drivers/nvdimm/pmem.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Persistent Memory Driver
   4  *
   5  * Copyright (c) 2014-2015, Intel Corporation.
   6  * Copyright (c) 2015, Christoph Hellwig <hch@lst.de>.
   7  * Copyright (c) 2015, Boaz Harrosh <boaz@plexistor.com>.
   8  */
   9
  10 #include <linux/blkdev.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/hdreg.h>
  13 #include <linux/init.h>
  14 #include <linux/platform_device.h>
  15 #include <linux/set_memory.h>
  16 #include <linux/module.h>
  17 #include <linux/moduleparam.h>
  18 #include <linux/badblocks.h>
  19 #include <linux/memremap.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/blk-mq.h>
  22 #include <linux/pfn_t.h>
  23 #include <linux/slab.h>
  24 #include <linux/uio.h>
  25 #include <linux/dax.h>
  26 #include <linux/nd.h>
  27 #include <linux/mm.h>
  28 #include <asm/cacheflush.h>
  29 #include "pmem.h"
  30 #include "btt.h"
  31 #include "pfn.h"
  32 #include "nd.h"
  33
  34 static struct device *to_dev(struct pmem_device *pmem)
  35 {
  36         /*
  37          * nvdimm bus services need a 'dev' parameter, and we record the device
  38          * at init in bb.dev.
  39          */
  40         return pmem->bb.dev;
  41 }
  42
  43 static struct nd_region *to_region(struct pmem_device *pmem)
  44 {
  45         return to_nd_region(to_dev(pmem)->parent);
  46 }
  47
  48 static phys_addr_t pmem_to_phys(struct pmem_device *pmem, phys_addr_t offset)
  49 {
  50         return pmem->phys_addr + offset;
  51 }
  52
  53 static sector_t to_sect(struct pmem_device *pmem, phys_addr_t offset)
  54 {
  55         return (offset - pmem->data_offset) >> SECTOR_SHIFT;
  56 }
  57
  58 static phys_addr_t to_offset(struct pmem_device *pmem, sector_t sector)
  59 {
  60         return (sector << SECTOR_SHIFT) + pmem->data_offset;
  61 }
  62
  63 static void pmem_mkpage_present(struct pmem_device *pmem, phys_addr_t offset,
  64                 unsigned int len)
  65 {
  66         phys_addr_t phys = pmem_to_phys(pmem, offset);
  67         unsigned long pfn_start, pfn_end, pfn;
  68
  69         /* only pmem in the linear map supports HWPoison */
  70         if (is_vmalloc_addr(pmem->virt_addr))
  71                 return;
  72
  73         pfn_start = PHYS_PFN(phys);
  74         pfn_end = pfn_start + PHYS_PFN(len);
  75         for (pfn = pfn_start; pfn < pfn_end; pfn++) {
  76                 struct page *page = pfn_to_page(pfn);
  77
  78                 /*
  79                  * Note, no need to hold a get_dev_pagemap() reference
  80                  * here since we're in the driver I/O path and
  81                  * outstanding I/O requests pin the dev_pagemap.
  82                  */
  83                 if (test_and_clear_pmem_poison(page))
  84                         clear_mce_nospec(pfn);
  85         }
  86 }
  87
  88 static void pmem_clear_bb(struct pmem_device *pmem, sector_t sector, long blks)
  89 {
  90         if (blks == 0)
  91                 return;
  92         badblocks_clear(&pmem->bb, sector, blks);
  93         if (pmem->bb_state)
  94                 sysfs_notify_dirent(pmem->bb_state);
  95 }
  96
  97 static long __pmem_clear_poison(struct pmem_device *pmem,
  98                 phys_addr_t offset, unsigned int len)
  99 {
 100         phys_addr_t phys = pmem_to_phys(pmem, offset);
 101         long cleared = nvdimm_clear_poison(to_dev(pmem), phys, len);
 102
 103         if (cleared > 0) {
 104                 pmem_mkpage_present(pmem, offset, cleared);
 105                 arch_invalidate_pmem(pmem->virt_addr + offset, len);
 106         }
 107         return cleared;
 108 }
 109
 110 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
 111                 phys_addr_t offset, unsigned int len)
 112 {
 113         long cleared = __pmem_clear_poison(pmem, offset, len);
 114
 115         if (cleared < 0)
 116                 return BLK_STS_IOERR;
 117
 118         pmem_clear_bb(pmem, to_sect(pmem, offset), cleared >> SECTOR_SHIFT);
 119         if (cleared < len)
 120                 return BLK_STS_IOERR;
 121         return BLK_STS_OK;
 122 }
 123
 124 static void write_pmem(void *pmem_addr, struct page *page,
 125                 unsigned int off, unsigned int len)
 126 {
 127         unsigned int chunk;
 128         void *mem;
 129
 130         while (len) {
 131                 mem = kmap_atomic(page);
 132                 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
 133                 memcpy_flushcache(pmem_addr, mem + off, chunk);
 134                 kunmap_atomic(mem);
 135                 len -= chunk;
 136                 off = 0;
 137                 page++;
 138                 pmem_addr += chunk;
 139         }
 140 }
 141
 142 static blk_status_t read_pmem(struct page *page, unsigned int off,
 143                 void *pmem_addr, unsigned int len)
 144 {
 145         unsigned int chunk;
 146         unsigned long rem;
 147         void *mem;
 148
 149         while (len) {
 150                 mem = kmap_atomic(page);
 151                 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
 152                 rem = copy_mc_to_kernel(mem + off, pmem_addr, chunk);
 153                 kunmap_atomic(mem);
 154                 if (rem)
 155                         return BLK_STS_IOERR;
 156                 len -= chunk;
 157                 off = 0;
 158                 page++;
 159                 pmem_addr += chunk;
 160         }
 161         return BLK_STS_OK;
 162 }
 163
 164 static blk_status_t pmem_do_read(struct pmem_device *pmem,
 165                         struct page *page, unsigned int page_off,
 166                         sector_t sector, unsigned int len)
 167 {
 168         blk_status_t rc;
 169         phys_addr_t pmem_off = to_offset(pmem, sector);
 170         void *pmem_addr = pmem->virt_addr + pmem_off;
 171
 172         if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
 173                 return BLK_STS_IOERR;
 174
 175         rc = read_pmem(page, page_off, pmem_addr, len);
 176         flush_dcache_page(page);
 177         return rc;
 178 }
 179
 180 static blk_status_t pmem_do_write(struct pmem_device *pmem,
 181                         struct page *page, unsigned int page_off,
 182                         sector_t sector, unsigned int len)
 183 {
 184         phys_addr_t pmem_off = to_offset(pmem, sector);
 185         void *pmem_addr = pmem->virt_addr + pmem_off;
 186
 187         if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) {
 188                 blk_status_t rc = pmem_clear_poison(pmem, pmem_off, len);
 189
 190                 if (rc != BLK_STS_OK)
 191                         return rc;
 192         }
 193
 194         flush_dcache_page(page);
 195         write_pmem(pmem_addr, page, page_off, len);
 196
 197         return BLK_STS_OK;
 198 }
 199
 200 static void pmem_submit_bio(struct bio *bio)
 201 {
 202         int ret = 0;
 203         blk_status_t rc = 0;
 204         bool do_acct;
 205         unsigned long start;
 206         struct bio_vec bvec;
 207         struct bvec_iter iter;
 208         struct pmem_device *pmem = bio->bi_bdev->bd_disk->private_data;
 209         struct nd_region *nd_region = to_region(pmem);
 210
 211         if (bio->bi_opf & REQ_PREFLUSH)
 212                 ret = nvdimm_flush(nd_region, bio);
 213
 214         do_acct = blk_queue_io_stat(bio->bi_bdev->bd_disk->queue);
 215         if (do_acct)
 216                 start = bio_start_io_acct(bio);
 217         bio_for_each_segment(bvec, bio, iter) {
 218                 if (op_is_write(bio_op(bio)))
 219                         rc = pmem_do_write(pmem, bvec.bv_page, bvec.bv_offset,
 220                                 iter.bi_sector, bvec.bv_len);
 221                 else
 222                         rc = pmem_do_read(pmem, bvec.bv_page, bvec.bv_offset,
 223                                 iter.bi_sector, bvec.bv_len);
 224                 if (rc) {
 225                         bio->bi_status = rc;
 226                         break;
 227                 }
 228         }
 229         if (do_acct)
 230                 bio_end_io_acct(bio, start);
 231
 232         if (bio->bi_opf & REQ_FUA)
 233                 ret = nvdimm_flush(nd_region, bio);
 234
 235         if (ret)
 236                 bio->bi_status = errno_to_blk_status(ret);
 237
 238         bio_endio(bio);
 239 }
 240
 241 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 242                        struct page *page, enum req_op op)
 243 {
 244         struct pmem_device *pmem = bdev->bd_disk->private_data;
 245         blk_status_t rc;
 246
 247         if (op_is_write(op))
 248                 rc = pmem_do_write(pmem, page, 0, sector, thp_size(page));
 249         else
 250                 rc = pmem_do_read(pmem, page, 0, sector, thp_size(page));
 251         /*
 252          * The ->rw_page interface is subtle and tricky.  The core
 253          * retries on any error, so we can only invoke page_endio() in
 254          * the successful completion case.  Otherwise, we'll see crashes
 255          * caused by double completion.
 256          */
 257         if (rc == 0)
 258                 page_endio(page, op_is_write(op), 0);
 259
 260         return blk_status_to_errno(rc);
 261 }
 262
 263 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
 264 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
 265                 long nr_pages, enum dax_access_mode mode, void **kaddr,
 266                 pfn_t *pfn)
 267 {
 268         resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
 269         sector_t sector = PFN_PHYS(pgoff) >> SECTOR_SHIFT;
 270         unsigned int num = PFN_PHYS(nr_pages) >> SECTOR_SHIFT;
 271         struct badblocks *bb = &pmem->bb;
 272         sector_t first_bad;
 273         int num_bad;
 274
 275         if (kaddr)
 276                 *kaddr = pmem->virt_addr + offset;
 277         if (pfn)
 278                 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 279
 280         if (bb->count &&
 281             badblocks_check(bb, sector, num, &first_bad, &num_bad)) {
 282                 long actual_nr;
 283
 284                 if (mode != DAX_RECOVERY_WRITE)
 285                         return -EIO;
 286
 287                 /*
 288                  * Set the recovery stride is set to kernel page size because
 289                  * the underlying driver and firmware clear poison functions
 290                  * don't appear to handle large chunk(such as 2MiB) reliably.
 291                  */
 292                 actual_nr = PHYS_PFN(
 293                         PAGE_ALIGN((first_bad - sector) << SECTOR_SHIFT));
 294                 dev_dbg(pmem->bb.dev, "start sector(%llu), nr_pages(%ld), first_bad(%llu), actual_nr(%ld)\n",
 295                                 sector, nr_pages, first_bad, actual_nr);
 296                 if (actual_nr)
 297                         return actual_nr;
 298                 return 1;
 299         }
 300
 301         /*
 302          * If badblocks are present but not in the range, limit known good range
 303          * to the requested range.
 304          */
 305         if (bb->count)
 306                 return nr_pages;
 307         return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
 308 }
 309
 310 static const struct block_device_operations pmem_fops = {
 311         .owner =                THIS_MODULE,
 312         .submit_bio =           pmem_submit_bio,
 313         .rw_page =              pmem_rw_page,
 314 };
 315
 316 static int pmem_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
 317                                     size_t nr_pages)
 318 {
 319         struct pmem_device *pmem = dax_get_private(dax_dev);
 320
 321         return blk_status_to_errno(pmem_do_write(pmem, ZERO_PAGE(0), 0,
 322                                    PFN_PHYS(pgoff) >> SECTOR_SHIFT,
 323                                    PAGE_SIZE));
 324 }
 325
 326 static long pmem_dax_direct_access(struct dax_device *dax_dev,
 327                 pgoff_t pgoff, long nr_pages, enum dax_access_mode mode,
 328                 void **kaddr, pfn_t *pfn)
 329 {
 330         struct pmem_device *pmem = dax_get_private(dax_dev);
 331
 332         return __pmem_direct_access(pmem, pgoff, nr_pages, mode, kaddr, pfn);
 333 }
 334
 335 /*
 336  * The recovery write thread started out as a normal pwrite thread and
 337  * when the filesystem was told about potential media error in the
 338  * range, filesystem turns the normal pwrite to a dax_recovery_write.
 339  *
 340  * The recovery write consists of clearing media poison, clearing page
 341  * HWPoison bit, reenable page-wide read-write permission, flush the
 342  * caches and finally write.  A competing pread thread will be held
 343  * off during the recovery process since data read back might not be
 344  * valid, and this is achieved by clearing the badblock records after
 345  * the recovery write is complete. Competing recovery write threads
 346  * are already serialized by writer lock held by dax_iomap_rw().
 347  */
 348 static size_t pmem_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
 349                 void *addr, size_t bytes, struct iov_iter *i)
 350 {
 351         struct pmem_device *pmem = dax_get_private(dax_dev);
 352         size_t olen, len, off;
 353         phys_addr_t pmem_off;
 354         struct device *dev = pmem->bb.dev;
 355         long cleared;
 356
 357         off = offset_in_page(addr);
 358         len = PFN_PHYS(PFN_UP(off + bytes));
 359         if (!is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) >> SECTOR_SHIFT, len))
 360                 return _copy_from_iter_flushcache(addr, bytes, i);
 361
 362         /*
 363          * Not page-aligned range cannot be recovered. This should not
 364          * happen unless something else went wrong.
 365          */
 366         if (off || !PAGE_ALIGNED(bytes)) {
 367                 dev_dbg(dev, "Found poison, but addr(%p) or bytes(%#zx) not page aligned\n",
 368                         addr, bytes);
 369                 return 0;
 370         }
 371
 372         pmem_off = PFN_PHYS(pgoff) + pmem->data_offset;
 373         cleared = __pmem_clear_poison(pmem, pmem_off, len);
 374         if (cleared > 0 && cleared < len) {
 375                 dev_dbg(dev, "poison cleared only %ld out of %zu bytes\n",
 376                         cleared, len);
 377                 return 0;
 378         }
 379         if (cleared < 0) {
 380                 dev_dbg(dev, "poison clear failed: %ld\n", cleared);
 381                 return 0;
 382         }
 383
 384         olen = _copy_from_iter_flushcache(addr, bytes, i);
 385         pmem_clear_bb(pmem, to_sect(pmem, pmem_off), cleared >> SECTOR_SHIFT);
 386
 387         return olen;
 388 }
 389
 390 static const struct dax_operations pmem_dax_ops = {
 391         .direct_access = pmem_dax_direct_access,
 392         .zero_page_range = pmem_dax_zero_page_range,
 393         .recovery_write = pmem_recovery_write,
 394 };
 395
 396 static ssize_t write_cache_show(struct device *dev,
 397                 struct device_attribute *attr, char *buf)
 398 {
 399         struct pmem_device *pmem = dev_to_disk(dev)->private_data;
 400
 401         return sprintf(buf, "%d\n", !!dax_write_cache_enabled(pmem->dax_dev));
 402 }
 403
 404 static ssize_t write_cache_store(struct device *dev,
 405                 struct device_attribute *attr, const char *buf, size_t len)
 406 {
 407         struct pmem_device *pmem = dev_to_disk(dev)->private_data;
 408         bool write_cache;
 409         int rc;
 410
 411         rc = strtobool(buf, &write_cache);
 412         if (rc)
 413                 return rc;
 414         dax_write_cache(pmem->dax_dev, write_cache);
 415         return len;
 416 }
 417 static DEVICE_ATTR_RW(write_cache);
 418
 419 static umode_t dax_visible(struct kobject *kobj, struct attribute *a, int n)
 420 {
 421 #ifndef CONFIG_ARCH_HAS_PMEM_API
 422         if (a == &dev_attr_write_cache.attr)
 423                 return 0;
 424 #endif
 425         return a->mode;
 426 }
 427
 428 static struct attribute *dax_attributes[] = {
 429         &dev_attr_write_cache.attr,
 430         NULL,
 431 };
 432
 433 static const struct attribute_group dax_attribute_group = {
 434         .name           = "dax",
 435         .attrs          = dax_attributes,
 436         .is_visible     = dax_visible,
 437 };
 438
 439 static const struct attribute_group *pmem_attribute_groups[] = {
 440         &dax_attribute_group,
 441         NULL,
 442 };
 443
 444 static void pmem_release_disk(void *__pmem)
 445 {
 446         struct pmem_device *pmem = __pmem;
 447
 448         dax_remove_host(pmem->disk);
 449         kill_dax(pmem->dax_dev);
 450         put_dax(pmem->dax_dev);
 451         del_gendisk(pmem->disk);
 452
 453         put_disk(pmem->disk);
 454 }
 455
 456 static int pmem_pagemap_memory_failure(struct dev_pagemap *pgmap,
 457                 unsigned long pfn, unsigned long nr_pages, int mf_flags)
 458 {
 459         struct pmem_device *pmem =
 460                         container_of(pgmap, struct pmem_device, pgmap);
 461         u64 offset = PFN_PHYS(pfn) - pmem->phys_addr - pmem->data_offset;
 462         u64 len = nr_pages << PAGE_SHIFT;
 463
 464         return dax_holder_notify_failure(pmem->dax_dev, offset, len, mf_flags);
 465 }
 466
 467 static const struct dev_pagemap_ops fsdax_pagemap_ops = {
 468         .memory_failure         = pmem_pagemap_memory_failure,
 469 };
 470
 471 static int pmem_attach_disk(struct device *dev,
 472                 struct nd_namespace_common *ndns)
 473 {
 474         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 475         struct nd_region *nd_region = to_nd_region(dev->parent);
 476         int nid = dev_to_node(dev), fua;
 477         struct resource *res = &nsio->res;
 478         struct range bb_range;
 479         struct nd_pfn *nd_pfn = NULL;
 480         struct dax_device *dax_dev;
 481         struct nd_pfn_sb *pfn_sb;
 482         struct pmem_device *pmem;
 483         struct request_queue *q;
 484         struct gendisk *disk;
 485         void *addr;
 486         int rc;
 487
 488         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 489         if (!pmem)
 490                 return -ENOMEM;
 491
 492         rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
 493         if (rc)
 494                 return rc;
 495
 496         /* while nsio_rw_bytes is active, parse a pfn info block if present */
 497         if (is_nd_pfn(dev)) {
 498                 nd_pfn = to_nd_pfn(dev);
 499                 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
 500                 if (rc)
 501                         return rc;
 502         }
 503
 504         /* we're attaching a block device, disable raw namespace access */
 505         devm_namespace_disable(dev, ndns);
 506
 507         dev_set_drvdata(dev, pmem);
 508         pmem->phys_addr = res->start;
 509         pmem->size = resource_size(res);
 510         fua = nvdimm_has_flush(nd_region);
 511         if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
 512                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 513                 fua = 0;
 514         }
 515
 516         if (!devm_request_mem_region(dev, res->start, resource_size(res),
 517                                 dev_name(&ndns->dev))) {
 518                 dev_warn(dev, "could not reserve region %pR\n", res);
 519                 return -EBUSY;
 520         }
 521
 522         disk = blk_alloc_disk(nid);
 523         if (!disk)
 524                 return -ENOMEM;
 525         q = disk->queue;
 526
 527         pmem->disk = disk;
 528         pmem->pgmap.owner = pmem;
 529         pmem->pfn_flags = PFN_DEV;
 530         if (is_nd_pfn(dev)) {
 531                 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
 532                 pmem->pgmap.ops = &fsdax_pagemap_ops;
 533                 addr = devm_memremap_pages(dev, &pmem->pgmap);
 534                 pfn_sb = nd_pfn->pfn_sb;
 535                 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
 536                 pmem->pfn_pad = resource_size(res) -
 537                         range_len(&pmem->pgmap.range);
 538                 pmem->pfn_flags |= PFN_MAP;
 539                 bb_range = pmem->pgmap.range;
 540                 bb_range.start += pmem->data_offset;
 541         } else if (pmem_should_map_pages(dev)) {
 542                 pmem->pgmap.range.start = res->start;
 543                 pmem->pgmap.range.end = res->end;
 544                 pmem->pgmap.nr_range = 1;
 545                 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
 546                 pmem->pgmap.ops = &fsdax_pagemap_ops;
 547                 addr = devm_memremap_pages(dev, &pmem->pgmap);
 548                 pmem->pfn_flags |= PFN_MAP;
 549                 bb_range = pmem->pgmap.range;
 550         } else {
 551                 addr = devm_memremap(dev, pmem->phys_addr,
 552                                 pmem->size, ARCH_MEMREMAP_PMEM);
 553                 bb_range.start =  res->start;
 554                 bb_range.end = res->end;
 555         }
 556
 557         if (IS_ERR(addr)) {
 558                 rc = PTR_ERR(addr);
 559                 goto out;
 560         }
 561         pmem->virt_addr = addr;
 562
 563         blk_queue_write_cache(q, true, fua);
 564         blk_queue_physical_block_size(q, PAGE_SIZE);
 565         blk_queue_logical_block_size(q, pmem_sector_size(ndns));
 566         blk_queue_max_hw_sectors(q, UINT_MAX);
 567         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 568         if (pmem->pfn_flags & PFN_MAP)
 569                 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
 570
 571         disk->fops              = &pmem_fops;
 572         disk->private_data      = pmem;
 573         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 574         set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
 575                         / 512);
 576         if (devm_init_badblocks(dev, &pmem->bb))
 577                 return -ENOMEM;
 578         nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_range);
 579         disk->bb = &pmem->bb;
 580
 581         dax_dev = alloc_dax(pmem, &pmem_dax_ops);
 582         if (IS_ERR(dax_dev)) {
 583                 rc = PTR_ERR(dax_dev);
 584                 goto out;
 585         }
 586         set_dax_nocache(dax_dev);
 587         set_dax_nomc(dax_dev);
 588         if (is_nvdimm_sync(nd_region))
 589                 set_dax_synchronous(dax_dev);
 590         rc = dax_add_host(dax_dev, disk);
 591         if (rc)
 592                 goto out_cleanup_dax;
 593         dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
 594         pmem->dax_dev = dax_dev;
 595
 596         rc = device_add_disk(dev, disk, pmem_attribute_groups);
 597         if (rc)
 598                 goto out_remove_host;
 599         if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
 600                 return -ENOMEM;
 601
 602         nvdimm_check_and_set_ro(disk);
 603
 604         pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
 605                                           "badblocks");
 606         if (!pmem->bb_state)
 607                 dev_warn(dev, "'badblocks' notification disabled\n");
 608         return 0;
 609
 610 out_remove_host:
 611         dax_remove_host(pmem->disk);
 612 out_cleanup_dax:
 613         kill_dax(pmem->dax_dev);
 614         put_dax(pmem->dax_dev);
 615 out:
 616         put_disk(pmem->disk);
 617         return rc;
 618 }
 619
 620 static int nd_pmem_probe(struct device *dev)
 621 {
 622         int ret;
 623         struct nd_namespace_common *ndns;
 624
 625         ndns = nvdimm_namespace_common_probe(dev);
 626         if (IS_ERR(ndns))
 627                 return PTR_ERR(ndns);
 628
 629         if (is_nd_btt(dev))
 630                 return nvdimm_namespace_attach_btt(ndns);
 631
 632         if (is_nd_pfn(dev))
 633                 return pmem_attach_disk(dev, ndns);
 634
 635         ret = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
 636         if (ret)
 637                 return ret;
 638
 639         ret = nd_btt_probe(dev, ndns);
 640         if (ret == 0)
 641                 return -ENXIO;
 642
 643         /*
 644          * We have two failure conditions here, there is no
 645          * info reserver block or we found a valid info reserve block
 646          * but failed to initialize the pfn superblock.
 647          *
 648          * For the first case consider namespace as a raw pmem namespace
 649          * and attach a disk.
 650          *
 651          * For the latter, consider this a success and advance the namespace
 652          * seed.
 653          */
 654         ret = nd_pfn_probe(dev, ndns);
 655         if (ret == 0)
 656                 return -ENXIO;
 657         else if (ret == -EOPNOTSUPP)
 658                 return ret;
 659
 660         ret = nd_dax_probe(dev, ndns);
 661         if (ret == 0)
 662                 return -ENXIO;
 663         else if (ret == -EOPNOTSUPP)
 664                 return ret;
 665
 666         /* probe complete, attach handles namespace enabling */
 667         devm_namespace_disable(dev, ndns);
 668
 669         return pmem_attach_disk(dev, ndns);
 670 }
 671
 672 static void nd_pmem_remove(struct device *dev)
 673 {
 674         struct pmem_device *pmem = dev_get_drvdata(dev);
 675
 676         if (is_nd_btt(dev))
 677                 nvdimm_namespace_detach_btt(to_nd_btt(dev));
 678         else {
 679                 /*
 680                  * Note, this assumes device_lock() context to not
 681                  * race nd_pmem_notify()
 682                  */
 683                 sysfs_put(pmem->bb_state);
 684                 pmem->bb_state = NULL;
 685         }
 686         nvdimm_flush(to_nd_region(dev->parent), NULL);
 687 }
 688
 689 static void nd_pmem_shutdown(struct device *dev)
 690 {
 691         nvdimm_flush(to_nd_region(dev->parent), NULL);
 692 }
 693
 694 static void pmem_revalidate_poison(struct device *dev)
 695 {
 696         struct nd_region *nd_region;
 697         resource_size_t offset = 0, end_trunc = 0;
 698         struct nd_namespace_common *ndns;
 699         struct nd_namespace_io *nsio;
 700         struct badblocks *bb;
 701         struct range range;
 702         struct kernfs_node *bb_state;
 703
 704         if (is_nd_btt(dev)) {
 705                 struct nd_btt *nd_btt = to_nd_btt(dev);
 706
 707                 ndns = nd_btt->ndns;
 708                 nd_region = to_nd_region(ndns->dev.parent);
 709                 nsio = to_nd_namespace_io(&ndns->dev);
 710                 bb = &nsio->bb;
 711                 bb_state = NULL;
 712         } else {
 713                 struct pmem_device *pmem = dev_get_drvdata(dev);
 714
 715                 nd_region = to_region(pmem);
 716                 bb = &pmem->bb;
 717                 bb_state = pmem->bb_state;
 718
 719                 if (is_nd_pfn(dev)) {
 720                         struct nd_pfn *nd_pfn = to_nd_pfn(dev);
 721                         struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 722
 723                         ndns = nd_pfn->ndns;
 724                         offset = pmem->data_offset +
 725                                         __le32_to_cpu(pfn_sb->start_pad);
 726                         end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
 727                 } else {
 728                         ndns = to_ndns(dev);
 729                 }
 730
 731                 nsio = to_nd_namespace_io(&ndns->dev);
 732         }
 733
 734         range.start = nsio->res.start + offset;
 735         range.end = nsio->res.end - end_trunc;
 736         nvdimm_badblocks_populate(nd_region, bb, &range);
 737         if (bb_state)
 738                 sysfs_notify_dirent(bb_state);
 739 }
 740
 741 static void pmem_revalidate_region(struct device *dev)
 742 {
 743         struct pmem_device *pmem;
 744
 745         if (is_nd_btt(dev)) {
 746                 struct nd_btt *nd_btt = to_nd_btt(dev);
 747                 struct btt *btt = nd_btt->btt;
 748
 749                 nvdimm_check_and_set_ro(btt->btt_disk);
 750                 return;
 751         }
 752
 753         pmem = dev_get_drvdata(dev);
 754         nvdimm_check_and_set_ro(pmem->disk);
 755 }
 756
 757 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
 758 {
 759         switch (event) {
 760         case NVDIMM_REVALIDATE_POISON:
 761                 pmem_revalidate_poison(dev);
 762                 break;
 763         case NVDIMM_REVALIDATE_REGION:
 764                 pmem_revalidate_region(dev);
 765                 break;
 766         default:
 767                 dev_WARN_ONCE(dev, 1, "notify: unknown event: %d\n", event);
 768                 break;
 769         }
 770 }
 771
 772 MODULE_ALIAS("pmem");
 773 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 774 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 775 static struct nd_device_driver nd_pmem_driver = {
 776         .probe = nd_pmem_probe,
 777         .remove = nd_pmem_remove,
 778         .notify = nd_pmem_notify,
 779         .shutdown = nd_pmem_shutdown,
 780         .drv = {
 781                 .name = "nd_pmem",
 782         },
 783         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 784 };
 785
 786 module_nd_driver(nd_pmem_driver);
 787
 788 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 789 MODULE_LICENSE("GPL v2");