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