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/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         u64                     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         int rc = 0;
  70         void *mem = kmap_atomic(page);
  71         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
  72         void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
  73
  74         if (rw == READ) {
  75                 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
  76                         rc = -EIO;
  77                 else {
  78                         memcpy_from_pmem(mem + off, pmem_addr, len);
  79                         flush_dcache_page(page);
  80                 }
  81         } else {
  82                 flush_dcache_page(page);
  83                 memcpy_to_pmem(pmem_addr, mem + off, len);
  84         }
  85
  86         kunmap_atomic(mem);
  87         return rc;
  88 }
  89
  90 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
  91 {
  92         int rc = 0;
  93         bool do_acct;
  94         unsigned long start;
  95         struct bio_vec bvec;
  96         struct bvec_iter iter;
  97         struct block_device *bdev = bio->bi_bdev;
  98         struct pmem_device *pmem = bdev->bd_disk->private_data;
  99
 100         do_acct = nd_iostat_start(bio, &start);
 101         bio_for_each_segment(bvec, bio, iter) {
 102                 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
 103                                 bvec.bv_offset, bio_data_dir(bio),
 104                                 iter.bi_sector);
 105                 if (rc) {
 106                         bio->bi_error = rc;
 107                         break;
 108                 }
 109         }
 110         if (do_acct)
 111                 nd_iostat_end(bio, start);
 112
 113         if (bio_data_dir(bio))
 114                 wmb_pmem();
 115
 116         bio_endio(bio);
 117         return BLK_QC_T_NONE;
 118 }
 119
 120 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 121                        struct page *page, int rw)
 122 {
 123         struct pmem_device *pmem = bdev->bd_disk->private_data;
 124         int rc;
 125
 126         rc = pmem_do_bvec(pmem, page, PAGE_CACHE_SIZE, 0, rw, sector);
 127         if (rw & WRITE)
 128                 wmb_pmem();
 129
 130         /*
 131          * The ->rw_page interface is subtle and tricky.  The core
 132          * retries on any error, so we can only invoke page_endio() in
 133          * the successful completion case.  Otherwise, we'll see crashes
 134          * caused by double completion.
 135          */
 136         if (rc == 0)
 137                 page_endio(page, rw & WRITE, 0);
 138
 139         return rc;
 140 }
 141
 142 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
 143                       void __pmem **kaddr, pfn_t *pfn)
 144 {
 145         struct pmem_device *pmem = bdev->bd_disk->private_data;
 146         resource_size_t offset = sector * 512 + pmem->data_offset;
 147
 148         *kaddr = pmem->virt_addr + offset;
 149         *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 150
 151         return pmem->size - offset;
 152 }
 153
 154 static const struct block_device_operations pmem_fops = {
 155         .owner =                THIS_MODULE,
 156         .rw_page =              pmem_rw_page,
 157         .direct_access =        pmem_direct_access,
 158         .revalidate_disk =      nvdimm_revalidate_disk,
 159 };
 160
 161 static struct pmem_device *pmem_alloc(struct device *dev,
 162                 struct resource *res, int id)
 163 {
 164         struct pmem_device *pmem;
 165         struct request_queue *q;
 166
 167         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 168         if (!pmem)
 169                 return ERR_PTR(-ENOMEM);
 170
 171         pmem->phys_addr = res->start;
 172         pmem->size = resource_size(res);
 173         if (!arch_has_wmb_pmem())
 174                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 175
 176         if (!devm_request_mem_region(dev, pmem->phys_addr, pmem->size,
 177                         dev_name(dev))) {
 178                 dev_warn(dev, "could not reserve region [0x%pa:0x%zx]\n",
 179                                 &pmem->phys_addr, pmem->size);
 180                 return ERR_PTR(-EBUSY);
 181         }
 182
 183         q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
 184         if (!q)
 185                 return ERR_PTR(-ENOMEM);
 186
 187         pmem->pfn_flags = PFN_DEV;
 188         if (pmem_should_map_pages(dev)) {
 189                 pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, res,
 190                                 &q->q_usage_counter, NULL);
 191                 pmem->pfn_flags |= PFN_MAP;
 192         } else
 193                 pmem->virt_addr = (void __pmem *) devm_memremap(dev,
 194                                 pmem->phys_addr, pmem->size,
 195                                 ARCH_MEMREMAP_PMEM);
 196
 197         if (IS_ERR(pmem->virt_addr)) {
 198                 blk_cleanup_queue(q);
 199                 return (void __force *) pmem->virt_addr;
 200         }
 201
 202         pmem->pmem_queue = q;
 203         return pmem;
 204 }
 205
 206 static void pmem_detach_disk(struct pmem_device *pmem)
 207 {
 208         if (!pmem->pmem_disk)
 209                 return;
 210
 211         del_gendisk(pmem->pmem_disk);
 212         put_disk(pmem->pmem_disk);
 213         blk_cleanup_queue(pmem->pmem_queue);
 214 }
 215
 216 static int pmem_attach_disk(struct device *dev,
 217                 struct nd_namespace_common *ndns, struct pmem_device *pmem)
 218 {
 219         int nid = dev_to_node(dev);
 220         struct gendisk *disk;
 221
 222         blk_queue_make_request(pmem->pmem_queue, pmem_make_request);
 223         blk_queue_physical_block_size(pmem->pmem_queue, PAGE_SIZE);
 224         blk_queue_max_hw_sectors(pmem->pmem_queue, UINT_MAX);
 225         blk_queue_bounce_limit(pmem->pmem_queue, BLK_BOUNCE_ANY);
 226         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, pmem->pmem_queue);
 227
 228         disk = alloc_disk_node(0, nid);
 229         if (!disk) {
 230                 blk_cleanup_queue(pmem->pmem_queue);
 231                 return -ENOMEM;
 232         }
 233
 234         disk->major             = pmem_major;
 235         disk->first_minor       = 0;
 236         disk->fops              = &pmem_fops;
 237         disk->private_data      = pmem;
 238         disk->queue             = pmem->pmem_queue;
 239         disk->flags             = GENHD_FL_EXT_DEVT;
 240         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 241         disk->driverfs_dev = dev;
 242         set_capacity(disk, (pmem->size - pmem->data_offset) / 512);
 243         pmem->pmem_disk = disk;
 244         devm_exit_badblocks(dev, &pmem->bb);
 245         if (devm_init_badblocks(dev, &pmem->bb))
 246                 return -ENOMEM;
 247         nvdimm_namespace_add_poison(ndns, &pmem->bb, pmem->data_offset);
 248
 249         disk->bb = &pmem->bb;
 250         add_disk(disk);
 251         revalidate_disk(disk);
 252
 253         return 0;
 254 }
 255
 256 static int pmem_rw_bytes(struct nd_namespace_common *ndns,
 257                 resource_size_t offset, void *buf, size_t size, int rw)
 258 {
 259         struct pmem_device *pmem = dev_get_drvdata(ndns->claim);
 260
 261         if (unlikely(offset + size > pmem->size)) {
 262                 dev_WARN_ONCE(&ndns->dev, 1, "request out of range\n");
 263                 return -EFAULT;
 264         }
 265
 266         if (rw == READ) {
 267                 unsigned int sz_align = ALIGN(size + (offset & (512 - 1)), 512);
 268
 269                 if (unlikely(is_bad_pmem(&pmem->bb, offset / 512, sz_align)))
 270                         return -EIO;
 271                 memcpy_from_pmem(buf, pmem->virt_addr + offset, size);
 272         } else {
 273                 memcpy_to_pmem(pmem->virt_addr + offset, buf, size);
 274                 wmb_pmem();
 275         }
 276
 277         return 0;
 278 }
 279
 280 static int nd_pfn_init(struct nd_pfn *nd_pfn)
 281 {
 282         struct nd_pfn_sb *pfn_sb = kzalloc(sizeof(*pfn_sb), GFP_KERNEL);
 283         struct pmem_device *pmem = dev_get_drvdata(&nd_pfn->dev);
 284         struct nd_namespace_common *ndns = nd_pfn->ndns;
 285         struct nd_region *nd_region;
 286         unsigned long npfns;
 287         phys_addr_t offset;
 288         u64 checksum;
 289         int rc;
 290
 291         if (!pfn_sb)
 292                 return -ENOMEM;
 293
 294         nd_pfn->pfn_sb = pfn_sb;
 295         rc = nd_pfn_validate(nd_pfn);
 296         if (rc == -ENODEV)
 297                 /* no info block, do init */;
 298         else
 299                 return rc;
 300
 301         nd_region = to_nd_region(nd_pfn->dev.parent);
 302         if (nd_region->ro) {
 303                 dev_info(&nd_pfn->dev,
 304                                 "%s is read-only, unable to init metadata\n",
 305                                 dev_name(&nd_region->dev));
 306                 goto err;
 307         }
 308
 309         memset(pfn_sb, 0, sizeof(*pfn_sb));
 310         npfns = (pmem->size - SZ_8K) / SZ_4K;
 311         /*
 312          * Note, we use 64 here for the standard size of struct page,
 313          * debugging options may cause it to be larger in which case the
 314          * implementation will limit the pfns advertised through
 315          * ->direct_access() to those that are included in the memmap.
 316          */
 317         if (nd_pfn->mode == PFN_MODE_PMEM)
 318                 offset = ALIGN(SZ_8K + 64 * npfns, nd_pfn->align);
 319         else if (nd_pfn->mode == PFN_MODE_RAM)
 320                 offset = ALIGN(SZ_8K, nd_pfn->align);
 321         else
 322                 goto err;
 323
 324         npfns = (pmem->size - offset) / SZ_4K;
 325         pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
 326         pfn_sb->dataoff = cpu_to_le64(offset);
 327         pfn_sb->npfns = cpu_to_le64(npfns);
 328         memcpy(pfn_sb->signature, PFN_SIG, PFN_SIG_LEN);
 329         memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
 330         memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
 331         pfn_sb->version_major = cpu_to_le16(1);
 332         checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
 333         pfn_sb->checksum = cpu_to_le64(checksum);
 334
 335         rc = nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb));
 336         if (rc)
 337                 goto err;
 338
 339         return 0;
 340  err:
 341         nd_pfn->pfn_sb = NULL;
 342         kfree(pfn_sb);
 343         return -ENXIO;
 344 }
 345
 346 static int nvdimm_namespace_detach_pfn(struct nd_namespace_common *ndns)
 347 {
 348         struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
 349         struct pmem_device *pmem;
 350
 351         /* free pmem disk */
 352         pmem = dev_get_drvdata(&nd_pfn->dev);
 353         pmem_detach_disk(pmem);
 354
 355         /* release nd_pfn resources */
 356         kfree(nd_pfn->pfn_sb);
 357         nd_pfn->pfn_sb = NULL;
 358
 359         return 0;
 360 }
 361
 362 static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
 363 {
 364         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 365         struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
 366         struct device *dev = &nd_pfn->dev;
 367         struct nd_region *nd_region;
 368         struct vmem_altmap *altmap;
 369         struct nd_pfn_sb *pfn_sb;
 370         struct pmem_device *pmem;
 371         struct request_queue *q;
 372         phys_addr_t offset;
 373         int rc;
 374         struct vmem_altmap __altmap = {
 375                 .base_pfn = __phys_to_pfn(nsio->res.start),
 376                 .reserve = __phys_to_pfn(SZ_8K),
 377         };
 378
 379         if (!nd_pfn->uuid || !nd_pfn->ndns)
 380                 return -ENODEV;
 381
 382         nd_region = to_nd_region(dev->parent);
 383         rc = nd_pfn_init(nd_pfn);
 384         if (rc)
 385                 return rc;
 386
 387         pfn_sb = nd_pfn->pfn_sb;
 388         offset = le64_to_cpu(pfn_sb->dataoff);
 389         nd_pfn->mode = le32_to_cpu(nd_pfn->pfn_sb->mode);
 390         if (nd_pfn->mode == PFN_MODE_RAM) {
 391                 if (offset < SZ_8K)
 392                         return -EINVAL;
 393                 nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
 394                 altmap = NULL;
 395         } else if (nd_pfn->mode == PFN_MODE_PMEM) {
 396                 nd_pfn->npfns = (resource_size(&nsio->res) - offset)
 397                         / PAGE_SIZE;
 398                 if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
 399                         dev_info(&nd_pfn->dev,
 400                                         "number of pfns truncated from %lld to %ld\n",
 401                                         le64_to_cpu(nd_pfn->pfn_sb->npfns),
 402                                         nd_pfn->npfns);
 403                 altmap = & __altmap;
 404                 altmap->free = __phys_to_pfn(offset - SZ_8K);
 405                 altmap->alloc = 0;
 406         } else {
 407                 rc = -ENXIO;
 408                 goto err;
 409         }
 410
 411         /* establish pfn range for lookup, and switch to direct map */
 412         pmem = dev_get_drvdata(dev);
 413         q = pmem->pmem_queue;
 414         devm_memunmap(dev, (void __force *) pmem->virt_addr);
 415         pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, &nsio->res,
 416                         &q->q_usage_counter, altmap);
 417         pmem->pfn_flags |= PFN_MAP;
 418         if (IS_ERR(pmem->virt_addr)) {
 419                 rc = PTR_ERR(pmem->virt_addr);
 420                 goto err;
 421         }
 422
 423         /* attach pmem disk in "pfn-mode" */
 424         pmem->data_offset = offset;
 425         rc = pmem_attach_disk(dev, ndns, pmem);
 426         if (rc)
 427                 goto err;
 428
 429         return rc;
 430  err:
 431         nvdimm_namespace_detach_pfn(ndns);
 432         return rc;
 433 }
 434
 435 static int nd_pmem_probe(struct device *dev)
 436 {
 437         struct nd_region *nd_region = to_nd_region(dev->parent);
 438         struct nd_namespace_common *ndns;
 439         struct nd_namespace_io *nsio;
 440         struct pmem_device *pmem;
 441
 442         ndns = nvdimm_namespace_common_probe(dev);
 443         if (IS_ERR(ndns))
 444                 return PTR_ERR(ndns);
 445
 446         nsio = to_nd_namespace_io(&ndns->dev);
 447         pmem = pmem_alloc(dev, &nsio->res, nd_region->id);
 448         if (IS_ERR(pmem))
 449                 return PTR_ERR(pmem);
 450
 451         pmem->ndns = ndns;
 452         dev_set_drvdata(dev, pmem);
 453         ndns->rw_bytes = pmem_rw_bytes;
 454         if (devm_init_badblocks(dev, &pmem->bb))
 455                 return -ENOMEM;
 456         nvdimm_namespace_add_poison(ndns, &pmem->bb, 0);
 457
 458         if (is_nd_btt(dev)) {
 459                 /* btt allocates its own request_queue */
 460                 blk_cleanup_queue(pmem->pmem_queue);
 461                 pmem->pmem_queue = NULL;
 462                 return nvdimm_namespace_attach_btt(ndns);
 463         }
 464
 465         if (is_nd_pfn(dev))
 466                 return nvdimm_namespace_attach_pfn(ndns);
 467
 468         if (nd_btt_probe(ndns, pmem) == 0 || nd_pfn_probe(ndns, pmem) == 0) {
 469                 /*
 470                  * We'll come back as either btt-pmem, or pfn-pmem, so
 471                  * drop the queue allocation for now.
 472                  */
 473                 blk_cleanup_queue(pmem->pmem_queue);
 474                 return -ENXIO;
 475         }
 476
 477         return pmem_attach_disk(dev, ndns, pmem);
 478 }
 479
 480 static int nd_pmem_remove(struct device *dev)
 481 {
 482         struct pmem_device *pmem = dev_get_drvdata(dev);
 483
 484         if (is_nd_btt(dev))
 485                 nvdimm_namespace_detach_btt(pmem->ndns);
 486         else if (is_nd_pfn(dev))
 487                 nvdimm_namespace_detach_pfn(pmem->ndns);
 488         else
 489                 pmem_detach_disk(pmem);
 490
 491         return 0;
 492 }
 493
 494 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
 495 {
 496         struct pmem_device *pmem = dev_get_drvdata(dev);
 497         struct nd_namespace_common *ndns = pmem->ndns;
 498
 499         if (event != NVDIMM_REVALIDATE_POISON)
 500                 return;
 501
 502         if (is_nd_btt(dev))
 503                 nvdimm_namespace_add_poison(ndns, &pmem->bb, 0);
 504         else
 505                 nvdimm_namespace_add_poison(ndns, &pmem->bb, pmem->data_offset);
 506 }
 507
 508 MODULE_ALIAS("pmem");
 509 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 510 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 511 static struct nd_device_driver nd_pmem_driver = {
 512         .probe = nd_pmem_probe,
 513         .remove = nd_pmem_remove,
 514         .notify = nd_pmem_notify,
 515         .drv = {
 516                 .name = "nd_pmem",
 517         },
 518         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 519 };
 520
 521 static int __init pmem_init(void)
 522 {
 523         int error;
 524
 525         pmem_major = register_blkdev(0, "pmem");
 526         if (pmem_major < 0)
 527                 return pmem_major;
 528
 529         error = nd_driver_register(&nd_pmem_driver);
 530         if (error) {
 531                 unregister_blkdev(pmem_major, "pmem");
 532                 return error;
 533         }
 534
 535         return 0;
 536 }
 537 module_init(pmem_init);
 538
 539 static void pmem_exit(void)
 540 {
 541         driver_unregister(&nd_pmem_driver.drv);
 542         unregister_blkdev(pmem_major, "pmem");
 543 }
 544 module_exit(pmem_exit);
 545
 546 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 547 MODULE_LICENSE("GPL v2");