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_queue))
63 return NULL;
64 return fs_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_debug("%s: error: unsupported blocksize for dax\n",
84 bdevname(bdev, buf));
85 return false;
86 }
87
88 err = bdev_dax_pgoff(bdev, start, PAGE_SIZE, &pgoff);
89 if (err) {
90 pr_debug("%s: error: unaligned partition for dax\n",
91 bdevname(bdev, buf));
92 return false;
93 }
94
95 last_page = PFN_DOWN((start + sectors - 1) * 512) * PAGE_SIZE / 512;
96 err = bdev_dax_pgoff(bdev, last_page, PAGE_SIZE, &pgoff_end);
97 if (err) {
98 pr_debug("%s: error: unaligned partition for dax\n",
99 bdevname(bdev, buf));
100 return false;
101 }
102
103 id = dax_read_lock();
104 len = dax_direct_access(dax_dev, pgoff, 1, &kaddr, &pfn);
105 len2 = dax_direct_access(dax_dev, pgoff_end, 1, &end_kaddr, &end_pfn);
106 dax_read_unlock(id);
107
108 if (len < 1 || len2 < 1) {
109 pr_debug("%s: error: dax access failed (%ld)\n",
110 bdevname(bdev, buf), len < 1 ? len : len2);
111 return false;
112 }
113
114 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED) && pfn_t_special(pfn)) {
115 /*
116 * An arch that has enabled the pmem api should also
117 * have its drivers support pfn_t_devmap()
118 *
119 * This is a developer warning and should not trigger in
120 * production. dax_flush() will crash since it depends
121 * on being able to do (page_address(pfn_to_page())).
122 */
123 WARN_ON(IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API));
124 dax_enabled = true;
125 } else if (pfn_t_devmap(pfn) && pfn_t_devmap(end_pfn)) {
126 struct dev_pagemap *pgmap, *end_pgmap;
127
128 pgmap = get_dev_pagemap(pfn_t_to_pfn(pfn), NULL);
129 end_pgmap = get_dev_pagemap(pfn_t_to_pfn(end_pfn), NULL);
130 if (pgmap && pgmap == end_pgmap && pgmap->type == MEMORY_DEVICE_FS_DAX
131 && pfn_t_to_page(pfn)->pgmap == pgmap
132 && pfn_t_to_page(end_pfn)->pgmap == pgmap
133 && pfn_t_to_pfn(pfn) == PHYS_PFN(__pa(kaddr))
134 && pfn_t_to_pfn(end_pfn) == PHYS_PFN(__pa(end_kaddr)))
135 dax_enabled = true;
136 put_dev_pagemap(pgmap);
137 put_dev_pagemap(end_pgmap);
138
139 }
140
141 if (!dax_enabled) {
142 pr_debug("%s: error: dax support not enabled\n",
143 bdevname(bdev, buf));
144 return false;
145 }
146 return true;
147}
148EXPORT_SYMBOL_GPL(__generic_fsdax_supported);
149
150/**
151 * __bdev_dax_supported() - Check if the device supports dax for filesystem
152 * @bdev: block device to check
153 * @blocksize: The block size of the device
154 *
155 * This is a library function for filesystems to check if the block device
156 * can be mounted with dax option.
157 *
158 * Return: true if supported, false if unsupported
159 */
160bool __bdev_dax_supported(struct block_device *bdev, int blocksize)
161{
162 struct dax_device *dax_dev;
163 struct request_queue *q;
164 char buf[BDEVNAME_SIZE];
165 bool ret;
166 int id;
167
168 q = bdev_get_queue(bdev);
169 if (!q || !blk_queue_dax(q)) {
170 pr_debug("%s: error: request queue doesn't support dax\n",
171 bdevname(bdev, buf));
172 return false;
173 }
174
175 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
176 if (!dax_dev) {
177 pr_debug("%s: error: device does not support dax\n",
178 bdevname(bdev, buf));
179 return false;
180 }
181
182 id = dax_read_lock();
183 ret = dax_supported(dax_dev, bdev, blocksize, 0,
184 i_size_read(bdev->bd_inode) / 512);
185 dax_read_unlock(id);
186
187 put_dax(dax_dev);
188
189 return ret;
190}
191EXPORT_SYMBOL_GPL(__bdev_dax_supported);
192#endif
193
194enum dax_device_flags {
195 /* !alive + rcu grace period == no new operations / mappings */
196 DAXDEV_ALIVE,
197 /* gate whether dax_flush() calls the low level flush routine */
198 DAXDEV_WRITE_CACHE,
199 /* flag to check if device supports synchronous flush */
200 DAXDEV_SYNC,
201};
202
203/**
204 * struct dax_device - anchor object for dax services
205 * @inode: core vfs
206 * @cdev: optional character interface for "device dax"
207 * @host: optional name for lookups where the device path is not available
208 * @private: dax driver private data
209 * @flags: state and boolean properties
210 */
211struct dax_device {
212 struct hlist_node list;
213 struct inode inode;
214 struct cdev cdev;
215 const char *host;
216 void *private;
217 unsigned long flags;
218 const struct dax_operations *ops;
219};
220
221static ssize_t write_cache_show(struct device *dev,
222 struct device_attribute *attr, char *buf)
223{
224 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
225 ssize_t rc;
226
227 WARN_ON_ONCE(!dax_dev);
228 if (!dax_dev)
229 return -ENXIO;
230
231 rc = sprintf(buf, "%d\n", !!dax_write_cache_enabled(dax_dev));
232 put_dax(dax_dev);
233 return rc;
234}
235
236static ssize_t write_cache_store(struct device *dev,
237 struct device_attribute *attr, const char *buf, size_t len)
238{
239 bool write_cache;
240 int rc = strtobool(buf, &write_cache);
241 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
242
243 WARN_ON_ONCE(!dax_dev);
244 if (!dax_dev)
245 return -ENXIO;
246
247 if (rc)
248 len = rc;
249 else
250 dax_write_cache(dax_dev, write_cache);
251
252 put_dax(dax_dev);
253 return len;
254}
255static DEVICE_ATTR_RW(write_cache);
256
257static umode_t dax_visible(struct kobject *kobj, struct attribute *a, int n)
258{
259 struct device *dev = container_of(kobj, typeof(*dev), kobj);
260 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
261
262 WARN_ON_ONCE(!dax_dev);
263 if (!dax_dev)
264 return 0;
265
266#ifndef CONFIG_ARCH_HAS_PMEM_API
267 if (a == &dev_attr_write_cache.attr)
268 return 0;
269#endif
270 return a->mode;
271}
272
273static struct attribute *dax_attributes[] = {
274 &dev_attr_write_cache.attr,
275 NULL,
276};
277
278struct attribute_group dax_attribute_group = {
279 .name = "dax",
280 .attrs = dax_attributes,
281 .is_visible = dax_visible,
282};
283EXPORT_SYMBOL_GPL(dax_attribute_group);
284
285/**
286 * dax_direct_access() - translate a device pgoff to an absolute pfn
287 * @dax_dev: a dax_device instance representing the logical memory range
288 * @pgoff: offset in pages from the start of the device to translate
289 * @nr_pages: number of consecutive pages caller can handle relative to @pfn
290 * @kaddr: output parameter that returns a virtual address mapping of pfn
291 * @pfn: output parameter that returns an absolute pfn translation of @pgoff
292 *
293 * Return: negative errno if an error occurs, otherwise the number of
294 * pages accessible at the device relative @pgoff.
295 */
296long dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff, long nr_pages,
297 void **kaddr, pfn_t *pfn)
298{
299 long avail;
300
301 if (!dax_dev)
302 return -EOPNOTSUPP;
303
304 if (!dax_alive(dax_dev))
305 return -ENXIO;
306
307 if (nr_pages < 0)
308 return nr_pages;
309
310 avail = dax_dev->ops->direct_access(dax_dev, pgoff, nr_pages,
311 kaddr, pfn);
312 if (!avail)
313 return -ERANGE;
314 return min(avail, nr_pages);
315}
316EXPORT_SYMBOL_GPL(dax_direct_access);
317
318bool dax_supported(struct dax_device *dax_dev, struct block_device *bdev,
319 int blocksize, sector_t start, sector_t len)
320{
321 if (!dax_alive(dax_dev))
322 return false;
323
324 return dax_dev->ops->dax_supported(dax_dev, bdev, blocksize, start, len);
325}
326
327size_t dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
328 size_t bytes, struct iov_iter *i)
329{
330 if (!dax_alive(dax_dev))
331 return 0;
332
333 return dax_dev->ops->copy_from_iter(dax_dev, pgoff, addr, bytes, i);
334}
335EXPORT_SYMBOL_GPL(dax_copy_from_iter);
336
337size_t dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
338 size_t bytes, struct iov_iter *i)
339{
340 if (!dax_alive(dax_dev))
341 return 0;
342
343 return dax_dev->ops->copy_to_iter(dax_dev, pgoff, addr, bytes, i);
344}
345EXPORT_SYMBOL_GPL(dax_copy_to_iter);
346
347#ifdef CONFIG_ARCH_HAS_PMEM_API
348void arch_wb_cache_pmem(void *addr, size_t size);
349void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
350{
351 if (unlikely(!dax_write_cache_enabled(dax_dev)))
352 return;
353
354 arch_wb_cache_pmem(addr, size);
355}
356#else
357void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
358{
359}
360#endif
361EXPORT_SYMBOL_GPL(dax_flush);
362
363void dax_write_cache(struct dax_device *dax_dev, bool wc)
364{
365 if (wc)
366 set_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
367 else
368 clear_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
369}
370EXPORT_SYMBOL_GPL(dax_write_cache);
371
372bool dax_write_cache_enabled(struct dax_device *dax_dev)
373{
374 return test_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
375}
376EXPORT_SYMBOL_GPL(dax_write_cache_enabled);
377
378bool __dax_synchronous(struct dax_device *dax_dev)
379{
380 return test_bit(DAXDEV_SYNC, &dax_dev->flags);
381}
382EXPORT_SYMBOL_GPL(__dax_synchronous);
383
384void __set_dax_synchronous(struct dax_device *dax_dev)
385{
386 set_bit(DAXDEV_SYNC, &dax_dev->flags);
387}
388EXPORT_SYMBOL_GPL(__set_dax_synchronous);
389
390bool dax_alive(struct dax_device *dax_dev)
391{
392 lockdep_assert_held(&dax_srcu);
393 return test_bit(DAXDEV_ALIVE, &dax_dev->flags);
394}
395EXPORT_SYMBOL_GPL(dax_alive);
396
397static int dax_host_hash(const char *host)
398{
399 return hashlen_hash(hashlen_string("DAX", host)) % DAX_HASH_SIZE;
400}
401
402/*
403 * Note, rcu is not protecting the liveness of dax_dev, rcu is ensuring
404 * that any fault handlers or operations that might have seen
405 * dax_alive(), have completed. Any operations that start after
406 * synchronize_srcu() has run will abort upon seeing !dax_alive().
407 */
408void kill_dax(struct dax_device *dax_dev)
409{
410 if (!dax_dev)
411 return;
412
413 clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
414
415 synchronize_srcu(&dax_srcu);
416
417 spin_lock(&dax_host_lock);
418 hlist_del_init(&dax_dev->list);
419 spin_unlock(&dax_host_lock);
420}
421EXPORT_SYMBOL_GPL(kill_dax);
422
423void run_dax(struct dax_device *dax_dev)
424{
425 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
426}
427EXPORT_SYMBOL_GPL(run_dax);
428
429static struct inode *dax_alloc_inode(struct super_block *sb)
430{
431 struct dax_device *dax_dev;
432 struct inode *inode;
433
434 dax_dev = kmem_cache_alloc(dax_cache, GFP_KERNEL);
435 if (!dax_dev)
436 return NULL;
437
438 inode = &dax_dev->inode;
439 inode->i_rdev = 0;
440 return inode;
441}
442
443static struct dax_device *to_dax_dev(struct inode *inode)
444{
445 return container_of(inode, struct dax_device, inode);
446}
447
448static void dax_free_inode(struct inode *inode)
449{
450 struct dax_device *dax_dev = to_dax_dev(inode);
451 kfree(dax_dev->host);
452 dax_dev->host = NULL;
453 if (inode->i_rdev)
454 ida_simple_remove(&dax_minor_ida, MINOR(inode->i_rdev));
455 kmem_cache_free(dax_cache, dax_dev);
456}
457
458static void dax_destroy_inode(struct inode *inode)
459{
460 struct dax_device *dax_dev = to_dax_dev(inode);
461 WARN_ONCE(test_bit(DAXDEV_ALIVE, &dax_dev->flags),
462 "kill_dax() must be called before final iput()\n");
463}
464
465static const struct super_operations dax_sops = {
466 .statfs = simple_statfs,
467 .alloc_inode = dax_alloc_inode,
468 .destroy_inode = dax_destroy_inode,
469 .free_inode = dax_free_inode,
470 .drop_inode = generic_delete_inode,
471};
472
473static int dax_init_fs_context(struct fs_context *fc)
474{
475 struct pseudo_fs_context *ctx = init_pseudo(fc, DAXFS_MAGIC);
476 if (!ctx)
477 return -ENOMEM;
478 ctx->ops = &dax_sops;
479 return 0;
480}
481
482static struct file_system_type dax_fs_type = {
483 .name = "dax",
484 .init_fs_context = dax_init_fs_context,
485 .kill_sb = kill_anon_super,
486};
487
488static int dax_test(struct inode *inode, void *data)
489{
490 dev_t devt = *(dev_t *) data;
491
492 return inode->i_rdev == devt;
493}
494
495static int dax_set(struct inode *inode, void *data)
496{
497 dev_t devt = *(dev_t *) data;
498
499 inode->i_rdev = devt;
500 return 0;
501}
502
503static struct dax_device *dax_dev_get(dev_t devt)
504{
505 struct dax_device *dax_dev;
506 struct inode *inode;
507
508 inode = iget5_locked(dax_superblock, hash_32(devt + DAXFS_MAGIC, 31),
509 dax_test, dax_set, &devt);
510
511 if (!inode)
512 return NULL;
513
514 dax_dev = to_dax_dev(inode);
515 if (inode->i_state & I_NEW) {
516 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
517 inode->i_cdev = &dax_dev->cdev;
518 inode->i_mode = S_IFCHR;
519 inode->i_flags = S_DAX;
520 mapping_set_gfp_mask(&inode->i_data, GFP_USER);
521 unlock_new_inode(inode);
522 }
523
524 return dax_dev;
525}
526
527static void dax_add_host(struct dax_device *dax_dev, const char *host)
528{
529 int hash;
530
531 /*
532 * Unconditionally init dax_dev since it's coming from a
533 * non-zeroed slab cache
534 */
535 INIT_HLIST_NODE(&dax_dev->list);
536 dax_dev->host = host;
537 if (!host)
538 return;
539
540 hash = dax_host_hash(host);
541 spin_lock(&dax_host_lock);
542 hlist_add_head(&dax_dev->list, &dax_host_list[hash]);
543 spin_unlock(&dax_host_lock);
544}
545
546struct dax_device *alloc_dax(void *private, const char *__host,
547 const struct dax_operations *ops, unsigned long flags)
548{
549 struct dax_device *dax_dev;
550 const char *host;
551 dev_t devt;
552 int minor;
553
554 host = kstrdup(__host, GFP_KERNEL);
555 if (__host && !host)
556 return NULL;
557
558 minor = ida_simple_get(&dax_minor_ida, 0, MINORMASK+1, GFP_KERNEL);
559 if (minor < 0)
560 goto err_minor;
561
562 devt = MKDEV(MAJOR(dax_devt), minor);
563 dax_dev = dax_dev_get(devt);
564 if (!dax_dev)
565 goto err_dev;
566
567 dax_add_host(dax_dev, host);
568 dax_dev->ops = ops;
569 dax_dev->private = private;
570 if (flags & DAXDEV_F_SYNC)
571 set_dax_synchronous(dax_dev);
572
573 return dax_dev;
574
575 err_dev:
576 ida_simple_remove(&dax_minor_ida, minor);
577 err_minor:
578 kfree(host);
579 return NULL;
580}
581EXPORT_SYMBOL_GPL(alloc_dax);
582
583void put_dax(struct dax_device *dax_dev)
584{
585 if (!dax_dev)
586 return;
587 iput(&dax_dev->inode);
588}
589EXPORT_SYMBOL_GPL(put_dax);
590
591/**
592 * dax_get_by_host() - temporary lookup mechanism for filesystem-dax
593 * @host: alternate name for the device registered by a dax driver
594 */
595struct dax_device *dax_get_by_host(const char *host)
596{
597 struct dax_device *dax_dev, *found = NULL;
598 int hash, id;
599
600 if (!host)
601 return NULL;
602
603 hash = dax_host_hash(host);
604
605 id = dax_read_lock();
606 spin_lock(&dax_host_lock);
607 hlist_for_each_entry(dax_dev, &dax_host_list[hash], list) {
608 if (!dax_alive(dax_dev)
609 || strcmp(host, dax_dev->host) != 0)
610 continue;
611
612 if (igrab(&dax_dev->inode))
613 found = dax_dev;
614 break;
615 }
616 spin_unlock(&dax_host_lock);
617 dax_read_unlock(id);
618
619 return found;
620}
621EXPORT_SYMBOL_GPL(dax_get_by_host);
622
623/**
624 * inode_dax: convert a public inode into its dax_dev
625 * @inode: An inode with i_cdev pointing to a dax_dev
626 *
627 * Note this is not equivalent to to_dax_dev() which is for private
628 * internal use where we know the inode filesystem type == dax_fs_type.
629 */
630struct dax_device *inode_dax(struct inode *inode)
631{
632 struct cdev *cdev = inode->i_cdev;
633
634 return container_of(cdev, struct dax_device, cdev);
635}
636EXPORT_SYMBOL_GPL(inode_dax);
637
638struct inode *dax_inode(struct dax_device *dax_dev)
639{
640 return &dax_dev->inode;
641}
642EXPORT_SYMBOL_GPL(dax_inode);
643
644void *dax_get_private(struct dax_device *dax_dev)
645{
646 if (!test_bit(DAXDEV_ALIVE, &dax_dev->flags))
647 return NULL;
648 return dax_dev->private;
649}
650EXPORT_SYMBOL_GPL(dax_get_private);
651
652static void init_once(void *_dax_dev)
653{
654 struct dax_device *dax_dev = _dax_dev;
655 struct inode *inode = &dax_dev->inode;
656
657 memset(dax_dev, 0, sizeof(*dax_dev));
658 inode_init_once(inode);
659}
660
661static int dax_fs_init(void)
662{
663 int rc;
664
665 dax_cache = kmem_cache_create("dax_cache", sizeof(struct dax_device), 0,
666 (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
667 SLAB_MEM_SPREAD|SLAB_ACCOUNT),
668 init_once);
669 if (!dax_cache)
670 return -ENOMEM;
671
672 dax_mnt = kern_mount(&dax_fs_type);
673 if (IS_ERR(dax_mnt)) {
674 rc = PTR_ERR(dax_mnt);
675 goto err_mount;
676 }
677 dax_superblock = dax_mnt->mnt_sb;
678
679 return 0;
680
681 err_mount:
682 kmem_cache_destroy(dax_cache);
683
684 return rc;
685}
686
687static void dax_fs_exit(void)
688{
689 kern_unmount(dax_mnt);
690 kmem_cache_destroy(dax_cache);
691}
692
693static int __init dax_core_init(void)
694{
695 int rc;
696
697 rc = dax_fs_init();
698 if (rc)
699 return rc;
700
701 rc = alloc_chrdev_region(&dax_devt, 0, MINORMASK+1, "dax");
702 if (rc)
703 goto err_chrdev;
704
705 rc = dax_bus_init();
706 if (rc)
707 goto err_bus;
708 return 0;
709
710err_bus:
711 unregister_chrdev_region(dax_devt, MINORMASK+1);
712err_chrdev:
713 dax_fs_exit();
714 return 0;
715}
716
717static void __exit dax_core_exit(void)
718{
719 unregister_chrdev_region(dax_devt, MINORMASK+1);
720 ida_destroy(&dax_minor_ida);
721 dax_fs_exit();
722}
723
724MODULE_AUTHOR("Intel Corporation");
725MODULE_LICENSE("GPL v2");
726subsys_initcall(dax_core_init);
727module_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 <linux/cacheinfo.h>
17#include "dax-private.h"
18
19/**
20 * struct dax_device - anchor object for dax services
21 * @inode: core vfs
22 * @cdev: optional character interface for "device dax"
23 * @private: dax driver private data
24 * @flags: state and boolean properties
25 * @ops: operations for this device
26 * @holder_data: holder of a dax_device: could be filesystem or mapped device
27 * @holder_ops: operations for the inner holder
28 */
29struct dax_device {
30 struct inode inode;
31 struct cdev cdev;
32 void *private;
33 unsigned long flags;
34 const struct dax_operations *ops;
35 void *holder_data;
36 const struct dax_holder_operations *holder_ops;
37};
38
39static dev_t dax_devt;
40DEFINE_STATIC_SRCU(dax_srcu);
41static struct vfsmount *dax_mnt;
42static DEFINE_IDA(dax_minor_ida);
43static struct kmem_cache *dax_cache __read_mostly;
44static struct super_block *dax_superblock __read_mostly;
45
46int dax_read_lock(void)
47{
48 return srcu_read_lock(&dax_srcu);
49}
50EXPORT_SYMBOL_GPL(dax_read_lock);
51
52void dax_read_unlock(int id)
53{
54 srcu_read_unlock(&dax_srcu, id);
55}
56EXPORT_SYMBOL_GPL(dax_read_unlock);
57
58#if defined(CONFIG_BLOCK) && defined(CONFIG_FS_DAX)
59#include <linux/blkdev.h>
60
61static DEFINE_XARRAY(dax_hosts);
62
63int dax_add_host(struct dax_device *dax_dev, struct gendisk *disk)
64{
65 return xa_insert(&dax_hosts, (unsigned long)disk, dax_dev, GFP_KERNEL);
66}
67EXPORT_SYMBOL_GPL(dax_add_host);
68
69void dax_remove_host(struct gendisk *disk)
70{
71 xa_erase(&dax_hosts, (unsigned long)disk);
72}
73EXPORT_SYMBOL_GPL(dax_remove_host);
74
75/**
76 * fs_dax_get_by_bdev() - temporary lookup mechanism for filesystem-dax
77 * @bdev: block device to find a dax_device for
78 * @start_off: returns the byte offset into the dax_device that @bdev starts
79 * @holder: filesystem or mapped device inside the dax_device
80 * @ops: operations for the inner holder
81 */
82struct dax_device *fs_dax_get_by_bdev(struct block_device *bdev, u64 *start_off,
83 void *holder, const struct dax_holder_operations *ops)
84{
85 struct dax_device *dax_dev;
86 u64 part_size;
87 int id;
88
89 if (!blk_queue_dax(bdev->bd_disk->queue))
90 return NULL;
91
92 *start_off = get_start_sect(bdev) * SECTOR_SIZE;
93 part_size = bdev_nr_sectors(bdev) * SECTOR_SIZE;
94 if (*start_off % PAGE_SIZE || part_size % PAGE_SIZE) {
95 pr_info("%pg: error: unaligned partition for dax\n", bdev);
96 return NULL;
97 }
98
99 id = dax_read_lock();
100 dax_dev = xa_load(&dax_hosts, (unsigned long)bdev->bd_disk);
101 if (!dax_dev || !dax_alive(dax_dev) || !igrab(&dax_dev->inode))
102 dax_dev = NULL;
103 else if (holder) {
104 if (!cmpxchg(&dax_dev->holder_data, NULL, holder))
105 dax_dev->holder_ops = ops;
106 else
107 dax_dev = NULL;
108 }
109 dax_read_unlock(id);
110
111 return dax_dev;
112}
113EXPORT_SYMBOL_GPL(fs_dax_get_by_bdev);
114
115void fs_put_dax(struct dax_device *dax_dev, void *holder)
116{
117 if (dax_dev && holder &&
118 cmpxchg(&dax_dev->holder_data, holder, NULL) == holder)
119 dax_dev->holder_ops = NULL;
120 put_dax(dax_dev);
121}
122EXPORT_SYMBOL_GPL(fs_put_dax);
123#endif /* CONFIG_BLOCK && CONFIG_FS_DAX */
124
125enum dax_device_flags {
126 /* !alive + rcu grace period == no new operations / mappings */
127 DAXDEV_ALIVE,
128 /* gate whether dax_flush() calls the low level flush routine */
129 DAXDEV_WRITE_CACHE,
130 /* flag to check if device supports synchronous flush */
131 DAXDEV_SYNC,
132 /* do not leave the caches dirty after writes */
133 DAXDEV_NOCACHE,
134 /* handle CPU fetch exceptions during reads */
135 DAXDEV_NOMC,
136};
137
138/**
139 * dax_direct_access() - translate a device pgoff to an absolute pfn
140 * @dax_dev: a dax_device instance representing the logical memory range
141 * @pgoff: offset in pages from the start of the device to translate
142 * @nr_pages: number of consecutive pages caller can handle relative to @pfn
143 * @mode: indicator on normal access or recovery write
144 * @kaddr: output parameter that returns a virtual address mapping of pfn
145 * @pfn: output parameter that returns an absolute pfn translation of @pgoff
146 *
147 * Return: negative errno if an error occurs, otherwise the number of
148 * pages accessible at the device relative @pgoff.
149 */
150long dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff, long nr_pages,
151 enum dax_access_mode mode, void **kaddr, pfn_t *pfn)
152{
153 long avail;
154
155 if (!dax_dev)
156 return -EOPNOTSUPP;
157
158 if (!dax_alive(dax_dev))
159 return -ENXIO;
160
161 if (nr_pages < 0)
162 return -EINVAL;
163
164 avail = dax_dev->ops->direct_access(dax_dev, pgoff, nr_pages,
165 mode, kaddr, pfn);
166 if (!avail)
167 return -ERANGE;
168 return min(avail, nr_pages);
169}
170EXPORT_SYMBOL_GPL(dax_direct_access);
171
172size_t dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
173 size_t bytes, struct iov_iter *i)
174{
175 if (!dax_alive(dax_dev))
176 return 0;
177
178 /*
179 * The userspace address for the memory copy has already been validated
180 * via access_ok() in vfs_write, so use the 'no check' version to bypass
181 * the HARDENED_USERCOPY overhead.
182 */
183 if (test_bit(DAXDEV_NOCACHE, &dax_dev->flags))
184 return _copy_from_iter_flushcache(addr, bytes, i);
185 return _copy_from_iter(addr, bytes, i);
186}
187
188size_t dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
189 size_t bytes, struct iov_iter *i)
190{
191 if (!dax_alive(dax_dev))
192 return 0;
193
194 /*
195 * The userspace address for the memory copy has already been validated
196 * via access_ok() in vfs_red, so use the 'no check' version to bypass
197 * the HARDENED_USERCOPY overhead.
198 */
199 if (test_bit(DAXDEV_NOMC, &dax_dev->flags))
200 return _copy_mc_to_iter(addr, bytes, i);
201 return _copy_to_iter(addr, bytes, i);
202}
203
204int dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
205 size_t nr_pages)
206{
207 int ret;
208
209 if (!dax_alive(dax_dev))
210 return -ENXIO;
211 /*
212 * There are no callers that want to zero more than one page as of now.
213 * Once users are there, this check can be removed after the
214 * device mapper code has been updated to split ranges across targets.
215 */
216 if (nr_pages != 1)
217 return -EIO;
218
219 ret = dax_dev->ops->zero_page_range(dax_dev, pgoff, nr_pages);
220 return dax_mem2blk_err(ret);
221}
222EXPORT_SYMBOL_GPL(dax_zero_page_range);
223
224size_t dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
225 void *addr, size_t bytes, struct iov_iter *iter)
226{
227 if (!dax_dev->ops->recovery_write)
228 return 0;
229 return dax_dev->ops->recovery_write(dax_dev, pgoff, addr, bytes, iter);
230}
231EXPORT_SYMBOL_GPL(dax_recovery_write);
232
233int dax_holder_notify_failure(struct dax_device *dax_dev, u64 off,
234 u64 len, int mf_flags)
235{
236 int rc, id;
237
238 id = dax_read_lock();
239 if (!dax_alive(dax_dev)) {
240 rc = -ENXIO;
241 goto out;
242 }
243
244 if (!dax_dev->holder_ops) {
245 rc = -EOPNOTSUPP;
246 goto out;
247 }
248
249 rc = dax_dev->holder_ops->notify_failure(dax_dev, off, len, mf_flags);
250out:
251 dax_read_unlock(id);
252 return rc;
253}
254EXPORT_SYMBOL_GPL(dax_holder_notify_failure);
255
256#ifdef CONFIG_ARCH_HAS_PMEM_API
257void arch_wb_cache_pmem(void *addr, size_t size);
258void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
259{
260 if (unlikely(!dax_write_cache_enabled(dax_dev)))
261 return;
262
263 arch_wb_cache_pmem(addr, size);
264}
265#else
266void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
267{
268}
269#endif
270EXPORT_SYMBOL_GPL(dax_flush);
271
272void dax_write_cache(struct dax_device *dax_dev, bool wc)
273{
274 if (wc)
275 set_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
276 else
277 clear_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
278}
279EXPORT_SYMBOL_GPL(dax_write_cache);
280
281bool dax_write_cache_enabled(struct dax_device *dax_dev)
282{
283 return test_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
284}
285EXPORT_SYMBOL_GPL(dax_write_cache_enabled);
286
287bool dax_synchronous(struct dax_device *dax_dev)
288{
289 return test_bit(DAXDEV_SYNC, &dax_dev->flags);
290}
291EXPORT_SYMBOL_GPL(dax_synchronous);
292
293void set_dax_synchronous(struct dax_device *dax_dev)
294{
295 set_bit(DAXDEV_SYNC, &dax_dev->flags);
296}
297EXPORT_SYMBOL_GPL(set_dax_synchronous);
298
299void set_dax_nocache(struct dax_device *dax_dev)
300{
301 set_bit(DAXDEV_NOCACHE, &dax_dev->flags);
302}
303EXPORT_SYMBOL_GPL(set_dax_nocache);
304
305void set_dax_nomc(struct dax_device *dax_dev)
306{
307 set_bit(DAXDEV_NOMC, &dax_dev->flags);
308}
309EXPORT_SYMBOL_GPL(set_dax_nomc);
310
311bool dax_alive(struct dax_device *dax_dev)
312{
313 lockdep_assert_held(&dax_srcu);
314 return test_bit(DAXDEV_ALIVE, &dax_dev->flags);
315}
316EXPORT_SYMBOL_GPL(dax_alive);
317
318/*
319 * Note, rcu is not protecting the liveness of dax_dev, rcu is ensuring
320 * that any fault handlers or operations that might have seen
321 * dax_alive(), have completed. Any operations that start after
322 * synchronize_srcu() has run will abort upon seeing !dax_alive().
323 *
324 * Note, because alloc_dax() returns an ERR_PTR() on error, callers
325 * typically store its result into a local variable in order to check
326 * the result. Therefore, care must be taken to populate the struct
327 * device dax_dev field make sure the dax_dev is not leaked.
328 */
329void kill_dax(struct dax_device *dax_dev)
330{
331 if (!dax_dev)
332 return;
333
334 if (dax_dev->holder_data != NULL)
335 dax_holder_notify_failure(dax_dev, 0, U64_MAX,
336 MF_MEM_PRE_REMOVE);
337
338 clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
339 synchronize_srcu(&dax_srcu);
340
341 /* clear holder data */
342 dax_dev->holder_ops = NULL;
343 dax_dev->holder_data = NULL;
344}
345EXPORT_SYMBOL_GPL(kill_dax);
346
347void run_dax(struct dax_device *dax_dev)
348{
349 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
350}
351EXPORT_SYMBOL_GPL(run_dax);
352
353static struct inode *dax_alloc_inode(struct super_block *sb)
354{
355 struct dax_device *dax_dev;
356 struct inode *inode;
357
358 dax_dev = alloc_inode_sb(sb, dax_cache, GFP_KERNEL);
359 if (!dax_dev)
360 return NULL;
361
362 inode = &dax_dev->inode;
363 inode->i_rdev = 0;
364 return inode;
365}
366
367static struct dax_device *to_dax_dev(struct inode *inode)
368{
369 return container_of(inode, struct dax_device, inode);
370}
371
372static void dax_free_inode(struct inode *inode)
373{
374 struct dax_device *dax_dev = to_dax_dev(inode);
375 if (inode->i_rdev)
376 ida_free(&dax_minor_ida, iminor(inode));
377 kmem_cache_free(dax_cache, dax_dev);
378}
379
380static void dax_destroy_inode(struct inode *inode)
381{
382 struct dax_device *dax_dev = to_dax_dev(inode);
383 WARN_ONCE(test_bit(DAXDEV_ALIVE, &dax_dev->flags),
384 "kill_dax() must be called before final iput()\n");
385}
386
387static const struct super_operations dax_sops = {
388 .statfs = simple_statfs,
389 .alloc_inode = dax_alloc_inode,
390 .destroy_inode = dax_destroy_inode,
391 .free_inode = dax_free_inode,
392 .drop_inode = generic_delete_inode,
393};
394
395static int dax_init_fs_context(struct fs_context *fc)
396{
397 struct pseudo_fs_context *ctx = init_pseudo(fc, DAXFS_MAGIC);
398 if (!ctx)
399 return -ENOMEM;
400 ctx->ops = &dax_sops;
401 return 0;
402}
403
404static struct file_system_type dax_fs_type = {
405 .name = "dax",
406 .init_fs_context = dax_init_fs_context,
407 .kill_sb = kill_anon_super,
408};
409
410static int dax_test(struct inode *inode, void *data)
411{
412 dev_t devt = *(dev_t *) data;
413
414 return inode->i_rdev == devt;
415}
416
417static int dax_set(struct inode *inode, void *data)
418{
419 dev_t devt = *(dev_t *) data;
420
421 inode->i_rdev = devt;
422 return 0;
423}
424
425static struct dax_device *dax_dev_get(dev_t devt)
426{
427 struct dax_device *dax_dev;
428 struct inode *inode;
429
430 inode = iget5_locked(dax_superblock, hash_32(devt + DAXFS_MAGIC, 31),
431 dax_test, dax_set, &devt);
432
433 if (!inode)
434 return NULL;
435
436 dax_dev = to_dax_dev(inode);
437 if (inode->i_state & I_NEW) {
438 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
439 inode->i_cdev = &dax_dev->cdev;
440 inode->i_mode = S_IFCHR;
441 inode->i_flags = S_DAX;
442 mapping_set_gfp_mask(&inode->i_data, GFP_USER);
443 unlock_new_inode(inode);
444 }
445
446 return dax_dev;
447}
448
449struct dax_device *alloc_dax(void *private, const struct dax_operations *ops)
450{
451 struct dax_device *dax_dev;
452 dev_t devt;
453 int minor;
454
455 /*
456 * Unavailable on architectures with virtually aliased data caches,
457 * except for device-dax (NULL operations pointer), which does
458 * not use aliased mappings from the kernel.
459 */
460 if (ops && cpu_dcache_is_aliasing())
461 return ERR_PTR(-EOPNOTSUPP);
462
463 if (WARN_ON_ONCE(ops && !ops->zero_page_range))
464 return ERR_PTR(-EINVAL);
465
466 minor = ida_alloc_max(&dax_minor_ida, MINORMASK, GFP_KERNEL);
467 if (minor < 0)
468 return ERR_PTR(-ENOMEM);
469
470 devt = MKDEV(MAJOR(dax_devt), minor);
471 dax_dev = dax_dev_get(devt);
472 if (!dax_dev)
473 goto err_dev;
474
475 dax_dev->ops = ops;
476 dax_dev->private = private;
477 return dax_dev;
478
479 err_dev:
480 ida_free(&dax_minor_ida, minor);
481 return ERR_PTR(-ENOMEM);
482}
483EXPORT_SYMBOL_GPL(alloc_dax);
484
485void put_dax(struct dax_device *dax_dev)
486{
487 if (!dax_dev)
488 return;
489 iput(&dax_dev->inode);
490}
491EXPORT_SYMBOL_GPL(put_dax);
492
493/**
494 * dax_holder() - obtain the holder of a dax device
495 * @dax_dev: a dax_device instance
496 *
497 * Return: the holder's data which represents the holder if registered,
498 * otherwize NULL.
499 */
500void *dax_holder(struct dax_device *dax_dev)
501{
502 return dax_dev->holder_data;
503}
504EXPORT_SYMBOL_GPL(dax_holder);
505
506/**
507 * inode_dax: convert a public inode into its dax_dev
508 * @inode: An inode with i_cdev pointing to a dax_dev
509 *
510 * Note this is not equivalent to to_dax_dev() which is for private
511 * internal use where we know the inode filesystem type == dax_fs_type.
512 */
513struct dax_device *inode_dax(struct inode *inode)
514{
515 struct cdev *cdev = inode->i_cdev;
516
517 return container_of(cdev, struct dax_device, cdev);
518}
519EXPORT_SYMBOL_GPL(inode_dax);
520
521struct inode *dax_inode(struct dax_device *dax_dev)
522{
523 return &dax_dev->inode;
524}
525EXPORT_SYMBOL_GPL(dax_inode);
526
527void *dax_get_private(struct dax_device *dax_dev)
528{
529 if (!test_bit(DAXDEV_ALIVE, &dax_dev->flags))
530 return NULL;
531 return dax_dev->private;
532}
533EXPORT_SYMBOL_GPL(dax_get_private);
534
535static void init_once(void *_dax_dev)
536{
537 struct dax_device *dax_dev = _dax_dev;
538 struct inode *inode = &dax_dev->inode;
539
540 memset(dax_dev, 0, sizeof(*dax_dev));
541 inode_init_once(inode);
542}
543
544static int dax_fs_init(void)
545{
546 int rc;
547
548 dax_cache = kmem_cache_create("dax_cache", sizeof(struct dax_device), 0,
549 SLAB_HWCACHE_ALIGN | SLAB_RECLAIM_ACCOUNT | SLAB_ACCOUNT,
550 init_once);
551 if (!dax_cache)
552 return -ENOMEM;
553
554 dax_mnt = kern_mount(&dax_fs_type);
555 if (IS_ERR(dax_mnt)) {
556 rc = PTR_ERR(dax_mnt);
557 goto err_mount;
558 }
559 dax_superblock = dax_mnt->mnt_sb;
560
561 return 0;
562
563 err_mount:
564 kmem_cache_destroy(dax_cache);
565
566 return rc;
567}
568
569static void dax_fs_exit(void)
570{
571 kern_unmount(dax_mnt);
572 rcu_barrier();
573 kmem_cache_destroy(dax_cache);
574}
575
576static int __init dax_core_init(void)
577{
578 int rc;
579
580 rc = dax_fs_init();
581 if (rc)
582 return rc;
583
584 rc = alloc_chrdev_region(&dax_devt, 0, MINORMASK+1, "dax");
585 if (rc)
586 goto err_chrdev;
587
588 rc = dax_bus_init();
589 if (rc)
590 goto err_bus;
591 return 0;
592
593err_bus:
594 unregister_chrdev_region(dax_devt, MINORMASK+1);
595err_chrdev:
596 dax_fs_exit();
597 return 0;
598}
599
600static void __exit dax_core_exit(void)
601{
602 dax_bus_exit();
603 unregister_chrdev_region(dax_devt, MINORMASK+1);
604 ida_destroy(&dax_minor_ida);
605 dax_fs_exit();
606}
607
608MODULE_AUTHOR("Intel Corporation");
609MODULE_DESCRIPTION("DAX: direct access to differentiated memory");
610MODULE_LICENSE("GPL v2");
611subsys_initcall(dax_core_init);
612module_exit(dax_core_exit);