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