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/slab.h>
  26 #include <linux/pmem.h>
  27 #include <linux/nd.h>
  28 #include "nd.h"
  29
  30 struct pmem_device {
  31         struct request_queue    *pmem_queue;
  32         struct gendisk          *pmem_disk;
  33
  34         /* One contiguous memory region per device */
  35         phys_addr_t             phys_addr;
  36         void __pmem             *virt_addr;
  37         size_t                  size;
  38 };
  39
  40 static int pmem_major;
  41
  42 static void pmem_do_bvec(struct pmem_device *pmem, struct page *page,
  43                         unsigned int len, unsigned int off, int rw,
  44                         sector_t sector)
  45 {
  46         void *mem = kmap_atomic(page);
  47         size_t pmem_off = sector << 9;
  48         void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
  49
  50         if (rw == READ) {
  51                 memcpy_from_pmem(mem + off, pmem_addr, len);
  52                 flush_dcache_page(page);
  53         } else {
  54                 flush_dcache_page(page);
  55                 memcpy_to_pmem(pmem_addr, mem + off, len);
  56         }
  57
  58         kunmap_atomic(mem);
  59 }
  60
  61 static void pmem_make_request(struct request_queue *q, struct bio *bio)
  62 {
  63         bool do_acct;
  64         unsigned long start;
  65         struct bio_vec bvec;
  66         struct bvec_iter iter;
  67         struct block_device *bdev = bio->bi_bdev;
  68         struct pmem_device *pmem = bdev->bd_disk->private_data;
  69
  70         do_acct = nd_iostat_start(bio, &start);
  71         bio_for_each_segment(bvec, bio, iter)
  72                 pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len, bvec.bv_offset,
  73                                 bio_data_dir(bio), iter.bi_sector);
  74         if (do_acct)
  75                 nd_iostat_end(bio, start);
  76
  77         if (bio_data_dir(bio))
  78                 wmb_pmem();
  79
  80         bio_endio(bio, 0);
  81 }
  82
  83 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
  84                        struct page *page, int rw)
  85 {
  86         struct pmem_device *pmem = bdev->bd_disk->private_data;
  87
  88         pmem_do_bvec(pmem, page, PAGE_CACHE_SIZE, 0, rw, sector);
  89         page_endio(page, rw & WRITE, 0);
  90
  91         return 0;
  92 }
  93
  94 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
  95                               void **kaddr, unsigned long *pfn, long size)
  96 {
  97         struct pmem_device *pmem = bdev->bd_disk->private_data;
  98         size_t offset = sector << 9;
  99
 100         if (!pmem)
 101                 return -ENODEV;
 102
 103         /* FIXME convert DAX to comprehend that this mapping has a lifetime */
 104         *kaddr = (void __force *) pmem->virt_addr + offset;
 105         *pfn = (pmem->phys_addr + offset) >> PAGE_SHIFT;
 106
 107         return pmem->size - offset;
 108 }
 109
 110 static const struct block_device_operations pmem_fops = {
 111         .owner =                THIS_MODULE,
 112         .rw_page =              pmem_rw_page,
 113         .direct_access =        pmem_direct_access,
 114         .revalidate_disk =      nvdimm_revalidate_disk,
 115 };
 116
 117 static struct pmem_device *pmem_alloc(struct device *dev,
 118                 struct resource *res, int id)
 119 {
 120         struct pmem_device *pmem;
 121
 122         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
 123         if (!pmem)
 124                 return ERR_PTR(-ENOMEM);
 125
 126         pmem->phys_addr = res->start;
 127         pmem->size = resource_size(res);
 128         if (!arch_has_pmem_api())
 129                 dev_warn(dev, "unable to guarantee persistence of writes\n");
 130
 131         if (!devm_request_mem_region(dev, pmem->phys_addr, pmem->size,
 132                         dev_name(dev))) {
 133                 dev_warn(dev, "could not reserve region [0x%pa:0x%zx]\n",
 134                                 &pmem->phys_addr, pmem->size);
 135                 return ERR_PTR(-EBUSY);
 136         }
 137
 138         pmem->virt_addr = memremap_pmem(dev, pmem->phys_addr, pmem->size);
 139         if (!pmem->virt_addr)
 140                 return ERR_PTR(-ENXIO);
 141
 142         return pmem;
 143 }
 144
 145 static void pmem_detach_disk(struct pmem_device *pmem)
 146 {
 147         del_gendisk(pmem->pmem_disk);
 148         put_disk(pmem->pmem_disk);
 149         blk_cleanup_queue(pmem->pmem_queue);
 150 }
 151
 152 static int pmem_attach_disk(struct nd_namespace_common *ndns,
 153                 struct pmem_device *pmem)
 154 {
 155         struct gendisk *disk;
 156
 157         pmem->pmem_queue = blk_alloc_queue(GFP_KERNEL);
 158         if (!pmem->pmem_queue)
 159                 return -ENOMEM;
 160
 161         blk_queue_make_request(pmem->pmem_queue, pmem_make_request);
 162         blk_queue_physical_block_size(pmem->pmem_queue, PAGE_SIZE);
 163         blk_queue_max_hw_sectors(pmem->pmem_queue, UINT_MAX);
 164         blk_queue_bounce_limit(pmem->pmem_queue, BLK_BOUNCE_ANY);
 165         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, pmem->pmem_queue);
 166
 167         disk = alloc_disk(0);
 168         if (!disk) {
 169                 blk_cleanup_queue(pmem->pmem_queue);
 170                 return -ENOMEM;
 171         }
 172
 173         disk->major             = pmem_major;
 174         disk->first_minor       = 0;
 175         disk->fops              = &pmem_fops;
 176         disk->private_data      = pmem;
 177         disk->queue             = pmem->pmem_queue;
 178         disk->flags             = GENHD_FL_EXT_DEVT;
 179         nvdimm_namespace_disk_name(ndns, disk->disk_name);
 180         disk->driverfs_dev = &ndns->dev;
 181         set_capacity(disk, pmem->size >> 9);
 182         pmem->pmem_disk = disk;
 183
 184         add_disk(disk);
 185         revalidate_disk(disk);
 186
 187         return 0;
 188 }
 189
 190 static int pmem_rw_bytes(struct nd_namespace_common *ndns,
 191                 resource_size_t offset, void *buf, size_t size, int rw)
 192 {
 193         struct pmem_device *pmem = dev_get_drvdata(ndns->claim);
 194
 195         if (unlikely(offset + size > pmem->size)) {
 196                 dev_WARN_ONCE(&ndns->dev, 1, "request out of range\n");
 197                 return -EFAULT;
 198         }
 199
 200         if (rw == READ)
 201                 memcpy_from_pmem(buf, pmem->virt_addr + offset, size);
 202         else {
 203                 memcpy_to_pmem(pmem->virt_addr + offset, buf, size);
 204                 wmb_pmem();
 205         }
 206
 207         return 0;
 208 }
 209
 210 static int nd_pmem_probe(struct device *dev)
 211 {
 212         struct nd_region *nd_region = to_nd_region(dev->parent);
 213         struct nd_namespace_common *ndns;
 214         struct nd_namespace_io *nsio;
 215         struct pmem_device *pmem;
 216
 217         ndns = nvdimm_namespace_common_probe(dev);
 218         if (IS_ERR(ndns))
 219                 return PTR_ERR(ndns);
 220
 221         nsio = to_nd_namespace_io(&ndns->dev);
 222         pmem = pmem_alloc(dev, &nsio->res, nd_region->id);
 223         if (IS_ERR(pmem))
 224                 return PTR_ERR(pmem);
 225
 226         dev_set_drvdata(dev, pmem);
 227         ndns->rw_bytes = pmem_rw_bytes;
 228
 229         if (is_nd_btt(dev))
 230                 return nvdimm_namespace_attach_btt(ndns);
 231
 232         if (nd_btt_probe(ndns, pmem) == 0)
 233                 /* we'll come back as btt-pmem */
 234                 return -ENXIO;
 235         return pmem_attach_disk(ndns, pmem);
 236 }
 237
 238 static int nd_pmem_remove(struct device *dev)
 239 {
 240         struct pmem_device *pmem = dev_get_drvdata(dev);
 241
 242         if (is_nd_btt(dev))
 243                 nvdimm_namespace_detach_btt(to_nd_btt(dev)->ndns);
 244         else
 245                 pmem_detach_disk(pmem);
 246
 247         return 0;
 248 }
 249
 250 MODULE_ALIAS("pmem");
 251 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
 252 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
 253 static struct nd_device_driver nd_pmem_driver = {
 254         .probe = nd_pmem_probe,
 255         .remove = nd_pmem_remove,
 256         .drv = {
 257                 .name = "nd_pmem",
 258         },
 259         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
 260 };
 261
 262 static int __init pmem_init(void)
 263 {
 264         int error;
 265
 266         pmem_major = register_blkdev(0, "pmem");
 267         if (pmem_major < 0)
 268                 return pmem_major;
 269
 270         error = nd_driver_register(&nd_pmem_driver);
 271         if (error) {
 272                 unregister_blkdev(pmem_major, "pmem");
 273                 return error;
 274         }
 275
 276         return 0;
 277 }
 278 module_init(pmem_init);
 279
 280 static void pmem_exit(void)
 281 {
 282         driver_unregister(&nd_pmem_driver.drv);
 283         unregister_blkdev(pmem_major, "pmem");
 284 }
 285 module_exit(pmem_exit);
 286
 287 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
 288 MODULE_LICENSE("GPL v2");