blob: 761495bf5eb91d97c6483646d89a67f10933864f [file] [log] [blame]
Matthew Wilcoxd475c632015-02-16 15:58:56 -08001/*
2 * fs/dax.c - Direct Access filesystem code
3 * Copyright (c) 2013-2014 Intel Corporation
4 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
5 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
6 *
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
10 *
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 */
16
17#include <linux/atomic.h>
18#include <linux/blkdev.h>
19#include <linux/buffer_head.h>
Ross Zwislerd77e92e2015-09-09 10:29:40 -060020#include <linux/dax.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080021#include <linux/fs.h>
22#include <linux/genhd.h>
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -080023#include <linux/highmem.h>
24#include <linux/memcontrol.h>
25#include <linux/mm.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080026#include <linux/mutex.h>
Ross Zwisler9973c982016-01-22 15:10:47 -080027#include <linux/pagevec.h>
Ross Zwisler2765cfb2015-08-18 13:55:40 -060028#include <linux/pmem.h>
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080029#include <linux/sched.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080030#include <linux/uio.h>
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -080031#include <linux/vmstat.h>
Dan Williams34c0fd52016-01-15 16:56:14 -080032#include <linux/pfn_t.h>
Dan Williams0e749e52016-01-15 16:55:53 -080033#include <linux/sizes.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080034
Jan Karae8043152016-05-12 18:29:16 +020035/*
36 * We use lowest available bit in exceptional entry for locking, other two
37 * bits to determine entry type. In total 3 special bits.
38 */
39#define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 3)
40#define RADIX_DAX_PTE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
41#define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
42#define RADIX_DAX_TYPE_MASK (RADIX_DAX_PTE | RADIX_DAX_PMD)
43#define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_TYPE_MASK)
NeilBrown78a9be02016-05-20 17:03:51 -070044#define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
45#define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
Jan Karae8043152016-05-12 18:29:16 +020046 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE) | \
47 RADIX_TREE_EXCEPTIONAL_ENTRY))
NeilBrowne4b27492016-05-11 11:58:47 +020048
Jan Karaac401cc2016-05-12 18:29:18 +020049/* We choose 4096 entries - same as per-zone page wait tables */
50#define DAX_WAIT_TABLE_BITS 12
51#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
52
53wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
54
55static int __init init_dax_wait_table(void)
56{
57 int i;
58
59 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
60 init_waitqueue_head(wait_table + i);
61 return 0;
62}
63fs_initcall(init_dax_wait_table);
64
65static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
66 pgoff_t index)
67{
68 unsigned long hash = hash_long((unsigned long)mapping ^ index,
69 DAX_WAIT_TABLE_BITS);
70 return wait_table + hash;
71}
NeilBrown78a9be02016-05-20 17:03:51 -070072
Dan Williamsb2e0d162016-01-15 16:55:59 -080073static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
74{
75 struct request_queue *q = bdev->bd_queue;
76 long rc = -EIO;
77
78 dax->addr = (void __pmem *) ERR_PTR(-EIO);
79 if (blk_queue_enter(q, true) != 0)
80 return rc;
81
82 rc = bdev_direct_access(bdev, dax);
83 if (rc < 0) {
84 dax->addr = (void __pmem *) ERR_PTR(rc);
85 blk_queue_exit(q);
86 return rc;
87 }
88 return rc;
89}
90
91static void dax_unmap_atomic(struct block_device *bdev,
92 const struct blk_dax_ctl *dax)
93{
94 if (IS_ERR(dax->addr))
95 return;
96 blk_queue_exit(bdev->bd_queue);
97}
98
Dan Williamsd1a5f2b42016-01-28 20:25:31 -080099struct page *read_dax_sector(struct block_device *bdev, sector_t n)
100{
101 struct page *page = alloc_pages(GFP_KERNEL, 0);
102 struct blk_dax_ctl dax = {
103 .size = PAGE_SIZE,
104 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
105 };
106 long rc;
107
108 if (!page)
109 return ERR_PTR(-ENOMEM);
110
111 rc = dax_map_atomic(bdev, &dax);
112 if (rc < 0)
113 return ERR_PTR(rc);
114 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
115 dax_unmap_atomic(bdev, &dax);
116 return page;
117}
118
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800119static bool buffer_written(struct buffer_head *bh)
120{
121 return buffer_mapped(bh) && !buffer_unwritten(bh);
122}
123
124/*
125 * When ext4 encounters a hole, it returns without modifying the buffer_head
126 * which means that we can't trust b_size. To cope with this, we set b_state
127 * to 0 before calling get_block and, if any bit is set, we know we can trust
128 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
129 * and would save us time calling get_block repeatedly.
130 */
131static bool buffer_size_valid(struct buffer_head *bh)
132{
133 return bh->b_state != 0;
134}
135
Dan Williamsb2e0d162016-01-15 16:55:59 -0800136
137static sector_t to_sector(const struct buffer_head *bh,
138 const struct inode *inode)
139{
140 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
141
142 return sector;
143}
144
Omar Sandovala95cd632015-03-16 04:33:51 -0700145static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
146 loff_t start, loff_t end, get_block_t get_block,
147 struct buffer_head *bh)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800148{
Dan Williamsb2e0d162016-01-15 16:55:59 -0800149 loff_t pos = start, max = start, bh_max = start;
150 bool hole = false, need_wmb = false;
151 struct block_device *bdev = NULL;
152 int rw = iov_iter_rw(iter), rc;
153 long map_len = 0;
154 struct blk_dax_ctl dax = {
155 .addr = (void __pmem *) ERR_PTR(-EIO),
156 };
Jan Kara069c77b2016-05-11 11:58:51 +0200157 unsigned blkbits = inode->i_blkbits;
158 sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
159 >> blkbits;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800160
Dan Williamsb2e0d162016-01-15 16:55:59 -0800161 if (rw == READ)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800162 end = min(end, i_size_read(inode));
163
164 while (pos < end) {
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600165 size_t len;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800166 if (pos == max) {
Jeff Moyere94f5a22015-08-14 16:15:31 -0400167 long page = pos >> PAGE_SHIFT;
168 sector_t block = page << (PAGE_SHIFT - blkbits);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800169 unsigned first = pos - (block << blkbits);
170 long size;
171
172 if (pos == bh_max) {
173 bh->b_size = PAGE_ALIGN(end - pos);
174 bh->b_state = 0;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800175 rc = get_block(inode, block, bh, rw == WRITE);
176 if (rc)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800177 break;
178 if (!buffer_size_valid(bh))
179 bh->b_size = 1 << blkbits;
180 bh_max = pos - first + bh->b_size;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800181 bdev = bh->b_bdev;
Jan Kara069c77b2016-05-11 11:58:51 +0200182 /*
183 * We allow uninitialized buffers for writes
184 * beyond EOF as those cannot race with faults
185 */
186 WARN_ON_ONCE(
187 (buffer_new(bh) && block < file_blks) ||
188 (rw == WRITE && buffer_unwritten(bh)));
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800189 } else {
190 unsigned done = bh->b_size -
191 (bh_max - (pos - first));
192 bh->b_blocknr += done >> blkbits;
193 bh->b_size -= done;
194 }
195
Dan Williamsb2e0d162016-01-15 16:55:59 -0800196 hole = rw == READ && !buffer_written(bh);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800197 if (hole) {
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800198 size = bh->b_size - first;
199 } else {
Dan Williamsb2e0d162016-01-15 16:55:59 -0800200 dax_unmap_atomic(bdev, &dax);
201 dax.sector = to_sector(bh, inode);
202 dax.size = bh->b_size;
203 map_len = dax_map_atomic(bdev, &dax);
204 if (map_len < 0) {
205 rc = map_len;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800206 break;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800207 }
Dan Williamsb2e0d162016-01-15 16:55:59 -0800208 dax.addr += first;
209 size = map_len - first;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800210 }
211 max = min(pos + size, end);
212 }
213
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600214 if (iov_iter_rw(iter) == WRITE) {
Dan Williamsb2e0d162016-01-15 16:55:59 -0800215 len = copy_from_iter_pmem(dax.addr, max - pos, iter);
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600216 need_wmb = true;
217 } else if (!hole)
Dan Williamsb2e0d162016-01-15 16:55:59 -0800218 len = copy_to_iter((void __force *) dax.addr, max - pos,
Ross Zwislere2e05392015-08-18 13:55:41 -0600219 iter);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800220 else
221 len = iov_iter_zero(max - pos, iter);
222
Al Virocadfbb62015-11-10 19:42:49 -0700223 if (!len) {
Dan Williamsb2e0d162016-01-15 16:55:59 -0800224 rc = -EFAULT;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800225 break;
Al Virocadfbb62015-11-10 19:42:49 -0700226 }
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800227
228 pos += len;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800229 if (!IS_ERR(dax.addr))
230 dax.addr += len;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800231 }
232
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600233 if (need_wmb)
234 wmb_pmem();
Dan Williamsb2e0d162016-01-15 16:55:59 -0800235 dax_unmap_atomic(bdev, &dax);
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600236
Dan Williamsb2e0d162016-01-15 16:55:59 -0800237 return (pos == start) ? rc : pos - start;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800238}
239
240/**
241 * dax_do_io - Perform I/O to a DAX file
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800242 * @iocb: The control block for this I/O
243 * @inode: The file which the I/O is directed at
244 * @iter: The addresses to do I/O from or to
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800245 * @get_block: The filesystem method used to translate file offsets to blocks
246 * @end_io: A filesystem callback for I/O completion
247 * @flags: See below
248 *
249 * This function uses the same locking scheme as do_blockdev_direct_IO:
250 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
251 * caller for writes. For reads, we take and release the i_mutex ourselves.
252 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
253 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
254 * is in progress.
255 */
Omar Sandovala95cd632015-03-16 04:33:51 -0700256ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -0700257 struct iov_iter *iter, get_block_t get_block,
Omar Sandovala95cd632015-03-16 04:33:51 -0700258 dio_iodone_t end_io, int flags)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800259{
260 struct buffer_head bh;
261 ssize_t retval = -EINVAL;
Christoph Hellwigc8b8e322016-04-07 08:51:58 -0700262 loff_t pos = iocb->ki_pos;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800263 loff_t end = pos + iov_iter_count(iter);
264
265 memset(&bh, 0, sizeof(bh));
Ross Zwislereab95db2016-01-22 15:10:59 -0800266 bh.b_bdev = inode->i_sb->s_bdev;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800267
Jan Karac3d98e32016-05-11 11:58:52 +0200268 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
Al Viro59551022016-01-22 15:40:57 -0500269 inode_lock(inode);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800270
271 /* Protects against truncate */
Matthew Wilcoxbbab37d2015-07-03 10:40:42 -0400272 if (!(flags & DIO_SKIP_DIO_COUNT))
273 inode_dio_begin(inode);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800274
Omar Sandovala95cd632015-03-16 04:33:51 -0700275 retval = dax_io(inode, iter, pos, end, get_block, &bh);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800276
Omar Sandovala95cd632015-03-16 04:33:51 -0700277 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
Al Viro59551022016-01-22 15:40:57 -0500278 inode_unlock(inode);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800279
Christoph Hellwig187372a2016-02-08 14:40:51 +1100280 if (end_io) {
281 int err;
282
283 err = end_io(iocb, pos, retval, bh.b_private);
284 if (err)
285 retval = err;
286 }
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800287
Matthew Wilcoxbbab37d2015-07-03 10:40:42 -0400288 if (!(flags & DIO_SKIP_DIO_COUNT))
289 inode_dio_end(inode);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800290 return retval;
291}
292EXPORT_SYMBOL_GPL(dax_do_io);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800293
294/*
Jan Karaac401cc2016-05-12 18:29:18 +0200295 * DAX radix tree locking
296 */
297struct exceptional_entry_key {
298 struct address_space *mapping;
299 unsigned long index;
300};
301
302struct wait_exceptional_entry_queue {
303 wait_queue_t wait;
304 struct exceptional_entry_key key;
305};
306
307static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
308 int sync, void *keyp)
309{
310 struct exceptional_entry_key *key = keyp;
311 struct wait_exceptional_entry_queue *ewait =
312 container_of(wait, struct wait_exceptional_entry_queue, wait);
313
314 if (key->mapping != ewait->key.mapping ||
315 key->index != ewait->key.index)
316 return 0;
317 return autoremove_wake_function(wait, mode, sync, NULL);
318}
319
320/*
321 * Check whether the given slot is locked. The function must be called with
322 * mapping->tree_lock held
323 */
324static inline int slot_locked(struct address_space *mapping, void **slot)
325{
326 unsigned long entry = (unsigned long)
327 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
328 return entry & RADIX_DAX_ENTRY_LOCK;
329}
330
331/*
332 * Mark the given slot is locked. The function must be called with
333 * mapping->tree_lock held
334 */
335static inline void *lock_slot(struct address_space *mapping, void **slot)
336{
337 unsigned long entry = (unsigned long)
338 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
339
340 entry |= RADIX_DAX_ENTRY_LOCK;
341 radix_tree_replace_slot(slot, (void *)entry);
342 return (void *)entry;
343}
344
345/*
346 * Mark the given slot is unlocked. The function must be called with
347 * mapping->tree_lock held
348 */
349static inline void *unlock_slot(struct address_space *mapping, void **slot)
350{
351 unsigned long entry = (unsigned long)
352 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
353
354 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
355 radix_tree_replace_slot(slot, (void *)entry);
356 return (void *)entry;
357}
358
359/*
360 * Lookup entry in radix tree, wait for it to become unlocked if it is
361 * exceptional entry and return it. The caller must call
362 * put_unlocked_mapping_entry() when he decided not to lock the entry or
363 * put_locked_mapping_entry() when he locked the entry and now wants to
364 * unlock it.
365 *
366 * The function must be called with mapping->tree_lock held.
367 */
368static void *get_unlocked_mapping_entry(struct address_space *mapping,
369 pgoff_t index, void ***slotp)
370{
371 void *ret, **slot;
372 struct wait_exceptional_entry_queue ewait;
373 wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
374
375 init_wait(&ewait.wait);
376 ewait.wait.func = wake_exceptional_entry_func;
377 ewait.key.mapping = mapping;
378 ewait.key.index = index;
379
380 for (;;) {
381 ret = __radix_tree_lookup(&mapping->page_tree, index, NULL,
382 &slot);
383 if (!ret || !radix_tree_exceptional_entry(ret) ||
384 !slot_locked(mapping, slot)) {
385 if (slotp)
386 *slotp = slot;
387 return ret;
388 }
389 prepare_to_wait_exclusive(wq, &ewait.wait,
390 TASK_UNINTERRUPTIBLE);
391 spin_unlock_irq(&mapping->tree_lock);
392 schedule();
393 finish_wait(wq, &ewait.wait);
394 spin_lock_irq(&mapping->tree_lock);
395 }
396}
397
398/*
399 * Find radix tree entry at given index. If it points to a page, return with
400 * the page locked. If it points to the exceptional entry, return with the
401 * radix tree entry locked. If the radix tree doesn't contain given index,
402 * create empty exceptional entry for the index and return with it locked.
403 *
404 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
405 * persistent memory the benefit is doubtful. We can add that later if we can
406 * show it helps.
407 */
408static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index)
409{
410 void *ret, **slot;
411
412restart:
413 spin_lock_irq(&mapping->tree_lock);
414 ret = get_unlocked_mapping_entry(mapping, index, &slot);
415 /* No entry for given index? Make sure radix tree is big enough. */
416 if (!ret) {
417 int err;
418
419 spin_unlock_irq(&mapping->tree_lock);
420 err = radix_tree_preload(
421 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
422 if (err)
423 return ERR_PTR(err);
424 ret = (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY |
425 RADIX_DAX_ENTRY_LOCK);
426 spin_lock_irq(&mapping->tree_lock);
427 err = radix_tree_insert(&mapping->page_tree, index, ret);
428 radix_tree_preload_end();
429 if (err) {
430 spin_unlock_irq(&mapping->tree_lock);
431 /* Someone already created the entry? */
432 if (err == -EEXIST)
433 goto restart;
434 return ERR_PTR(err);
435 }
436 /* Good, we have inserted empty locked entry into the tree. */
437 mapping->nrexceptional++;
438 spin_unlock_irq(&mapping->tree_lock);
439 return ret;
440 }
441 /* Normal page in radix tree? */
442 if (!radix_tree_exceptional_entry(ret)) {
443 struct page *page = ret;
444
445 get_page(page);
446 spin_unlock_irq(&mapping->tree_lock);
447 lock_page(page);
448 /* Page got truncated? Retry... */
449 if (unlikely(page->mapping != mapping)) {
450 unlock_page(page);
451 put_page(page);
452 goto restart;
453 }
454 return page;
455 }
456 ret = lock_slot(mapping, slot);
457 spin_unlock_irq(&mapping->tree_lock);
458 return ret;
459}
460
461void dax_wake_mapping_entry_waiter(struct address_space *mapping,
462 pgoff_t index, bool wake_all)
463{
464 wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
465
466 /*
467 * Checking for locked entry and prepare_to_wait_exclusive() happens
468 * under mapping->tree_lock, ditto for entry handling in our callers.
469 * So at this point all tasks that could have seen our entry locked
470 * must be in the waitqueue and the following check will see them.
471 */
472 if (waitqueue_active(wq)) {
473 struct exceptional_entry_key key;
474
475 key.mapping = mapping;
476 key.index = index;
477 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
478 }
479}
480
Jan Karabc2466e2016-05-12 18:29:19 +0200481void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
Jan Karaac401cc2016-05-12 18:29:18 +0200482{
483 void *ret, **slot;
484
485 spin_lock_irq(&mapping->tree_lock);
486 ret = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
487 if (WARN_ON_ONCE(!ret || !radix_tree_exceptional_entry(ret) ||
488 !slot_locked(mapping, slot))) {
489 spin_unlock_irq(&mapping->tree_lock);
490 return;
491 }
492 unlock_slot(mapping, slot);
493 spin_unlock_irq(&mapping->tree_lock);
494 dax_wake_mapping_entry_waiter(mapping, index, false);
495}
496
497static void put_locked_mapping_entry(struct address_space *mapping,
498 pgoff_t index, void *entry)
499{
500 if (!radix_tree_exceptional_entry(entry)) {
501 unlock_page(entry);
502 put_page(entry);
503 } else {
Jan Karabc2466e2016-05-12 18:29:19 +0200504 dax_unlock_mapping_entry(mapping, index);
Jan Karaac401cc2016-05-12 18:29:18 +0200505 }
506}
507
508/*
509 * Called when we are done with radix tree entry we looked up via
510 * get_unlocked_mapping_entry() and which we didn't lock in the end.
511 */
512static void put_unlocked_mapping_entry(struct address_space *mapping,
513 pgoff_t index, void *entry)
514{
515 if (!radix_tree_exceptional_entry(entry))
516 return;
517
518 /* We have to wake up next waiter for the radix tree entry lock */
519 dax_wake_mapping_entry_waiter(mapping, index, false);
520}
521
522/*
523 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
524 * entry to get unlocked before deleting it.
525 */
526int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
527{
528 void *entry;
529
530 spin_lock_irq(&mapping->tree_lock);
531 entry = get_unlocked_mapping_entry(mapping, index, NULL);
532 /*
533 * This gets called from truncate / punch_hole path. As such, the caller
534 * must hold locks protecting against concurrent modifications of the
535 * radix tree (usually fs-private i_mmap_sem for writing). Since the
536 * caller has seen exceptional entry for this index, we better find it
537 * at that index as well...
538 */
539 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) {
540 spin_unlock_irq(&mapping->tree_lock);
541 return 0;
542 }
543 radix_tree_delete(&mapping->page_tree, index);
544 mapping->nrexceptional--;
545 spin_unlock_irq(&mapping->tree_lock);
546 dax_wake_mapping_entry_waiter(mapping, index, true);
547
548 return 1;
549}
550
551/*
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800552 * The user has performed a load from a hole in the file. Allocating
553 * a new page in the file would cause excessive storage usage for
554 * workloads with sparse files. We allocate a page cache page instead.
555 * We'll kick it out of the page cache if it's ever written to,
556 * otherwise it will simply fall out of the page cache under memory
557 * pressure without ever having been dirtied.
558 */
Jan Karaac401cc2016-05-12 18:29:18 +0200559static int dax_load_hole(struct address_space *mapping, void *entry,
560 struct vm_fault *vmf)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800561{
Jan Karaac401cc2016-05-12 18:29:18 +0200562 struct page *page;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800563
Jan Karaac401cc2016-05-12 18:29:18 +0200564 /* Hole page already exists? Return it... */
565 if (!radix_tree_exceptional_entry(entry)) {
566 vmf->page = entry;
567 return VM_FAULT_LOCKED;
568 }
569
570 /* This will replace locked radix tree entry with a hole page */
571 page = find_or_create_page(mapping, vmf->pgoff,
572 vmf->gfp_mask | __GFP_ZERO);
573 if (!page) {
574 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
575 return VM_FAULT_OOM;
576 }
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800577 vmf->page = page;
578 return VM_FAULT_LOCKED;
579}
580
Dan Williamsb2e0d162016-01-15 16:55:59 -0800581static int copy_user_bh(struct page *to, struct inode *inode,
582 struct buffer_head *bh, unsigned long vaddr)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800583{
Dan Williamsb2e0d162016-01-15 16:55:59 -0800584 struct blk_dax_ctl dax = {
585 .sector = to_sector(bh, inode),
586 .size = bh->b_size,
587 };
588 struct block_device *bdev = bh->b_bdev;
Ross Zwislere2e05392015-08-18 13:55:41 -0600589 void *vto;
590
Dan Williamsb2e0d162016-01-15 16:55:59 -0800591 if (dax_map_atomic(bdev, &dax) < 0)
592 return PTR_ERR(dax.addr);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800593 vto = kmap_atomic(to);
Dan Williamsb2e0d162016-01-15 16:55:59 -0800594 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800595 kunmap_atomic(vto);
Dan Williamsb2e0d162016-01-15 16:55:59 -0800596 dax_unmap_atomic(bdev, &dax);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800597 return 0;
598}
599
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300600#define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
Ross Zwisler9973c982016-01-22 15:10:47 -0800601
Jan Karaac401cc2016-05-12 18:29:18 +0200602static void *dax_insert_mapping_entry(struct address_space *mapping,
603 struct vm_fault *vmf,
604 void *entry, sector_t sector)
Ross Zwisler9973c982016-01-22 15:10:47 -0800605{
606 struct radix_tree_root *page_tree = &mapping->page_tree;
Jan Karaac401cc2016-05-12 18:29:18 +0200607 int error = 0;
608 bool hole_fill = false;
609 void *new_entry;
610 pgoff_t index = vmf->pgoff;
Ross Zwisler9973c982016-01-22 15:10:47 -0800611
Jan Karaac401cc2016-05-12 18:29:18 +0200612 if (vmf->flags & FAULT_FLAG_WRITE)
Dmitry Monakhovd2b2a282016-02-05 15:36:55 -0800613 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Ross Zwisler9973c982016-01-22 15:10:47 -0800614
Jan Karaac401cc2016-05-12 18:29:18 +0200615 /* Replacing hole page with block mapping? */
616 if (!radix_tree_exceptional_entry(entry)) {
617 hole_fill = true;
618 /*
619 * Unmap the page now before we remove it from page cache below.
620 * The page is locked so it cannot be faulted in again.
621 */
622 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
623 PAGE_SIZE, 0);
624 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
625 if (error)
626 return ERR_PTR(error);
Ross Zwisler9973c982016-01-22 15:10:47 -0800627 }
628
Jan Karaac401cc2016-05-12 18:29:18 +0200629 spin_lock_irq(&mapping->tree_lock);
630 new_entry = (void *)((unsigned long)RADIX_DAX_ENTRY(sector, false) |
631 RADIX_DAX_ENTRY_LOCK);
632 if (hole_fill) {
633 __delete_from_page_cache(entry, NULL);
634 /* Drop pagecache reference */
635 put_page(entry);
636 error = radix_tree_insert(page_tree, index, new_entry);
637 if (error) {
638 new_entry = ERR_PTR(error);
Ross Zwisler9973c982016-01-22 15:10:47 -0800639 goto unlock;
640 }
Jan Karaac401cc2016-05-12 18:29:18 +0200641 mapping->nrexceptional++;
642 } else {
643 void **slot;
644 void *ret;
Ross Zwisler9973c982016-01-22 15:10:47 -0800645
Jan Karaac401cc2016-05-12 18:29:18 +0200646 ret = __radix_tree_lookup(page_tree, index, NULL, &slot);
647 WARN_ON_ONCE(ret != entry);
648 radix_tree_replace_slot(slot, new_entry);
Ross Zwisler9973c982016-01-22 15:10:47 -0800649 }
Jan Karaac401cc2016-05-12 18:29:18 +0200650 if (vmf->flags & FAULT_FLAG_WRITE)
Ross Zwisler9973c982016-01-22 15:10:47 -0800651 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
652 unlock:
653 spin_unlock_irq(&mapping->tree_lock);
Jan Karaac401cc2016-05-12 18:29:18 +0200654 if (hole_fill) {
655 radix_tree_preload_end();
656 /*
657 * We don't need hole page anymore, it has been replaced with
658 * locked radix tree entry now.
659 */
660 if (mapping->a_ops->freepage)
661 mapping->a_ops->freepage(entry);
662 unlock_page(entry);
663 put_page(entry);
664 }
665 return new_entry;
Ross Zwisler9973c982016-01-22 15:10:47 -0800666}
667
668static int dax_writeback_one(struct block_device *bdev,
669 struct address_space *mapping, pgoff_t index, void *entry)
670{
671 struct radix_tree_root *page_tree = &mapping->page_tree;
672 int type = RADIX_DAX_TYPE(entry);
673 struct radix_tree_node *node;
674 struct blk_dax_ctl dax;
675 void **slot;
676 int ret = 0;
677
678 spin_lock_irq(&mapping->tree_lock);
679 /*
680 * Regular page slots are stabilized by the page lock even
681 * without the tree itself locked. These unlocked entries
682 * need verification under the tree lock.
683 */
684 if (!__radix_tree_lookup(page_tree, index, &node, &slot))
685 goto unlock;
686 if (*slot != entry)
687 goto unlock;
688
689 /* another fsync thread may have already written back this entry */
690 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
691 goto unlock;
692
693 if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
694 ret = -EIO;
695 goto unlock;
696 }
697
698 dax.sector = RADIX_DAX_SECTOR(entry);
699 dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
700 spin_unlock_irq(&mapping->tree_lock);
701
702 /*
703 * We cannot hold tree_lock while calling dax_map_atomic() because it
704 * eventually calls cond_resched().
705 */
706 ret = dax_map_atomic(bdev, &dax);
707 if (ret < 0)
708 return ret;
709
710 if (WARN_ON_ONCE(ret < dax.size)) {
711 ret = -EIO;
712 goto unmap;
713 }
714
715 wb_cache_pmem(dax.addr, dax.size);
716
717 spin_lock_irq(&mapping->tree_lock);
718 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
719 spin_unlock_irq(&mapping->tree_lock);
720 unmap:
721 dax_unmap_atomic(bdev, &dax);
722 return ret;
723
724 unlock:
725 spin_unlock_irq(&mapping->tree_lock);
726 return ret;
727}
728
729/*
730 * Flush the mapping to the persistent domain within the byte range of [start,
731 * end]. This is required by data integrity operations to ensure file data is
732 * on persistent storage prior to completion of the operation.
733 */
Ross Zwisler7f6d5b52016-02-26 15:19:55 -0800734int dax_writeback_mapping_range(struct address_space *mapping,
735 struct block_device *bdev, struct writeback_control *wbc)
Ross Zwisler9973c982016-01-22 15:10:47 -0800736{
737 struct inode *inode = mapping->host;
Ross Zwisler9973c982016-01-22 15:10:47 -0800738 pgoff_t start_index, end_index, pmd_index;
739 pgoff_t indices[PAGEVEC_SIZE];
740 struct pagevec pvec;
741 bool done = false;
742 int i, ret = 0;
743 void *entry;
744
745 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
746 return -EIO;
747
Ross Zwisler7f6d5b52016-02-26 15:19:55 -0800748 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
749 return 0;
750
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300751 start_index = wbc->range_start >> PAGE_SHIFT;
752 end_index = wbc->range_end >> PAGE_SHIFT;
Ross Zwisler9973c982016-01-22 15:10:47 -0800753 pmd_index = DAX_PMD_INDEX(start_index);
754
755 rcu_read_lock();
756 entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
757 rcu_read_unlock();
758
759 /* see if the start of our range is covered by a PMD entry */
760 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
761 start_index = pmd_index;
762
763 tag_pages_for_writeback(mapping, start_index, end_index);
764
765 pagevec_init(&pvec, 0);
766 while (!done) {
767 pvec.nr = find_get_entries_tag(mapping, start_index,
768 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
769 pvec.pages, indices);
770
771 if (pvec.nr == 0)
772 break;
773
774 for (i = 0; i < pvec.nr; i++) {
775 if (indices[i] > end_index) {
776 done = true;
777 break;
778 }
779
780 ret = dax_writeback_one(bdev, mapping, indices[i],
781 pvec.pages[i]);
782 if (ret < 0)
783 return ret;
784 }
785 }
786 wmb_pmem();
787 return 0;
788}
789EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
790
Jan Karaac401cc2016-05-12 18:29:18 +0200791static int dax_insert_mapping(struct address_space *mapping,
792 struct buffer_head *bh, void **entryp,
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800793 struct vm_area_struct *vma, struct vm_fault *vmf)
794{
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800795 unsigned long vaddr = (unsigned long)vmf->virtual_address;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800796 struct block_device *bdev = bh->b_bdev;
797 struct blk_dax_ctl dax = {
Jan Karaac401cc2016-05-12 18:29:18 +0200798 .sector = to_sector(bh, mapping->host),
Dan Williamsb2e0d162016-01-15 16:55:59 -0800799 .size = bh->b_size,
800 };
Jan Karaac401cc2016-05-12 18:29:18 +0200801 void *ret;
802 void *entry = *entryp;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800803
Jan Kara4d9a2c82016-05-12 18:29:20 +0200804 if (dax_map_atomic(bdev, &dax) < 0)
805 return PTR_ERR(dax.addr);
Dan Williamsb2e0d162016-01-15 16:55:59 -0800806 dax_unmap_atomic(bdev, &dax);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800807
Jan Karaac401cc2016-05-12 18:29:18 +0200808 ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector);
Jan Kara4d9a2c82016-05-12 18:29:20 +0200809 if (IS_ERR(ret))
810 return PTR_ERR(ret);
Jan Karaac401cc2016-05-12 18:29:18 +0200811 *entryp = ret;
Ross Zwisler9973c982016-01-22 15:10:47 -0800812
Jan Kara4d9a2c82016-05-12 18:29:20 +0200813 return vm_insert_mixed(vma, vaddr, dax.pfn);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800814}
815
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000816/**
817 * __dax_fault - handle a page fault on a DAX file
818 * @vma: The virtual memory area where the fault occurred
819 * @vmf: The description of the fault
820 * @get_block: The filesystem method used to translate file offsets to blocks
821 *
822 * When a page fault occurs, filesystems may call this helper in their
823 * fault handler for DAX files. __dax_fault() assumes the caller has done all
824 * the necessary locking for the page fault to proceed successfully.
825 */
826int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
Jan Kara02fbd132016-05-11 11:58:48 +0200827 get_block_t get_block)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800828{
829 struct file *file = vma->vm_file;
830 struct address_space *mapping = file->f_mapping;
831 struct inode *inode = mapping->host;
Jan Karaac401cc2016-05-12 18:29:18 +0200832 void *entry;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800833 struct buffer_head bh;
834 unsigned long vaddr = (unsigned long)vmf->virtual_address;
835 unsigned blkbits = inode->i_blkbits;
836 sector_t block;
837 pgoff_t size;
838 int error;
839 int major = 0;
840
Jan Karaac401cc2016-05-12 18:29:18 +0200841 /*
842 * Check whether offset isn't beyond end of file now. Caller is supposed
843 * to hold locks serializing us with truncate / punch hole so this is
844 * a reliable test.
845 */
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800846 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
847 if (vmf->pgoff >= size)
848 return VM_FAULT_SIGBUS;
849
850 memset(&bh, 0, sizeof(bh));
851 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
Ross Zwislereab95db2016-01-22 15:10:59 -0800852 bh.b_bdev = inode->i_sb->s_bdev;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800853 bh.b_size = PAGE_SIZE;
854
Jan Karaac401cc2016-05-12 18:29:18 +0200855 entry = grab_mapping_entry(mapping, vmf->pgoff);
856 if (IS_ERR(entry)) {
857 error = PTR_ERR(entry);
858 goto out;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800859 }
860
861 error = get_block(inode, block, &bh, 0);
862 if (!error && (bh.b_size < PAGE_SIZE))
863 error = -EIO; /* fs corruption? */
864 if (error)
Jan Karaac401cc2016-05-12 18:29:18 +0200865 goto unlock_entry;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800866
867 if (vmf->cow_page) {
868 struct page *new_page = vmf->cow_page;
869 if (buffer_written(&bh))
Dan Williamsb2e0d162016-01-15 16:55:59 -0800870 error = copy_user_bh(new_page, inode, &bh, vaddr);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800871 else
872 clear_user_highpage(new_page, vaddr);
873 if (error)
Jan Karaac401cc2016-05-12 18:29:18 +0200874 goto unlock_entry;
875 if (!radix_tree_exceptional_entry(entry)) {
876 vmf->page = entry;
Jan Karabc2466e2016-05-12 18:29:19 +0200877 return VM_FAULT_LOCKED;
Jan Karaac401cc2016-05-12 18:29:18 +0200878 }
Jan Karabc2466e2016-05-12 18:29:19 +0200879 vmf->entry = entry;
880 return VM_FAULT_DAX_LOCKED;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800881 }
882
Jan Karaac401cc2016-05-12 18:29:18 +0200883 if (!buffer_mapped(&bh)) {
884 if (vmf->flags & FAULT_FLAG_WRITE) {
885 error = get_block(inode, block, &bh, 1);
886 count_vm_event(PGMAJFAULT);
887 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
888 major = VM_FAULT_MAJOR;
889 if (!error && (bh.b_size < PAGE_SIZE))
890 error = -EIO;
891 if (error)
892 goto unlock_entry;
893 } else {
894 return dax_load_hole(mapping, entry, vmf);
895 }
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800896 }
897
Jan Kara02fbd132016-05-11 11:58:48 +0200898 /* Filesystem should not return unwritten buffers to us! */
Jan Kara2b109452016-05-11 11:58:50 +0200899 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
Jan Karaac401cc2016-05-12 18:29:18 +0200900 error = dax_insert_mapping(mapping, &bh, &entry, vma, vmf);
901 unlock_entry:
902 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800903 out:
904 if (error == -ENOMEM)
905 return VM_FAULT_OOM | major;
906 /* -EBUSY is fine, somebody else faulted on the same PTE */
907 if ((error < 0) && (error != -EBUSY))
908 return VM_FAULT_SIGBUS | major;
909 return VM_FAULT_NOPAGE | major;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800910}
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000911EXPORT_SYMBOL(__dax_fault);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800912
913/**
914 * dax_fault - handle a page fault on a DAX file
915 * @vma: The virtual memory area where the fault occurred
916 * @vmf: The description of the fault
917 * @get_block: The filesystem method used to translate file offsets to blocks
918 *
919 * When a page fault occurs, filesystems may call this helper in their
920 * fault handler for DAX files.
921 */
922int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
Jan Kara02fbd132016-05-11 11:58:48 +0200923 get_block_t get_block)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800924{
925 int result;
926 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
927
928 if (vmf->flags & FAULT_FLAG_WRITE) {
929 sb_start_pagefault(sb);
930 file_update_time(vma->vm_file);
931 }
Jan Kara02fbd132016-05-11 11:58:48 +0200932 result = __dax_fault(vma, vmf, get_block);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800933 if (vmf->flags & FAULT_FLAG_WRITE)
934 sb_end_pagefault(sb);
935
936 return result;
937}
938EXPORT_SYMBOL_GPL(dax_fault);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800939
Jan Kara348e9672016-05-12 18:29:15 +0200940#if defined(CONFIG_TRANSPARENT_HUGEPAGE)
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700941/*
942 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
943 * more often than one might expect in the below function.
944 */
945#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
946
Dan Williamscbb38e42016-01-15 16:56:58 -0800947static void __dax_dbg(struct buffer_head *bh, unsigned long address,
948 const char *reason, const char *fn)
949{
950 if (bh) {
951 char bname[BDEVNAME_SIZE];
952 bdevname(bh->b_bdev, bname);
953 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
954 "length %zd fallback: %s\n", fn, current->comm,
955 address, bname, bh->b_state, (u64)bh->b_blocknr,
956 bh->b_size, reason);
957 } else {
958 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
959 current->comm, address, reason);
960 }
961}
962
963#define dax_pmd_dbg(bh, address, reason) __dax_dbg(bh, address, reason, "dax_pmd")
964
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700965int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
Jan Kara02fbd132016-05-11 11:58:48 +0200966 pmd_t *pmd, unsigned int flags, get_block_t get_block)
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700967{
968 struct file *file = vma->vm_file;
969 struct address_space *mapping = file->f_mapping;
970 struct inode *inode = mapping->host;
971 struct buffer_head bh;
972 unsigned blkbits = inode->i_blkbits;
973 unsigned long pmd_addr = address & PMD_MASK;
974 bool write = flags & FAULT_FLAG_WRITE;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800975 struct block_device *bdev;
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700976 pgoff_t size, pgoff;
Dan Williamsb2e0d162016-01-15 16:55:59 -0800977 sector_t block;
Jan Karaac401cc2016-05-12 18:29:18 +0200978 int result = 0;
Ross Zwisler9973c982016-01-22 15:10:47 -0800979 bool alloc = false;
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700980
Dan Williamsc046c322016-01-15 16:57:01 -0800981 /* dax pmd mappings require pfn_t_devmap() */
Dan Williamsee82c9e2015-11-15 16:06:32 -0800982 if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
983 return VM_FAULT_FALLBACK;
984
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700985 /* Fall back to PTEs if we're going to COW */
Toshi Kani59bf4fb2016-01-15 16:56:05 -0800986 if (write && !(vma->vm_flags & VM_SHARED)) {
987 split_huge_pmd(vma, pmd, address);
Dan Williamscbb38e42016-01-15 16:56:58 -0800988 dax_pmd_dbg(NULL, address, "cow write");
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700989 return VM_FAULT_FALLBACK;
Toshi Kani59bf4fb2016-01-15 16:56:05 -0800990 }
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700991 /* If the PMD would extend outside the VMA */
Dan Williamscbb38e42016-01-15 16:56:58 -0800992 if (pmd_addr < vma->vm_start) {
993 dax_pmd_dbg(NULL, address, "vma start unaligned");
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700994 return VM_FAULT_FALLBACK;
Dan Williamscbb38e42016-01-15 16:56:58 -0800995 }
996 if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
997 dax_pmd_dbg(NULL, address, "vma end unaligned");
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700998 return VM_FAULT_FALLBACK;
Dan Williamscbb38e42016-01-15 16:56:58 -0800999 }
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001000
Matthew Wilcox3fdd1b472015-09-08 14:59:39 -07001001 pgoff = linear_page_index(vma, pmd_addr);
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001002 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1003 if (pgoff >= size)
1004 return VM_FAULT_SIGBUS;
1005 /* If the PMD would cover blocks out of the file */
Dan Williamscbb38e42016-01-15 16:56:58 -08001006 if ((pgoff | PG_PMD_COLOUR) >= size) {
1007 dax_pmd_dbg(NULL, address,
1008 "offset + huge page size > file size");
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001009 return VM_FAULT_FALLBACK;
Dan Williamscbb38e42016-01-15 16:56:58 -08001010 }
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001011
1012 memset(&bh, 0, sizeof(bh));
Ross Zwislerd4bbe702016-01-22 15:10:31 -08001013 bh.b_bdev = inode->i_sb->s_bdev;
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001014 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
1015
1016 bh.b_size = PMD_SIZE;
Ross Zwisler9973c982016-01-22 15:10:47 -08001017
1018 if (get_block(inode, block, &bh, 0) != 0)
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001019 return VM_FAULT_SIGBUS;
Ross Zwisler9973c982016-01-22 15:10:47 -08001020
1021 if (!buffer_mapped(&bh) && write) {
1022 if (get_block(inode, block, &bh, 1) != 0)
1023 return VM_FAULT_SIGBUS;
1024 alloc = true;
Jan Kara2b109452016-05-11 11:58:50 +02001025 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
Ross Zwisler9973c982016-01-22 15:10:47 -08001026 }
1027
Dan Williamsb2e0d162016-01-15 16:55:59 -08001028 bdev = bh.b_bdev;
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001029
1030 /*
1031 * If the filesystem isn't willing to tell us the length of a hole,
1032 * just fall back to PTEs. Calling get_block 512 times in a loop
1033 * would be silly.
1034 */
Dan Williamscbb38e42016-01-15 16:56:58 -08001035 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
1036 dax_pmd_dbg(&bh, address, "allocated block too small");
Ross Zwisler9973c982016-01-22 15:10:47 -08001037 return VM_FAULT_FALLBACK;
Dan Williamscbb38e42016-01-15 16:56:58 -08001038 }
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001039
Ross Zwisler9973c982016-01-22 15:10:47 -08001040 /*
1041 * If we allocated new storage, make sure no process has any
1042 * zero pages covering this hole
1043 */
1044 if (alloc) {
1045 loff_t lstart = pgoff << PAGE_SHIFT;
1046 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
1047
1048 truncate_pagecache_range(inode, lstart, lend);
1049 }
1050
Jan Karab9953532016-05-12 18:29:14 +02001051 if (!write && !buffer_mapped(&bh)) {
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001052 spinlock_t *ptl;
Kirill A. Shutemovd295e342015-09-08 14:59:34 -07001053 pmd_t entry;
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001054 struct page *zero_page = get_huge_zero_page();
Kirill A. Shutemovd295e342015-09-08 14:59:34 -07001055
Dan Williamscbb38e42016-01-15 16:56:58 -08001056 if (unlikely(!zero_page)) {
1057 dax_pmd_dbg(&bh, address, "no zero page");
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001058 goto fallback;
Dan Williamscbb38e42016-01-15 16:56:58 -08001059 }
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001060
Kirill A. Shutemovd295e342015-09-08 14:59:34 -07001061 ptl = pmd_lock(vma->vm_mm, pmd);
1062 if (!pmd_none(*pmd)) {
1063 spin_unlock(ptl);
Dan Williamscbb38e42016-01-15 16:56:58 -08001064 dax_pmd_dbg(&bh, address, "pmd already present");
Kirill A. Shutemovd295e342015-09-08 14:59:34 -07001065 goto fallback;
1066 }
1067
Dan Williamscbb38e42016-01-15 16:56:58 -08001068 dev_dbg(part_to_dev(bdev->bd_part),
1069 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
1070 __func__, current->comm, address,
1071 (unsigned long long) to_sector(&bh, inode));
1072
Kirill A. Shutemovd295e342015-09-08 14:59:34 -07001073 entry = mk_pmd(zero_page, vma->vm_page_prot);
1074 entry = pmd_mkhuge(entry);
1075 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001076 result = VM_FAULT_NOPAGE;
Kirill A. Shutemovd295e342015-09-08 14:59:34 -07001077 spin_unlock(ptl);
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001078 } else {
Dan Williamsb2e0d162016-01-15 16:55:59 -08001079 struct blk_dax_ctl dax = {
1080 .sector = to_sector(&bh, inode),
1081 .size = PMD_SIZE,
1082 };
1083 long length = dax_map_atomic(bdev, &dax);
1084
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001085 if (length < 0) {
Dan Williams8b3db9792016-02-24 14:02:06 -08001086 dax_pmd_dbg(&bh, address, "dax-error fallback");
1087 goto fallback;
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001088 }
Dan Williamscbb38e42016-01-15 16:56:58 -08001089 if (length < PMD_SIZE) {
1090 dax_pmd_dbg(&bh, address, "dax-length too small");
1091 dax_unmap_atomic(bdev, &dax);
1092 goto fallback;
1093 }
1094 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
1095 dax_pmd_dbg(&bh, address, "pfn unaligned");
Dan Williamsb2e0d162016-01-15 16:55:59 -08001096 dax_unmap_atomic(bdev, &dax);
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001097 goto fallback;
Dan Williamsb2e0d162016-01-15 16:55:59 -08001098 }
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001099
Dan Williamsc046c322016-01-15 16:57:01 -08001100 if (!pfn_t_devmap(dax.pfn)) {
Dan Williamsb2e0d162016-01-15 16:55:59 -08001101 dax_unmap_atomic(bdev, &dax);
Dan Williamscbb38e42016-01-15 16:56:58 -08001102 dax_pmd_dbg(&bh, address, "pfn not in memmap");
Dan Williams152d7bd2015-11-12 18:33:54 -08001103 goto fallback;
Dan Williamsb2e0d162016-01-15 16:55:59 -08001104 }
Dan Williamsb2e0d162016-01-15 16:55:59 -08001105 dax_unmap_atomic(bdev, &dax);
Ross Zwisler0f90cc62015-10-15 15:28:32 -07001106
Ross Zwisler9973c982016-01-22 15:10:47 -08001107 /*
1108 * For PTE faults we insert a radix tree entry for reads, and
1109 * leave it clean. Then on the first write we dirty the radix
1110 * tree entry via the dax_pfn_mkwrite() path. This sequence
1111 * allows the dax_pfn_mkwrite() call to be simpler and avoid a
1112 * call into get_block() to translate the pgoff to a sector in
1113 * order to be able to create a new radix tree entry.
1114 *
1115 * The PMD path doesn't have an equivalent to
1116 * dax_pfn_mkwrite(), though, so for a read followed by a
1117 * write we traverse all the way through __dax_pmd_fault()
1118 * twice. This means we can just skip inserting a radix tree
1119 * entry completely on the initial read and just wait until
1120 * the write to insert a dirty entry.
1121 */
1122 if (write) {
Jan Karaac401cc2016-05-12 18:29:18 +02001123 /*
1124 * We should insert radix-tree entry and dirty it here.
1125 * For now this is broken...
1126 */
Ross Zwisler9973c982016-01-22 15:10:47 -08001127 }
1128
Dan Williamscbb38e42016-01-15 16:56:58 -08001129 dev_dbg(part_to_dev(bdev->bd_part),
1130 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1131 __func__, current->comm, address,
1132 pfn_t_to_pfn(dax.pfn),
1133 (unsigned long long) dax.sector);
Dan Williams34c0fd52016-01-15 16:56:14 -08001134 result |= vmf_insert_pfn_pmd(vma, address, pmd,
Dan Williamsf25748e32016-01-15 16:56:43 -08001135 dax.pfn, write);
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001136 }
1137
1138 out:
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001139 return result;
1140
1141 fallback:
1142 count_vm_event(THP_FAULT_FALLBACK);
1143 result = VM_FAULT_FALLBACK;
1144 goto out;
1145}
1146EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1147
1148/**
1149 * dax_pmd_fault - handle a PMD fault on a DAX file
1150 * @vma: The virtual memory area where the fault occurred
1151 * @vmf: The description of the fault
1152 * @get_block: The filesystem method used to translate file offsets to blocks
1153 *
1154 * When a page fault occurs, filesystems may call this helper in their
1155 * pmd_fault handler for DAX files.
1156 */
1157int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
Jan Kara02fbd132016-05-11 11:58:48 +02001158 pmd_t *pmd, unsigned int flags, get_block_t get_block)
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001159{
1160 int result;
1161 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1162
1163 if (flags & FAULT_FLAG_WRITE) {
1164 sb_start_pagefault(sb);
1165 file_update_time(vma->vm_file);
1166 }
Jan Kara02fbd132016-05-11 11:58:48 +02001167 result = __dax_pmd_fault(vma, address, pmd, flags, get_block);
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001168 if (flags & FAULT_FLAG_WRITE)
1169 sb_end_pagefault(sb);
1170
1171 return result;
1172}
1173EXPORT_SYMBOL_GPL(dax_pmd_fault);
Valentin Rothbergdd8a2b62015-09-08 14:59:09 -07001174#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
Matthew Wilcox844f35d2015-09-08 14:58:57 -07001175
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001176/**
Boaz Harrosh0e3b2102015-04-15 16:15:14 -07001177 * dax_pfn_mkwrite - handle first write to DAX page
1178 * @vma: The virtual memory area where the fault occurred
1179 * @vmf: The description of the fault
Boaz Harrosh0e3b2102015-04-15 16:15:14 -07001180 */
1181int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1182{
Ross Zwisler9973c982016-01-22 15:10:47 -08001183 struct file *file = vma->vm_file;
Jan Karaac401cc2016-05-12 18:29:18 +02001184 struct address_space *mapping = file->f_mapping;
1185 void *entry;
1186 pgoff_t index = vmf->pgoff;
Boaz Harrosh0e3b2102015-04-15 16:15:14 -07001187
Jan Karaac401cc2016-05-12 18:29:18 +02001188 spin_lock_irq(&mapping->tree_lock);
1189 entry = get_unlocked_mapping_entry(mapping, index, NULL);
1190 if (!entry || !radix_tree_exceptional_entry(entry))
1191 goto out;
1192 radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
1193 put_unlocked_mapping_entry(mapping, index, entry);
1194out:
1195 spin_unlock_irq(&mapping->tree_lock);
Boaz Harrosh0e3b2102015-04-15 16:15:14 -07001196 return VM_FAULT_NOPAGE;
1197}
1198EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1199
Vishal Verma4b0228f2016-04-21 15:13:46 -04001200static bool dax_range_is_aligned(struct block_device *bdev,
1201 unsigned int offset, unsigned int length)
1202{
1203 unsigned short sector_size = bdev_logical_block_size(bdev);
1204
1205 if (!IS_ALIGNED(offset, sector_size))
1206 return false;
1207 if (!IS_ALIGNED(length, sector_size))
1208 return false;
1209
1210 return true;
1211}
1212
Christoph Hellwig679c8bd2016-05-09 10:47:04 +02001213int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
1214 unsigned int offset, unsigned int length)
1215{
1216 struct blk_dax_ctl dax = {
1217 .sector = sector,
1218 .size = PAGE_SIZE,
1219 };
1220
Vishal Verma4b0228f2016-04-21 15:13:46 -04001221 if (dax_range_is_aligned(bdev, offset, length)) {
1222 sector_t start_sector = dax.sector + (offset >> 9);
1223
1224 return blkdev_issue_zeroout(bdev, start_sector,
1225 length >> 9, GFP_NOFS, true);
1226 } else {
1227 if (dax_map_atomic(bdev, &dax) < 0)
1228 return PTR_ERR(dax.addr);
1229 clear_pmem(dax.addr + offset, length);
1230 wmb_pmem();
1231 dax_unmap_atomic(bdev, &dax);
1232 }
Christoph Hellwig679c8bd2016-05-09 10:47:04 +02001233 return 0;
1234}
1235EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1236
Boaz Harrosh0e3b2102015-04-15 16:15:14 -07001237/**
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001238 * dax_zero_page_range - zero a range within a page of a DAX file
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001239 * @inode: The file being truncated
1240 * @from: The file offset that is being truncated to
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001241 * @length: The number of bytes to zero
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001242 * @get_block: The filesystem method used to translate file offsets to blocks
1243 *
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001244 * This function can be called by a filesystem when it is zeroing part of a
1245 * page in a DAX file. This is intended for hole-punch operations. If
1246 * you are truncating a file, the helper function dax_truncate_page() may be
1247 * more convenient.
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001248 */
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001249int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1250 get_block_t get_block)
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001251{
1252 struct buffer_head bh;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001253 pgoff_t index = from >> PAGE_SHIFT;
1254 unsigned offset = from & (PAGE_SIZE-1);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001255 int err;
1256
1257 /* Block boundary? Nothing to do */
1258 if (!length)
1259 return 0;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001260 BUG_ON((offset + length) > PAGE_SIZE);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001261
1262 memset(&bh, 0, sizeof(bh));
Ross Zwislereab95db2016-01-22 15:10:59 -08001263 bh.b_bdev = inode->i_sb->s_bdev;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001264 bh.b_size = PAGE_SIZE;
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001265 err = get_block(inode, index, &bh, 0);
Christoph Hellwig679c8bd2016-05-09 10:47:04 +02001266 if (err < 0 || !buffer_written(&bh))
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001267 return err;
Dan Williamsb2e0d162016-01-15 16:55:59 -08001268
Christoph Hellwig679c8bd2016-05-09 10:47:04 +02001269 return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode),
1270 offset, length);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001271}
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001272EXPORT_SYMBOL_GPL(dax_zero_page_range);
1273
1274/**
1275 * dax_truncate_page - handle a partial page being truncated in a DAX file
1276 * @inode: The file being truncated
1277 * @from: The file offset that is being truncated to
1278 * @get_block: The filesystem method used to translate file offsets to blocks
1279 *
1280 * Similar to block_truncate_page(), this function can be called by a
1281 * filesystem when it is truncating a DAX file to handle the partial page.
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001282 */
1283int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1284{
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001285 unsigned length = PAGE_ALIGN(from) - from;
Matthew Wilcox25726bc2015-02-16 15:59:35 -08001286 return dax_zero_page_range(inode, from, length, get_block);
1287}
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -08001288EXPORT_SYMBOL_GPL(dax_truncate_page);