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 <asm/cacheflush.h>
11 #include <linux/blkdev.h>
12 #include <linux/hdreg.h>
13 #include <linux/init.h>
14 #include <linux/platform_device.h>
15 #include <linux/set_memory.h>
16 #include <linux/module.h>
17 #include <linux/moduleparam.h>
18 #include <linux/badblocks.h>
19 #include <linux/memremap.h>
20 #include <linux/vmalloc.h>
21 #include <linux/blk-mq.h>
22 #include <linux/pfn_t.h>
23 #include <linux/slab.h>
24 #include <linux/uio.h>
25 #include <linux/dax.h>
27 #include <linux/backing-dev.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 = memcpy_mcsafe(mem + off, pmem_addr, chunk);
131 return BLK_STS_IOERR;
140 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
141 unsigned int len, unsigned int off, unsigned int op,
144 blk_status_t rc = BLK_STS_OK;
145 bool bad_pmem = false;
146 phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
147 void *pmem_addr = pmem->virt_addr + pmem_off;
149 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
152 if (!op_is_write(op)) {
153 if (unlikely(bad_pmem))
156 rc = read_pmem(page, off, pmem_addr, len);
157 flush_dcache_page(page);
161 * Note that we write the data both before and after
162 * clearing poison. The write before clear poison
163 * handles situations where the latest written data is
164 * preserved and the clear poison operation simply marks
165 * the address range as valid without changing the data.
166 * In this case application software can assume that an
167 * interrupted write will either return the new good
170 * However, if pmem_clear_poison() leaves the data in an
171 * indeterminate state we need to perform the write
172 * after clear poison.
174 flush_dcache_page(page);
175 write_pmem(pmem_addr, page, off, len);
176 if (unlikely(bad_pmem)) {
177 rc = pmem_clear_poison(pmem, pmem_off, len);
178 write_pmem(pmem_addr, page, off, len);
185 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
191 struct bvec_iter iter;
192 struct pmem_device *pmem = q->queuedata;
193 struct nd_region *nd_region = to_region(pmem);
195 if (bio->bi_opf & REQ_PREFLUSH)
196 nvdimm_flush(nd_region);
198 do_acct = nd_iostat_start(bio, &start);
199 bio_for_each_segment(bvec, bio, iter) {
200 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
201 bvec.bv_offset, bio_op(bio), iter.bi_sector);
208 nd_iostat_end(bio, start);
210 if (bio->bi_opf & REQ_FUA)
211 nvdimm_flush(nd_region);
214 return BLK_QC_T_NONE;
217 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
218 struct page *page, unsigned int op)
220 struct pmem_device *pmem = bdev->bd_queue->queuedata;
223 rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
227 * The ->rw_page interface is subtle and tricky. The core
228 * retries on any error, so we can only invoke page_endio() in
229 * the successful completion case. Otherwise, we'll see crashes
230 * caused by double completion.
233 page_endio(page, op_is_write(op), 0);
235 return blk_status_to_errno(rc);
238 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
239 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
240 long nr_pages, void **kaddr, pfn_t *pfn)
242 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
244 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
245 PFN_PHYS(nr_pages))))
249 *kaddr = pmem->virt_addr + offset;
251 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
254 * If badblocks are present, limit known good range to the
257 if (unlikely(pmem->bb.count))
259 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
262 static const struct block_device_operations pmem_fops = {
263 .owner = THIS_MODULE,
264 .rw_page = pmem_rw_page,
265 .revalidate_disk = nvdimm_revalidate_disk,
268 static long pmem_dax_direct_access(struct dax_device *dax_dev,
269 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
271 struct pmem_device *pmem = dax_get_private(dax_dev);
273 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
277 * Use the 'no check' versions of copy_from_iter_flushcache() and
278 * copy_to_iter_mcsafe() to bypass HARDENED_USERCOPY overhead. Bounds
279 * checking, both file offset and device offset, is handled by
282 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
283 void *addr, size_t bytes, struct iov_iter *i)
285 return _copy_from_iter_flushcache(addr, bytes, i);
288 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
289 void *addr, size_t bytes, struct iov_iter *i)
291 return _copy_to_iter_mcsafe(addr, bytes, i);
294 static const struct dax_operations pmem_dax_ops = {
295 .direct_access = pmem_dax_direct_access,
296 .dax_supported = generic_fsdax_supported,
297 .copy_from_iter = pmem_copy_from_iter,
298 .copy_to_iter = pmem_copy_to_iter,
301 static const struct attribute_group *pmem_attribute_groups[] = {
302 &dax_attribute_group,
306 static void pmem_release_queue(void *q)
308 blk_cleanup_queue(q);
311 static void pmem_freeze_queue(struct percpu_ref *ref)
313 struct request_queue *q;
315 q = container_of(ref, typeof(*q), q_usage_counter);
316 blk_freeze_queue_start(q);
319 static void pmem_release_disk(void *__pmem)
321 struct pmem_device *pmem = __pmem;
323 kill_dax(pmem->dax_dev);
324 put_dax(pmem->dax_dev);
325 del_gendisk(pmem->disk);
326 put_disk(pmem->disk);
329 static void pmem_release_pgmap_ops(void *__pgmap)
331 dev_pagemap_put_ops();
334 static void fsdax_pagefree(struct page *page, void *data)
336 wake_up_var(&page->_refcount);
339 static int setup_pagemap_fsdax(struct device *dev, struct dev_pagemap *pgmap)
341 dev_pagemap_get_ops();
342 if (devm_add_action_or_reset(dev, pmem_release_pgmap_ops, pgmap))
344 pgmap->type = MEMORY_DEVICE_FS_DAX;
345 pgmap->page_free = fsdax_pagefree;
350 static int pmem_attach_disk(struct device *dev,
351 struct nd_namespace_common *ndns)
353 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
354 struct nd_region *nd_region = to_nd_region(dev->parent);
355 int nid = dev_to_node(dev), fua;
356 struct resource *res = &nsio->res;
357 struct resource bb_res;
358 struct nd_pfn *nd_pfn = NULL;
359 struct dax_device *dax_dev;
360 struct nd_pfn_sb *pfn_sb;
361 struct pmem_device *pmem;
362 struct request_queue *q;
363 struct device *gendev;
364 struct gendisk *disk;
368 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
372 /* while nsio_rw_bytes is active, parse a pfn info block if present */
373 if (is_nd_pfn(dev)) {
374 nd_pfn = to_nd_pfn(dev);
375 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
380 /* we're attaching a block device, disable raw namespace access */
381 devm_nsio_disable(dev, nsio);
383 dev_set_drvdata(dev, pmem);
384 pmem->phys_addr = res->start;
385 pmem->size = resource_size(res);
386 fua = nvdimm_has_flush(nd_region);
387 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
388 dev_warn(dev, "unable to guarantee persistence of writes\n");
392 if (!devm_request_mem_region(dev, res->start, resource_size(res),
393 dev_name(&ndns->dev))) {
394 dev_warn(dev, "could not reserve region %pR\n", res);
398 q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
402 if (devm_add_action_or_reset(dev, pmem_release_queue, q))
405 pmem->pfn_flags = PFN_DEV;
406 pmem->pgmap.ref = &q->q_usage_counter;
407 pmem->pgmap.kill = pmem_freeze_queue;
408 if (is_nd_pfn(dev)) {
409 if (setup_pagemap_fsdax(dev, &pmem->pgmap))
411 addr = devm_memremap_pages(dev, &pmem->pgmap);
412 pfn_sb = nd_pfn->pfn_sb;
413 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
414 pmem->pfn_pad = resource_size(res) -
415 resource_size(&pmem->pgmap.res);
416 pmem->pfn_flags |= PFN_MAP;
417 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
418 bb_res.start += pmem->data_offset;
419 } else if (pmem_should_map_pages(dev)) {
420 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
421 pmem->pgmap.altmap_valid = false;
422 if (setup_pagemap_fsdax(dev, &pmem->pgmap))
424 addr = devm_memremap_pages(dev, &pmem->pgmap);
425 pmem->pfn_flags |= PFN_MAP;
426 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
428 addr = devm_memremap(dev, pmem->phys_addr,
429 pmem->size, ARCH_MEMREMAP_PMEM);
430 memcpy(&bb_res, &nsio->res, sizeof(bb_res));
434 return PTR_ERR(addr);
435 pmem->virt_addr = addr;
437 blk_queue_write_cache(q, true, fua);
438 blk_queue_make_request(q, pmem_make_request);
439 blk_queue_physical_block_size(q, PAGE_SIZE);
440 blk_queue_logical_block_size(q, pmem_sector_size(ndns));
441 blk_queue_max_hw_sectors(q, UINT_MAX);
442 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
443 if (pmem->pfn_flags & PFN_MAP)
444 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
447 disk = alloc_disk_node(0, nid);
452 disk->fops = &pmem_fops;
454 disk->flags = GENHD_FL_EXT_DEVT;
455 disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
456 nvdimm_namespace_disk_name(ndns, disk->disk_name);
457 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
459 if (devm_init_badblocks(dev, &pmem->bb))
461 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res);
462 disk->bb = &pmem->bb;
464 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops);
469 dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
470 pmem->dax_dev = dax_dev;
472 gendev = disk_to_dev(disk);
473 gendev->groups = pmem_attribute_groups;
475 device_add_disk(dev, disk, NULL);
476 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
479 revalidate_disk(disk);
481 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
484 dev_warn(dev, "'badblocks' notification disabled\n");
489 static int nd_pmem_probe(struct device *dev)
491 struct nd_namespace_common *ndns;
493 ndns = nvdimm_namespace_common_probe(dev);
495 return PTR_ERR(ndns);
497 if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
501 return nvdimm_namespace_attach_btt(ndns);
504 return pmem_attach_disk(dev, ndns);
506 /* if we find a valid info-block we'll come back as that personality */
507 if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
508 || nd_dax_probe(dev, ndns) == 0)
511 /* ...otherwise we're just a raw pmem device */
512 return pmem_attach_disk(dev, ndns);
515 static int nd_pmem_remove(struct device *dev)
517 struct pmem_device *pmem = dev_get_drvdata(dev);
520 nvdimm_namespace_detach_btt(to_nd_btt(dev));
523 * Note, this assumes device_lock() context to not race
526 sysfs_put(pmem->bb_state);
527 pmem->bb_state = NULL;
529 nvdimm_flush(to_nd_region(dev->parent));
534 static void nd_pmem_shutdown(struct device *dev)
536 nvdimm_flush(to_nd_region(dev->parent));
539 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
541 struct nd_region *nd_region;
542 resource_size_t offset = 0, end_trunc = 0;
543 struct nd_namespace_common *ndns;
544 struct nd_namespace_io *nsio;
546 struct badblocks *bb;
547 struct kernfs_node *bb_state;
549 if (event != NVDIMM_REVALIDATE_POISON)
552 if (is_nd_btt(dev)) {
553 struct nd_btt *nd_btt = to_nd_btt(dev);
556 nd_region = to_nd_region(ndns->dev.parent);
557 nsio = to_nd_namespace_io(&ndns->dev);
561 struct pmem_device *pmem = dev_get_drvdata(dev);
563 nd_region = to_region(pmem);
565 bb_state = pmem->bb_state;
567 if (is_nd_pfn(dev)) {
568 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
569 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
572 offset = pmem->data_offset +
573 __le32_to_cpu(pfn_sb->start_pad);
574 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
579 nsio = to_nd_namespace_io(&ndns->dev);
582 res.start = nsio->res.start + offset;
583 res.end = nsio->res.end - end_trunc;
584 nvdimm_badblocks_populate(nd_region, bb, &res);
586 sysfs_notify_dirent(bb_state);
589 MODULE_ALIAS("pmem");
590 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
591 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
592 static struct nd_device_driver nd_pmem_driver = {
593 .probe = nd_pmem_probe,
594 .remove = nd_pmem_remove,
595 .notify = nd_pmem_notify,
596 .shutdown = nd_pmem_shutdown,
600 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
603 module_nd_driver(nd_pmem_driver);
605 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
606 MODULE_LICENSE("GPL v2");