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