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>
26 #include <linux/backing-dev.h>
28 #include <asm/cacheflush.h>
33 static struct device *to_dev(struct pmem_device *pmem)
36 * nvdimm bus services need a 'dev' parameter, and we record the device
42 static struct nd_region *to_region(struct pmem_device *pmem)
44 return to_nd_region(to_dev(pmem)->parent);
47 static void hwpoison_clear(struct pmem_device *pmem,
48 phys_addr_t phys, unsigned int len)
50 unsigned long pfn_start, pfn_end, pfn;
52 /* only pmem in the linear map supports HWPoison */
53 if (is_vmalloc_addr(pmem->virt_addr))
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);
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.
66 if (test_and_clear_pmem_poison(page))
67 clear_mce_nospec(pfn);
71 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
72 phys_addr_t offset, unsigned int len)
74 struct device *dev = to_dev(pmem);
77 blk_status_t rc = BLK_STS_OK;
79 sector = (offset - pmem->data_offset) / 512;
81 cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
84 if (cleared > 0 && cleared / 512) {
85 hwpoison_clear(pmem, pmem->phys_addr + offset, cleared);
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);
92 sysfs_notify_dirent(pmem->bb_state);
95 arch_invalidate_pmem(pmem->virt_addr + offset, len);
100 static void write_pmem(void *pmem_addr, struct page *page,
101 unsigned int off, unsigned int len)
107 mem = kmap_atomic(page);
108 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
109 memcpy_flushcache(pmem_addr, mem + off, chunk);
118 static blk_status_t read_pmem(struct page *page, unsigned int off,
119 void *pmem_addr, unsigned int 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);
131 return BLK_STS_IOERR;
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)
145 phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
146 void *pmem_addr = pmem->virt_addr + pmem_off;
148 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
149 return BLK_STS_IOERR;
151 rc = read_pmem(page, page_off, pmem_addr, len);
152 flush_dcache_page(page);
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)
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;
165 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
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
178 * However, if pmem_clear_poison() leaves the data in an
179 * indeterminate state we need to perform the write
180 * after clear poison.
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);
192 static blk_qc_t pmem_submit_bio(struct bio *bio)
199 struct bvec_iter iter;
200 struct pmem_device *pmem = bio->bi_disk->private_data;
201 struct nd_region *nd_region = to_region(pmem);
203 if (bio->bi_opf & REQ_PREFLUSH)
204 ret = nvdimm_flush(nd_region, bio);
206 do_acct = blk_queue_io_stat(bio->bi_disk->queue);
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);
214 rc = pmem_do_read(pmem, bvec.bv_page, bvec.bv_offset,
215 iter.bi_sector, bvec.bv_len);
222 bio_end_io_acct(bio, start);
224 if (bio->bi_opf & REQ_FUA)
225 ret = nvdimm_flush(nd_region, bio);
228 bio->bi_status = errno_to_blk_status(ret);
231 return BLK_QC_T_NONE;
234 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
235 struct page *page, unsigned int op)
237 struct pmem_device *pmem = bdev->bd_disk->private_data;
241 rc = pmem_do_write(pmem, page, 0, sector, thp_size(page));
243 rc = pmem_do_read(pmem, page, 0, sector, thp_size(page));
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.
251 page_endio(page, op_is_write(op), 0);
253 return blk_status_to_errno(rc);
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)
260 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
262 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
263 PFN_PHYS(nr_pages))))
267 *kaddr = pmem->virt_addr + offset;
269 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
272 * If badblocks are present, limit known good range to the
275 if (unlikely(pmem->bb.count))
277 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
280 static const struct block_device_operations pmem_fops = {
281 .owner = THIS_MODULE,
282 .submit_bio = pmem_submit_bio,
283 .rw_page = pmem_rw_page,
284 .revalidate_disk = nvdimm_revalidate_disk,
287 static int pmem_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
290 struct pmem_device *pmem = dax_get_private(dax_dev);
292 return blk_status_to_errno(pmem_do_write(pmem, ZERO_PAGE(0), 0,
293 PFN_PHYS(pgoff) >> SECTOR_SHIFT,
297 static long pmem_dax_direct_access(struct dax_device *dax_dev,
298 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
300 struct pmem_device *pmem = dax_get_private(dax_dev);
302 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
306 * Use the 'no check' versions of copy_from_iter_flushcache() and
307 * copy_mc_to_iter() to bypass HARDENED_USERCOPY overhead. Bounds
308 * checking, both file offset and device offset, is handled by
311 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
312 void *addr, size_t bytes, struct iov_iter *i)
314 return _copy_from_iter_flushcache(addr, bytes, i);
317 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
318 void *addr, size_t bytes, struct iov_iter *i)
320 return _copy_mc_to_iter(addr, bytes, i);
323 static const struct dax_operations pmem_dax_ops = {
324 .direct_access = pmem_dax_direct_access,
325 .dax_supported = generic_fsdax_supported,
326 .copy_from_iter = pmem_copy_from_iter,
327 .copy_to_iter = pmem_copy_to_iter,
328 .zero_page_range = pmem_dax_zero_page_range,
331 static const struct attribute_group *pmem_attribute_groups[] = {
332 &dax_attribute_group,
336 static void pmem_pagemap_cleanup(struct dev_pagemap *pgmap)
338 struct request_queue *q =
339 container_of(pgmap->ref, struct request_queue, q_usage_counter);
341 blk_cleanup_queue(q);
344 static void pmem_release_queue(void *pgmap)
346 pmem_pagemap_cleanup(pgmap);
349 static void pmem_pagemap_kill(struct dev_pagemap *pgmap)
351 struct request_queue *q =
352 container_of(pgmap->ref, struct request_queue, q_usage_counter);
354 blk_freeze_queue_start(q);
357 static void pmem_release_disk(void *__pmem)
359 struct pmem_device *pmem = __pmem;
361 kill_dax(pmem->dax_dev);
362 put_dax(pmem->dax_dev);
363 del_gendisk(pmem->disk);
364 put_disk(pmem->disk);
367 static const struct dev_pagemap_ops fsdax_pagemap_ops = {
368 .kill = pmem_pagemap_kill,
369 .cleanup = pmem_pagemap_cleanup,
372 static int pmem_attach_disk(struct device *dev,
373 struct nd_namespace_common *ndns)
375 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
376 struct nd_region *nd_region = to_nd_region(dev->parent);
377 int nid = dev_to_node(dev), fua;
378 struct resource *res = &nsio->res;
379 struct resource bb_res;
380 struct nd_pfn *nd_pfn = NULL;
381 struct dax_device *dax_dev;
382 struct nd_pfn_sb *pfn_sb;
383 struct pmem_device *pmem;
384 struct request_queue *q;
385 struct device *gendev;
386 struct gendisk *disk;
389 unsigned long flags = 0UL;
391 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
395 rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
399 /* while nsio_rw_bytes is active, parse a pfn info block if present */
400 if (is_nd_pfn(dev)) {
401 nd_pfn = to_nd_pfn(dev);
402 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
407 /* we're attaching a block device, disable raw namespace access */
408 devm_namespace_disable(dev, ndns);
410 dev_set_drvdata(dev, pmem);
411 pmem->phys_addr = res->start;
412 pmem->size = resource_size(res);
413 fua = nvdimm_has_flush(nd_region);
414 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
415 dev_warn(dev, "unable to guarantee persistence of writes\n");
419 if (!devm_request_mem_region(dev, res->start, resource_size(res),
420 dev_name(&ndns->dev))) {
421 dev_warn(dev, "could not reserve region %pR\n", res);
425 q = blk_alloc_queue(dev_to_node(dev));
429 pmem->pfn_flags = PFN_DEV;
430 pmem->pgmap.ref = &q->q_usage_counter;
431 if (is_nd_pfn(dev)) {
432 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
433 pmem->pgmap.ops = &fsdax_pagemap_ops;
434 addr = devm_memremap_pages(dev, &pmem->pgmap);
435 pfn_sb = nd_pfn->pfn_sb;
436 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
437 pmem->pfn_pad = resource_size(res) -
438 resource_size(&pmem->pgmap.res);
439 pmem->pfn_flags |= PFN_MAP;
440 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
441 bb_res.start += pmem->data_offset;
442 } else if (pmem_should_map_pages(dev)) {
443 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
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 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
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 memcpy(&bb_res, &nsio->res, sizeof(bb_res));
459 return PTR_ERR(addr);
460 pmem->virt_addr = addr;
462 blk_queue_write_cache(q, true, fua);
463 blk_queue_physical_block_size(q, PAGE_SIZE);
464 blk_queue_logical_block_size(q, pmem_sector_size(ndns));
465 blk_queue_max_hw_sectors(q, UINT_MAX);
466 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
467 if (pmem->pfn_flags & PFN_MAP)
468 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
470 disk = alloc_disk_node(0, nid);
475 disk->fops = &pmem_fops;
477 disk->flags = GENHD_FL_EXT_DEVT;
478 disk->private_data = pmem;
479 disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
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_res);
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 revalidate_disk(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 int 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);
585 static void nd_pmem_shutdown(struct device *dev)
587 nvdimm_flush(to_nd_region(dev->parent), NULL);
590 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
592 struct nd_region *nd_region;
593 resource_size_t offset = 0, end_trunc = 0;
594 struct nd_namespace_common *ndns;
595 struct nd_namespace_io *nsio;
597 struct badblocks *bb;
598 struct kernfs_node *bb_state;
600 if (event != NVDIMM_REVALIDATE_POISON)
603 if (is_nd_btt(dev)) {
604 struct nd_btt *nd_btt = to_nd_btt(dev);
607 nd_region = to_nd_region(ndns->dev.parent);
608 nsio = to_nd_namespace_io(&ndns->dev);
612 struct pmem_device *pmem = dev_get_drvdata(dev);
614 nd_region = to_region(pmem);
616 bb_state = pmem->bb_state;
618 if (is_nd_pfn(dev)) {
619 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
620 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
623 offset = pmem->data_offset +
624 __le32_to_cpu(pfn_sb->start_pad);
625 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
630 nsio = to_nd_namespace_io(&ndns->dev);
633 res.start = nsio->res.start + offset;
634 res.end = nsio->res.end - end_trunc;
635 nvdimm_badblocks_populate(nd_region, bb, &res);
637 sysfs_notify_dirent(bb_state);
640 MODULE_ALIAS("pmem");
641 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
642 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
643 static struct nd_device_driver nd_pmem_driver = {
644 .probe = nd_pmem_probe,
645 .remove = nd_pmem_remove,
646 .notify = nd_pmem_notify,
647 .shutdown = nd_pmem_shutdown,
651 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
654 module_nd_driver(nd_pmem_driver);
656 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
657 MODULE_LICENSE("GPL v2");