Loading...
Note: File does not exist in v3.1.
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 <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>
26#include <linux/nd.h>
27#include <linux/backing-dev.h>
28#include "pmem.h"
29#include "pfn.h"
30#include "nd.h"
31#include "nd-core.h"
32
33static 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
42static struct nd_region *to_region(struct pmem_device *pmem)
43{
44 return to_nd_region(to_dev(pmem)->parent);
45}
46
47static 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
71static 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
100static 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
118static 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
140static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
141 unsigned int len, unsigned int off, unsigned int op,
142 sector_t sector)
143{
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;
148
149 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
150 bad_pmem = true;
151
152 if (!op_is_write(op)) {
153 if (unlikely(bad_pmem))
154 rc = BLK_STS_IOERR;
155 else {
156 rc = read_pmem(page, off, pmem_addr, len);
157 flush_dcache_page(page);
158 }
159 } else {
160 /*
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
168 * data or an error.
169 *
170 * However, if pmem_clear_poison() leaves the data in an
171 * indeterminate state we need to perform the write
172 * after clear poison.
173 */
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);
179 }
180 }
181
182 return rc;
183}
184
185static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
186{
187 int ret = 0;
188 blk_status_t rc = 0;
189 bool do_acct;
190 unsigned long start;
191 struct bio_vec bvec;
192 struct bvec_iter iter;
193 struct pmem_device *pmem = q->queuedata;
194 struct nd_region *nd_region = to_region(pmem);
195
196 if (bio->bi_opf & REQ_PREFLUSH)
197 ret = nvdimm_flush(nd_region, bio);
198
199 do_acct = nd_iostat_start(bio, &start);
200 bio_for_each_segment(bvec, bio, iter) {
201 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
202 bvec.bv_offset, bio_op(bio), iter.bi_sector);
203 if (rc) {
204 bio->bi_status = rc;
205 break;
206 }
207 }
208 if (do_acct)
209 nd_iostat_end(bio, start);
210
211 if (bio->bi_opf & REQ_FUA)
212 ret = nvdimm_flush(nd_region, bio);
213
214 if (ret)
215 bio->bi_status = errno_to_blk_status(ret);
216
217 bio_endio(bio);
218 return BLK_QC_T_NONE;
219}
220
221static int pmem_rw_page(struct block_device *bdev, sector_t sector,
222 struct page *page, unsigned int op)
223{
224 struct pmem_device *pmem = bdev->bd_queue->queuedata;
225 blk_status_t rc;
226
227 rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
228 0, op, sector);
229
230 /*
231 * The ->rw_page interface is subtle and tricky. The core
232 * retries on any error, so we can only invoke page_endio() in
233 * the successful completion case. Otherwise, we'll see crashes
234 * caused by double completion.
235 */
236 if (rc == 0)
237 page_endio(page, op_is_write(op), 0);
238
239 return blk_status_to_errno(rc);
240}
241
242/* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
243__weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
244 long nr_pages, void **kaddr, pfn_t *pfn)
245{
246 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
247
248 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
249 PFN_PHYS(nr_pages))))
250 return -EIO;
251
252 if (kaddr)
253 *kaddr = pmem->virt_addr + offset;
254 if (pfn)
255 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
256
257 /*
258 * If badblocks are present, limit known good range to the
259 * requested range.
260 */
261 if (unlikely(pmem->bb.count))
262 return nr_pages;
263 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
264}
265
266static const struct block_device_operations pmem_fops = {
267 .owner = THIS_MODULE,
268 .rw_page = pmem_rw_page,
269 .revalidate_disk = nvdimm_revalidate_disk,
270};
271
272static long pmem_dax_direct_access(struct dax_device *dax_dev,
273 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
274{
275 struct pmem_device *pmem = dax_get_private(dax_dev);
276
277 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
278}
279
280/*
281 * Use the 'no check' versions of copy_from_iter_flushcache() and
282 * copy_to_iter_mcsafe() to bypass HARDENED_USERCOPY overhead. Bounds
283 * checking, both file offset and device offset, is handled by
284 * dax_iomap_actor()
285 */
286static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
287 void *addr, size_t bytes, struct iov_iter *i)
288{
289 return _copy_from_iter_flushcache(addr, bytes, i);
290}
291
292static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
293 void *addr, size_t bytes, struct iov_iter *i)
294{
295 return _copy_to_iter_mcsafe(addr, bytes, i);
296}
297
298static const struct dax_operations pmem_dax_ops = {
299 .direct_access = pmem_dax_direct_access,
300 .dax_supported = generic_fsdax_supported,
301 .copy_from_iter = pmem_copy_from_iter,
302 .copy_to_iter = pmem_copy_to_iter,
303};
304
305static const struct attribute_group *pmem_attribute_groups[] = {
306 &dax_attribute_group,
307 NULL,
308};
309
310static void pmem_pagemap_cleanup(struct dev_pagemap *pgmap)
311{
312 struct request_queue *q =
313 container_of(pgmap->ref, struct request_queue, q_usage_counter);
314
315 blk_cleanup_queue(q);
316}
317
318static void pmem_release_queue(void *pgmap)
319{
320 pmem_pagemap_cleanup(pgmap);
321}
322
323static void pmem_pagemap_kill(struct dev_pagemap *pgmap)
324{
325 struct request_queue *q =
326 container_of(pgmap->ref, struct request_queue, q_usage_counter);
327
328 blk_freeze_queue_start(q);
329}
330
331static void pmem_release_disk(void *__pmem)
332{
333 struct pmem_device *pmem = __pmem;
334
335 kill_dax(pmem->dax_dev);
336 put_dax(pmem->dax_dev);
337 del_gendisk(pmem->disk);
338 put_disk(pmem->disk);
339}
340
341static void pmem_pagemap_page_free(struct page *page)
342{
343 wake_up_var(&page->_refcount);
344}
345
346static const struct dev_pagemap_ops fsdax_pagemap_ops = {
347 .page_free = pmem_pagemap_page_free,
348 .kill = pmem_pagemap_kill,
349 .cleanup = pmem_pagemap_cleanup,
350};
351
352static int pmem_attach_disk(struct device *dev,
353 struct nd_namespace_common *ndns)
354{
355 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
356 struct nd_region *nd_region = to_nd_region(dev->parent);
357 int nid = dev_to_node(dev), fua;
358 struct resource *res = &nsio->res;
359 struct resource bb_res;
360 struct nd_pfn *nd_pfn = NULL;
361 struct dax_device *dax_dev;
362 struct nd_pfn_sb *pfn_sb;
363 struct pmem_device *pmem;
364 struct request_queue *q;
365 struct device *gendev;
366 struct gendisk *disk;
367 void *addr;
368 int rc;
369 unsigned long flags = 0UL;
370
371 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
372 if (!pmem)
373 return -ENOMEM;
374
375 /* while nsio_rw_bytes is active, parse a pfn info block if present */
376 if (is_nd_pfn(dev)) {
377 nd_pfn = to_nd_pfn(dev);
378 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
379 if (rc)
380 return rc;
381 }
382
383 /* we're attaching a block device, disable raw namespace access */
384 devm_nsio_disable(dev, nsio);
385
386 dev_set_drvdata(dev, pmem);
387 pmem->phys_addr = res->start;
388 pmem->size = resource_size(res);
389 fua = nvdimm_has_flush(nd_region);
390 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
391 dev_warn(dev, "unable to guarantee persistence of writes\n");
392 fua = 0;
393 }
394
395 if (!devm_request_mem_region(dev, res->start, resource_size(res),
396 dev_name(&ndns->dev))) {
397 dev_warn(dev, "could not reserve region %pR\n", res);
398 return -EBUSY;
399 }
400
401 q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
402 if (!q)
403 return -ENOMEM;
404
405 pmem->pfn_flags = PFN_DEV;
406 pmem->pgmap.ref = &q->q_usage_counter;
407 if (is_nd_pfn(dev)) {
408 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
409 pmem->pgmap.ops = &fsdax_pagemap_ops;
410 addr = devm_memremap_pages(dev, &pmem->pgmap);
411 pfn_sb = nd_pfn->pfn_sb;
412 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
413 pmem->pfn_pad = resource_size(res) -
414 resource_size(&pmem->pgmap.res);
415 pmem->pfn_flags |= PFN_MAP;
416 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
417 bb_res.start += pmem->data_offset;
418 } else if (pmem_should_map_pages(dev)) {
419 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
420 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
421 pmem->pgmap.ops = &fsdax_pagemap_ops;
422 addr = devm_memremap_pages(dev, &pmem->pgmap);
423 pmem->pfn_flags |= PFN_MAP;
424 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
425 } else {
426 if (devm_add_action_or_reset(dev, pmem_release_queue,
427 &pmem->pgmap))
428 return -ENOMEM;
429 addr = devm_memremap(dev, pmem->phys_addr,
430 pmem->size, ARCH_MEMREMAP_PMEM);
431 memcpy(&bb_res, &nsio->res, sizeof(bb_res));
432 }
433
434 if (IS_ERR(addr))
435 return PTR_ERR(addr);
436 pmem->virt_addr = addr;
437
438 blk_queue_write_cache(q, true, fua);
439 blk_queue_make_request(q, pmem_make_request);
440 blk_queue_physical_block_size(q, PAGE_SIZE);
441 blk_queue_logical_block_size(q, pmem_sector_size(ndns));
442 blk_queue_max_hw_sectors(q, UINT_MAX);
443 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
444 if (pmem->pfn_flags & PFN_MAP)
445 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
446 q->queuedata = pmem;
447
448 disk = alloc_disk_node(0, nid);
449 if (!disk)
450 return -ENOMEM;
451 pmem->disk = disk;
452
453 disk->fops = &pmem_fops;
454 disk->queue = q;
455 disk->flags = GENHD_FL_EXT_DEVT;
456 disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
457 nvdimm_namespace_disk_name(ndns, disk->disk_name);
458 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
459 / 512);
460 if (devm_init_badblocks(dev, &pmem->bb))
461 return -ENOMEM;
462 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res);
463 disk->bb = &pmem->bb;
464
465 if (is_nvdimm_sync(nd_region))
466 flags = DAXDEV_F_SYNC;
467 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops, flags);
468 if (!dax_dev) {
469 put_disk(disk);
470 return -ENOMEM;
471 }
472 dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
473 pmem->dax_dev = dax_dev;
474 gendev = disk_to_dev(disk);
475 gendev->groups = pmem_attribute_groups;
476
477 device_add_disk(dev, disk, NULL);
478 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
479 return -ENOMEM;
480
481 revalidate_disk(disk);
482
483 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
484 "badblocks");
485 if (!pmem->bb_state)
486 dev_warn(dev, "'badblocks' notification disabled\n");
487
488 return 0;
489}
490
491static int nd_pmem_probe(struct device *dev)
492{
493 int ret;
494 struct nd_namespace_common *ndns;
495
496 ndns = nvdimm_namespace_common_probe(dev);
497 if (IS_ERR(ndns))
498 return PTR_ERR(ndns);
499
500 if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
501 return -ENXIO;
502
503 if (is_nd_btt(dev))
504 return nvdimm_namespace_attach_btt(ndns);
505
506 if (is_nd_pfn(dev))
507 return pmem_attach_disk(dev, ndns);
508
509 ret = nd_btt_probe(dev, ndns);
510 if (ret == 0)
511 return -ENXIO;
512
513 /*
514 * We have two failure conditions here, there is no
515 * info reserver block or we found a valid info reserve block
516 * but failed to initialize the pfn superblock.
517 *
518 * For the first case consider namespace as a raw pmem namespace
519 * and attach a disk.
520 *
521 * For the latter, consider this a success and advance the namespace
522 * seed.
523 */
524 ret = nd_pfn_probe(dev, ndns);
525 if (ret == 0)
526 return -ENXIO;
527 else if (ret == -EOPNOTSUPP)
528 return ret;
529
530 ret = nd_dax_probe(dev, ndns);
531 if (ret == 0)
532 return -ENXIO;
533 else if (ret == -EOPNOTSUPP)
534 return ret;
535 return pmem_attach_disk(dev, ndns);
536}
537
538static int nd_pmem_remove(struct device *dev)
539{
540 struct pmem_device *pmem = dev_get_drvdata(dev);
541
542 if (is_nd_btt(dev))
543 nvdimm_namespace_detach_btt(to_nd_btt(dev));
544 else {
545 /*
546 * Note, this assumes nd_device_lock() context to not
547 * race nd_pmem_notify()
548 */
549 sysfs_put(pmem->bb_state);
550 pmem->bb_state = NULL;
551 }
552 nvdimm_flush(to_nd_region(dev->parent), NULL);
553
554 return 0;
555}
556
557static void nd_pmem_shutdown(struct device *dev)
558{
559 nvdimm_flush(to_nd_region(dev->parent), NULL);
560}
561
562static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
563{
564 struct nd_region *nd_region;
565 resource_size_t offset = 0, end_trunc = 0;
566 struct nd_namespace_common *ndns;
567 struct nd_namespace_io *nsio;
568 struct resource res;
569 struct badblocks *bb;
570 struct kernfs_node *bb_state;
571
572 if (event != NVDIMM_REVALIDATE_POISON)
573 return;
574
575 if (is_nd_btt(dev)) {
576 struct nd_btt *nd_btt = to_nd_btt(dev);
577
578 ndns = nd_btt->ndns;
579 nd_region = to_nd_region(ndns->dev.parent);
580 nsio = to_nd_namespace_io(&ndns->dev);
581 bb = &nsio->bb;
582 bb_state = NULL;
583 } else {
584 struct pmem_device *pmem = dev_get_drvdata(dev);
585
586 nd_region = to_region(pmem);
587 bb = &pmem->bb;
588 bb_state = pmem->bb_state;
589
590 if (is_nd_pfn(dev)) {
591 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
592 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
593
594 ndns = nd_pfn->ndns;
595 offset = pmem->data_offset +
596 __le32_to_cpu(pfn_sb->start_pad);
597 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
598 } else {
599 ndns = to_ndns(dev);
600 }
601
602 nsio = to_nd_namespace_io(&ndns->dev);
603 }
604
605 res.start = nsio->res.start + offset;
606 res.end = nsio->res.end - end_trunc;
607 nvdimm_badblocks_populate(nd_region, bb, &res);
608 if (bb_state)
609 sysfs_notify_dirent(bb_state);
610}
611
612MODULE_ALIAS("pmem");
613MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
614MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
615static struct nd_device_driver nd_pmem_driver = {
616 .probe = nd_pmem_probe,
617 .remove = nd_pmem_remove,
618 .notify = nd_pmem_notify,
619 .shutdown = nd_pmem_shutdown,
620 .drv = {
621 .name = "nd_pmem",
622 },
623 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
624};
625
626module_nd_driver(nd_pmem_driver);
627
628MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
629MODULE_LICENSE("GPL v2");