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_nd_pmem(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_pmem(dev->parent) || is_nd_blk(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_pmem(dev) || is_nd_blk(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_pmem(dev) || is_nd_blk(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         return 0;
 202 }
 203
 204 static void nvdimm_clear_badblocks_regions(struct nvdimm_bus *nvdimm_bus,
 205                 phys_addr_t phys, u64 cleared)
 206 {
 207         struct clear_badblocks_context ctx = {
 208                 .phys = phys,
 209                 .cleared = cleared,
 210         };
 211
 212         device_for_each_child(&nvdimm_bus->dev, &ctx,
 213                         nvdimm_clear_badblocks_region);
 214 }
 215
 216 static void nvdimm_account_cleared_poison(struct nvdimm_bus *nvdimm_bus,
 217                 phys_addr_t phys, u64 cleared)
 218 {
 219         if (cleared > 0)
 220                 nvdimm_forget_poison(nvdimm_bus, phys, cleared);
 221
 222         if (cleared > 0 && cleared / 512)
 223                 nvdimm_clear_badblocks_regions(nvdimm_bus, phys, cleared);
 224 }
 225
 226 long nvdimm_clear_poison(struct device *dev, phys_addr_t phys,
 227                 unsigned int len)
 228 {
 229         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 230         struct nvdimm_bus_descriptor *nd_desc;
 231         struct nd_cmd_clear_error clear_err;
 232         struct nd_cmd_ars_cap ars_cap;
 233         u32 clear_err_unit, mask;
 234         int cmd_rc, rc;
 235
 236         if (!nvdimm_bus)
 237                 return -ENXIO;
 238
 239         nd_desc = nvdimm_bus->nd_desc;
 240         /*
 241          * if ndctl does not exist, it's PMEM_LEGACY and
 242          * we want to just pretend everything is handled.
 243          */
 244         if (!nd_desc->ndctl)
 245                 return len;
 246
 247         memset(&ars_cap, 0, sizeof(ars_cap));
 248         ars_cap.address = phys;
 249         ars_cap.length = len;
 250         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_ARS_CAP, &ars_cap,
 251                         sizeof(ars_cap), &cmd_rc);
 252         if (rc < 0)
 253                 return rc;
 254         if (cmd_rc < 0)
 255                 return cmd_rc;
 256         clear_err_unit = ars_cap.clear_err_unit;
 257         if (!clear_err_unit || !is_power_of_2(clear_err_unit))
 258                 return -ENXIO;
 259
 260         mask = clear_err_unit - 1;
 261         if ((phys | len) & mask)
 262                 return -ENXIO;
 263         memset(&clear_err, 0, sizeof(clear_err));
 264         clear_err.address = phys;
 265         clear_err.length = len;
 266         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_CLEAR_ERROR, &clear_err,
 267                         sizeof(clear_err), &cmd_rc);
 268         if (rc < 0)
 269                 return rc;
 270         if (cmd_rc < 0)
 271                 return cmd_rc;
 272
 273         nvdimm_account_cleared_poison(nvdimm_bus, phys, clear_err.cleared);
 274
 275         return clear_err.cleared;
 276 }
 277 EXPORT_SYMBOL_GPL(nvdimm_clear_poison);
 278
 279 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv);
 280
 281 static struct bus_type nvdimm_bus_type = {
 282         .name = "nd",
 283         .uevent = nvdimm_bus_uevent,
 284         .match = nvdimm_bus_match,
 285         .probe = nvdimm_bus_probe,
 286         .remove = nvdimm_bus_remove,
 287         .shutdown = nvdimm_bus_shutdown,
 288 };
 289
 290 static void nvdimm_bus_release(struct device *dev)
 291 {
 292         struct nvdimm_bus *nvdimm_bus;
 293
 294         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 295         ida_simple_remove(&nd_ida, nvdimm_bus->id);
 296         kfree(nvdimm_bus);
 297 }
 298
 299 static bool is_nvdimm_bus(struct device *dev)
 300 {
 301         return dev->release == nvdimm_bus_release;
 302 }
 303
 304 struct nvdimm_bus *walk_to_nvdimm_bus(struct device *nd_dev)
 305 {
 306         struct device *dev;
 307
 308         for (dev = nd_dev; dev; dev = dev->parent)
 309                 if (is_nvdimm_bus(dev))
 310                         break;
 311         dev_WARN_ONCE(nd_dev, !dev, "invalid dev, not on nd bus\n");
 312         if (dev)
 313                 return to_nvdimm_bus(dev);
 314         return NULL;
 315 }
 316
 317 struct nvdimm_bus *to_nvdimm_bus(struct device *dev)
 318 {
 319         struct nvdimm_bus *nvdimm_bus;
 320
 321         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 322         WARN_ON(!is_nvdimm_bus(dev));
 323         return nvdimm_bus;
 324 }
 325 EXPORT_SYMBOL_GPL(to_nvdimm_bus);
 326
 327 struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
 328                 struct nvdimm_bus_descriptor *nd_desc)
 329 {
 330         struct nvdimm_bus *nvdimm_bus;
 331         int rc;
 332
 333         nvdimm_bus = kzalloc(sizeof(*nvdimm_bus), GFP_KERNEL);
 334         if (!nvdimm_bus)
 335                 return NULL;
 336         INIT_LIST_HEAD(&nvdimm_bus->list);
 337         INIT_LIST_HEAD(&nvdimm_bus->mapping_list);
 338         INIT_LIST_HEAD(&nvdimm_bus->poison_list);
 339         init_waitqueue_head(&nvdimm_bus->probe_wait);
 340         nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL);
 341         mutex_init(&nvdimm_bus->reconfig_mutex);
 342         spin_lock_init(&nvdimm_bus->poison_lock);
 343         if (nvdimm_bus->id < 0) {
 344                 kfree(nvdimm_bus);
 345                 return NULL;
 346         }
 347         nvdimm_bus->nd_desc = nd_desc;
 348         nvdimm_bus->dev.parent = parent;
 349         nvdimm_bus->dev.release = nvdimm_bus_release;
 350         nvdimm_bus->dev.groups = nd_desc->attr_groups;
 351         nvdimm_bus->dev.bus = &nvdimm_bus_type;
 352         dev_set_name(&nvdimm_bus->dev, "ndbus%d", nvdimm_bus->id);
 353         rc = device_register(&nvdimm_bus->dev);
 354         if (rc) {
 355                 dev_dbg(&nvdimm_bus->dev, "registration failed: %d\n", rc);
 356                 goto err;
 357         }
 358
 359         return nvdimm_bus;
 360  err:
 361         put_device(&nvdimm_bus->dev);
 362         return NULL;
 363 }
 364 EXPORT_SYMBOL_GPL(nvdimm_bus_register);
 365
 366 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus)
 367 {
 368         if (!nvdimm_bus)
 369                 return;
 370         device_unregister(&nvdimm_bus->dev);
 371 }
 372 EXPORT_SYMBOL_GPL(nvdimm_bus_unregister);
 373
 374 static int child_unregister(struct device *dev, void *data)
 375 {
 376         /*
 377          * the singular ndctl class device per bus needs to be
 378          * "device_destroy"ed, so skip it here
 379          *
 380          * i.e. remove classless children
 381          */
 382         if (dev->class)
 383                 /* pass */;
 384         else
 385                 nd_device_unregister(dev, ND_SYNC);
 386         return 0;
 387 }
 388
 389 static void free_poison_list(struct list_head *poison_list)
 390 {
 391         struct nd_poison *pl, *next;
 392
 393         list_for_each_entry_safe(pl, next, poison_list, list) {
 394                 list_del(&pl->list);
 395                 kfree(pl);
 396         }
 397         list_del_init(poison_list);
 398 }
 399
 400 static int nd_bus_remove(struct device *dev)
 401 {
 402         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 403
 404         mutex_lock(&nvdimm_bus_list_mutex);
 405         list_del_init(&nvdimm_bus->list);
 406         mutex_unlock(&nvdimm_bus_list_mutex);
 407
 408         nd_synchronize();
 409         device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister);
 410
 411         spin_lock(&nvdimm_bus->poison_lock);
 412         free_poison_list(&nvdimm_bus->poison_list);
 413         spin_unlock(&nvdimm_bus->poison_lock);
 414
 415         nvdimm_bus_destroy_ndctl(nvdimm_bus);
 416
 417         return 0;
 418 }
 419
 420 static int nd_bus_probe(struct device *dev)
 421 {
 422         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 423         int rc;
 424
 425         rc = nvdimm_bus_create_ndctl(nvdimm_bus);
 426         if (rc)
 427                 return rc;
 428
 429         mutex_lock(&nvdimm_bus_list_mutex);
 430         list_add_tail(&nvdimm_bus->list, &nvdimm_bus_list);
 431         mutex_unlock(&nvdimm_bus_list_mutex);
 432
 433         /* enable bus provider attributes to look up their local context */
 434         dev_set_drvdata(dev, nvdimm_bus->nd_desc);
 435
 436         return 0;
 437 }
 438
 439 static struct nd_device_driver nd_bus_driver = {
 440         .probe = nd_bus_probe,
 441         .remove = nd_bus_remove,
 442         .drv = {
 443                 .name = "nd_bus",
 444                 .suppress_bind_attrs = true,
 445                 .bus = &nvdimm_bus_type,
 446                 .owner = THIS_MODULE,
 447                 .mod_name = KBUILD_MODNAME,
 448         },
 449 };
 450
 451 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv)
 452 {
 453         struct nd_device_driver *nd_drv = to_nd_device_driver(drv);
 454
 455         if (is_nvdimm_bus(dev) && nd_drv == &nd_bus_driver)
 456                 return true;
 457
 458         return !!test_bit(to_nd_device_type(dev), &nd_drv->type);
 459 }
 460
 461 static ASYNC_DOMAIN_EXCLUSIVE(nd_async_domain);
 462
 463 void nd_synchronize(void)
 464 {
 465         async_synchronize_full_domain(&nd_async_domain);
 466 }
 467 EXPORT_SYMBOL_GPL(nd_synchronize);
 468
 469 static void nd_async_device_register(void *d, async_cookie_t cookie)
 470 {
 471         struct device *dev = d;
 472
 473         if (device_add(dev) != 0) {
 474                 dev_err(dev, "%s: failed\n", __func__);
 475                 put_device(dev);
 476         }
 477         put_device(dev);
 478 }
 479
 480 static void nd_async_device_unregister(void *d, async_cookie_t cookie)
 481 {
 482         struct device *dev = d;
 483
 484         /* flush bus operations before delete */
 485         nvdimm_bus_lock(dev);
 486         nvdimm_bus_unlock(dev);
 487
 488         device_unregister(dev);
 489         put_device(dev);
 490 }
 491
 492 void __nd_device_register(struct device *dev)
 493 {
 494         if (!dev)
 495                 return;
 496         dev->bus = &nvdimm_bus_type;
 497         get_device(dev);
 498         async_schedule_domain(nd_async_device_register, dev,
 499                         &nd_async_domain);
 500 }
 501
 502 void nd_device_register(struct device *dev)
 503 {
 504         device_initialize(dev);
 505         __nd_device_register(dev);
 506 }
 507 EXPORT_SYMBOL(nd_device_register);
 508
 509 void nd_device_unregister(struct device *dev, enum nd_async_mode mode)
 510 {
 511         switch (mode) {
 512         case ND_ASYNC:
 513                 get_device(dev);
 514                 async_schedule_domain(nd_async_device_unregister, dev,
 515                                 &nd_async_domain);
 516                 break;
 517         case ND_SYNC:
 518                 nd_synchronize();
 519                 device_unregister(dev);
 520                 break;
 521         }
 522 }
 523 EXPORT_SYMBOL(nd_device_unregister);
 524
 525 /**
 526  * __nd_driver_register() - register a region or a namespace driver
 527  * @nd_drv: driver to register
 528  * @owner: automatically set by nd_driver_register() macro
 529  * @mod_name: automatically set by nd_driver_register() macro
 530  */
 531 int __nd_driver_register(struct nd_device_driver *nd_drv, struct module *owner,
 532                 const char *mod_name)
 533 {
 534         struct device_driver *drv = &nd_drv->drv;
 535
 536         if (!nd_drv->type) {
 537                 pr_debug("driver type bitmask not set (%pf)\n",
 538                                 __builtin_return_address(0));
 539                 return -EINVAL;
 540         }
 541
 542         if (!nd_drv->probe) {
 543                 pr_debug("%s ->probe() must be specified\n", mod_name);
 544                 return -EINVAL;
 545         }
 546
 547         drv->bus = &nvdimm_bus_type;
 548         drv->owner = owner;
 549         drv->mod_name = mod_name;
 550
 551         return driver_register(drv);
 552 }
 553 EXPORT_SYMBOL(__nd_driver_register);
 554
 555 int nvdimm_revalidate_disk(struct gendisk *disk)
 556 {
 557         struct device *dev = disk_to_dev(disk)->parent;
 558         struct nd_region *nd_region = to_nd_region(dev->parent);
 559         const char *pol = nd_region->ro ? "only" : "write";
 560
 561         if (nd_region->ro == get_disk_ro(disk))
 562                 return 0;
 563
 564         dev_info(dev, "%s read-%s, marking %s read-%s\n",
 565                         dev_name(&nd_region->dev), pol, disk->disk_name, pol);
 566         set_disk_ro(disk, nd_region->ro);
 567
 568         return 0;
 569
 570 }
 571 EXPORT_SYMBOL(nvdimm_revalidate_disk);
 572
 573 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr,
 574                 char *buf)
 575 {
 576         return sprintf(buf, ND_DEVICE_MODALIAS_FMT "\n",
 577                         to_nd_device_type(dev));
 578 }
 579 static DEVICE_ATTR_RO(modalias);
 580
 581 static ssize_t devtype_show(struct device *dev, struct device_attribute *attr,
 582                 char *buf)
 583 {
 584         return sprintf(buf, "%s\n", dev->type->name);
 585 }
 586 static DEVICE_ATTR_RO(devtype);
 587
 588 static struct attribute *nd_device_attributes[] = {
 589         &dev_attr_modalias.attr,
 590         &dev_attr_devtype.attr,
 591         NULL,
 592 };
 593
 594 /**
 595  * nd_device_attribute_group - generic attributes for all devices on an nd bus
 596  */
 597 struct attribute_group nd_device_attribute_group = {
 598         .attrs = nd_device_attributes,
 599 };
 600 EXPORT_SYMBOL_GPL(nd_device_attribute_group);
 601
 602 static ssize_t numa_node_show(struct device *dev,
 603                 struct device_attribute *attr, char *buf)
 604 {
 605         return sprintf(buf, "%d\n", dev_to_node(dev));
 606 }
 607 static DEVICE_ATTR_RO(numa_node);
 608
 609 static struct attribute *nd_numa_attributes[] = {
 610         &dev_attr_numa_node.attr,
 611         NULL,
 612 };
 613
 614 static umode_t nd_numa_attr_visible(struct kobject *kobj, struct attribute *a,
 615                 int n)
 616 {
 617         if (!IS_ENABLED(CONFIG_NUMA))
 618                 return 0;
 619
 620         return a->mode;
 621 }
 622
 623 /**
 624  * nd_numa_attribute_group - NUMA attributes for all devices on an nd bus
 625  */
 626 struct attribute_group nd_numa_attribute_group = {
 627         .attrs = nd_numa_attributes,
 628         .is_visible = nd_numa_attr_visible,
 629 };
 630 EXPORT_SYMBOL_GPL(nd_numa_attribute_group);
 631
 632 int nvdimm_bus_create_ndctl(struct nvdimm_bus *nvdimm_bus)
 633 {
 634         dev_t devt = MKDEV(nvdimm_bus_major, nvdimm_bus->id);
 635         struct device *dev;
 636
 637         dev = device_create(nd_class, &nvdimm_bus->dev, devt, nvdimm_bus,
 638                         "ndctl%d", nvdimm_bus->id);
 639
 640         if (IS_ERR(dev))
 641                 dev_dbg(&nvdimm_bus->dev, "failed to register ndctl%d: %ld\n",
 642                                 nvdimm_bus->id, PTR_ERR(dev));
 643         return PTR_ERR_OR_ZERO(dev);
 644 }
 645
 646 void nvdimm_bus_destroy_ndctl(struct nvdimm_bus *nvdimm_bus)
 647 {
 648         device_destroy(nd_class, MKDEV(nvdimm_bus_major, nvdimm_bus->id));
 649 }
 650
 651 static const struct nd_cmd_desc __nd_cmd_dimm_descs[] = {
 652         [ND_CMD_IMPLEMENTED] = { },
 653         [ND_CMD_SMART] = {
 654                 .out_num = 2,
 655                 .out_sizes = { 4, 128, },
 656         },
 657         [ND_CMD_SMART_THRESHOLD] = {
 658                 .out_num = 2,
 659                 .out_sizes = { 4, 8, },
 660         },
 661         [ND_CMD_DIMM_FLAGS] = {
 662                 .out_num = 2,
 663                 .out_sizes = { 4, 4 },
 664         },
 665         [ND_CMD_GET_CONFIG_SIZE] = {
 666                 .out_num = 3,
 667                 .out_sizes = { 4, 4, 4, },
 668         },
 669         [ND_CMD_GET_CONFIG_DATA] = {
 670                 .in_num = 2,
 671                 .in_sizes = { 4, 4, },
 672                 .out_num = 2,
 673                 .out_sizes = { 4, UINT_MAX, },
 674         },
 675         [ND_CMD_SET_CONFIG_DATA] = {
 676                 .in_num = 3,
 677                 .in_sizes = { 4, 4, UINT_MAX, },
 678                 .out_num = 1,
 679                 .out_sizes = { 4, },
 680         },
 681         [ND_CMD_VENDOR] = {
 682                 .in_num = 3,
 683                 .in_sizes = { 4, 4, UINT_MAX, },
 684                 .out_num = 3,
 685                 .out_sizes = { 4, 4, UINT_MAX, },
 686         },
 687         [ND_CMD_CALL] = {
 688                 .in_num = 2,
 689                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 690                 .out_num = 1,
 691                 .out_sizes = { UINT_MAX, },
 692         },
 693 };
 694
 695 const struct nd_cmd_desc *nd_cmd_dimm_desc(int cmd)
 696 {
 697         if (cmd < ARRAY_SIZE(__nd_cmd_dimm_descs))
 698                 return &__nd_cmd_dimm_descs[cmd];
 699         return NULL;
 700 }
 701 EXPORT_SYMBOL_GPL(nd_cmd_dimm_desc);
 702
 703 static const struct nd_cmd_desc __nd_cmd_bus_descs[] = {
 704         [ND_CMD_IMPLEMENTED] = { },
 705         [ND_CMD_ARS_CAP] = {
 706                 .in_num = 2,
 707                 .in_sizes = { 8, 8, },
 708                 .out_num = 4,
 709                 .out_sizes = { 4, 4, 4, 4, },
 710         },
 711         [ND_CMD_ARS_START] = {
 712                 .in_num = 5,
 713                 .in_sizes = { 8, 8, 2, 1, 5, },
 714                 .out_num = 2,
 715                 .out_sizes = { 4, 4, },
 716         },
 717         [ND_CMD_ARS_STATUS] = {
 718                 .out_num = 3,
 719                 .out_sizes = { 4, 4, UINT_MAX, },
 720         },
 721         [ND_CMD_CLEAR_ERROR] = {
 722                 .in_num = 2,
 723                 .in_sizes = { 8, 8, },
 724                 .out_num = 3,
 725                 .out_sizes = { 4, 4, 8, },
 726         },
 727         [ND_CMD_CALL] = {
 728                 .in_num = 2,
 729                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 730                 .out_num = 1,
 731                 .out_sizes = { UINT_MAX, },
 732         },
 733 };
 734
 735 const struct nd_cmd_desc *nd_cmd_bus_desc(int cmd)
 736 {
 737         if (cmd < ARRAY_SIZE(__nd_cmd_bus_descs))
 738                 return &__nd_cmd_bus_descs[cmd];
 739         return NULL;
 740 }
 741 EXPORT_SYMBOL_GPL(nd_cmd_bus_desc);
 742
 743 u32 nd_cmd_in_size(struct nvdimm *nvdimm, int cmd,
 744                 const struct nd_cmd_desc *desc, int idx, void *buf)
 745 {
 746         if (idx >= desc->in_num)
 747                 return UINT_MAX;
 748
 749         if (desc->in_sizes[idx] < UINT_MAX)
 750                 return desc->in_sizes[idx];
 751
 752         if (nvdimm && cmd == ND_CMD_SET_CONFIG_DATA && idx == 2) {
 753                 struct nd_cmd_set_config_hdr *hdr = buf;
 754
 755                 return hdr->in_length;
 756         } else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) {
 757                 struct nd_cmd_vendor_hdr *hdr = buf;
 758
 759                 return hdr->in_length;
 760         } else if (cmd == ND_CMD_CALL) {
 761                 struct nd_cmd_pkg *pkg = buf;
 762
 763                 return pkg->nd_size_in;
 764         }
 765
 766         return UINT_MAX;
 767 }
 768 EXPORT_SYMBOL_GPL(nd_cmd_in_size);
 769
 770 u32 nd_cmd_out_size(struct nvdimm *nvdimm, int cmd,
 771                 const struct nd_cmd_desc *desc, int idx, const u32 *in_field,
 772                 const u32 *out_field, unsigned long remainder)
 773 {
 774         if (idx >= desc->out_num)
 775                 return UINT_MAX;
 776
 777         if (desc->out_sizes[idx] < UINT_MAX)
 778                 return desc->out_sizes[idx];
 779
 780         if (nvdimm && cmd == ND_CMD_GET_CONFIG_DATA && idx == 1)
 781                 return in_field[1];
 782         else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2)
 783                 return out_field[1];
 784         else if (!nvdimm && cmd == ND_CMD_ARS_STATUS && idx == 2) {
 785                 /*
 786                  * Per table 9-276 ARS Data in ACPI 6.1, out_field[1] is
 787                  * "Size of Output Buffer in bytes, including this
 788                  * field."
 789                  */
 790                 if (out_field[1] < 4)
 791                         return 0;
 792                 /*
 793                  * ACPI 6.1 is ambiguous if 'status' is included in the
 794                  * output size. If we encounter an output size that
 795                  * overshoots the remainder by 4 bytes, assume it was
 796                  * including 'status'.
 797                  */
 798                 if (out_field[1] - 8 == remainder)
 799                         return remainder;
 800                 return out_field[1] - 4;
 801         } else if (cmd == ND_CMD_CALL) {
 802                 struct nd_cmd_pkg *pkg = (struct nd_cmd_pkg *) in_field;
 803
 804                 return pkg->nd_size_out;
 805         }
 806
 807
 808         return UINT_MAX;
 809 }
 810 EXPORT_SYMBOL_GPL(nd_cmd_out_size);
 811
 812 void wait_nvdimm_bus_probe_idle(struct device *dev)
 813 {
 814         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 815
 816         do {
 817                 if (nvdimm_bus->probe_active == 0)
 818                         break;
 819                 nvdimm_bus_unlock(&nvdimm_bus->dev);
 820                 wait_event(nvdimm_bus->probe_wait,
 821                                 nvdimm_bus->probe_active == 0);
 822                 nvdimm_bus_lock(&nvdimm_bus->dev);
 823         } while (true);
 824 }
 825
 826 static int nd_pmem_forget_poison_check(struct device *dev, void *data)
 827 {
 828         struct nd_cmd_clear_error *clear_err =
 829                 (struct nd_cmd_clear_error *)data;
 830         struct nd_btt *nd_btt = is_nd_btt(dev) ? to_nd_btt(dev) : NULL;
 831         struct nd_pfn *nd_pfn = is_nd_pfn(dev) ? to_nd_pfn(dev) : NULL;
 832         struct nd_dax *nd_dax = is_nd_dax(dev) ? to_nd_dax(dev) : NULL;
 833         struct nd_namespace_common *ndns = NULL;
 834         struct nd_namespace_io *nsio;
 835         resource_size_t offset = 0, end_trunc = 0, start, end, pstart, pend;
 836
 837         if (nd_dax || !dev->driver)
 838                 return 0;
 839
 840         start = clear_err->address;
 841         end = clear_err->address + clear_err->cleared - 1;
 842
 843         if (nd_btt || nd_pfn || nd_dax) {
 844                 if (nd_btt)
 845                         ndns = nd_btt->ndns;
 846                 else if (nd_pfn)
 847                         ndns = nd_pfn->ndns;
 848                 else if (nd_dax)
 849                         ndns = nd_dax->nd_pfn.ndns;
 850
 851                 if (!ndns)
 852                         return 0;
 853         } else
 854                 ndns = to_ndns(dev);
 855
 856         nsio = to_nd_namespace_io(&ndns->dev);
 857         pstart = nsio->res.start + offset;
 858         pend = nsio->res.end - end_trunc;
 859
 860         if ((pstart >= start) && (pend <= end))
 861                 return -EBUSY;
 862
 863         return 0;
 864
 865 }
 866
 867 static int nd_ns_forget_poison_check(struct device *dev, void *data)
 868 {
 869         return device_for_each_child(dev, data, nd_pmem_forget_poison_check);
 870 }
 871
 872 /* set_config requires an idle interleave set */
 873 static int nd_cmd_clear_to_send(struct nvdimm_bus *nvdimm_bus,
 874                 struct nvdimm *nvdimm, unsigned int cmd, void *data)
 875 {
 876         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 877
 878         /* ask the bus provider if it would like to block this request */
 879         if (nd_desc->clear_to_send) {
 880                 int rc = nd_desc->clear_to_send(nd_desc, nvdimm, cmd);
 881
 882                 if (rc)
 883                         return rc;
 884         }
 885
 886         /* require clear error to go through the pmem driver */
 887         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR)
 888                 return device_for_each_child(&nvdimm_bus->dev, data,
 889                                 nd_ns_forget_poison_check);
 890
 891         if (!nvdimm || cmd != ND_CMD_SET_CONFIG_DATA)
 892                 return 0;
 893
 894         /* prevent label manipulation while the kernel owns label updates */
 895         wait_nvdimm_bus_probe_idle(&nvdimm_bus->dev);
 896         if (atomic_read(&nvdimm->busy))
 897                 return -EBUSY;
 898         return 0;
 899 }
 900
 901 static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm,
 902                 int read_only, unsigned int ioctl_cmd, unsigned long arg)
 903 {
 904         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 905         size_t buf_len = 0, in_len = 0, out_len = 0;
 906         static char out_env[ND_CMD_MAX_ENVELOPE];
 907         static char in_env[ND_CMD_MAX_ENVELOPE];
 908         const struct nd_cmd_desc *desc = NULL;
 909         unsigned int cmd = _IOC_NR(ioctl_cmd);
 910         void __user *p = (void __user *) arg;
 911         struct device *dev = &nvdimm_bus->dev;
 912         struct nd_cmd_pkg pkg;
 913         const char *cmd_name, *dimm_name;
 914         unsigned long cmd_mask;
 915         void *buf;
 916         int rc, i, cmd_rc;
 917
 918         if (nvdimm) {
 919                 desc = nd_cmd_dimm_desc(cmd);
 920                 cmd_name = nvdimm_cmd_name(cmd);
 921                 cmd_mask = nvdimm->cmd_mask;
 922                 dimm_name = dev_name(&nvdimm->dev);
 923         } else {
 924                 desc = nd_cmd_bus_desc(cmd);
 925                 cmd_name = nvdimm_bus_cmd_name(cmd);
 926                 cmd_mask = nd_desc->cmd_mask;
 927                 dimm_name = "bus";
 928         }
 929
 930         if (cmd == ND_CMD_CALL) {
 931                 if (copy_from_user(&pkg, p, sizeof(pkg)))
 932                         return -EFAULT;
 933         }
 934
 935         if (!desc || (desc->out_num + desc->in_num == 0) ||
 936                         !test_bit(cmd, &cmd_mask))
 937                 return -ENOTTY;
 938
 939         /* fail write commands (when read-only) */
 940         if (read_only)
 941                 switch (cmd) {
 942                 case ND_CMD_VENDOR:
 943                 case ND_CMD_SET_CONFIG_DATA:
 944                 case ND_CMD_ARS_START:
 945                 case ND_CMD_CLEAR_ERROR:
 946                 case ND_CMD_CALL:
 947                         dev_dbg(&nvdimm_bus->dev, "'%s' command while read-only.\n",
 948                                         nvdimm ? nvdimm_cmd_name(cmd)
 949                                         : nvdimm_bus_cmd_name(cmd));
 950                         return -EPERM;
 951                 default:
 952                         break;
 953                 }
 954
 955         /* process an input envelope */
 956         for (i = 0; i < desc->in_num; i++) {
 957                 u32 in_size, copy;
 958
 959                 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env);
 960                 if (in_size == UINT_MAX) {
 961                         dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n",
 962                                         __func__, dimm_name, cmd_name, i);
 963                         return -ENXIO;
 964                 }
 965                 if (in_len < sizeof(in_env))
 966                         copy = min_t(u32, sizeof(in_env) - in_len, in_size);
 967                 else
 968                         copy = 0;
 969                 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy))
 970                         return -EFAULT;
 971                 in_len += in_size;
 972         }
 973
 974         if (cmd == ND_CMD_CALL) {
 975                 dev_dbg(dev, "%s:%s, idx: %llu, in: %zu, out: %zu, len %zu\n",
 976                                 __func__, dimm_name, pkg.nd_command,
 977                                 in_len, out_len, buf_len);
 978
 979                 for (i = 0; i < ARRAY_SIZE(pkg.nd_reserved2); i++)
 980                         if (pkg.nd_reserved2[i])
 981                                 return -EINVAL;
 982         }
 983
 984         /* process an output envelope */
 985         for (i = 0; i < desc->out_num; i++) {
 986                 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i,
 987                                 (u32 *) in_env, (u32 *) out_env, 0);
 988                 u32 copy;
 989
 990                 if (out_size == UINT_MAX) {
 991                         dev_dbg(dev, "%s:%s unknown output size cmd: %s field: %d\n",
 992                                         __func__, dimm_name, cmd_name, i);
 993                         return -EFAULT;
 994                 }
 995                 if (out_len < sizeof(out_env))
 996                         copy = min_t(u32, sizeof(out_env) - out_len, out_size);
 997                 else
 998                         copy = 0;
 999                 if (copy && copy_from_user(&out_env[out_len],
1000                                         p + in_len + out_len, copy))
1001                         return -EFAULT;
1002                 out_len += out_size;
1003         }
1004
1005         buf_len = out_len + in_len;
1006         if (buf_len > ND_IOCTL_MAX_BUFLEN) {
1007                 dev_dbg(dev, "%s:%s cmd: %s buf_len: %zu > %d\n", __func__,
1008                                 dimm_name, cmd_name, buf_len,
1009                                 ND_IOCTL_MAX_BUFLEN);
1010                 return -EINVAL;
1011         }
1012
1013         buf = vmalloc(buf_len);
1014         if (!buf)
1015                 return -ENOMEM;
1016
1017         if (copy_from_user(buf, p, buf_len)) {
1018                 rc = -EFAULT;
1019                 goto out;
1020         }
1021
1022         nvdimm_bus_lock(&nvdimm_bus->dev);
1023         rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, cmd, buf);
1024         if (rc)
1025                 goto out_unlock;
1026
1027         rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc);
1028         if (rc < 0)
1029                 goto out_unlock;
1030
1031         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) {
1032                 struct nd_cmd_clear_error *clear_err = buf;
1033
1034                 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address,
1035                                 clear_err->cleared);
1036         }
1037         nvdimm_bus_unlock(&nvdimm_bus->dev);
1038
1039         if (copy_to_user(p, buf, buf_len))
1040                 rc = -EFAULT;
1041
1042         vfree(buf);
1043         return rc;
1044
1045  out_unlock:
1046         nvdimm_bus_unlock(&nvdimm_bus->dev);
1047  out:
1048         vfree(buf);
1049         return rc;
1050 }
1051
1052 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1053 {
1054         long id = (long) file->private_data;
1055         int rc = -ENXIO, ro;
1056         struct nvdimm_bus *nvdimm_bus;
1057
1058         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1059         mutex_lock(&nvdimm_bus_list_mutex);
1060         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1061                 if (nvdimm_bus->id == id) {
1062                         rc = __nd_ioctl(nvdimm_bus, NULL, ro, cmd, arg);
1063                         break;
1064                 }
1065         }
1066         mutex_unlock(&nvdimm_bus_list_mutex);
1067
1068         return rc;
1069 }
1070
1071 static int match_dimm(struct device *dev, void *data)
1072 {
1073         long id = (long) data;
1074
1075         if (is_nvdimm(dev)) {
1076                 struct nvdimm *nvdimm = to_nvdimm(dev);
1077
1078                 return nvdimm->id == id;
1079         }
1080
1081         return 0;
1082 }
1083
1084 static long nvdimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1085 {
1086         int rc = -ENXIO, ro;
1087         struct nvdimm_bus *nvdimm_bus;
1088
1089         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1090         mutex_lock(&nvdimm_bus_list_mutex);
1091         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1092                 struct device *dev = device_find_child(&nvdimm_bus->dev,
1093                                 file->private_data, match_dimm);
1094                 struct nvdimm *nvdimm;
1095
1096                 if (!dev)
1097                         continue;
1098
1099                 nvdimm = to_nvdimm(dev);
1100                 rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg);
1101                 put_device(dev);
1102                 break;
1103         }
1104         mutex_unlock(&nvdimm_bus_list_mutex);
1105
1106         return rc;
1107 }
1108
1109 static int nd_open(struct inode *inode, struct file *file)
1110 {
1111         long minor = iminor(inode);
1112
1113         file->private_data = (void *) minor;
1114         return 0;
1115 }
1116
1117 static const struct file_operations nvdimm_bus_fops = {
1118         .owner = THIS_MODULE,
1119         .open = nd_open,
1120         .unlocked_ioctl = nd_ioctl,
1121         .compat_ioctl = nd_ioctl,
1122         .llseek = noop_llseek,
1123 };
1124
1125 static const struct file_operations nvdimm_fops = {
1126         .owner = THIS_MODULE,
1127         .open = nd_open,
1128         .unlocked_ioctl = nvdimm_ioctl,
1129         .compat_ioctl = nvdimm_ioctl,
1130         .llseek = noop_llseek,
1131 };
1132
1133 int __init nvdimm_bus_init(void)
1134 {
1135         int rc;
1136
1137         BUILD_BUG_ON(sizeof(struct nd_smart_payload) != 128);
1138         BUILD_BUG_ON(sizeof(struct nd_smart_threshold_payload) != 8);
1139
1140         rc = bus_register(&nvdimm_bus_type);
1141         if (rc)
1142                 return rc;
1143
1144         rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops);
1145         if (rc < 0)
1146                 goto err_bus_chrdev;
1147         nvdimm_bus_major = rc;
1148
1149         rc = register_chrdev(0, "dimmctl", &nvdimm_fops);
1150         if (rc < 0)
1151                 goto err_dimm_chrdev;
1152         nvdimm_major = rc;
1153
1154         nd_class = class_create(THIS_MODULE, "nd");
1155         if (IS_ERR(nd_class)) {
1156                 rc = PTR_ERR(nd_class);
1157                 goto err_class;
1158         }
1159
1160         rc = driver_register(&nd_bus_driver.drv);
1161         if (rc)
1162                 goto err_nd_bus;
1163
1164         return 0;
1165
1166  err_nd_bus:
1167         class_destroy(nd_class);
1168  err_class:
1169         unregister_chrdev(nvdimm_major, "dimmctl");
1170  err_dimm_chrdev:
1171         unregister_chrdev(nvdimm_bus_major, "ndctl");
1172  err_bus_chrdev:
1173         bus_unregister(&nvdimm_bus_type);
1174
1175         return rc;
1176 }
1177
1178 void nvdimm_bus_exit(void)
1179 {
1180         driver_unregister(&nd_bus_driver.drv);
1181         class_destroy(nd_class);
1182         unregister_chrdev(nvdimm_bus_major, "ndctl");
1183         unregister_chrdev(nvdimm_major, "dimmctl");
1184         bus_unregister(&nvdimm_bus_type);
1185         ida_destroy(&nd_ida);
1186 }