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/memory_hotplug.h>
  25 #include <linux/moduleparam.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/slab.h>
  28 #include <linux/pmem.h>
  29 #include <linux/nd.h>
  30 #include "pfn.h"
  31 #include "nd.h"
  32
  33 struct pmem_device {
  34         struct request_queue    *pmem_queue;
  35         struct gendisk          *pmem_disk;
  36         struct nd_namespace_common *ndns;
  37
  38         /* One contiguous memory region per device */
  39         phys_addr_t             phys_addr;
  40         /* when non-zero this device is hosting a 'pfn' instance */
  41         phys_addr_t             data_offset;
  42         void __pmem             *virt_addr;
  43         size_t                  size;
  44 };
  45
  46 static int pmem_major;
  47
  48 static void pmem_do_bvec(struct pmem_device *pmem, struct page *page,
  49                         unsigned int len, unsigned int off, int rw,
  50                         sector_t sector)
  51 {
  52         void *mem = kmap_atomic(page);
  53         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
  54         void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
  55
  56         if (rw == READ) {
  57                 memcpy_from_pmem(mem + off, pmem_addr, len);
  58                 flush_dcache_page(page);
  59         } else {
  60                 flush_dcache_page(page);
  61                 memcpy_to_pmem(pmem_addr, mem + off, len);
  62         }
  63
  64         kunmap_atomic(mem);
  65 }
  66
  67 static void pmem_make_request(struct request_queue *q, struct bio *bio)
  68 {
  69         bool do_acct;
  70         unsigned long start;
  71         struct bio_vec bvec;
  72         struct bvec_iter iter;
  73         struct block_device *bdev = bio->bi_bdev;
  74         struct pmem_device *pmem = bdev->bd_disk->private_data;
  75
  76         do_acct = nd_iostat_start(bio, &start);
  77         bio_for_each_segment(bvec, bio, iter)
  78                 pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len, bvec.bv_offset,
  79                                 bio_data_dir(bio), iter.bi_sector);
  80         if (do_acct)
  81                 nd_iostat_end(bio, start);
  82
  83         if (bio_data_dir(bio))
  84                 wmb_pmem();
  85
  86         bio_endio(bio);
  87 }
  88
  89 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
  90                        struct page *page, int rw)
  91 {
  92         struct pmem_device *pmem = bdev->bd_disk->private_data;
  93
  94         pmem_do_bvec(pmem, page, PAGE_CACHE_SIZE, 0, rw, sector);
  95         if (rw & WRITE)
  96                 wmb_pmem();
  97         page_endio(page, rw & WRITE, 0);
  98
  99         return 0;
 100 }
 101
 102 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
 103                       void __pmem **kaddr, unsigned long *pfn)
 104 {
 105         struct pmem_device *pmem = bdev->bd_disk->private_data;
 106         resource_size_t offset = sector * 512 + pmem->data_offset;
 107         resource_size_t size;
 108
 109         if (pmem->data_offset) {
 110                 /*
 111                  * Limit the direct_access() size to what is covered by
 112                  * the memmap
 113                  */
 114                 size = (pmem->size - offset) & ~ND_PFN_MASK;
 115         } else
 116                 size = pmem->size - offset;
 117
 118         /* FIXME convert DAX to comprehend that this mapping has a lifetime */
 119         *kaddr = pmem->virt_addr + offset;
 120         *pfn = (pmem->phys_addr + offset) >> PAGE_SHIFT;
 121
 122         return size;
 123 }
 124
 125 static const struct block_device_operations pmem_fops = {
 126         .owner =                THIS_MODULE,
 127         .rw_page =              pmem_rw_page,
 128         .direct_access =        pmem_direct_access,
 129         .revalidate_disk =      nvdimm_revalidate_disk,
 130 };
 131
 132 static struct pmem_device *pmem_alloc(struct device *dev,
 133                 struct resource *res, int id)
 134 {
 135         struct pmem_device *pmem;
 136
 137         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 138         if (!pmem)
 139                 return ERR_PTR(-ENOMEM);
 140
 141         pmem->phys_addr = res->start;
 142         pmem->size = resource_size(res);
 143         if (!arch_has_wmb_pmem())
 144                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 145
 146         if (!devm_request_mem_region(dev, pmem->phys_addr, pmem->size,
 147                         dev_name(dev))) {
 148                 dev_warn(dev, "could not reserve region [0x%pa:0x%zx]\n",
 149                                 &pmem->phys_addr, pmem->size);
 150                 return ERR_PTR(-EBUSY);
 151         }
 152
 153         if (pmem_should_map_pages(dev))
 154                 pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, res);
 155         else
 156                 pmem->virt_addr = (void __pmem *) devm_memremap(dev,
 157                                 pmem->phys_addr, pmem->size,
 158                                 ARCH_MEMREMAP_PMEM);
 159
 160         if (IS_ERR(pmem->virt_addr))
 161                 return (void __force *) pmem->virt_addr;
 162
 163         return pmem;
 164 }
 165
 166 static void pmem_detach_disk(struct pmem_device *pmem)
 167 {
 168         if (!pmem->pmem_disk)
 169                 return;
 170
 171         del_gendisk(pmem->pmem_disk);
 172         put_disk(pmem->pmem_disk);
 173         blk_cleanup_queue(pmem->pmem_queue);
 174 }
 175
 176 static int pmem_attach_disk(struct device *dev,
 177                 struct nd_namespace_common *ndns, struct pmem_device *pmem)
 178 {
 179         struct gendisk *disk;
 180
 181         pmem->pmem_queue = blk_alloc_queue(GFP_KERNEL);
 182         if (!pmem->pmem_queue)
 183                 return -ENOMEM;
 184
 185         blk_queue_make_request(pmem->pmem_queue, pmem_make_request);
 186         blk_queue_physical_block_size(pmem->pmem_queue, PAGE_SIZE);
 187         blk_queue_max_hw_sectors(pmem->pmem_queue, UINT_MAX);
 188         blk_queue_bounce_limit(pmem->pmem_queue, BLK_BOUNCE_ANY);
 189         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, pmem->pmem_queue);
 190
 191         disk = alloc_disk(0);
 192         if (!disk) {
 193                 blk_cleanup_queue(pmem->pmem_queue);
 194                 return -ENOMEM;
 195         }
 196
 197         disk->major             = pmem_major;
 198         disk->first_minor       = 0;
 199         disk->fops              = &pmem_fops;
 200         disk->private_data      = pmem;
 201         disk->queue             = pmem->pmem_queue;
 202         disk->flags             = GENHD_FL_EXT_DEVT;
 203         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 204         disk->driverfs_dev = dev;
 205         set_capacity(disk, (pmem->size - pmem->data_offset) / 512);
 206         pmem->pmem_disk = disk;
 207
 208         add_disk(disk);
 209         revalidate_disk(disk);
 210
 211         return 0;
 212 }
 213
 214 static int pmem_rw_bytes(struct nd_namespace_common *ndns,
 215                 resource_size_t offset, void *buf, size_t size, int rw)
 216 {
 217         struct pmem_device *pmem = dev_get_drvdata(ndns->claim);
 218
 219         if (unlikely(offset + size > pmem->size)) {
 220                 dev_WARN_ONCE(&ndns->dev, 1, "request out of range\n");
 221                 return -EFAULT;
 222         }
 223
 224         if (rw == READ)
 225                 memcpy_from_pmem(buf, pmem->virt_addr + offset, size);
 226         else {
 227                 memcpy_to_pmem(pmem->virt_addr + offset, buf, size);
 228                 wmb_pmem();
 229         }
 230
 231         return 0;
 232 }
 233
 234 static int nd_pfn_init(struct nd_pfn *nd_pfn)
 235 {
 236         struct nd_pfn_sb *pfn_sb = kzalloc(sizeof(*pfn_sb), GFP_KERNEL);
 237         struct pmem_device *pmem = dev_get_drvdata(&nd_pfn->dev);
 238         struct nd_namespace_common *ndns = nd_pfn->ndns;
 239         struct nd_region *nd_region;
 240         unsigned long npfns;
 241         phys_addr_t offset;
 242         u64 checksum;
 243         int rc;
 244
 245         if (!pfn_sb)
 246                 return -ENOMEM;
 247
 248         nd_pfn->pfn_sb = pfn_sb;
 249         rc = nd_pfn_validate(nd_pfn);
 250         if (rc == 0 || rc == -EBUSY)
 251                 return rc;
 252
 253         /* section alignment for simple hotplug */
 254         if (nvdimm_namespace_capacity(ndns) < ND_PFN_ALIGN
 255                         || pmem->phys_addr & ND_PFN_MASK)
 256                 return -ENODEV;
 257
 258         nd_region = to_nd_region(nd_pfn->dev.parent);
 259         if (nd_region->ro) {
 260                 dev_info(&nd_pfn->dev,
 261                                 "%s is read-only, unable to init metadata\n",
 262                                 dev_name(&nd_region->dev));
 263                 goto err;
 264         }
 265
 266         memset(pfn_sb, 0, sizeof(*pfn_sb));
 267         npfns = (pmem->size - SZ_8K) / SZ_4K;
 268         /*
 269          * Note, we use 64 here for the standard size of struct page,
 270          * debugging options may cause it to be larger in which case the
 271          * implementation will limit the pfns advertised through
 272          * ->direct_access() to those that are included in the memmap.
 273          */
 274         if (nd_pfn->mode == PFN_MODE_PMEM)
 275                 offset = ALIGN(SZ_8K + 64 * npfns, PMD_SIZE);
 276         else if (nd_pfn->mode == PFN_MODE_RAM)
 277                 offset = SZ_8K;
 278         else
 279                 goto err;
 280
 281         npfns = (pmem->size - offset) / SZ_4K;
 282         pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
 283         pfn_sb->dataoff = cpu_to_le64(offset);
 284         pfn_sb->npfns = cpu_to_le64(npfns);
 285         memcpy(pfn_sb->signature, PFN_SIG, PFN_SIG_LEN);
 286         memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
 287         pfn_sb->version_major = cpu_to_le16(1);
 288         checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
 289         pfn_sb->checksum = cpu_to_le64(checksum);
 290
 291         rc = nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb));
 292         if (rc)
 293                 goto err;
 294
 295         return 0;
 296  err:
 297         nd_pfn->pfn_sb = NULL;
 298         kfree(pfn_sb);
 299         return -ENXIO;
 300 }
 301
 302 static int nvdimm_namespace_detach_pfn(struct nd_namespace_common *ndns)
 303 {
 304         struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
 305         struct pmem_device *pmem;
 306
 307         /* free pmem disk */
 308         pmem = dev_get_drvdata(&nd_pfn->dev);
 309         pmem_detach_disk(pmem);
 310
 311         /* release nd_pfn resources */
 312         kfree(nd_pfn->pfn_sb);
 313         nd_pfn->pfn_sb = NULL;
 314
 315         return 0;
 316 }
 317
 318 static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
 319 {
 320         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 321         struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
 322         struct device *dev = &nd_pfn->dev;
 323         struct vmem_altmap *altmap;
 324         struct nd_region *nd_region;
 325         struct nd_pfn_sb *pfn_sb;
 326         struct pmem_device *pmem;
 327         phys_addr_t offset;
 328         int rc;
 329
 330         if (!nd_pfn->uuid || !nd_pfn->ndns)
 331                 return -ENODEV;
 332
 333         nd_region = to_nd_region(dev->parent);
 334         rc = nd_pfn_init(nd_pfn);
 335         if (rc)
 336                 return rc;
 337
 338         if (PAGE_SIZE != SZ_4K) {
 339                 dev_err(dev, "only supported on systems with 4K PAGE_SIZE\n");
 340                 return -ENXIO;
 341         }
 342         if (nsio->res.start & ND_PFN_MASK) {
 343                 dev_err(dev, "%s not memory hotplug section aligned\n",
 344                                 dev_name(&ndns->dev));
 345                 return -ENXIO;
 346         }
 347
 348         pfn_sb = nd_pfn->pfn_sb;
 349         offset = le64_to_cpu(pfn_sb->dataoff);
 350         nd_pfn->mode = le32_to_cpu(nd_pfn->pfn_sb->mode);
 351         if (nd_pfn->mode == PFN_MODE_RAM) {
 352                 if (offset != SZ_8K)
 353                         return -EINVAL;
 354                 nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
 355                 altmap = NULL;
 356         } else {
 357                 rc = -ENXIO;
 358                 goto err;
 359         }
 360
 361         /* establish pfn range for lookup, and switch to direct map */
 362         pmem = dev_get_drvdata(dev);
 363         devm_memunmap(dev, (void __force *) pmem->virt_addr);
 364         pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, &nsio->res);
 365         if (IS_ERR(pmem->virt_addr)) {
 366                 rc = PTR_ERR(pmem->virt_addr);
 367                 goto err;
 368         }
 369
 370         /* attach pmem disk in "pfn-mode" */
 371         pmem->data_offset = offset;
 372         rc = pmem_attach_disk(dev, ndns, pmem);
 373         if (rc)
 374                 goto err;
 375
 376         return rc;
 377  err:
 378         nvdimm_namespace_detach_pfn(ndns);
 379         return rc;
 380 }
 381
 382 static int nd_pmem_probe(struct device *dev)
 383 {
 384         struct nd_region *nd_region = to_nd_region(dev->parent);
 385         struct nd_namespace_common *ndns;
 386         struct nd_namespace_io *nsio;
 387         struct pmem_device *pmem;
 388
 389         ndns = nvdimm_namespace_common_probe(dev);
 390         if (IS_ERR(ndns))
 391                 return PTR_ERR(ndns);
 392
 393         nsio = to_nd_namespace_io(&ndns->dev);
 394         pmem = pmem_alloc(dev, &nsio->res, nd_region->id);
 395         if (IS_ERR(pmem))
 396                 return PTR_ERR(pmem);
 397
 398         pmem->ndns = ndns;
 399         dev_set_drvdata(dev, pmem);
 400         ndns->rw_bytes = pmem_rw_bytes;
 401
 402         if (is_nd_btt(dev))
 403                 return nvdimm_namespace_attach_btt(ndns);
 404
 405         if (is_nd_pfn(dev))
 406                 return nvdimm_namespace_attach_pfn(ndns);
 407
 408         if (nd_btt_probe(ndns, pmem) == 0) {
 409                 /* we'll come back as btt-pmem */
 410                 return -ENXIO;
 411         }
 412
 413         if (nd_pfn_probe(ndns, pmem) == 0) {
 414                 /* we'll come back as pfn-pmem */
 415                 return -ENXIO;
 416         }
 417
 418         return pmem_attach_disk(dev, ndns, pmem);
 419 }
 420
 421 static int nd_pmem_remove(struct device *dev)
 422 {
 423         struct pmem_device *pmem = dev_get_drvdata(dev);
 424
 425         if (is_nd_btt(dev))
 426                 nvdimm_namespace_detach_btt(pmem->ndns);
 427         else if (is_nd_pfn(dev))
 428                 nvdimm_namespace_detach_pfn(pmem->ndns);
 429         else
 430                 pmem_detach_disk(pmem);
 431
 432         return 0;
 433 }
 434
 435 MODULE_ALIAS("pmem");
 436 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 437 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 438 static struct nd_device_driver nd_pmem_driver = {
 439         .probe = nd_pmem_probe,
 440         .remove = nd_pmem_remove,
 441         .drv = {
 442                 .name = "nd_pmem",
 443         },
 444         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 445 };
 446
 447 static int __init pmem_init(void)
 448 {
 449         int error;
 450
 451         pmem_major = register_blkdev(0, "pmem");
 452         if (pmem_major < 0)
 453                 return pmem_major;
 454
 455         error = nd_driver_register(&nd_pmem_driver);
 456         if (error) {
 457                 unregister_blkdev(pmem_major, "pmem");
 458                 return error;
 459         }
 460
 461         return 0;
 462 }
 463 module_init(pmem_init);
 464
 465 static void pmem_exit(void)
 466 {
 467         driver_unregister(&nd_pmem_driver.drv);
 468         unregister_blkdev(pmem_major, "pmem");
 469 }
 470 module_exit(pmem_exit);
 471
 472 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 473 MODULE_LICENSE("GPL v2");