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