]> asedeno.scripts.mit.edu Git - linux.git/blob - drivers/nvdimm/pmem.c
Merge tag 'armsoc-dt' of git://git.kernel.org/pub/scm/linux/kernel/git/soc/soc
[linux.git] / drivers / nvdimm / pmem.c
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
32 static struct device *to_dev(struct pmem_device *pmem)
33 {
34         /*
35          * nvdimm bus services need a 'dev' parameter, and we record the device
36          * at init in bb.dev.
37          */
38         return pmem->bb.dev;
39 }
40
41 static struct nd_region *to_region(struct pmem_device *pmem)
42 {
43         return to_nd_region(to_dev(pmem)->parent);
44 }
45
46 static void hwpoison_clear(struct pmem_device *pmem,
47                 phys_addr_t phys, unsigned int len)
48 {
49         unsigned long pfn_start, pfn_end, pfn;
50
51         /* only pmem in the linear map supports HWPoison */
52         if (is_vmalloc_addr(pmem->virt_addr))
53                 return;
54
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);
59
60                 /*
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.
64                  */
65                 if (test_and_clear_pmem_poison(page))
66                         clear_mce_nospec(pfn);
67         }
68 }
69
70 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
71                 phys_addr_t offset, unsigned int len)
72 {
73         struct device *dev = to_dev(pmem);
74         sector_t sector;
75         long cleared;
76         blk_status_t rc = BLK_STS_OK;
77
78         sector = (offset - pmem->data_offset) / 512;
79
80         cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
81         if (cleared < len)
82                 rc = BLK_STS_IOERR;
83         if (cleared > 0 && cleared / 512) {
84                 hwpoison_clear(pmem, pmem->phys_addr + offset, cleared);
85                 cleared /= 512;
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);
90                 if (pmem->bb_state)
91                         sysfs_notify_dirent(pmem->bb_state);
92         }
93
94         arch_invalidate_pmem(pmem->virt_addr + offset, len);
95
96         return rc;
97 }
98
99 static void write_pmem(void *pmem_addr, struct page *page,
100                 unsigned int off, unsigned int len)
101 {
102         unsigned int chunk;
103         void *mem;
104
105         while (len) {
106                 mem = kmap_atomic(page);
107                 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
108                 memcpy_flushcache(pmem_addr, mem + off, chunk);
109                 kunmap_atomic(mem);
110                 len -= chunk;
111                 off = 0;
112                 page++;
113                 pmem_addr += chunk;
114         }
115 }
116
117 static blk_status_t read_pmem(struct page *page, unsigned int off,
118                 void *pmem_addr, unsigned int len)
119 {
120         unsigned int chunk;
121         unsigned long rem;
122         void *mem;
123
124         while (len) {
125                 mem = kmap_atomic(page);
126                 chunk = min_t(unsigned int, len, PAGE_SIZE - off);
127                 rem = memcpy_mcsafe(mem + off, pmem_addr, chunk);
128                 kunmap_atomic(mem);
129                 if (rem)
130                         return BLK_STS_IOERR;
131                 len -= chunk;
132                 off = 0;
133                 page++;
134                 pmem_addr += chunk;
135         }
136         return BLK_STS_OK;
137 }
138
139 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
140                         unsigned int len, unsigned int off, unsigned int op,
141                         sector_t sector)
142 {
143         blk_status_t rc = BLK_STS_OK;
144         bool bad_pmem = false;
145         phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
146         void *pmem_addr = pmem->virt_addr + pmem_off;
147
148         if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
149                 bad_pmem = true;
150
151         if (!op_is_write(op)) {
152                 if (unlikely(bad_pmem))
153                         rc = BLK_STS_IOERR;
154                 else {
155                         rc = read_pmem(page, off, pmem_addr, len);
156                         flush_dcache_page(page);
157                 }
158         } else {
159                 /*
160                  * Note that we write the data both before and after
161                  * clearing poison.  The write before clear poison
162                  * handles situations where the latest written data is
163                  * preserved and the clear poison operation simply marks
164                  * the address range as valid without changing the data.
165                  * In this case application software can assume that an
166                  * interrupted write will either return the new good
167                  * data or an error.
168                  *
169                  * However, if pmem_clear_poison() leaves the data in an
170                  * indeterminate state we need to perform the write
171                  * after clear poison.
172                  */
173                 flush_dcache_page(page);
174                 write_pmem(pmem_addr, page, off, len);
175                 if (unlikely(bad_pmem)) {
176                         rc = pmem_clear_poison(pmem, pmem_off, len);
177                         write_pmem(pmem_addr, page, off, len);
178                 }
179         }
180
181         return rc;
182 }
183
184 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
185 {
186         int ret = 0;
187         blk_status_t rc = 0;
188         bool do_acct;
189         unsigned long start;
190         struct bio_vec bvec;
191         struct bvec_iter iter;
192         struct pmem_device *pmem = q->queuedata;
193         struct nd_region *nd_region = to_region(pmem);
194
195         if (bio->bi_opf & REQ_PREFLUSH)
196                 ret = nvdimm_flush(nd_region, bio);
197
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);
202                 if (rc) {
203                         bio->bi_status = rc;
204                         break;
205                 }
206         }
207         if (do_acct)
208                 nd_iostat_end(bio, start);
209
210         if (bio->bi_opf & REQ_FUA)
211                 ret = nvdimm_flush(nd_region, bio);
212
213         if (ret)
214                 bio->bi_status = errno_to_blk_status(ret);
215
216         bio_endio(bio);
217         return BLK_QC_T_NONE;
218 }
219
220 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
221                        struct page *page, unsigned int op)
222 {
223         struct pmem_device *pmem = bdev->bd_queue->queuedata;
224         blk_status_t rc;
225
226         rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
227                           0, op, sector);
228
229         /*
230          * The ->rw_page interface is subtle and tricky.  The core
231          * retries on any error, so we can only invoke page_endio() in
232          * the successful completion case.  Otherwise, we'll see crashes
233          * caused by double completion.
234          */
235         if (rc == 0)
236                 page_endio(page, op_is_write(op), 0);
237
238         return blk_status_to_errno(rc);
239 }
240
241 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
242 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
243                 long nr_pages, void **kaddr, pfn_t *pfn)
244 {
245         resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
246
247         if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
248                                         PFN_PHYS(nr_pages))))
249                 return -EIO;
250
251         if (kaddr)
252                 *kaddr = pmem->virt_addr + offset;
253         if (pfn)
254                 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
255
256         /*
257          * If badblocks are present, limit known good range to the
258          * requested range.
259          */
260         if (unlikely(pmem->bb.count))
261                 return nr_pages;
262         return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
263 }
264
265 static const struct block_device_operations pmem_fops = {
266         .owner =                THIS_MODULE,
267         .rw_page =              pmem_rw_page,
268         .revalidate_disk =      nvdimm_revalidate_disk,
269 };
270
271 static long pmem_dax_direct_access(struct dax_device *dax_dev,
272                 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
273 {
274         struct pmem_device *pmem = dax_get_private(dax_dev);
275
276         return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
277 }
278
279 /*
280  * Use the 'no check' versions of copy_from_iter_flushcache() and
281  * copy_to_iter_mcsafe() to bypass HARDENED_USERCOPY overhead. Bounds
282  * checking, both file offset and device offset, is handled by
283  * dax_iomap_actor()
284  */
285 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
286                 void *addr, size_t bytes, struct iov_iter *i)
287 {
288         return _copy_from_iter_flushcache(addr, bytes, i);
289 }
290
291 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
292                 void *addr, size_t bytes, struct iov_iter *i)
293 {
294         return _copy_to_iter_mcsafe(addr, bytes, i);
295 }
296
297 static const struct dax_operations pmem_dax_ops = {
298         .direct_access = pmem_dax_direct_access,
299         .dax_supported = generic_fsdax_supported,
300         .copy_from_iter = pmem_copy_from_iter,
301         .copy_to_iter = pmem_copy_to_iter,
302 };
303
304 static const struct attribute_group *pmem_attribute_groups[] = {
305         &dax_attribute_group,
306         NULL,
307 };
308
309 static void pmem_pagemap_cleanup(struct dev_pagemap *pgmap)
310 {
311         struct request_queue *q =
312                 container_of(pgmap->ref, struct request_queue, q_usage_counter);
313
314         blk_cleanup_queue(q);
315 }
316
317 static void pmem_release_queue(void *pgmap)
318 {
319         pmem_pagemap_cleanup(pgmap);
320 }
321
322 static void pmem_pagemap_kill(struct dev_pagemap *pgmap)
323 {
324         struct request_queue *q =
325                 container_of(pgmap->ref, struct request_queue, q_usage_counter);
326
327         blk_freeze_queue_start(q);
328 }
329
330 static void pmem_release_disk(void *__pmem)
331 {
332         struct pmem_device *pmem = __pmem;
333
334         kill_dax(pmem->dax_dev);
335         put_dax(pmem->dax_dev);
336         del_gendisk(pmem->disk);
337         put_disk(pmem->disk);
338 }
339
340 static const struct dev_pagemap_ops fsdax_pagemap_ops = {
341         .kill                   = pmem_pagemap_kill,
342         .cleanup                = pmem_pagemap_cleanup,
343 };
344
345 static int pmem_attach_disk(struct device *dev,
346                 struct nd_namespace_common *ndns)
347 {
348         struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
349         struct nd_region *nd_region = to_nd_region(dev->parent);
350         int nid = dev_to_node(dev), fua;
351         struct resource *res = &nsio->res;
352         struct resource bb_res;
353         struct nd_pfn *nd_pfn = NULL;
354         struct dax_device *dax_dev;
355         struct nd_pfn_sb *pfn_sb;
356         struct pmem_device *pmem;
357         struct request_queue *q;
358         struct device *gendev;
359         struct gendisk *disk;
360         void *addr;
361         int rc;
362         unsigned long flags = 0UL;
363
364         pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
365         if (!pmem)
366                 return -ENOMEM;
367
368         rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
369         if (rc)
370                 return rc;
371
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);
376                 if (rc)
377                         return rc;
378         }
379
380         /* we're attaching a block device, disable raw namespace access */
381         devm_namespace_disable(dev, ndns);
382
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");
389                 fua = 0;
390         }
391
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);
395                 return -EBUSY;
396         }
397
398         q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
399         if (!q)
400                 return -ENOMEM;
401
402         pmem->pfn_flags = PFN_DEV;
403         pmem->pgmap.ref = &q->q_usage_counter;
404         if (is_nd_pfn(dev)) {
405                 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
406                 pmem->pgmap.ops = &fsdax_pagemap_ops;
407                 addr = devm_memremap_pages(dev, &pmem->pgmap);
408                 pfn_sb = nd_pfn->pfn_sb;
409                 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
410                 pmem->pfn_pad = resource_size(res) -
411                         resource_size(&pmem->pgmap.res);
412                 pmem->pfn_flags |= PFN_MAP;
413                 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
414                 bb_res.start += pmem->data_offset;
415         } else if (pmem_should_map_pages(dev)) {
416                 memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
417                 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX;
418                 pmem->pgmap.ops = &fsdax_pagemap_ops;
419                 addr = devm_memremap_pages(dev, &pmem->pgmap);
420                 pmem->pfn_flags |= PFN_MAP;
421                 memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
422         } else {
423                 if (devm_add_action_or_reset(dev, pmem_release_queue,
424                                         &pmem->pgmap))
425                         return -ENOMEM;
426                 addr = devm_memremap(dev, pmem->phys_addr,
427                                 pmem->size, ARCH_MEMREMAP_PMEM);
428                 memcpy(&bb_res, &nsio->res, sizeof(bb_res));
429         }
430
431         if (IS_ERR(addr))
432                 return PTR_ERR(addr);
433         pmem->virt_addr = addr;
434
435         blk_queue_write_cache(q, true, fua);
436         blk_queue_make_request(q, pmem_make_request);
437         blk_queue_physical_block_size(q, PAGE_SIZE);
438         blk_queue_logical_block_size(q, pmem_sector_size(ndns));
439         blk_queue_max_hw_sectors(q, UINT_MAX);
440         blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
441         if (pmem->pfn_flags & PFN_MAP)
442                 blk_queue_flag_set(QUEUE_FLAG_DAX, q);
443         q->queuedata = pmem;
444
445         disk = alloc_disk_node(0, nid);
446         if (!disk)
447                 return -ENOMEM;
448         pmem->disk = disk;
449
450         disk->fops              = &pmem_fops;
451         disk->queue             = q;
452         disk->flags             = GENHD_FL_EXT_DEVT;
453         disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
454         nvdimm_namespace_disk_name(ndns, disk->disk_name);
455         set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
456                         / 512);
457         if (devm_init_badblocks(dev, &pmem->bb))
458                 return -ENOMEM;
459         nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res);
460         disk->bb = &pmem->bb;
461
462         if (is_nvdimm_sync(nd_region))
463                 flags = DAXDEV_F_SYNC;
464         dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops, flags);
465         if (!dax_dev) {
466                 put_disk(disk);
467                 return -ENOMEM;
468         }
469         dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
470         pmem->dax_dev = dax_dev;
471         gendev = disk_to_dev(disk);
472         gendev->groups = pmem_attribute_groups;
473
474         device_add_disk(dev, disk, NULL);
475         if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
476                 return -ENOMEM;
477
478         revalidate_disk(disk);
479
480         pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
481                                           "badblocks");
482         if (!pmem->bb_state)
483                 dev_warn(dev, "'badblocks' notification disabled\n");
484
485         return 0;
486 }
487
488 static int nd_pmem_probe(struct device *dev)
489 {
490         int ret;
491         struct nd_namespace_common *ndns;
492
493         ndns = nvdimm_namespace_common_probe(dev);
494         if (IS_ERR(ndns))
495                 return PTR_ERR(ndns);
496
497         if (is_nd_btt(dev))
498                 return nvdimm_namespace_attach_btt(ndns);
499
500         if (is_nd_pfn(dev))
501                 return pmem_attach_disk(dev, ndns);
502
503         ret = devm_namespace_enable(dev, ndns, nd_info_block_reserve());
504         if (ret)
505                 return ret;
506
507         ret = nd_btt_probe(dev, ndns);
508         if (ret == 0)
509                 return -ENXIO;
510
511         /*
512          * We have two failure conditions here, there is no
513          * info reserver block or we found a valid info reserve block
514          * but failed to initialize the pfn superblock.
515          *
516          * For the first case consider namespace as a raw pmem namespace
517          * and attach a disk.
518          *
519          * For the latter, consider this a success and advance the namespace
520          * seed.
521          */
522         ret = nd_pfn_probe(dev, ndns);
523         if (ret == 0)
524                 return -ENXIO;
525         else if (ret == -EOPNOTSUPP)
526                 return ret;
527
528         ret = nd_dax_probe(dev, ndns);
529         if (ret == 0)
530                 return -ENXIO;
531         else if (ret == -EOPNOTSUPP)
532                 return ret;
533
534         /* probe complete, attach handles namespace enabling */
535         devm_namespace_disable(dev, ndns);
536
537         return pmem_attach_disk(dev, ndns);
538 }
539
540 static int nd_pmem_remove(struct device *dev)
541 {
542         struct pmem_device *pmem = dev_get_drvdata(dev);
543
544         if (is_nd_btt(dev))
545                 nvdimm_namespace_detach_btt(to_nd_btt(dev));
546         else {
547                 /*
548                  * Note, this assumes nd_device_lock() context to not
549                  * race nd_pmem_notify()
550                  */
551                 sysfs_put(pmem->bb_state);
552                 pmem->bb_state = NULL;
553         }
554         nvdimm_flush(to_nd_region(dev->parent), NULL);
555
556         return 0;
557 }
558
559 static void nd_pmem_shutdown(struct device *dev)
560 {
561         nvdimm_flush(to_nd_region(dev->parent), NULL);
562 }
563
564 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
565 {
566         struct nd_region *nd_region;
567         resource_size_t offset = 0, end_trunc = 0;
568         struct nd_namespace_common *ndns;
569         struct nd_namespace_io *nsio;
570         struct resource res;
571         struct badblocks *bb;
572         struct kernfs_node *bb_state;
573
574         if (event != NVDIMM_REVALIDATE_POISON)
575                 return;
576
577         if (is_nd_btt(dev)) {
578                 struct nd_btt *nd_btt = to_nd_btt(dev);
579
580                 ndns = nd_btt->ndns;
581                 nd_region = to_nd_region(ndns->dev.parent);
582                 nsio = to_nd_namespace_io(&ndns->dev);
583                 bb = &nsio->bb;
584                 bb_state = NULL;
585         } else {
586                 struct pmem_device *pmem = dev_get_drvdata(dev);
587
588                 nd_region = to_region(pmem);
589                 bb = &pmem->bb;
590                 bb_state = pmem->bb_state;
591
592                 if (is_nd_pfn(dev)) {
593                         struct nd_pfn *nd_pfn = to_nd_pfn(dev);
594                         struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
595
596                         ndns = nd_pfn->ndns;
597                         offset = pmem->data_offset +
598                                         __le32_to_cpu(pfn_sb->start_pad);
599                         end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
600                 } else {
601                         ndns = to_ndns(dev);
602                 }
603
604                 nsio = to_nd_namespace_io(&ndns->dev);
605         }
606
607         res.start = nsio->res.start + offset;
608         res.end = nsio->res.end - end_trunc;
609         nvdimm_badblocks_populate(nd_region, bb, &res);
610         if (bb_state)
611                 sysfs_notify_dirent(bb_state);
612 }
613
614 MODULE_ALIAS("pmem");
615 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
616 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
617 static struct nd_device_driver nd_pmem_driver = {
618         .probe = nd_pmem_probe,
619         .remove = nd_pmem_remove,
620         .notify = nd_pmem_notify,
621         .shutdown = nd_pmem_shutdown,
622         .drv = {
623                 .name = "nd_pmem",
624         },
625         .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
626 };
627
628 module_nd_driver(nd_pmem_driver);
629
630 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
631 MODULE_LICENSE("GPL v2");