set by nd_driver_register() macro
 532  * @mod_name: automatically set by nd_driver_register() macro
 533  */
 534 int __nd_driver_register(struct nd_device_driver *nd_drv, struct module *owner,
 535                 const char *mod_name)
 536 {
 537         struct device_driver *drv = &nd_drv->drv;
 538
 539         if (!nd_drv->type) {
 540                 pr_debug("driver type bitmask not set (%pf)\n",
 541                                 __builtin_return_address(0));
 542                 return -EINVAL;
 543         }
 544
 545         if (!nd_drv->probe) {
 546                 pr_debug("%s ->probe() must be specified\n", mod_name);
 547                 return -EINVAL;
 548         }
 549
 550         drv->bus = &nvdimm_bus_type;
 551         drv->owner = owner;
 552         drv->mod_name = mod_name;
 553
 554         return driver_register(drv);
 555 }
 556 EXPORT_SYMBOL(__nd_driver_register);
 557
 558 int nvdimm_revalidate_disk(struct gendisk *disk)
 559 {
 560         struct device *dev = disk_to_dev(disk)->parent;
 561         struct nd_region *nd_region = to_nd_region(dev->parent);
 562         const char *pol = nd_region->ro ? "only" : "write";
 563
 564         if (nd_region->ro == get_disk_ro(disk))
 565                 return 0;
 566
 567         dev_info(dev, "%s read-%s, marking %s read-%s\n",
 568                         dev_name(&nd_region->dev), pol, disk->disk_name, pol);
 569         set_disk_ro(disk, nd_region->ro);
 570
 571         return 0;
 572
 573 }
 574 EXPORT_SYMBOL(nvdimm_revalidate_disk);
 575
 576 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr,
 577                 char *buf)
 578 {
 579         return sprintf(buf, ND_DEVICE_MODALIAS_FMT "\n",
 580                         to_nd_device_type(dev));
 581 }
 582 static DEVICE_ATTR_RO(modalias);
 583
 584 static ssize_t devtype_show(struct device *dev, struct device_attribute *attr,
 585                 char *buf)
 586 {
 587         return sprintf(buf, "%s\n", dev->type->name);
 588 }
 589 static DEVICE_ATTR_RO(devtype);
 590
 591 static struct attribute *nd_device_attributes[] = {
 592         &dev_attr_modalias.attr,
 593         &dev_attr_devtype.attr,
 594         NULL,
 595 };
 596
 597 /**
 598  * nd_device_attribute_group - generic attributes for all devices on an nd bus
 599  */
 600 struct attribute_group nd_device_attribute_group = {
 601         .attrs = nd_device_attributes,
 602 };
 603 EXPORT_SYMBOL_GPL(nd_device_attribute_group);
 604
 605 static ssize_t numa_node_show(struct device *dev,
 606                 struct device_attribute *attr, char *buf)
 607 {
 608         return sprintf(buf, "%d\n", dev_to_node(dev));
 609 }
 610 static DEVICE_ATTR_RO(numa_node);
 611
 612 static struct attribute *nd_numa_attributes[] = {
 613         &dev_attr_numa_node.attr,
 614         NULL,
 615 };
 616
 617 static umode_t nd_numa_attr_visible(struct kobject *kobj, struct attribute *a,
 618                 int n)
 619 {
 620         if (!IS_ENABLED(CONFIG_NUMA))
 621                 return 0;
 622
 623         return a->mode;
 624 }
 625
 626 /**
 627  * nd_numa_attribute_group - NUMA attributes for all devices on an nd bus
 628  */
 629 struct attribute_group nd_numa_attribute_group = {
 630         .attrs = nd_numa_attributes,
 631         .is_visible = nd_numa_attr_visible,
 632 };
 633 EXPORT_SYMBOL_GPL(nd_numa_attribute_group);
 634
 635 int nvdimm_bus_create_ndctl(struct nvdimm_bus *nvdimm_bus)
 636 {
 637         dev_t devt = MKDEV(nvdimm_bus_major, nvdimm_bus->id);
 638         struct device *dev;
 639
 640         dev = device_create(nd_class, &nvdimm_bus->dev, devt, nvdimm_bus,
 641                         "ndctl%d", nvdimm_bus->id);
 642
 643         if (IS_ERR(dev))
 644                 dev_dbg(&nvdimm_bus->dev, "failed to register ndctl%d: %ld\n",
 645                                 nvdimm_bus->id, PTR_ERR(dev));
 646         return PTR_ERR_OR_ZERO(dev);
 647 }
 648
 649 void nvdimm_bus_destroy_ndctl(struct nvdimm_bus *nvdimm_bus)
 650 {
 651         device_destroy(nd_class, MKDEV(nvdimm_bus_major, nvdimm_bus->id));
 652 }
 653
 654 static const struct nd_cmd_desc __nd_cmd_dimm_descs[] = {
 655         [ND_CMD_IMPLEMENTED] = { },
 656         [ND_CMD_SMART] = {
 657                 .out_num = 2,
 658                 .out_sizes = { 4, 128, },
 659         },
 660         [ND_CMD_SMART_THRESHOLD] = {
 661                 .out_num = 2,
 662                 .out_sizes = { 4, 8, },
 663         },
 664         [ND_CMD_DIMM_FLAGS] = {
 665                 .out_num = 2,
 666                 .out_sizes = { 4, 4 },
 667         },
 668         [ND_CMD_GET_CONFIG_SIZE] = {
 669                 .out_num = 3,
 670                 .out_sizes = { 4, 4, 4, },
 671         },
 672         [ND_CMD_GET_CONFIG_DATA] = {
 673                 .in_num = 2,
 674                 .in_sizes = { 4, 4, },
 675                 .out_num = 2,
 676                 .out_sizes = { 4, UINT_MAX, },
 677         },
 678         [ND_CMD_SET_CONFIG_DATA] = {
 679                 .in_num = 3,
 680                 .in_sizes = { 4, 4, UINT_MAX, },
 681                 .out_num = 1,
 682                 .out_sizes = { 4, },
 683         },
 684         [ND_CMD_VENDOR] = {
 685                 .in_num = 3,
 686                 .in_sizes = { 4, 4, UINT_MAX, },
 687                 .out_num = 3,
 688                 .out_sizes = { 4, 4, UINT_MAX, },
 689         },
 690         [ND_CMD_CALL] = {
 691                 .in_num = 2,
 692                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 693                 .out_num = 1,
 694                 .out_sizes = { UINT_MAX, },
 695         },
 696 };
 697
 698 const struct nd_cmd_desc *nd_cmd_dimm_desc(int cmd)
 699 {
 700         if (cmd < ARRAY_SIZE(__nd_cmd_dimm_descs))
 701                 return &__nd_cmd_dimm_descs[cmd];
 702         return NULL;
 703 }
 704 EXPORT_SYMBOL_GPL(nd_cmd_dimm_desc);
 705
 706 static const struct nd_cmd_desc __nd_cmd_bus_descs[] = {
 707         [ND_CMD_IMPLEMENTED] = { },
 708         [ND_CMD_ARS_CAP] = {
 709                 .in_num = 2,
 710                 .in_sizes = { 8, 8, },
 711                 .out_num = 4,
 712                 .out_sizes = { 4, 4, 4, 4, },
 713         },
 714         [ND_CMD_ARS_START] = {
 715                 .in_num = 5,
 716                 .in_sizes = { 8, 8, 2, 1, 5, },
 717                 .out_num = 2,
 718                 .out_sizes = { 4, 4, },
 719         },
 720         [ND_CMD_ARS_STATUS] = {
 721                 .out_num = 3,
 722                 .out_sizes = { 4, 4, UINT_MAX, },
 723         },
 724         [ND_CMD_CLEAR_ERROR] = {
 725                 .in_num = 2,
 726                 .in_sizes = { 8, 8, },
 727                 .out_num = 3,
 728                 .out_sizes = { 4, 4, 8, },
 729         },
 730         [ND_CMD_CALL] = {
 731                 .in_num = 2,
 732                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 733                 .out_num = 1,
 734                 .out_sizes = { UINT_MAX, },
 735         },
 736 };
 737
 738 const struct nd_cmd_desc *nd_cmd_bus_desc(int cmd)
 739 {
 740         if (cmd < ARRAY_SIZE(__nd_cmd_bus_descs))
 741                 return &__nd_cmd_bus_descs[cmd];
 742         return NULL;
 743 }
 744 EXPORT_SYMBOL_GPL(nd_cmd_bus_desc);
 745
 746 u32 nd_cmd_in_size(struct nvdimm *nvdimm, int cmd,
 747                 const struct nd_cmd_desc *desc, int idx, void *buf)
 748 {
 749         if (idx >= desc->in_num)
 750                 return UINT_MAX;
 751
 752         if (desc->in_sizes[idx] < UINT_MAX)
 753                 return desc->in_sizes[idx];
 754
 755         if (nvdimm && cmd == ND_CMD_SET_CONFIG_DATA && idx == 2) {
 756                 struct nd_cmd_set_config_hdr *hdr = buf;
 757
 758                 return hdr->in_length;
 759         } else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) {
 760                 struct nd_cmd_vendor_hdr *hdr = buf;
 761
 762                 return hdr->in_length;
 763         } else if (cmd == ND_CMD_CALL) {
 764                 struct nd_cmd_pkg *pkg = buf;
 765
 766                 return pkg->nd_size_in;
 767         }
 768
 769         return UINT_MAX;
 770 }
 771 EXPORT_SYMBOL_GPL(nd_cmd_in_size);
 772
 773 u32 nd_cmd_out_size(struct nvdimm *nvdimm, int cmd,
 774                 const struct nd_cmd_desc *desc, int idx, const u32 *in_field,
 775                 const u32 *out_field, unsigned long remainder)
 776 {
 777         if (idx >= desc->out_num)
 778                 return UINT_MAX;
 779
 780         if (desc->out_sizes[idx] < UINT_MAX)
 781                 return desc->out_sizes[idx];
 782
 783         if (nvdimm && cmd == ND_CMD_GET_CONFIG_DATA && idx == 1)
 784                 return in_field[1];
 785         else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2)
 786                 return out_field[1];
 787         else if (!nvdimm && cmd == ND_CMD_ARS_STATUS && idx == 2) {
 788                 /*
 789                  * Per table 9-276 ARS Data in ACPI 6.1, out_field[1] is
 790                  * "Size of Output Buffer in bytes, including this
 791                  * field."
 792                  */
 793                 if (out_field[1] < 4)
 794                         return 0;
 795                 /*
 796                  * ACPI 6.1 is ambiguous if 'status' is included in the
 797                  * output size. If we encounter an output size that
 798                  * overshoots the remainder by 4 bytes, assume it was
 799                  * including 'status'.
 800                  */
 801                 if (out_field[1] - 8 == remainder)
 802                         return remainder;
 803                 return out_field[1] - 4;
 804         } else if (cmd == ND_CMD_CALL) {
 805                 struct nd_cmd_pkg *pkg = (struct nd_cmd_pkg *) in_field;
 806
 807                 return pkg->nd_size_out;
 808         }
 809
 810
 811         return UINT_MAX;
 812 }
 813 EXPORT_SYMBOL_GPL(nd_cmd_out_size);
 814
 815 void wait_nvdimm_bus_probe_idle(struct device *dev)
 816 {
 817         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 818
 819         do {
 820                 if (nvdimm_bus->probe_active == 0)
 821                         break;
 822                 nvdimm_bus_unlock(&nvdimm_bus->dev);
 823                 wait_event(nvdimm_bus->probe_wait,
 824                                 nvdimm_bus->probe_active == 0);
 825                 nvdimm_bus_lock(&nvdimm_bus->dev);
 826         } while (true);
 827 }
 828
 829 static int nd_pmem_forget_poison_check(struct device *dev, void *data)
 830 {
 831         struct nd_cmd_clear_error *clear_err =
 832                 (struct nd_cmd_clear_error *)data;
 833         struct nd_btt *nd_btt = is_nd_btt(dev) ? to_nd_btt(dev) : NULL;
 834         struct nd_pfn *nd_pfn = is_nd_pfn(dev) ? to_nd_pfn(dev) : NULL;
 835         struct nd_dax *nd_dax = is_nd_dax(dev) ? to_nd_dax(dev) : NULL;
 836         struct nd_namespace_common *ndns = NULL;
 837         struct nd_namespace_io *nsio;
 838         resource_size_t offset = 0, end_trunc = 0, start, end, pstart, pend;
 839
 840         if (nd_dax || !dev->driver)
 841                 return 0;
 842
 843         start = clear_err->address;
 844         end = clear_err->address + clear_err->cleared - 1;
 845
 846         if (nd_btt || nd_pfn || nd_dax) {
 847                 if (nd_btt)
 848                         ndns = nd_btt->ndns;
 849                 else if (nd_pfn)
 850                         ndns = nd_pfn->ndns;
 851                 else if (nd_dax)
 852                         ndns = nd_dax->nd_pfn.ndns;
 853
 854                 if (!ndns)
 855                         return 0;
 856         } else
 857                 ndns = to_ndns(dev);
 858
 859         nsio = to_nd_namespace_io(&ndns->dev);
 860         pstart = nsio->res.start + offset;
 861         pend = nsio->res.end - end_trunc;
 862
 863         if ((pstart >= start) && (pend <= end))
 864                 return -EBUSY;
 865
 866         return 0;
 867
 868 }
 869
 870 static int nd_ns_forget_poison_check(struct device *dev, void *data)
 871 {
 872         return device_for_each_child(dev, data, nd_pmem_forget_poison_check);
 873 }
 874
 875 /* set_config requires an idle interleave set */
 876 static int nd_cmd_clear_to_send(struct nvdimm_bus *nvdimm_bus,
 877                 struct nvdimm *nvdimm, unsigned int cmd, void *data)
 878 {
 879         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 880
 881         /* ask the bus provider if it would like to block this request */
 882         if (nd_desc->clear_to_send) {
 883                 int rc = nd_desc->clear_to_send(nd_desc, nvdimm, cmd);
 884
 885                 if (rc)
 886                         return rc;
 887         }
 888
 889         /* require clear error to go through the pmem driver */
 890         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR)
 891                 return device_for_each_child(&nvdimm_bus->dev, data,
 892                                 nd_ns_forget_poison_check);
 893
 894         if (!nvdimm || cmd != ND_CMD_SET_CONFIG_DATA)
 895                 return 0;
 896
 897         /* prevent label manipulation while the kernel owns label updates */
 898         wait_nvdimm_bus_probe_idle(&nvdimm_bus->dev);
 899         if (atomic_read(&nvdimm->busy))
 900                 return -EBUSY;
 901         return 0;
 902 }
 903
 904 static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm,
 905                 int read_only, unsigned int ioctl_cmd, unsigned long arg)
 906 {
 907         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 908         size_t buf_len = 0, in_len = 0, out_len = 0;
 909         static char out_env[ND_CMD_MAX_ENVELOPE];
 910         static char in_env[ND_CMD_MAX_ENVELOPE];
 911         const struct nd_cmd_desc *desc = NULL;
 912         unsigned int cmd = _IOC_NR(ioctl_cmd);
 913         unsigned int func = cmd;
 914         void __user *p = (void __user *) arg;
 915         struct device *dev = &nvdimm_bus->dev;
 916         struct nd_cmd_pkg pkg;
 917         const char *cmd_name, *dimm_name;
 918         unsigned long cmd_mask;
 919         void *buf;
 920         int rc, i, cmd_rc;
 921
 922         if (nvdimm) {
 923                 desc = nd_cmd_dimm_desc(cmd);
 924                 cmd_name = nvdimm_cmd_name(cmd);
 925                 cmd_mask = nvdimm->cmd_mask;
 926                 dimm_name = dev_name(&nvdimm->dev);
 927         } else {
 928                 desc = nd_cmd_bus_desc(cmd);
 929                 cmd_name = nvdimm_bus_cmd_name(cmd);
 930                 cmd_mask = nd_desc->cmd_mask;
 931                 dimm_name = "bus";
 932         }
 933
 934         if (cmd == ND_CMD_CALL) {
 935                 if (copy_from_user(&pkg, p, sizeof(pkg)))
 936                         return -EFAULT;
 937         }
 938
 939         if (!desc || (desc->out_num + desc->in_num == 0) ||
 940                         !test_bit(cmd, &cmd_mask))
 941                 return -ENOTTY;
 942
 943         /* fail write commands (when read-only) */
 944         if (read_only)
 945                 switch (cmd) {
 946                 case ND_CMD_VENDOR:
 947                 case ND_CMD_SET_CONFIG_DATA:
 948                 case ND_CMD_ARS_START:
 949                 case ND_CMD_CLEAR_ERROR:
 950                 case ND_CMD_CALL:
 951                         dev_dbg(&nvdimm_bus->dev, "'%s' command while read-only.\n",
 952                                         nvdimm ? nvdimm_cmd_name(cmd)
 953                                         : nvdimm_bus_cmd_name(cmd));
 954                         return -EPERM;
 955                 default:
 956                         break;
 957                 }
 958
 959         /* process an input envelope */
 960         for (i = 0; i < desc->in_num; i++) {
 961                 u32 in_size, copy;
 962
 963                 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env);
 964                 if (in_size == UINT_MAX) {
 965                         dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n",
 966                                         __func__, dimm_name, cmd_name, i);
 967                         return -ENXIO;
 968                 }
 969                 if (in_len < sizeof(in_env))
 970                         copy = min_t(u32, sizeof(in_env) - in_len, in_size);
 971                 else
 972                         copy = 0;
 973                 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy))
 974                         return -EFAULT;
 975                 in_len += in_size;
 976         }
 977
 978         if (cmd == ND_CMD_CALL) {
 979                 func = pkg.nd_command;
 980                 dev_dbg(dev, "%s:%s, idx: %llu, in: %zu, out: %zu, len %zu\n",
 981                                 __func__, dimm_name, pkg.nd_command,
 982                                 in_len, out_len, buf_len);
 983
 984                 for (i = 0; i < ARRAY_SIZE(pkg.nd_reserved2); i++)
 985                         if (pkg.nd_reserved2[i])
 986                                 return -EINVAL;
 987         }
 988
 989         /* process an output envelope */
 990         for (i = 0; i < desc->out_num; i++) {
 991                 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i,
 992                                 (u32 *) in_env, (u32 *) out_env, 0);
 993                 u32 copy;
 994
 995                 if (out_size == UINT_MAX) {
 996                         dev_dbg(dev, "%s:%s unknown output size cmd: %s field: %d\n",
 997                                         __func__, dimm_name, cmd_name, i);
 998                         return -EFAULT;
 999                 }
1000                 if (out_len < sizeof(out_env))
1001                         copy = min_t(u32, sizeof(out_env) - out_len, out_size);
1002                 else
1003                         copy = 0;
1004                 if (copy && copy_from_user(&out_env[out_len],
1005                                         p + in_len + out_len, copy))
1006                         return -EFAULT;
1007                 out_len += out_size;
1008         }
1009
1010         buf_len = out_len + in_len;
1011         if (buf_len > ND_IOCTL_MAX_BUFLEN) {
1012                 dev_dbg(dev, "%s:%s cmd: %s buf_len: %zu > %d\n", __func__,
1013                                 dimm_name, cmd_name, buf_len,
1014                                 ND_IOCTL_MAX_BUFLEN);
1015                 return -EINVAL;
1016         }
1017
1018         buf = vmalloc(buf_len);
1019         if (!buf)
1020                 return -ENOMEM;
1021
1022         if (copy_from_user(buf, p, buf_len)) {
1023                 rc = -EFAULT;
1024                 goto out;
1025         }
1026
1027         nvdimm_bus_lock(&nvdimm_bus->dev);
1028         rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, func, buf);
1029         if (rc)
1030                 goto out_unlock;
1031
1032         rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc);
1033         if (rc < 0)
1034                 goto out_unlock;
1035
1036         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) {
1037                 struct nd_cmd_clear_error *clear_err = buf;
1038
1039                 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address,
1040                                 clear_err->cleared);
1041         }
1042         nvdimm_bus_unlock(&nvdimm_bus->dev);
1043
1044         if (copy_to_user(p, buf, buf_len))
1045                 rc = -EFAULT;
1046
1047         vfree(buf);
1048         return rc;
1049
1050  out_unlock:
1051         nvdimm_bus_unlock(&nvdimm_bus->dev);
1052  out:
1053         vfree(buf);
1054         return rc;
1055 }
1056
1057 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1058 {
1059         long id = (long) file->private_data;
1060         int rc = -ENXIO, ro;
1061         struct nvdimm_bus *nvdimm_bus;
1062
1063         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1064         mutex_lock(&nvdimm_bus_list_mutex);
1065         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1066                 if (nvdimm_bus->id == id) {
1067                         rc = __nd_ioctl(nvdimm_bus, NULL, ro, cmd, arg);
1068                         break;
1069                 }
1070         }
1071         mutex_unlock(&nvdimm_bus_list_mutex);
1072
1073         return rc;
1074 }
1075
1076 static int match_dimm(struct device *dev, void *data)
1077 {
1078         long id = (long) data;
1079
1080         if (is_nvdimm(dev)) {
1081                 struct nvdimm *nvdimm = to_nvdimm(dev);
1082
1083                 return nvdimm->id == id;
1084         }
1085
1086         return 0;
1087 }
1088
1089 static long nvdimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1090 {
1091         int rc = -ENXIO, ro;
1092         struct nvdimm_bus *nvdimm_bus;
1093
1094         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1095         mutex_lock(&nvdimm_bus_list_mutex);
1096         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1097                 struct device *dev = device_find_child(&nvdimm_bus->dev,
1098                                 file->private_data, match_dimm);
1099                 struct nvdimm *nvdimm;
1100
1101                 if (!dev)
1102                         continue;
1103
1104                 nvdimm = to_nvdimm(dev);
1105                 rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg);
1106                 put_device(dev);
1107                 break;
1108         }
1109         mutex_unlock(&nvdimm_bus_list_mutex);
1110
1111         return rc;
1112 }
1113
1114 static int nd_open(struct inode *inode, struct file *file)
1115 {
1116         long minor = iminor(inode);
1117
1118         file->private_data = (void *) minor;
1119         return 0;
1120 }
1121
1122 static const struct file_operations nvdimm_bus_fops = {
1123         .owner = THIS_MODULE,
1124         .open = nd_open,
1125         .unlocked_ioctl = nd_ioctl,
1126         .compat_ioctl = nd_ioctl,
1127         .llseek = noop_llseek,
1128 };
1129
1130 static const struct file_operations nvdimm_fops = {
1131         .owner = THIS_MODULE,
1132         .open = nd_open,
1133         .unlocked_ioctl = nvdimm_ioctl,
1134         .compat_ioctl = nvdimm_ioctl,
1135         .llseek = noop_llseek,
1136 };
1137
1138 int __init nvdimm_bus_init(void)
1139 {
1140         int rc;
1141
1142         BUILD_BUG_ON(sizeof(struct nd_smart_payload) != 128);
1143         BUILD_BUG_ON(sizeof(struct nd_smart_threshold_payload) != 8);
1144
1145         rc = bus_register(&nvdimm_bus_type);
1146         if (rc)
1147                 return rc;
1148
1149         rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops);
1150         if (rc < 0)
1151                 goto err_bus_chrdev;
1152         nvdimm_bus_major = rc;
1153
1154         rc = register_chrdev(0, "dimmctl", &nvdimm_fops);
1155         if (rc < 0)
1156                 goto err_dimm_chrdev;
1157         nvdimm_major = rc;
1158
1159         nd_class = class_create(THIS_MODULE, "nd");
1160         if (IS_ERR(nd_class)) {
1161                 rc = PTR_ERR(nd_class);
1162                 goto err_class;
1163         }
1164
1165         rc = driver_register(&nd_bus_driver.drv);
1166         if (rc)
1167                 goto err_nd_bus;
1168
1169         return 0;
1170
1171  err_nd_bus:
1172         class_destroy(nd_class);
1173  err_class:
1174         unregister_chrdev(nvdimm_major, "dimmctl");
1175  err_dimm_chrdev:
1176         unregister_chrdev(nvdimm_bus_major, "ndctl");
1177  err_bus_chrdev:
1178         bus_unregister(&nvdimm_bus_type);
1179
1180         return rc;
1181 }
1182
1183 void nvdimm_bus_exit(void)
1184 {
1185         driver_unregister(&nd_bus_driver.drv);
1186         class_destroy(nd_class);
1187         unregister_chrdev(nvdimm_bus_major, "ndctl");
1188         unregister_chrdev(nvdimm_major, "dimmctl");
1189         bus_unregister(&nvdimm_bus_type);
1190         ida_destroy(&nd_ida);
1191 }