drivers/dax/bus.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright(c) 2017-2018 Intel Corporation. All rights reserved. */
   3 #include <linux/memremap.h>
   4 #include <linux/device.h>
   5 #include <linux/mutex.h>
   6 #include <linux/list.h>
   7 #include <linux/slab.h>
   8 #include <linux/dax.h>
   9 #include <linux/io.h>
  10 #include "dax-private.h"
  11 #include "bus.h"
  12
  13 static struct class *dax_class;
  14
  15 static DEFINE_MUTEX(dax_bus_lock);
  16
  17 #define DAX_NAME_LEN 30
  18 struct dax_id {
  19         struct list_head list;
  20         char dev_name[DAX_NAME_LEN];
  21 };
  22
  23 static int dax_bus_uevent(struct device *dev, struct kobj_uevent_env *env)
  24 {
  25         /*
  26          * We only ever expect to handle device-dax instances, i.e. the
  27          * @type argument to MODULE_ALIAS_DAX_DEVICE() is always zero
  28          */
  29         return add_uevent_var(env, "MODALIAS=" DAX_DEVICE_MODALIAS_FMT, 0);
  30 }
  31
  32 static struct dax_device_driver *to_dax_drv(struct device_driver *drv)
  33 {
  34         return container_of(drv, struct dax_device_driver, drv);
  35 }
  36
  37 static struct dax_id *__dax_match_id(struct dax_device_driver *dax_drv,
  38                 const char *dev_name)
  39 {
  40         struct dax_id *dax_id;
  41
  42         lockdep_assert_held(&dax_bus_lock);
  43
  44         list_for_each_entry(dax_id, &dax_drv->ids, list)
  45                 if (sysfs_streq(dax_id->dev_name, dev_name))
  46                         return dax_id;
  47         return NULL;
  48 }
  49
  50 static int dax_match_id(struct dax_device_driver *dax_drv, struct device *dev)
  51 {
  52         int match;
  53
  54         mutex_lock(&dax_bus_lock);
  55         match = !!__dax_match_id(dax_drv, dev_name(dev));
  56         mutex_unlock(&dax_bus_lock);
  57
  58         return match;
  59 }
  60
  61 enum id_action {
  62         ID_REMOVE,
  63         ID_ADD,
  64 };
  65
  66 static ssize_t do_id_store(struct device_driver *drv, const char *buf,
  67                 size_t count, enum id_action action)
  68 {
  69         struct dax_device_driver *dax_drv = to_dax_drv(drv);
  70         unsigned int region_id, id;
  71         char devname[DAX_NAME_LEN];
  72         struct dax_id *dax_id;
  73         ssize_t rc = count;
  74         int fields;
  75
  76         fields = sscanf(buf, "dax%d.%d", &region_id, &id);
  77         if (fields != 2)
  78                 return -EINVAL;
  79         sprintf(devname, "dax%d.%d", region_id, id);
  80         if (!sysfs_streq(buf, devname))
  81                 return -EINVAL;
  82
  83         mutex_lock(&dax_bus_lock);
  84         dax_id = __dax_match_id(dax_drv, buf);
  85         if (!dax_id) {
  86                 if (action == ID_ADD) {
  87                         dax_id = kzalloc(sizeof(*dax_id), GFP_KERNEL);
  88                         if (dax_id) {
  89                                 strncpy(dax_id->dev_name, buf, DAX_NAME_LEN);
  90                                 list_add(&dax_id->list, &dax_drv->ids);
  91                         } else
  92                                 rc = -ENOMEM;
  93                 } else
  94                         /* nothing to remove */;
  95         } else if (action == ID_REMOVE) {
  96                 list_del(&dax_id->list);
  97                 kfree(dax_id);
  98         } else
  99                 /* dax_id already added */;
 100         mutex_unlock(&dax_bus_lock);
 101
 102         if (rc < 0)
 103                 return rc;
 104         if (action == ID_ADD)
 105                 rc = driver_attach(drv);
 106         if (rc)
 107                 return rc;
 108         return count;
 109 }
 110
 111 static ssize_t new_id_store(struct device_driver *drv, const char *buf,
 112                 size_t count)
 113 {
 114         return do_id_store(drv, buf, count, ID_ADD);
 115 }
 116 static DRIVER_ATTR_WO(new_id);
 117
 118 static ssize_t remove_id_store(struct device_driver *drv, const char *buf,
 119                 size_t count)
 120 {
 121         return do_id_store(drv, buf, count, ID_REMOVE);
 122 }
 123 static DRIVER_ATTR_WO(remove_id);
 124
 125 static struct attribute *dax_drv_attrs[] = {
 126         &driver_attr_new_id.attr,
 127         &driver_attr_remove_id.attr,
 128         NULL,
 129 };
 130 ATTRIBUTE_GROUPS(dax_drv);
 131
 132 static int dax_bus_match(struct device *dev, struct device_driver *drv);
 133
 134 static bool is_static(struct dax_region *dax_region)
 135 {
 136         return (dax_region->res.flags & IORESOURCE_DAX_STATIC) != 0;
 137 }
 138
 139 static u64 dev_dax_size(struct dev_dax *dev_dax)
 140 {
 141         u64 size = 0;
 142         int i;
 143
 144         device_lock_assert(&dev_dax->dev);
 145
 146         for (i = 0; i < dev_dax->nr_range; i++)
 147                 size += range_len(&dev_dax->ranges[i].range);
 148
 149         return size;
 150 }
 151
 152 static int dax_bus_probe(struct device *dev)
 153 {
 154         struct dax_device_driver *dax_drv = to_dax_drv(dev->driver);
 155         struct dev_dax *dev_dax = to_dev_dax(dev);
 156         struct dax_region *dax_region = dev_dax->region;
 157         int rc;
 158
 159         if (dev_dax_size(dev_dax) == 0 || dev_dax->id < 0)
 160                 return -ENXIO;
 161
 162         rc = dax_drv->probe(dev_dax);
 163
 164         if (rc || is_static(dax_region))
 165                 return rc;
 166
 167         /*
 168          * Track new seed creation only after successful probe of the
 169          * previous seed.
 170          */
 171         if (dax_region->seed == dev)
 172                 dax_region->seed = NULL;
 173
 174         return 0;
 175 }
 176
 177 static int dax_bus_remove(struct device *dev)
 178 {
 179         struct dax_device_driver *dax_drv = to_dax_drv(dev->driver);
 180         struct dev_dax *dev_dax = to_dev_dax(dev);
 181
 182         return dax_drv->remove(dev_dax);
 183 }
 184
 185 static struct bus_type dax_bus_type = {
 186         .name = "dax",
 187         .uevent = dax_bus_uevent,
 188         .match = dax_bus_match,
 189         .probe = dax_bus_probe,
 190         .remove = dax_bus_remove,
 191         .drv_groups = dax_drv_groups,
 192 };
 193
 194 static int dax_bus_match(struct device *dev, struct device_driver *drv)
 195 {
 196         struct dax_device_driver *dax_drv = to_dax_drv(drv);
 197
 198         /*
 199          * All but the 'device-dax' driver, which has 'match_always'
 200          * set, requires an exact id match.
 201          */
 202         if (dax_drv->match_always)
 203                 return 1;
 204
 205         return dax_match_id(dax_drv, dev);
 206 }
 207
 208 /*
 209  * Rely on the fact that drvdata is set before the attributes are
 210  * registered, and that the attributes are unregistered before drvdata
 211  * is cleared to assume that drvdata is always valid.
 212  */
 213 static ssize_t id_show(struct device *dev,
 214                 struct device_attribute *attr, char *buf)
 215 {
 216         struct dax_region *dax_region = dev_get_drvdata(dev);
 217
 218         return sprintf(buf, "%d\n", dax_region->id);
 219 }
 220 static DEVICE_ATTR_RO(id);
 221
 222 static ssize_t region_size_show(struct device *dev,
 223                 struct device_attribute *attr, char *buf)
 224 {
 225         struct dax_region *dax_region = dev_get_drvdata(dev);
 226
 227         return sprintf(buf, "%llu\n", (unsigned long long)
 228                         resource_size(&dax_region->res));
 229 }
 230 static struct device_attribute dev_attr_region_size = __ATTR(size, 0444,
 231                 region_size_show, NULL);
 232
 233 static ssize_t align_show(struct device *dev,
 234                 struct device_attribute *attr, char *buf)
 235 {
 236         struct dax_region *dax_region = dev_get_drvdata(dev);
 237
 238         return sprintf(buf, "%u\n", dax_region->align);
 239 }
 240 static DEVICE_ATTR_RO(align);
 241
 242 #define for_each_dax_region_resource(dax_region, res) \
 243         for (res = (dax_region)->res.child; res; res = res->sibling)
 244
 245 static unsigned long long dax_region_avail_size(struct dax_region *dax_region)
 246 {
 247         resource_size_t size = resource_size(&dax_region->res);
 248         struct resource *res;
 249
 250         device_lock_assert(dax_region->dev);
 251
 252         for_each_dax_region_resource(dax_region, res)
 253                 size -= resource_size(res);
 254         return size;
 255 }
 256
 257 static ssize_t available_size_show(struct device *dev,
 258                 struct device_attribute *attr, char *buf)
 259 {
 260         struct dax_region *dax_region = dev_get_drvdata(dev);
 261         unsigned long long size;
 262
 263         device_lock(dev);
 264         size = dax_region_avail_size(dax_region);
 265         device_unlock(dev);
 266
 267         return sprintf(buf, "%llu\n", size);
 268 }
 269 static DEVICE_ATTR_RO(available_size);
 270
 271 static ssize_t seed_show(struct device *dev,
 272                 struct device_attribute *attr, char *buf)
 273 {
 274         struct dax_region *dax_region = dev_get_drvdata(dev);
 275         struct device *seed;
 276         ssize_t rc;
 277
 278         if (is_static(dax_region))
 279                 return -EINVAL;
 280
 281         device_lock(dev);
 282         seed = dax_region->seed;
 283         rc = sprintf(buf, "%s\n", seed ? dev_name(seed) : "");
 284         device_unlock(dev);
 285
 286         return rc;
 287 }
 288 static DEVICE_ATTR_RO(seed);
 289
 290 static ssize_t create_show(struct device *dev,
 291                 struct device_attribute *attr, char *buf)
 292 {
 293         struct dax_region *dax_region = dev_get_drvdata(dev);
 294         struct device *youngest;
 295         ssize_t rc;
 296
 297         if (is_static(dax_region))
 298                 return -EINVAL;
 299
 300         device_lock(dev);
 301         youngest = dax_region->youngest;
 302         rc = sprintf(buf, "%s\n", youngest ? dev_name(youngest) : "");
 303         device_unlock(dev);
 304
 305         return rc;
 306 }
 307
 308 static ssize_t create_store(struct device *dev, struct device_attribute *attr,
 309                 const char *buf, size_t len)
 310 {
 311         struct dax_region *dax_region = dev_get_drvdata(dev);
 312         unsigned long long avail;
 313         ssize_t rc;
 314         int val;
 315
 316         if (is_static(dax_region))
 317                 return -EINVAL;
 318
 319         rc = kstrtoint(buf, 0, &val);
 320         if (rc)
 321                 return rc;
 322         if (val != 1)
 323                 return -EINVAL;
 324
 325         device_lock(dev);
 326         avail = dax_region_avail_size(dax_region);
 327         if (avail == 0)
 328                 rc = -ENOSPC;
 329         else {
 330                 struct dev_dax_data data = {
 331                         .dax_region = dax_region,
 332                         .size = 0,
 333                         .id = -1,
 334                 };
 335                 struct dev_dax *dev_dax = devm_create_dev_dax(&data);
 336
 337                 if (IS_ERR(dev_dax))
 338                         rc = PTR_ERR(dev_dax);
 339                 else {
 340                         /*
 341                          * In support of crafting multiple new devices
 342                          * simultaneously multiple seeds can be created,
 343                          * but only the first one that has not been
 344                          * successfully bound is tracked as the region
 345                          * seed.
 346                          */
 347                         if (!dax_region->seed)
 348                                 dax_region->seed = &dev_dax->dev;
 349                         dax_region->youngest = &dev_dax->dev;
 350                         rc = len;
 351                 }
 352         }
 353         device_unlock(dev);
 354
 355         return rc;
 356 }
 357 static DEVICE_ATTR_RW(create);
 358
 359 void kill_dev_dax(struct dev_dax *dev_dax)
 360 {
 361         struct dax_device *dax_dev = dev_dax->dax_dev;
 362         struct inode *inode = dax_inode(dax_dev);
 363
 364         kill_dax(dax_dev);
 365         unmap_mapping_range(inode->i_mapping, 0, 0, 1);
 366 }
 367 EXPORT_SYMBOL_GPL(kill_dev_dax);
 368
 369 static void free_dev_dax_ranges(struct dev_dax *dev_dax)
 370 {
 371         struct dax_region *dax_region = dev_dax->region;
 372         int i;
 373
 374         device_lock_assert(dax_region->dev);
 375         for (i = 0; i < dev_dax->nr_range; i++) {
 376                 struct range *range = &dev_dax->ranges[i].range;
 377
 378                 __release_region(&dax_region->res, range->start,
 379                                 range_len(range));
 380         }
 381         dev_dax->nr_range = 0;
 382 }
 383
 384 static void unregister_dev_dax(void *dev)
 385 {
 386         struct dev_dax *dev_dax = to_dev_dax(dev);
 387
 388         dev_dbg(dev, "%s\n", __func__);
 389
 390         kill_dev_dax(dev_dax);
 391         free_dev_dax_ranges(dev_dax);
 392         device_del(dev);
 393         put_device(dev);
 394 }
 395
 396 /* a return value >= 0 indicates this invocation invalidated the id */
 397 static int __free_dev_dax_id(struct dev_dax *dev_dax)
 398 {
 399         struct dax_region *dax_region = dev_dax->region;
 400         struct device *dev = &dev_dax->dev;
 401         int rc = dev_dax->id;
 402
 403         device_lock_assert(dev);
 404
 405         if (is_static(dax_region) || dev_dax->id < 0)
 406                 return -1;
 407         ida_free(&dax_region->ida, dev_dax->id);
 408         dev_dax->id = -1;
 409         return rc;
 410 }
 411
 412 static int free_dev_dax_id(struct dev_dax *dev_dax)
 413 {
 414         struct device *dev = &dev_dax->dev;
 415         int rc;
 416
 417         device_lock(dev);
 418         rc = __free_dev_dax_id(dev_dax);
 419         device_unlock(dev);
 420         return rc;
 421 }
 422
 423 static ssize_t delete_store(struct device *dev, struct device_attribute *attr,
 424                 const char *buf, size_t len)
 425 {
 426         struct dax_region *dax_region = dev_get_drvdata(dev);
 427         struct dev_dax *dev_dax;
 428         struct device *victim;
 429         bool do_del = false;
 430         int rc;
 431
 432         if (is_static(dax_region))
 433                 return -EINVAL;
 434
 435         victim = device_find_child_by_name(dax_region->dev, buf);
 436         if (!victim)
 437                 return -ENXIO;
 438
 439         device_lock(dev);
 440         device_lock(victim);
 441         dev_dax = to_dev_dax(victim);
 442         if (victim->driver || dev_dax_size(dev_dax))
 443                 rc = -EBUSY;
 444         else {
 445                 /*
 446                  * Invalidate the device so it does not become active
 447                  * again, but always preserve device-id-0 so that
 448                  * /sys/bus/dax/ is guaranteed to be populated while any
 449                  * dax_region is registered.
 450                  */
 451                 if (dev_dax->id > 0) {
 452                         do_del = __free_dev_dax_id(dev_dax) >= 0;
 453                         rc = len;
 454                         if (dax_region->seed == victim)
 455                                 dax_region->seed = NULL;
 456                         if (dax_region->youngest == victim)
 457                                 dax_region->youngest = NULL;
 458                 } else
 459                         rc = -EBUSY;
 460         }
 461         device_unlock(victim);
 462
 463         /* won the race to invalidate the device, clean it up */
 464         if (do_del)
 465                 devm_release_action(dev, unregister_dev_dax, victim);
 466         device_unlock(dev);
 467         put_device(victim);
 468
 469         return rc;
 470 }
 471 static DEVICE_ATTR_WO(delete);
 472
 473 static umode_t dax_region_visible(struct kobject *kobj, struct attribute *a,
 474                 int n)
 475 {
 476         struct device *dev = container_of(kobj, struct device, kobj);
 477         struct dax_region *dax_region = dev_get_drvdata(dev);
 478
 479         if (is_static(dax_region))
 480                 if (a == &dev_attr_available_size.attr
 481                                 || a == &dev_attr_create.attr
 482                                 || a == &dev_attr_seed.attr
 483                                 || a == &dev_attr_delete.attr)
 484                         return 0;
 485         return a->mode;
 486 }
 487
 488 static struct attribute *dax_region_attributes[] = {
 489         &dev_attr_available_size.attr,
 490         &dev_attr_region_size.attr,
 491         &dev_attr_align.attr,
 492         &dev_attr_create.attr,
 493         &dev_attr_seed.attr,
 494         &dev_attr_delete.attr,
 495         &dev_attr_id.attr,
 496         NULL,
 497 };
 498
 499 static const struct attribute_group dax_region_attribute_group = {
 500         .name = "dax_region",
 501         .attrs = dax_region_attributes,
 502         .is_visible = dax_region_visible,
 503 };
 504
 505 static const struct attribute_group *dax_region_attribute_groups[] = {
 506         &dax_region_attribute_group,
 507         NULL,
 508 };
 509
 510 static void dax_region_free(struct kref *kref)
 511 {
 512         struct dax_region *dax_region;
 513
 514         dax_region = container_of(kref, struct dax_region, kref);
 515         kfree(dax_region);
 516 }
 517
 518 void dax_region_put(struct dax_region *dax_region)
 519 {
 520         kref_put(&dax_region->kref, dax_region_free);
 521 }
 522 EXPORT_SYMBOL_GPL(dax_region_put);
 523
 524 static void dax_region_unregister(void *region)
 525 {
 526         struct dax_region *dax_region = region;
 527
 528         sysfs_remove_groups(&dax_region->dev->kobj,
 529                         dax_region_attribute_groups);
 530         dax_region_put(dax_region);
 531 }
 532
 533 struct dax_region *alloc_dax_region(struct device *parent, int region_id,
 534                 struct range *range, int target_node, unsigned int align,
 535                 unsigned long flags)
 536 {
 537         struct dax_region *dax_region;
 538
 539         /*
 540          * The DAX core assumes that it can store its private data in
 541          * parent->driver_data. This WARN is a reminder / safeguard for
 542          * developers of device-dax drivers.
 543          */
 544         if (dev_get_drvdata(parent)) {
 545                 dev_WARN(parent, "dax core failed to setup private data\n");
 546                 return NULL;
 547         }
 548
 549         if (!IS_ALIGNED(range->start, align)
 550                         || !IS_ALIGNED(range_len(range), align))
 551                 return NULL;
 552
 553         dax_region = kzalloc(sizeof(*dax_region), GFP_KERNEL);
 554         if (!dax_region)
 555                 return NULL;
 556
 557         dev_set_drvdata(parent, dax_region);
 558         kref_init(&dax_region->kref);
 559         dax_region->id = region_id;
 560         dax_region->align = align;
 561         dax_region->dev = parent;
 562         dax_region->target_node = target_node;
 563         ida_init(&dax_region->ida);
 564         dax_region->res = (struct resource) {
 565                 .start = range->start,
 566                 .end = range->end,
 567                 .flags = IORESOURCE_MEM | flags,
 568         };
 569
 570         if (sysfs_create_groups(&parent->kobj, dax_region_attribute_groups)) {
 571                 kfree(dax_region);
 572                 return NULL;
 573         }
 574
 575         kref_get(&dax_region->kref);
 576         if (devm_add_action_or_reset(parent, dax_region_unregister, dax_region))
 577                 return NULL;
 578         return dax_region;
 579 }
 580 EXPORT_SYMBOL_GPL(alloc_dax_region);
 581
 582 static void dax_mapping_release(struct device *dev)
 583 {
 584         struct dax_mapping *mapping = to_dax_mapping(dev);
 585         struct dev_dax *dev_dax = to_dev_dax(dev->parent);
 586
 587         ida_free(&dev_dax->ida, mapping->id);
 588         kfree(mapping);
 589 }
 590
 591 static void unregister_dax_mapping(void *data)
 592 {
 593         struct device *dev = data;
 594         struct dax_mapping *mapping = to_dax_mapping(dev);
 595         struct dev_dax *dev_dax = to_dev_dax(dev->parent);
 596         struct dax_region *dax_region = dev_dax->region;
 597
 598         dev_dbg(dev, "%s\n", __func__);
 599
 600         device_lock_assert(dax_region->dev);
 601
 602         dev_dax->ranges[mapping->range_id].mapping = NULL;
 603         mapping->range_id = -1;
 604
 605         device_del(dev);
 606         put_device(dev);
 607 }
 608
 609 static struct dev_dax_range *get_dax_range(struct device *dev)
 610 {
 611         struct dax_mapping *mapping = to_dax_mapping(dev);
 612         struct dev_dax *dev_dax = to_dev_dax(dev->parent);
 613         struct dax_region *dax_region = dev_dax->region;
 614
 615         device_lock(dax_region->dev);
 616         if (mapping->range_id < 0) {
 617                 device_unlock(dax_region->dev);
 618                 return NULL;
 619         }
 620
 621         return &dev_dax->ranges[mapping->range_id];
 622 }
 623
 624 static void put_dax_range(struct dev_dax_range *dax_range)
 625 {
 626         struct dax_mapping *mapping = dax_range->mapping;
 627         struct dev_dax *dev_dax = to_dev_dax(mapping->dev.parent);
 628         struct dax_region *dax_region = dev_dax->region;
 629
 630         device_unlock(dax_region->dev);
 631 }
 632
 633 static ssize_t start_show(struct device *dev,
 634                 struct device_attribute *attr, char *buf)
 635 {
 636         struct dev_dax_range *dax_range;
 637         ssize_t rc;
 638
 639         dax_range = get_dax_range(dev);
 640         if (!dax_range)
 641                 return -ENXIO;
 642         rc = sprintf(buf, "%#llx\n", dax_range->range.start);
 643         put_dax_range(dax_range);
 644
 645         return rc;
 646 }
 647 static DEVICE_ATTR(start, 0400, start_show, NULL);
 648
 649 static ssize_t end_show(struct device *dev,
 650                 struct device_attribute *attr, char *buf)
 651 {
 652         struct dev_dax_range *dax_range;
 653         ssize_t rc;
 654
 655         dax_range = get_dax_range(dev);
 656         if (!dax_range)
 657                 return -ENXIO;
 658         rc = sprintf(buf, "%#llx\n", dax_range->range.end);
 659         put_dax_range(dax_range);
 660
 661         return rc;
 662 }
 663 static DEVICE_ATTR(end, 0400, end_show, NULL);
 664
 665 static ssize_t pgoff_show(struct device *dev,
 666                 struct device_attribute *attr, char *buf)
 667 {
 668         struct dev_dax_range *dax_range;
 669         ssize_t rc;
 670
 671         dax_range = get_dax_range(dev);
 672         if (!dax_range)
 673                 return -ENXIO;
 674         rc = sprintf(buf, "%#lx\n", dax_range->pgoff);
 675         put_dax_range(dax_range);
 676
 677         return rc;
 678 }
 679 static DEVICE_ATTR(page_offset, 0400, pgoff_show, NULL);
 680
 681 static struct attribute *dax_mapping_attributes[] = {
 682         &dev_attr_start.attr,
 683         &dev_attr_end.attr,
 684         &dev_attr_page_offset.attr,
 685         NULL,
 686 };
 687
 688 static const struct attribute_group dax_mapping_attribute_group = {
 689         .attrs = dax_mapping_attributes,
 690 };
 691
 692 static const struct attribute_group *dax_mapping_attribute_groups[] = {
 693         &dax_mapping_attribute_group,
 694         NULL,
 695 };
 696
 697 static struct device_type dax_mapping_type = {
 698         .release = dax_mapping_release,
 699         .groups = dax_mapping_attribute_groups,
 700 };
 701
 702 static int devm_register_dax_mapping(struct dev_dax *dev_dax, int range_id)
 703 {
 704         struct dax_region *dax_region = dev_dax->region;
 705         struct dax_mapping *mapping;
 706         struct device *dev;
 707         int rc;
 708
 709         device_lock_assert(dax_region->dev);
 710
 711         if (dev_WARN_ONCE(&dev_dax->dev, !dax_region->dev->driver,
 712                                 "region disabled\n"))
 713                 return -ENXIO;
 714
 715         mapping = kzalloc(sizeof(*mapping), GFP_KERNEL);
 716         if (!mapping)
 717                 return -ENOMEM;
 718         mapping->range_id = range_id;
 719         mapping->id = ida_alloc(&dev_dax->ida, GFP_KERNEL);
 720         if (mapping->id < 0) {
 721                 kfree(mapping);
 722                 return -ENOMEM;
 723         }
 724         dev_dax->ranges[range_id].mapping = mapping;
 725         dev = &mapping->dev;
 726         device_initialize(dev);
 727         dev->parent = &dev_dax->dev;
 728         dev->type = &dax_mapping_type;
 729         dev_set_name(dev, "mapping%d", mapping->id);
 730         rc = device_add(dev);
 731         if (rc) {
 732                 put_device(dev);
 733                 return rc;
 734         }
 735
 736         rc = devm_add_action_or_reset(dax_region->dev, unregister_dax_mapping,
 737                         dev);
 738         if (rc)
 739                 return rc;
 740         return 0;
 741 }
 742
 743 static int alloc_dev_dax_range(struct dev_dax *dev_dax, u64 start,
 744                 resource_size_t size)
 745 {
 746         struct dax_region *dax_region = dev_dax->region;
 747         struct resource *res = &dax_region->res;
 748         struct device *dev = &dev_dax->dev;
 749         struct dev_dax_range *ranges;
 750         unsigned long pgoff = 0;
 751         struct resource *alloc;
 752         int i, rc;
 753
 754         device_lock_assert(dax_region->dev);
 755
 756         /* handle the seed alloc special case */
 757         if (!size) {
 758                 if (dev_WARN_ONCE(dev, dev_dax->nr_range,
 759                                         "0-size allocation must be first\n"))
 760                         return -EBUSY;
 761                 /* nr_range == 0 is elsewhere special cased as 0-size device */
 762                 return 0;
 763         }
 764
 765         ranges = krealloc(dev_dax->ranges, sizeof(*ranges)
 766                         * (dev_dax->nr_range + 1), GFP_KERNEL);
 767         if (!ranges)
 768                 return -ENOMEM;
 769
 770         alloc = __request_region(res, start, size, dev_name(dev), 0);
 771         if (!alloc) {
 772                 /*
 773                  * If this was an empty set of ranges nothing else
 774                  * will release @ranges, so do it now.
 775                  */
 776                 if (!dev_dax->nr_range) {
 777                         kfree(ranges);
 778                         ranges = NULL;
 779                 }
 780                 dev_dax->ranges = ranges;
 781                 return -ENOMEM;
 782         }
 783
 784         for (i = 0; i < dev_dax->nr_range; i++)
 785                 pgoff += PHYS_PFN(range_len(&ranges[i].range));
 786         dev_dax->ranges = ranges;
 787         ranges[dev_dax->nr_range++] = (struct dev_dax_range) {
 788                 .pgoff = pgoff,
 789                 .range = {
 790                         .start = alloc->start,
 791                         .end = alloc->end,
 792                 },
 793         };
 794
 795         dev_dbg(dev, "alloc range[%d]: %pa:%pa\n", dev_dax->nr_range - 1,
 796                         &alloc->start, &alloc->end);
 797         /*
 798          * A dev_dax instance must be registered before mapping device
 799          * children can be added. Defer to devm_create_dev_dax() to add
 800          * the initial mapping device.
 801          */
 802         if (!device_is_registered(&dev_dax->dev))
 803                 return 0;
 804
 805         rc = devm_register_dax_mapping(dev_dax, dev_dax->nr_range - 1);
 806         if (rc) {
 807                 dev_dbg(dev, "delete range[%d]: %pa:%pa\n", dev_dax->nr_range - 1,
 808                                 &alloc->start, &alloc->end);
 809                 dev_dax->nr_range--;
 810                 __release_region(res, alloc->start, resource_size(alloc));
 811                 return rc;
 812         }
 813
 814         return 0;
 815 }
 816
 817 static int adjust_dev_dax_range(struct dev_dax *dev_dax, struct resource *res, resource_size_t size)
 818 {
 819         int last_range = dev_dax->nr_range - 1;
 820         struct dev_dax_range *dax_range = &dev_dax->ranges[last_range];
 821         struct dax_region *dax_region = dev_dax->region;
 822         bool is_shrink = resource_size(res) > size;
 823         struct range *range = &dax_range->range;
 824         struct device *dev = &dev_dax->dev;
 825         int rc;
 826
 827         device_lock_assert(dax_region->dev);
 828
 829         if (dev_WARN_ONCE(dev, !size, "deletion is handled by dev_dax_shrink\n"))
 830                 return -EINVAL;
 831
 832         rc = adjust_resource(res, range->start, size);
 833         if (rc)
 834                 return rc;
 835
 836         *range = (struct range) {
 837                 .start = range->start,
 838                 .end = range->start + size - 1,
 839         };
 840
 841         dev_dbg(dev, "%s range[%d]: %#llx:%#llx\n", is_shrink ? "shrink" : "extend",
 842                         last_range, (unsigned long long) range->start,
 843                         (unsigned long long) range->end);
 844
 845         return 0;
 846 }
 847
 848 static ssize_t size_show(struct device *dev,
 849                 struct device_attribute *attr, char *buf)
 850 {
 851         struct dev_dax *dev_dax = to_dev_dax(dev);
 852         unsigned long long size;
 853
 854         device_lock(dev);
 855         size = dev_dax_size(dev_dax);
 856         device_unlock(dev);
 857
 858         return sprintf(buf, "%llu\n", size);
 859 }
 860
 861 static bool alloc_is_aligned(struct dax_region *dax_region,
 862                 resource_size_t size)
 863 {
 864         /*
 865          * The minimum mapping granularity for a device instance is a
 866          * single subsection, unless the arch says otherwise.
 867          */
 868         return IS_ALIGNED(size, max_t(unsigned long, dax_region->align,
 869                                 memremap_compat_align()));
 870 }
 871
 872 static int dev_dax_shrink(struct dev_dax *dev_dax, resource_size_t size)
 873 {
 874         resource_size_t to_shrink = dev_dax_size(dev_dax) - size;
 875         struct dax_region *dax_region = dev_dax->region;
 876         struct device *dev = &dev_dax->dev;
 877         int i;
 878
 879         for (i = dev_dax->nr_range - 1; i >= 0; i--) {
 880                 struct range *range = &dev_dax->ranges[i].range;
 881                 struct dax_mapping *mapping = dev_dax->ranges[i].mapping;
 882                 struct resource *adjust = NULL, *res;
 883                 resource_size_t shrink;
 884
 885                 shrink = min_t(u64, to_shrink, range_len(range));
 886                 if (shrink >= range_len(range)) {
 887                         devm_release_action(dax_region->dev,
 888                                         unregister_dax_mapping, &mapping->dev);
 889                         __release_region(&dax_region->res, range->start,
 890                                         range_len(range));
 891                         dev_dax->nr_range--;
 892                         dev_dbg(dev, "delete range[%d]: %#llx:%#llx\n", i,
 893                                         (unsigned long long) range->start,
 894                                         (unsigned long long) range->end);
 895                         to_shrink -= shrink;
 896                         if (!to_shrink)
 897                                 break;
 898                         continue;
 899                 }
 900
 901                 for_each_dax_region_resource(dax_region, res)
 902                         if (strcmp(res->name, dev_name(dev)) == 0
 903                                         && res->start == range->start) {
 904                                 adjust = res;
 905                                 break;
 906                         }
 907
 908                 if (dev_WARN_ONCE(dev, !adjust || i != dev_dax->nr_range - 1,
 909                                         "failed to find matching resource\n"))
 910                         return -ENXIO;
 911                 return adjust_dev_dax_range(dev_dax, adjust, range_len(range)
 912                                 - shrink);
 913         }
 914         return 0;
 915 }
 916
 917 /*
 918  * Only allow adjustments that preserve the relative pgoff of existing
 919  * allocations. I.e. the dev_dax->ranges array is ordered by increasing pgoff.
 920  */
 921 static bool adjust_ok(struct dev_dax *dev_dax, struct resource *res)
 922 {
 923         struct dev_dax_range *last;
 924         int i;
 925
 926         if (dev_dax->nr_range == 0)
 927                 return false;
 928         if (strcmp(res->name, dev_name(&dev_dax->dev)) != 0)
 929                 return false;
 930         last = &dev_dax->ranges[dev_dax->nr_range - 1];
 931         if (last->range.start != res->start || last->range.end != res->end)
 932                 return false;
 933         for (i = 0; i < dev_dax->nr_range - 1; i++) {
 934                 struct dev_dax_range *dax_range = &dev_dax->ranges[i];
 935
 936                 if (dax_range->pgoff > last->pgoff)
 937                         return false;
 938         }
 939
 940         return true;
 941 }
 942
 943 static ssize_t dev_dax_resize(struct dax_region *dax_region,
 944                 struct dev_dax *dev_dax, resource_size_t size)
 945 {
 946         resource_size_t avail = dax_region_avail_size(dax_region), to_alloc;
 947         resource_size_t dev_size = dev_dax_size(dev_dax);
 948         struct resource *region_res = &dax_region->res;
 949         struct device *dev = &dev_dax->dev;
 950         struct resource *res, *first;
 951         resource_size_t alloc = 0;
 952         int rc;
 953
 954         if (dev->driver)
 955                 return -EBUSY;
 956         if (size == dev_size)
 957                 return 0;
 958         if (size > dev_size && size - dev_size > avail)
 959                 return -ENOSPC;
 960         if (size < dev_size)
 961                 return dev_dax_shrink(dev_dax, size);
 962
 963         to_alloc = size - dev_size;
 964         if (dev_WARN_ONCE(dev, !alloc_is_aligned(dax_region, to_alloc),
 965                         "resize of %pa misaligned\n", &to_alloc))
 966                 return -ENXIO;
 967
 968         /*
 969          * Expand the device into the unused portion of the region. This
 970          * may involve adjusting the end of an existing resource, or
 971          * allocating a new resource.
 972          */
 973 retry:
 974         first = region_res->child;
 975         if (!first)
 976                 return alloc_dev_dax_range(dev_dax, dax_region->res.start, to_alloc);
 977
 978         rc = -ENOSPC;
 979         for (res = first; res; res = res->sibling) {
 980                 struct resource *next = res->sibling;
 981
 982                 /* space at the beginning of the region */
 983                 if (res == first && res->start > dax_region->res.start) {
 984                         alloc = min(res->start - dax_region->res.start, to_alloc);
 985                         rc = alloc_dev_dax_range(dev_dax, dax_region->res.start, alloc);
 986                         break;
 987                 }
 988
 989                 alloc = 0;
 990                 /* space between allocations */
 991                 if (next && next->start > res->end + 1)
 992                         alloc = min(next->start - (res->end + 1), to_alloc);
 993
 994                 /* space at the end of the region */
 995                 if (!alloc && !next && res->end < region_res->end)
 996                         alloc = min(region_res->end - res->end, to_alloc);
 997
 998                 if (!alloc)
 999                         continue;
1000
1001                 if (adjust_ok(dev_dax, res)) {
1002                         rc = adjust_dev_dax_range(dev_dax, res, resource_size(res) + alloc);
1003                         break;
1004                 }
1005                 rc = alloc_dev_dax_range(dev_dax, res->end + 1, alloc);
1006                 break;
1007         }
1008         if (rc)
1009                 return rc;
1010         to_alloc -= alloc;
1011         if (to_alloc)
1012                 goto retry;
1013         return 0;
1014 }
1015
1016 static ssize_t size_store(struct device *dev, struct device_attribute *attr,
1017                 const char *buf, size_t len)
1018 {
1019         ssize_t rc;
1020         unsigned long long val;
1021         struct dev_dax *dev_dax = to_dev_dax(dev);
1022         struct dax_region *dax_region = dev_dax->region;
1023
1024         rc = kstrtoull(buf, 0, &val);
1025         if (rc)
1026                 return rc;
1027
1028         if (!alloc_is_aligned(dax_region, val)) {
1029                 dev_dbg(dev, "%s: size: %lld misaligned\n", __func__, val);
1030                 return -EINVAL;
1031         }
1032
1033         device_lock(dax_region->dev);
1034         if (!dax_region->dev->driver) {
1035                 device_unlock(dax_region->dev);
1036                 return -ENXIO;
1037         }
1038         device_lock(dev);
1039         rc = dev_dax_resize(dax_region, dev_dax, val);
1040         device_unlock(dev);
1041         device_unlock(dax_region->dev);
1042
1043         return rc == 0 ? len : rc;
1044 }
1045 static DEVICE_ATTR_RW(size);
1046
1047 static int dev_dax_target_node(struct dev_dax *dev_dax)
1048 {
1049         struct dax_region *dax_region = dev_dax->region;
1050
1051         return dax_region->target_node;
1052 }
1053
1054 static ssize_t target_node_show(struct device *dev,
1055                 struct device_attribute *attr, char *buf)
1056 {
1057         struct dev_dax *dev_dax = to_dev_dax(dev);
1058
1059         return sprintf(buf, "%d\n", dev_dax_target_node(dev_dax));
1060 }
1061 static DEVICE_ATTR_RO(target_node);
1062
1063 static ssize_t resource_show(struct device *dev,
1064                 struct device_attribute *attr, char *buf)
1065 {
1066         struct dev_dax *dev_dax = to_dev_dax(dev);
1067         struct dax_region *dax_region = dev_dax->region;
1068         unsigned long long start;
1069
1070         if (dev_dax->nr_range < 1)
1071                 start = dax_region->res.start;
1072         else
1073                 start = dev_dax->ranges[0].range.start;
1074
1075         return sprintf(buf, "%#llx\n", start);
1076 }
1077 static DEVICE_ATTR(resource, 0400, resource_show, NULL);
1078
1079 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr,
1080                 char *buf)
1081 {
1082         /*
1083          * We only ever expect to handle device-dax instances, i.e. the
1084          * @type argument to MODULE_ALIAS_DAX_DEVICE() is always zero
1085          */
1086         return sprintf(buf, DAX_DEVICE_MODALIAS_FMT "\n", 0);
1087 }
1088 static DEVICE_ATTR_RO(modalias);
1089
1090 static ssize_t numa_node_show(struct device *dev,
1091                 struct device_attribute *attr, char *buf)
1092 {
1093         return sprintf(buf, "%d\n", dev_to_node(dev));
1094 }
1095 static DEVICE_ATTR_RO(numa_node);
1096
1097 static umode_t dev_dax_visible(struct kobject *kobj, struct attribute *a, int n)
1098 {
1099         struct device *dev = container_of(kobj, struct device, kobj);
1100         struct dev_dax *dev_dax = to_dev_dax(dev);
1101         struct dax_region *dax_region = dev_dax->region;
1102
1103         if (a == &dev_attr_target_node.attr && dev_dax_target_node(dev_dax) < 0)
1104                 return 0;
1105         if (a == &dev_attr_numa_node.attr && !IS_ENABLED(CONFIG_NUMA))
1106                 return 0;
1107         if (a == &dev_attr_size.attr && is_static(dax_region))
1108                 return 0444;
1109         return a->mode;
1110 }
1111
1112 static struct attribute *dev_dax_attributes[] = {
1113         &dev_attr_modalias.attr,
1114         &dev_attr_size.attr,
1115         &dev_attr_target_node.attr,
1116         &dev_attr_resource.attr,
1117         &dev_attr_numa_node.attr,
1118         NULL,
1119 };
1120
1121 static const struct attribute_group dev_dax_attribute_group = {
1122         .attrs = dev_dax_attributes,
1123         .is_visible = dev_dax_visible,
1124 };
1125
1126 static const struct attribute_group *dax_attribute_groups[] = {
1127         &dev_dax_attribute_group,
1128         NULL,
1129 };
1130
1131 static void dev_dax_release(struct device *dev)
1132 {
1133         struct dev_dax *dev_dax = to_dev_dax(dev);
1134         struct dax_region *dax_region = dev_dax->region;
1135         struct dax_device *dax_dev = dev_dax->dax_dev;
1136
1137         put_dax(dax_dev);
1138         free_dev_dax_id(dev_dax);
1139         dax_region_put(dax_region);
1140         kfree(dev_dax->ranges);
1141         kfree(dev_dax->pgmap);
1142         kfree(dev_dax);
1143 }
1144
1145 static const struct device_type dev_dax_type = {
1146         .release = dev_dax_release,
1147         .groups = dax_attribute_groups,
1148 };
1149
1150 struct dev_dax *devm_create_dev_dax(struct dev_dax_data *data)
1151 {
1152         struct dax_region *dax_region = data->dax_region;
1153         struct device *parent = dax_region->dev;
1154         struct dax_device *dax_dev;
1155         struct dev_dax *dev_dax;
1156         struct inode *inode;
1157         struct device *dev;
1158         int rc;
1159
1160         dev_dax = kzalloc(sizeof(*dev_dax), GFP_KERNEL);
1161         if (!dev_dax)
1162                 return ERR_PTR(-ENOMEM);
1163
1164         if (is_static(dax_region)) {
1165                 if (dev_WARN_ONCE(parent, data->id < 0,
1166                                 "dynamic id specified to static region\n")) {
1167                         rc = -EINVAL;
1168                         goto err_id;
1169                 }
1170
1171                 dev_dax->id = data->id;
1172         } else {
1173                 if (dev_WARN_ONCE(parent, data->id >= 0,
1174                                 "static id specified to dynamic region\n")) {
1175                         rc = -EINVAL;
1176                         goto err_id;
1177                 }
1178
1179                 rc = ida_alloc(&dax_region->ida, GFP_KERNEL);
1180                 if (rc < 0)
1181                         goto err_id;
1182                 dev_dax->id = rc;
1183         }
1184
1185         dev_dax->region = dax_region;
1186         dev = &dev_dax->dev;
1187         device_initialize(dev);
1188         dev_set_name(dev, "dax%d.%d", dax_region->id, dev_dax->id);
1189
1190         rc = alloc_dev_dax_range(dev_dax, dax_region->res.start, data->size);
1191         if (rc)
1192                 goto err_range;
1193
1194         if (data->pgmap) {
1195                 dev_WARN_ONCE(parent, !is_static(dax_region),
1196                         "custom dev_pagemap requires a static dax_region\n");
1197
1198                 dev_dax->pgmap = kmemdup(data->pgmap,
1199                                 sizeof(struct dev_pagemap), GFP_KERNEL);
1200                 if (!dev_dax->pgmap) {
1201                         rc = -ENOMEM;
1202                         goto err_pgmap;
1203                 }
1204         }
1205
1206         /*
1207          * No 'host' or dax_operations since there is no access to this
1208          * device outside of mmap of the resulting character device.
1209          */
1210         dax_dev = alloc_dax(dev_dax, NULL, NULL, DAXDEV_F_SYNC);
1211         if (IS_ERR(dax_dev)) {
1212                 rc = PTR_ERR(dax_dev);
1213                 goto err_alloc_dax;
1214         }
1215
1216         /* a device_dax instance is dead while the driver is not attached */
1217         kill_dax(dax_dev);
1218
1219         dev_dax->dax_dev = dax_dev;
1220         dev_dax->target_node = dax_region->target_node;
1221         dev_dax->align = dax_region->align;
1222         ida_init(&dev_dax->ida);
1223         kref_get(&dax_region->kref);
1224
1225         inode = dax_inode(dax_dev);
1226         dev->devt = inode->i_rdev;
1227         if (data->subsys == DEV_DAX_BUS)
1228                 dev->bus = &dax_bus_type;
1229         else
1230                 dev->class = dax_class;
1231         dev->parent = parent;
1232         dev->type = &dev_dax_type;
1233
1234         rc = device_add(dev);
1235         if (rc) {
1236                 kill_dev_dax(dev_dax);
1237                 put_device(dev);
1238                 return ERR_PTR(rc);
1239         }
1240
1241         rc = devm_add_action_or_reset(dax_region->dev, unregister_dev_dax, dev);
1242         if (rc)
1243                 return ERR_PTR(rc);
1244
1245         /* register mapping device for the initial allocation range */
1246         if (dev_dax->nr_range && range_len(&dev_dax->ranges[0].range)) {
1247                 rc = devm_register_dax_mapping(dev_dax, 0);
1248                 if (rc)
1249                         return ERR_PTR(rc);
1250         }
1251
1252         return dev_dax;
1253
1254 err_alloc_dax:
1255         kfree(dev_dax->pgmap);
1256 err_pgmap:
1257         free_dev_dax_ranges(dev_dax);
1258 err_range:
1259         free_dev_dax_id(dev_dax);
1260 err_id:
1261         kfree(dev_dax);
1262
1263         return ERR_PTR(rc);
1264 }
1265 EXPORT_SYMBOL_GPL(devm_create_dev_dax);
1266
1267 static int match_always_count;
1268
1269 int __dax_driver_register(struct dax_device_driver *dax_drv,
1270                 struct module *module, const char *mod_name)
1271 {
1272         struct device_driver *drv = &dax_drv->drv;
1273         int rc = 0;
1274
1275         INIT_LIST_HEAD(&dax_drv->ids);
1276         drv->owner = module;
1277         drv->name = mod_name;
1278         drv->mod_name = mod_name;
1279         drv->bus = &dax_bus_type;
1280
1281         /* there can only be one default driver */
1282         mutex_lock(&dax_bus_lock);
1283         match_always_count += dax_drv->match_always;
1284         if (match_always_count > 1) {
1285                 match_always_count--;
1286                 WARN_ON(1);
1287                 rc = -EINVAL;
1288         }
1289         mutex_unlock(&dax_bus_lock);
1290         if (rc)
1291                 return rc;
1292         return driver_register(drv);
1293 }
1294 EXPORT_SYMBOL_GPL(__dax_driver_register);
1295
1296 void dax_driver_unregister(struct dax_device_driver *dax_drv)
1297 {
1298         struct device_driver *drv = &dax_drv->drv;
1299         struct dax_id *dax_id, *_id;
1300
1301         mutex_lock(&dax_bus_lock);
1302         match_always_count -= dax_drv->match_always;
1303         list_for_each_entry_safe(dax_id, _id, &dax_drv->ids, list) {
1304                 list_del(&dax_id->list);
1305                 kfree(dax_id);
1306         }
1307         mutex_unlock(&dax_bus_lock);
1308         driver_unregister(drv);
1309 }
1310 EXPORT_SYMBOL_GPL(dax_driver_unregister);
1311
1312 int __init dax_bus_init(void)
1313 {
1314         int rc;
1315
1316         if (IS_ENABLED(CONFIG_DEV_DAX_PMEM_COMPAT)) {
1317                 dax_class = class_create(THIS_MODULE, "dax");
1318                 if (IS_ERR(dax_class))
1319                         return PTR_ERR(dax_class);
1320         }
1321
1322         rc = bus_register(&dax_bus_type);
1323         if (rc)
1324                 class_destroy(dax_class);
1325         return rc;
1326 }
1327
1328 void __exit dax_bus_exit(void)
1329 {
1330         bus_unregister(&dax_bus_type);
1331         class_destroy(dax_class);
1332 }