1 // SPDX-License-Identifier: GPL-2.0-only
3 * Persistent Memory Driver
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>.
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>
27 #include <asm/cacheflush.h>
32 static struct device *to_dev(struct pmem_device *pmem)
35 * nvdimm bus services need a 'dev' parameter, and we record the device
41 static struct nd_region *to_region(struct pmem_device *pmem)
43 return to_nd_region(to_dev(pmem)->parent);
46 static void hwpoison_clear(struct pmem_device *pmem,
47 phys_addr_t phys, unsigned int len)
49 unsigned long pfn_start, pfn_end, pfn;
51 /* only pmem in the linear map supports HWPoison */
52 if (is_vmalloc_addr(pmem->virt_addr))
55 pfn_start = PHYS_PFN(phys);
56 pfn_end = pfn_start + PHYS_PFN(len);
57 for (pfn = pfn_start; pfn < pfn_end; pfn++) {
58 struct page *page = pfn_to_page(pfn);
61 * Note, no need to hold a get_dev_pagemap() reference
62 * here since we're in the driver I/O path and
63 * outstanding I/O requests pin the dev_pagemap.
65 if (test_and_clear_pmem_poison(page))
66 clear_mce_nospec(pfn);
70 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
71 phys_addr_t offset, unsigned int len)
73 struct device *dev = to_dev(pmem);
76 blk_status_t rc = BLK_STS_OK;
78 sector = (offset - pmem->data_offset) / 512;
80 cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
83 if (cleared > 0 && cleared / 512) {
84 hwpoison_clear(pmem, pmem->phys_addr + offset, cleared);
86 dev_dbg(dev, "%#llx clear %ld sector%s\n",
87 (unsigned long long) sector, cleared,
88 cleared > 1 ? "s" : "");
89 badblocks_clear(&pmem->bb, sector, cleared);
91 sysfs_notify_dirent(pmem->bb_state);
94 arch_invalidate_pmem(pmem->virt_addr + offset, len);
99 static void write_pmem(void *pmem_addr, struct page *page,
100 unsigned int off, unsigned int len)
106 mem = kmap_atomic(page);
107 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
108 memcpy_flushcache(pmem_addr, mem + off, chunk);
117 static blk_status_t read_pmem(struct page *page, unsigned int off,
118 void *pmem_addr, unsigned int len)
125 mem = kmap_atomic(page);
126 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
127 rem = copy_mc_to_kernel(mem + off, pmem_addr, chunk);
130 return BLK_STS_IOERR;
139 static blk_status_t pmem_do_read(struct pmem_device *pmem,
140 struct page *page, unsigned int page_off,
141 sector_t sector, unsigned int len)
144 phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
145 void *pmem_addr = pmem->virt_addr + pmem_off;
147 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
148 return BLK_STS_IOERR;
150 rc = read_pmem(page, page_off, pmem_addr, len);
151 flush_dcache_page(page);
155 static blk_status_t pmem_do_write(struct pmem_device *pmem,
156 struct page *page, unsigned int page_off,
157 sector_t sector, unsigned int len)
159 blk_status_t rc = BLK_STS_OK;
160 bool bad_pmem = false;
161 phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
162 void *pmem_addr = pmem->virt_addr + pmem_off;
164 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
168 * Note that we write the data both before and after
169 * clearing poison. The write before clear poison
170 * handles situations where the latest written data is
171 * preserved and the clear poison operation simply marks
172 * the address range as valid without changing the data.
173 * In this case application software can assume that an
174 * interrupted write will either return the new good
177 * However, if pmem_clear_poison() leaves the data in an
178 * indeterminate state we need to perform the write
179 * after clear poison.
181 flush_dcache_page(page);
182 write_pmem(pmem_addr, page, page_off, len);
183 if (unlikely(bad_pmem)) {
184 rc = pmem_clear_poison(pmem, pmem_off, len);
185 write_pmem(pmem_addr, page, page_off, len);
191 static blk_qc_t pmem_submit_bio(struct bio *bio)
198 struct bvec_iter iter;
199 struct pmem_device *pmem = bio->bi_bdev->bd_disk->private_data;
200 struct nd_region *nd_region = to_region(pmem);
202 if (bio->bi_opf & REQ_PREFLUSH)
203 ret = nvdimm_flush(nd_region, bio);
205 do_acct = blk_queue_io_stat(bio->bi_bdev->bd_disk->queue);
207 start = bio_start_io_acct(bio);
208 bio_for_each_segment(bvec, bio, iter) {
209 if (op_is_write(bio_op(bio)))
210 rc = pmem_do_write(pmem, bvec.bv_page, bvec.bv_offset,
211 iter.bi_sector, bvec.bv_len);
213 rc = pmem_do_read(pmem, bvec.bv_page, bvec.bv_offset,
214 iter.bi_sector, bvec.bv_len);
221 bio_end_io_acct(bio, start);
223 if (bio->bi_opf & REQ_FUA)
224 ret = nvdimm_flush(nd_region, bio);
227 bio->bi_status = errno_to_blk_status(ret);
230 return BLK_QC_T_NONE;
233 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
234 struct page *page, unsigned int op)
236 struct pmem_device *pmem = bdev->bd_disk->private_data;
240 rc = pmem_do_write(pmem, page, 0, sector, thp_size(page));
242 rc = pmem_do_read(pmem, page, 0, sector, thp_size(page));
244 * The ->rw_page interface is subtle and tricky. The core
245 * retries on any error, so we can only invoke page_endio() in
246 * the successful completion case. Otherwise, we'll see crashes
247 * caused by double completion.
250 page_endio(page, op_is_write(op), 0);
252 return blk_status_to_errno(rc);
255 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
256 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
257 long nr_pages, void **kaddr, pfn_t *pfn)
259 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
261 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
262 PFN_PHYS(nr_pages))))
266 *kaddr = pmem->virt_addr + offset;
268 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
271 * If badblocks are present, limit known good range to the
274 if (unlikely(pmem->bb.count))
276 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
279 static const struct block_device_operations pmem_fops = {
280 .owner = THIS_MODULE,
281 .submit_bio = pmem_submit_bio,
282 .rw_page = pmem_rw_page,
285 static int pmem_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
288 struct pmem_device *pmem = dax_get_private(dax_dev);
290 return blk_status_to_errno(pmem_do_write(pmem, ZERO_PAGE(0), 0,
291 PFN_PHYS(pgoff) >> SECTOR_SHIFT,
295 static long pmem_dax_direct_access(struct dax_device *dax_dev,
296 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
298 struct pmem_device *pmem = dax_get_private(dax_dev);
300 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
304 * Use the 'no check' versions of copy_from_iter_flushcache() and
305 * copy_mc_to_iter() to bypass HARDENED_USERCOPY overhead. Bounds
306 * checking, both file offset and device offset, is handled by
309 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
310 void *addr, size_t bytes, struct iov_iter *i)
312 return _copy_from_iter_flushcache(addr, bytes, i);
315 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
316 void *addr, size_t bytes, struct iov_iter *i)
318 return _copy_mc_to_iter(addr, bytes, i);
321 static const struct dax_operations pmem_dax_ops = {
322 .direct_access = pmem_dax_direct_access,
323 .dax_supported = generic_fsdax_supported,
324 .copy_from_iter = pmem_copy_from_iter,
325 .copy_to_iter = pmem_copy_to_iter,
326 .zero_page_range = pmem_dax_zero_page_range,
329 static const struct attribute_group *pmem_attribute_groups[] = {
330 &dax_attribute_group,
334 static void pmem_pagemap_cleanup(struct dev_pagemap *pgmap)
336 struct request_queue *q =
337 container_of(pgmap->ref, struct request_queue, q_usage_counter);
339 blk_cleanup_queue(q);
342 static void pmem_release_queue(void *pgmap)
344 pmem_pagemap_cleanup(pgmap);
347 static void pmem_pagemap_kill(struct dev_pagemap *pgmap)
349 struct request_queue *q =
350 container_of(pgmap->ref, struct request_queue, q_usage_counter);
352 blk_freeze_queue_start(q);
355 static void pmem_release_disk(void *__pmem)
357 struct pmem_device *pmem = __pmem;
359 kill_dax(pmem->dax_dev);
360 put_dax(pmem->dax_dev);
361 del_gendisk(pmem->disk);
362 put_disk(pmem->disk);
365 static const struct dev_pagemap_ops fsdax_pagemap_ops = {
366 .kill = pmem_pagemap_kill,
367 .cleanup = pmem_pagemap_cleanup,
370 static int pmem_attach_disk(struct device *dev,
371 struct nd_namespace_common *ndns)
373 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
374 struct nd_region *nd_region = to_nd_region(dev->parent);
375 int nid = dev_to_node(dev), fua;
376 struct resource *res = &nsio->res;
377 struct range bb_range;
378 struct nd_pfn *nd_pfn = NULL;
379 struct dax_device *dax_dev;
380 struct nd_pfn_sb *pfn_sb;
381 struct pmem_device *pmem;
382 struct request_queue *q;
383 struct device *gendev;
384 struct gendisk *disk;
387 unsigned long flags = 0UL;
389 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
393 rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
397 /* while nsio_rw_bytes is active, parse a pfn info block if present */
398 if (is_nd_pfn(dev)) {
399 nd_pfn = to_nd_pfn(dev);
400 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
405 /* we're attaching a block device, disable raw namespace access */
406 devm_namespace_disable(dev, ndns);
408 dev_set_drvdata(dev, pmem);
409 pmem->phys_addr = res->start;
410 pmem->size = resource_size(res);
411 fua = nvdimm_has_flush(nd_region);
412 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
413 dev_warn(dev, "unable to guarantee persistence of writes\n");
417 if (!devm_request_mem_region(dev, res->start, resource_size(res),
418 dev_name(&ndns->dev))) {
419 dev_warn(dev, "could not reserve region %pR\n", res);
423 q = blk_alloc_queue(dev_to_node(dev));
427 pmem->pfn_flags = PFN_DEV;
428 pmem->pgmap.ref = &q->q_usage_counter;
429 if (is_nd_pfn(dev)) {
430 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
431 pmem->pgmap.ops = &fsdax_pagemap_ops;
432 addr = devm_memremap_pages(dev, &pmem->pgmap);
433 pfn_sb = nd_pfn->pfn_sb;
434 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
435 pmem->pfn_pad = resource_size(res) -
436 range_len(&pmem->pgmap.range);
437 pmem->pfn_flags |= PFN_MAP;
438 bb_range = pmem->pgmap.range;
439 bb_range.start += pmem->data_offset;
440 } else if (pmem_should_map_pages(dev)) {
441 pmem->pgmap.range.start = res->start;
442 pmem->pgmap.range.end = res->end;
443 pmem->pgmap.nr_range = 1;
444 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
445 pmem->pgmap.ops = &fsdax_pagemap_ops;
446 addr = devm_memremap_pages(dev, &pmem->pgmap);
447 pmem->pfn_flags |= PFN_MAP;
448 bb_range = pmem->pgmap.range;
450 if (devm_add_action_or_reset(dev, pmem_release_queue,
453 addr = devm_memremap(dev, pmem->phys_addr,
454 pmem->size, ARCH_MEMREMAP_PMEM);
455 bb_range.start = res->start;
456 bb_range.end = res->end;
460 return PTR_ERR(addr);
461 pmem->virt_addr = addr;
463 blk_queue_write_cache(q, true, fua);
464 blk_queue_physical_block_size(q, PAGE_SIZE);
465 blk_queue_logical_block_size(q, pmem_sector_size(ndns));
466 blk_queue_max_hw_sectors(q, UINT_MAX);
467 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
468 if (pmem->pfn_flags & PFN_MAP)
469 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
471 disk = alloc_disk_node(0, nid);
476 disk->fops = &pmem_fops;
478 disk->flags = GENHD_FL_EXT_DEVT;
479 disk->private_data = pmem;
480 nvdimm_namespace_disk_name(ndns, disk->disk_name);
481 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
483 if (devm_init_badblocks(dev, &pmem->bb))
485 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_range);
486 disk->bb = &pmem->bb;
488 if (is_nvdimm_sync(nd_region))
489 flags = DAXDEV_F_SYNC;
490 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops, flags);
491 if (IS_ERR(dax_dev)) {
493 return PTR_ERR(dax_dev);
495 dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
496 pmem->dax_dev = dax_dev;
497 gendev = disk_to_dev(disk);
498 gendev->groups = pmem_attribute_groups;
500 device_add_disk(dev, disk, NULL);
501 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
504 nvdimm_check_and_set_ro(disk);
506 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
509 dev_warn(dev, "'badblocks' notification disabled\n");
514 static int nd_pmem_probe(struct device *dev)
517 struct nd_namespace_common *ndns;
519 ndns = nvdimm_namespace_common_probe(dev);
521 return PTR_ERR(ndns);
524 return nvdimm_namespace_attach_btt(ndns);
527 return pmem_attach_disk(dev, ndns);
529 ret = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
533 ret = nd_btt_probe(dev, ndns);
538 * We have two failure conditions here, there is no
539 * info reserver block or we found a valid info reserve block
540 * but failed to initialize the pfn superblock.
542 * For the first case consider namespace as a raw pmem namespace
545 * For the latter, consider this a success and advance the namespace
548 ret = nd_pfn_probe(dev, ndns);
551 else if (ret == -EOPNOTSUPP)
554 ret = nd_dax_probe(dev, ndns);
557 else if (ret == -EOPNOTSUPP)
560 /* probe complete, attach handles namespace enabling */
561 devm_namespace_disable(dev, ndns);
563 return pmem_attach_disk(dev, ndns);
566 static void nd_pmem_remove(struct device *dev)
568 struct pmem_device *pmem = dev_get_drvdata(dev);
571 nvdimm_namespace_detach_btt(to_nd_btt(dev));
574 * Note, this assumes nd_device_lock() context to not
575 * race nd_pmem_notify()
577 sysfs_put(pmem->bb_state);
578 pmem->bb_state = NULL;
580 nvdimm_flush(to_nd_region(dev->parent), NULL);
583 static void nd_pmem_shutdown(struct device *dev)
585 nvdimm_flush(to_nd_region(dev->parent), NULL);
588 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
590 struct nd_region *nd_region;
591 resource_size_t offset = 0, end_trunc = 0;
592 struct nd_namespace_common *ndns;
593 struct nd_namespace_io *nsio;
594 struct badblocks *bb;
596 struct kernfs_node *bb_state;
598 if (event != NVDIMM_REVALIDATE_POISON)
601 if (is_nd_btt(dev)) {
602 struct nd_btt *nd_btt = to_nd_btt(dev);
605 nd_region = to_nd_region(ndns->dev.parent);
606 nsio = to_nd_namespace_io(&ndns->dev);
610 struct pmem_device *pmem = dev_get_drvdata(dev);
612 nd_region = to_region(pmem);
614 bb_state = pmem->bb_state;
616 if (is_nd_pfn(dev)) {
617 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
618 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
621 offset = pmem->data_offset +
622 __le32_to_cpu(pfn_sb->start_pad);
623 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
628 nsio = to_nd_namespace_io(&ndns->dev);
631 range.start = nsio->res.start + offset;
632 range.end = nsio->res.end - end_trunc;
633 nvdimm_badblocks_populate(nd_region, bb, &range);
635 sysfs_notify_dirent(bb_state);
638 MODULE_ALIAS("pmem");
639 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
640 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
641 static struct nd_device_driver nd_pmem_driver = {
642 .probe = nd_pmem_probe,
643 .remove = nd_pmem_remove,
644 .notify = nd_pmem_notify,
645 .shutdown = nd_pmem_shutdown,
649 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
652 module_nd_driver(nd_pmem_driver);
654 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
655 MODULE_LICENSE("GPL v2");