drivers/nvdimm/bus.c

   1 /*
   2  * Copyright(c) 2013-2015 Intel Corporation. All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of version 2 of the GNU General Public License as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  */
  13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14 #include <linux/libnvdimm.h>
  15 #include <linux/sched/mm.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/uaccess.h>
  18 #include <linux/module.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/fcntl.h>
  21 #include <linux/async.h>
  22 #include <linux/genhd.h>
  23 #include <linux/ndctl.h>
  24 #include <linux/sched.h>
  25 #include <linux/slab.h>
  26 #include <linux/fs.h>
  27 #include <linux/io.h>
  28 #include <linux/mm.h>
  29 #include <linux/nd.h>
  30 #include "nd-core.h"
  31 #include "nd.h"
  32 #include "pfn.h"
  33
  34 int nvdimm_major;
  35 static int nvdimm_bus_major;
  36 static struct class *nd_class;
  37 static DEFINE_IDA(nd_ida);
  38
  39 static int to_nd_device_type(struct device *dev)
  40 {
  41         if (is_nvdimm(dev))
  42                 return ND_DEVICE_DIMM;
  43         else if (is_memory(dev))
  44                 return ND_DEVICE_REGION_PMEM;
  45         else if (is_nd_blk(dev))
  46                 return ND_DEVICE_REGION_BLK;
  47         else if (is_nd_dax(dev))
  48                 return ND_DEVICE_DAX_PMEM;
  49         else if (is_nd_region(dev->parent))
  50                 return nd_region_to_nstype(to_nd_region(dev->parent));
  51
  52         return 0;
  53 }
  54
  55 static int nvdimm_bus_uevent(struct device *dev, struct kobj_uevent_env *env)
  56 {
  57         return add_uevent_var(env, "MODALIAS=" ND_DEVICE_MODALIAS_FMT,
  58                         to_nd_device_type(dev));
  59 }
  60
  61 static struct module *to_bus_provider(struct device *dev)
  62 {
  63         /* pin bus providers while regions are enabled */
  64         if (is_nd_region(dev)) {
  65                 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
  66
  67                 return nvdimm_bus->nd_desc->module;
  68         }
  69         return NULL;
  70 }
  71
  72 static void nvdimm_bus_probe_start(struct nvdimm_bus *nvdimm_bus)
  73 {
  74         nvdimm_bus_lock(&nvdimm_bus->dev);
  75         nvdimm_bus->probe_active++;
  76         nvdimm_bus_unlock(&nvdimm_bus->dev);
  77 }
  78
  79 static void nvdimm_bus_probe_end(struct nvdimm_bus *nvdimm_bus)
  80 {
  81         nvdimm_bus_lock(&nvdimm_bus->dev);
  82         if (--nvdimm_bus->probe_active == 0)
  83                 wake_up(&nvdimm_bus->probe_wait);
  84         nvdimm_bus_unlock(&nvdimm_bus->dev);
  85 }
  86
  87 static int nvdimm_bus_probe(struct device *dev)
  88 {
  89         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
  90         struct module *provider = to_bus_provider(dev);
  91         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
  92         int rc;
  93
  94         if (!try_module_get(provider))
  95                 return -ENXIO;
  96
  97         dev_dbg(&nvdimm_bus->dev, "START: %s.probe(%s)\n",
  98                         dev->driver->name, dev_name(dev));
  99
 100         nvdimm_bus_probe_start(nvdimm_bus);
 101         rc = nd_drv->probe(dev);
 102         if (rc == 0)
 103                 nd_region_probe_success(nvdimm_bus, dev);
 104         else
 105                 nd_region_disable(nvdimm_bus, dev);
 106         nvdimm_bus_probe_end(nvdimm_bus);
 107
 108         dev_dbg(&nvdimm_bus->dev, "END: %s.probe(%s) = %d\n", dev->driver->name,
 109                         dev_name(dev), rc);
 110
 111         if (rc != 0)
 112                 module_put(provider);
 113         return rc;
 114 }
 115
 116 static int nvdimm_bus_remove(struct device *dev)
 117 {
 118         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
 119         struct module *provider = to_bus_provider(dev);
 120         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 121         int rc = 0;
 122
 123         if (nd_drv->remove)
 124                 rc = nd_drv->remove(dev);
 125         nd_region_disable(nvdimm_bus, dev);
 126
 127         dev_dbg(&nvdimm_bus->dev, "%s.remove(%s) = %d\n", dev->driver->name,
 128                         dev_name(dev), rc);
 129         module_put(provider);
 130         return rc;
 131 }
 132
 133 static void nvdimm_bus_shutdown(struct device *dev)
 134 {
 135         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 136         struct nd_device_driver *nd_drv = NULL;
 137
 138         if (dev->driver)
 139                 nd_drv = to_nd_device_driver(dev->driver);
 140
 141         if (nd_drv && nd_drv->shutdown) {
 142                 nd_drv->shutdown(dev);
 143                 dev_dbg(&nvdimm_bus->dev, "%s.shutdown(%s)\n",
 144                                 dev->driver->name, dev_name(dev));
 145         }
 146 }
 147
 148 void nd_device_notify(struct device *dev, enum nvdimm_event event)
 149 {
 150         device_lock(dev);
 151         if (dev->driver) {
 152                 struct nd_device_driver *nd_drv;
 153
 154                 nd_drv = to_nd_device_driver(dev->driver);
 155                 if (nd_drv->notify)
 156                         nd_drv->notify(dev, event);
 157         }
 158         device_unlock(dev);
 159 }
 160 EXPORT_SYMBOL(nd_device_notify);
 161
 162 void nvdimm_region_notify(struct nd_region *nd_region, enum nvdimm_event event)
 163 {
 164         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(&nd_region->dev);
 165
 166         if (!nvdimm_bus)
 167                 return;
 168
 169         /* caller is responsible for holding a reference on the device */
 170         nd_device_notify(&nd_region->dev, event);
 171 }
 172 EXPORT_SYMBOL_GPL(nvdimm_region_notify);
 173
 174 struct clear_badblocks_context {
 175         resource_size_t phys, cleared;
 176 };
 177
 178 static int nvdimm_clear_badblocks_region(struct device *dev, void *data)
 179 {
 180         struct clear_badblocks_context *ctx = data;
 181         struct nd_region *nd_region;
 182         resource_size_t ndr_end;
 183         sector_t sector;
 184
 185         /* make sure device is a region */
 186         if (!is_nd_pmem(dev))
 187                 return 0;
 188
 189         nd_region = to_nd_region(dev);
 190         ndr_end = nd_region->ndr_start + nd_region->ndr_size - 1;
 191
 192         /* make sure we are in the region */
 193         if (ctx->phys < nd_region->ndr_start
 194                         || (ctx->phys + ctx->cleared) > ndr_end)
 195                 return 0;
 196
 197         sector = (ctx->phys - nd_region->ndr_start) / 512;
 198         badblocks_clear(&nd_region->bb, sector, ctx->cleared / 512);
 199
 200         if (nd_region->bb_state)
 201                 sysfs_notify_dirent(nd_region->bb_state);
 202
 203         return 0;
 204 }
 205
 206 static void nvdimm_clear_badblocks_regions(struct nvdimm_bus *nvdimm_bus,
 207                 phys_addr_t phys, u64 cleared)
 208 {
 209         struct clear_badblocks_context ctx = {
 210                 .phys = phys,
 211                 .cleared = cleared,
 212         };
 213
 214         device_for_each_child(&nvdimm_bus->dev, &ctx,
 215                         nvdimm_clear_badblocks_region);
 216 }
 217
 218 static void nvdimm_account_cleared_poison(struct nvdimm_bus *nvdimm_bus,
 219                 phys_addr_t phys, u64 cleared)
 220 {
 221         if (cleared > 0)
 222                 badrange_forget(&nvdimm_bus->badrange, phys, cleared);
 223
 224         if (cleared > 0 && cleared / 512)
 225                 nvdimm_clear_badblocks_regions(nvdimm_bus, phys, cleared);
 226 }
 227
 228 long nvdimm_clear_poison(struct device *dev, phys_addr_t phys,
 229                 unsigned int len)
 230 {
 231         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 232         struct nvdimm_bus_descriptor *nd_desc;
 233         struct nd_cmd_clear_error clear_err;
 234         struct nd_cmd_ars_cap ars_cap;
 235         u32 clear_err_unit, mask;
 236         unsigned int noio_flag;
 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         noio_flag = memalloc_noio_save();
 254         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_ARS_CAP, &ars_cap,
 255                         sizeof(ars_cap), &cmd_rc);
 256         memalloc_noio_restore(noio_flag);
 257         if (rc < 0)
 258                 return rc;
 259         if (cmd_rc < 0)
 260                 return cmd_rc;
 261         clear_err_unit = ars_cap.clear_err_unit;
 262         if (!clear_err_unit || !is_power_of_2(clear_err_unit))
 263                 return -ENXIO;
 264
 265         mask = clear_err_unit - 1;
 266         if ((phys | len) & mask)
 267                 return -ENXIO;
 268         memset(&clear_err, 0, sizeof(clear_err));
 269         clear_err.address = phys;
 270         clear_err.length = len;
 271         noio_flag = memalloc_noio_save();
 272         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_CLEAR_ERROR, &clear_err,
 273                         sizeof(clear_err), &cmd_rc);
 274         memalloc_noio_restore(noio_flag);
 275         if (rc < 0)
 276                 return rc;
 277         if (cmd_rc < 0)
 278                 return cmd_rc;
 279
 280         nvdimm_account_cleared_poison(nvdimm_bus, phys, clear_err.cleared);
 281
 282         return clear_err.cleared;
 283 }
 284 EXPORT_SYMBOL_GPL(nvdimm_clear_poison);
 285
 286 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv);
 287
 288 static struct bus_type nvdimm_bus_type = {
 289         .name = "nd",
 290         .uevent = nvdimm_bus_uevent,
 291         .match = nvdimm_bus_match,
 292         .probe = nvdimm_bus_probe,
 293         .remove = nvdimm_bus_remove,
 294         .shutdown = nvdimm_bus_shutdown,
 295 };
 296
 297 static void nvdimm_bus_release(struct device *dev)
 298 {
 299         struct nvdimm_bus *nvdimm_bus;
 300
 301         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 302         ida_simple_remove(&nd_ida, nvdimm_bus->id);
 303         kfree(nvdimm_bus);
 304 }
 305
 306 static bool is_nvdimm_bus(struct device *dev)
 307 {
 308         return dev->release == nvdimm_bus_release;
 309 }
 310
 311 struct nvdimm_bus *walk_to_nvdimm_bus(struct device *nd_dev)
 312 {
 313         struct device *dev;
 314
 315         for (dev = nd_dev; dev; dev = dev->parent)
 316                 if (is_nvdimm_bus(dev))
 317                         break;
 318         dev_WARN_ONCE(nd_dev, !dev, "invalid dev, not on nd bus\n");
 319         if (dev)
 320                 return to_nvdimm_bus(dev);
 321         return NULL;
 322 }
 323
 324 struct nvdimm_bus *to_nvdimm_bus(struct device *dev)
 325 {
 326         struct nvdimm_bus *nvdimm_bus;
 327
 328         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 329         WARN_ON(!is_nvdimm_bus(dev));
 330         return nvdimm_bus;
 331 }
 332 EXPORT_SYMBOL_GPL(to_nvdimm_bus);
 333
 334 struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
 335                 struct nvdimm_bus_descriptor *nd_desc)
 336 {
 337         struct nvdimm_bus *nvdimm_bus;
 338         int rc;
 339
 340         nvdimm_bus = kzalloc(sizeof(*nvdimm_bus), GFP_KERNEL);
 341         if (!nvdimm_bus)
 342                 return NULL;
 343         INIT_LIST_HEAD(&nvdimm_bus->list);
 344         INIT_LIST_HEAD(&nvdimm_bus->mapping_list);
 345         init_waitqueue_head(&nvdimm_bus->probe_wait);
 346         nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL);
 347         mutex_init(&nvdimm_bus->reconfig_mutex);
 348         badrange_init(&nvdimm_bus->badrange);
 349         if (nvdimm_bus->id < 0) {
 350                 kfree(nvdimm_bus);
 351                 return NULL;
 352         }
 353         nvdimm_bus->nd_desc = nd_desc;
 354         nvdimm_bus->dev.parent = parent;
 355         nvdimm_bus->dev.release = nvdimm_bus_release;
 356         nvdimm_bus->dev.groups = nd_desc->attr_groups;
 357         nvdimm_bus->dev.bus = &nvdimm_bus_type;
 358         nvdimm_bus->dev.of_node = nd_desc->of_node;
 359         dev_set_name(&nvdimm_bus->dev, "ndbus%d", nvdimm_bus->id);
 360         rc = device_register(&nvdimm_bus->dev);
 361         if (rc) {
 362                 dev_dbg(&nvdimm_bus->dev, "registration failed: %d\n", rc);
 363                 goto err;
 364         }
 365
 366         return nvdimm_bus;
 367  err:
 368         put_device(&nvdimm_bus->dev);
 369         return NULL;
 370 }
 371 EXPORT_SYMBOL_GPL(nvdimm_bus_register);
 372
 373 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus)
 374 {
 375         if (!nvdimm_bus)
 376                 return;
 377         device_unregister(&nvdimm_bus->dev);
 378 }
 379 EXPORT_SYMBOL_GPL(nvdimm_bus_unregister);
 380
 381 static int child_unregister(struct device *dev, void *data)
 382 {
 383         /*
 384          * the singular ndctl class device per bus needs to be
 385          * "device_destroy"ed, so skip it here
 386          *
 387          * i.e. remove classless children
 388          */
 389         if (dev->class)
 390                 /* pass */;
 391         else
 392                 nd_device_unregister(dev, ND_SYNC);
 393         return 0;
 394 }
 395
 396 static void free_badrange_list(struct list_head *badrange_list)
 397 {
 398         struct badrange_entry *bre, *next;
 399
 400         list_for_each_entry_safe(bre, next, badrange_list, list) {
 401                 list_del(&bre->list);
 402                 kfree(bre);
 403         }
 404         list_del_init(badrange_list);
 405 }
 406
 407 static int nd_bus_remove(struct device *dev)
 408 {
 409         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 410
 411         mutex_lock(&nvdimm_bus_list_mutex);
 412         list_del_init(&nvdimm_bus->list);
 413         mutex_unlock(&nvdimm_bus_list_mutex);
 414
 415         nd_synchronize();
 416         device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister);
 417
 418         spin_lock(&nvdimm_bus->badrange.lock);
 419         free_badrange_list(&nvdimm_bus->badrange.list);
 420         spin_unlock(&nvdimm_bus->badrange.lock);
 421
 422         nvdimm_bus_destroy_ndctl(nvdimm_bus);
 423
 424         return 0;
 425 }
 426
 427 static int nd_bus_probe(struct device *dev)
 428 {
 429         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 430         int rc;
 431
 432         rc = nvdimm_bus_create_ndctl(nvdimm_bus);
 433         if (rc)
 434                 return rc;
 435
 436         mutex_lock(&nvdimm_bus_list_mutex);
 437         list_add_tail(&nvdimm_bus->list, &nvdimm_bus_list);
 438         mutex_unlock(&nvdimm_bus_list_mutex);
 439
 440         /* enable bus provider attributes to look up their local context */
 441         dev_set_drvdata(dev, nvdimm_bus->nd_desc);
 442
 443         return 0;
 444 }
 445
 446 static struct nd_device_driver nd_bus_driver = {
 447         .probe = nd_bus_probe,
 448         .remove = nd_bus_remove,
 449         .drv = {
 450                 .name = "nd_bus",
 451                 .suppress_bind_attrs = true,
 452                 .bus = &nvdimm_bus_type,
 453                 .owner = THIS_MODULE,
 454                 .mod_name = KBUILD_MODNAME,
 455         },
 456 };
 457
 458 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv)
 459 {
 460         struct nd_device_driver *nd_drv = to_nd_device_driver(drv);
 461
 462         if (is_nvdimm_bus(dev) && nd_drv == &nd_bus_driver)
 463                 return true;
 464
 465         return !!test_bit(to_nd_device_type(dev), &nd_drv->type);
 466 }
 467
 468 static ASYNC_DOMAIN_EXCLUSIVE(nd_async_domain);
 469
 470 void nd_synchronize(void)
 471 {
 472         async_synchronize_full_domain(&nd_async_domain);
 473 }
 474 EXPORT_SYMBOL_GPL(nd_synchronize);
 475
 476 static void nd_async_device_register(void *d, async_cookie_t cookie)
 477 {
 478         struct device *dev = d;
 479
 480         if (device_add(dev) != 0) {
 481                 dev_err(dev, "%s: failed\n", __func__);
 482                 put_device(dev);
 483         }
 484         put_device(dev);
 485         if (dev->parent)
 486                 put_device(dev->parent);
 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
 506         /*
 507          * Ensure that region devices always have their NUMA node set as
 508          * early as possible. This way we are able to make certain that
 509          * any memory associated with the creation and the creation
 510          * itself of the region is associated with the correct node.
 511          */
 512         if (is_nd_region(dev))
 513                 set_dev_node(dev, to_nd_region(dev)->numa_node);
 514
 515         dev->bus = &nvdimm_bus_type;
 516         if (dev->parent)
 517                 get_device(dev->parent);
 518         get_device(dev);
 519         async_schedule_domain(nd_async_device_register, dev,
 520                         &nd_async_domain);
 521 }
 522
 523 void nd_device_register(struct device *dev)
 524 {
 525         device_initialize(dev);
 526         __nd_device_register(dev);
 527 }
 528 EXPORT_SYMBOL(nd_device_register);
 529
 530 void nd_device_unregister(struct device *dev, enum nd_async_mode mode)
 531 {
 532         switch (mode) {
 533         case ND_ASYNC:
 534                 get_device(dev);
 535                 async_schedule_domain(nd_async_device_unregister, dev,
 536                                 &nd_async_domain);
 537                 break;
 538         case ND_SYNC:
 539                 nd_synchronize();
 540                 device_unregister(dev);
 541                 break;
 542         }
 543 }
 544 EXPORT_SYMBOL(nd_device_unregister);
 545
 546 /**
 547  * __nd_driver_register() - register a region or a namespace driver
 548  * @nd_drv: driver to register
 549  * @owner: automatically set by nd_driver_register() macro
 550  * @mod_name: automatically set by nd_driver_register() macro
 551  */
 552 int __nd_driver_register(struct nd_device_driver *nd_drv, struct module *owner,
 553                 const char *mod_name)
 554 {
 555         struct device_driver *drv = &nd_drv->drv;
 556
 557         if (!nd_drv->type) {
 558                 pr_debug("driver type bitmask not set (%pf)\n",
 559                                 __builtin_return_address(0));
 560                 return -EINVAL;
 561         }
 562
 563         if (!nd_drv->probe) {
 564                 pr_debug("%s ->probe() must be specified\n", mod_name);
 565                 return -EINVAL;
 566         }
 567
 568         drv->bus = &nvdimm_bus_type;
 569         drv->owner = owner;
 570         drv->mod_name = mod_name;
 571
 572         return driver_register(drv);
 573 }
 574 EXPORT_SYMBOL(__nd_driver_register);
 575
 576 int nvdimm_revalidate_disk(struct gendisk *disk)
 577 {
 578         struct device *dev = disk_to_dev(disk)->parent;
 579         struct nd_region *nd_region = to_nd_region(dev->parent);
 580         int disk_ro = get_disk_ro(disk);
 581
 582         /*
 583          * Upgrade to read-only if the region is read-only preserve as
 584          * read-only if the disk is already read-only.
 585          */
 586         if (disk_ro || nd_region->ro == disk_ro)
 587                 return 0;
 588
 589         dev_info(dev, "%s read-only, marking %s read-only\n",
 590                         dev_name(&nd_region->dev), disk->disk_name);
 591         set_disk_ro(disk, 1);
 592
 593         return 0;
 594
 595 }
 596 EXPORT_SYMBOL(nvdimm_revalidate_disk);
 597
 598 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr,
 599                 char *buf)
 600 {
 601         return sprintf(buf, ND_DEVICE_MODALIAS_FMT "\n",
 602                         to_nd_device_type(dev));
 603 }
 604 static DEVICE_ATTR_RO(modalias);
 605
 606 static ssize_t devtype_show(struct device *dev, struct device_attribute *attr,
 607                 char *buf)
 608 {
 609         return sprintf(buf, "%s\n", dev->type->name);
 610 }
 611 static DEVICE_ATTR_RO(devtype);
 612
 613 static struct attribute *nd_device_attributes[] = {
 614         &dev_attr_modalias.attr,
 615         &dev_attr_devtype.attr,
 616         NULL,
 617 };
 618
 619 /**
 620  * nd_device_attribute_group - generic attributes for all devices on an nd bus
 621  */
 622 struct attribute_group nd_device_attribute_group = {
 623         .attrs = nd_device_attributes,
 624 };
 625 EXPORT_SYMBOL_GPL(nd_device_attribute_group);
 626
 627 static ssize_t numa_node_show(struct device *dev,
 628                 struct device_attribute *attr, char *buf)
 629 {
 630         return sprintf(buf, "%d\n", dev_to_node(dev));
 631 }
 632 static DEVICE_ATTR_RO(numa_node);
 633
 634 static struct attribute *nd_numa_attributes[] = {
 635         &dev_attr_numa_node.attr,
 636         NULL,
 637 };
 638
 639 static umode_t nd_numa_attr_visible(struct kobject *kobj, struct attribute *a,
 640                 int n)
 641 {
 642         if (!IS_ENABLED(CONFIG_NUMA))
 643                 return 0;
 644
 645         return a->mode;
 646 }
 647
 648 /**
 649  * nd_numa_attribute_group - NUMA attributes for all devices on an nd bus
 650  */
 651 struct attribute_group nd_numa_attribute_group = {
 652         .attrs = nd_numa_attributes,
 653         .is_visible = nd_numa_attr_visible,
 654 };
 655 EXPORT_SYMBOL_GPL(nd_numa_attribute_group);
 656
 657 int nvdimm_bus_create_ndctl(struct nvdimm_bus *nvdimm_bus)
 658 {
 659         dev_t devt = MKDEV(nvdimm_bus_major, nvdimm_bus->id);
 660         struct device *dev;
 661
 662         dev = device_create(nd_class, &nvdimm_bus->dev, devt, nvdimm_bus,
 663                         "ndctl%d", nvdimm_bus->id);
 664
 665         if (IS_ERR(dev))
 666                 dev_dbg(&nvdimm_bus->dev, "failed to register ndctl%d: %ld\n",
 667                                 nvdimm_bus->id, PTR_ERR(dev));
 668         return PTR_ERR_OR_ZERO(dev);
 669 }
 670
 671 void nvdimm_bus_destroy_ndctl(struct nvdimm_bus *nvdimm_bus)
 672 {
 673         device_destroy(nd_class, MKDEV(nvdimm_bus_major, nvdimm_bus->id));
 674 }
 675
 676 static const struct nd_cmd_desc __nd_cmd_dimm_descs[] = {
 677         [ND_CMD_IMPLEMENTED] = { },
 678         [ND_CMD_SMART] = {
 679                 .out_num = 2,
 680                 .out_sizes = { 4, 128, },
 681         },
 682         [ND_CMD_SMART_THRESHOLD] = {
 683                 .out_num = 2,
 684                 .out_sizes = { 4, 8, },
 685         },
 686         [ND_CMD_DIMM_FLAGS] = {
 687                 .out_num = 2,
 688                 .out_sizes = { 4, 4 },
 689         },
 690         [ND_CMD_GET_CONFIG_SIZE] = {
 691                 .out_num = 3,
 692                 .out_sizes = { 4, 4, 4, },
 693         },
 694         [ND_CMD_GET_CONFIG_DATA] = {
 695                 .in_num = 2,
 696                 .in_sizes = { 4, 4, },
 697                 .out_num = 2,
 698                 .out_sizes = { 4, UINT_MAX, },
 699         },
 700         [ND_CMD_SET_CONFIG_DATA] = {
 701                 .in_num = 3,
 702                 .in_sizes = { 4, 4, UINT_MAX, },
 703                 .out_num = 1,
 704                 .out_sizes = { 4, },
 705         },
 706         [ND_CMD_VENDOR] = {
 707                 .in_num = 3,
 708                 .in_sizes = { 4, 4, UINT_MAX, },
 709                 .out_num = 3,
 710                 .out_sizes = { 4, 4, UINT_MAX, },
 711         },
 712         [ND_CMD_CALL] = {
 713                 .in_num = 2,
 714                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 715                 .out_num = 1,
 716                 .out_sizes = { UINT_MAX, },
 717         },
 718 };
 719
 720 const struct nd_cmd_desc *nd_cmd_dimm_desc(int cmd)
 721 {
 722         if (cmd < ARRAY_SIZE(__nd_cmd_dimm_descs))
 723                 return &__nd_cmd_dimm_descs[cmd];
 724         return NULL;
 725 }
 726 EXPORT_SYMBOL_GPL(nd_cmd_dimm_desc);
 727
 728 static const struct nd_cmd_desc __nd_cmd_bus_descs[] = {
 729         [ND_CMD_IMPLEMENTED] = { },
 730         [ND_CMD_ARS_CAP] = {
 731                 .in_num = 2,
 732                 .in_sizes = { 8, 8, },
 733                 .out_num = 4,
 734                 .out_sizes = { 4, 4, 4, 4, },
 735         },
 736         [ND_CMD_ARS_START] = {
 737                 .in_num = 5,
 738                 .in_sizes = { 8, 8, 2, 1, 5, },
 739                 .out_num = 2,
 740                 .out_sizes = { 4, 4, },
 741         },
 742         [ND_CMD_ARS_STATUS] = {
 743                 .out_num = 3,
 744                 .out_sizes = { 4, 4, UINT_MAX, },
 745         },
 746         [ND_CMD_CLEAR_ERROR] = {
 747                 .in_num = 2,
 748                 .in_sizes = { 8, 8, },
 749                 .out_num = 3,
 750                 .out_sizes = { 4, 4, 8, },
 751         },
 752         [ND_CMD_CALL] = {
 753                 .in_num = 2,
 754                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 755                 .out_num = 1,
 756                 .out_sizes = { UINT_MAX, },
 757         },
 758 };
 759
 760 const struct nd_cmd_desc *nd_cmd_bus_desc(int cmd)
 761 {
 762         if (cmd < ARRAY_SIZE(__nd_cmd_bus_descs))
 763                 return &__nd_cmd_bus_descs[cmd];
 764         return NULL;
 765 }
 766 EXPORT_SYMBOL_GPL(nd_cmd_bus_desc);
 767
 768 u32 nd_cmd_in_size(struct nvdimm *nvdimm, int cmd,
 769                 const struct nd_cmd_desc *desc, int idx, void *buf)
 770 {
 771         if (idx >= desc->in_num)
 772                 return UINT_MAX;
 773
 774         if (desc->in_sizes[idx] < UINT_MAX)
 775                 return desc->in_sizes[idx];
 776
 777         if (nvdimm && cmd == ND_CMD_SET_CONFIG_DATA && idx == 2) {
 778                 struct nd_cmd_set_config_hdr *hdr = buf;
 779
 780                 return hdr->in_length;
 781         } else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) {
 782                 struct nd_cmd_vendor_hdr *hdr = buf;
 783
 784                 return hdr->in_length;
 785         } else if (cmd == ND_CMD_CALL) {
 786                 struct nd_cmd_pkg *pkg = buf;
 787
 788                 return pkg->nd_size_in;
 789         }
 790
 791         return UINT_MAX;
 792 }
 793 EXPORT_SYMBOL_GPL(nd_cmd_in_size);
 794
 795 u32 nd_cmd_out_size(struct nvdimm *nvdimm, int cmd,
 796                 const struct nd_cmd_desc *desc, int idx, const u32 *in_field,
 797                 const u32 *out_field, unsigned long remainder)
 798 {
 799         if (idx >= desc->out_num)
 800                 return UINT_MAX;
 801
 802         if (desc->out_sizes[idx] < UINT_MAX)
 803                 return desc->out_sizes[idx];
 804
 805         if (nvdimm && cmd == ND_CMD_GET_CONFIG_DATA && idx == 1)
 806                 return in_field[1];
 807         else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2)
 808                 return out_field[1];
 809         else if (!nvdimm && cmd == ND_CMD_ARS_STATUS && idx == 2) {
 810                 /*
 811                  * Per table 9-276 ARS Data in ACPI 6.1, out_field[1] is
 812                  * "Size of Output Buffer in bytes, including this
 813                  * field."
 814                  */
 815                 if (out_field[1] < 4)
 816                         return 0;
 817                 /*
 818                  * ACPI 6.1 is ambiguous if 'status' is included in the
 819                  * output size. If we encounter an output size that
 820                  * overshoots the remainder by 4 bytes, assume it was
 821                  * including 'status'.
 822                  */
 823                 if (out_field[1] - 4 == remainder)
 824                         return remainder;
 825                 return out_field[1] - 8;
 826         } else if (cmd == ND_CMD_CALL) {
 827                 struct nd_cmd_pkg *pkg = (struct nd_cmd_pkg *) in_field;
 828
 829                 return pkg->nd_size_out;
 830         }
 831
 832
 833         return UINT_MAX;
 834 }
 835 EXPORT_SYMBOL_GPL(nd_cmd_out_size);
 836
 837 void wait_nvdimm_bus_probe_idle(struct device *dev)
 838 {
 839         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 840
 841         do {
 842                 if (nvdimm_bus->probe_active == 0)
 843                         break;
 844                 nvdimm_bus_unlock(&nvdimm_bus->dev);
 845                 wait_event(nvdimm_bus->probe_wait,
 846                                 nvdimm_bus->probe_active == 0);
 847                 nvdimm_bus_lock(&nvdimm_bus->dev);
 848         } while (true);
 849 }
 850
 851 static int nd_pmem_forget_poison_check(struct device *dev, void *data)
 852 {
 853         struct nd_cmd_clear_error *clear_err =
 854                 (struct nd_cmd_clear_error *)data;
 855         struct nd_btt *nd_btt = is_nd_btt(dev) ? to_nd_btt(dev) : NULL;
 856         struct nd_pfn *nd_pfn = is_nd_pfn(dev) ? to_nd_pfn(dev) : NULL;
 857         struct nd_dax *nd_dax = is_nd_dax(dev) ? to_nd_dax(dev) : NULL;
 858         struct nd_namespace_common *ndns = NULL;
 859         struct nd_namespace_io *nsio;
 860         resource_size_t offset = 0, end_trunc = 0, start, end, pstart, pend;
 861
 862         if (nd_dax || !dev->driver)
 863                 return 0;
 864
 865         start = clear_err->address;
 866         end = clear_err->address + clear_err->cleared - 1;
 867
 868         if (nd_btt || nd_pfn || nd_dax) {
 869                 if (nd_btt)
 870                         ndns = nd_btt->ndns;
 871                 else if (nd_pfn)
 872                         ndns = nd_pfn->ndns;
 873                 else if (nd_dax)
 874                         ndns = nd_dax->nd_pfn.ndns;
 875
 876                 if (!ndns)
 877                         return 0;
 878         } else
 879                 ndns = to_ndns(dev);
 880
 881         nsio = to_nd_namespace_io(&ndns->dev);
 882         pstart = nsio->res.start + offset;
 883         pend = nsio->res.end - end_trunc;
 884
 885         if ((pstart >= start) && (pend <= end))
 886                 return -EBUSY;
 887
 888         return 0;
 889
 890 }
 891
 892 static int nd_ns_forget_poison_check(struct device *dev, void *data)
 893 {
 894         return device_for_each_child(dev, data, nd_pmem_forget_poison_check);
 895 }
 896
 897 /* set_config requires an idle interleave set */
 898 static int nd_cmd_clear_to_send(struct nvdimm_bus *nvdimm_bus,
 899                 struct nvdimm *nvdimm, unsigned int cmd, void *data)
 900 {
 901         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 902
 903         /* ask the bus provider if it would like to block this request */
 904         if (nd_desc->clear_to_send) {
 905                 int rc = nd_desc->clear_to_send(nd_desc, nvdimm, cmd);
 906
 907                 if (rc)
 908                         return rc;
 909         }
 910
 911         /* require clear error to go through the pmem driver */
 912         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR)
 913                 return device_for_each_child(&nvdimm_bus->dev, data,
 914                                 nd_ns_forget_poison_check);
 915
 916         if (!nvdimm || cmd != ND_CMD_SET_CONFIG_DATA)
 917                 return 0;
 918
 919         /* prevent label manipulation while the kernel owns label updates */
 920         wait_nvdimm_bus_probe_idle(&nvdimm_bus->dev);
 921         if (atomic_read(&nvdimm->busy))
 922                 return -EBUSY;
 923         return 0;
 924 }
 925
 926 static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm,
 927                 int read_only, unsigned int ioctl_cmd, unsigned long arg)
 928 {
 929         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 930         static char out_env[ND_CMD_MAX_ENVELOPE];
 931         static char in_env[ND_CMD_MAX_ENVELOPE];
 932         const struct nd_cmd_desc *desc = NULL;
 933         unsigned int cmd = _IOC_NR(ioctl_cmd);
 934         struct device *dev = &nvdimm_bus->dev;
 935         void __user *p = (void __user *) arg;
 936         const char *cmd_name, *dimm_name;
 937         u32 in_len = 0, out_len = 0;
 938         unsigned int func = cmd;
 939         unsigned long cmd_mask;
 940         struct nd_cmd_pkg pkg;
 941         int rc, i, cmd_rc;
 942         u64 buf_len = 0;
 943         void *buf;
 944
 945         if (nvdimm) {
 946                 desc = nd_cmd_dimm_desc(cmd);
 947                 cmd_name = nvdimm_cmd_name(cmd);
 948                 cmd_mask = nvdimm->cmd_mask;
 949                 dimm_name = dev_name(&nvdimm->dev);
 950         } else {
 951                 desc = nd_cmd_bus_desc(cmd);
 952                 cmd_name = nvdimm_bus_cmd_name(cmd);
 953                 cmd_mask = nd_desc->cmd_mask;
 954                 dimm_name = "bus";
 955         }
 956
 957         if (cmd == ND_CMD_CALL) {
 958                 if (copy_from_user(&pkg, p, sizeof(pkg)))
 959                         return -EFAULT;
 960         }
 961
 962         if (!desc || (desc->out_num + desc->in_num == 0) ||
 963                         !test_bit(cmd, &cmd_mask))
 964                 return -ENOTTY;
 965
 966         /* fail write commands (when read-only) */
 967         if (read_only)
 968                 switch (cmd) {
 969                 case ND_CMD_VENDOR:
 970                 case ND_CMD_SET_CONFIG_DATA:
 971                 case ND_CMD_ARS_START:
 972                 case ND_CMD_CLEAR_ERROR:
 973                 case ND_CMD_CALL:
 974                         dev_dbg(&nvdimm_bus->dev, "'%s' command while read-only.\n",
 975                                         nvdimm ? nvdimm_cmd_name(cmd)
 976                                         : nvdimm_bus_cmd_name(cmd));
 977                         return -EPERM;
 978                 default:
 979                         break;
 980                 }
 981
 982         /* process an input envelope */
 983         for (i = 0; i < desc->in_num; i++) {
 984                 u32 in_size, copy;
 985
 986                 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env);
 987                 if (in_size == UINT_MAX) {
 988                         dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n",
 989                                         __func__, dimm_name, cmd_name, i);
 990                         return -ENXIO;
 991                 }
 992                 if (in_len < sizeof(in_env))
 993                         copy = min_t(u32, sizeof(in_env) - in_len, in_size);
 994                 else
 995                         copy = 0;
 996                 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy))
 997                         return -EFAULT;
 998                 in_len += in_size;
 999         }
1000
1001         if (cmd == ND_CMD_CALL) {
1002                 func = pkg.nd_command;
1003                 dev_dbg(dev, "%s, idx: %llu, in: %u, out: %u, len %llu\n",
1004                                 dimm_name, pkg.nd_command,
1005                                 in_len, out_len, buf_len);
1006         }
1007
1008         /* process an output envelope */
1009         for (i = 0; i < desc->out_num; i++) {
1010                 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i,
1011                                 (u32 *) in_env, (u32 *) out_env, 0);
1012                 u32 copy;
1013
1014                 if (out_size == UINT_MAX) {
1015                         dev_dbg(dev, "%s unknown output size cmd: %s field: %d\n",
1016                                         dimm_name, cmd_name, i);
1017                         return -EFAULT;
1018                 }
1019                 if (out_len < sizeof(out_env))
1020                         copy = min_t(u32, sizeof(out_env) - out_len, out_size);
1021                 else
1022                         copy = 0;
1023                 if (copy && copy_from_user(&out_env[out_len],
1024                                         p + in_len + out_len, copy))
1025                         return -EFAULT;
1026                 out_len += out_size;
1027         }
1028
1029         buf_len = (u64) out_len + (u64) in_len;
1030         if (buf_len > ND_IOCTL_MAX_BUFLEN) {
1031                 dev_dbg(dev, "%s cmd: %s buf_len: %llu > %d\n", dimm_name,
1032                                 cmd_name, buf_len, ND_IOCTL_MAX_BUFLEN);
1033                 return -EINVAL;
1034         }
1035
1036         buf = vmalloc(buf_len);
1037         if (!buf)
1038                 return -ENOMEM;
1039
1040         if (copy_from_user(buf, p, buf_len)) {
1041                 rc = -EFAULT;
1042                 goto out;
1043         }
1044
1045         nvdimm_bus_lock(&nvdimm_bus->dev);
1046         rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, func, buf);
1047         if (rc)
1048                 goto out_unlock;
1049
1050         rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc);
1051         if (rc < 0)
1052                 goto out_unlock;
1053
1054         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) {
1055                 struct nd_cmd_clear_error *clear_err = buf;
1056
1057                 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address,
1058                                 clear_err->cleared);
1059         }
1060         nvdimm_bus_unlock(&nvdimm_bus->dev);
1061
1062         if (copy_to_user(p, buf, buf_len))
1063                 rc = -EFAULT;
1064
1065         vfree(buf);
1066         return rc;
1067
1068  out_unlock:
1069         nvdimm_bus_unlock(&nvdimm_bus->dev);
1070  out:
1071         vfree(buf);
1072         return rc;
1073 }
1074
1075 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1076 {
1077         long id = (long) file->private_data;
1078         int rc = -ENXIO, ro;
1079         struct nvdimm_bus *nvdimm_bus;
1080
1081         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1082         mutex_lock(&nvdimm_bus_list_mutex);
1083         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1084                 if (nvdimm_bus->id == id) {
1085                         rc = __nd_ioctl(nvdimm_bus, NULL, ro, cmd, arg);
1086                         break;
1087                 }
1088         }
1089         mutex_unlock(&nvdimm_bus_list_mutex);
1090
1091         return rc;
1092 }
1093
1094 static int match_dimm(struct device *dev, void *data)
1095 {
1096         long id = (long) data;
1097
1098         if (is_nvdimm(dev)) {
1099                 struct nvdimm *nvdimm = to_nvdimm(dev);
1100
1101                 return nvdimm->id == id;
1102         }
1103
1104         return 0;
1105 }
1106
1107 static long nvdimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1108 {
1109         int rc = -ENXIO, ro;
1110         struct nvdimm_bus *nvdimm_bus;
1111
1112         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1113         mutex_lock(&nvdimm_bus_list_mutex);
1114         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1115                 struct device *dev = device_find_child(&nvdimm_bus->dev,
1116                                 file->private_data, match_dimm);
1117                 struct nvdimm *nvdimm;
1118
1119                 if (!dev)
1120                         continue;
1121
1122                 nvdimm = to_nvdimm(dev);
1123                 rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg);
1124                 put_device(dev);
1125                 break;
1126         }
1127         mutex_unlock(&nvdimm_bus_list_mutex);
1128
1129         return rc;
1130 }
1131
1132 static int nd_open(struct inode *inode, struct file *file)
1133 {
1134         long minor = iminor(inode);
1135
1136         file->private_data = (void *) minor;
1137         return 0;
1138 }
1139
1140 static const struct file_operations nvdimm_bus_fops = {
1141         .owner = THIS_MODULE,
1142         .open = nd_open,
1143         .unlocked_ioctl = nd_ioctl,
1144         .compat_ioctl = nd_ioctl,
1145         .llseek = noop_llseek,
1146 };
1147
1148 static const struct file_operations nvdimm_fops = {
1149         .owner = THIS_MODULE,
1150         .open = nd_open,
1151         .unlocked_ioctl = nvdimm_ioctl,
1152         .compat_ioctl = nvdimm_ioctl,
1153         .llseek = noop_llseek,
1154 };
1155
1156 int __init nvdimm_bus_init(void)
1157 {
1158         int rc;
1159
1160         rc = bus_register(&nvdimm_bus_type);
1161         if (rc)
1162                 return rc;
1163
1164         rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops);
1165         if (rc < 0)
1166                 goto err_bus_chrdev;
1167         nvdimm_bus_major = rc;
1168
1169         rc = register_chrdev(0, "dimmctl", &nvdimm_fops);
1170         if (rc < 0)
1171                 goto err_dimm_chrdev;
1172         nvdimm_major = rc;
1173
1174         nd_class = class_create(THIS_MODULE, "nd");
1175         if (IS_ERR(nd_class)) {
1176                 rc = PTR_ERR(nd_class);
1177                 goto err_class;
1178         }
1179
1180         rc = driver_register(&nd_bus_driver.drv);
1181         if (rc)
1182                 goto err_nd_bus;
1183
1184         return 0;
1185
1186  err_nd_bus:
1187         class_destroy(nd_class);
1188  err_class:
1189         unregister_chrdev(nvdimm_major, "dimmctl");
1190  err_dimm_chrdev:
1191         unregister_chrdev(nvdimm_bus_major, "ndctl");
1192  err_bus_chrdev:
1193         bus_unregister(&nvdimm_bus_type);
1194
1195         return rc;
1196 }
1197
1198 void nvdimm_bus_exit(void)
1199 {
1200         driver_unregister(&nd_bus_driver.drv);
1201         class_destroy(nd_class);
1202         unregister_chrdev(nvdimm_bus_major, "ndctl");
1203         unregister_chrdev(nvdimm_major, "dimmctl");
1204         bus_unregister(&nvdimm_bus_type);
1205         ida_destroy(&nd_ida);
1206 }