blob: d1e5cb7311a1de295ecd4348360a9a9947e71142 [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 Zwisler2765cfb2015-08-18 13:55:40 -060027#include <linux/pmem.h>
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080028#include <linux/sched.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080029#include <linux/uio.h>
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -080030#include <linux/vmstat.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080031
Dave Chinner1ca19152015-11-03 12:37:00 +110032/*
33 * dax_clear_blocks() is called from within transaction context from XFS,
34 * and hence this means the stack from this point must follow GFP_NOFS
35 * semantics for all operations.
36 */
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080037int dax_clear_blocks(struct inode *inode, sector_t block, long size)
38{
39 struct block_device *bdev = inode->i_sb->s_bdev;
40 sector_t sector = block << (inode->i_blkbits - 9);
41
42 might_sleep();
43 do {
Ross Zwislere2e05392015-08-18 13:55:41 -060044 void __pmem *addr;
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080045 unsigned long pfn;
46 long count;
47
48 count = bdev_direct_access(bdev, sector, &addr, &pfn, size);
49 if (count < 0)
50 return count;
51 BUG_ON(size < count);
52 while (count > 0) {
53 unsigned pgsz = PAGE_SIZE - offset_in_page(addr);
54 if (pgsz > count)
55 pgsz = count;
Ross Zwislere2e05392015-08-18 13:55:41 -060056 clear_pmem(addr, pgsz);
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080057 addr += pgsz;
58 size -= pgsz;
59 count -= pgsz;
60 BUG_ON(pgsz & 511);
61 sector += pgsz / 512;
62 cond_resched();
63 }
64 } while (size);
65
Ross Zwisler2765cfb2015-08-18 13:55:40 -060066 wmb_pmem();
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080067 return 0;
68}
69EXPORT_SYMBOL_GPL(dax_clear_blocks);
70
Ross Zwislere2e05392015-08-18 13:55:41 -060071static long dax_get_addr(struct buffer_head *bh, void __pmem **addr,
72 unsigned blkbits)
Matthew Wilcoxd475c632015-02-16 15:58:56 -080073{
74 unsigned long pfn;
75 sector_t sector = bh->b_blocknr << (blkbits - 9);
76 return bdev_direct_access(bh->b_bdev, sector, addr, &pfn, bh->b_size);
77}
78
Ross Zwisler2765cfb2015-08-18 13:55:40 -060079/* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
Ross Zwislere2e05392015-08-18 13:55:41 -060080static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
81 loff_t pos, loff_t end)
Matthew Wilcoxd475c632015-02-16 15:58:56 -080082{
83 loff_t final = end - pos + first; /* The final byte of the buffer */
84
85 if (first > 0)
Ross Zwislere2e05392015-08-18 13:55:41 -060086 clear_pmem(addr, first);
Matthew Wilcoxd475c632015-02-16 15:58:56 -080087 if (final < size)
Ross Zwislere2e05392015-08-18 13:55:41 -060088 clear_pmem(addr + final, size - final);
Matthew Wilcoxd475c632015-02-16 15:58:56 -080089}
90
91static bool buffer_written(struct buffer_head *bh)
92{
93 return buffer_mapped(bh) && !buffer_unwritten(bh);
94}
95
96/*
97 * When ext4 encounters a hole, it returns without modifying the buffer_head
98 * which means that we can't trust b_size. To cope with this, we set b_state
99 * to 0 before calling get_block and, if any bit is set, we know we can trust
100 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
101 * and would save us time calling get_block repeatedly.
102 */
103static bool buffer_size_valid(struct buffer_head *bh)
104{
105 return bh->b_state != 0;
106}
107
Omar Sandovala95cd632015-03-16 04:33:51 -0700108static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
109 loff_t start, loff_t end, get_block_t get_block,
110 struct buffer_head *bh)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800111{
112 ssize_t retval = 0;
113 loff_t pos = start;
114 loff_t max = start;
115 loff_t bh_max = start;
Ross Zwislere2e05392015-08-18 13:55:41 -0600116 void __pmem *addr;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800117 bool hole = false;
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600118 bool need_wmb = false;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800119
Omar Sandovala95cd632015-03-16 04:33:51 -0700120 if (iov_iter_rw(iter) != WRITE)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800121 end = min(end, i_size_read(inode));
122
123 while (pos < end) {
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600124 size_t len;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800125 if (pos == max) {
126 unsigned blkbits = inode->i_blkbits;
Jeff Moyere94f5a22015-08-14 16:15:31 -0400127 long page = pos >> PAGE_SHIFT;
128 sector_t block = page << (PAGE_SHIFT - blkbits);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800129 unsigned first = pos - (block << blkbits);
130 long size;
131
132 if (pos == bh_max) {
133 bh->b_size = PAGE_ALIGN(end - pos);
134 bh->b_state = 0;
135 retval = get_block(inode, block, bh,
Omar Sandovala95cd632015-03-16 04:33:51 -0700136 iov_iter_rw(iter) == WRITE);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800137 if (retval)
138 break;
139 if (!buffer_size_valid(bh))
140 bh->b_size = 1 << blkbits;
141 bh_max = pos - first + bh->b_size;
142 } else {
143 unsigned done = bh->b_size -
144 (bh_max - (pos - first));
145 bh->b_blocknr += done >> blkbits;
146 bh->b_size -= done;
147 }
148
Omar Sandovala95cd632015-03-16 04:33:51 -0700149 hole = iov_iter_rw(iter) != WRITE && !buffer_written(bh);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800150 if (hole) {
151 addr = NULL;
152 size = bh->b_size - first;
153 } else {
154 retval = dax_get_addr(bh, &addr, blkbits);
155 if (retval < 0)
156 break;
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600157 if (buffer_unwritten(bh) || buffer_new(bh)) {
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800158 dax_new_buf(addr, retval, first, pos,
159 end);
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600160 need_wmb = true;
161 }
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800162 addr += first;
163 size = retval - first;
164 }
165 max = min(pos + size, end);
166 }
167
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600168 if (iov_iter_rw(iter) == WRITE) {
Ross Zwislere2e05392015-08-18 13:55:41 -0600169 len = copy_from_iter_pmem(addr, max - pos, iter);
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600170 need_wmb = true;
171 } else if (!hole)
Ross Zwislere2e05392015-08-18 13:55:41 -0600172 len = copy_to_iter((void __force *)addr, max - pos,
173 iter);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800174 else
175 len = iov_iter_zero(max - pos, iter);
176
Al Virocadfbb62015-11-10 19:42:49 -0700177 if (!len) {
178 retval = -EFAULT;
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800179 break;
Al Virocadfbb62015-11-10 19:42:49 -0700180 }
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800181
182 pos += len;
183 addr += len;
184 }
185
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600186 if (need_wmb)
187 wmb_pmem();
188
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800189 return (pos == start) ? retval : pos - start;
190}
191
192/**
193 * dax_do_io - Perform I/O to a DAX file
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800194 * @iocb: The control block for this I/O
195 * @inode: The file which the I/O is directed at
196 * @iter: The addresses to do I/O from or to
197 * @pos: The file offset where the I/O starts
198 * @get_block: The filesystem method used to translate file offsets to blocks
199 * @end_io: A filesystem callback for I/O completion
200 * @flags: See below
201 *
202 * This function uses the same locking scheme as do_blockdev_direct_IO:
203 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
204 * caller for writes. For reads, we take and release the i_mutex ourselves.
205 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
206 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
207 * is in progress.
208 */
Omar Sandovala95cd632015-03-16 04:33:51 -0700209ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
210 struct iov_iter *iter, loff_t pos, get_block_t get_block,
211 dio_iodone_t end_io, int flags)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800212{
213 struct buffer_head bh;
214 ssize_t retval = -EINVAL;
215 loff_t end = pos + iov_iter_count(iter);
216
217 memset(&bh, 0, sizeof(bh));
218
Omar Sandovala95cd632015-03-16 04:33:51 -0700219 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800220 struct address_space *mapping = inode->i_mapping;
221 mutex_lock(&inode->i_mutex);
222 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
223 if (retval) {
224 mutex_unlock(&inode->i_mutex);
225 goto out;
226 }
227 }
228
229 /* Protects against truncate */
Matthew Wilcoxbbab37d2015-07-03 10:40:42 -0400230 if (!(flags & DIO_SKIP_DIO_COUNT))
231 inode_dio_begin(inode);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800232
Omar Sandovala95cd632015-03-16 04:33:51 -0700233 retval = dax_io(inode, iter, pos, end, get_block, &bh);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800234
Omar Sandovala95cd632015-03-16 04:33:51 -0700235 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800236 mutex_unlock(&inode->i_mutex);
237
238 if ((retval > 0) && end_io)
239 end_io(iocb, pos, retval, bh.b_private);
240
Matthew Wilcoxbbab37d2015-07-03 10:40:42 -0400241 if (!(flags & DIO_SKIP_DIO_COUNT))
242 inode_dio_end(inode);
Matthew Wilcoxd475c632015-02-16 15:58:56 -0800243 out:
244 return retval;
245}
246EXPORT_SYMBOL_GPL(dax_do_io);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800247
248/*
249 * The user has performed a load from a hole in the file. Allocating
250 * a new page in the file would cause excessive storage usage for
251 * workloads with sparse files. We allocate a page cache page instead.
252 * We'll kick it out of the page cache if it's ever written to,
253 * otherwise it will simply fall out of the page cache under memory
254 * pressure without ever having been dirtied.
255 */
256static int dax_load_hole(struct address_space *mapping, struct page *page,
257 struct vm_fault *vmf)
258{
259 unsigned long size;
260 struct inode *inode = mapping->host;
261 if (!page)
262 page = find_or_create_page(mapping, vmf->pgoff,
263 GFP_KERNEL | __GFP_ZERO);
264 if (!page)
265 return VM_FAULT_OOM;
266 /* Recheck i_size under page lock to avoid truncate race */
267 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
268 if (vmf->pgoff >= size) {
269 unlock_page(page);
270 page_cache_release(page);
271 return VM_FAULT_SIGBUS;
272 }
273
274 vmf->page = page;
275 return VM_FAULT_LOCKED;
276}
277
278static int copy_user_bh(struct page *to, struct buffer_head *bh,
279 unsigned blkbits, unsigned long vaddr)
280{
Ross Zwislere2e05392015-08-18 13:55:41 -0600281 void __pmem *vfrom;
282 void *vto;
283
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800284 if (dax_get_addr(bh, &vfrom, blkbits) < 0)
285 return -EIO;
286 vto = kmap_atomic(to);
Ross Zwislere2e05392015-08-18 13:55:41 -0600287 copy_user_page(vto, (void __force *)vfrom, vaddr, to);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800288 kunmap_atomic(vto);
289 return 0;
290}
291
292static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
293 struct vm_area_struct *vma, struct vm_fault *vmf)
294{
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700295 struct address_space *mapping = inode->i_mapping;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800296 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
297 unsigned long vaddr = (unsigned long)vmf->virtual_address;
Ross Zwislere2e05392015-08-18 13:55:41 -0600298 void __pmem *addr;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800299 unsigned long pfn;
300 pgoff_t size;
301 int error;
302
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700303 i_mmap_lock_read(mapping);
304
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800305 /*
306 * Check truncate didn't happen while we were allocating a block.
307 * If it did, this block may or may not be still allocated to the
308 * file. We can't tell the filesystem to free it because we can't
309 * take i_mutex here. In the worst case, the file still has blocks
310 * allocated past the end of the file.
311 */
312 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
313 if (unlikely(vmf->pgoff >= size)) {
314 error = -EIO;
315 goto out;
316 }
317
318 error = bdev_direct_access(bh->b_bdev, sector, &addr, &pfn, bh->b_size);
319 if (error < 0)
320 goto out;
321 if (error < PAGE_SIZE) {
322 error = -EIO;
323 goto out;
324 }
325
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600326 if (buffer_unwritten(bh) || buffer_new(bh)) {
Ross Zwislere2e05392015-08-18 13:55:41 -0600327 clear_pmem(addr, PAGE_SIZE);
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600328 wmb_pmem();
329 }
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800330
331 error = vm_insert_mixed(vma, vaddr, pfn);
332
333 out:
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700334 i_mmap_unlock_read(mapping);
335
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800336 return error;
337}
338
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000339/**
340 * __dax_fault - handle a page fault on a DAX file
341 * @vma: The virtual memory area where the fault occurred
342 * @vmf: The description of the fault
343 * @get_block: The filesystem method used to translate file offsets to blocks
Dave Chinnerb2442c52015-07-29 11:48:00 +1000344 * @complete_unwritten: The filesystem method used to convert unwritten blocks
345 * to written so the data written to them is exposed. This is required for
346 * required by write faults for filesystems that will return unwritten
347 * extent mappings from @get_block, but it is optional for reads as
348 * dax_insert_mapping() will always zero unwritten blocks. If the fs does
349 * not support unwritten extents, the it should pass NULL.
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000350 *
351 * When a page fault occurs, filesystems may call this helper in their
352 * fault handler for DAX files. __dax_fault() assumes the caller has done all
353 * the necessary locking for the page fault to proceed successfully.
354 */
355int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
Dave Chinnere842f292015-06-04 09:18:18 +1000356 get_block_t get_block, dax_iodone_t complete_unwritten)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800357{
358 struct file *file = vma->vm_file;
359 struct address_space *mapping = file->f_mapping;
360 struct inode *inode = mapping->host;
361 struct page *page;
362 struct buffer_head bh;
363 unsigned long vaddr = (unsigned long)vmf->virtual_address;
364 unsigned blkbits = inode->i_blkbits;
365 sector_t block;
366 pgoff_t size;
367 int error;
368 int major = 0;
369
370 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
371 if (vmf->pgoff >= size)
372 return VM_FAULT_SIGBUS;
373
374 memset(&bh, 0, sizeof(bh));
375 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
376 bh.b_size = PAGE_SIZE;
377
378 repeat:
379 page = find_get_page(mapping, vmf->pgoff);
380 if (page) {
381 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
382 page_cache_release(page);
383 return VM_FAULT_RETRY;
384 }
385 if (unlikely(page->mapping != mapping)) {
386 unlock_page(page);
387 page_cache_release(page);
388 goto repeat;
389 }
390 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
391 if (unlikely(vmf->pgoff >= size)) {
392 /*
393 * We have a struct page covering a hole in the file
394 * from a read fault and we've raced with a truncate
395 */
396 error = -EIO;
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700397 goto unlock_page;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800398 }
399 }
400
401 error = get_block(inode, block, &bh, 0);
402 if (!error && (bh.b_size < PAGE_SIZE))
403 error = -EIO; /* fs corruption? */
404 if (error)
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700405 goto unlock_page;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800406
407 if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
408 if (vmf->flags & FAULT_FLAG_WRITE) {
409 error = get_block(inode, block, &bh, 1);
410 count_vm_event(PGMAJFAULT);
411 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
412 major = VM_FAULT_MAJOR;
413 if (!error && (bh.b_size < PAGE_SIZE))
414 error = -EIO;
415 if (error)
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700416 goto unlock_page;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800417 } else {
418 return dax_load_hole(mapping, page, vmf);
419 }
420 }
421
422 if (vmf->cow_page) {
423 struct page *new_page = vmf->cow_page;
424 if (buffer_written(&bh))
425 error = copy_user_bh(new_page, &bh, blkbits, vaddr);
426 else
427 clear_user_highpage(new_page, vaddr);
428 if (error)
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700429 goto unlock_page;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800430 vmf->page = page;
431 if (!page) {
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700432 i_mmap_lock_read(mapping);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800433 /* Check we didn't race with truncate */
434 size = (i_size_read(inode) + PAGE_SIZE - 1) >>
435 PAGE_SHIFT;
436 if (vmf->pgoff >= size) {
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700437 i_mmap_unlock_read(mapping);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800438 error = -EIO;
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700439 goto out;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800440 }
441 }
442 return VM_FAULT_LOCKED;
443 }
444
445 /* Check we didn't race with a read fault installing a new page */
446 if (!page && major)
447 page = find_lock_page(mapping, vmf->pgoff);
448
449 if (page) {
450 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
451 PAGE_CACHE_SIZE, 0);
452 delete_from_page_cache(page);
453 unlock_page(page);
454 page_cache_release(page);
455 }
456
Dave Chinnere842f292015-06-04 09:18:18 +1000457 /*
458 * If we successfully insert the new mapping over an unwritten extent,
459 * we need to ensure we convert the unwritten extent. If there is an
460 * error inserting the mapping, the filesystem needs to leave it as
461 * unwritten to prevent exposure of the stale underlying data to
462 * userspace, but we still need to call the completion function so
463 * the private resources on the mapping buffer can be released. We
464 * indicate what the callback should do via the uptodate variable, same
465 * as for normal BH based IO completions.
466 */
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800467 error = dax_insert_mapping(inode, &bh, vma, vmf);
Dave Chinnerb2442c52015-07-29 11:48:00 +1000468 if (buffer_unwritten(&bh)) {
469 if (complete_unwritten)
470 complete_unwritten(&bh, !error);
471 else
472 WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
473 }
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800474
475 out:
476 if (error == -ENOMEM)
477 return VM_FAULT_OOM | major;
478 /* -EBUSY is fine, somebody else faulted on the same PTE */
479 if ((error < 0) && (error != -EBUSY))
480 return VM_FAULT_SIGBUS | major;
481 return VM_FAULT_NOPAGE | major;
482
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700483 unlock_page:
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800484 if (page) {
485 unlock_page(page);
486 page_cache_release(page);
487 }
488 goto out;
489}
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000490EXPORT_SYMBOL(__dax_fault);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800491
492/**
493 * dax_fault - handle a page fault on a DAX file
494 * @vma: The virtual memory area where the fault occurred
495 * @vmf: The description of the fault
496 * @get_block: The filesystem method used to translate file offsets to blocks
497 *
498 * When a page fault occurs, filesystems may call this helper in their
499 * fault handler for DAX files.
500 */
501int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
Dave Chinnere842f292015-06-04 09:18:18 +1000502 get_block_t get_block, dax_iodone_t complete_unwritten)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800503{
504 int result;
505 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
506
507 if (vmf->flags & FAULT_FLAG_WRITE) {
508 sb_start_pagefault(sb);
509 file_update_time(vma->vm_file);
510 }
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000511 result = __dax_fault(vma, vmf, get_block, complete_unwritten);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800512 if (vmf->flags & FAULT_FLAG_WRITE)
513 sb_end_pagefault(sb);
514
515 return result;
516}
517EXPORT_SYMBOL_GPL(dax_fault);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800518
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700519#ifdef CONFIG_TRANSPARENT_HUGEPAGE
520/*
521 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
522 * more often than one might expect in the below function.
523 */
524#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
525
526int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
527 pmd_t *pmd, unsigned int flags, get_block_t get_block,
528 dax_iodone_t complete_unwritten)
529{
530 struct file *file = vma->vm_file;
531 struct address_space *mapping = file->f_mapping;
532 struct inode *inode = mapping->host;
533 struct buffer_head bh;
534 unsigned blkbits = inode->i_blkbits;
535 unsigned long pmd_addr = address & PMD_MASK;
536 bool write = flags & FAULT_FLAG_WRITE;
537 long length;
Ross Zwislerd77e92e2015-09-09 10:29:40 -0600538 void __pmem *kaddr;
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700539 pgoff_t size, pgoff;
540 sector_t block, sector;
541 unsigned long pfn;
542 int result = 0;
543
544 /* Fall back to PTEs if we're going to COW */
545 if (write && !(vma->vm_flags & VM_SHARED))
546 return VM_FAULT_FALLBACK;
547 /* If the PMD would extend outside the VMA */
548 if (pmd_addr < vma->vm_start)
549 return VM_FAULT_FALLBACK;
550 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
551 return VM_FAULT_FALLBACK;
552
Matthew Wilcox3fdd1b472015-09-08 14:59:39 -0700553 pgoff = linear_page_index(vma, pmd_addr);
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700554 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
555 if (pgoff >= size)
556 return VM_FAULT_SIGBUS;
557 /* If the PMD would cover blocks out of the file */
558 if ((pgoff | PG_PMD_COLOUR) >= size)
559 return VM_FAULT_FALLBACK;
560
561 memset(&bh, 0, sizeof(bh));
562 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
563
564 bh.b_size = PMD_SIZE;
565 length = get_block(inode, block, &bh, write);
566 if (length)
567 return VM_FAULT_SIGBUS;
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700568 i_mmap_lock_read(mapping);
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700569
570 /*
571 * If the filesystem isn't willing to tell us the length of a hole,
572 * just fall back to PTEs. Calling get_block 512 times in a loop
573 * would be silly.
574 */
575 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE)
576 goto fallback;
577
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -0700578 /*
579 * If we allocated new storage, make sure no process has any
580 * zero pages covering this hole
581 */
582 if (buffer_new(&bh)) {
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700583 i_mmap_unlock_read(mapping);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -0700584 unmap_mapping_range(mapping, pgoff << PAGE_SHIFT, PMD_SIZE, 0);
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700585 i_mmap_lock_read(mapping);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -0700586 }
587
Matthew Wilcox84c4e5e2015-09-08 14:59:17 -0700588 /*
589 * If a truncate happened while we were allocating blocks, we may
590 * leave blocks allocated to the file that are beyond EOF. We can't
591 * take i_mutex here, so just leave them hanging; they'll be freed
592 * when the file is deleted.
593 */
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700594 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
595 if (pgoff >= size) {
596 result = VM_FAULT_SIGBUS;
597 goto out;
598 }
599 if ((pgoff | PG_PMD_COLOUR) >= size)
600 goto fallback;
601
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700602 if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700603 spinlock_t *ptl;
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700604 pmd_t entry;
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700605 struct page *zero_page = get_huge_zero_page();
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700606
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700607 if (unlikely(!zero_page))
608 goto fallback;
609
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700610 ptl = pmd_lock(vma->vm_mm, pmd);
611 if (!pmd_none(*pmd)) {
612 spin_unlock(ptl);
613 goto fallback;
614 }
615
616 entry = mk_pmd(zero_page, vma->vm_page_prot);
617 entry = pmd_mkhuge(entry);
618 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700619 result = VM_FAULT_NOPAGE;
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700620 spin_unlock(ptl);
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700621 } else {
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700622 sector = bh.b_blocknr << (blkbits - 9);
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700623 length = bdev_direct_access(bh.b_bdev, sector, &kaddr, &pfn,
624 bh.b_size);
625 if (length < 0) {
626 result = VM_FAULT_SIGBUS;
627 goto out;
628 }
629 if ((length < PMD_SIZE) || (pfn & PG_PMD_COLOUR))
630 goto fallback;
631
Dan Williams152d7bd2015-11-12 18:33:54 -0800632 /*
633 * TODO: teach vmf_insert_pfn_pmd() to support
634 * 'pte_special' for pmds
635 */
636 if (pfn_valid(pfn))
637 goto fallback;
638
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700639 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
640 int i;
641 for (i = 0; i < PTRS_PER_PMD; i++)
642 clear_pmem(kaddr + i * PAGE_SIZE, PAGE_SIZE);
643 wmb_pmem();
644 count_vm_event(PGMAJFAULT);
645 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
646 result |= VM_FAULT_MAJOR;
647 }
648
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700649 result |= vmf_insert_pfn_pmd(vma, address, pmd, pfn, write);
650 }
651
652 out:
Ross Zwisler0f90cc62015-10-15 15:28:32 -0700653 i_mmap_unlock_read(mapping);
654
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700655 if (buffer_unwritten(&bh))
656 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
657
658 return result;
659
660 fallback:
661 count_vm_event(THP_FAULT_FALLBACK);
662 result = VM_FAULT_FALLBACK;
663 goto out;
664}
665EXPORT_SYMBOL_GPL(__dax_pmd_fault);
666
667/**
668 * dax_pmd_fault - handle a PMD fault on a DAX file
669 * @vma: The virtual memory area where the fault occurred
670 * @vmf: The description of the fault
671 * @get_block: The filesystem method used to translate file offsets to blocks
672 *
673 * When a page fault occurs, filesystems may call this helper in their
674 * pmd_fault handler for DAX files.
675 */
676int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
677 pmd_t *pmd, unsigned int flags, get_block_t get_block,
678 dax_iodone_t complete_unwritten)
679{
680 int result;
681 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
682
683 if (flags & FAULT_FLAG_WRITE) {
684 sb_start_pagefault(sb);
685 file_update_time(vma->vm_file);
686 }
687 result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
688 complete_unwritten);
689 if (flags & FAULT_FLAG_WRITE)
690 sb_end_pagefault(sb);
691
692 return result;
693}
694EXPORT_SYMBOL_GPL(dax_pmd_fault);
Valentin Rothbergdd8a2b62015-09-08 14:59:09 -0700695#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
Matthew Wilcox844f35d2015-09-08 14:58:57 -0700696
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800697/**
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700698 * dax_pfn_mkwrite - handle first write to DAX page
699 * @vma: The virtual memory area where the fault occurred
700 * @vmf: The description of the fault
701 *
702 */
703int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
704{
705 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
706
707 sb_start_pagefault(sb);
708 file_update_time(vma->vm_file);
709 sb_end_pagefault(sb);
710 return VM_FAULT_NOPAGE;
711}
712EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
713
714/**
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800715 * dax_zero_page_range - zero a range within a page of a DAX file
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800716 * @inode: The file being truncated
717 * @from: The file offset that is being truncated to
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800718 * @length: The number of bytes to zero
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800719 * @get_block: The filesystem method used to translate file offsets to blocks
720 *
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800721 * This function can be called by a filesystem when it is zeroing part of a
722 * page in a DAX file. This is intended for hole-punch operations. If
723 * you are truncating a file, the helper function dax_truncate_page() may be
724 * more convenient.
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800725 *
726 * We work in terms of PAGE_CACHE_SIZE here for commonality with
727 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
728 * took care of disposing of the unnecessary blocks. Even if the filesystem
729 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800730 * since the file might be mmapped.
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800731 */
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800732int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
733 get_block_t get_block)
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800734{
735 struct buffer_head bh;
736 pgoff_t index = from >> PAGE_CACHE_SHIFT;
737 unsigned offset = from & (PAGE_CACHE_SIZE-1);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800738 int err;
739
740 /* Block boundary? Nothing to do */
741 if (!length)
742 return 0;
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800743 BUG_ON((offset + length) > PAGE_CACHE_SIZE);
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800744
745 memset(&bh, 0, sizeof(bh));
746 bh.b_size = PAGE_CACHE_SIZE;
747 err = get_block(inode, index, &bh, 0);
748 if (err < 0)
749 return err;
750 if (buffer_written(&bh)) {
Ross Zwislere2e05392015-08-18 13:55:41 -0600751 void __pmem *addr;
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800752 err = dax_get_addr(&bh, &addr, inode->i_blkbits);
753 if (err < 0)
754 return err;
Ross Zwislere2e05392015-08-18 13:55:41 -0600755 clear_pmem(addr + offset, length);
Ross Zwisler2765cfb2015-08-18 13:55:40 -0600756 wmb_pmem();
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800757 }
758
759 return 0;
760}
Matthew Wilcox25726bc2015-02-16 15:59:35 -0800761EXPORT_SYMBOL_GPL(dax_zero_page_range);
762
763/**
764 * dax_truncate_page - handle a partial page being truncated in a DAX file
765 * @inode: The file being truncated
766 * @from: The file offset that is being truncated to
767 * @get_block: The filesystem method used to translate file offsets to blocks
768 *
769 * Similar to block_truncate_page(), this function can be called by a
770 * filesystem when it is truncating a DAX file to handle the partial page.
771 *
772 * We work in terms of PAGE_CACHE_SIZE here for commonality with
773 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
774 * took care of disposing of the unnecessary blocks. Even if the filesystem
775 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
776 * since the file might be mmapped.
777 */
778int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
779{
780 unsigned length = PAGE_CACHE_ALIGN(from) - from;
781 return dax_zero_page_range(inode, from, length, get_block);
782}
Matthew Wilcox4c0ccfe2015-02-16 15:59:06 -0800783EXPORT_SYMBOL_GPL(dax_truncate_page);