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/blk-mq.h> 29 #include <linux/pfn_t.h> 30 #include <linux/slab.h> 31 #include <linux/uio.h> 32 #include <linux/dax.h> 33 #include <linux/nd.h> 34 #include <linux/backing-dev.h> 35 #include "pmem.h" 36 #include "pfn.h" 37 #include "nd.h" 38 #include "nd-core.h" 39 40 static struct device *to_dev(struct pmem_device *pmem) 41 { 42 /* 43 * nvdimm bus services need a 'dev' parameter, and we record the device 44 * at init in bb.dev. 45 */ 46 return pmem->bb.dev; 47 } 48 49 static struct nd_region *to_region(struct pmem_device *pmem) 50 { 51 return to_nd_region(to_dev(pmem)->parent); 52 } 53 54 static blk_status_t pmem_clear_poison(struct pmem_device *pmem, 55 phys_addr_t offset, unsigned int len) 56 { 57 struct device *dev = to_dev(pmem); 58 sector_t sector; 59 long cleared; 60 blk_status_t rc = BLK_STS_OK; 61 62 sector = (offset - pmem->data_offset) / 512; 63 64 cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len); 65 if (cleared < len) 66 rc = BLK_STS_IOERR; 67 if (cleared > 0 && cleared / 512) { 68 cleared /= 512; 69 dev_dbg(dev, "%s: %#llx clear %ld sector%s\n", __func__, 70 (unsigned long long) sector, cleared, 71 cleared > 1 ? "s" : ""); 72 badblocks_clear(&pmem->bb, sector, cleared); 73 if (pmem->bb_state) 74 sysfs_notify_dirent(pmem->bb_state); 75 } 76 77 arch_invalidate_pmem(pmem->virt_addr + offset, len); 78 79 return rc; 80 } 81 82 static void write_pmem(void *pmem_addr, struct page *page, 83 unsigned int off, unsigned int len) 84 { 85 unsigned int chunk; 86 void *mem; 87 88 while (len) { 89 mem = kmap_atomic(page); 90 chunk = min_t(unsigned int, len, PAGE_SIZE); 91 memcpy_flushcache(pmem_addr, mem + off, chunk); 92 kunmap_atomic(mem); 93 len -= chunk; 94 off = 0; 95 page++; 96 pmem_addr += PAGE_SIZE; 97 } 98 } 99 100 static blk_status_t read_pmem(struct page *page, unsigned int off, 101 void *pmem_addr, unsigned int len) 102 { 103 unsigned int chunk; 104 int rc; 105 void *mem; 106 107 while (len) { 108 mem = kmap_atomic(page); 109 chunk = min_t(unsigned int, len, PAGE_SIZE); 110 rc = memcpy_mcsafe(mem + off, pmem_addr, chunk); 111 kunmap_atomic(mem); 112 if (rc) 113 return BLK_STS_IOERR; 114 len -= chunk; 115 off = 0; 116 page++; 117 pmem_addr += PAGE_SIZE; 118 } 119 return BLK_STS_OK; 120 } 121 122 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page, 123 unsigned int len, unsigned int off, bool is_write, 124 sector_t sector) 125 { 126 blk_status_t rc = BLK_STS_OK; 127 bool bad_pmem = false; 128 phys_addr_t pmem_off = sector * 512 + pmem->data_offset; 129 void *pmem_addr = pmem->virt_addr + pmem_off; 130 131 if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) 132 bad_pmem = true; 133 134 if (!is_write) { 135 if (unlikely(bad_pmem)) 136 rc = BLK_STS_IOERR; 137 else { 138 rc = read_pmem(page, off, pmem_addr, len); 139 flush_dcache_page(page); 140 } 141 } else { 142 /* 143 * Note that we write the data both before and after 144 * clearing poison. The write before clear poison 145 * handles situations where the latest written data is 146 * preserved and the clear poison operation simply marks 147 * the address range as valid without changing the data. 148 * In this case application software can assume that an 149 * interrupted write will either return the new good 150 * data or an error. 151 * 152 * However, if pmem_clear_poison() leaves the data in an 153 * indeterminate state we need to perform the write 154 * after clear poison. 155 */ 156 flush_dcache_page(page); 157 write_pmem(pmem_addr, page, off, len); 158 if (unlikely(bad_pmem)) { 159 rc = pmem_clear_poison(pmem, pmem_off, len); 160 write_pmem(pmem_addr, page, off, len); 161 } 162 } 163 164 return rc; 165 } 166 167 /* account for REQ_FLUSH rename, replace with REQ_PREFLUSH after v4.8-rc1 */ 168 #ifndef REQ_FLUSH 169 #define REQ_FLUSH REQ_PREFLUSH 170 #endif 171 172 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio) 173 { 174 blk_status_t rc = 0; 175 bool do_acct; 176 unsigned long start; 177 struct bio_vec bvec; 178 struct bvec_iter iter; 179 struct pmem_device *pmem = q->queuedata; 180 struct nd_region *nd_region = to_region(pmem); 181 182 if (bio->bi_opf & REQ_FLUSH) 183 nvdimm_flush(nd_region); 184 185 do_acct = nd_iostat_start(bio, &start); 186 bio_for_each_segment(bvec, bio, iter) { 187 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len, 188 bvec.bv_offset, op_is_write(bio_op(bio)), 189 iter.bi_sector); 190 if (rc) { 191 bio->bi_status = rc; 192 break; 193 } 194 } 195 if (do_acct) 196 nd_iostat_end(bio, start); 197 198 if (bio->bi_opf & REQ_FUA) 199 nvdimm_flush(nd_region); 200 201 bio_endio(bio); 202 return BLK_QC_T_NONE; 203 } 204 205 static int pmem_rw_page(struct block_device *bdev, sector_t sector, 206 struct page *page, bool is_write) 207 { 208 struct pmem_device *pmem = bdev->bd_queue->queuedata; 209 blk_status_t rc; 210 211 rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE, 212 0, is_write, sector); 213 214 /* 215 * The ->rw_page interface is subtle and tricky. The core 216 * retries on any error, so we can only invoke page_endio() in 217 * the successful completion case. Otherwise, we'll see crashes 218 * caused by double completion. 219 */ 220 if (rc == 0) 221 page_endio(page, is_write, 0); 222 223 return blk_status_to_errno(rc); 224 } 225 226 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */ 227 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff, 228 long nr_pages, void **kaddr, pfn_t *pfn) 229 { 230 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset; 231 232 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512, 233 PFN_PHYS(nr_pages)))) 234 return -EIO; 235 *kaddr = pmem->virt_addr + offset; 236 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags); 237 238 /* 239 * If badblocks are present, limit known good range to the 240 * requested range. 241 */ 242 if (unlikely(pmem->bb.count)) 243 return nr_pages; 244 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset); 245 } 246 247 static const struct block_device_operations pmem_fops = { 248 .owner = THIS_MODULE, 249 .rw_page = pmem_rw_page, 250 .revalidate_disk = nvdimm_revalidate_disk, 251 }; 252 253 static long pmem_dax_direct_access(struct dax_device *dax_dev, 254 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn) 255 { 256 struct pmem_device *pmem = dax_get_private(dax_dev); 257 258 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn); 259 } 260 261 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, 262 void *addr, size_t bytes, struct iov_iter *i) 263 { 264 return copy_from_iter_flushcache(addr, bytes, i); 265 } 266 267 static const struct dax_operations pmem_dax_ops = { 268 .direct_access = pmem_dax_direct_access, 269 .copy_from_iter = pmem_copy_from_iter, 270 }; 271 272 static const struct attribute_group *pmem_attribute_groups[] = { 273 &dax_attribute_group, 274 NULL, 275 }; 276 277 static void pmem_release_queue(void *q) 278 { 279 blk_cleanup_queue(q); 280 } 281 282 static void pmem_freeze_queue(void *q) 283 { 284 blk_freeze_queue_start(q); 285 } 286 287 static void pmem_release_disk(void *__pmem) 288 { 289 struct pmem_device *pmem = __pmem; 290 291 kill_dax(pmem->dax_dev); 292 put_dax(pmem->dax_dev); 293 del_gendisk(pmem->disk); 294 put_disk(pmem->disk); 295 } 296 297 static int pmem_attach_disk(struct device *dev, 298 struct nd_namespace_common *ndns) 299 { 300 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev); 301 struct nd_region *nd_region = to_nd_region(dev->parent); 302 int nid = dev_to_node(dev), fua, wbc; 303 struct resource *res = &nsio->res; 304 struct resource bb_res; 305 struct nd_pfn *nd_pfn = NULL; 306 struct dax_device *dax_dev; 307 struct nd_pfn_sb *pfn_sb; 308 struct pmem_device *pmem; 309 struct request_queue *q; 310 struct device *gendev; 311 struct gendisk *disk; 312 void *addr; 313 int rc; 314 315 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL); 316 if (!pmem) 317 return -ENOMEM; 318 319 /* while nsio_rw_bytes is active, parse a pfn info block if present */ 320 if (is_nd_pfn(dev)) { 321 nd_pfn = to_nd_pfn(dev); 322 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap); 323 if (rc) 324 return rc; 325 } 326 327 /* we're attaching a block device, disable raw namespace access */ 328 devm_nsio_disable(dev, nsio); 329 330 dev_set_drvdata(dev, pmem); 331 pmem->phys_addr = res->start; 332 pmem->size = resource_size(res); 333 fua = nvdimm_has_flush(nd_region); 334 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) { 335 dev_warn(dev, "unable to guarantee persistence of writes\n"); 336 fua = 0; 337 } 338 wbc = nvdimm_has_cache(nd_region); 339 340 if (!devm_request_mem_region(dev, res->start, resource_size(res), 341 dev_name(&ndns->dev))) { 342 dev_warn(dev, "could not reserve region %pR\n", res); 343 return -EBUSY; 344 } 345 346 q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev)); 347 if (!q) 348 return -ENOMEM; 349 350 if (devm_add_action_or_reset(dev, pmem_release_queue, q)) 351 return -ENOMEM; 352 353 pmem->pfn_flags = PFN_DEV; 354 pmem->pgmap.ref = &q->q_usage_counter; 355 if (is_nd_pfn(dev)) { 356 addr = devm_memremap_pages(dev, &pmem->pgmap); 357 pfn_sb = nd_pfn->pfn_sb; 358 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff); 359 pmem->pfn_pad = resource_size(res) - 360 resource_size(&pmem->pgmap.res); 361 pmem->pfn_flags |= PFN_MAP; 362 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res)); 363 bb_res.start += pmem->data_offset; 364 } else if (pmem_should_map_pages(dev)) { 365 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res)); 366 pmem->pgmap.altmap_valid = false; 367 addr = devm_memremap_pages(dev, &pmem->pgmap); 368 pmem->pfn_flags |= PFN_MAP; 369 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res)); 370 } else 371 addr = devm_memremap(dev, pmem->phys_addr, 372 pmem->size, ARCH_MEMREMAP_PMEM); 373 374 /* 375 * At release time the queue must be frozen before 376 * devm_memremap_pages is unwound 377 */ 378 if (devm_add_action_or_reset(dev, pmem_freeze_queue, q)) 379 return -ENOMEM; 380 381 if (IS_ERR(addr)) 382 return PTR_ERR(addr); 383 pmem->virt_addr = addr; 384 385 blk_queue_write_cache(q, wbc, fua); 386 blk_queue_make_request(q, pmem_make_request); 387 blk_queue_physical_block_size(q, PAGE_SIZE); 388 blk_queue_logical_block_size(q, pmem_sector_size(ndns)); 389 blk_queue_max_hw_sectors(q, UINT_MAX); 390 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q); 391 queue_flag_set_unlocked(QUEUE_FLAG_DAX, q); 392 q->queuedata = pmem; 393 394 disk = alloc_disk_node(0, nid); 395 if (!disk) 396 return -ENOMEM; 397 pmem->disk = disk; 398 399 disk->fops = &pmem_fops; 400 disk->queue = q; 401 disk->flags = GENHD_FL_EXT_DEVT; 402 disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO; 403 nvdimm_namespace_disk_name(ndns, disk->disk_name); 404 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset) 405 / 512); 406 if (devm_init_badblocks(dev, &pmem->bb)) 407 return -ENOMEM; 408 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res); 409 disk->bb = &pmem->bb; 410 411 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops); 412 if (!dax_dev) { 413 put_disk(disk); 414 return -ENOMEM; 415 } 416 dax_write_cache(dax_dev, wbc); 417 pmem->dax_dev = dax_dev; 418 419 gendev = disk_to_dev(disk); 420 gendev->groups = pmem_attribute_groups; 421 422 device_add_disk(dev, disk); 423 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem)) 424 return -ENOMEM; 425 426 revalidate_disk(disk); 427 428 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd, 429 "badblocks"); 430 if (!pmem->bb_state) 431 dev_warn(dev, "'badblocks' notification disabled\n"); 432 433 return 0; 434 } 435 436 static int nd_pmem_probe(struct device *dev) 437 { 438 struct nd_namespace_common *ndns; 439 440 ndns = nvdimm_namespace_common_probe(dev); 441 if (IS_ERR(ndns)) 442 return PTR_ERR(ndns); 443 444 if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev))) 445 return -ENXIO; 446 447 if (is_nd_btt(dev)) 448 return nvdimm_namespace_attach_btt(ndns); 449 450 if (is_nd_pfn(dev)) 451 return pmem_attach_disk(dev, ndns); 452 453 /* if we find a valid info-block we'll come back as that personality */ 454 if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0 455 || nd_dax_probe(dev, ndns) == 0) 456 return -ENXIO; 457 458 /* ...otherwise we're just a raw pmem device */ 459 return pmem_attach_disk(dev, ndns); 460 } 461 462 static int nd_pmem_remove(struct device *dev) 463 { 464 struct pmem_device *pmem = dev_get_drvdata(dev); 465 466 if (is_nd_btt(dev)) 467 nvdimm_namespace_detach_btt(to_nd_btt(dev)); 468 else { 469 /* 470 * Note, this assumes device_lock() context to not race 471 * nd_pmem_notify() 472 */ 473 sysfs_put(pmem->bb_state); 474 pmem->bb_state = NULL; 475 } 476 nvdimm_flush(to_nd_region(dev->parent)); 477 478 return 0; 479 } 480 481 static void nd_pmem_shutdown(struct device *dev) 482 { 483 nvdimm_flush(to_nd_region(dev->parent)); 484 } 485 486 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event) 487 { 488 struct nd_region *nd_region; 489 resource_size_t offset = 0, end_trunc = 0; 490 struct nd_namespace_common *ndns; 491 struct nd_namespace_io *nsio; 492 struct resource res; 493 struct badblocks *bb; 494 struct kernfs_node *bb_state; 495 496 if (event != NVDIMM_REVALIDATE_POISON) 497 return; 498 499 if (is_nd_btt(dev)) { 500 struct nd_btt *nd_btt = to_nd_btt(dev); 501 502 ndns = nd_btt->ndns; 503 nd_region = to_nd_region(ndns->dev.parent); 504 nsio = to_nd_namespace_io(&ndns->dev); 505 bb = &nsio->bb; 506 bb_state = NULL; 507 } else { 508 struct pmem_device *pmem = dev_get_drvdata(dev); 509 510 nd_region = to_region(pmem); 511 bb = &pmem->bb; 512 bb_state = pmem->bb_state; 513 514 if (is_nd_pfn(dev)) { 515 struct nd_pfn *nd_pfn = to_nd_pfn(dev); 516 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; 517 518 ndns = nd_pfn->ndns; 519 offset = pmem->data_offset + 520 __le32_to_cpu(pfn_sb->start_pad); 521 end_trunc = __le32_to_cpu(pfn_sb->end_trunc); 522 } else { 523 ndns = to_ndns(dev); 524 } 525 526 nsio = to_nd_namespace_io(&ndns->dev); 527 } 528 529 res.start = nsio->res.start + offset; 530 res.end = nsio->res.end - end_trunc; 531 nvdimm_badblocks_populate(nd_region, bb, &res); 532 if (bb_state) 533 sysfs_notify_dirent(bb_state); 534 } 535 536 MODULE_ALIAS("pmem"); 537 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO); 538 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM); 539 static struct nd_device_driver nd_pmem_driver = { 540 .probe = nd_pmem_probe, 541 .remove = nd_pmem_remove, 542 .notify = nd_pmem_notify, 543 .shutdown = nd_pmem_shutdown, 544 .drv = { 545 .name = "nd_pmem", 546 }, 547 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM, 548 }; 549 550 static int __init pmem_init(void) 551 { 552 return nd_driver_register(&nd_pmem_driver); 553 } 554 module_init(pmem_init); 555 556 static void pmem_exit(void) 557 { 558 driver_unregister(&nd_pmem_driver.drv); 559 } 560 module_exit(pmem_exit); 561 562 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>"); 563 MODULE_LICENSE("GPL v2"); 564