drivers/nvdimm/pmem.c

   1 /*
   2  * Persistent Memory Driver
   3  *
   4  * Copyright (c) 2014-2015, Intel Corporation.
   5  * Copyright (c) 2015, Christoph Hellwig <hch@lst.de>.
   6  * Copyright (c) 2015, Boaz Harrosh <boaz@plexistor.com>.
   7  *
   8  * This program is free software; you can redistribute it and/or modify it
   9  * under the terms and conditions of the GNU General Public License,
  10  * version 2, as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope it will be useful, but WITHOUT
  13  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  14  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  15  * more details.
  16  */
  17
  18 #include <asm/cacheflush.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/hdreg.h>
  21 #include <linux/init.h>
  22 #include <linux/platform_device.h>
  23 #include <linux/module.h>
  24 #include <linux/moduleparam.h>
  25 #include <linux/badblocks.h>
  26 #include <linux/memremap.h>
  27 #include <linux/vmalloc.h>
  28 #include <linux/blk-mq.h>
  29 #include <linux/pfn_t.h>
  30 #include <linux/slab.h>
  31 #include <linux/uio.h>
  32 #include <linux/dax.h>
  33 #include <linux/nd.h>
  34 #include "pmem.h"
  35 #include "pfn.h"
  36 #include "nd.h"
  37
  38 static struct device *to_dev(struct pmem_device *pmem)
  39 {
  40         /*
  41          * nvdimm bus services need a 'dev' parameter, and we record the device
  42          * at init in bb.dev.
  43          */
  44         return pmem->bb.dev;
  45 }
  46
  47 static struct nd_region *to_region(struct pmem_device *pmem)
  48 {
  49         return to_nd_region(to_dev(pmem)->parent);
  50 }
  51
  52 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
  53                 phys_addr_t offset, unsigned int len)
  54 {
  55         struct device *dev = to_dev(pmem);
  56         sector_t sector;
  57         long cleared;
  58         blk_status_t rc = BLK_STS_OK;
  59
  60         sector = (offset - pmem->data_offset) / 512;
  61
  62         cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
  63         if (cleared < len)
  64                 rc = BLK_STS_IOERR;
  65         if (cleared > 0 && cleared / 512) {
  66                 cleared /= 512;
  67                 dev_dbg(dev, "%s: %#llx clear %ld sector%s\n", __func__,
  68                                 (unsigned long long) sector, cleared,
  69                                 cleared > 1 ? "s" : "");
  70                 badblocks_clear(&pmem->bb, sector, cleared);
  71                 if (pmem->bb_state)
  72                         sysfs_notify_dirent(pmem->bb_state);
  73         }
  74
  75         arch_invalidate_pmem(pmem->virt_addr + offset, len);
  76
  77         return rc;
  78 }
  79
  80 static void write_pmem(void *pmem_addr, struct page *page,
  81                 unsigned int off, unsigned int len)
  82 {
  83         unsigned int chunk;
  84         void *mem;
  85
  86         while (len) {
  87                 mem = kmap_atomic(page);
  88                 chunk = min_t(unsigned int, len, PAGE_SIZE);
  89                 memcpy_flushcache(pmem_addr, mem + off, chunk);
  90                 kunmap_atomic(mem);
  91                 len -= chunk;
  92                 off = 0;
  93                 page++;
  94                 pmem_addr += PAGE_SIZE;
  95         }
  96 }
  97
  98 static blk_status_t read_pmem(struct page *page, unsigned int off,
  99                 void *pmem_addr, unsigned int len)
 100 {
 101         unsigned int chunk;
 102         int rc;
 103         void *mem;
 104
 105         while (len) {
 106                 mem = kmap_atomic(page);
 107                 chunk = min_t(unsigned int, len, PAGE_SIZE);
 108                 rc = memcpy_mcsafe(mem + off, pmem_addr, chunk);
 109                 kunmap_atomic(mem);
 110                 if (rc)
 111                         return BLK_STS_IOERR;
 112                 len -= chunk;
 113                 off = 0;
 114                 page++;
 115                 pmem_addr += PAGE_SIZE;
 116         }
 117         return BLK_STS_OK;
 118 }
 119
 120 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
 121                         unsigned int len, unsigned int off, bool is_write,
 122                         sector_t sector)
 123 {
 124         blk_status_t rc = BLK_STS_OK;
 125         bool bad_pmem = false;
 126         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
 127         void *pmem_addr = pmem->virt_addr + pmem_off;
 128
 129         if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
 130                 bad_pmem = true;
 131
 132         if (!is_write) {
 133                 if (unlikely(bad_pmem))
 134                         rc = BLK_STS_IOERR;
 135                 else {
 136                         rc = read_pmem(page, off, pmem_addr, len);
 137                         flush_dcache_page(page);
 138                 }
 139         } else {
 140                 /*
 141                  * Note that we write the data both before and after
 142                  * clearing poison.  The write before clear poison
 143                  * handles situations where the latest written data is
 144                  * preserved and the clear poison operation simply marks
 145                  * the address range as valid without changing the data.
 146                  * In this case application software can assume that an
 147                  * interrupted write will either return the new good
 148                  * data or an error.
 149                  *
 150                  * However, if pmem_clear_poison() leaves the data in an
 151                  * indeterminate state we need to perform the write
 152                  * after clear poison.
 153                  */
 154                 flush_dcache_page(page);
 155                 write_pmem(pmem_addr, page, off, len);
 156                 if (unlikely(bad_pmem)) {
 157                         rc = pmem_clear_poison(pmem, pmem_off, len);
 158                         write_pmem(pmem_addr, page, off, len);
 159                 }
 160         }
 161
 162         return rc;
 163 }
 164
 165 /* account for REQ_FLUSH rename, replace with REQ_PREFLUSH after v4.8-rc1 */
 166 #ifndef REQ_FLUSH
 167 #define REQ_FLUSH REQ_PREFLUSH
 168 #endif
 169
 170 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
 171 {
 172         blk_status_t rc = 0;
 173         bool do_acct;
 174         unsigned long start;
 175         struct bio_vec bvec;
 176         struct bvec_iter iter;
 177         struct pmem_device *pmem = q->queuedata;
 178         struct nd_region *nd_region = to_region(pmem);
 179
 180         if (bio->bi_opf & REQ_FLUSH)
 181                 nvdimm_flush(nd_region);
 182
 183         do_acct = nd_iostat_start(bio, &start);
 184         bio_for_each_segment(bvec, bio, iter) {
 185                 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
 186                                 bvec.bv_offset, op_is_write(bio_op(bio)),
 187                                 iter.bi_sector);
 188                 if (rc) {
 189                         bio->bi_status = rc;
 190                         break;
 191                 }
 192         }
 193         if (do_acct)
 194                 nd_iostat_end(bio, start);
 195
 196         if (bio->bi_opf & REQ_FUA)
 197                 nvdimm_flush(nd_region);
 198
 199         bio_endio(bio);
 200         return BLK_QC_T_NONE;
 201 }
 202
 203 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 204                        struct page *page, bool is_write)
 205 {
 206         struct pmem_device *pmem = bdev->bd_queue->queuedata;
 207         blk_status_t rc;
 208
 209         rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
 210                           0, is_write, sector);
 211
 212         /*
 213          * The ->rw_page interface is subtle and tricky.  The core
 214          * retries on any error, so we can only invoke page_endio() in
 215          * the successful completion case.  Otherwise, we'll see crashes
 216          * caused by double completion.
 217          */
 218         if (rc == 0)
 219                 page_endio(page, is_write, 0);
 220
 221         return blk_status_to_errno(rc);
 222 }
 223
 224 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
 225 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
 226                 long nr_pages, void **kaddr, pfn_t *pfn)
 227 {
 228         resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
 229
 230         if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
 231                                         PFN_PHYS(nr_pages))))
 232                 return -EIO;
 233         *kaddr = pmem->virt_addr + offset;
 234         *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 235
 236         /*
 237          * If badblocks are present, limit known good range to the
 238          * requested range.
 239          */
 240         if (unlikely(pmem->bb.count))
 241                 return nr_pages;
 242         return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
 243 }
 244
 245 static const struct block_device_operations pmem_fops = {
 246         .owner =                THIS_MODULE,
 247         .rw_page =              pmem_rw_page,
 248         .revalidate_disk =      nvdimm_revalidate_disk,
 249 };
 250
 251 static long pmem_dax_direct_access(struct dax_device *dax_dev,
 252                 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
 253 {
 254         struct pmem_device *pmem = dax_get_private(dax_dev);
 255
 256         return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
 257 }
 258
 259 /*
 260  * Use the 'no check' versions of copy_from_iter_flushcache() and
 261  * copy_to_iter_mcsafe() to bypass HARDENED_USERCOPY overhead. Bounds
 262  * checking, both file offset and device offset, is handled by
 263  * dax_iomap_actor()
 264  */
 265 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
 266                 void *addr, size_t bytes, struct iov_iter *i)
 267 {
 268         return _copy_from_iter_flushcache(addr, bytes, i);
 269 }
 270
 271 static const struct dax_operations pmem_dax_ops = {
 272         .direct_access = pmem_dax_direct_access,
 273         .copy_from_iter = pmem_copy_from_iter,
 274 };
 275
 276 static const struct attribute_group *pmem_attribute_groups[] = {
 277         &dax_attribute_group,
 278         NULL,
 279 };
 280
 281 static void pmem_release_queue(void *q)
 282 {
 283         blk_cleanup_queue(q);
 284 }
 285
 286 static void pmem_freeze_queue(void *q)
 287 {
 288         blk_freeze_queue_start(q);
 289 }
 290
 291 static void pmem_release_disk(void *__pmem)
 292 {
 293         struct pmem_device *pmem = __pmem;
 294
 295         kill_dax(pmem->dax_dev);
 296         put_dax(pmem->dax_dev);
 297         del_gendisk(pmem->disk);
 298         put_disk(pmem->disk);
 299 }
 300
 301 static int pmem_attach_disk(struct device *dev,
 302                 struct nd_namespace_common *ndns)
 303 {
 304         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 305         struct nd_region *nd_region = to_nd_region(dev->parent);
 306         struct vmem_altmap __altmap, *altmap = NULL;
 307         int nid = dev_to_node(dev), fua, wbc;
 308         struct resource *res = &nsio->res;
 309         struct nd_pfn *nd_pfn = NULL;
 310         struct dax_device *dax_dev;
 311         struct nd_pfn_sb *pfn_sb;
 312         struct pmem_device *pmem;
 313         struct resource pfn_res;
 314         struct request_queue *q;
 315         struct device *gendev;
 316         struct gendisk *disk;
 317         void *addr;
 318
 319         /* while nsio_rw_bytes is active, parse a pfn info block if present */
 320         if (is_nd_pfn(dev)) {
 321                 nd_pfn = to_nd_pfn(dev);
 322                 altmap = nvdimm_setup_pfn(nd_pfn, &pfn_res, &__altmap);
 323                 if (IS_ERR(altmap))
 324                         return PTR_ERR(altmap);
 325         }
 326
 327         /* we're attaching a block device, disable raw namespace access */
 328         devm_nsio_disable(dev, nsio);
 329
 330         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 331         if (!pmem)
 332                 return -ENOMEM;
 333
 334         dev_set_drvdata(dev, pmem);
 335         pmem->phys_addr = res->start;
 336         pmem->size = resource_size(res);
 337         fua = nvdimm_has_flush(nd_region);
 338         if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
 339                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 340                 fua = 0;
 341         }
 342         wbc = nvdimm_has_cache(nd_region);
 343
 344         if (!devm_request_mem_region(dev, res->start, resource_size(res),
 345                                 dev_name(&ndns->dev))) {
 346                 dev_warn(dev, "could not reserve region %pR\n", res);
 347                 return -EBUSY;
 348         }
 349
 350         q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
 351         if (!q)
 352                 return -ENOMEM;
 353
 354         if (devm_add_action_or_reset(dev, pmem_release_queue, q))
 355                 return -ENOMEM;
 356
 357         pmem->pfn_flags = PFN_DEV;
 358         if (is_nd_pfn(dev)) {
 359                 addr = devm_memremap_pages(dev, &pfn_res, &q->q_usage_counter,
 360                                 altmap);
 361                 pfn_sb = nd_pfn->pfn_sb;
 362                 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
 363                 pmem->pfn_pad = resource_size(res) - resource_size(&pfn_res);
 364                 pmem->pfn_flags |= PFN_MAP;
 365                 res = &pfn_res; /* for badblocks populate */
 366                 res->start += pmem->data_offset;
 367         } else if (pmem_should_map_pages(dev)) {
 368                 addr = devm_memremap_pages(dev, &nsio->res,
 369                                 &q->q_usage_counter, NULL);
 370                 pmem->pfn_flags |= PFN_MAP;
 371         } else
 372                 addr = devm_memremap(dev, pmem->phys_addr,
 373                                 pmem->size, ARCH_MEMREMAP_PMEM);
 374
 375         /*
 376          * At release time the queue must be frozen before
 377          * devm_memremap_pages is unwound
 378          */
 379         if (devm_add_action_or_reset(dev, pmem_freeze_queue, q))
 380                 return -ENOMEM;
 381
 382         if (IS_ERR(addr))
 383                 return PTR_ERR(addr);
 384         pmem->virt_addr = addr;
 385
 386         blk_queue_write_cache(q, wbc, fua);
 387         blk_queue_make_request(q, pmem_make_request);
 388         blk_queue_physical_block_size(q, PAGE_SIZE);
 389         blk_queue_logical_block_size(q, pmem_sector_size(ndns));
 390         blk_queue_max_hw_sectors(q, UINT_MAX);
 391         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
 392         queue_flag_set_unlocked(QUEUE_FLAG_DAX, q);
 393         q->queuedata = pmem;
 394
 395         disk = alloc_disk_node(0, nid);
 396         if (!disk)
 397                 return -ENOMEM;
 398         pmem->disk = disk;
 399
 400         disk->fops              = &pmem_fops;
 401         disk->queue             = q;
 402         disk->flags             = GENHD_FL_EXT_DEVT;
 403         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 404         set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
 405                         / 512);
 406         if (devm_init_badblocks(dev, &pmem->bb))
 407                 return -ENOMEM;
 408         nvdimm_badblocks_populate(nd_region, &pmem->bb, res);
 409         disk->bb = &pmem->bb;
 410
 411         dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops);
 412         if (!dax_dev) {
 413                 put_disk(disk);
 414                 return -ENOMEM;
 415         }
 416         dax_write_cache(dax_dev, wbc);
 417         pmem->dax_dev = dax_dev;
 418
 419         gendev = disk_to_dev(disk);
 420         gendev->groups = pmem_attribute_groups;
 421
 422         device_add_disk(dev, disk);
 423         if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
 424                 return -ENOMEM;
 425
 426         revalidate_disk(disk);
 427
 428         pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
 429                                           "badblocks");
 430         if (!pmem->bb_state)
 431                 dev_warn(dev, "'badblocks' notification disabled\n");
 432
 433         return 0;
 434 }
 435
 436 static int nd_pmem_probe(struct device *dev)
 437 {
 438         struct nd_namespace_common *ndns;
 439
 440         ndns = nvdimm_namespace_common_probe(dev);
 441         if (IS_ERR(ndns))
 442                 return PTR_ERR(ndns);
 443
 444         if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
 445                 return -ENXIO;
 446
 447         if (is_nd_btt(dev))
 448                 return nvdimm_namespace_attach_btt(ndns);
 449
 450         if (is_nd_pfn(dev))
 451                 return pmem_attach_disk(dev, ndns);
 452
 453         /* if we find a valid info-block we'll come back as that personality */
 454         if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
 455                         || nd_dax_probe(dev, ndns) == 0)
 456                 return -ENXIO;
 457
 458         /* ...otherwise we're just a raw pmem device */
 459         return pmem_attach_disk(dev, ndns);
 460 }
 461
 462 static int nd_pmem_remove(struct device *dev)
 463 {
 464         struct pmem_device *pmem = dev_get_drvdata(dev);
 465
 466         if (is_nd_btt(dev))
 467                 nvdimm_namespace_detach_btt(to_nd_btt(dev));
 468         else {
 469                 /*
 470                  * Note, this assumes device_lock() context to not race
 471                  * nd_pmem_notify()
 472                  */
 473                 sysfs_put(pmem->bb_state);
 474                 pmem->bb_state = NULL;
 475         }
 476         nvdimm_flush(to_nd_region(dev->parent));
 477
 478         return 0;
 479 }
 480
 481 static void nd_pmem_shutdown(struct device *dev)
 482 {
 483         nvdimm_flush(to_nd_region(dev->parent));
 484 }
 485
 486 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
 487 {
 488         struct nd_region *nd_region;
 489         resource_size_t offset = 0, end_trunc = 0;
 490         struct nd_namespace_common *ndns;
 491         struct nd_namespace_io *nsio;
 492         struct resource res;
 493         struct badblocks *bb;
 494         struct kernfs_node *bb_state;
 495
 496         if (event != NVDIMM_REVALIDATE_POISON)
 497                 return;
 498
 499         if (is_nd_btt(dev)) {
 500                 struct nd_btt *nd_btt = to_nd_btt(dev);
 501
 502                 ndns = nd_btt->ndns;
 503                 nd_region = to_nd_region(ndns->dev.parent);
 504                 nsio = to_nd_namespace_io(&ndns->dev);
 505                 bb = &nsio->bb;
 506                 bb_state = NULL;
 507         } else {
 508                 struct pmem_device *pmem = dev_get_drvdata(dev);
 509
 510                 nd_region = to_region(pmem);
 511                 bb = &pmem->bb;
 512                 bb_state = pmem->bb_state;
 513
 514                 if (is_nd_pfn(dev)) {
 515                         struct nd_pfn *nd_pfn = to_nd_pfn(dev);
 516                         struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 517
 518                         ndns = nd_pfn->ndns;
 519                         offset = pmem->data_offset +
 520                                         __le32_to_cpu(pfn_sb->start_pad);
 521                         end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
 522                 } else {
 523                         ndns = to_ndns(dev);
 524                 }
 525
 526                 nsio = to_nd_namespace_io(&ndns->dev);
 527         }
 528
 529         res.start = nsio->res.start + offset;
 530         res.end = nsio->res.end - end_trunc;
 531         nvdimm_badblocks_populate(nd_region, bb, &res);
 532         if (bb_state)
 533                 sysfs_notify_dirent(bb_state);
 534 }
 535
 536 MODULE_ALIAS("pmem");
 537 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 538 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 539 static struct nd_device_driver nd_pmem_driver = {
 540         .probe = nd_pmem_probe,
 541         .remove = nd_pmem_remove,
 542         .notify = nd_pmem_notify,
 543         .shutdown = nd_pmem_shutdown,
 544         .drv = {
 545                 .name = "nd_pmem",
 546         },
 547         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 548 };
 549
 550 static int __init pmem_init(void)
 551 {
 552         return nd_driver_register(&nd_pmem_driver);
 553 }
 554 module_init(pmem_init);
 555
 556 static void pmem_exit(void)
 557 {
 558         driver_unregister(&nd_pmem_driver.drv);
 559 }
 560 module_exit(pmem_exit);
 561
 562 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 563 MODULE_LICENSE("GPL v2");