blob: 77397b5a96ef9c8f90a8e2b9c7afb50afdbae025 [file] [log] [blame]
Christoph Hellwigae259a92016-06-21 09:23:11 +10001/*
2 * Copyright (C) 2010 Red Hat, Inc.
3 * Copyright (c) 2016 Christoph Hellwig.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 */
14#include <linux/module.h>
15#include <linux/compiler.h>
16#include <linux/fs.h>
17#include <linux/iomap.h>
18#include <linux/uaccess.h>
19#include <linux/gfp.h>
20#include <linux/mm.h>
21#include <linux/swap.h>
22#include <linux/pagemap.h>
Christoph Hellwig8a78cb12018-06-01 09:04:40 -070023#include <linux/pagevec.h>
Christoph Hellwigae259a92016-06-21 09:23:11 +100024#include <linux/file.h>
25#include <linux/uio.h>
26#include <linux/backing-dev.h>
27#include <linux/buffer_head.h>
Christoph Hellwigff6a9292016-11-30 14:36:01 +110028#include <linux/task_io_accounting_ops.h>
Christoph Hellwig9a286f02016-06-21 09:31:39 +100029#include <linux/dax.h>
Ingo Molnarf361bf42017-02-03 23:47:37 +010030#include <linux/sched/signal.h>
Darrick J. Wong67482122018-05-10 08:38:15 -070031#include <linux/swap.h>
Ingo Molnarf361bf42017-02-03 23:47:37 +010032
Christoph Hellwigae259a92016-06-21 09:23:11 +100033#include "internal.h"
34
Christoph Hellwigae259a92016-06-21 09:23:11 +100035/*
36 * Execute a iomap write on a segment of the mapping that spans a
37 * contiguous range of pages that have identical block mapping state.
38 *
39 * This avoids the need to map pages individually, do individual allocations
40 * for each page and most importantly avoid the need for filesystem specific
41 * locking per page. Instead, all the operations are amortised over the entire
42 * range of pages. It is assumed that the filesystems will lock whatever
43 * resources they require in the iomap_begin call, and release them in the
44 * iomap_end call.
45 */
Christoph Hellwigbefb5032016-09-19 11:24:49 +100046loff_t
Christoph Hellwigae259a92016-06-21 09:23:11 +100047iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -080048 const struct iomap_ops *ops, void *data, iomap_actor_t actor)
Christoph Hellwigae259a92016-06-21 09:23:11 +100049{
50 struct iomap iomap = { 0 };
51 loff_t written = 0, ret;
52
53 /*
54 * Need to map a range from start position for length bytes. This can
55 * span multiple pages - it is only guaranteed to return a range of a
56 * single type of pages (e.g. all into a hole, all mapped or all
57 * unwritten). Failure at this point has nothing to undo.
58 *
59 * If allocation is required for this range, reserve the space now so
60 * that the allocation is guaranteed to succeed later on. Once we copy
61 * the data into the page cache pages, then we cannot fail otherwise we
62 * expose transient stale data. If the reserve fails, we can safely
63 * back out at this point as there is nothing to undo.
64 */
65 ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
66 if (ret)
67 return ret;
68 if (WARN_ON(iomap.offset > pos))
69 return -EIO;
Darrick J. Wong0c6dda72018-01-26 11:11:20 -080070 if (WARN_ON(iomap.length == 0))
71 return -EIO;
Christoph Hellwigae259a92016-06-21 09:23:11 +100072
73 /*
74 * Cut down the length to the one actually provided by the filesystem,
75 * as it might not be able to give us the whole size that we requested.
76 */
77 if (iomap.offset + iomap.length < pos + length)
78 length = iomap.offset + iomap.length - pos;
79
80 /*
81 * Now that we have guaranteed that the space allocation will succeed.
82 * we can do the copy-in page by page without having to worry about
83 * failures exposing transient data.
84 */
85 written = actor(inode, pos, length, data, &iomap);
86
87 /*
88 * Now the data has been copied, commit the range we've copied. This
89 * should not fail unless the filesystem has had a fatal error.
90 */
Christoph Hellwigf20ac7a2016-08-17 08:42:34 +100091 if (ops->iomap_end) {
92 ret = ops->iomap_end(inode, pos, length,
93 written > 0 ? written : 0,
94 flags, &iomap);
95 }
Christoph Hellwigae259a92016-06-21 09:23:11 +100096
97 return written ? written : ret;
98}
99
Christoph Hellwig57fc5052018-06-01 09:03:08 -0700100static sector_t
101iomap_sector(struct iomap *iomap, loff_t pos)
102{
103 return (iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT;
104}
105
Christoph Hellwigae259a92016-06-21 09:23:11 +1000106static void
107iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
108{
109 loff_t i_size = i_size_read(inode);
110
111 /*
112 * Only truncate newly allocated pages beyoned EOF, even if the
113 * write started inside the existing inode size.
114 */
115 if (pos + len > i_size)
116 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
117}
118
119static int
120iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
121 struct page **pagep, struct iomap *iomap)
122{
123 pgoff_t index = pos >> PAGE_SHIFT;
124 struct page *page;
125 int status = 0;
126
127 BUG_ON(pos + len > iomap->offset + iomap->length);
128
Michal Hockod1908f52017-02-03 13:13:26 -0800129 if (fatal_signal_pending(current))
130 return -EINTR;
131
Christoph Hellwigae259a92016-06-21 09:23:11 +1000132 page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
133 if (!page)
134 return -ENOMEM;
135
136 status = __block_write_begin_int(page, pos, len, NULL, iomap);
137 if (unlikely(status)) {
138 unlock_page(page);
139 put_page(page);
140 page = NULL;
141
142 iomap_write_failed(inode, pos, len);
143 }
144
145 *pagep = page;
146 return status;
147}
148
149static int
150iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
151 unsigned copied, struct page *page)
152{
153 int ret;
154
155 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
156 copied, page, NULL);
157 if (ret < len)
158 iomap_write_failed(inode, pos, len);
159 return ret;
160}
161
162static loff_t
163iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
164 struct iomap *iomap)
165{
166 struct iov_iter *i = data;
167 long status = 0;
168 ssize_t written = 0;
169 unsigned int flags = AOP_FLAG_NOFS;
170
Christoph Hellwigae259a92016-06-21 09:23:11 +1000171 do {
172 struct page *page;
173 unsigned long offset; /* Offset into pagecache page */
174 unsigned long bytes; /* Bytes to write to page */
175 size_t copied; /* Bytes copied from user */
176
177 offset = (pos & (PAGE_SIZE - 1));
178 bytes = min_t(unsigned long, PAGE_SIZE - offset,
179 iov_iter_count(i));
180again:
181 if (bytes > length)
182 bytes = length;
183
184 /*
185 * Bring in the user page that we will copy from _first_.
186 * Otherwise there's a nasty deadlock on copying from the
187 * same page as we're writing to, without it being marked
188 * up-to-date.
189 *
190 * Not only is this an optimisation, but it is also required
191 * to check that the address is actually valid, when atomic
192 * usercopies are used, below.
193 */
194 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
195 status = -EFAULT;
196 break;
197 }
198
199 status = iomap_write_begin(inode, pos, bytes, flags, &page,
200 iomap);
201 if (unlikely(status))
202 break;
203
204 if (mapping_writably_mapped(inode->i_mapping))
205 flush_dcache_page(page);
206
Christoph Hellwigae259a92016-06-21 09:23:11 +1000207 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000208
209 flush_dcache_page(page);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000210
211 status = iomap_write_end(inode, pos, bytes, copied, page);
212 if (unlikely(status < 0))
213 break;
214 copied = status;
215
216 cond_resched();
217
218 iov_iter_advance(i, copied);
219 if (unlikely(copied == 0)) {
220 /*
221 * If we were unable to copy any data at all, we must
222 * fall back to a single segment length write.
223 *
224 * If we didn't fallback here, we could livelock
225 * because not all segments in the iov can be copied at
226 * once without a pagefault.
227 */
228 bytes = min_t(unsigned long, PAGE_SIZE - offset,
229 iov_iter_single_seg_count(i));
230 goto again;
231 }
232 pos += copied;
233 written += copied;
234 length -= copied;
235
236 balance_dirty_pages_ratelimited(inode->i_mapping);
237 } while (iov_iter_count(i) && length);
238
239 return written ? written : status;
240}
241
242ssize_t
243iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -0800244 const struct iomap_ops *ops)
Christoph Hellwigae259a92016-06-21 09:23:11 +1000245{
246 struct inode *inode = iocb->ki_filp->f_mapping->host;
247 loff_t pos = iocb->ki_pos, ret = 0, written = 0;
248
249 while (iov_iter_count(iter)) {
250 ret = iomap_apply(inode, pos, iov_iter_count(iter),
251 IOMAP_WRITE, ops, iter, iomap_write_actor);
252 if (ret <= 0)
253 break;
254 pos += ret;
255 written += ret;
256 }
257
258 return written ? written : ret;
259}
260EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
261
Christoph Hellwig5f4e5752016-09-19 10:12:45 +1000262static struct page *
263__iomap_read_page(struct inode *inode, loff_t offset)
264{
265 struct address_space *mapping = inode->i_mapping;
266 struct page *page;
267
268 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
269 if (IS_ERR(page))
270 return page;
271 if (!PageUptodate(page)) {
272 put_page(page);
273 return ERR_PTR(-EIO);
274 }
275 return page;
276}
277
278static loff_t
279iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
280 struct iomap *iomap)
281{
282 long status = 0;
283 ssize_t written = 0;
284
285 do {
286 struct page *page, *rpage;
287 unsigned long offset; /* Offset into pagecache page */
288 unsigned long bytes; /* Bytes to write to page */
289
290 offset = (pos & (PAGE_SIZE - 1));
Christoph Hellwige28ae8e2017-08-11 12:45:35 -0700291 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
Christoph Hellwig5f4e5752016-09-19 10:12:45 +1000292
293 rpage = __iomap_read_page(inode, pos);
294 if (IS_ERR(rpage))
295 return PTR_ERR(rpage);
296
297 status = iomap_write_begin(inode, pos, bytes,
Tetsuo Handac718a972017-05-08 15:58:59 -0700298 AOP_FLAG_NOFS, &page, iomap);
Christoph Hellwig5f4e5752016-09-19 10:12:45 +1000299 put_page(rpage);
300 if (unlikely(status))
301 return status;
302
303 WARN_ON_ONCE(!PageUptodate(page));
304
305 status = iomap_write_end(inode, pos, bytes, bytes, page);
306 if (unlikely(status <= 0)) {
307 if (WARN_ON_ONCE(status == 0))
308 return -EIO;
309 return status;
310 }
311
312 cond_resched();
313
314 pos += status;
315 written += status;
316 length -= status;
317
318 balance_dirty_pages_ratelimited(inode->i_mapping);
319 } while (length);
320
321 return written;
322}
323
324int
325iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -0800326 const struct iomap_ops *ops)
Christoph Hellwig5f4e5752016-09-19 10:12:45 +1000327{
328 loff_t ret;
329
330 while (len) {
331 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
332 iomap_dirty_actor);
333 if (ret <= 0)
334 return ret;
335 pos += ret;
336 len -= ret;
337 }
338
339 return 0;
340}
341EXPORT_SYMBOL_GPL(iomap_file_dirty);
342
Christoph Hellwigae259a92016-06-21 09:23:11 +1000343static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
344 unsigned bytes, struct iomap *iomap)
345{
346 struct page *page;
347 int status;
348
Tetsuo Handac718a972017-05-08 15:58:59 -0700349 status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
350 iomap);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000351 if (status)
352 return status;
353
354 zero_user(page, offset, bytes);
355 mark_page_accessed(page);
356
357 return iomap_write_end(inode, pos, bytes, bytes, page);
358}
359
Christoph Hellwig9a286f02016-06-21 09:31:39 +1000360static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
361 struct iomap *iomap)
362{
Christoph Hellwig57fc5052018-06-01 09:03:08 -0700363 return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
364 iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
Christoph Hellwig9a286f02016-06-21 09:31:39 +1000365}
366
Christoph Hellwigae259a92016-06-21 09:23:11 +1000367static loff_t
368iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
369 void *data, struct iomap *iomap)
370{
371 bool *did_zero = data;
372 loff_t written = 0;
373 int status;
374
375 /* already zeroed? we're done. */
376 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
377 return count;
378
379 do {
380 unsigned offset, bytes;
381
382 offset = pos & (PAGE_SIZE - 1); /* Within page */
Christoph Hellwige28ae8e2017-08-11 12:45:35 -0700383 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000384
Christoph Hellwig9a286f02016-06-21 09:31:39 +1000385 if (IS_DAX(inode))
386 status = iomap_dax_zero(pos, offset, bytes, iomap);
387 else
388 status = iomap_zero(inode, pos, offset, bytes, iomap);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000389 if (status < 0)
390 return status;
391
392 pos += bytes;
393 count -= bytes;
394 written += bytes;
395 if (did_zero)
396 *did_zero = true;
397 } while (count > 0);
398
399 return written;
400}
401
402int
403iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -0800404 const struct iomap_ops *ops)
Christoph Hellwigae259a92016-06-21 09:23:11 +1000405{
406 loff_t ret;
407
408 while (len > 0) {
409 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
410 ops, did_zero, iomap_zero_range_actor);
411 if (ret <= 0)
412 return ret;
413
414 pos += ret;
415 len -= ret;
416 }
417
418 return 0;
419}
420EXPORT_SYMBOL_GPL(iomap_zero_range);
421
422int
423iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -0800424 const struct iomap_ops *ops)
Christoph Hellwigae259a92016-06-21 09:23:11 +1000425{
Fabian Frederick93407472017-02-27 14:28:32 -0800426 unsigned int blocksize = i_blocksize(inode);
427 unsigned int off = pos & (blocksize - 1);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000428
429 /* Block boundary? Nothing to do */
430 if (!off)
431 return 0;
432 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
433}
434EXPORT_SYMBOL_GPL(iomap_truncate_page);
435
436static loff_t
437iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
438 void *data, struct iomap *iomap)
439{
440 struct page *page = data;
441 int ret;
442
Jan Karac663e292016-10-24 14:20:25 +1100443 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000444 if (ret)
445 return ret;
446
447 block_commit_write(page, 0, length);
448 return length;
449}
450
Dave Jiang11bac802017-02-24 14:56:41 -0800451int iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
Christoph Hellwigae259a92016-06-21 09:23:11 +1000452{
453 struct page *page = vmf->page;
Dave Jiang11bac802017-02-24 14:56:41 -0800454 struct inode *inode = file_inode(vmf->vma->vm_file);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000455 unsigned long length;
456 loff_t offset, size;
457 ssize_t ret;
458
459 lock_page(page);
460 size = i_size_read(inode);
461 if ((page->mapping != inode->i_mapping) ||
462 (page_offset(page) > size)) {
463 /* We overload EFAULT to mean page got truncated */
464 ret = -EFAULT;
465 goto out_unlock;
466 }
467
468 /* page is wholly or partially inside EOF */
469 if (((page->index + 1) << PAGE_SHIFT) > size)
470 length = size & ~PAGE_MASK;
471 else
472 length = PAGE_SIZE;
473
474 offset = page_offset(page);
475 while (length > 0) {
Jan Kara9484ab12016-11-10 10:26:50 +1100476 ret = iomap_apply(inode, offset, length,
477 IOMAP_WRITE | IOMAP_FAULT, ops, page,
478 iomap_page_mkwrite_actor);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000479 if (unlikely(ret <= 0))
480 goto out_unlock;
481 offset += ret;
482 length -= ret;
483 }
484
485 set_page_dirty(page);
486 wait_for_stable_page(page);
Christoph Hellwige7647fb2017-08-29 10:08:41 -0700487 return VM_FAULT_LOCKED;
Christoph Hellwigae259a92016-06-21 09:23:11 +1000488out_unlock:
489 unlock_page(page);
Christoph Hellwige7647fb2017-08-29 10:08:41 -0700490 return block_page_mkwrite_return(ret);
Christoph Hellwigae259a92016-06-21 09:23:11 +1000491}
492EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000493
494struct fiemap_ctx {
495 struct fiemap_extent_info *fi;
496 struct iomap prev;
497};
498
499static int iomap_to_fiemap(struct fiemap_extent_info *fi,
500 struct iomap *iomap, u32 flags)
501{
502 switch (iomap->type) {
503 case IOMAP_HOLE:
504 /* skip holes */
505 return 0;
506 case IOMAP_DELALLOC:
507 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
508 break;
Christoph Hellwig19319b52018-06-01 09:03:06 -0700509 case IOMAP_MAPPED:
510 break;
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000511 case IOMAP_UNWRITTEN:
512 flags |= FIEMAP_EXTENT_UNWRITTEN;
513 break;
Christoph Hellwig19319b52018-06-01 09:03:06 -0700514 case IOMAP_INLINE:
515 flags |= FIEMAP_EXTENT_DATA_INLINE;
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000516 break;
517 }
518
Christoph Hellwig17de0a92016-08-29 11:33:58 +1000519 if (iomap->flags & IOMAP_F_MERGED)
520 flags |= FIEMAP_EXTENT_MERGED;
Darrick J. Wonge43c4602016-09-19 10:13:02 +1000521 if (iomap->flags & IOMAP_F_SHARED)
522 flags |= FIEMAP_EXTENT_SHARED;
Christoph Hellwig17de0a92016-08-29 11:33:58 +1000523
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000524 return fiemap_fill_next_extent(fi, iomap->offset,
Andreas Gruenbacher19fe5f62017-10-01 17:55:54 -0400525 iomap->addr != IOMAP_NULL_ADDR ? iomap->addr : 0,
Christoph Hellwig17de0a92016-08-29 11:33:58 +1000526 iomap->length, flags);
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000527}
528
529static loff_t
530iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
531 struct iomap *iomap)
532{
533 struct fiemap_ctx *ctx = data;
534 loff_t ret = length;
535
536 if (iomap->type == IOMAP_HOLE)
537 return length;
538
539 ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
540 ctx->prev = *iomap;
541 switch (ret) {
542 case 0: /* success */
543 return length;
544 case 1: /* extent array full */
545 return 0;
546 default:
547 return ret;
548 }
549}
550
551int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -0800552 loff_t start, loff_t len, const struct iomap_ops *ops)
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000553{
554 struct fiemap_ctx ctx;
555 loff_t ret;
556
557 memset(&ctx, 0, sizeof(ctx));
558 ctx.fi = fi;
559 ctx.prev.type = IOMAP_HOLE;
560
561 ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
562 if (ret)
563 return ret;
564
Dave Chinner8896b8f2016-08-17 08:41:10 +1000565 if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
566 ret = filemap_write_and_wait(inode->i_mapping);
567 if (ret)
568 return ret;
569 }
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000570
571 while (len > 0) {
Christoph Hellwigd33fd772016-10-20 15:51:28 +1100572 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000573 iomap_fiemap_actor);
Dave Chinnerac2dc052016-08-17 08:41:34 +1000574 /* inode with no (attribute) mapping will give ENOENT */
575 if (ret == -ENOENT)
576 break;
Christoph Hellwig8be9f562016-06-21 09:38:45 +1000577 if (ret < 0)
578 return ret;
579 if (ret == 0)
580 break;
581
582 start += ret;
583 len -= ret;
584 }
585
586 if (ctx.prev.type != IOMAP_HOLE) {
587 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
588 if (ret < 0)
589 return ret;
590 }
591
592 return 0;
593}
594EXPORT_SYMBOL_GPL(iomap_fiemap);
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100595
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700596/*
597 * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700598 * Returns true if found and updates @lastoff to the offset in file.
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700599 */
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700600static bool
601page_seek_hole_data(struct inode *inode, struct page *page, loff_t *lastoff,
602 int whence)
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700603{
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700604 const struct address_space_operations *ops = inode->i_mapping->a_ops;
605 unsigned int bsize = i_blocksize(inode), off;
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700606 bool seek_data = whence == SEEK_DATA;
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700607 loff_t poff = page_offset(page);
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700608
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700609 if (WARN_ON_ONCE(*lastoff >= poff + PAGE_SIZE))
610 return false;
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700611
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700612 if (*lastoff < poff) {
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700613 /*
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700614 * Last offset smaller than the start of the page means we found
615 * a hole:
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700616 */
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700617 if (whence == SEEK_HOLE)
618 return true;
619 *lastoff = poff;
620 }
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700621
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700622 /*
623 * Just check the page unless we can and should check block ranges:
624 */
625 if (bsize == PAGE_SIZE || !ops->is_partially_uptodate)
626 return PageUptodate(page) == seek_data;
627
628 lock_page(page);
629 if (unlikely(page->mapping != inode->i_mapping))
630 goto out_unlock_not_found;
631
632 for (off = 0; off < PAGE_SIZE; off += bsize) {
633 if ((*lastoff & ~PAGE_MASK) >= off + bsize)
634 continue;
635 if (ops->is_partially_uptodate(page, off, bsize) == seek_data) {
636 unlock_page(page);
637 return true;
638 }
639 *lastoff = poff + off + bsize;
640 }
641
642out_unlock_not_found:
643 unlock_page(page);
644 return false;
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700645}
646
647/*
648 * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
649 *
650 * Within unwritten extents, the page cache determines which parts are holes
Christoph Hellwigbd56b3e2018-06-01 09:05:14 -0700651 * and which are data: uptodate buffer heads count as data; everything else
652 * counts as a hole.
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700653 *
654 * Returns the resulting offset on successs, and -ENOENT otherwise.
655 */
656static loff_t
657page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
658 int whence)
659{
660 pgoff_t index = offset >> PAGE_SHIFT;
661 pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
662 loff_t lastoff = offset;
663 struct pagevec pvec;
664
665 if (length <= 0)
666 return -ENOENT;
667
668 pagevec_init(&pvec);
669
670 do {
671 unsigned nr_pages, i;
672
673 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
674 end - 1);
675 if (nr_pages == 0)
676 break;
677
678 for (i = 0; i < nr_pages; i++) {
679 struct page *page = pvec.pages[i];
680
Christoph Hellwigafd9d6a2018-06-01 09:05:15 -0700681 if (page_seek_hole_data(inode, page, &lastoff, whence))
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700682 goto check_range;
Christoph Hellwig8a78cb12018-06-01 09:04:40 -0700683 lastoff = page_offset(page) + PAGE_SIZE;
684 }
685 pagevec_release(&pvec);
686 } while (index < end);
687
688 /* When no page at lastoff and we are not done, we found a hole. */
689 if (whence != SEEK_HOLE)
690 goto not_found;
691
692check_range:
693 if (lastoff < offset + length)
694 goto out;
695not_found:
696 lastoff = -ENOENT;
697out:
698 pagevec_release(&pvec);
699 return lastoff;
700}
701
702
Andreas Gruenbacher0ed3b0d2017-06-29 11:43:21 -0700703static loff_t
704iomap_seek_hole_actor(struct inode *inode, loff_t offset, loff_t length,
705 void *data, struct iomap *iomap)
706{
707 switch (iomap->type) {
708 case IOMAP_UNWRITTEN:
709 offset = page_cache_seek_hole_data(inode, offset, length,
710 SEEK_HOLE);
711 if (offset < 0)
712 return length;
713 /* fall through */
714 case IOMAP_HOLE:
715 *(loff_t *)data = offset;
716 return 0;
717 default:
718 return length;
719 }
720}
721
722loff_t
723iomap_seek_hole(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
724{
725 loff_t size = i_size_read(inode);
726 loff_t length = size - offset;
727 loff_t ret;
728
Darrick J. Wongd6ab17f2017-07-12 10:26:47 -0700729 /* Nothing to be found before or beyond the end of the file. */
730 if (offset < 0 || offset >= size)
Andreas Gruenbacher0ed3b0d2017-06-29 11:43:21 -0700731 return -ENXIO;
732
733 while (length > 0) {
734 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
735 &offset, iomap_seek_hole_actor);
736 if (ret < 0)
737 return ret;
738 if (ret == 0)
739 break;
740
741 offset += ret;
742 length -= ret;
743 }
744
745 return offset;
746}
747EXPORT_SYMBOL_GPL(iomap_seek_hole);
748
749static loff_t
750iomap_seek_data_actor(struct inode *inode, loff_t offset, loff_t length,
751 void *data, struct iomap *iomap)
752{
753 switch (iomap->type) {
754 case IOMAP_HOLE:
755 return length;
756 case IOMAP_UNWRITTEN:
757 offset = page_cache_seek_hole_data(inode, offset, length,
758 SEEK_DATA);
759 if (offset < 0)
760 return length;
761 /*FALLTHRU*/
762 default:
763 *(loff_t *)data = offset;
764 return 0;
765 }
766}
767
768loff_t
769iomap_seek_data(struct inode *inode, loff_t offset, const struct iomap_ops *ops)
770{
771 loff_t size = i_size_read(inode);
772 loff_t length = size - offset;
773 loff_t ret;
774
Darrick J. Wongd6ab17f2017-07-12 10:26:47 -0700775 /* Nothing to be found before or beyond the end of the file. */
776 if (offset < 0 || offset >= size)
Andreas Gruenbacher0ed3b0d2017-06-29 11:43:21 -0700777 return -ENXIO;
778
779 while (length > 0) {
780 ret = iomap_apply(inode, offset, length, IOMAP_REPORT, ops,
781 &offset, iomap_seek_data_actor);
782 if (ret < 0)
783 return ret;
784 if (ret == 0)
785 break;
786
787 offset += ret;
788 length -= ret;
789 }
790
791 if (length <= 0)
792 return -ENXIO;
793 return offset;
794}
795EXPORT_SYMBOL_GPL(iomap_seek_data);
796
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100797/*
798 * Private flags for iomap_dio, must not overlap with the public ones in
799 * iomap.h:
800 */
Dave Chinner3460cac2018-05-02 12:54:53 -0700801#define IOMAP_DIO_WRITE_FUA (1 << 28)
Dave Chinner4f8ff442018-05-02 12:54:52 -0700802#define IOMAP_DIO_NEED_SYNC (1 << 29)
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100803#define IOMAP_DIO_WRITE (1 << 30)
804#define IOMAP_DIO_DIRTY (1 << 31)
805
806struct iomap_dio {
807 struct kiocb *iocb;
808 iomap_dio_end_io_t *end_io;
809 loff_t i_size;
810 loff_t size;
811 atomic_t ref;
812 unsigned flags;
813 int error;
814
815 union {
816 /* used during submission and for synchronous completion: */
817 struct {
818 struct iov_iter *iter;
819 struct task_struct *waiter;
820 struct request_queue *last_queue;
821 blk_qc_t cookie;
822 } submit;
823
824 /* used for aio completion: */
825 struct {
826 struct work_struct work;
827 } aio;
828 };
829};
830
831static ssize_t iomap_dio_complete(struct iomap_dio *dio)
832{
833 struct kiocb *iocb = dio->iocb;
Lukas Czerner332391a2017-09-21 08:16:29 -0600834 struct inode *inode = file_inode(iocb->ki_filp);
Eryu Guan5e25c262017-10-13 09:47:46 -0700835 loff_t offset = iocb->ki_pos;
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100836 ssize_t ret;
837
838 if (dio->end_io) {
839 ret = dio->end_io(iocb,
840 dio->error ? dio->error : dio->size,
841 dio->flags);
842 } else {
843 ret = dio->error;
844 }
845
846 if (likely(!ret)) {
847 ret = dio->size;
848 /* check for short read */
Eryu Guan5e25c262017-10-13 09:47:46 -0700849 if (offset + ret > dio->i_size &&
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100850 !(dio->flags & IOMAP_DIO_WRITE))
Eryu Guan5e25c262017-10-13 09:47:46 -0700851 ret = dio->i_size - offset;
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100852 iocb->ki_pos += ret;
853 }
854
Eryu Guan5e25c262017-10-13 09:47:46 -0700855 /*
856 * Try again to invalidate clean pages which might have been cached by
857 * non-direct readahead, or faulted in by get_user_pages() if the source
858 * of the write was an mmap'ed region of the file we're writing. Either
859 * one is a pretty crazy thing to do, so we don't support it 100%. If
860 * this invalidation fails, tough, the write still worked...
861 *
862 * And this page cache invalidation has to be after dio->end_io(), as
863 * some filesystems convert unwritten extents to real allocations in
864 * end_io() when necessary, otherwise a racing buffer read would cache
865 * zeros from unwritten extents.
866 */
867 if (!dio->error &&
868 (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
869 int err;
870 err = invalidate_inode_pages2_range(inode->i_mapping,
871 offset >> PAGE_SHIFT,
872 (offset + dio->size - 1) >> PAGE_SHIFT);
Darrick J. Wong5a9d9292018-01-08 10:41:39 -0800873 if (err)
874 dio_warn_stale_pagecache(iocb->ki_filp);
Eryu Guan5e25c262017-10-13 09:47:46 -0700875 }
876
Dave Chinner4f8ff442018-05-02 12:54:52 -0700877 /*
878 * If this is a DSYNC write, make sure we push it to stable storage now
879 * that we've written data.
880 */
881 if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
882 ret = generic_write_sync(iocb, ret);
883
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100884 inode_dio_end(file_inode(iocb->ki_filp));
885 kfree(dio);
886
887 return ret;
888}
889
890static void iomap_dio_complete_work(struct work_struct *work)
891{
892 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
893 struct kiocb *iocb = dio->iocb;
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100894
Dave Chinner4f8ff442018-05-02 12:54:52 -0700895 iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100896}
897
898/*
899 * Set an error in the dio if none is set yet. We have to use cmpxchg
900 * as the submission context and the completion context(s) can race to
901 * update the error.
902 */
903static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
904{
905 cmpxchg(&dio->error, 0, ret);
906}
907
908static void iomap_dio_bio_end_io(struct bio *bio)
909{
910 struct iomap_dio *dio = bio->bi_private;
911 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
912
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200913 if (bio->bi_status)
914 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100915
916 if (atomic_dec_and_test(&dio->ref)) {
917 if (is_sync_kiocb(dio->iocb)) {
918 struct task_struct *waiter = dio->submit.waiter;
919
920 WRITE_ONCE(dio->submit.waiter, NULL);
921 wake_up_process(waiter);
922 } else if (dio->flags & IOMAP_DIO_WRITE) {
923 struct inode *inode = file_inode(dio->iocb->ki_filp);
924
925 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
926 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
927 } else {
928 iomap_dio_complete_work(&dio->aio.work);
929 }
930 }
931
932 if (should_dirty) {
933 bio_check_pages_dirty(bio);
934 } else {
935 struct bio_vec *bvec;
936 int i;
937
938 bio_for_each_segment_all(bvec, bio, i)
939 put_page(bvec->bv_page);
940 bio_put(bio);
941 }
942}
943
944static blk_qc_t
945iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
946 unsigned len)
947{
948 struct page *page = ZERO_PAGE(0);
949 struct bio *bio;
950
951 bio = bio_alloc(GFP_KERNEL, 1);
Christoph Hellwig74d46992017-08-23 19:10:32 +0200952 bio_set_dev(bio, iomap->bdev);
Christoph Hellwig57fc5052018-06-01 09:03:08 -0700953 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100954 bio->bi_private = dio;
955 bio->bi_end_io = iomap_dio_bio_end_io;
956
957 get_page(page);
Christoph Hellwig6533b4e2018-06-01 09:03:07 -0700958 __bio_add_page(bio, page, len, 0);
Linus Torvalds5cc60ae2016-12-14 21:35:31 -0800959 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100960
961 atomic_inc(&dio->ref);
962 return submit_bio(bio);
963}
964
965static loff_t
966iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
967 void *data, struct iomap *iomap)
968{
969 struct iomap_dio *dio = data;
Fabian Frederick93407472017-02-27 14:28:32 -0800970 unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
971 unsigned int fs_block_size = i_blocksize(inode), pad;
972 unsigned int align = iov_iter_alignment(dio->submit.iter);
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100973 struct iov_iter iter;
974 struct bio *bio;
975 bool need_zeroout = false;
Dave Chinner3460cac2018-05-02 12:54:53 -0700976 bool use_fua = false;
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100977 int nr_pages, ret;
Al Virocfe057f2017-09-11 21:17:09 +0100978 size_t copied = 0;
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100979
980 if ((pos | length | align) & ((1 << blkbits) - 1))
981 return -EINVAL;
982
983 switch (iomap->type) {
984 case IOMAP_HOLE:
985 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
986 return -EIO;
987 /*FALLTHRU*/
988 case IOMAP_UNWRITTEN:
989 if (!(dio->flags & IOMAP_DIO_WRITE)) {
Al Virocfe057f2017-09-11 21:17:09 +0100990 length = iov_iter_zero(length, dio->submit.iter);
Christoph Hellwigff6a9292016-11-30 14:36:01 +1100991 dio->size += length;
992 return length;
993 }
994 dio->flags |= IOMAP_DIO_UNWRITTEN;
995 need_zeroout = true;
996 break;
997 case IOMAP_MAPPED:
998 if (iomap->flags & IOMAP_F_SHARED)
999 dio->flags |= IOMAP_DIO_COW;
Dave Chinner3460cac2018-05-02 12:54:53 -07001000 if (iomap->flags & IOMAP_F_NEW) {
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001001 need_zeroout = true;
Dave Chinner3460cac2018-05-02 12:54:53 -07001002 } else {
1003 /*
1004 * Use a FUA write if we need datasync semantics, this
1005 * is a pure data IO that doesn't require any metadata
1006 * updates and the underlying device supports FUA. This
1007 * allows us to avoid cache flushes on IO completion.
1008 */
1009 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
1010 (dio->flags & IOMAP_DIO_WRITE_FUA) &&
1011 blk_queue_fua(bdev_get_queue(iomap->bdev)))
1012 use_fua = true;
1013 }
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001014 break;
1015 default:
1016 WARN_ON_ONCE(1);
1017 return -EIO;
1018 }
1019
1020 /*
1021 * Operate on a partial iter trimmed to the extent we were called for.
1022 * We'll update the iter in the dio once we're done with this extent.
1023 */
1024 iter = *dio->submit.iter;
1025 iov_iter_truncate(&iter, length);
1026
1027 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1028 if (nr_pages <= 0)
1029 return nr_pages;
1030
1031 if (need_zeroout) {
1032 /* zero out from the start of the block to the write offset */
1033 pad = pos & (fs_block_size - 1);
1034 if (pad)
1035 iomap_dio_zero(dio, iomap, pos - pad, pad);
1036 }
1037
1038 do {
Al Virocfe057f2017-09-11 21:17:09 +01001039 size_t n;
1040 if (dio->error) {
1041 iov_iter_revert(dio->submit.iter, copied);
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001042 return 0;
Al Virocfe057f2017-09-11 21:17:09 +01001043 }
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001044
1045 bio = bio_alloc(GFP_KERNEL, nr_pages);
Christoph Hellwig74d46992017-08-23 19:10:32 +02001046 bio_set_dev(bio, iomap->bdev);
Christoph Hellwig57fc5052018-06-01 09:03:08 -07001047 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
Jens Axboe45d06cf2017-06-27 11:01:22 -06001048 bio->bi_write_hint = dio->iocb->ki_hint;
Adam Manzanares087e5662018-05-22 10:52:21 -07001049 bio->bi_ioprio = dio->iocb->ki_ioprio;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001050 bio->bi_private = dio;
1051 bio->bi_end_io = iomap_dio_bio_end_io;
1052
1053 ret = bio_iov_iter_get_pages(bio, &iter);
1054 if (unlikely(ret)) {
1055 bio_put(bio);
Al Virocfe057f2017-09-11 21:17:09 +01001056 return copied ? copied : ret;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001057 }
1058
Al Virocfe057f2017-09-11 21:17:09 +01001059 n = bio->bi_iter.bi_size;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001060 if (dio->flags & IOMAP_DIO_WRITE) {
Dave Chinner3460cac2018-05-02 12:54:53 -07001061 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1062 if (use_fua)
1063 bio->bi_opf |= REQ_FUA;
1064 else
1065 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
Al Virocfe057f2017-09-11 21:17:09 +01001066 task_io_account_write(n);
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001067 } else {
Dave Chinner3460cac2018-05-02 12:54:53 -07001068 bio->bi_opf = REQ_OP_READ;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001069 if (dio->flags & IOMAP_DIO_DIRTY)
1070 bio_set_pages_dirty(bio);
1071 }
1072
Al Virocfe057f2017-09-11 21:17:09 +01001073 iov_iter_advance(dio->submit.iter, n);
1074
1075 dio->size += n;
1076 pos += n;
1077 copied += n;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001078
1079 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
1080
1081 atomic_inc(&dio->ref);
1082
1083 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
1084 dio->submit.cookie = submit_bio(bio);
1085 } while (nr_pages);
1086
1087 if (need_zeroout) {
1088 /* zero out from the end of the write to the end of the block */
1089 pad = pos & (fs_block_size - 1);
1090 if (pad)
1091 iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
1092 }
Al Virocfe057f2017-09-11 21:17:09 +01001093 return copied;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001094}
1095
Dave Chinner4f8ff442018-05-02 12:54:52 -07001096/*
1097 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
Dave Chinner3460cac2018-05-02 12:54:53 -07001098 * is being issued as AIO or not. This allows us to optimise pure data writes
1099 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
1100 * REQ_FLUSH post write. This is slightly tricky because a single request here
1101 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
1102 * may be pure data writes. In that case, we still need to do a full data sync
1103 * completion.
Dave Chinner4f8ff442018-05-02 12:54:52 -07001104 */
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001105ssize_t
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -08001106iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
1107 const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001108{
1109 struct address_space *mapping = iocb->ki_filp->f_mapping;
1110 struct inode *inode = file_inode(iocb->ki_filp);
1111 size_t count = iov_iter_count(iter);
Eryu Guanc771c142017-03-02 15:02:06 -08001112 loff_t pos = iocb->ki_pos, start = pos;
1113 loff_t end = iocb->ki_pos + count - 1, ret = 0;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001114 unsigned int flags = IOMAP_DIRECT;
1115 struct blk_plug plug;
1116 struct iomap_dio *dio;
1117
1118 lockdep_assert_held(&inode->i_rwsem);
1119
1120 if (!count)
1121 return 0;
1122
1123 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1124 if (!dio)
1125 return -ENOMEM;
1126
1127 dio->iocb = iocb;
1128 atomic_set(&dio->ref, 1);
1129 dio->size = 0;
1130 dio->i_size = i_size_read(inode);
1131 dio->end_io = end_io;
1132 dio->error = 0;
1133 dio->flags = 0;
1134
1135 dio->submit.iter = iter;
1136 if (is_sync_kiocb(iocb)) {
1137 dio->submit.waiter = current;
1138 dio->submit.cookie = BLK_QC_T_NONE;
1139 dio->submit.last_queue = NULL;
1140 }
1141
1142 if (iov_iter_rw(iter) == READ) {
1143 if (pos >= dio->i_size)
1144 goto out_free_dio;
1145
1146 if (iter->type == ITER_IOVEC)
1147 dio->flags |= IOMAP_DIO_DIRTY;
1148 } else {
Dave Chinner3460cac2018-05-02 12:54:53 -07001149 flags |= IOMAP_WRITE;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001150 dio->flags |= IOMAP_DIO_WRITE;
Dave Chinner3460cac2018-05-02 12:54:53 -07001151
1152 /* for data sync or sync, we need sync completion processing */
Dave Chinner4f8ff442018-05-02 12:54:52 -07001153 if (iocb->ki_flags & IOCB_DSYNC)
1154 dio->flags |= IOMAP_DIO_NEED_SYNC;
Dave Chinner3460cac2018-05-02 12:54:53 -07001155
1156 /*
1157 * For datasync only writes, we optimistically try using FUA for
1158 * this IO. Any non-FUA write that occurs will clear this flag,
1159 * hence we know before completion whether a cache flush is
1160 * necessary.
1161 */
1162 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
1163 dio->flags |= IOMAP_DIO_WRITE_FUA;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001164 }
1165
Goldwyn Rodriguesa38d1242017-06-20 07:05:45 -05001166 if (iocb->ki_flags & IOCB_NOWAIT) {
1167 if (filemap_range_has_page(mapping, start, end)) {
1168 ret = -EAGAIN;
1169 goto out_free_dio;
1170 }
1171 flags |= IOMAP_NOWAIT;
1172 }
1173
Andrey Ryabinin55635ba2017-05-03 14:55:59 -07001174 ret = filemap_write_and_wait_range(mapping, start, end);
1175 if (ret)
1176 goto out_free_dio;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001177
Darrick J. Wong5a9d9292018-01-08 10:41:39 -08001178 /*
1179 * Try to invalidate cache pages for the range we're direct
1180 * writing. If this invalidation fails, tough, the write will
1181 * still work, but racing two incompatible write paths is a
1182 * pretty crazy thing to do, so we don't support it 100%.
1183 */
Andrey Ryabinin55635ba2017-05-03 14:55:59 -07001184 ret = invalidate_inode_pages2_range(mapping,
1185 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
Darrick J. Wong5a9d9292018-01-08 10:41:39 -08001186 if (ret)
1187 dio_warn_stale_pagecache(iocb->ki_filp);
Andrey Ryabinin55635ba2017-05-03 14:55:59 -07001188 ret = 0;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001189
Chandan Rajendra546e7be2017-09-22 11:47:33 -07001190 if (iov_iter_rw(iter) == WRITE && !is_sync_kiocb(iocb) &&
1191 !inode->i_sb->s_dio_done_wq) {
1192 ret = sb_init_dio_done_wq(inode->i_sb);
1193 if (ret < 0)
1194 goto out_free_dio;
1195 }
1196
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001197 inode_dio_begin(inode);
1198
1199 blk_start_plug(&plug);
1200 do {
1201 ret = iomap_apply(inode, pos, count, flags, ops, dio,
1202 iomap_dio_actor);
1203 if (ret <= 0) {
1204 /* magic error code to fall back to buffered I/O */
1205 if (ret == -ENOTBLK)
1206 ret = 0;
1207 break;
1208 }
1209 pos += ret;
Chandan Rajendraa008c312017-04-12 11:03:20 -07001210
1211 if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
1212 break;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001213 } while ((count = iov_iter_count(iter)) > 0);
1214 blk_finish_plug(&plug);
1215
1216 if (ret < 0)
1217 iomap_dio_set_error(dio, ret);
1218
Dave Chinner3460cac2018-05-02 12:54:53 -07001219 /*
1220 * If all the writes we issued were FUA, we don't need to flush the
1221 * cache on IO completion. Clear the sync flag for this case.
1222 */
1223 if (dio->flags & IOMAP_DIO_WRITE_FUA)
1224 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
1225
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001226 if (!atomic_dec_and_test(&dio->ref)) {
1227 if (!is_sync_kiocb(iocb))
1228 return -EIOCBQUEUED;
1229
1230 for (;;) {
1231 set_current_state(TASK_UNINTERRUPTIBLE);
1232 if (!READ_ONCE(dio->submit.waiter))
1233 break;
1234
1235 if (!(iocb->ki_flags & IOCB_HIPRI) ||
1236 !dio->submit.last_queue ||
Christoph Hellwigea435e12017-11-02 21:29:54 +03001237 !blk_poll(dio->submit.last_queue,
Linus Torvalds5cc60ae2016-12-14 21:35:31 -08001238 dio->submit.cookie))
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001239 io_schedule();
1240 }
1241 __set_current_state(TASK_RUNNING);
1242 }
1243
Eryu Guanc771c142017-03-02 15:02:06 -08001244 ret = iomap_dio_complete(dio);
1245
Eryu Guanc771c142017-03-02 15:02:06 -08001246 return ret;
Christoph Hellwigff6a9292016-11-30 14:36:01 +11001247
1248out_free_dio:
1249 kfree(dio);
1250 return ret;
1251}
1252EXPORT_SYMBOL_GPL(iomap_dio_rw);
Darrick J. Wong67482122018-05-10 08:38:15 -07001253
1254/* Swapfile activation */
1255
1256#ifdef CONFIG_SWAP
1257struct iomap_swapfile_info {
1258 struct iomap iomap; /* accumulated iomap */
1259 struct swap_info_struct *sis;
1260 uint64_t lowest_ppage; /* lowest physical addr seen (pages) */
1261 uint64_t highest_ppage; /* highest physical addr seen (pages) */
1262 unsigned long nr_pages; /* number of pages collected */
1263 int nr_extents; /* extent count */
1264};
1265
1266/*
1267 * Collect physical extents for this swap file. Physical extents reported to
1268 * the swap code must be trimmed to align to a page boundary. The logical
1269 * offset within the file is irrelevant since the swapfile code maps logical
1270 * page numbers of the swap device to the physical page-aligned extents.
1271 */
1272static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
1273{
1274 struct iomap *iomap = &isi->iomap;
1275 unsigned long nr_pages;
1276 uint64_t first_ppage;
1277 uint64_t first_ppage_reported;
1278 uint64_t next_ppage;
1279 int error;
1280
1281 /*
1282 * Round the start up and the end down so that the physical
1283 * extent aligns to a page boundary.
1284 */
1285 first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
1286 next_ppage = ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
1287 PAGE_SHIFT;
1288
1289 /* Skip too-short physical extents. */
1290 if (first_ppage >= next_ppage)
1291 return 0;
1292 nr_pages = next_ppage - first_ppage;
1293
1294 /*
1295 * Calculate how much swap space we're adding; the first page contains
1296 * the swap header and doesn't count. The mm still wants that first
1297 * page fed to add_swap_extent, however.
1298 */
1299 first_ppage_reported = first_ppage;
1300 if (iomap->offset == 0)
1301 first_ppage_reported++;
1302 if (isi->lowest_ppage > first_ppage_reported)
1303 isi->lowest_ppage = first_ppage_reported;
1304 if (isi->highest_ppage < (next_ppage - 1))
1305 isi->highest_ppage = next_ppage - 1;
1306
1307 /* Add extent, set up for the next call. */
1308 error = add_swap_extent(isi->sis, isi->nr_pages, nr_pages, first_ppage);
1309 if (error < 0)
1310 return error;
1311 isi->nr_extents += error;
1312 isi->nr_pages += nr_pages;
1313 return 0;
1314}
1315
1316/*
1317 * Accumulate iomaps for this swap file. We have to accumulate iomaps because
1318 * swap only cares about contiguous page-aligned physical extents and makes no
1319 * distinction between written and unwritten extents.
1320 */
1321static loff_t iomap_swapfile_activate_actor(struct inode *inode, loff_t pos,
1322 loff_t count, void *data, struct iomap *iomap)
1323{
1324 struct iomap_swapfile_info *isi = data;
1325 int error;
1326
Christoph Hellwig19319b52018-06-01 09:03:06 -07001327 switch (iomap->type) {
1328 case IOMAP_MAPPED:
1329 case IOMAP_UNWRITTEN:
1330 /* Only real or unwritten extents. */
1331 break;
1332 case IOMAP_INLINE:
1333 /* No inline data. */
Omar Sandovalec601922018-05-16 11:13:34 -07001334 pr_err("swapon: file is inline\n");
1335 return -EINVAL;
Christoph Hellwig19319b52018-06-01 09:03:06 -07001336 default:
Omar Sandovalec601922018-05-16 11:13:34 -07001337 pr_err("swapon: file has unallocated extents\n");
1338 return -EINVAL;
1339 }
Darrick J. Wong67482122018-05-10 08:38:15 -07001340
Omar Sandovalec601922018-05-16 11:13:34 -07001341 /* No uncommitted metadata or shared blocks. */
1342 if (iomap->flags & IOMAP_F_DIRTY) {
1343 pr_err("swapon: file is not committed\n");
1344 return -EINVAL;
1345 }
1346 if (iomap->flags & IOMAP_F_SHARED) {
1347 pr_err("swapon: file has shared extents\n");
1348 return -EINVAL;
1349 }
Darrick J. Wong67482122018-05-10 08:38:15 -07001350
Omar Sandovalec601922018-05-16 11:13:34 -07001351 /* Only one bdev per swap file. */
1352 if (iomap->bdev != isi->sis->bdev) {
1353 pr_err("swapon: file is on multiple devices\n");
1354 return -EINVAL;
1355 }
Darrick J. Wong67482122018-05-10 08:38:15 -07001356
1357 if (isi->iomap.length == 0) {
1358 /* No accumulated extent, so just store it. */
1359 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
1360 } else if (isi->iomap.addr + isi->iomap.length == iomap->addr) {
1361 /* Append this to the accumulated extent. */
1362 isi->iomap.length += iomap->length;
1363 } else {
1364 /* Otherwise, add the retained iomap and store this one. */
1365 error = iomap_swapfile_add_extent(isi);
1366 if (error)
1367 return error;
1368 memcpy(&isi->iomap, iomap, sizeof(isi->iomap));
1369 }
Darrick J. Wong67482122018-05-10 08:38:15 -07001370 return count;
Darrick J. Wong67482122018-05-10 08:38:15 -07001371}
1372
1373/*
1374 * Iterate a swap file's iomaps to construct physical extents that can be
1375 * passed to the swapfile subsystem.
1376 */
1377int iomap_swapfile_activate(struct swap_info_struct *sis,
1378 struct file *swap_file, sector_t *pagespan,
1379 const struct iomap_ops *ops)
1380{
1381 struct iomap_swapfile_info isi = {
1382 .sis = sis,
1383 .lowest_ppage = (sector_t)-1ULL,
1384 };
1385 struct address_space *mapping = swap_file->f_mapping;
1386 struct inode *inode = mapping->host;
1387 loff_t pos = 0;
1388 loff_t len = ALIGN_DOWN(i_size_read(inode), PAGE_SIZE);
1389 loff_t ret;
1390
Darrick J. Wong117a1482018-06-05 09:53:05 -07001391 /*
1392 * Persist all file mapping metadata so that we won't have any
1393 * IOMAP_F_DIRTY iomaps.
1394 */
1395 ret = vfs_fsync(swap_file, 1);
Darrick J. Wong67482122018-05-10 08:38:15 -07001396 if (ret)
1397 return ret;
1398
1399 while (len > 0) {
1400 ret = iomap_apply(inode, pos, len, IOMAP_REPORT,
1401 ops, &isi, iomap_swapfile_activate_actor);
1402 if (ret <= 0)
1403 return ret;
1404
1405 pos += ret;
1406 len -= ret;
1407 }
1408
1409 if (isi.iomap.length) {
1410 ret = iomap_swapfile_add_extent(&isi);
1411 if (ret)
1412 return ret;
1413 }
1414
1415 *pagespan = 1 + isi.highest_ppage - isi.lowest_ppage;
1416 sis->max = isi.nr_pages;
1417 sis->pages = isi.nr_pages - 1;
1418 sis->highest_bit = isi.nr_pages - 1;
1419 return isi.nr_extents;
1420}
1421EXPORT_SYMBOL_GPL(iomap_swapfile_activate);
1422#endif /* CONFIG_SWAP */
Christoph Hellwig89eb1902018-06-01 09:03:08 -07001423
1424static loff_t
1425iomap_bmap_actor(struct inode *inode, loff_t pos, loff_t length,
1426 void *data, struct iomap *iomap)
1427{
1428 sector_t *bno = data, addr;
1429
1430 if (iomap->type == IOMAP_MAPPED) {
1431 addr = (pos - iomap->offset + iomap->addr) >> inode->i_blkbits;
1432 if (addr > INT_MAX)
1433 WARN(1, "would truncate bmap result\n");
1434 else
1435 *bno = addr;
1436 }
1437 return 0;
1438}
1439
1440/* legacy ->bmap interface. 0 is the error return (!) */
1441sector_t
1442iomap_bmap(struct address_space *mapping, sector_t bno,
1443 const struct iomap_ops *ops)
1444{
1445 struct inode *inode = mapping->host;
1446 loff_t pos = bno >> inode->i_blkbits;
1447 unsigned blocksize = i_blocksize(inode);
1448
1449 if (filemap_write_and_wait(mapping))
1450 return 0;
1451
1452 bno = 0;
1453 iomap_apply(inode, pos, blocksize, 0, ops, &bno, iomap_bmap_actor);
1454 return bno;
1455}
1456EXPORT_SYMBOL_GPL(iomap_bmap);