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 33 static 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 42 static struct nd_region *to_region(struct pmem_device *pmem) 43 { 44 return to_nd_region(to_dev(pmem)->parent); 45 } 46 47 static 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 71 static 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 } 99 100 static 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 } 117 118 static 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 = memcpy_mcsafe(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 } 139 140 static 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 } 155 156 static 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 } 191 192 static 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 } 233 234 static 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, 242 hpage_nr_pages(page) * PAGE_SIZE); 243 else 244 rc = pmem_do_read(pmem, page, 0, sector, 245 hpage_nr_pages(page) * PAGE_SIZE); 246 /* 247 * The ->rw_page interface is subtle and tricky. The core 248 * retries on any error, so we can only invoke page_endio() in 249 * the successful completion case. Otherwise, we'll see crashes 250 * caused by double completion. 251 */ 252 if (rc == 0) 253 page_endio(page, op_is_write(op), 0); 254 255 return blk_status_to_errno(rc); 256 } 257 258 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */ 259 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff, 260 long nr_pages, void **kaddr, pfn_t *pfn) 261 { 262 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset; 263 264 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512, 265 PFN_PHYS(nr_pages)))) 266 return -EIO; 267 268 if (kaddr) 269 *kaddr = pmem->virt_addr + offset; 270 if (pfn) 271 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags); 272 273 /* 274 * If badblocks are present, limit known good range to the 275 * requested range. 276 */ 277 if (unlikely(pmem->bb.count)) 278 return nr_pages; 279 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset); 280 } 281 282 static const struct block_device_operations pmem_fops = { 283 .owner = THIS_MODULE, 284 .submit_bio = pmem_submit_bio, 285 .rw_page = pmem_rw_page, 286 .revalidate_disk = nvdimm_revalidate_disk, 287 }; 288 289 static int pmem_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff, 290 size_t nr_pages) 291 { 292 struct pmem_device *pmem = dax_get_private(dax_dev); 293 294 return blk_status_to_errno(pmem_do_write(pmem, ZERO_PAGE(0), 0, 295 PFN_PHYS(pgoff) >> SECTOR_SHIFT, 296 PAGE_SIZE)); 297 } 298 299 static long pmem_dax_direct_access(struct dax_device *dax_dev, 300 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn) 301 { 302 struct pmem_device *pmem = dax_get_private(dax_dev); 303 304 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn); 305 } 306 307 /* 308 * Use the 'no check' versions of copy_from_iter_flushcache() and 309 * copy_to_iter_mcsafe() to bypass HARDENED_USERCOPY overhead. Bounds 310 * checking, both file offset and device offset, is handled by 311 * dax_iomap_actor() 312 */ 313 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, 314 void *addr, size_t bytes, struct iov_iter *i) 315 { 316 return _copy_from_iter_flushcache(addr, bytes, i); 317 } 318 319 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, 320 void *addr, size_t bytes, struct iov_iter *i) 321 { 322 return _copy_to_iter_mcsafe(addr, bytes, i); 323 } 324 325 static const struct dax_operations pmem_dax_ops = { 326 .direct_access = pmem_dax_direct_access, 327 .dax_supported = generic_fsdax_supported, 328 .copy_from_iter = pmem_copy_from_iter, 329 .copy_to_iter = pmem_copy_to_iter, 330 .zero_page_range = pmem_dax_zero_page_range, 331 }; 332 333 static const struct attribute_group *pmem_attribute_groups[] = { 334 &dax_attribute_group, 335 NULL, 336 }; 337 338 static void pmem_pagemap_cleanup(struct dev_pagemap *pgmap) 339 { 340 struct request_queue *q = 341 container_of(pgmap->ref, struct request_queue, q_usage_counter); 342 343 blk_cleanup_queue(q); 344 } 345 346 static void pmem_release_queue(void *pgmap) 347 { 348 pmem_pagemap_cleanup(pgmap); 349 } 350 351 static void pmem_pagemap_kill(struct dev_pagemap *pgmap) 352 { 353 struct request_queue *q = 354 container_of(pgmap->ref, struct request_queue, q_usage_counter); 355 356 blk_freeze_queue_start(q); 357 } 358 359 static void pmem_release_disk(void *__pmem) 360 { 361 struct pmem_device *pmem = __pmem; 362 363 kill_dax(pmem->dax_dev); 364 put_dax(pmem->dax_dev); 365 del_gendisk(pmem->disk); 366 put_disk(pmem->disk); 367 } 368 369 static const struct dev_pagemap_ops fsdax_pagemap_ops = { 370 .kill = pmem_pagemap_kill, 371 .cleanup = pmem_pagemap_cleanup, 372 }; 373 374 static int pmem_attach_disk(struct device *dev, 375 struct nd_namespace_common *ndns) 376 { 377 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev); 378 struct nd_region *nd_region = to_nd_region(dev->parent); 379 int nid = dev_to_node(dev), fua; 380 struct resource *res = &nsio->res; 381 struct resource bb_res; 382 struct nd_pfn *nd_pfn = NULL; 383 struct dax_device *dax_dev; 384 struct nd_pfn_sb *pfn_sb; 385 struct pmem_device *pmem; 386 struct request_queue *q; 387 struct device *gendev; 388 struct gendisk *disk; 389 void *addr; 390 int rc; 391 unsigned long flags = 0UL; 392 393 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL); 394 if (!pmem) 395 return -ENOMEM; 396 397 rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve()); 398 if (rc) 399 return rc; 400 401 /* while nsio_rw_bytes is active, parse a pfn info block if present */ 402 if (is_nd_pfn(dev)) { 403 nd_pfn = to_nd_pfn(dev); 404 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap); 405 if (rc) 406 return rc; 407 } 408 409 /* we're attaching a block device, disable raw namespace access */ 410 devm_namespace_disable(dev, ndns); 411 412 dev_set_drvdata(dev, pmem); 413 pmem->phys_addr = res->start; 414 pmem->size = resource_size(res); 415 fua = nvdimm_has_flush(nd_region); 416 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) { 417 dev_warn(dev, "unable to guarantee persistence of writes\n"); 418 fua = 0; 419 } 420 421 if (!devm_request_mem_region(dev, res->start, resource_size(res), 422 dev_name(&ndns->dev))) { 423 dev_warn(dev, "could not reserve region %pR\n", res); 424 return -EBUSY; 425 } 426 427 q = blk_alloc_queue(dev_to_node(dev)); 428 if (!q) 429 return -ENOMEM; 430 431 pmem->pfn_flags = PFN_DEV; 432 pmem->pgmap.ref = &q->q_usage_counter; 433 if (is_nd_pfn(dev)) { 434 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX; 435 pmem->pgmap.ops = &fsdax_pagemap_ops; 436 addr = devm_memremap_pages(dev, &pmem->pgmap); 437 pfn_sb = nd_pfn->pfn_sb; 438 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff); 439 pmem->pfn_pad = resource_size(res) - 440 resource_size(&pmem->pgmap.res); 441 pmem->pfn_flags |= PFN_MAP; 442 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res)); 443 bb_res.start += pmem->data_offset; 444 } else if (pmem_should_map_pages(dev)) { 445 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res)); 446 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX; 447 pmem->pgmap.ops = &fsdax_pagemap_ops; 448 addr = devm_memremap_pages(dev, &pmem->pgmap); 449 pmem->pfn_flags |= PFN_MAP; 450 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res)); 451 } else { 452 if (devm_add_action_or_reset(dev, pmem_release_queue, 453 &pmem->pgmap)) 454 return -ENOMEM; 455 addr = devm_memremap(dev, pmem->phys_addr, 456 pmem->size, ARCH_MEMREMAP_PMEM); 457 memcpy(&bb_res, &nsio->res, sizeof(bb_res)); 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 disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO; 482 nvdimm_namespace_disk_name(ndns, disk->disk_name); 483 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset) 484 / 512); 485 if (devm_init_badblocks(dev, &pmem->bb)) 486 return -ENOMEM; 487 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res); 488 disk->bb = &pmem->bb; 489 490 if (is_nvdimm_sync(nd_region)) 491 flags = DAXDEV_F_SYNC; 492 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops, flags); 493 if (IS_ERR(dax_dev)) { 494 put_disk(disk); 495 return PTR_ERR(dax_dev); 496 } 497 dax_write_cache(dax_dev, nvdimm_has_cache(nd_region)); 498 pmem->dax_dev = dax_dev; 499 gendev = disk_to_dev(disk); 500 gendev->groups = pmem_attribute_groups; 501 502 device_add_disk(dev, disk, NULL); 503 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem)) 504 return -ENOMEM; 505 506 revalidate_disk(disk); 507 508 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd, 509 "badblocks"); 510 if (!pmem->bb_state) 511 dev_warn(dev, "'badblocks' notification disabled\n"); 512 513 return 0; 514 } 515 516 static int nd_pmem_probe(struct device *dev) 517 { 518 int ret; 519 struct nd_namespace_common *ndns; 520 521 ndns = nvdimm_namespace_common_probe(dev); 522 if (IS_ERR(ndns)) 523 return PTR_ERR(ndns); 524 525 if (is_nd_btt(dev)) 526 return nvdimm_namespace_attach_btt(ndns); 527 528 if (is_nd_pfn(dev)) 529 return pmem_attach_disk(dev, ndns); 530 531 ret = devm_namespace_enable(dev, ndns, nd_info_block_reserve()); 532 if (ret) 533 return ret; 534 535 ret = nd_btt_probe(dev, ndns); 536 if (ret == 0) 537 return -ENXIO; 538 539 /* 540 * We have two failure conditions here, there is no 541 * info reserver block or we found a valid info reserve block 542 * but failed to initialize the pfn superblock. 543 * 544 * For the first case consider namespace as a raw pmem namespace 545 * and attach a disk. 546 * 547 * For the latter, consider this a success and advance the namespace 548 * seed. 549 */ 550 ret = nd_pfn_probe(dev, ndns); 551 if (ret == 0) 552 return -ENXIO; 553 else if (ret == -EOPNOTSUPP) 554 return ret; 555 556 ret = nd_dax_probe(dev, ndns); 557 if (ret == 0) 558 return -ENXIO; 559 else if (ret == -EOPNOTSUPP) 560 return ret; 561 562 /* probe complete, attach handles namespace enabling */ 563 devm_namespace_disable(dev, ndns); 564 565 return pmem_attach_disk(dev, ndns); 566 } 567 568 static int nd_pmem_remove(struct device *dev) 569 { 570 struct pmem_device *pmem = dev_get_drvdata(dev); 571 572 if (is_nd_btt(dev)) 573 nvdimm_namespace_detach_btt(to_nd_btt(dev)); 574 else { 575 /* 576 * Note, this assumes nd_device_lock() context to not 577 * race nd_pmem_notify() 578 */ 579 sysfs_put(pmem->bb_state); 580 pmem->bb_state = NULL; 581 } 582 nvdimm_flush(to_nd_region(dev->parent), NULL); 583 584 return 0; 585 } 586 587 static void nd_pmem_shutdown(struct device *dev) 588 { 589 nvdimm_flush(to_nd_region(dev->parent), NULL); 590 } 591 592 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event) 593 { 594 struct nd_region *nd_region; 595 resource_size_t offset = 0, end_trunc = 0; 596 struct nd_namespace_common *ndns; 597 struct nd_namespace_io *nsio; 598 struct resource res; 599 struct badblocks *bb; 600 struct kernfs_node *bb_state; 601 602 if (event != NVDIMM_REVALIDATE_POISON) 603 return; 604 605 if (is_nd_btt(dev)) { 606 struct nd_btt *nd_btt = to_nd_btt(dev); 607 608 ndns = nd_btt->ndns; 609 nd_region = to_nd_region(ndns->dev.parent); 610 nsio = to_nd_namespace_io(&ndns->dev); 611 bb = &nsio->bb; 612 bb_state = NULL; 613 } else { 614 struct pmem_device *pmem = dev_get_drvdata(dev); 615 616 nd_region = to_region(pmem); 617 bb = &pmem->bb; 618 bb_state = pmem->bb_state; 619 620 if (is_nd_pfn(dev)) { 621 struct nd_pfn *nd_pfn = to_nd_pfn(dev); 622 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; 623 624 ndns = nd_pfn->ndns; 625 offset = pmem->data_offset + 626 __le32_to_cpu(pfn_sb->start_pad); 627 end_trunc = __le32_to_cpu(pfn_sb->end_trunc); 628 } else { 629 ndns = to_ndns(dev); 630 } 631 632 nsio = to_nd_namespace_io(&ndns->dev); 633 } 634 635 res.start = nsio->res.start + offset; 636 res.end = nsio->res.end - end_trunc; 637 nvdimm_badblocks_populate(nd_region, bb, &res); 638 if (bb_state) 639 sysfs_notify_dirent(bb_state); 640 } 641 642 MODULE_ALIAS("pmem"); 643 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO); 644 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM); 645 static struct nd_device_driver nd_pmem_driver = { 646 .probe = nd_pmem_probe, 647 .remove = nd_pmem_remove, 648 .notify = nd_pmem_notify, 649 .shutdown = nd_pmem_shutdown, 650 .drv = { 651 .name = "nd_pmem", 652 }, 653 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM, 654 }; 655 656 module_nd_driver(nd_pmem_driver); 657 658 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>"); 659 MODULE_LICENSE("GPL v2"); 660