1// SPDX-License-Identifier: GPL-2.0-only 2/* 3 * Persistent Memory Driver 4 * 5 * Copyright (c) 2014-2015, Intel Corporation. 6 * Copyright (c) 2015, Christoph Hellwig <hch@lst.de>. 7 * Copyright (c) 2015, Boaz Harrosh <boaz@plexistor.com>. 8 */ 9 10#include <linux/blkdev.h> 11#include <linux/hdreg.h> 12#include <linux/init.h> 13#include <linux/platform_device.h> 14#include <linux/set_memory.h> 15#include <linux/module.h> 16#include <linux/moduleparam.h> 17#include <linux/badblocks.h> 18#include <linux/memremap.h> 19#include <linux/vmalloc.h> 20#include <linux/blk-mq.h> 21#include <linux/pfn_t.h> 22#include <linux/slab.h> 23#include <linux/uio.h> 24#include <linux/dax.h> 25#include <linux/nd.h> 26#include <linux/backing-dev.h> 27#include <linux/mm.h> 28#include <asm/cacheflush.h> 29#include "pmem.h" 30#include "pfn.h" 31#include "nd.h" 32 33static struct device *to_dev(struct pmem_device *pmem) 34{ 35 /* 36 * nvdimm bus services need a 'dev' parameter, and we record the device 37 * at init in bb.dev. 38 */ 39 return pmem->bb.dev; 40} 41 42static struct nd_region *to_region(struct pmem_device *pmem) 43{ 44 return to_nd_region(to_dev(pmem)->parent); 45} 46 47static void hwpoison_clear(struct pmem_device *pmem, 48 phys_addr_t phys, unsigned int len) 49{ 50 unsigned long pfn_start, pfn_end, pfn; 51 52 /* only pmem in the linear map supports HWPoison */ 53 if (is_vmalloc_addr(pmem->virt_addr)) 54 return; 55 56 pfn_start = PHYS_PFN(phys); 57 pfn_end = pfn_start + PHYS_PFN(len); 58 for (pfn = pfn_start; pfn < pfn_end; pfn++) { 59 struct page *page = pfn_to_page(pfn); 60 61 /* 62 * Note, no need to hold a get_dev_pagemap() reference 63 * here since we're in the driver I/O path and 64 * outstanding I/O requests pin the dev_pagemap. 65 */ 66 if (test_and_clear_pmem_poison(page)) 67 clear_mce_nospec(pfn); 68 } 69} 70 71static blk_status_t pmem_clear_poison(struct pmem_device *pmem, 72 phys_addr_t offset, unsigned int len) 73{ 74 struct device *dev = to_dev(pmem); 75 sector_t sector; 76 long cleared; 77 blk_status_t rc = BLK_STS_OK; 78 79 sector = (offset - pmem->data_offset) / 512; 80 81 cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len); 82 if (cleared < len) 83 rc = BLK_STS_IOERR; 84 if (cleared > 0 && cleared / 512) { 85 hwpoison_clear(pmem, pmem->phys_addr + offset, cleared); 86 cleared /= 512; 87 dev_dbg(dev, "%#llx clear %ld sector%s\n", 88 (unsigned long long) sector, cleared, 89 cleared > 1 ? "s" : ""); 90 badblocks_clear(&pmem->bb, sector, cleared); 91 if (pmem->bb_state) 92 sysfs_notify_dirent(pmem->bb_state); 93 } 94 95 arch_invalidate_pmem(pmem->virt_addr + offset, len); 96 97 return rc; 98} 99 100static void write_pmem(void *pmem_addr, struct page *page, 101 unsigned int off, unsigned int len) 102{ 103 unsigned int chunk; 104 void *mem; 105 106 while (len) { 107 mem = kmap_atomic(page); 108 chunk = min_t(unsigned int, len, PAGE_SIZE - off); 109 memcpy_flushcache(pmem_addr, mem + off, chunk); 110 kunmap_atomic(mem); 111 len -= chunk; 112 off = 0; 113 page++; 114 pmem_addr += chunk; 115 } 116} 117 118static blk_status_t read_pmem(struct page *page, unsigned int off, 119 void *pmem_addr, unsigned int len) 120{ 121 unsigned int chunk; 122 unsigned long rem; 123 void *mem; 124 125 while (len) { 126 mem = kmap_atomic(page); 127 chunk = min_t(unsigned int, len, PAGE_SIZE - off); 128 rem = copy_mc_to_kernel(mem + off, pmem_addr, chunk); 129 kunmap_atomic(mem); 130 if (rem) 131 return BLK_STS_IOERR; 132 len -= chunk; 133 off = 0; 134 page++; 135 pmem_addr += chunk; 136 } 137 return BLK_STS_OK; 138} 139 140static blk_status_t pmem_do_read(struct pmem_device *pmem, 141 struct page *page, unsigned int page_off, 142 sector_t sector, unsigned int len) 143{ 144 blk_status_t rc; 145 phys_addr_t pmem_off = sector * 512 + pmem->data_offset; 146 void *pmem_addr = pmem->virt_addr + pmem_off; 147 148 if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) 149 return BLK_STS_IOERR; 150 151 rc = read_pmem(page, page_off, pmem_addr, len); 152 flush_dcache_page(page); 153 return rc; 154} 155 156static blk_status_t pmem_do_write(struct pmem_device *pmem, 157 struct page *page, unsigned int page_off, 158 sector_t sector, unsigned int len) 159{ 160 blk_status_t rc = BLK_STS_OK; 161 bool bad_pmem = false; 162 phys_addr_t pmem_off = sector * 512 + pmem->data_offset; 163 void *pmem_addr = pmem->virt_addr + pmem_off; 164 165 if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) 166 bad_pmem = true; 167 168 /* 169 * Note that we write the data both before and after 170 * clearing poison. The write before clear poison 171 * handles situations where the latest written data is 172 * preserved and the clear poison operation simply marks 173 * the address range as valid without changing the data. 174 * In this case application software can assume that an 175 * interrupted write will either return the new good 176 * data or an error. 177 * 178 * However, if pmem_clear_poison() leaves the data in an 179 * indeterminate state we need to perform the write 180 * after clear poison. 181 */ 182 flush_dcache_page(page); 183 write_pmem(pmem_addr, page, page_off, len); 184 if (unlikely(bad_pmem)) { 185 rc = pmem_clear_poison(pmem, pmem_off, len); 186 write_pmem(pmem_addr, page, page_off, len); 187 } 188 189 return rc; 190} 191 192static blk_qc_t pmem_submit_bio(struct bio *bio) 193{ 194 int ret = 0; 195 blk_status_t rc = 0; 196 bool do_acct; 197 unsigned long start; 198 struct bio_vec bvec; 199 struct bvec_iter iter; 200 struct pmem_device *pmem = bio->bi_disk->private_data; 201 struct nd_region *nd_region = to_region(pmem); 202 203 if (bio->bi_opf & REQ_PREFLUSH) 204 ret = nvdimm_flush(nd_region, bio); 205 206 do_acct = blk_queue_io_stat(bio->bi_disk->queue); 207 if (do_acct) 208 start = bio_start_io_acct(bio); 209 bio_for_each_segment(bvec, bio, iter) { 210 if (op_is_write(bio_op(bio))) 211 rc = pmem_do_write(pmem, bvec.bv_page, bvec.bv_offset, 212 iter.bi_sector, bvec.bv_len); 213 else 214 rc = pmem_do_read(pmem, bvec.bv_page, bvec.bv_offset, 215 iter.bi_sector, bvec.bv_len); 216 if (rc) { 217 bio->bi_status = rc; 218 break; 219 } 220 } 221 if (do_acct) 222 bio_end_io_acct(bio, start); 223 224 if (bio->bi_opf & REQ_FUA) 225 ret = nvdimm_flush(nd_region, bio); 226 227 if (ret) 228 bio->bi_status = errno_to_blk_status(ret); 229 230 bio_endio(bio); 231 return BLK_QC_T_NONE; 232} 233 234static int pmem_rw_page(struct block_device *bdev, sector_t sector, 235 struct page *page, unsigned int op) 236{ 237 struct pmem_device *pmem = bdev->bd_disk->private_data; 238 blk_status_t rc; 239 240 if (op_is_write(op)) 241 rc = pmem_do_write(pmem, page, 0, sector, thp_size(page)); 242 else 243 rc = pmem_do_read(pmem, page, 0, sector, thp_size(page)); 244 /* 245 * The ->rw_page interface is subtle and tricky. The core 246 * retries on any error, so we can only invoke page_endio() in 247 * the successful completion case. Otherwise, we'll see crashes 248 * caused by double completion. 249 */ 250 if (rc == 0) 251 page_endio(page, op_is_write(op), 0); 252 253 return blk_status_to_errno(rc); 254} 255 256/* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */ 257__weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff, 258 long nr_pages, void **kaddr, pfn_t *pfn) 259{ 260 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset; 261 262 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512, 263 PFN_PHYS(nr_pages)))) 264 return -EIO; 265 266 if (kaddr) 267 *kaddr = pmem->virt_addr + offset; 268 if (pfn) 269 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags); 270 271 /* 272 * If badblocks are present, limit known good range to the 273 * requested range. 274 */ 275 if (unlikely(pmem->bb.count)) 276 return nr_pages; 277 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset); 278} 279 280static const struct block_device_operations pmem_fops = { 281 .owner = THIS_MODULE, 282 .submit_bio = pmem_submit_bio, 283 .rw_page = pmem_rw_page, 284}; 285 286static int pmem_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff, 287 size_t nr_pages) 288{ 289 struct pmem_device *pmem = dax_get_private(dax_dev); 290 291 return blk_status_to_errno(pmem_do_write(pmem, ZERO_PAGE(0), 0, 292 PFN_PHYS(pgoff) >> SECTOR_SHIFT, 293 PAGE_SIZE)); 294} 295 296static long pmem_dax_direct_access(struct dax_device *dax_dev, 297 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn) 298{ 299 struct pmem_device *pmem = dax_get_private(dax_dev); 300 301 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn); 302} 303 304/* 305 * Use the 'no check' versions of copy_from_iter_flushcache() and 306 * copy_mc_to_iter() to bypass HARDENED_USERCOPY overhead. Bounds 307 * checking, both file offset and device offset, is handled by 308 * dax_iomap_actor() 309 */ 310static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, 311 void *addr, size_t bytes, struct iov_iter *i) 312{ 313 return _copy_from_iter_flushcache(addr, bytes, i); 314} 315 316static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, 317 void *addr, size_t bytes, struct iov_iter *i) 318{ 319 return _copy_mc_to_iter(addr, bytes, i); 320} 321 322static const struct dax_operations pmem_dax_ops = { 323 .direct_access = pmem_dax_direct_access, 324 .dax_supported = generic_fsdax_supported, 325 .copy_from_iter = pmem_copy_from_iter, 326 .copy_to_iter = pmem_copy_to_iter, 327 .zero_page_range = pmem_dax_zero_page_range, 328}; 329 330static const struct attribute_group *pmem_attribute_groups[] = { 331 &dax_attribute_group, 332 NULL, 333}; 334 335static void pmem_pagemap_cleanup(struct dev_pagemap *pgmap) 336{ 337 struct request_queue *q = 338 container_of(pgmap->ref, struct request_queue, q_usage_counter); 339 340 blk_cleanup_queue(q); 341} 342 343static void pmem_release_queue(void *pgmap) 344{ 345 pmem_pagemap_cleanup(pgmap); 346} 347 348static void pmem_pagemap_kill(struct dev_pagemap *pgmap) 349{ 350 struct request_queue *q = 351 container_of(pgmap->ref, struct request_queue, q_usage_counter); 352 353 blk_freeze_queue_start(q); 354} 355 356static void pmem_release_disk(void *__pmem) 357{ 358 struct pmem_device *pmem = __pmem; 359 360 kill_dax(pmem->dax_dev); 361 put_dax(pmem->dax_dev); 362 del_gendisk(pmem->disk); 363 put_disk(pmem->disk); 364} 365 366static const struct dev_pagemap_ops fsdax_pagemap_ops = { 367 .kill = pmem_pagemap_kill, 368 .cleanup = pmem_pagemap_cleanup, 369}; 370 371static int pmem_attach_disk(struct device *dev, 372 struct nd_namespace_common *ndns) 373{ 374 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev); 375 struct nd_region *nd_region = to_nd_region(dev->parent); 376 int nid = dev_to_node(dev), fua; 377 struct resource *res = &nsio->res; 378 struct range bb_range; 379 struct nd_pfn *nd_pfn = NULL; 380 struct dax_device *dax_dev; 381 struct nd_pfn_sb *pfn_sb; 382 struct pmem_device *pmem; 383 struct request_queue *q; 384 struct device *gendev; 385 struct gendisk *disk; 386 void *addr; 387 int rc; 388 unsigned long flags = 0UL; 389 390 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL); 391 if (!pmem) 392 return -ENOMEM; 393 394 rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve()); 395 if (rc) 396 return rc; 397 398 /* while nsio_rw_bytes is active, parse a pfn info block if present */ 399 if (is_nd_pfn(dev)) { 400 nd_pfn = to_nd_pfn(dev); 401 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap); 402 if (rc) 403 return rc; 404 } 405 406 /* we're attaching a block device, disable raw namespace access */ 407 devm_namespace_disable(dev, ndns); 408 409 dev_set_drvdata(dev, pmem); 410 pmem->phys_addr = res->start; 411 pmem->size = resource_size(res); 412 fua = nvdimm_has_flush(nd_region); 413 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) { 414 dev_warn(dev, "unable to guarantee persistence of writes\n"); 415 fua = 0; 416 } 417 418 if (!devm_request_mem_region(dev, res->start, resource_size(res), 419 dev_name(&ndns->dev))) { 420 dev_warn(dev, "could not reserve region %pR\n", res); 421 return -EBUSY; 422 } 423 424 q = blk_alloc_queue(dev_to_node(dev)); 425 if (!q) 426 return -ENOMEM; 427 428 pmem->pfn_flags = PFN_DEV; 429 pmem->pgmap.ref = &q->q_usage_counter; 430 if (is_nd_pfn(dev)) { 431 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX; 432 pmem->pgmap.ops = &fsdax_pagemap_ops; 433 addr = devm_memremap_pages(dev, &pmem->pgmap); 434 pfn_sb = nd_pfn->pfn_sb; 435 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff); 436 pmem->pfn_pad = resource_size(res) - 437 range_len(&pmem->pgmap.range); 438 pmem->pfn_flags |= PFN_MAP; 439 bb_range = pmem->pgmap.range; 440 bb_range.start += pmem->data_offset; 441 } else if (pmem_should_map_pages(dev)) { 442 pmem->pgmap.range.start = res->start; 443 pmem->pgmap.range.end = res->end; 444 pmem->pgmap.nr_range = 1; 445 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX; 446 pmem->pgmap.ops = &fsdax_pagemap_ops; 447 addr = devm_memremap_pages(dev, &pmem->pgmap); 448 pmem->pfn_flags |= PFN_MAP; 449 bb_range = pmem->pgmap.range; 450 } else { 451 addr = devm_memremap(dev, pmem->phys_addr, 452 pmem->size, ARCH_MEMREMAP_PMEM); 453 if (devm_add_action_or_reset(dev, pmem_release_queue, 454 &pmem->pgmap)) 455 return -ENOMEM; 456 bb_range.start = res->start; 457 bb_range.end = res->end; 458 } 459 460 if (IS_ERR(addr)) 461 return PTR_ERR(addr); 462 pmem->virt_addr = addr; 463 464 blk_queue_write_cache(q, true, fua); 465 blk_queue_physical_block_size(q, PAGE_SIZE); 466 blk_queue_logical_block_size(q, pmem_sector_size(ndns)); 467 blk_queue_max_hw_sectors(q, UINT_MAX); 468 blk_queue_flag_set(QUEUE_FLAG_NONROT, q); 469 if (pmem->pfn_flags & PFN_MAP) 470 blk_queue_flag_set(QUEUE_FLAG_DAX, q); 471 472 disk = alloc_disk_node(0, nid); 473 if (!disk) 474 return -ENOMEM; 475 pmem->disk = disk; 476 477 disk->fops = &pmem_fops; 478 disk->queue = q; 479 disk->flags = GENHD_FL_EXT_DEVT; 480 disk->private_data = pmem; 481 nvdimm_namespace_disk_name(ndns, disk->disk_name); 482 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset) 483 / 512); 484 if (devm_init_badblocks(dev, &pmem->bb)) 485 return -ENOMEM; 486 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_range); 487 disk->bb = &pmem->bb; 488 489 if (is_nvdimm_sync(nd_region)) 490 flags = DAXDEV_F_SYNC; 491 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops, flags); 492 if (IS_ERR(dax_dev)) { 493 put_disk(disk); 494 return PTR_ERR(dax_dev); 495 } 496 dax_write_cache(dax_dev, nvdimm_has_cache(nd_region)); 497 pmem->dax_dev = dax_dev; 498 gendev = disk_to_dev(disk); 499 gendev->groups = pmem_attribute_groups; 500 501 device_add_disk(dev, disk, NULL); 502 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem)) 503 return -ENOMEM; 504 505 nvdimm_check_and_set_ro(disk); 506 507 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd, 508 "badblocks"); 509 if (!pmem->bb_state) 510 dev_warn(dev, "'badblocks' notification disabled\n"); 511 512 return 0; 513} 514 515static int nd_pmem_probe(struct device *dev) 516{ 517 int ret; 518 struct nd_namespace_common *ndns; 519 520 ndns = nvdimm_namespace_common_probe(dev); 521 if (IS_ERR(ndns)) 522 return PTR_ERR(ndns); 523 524 if (is_nd_btt(dev)) 525 return nvdimm_namespace_attach_btt(ndns); 526 527 if (is_nd_pfn(dev)) 528 return pmem_attach_disk(dev, ndns); 529 530 ret = devm_namespace_enable(dev, ndns, nd_info_block_reserve()); 531 if (ret) 532 return ret; 533 534 ret = nd_btt_probe(dev, ndns); 535 if (ret == 0) 536 return -ENXIO; 537 538 /* 539 * We have two failure conditions here, there is no 540 * info reserver block or we found a valid info reserve block 541 * but failed to initialize the pfn superblock. 542 * 543 * For the first case consider namespace as a raw pmem namespace 544 * and attach a disk. 545 * 546 * For the latter, consider this a success and advance the namespace 547 * seed. 548 */ 549 ret = nd_pfn_probe(dev, ndns); 550 if (ret == 0) 551 return -ENXIO; 552 else if (ret == -EOPNOTSUPP) 553 return ret; 554 555 ret = nd_dax_probe(dev, ndns); 556 if (ret == 0) 557 return -ENXIO; 558 else if (ret == -EOPNOTSUPP) 559 return ret; 560 561 /* probe complete, attach handles namespace enabling */ 562 devm_namespace_disable(dev, ndns); 563 564 return pmem_attach_disk(dev, ndns); 565} 566 567static int nd_pmem_remove(struct device *dev) 568{ 569 struct pmem_device *pmem = dev_get_drvdata(dev); 570 571 if (is_nd_btt(dev)) 572 nvdimm_namespace_detach_btt(to_nd_btt(dev)); 573 else { 574 /* 575 * Note, this assumes nd_device_lock() context to not 576 * race nd_pmem_notify() 577 */ 578 sysfs_put(pmem->bb_state); 579 pmem->bb_state = NULL; 580 } 581 nvdimm_flush(to_nd_region(dev->parent), NULL); 582 583 return 0; 584} 585 586static void nd_pmem_shutdown(struct device *dev) 587{ 588 nvdimm_flush(to_nd_region(dev->parent), NULL); 589} 590 591static void nd_pmem_notify(struct device *dev, enum nvdimm_event event) 592{ 593 struct nd_region *nd_region; 594 resource_size_t offset = 0, end_trunc = 0; 595 struct nd_namespace_common *ndns; 596 struct nd_namespace_io *nsio; 597 struct badblocks *bb; 598 struct range range; 599 struct kernfs_node *bb_state; 600 601 if (event != NVDIMM_REVALIDATE_POISON) 602 return; 603 604 if (is_nd_btt(dev)) { 605 struct nd_btt *nd_btt = to_nd_btt(dev); 606 607 ndns = nd_btt->ndns; 608 nd_region = to_nd_region(ndns->dev.parent); 609 nsio = to_nd_namespace_io(&ndns->dev); 610 bb = &nsio->bb; 611 bb_state = NULL; 612 } else { 613 struct pmem_device *pmem = dev_get_drvdata(dev); 614 615 nd_region = to_region(pmem); 616 bb = &pmem->bb; 617 bb_state = pmem->bb_state; 618 619 if (is_nd_pfn(dev)) { 620 struct nd_pfn *nd_pfn = to_nd_pfn(dev); 621 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; 622 623 ndns = nd_pfn->ndns; 624 offset = pmem->data_offset + 625 __le32_to_cpu(pfn_sb->start_pad); 626 end_trunc = __le32_to_cpu(pfn_sb->end_trunc); 627 } else { 628 ndns = to_ndns(dev); 629 } 630 631 nsio = to_nd_namespace_io(&ndns->dev); 632 } 633 634 range.start = nsio->res.start + offset; 635 range.end = nsio->res.end - end_trunc; 636 nvdimm_badblocks_populate(nd_region, bb, &range); 637 if (bb_state) 638 sysfs_notify_dirent(bb_state); 639} 640 641MODULE_ALIAS("pmem"); 642MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO); 643MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM); 644static struct nd_device_driver nd_pmem_driver = { 645 .probe = nd_pmem_probe, 646 .remove = nd_pmem_remove, 647 .notify = nd_pmem_notify, 648 .shutdown = nd_pmem_shutdown, 649 .drv = { 650 .name = "nd_pmem", 651 }, 652 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM, 653}; 654 655module_nd_driver(nd_pmem_driver); 656 657MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>"); 658MODULE_LICENSE("GPL v2"); 659