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         if (dev->parent)
 488                 put_device(dev->parent);
 489 }
 490
 491 static void nd_async_device_unregister(void *d, async_cookie_t cookie)
 492 {
 493         struct device *dev = d;
 494
 495         /* flush bus operations before delete */
 496         nvdimm_bus_lock(dev);
 497         nvdimm_bus_unlock(dev);
 498
 499         device_unregister(dev);
 500         put_device(dev);
 501 }
 502
 503 void __nd_device_register(struct device *dev)
 504 {
 505         if (!dev)
 506                 return;
 507         dev->bus = &nvdimm_bus_type;
 508         if (dev->parent)
 509                 get_device(dev->parent);
 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] - 4 == remainder)
 816                         return remainder;
 817                 return out_field[1] - 8;
 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 ||
 955             (desc->out_num + desc->in_num == 0) ||
 956             cmd > ND_CMD_CALL ||
 957             !test_bit(cmd, &cmd_mask))
 958                 return -ENOTTY;
 959
 960         /* fail write commands (when read-only) */
 961         if (read_only)
 962                 switch (cmd) {
 963                 case ND_CMD_VENDOR:
 964                 case ND_CMD_SET_CONFIG_DATA:
 965                 case ND_CMD_ARS_START:
 966                 case ND_CMD_CLEAR_ERROR:
 967                 case ND_CMD_CALL:
 968                         dev_dbg(&nvdimm_bus->dev, "'%s' command while read-only.\n",
 969                                         nvdimm ? nvdimm_cmd_name(cmd)
 970                                         : nvdimm_bus_cmd_name(cmd));
 971                         return -EPERM;
 972                 default:
 973                         break;
 974                 }
 975
 976         /* process an input envelope */
 977         for (i = 0; i < desc->in_num; i++) {
 978                 u32 in_size, copy;
 979
 980                 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env);
 981                 if (in_size == UINT_MAX) {
 982                         dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n",
 983                                         __func__, dimm_name, cmd_name, i);
 984                         return -ENXIO;
 985                 }
 986                 if (in_len < sizeof(in_env))
 987                         copy = min_t(u32, sizeof(in_env) - in_len, in_size);
 988                 else
 989                         copy = 0;
 990                 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy))
 991                         return -EFAULT;
 992                 in_len += in_size;
 993         }
 994
 995         if (cmd == ND_CMD_CALL) {
 996                 func = pkg.nd_command;
 997                 dev_dbg(dev, "%s:%s, idx: %llu, in: %u, out: %u, len %llu\n",
 998                                 __func__, dimm_name, pkg.nd_command,
 999                                 in_len, out_len, buf_len);
1000         }
1001
1002         /* process an output envelope */
1003         for (i = 0; i < desc->out_num; i++) {
1004                 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i,
1005                                 (u32 *) in_env, (u32 *) out_env, 0);
1006                 u32 copy;
1007
1008                 if (out_size == UINT_MAX) {
1009                         dev_dbg(dev, "%s:%s unknown output size cmd: %s field: %d\n",
1010                                         __func__, dimm_name, cmd_name, i);
1011                         return -EFAULT;
1012                 }
1013                 if (out_len < sizeof(out_env))
1014                         copy = min_t(u32, sizeof(out_env) - out_len, out_size);
1015                 else
1016                         copy = 0;
1017                 if (copy && copy_from_user(&out_env[out_len],
1018                                         p + in_len + out_len, copy))
1019                         return -EFAULT;
1020                 out_len += out_size;
1021         }
1022
1023         buf_len = (u64) out_len + (u64) in_len;
1024         if (buf_len > ND_IOCTL_MAX_BUFLEN) {
1025                 dev_dbg(dev, "%s:%s cmd: %s buf_len: %llu > %d\n", __func__,
1026                                 dimm_name, cmd_name, buf_len,
1027                                 ND_IOCTL_MAX_BUFLEN);
1028                 return -EINVAL;
1029         }
1030
1031         buf = vmalloc(buf_len);
1032         if (!buf)
1033                 return -ENOMEM;
1034
1035         if (copy_from_user(buf, p, buf_len)) {
1036                 rc = -EFAULT;
1037                 goto out;
1038         }
1039
1040         nvdimm_bus_lock(&nvdimm_bus->dev);
1041         rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, func, buf);
1042         if (rc)
1043                 goto out_unlock;
1044
1045         rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc);
1046         if (rc < 0)
1047                 goto out_unlock;
1048
1049         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) {
1050                 struct nd_cmd_clear_error *clear_err = buf;
1051
1052                 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address,
1053                                 clear_err->cleared);
1054         }
1055         nvdimm_bus_unlock(&nvdimm_bus->dev);
1056
1057         if (copy_to_user(p, buf, buf_len))
1058                 rc = -EFAULT;
1059
1060         vfree(buf);
1061         return rc;
1062
1063  out_unlock:
1064         nvdimm_bus_unlock(&nvdimm_bus->dev);
1065  out:
1066         vfree(buf);
1067         return rc;
1068 }
1069
1070 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1071 {
1072         long id = (long) file->private_data;
1073         int rc = -ENXIO, ro;
1074         struct nvdimm_bus *nvdimm_bus;
1075
1076         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1077         mutex_lock(&nvdimm_bus_list_mutex);
1078         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1079                 if (nvdimm_bus->id == id) {
1080                         rc = __nd_ioctl(nvdimm_bus, NULL, ro, cmd, arg);
1081                         break;
1082                 }
1083         }
1084         mutex_unlock(&nvdimm_bus_list_mutex);
1085
1086         return rc;
1087 }
1088
1089 static int match_dimm(struct device *dev, void *data)
1090 {
1091         long id = (long) data;
1092
1093         if (is_nvdimm(dev)) {
1094                 struct nvdimm *nvdimm = to_nvdimm(dev);
1095
1096                 return nvdimm->id == id;
1097         }
1098
1099         return 0;
1100 }
1101
1102 static long nvdimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1103 {
1104         int rc = -ENXIO, ro;
1105         struct nvdimm_bus *nvdimm_bus;
1106
1107         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1108         mutex_lock(&nvdimm_bus_list_mutex);
1109         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1110                 struct device *dev = device_find_child(&nvdimm_bus->dev,
1111                                 file->private_data, match_dimm);
1112                 struct nvdimm *nvdimm;
1113
1114                 if (!dev)
1115                         continue;
1116
1117                 nvdimm = to_nvdimm(dev);
1118                 rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg);
1119                 put_device(dev);
1120                 break;
1121         }
1122         mutex_unlock(&nvdimm_bus_list_mutex);
1123
1124         return rc;
1125 }
1126
1127 static int nd_open(struct inode *inode, struct file *file)
1128 {
1129         long minor = iminor(inode);
1130
1131         file->private_data = (void *) minor;
1132         return 0;
1133 }
1134
1135 static const struct file_operations nvdimm_bus_fops = {
1136         .owner = THIS_MODULE,
1137         .open = nd_open,
1138         .unlocked_ioctl = nd_ioctl,
1139         .compat_ioctl = nd_ioctl,
1140         .llseek = noop_llseek,
1141 };
1142
1143 static const struct file_operations nvdimm_fops = {
1144         .owner = THIS_MODULE,
1145         .open = nd_open,
1146         .unlocked_ioctl = nvdimm_ioctl,
1147         .compat_ioctl = nvdimm_ioctl,
1148         .llseek = noop_llseek,
1149 };
1150
1151 int __init nvdimm_bus_init(void)
1152 {
1153         int rc;
1154
1155         BUILD_BUG_ON(sizeof(struct nd_smart_payload) != 128);
1156         BUILD_BUG_ON(sizeof(struct nd_smart_threshold_payload) != 8);
1157
1158         rc = bus_register(&nvdimm_bus_type);
1159         if (rc)
1160                 return rc;
1161
1162         rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops);
1163         if (rc < 0)
1164                 goto err_bus_chrdev;
1165         nvdimm_bus_major = rc;
1166
1167         rc = register_chrdev(0, "dimmctl", &nvdimm_fops);
1168         if (rc < 0)
1169                 goto err_dimm_chrdev;
1170         nvdimm_major = rc;
1171
1172         nd_class = class_create(THIS_MODULE, "nd");
1173         if (IS_ERR(nd_class)) {
1174                 rc = PTR_ERR(nd_class);
1175                 goto err_class;
1176         }
1177
1178         rc = driver_register(&nd_bus_driver.drv);
1179         if (rc)
1180                 goto err_nd_bus;
1181
1182         return 0;
1183
1184  err_nd_bus:
1185         class_destroy(nd_class);
1186  err_class:
1187         unregister_chrdev(nvdimm_major, "dimmctl");
1188  err_dimm_chrdev:
1189         unregister_chrdev(nvdimm_bus_major, "ndctl");
1190  err_bus_chrdev:
1191         bus_unregister(&nvdimm_bus_type);
1192
1193         return rc;
1194 }
1195
1196 void nvdimm_bus_exit(void)
1197 {
1198         driver_unregister(&nd_bus_driver.drv);
1199         class_destroy(nd_class);
1200         unregister_chrdev(nvdimm_bus_major, "ndctl");
1201         unregister_chrdev(nvdimm_major, "dimmctl");
1202         bus_unregister(&nvdimm_bus_type);
1203         ida_destroy(&nd_ida);
1204 }