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