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 "pfn.h"
  33 #include "nd.h"
  34
  35 struct pmem_device {
  36         /* One contiguous memory region per device */
  37         phys_addr_t             phys_addr;
  38         /* when non-zero this device is hosting a 'pfn' instance */
  39         phys_addr_t             data_offset;
  40         u64                     pfn_flags;
  41         void __pmem             *virt_addr;
  42         /* immutable base size of the namespace */
  43         size_t                  size;
  44         /* trim size when namespace capacity has been section aligned */
  45         u32                     pfn_pad;
  46         struct badblocks        bb;
  47 };
  48
  49 static void pmem_clear_poison(struct pmem_device *pmem, phys_addr_t offset,
  50                 unsigned int len)
  51 {
  52         struct device *dev = pmem->bb.dev;
  53         sector_t sector;
  54         long cleared;
  55
  56         sector = (offset - pmem->data_offset) / 512;
  57         cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
  58
  59         if (cleared > 0 && cleared / 512) {
  60                 dev_dbg(dev, "%s: %llx clear %ld sector%s\n",
  61                                 __func__, (unsigned long long) sector,
  62                                 cleared / 512, cleared / 512 > 1 ? "s" : "");
  63                 badblocks_clear(&pmem->bb, sector, cleared / 512);
  64         }
  65         invalidate_pmem(pmem->virt_addr + offset, len);
  66 }
  67
  68 static int pmem_do_bvec(struct pmem_device *pmem, struct page *page,
  69                         unsigned int len, unsigned int off, int rw,
  70                         sector_t sector)
  71 {
  72         int rc = 0;
  73         bool bad_pmem = false;
  74         void *mem = kmap_atomic(page);
  75         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
  76         void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
  77
  78         if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
  79                 bad_pmem = true;
  80
  81         if (rw == READ) {
  82                 if (unlikely(bad_pmem))
  83                         rc = -EIO;
  84                 else {
  85                         rc = memcpy_from_pmem(mem + off, pmem_addr, len);
  86                         flush_dcache_page(page);
  87                 }
  88         } else {
  89                 /*
  90                  * Note that we write the data both before and after
  91                  * clearing poison.  The write before clear poison
  92                  * handles situations where the latest written data is
  93                  * preserved and the clear poison operation simply marks
  94                  * the address range as valid without changing the data.
  95                  * In this case application software can assume that an
  96                  * interrupted write will either return the new good
  97                  * data or an error.
  98                  *
  99                  * However, if pmem_clear_poison() leaves the data in an
 100                  * indeterminate state we need to perform the write
 101                  * after clear poison.
 102                  */
 103                 flush_dcache_page(page);
 104                 memcpy_to_pmem(pmem_addr, mem + off, len);
 105                 if (unlikely(bad_pmem)) {
 106                         pmem_clear_poison(pmem, pmem_off, len);
 107                         memcpy_to_pmem(pmem_addr, mem + off, len);
 108                 }
 109         }
 110
 111         kunmap_atomic(mem);
 112         return rc;
 113 }
 114
 115 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
 116 {
 117         int rc = 0;
 118         bool do_acct;
 119         unsigned long start;
 120         struct bio_vec bvec;
 121         struct bvec_iter iter;
 122         struct pmem_device *pmem = q->queuedata;
 123
 124         do_acct = nd_iostat_start(bio, &start);
 125         bio_for_each_segment(bvec, bio, iter) {
 126                 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
 127                                 bvec.bv_offset, bio_data_dir(bio),
 128                                 iter.bi_sector);
 129                 if (rc) {
 130                         bio->bi_error = rc;
 131                         break;
 132                 }
 133         }
 134         if (do_acct)
 135                 nd_iostat_end(bio, start);
 136
 137         if (bio_data_dir(bio))
 138                 wmb_pmem();
 139
 140         bio_endio(bio);
 141         return BLK_QC_T_NONE;
 142 }
 143
 144 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 145                        struct page *page, int rw)
 146 {
 147         struct pmem_device *pmem = bdev->bd_queue->queuedata;
 148         int rc;
 149
 150         rc = pmem_do_bvec(pmem, page, PAGE_SIZE, 0, rw, sector);
 151         if (rw & WRITE)
 152                 wmb_pmem();
 153
 154         /*
 155          * The ->rw_page interface is subtle and tricky.  The core
 156          * retries on any error, so we can only invoke page_endio() in
 157          * the successful completion case.  Otherwise, we'll see crashes
 158          * caused by double completion.
 159          */
 160         if (rc == 0)
 161                 page_endio(page, rw & WRITE, 0);
 162
 163         return rc;
 164 }
 165
 166 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
 167                       void __pmem **kaddr, pfn_t *pfn, long size)
 168 {
 169         struct pmem_device *pmem = bdev->bd_queue->queuedata;
 170         resource_size_t offset = sector * 512 + pmem->data_offset;
 171
 172         if (unlikely(is_bad_pmem(&pmem->bb, sector, size)))
 173                 return -EIO;
 174         *kaddr = pmem->virt_addr + offset;
 175         *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
 176
 177         /*
 178          * If badblocks are present, limit known good range to the
 179          * requested range.
 180          */
 181         if (unlikely(pmem->bb.count))
 182                 return size;
 183         return pmem->size - pmem->pfn_pad - offset;
 184 }
 185
 186 static const struct block_device_operations pmem_fops = {
 187         .owner =                THIS_MODULE,
 188         .rw_page =              pmem_rw_page,
 189         .direct_access =        pmem_direct_access,
 190         .revalidate_disk =      nvdimm_revalidate_disk,
 191 };
 192
 193 static void pmem_release_queue(void *q)
 194 {
 195         blk_cleanup_queue(q);
 196 }
 197
 198 void pmem_release_disk(void *disk)
 199 {
 200         del_gendisk(disk);
 201         put_disk(disk);
 202 }
 203
 204 static int pmem_attach_disk(struct device *dev,
 205                 struct nd_namespace_common *ndns)
 206 {
 207         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
 208         struct vmem_altmap __altmap, *altmap = NULL;
 209         struct resource *res = &nsio->res;
 210         struct nd_pfn *nd_pfn = NULL;
 211         int nid = dev_to_node(dev);
 212         struct nd_pfn_sb *pfn_sb;
 213         struct pmem_device *pmem;
 214         struct resource pfn_res;
 215         struct request_queue *q;
 216         struct gendisk *disk;
 217         void *addr;
 218
 219         /* while nsio_rw_bytes is active, parse a pfn info block if present */
 220         if (is_nd_pfn(dev)) {
 221                 nd_pfn = to_nd_pfn(dev);
 222                 altmap = nvdimm_setup_pfn(nd_pfn, &pfn_res, &__altmap);
 223                 if (IS_ERR(altmap))
 224                         return PTR_ERR(altmap);
 225         }
 226
 227         /* we're attaching a block device, disable raw namespace access */
 228         devm_nsio_disable(dev, nsio);
 229
 230         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 231         if (!pmem)
 232                 return -ENOMEM;
 233
 234         dev_set_drvdata(dev, pmem);
 235         pmem->phys_addr = res->start;
 236         pmem->size = resource_size(res);
 237         if (!arch_has_wmb_pmem())
 238                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 239
 240         if (!devm_request_mem_region(dev, res->start, resource_size(res),
 241                                 dev_name(dev))) {
 242                 dev_warn(dev, "could not reserve region %pR\n", res);
 243                 return -EBUSY;
 244         }
 245
 246         q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
 247         if (!q)
 248                 return -ENOMEM;
 249
 250         pmem->pfn_flags = PFN_DEV;
 251         if (is_nd_pfn(dev)) {
 252                 addr = devm_memremap_pages(dev, &pfn_res, &q->q_usage_counter,
 253                                 altmap);
 254                 pfn_sb = nd_pfn->pfn_sb;
 255                 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
 256                 pmem->pfn_pad = resource_size(res) - resource_size(&pfn_res);
 257                 pmem->pfn_flags |= PFN_MAP;
 258                 res = &pfn_res; /* for badblocks populate */
 259                 res->start += pmem->data_offset;
 260         } else if (pmem_should_map_pages(dev)) {
 261                 addr = devm_memremap_pages(dev, &nsio->res,
 262                                 &q->q_usage_counter, NULL);
 263                 pmem->pfn_flags |= PFN_MAP;
 264         } else
 265                 addr = devm_memremap(dev, pmem->phys_addr,
 266                                 pmem->size, ARCH_MEMREMAP_PMEM);
 267
 268         /*
 269          * At release time the queue must be dead before
 270          * devm_memremap_pages is unwound
 271          */
 272         if (devm_add_action(dev, pmem_release_queue, q)) {
 273                 blk_cleanup_queue(q);
 274                 return -ENOMEM;
 275         }
 276
 277         if (IS_ERR(addr))
 278                 return PTR_ERR(addr);
 279         pmem->virt_addr = (void __pmem *) addr;
 280
 281         blk_queue_make_request(q, pmem_make_request);
 282         blk_queue_physical_block_size(q, PAGE_SIZE);
 283         blk_queue_max_hw_sectors(q, UINT_MAX);
 284         blk_queue_bounce_limit(q, BLK_BOUNCE_ANY);
 285         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
 286         queue_flag_set_unlocked(QUEUE_FLAG_DAX, q);
 287         q->queuedata = pmem;
 288
 289         disk = alloc_disk_node(0, nid);
 290         if (!disk)
 291                 return -ENOMEM;
 292         if (devm_add_action(dev, pmem_release_disk, disk)) {
 293                 put_disk(disk);
 294                 return -ENOMEM;
 295         }
 296
 297         disk->fops              = &pmem_fops;
 298         disk->queue             = q;
 299         disk->flags             = GENHD_FL_EXT_DEVT;
 300         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 301         disk->driverfs_dev = dev;
 302         set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
 303                         / 512);
 304         if (devm_init_badblocks(dev, &pmem->bb))
 305                 return -ENOMEM;
 306         nvdimm_badblocks_populate(to_nd_region(dev->parent), &pmem->bb, res);
 307         disk->bb = &pmem->bb;
 308         add_disk(disk);
 309         revalidate_disk(disk);
 310
 311         return 0;
 312 }
 313
 314 static int nd_pmem_probe(struct device *dev)
 315 {
 316         struct nd_namespace_common *ndns;
 317
 318         ndns = nvdimm_namespace_common_probe(dev);
 319         if (IS_ERR(ndns))
 320                 return PTR_ERR(ndns);
 321
 322         if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
 323                 return -ENXIO;
 324
 325         if (is_nd_btt(dev))
 326                 return nvdimm_namespace_attach_btt(ndns);
 327
 328         if (is_nd_pfn(dev))
 329                 return pmem_attach_disk(dev, ndns);
 330
 331         /* if we find a valid info-block we'll come back as that personality */
 332         if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
 333                         || nd_dax_probe(dev, ndns) == 0)
 334                 return -ENXIO;
 335
 336         /* ...otherwise we're just a raw pmem device */
 337         return pmem_attach_disk(dev, ndns);
 338 }
 339
 340 static int nd_pmem_remove(struct device *dev)
 341 {
 342         if (is_nd_btt(dev))
 343                 nvdimm_namespace_detach_btt(to_nd_btt(dev));
 344         return 0;
 345 }
 346
 347 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
 348 {
 349         struct nd_region *nd_region = to_nd_region(dev->parent);
 350         struct pmem_device *pmem = dev_get_drvdata(dev);
 351         resource_size_t offset = 0, end_trunc = 0;
 352         struct nd_namespace_common *ndns;
 353         struct nd_namespace_io *nsio;
 354         struct resource res;
 355
 356         if (event != NVDIMM_REVALIDATE_POISON)
 357                 return;
 358
 359         if (is_nd_btt(dev)) {
 360                 struct nd_btt *nd_btt = to_nd_btt(dev);
 361
 362                 ndns = nd_btt->ndns;
 363         } else if (is_nd_pfn(dev)) {
 364                 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
 365                 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 366
 367                 ndns = nd_pfn->ndns;
 368                 offset = pmem->data_offset + __le32_to_cpu(pfn_sb->start_pad);
 369                 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
 370         } else
 371                 ndns = to_ndns(dev);
 372
 373         nsio = to_nd_namespace_io(&ndns->dev);
 374         res.start = nsio->res.start + offset;
 375         res.end = nsio->res.end - end_trunc;
 376         nvdimm_badblocks_populate(nd_region, &pmem->bb, &res);
 377 }
 378
 379 MODULE_ALIAS("pmem");
 380 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 381 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 382 static struct nd_device_driver nd_pmem_driver = {
 383         .probe = nd_pmem_probe,
 384         .remove = nd_pmem_remove,
 385         .notify = nd_pmem_notify,
 386         .drv = {
 387                 .name = "nd_pmem",
 388         },
 389         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 390 };
 391
 392 static int __init pmem_init(void)
 393 {
 394         return nd_driver_register(&nd_pmem_driver);
 395 }
 396 module_init(pmem_init);
 397
 398 static void pmem_exit(void)
 399 {
 400         driver_unregister(&nd_pmem_driver.drv);
 401 }
 402 module_exit(pmem_exit);
 403
 404 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 405 MODULE_LICENSE("GPL v2");