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