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