Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright(c) 2017 Intel Corporation. All rights reserved.
4 */
5#include <linux/pagemap.h>
6#include <linux/module.h>
7#include <linux/mount.h>
8#include <linux/pseudo_fs.h>
9#include <linux/magic.h>
10#include <linux/genhd.h>
11#include <linux/pfn_t.h>
12#include <linux/cdev.h>
13#include <linux/hash.h>
14#include <linux/slab.h>
15#include <linux/uio.h>
16#include <linux/dax.h>
17#include <linux/fs.h>
18#include "dax-private.h"
19
20static dev_t dax_devt;
21DEFINE_STATIC_SRCU(dax_srcu);
22static struct vfsmount *dax_mnt;
23static DEFINE_IDA(dax_minor_ida);
24static struct kmem_cache *dax_cache __read_mostly;
25static struct super_block *dax_superblock __read_mostly;
26
27#define DAX_HASH_SIZE (PAGE_SIZE / sizeof(struct hlist_head))
28static struct hlist_head dax_host_list[DAX_HASH_SIZE];
29static DEFINE_SPINLOCK(dax_host_lock);
30
31int dax_read_lock(void)
32{
33 return srcu_read_lock(&dax_srcu);
34}
35EXPORT_SYMBOL_GPL(dax_read_lock);
36
37void dax_read_unlock(int id)
38{
39 srcu_read_unlock(&dax_srcu, id);
40}
41EXPORT_SYMBOL_GPL(dax_read_unlock);
42
43#ifdef CONFIG_BLOCK
44#include <linux/blkdev.h>
45
46int bdev_dax_pgoff(struct block_device *bdev, sector_t sector, size_t size,
47 pgoff_t *pgoff)
48{
49 phys_addr_t phys_off = (get_start_sect(bdev) + sector) * 512;
50
51 if (pgoff)
52 *pgoff = PHYS_PFN(phys_off);
53 if (phys_off % PAGE_SIZE || size % PAGE_SIZE)
54 return -EINVAL;
55 return 0;
56}
57EXPORT_SYMBOL(bdev_dax_pgoff);
58
59#if IS_ENABLED(CONFIG_FS_DAX)
60struct dax_device *fs_dax_get_by_bdev(struct block_device *bdev)
61{
62 if (!blk_queue_dax(bdev->bd_disk->queue))
63 return NULL;
64 return dax_get_by_host(bdev->bd_disk->disk_name);
65}
66EXPORT_SYMBOL_GPL(fs_dax_get_by_bdev);
67#endif
68
69bool __generic_fsdax_supported(struct dax_device *dax_dev,
70 struct block_device *bdev, int blocksize, sector_t start,
71 sector_t sectors)
72{
73 bool dax_enabled = false;
74 pgoff_t pgoff, pgoff_end;
75 char buf[BDEVNAME_SIZE];
76 void *kaddr, *end_kaddr;
77 pfn_t pfn, end_pfn;
78 sector_t last_page;
79 long len, len2;
80 int err, id;
81
82 if (blocksize != PAGE_SIZE) {
83 pr_info("%s: error: unsupported blocksize for dax\n",
84 bdevname(bdev, buf));
85 return false;
86 }
87
88 if (!dax_dev) {
89 pr_debug("%s: error: dax unsupported by block device\n",
90 bdevname(bdev, buf));
91 return false;
92 }
93
94 err = bdev_dax_pgoff(bdev, start, PAGE_SIZE, &pgoff);
95 if (err) {
96 pr_info("%s: error: unaligned partition for dax\n",
97 bdevname(bdev, buf));
98 return false;
99 }
100
101 last_page = PFN_DOWN((start + sectors - 1) * 512) * PAGE_SIZE / 512;
102 err = bdev_dax_pgoff(bdev, last_page, PAGE_SIZE, &pgoff_end);
103 if (err) {
104 pr_info("%s: error: unaligned partition for dax\n",
105 bdevname(bdev, buf));
106 return false;
107 }
108
109 id = dax_read_lock();
110 len = dax_direct_access(dax_dev, pgoff, 1, &kaddr, &pfn);
111 len2 = dax_direct_access(dax_dev, pgoff_end, 1, &end_kaddr, &end_pfn);
112
113 if (len < 1 || len2 < 1) {
114 pr_info("%s: error: dax access failed (%ld)\n",
115 bdevname(bdev, buf), len < 1 ? len : len2);
116 dax_read_unlock(id);
117 return false;
118 }
119
120 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED) && pfn_t_special(pfn)) {
121 /*
122 * An arch that has enabled the pmem api should also
123 * have its drivers support pfn_t_devmap()
124 *
125 * This is a developer warning and should not trigger in
126 * production. dax_flush() will crash since it depends
127 * on being able to do (page_address(pfn_to_page())).
128 */
129 WARN_ON(IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API));
130 dax_enabled = true;
131 } else if (pfn_t_devmap(pfn) && pfn_t_devmap(end_pfn)) {
132 struct dev_pagemap *pgmap, *end_pgmap;
133
134 pgmap = get_dev_pagemap(pfn_t_to_pfn(pfn), NULL);
135 end_pgmap = get_dev_pagemap(pfn_t_to_pfn(end_pfn), NULL);
136 if (pgmap && pgmap == end_pgmap && pgmap->type == MEMORY_DEVICE_FS_DAX
137 && pfn_t_to_page(pfn)->pgmap == pgmap
138 && pfn_t_to_page(end_pfn)->pgmap == pgmap
139 && pfn_t_to_pfn(pfn) == PHYS_PFN(__pa(kaddr))
140 && pfn_t_to_pfn(end_pfn) == PHYS_PFN(__pa(end_kaddr)))
141 dax_enabled = true;
142 put_dev_pagemap(pgmap);
143 put_dev_pagemap(end_pgmap);
144
145 }
146 dax_read_unlock(id);
147
148 if (!dax_enabled) {
149 pr_info("%s: error: dax support not enabled\n",
150 bdevname(bdev, buf));
151 return false;
152 }
153 return true;
154}
155EXPORT_SYMBOL_GPL(__generic_fsdax_supported);
156
157/**
158 * __bdev_dax_supported() - Check if the device supports dax for filesystem
159 * @bdev: block device to check
160 * @blocksize: The block size of the device
161 *
162 * This is a library function for filesystems to check if the block device
163 * can be mounted with dax option.
164 *
165 * Return: true if supported, false if unsupported
166 */
167bool __bdev_dax_supported(struct block_device *bdev, int blocksize)
168{
169 struct dax_device *dax_dev;
170 struct request_queue *q;
171 char buf[BDEVNAME_SIZE];
172 bool ret;
173 int id;
174
175 q = bdev_get_queue(bdev);
176 if (!q || !blk_queue_dax(q)) {
177 pr_debug("%s: error: request queue doesn't support dax\n",
178 bdevname(bdev, buf));
179 return false;
180 }
181
182 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
183 if (!dax_dev) {
184 pr_debug("%s: error: device does not support dax\n",
185 bdevname(bdev, buf));
186 return false;
187 }
188
189 id = dax_read_lock();
190 ret = dax_supported(dax_dev, bdev, blocksize, 0,
191 i_size_read(bdev->bd_inode) / 512);
192 dax_read_unlock(id);
193
194 put_dax(dax_dev);
195
196 return ret;
197}
198EXPORT_SYMBOL_GPL(__bdev_dax_supported);
199#endif
200
201enum dax_device_flags {
202 /* !alive + rcu grace period == no new operations / mappings */
203 DAXDEV_ALIVE,
204 /* gate whether dax_flush() calls the low level flush routine */
205 DAXDEV_WRITE_CACHE,
206 /* flag to check if device supports synchronous flush */
207 DAXDEV_SYNC,
208};
209
210/**
211 * struct dax_device - anchor object for dax services
212 * @inode: core vfs
213 * @cdev: optional character interface for "device dax"
214 * @host: optional name for lookups where the device path is not available
215 * @private: dax driver private data
216 * @flags: state and boolean properties
217 */
218struct dax_device {
219 struct hlist_node list;
220 struct inode inode;
221 struct cdev cdev;
222 const char *host;
223 void *private;
224 unsigned long flags;
225 const struct dax_operations *ops;
226};
227
228static ssize_t write_cache_show(struct device *dev,
229 struct device_attribute *attr, char *buf)
230{
231 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
232 ssize_t rc;
233
234 WARN_ON_ONCE(!dax_dev);
235 if (!dax_dev)
236 return -ENXIO;
237
238 rc = sprintf(buf, "%d\n", !!dax_write_cache_enabled(dax_dev));
239 put_dax(dax_dev);
240 return rc;
241}
242
243static ssize_t write_cache_store(struct device *dev,
244 struct device_attribute *attr, const char *buf, size_t len)
245{
246 bool write_cache;
247 int rc = strtobool(buf, &write_cache);
248 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
249
250 WARN_ON_ONCE(!dax_dev);
251 if (!dax_dev)
252 return -ENXIO;
253
254 if (rc)
255 len = rc;
256 else
257 dax_write_cache(dax_dev, write_cache);
258
259 put_dax(dax_dev);
260 return len;
261}
262static DEVICE_ATTR_RW(write_cache);
263
264static umode_t dax_visible(struct kobject *kobj, struct attribute *a, int n)
265{
266 struct device *dev = container_of(kobj, typeof(*dev), kobj);
267 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
268
269 WARN_ON_ONCE(!dax_dev);
270 if (!dax_dev)
271 return 0;
272
273#ifndef CONFIG_ARCH_HAS_PMEM_API
274 if (a == &dev_attr_write_cache.attr)
275 return 0;
276#endif
277 return a->mode;
278}
279
280static struct attribute *dax_attributes[] = {
281 &dev_attr_write_cache.attr,
282 NULL,
283};
284
285struct attribute_group dax_attribute_group = {
286 .name = "dax",
287 .attrs = dax_attributes,
288 .is_visible = dax_visible,
289};
290EXPORT_SYMBOL_GPL(dax_attribute_group);
291
292/**
293 * dax_direct_access() - translate a device pgoff to an absolute pfn
294 * @dax_dev: a dax_device instance representing the logical memory range
295 * @pgoff: offset in pages from the start of the device to translate
296 * @nr_pages: number of consecutive pages caller can handle relative to @pfn
297 * @kaddr: output parameter that returns a virtual address mapping of pfn
298 * @pfn: output parameter that returns an absolute pfn translation of @pgoff
299 *
300 * Return: negative errno if an error occurs, otherwise the number of
301 * pages accessible at the device relative @pgoff.
302 */
303long dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff, long nr_pages,
304 void **kaddr, pfn_t *pfn)
305{
306 long avail;
307
308 if (!dax_dev)
309 return -EOPNOTSUPP;
310
311 if (!dax_alive(dax_dev))
312 return -ENXIO;
313
314 if (nr_pages < 0)
315 return nr_pages;
316
317 avail = dax_dev->ops->direct_access(dax_dev, pgoff, nr_pages,
318 kaddr, pfn);
319 if (!avail)
320 return -ERANGE;
321 return min(avail, nr_pages);
322}
323EXPORT_SYMBOL_GPL(dax_direct_access);
324
325bool dax_supported(struct dax_device *dax_dev, struct block_device *bdev,
326 int blocksize, sector_t start, sector_t len)
327{
328 if (!dax_dev)
329 return false;
330
331 if (!dax_alive(dax_dev))
332 return false;
333
334 return dax_dev->ops->dax_supported(dax_dev, bdev, blocksize, start, len);
335}
336EXPORT_SYMBOL_GPL(dax_supported);
337
338size_t dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
339 size_t bytes, struct iov_iter *i)
340{
341 if (!dax_alive(dax_dev))
342 return 0;
343
344 return dax_dev->ops->copy_from_iter(dax_dev, pgoff, addr, bytes, i);
345}
346EXPORT_SYMBOL_GPL(dax_copy_from_iter);
347
348size_t dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
349 size_t bytes, struct iov_iter *i)
350{
351 if (!dax_alive(dax_dev))
352 return 0;
353
354 return dax_dev->ops->copy_to_iter(dax_dev, pgoff, addr, bytes, i);
355}
356EXPORT_SYMBOL_GPL(dax_copy_to_iter);
357
358int dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
359 size_t nr_pages)
360{
361 if (!dax_alive(dax_dev))
362 return -ENXIO;
363 /*
364 * There are no callers that want to zero more than one page as of now.
365 * Once users are there, this check can be removed after the
366 * device mapper code has been updated to split ranges across targets.
367 */
368 if (nr_pages != 1)
369 return -EIO;
370
371 return dax_dev->ops->zero_page_range(dax_dev, pgoff, nr_pages);
372}
373EXPORT_SYMBOL_GPL(dax_zero_page_range);
374
375#ifdef CONFIG_ARCH_HAS_PMEM_API
376void arch_wb_cache_pmem(void *addr, size_t size);
377void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
378{
379 if (unlikely(!dax_write_cache_enabled(dax_dev)))
380 return;
381
382 arch_wb_cache_pmem(addr, size);
383}
384#else
385void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
386{
387}
388#endif
389EXPORT_SYMBOL_GPL(dax_flush);
390
391void dax_write_cache(struct dax_device *dax_dev, bool wc)
392{
393 if (wc)
394 set_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
395 else
396 clear_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
397}
398EXPORT_SYMBOL_GPL(dax_write_cache);
399
400bool dax_write_cache_enabled(struct dax_device *dax_dev)
401{
402 return test_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
403}
404EXPORT_SYMBOL_GPL(dax_write_cache_enabled);
405
406bool __dax_synchronous(struct dax_device *dax_dev)
407{
408 return test_bit(DAXDEV_SYNC, &dax_dev->flags);
409}
410EXPORT_SYMBOL_GPL(__dax_synchronous);
411
412void __set_dax_synchronous(struct dax_device *dax_dev)
413{
414 set_bit(DAXDEV_SYNC, &dax_dev->flags);
415}
416EXPORT_SYMBOL_GPL(__set_dax_synchronous);
417
418bool dax_alive(struct dax_device *dax_dev)
419{
420 lockdep_assert_held(&dax_srcu);
421 return test_bit(DAXDEV_ALIVE, &dax_dev->flags);
422}
423EXPORT_SYMBOL_GPL(dax_alive);
424
425static int dax_host_hash(const char *host)
426{
427 return hashlen_hash(hashlen_string("DAX", host)) % DAX_HASH_SIZE;
428}
429
430/*
431 * Note, rcu is not protecting the liveness of dax_dev, rcu is ensuring
432 * that any fault handlers or operations that might have seen
433 * dax_alive(), have completed. Any operations that start after
434 * synchronize_srcu() has run will abort upon seeing !dax_alive().
435 */
436void kill_dax(struct dax_device *dax_dev)
437{
438 if (!dax_dev)
439 return;
440
441 clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
442
443 synchronize_srcu(&dax_srcu);
444
445 spin_lock(&dax_host_lock);
446 hlist_del_init(&dax_dev->list);
447 spin_unlock(&dax_host_lock);
448}
449EXPORT_SYMBOL_GPL(kill_dax);
450
451void run_dax(struct dax_device *dax_dev)
452{
453 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
454}
455EXPORT_SYMBOL_GPL(run_dax);
456
457static struct inode *dax_alloc_inode(struct super_block *sb)
458{
459 struct dax_device *dax_dev;
460 struct inode *inode;
461
462 dax_dev = kmem_cache_alloc(dax_cache, GFP_KERNEL);
463 if (!dax_dev)
464 return NULL;
465
466 inode = &dax_dev->inode;
467 inode->i_rdev = 0;
468 return inode;
469}
470
471static struct dax_device *to_dax_dev(struct inode *inode)
472{
473 return container_of(inode, struct dax_device, inode);
474}
475
476static void dax_free_inode(struct inode *inode)
477{
478 struct dax_device *dax_dev = to_dax_dev(inode);
479 kfree(dax_dev->host);
480 dax_dev->host = NULL;
481 if (inode->i_rdev)
482 ida_simple_remove(&dax_minor_ida, MINOR(inode->i_rdev));
483 kmem_cache_free(dax_cache, dax_dev);
484}
485
486static void dax_destroy_inode(struct inode *inode)
487{
488 struct dax_device *dax_dev = to_dax_dev(inode);
489 WARN_ONCE(test_bit(DAXDEV_ALIVE, &dax_dev->flags),
490 "kill_dax() must be called before final iput()\n");
491}
492
493static const struct super_operations dax_sops = {
494 .statfs = simple_statfs,
495 .alloc_inode = dax_alloc_inode,
496 .destroy_inode = dax_destroy_inode,
497 .free_inode = dax_free_inode,
498 .drop_inode = generic_delete_inode,
499};
500
501static int dax_init_fs_context(struct fs_context *fc)
502{
503 struct pseudo_fs_context *ctx = init_pseudo(fc, DAXFS_MAGIC);
504 if (!ctx)
505 return -ENOMEM;
506 ctx->ops = &dax_sops;
507 return 0;
508}
509
510static struct file_system_type dax_fs_type = {
511 .name = "dax",
512 .init_fs_context = dax_init_fs_context,
513 .kill_sb = kill_anon_super,
514};
515
516static int dax_test(struct inode *inode, void *data)
517{
518 dev_t devt = *(dev_t *) data;
519
520 return inode->i_rdev == devt;
521}
522
523static int dax_set(struct inode *inode, void *data)
524{
525 dev_t devt = *(dev_t *) data;
526
527 inode->i_rdev = devt;
528 return 0;
529}
530
531static struct dax_device *dax_dev_get(dev_t devt)
532{
533 struct dax_device *dax_dev;
534 struct inode *inode;
535
536 inode = iget5_locked(dax_superblock, hash_32(devt + DAXFS_MAGIC, 31),
537 dax_test, dax_set, &devt);
538
539 if (!inode)
540 return NULL;
541
542 dax_dev = to_dax_dev(inode);
543 if (inode->i_state & I_NEW) {
544 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
545 inode->i_cdev = &dax_dev->cdev;
546 inode->i_mode = S_IFCHR;
547 inode->i_flags = S_DAX;
548 mapping_set_gfp_mask(&inode->i_data, GFP_USER);
549 unlock_new_inode(inode);
550 }
551
552 return dax_dev;
553}
554
555static void dax_add_host(struct dax_device *dax_dev, const char *host)
556{
557 int hash;
558
559 /*
560 * Unconditionally init dax_dev since it's coming from a
561 * non-zeroed slab cache
562 */
563 INIT_HLIST_NODE(&dax_dev->list);
564 dax_dev->host = host;
565 if (!host)
566 return;
567
568 hash = dax_host_hash(host);
569 spin_lock(&dax_host_lock);
570 hlist_add_head(&dax_dev->list, &dax_host_list[hash]);
571 spin_unlock(&dax_host_lock);
572}
573
574struct dax_device *alloc_dax(void *private, const char *__host,
575 const struct dax_operations *ops, unsigned long flags)
576{
577 struct dax_device *dax_dev;
578 const char *host;
579 dev_t devt;
580 int minor;
581
582 if (ops && !ops->zero_page_range) {
583 pr_debug("%s: error: device does not provide dax"
584 " operation zero_page_range()\n",
585 __host ? __host : "Unknown");
586 return ERR_PTR(-EINVAL);
587 }
588
589 host = kstrdup(__host, GFP_KERNEL);
590 if (__host && !host)
591 return ERR_PTR(-ENOMEM);
592
593 minor = ida_simple_get(&dax_minor_ida, 0, MINORMASK+1, GFP_KERNEL);
594 if (minor < 0)
595 goto err_minor;
596
597 devt = MKDEV(MAJOR(dax_devt), minor);
598 dax_dev = dax_dev_get(devt);
599 if (!dax_dev)
600 goto err_dev;
601
602 dax_add_host(dax_dev, host);
603 dax_dev->ops = ops;
604 dax_dev->private = private;
605 if (flags & DAXDEV_F_SYNC)
606 set_dax_synchronous(dax_dev);
607
608 return dax_dev;
609
610 err_dev:
611 ida_simple_remove(&dax_minor_ida, minor);
612 err_minor:
613 kfree(host);
614 return ERR_PTR(-ENOMEM);
615}
616EXPORT_SYMBOL_GPL(alloc_dax);
617
618void put_dax(struct dax_device *dax_dev)
619{
620 if (!dax_dev)
621 return;
622 iput(&dax_dev->inode);
623}
624EXPORT_SYMBOL_GPL(put_dax);
625
626/**
627 * dax_get_by_host() - temporary lookup mechanism for filesystem-dax
628 * @host: alternate name for the device registered by a dax driver
629 */
630struct dax_device *dax_get_by_host(const char *host)
631{
632 struct dax_device *dax_dev, *found = NULL;
633 int hash, id;
634
635 if (!host)
636 return NULL;
637
638 hash = dax_host_hash(host);
639
640 id = dax_read_lock();
641 spin_lock(&dax_host_lock);
642 hlist_for_each_entry(dax_dev, &dax_host_list[hash], list) {
643 if (!dax_alive(dax_dev)
644 || strcmp(host, dax_dev->host) != 0)
645 continue;
646
647 if (igrab(&dax_dev->inode))
648 found = dax_dev;
649 break;
650 }
651 spin_unlock(&dax_host_lock);
652 dax_read_unlock(id);
653
654 return found;
655}
656EXPORT_SYMBOL_GPL(dax_get_by_host);
657
658/**
659 * inode_dax: convert a public inode into its dax_dev
660 * @inode: An inode with i_cdev pointing to a dax_dev
661 *
662 * Note this is not equivalent to to_dax_dev() which is for private
663 * internal use where we know the inode filesystem type == dax_fs_type.
664 */
665struct dax_device *inode_dax(struct inode *inode)
666{
667 struct cdev *cdev = inode->i_cdev;
668
669 return container_of(cdev, struct dax_device, cdev);
670}
671EXPORT_SYMBOL_GPL(inode_dax);
672
673struct inode *dax_inode(struct dax_device *dax_dev)
674{
675 return &dax_dev->inode;
676}
677EXPORT_SYMBOL_GPL(dax_inode);
678
679void *dax_get_private(struct dax_device *dax_dev)
680{
681 if (!test_bit(DAXDEV_ALIVE, &dax_dev->flags))
682 return NULL;
683 return dax_dev->private;
684}
685EXPORT_SYMBOL_GPL(dax_get_private);
686
687static void init_once(void *_dax_dev)
688{
689 struct dax_device *dax_dev = _dax_dev;
690 struct inode *inode = &dax_dev->inode;
691
692 memset(dax_dev, 0, sizeof(*dax_dev));
693 inode_init_once(inode);
694}
695
696static int dax_fs_init(void)
697{
698 int rc;
699
700 dax_cache = kmem_cache_create("dax_cache", sizeof(struct dax_device), 0,
701 (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
702 SLAB_MEM_SPREAD|SLAB_ACCOUNT),
703 init_once);
704 if (!dax_cache)
705 return -ENOMEM;
706
707 dax_mnt = kern_mount(&dax_fs_type);
708 if (IS_ERR(dax_mnt)) {
709 rc = PTR_ERR(dax_mnt);
710 goto err_mount;
711 }
712 dax_superblock = dax_mnt->mnt_sb;
713
714 return 0;
715
716 err_mount:
717 kmem_cache_destroy(dax_cache);
718
719 return rc;
720}
721
722static void dax_fs_exit(void)
723{
724 kern_unmount(dax_mnt);
725 kmem_cache_destroy(dax_cache);
726}
727
728static int __init dax_core_init(void)
729{
730 int rc;
731
732 rc = dax_fs_init();
733 if (rc)
734 return rc;
735
736 rc = alloc_chrdev_region(&dax_devt, 0, MINORMASK+1, "dax");
737 if (rc)
738 goto err_chrdev;
739
740 rc = dax_bus_init();
741 if (rc)
742 goto err_bus;
743 return 0;
744
745err_bus:
746 unregister_chrdev_region(dax_devt, MINORMASK+1);
747err_chrdev:
748 dax_fs_exit();
749 return 0;
750}
751
752static void __exit dax_core_exit(void)
753{
754 unregister_chrdev_region(dax_devt, MINORMASK+1);
755 ida_destroy(&dax_minor_ida);
756 dax_fs_exit();
757}
758
759MODULE_AUTHOR("Intel Corporation");
760MODULE_LICENSE("GPL v2");
761subsys_initcall(dax_core_init);
762module_exit(dax_core_exit);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright(c) 2017 Intel Corporation. All rights reserved.
4 */
5#include <linux/pagemap.h>
6#include <linux/module.h>
7#include <linux/mount.h>
8#include <linux/pseudo_fs.h>
9#include <linux/magic.h>
10#include <linux/pfn_t.h>
11#include <linux/cdev.h>
12#include <linux/slab.h>
13#include <linux/uio.h>
14#include <linux/dax.h>
15#include <linux/fs.h>
16#include "dax-private.h"
17
18/**
19 * struct dax_device - anchor object for dax services
20 * @inode: core vfs
21 * @cdev: optional character interface for "device dax"
22 * @private: dax driver private data
23 * @flags: state and boolean properties
24 * @ops: operations for this device
25 * @holder_data: holder of a dax_device: could be filesystem or mapped device
26 * @holder_ops: operations for the inner holder
27 */
28struct dax_device {
29 struct inode inode;
30 struct cdev cdev;
31 void *private;
32 unsigned long flags;
33 const struct dax_operations *ops;
34 void *holder_data;
35 const struct dax_holder_operations *holder_ops;
36};
37
38static dev_t dax_devt;
39DEFINE_STATIC_SRCU(dax_srcu);
40static struct vfsmount *dax_mnt;
41static DEFINE_IDA(dax_minor_ida);
42static struct kmem_cache *dax_cache __read_mostly;
43static struct super_block *dax_superblock __read_mostly;
44
45int dax_read_lock(void)
46{
47 return srcu_read_lock(&dax_srcu);
48}
49EXPORT_SYMBOL_GPL(dax_read_lock);
50
51void dax_read_unlock(int id)
52{
53 srcu_read_unlock(&dax_srcu, id);
54}
55EXPORT_SYMBOL_GPL(dax_read_unlock);
56
57#if defined(CONFIG_BLOCK) && defined(CONFIG_FS_DAX)
58#include <linux/blkdev.h>
59
60static DEFINE_XARRAY(dax_hosts);
61
62int dax_add_host(struct dax_device *dax_dev, struct gendisk *disk)
63{
64 return xa_insert(&dax_hosts, (unsigned long)disk, dax_dev, GFP_KERNEL);
65}
66EXPORT_SYMBOL_GPL(dax_add_host);
67
68void dax_remove_host(struct gendisk *disk)
69{
70 xa_erase(&dax_hosts, (unsigned long)disk);
71}
72EXPORT_SYMBOL_GPL(dax_remove_host);
73
74/**
75 * fs_dax_get_by_bdev() - temporary lookup mechanism for filesystem-dax
76 * @bdev: block device to find a dax_device for
77 * @start_off: returns the byte offset into the dax_device that @bdev starts
78 * @holder: filesystem or mapped device inside the dax_device
79 * @ops: operations for the inner holder
80 */
81struct dax_device *fs_dax_get_by_bdev(struct block_device *bdev, u64 *start_off,
82 void *holder, const struct dax_holder_operations *ops)
83{
84 struct dax_device *dax_dev;
85 u64 part_size;
86 int id;
87
88 if (!blk_queue_dax(bdev->bd_disk->queue))
89 return NULL;
90
91 *start_off = get_start_sect(bdev) * SECTOR_SIZE;
92 part_size = bdev_nr_sectors(bdev) * SECTOR_SIZE;
93 if (*start_off % PAGE_SIZE || part_size % PAGE_SIZE) {
94 pr_info("%pg: error: unaligned partition for dax\n", bdev);
95 return NULL;
96 }
97
98 id = dax_read_lock();
99 dax_dev = xa_load(&dax_hosts, (unsigned long)bdev->bd_disk);
100 if (!dax_dev || !dax_alive(dax_dev) || !igrab(&dax_dev->inode))
101 dax_dev = NULL;
102 else if (holder) {
103 if (!cmpxchg(&dax_dev->holder_data, NULL, holder))
104 dax_dev->holder_ops = ops;
105 else
106 dax_dev = NULL;
107 }
108 dax_read_unlock(id);
109
110 return dax_dev;
111}
112EXPORT_SYMBOL_GPL(fs_dax_get_by_bdev);
113
114void fs_put_dax(struct dax_device *dax_dev, void *holder)
115{
116 if (dax_dev && holder &&
117 cmpxchg(&dax_dev->holder_data, holder, NULL) == holder)
118 dax_dev->holder_ops = NULL;
119 put_dax(dax_dev);
120}
121EXPORT_SYMBOL_GPL(fs_put_dax);
122#endif /* CONFIG_BLOCK && CONFIG_FS_DAX */
123
124enum dax_device_flags {
125 /* !alive + rcu grace period == no new operations / mappings */
126 DAXDEV_ALIVE,
127 /* gate whether dax_flush() calls the low level flush routine */
128 DAXDEV_WRITE_CACHE,
129 /* flag to check if device supports synchronous flush */
130 DAXDEV_SYNC,
131 /* do not leave the caches dirty after writes */
132 DAXDEV_NOCACHE,
133 /* handle CPU fetch exceptions during reads */
134 DAXDEV_NOMC,
135};
136
137/**
138 * dax_direct_access() - translate a device pgoff to an absolute pfn
139 * @dax_dev: a dax_device instance representing the logical memory range
140 * @pgoff: offset in pages from the start of the device to translate
141 * @nr_pages: number of consecutive pages caller can handle relative to @pfn
142 * @mode: indicator on normal access or recovery write
143 * @kaddr: output parameter that returns a virtual address mapping of pfn
144 * @pfn: output parameter that returns an absolute pfn translation of @pgoff
145 *
146 * Return: negative errno if an error occurs, otherwise the number of
147 * pages accessible at the device relative @pgoff.
148 */
149long dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff, long nr_pages,
150 enum dax_access_mode mode, void **kaddr, pfn_t *pfn)
151{
152 long avail;
153
154 if (!dax_dev)
155 return -EOPNOTSUPP;
156
157 if (!dax_alive(dax_dev))
158 return -ENXIO;
159
160 if (nr_pages < 0)
161 return -EINVAL;
162
163 avail = dax_dev->ops->direct_access(dax_dev, pgoff, nr_pages,
164 mode, kaddr, pfn);
165 if (!avail)
166 return -ERANGE;
167 return min(avail, nr_pages);
168}
169EXPORT_SYMBOL_GPL(dax_direct_access);
170
171size_t dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
172 size_t bytes, struct iov_iter *i)
173{
174 if (!dax_alive(dax_dev))
175 return 0;
176
177 /*
178 * The userspace address for the memory copy has already been validated
179 * via access_ok() in vfs_write, so use the 'no check' version to bypass
180 * the HARDENED_USERCOPY overhead.
181 */
182 if (test_bit(DAXDEV_NOCACHE, &dax_dev->flags))
183 return _copy_from_iter_flushcache(addr, bytes, i);
184 return _copy_from_iter(addr, bytes, i);
185}
186
187size_t dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
188 size_t bytes, struct iov_iter *i)
189{
190 if (!dax_alive(dax_dev))
191 return 0;
192
193 /*
194 * The userspace address for the memory copy has already been validated
195 * via access_ok() in vfs_red, so use the 'no check' version to bypass
196 * the HARDENED_USERCOPY overhead.
197 */
198 if (test_bit(DAXDEV_NOMC, &dax_dev->flags))
199 return _copy_mc_to_iter(addr, bytes, i);
200 return _copy_to_iter(addr, bytes, i);
201}
202
203int dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
204 size_t nr_pages)
205{
206 if (!dax_alive(dax_dev))
207 return -ENXIO;
208 /*
209 * There are no callers that want to zero more than one page as of now.
210 * Once users are there, this check can be removed after the
211 * device mapper code has been updated to split ranges across targets.
212 */
213 if (nr_pages != 1)
214 return -EIO;
215
216 return dax_dev->ops->zero_page_range(dax_dev, pgoff, nr_pages);
217}
218EXPORT_SYMBOL_GPL(dax_zero_page_range);
219
220size_t dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
221 void *addr, size_t bytes, struct iov_iter *iter)
222{
223 if (!dax_dev->ops->recovery_write)
224 return 0;
225 return dax_dev->ops->recovery_write(dax_dev, pgoff, addr, bytes, iter);
226}
227EXPORT_SYMBOL_GPL(dax_recovery_write);
228
229int dax_holder_notify_failure(struct dax_device *dax_dev, u64 off,
230 u64 len, int mf_flags)
231{
232 int rc, id;
233
234 id = dax_read_lock();
235 if (!dax_alive(dax_dev)) {
236 rc = -ENXIO;
237 goto out;
238 }
239
240 if (!dax_dev->holder_ops) {
241 rc = -EOPNOTSUPP;
242 goto out;
243 }
244
245 rc = dax_dev->holder_ops->notify_failure(dax_dev, off, len, mf_flags);
246out:
247 dax_read_unlock(id);
248 return rc;
249}
250EXPORT_SYMBOL_GPL(dax_holder_notify_failure);
251
252#ifdef CONFIG_ARCH_HAS_PMEM_API
253void arch_wb_cache_pmem(void *addr, size_t size);
254void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
255{
256 if (unlikely(!dax_write_cache_enabled(dax_dev)))
257 return;
258
259 arch_wb_cache_pmem(addr, size);
260}
261#else
262void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
263{
264}
265#endif
266EXPORT_SYMBOL_GPL(dax_flush);
267
268void dax_write_cache(struct dax_device *dax_dev, bool wc)
269{
270 if (wc)
271 set_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
272 else
273 clear_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
274}
275EXPORT_SYMBOL_GPL(dax_write_cache);
276
277bool dax_write_cache_enabled(struct dax_device *dax_dev)
278{
279 return test_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
280}
281EXPORT_SYMBOL_GPL(dax_write_cache_enabled);
282
283bool dax_synchronous(struct dax_device *dax_dev)
284{
285 return test_bit(DAXDEV_SYNC, &dax_dev->flags);
286}
287EXPORT_SYMBOL_GPL(dax_synchronous);
288
289void set_dax_synchronous(struct dax_device *dax_dev)
290{
291 set_bit(DAXDEV_SYNC, &dax_dev->flags);
292}
293EXPORT_SYMBOL_GPL(set_dax_synchronous);
294
295void set_dax_nocache(struct dax_device *dax_dev)
296{
297 set_bit(DAXDEV_NOCACHE, &dax_dev->flags);
298}
299EXPORT_SYMBOL_GPL(set_dax_nocache);
300
301void set_dax_nomc(struct dax_device *dax_dev)
302{
303 set_bit(DAXDEV_NOMC, &dax_dev->flags);
304}
305EXPORT_SYMBOL_GPL(set_dax_nomc);
306
307bool dax_alive(struct dax_device *dax_dev)
308{
309 lockdep_assert_held(&dax_srcu);
310 return test_bit(DAXDEV_ALIVE, &dax_dev->flags);
311}
312EXPORT_SYMBOL_GPL(dax_alive);
313
314/*
315 * Note, rcu is not protecting the liveness of dax_dev, rcu is ensuring
316 * that any fault handlers or operations that might have seen
317 * dax_alive(), have completed. Any operations that start after
318 * synchronize_srcu() has run will abort upon seeing !dax_alive().
319 */
320void kill_dax(struct dax_device *dax_dev)
321{
322 if (!dax_dev)
323 return;
324
325 if (dax_dev->holder_data != NULL)
326 dax_holder_notify_failure(dax_dev, 0, U64_MAX, 0);
327
328 clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
329 synchronize_srcu(&dax_srcu);
330
331 /* clear holder data */
332 dax_dev->holder_ops = NULL;
333 dax_dev->holder_data = NULL;
334}
335EXPORT_SYMBOL_GPL(kill_dax);
336
337void run_dax(struct dax_device *dax_dev)
338{
339 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
340}
341EXPORT_SYMBOL_GPL(run_dax);
342
343static struct inode *dax_alloc_inode(struct super_block *sb)
344{
345 struct dax_device *dax_dev;
346 struct inode *inode;
347
348 dax_dev = alloc_inode_sb(sb, dax_cache, GFP_KERNEL);
349 if (!dax_dev)
350 return NULL;
351
352 inode = &dax_dev->inode;
353 inode->i_rdev = 0;
354 return inode;
355}
356
357static struct dax_device *to_dax_dev(struct inode *inode)
358{
359 return container_of(inode, struct dax_device, inode);
360}
361
362static void dax_free_inode(struct inode *inode)
363{
364 struct dax_device *dax_dev = to_dax_dev(inode);
365 if (inode->i_rdev)
366 ida_free(&dax_minor_ida, iminor(inode));
367 kmem_cache_free(dax_cache, dax_dev);
368}
369
370static void dax_destroy_inode(struct inode *inode)
371{
372 struct dax_device *dax_dev = to_dax_dev(inode);
373 WARN_ONCE(test_bit(DAXDEV_ALIVE, &dax_dev->flags),
374 "kill_dax() must be called before final iput()\n");
375}
376
377static const struct super_operations dax_sops = {
378 .statfs = simple_statfs,
379 .alloc_inode = dax_alloc_inode,
380 .destroy_inode = dax_destroy_inode,
381 .free_inode = dax_free_inode,
382 .drop_inode = generic_delete_inode,
383};
384
385static int dax_init_fs_context(struct fs_context *fc)
386{
387 struct pseudo_fs_context *ctx = init_pseudo(fc, DAXFS_MAGIC);
388 if (!ctx)
389 return -ENOMEM;
390 ctx->ops = &dax_sops;
391 return 0;
392}
393
394static struct file_system_type dax_fs_type = {
395 .name = "dax",
396 .init_fs_context = dax_init_fs_context,
397 .kill_sb = kill_anon_super,
398};
399
400static int dax_test(struct inode *inode, void *data)
401{
402 dev_t devt = *(dev_t *) data;
403
404 return inode->i_rdev == devt;
405}
406
407static int dax_set(struct inode *inode, void *data)
408{
409 dev_t devt = *(dev_t *) data;
410
411 inode->i_rdev = devt;
412 return 0;
413}
414
415static struct dax_device *dax_dev_get(dev_t devt)
416{
417 struct dax_device *dax_dev;
418 struct inode *inode;
419
420 inode = iget5_locked(dax_superblock, hash_32(devt + DAXFS_MAGIC, 31),
421 dax_test, dax_set, &devt);
422
423 if (!inode)
424 return NULL;
425
426 dax_dev = to_dax_dev(inode);
427 if (inode->i_state & I_NEW) {
428 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
429 inode->i_cdev = &dax_dev->cdev;
430 inode->i_mode = S_IFCHR;
431 inode->i_flags = S_DAX;
432 mapping_set_gfp_mask(&inode->i_data, GFP_USER);
433 unlock_new_inode(inode);
434 }
435
436 return dax_dev;
437}
438
439struct dax_device *alloc_dax(void *private, const struct dax_operations *ops)
440{
441 struct dax_device *dax_dev;
442 dev_t devt;
443 int minor;
444
445 if (WARN_ON_ONCE(ops && !ops->zero_page_range))
446 return ERR_PTR(-EINVAL);
447
448 minor = ida_alloc_max(&dax_minor_ida, MINORMASK, GFP_KERNEL);
449 if (minor < 0)
450 return ERR_PTR(-ENOMEM);
451
452 devt = MKDEV(MAJOR(dax_devt), minor);
453 dax_dev = dax_dev_get(devt);
454 if (!dax_dev)
455 goto err_dev;
456
457 dax_dev->ops = ops;
458 dax_dev->private = private;
459 return dax_dev;
460
461 err_dev:
462 ida_free(&dax_minor_ida, minor);
463 return ERR_PTR(-ENOMEM);
464}
465EXPORT_SYMBOL_GPL(alloc_dax);
466
467void put_dax(struct dax_device *dax_dev)
468{
469 if (!dax_dev)
470 return;
471 iput(&dax_dev->inode);
472}
473EXPORT_SYMBOL_GPL(put_dax);
474
475/**
476 * dax_holder() - obtain the holder of a dax device
477 * @dax_dev: a dax_device instance
478 *
479 * Return: the holder's data which represents the holder if registered,
480 * otherwize NULL.
481 */
482void *dax_holder(struct dax_device *dax_dev)
483{
484 return dax_dev->holder_data;
485}
486EXPORT_SYMBOL_GPL(dax_holder);
487
488/**
489 * inode_dax: convert a public inode into its dax_dev
490 * @inode: An inode with i_cdev pointing to a dax_dev
491 *
492 * Note this is not equivalent to to_dax_dev() which is for private
493 * internal use where we know the inode filesystem type == dax_fs_type.
494 */
495struct dax_device *inode_dax(struct inode *inode)
496{
497 struct cdev *cdev = inode->i_cdev;
498
499 return container_of(cdev, struct dax_device, cdev);
500}
501EXPORT_SYMBOL_GPL(inode_dax);
502
503struct inode *dax_inode(struct dax_device *dax_dev)
504{
505 return &dax_dev->inode;
506}
507EXPORT_SYMBOL_GPL(dax_inode);
508
509void *dax_get_private(struct dax_device *dax_dev)
510{
511 if (!test_bit(DAXDEV_ALIVE, &dax_dev->flags))
512 return NULL;
513 return dax_dev->private;
514}
515EXPORT_SYMBOL_GPL(dax_get_private);
516
517static void init_once(void *_dax_dev)
518{
519 struct dax_device *dax_dev = _dax_dev;
520 struct inode *inode = &dax_dev->inode;
521
522 memset(dax_dev, 0, sizeof(*dax_dev));
523 inode_init_once(inode);
524}
525
526static int dax_fs_init(void)
527{
528 int rc;
529
530 dax_cache = kmem_cache_create("dax_cache", sizeof(struct dax_device), 0,
531 (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
532 SLAB_MEM_SPREAD|SLAB_ACCOUNT),
533 init_once);
534 if (!dax_cache)
535 return -ENOMEM;
536
537 dax_mnt = kern_mount(&dax_fs_type);
538 if (IS_ERR(dax_mnt)) {
539 rc = PTR_ERR(dax_mnt);
540 goto err_mount;
541 }
542 dax_superblock = dax_mnt->mnt_sb;
543
544 return 0;
545
546 err_mount:
547 kmem_cache_destroy(dax_cache);
548
549 return rc;
550}
551
552static void dax_fs_exit(void)
553{
554 kern_unmount(dax_mnt);
555 rcu_barrier();
556 kmem_cache_destroy(dax_cache);
557}
558
559static int __init dax_core_init(void)
560{
561 int rc;
562
563 rc = dax_fs_init();
564 if (rc)
565 return rc;
566
567 rc = alloc_chrdev_region(&dax_devt, 0, MINORMASK+1, "dax");
568 if (rc)
569 goto err_chrdev;
570
571 rc = dax_bus_init();
572 if (rc)
573 goto err_bus;
574 return 0;
575
576err_bus:
577 unregister_chrdev_region(dax_devt, MINORMASK+1);
578err_chrdev:
579 dax_fs_exit();
580 return 0;
581}
582
583static void __exit dax_core_exit(void)
584{
585 dax_bus_exit();
586 unregister_chrdev_region(dax_devt, MINORMASK+1);
587 ida_destroy(&dax_minor_ida);
588 dax_fs_exit();
589}
590
591MODULE_AUTHOR("Intel Corporation");
592MODULE_LICENSE("GPL v2");
593subsys_initcall(dax_core_init);
594module_exit(dax_core_exit);