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/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);
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 sector_t start_sect = bdev ? get_start_sect(bdev) : 0;
50 phys_addr_t phys_off = (start_sect + sector) * 512;
51
52 if (pgoff)
53 *pgoff = PHYS_PFN(phys_off);
54 if (phys_off % PAGE_SIZE || size % PAGE_SIZE)
55 return -EINVAL;
56 return 0;
57}
58EXPORT_SYMBOL(bdev_dax_pgoff);
59
60#if IS_ENABLED(CONFIG_FS_DAX)
61struct dax_device *fs_dax_get_by_bdev(struct block_device *bdev)
62{
63 if (!blk_queue_dax(bdev->bd_disk->queue))
64 return NULL;
65 return dax_get_by_host(bdev->bd_disk->disk_name);
66}
67EXPORT_SYMBOL_GPL(fs_dax_get_by_bdev);
68#endif
69
70bool __generic_fsdax_supported(struct dax_device *dax_dev,
71 struct block_device *bdev, int blocksize, sector_t start,
72 sector_t sectors)
73{
74 bool dax_enabled = false;
75 pgoff_t pgoff, pgoff_end;
76 char buf[BDEVNAME_SIZE];
77 void *kaddr, *end_kaddr;
78 pfn_t pfn, end_pfn;
79 sector_t last_page;
80 long len, len2;
81 int err, id;
82
83 if (blocksize != PAGE_SIZE) {
84 pr_info("%s: error: unsupported blocksize for dax\n",
85 bdevname(bdev, buf));
86 return false;
87 }
88
89 if (!dax_dev) {
90 pr_debug("%s: error: dax unsupported by block device\n",
91 bdevname(bdev, buf));
92 return false;
93 }
94
95 err = bdev_dax_pgoff(bdev, start, PAGE_SIZE, &pgoff);
96 if (err) {
97 pr_info("%s: error: unaligned partition for dax\n",
98 bdevname(bdev, buf));
99 return false;
100 }
101
102 last_page = PFN_DOWN((start + sectors - 1) * 512) * PAGE_SIZE / 512;
103 err = bdev_dax_pgoff(bdev, last_page, PAGE_SIZE, &pgoff_end);
104 if (err) {
105 pr_info("%s: error: unaligned partition for dax\n",
106 bdevname(bdev, buf));
107 return false;
108 }
109
110 id = dax_read_lock();
111 len = dax_direct_access(dax_dev, pgoff, 1, &kaddr, &pfn);
112 len2 = dax_direct_access(dax_dev, pgoff_end, 1, &end_kaddr, &end_pfn);
113
114 if (len < 1 || len2 < 1) {
115 pr_info("%s: error: dax access failed (%ld)\n",
116 bdevname(bdev, buf), len < 1 ? len : len2);
117 dax_read_unlock(id);
118 return false;
119 }
120
121 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED) && pfn_t_special(pfn)) {
122 /*
123 * An arch that has enabled the pmem api should also
124 * have its drivers support pfn_t_devmap()
125 *
126 * This is a developer warning and should not trigger in
127 * production. dax_flush() will crash since it depends
128 * on being able to do (page_address(pfn_to_page())).
129 */
130 WARN_ON(IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API));
131 dax_enabled = true;
132 } else if (pfn_t_devmap(pfn) && pfn_t_devmap(end_pfn)) {
133 struct dev_pagemap *pgmap, *end_pgmap;
134
135 pgmap = get_dev_pagemap(pfn_t_to_pfn(pfn), NULL);
136 end_pgmap = get_dev_pagemap(pfn_t_to_pfn(end_pfn), NULL);
137 if (pgmap && pgmap == end_pgmap && pgmap->type == MEMORY_DEVICE_FS_DAX
138 && pfn_t_to_page(pfn)->pgmap == pgmap
139 && pfn_t_to_page(end_pfn)->pgmap == pgmap
140 && pfn_t_to_pfn(pfn) == PHYS_PFN(__pa(kaddr))
141 && pfn_t_to_pfn(end_pfn) == PHYS_PFN(__pa(end_kaddr)))
142 dax_enabled = true;
143 put_dev_pagemap(pgmap);
144 put_dev_pagemap(end_pgmap);
145
146 }
147 dax_read_unlock(id);
148
149 if (!dax_enabled) {
150 pr_info("%s: error: dax support not enabled\n",
151 bdevname(bdev, buf));
152 return false;
153 }
154 return true;
155}
156EXPORT_SYMBOL_GPL(__generic_fsdax_supported);
157
158/**
159 * __bdev_dax_supported() - Check if the device supports dax for filesystem
160 * @bdev: block device to check
161 * @blocksize: The block size of the device
162 *
163 * This is a library function for filesystems to check if the block device
164 * can be mounted with dax option.
165 *
166 * Return: true if supported, false if unsupported
167 */
168bool __bdev_dax_supported(struct block_device *bdev, int blocksize)
169{
170 struct dax_device *dax_dev;
171 struct request_queue *q;
172 char buf[BDEVNAME_SIZE];
173 bool ret;
174 int id;
175
176 q = bdev_get_queue(bdev);
177 if (!q || !blk_queue_dax(q)) {
178 pr_debug("%s: error: request queue doesn't support dax\n",
179 bdevname(bdev, buf));
180 return false;
181 }
182
183 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
184 if (!dax_dev) {
185 pr_debug("%s: error: device does not support dax\n",
186 bdevname(bdev, buf));
187 return false;
188 }
189
190 id = dax_read_lock();
191 ret = dax_supported(dax_dev, bdev, blocksize, 0,
192 i_size_read(bdev->bd_inode) / 512);
193 dax_read_unlock(id);
194
195 put_dax(dax_dev);
196
197 return ret;
198}
199EXPORT_SYMBOL_GPL(__bdev_dax_supported);
200#endif
201
202enum dax_device_flags {
203 /* !alive + rcu grace period == no new operations / mappings */
204 DAXDEV_ALIVE,
205 /* gate whether dax_flush() calls the low level flush routine */
206 DAXDEV_WRITE_CACHE,
207 /* flag to check if device supports synchronous flush */
208 DAXDEV_SYNC,
209};
210
211/**
212 * struct dax_device - anchor object for dax services
213 * @inode: core vfs
214 * @cdev: optional character interface for "device dax"
215 * @host: optional name for lookups where the device path is not available
216 * @private: dax driver private data
217 * @flags: state and boolean properties
218 */
219struct dax_device {
220 struct hlist_node list;
221 struct inode inode;
222 struct cdev cdev;
223 const char *host;
224 void *private;
225 unsigned long flags;
226 const struct dax_operations *ops;
227};
228
229static ssize_t write_cache_show(struct device *dev,
230 struct device_attribute *attr, char *buf)
231{
232 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
233 ssize_t rc;
234
235 WARN_ON_ONCE(!dax_dev);
236 if (!dax_dev)
237 return -ENXIO;
238
239 rc = sprintf(buf, "%d\n", !!dax_write_cache_enabled(dax_dev));
240 put_dax(dax_dev);
241 return rc;
242}
243
244static ssize_t write_cache_store(struct device *dev,
245 struct device_attribute *attr, const char *buf, size_t len)
246{
247 bool write_cache;
248 int rc = strtobool(buf, &write_cache);
249 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
250
251 WARN_ON_ONCE(!dax_dev);
252 if (!dax_dev)
253 return -ENXIO;
254
255 if (rc)
256 len = rc;
257 else
258 dax_write_cache(dax_dev, write_cache);
259
260 put_dax(dax_dev);
261 return len;
262}
263static DEVICE_ATTR_RW(write_cache);
264
265static umode_t dax_visible(struct kobject *kobj, struct attribute *a, int n)
266{
267 struct device *dev = container_of(kobj, typeof(*dev), kobj);
268 struct dax_device *dax_dev = dax_get_by_host(dev_name(dev));
269
270 WARN_ON_ONCE(!dax_dev);
271 if (!dax_dev)
272 return 0;
273
274#ifndef CONFIG_ARCH_HAS_PMEM_API
275 if (a == &dev_attr_write_cache.attr)
276 return 0;
277#endif
278 return a->mode;
279}
280
281static struct attribute *dax_attributes[] = {
282 &dev_attr_write_cache.attr,
283 NULL,
284};
285
286struct attribute_group dax_attribute_group = {
287 .name = "dax",
288 .attrs = dax_attributes,
289 .is_visible = dax_visible,
290};
291EXPORT_SYMBOL_GPL(dax_attribute_group);
292
293/**
294 * dax_direct_access() - translate a device pgoff to an absolute pfn
295 * @dax_dev: a dax_device instance representing the logical memory range
296 * @pgoff: offset in pages from the start of the device to translate
297 * @nr_pages: number of consecutive pages caller can handle relative to @pfn
298 * @kaddr: output parameter that returns a virtual address mapping of pfn
299 * @pfn: output parameter that returns an absolute pfn translation of @pgoff
300 *
301 * Return: negative errno if an error occurs, otherwise the number of
302 * pages accessible at the device relative @pgoff.
303 */
304long dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff, long nr_pages,
305 void **kaddr, pfn_t *pfn)
306{
307 long avail;
308
309 if (!dax_dev)
310 return -EOPNOTSUPP;
311
312 if (!dax_alive(dax_dev))
313 return -ENXIO;
314
315 if (nr_pages < 0)
316 return -EINVAL;
317
318 avail = dax_dev->ops->direct_access(dax_dev, pgoff, nr_pages,
319 kaddr, pfn);
320 if (!avail)
321 return -ERANGE;
322 return min(avail, nr_pages);
323}
324EXPORT_SYMBOL_GPL(dax_direct_access);
325
326bool dax_supported(struct dax_device *dax_dev, struct block_device *bdev,
327 int blocksize, sector_t start, sector_t len)
328{
329 if (!dax_dev)
330 return false;
331
332 if (!dax_alive(dax_dev))
333 return false;
334
335 return dax_dev->ops->dax_supported(dax_dev, bdev, blocksize, start, len);
336}
337EXPORT_SYMBOL_GPL(dax_supported);
338
339size_t dax_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
340 size_t bytes, struct iov_iter *i)
341{
342 if (!dax_alive(dax_dev))
343 return 0;
344
345 return dax_dev->ops->copy_from_iter(dax_dev, pgoff, addr, bytes, i);
346}
347EXPORT_SYMBOL_GPL(dax_copy_from_iter);
348
349size_t dax_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, void *addr,
350 size_t bytes, struct iov_iter *i)
351{
352 if (!dax_alive(dax_dev))
353 return 0;
354
355 return dax_dev->ops->copy_to_iter(dax_dev, pgoff, addr, bytes, i);
356}
357EXPORT_SYMBOL_GPL(dax_copy_to_iter);
358
359int dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
360 size_t nr_pages)
361{
362 if (!dax_alive(dax_dev))
363 return -ENXIO;
364 /*
365 * There are no callers that want to zero more than one page as of now.
366 * Once users are there, this check can be removed after the
367 * device mapper code has been updated to split ranges across targets.
368 */
369 if (nr_pages != 1)
370 return -EIO;
371
372 return dax_dev->ops->zero_page_range(dax_dev, pgoff, nr_pages);
373}
374EXPORT_SYMBOL_GPL(dax_zero_page_range);
375
376#ifdef CONFIG_ARCH_HAS_PMEM_API
377void arch_wb_cache_pmem(void *addr, size_t size);
378void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
379{
380 if (unlikely(!dax_write_cache_enabled(dax_dev)))
381 return;
382
383 arch_wb_cache_pmem(addr, size);
384}
385#else
386void dax_flush(struct dax_device *dax_dev, void *addr, size_t size)
387{
388}
389#endif
390EXPORT_SYMBOL_GPL(dax_flush);
391
392void dax_write_cache(struct dax_device *dax_dev, bool wc)
393{
394 if (wc)
395 set_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
396 else
397 clear_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
398}
399EXPORT_SYMBOL_GPL(dax_write_cache);
400
401bool dax_write_cache_enabled(struct dax_device *dax_dev)
402{
403 return test_bit(DAXDEV_WRITE_CACHE, &dax_dev->flags);
404}
405EXPORT_SYMBOL_GPL(dax_write_cache_enabled);
406
407bool __dax_synchronous(struct dax_device *dax_dev)
408{
409 return test_bit(DAXDEV_SYNC, &dax_dev->flags);
410}
411EXPORT_SYMBOL_GPL(__dax_synchronous);
412
413void __set_dax_synchronous(struct dax_device *dax_dev)
414{
415 set_bit(DAXDEV_SYNC, &dax_dev->flags);
416}
417EXPORT_SYMBOL_GPL(__set_dax_synchronous);
418
419bool dax_alive(struct dax_device *dax_dev)
420{
421 lockdep_assert_held(&dax_srcu);
422 return test_bit(DAXDEV_ALIVE, &dax_dev->flags);
423}
424EXPORT_SYMBOL_GPL(dax_alive);
425
426static int dax_host_hash(const char *host)
427{
428 return hashlen_hash(hashlen_string("DAX", host)) % DAX_HASH_SIZE;
429}
430
431/*
432 * Note, rcu is not protecting the liveness of dax_dev, rcu is ensuring
433 * that any fault handlers or operations that might have seen
434 * dax_alive(), have completed. Any operations that start after
435 * synchronize_srcu() has run will abort upon seeing !dax_alive().
436 */
437void kill_dax(struct dax_device *dax_dev)
438{
439 if (!dax_dev)
440 return;
441
442 clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
443
444 synchronize_srcu(&dax_srcu);
445
446 spin_lock(&dax_host_lock);
447 hlist_del_init(&dax_dev->list);
448 spin_unlock(&dax_host_lock);
449}
450EXPORT_SYMBOL_GPL(kill_dax);
451
452void run_dax(struct dax_device *dax_dev)
453{
454 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
455}
456EXPORT_SYMBOL_GPL(run_dax);
457
458static struct inode *dax_alloc_inode(struct super_block *sb)
459{
460 struct dax_device *dax_dev;
461 struct inode *inode;
462
463 dax_dev = kmem_cache_alloc(dax_cache, GFP_KERNEL);
464 if (!dax_dev)
465 return NULL;
466
467 inode = &dax_dev->inode;
468 inode->i_rdev = 0;
469 return inode;
470}
471
472static struct dax_device *to_dax_dev(struct inode *inode)
473{
474 return container_of(inode, struct dax_device, inode);
475}
476
477static void dax_free_inode(struct inode *inode)
478{
479 struct dax_device *dax_dev = to_dax_dev(inode);
480 kfree(dax_dev->host);
481 dax_dev->host = NULL;
482 if (inode->i_rdev)
483 ida_simple_remove(&dax_minor_ida, iminor(inode));
484 kmem_cache_free(dax_cache, dax_dev);
485}
486
487static void dax_destroy_inode(struct inode *inode)
488{
489 struct dax_device *dax_dev = to_dax_dev(inode);
490 WARN_ONCE(test_bit(DAXDEV_ALIVE, &dax_dev->flags),
491 "kill_dax() must be called before final iput()\n");
492}
493
494static const struct super_operations dax_sops = {
495 .statfs = simple_statfs,
496 .alloc_inode = dax_alloc_inode,
497 .destroy_inode = dax_destroy_inode,
498 .free_inode = dax_free_inode,
499 .drop_inode = generic_delete_inode,
500};
501
502static int dax_init_fs_context(struct fs_context *fc)
503{
504 struct pseudo_fs_context *ctx = init_pseudo(fc, DAXFS_MAGIC);
505 if (!ctx)
506 return -ENOMEM;
507 ctx->ops = &dax_sops;
508 return 0;
509}
510
511static struct file_system_type dax_fs_type = {
512 .name = "dax",
513 .init_fs_context = dax_init_fs_context,
514 .kill_sb = kill_anon_super,
515};
516
517static int dax_test(struct inode *inode, void *data)
518{
519 dev_t devt = *(dev_t *) data;
520
521 return inode->i_rdev == devt;
522}
523
524static int dax_set(struct inode *inode, void *data)
525{
526 dev_t devt = *(dev_t *) data;
527
528 inode->i_rdev = devt;
529 return 0;
530}
531
532static struct dax_device *dax_dev_get(dev_t devt)
533{
534 struct dax_device *dax_dev;
535 struct inode *inode;
536
537 inode = iget5_locked(dax_superblock, hash_32(devt + DAXFS_MAGIC, 31),
538 dax_test, dax_set, &devt);
539
540 if (!inode)
541 return NULL;
542
543 dax_dev = to_dax_dev(inode);
544 if (inode->i_state & I_NEW) {
545 set_bit(DAXDEV_ALIVE, &dax_dev->flags);
546 inode->i_cdev = &dax_dev->cdev;
547 inode->i_mode = S_IFCHR;
548 inode->i_flags = S_DAX;
549 mapping_set_gfp_mask(&inode->i_data, GFP_USER);
550 unlock_new_inode(inode);
551 }
552
553 return dax_dev;
554}
555
556static void dax_add_host(struct dax_device *dax_dev, const char *host)
557{
558 int hash;
559
560 /*
561 * Unconditionally init dax_dev since it's coming from a
562 * non-zeroed slab cache
563 */
564 INIT_HLIST_NODE(&dax_dev->list);
565 dax_dev->host = host;
566 if (!host)
567 return;
568
569 hash = dax_host_hash(host);
570 spin_lock(&dax_host_lock);
571 hlist_add_head(&dax_dev->list, &dax_host_list[hash]);
572 spin_unlock(&dax_host_lock);
573}
574
575struct dax_device *alloc_dax(void *private, const char *__host,
576 const struct dax_operations *ops, unsigned long flags)
577{
578 struct dax_device *dax_dev;
579 const char *host;
580 dev_t devt;
581 int minor;
582
583 if (ops && !ops->zero_page_range) {
584 pr_debug("%s: error: device does not provide dax"
585 " operation zero_page_range()\n",
586 __host ? __host : "Unknown");
587 return ERR_PTR(-EINVAL);
588 }
589
590 host = kstrdup(__host, GFP_KERNEL);
591 if (__host && !host)
592 return ERR_PTR(-ENOMEM);
593
594 minor = ida_simple_get(&dax_minor_ida, 0, MINORMASK+1, GFP_KERNEL);
595 if (minor < 0)
596 goto err_minor;
597
598 devt = MKDEV(MAJOR(dax_devt), minor);
599 dax_dev = dax_dev_get(devt);
600 if (!dax_dev)
601 goto err_dev;
602
603 dax_add_host(dax_dev, host);
604 dax_dev->ops = ops;
605 dax_dev->private = private;
606 if (flags & DAXDEV_F_SYNC)
607 set_dax_synchronous(dax_dev);
608
609 return dax_dev;
610
611 err_dev:
612 ida_simple_remove(&dax_minor_ida, minor);
613 err_minor:
614 kfree(host);
615 return ERR_PTR(-ENOMEM);
616}
617EXPORT_SYMBOL_GPL(alloc_dax);
618
619void put_dax(struct dax_device *dax_dev)
620{
621 if (!dax_dev)
622 return;
623 iput(&dax_dev->inode);
624}
625EXPORT_SYMBOL_GPL(put_dax);
626
627/**
628 * dax_get_by_host() - temporary lookup mechanism for filesystem-dax
629 * @host: alternate name for the device registered by a dax driver
630 */
631struct dax_device *dax_get_by_host(const char *host)
632{
633 struct dax_device *dax_dev, *found = NULL;
634 int hash, id;
635
636 if (!host)
637 return NULL;
638
639 hash = dax_host_hash(host);
640
641 id = dax_read_lock();
642 spin_lock(&dax_host_lock);
643 hlist_for_each_entry(dax_dev, &dax_host_list[hash], list) {
644 if (!dax_alive(dax_dev)
645 || strcmp(host, dax_dev->host) != 0)
646 continue;
647
648 if (igrab(&dax_dev->inode))
649 found = dax_dev;
650 break;
651 }
652 spin_unlock(&dax_host_lock);
653 dax_read_unlock(id);
654
655 return found;
656}
657EXPORT_SYMBOL_GPL(dax_get_by_host);
658
659/**
660 * inode_dax: convert a public inode into its dax_dev
661 * @inode: An inode with i_cdev pointing to a dax_dev
662 *
663 * Note this is not equivalent to to_dax_dev() which is for private
664 * internal use where we know the inode filesystem type == dax_fs_type.
665 */
666struct dax_device *inode_dax(struct inode *inode)
667{
668 struct cdev *cdev = inode->i_cdev;
669
670 return container_of(cdev, struct dax_device, cdev);
671}
672EXPORT_SYMBOL_GPL(inode_dax);
673
674struct inode *dax_inode(struct dax_device *dax_dev)
675{
676 return &dax_dev->inode;
677}
678EXPORT_SYMBOL_GPL(dax_inode);
679
680void *dax_get_private(struct dax_device *dax_dev)
681{
682 if (!test_bit(DAXDEV_ALIVE, &dax_dev->flags))
683 return NULL;
684 return dax_dev->private;
685}
686EXPORT_SYMBOL_GPL(dax_get_private);
687
688static void init_once(void *_dax_dev)
689{
690 struct dax_device *dax_dev = _dax_dev;
691 struct inode *inode = &dax_dev->inode;
692
693 memset(dax_dev, 0, sizeof(*dax_dev));
694 inode_init_once(inode);
695}
696
697static int dax_fs_init(void)
698{
699 int rc;
700
701 dax_cache = kmem_cache_create("dax_cache", sizeof(struct dax_device), 0,
702 (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
703 SLAB_MEM_SPREAD|SLAB_ACCOUNT),
704 init_once);
705 if (!dax_cache)
706 return -ENOMEM;
707
708 dax_mnt = kern_mount(&dax_fs_type);
709 if (IS_ERR(dax_mnt)) {
710 rc = PTR_ERR(dax_mnt);
711 goto err_mount;
712 }
713 dax_superblock = dax_mnt->mnt_sb;
714
715 return 0;
716
717 err_mount:
718 kmem_cache_destroy(dax_cache);
719
720 return rc;
721}
722
723static void dax_fs_exit(void)
724{
725 kern_unmount(dax_mnt);
726 kmem_cache_destroy(dax_cache);
727}
728
729static int __init dax_core_init(void)
730{
731 int rc;
732
733 rc = dax_fs_init();
734 if (rc)
735 return rc;
736
737 rc = alloc_chrdev_region(&dax_devt, 0, MINORMASK+1, "dax");
738 if (rc)
739 goto err_chrdev;
740
741 rc = dax_bus_init();
742 if (rc)
743 goto err_bus;
744 return 0;
745
746err_bus:
747 unregister_chrdev_region(dax_devt, MINORMASK+1);
748err_chrdev:
749 dax_fs_exit();
750 return 0;
751}
752
753static void __exit dax_core_exit(void)
754{
755 dax_bus_exit();
756 unregister_chrdev_region(dax_devt, MINORMASK+1);
757 ida_destroy(&dax_minor_ida);
758 dax_fs_exit();
759}
760
761MODULE_AUTHOR("Intel Corporation");
762MODULE_LICENSE("GPL v2");
763subsys_initcall(dax_core_init);
764module_exit(dax_core_exit);