drivers/nvdimm/pmem.c

   1 /*
   2  * Persistent Memory Driver
   3  *
   4  * Copyright (c) 2014-2015, Intel Corporation.
   5  * Copyright (c) 2015, Christoph Hellwig <[email protected]>.
   6  * Copyright (c) 2015, Boaz Harrosh <[email protected]>.
   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/pfn_t.h>
  29 #include <linux/slab.h>
  30 #include <linux/pmem.h>
  31 #include <linux/nd.h>
  32 #include "pfn.h"
  33 #include "nd.h"
  34
  35 struct pmem_device {
  36         struct request_queue    *pmem_queue;
  37         struct gendisk          *pmem_disk;
  38         struct nd_namespace_common *ndns;
  39
  40         /* One contiguous memory region per device */
  41         phys_addr_t             phys_addr;
  42         /* when non-zero this device is hosting a 'pfn' instance */
  43         phys_addr_t             data_offset;
  44         unsigned long           pfn_flags;
  45         void __pmem             *virt_addr;
  46         size_t                  size;
  47         struct badblocks        bb;
  48 };
  49
  50 static int pmem_major;
  51
  52 static bool is_bad_pmem(struct badblocks *bb, sector_t sector, unsigned int len)
  53 {
  54         if (bb->count) {
  55                 sector_t first_bad;
  56                 int num_bad;
  57
  58                 return !!badblocks_check(bb, sector, len / 512, &first_bad,
  59                                 &num_bad);
  60         }
  61
  62         return false;
  63 }
  64
  65 static int pmem_do_bvec(struct pmem_device *pmem, struct page *page,
  66                         unsigned int len, unsigned int off, int rw,
  67                         sector_t sector)
  68 {
  69         void *mem = kmap_atomic(page);
  70         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
  71         void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
  72
  73         if (rw == READ) {
  74                 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
  75                         return -EIO;
  76                 memcpy_from_pmem(mem + off, pmem_addr, len);
  77                 flush_dcache_page(page);
  78         } else {
  79                 flush_dcache_page(page);
  80                 memcpy_to_pmem(pmem_addr, mem + off, len);
  81         }
  82
  83         kunmap_atomic(mem);
  84         return 0;
  85 }
  86
  87 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
  88 {
  89         int rc = 0;
  90         bool do_acct;
  91         unsigned long start;
  92         struct bio_vec bvec;
  93         struct bvec_iter iter;
  94         struct block_device *bdev = bio->bi_bdev;
  95         struct pmem_device *pmem = bdev->bd_disk->private_data;
  96
  97         do_acct = nd_iostat_start(bio, &start);
  98         bio_for_each_segment(bvec, bio, iter) {
  99                 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
 100                                 bvec.bv_offset, bio_data_dir(bio),
 101                                 iter.bi_sector);
 102                 if (rc) {
 103                         bio->bi_error = rc;
 104                         break;
 105                 }
 106         }
 107         if (do_acct)
 108                 nd_iostat_end(bio, start);
 109
 110         if (bio_data_dir(bio))
 111                 wmb_pmem();
 112
 113         bio_endio(bio);
 114         return BLK_QC_T_NONE;
 115 }
 116
 117 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 118                        struct page *page, int rw)
 119 {
 120         struct pmem_device *pmem = bdev->bd_disk->private_data;
 121         int rc;
 122
 123         rc = pmem_do_bvec(pmem, page, PAGE_CACHE_SIZE, 0, rw, sector);
 124         if (rw & WRITE)
 125                 wmb_pmem();
 126
 127         /*
 128          * The ->rw_page interface is subtle and tricky.  The core
 129          * retries on any error, so we can only invoke page_endio() in
 130          * the successful completion case.  Otherwise, we'll see crashes
 131          * caused by double completion.
 132          */
 133         if (rc == 0)
 134                 page_endio(page, rw & WRITE, 0);
 135
 136         return rc;
 137 }
 138
 139 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
 140                       void __pmem **kaddr, pfn_t *pfn)
 141 {
 142         struct pmem_device *pmem = bdev->bd_disk->private_data;
 143         resource_size_t offset = sector * 512 + pmem->data_offset;
 144
 145         *kaddr = pmem->virt_addr + offset;
 146         *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 147
 148         return pmem->size - offset;
 149 }
 150
 151 static const struct block_device_operations pmem_fops = {
 152         .owner =                THIS_MODULE,
 153         .rw_page =              pmem_rw_page,
 154         .direct_access =        pmem_direct_access,
 155         .revalidate_disk =      nvdimm_revalidate_disk,
 156 };
 157
 158 static struct pmem_device *pmem_alloc(struct device *dev,
 159                 struct resource *res, int id)
 160 {
 161         struct pmem_device *pmem;
 162
 163         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 164         if (!pmem)
 165                 return ERR_PTR(-ENOMEM);
 166
 167         pmem->phys_addr = res->start;
 168         pmem->size = resource_size(res);
 169         if (!arch_has_wmb_pmem())
 170                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 171
 172         if (!devm_request_mem_region(dev, pmem->phys_addr, pmem->size,
 173                         dev_name(dev))) {
 174                 dev_warn(dev, "could not reserve region [0x%pa:0x%zx]\n",
 175                                 &pmem->phys_addr, pmem->size);
 176                 return ERR_PTR(-EBUSY);
 177         }
 178
 179         pmem->pfn_flags = PFN_DEV;
 180         if (pmem_should_map_pages(dev)) {
 181                 pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, res);
 182                 pmem->pfn_flags |= PFN_MAP;
 183         } else
 184                 pmem->virt_addr = (void __pmem *) devm_memremap(dev,
 185                                 pmem->phys_addr, pmem->size,
 186                                 ARCH_MEMREMAP_PMEM);
 187
 188         if (IS_ERR(pmem->virt_addr))
 189                 return (void __force *) pmem->virt_addr;
 190
 191         return pmem;
 192 }
 193
 194 static void pmem_detach_disk(struct pmem_device *pmem)
 195 {
 196         if (!pmem->pmem_disk)
 197                 return;
 198
 199         del_gendisk(pmem->pmem_disk);
 200         put_disk(pmem->pmem_disk);
 201         blk_cleanup_queue(pmem->pmem_queue);
 202 }
 203
 204 static int pmem_attach_disk(struct device *dev,
 205                 struct nd_namespace_common *ndns, struct pmem_device *pmem)
 206 {
 207         int nid = dev_to_node(dev);
 208         struct gendisk *disk;
 209
 210         pmem->pmem_queue = blk_alloc_queue_node(GFP_KERNEL, nid);
 211         if (!pmem->pmem_queue)
 212                 return -ENOMEM;
 213
 214         blk_queue_make_request(pmem->pmem_queue, pmem_make_request);
 215         blk_queue_physical_block_size(pmem->pmem_queue, PAGE_SIZE);
 216         blk_queue_max_hw_sectors(pmem->pmem_queue, UINT_MAX);
 217         blk_queue_bounce_limit(pmem->pmem_queue, BLK_BOUNCE_ANY);
 218         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, pmem->pmem_queue);
 219
 220         disk = alloc_disk_node(0, nid);
 221         if (!disk) {
 222                 blk_cleanup_queue(pmem->pmem_queue);
 223                 return -ENOMEM;
 224         }
 225
 226         disk->major             = pmem_major;
 227         disk->first_minor       = 0;
 228         disk->fops              = &pmem_fops;
 229         disk->private_data      = pmem;
 230         disk->queue             = pmem->pmem_queue;
 231         disk->flags             = GENHD_FL_EXT_DEVT;
 232         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 233         disk->driverfs_dev = dev;
 234         set_capacity(disk, (pmem->size - pmem->data_offset) / 512);
 235         pmem->pmem_disk = disk;
 236         devm_exit_badblocks(dev, &pmem->bb);
 237         if (devm_init_badblocks(dev, &pmem->bb))
 238                 return -ENOMEM;
 239         nvdimm_namespace_add_poison(ndns, &pmem->bb, pmem->data_offset);
 240
 241         disk->bb = &pmem->bb;
 242         add_disk(disk);
 243         revalidate_disk(disk);
 244
 245         return 0;
 246 }
 247
 248 static int pmem_rw_bytes(struct nd_namespace_common *ndns,
 249                 resource_size_t offset, void *buf, size_t size, int rw)
 250 {
 251         struct pmem_device *pmem = dev_get_drvdata(ndns->claim);
 252
 253         if (unlikely(offset + size > pmem->size)) {
 254                 dev_WARN_ONCE(&ndns->dev, 1, "request out of range\n");
 255                 return -EFAULT;
 256         }
 257
 258         if (rw == READ) {
 259                 unsigned int sz_align = ALIGN(size + (offset & (512 - 1)), 512);
 260
 261                 if (unlikely(is_bad_pmem(&pmem->bb, offset / 512, sz_align)))
 262                         return -EIO;
 263                 memcpy_from_pmem(buf, pmem->virt_addr + offset, size);
 264         } else {
 265                 memcpy_to_pmem(pmem->virt_addr + offset, buf, size);
 266                 wmb_pmem();
 267         }
 268
 269         return 0;
 270 }
 271
 272 static int nd_pfn_init(struct nd_pfn *nd_pfn)
 273 {
 274         struct nd_pfn_sb *pfn_sb = kzalloc(sizeof(*pfn_sb), GFP_KERNEL);
 275         struct pmem_device *pmem = dev_get_drvdata(&nd_pfn->dev);
 276         struct nd_namespace_common *ndns = nd_pfn->ndns;
 277         struct nd_region *nd_region;
 278         unsigned long npfns;
 279         phys_addr_t offset;
 280         u64 checksum;
 281         int rc;
 282
 283         if (!pfn_sb)
 284                 return -ENOMEM;
 285
 286         nd_pfn->pfn_sb = pfn_sb;
 287         rc = nd_pfn_validate(nd_pfn);
 288         if (rc == -ENODEV)
 289                 /* no info block, do init */;
 290         else
 291                 return rc;
 292
 293         nd_region = to_nd_region(nd_pfn->dev.parent);
 294         if (nd_region->ro) {
 295                 dev_info(&nd_pfn->dev,
 296                                 "%s is read-only, unable to init metadata\n",
 297                                 dev_name(&nd_region->dev));
 298                 goto err;
 299         }
 300
 301         memset(pfn_sb, 0, sizeof(*pfn_sb));
 302         npfns = (pmem->size - SZ_8K) / SZ_4K;
 303         /*
 304          * Note, we use 64 here for the standard size of struct page,
 305          * debugging options may cause it to be larger in which case the
 306          * implementation will limit the pfns advertised through
 307          * ->direct_access() to those that are included in the memmap.
 308          */
 309         if (nd_pfn->mode == PFN_MODE_PMEM)
 310                 offset = ALIGN(SZ_8K + 64 * npfns, nd_pfn->align);
 311         else if (nd_pfn->mode == PFN_MODE_RAM)
 312                 offset = ALIGN(SZ_8K, nd_pfn->align);
 313         else
 314                 goto err;
 315
 316         npfns = (pmem->size - offset) / SZ_4K;
 317         pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
 318         pfn_sb->dataoff = cpu_to_le64(offset);
 319         pfn_sb->npfns = cpu_to_le64(npfns);
 320         memcpy(pfn_sb->signature, PFN_SIG, PFN_SIG_LEN);
 321         memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
 322         memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
 323         pfn_sb->version_major = cpu_to_le16(1);
 324         checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
 325         pfn_sb->checksum = cpu_to_le64(checksum);
 326
 327         rc = nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb));
 328         if (rc)
 329                 goto err;
 330
 331         return 0;
 332  err:
 333         nd_pfn->pfn_sb = NULL;
 334         kfree(pfn_sb);
 335         return -ENXIO;
 336 }
 337
 338 static int nvdimm_namespace_detach_pfn(struct nd_namespace_common *ndns)
 339 {
 340         struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
 341         struct pmem_device *pmem;
 342
 343         /* free pmem disk */
 344         pmem = dev_get_drvdata(&nd_pfn->dev);
 345         pmem_detach_disk(pmem);
 346
 347         /* release nd_pfn resources */
 348         kfree(nd_pfn->pfn_sb);
 349         nd_pfn->pfn_sb = NULL;
 350
 351         return 0;
 352 }
 353
 354 static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
 355 {
 356         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 357         struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
 358         struct device *dev = &nd_pfn->dev;
 359         struct vmem_altmap *altmap;
 360         struct nd_region *nd_region;
 361         struct nd_pfn_sb *pfn_sb;
 362         struct pmem_device *pmem;
 363         phys_addr_t offset;
 364         int rc;
 365
 366         if (!nd_pfn->uuid || !nd_pfn->ndns)
 367                 return -ENODEV;
 368
 369         nd_region = to_nd_region(dev->parent);
 370         rc = nd_pfn_init(nd_pfn);
 371         if (rc)
 372                 return rc;
 373
 374         pfn_sb = nd_pfn->pfn_sb;
 375         offset = le64_to_cpu(pfn_sb->dataoff);
 376         nd_pfn->mode = le32_to_cpu(nd_pfn->pfn_sb->mode);
 377         if (nd_pfn->mode == PFN_MODE_RAM) {
 378                 if (offset < SZ_8K)
 379                         return -EINVAL;
 380                 nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
 381                 altmap = NULL;
 382         } else {
 383                 rc = -ENXIO;
 384                 goto err;
 385         }
 386
 387         /* establish pfn range for lookup, and switch to direct map */
 388         pmem = dev_get_drvdata(dev);
 389         devm_memunmap(dev, (void __force *) pmem->virt_addr);
 390         pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, &nsio->res);
 391         pmem->pfn_flags |= PFN_MAP;
 392         if (IS_ERR(pmem->virt_addr)) {
 393                 rc = PTR_ERR(pmem->virt_addr);
 394                 goto err;
 395         }
 396
 397         /* attach pmem disk in "pfn-mode" */
 398         pmem->data_offset = offset;
 399         rc = pmem_attach_disk(dev, ndns, pmem);
 400         if (rc)
 401                 goto err;
 402
 403         return rc;
 404  err:
 405         nvdimm_namespace_detach_pfn(ndns);
 406         return rc;
 407 }
 408
 409 static int nd_pmem_probe(struct device *dev)
 410 {
 411         struct nd_region *nd_region = to_nd_region(dev->parent);
 412         struct nd_namespace_common *ndns;
 413         struct nd_namespace_io *nsio;
 414         struct pmem_device *pmem;
 415
 416         ndns = nvdimm_namespace_common_probe(dev);
 417         if (IS_ERR(ndns))
 418                 return PTR_ERR(ndns);
 419
 420         nsio = to_nd_namespace_io(&ndns->dev);
 421         pmem = pmem_alloc(dev, &nsio->res, nd_region->id);
 422         if (IS_ERR(pmem))
 423                 return PTR_ERR(pmem);
 424
 425         pmem->ndns = ndns;
 426         dev_set_drvdata(dev, pmem);
 427         ndns->rw_bytes = pmem_rw_bytes;
 428         if (devm_init_badblocks(dev, &pmem->bb))
 429                 return -ENOMEM;
 430         nvdimm_namespace_add_poison(ndns, &pmem->bb, 0);
 431
 432         if (is_nd_btt(dev))
 433                 return nvdimm_namespace_attach_btt(ndns);
 434
 435         if (is_nd_pfn(dev))
 436                 return nvdimm_namespace_attach_pfn(ndns);
 437
 438         if (nd_btt_probe(ndns, pmem) == 0) {
 439                 /* we'll come back as btt-pmem */
 440                 return -ENXIO;
 441         }
 442
 443         if (nd_pfn_probe(ndns, pmem) == 0) {
 444                 /* we'll come back as pfn-pmem */
 445                 return -ENXIO;
 446         }
 447
 448         return pmem_attach_disk(dev, ndns, pmem);
 449 }
 450
 451 static int nd_pmem_remove(struct device *dev)
 452 {
 453         struct pmem_device *pmem = dev_get_drvdata(dev);
 454
 455         if (is_nd_btt(dev))
 456                 nvdimm_namespace_detach_btt(pmem->ndns);
 457         else if (is_nd_pfn(dev))
 458                 nvdimm_namespace_detach_pfn(pmem->ndns);
 459         else
 460                 pmem_detach_disk(pmem);
 461
 462         return 0;
 463 }
 464
 465 MODULE_ALIAS("pmem");
 466 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 467 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 468 static struct nd_device_driver nd_pmem_driver = {
 469         .probe = nd_pmem_probe,
 470         .remove = nd_pmem_remove,
 471         .drv = {
 472                 .name = "nd_pmem",
 473         },
 474         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 475 };
 476
 477 static int __init pmem_init(void)
 478 {
 479         int error;
 480
 481         pmem_major = register_blkdev(0, "pmem");
 482         if (pmem_major < 0)
 483                 return pmem_major;
 484
 485         error = nd_driver_register(&nd_pmem_driver);
 486         if (error) {
 487                 unregister_blkdev(pmem_major, "pmem");
 488                 return error;
 489         }
 490
 491         return 0;
 492 }
 493 module_init(pmem_init);
 494
 495 static void pmem_exit(void)
 496 {
 497         driver_unregister(&nd_pmem_driver.drv);
 498         unregister_blkdev(pmem_major, "pmem");
 499 }
 500 module_exit(pmem_exit);
 501
 502 MODULE_AUTHOR("Ross Zwisler <[email protected]>");
 503 MODULE_LICENSE("GPL v2");