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         /*
  58          * Ensure that region devices always have their numa node set as
  59          * early as possible.
  60          */
  61         if (is_nd_region(dev))
  62                 set_dev_node(dev, to_nd_region(dev)->numa_node);
  63         return add_uevent_var(env, "MODALIAS=" ND_DEVICE_MODALIAS_FMT,
  64                         to_nd_device_type(dev));
  65 }
  66
  67 static struct module *to_bus_provider(struct device *dev)
  68 {
  69         /* pin bus providers while regions are enabled */
  70         if (is_nd_region(dev)) {
  71                 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
  72
  73                 return nvdimm_bus->nd_desc->module;
  74         }
  75         return NULL;
  76 }
  77
  78 static void nvdimm_bus_probe_start(struct nvdimm_bus *nvdimm_bus)
  79 {
  80         nvdimm_bus_lock(&nvdimm_bus->dev);
  81         nvdimm_bus->probe_active++;
  82         nvdimm_bus_unlock(&nvdimm_bus->dev);
  83 }
  84
  85 static void nvdimm_bus_probe_end(struct nvdimm_bus *nvdimm_bus)
  86 {
  87         nvdimm_bus_lock(&nvdimm_bus->dev);
  88         if (--nvdimm_bus->probe_active == 0)
  89                 wake_up(&nvdimm_bus->probe_wait);
  90         nvdimm_bus_unlock(&nvdimm_bus->dev);
  91 }
  92
  93 static int nvdimm_bus_probe(struct device *dev)
  94 {
  95         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
  96         struct module *provider = to_bus_provider(dev);
  97         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
  98         int rc;
  99
 100         if (!try_module_get(provider))
 101                 return -ENXIO;
 102
 103         dev_dbg(&nvdimm_bus->dev, "START: %s.probe(%s)\n",
 104                         dev->driver->name, dev_name(dev));
 105
 106         nvdimm_bus_probe_start(nvdimm_bus);
 107         rc = nd_drv->probe(dev);
 108         if (rc == 0)
 109                 nd_region_probe_success(nvdimm_bus, dev);
 110         else
 111                 nd_region_disable(nvdimm_bus, dev);
 112         nvdimm_bus_probe_end(nvdimm_bus);
 113
 114         dev_dbg(&nvdimm_bus->dev, "END: %s.probe(%s) = %d\n", dev->driver->name,
 115                         dev_name(dev), rc);
 116
 117         if (rc != 0)
 118                 module_put(provider);
 119         return rc;
 120 }
 121
 122 static int nvdimm_bus_remove(struct device *dev)
 123 {
 124         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
 125         struct module *provider = to_bus_provider(dev);
 126         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 127         int rc = 0;
 128
 129         if (nd_drv->remove)
 130                 rc = nd_drv->remove(dev);
 131         nd_region_disable(nvdimm_bus, dev);
 132
 133         dev_dbg(&nvdimm_bus->dev, "%s.remove(%s) = %d\n", dev->driver->name,
 134                         dev_name(dev), rc);
 135         module_put(provider);
 136         return rc;
 137 }
 138
 139 static void nvdimm_bus_shutdown(struct device *dev)
 140 {
 141         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 142         struct nd_device_driver *nd_drv = NULL;
 143
 144         if (dev->driver)
 145                 nd_drv = to_nd_device_driver(dev->driver);
 146
 147         if (nd_drv && nd_drv->shutdown) {
 148                 nd_drv->shutdown(dev);
 149                 dev_dbg(&nvdimm_bus->dev, "%s.shutdown(%s)\n",
 150                                 dev->driver->name, dev_name(dev));
 151         }
 152 }
 153
 154 void nd_device_notify(struct device *dev, enum nvdimm_event event)
 155 {
 156         device_lock(dev);
 157         if (dev->driver) {
 158                 struct nd_device_driver *nd_drv;
 159
 160                 nd_drv = to_nd_device_driver(dev->driver);
 161                 if (nd_drv->notify)
 162                         nd_drv->notify(dev, event);
 163         }
 164         device_unlock(dev);
 165 }
 166 EXPORT_SYMBOL(nd_device_notify);
 167
 168 void nvdimm_region_notify(struct nd_region *nd_region, enum nvdimm_event event)
 169 {
 170         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(&nd_region->dev);
 171
 172         if (!nvdimm_bus)
 173                 return;
 174
 175         /* caller is responsible for holding a reference on the device */
 176         nd_device_notify(&nd_region->dev, event);
 177 }
 178 EXPORT_SYMBOL_GPL(nvdimm_region_notify);
 179
 180 struct clear_badblocks_context {
 181         resource_size_t phys, cleared;
 182 };
 183
 184 static int nvdimm_clear_badblocks_region(struct device *dev, void *data)
 185 {
 186         struct clear_badblocks_context *ctx = data;
 187         struct nd_region *nd_region;
 188         resource_size_t ndr_end;
 189         sector_t sector;
 190
 191         /* make sure device is a region */
 192         if (!is_nd_pmem(dev))
 193                 return 0;
 194
 195         nd_region = to_nd_region(dev);
 196         ndr_end = nd_region->ndr_start + nd_region->ndr_size - 1;
 197
 198         /* make sure we are in the region */
 199         if (ctx->phys < nd_region->ndr_start
 200                         || (ctx->phys + ctx->cleared) > ndr_end)
 201                 return 0;
 202
 203         sector = (ctx->phys - nd_region->ndr_start) / 512;
 204         badblocks_clear(&nd_region->bb, sector, ctx->cleared / 512);
 205
 206         if (nd_region->bb_state)
 207                 sysfs_notify_dirent(nd_region->bb_state);
 208
 209         return 0;
 210 }
 211
 212 static void nvdimm_clear_badblocks_regions(struct nvdimm_bus *nvdimm_bus,
 213                 phys_addr_t phys, u64 cleared)
 214 {
 215         struct clear_badblocks_context ctx = {
 216                 .phys = phys,
 217                 .cleared = cleared,
 218         };
 219
 220         device_for_each_child(&nvdimm_bus->dev, &ctx,
 221                         nvdimm_clear_badblocks_region);
 222 }
 223
 224 static void nvdimm_account_cleared_poison(struct nvdimm_bus *nvdimm_bus,
 225                 phys_addr_t phys, u64 cleared)
 226 {
 227         if (cleared > 0)
 228                 badrange_forget(&nvdimm_bus->badrange, phys, cleared);
 229
 230         if (cleared > 0 && cleared / 512)
 231                 nvdimm_clear_badblocks_regions(nvdimm_bus, phys, cleared);
 232 }
 233
 234 long nvdimm_clear_poison(struct device *dev, phys_addr_t phys,
 235                 unsigned int len)
 236 {
 237         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 238         struct nvdimm_bus_descriptor *nd_desc;
 239         struct nd_cmd_clear_error clear_err;
 240         struct nd_cmd_ars_cap ars_cap;
 241         u32 clear_err_unit, mask;
 242         unsigned int noio_flag;
 243         int cmd_rc, rc;
 244
 245         if (!nvdimm_bus)
 246                 return -ENXIO;
 247
 248         nd_desc = nvdimm_bus->nd_desc;
 249         /*
 250          * if ndctl does not exist, it's PMEM_LEGACY and
 251          * we want to just pretend everything is handled.
 252          */
 253         if (!nd_desc->ndctl)
 254                 return len;
 255
 256         memset(&ars_cap, 0, sizeof(ars_cap));
 257         ars_cap.address = phys;
 258         ars_cap.length = len;
 259         noio_flag = memalloc_noio_save();
 260         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_ARS_CAP, &ars_cap,
 261                         sizeof(ars_cap), &cmd_rc);
 262         memalloc_noio_restore(noio_flag);
 263         if (rc < 0)
 264                 return rc;
 265         if (cmd_rc < 0)
 266                 return cmd_rc;
 267         clear_err_unit = ars_cap.clear_err_unit;
 268         if (!clear_err_unit || !is_power_of_2(clear_err_unit))
 269                 return -ENXIO;
 270
 271         mask = clear_err_unit - 1;
 272         if ((phys | len) & mask)
 273                 return -ENXIO;
 274         memset(&clear_err, 0, sizeof(clear_err));
 275         clear_err.address = phys;
 276         clear_err.length = len;
 277         noio_flag = memalloc_noio_save();
 278         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_CLEAR_ERROR, &clear_err,
 279                         sizeof(clear_err), &cmd_rc);
 280         memalloc_noio_restore(noio_flag);
 281         if (rc < 0)
 282                 return rc;
 283         if (cmd_rc < 0)
 284                 return cmd_rc;
 285
 286         nvdimm_account_cleared_poison(nvdimm_bus, phys, clear_err.cleared);
 287
 288         return clear_err.cleared;
 289 }
 290 EXPORT_SYMBOL_GPL(nvdimm_clear_poison);
 291
 292 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv);
 293
 294 static struct bus_type nvdimm_bus_type = {
 295         .name = "nd",
 296         .uevent = nvdimm_bus_uevent,
 297         .match = nvdimm_bus_match,
 298         .probe = nvdimm_bus_probe,
 299         .remove = nvdimm_bus_remove,
 300         .shutdown = nvdimm_bus_shutdown,
 301 };
 302
 303 static void nvdimm_bus_release(struct device *dev)
 304 {
 305         struct nvdimm_bus *nvdimm_bus;
 306
 307         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 308         ida_simple_remove(&nd_ida, nvdimm_bus->id);
 309         kfree(nvdimm_bus);
 310 }
 311
 312 static bool is_nvdimm_bus(struct device *dev)
 313 {
 314         return dev->release == nvdimm_bus_release;
 315 }
 316
 317 struct nvdimm_bus *walk_to_nvdimm_bus(struct device *nd_dev)
 318 {
 319         struct device *dev;
 320
 321         for (dev = nd_dev; dev; dev = dev->parent)
 322                 if (is_nvdimm_bus(dev))
 323                         break;
 324         dev_WARN_ONCE(nd_dev, !dev, "invalid dev, not on nd bus\n");
 325         if (dev)
 326                 return to_nvdimm_bus(dev);
 327         return NULL;
 328 }
 329
 330 struct nvdimm_bus *to_nvdimm_bus(struct device *dev)
 331 {
 332         struct nvdimm_bus *nvdimm_bus;
 333
 334         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
 335         WARN_ON(!is_nvdimm_bus(dev));
 336         return nvdimm_bus;
 337 }
 338 EXPORT_SYMBOL_GPL(to_nvdimm_bus);
 339
 340 struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
 341                 struct nvdimm_bus_descriptor *nd_desc)
 342 {
 343         struct nvdimm_bus *nvdimm_bus;
 344         int rc;
 345
 346         nvdimm_bus = kzalloc(sizeof(*nvdimm_bus), GFP_KERNEL);
 347         if (!nvdimm_bus)
 348                 return NULL;
 349         INIT_LIST_HEAD(&nvdimm_bus->list);
 350         INIT_LIST_HEAD(&nvdimm_bus->mapping_list);
 351         init_waitqueue_head(&nvdimm_bus->probe_wait);
 352         nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL);
 353         mutex_init(&nvdimm_bus->reconfig_mutex);
 354         badrange_init(&nvdimm_bus->badrange);
 355         if (nvdimm_bus->id < 0) {
 356                 kfree(nvdimm_bus);
 357                 return NULL;
 358         }
 359         nvdimm_bus->nd_desc = nd_desc;
 360         nvdimm_bus->dev.parent = parent;
 361         nvdimm_bus->dev.release = nvdimm_bus_release;
 362         nvdimm_bus->dev.groups = nd_desc->attr_groups;
 363         nvdimm_bus->dev.bus = &nvdimm_bus_type;
 364         nvdimm_bus->dev.of_node = nd_desc->of_node;
 365         dev_set_name(&nvdimm_bus->dev, "ndbus%d", nvdimm_bus->id);
 366         rc = device_register(&nvdimm_bus->dev);
 367         if (rc) {
 368                 dev_dbg(&nvdimm_bus->dev, "registration failed: %d\n", rc);
 369                 goto err;
 370         }
 371
 372         return nvdimm_bus;
 373  err:
 374         put_device(&nvdimm_bus->dev);
 375         return NULL;
 376 }
 377 EXPORT_SYMBOL_GPL(nvdimm_bus_register);
 378
 379 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus)
 380 {
 381         if (!nvdimm_bus)
 382                 return;
 383         device_unregister(&nvdimm_bus->dev);
 384 }
 385 EXPORT_SYMBOL_GPL(nvdimm_bus_unregister);
 386
 387 static int child_unregister(struct device *dev, void *data)
 388 {
 389         /*
 390          * the singular ndctl class device per bus needs to be
 391          * "device_destroy"ed, so skip it here
 392          *
 393          * i.e. remove classless children
 394          */
 395         if (dev->class)
 396                 /* pass */;
 397         else
 398                 nd_device_unregister(dev, ND_SYNC);
 399         return 0;
 400 }
 401
 402 static void free_badrange_list(struct list_head *badrange_list)
 403 {
 404         struct badrange_entry *bre, *next;
 405
 406         list_for_each_entry_safe(bre, next, badrange_list, list) {
 407                 list_del(&bre->list);
 408                 kfree(bre);
 409         }
 410         list_del_init(badrange_list);
 411 }
 412
 413 static int nd_bus_remove(struct device *dev)
 414 {
 415         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 416
 417         mutex_lock(&nvdimm_bus_list_mutex);
 418         list_del_init(&nvdimm_bus->list);
 419         mutex_unlock(&nvdimm_bus_list_mutex);
 420
 421         nd_synchronize();
 422         device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister);
 423
 424         spin_lock(&nvdimm_bus->badrange.lock);
 425         free_badrange_list(&nvdimm_bus->badrange.list);
 426         spin_unlock(&nvdimm_bus->badrange.lock);
 427
 428         nvdimm_bus_destroy_ndctl(nvdimm_bus);
 429
 430         return 0;
 431 }
 432
 433 static int nd_bus_probe(struct device *dev)
 434 {
 435         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
 436         int rc;
 437
 438         rc = nvdimm_bus_create_ndctl(nvdimm_bus);
 439         if (rc)
 440                 return rc;
 441
 442         mutex_lock(&nvdimm_bus_list_mutex);
 443         list_add_tail(&nvdimm_bus->list, &nvdimm_bus_list);
 444         mutex_unlock(&nvdimm_bus_list_mutex);
 445
 446         /* enable bus provider attributes to look up their local context */
 447         dev_set_drvdata(dev, nvdimm_bus->nd_desc);
 448
 449         return 0;
 450 }
 451
 452 static struct nd_device_driver nd_bus_driver = {
 453         .probe = nd_bus_probe,
 454         .remove = nd_bus_remove,
 455         .drv = {
 456                 .name = "nd_bus",
 457                 .suppress_bind_attrs = true,
 458                 .bus = &nvdimm_bus_type,
 459                 .owner = THIS_MODULE,
 460                 .mod_name = KBUILD_MODNAME,
 461         },
 462 };
 463
 464 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv)
 465 {
 466         struct nd_device_driver *nd_drv = to_nd_device_driver(drv);
 467
 468         if (is_nvdimm_bus(dev) && nd_drv == &nd_bus_driver)
 469                 return true;
 470
 471         return !!test_bit(to_nd_device_type(dev), &nd_drv->type);
 472 }
 473
 474 static ASYNC_DOMAIN_EXCLUSIVE(nd_async_domain);
 475
 476 void nd_synchronize(void)
 477 {
 478         async_synchronize_full_domain(&nd_async_domain);
 479 }
 480 EXPORT_SYMBOL_GPL(nd_synchronize);
 481
 482 static void nd_async_device_register(void *d, async_cookie_t cookie)
 483 {
 484         struct device *dev = d;
 485
 486         if (device_add(dev) != 0) {
 487                 dev_err(dev, "%s: failed\n", __func__);
 488                 put_device(dev);
 489         }
 490         put_device(dev);
 491 }
 492
 493 static void nd_async_device_unregister(void *d, async_cookie_t cookie)
 494 {
 495         struct device *dev = d;
 496
 497         /* flush bus operations before delete */
 498         nvdimm_bus_lock(dev);
 499         nvdimm_bus_unlock(dev);
 500
 501         device_unregister(dev);
 502         put_device(dev);
 503 }
 504
 505 void __nd_device_register(struct device *dev)
 506 {
 507         if (!dev)
 508                 return;
 509         dev->bus = &nvdimm_bus_type;
 510         get_device(dev);
 511         async_schedule_domain(nd_async_device_register, dev,
 512                         &nd_async_domain);
 513 }
 514
 515 void nd_device_register(struct device *dev)
 516 {
 517         device_initialize(dev);
 518         __nd_device_register(dev);
 519 }
 520 EXPORT_SYMBOL(nd_device_register);
 521
 522 void nd_device_unregister(struct device *dev, enum nd_async_mode mode)
 523 {
 524         switch (mode) {
 525         case ND_ASYNC:
 526                 get_device(dev);
 527                 async_schedule_domain(nd_async_device_unregister, dev,
 528                                 &nd_async_domain);
 529                 break;
 530         case ND_SYNC:
 531                 nd_synchronize();
 532                 device_unregister(dev);
 533                 break;
 534         }
 535 }
 536 EXPORT_SYMBOL(nd_device_unregister);
 537
 538 /**
 539  * __nd_driver_register() - register a region or a namespace driver
 540  * @nd_drv: driver to register
 541  * @owner: automatically set by nd_driver_register() macro
 542  * @mod_name: automatically set by nd_driver_register() macro
 543  */
 544 int __nd_driver_register(struct nd_device_driver *nd_drv, struct module *owner,
 545                 const char *mod_name)
 546 {
 547         struct device_driver *drv = &nd_drv->drv;
 548
 549         if (!nd_drv->type) {
 550                 pr_debug("driver type bitmask not set (%pf)\n",
 551                                 __builtin_return_address(0));
 552                 return -EINVAL;
 553         }
 554
 555         if (!nd_drv->probe) {
 556                 pr_debug("%s ->probe() must be specified\n", mod_name);
 557                 return -EINVAL;
 558         }
 559
 560         drv->bus = &nvdimm_bus_type;
 561         drv->owner = owner;
 562         drv->mod_name = mod_name;
 563
 564         return driver_register(drv);
 565 }
 566 EXPORT_SYMBOL(__nd_driver_register);
 567
 568 int nvdimm_revalidate_disk(struct gendisk *disk)
 569 {
 570         struct device *dev = disk_to_dev(disk)->parent;
 571         struct nd_region *nd_region = to_nd_region(dev->parent);
 572         int disk_ro = get_disk_ro(disk);
 573
 574         /*
 575          * Upgrade to read-only if the region is read-only preserve as
 576          * read-only if the disk is already read-only.
 577          */
 578         if (disk_ro || nd_region->ro == disk_ro)
 579                 return 0;
 580
 581         dev_info(dev, "%s read-only, marking %s read-only\n",
 582                         dev_name(&nd_region->dev), disk->disk_name);
 583         set_disk_ro(disk, 1);
 584
 585         return 0;
 586
 587 }
 588 EXPORT_SYMBOL(nvdimm_revalidate_disk);
 589
 590 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr,
 591                 char *buf)
 592 {
 593         return sprintf(buf, ND_DEVICE_MODALIAS_FMT "\n",
 594                         to_nd_device_type(dev));
 595 }
 596 static DEVICE_ATTR_RO(modalias);
 597
 598 static ssize_t devtype_show(struct device *dev, struct device_attribute *attr,
 599                 char *buf)
 600 {
 601         return sprintf(buf, "%s\n", dev->type->name);
 602 }
 603 static DEVICE_ATTR_RO(devtype);
 604
 605 static struct attribute *nd_device_attributes[] = {
 606         &dev_attr_modalias.attr,
 607         &dev_attr_devtype.attr,
 608         NULL,
 609 };
 610
 611 /**
 612  * nd_device_attribute_group - generic attributes for all devices on an nd bus
 613  */
 614 struct attribute_group nd_device_attribute_group = {
 615         .attrs = nd_device_attributes,
 616 };
 617 EXPORT_SYMBOL_GPL(nd_device_attribute_group);
 618
 619 static ssize_t numa_node_show(struct device *dev,
 620                 struct device_attribute *attr, char *buf)
 621 {
 622         return sprintf(buf, "%d\n", dev_to_node(dev));
 623 }
 624 static DEVICE_ATTR_RO(numa_node);
 625
 626 static struct attribute *nd_numa_attributes[] = {
 627         &dev_attr_numa_node.attr,
 628         NULL,
 629 };
 630
 631 static umode_t nd_numa_attr_visible(struct kobject *kobj, struct attribute *a,
 632                 int n)
 633 {
 634         if (!IS_ENABLED(CONFIG_NUMA))
 635                 return 0;
 636
 637         return a->mode;
 638 }
 639
 640 /**
 641  * nd_numa_attribute_group - NUMA attributes for all devices on an nd bus
 642  */
 643 struct attribute_group nd_numa_attribute_group = {
 644         .attrs = nd_numa_attributes,
 645         .is_visible = nd_numa_attr_visible,
 646 };
 647 EXPORT_SYMBOL_GPL(nd_numa_attribute_group);
 648
 649 int nvdimm_bus_create_ndctl(struct nvdimm_bus *nvdimm_bus)
 650 {
 651         dev_t devt = MKDEV(nvdimm_bus_major, nvdimm_bus->id);
 652         struct device *dev;
 653
 654         dev = device_create(nd_class, &nvdimm_bus->dev, devt, nvdimm_bus,
 655                         "ndctl%d", nvdimm_bus->id);
 656
 657         if (IS_ERR(dev))
 658                 dev_dbg(&nvdimm_bus->dev, "failed to register ndctl%d: %ld\n",
 659                                 nvdimm_bus->id, PTR_ERR(dev));
 660         return PTR_ERR_OR_ZERO(dev);
 661 }
 662
 663 void nvdimm_bus_destroy_ndctl(struct nvdimm_bus *nvdimm_bus)
 664 {
 665         device_destroy(nd_class, MKDEV(nvdimm_bus_major, nvdimm_bus->id));
 666 }
 667
 668 static const struct nd_cmd_desc __nd_cmd_dimm_descs[] = {
 669         [ND_CMD_IMPLEMENTED] = { },
 670         [ND_CMD_SMART] = {
 671                 .out_num = 2,
 672                 .out_sizes = { 4, 128, },
 673         },
 674         [ND_CMD_SMART_THRESHOLD] = {
 675                 .out_num = 2,
 676                 .out_sizes = { 4, 8, },
 677         },
 678         [ND_CMD_DIMM_FLAGS] = {
 679                 .out_num = 2,
 680                 .out_sizes = { 4, 4 },
 681         },
 682         [ND_CMD_GET_CONFIG_SIZE] = {
 683                 .out_num = 3,
 684                 .out_sizes = { 4, 4, 4, },
 685         },
 686         [ND_CMD_GET_CONFIG_DATA] = {
 687                 .in_num = 2,
 688                 .in_sizes = { 4, 4, },
 689                 .out_num = 2,
 690                 .out_sizes = { 4, UINT_MAX, },
 691         },
 692         [ND_CMD_SET_CONFIG_DATA] = {
 693                 .in_num = 3,
 694                 .in_sizes = { 4, 4, UINT_MAX, },
 695                 .out_num = 1,
 696                 .out_sizes = { 4, },
 697         },
 698         [ND_CMD_VENDOR] = {
 699                 .in_num = 3,
 700                 .in_sizes = { 4, 4, UINT_MAX, },
 701                 .out_num = 3,
 702                 .out_sizes = { 4, 4, UINT_MAX, },
 703         },
 704         [ND_CMD_CALL] = {
 705                 .in_num = 2,
 706                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 707                 .out_num = 1,
 708                 .out_sizes = { UINT_MAX, },
 709         },
 710 };
 711
 712 const struct nd_cmd_desc *nd_cmd_dimm_desc(int cmd)
 713 {
 714         if (cmd < ARRAY_SIZE(__nd_cmd_dimm_descs))
 715                 return &__nd_cmd_dimm_descs[cmd];
 716         return NULL;
 717 }
 718 EXPORT_SYMBOL_GPL(nd_cmd_dimm_desc);
 719
 720 static const struct nd_cmd_desc __nd_cmd_bus_descs[] = {
 721         [ND_CMD_IMPLEMENTED] = { },
 722         [ND_CMD_ARS_CAP] = {
 723                 .in_num = 2,
 724                 .in_sizes = { 8, 8, },
 725                 .out_num = 4,
 726                 .out_sizes = { 4, 4, 4, 4, },
 727         },
 728         [ND_CMD_ARS_START] = {
 729                 .in_num = 5,
 730                 .in_sizes = { 8, 8, 2, 1, 5, },
 731                 .out_num = 2,
 732                 .out_sizes = { 4, 4, },
 733         },
 734         [ND_CMD_ARS_STATUS] = {
 735                 .out_num = 3,
 736                 .out_sizes = { 4, 4, UINT_MAX, },
 737         },
 738         [ND_CMD_CLEAR_ERROR] = {
 739                 .in_num = 2,
 740                 .in_sizes = { 8, 8, },
 741                 .out_num = 3,
 742                 .out_sizes = { 4, 4, 8, },
 743         },
 744         [ND_CMD_CALL] = {
 745                 .in_num = 2,
 746                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
 747                 .out_num = 1,
 748                 .out_sizes = { UINT_MAX, },
 749         },
 750 };
 751
 752 const struct nd_cmd_desc *nd_cmd_bus_desc(int cmd)
 753 {
 754         if (cmd < ARRAY_SIZE(__nd_cmd_bus_descs))
 755                 return &__nd_cmd_bus_descs[cmd];
 756         return NULL;
 757 }
 758 EXPORT_SYMBOL_GPL(nd_cmd_bus_desc);
 759
 760 u32 nd_cmd_in_size(struct nvdimm *nvdimm, int cmd,
 761                 const struct nd_cmd_desc *desc, int idx, void *buf)
 762 {
 763         if (idx >= desc->in_num)
 764                 return UINT_MAX;
 765
 766         if (desc->in_sizes[idx] < UINT_MAX)
 767                 return desc->in_sizes[idx];
 768
 769         if (nvdimm && cmd == ND_CMD_SET_CONFIG_DATA && idx == 2) {
 770                 struct nd_cmd_set_config_hdr *hdr = buf;
 771
 772                 return hdr->in_length;
 773         } else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) {
 774                 struct nd_cmd_vendor_hdr *hdr = buf;
 775
 776                 return hdr->in_length;
 777         } else if (cmd == ND_CMD_CALL) {
 778                 struct nd_cmd_pkg *pkg = buf;
 779
 780                 return pkg->nd_size_in;
 781         }
 782
 783         return UINT_MAX;
 784 }
 785 EXPORT_SYMBOL_GPL(nd_cmd_in_size);
 786
 787 u32 nd_cmd_out_size(struct nvdimm *nvdimm, int cmd,
 788                 const struct nd_cmd_desc *desc, int idx, const u32 *in_field,
 789                 const u32 *out_field, unsigned long remainder)
 790 {
 791         if (idx >= desc->out_num)
 792                 return UINT_MAX;
 793
 794         if (desc->out_sizes[idx] < UINT_MAX)
 795                 return desc->out_sizes[idx];
 796
 797         if (nvdimm && cmd == ND_CMD_GET_CONFIG_DATA && idx == 1)
 798                 return in_field[1];
 799         else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2)
 800                 return out_field[1];
 801         else if (!nvdimm && cmd == ND_CMD_ARS_STATUS && idx == 2) {
 802                 /*
 803                  * Per table 9-276 ARS Data in ACPI 6.1, out_field[1] is
 804                  * "Size of Output Buffer in bytes, including this
 805                  * field."
 806                  */
 807                 if (out_field[1] < 4)
 808                         return 0;
 809                 /*
 810                  * ACPI 6.1 is ambiguous if 'status' is included in the
 811                  * output size. If we encounter an output size that
 812                  * overshoots the remainder by 4 bytes, assume it was
 813                  * including 'status'.
 814                  */
 815                 if (out_field[1] - 8 == remainder)
 816                         return remainder;
 817                 return out_field[1] - 4;
 818         } else if (cmd == ND_CMD_CALL) {
 819                 struct nd_cmd_pkg *pkg = (struct nd_cmd_pkg *) in_field;
 820
 821                 return pkg->nd_size_out;
 822         }
 823
 824
 825         return UINT_MAX;
 826 }
 827 EXPORT_SYMBOL_GPL(nd_cmd_out_size);
 828
 829 void wait_nvdimm_bus_probe_idle(struct device *dev)
 830 {
 831         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
 832
 833         do {
 834                 if (nvdimm_bus->probe_active == 0)
 835                         break;
 836                 nvdimm_bus_unlock(&nvdimm_bus->dev);
 837                 wait_event(nvdimm_bus->probe_wait,
 838                                 nvdimm_bus->probe_active == 0);
 839                 nvdimm_bus_lock(&nvdimm_bus->dev);
 840         } while (true);
 841 }
 842
 843 static int nd_pmem_forget_poison_check(struct device *dev, void *data)
 844 {
 845         struct nd_cmd_clear_error *clear_err =
 846                 (struct nd_cmd_clear_error *)data;
 847         struct nd_btt *nd_btt = is_nd_btt(dev) ? to_nd_btt(dev) : NULL;
 848         struct nd_pfn *nd_pfn = is_nd_pfn(dev) ? to_nd_pfn(dev) : NULL;
 849         struct nd_dax *nd_dax = is_nd_dax(dev) ? to_nd_dax(dev) : NULL;
 850         struct nd_namespace_common *ndns = NULL;
 851         struct nd_namespace_io *nsio;
 852         resource_size_t offset = 0, end_trunc = 0, start, end, pstart, pend;
 853
 854         if (nd_dax || !dev->driver)
 855                 return 0;
 856
 857         start = clear_err->address;
 858         end = clear_err->address + clear_err->cleared - 1;
 859
 860         if (nd_btt || nd_pfn || nd_dax) {
 861                 if (nd_btt)
 862                         ndns = nd_btt->ndns;
 863                 else if (nd_pfn)
 864                         ndns = nd_pfn->ndns;
 865                 else if (nd_dax)
 866                         ndns = nd_dax->nd_pfn.ndns;
 867
 868                 if (!ndns)
 869                         return 0;
 870         } else
 871                 ndns = to_ndns(dev);
 872
 873         nsio = to_nd_namespace_io(&ndns->dev);
 874         pstart = nsio->res.start + offset;
 875         pend = nsio->res.end - end_trunc;
 876
 877         if ((pstart >= start) && (pend <= end))
 878                 return -EBUSY;
 879
 880         return 0;
 881
 882 }
 883
 884 static int nd_ns_forget_poison_check(struct device *dev, void *data)
 885 {
 886         return device_for_each_child(dev, data, nd_pmem_forget_poison_check);
 887 }
 888
 889 /* set_config requires an idle interleave set */
 890 static int nd_cmd_clear_to_send(struct nvdimm_bus *nvdimm_bus,
 891                 struct nvdimm *nvdimm, unsigned int cmd, void *data)
 892 {
 893         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 894
 895         /* ask the bus provider if it would like to block this request */
 896         if (nd_desc->clear_to_send) {
 897                 int rc = nd_desc->clear_to_send(nd_desc, nvdimm, cmd);
 898
 899                 if (rc)
 900                         return rc;
 901         }
 902
 903         /* require clear error to go through the pmem driver */
 904         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR)
 905                 return device_for_each_child(&nvdimm_bus->dev, data,
 906                                 nd_ns_forget_poison_check);
 907
 908         if (!nvdimm || cmd != ND_CMD_SET_CONFIG_DATA)
 909                 return 0;
 910
 911         /* prevent label manipulation while the kernel owns label updates */
 912         wait_nvdimm_bus_probe_idle(&nvdimm_bus->dev);
 913         if (atomic_read(&nvdimm->busy))
 914                 return -EBUSY;
 915         return 0;
 916 }
 917
 918 static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm,
 919                 int read_only, unsigned int ioctl_cmd, unsigned long arg)
 920 {
 921         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
 922         static char out_env[ND_CMD_MAX_ENVELOPE];
 923         static char in_env[ND_CMD_MAX_ENVELOPE];
 924         const struct nd_cmd_desc *desc = NULL;
 925         unsigned int cmd = _IOC_NR(ioctl_cmd);
 926         struct device *dev = &nvdimm_bus->dev;
 927         void __user *p = (void __user *) arg;
 928         const char *cmd_name, *dimm_name;
 929         u32 in_len = 0, out_len = 0;
 930         unsigned int func = cmd;
 931         unsigned long cmd_mask;
 932         struct nd_cmd_pkg pkg;
 933         int rc, i, cmd_rc;
 934         u64 buf_len = 0;
 935         void *buf;
 936
 937         if (nvdimm) {
 938                 desc = nd_cmd_dimm_desc(cmd);
 939                 cmd_name = nvdimm_cmd_name(cmd);
 940                 cmd_mask = nvdimm->cmd_mask;
 941                 dimm_name = dev_name(&nvdimm->dev);
 942         } else {
 943                 desc = nd_cmd_bus_desc(cmd);
 944                 cmd_name = nvdimm_bus_cmd_name(cmd);
 945                 cmd_mask = nd_desc->cmd_mask;
 946                 dimm_name = "bus";
 947         }
 948
 949         if (cmd == ND_CMD_CALL) {
 950                 if (copy_from_user(&pkg, p, sizeof(pkg)))
 951                         return -EFAULT;
 952         }
 953
 954         if (!desc || (desc->out_num + desc->in_num == 0) ||
 955                         !test_bit(cmd, &cmd_mask))
 956                 return -ENOTTY;
 957
 958         /* fail write commands (when read-only) */
 959         if (read_only)
 960                 switch (cmd) {
 961                 case ND_CMD_VENDOR:
 962                 case ND_CMD_SET_CONFIG_DATA:
 963                 case ND_CMD_ARS_START:
 964                 case ND_CMD_CLEAR_ERROR:
 965                 case ND_CMD_CALL:
 966                         dev_dbg(&nvdimm_bus->dev, "'%s' command while read-only.\n",
 967                                         nvdimm ? nvdimm_cmd_name(cmd)
 968                                         : nvdimm_bus_cmd_name(cmd));
 969                         return -EPERM;
 970                 default:
 971                         break;
 972                 }
 973
 974         /* process an input envelope */
 975         for (i = 0; i < desc->in_num; i++) {
 976                 u32 in_size, copy;
 977
 978                 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env);
 979                 if (in_size == UINT_MAX) {
 980                         dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n",
 981                                         __func__, dimm_name, cmd_name, i);
 982                         return -ENXIO;
 983                 }
 984                 if (in_len < sizeof(in_env))
 985                         copy = min_t(u32, sizeof(in_env) - in_len, in_size);
 986                 else
 987                         copy = 0;
 988                 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy))
 989                         return -EFAULT;
 990                 in_len += in_size;
 991         }
 992
 993         if (cmd == ND_CMD_CALL) {
 994                 func = pkg.nd_command;
 995                 dev_dbg(dev, "%s, idx: %llu, in: %u, out: %u, len %llu\n",
 996                                 dimm_name, pkg.nd_command,
 997                                 in_len, out_len, buf_len);
 998         }
 999
1000         /* process an output envelope */
1001         for (i = 0; i < desc->out_num; i++) {
1002                 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i,
1003                                 (u32 *) in_env, (u32 *) out_env, 0);
1004                 u32 copy;
1005
1006                 if (out_size == UINT_MAX) {
1007                         dev_dbg(dev, "%s unknown output size cmd: %s field: %d\n",
1008                                         dimm_name, cmd_name, i);
1009                         return -EFAULT;
1010                 }
1011                 if (out_len < sizeof(out_env))
1012                         copy = min_t(u32, sizeof(out_env) - out_len, out_size);
1013                 else
1014                         copy = 0;
1015                 if (copy && copy_from_user(&out_env[out_len],
1016                                         p + in_len + out_len, copy))
1017                         return -EFAULT;
1018                 out_len += out_size;
1019         }
1020
1021         buf_len = (u64) out_len + (u64) in_len;
1022         if (buf_len > ND_IOCTL_MAX_BUFLEN) {
1023                 dev_dbg(dev, "%s cmd: %s buf_len: %llu > %d\n", dimm_name,
1024                                 cmd_name, buf_len, ND_IOCTL_MAX_BUFLEN);
1025                 return -EINVAL;
1026         }
1027
1028         buf = vmalloc(buf_len);
1029         if (!buf)
1030                 return -ENOMEM;
1031
1032         if (copy_from_user(buf, p, buf_len)) {
1033                 rc = -EFAULT;
1034                 goto out;
1035         }
1036
1037         nvdimm_bus_lock(&nvdimm_bus->dev);
1038         rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, func, buf);
1039         if (rc)
1040                 goto out_unlock;
1041
1042         rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc);
1043         if (rc < 0)
1044                 goto out_unlock;
1045
1046         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) {
1047                 struct nd_cmd_clear_error *clear_err = buf;
1048
1049                 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address,
1050                                 clear_err->cleared);
1051         }
1052         nvdimm_bus_unlock(&nvdimm_bus->dev);
1053
1054         if (copy_to_user(p, buf, buf_len))
1055                 rc = -EFAULT;
1056
1057         vfree(buf);
1058         return rc;
1059
1060  out_unlock:
1061         nvdimm_bus_unlock(&nvdimm_bus->dev);
1062  out:
1063         vfree(buf);
1064         return rc;
1065 }
1066
1067 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1068 {
1069         long id = (long) file->private_data;
1070         int rc = -ENXIO, ro;
1071         struct nvdimm_bus *nvdimm_bus;
1072
1073         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1074         mutex_lock(&nvdimm_bus_list_mutex);
1075         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1076                 if (nvdimm_bus->id == id) {
1077                         rc = __nd_ioctl(nvdimm_bus, NULL, ro, cmd, arg);
1078                         break;
1079                 }
1080         }
1081         mutex_unlock(&nvdimm_bus_list_mutex);
1082
1083         return rc;
1084 }
1085
1086 static int match_dimm(struct device *dev, void *data)
1087 {
1088         long id = (long) data;
1089
1090         if (is_nvdimm(dev)) {
1091                 struct nvdimm *nvdimm = to_nvdimm(dev);
1092
1093                 return nvdimm->id == id;
1094         }
1095
1096         return 0;
1097 }
1098
1099 static long nvdimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1100 {
1101         int rc = -ENXIO, ro;
1102         struct nvdimm_bus *nvdimm_bus;
1103
1104         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1105         mutex_lock(&nvdimm_bus_list_mutex);
1106         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1107                 struct device *dev = device_find_child(&nvdimm_bus->dev,
1108                                 file->private_data, match_dimm);
1109                 struct nvdimm *nvdimm;
1110
1111                 if (!dev)
1112                         continue;
1113
1114                 nvdimm = to_nvdimm(dev);
1115                 rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg);
1116                 put_device(dev);
1117                 break;
1118         }
1119         mutex_unlock(&nvdimm_bus_list_mutex);
1120
1121         return rc;
1122 }
1123
1124 static int nd_open(struct inode *inode, struct file *file)
1125 {
1126         long minor = iminor(inode);
1127
1128         file->private_data = (void *) minor;
1129         return 0;
1130 }
1131
1132 static const struct file_operations nvdimm_bus_fops = {
1133         .owner = THIS_MODULE,
1134         .open = nd_open,
1135         .unlocked_ioctl = nd_ioctl,
1136         .compat_ioctl = nd_ioctl,
1137         .llseek = noop_llseek,
1138 };
1139
1140 static const struct file_operations nvdimm_fops = {
1141         .owner = THIS_MODULE,
1142         .open = nd_open,
1143         .unlocked_ioctl = nvdimm_ioctl,
1144         .compat_ioctl = nvdimm_ioctl,
1145         .llseek = noop_llseek,
1146 };
1147
1148 int __init nvdimm_bus_init(void)
1149 {
1150         int rc;
1151
1152         rc = bus_register(&nvdimm_bus_type);
1153         if (rc)
1154                 return rc;
1155
1156         rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops);
1157         if (rc < 0)
1158                 goto err_bus_chrdev;
1159         nvdimm_bus_major = rc;
1160
1161         rc = register_chrdev(0, "dimmctl", &nvdimm_fops);
1162         if (rc < 0)
1163                 goto err_dimm_chrdev;
1164         nvdimm_major = rc;
1165
1166         nd_class = class_create(THIS_MODULE, "nd");
1167         if (IS_ERR(nd_class)) {
1168                 rc = PTR_ERR(nd_class);
1169                 goto err_class;
1170         }
1171
1172         rc = driver_register(&nd_bus_driver.drv);
1173         if (rc)
1174                 goto err_nd_bus;
1175
1176         return 0;
1177
1178  err_nd_bus:
1179         class_destroy(nd_class);
1180  err_class:
1181         unregister_chrdev(nvdimm_major, "dimmctl");
1182  err_dimm_chrdev:
1183         unregister_chrdev(nvdimm_bus_major, "ndctl");
1184  err_bus_chrdev:
1185         bus_unregister(&nvdimm_bus_type);
1186
1187         return rc;
1188 }
1189
1190 void nvdimm_bus_exit(void)
1191 {
1192         driver_unregister(&nd_bus_driver.drv);
1193         class_destroy(nd_class);
1194         unregister_chrdev(nvdimm_bus_major, "ndctl");
1195         unregister_chrdev(nvdimm_major, "dimmctl");
1196         bus_unregister(&nvdimm_bus_type);
1197         ida_destroy(&nd_ida);
1198 }