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 <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);
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);