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