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