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 #include "nd-core.h"
  39
  40 static struct device *to_dev(struct pmem_device *pmem)
  41 {
  42         /*
  43          * nvdimm bus services need a 'dev' parameter, and we record the device
  44          * at init in bb.dev.
  45          */
  46         return pmem->bb.dev;
  47 }
  48
  49 static struct nd_region *to_region(struct pmem_device *pmem)
  50 {
  51         return to_nd_region(to_dev(pmem)->parent);
  52 }
  53
  54 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
  55                 phys_addr_t offset, unsigned int len)
  56 {
  57         struct device *dev = to_dev(pmem);
  58         sector_t sector;
  59         long cleared;
  60         blk_status_t rc = BLK_STS_OK;
  61
  62         sector = (offset - pmem->data_offset) / 512;
  63
  64         cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
  65         if (cleared < len)
  66                 rc = BLK_STS_IOERR;
  67         if (cleared > 0 && cleared / 512) {
  68                 cleared /= 512;
  69                 dev_dbg(dev, "%#llx clear %ld sector%s\n",
  70                                 (unsigned long long) sector, cleared,
  71                                 cleared > 1 ? "s" : "");
  72                 badblocks_clear(&pmem->bb, sector, cleared);
  73                 if (pmem->bb_state)
  74                         sysfs_notify_dirent(pmem->bb_state);
  75         }
  76
  77         arch_invalidate_pmem(pmem->virt_addr + offset, len);
  78
  79         return rc;
  80 }
  81
  82 static void write_pmem(void *pmem_addr, struct page *page,
  83                 unsigned int off, unsigned int len)
  84 {
  85         unsigned int chunk;
  86         void *mem;
  87
  88         while (len) {
  89                 mem = kmap_atomic(page);
  90                 chunk = min_t(unsigned int, len, PAGE_SIZE);
  91                 memcpy_flushcache(pmem_addr, mem + off, chunk);
  92                 kunmap_atomic(mem);
  93                 len -= chunk;
  94                 off = 0;
  95                 page++;
  96                 pmem_addr += PAGE_SIZE;
  97         }
  98 }
  99
 100 static blk_status_t read_pmem(struct page *page, unsigned int off,
 101                 void *pmem_addr, unsigned int len)
 102 {
 103         unsigned int chunk;
 104         unsigned long rem;
 105         void *mem;
 106
 107         while (len) {
 108                 mem = kmap_atomic(page);
 109                 chunk = min_t(unsigned int, len, PAGE_SIZE);
 110                 rem = memcpy_mcsafe(mem + off, pmem_addr, chunk);
 111                 kunmap_atomic(mem);
 112                 if (rem)
 113                         return BLK_STS_IOERR;
 114                 len -= chunk;
 115                 off = 0;
 116                 page++;
 117                 pmem_addr += PAGE_SIZE;
 118         }
 119         return BLK_STS_OK;
 120 }
 121
 122 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
 123                         unsigned int len, unsigned int off, unsigned int op,
 124                         sector_t sector)
 125 {
 126         blk_status_t rc = BLK_STS_OK;
 127         bool bad_pmem = false;
 128         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
 129         void *pmem_addr = pmem->virt_addr + pmem_off;
 130
 131         if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
 132                 bad_pmem = true;
 133
 134         if (!op_is_write(op)) {
 135                 if (unlikely(bad_pmem))
 136                         rc = BLK_STS_IOERR;
 137                 else {
 138                         rc = read_pmem(page, off, pmem_addr, len);
 139                         flush_dcache_page(page);
 140                 }
 141         } else {
 142                 /*
 143                  * Note that we write the data both before and after
 144                  * clearing poison.  The write before clear poison
 145                  * handles situations where the latest written data is
 146                  * preserved and the clear poison operation simply marks
 147                  * the address range as valid without changing the data.
 148                  * In this case application software can assume that an
 149                  * interrupted write will either return the new good
 150                  * data or an error.
 151                  *
 152                  * However, if pmem_clear_poison() leaves the data in an
 153                  * indeterminate state we need to perform the write
 154                  * after clear poison.
 155                  */
 156                 flush_dcache_page(page);
 157                 write_pmem(pmem_addr, page, off, len);
 158                 if (unlikely(bad_pmem)) {
 159                         rc = pmem_clear_poison(pmem, pmem_off, len);
 160                         write_pmem(pmem_addr, page, off, len);
 161                 }
 162         }
 163
 164         return rc;
 165 }
 166
 167 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
 168 {
 169         blk_status_t rc = 0;
 170         bool do_acct;
 171         unsigned long start;
 172         struct bio_vec bvec;
 173         struct bvec_iter iter;
 174         struct pmem_device *pmem = q->queuedata;
 175         struct nd_region *nd_region = to_region(pmem);
 176
 177         if (bio->bi_opf & REQ_PREFLUSH)
 178                 nvdimm_flush(nd_region);
 179
 180         do_acct = nd_iostat_start(bio, &start);
 181         bio_for_each_segment(bvec, bio, iter) {
 182                 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
 183                                 bvec.bv_offset, bio_op(bio), iter.bi_sector);
 184                 if (rc) {
 185                         bio->bi_status = rc;
 186                         break;
 187                 }
 188         }
 189         if (do_acct)
 190                 nd_iostat_end(bio, start);
 191
 192         if (bio->bi_opf & REQ_FUA)
 193                 nvdimm_flush(nd_region);
 194
 195         bio_endio(bio);
 196         return BLK_QC_T_NONE;
 197 }
 198
 199 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 200                        struct page *page, unsigned int op)
 201 {
 202         struct pmem_device *pmem = bdev->bd_queue->queuedata;
 203         blk_status_t rc;
 204
 205         rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
 206                           0, op, sector);
 207
 208         /*
 209          * The ->rw_page interface is subtle and tricky.  The core
 210          * retries on any error, so we can only invoke page_endio() in
 211          * the successful completion case.  Otherwise, we'll see crashes
 212          * caused by double completion.
 213          */
 214         if (rc == 0)
 215                 page_endio(page, op_is_write(op), 0);
 216
 217         return blk_status_to_errno(rc);
 218 }
 219
 220 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
 221 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
 222                 long nr_pages, void **kaddr, pfn_t *pfn)
 223 {
 224         resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
 225
 226         if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
 227                                         PFN_PHYS(nr_pages))))
 228                 return -EIO;
 229
 230         if (kaddr)
 231                 *kaddr = pmem->virt_addr + offset;
 232         if (pfn)
 233                 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 234
 235         /*
 236          * If badblocks are present, limit known good range to the
 237          * requested range.
 238          */
 239         if (unlikely(pmem->bb.count))
 240                 return nr_pages;
 241         return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
 242 }
 243
 244 static const struct block_device_operations pmem_fops = {
 245         .owner =                THIS_MODULE,
 246         .rw_page =              pmem_rw_page,
 247         .revalidate_disk =      nvdimm_revalidate_disk,
 248 };
 249
 250 static long pmem_dax_direct_access(struct dax_device *dax_dev,
 251                 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
 252 {
 253         struct pmem_device *pmem = dax_get_private(dax_dev);
 254
 255         return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
 256 }
 257
 258 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
 259                 void *addr, size_t bytes, struct iov_iter *i)
 260 {
 261         return copy_from_iter_flushcache(addr, bytes, i);
 262 }
 263
 264 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
 265                 void *addr, size_t bytes, struct iov_iter *i)
 266 {
 267         return copy_to_iter_mcsafe(addr, bytes, i);
 268 }
 269
 270 static const struct dax_operations pmem_dax_ops = {
 271         .direct_access = pmem_dax_direct_access,
 272         .copy_from_iter = pmem_copy_from_iter,
 273         .copy_to_iter = pmem_copy_to_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 void pmem_release_pgmap_ops(void *__pgmap)
 302 {
 303         dev_pagemap_put_ops();
 304 }
 305
 306 static void fsdax_pagefree(struct page *page, void *data)
 307 {
 308         wake_up_var(&page->_refcount);
 309 }
 310
 311 static int setup_pagemap_fsdax(struct device *dev, struct dev_pagemap *pgmap)
 312 {
 313         dev_pagemap_get_ops();
 314         if (devm_add_action_or_reset(dev, pmem_release_pgmap_ops, pgmap))
 315                 return -ENOMEM;
 316         pgmap->type = MEMORY_DEVICE_FS_DAX;
 317         pgmap->page_free = fsdax_pagefree;
 318
 319         return 0;
 320 }
 321
 322 static int pmem_attach_disk(struct device *dev,
 323                 struct nd_namespace_common *ndns)
 324 {
 325         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 326         struct nd_region *nd_region = to_nd_region(dev->parent);
 327         int nid = dev_to_node(dev), fua;
 328         struct resource *res = &nsio->res;
 329         struct resource bb_res;
 330         struct nd_pfn *nd_pfn = NULL;
 331         struct dax_device *dax_dev;
 332         struct nd_pfn_sb *pfn_sb;
 333         struct pmem_device *pmem;
 334         struct request_queue *q;
 335         struct device *gendev;
 336         struct gendisk *disk;
 337         void *addr;
 338         int rc;
 339
 340         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 341         if (!pmem)
 342                 return -ENOMEM;
 343
 344         /* while nsio_rw_bytes is active, parse a pfn info block if present */
 345         if (is_nd_pfn(dev)) {
 346                 nd_pfn = to_nd_pfn(dev);
 347                 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
 348                 if (rc)
 349                         return rc;
 350         }
 351
 352         /* we're attaching a block device, disable raw namespace access */
 353         devm_nsio_disable(dev, nsio);
 354
 355         dev_set_drvdata(dev, pmem);
 356         pmem->phys_addr = res->start;
 357         pmem->size = resource_size(res);
 358         fua = nvdimm_has_flush(nd_region);
 359         if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
 360                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 361                 fua = 0;
 362         }
 363
 364         if (!devm_request_mem_region(dev, res->start, resource_size(res),
 365                                 dev_name(&ndns->dev))) {
 366                 dev_warn(dev, "could not reserve region %pR\n", res);
 367                 return -EBUSY;
 368         }
 369
 370         q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev), NULL);
 371         if (!q)
 372                 return -ENOMEM;
 373
 374         if (devm_add_action_or_reset(dev, pmem_release_queue, q))
 375                 return -ENOMEM;
 376
 377         pmem->pfn_flags = PFN_DEV;
 378         pmem->pgmap.ref = &q->q_usage_counter;
 379         if (is_nd_pfn(dev)) {
 380                 if (setup_pagemap_fsdax(dev, &pmem->pgmap))
 381                         return -ENOMEM;
 382                 addr = devm_memremap_pages(dev, &pmem->pgmap);
 383                 pfn_sb = nd_pfn->pfn_sb;
 384                 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
 385                 pmem->pfn_pad = resource_size(res) -
 386                         resource_size(&pmem->pgmap.res);
 387                 pmem->pfn_flags |= PFN_MAP;
 388                 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
 389                 bb_res.start += pmem->data_offset;
 390         } else if (pmem_should_map_pages(dev)) {
 391                 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
 392                 pmem->pgmap.altmap_valid = false;
 393                 if (setup_pagemap_fsdax(dev, &pmem->pgmap))
 394                         return -ENOMEM;
 395                 addr = devm_memremap_pages(dev, &pmem->pgmap);
 396                 pmem->pfn_flags |= PFN_MAP;
 397                 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
 398         } else
 399                 addr = devm_memremap(dev, pmem->phys_addr,
 400                                 pmem->size, ARCH_MEMREMAP_PMEM);
 401
 402         /*
 403          * At release time the queue must be frozen before
 404          * devm_memremap_pages is unwound
 405          */
 406         if (devm_add_action_or_reset(dev, pmem_freeze_queue, q))
 407                 return -ENOMEM;
 408
 409         if (IS_ERR(addr))
 410                 return PTR_ERR(addr);
 411         pmem->virt_addr = addr;
 412
 413         blk_queue_write_cache(q, true, fua);
 414         blk_queue_make_request(q, pmem_make_request);
 415         blk_queue_physical_block_size(q, PAGE_SIZE);
 416         blk_queue_logical_block_size(q, pmem_sector_size(ndns));
 417         blk_queue_max_hw_sectors(q, UINT_MAX);
 418         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 419         if (pmem->pfn_flags & PFN_MAP)
 420                 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
 421         q->queuedata = pmem;
 422
 423         disk = alloc_disk_node(0, nid);
 424         if (!disk)
 425                 return -ENOMEM;
 426         pmem->disk = disk;
 427
 428         disk->fops              = &pmem_fops;
 429         disk->queue             = q;
 430         disk->flags             = GENHD_FL_EXT_DEVT;
 431         disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
 432         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 433         set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
 434                         / 512);
 435         if (devm_init_badblocks(dev, &pmem->bb))
 436                 return -ENOMEM;
 437         nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res);
 438         disk->bb = &pmem->bb;
 439
 440         dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops);
 441         if (!dax_dev) {
 442                 put_disk(disk);
 443                 return -ENOMEM;
 444         }
 445         dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
 446         pmem->dax_dev = dax_dev;
 447
 448         gendev = disk_to_dev(disk);
 449         gendev->groups = pmem_attribute_groups;
 450
 451         device_add_disk(dev, disk);
 452         if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
 453                 return -ENOMEM;
 454
 455         revalidate_disk(disk);
 456
 457         pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
 458                                           "badblocks");
 459         if (!pmem->bb_state)
 460                 dev_warn(dev, "'badblocks' notification disabled\n");
 461
 462         return 0;
 463 }
 464
 465 static int nd_pmem_probe(struct device *dev)
 466 {
 467         struct nd_namespace_common *ndns;
 468
 469         ndns = nvdimm_namespace_common_probe(dev);
 470         if (IS_ERR(ndns))
 471                 return PTR_ERR(ndns);
 472
 473         if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
 474                 return -ENXIO;
 475
 476         if (is_nd_btt(dev))
 477                 return nvdimm_namespace_attach_btt(ndns);
 478
 479         if (is_nd_pfn(dev))
 480                 return pmem_attach_disk(dev, ndns);
 481
 482         /* if we find a valid info-block we'll come back as that personality */
 483         if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
 484                         || nd_dax_probe(dev, ndns) == 0)
 485                 return -ENXIO;
 486
 487         /* ...otherwise we're just a raw pmem device */
 488         return pmem_attach_disk(dev, ndns);
 489 }
 490
 491 static int nd_pmem_remove(struct device *dev)
 492 {
 493         struct pmem_device *pmem = dev_get_drvdata(dev);
 494
 495         if (is_nd_btt(dev))
 496                 nvdimm_namespace_detach_btt(to_nd_btt(dev));
 497         else {
 498                 /*
 499                  * Note, this assumes device_lock() context to not race
 500                  * nd_pmem_notify()
 501                  */
 502                 sysfs_put(pmem->bb_state);
 503                 pmem->bb_state = NULL;
 504         }
 505         nvdimm_flush(to_nd_region(dev->parent));
 506
 507         return 0;
 508 }
 509
 510 static void nd_pmem_shutdown(struct device *dev)
 511 {
 512         nvdimm_flush(to_nd_region(dev->parent));
 513 }
 514
 515 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
 516 {
 517         struct nd_region *nd_region;
 518         resource_size_t offset = 0, end_trunc = 0;
 519         struct nd_namespace_common *ndns;
 520         struct nd_namespace_io *nsio;
 521         struct resource res;
 522         struct badblocks *bb;
 523         struct kernfs_node *bb_state;
 524
 525         if (event != NVDIMM_REVALIDATE_POISON)
 526                 return;
 527
 528         if (is_nd_btt(dev)) {
 529                 struct nd_btt *nd_btt = to_nd_btt(dev);
 530
 531                 ndns = nd_btt->ndns;
 532                 nd_region = to_nd_region(ndns->dev.parent);
 533                 nsio = to_nd_namespace_io(&ndns->dev);
 534                 bb = &nsio->bb;
 535                 bb_state = NULL;
 536         } else {
 537                 struct pmem_device *pmem = dev_get_drvdata(dev);
 538
 539                 nd_region = to_region(pmem);
 540                 bb = &pmem->bb;
 541                 bb_state = pmem->bb_state;
 542
 543                 if (is_nd_pfn(dev)) {
 544                         struct nd_pfn *nd_pfn = to_nd_pfn(dev);
 545                         struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 546
 547                         ndns = nd_pfn->ndns;
 548                         offset = pmem->data_offset +
 549                                         __le32_to_cpu(pfn_sb->start_pad);
 550                         end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
 551                 } else {
 552                         ndns = to_ndns(dev);
 553                 }
 554
 555                 nsio = to_nd_namespace_io(&ndns->dev);
 556         }
 557
 558         res.start = nsio->res.start + offset;
 559         res.end = nsio->res.end - end_trunc;
 560         nvdimm_badblocks_populate(nd_region, bb, &res);
 561         if (bb_state)
 562                 sysfs_notify_dirent(bb_state);
 563 }
 564
 565 MODULE_ALIAS("pmem");
 566 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 567 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 568 static struct nd_device_driver nd_pmem_driver = {
 569         .probe = nd_pmem_probe,
 570         .remove = nd_pmem_remove,
 571         .notify = nd_pmem_notify,
 572         .shutdown = nd_pmem_shutdown,
 573         .drv = {
 574                 .name = "nd_pmem",
 575         },
 576         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 577 };
 578
 579 module_nd_driver(nd_pmem_driver);
 580
 581 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 582 MODULE_LICENSE("GPL v2");