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