1/* handling of writes to regular files and writing back to the server
2 *
3 * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
4 * Written by David Howells (dhowells@redhat.com)
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 */
11#include <linux/backing-dev.h>
12#include <linux/slab.h>
13#include <linux/fs.h>
14#include <linux/pagemap.h>
15#include <linux/writeback.h>
16#include <linux/pagevec.h>
17#include "internal.h"
18
19static int afs_write_back_from_locked_page(struct afs_writeback *wb,
20					   struct page *page);
21
22/*
23 * mark a page as having been made dirty and thus needing writeback
24 */
25int afs_set_page_dirty(struct page *page)
26{
27	_enter("");
28	return __set_page_dirty_nobuffers(page);
29}
30
31/*
32 * unlink a writeback record because its usage has reached zero
33 * - must be called with the wb->vnode->writeback_lock held
34 */
35static void afs_unlink_writeback(struct afs_writeback *wb)
36{
37	struct afs_writeback *front;
38	struct afs_vnode *vnode = wb->vnode;
39
40	list_del_init(&wb->link);
41	if (!list_empty(&vnode->writebacks)) {
42		/* if an fsync rises to the front of the queue then wake it
43		 * up */
44		front = list_entry(vnode->writebacks.next,
45				   struct afs_writeback, link);
46		if (front->state == AFS_WBACK_SYNCING) {
47			_debug("wake up sync");
48			front->state = AFS_WBACK_COMPLETE;
49			wake_up(&front->waitq);
50		}
51	}
52}
53
54/*
55 * free a writeback record
56 */
57static void afs_free_writeback(struct afs_writeback *wb)
58{
59	_enter("");
60	key_put(wb->key);
61	kfree(wb);
62}
63
64/*
65 * dispose of a reference to a writeback record
66 */
67void afs_put_writeback(struct afs_writeback *wb)
68{
69	struct afs_vnode *vnode = wb->vnode;
70
71	_enter("{%d}", wb->usage);
72
73	spin_lock(&vnode->writeback_lock);
74	if (--wb->usage == 0)
75		afs_unlink_writeback(wb);
76	else
77		wb = NULL;
78	spin_unlock(&vnode->writeback_lock);
79	if (wb)
80		afs_free_writeback(wb);
81}
82
83/*
84 * partly or wholly fill a page that's under preparation for writing
85 */
86static int afs_fill_page(struct afs_vnode *vnode, struct key *key,
87			 loff_t pos, struct page *page)
88{
89	loff_t i_size;
90	int ret;
91	int len;
92
93	_enter(",,%llu", (unsigned long long)pos);
94
95	i_size = i_size_read(&vnode->vfs_inode);
96	if (pos + PAGE_CACHE_SIZE > i_size)
97		len = i_size - pos;
98	else
99		len = PAGE_CACHE_SIZE;
100
101	ret = afs_vnode_fetch_data(vnode, key, pos, len, page);
102	if (ret < 0) {
103		if (ret == -ENOENT) {
104			_debug("got NOENT from server"
105			       " - marking file deleted and stale");
106			set_bit(AFS_VNODE_DELETED, &vnode->flags);
107			ret = -ESTALE;
108		}
109	}
110
111	_leave(" = %d", ret);
112	return ret;
113}
114
115/*
116 * prepare to perform part of a write to a page
117 */
118int afs_write_begin(struct file *file, struct address_space *mapping,
119		    loff_t pos, unsigned len, unsigned flags,
120		    struct page **pagep, void **fsdata)
121{
122	struct afs_writeback *candidate, *wb;
123	struct afs_vnode *vnode = AFS_FS_I(file_inode(file));
124	struct page *page;
125	struct key *key = file->private_data;
126	unsigned from = pos & (PAGE_CACHE_SIZE - 1);
127	unsigned to = from + len;
128	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
129	int ret;
130
131	_enter("{%x:%u},{%lx},%u,%u",
132	       vnode->fid.vid, vnode->fid.vnode, index, from, to);
133
134	candidate = kzalloc(sizeof(*candidate), GFP_KERNEL);
135	if (!candidate)
136		return -ENOMEM;
137	candidate->vnode = vnode;
138	candidate->first = candidate->last = index;
139	candidate->offset_first = from;
140	candidate->to_last = to;
141	INIT_LIST_HEAD(&candidate->link);
142	candidate->usage = 1;
143	candidate->state = AFS_WBACK_PENDING;
144	init_waitqueue_head(&candidate->waitq);
145
146	page = grab_cache_page_write_begin(mapping, index, flags);
147	if (!page) {
148		kfree(candidate);
149		return -ENOMEM;
150	}
151	*pagep = page;
152	/* page won't leak in error case: it eventually gets cleaned off LRU */
153
154	if (!PageUptodate(page) && len != PAGE_CACHE_SIZE) {
155		ret = afs_fill_page(vnode, key, index << PAGE_CACHE_SHIFT, page);
156		if (ret < 0) {
157			kfree(candidate);
158			_leave(" = %d [prep]", ret);
159			return ret;
160		}
161		SetPageUptodate(page);
162	}
163
164try_again:
165	spin_lock(&vnode->writeback_lock);
166
167	/* see if this page is already pending a writeback under a suitable key
168	 * - if so we can just join onto that one */
169	wb = (struct afs_writeback *) page_private(page);
170	if (wb) {
171		if (wb->key == key && wb->state == AFS_WBACK_PENDING)
172			goto subsume_in_current_wb;
173		goto flush_conflicting_wb;
174	}
175
176	if (index > 0) {
177		/* see if we can find an already pending writeback that we can
178		 * append this page to */
179		list_for_each_entry(wb, &vnode->writebacks, link) {
180			if (wb->last == index - 1 && wb->key == key &&
181			    wb->state == AFS_WBACK_PENDING)
182				goto append_to_previous_wb;
183		}
184	}
185
186	list_add_tail(&candidate->link, &vnode->writebacks);
187	candidate->key = key_get(key);
188	spin_unlock(&vnode->writeback_lock);
189	SetPagePrivate(page);
190	set_page_private(page, (unsigned long) candidate);
191	_leave(" = 0 [new]");
192	return 0;
193
194subsume_in_current_wb:
195	_debug("subsume");
196	ASSERTRANGE(wb->first, <=, index, <=, wb->last);
197	if (index == wb->first && from < wb->offset_first)
198		wb->offset_first = from;
199	if (index == wb->last && to > wb->to_last)
200		wb->to_last = to;
201	spin_unlock(&vnode->writeback_lock);
202	kfree(candidate);
203	_leave(" = 0 [sub]");
204	return 0;
205
206append_to_previous_wb:
207	_debug("append into %lx-%lx", wb->first, wb->last);
208	wb->usage++;
209	wb->last++;
210	wb->to_last = to;
211	spin_unlock(&vnode->writeback_lock);
212	SetPagePrivate(page);
213	set_page_private(page, (unsigned long) wb);
214	kfree(candidate);
215	_leave(" = 0 [app]");
216	return 0;
217
218	/* the page is currently bound to another context, so if it's dirty we
219	 * need to flush it before we can use the new context */
220flush_conflicting_wb:
221	_debug("flush conflict");
222	if (wb->state == AFS_WBACK_PENDING)
223		wb->state = AFS_WBACK_CONFLICTING;
224	spin_unlock(&vnode->writeback_lock);
225	if (PageDirty(page)) {
226		ret = afs_write_back_from_locked_page(wb, page);
227		if (ret < 0) {
228			afs_put_writeback(candidate);
229			_leave(" = %d", ret);
230			return ret;
231		}
232	}
233
234	/* the page holds a ref on the writeback record */
235	afs_put_writeback(wb);
236	set_page_private(page, 0);
237	ClearPagePrivate(page);
238	goto try_again;
239}
240
241/*
242 * finalise part of a write to a page
243 */
244int afs_write_end(struct file *file, struct address_space *mapping,
245		  loff_t pos, unsigned len, unsigned copied,
246		  struct page *page, void *fsdata)
247{
248	struct afs_vnode *vnode = AFS_FS_I(file_inode(file));
249	loff_t i_size, maybe_i_size;
250
251	_enter("{%x:%u},{%lx}",
252	       vnode->fid.vid, vnode->fid.vnode, page->index);
253
254	maybe_i_size = pos + copied;
255
256	i_size = i_size_read(&vnode->vfs_inode);
257	if (maybe_i_size > i_size) {
258		spin_lock(&vnode->writeback_lock);
259		i_size = i_size_read(&vnode->vfs_inode);
260		if (maybe_i_size > i_size)
261			i_size_write(&vnode->vfs_inode, maybe_i_size);
262		spin_unlock(&vnode->writeback_lock);
263	}
264
265	set_page_dirty(page);
266	if (PageDirty(page))
267		_debug("dirtied");
268	unlock_page(page);
269	page_cache_release(page);
270
271	return copied;
272}
273
274/*
275 * kill all the pages in the given range
276 */
277static void afs_kill_pages(struct afs_vnode *vnode, bool error,
278			   pgoff_t first, pgoff_t last)
279{
280	struct pagevec pv;
281	unsigned count, loop;
282
283	_enter("{%x:%u},%lx-%lx",
284	       vnode->fid.vid, vnode->fid.vnode, first, last);
285
286	pagevec_init(&pv, 0);
287
288	do {
289		_debug("kill %lx-%lx", first, last);
290
291		count = last - first + 1;
292		if (count > PAGEVEC_SIZE)
293			count = PAGEVEC_SIZE;
294		pv.nr = find_get_pages_contig(vnode->vfs_inode.i_mapping,
295					      first, count, pv.pages);
296		ASSERTCMP(pv.nr, ==, count);
297
298		for (loop = 0; loop < count; loop++) {
299			ClearPageUptodate(pv.pages[loop]);
300			if (error)
301				SetPageError(pv.pages[loop]);
302			end_page_writeback(pv.pages[loop]);
303		}
304
305		__pagevec_release(&pv);
306	} while (first < last);
307
308	_leave("");
309}
310
311/*
312 * synchronously write back the locked page and any subsequent non-locked dirty
313 * pages also covered by the same writeback record
314 */
315static int afs_write_back_from_locked_page(struct afs_writeback *wb,
316					   struct page *primary_page)
317{
318	struct page *pages[8], *page;
319	unsigned long count;
320	unsigned n, offset, to;
321	pgoff_t start, first, last;
322	int loop, ret;
323
324	_enter(",%lx", primary_page->index);
325
326	count = 1;
327	if (!clear_page_dirty_for_io(primary_page))
328		BUG();
329	if (test_set_page_writeback(primary_page))
330		BUG();
331
332	/* find all consecutive lockable dirty pages, stopping when we find a
333	 * page that is not immediately lockable, is not dirty or is missing,
334	 * or we reach the end of the range */
335	start = primary_page->index;
336	if (start >= wb->last)
337		goto no_more;
338	start++;
339	do {
340		_debug("more %lx [%lx]", start, count);
341		n = wb->last - start + 1;
342		if (n > ARRAY_SIZE(pages))
343			n = ARRAY_SIZE(pages);
344		n = find_get_pages_contig(wb->vnode->vfs_inode.i_mapping,
345					  start, n, pages);
346		_debug("fgpc %u", n);
347		if (n == 0)
348			goto no_more;
349		if (pages[0]->index != start) {
350			do {
351				put_page(pages[--n]);
352			} while (n > 0);
353			goto no_more;
354		}
355
356		for (loop = 0; loop < n; loop++) {
357			page = pages[loop];
358			if (page->index > wb->last)
359				break;
360			if (!trylock_page(page))
361				break;
362			if (!PageDirty(page) ||
363			    page_private(page) != (unsigned long) wb) {
364				unlock_page(page);
365				break;
366			}
367			if (!clear_page_dirty_for_io(page))
368				BUG();
369			if (test_set_page_writeback(page))
370				BUG();
371			unlock_page(page);
372			put_page(page);
373		}
374		count += loop;
375		if (loop < n) {
376			for (; loop < n; loop++)
377				put_page(pages[loop]);
378			goto no_more;
379		}
380
381		start += loop;
382	} while (start <= wb->last && count < 65536);
383
384no_more:
385	/* we now have a contiguous set of dirty pages, each with writeback set
386	 * and the dirty mark cleared; the first page is locked and must remain
387	 * so, all the rest are unlocked */
388	first = primary_page->index;
389	last = first + count - 1;
390
391	offset = (first == wb->first) ? wb->offset_first : 0;
392	to = (last == wb->last) ? wb->to_last : PAGE_SIZE;
393
394	_debug("write back %lx[%u..] to %lx[..%u]", first, offset, last, to);
395
396	ret = afs_vnode_store_data(wb, first, last, offset, to);
397	if (ret < 0) {
398		switch (ret) {
399		case -EDQUOT:
400		case -ENOSPC:
401			set_bit(AS_ENOSPC,
402				&wb->vnode->vfs_inode.i_mapping->flags);
403			break;
404		case -EROFS:
405		case -EIO:
406		case -EREMOTEIO:
407		case -EFBIG:
408		case -ENOENT:
409		case -ENOMEDIUM:
410		case -ENXIO:
411			afs_kill_pages(wb->vnode, true, first, last);
412			set_bit(AS_EIO, &wb->vnode->vfs_inode.i_mapping->flags);
413			break;
414		case -EACCES:
415		case -EPERM:
416		case -ENOKEY:
417		case -EKEYEXPIRED:
418		case -EKEYREJECTED:
419		case -EKEYREVOKED:
420			afs_kill_pages(wb->vnode, false, first, last);
421			break;
422		default:
423			break;
424		}
425	} else {
426		ret = count;
427	}
428
429	_leave(" = %d", ret);
430	return ret;
431}
432
433/*
434 * write a page back to the server
435 * - the caller locked the page for us
436 */
437int afs_writepage(struct page *page, struct writeback_control *wbc)
438{
439	struct afs_writeback *wb;
440	int ret;
441
442	_enter("{%lx},", page->index);
443
444	wb = (struct afs_writeback *) page_private(page);
445	ASSERT(wb != NULL);
446
447	ret = afs_write_back_from_locked_page(wb, page);
448	unlock_page(page);
449	if (ret < 0) {
450		_leave(" = %d", ret);
451		return 0;
452	}
453
454	wbc->nr_to_write -= ret;
455
456	_leave(" = 0");
457	return 0;
458}
459
460/*
461 * write a region of pages back to the server
462 */
463static int afs_writepages_region(struct address_space *mapping,
464				 struct writeback_control *wbc,
465				 pgoff_t index, pgoff_t end, pgoff_t *_next)
466{
467	struct afs_writeback *wb;
468	struct page *page;
469	int ret, n;
470
471	_enter(",,%lx,%lx,", index, end);
472
473	do {
474		n = find_get_pages_tag(mapping, &index, PAGECACHE_TAG_DIRTY,
475				       1, &page);
476		if (!n)
477			break;
478
479		_debug("wback %lx", page->index);
480
481		if (page->index > end) {
482			*_next = index;
483			page_cache_release(page);
484			_leave(" = 0 [%lx]", *_next);
485			return 0;
486		}
487
488		/* at this point we hold neither mapping->tree_lock nor lock on
489		 * the page itself: the page may be truncated or invalidated
490		 * (changing page->mapping to NULL), or even swizzled back from
491		 * swapper_space to tmpfs file mapping
492		 */
493		lock_page(page);
494
495		if (page->mapping != mapping) {
496			unlock_page(page);
497			page_cache_release(page);
498			continue;
499		}
500
501		if (wbc->sync_mode != WB_SYNC_NONE)
502			wait_on_page_writeback(page);
503
504		if (PageWriteback(page) || !PageDirty(page)) {
505			unlock_page(page);
506			continue;
507		}
508
509		wb = (struct afs_writeback *) page_private(page);
510		ASSERT(wb != NULL);
511
512		spin_lock(&wb->vnode->writeback_lock);
513		wb->state = AFS_WBACK_WRITING;
514		spin_unlock(&wb->vnode->writeback_lock);
515
516		ret = afs_write_back_from_locked_page(wb, page);
517		unlock_page(page);
518		page_cache_release(page);
519		if (ret < 0) {
520			_leave(" = %d", ret);
521			return ret;
522		}
523
524		wbc->nr_to_write -= ret;
525
526		cond_resched();
527	} while (index < end && wbc->nr_to_write > 0);
528
529	*_next = index;
530	_leave(" = 0 [%lx]", *_next);
531	return 0;
532}
533
534/*
535 * write some of the pending data back to the server
536 */
537int afs_writepages(struct address_space *mapping,
538		   struct writeback_control *wbc)
539{
540	pgoff_t start, end, next;
541	int ret;
542
543	_enter("");
544
545	if (wbc->range_cyclic) {
546		start = mapping->writeback_index;
547		end = -1;
548		ret = afs_writepages_region(mapping, wbc, start, end, &next);
549		if (start > 0 && wbc->nr_to_write > 0 && ret == 0)
550			ret = afs_writepages_region(mapping, wbc, 0, start,
551						    &next);
552		mapping->writeback_index = next;
553	} else if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) {
554		end = (pgoff_t)(LLONG_MAX >> PAGE_CACHE_SHIFT);
555		ret = afs_writepages_region(mapping, wbc, 0, end, &next);
556		if (wbc->nr_to_write > 0)
557			mapping->writeback_index = next;
558	} else {
559		start = wbc->range_start >> PAGE_CACHE_SHIFT;
560		end = wbc->range_end >> PAGE_CACHE_SHIFT;
561		ret = afs_writepages_region(mapping, wbc, start, end, &next);
562	}
563
564	_leave(" = %d", ret);
565	return ret;
566}
567
568/*
569 * completion of write to server
570 */
571void afs_pages_written_back(struct afs_vnode *vnode, struct afs_call *call)
572{
573	struct afs_writeback *wb = call->wb;
574	struct pagevec pv;
575	unsigned count, loop;
576	pgoff_t first = call->first, last = call->last;
577	bool free_wb;
578
579	_enter("{%x:%u},{%lx-%lx}",
580	       vnode->fid.vid, vnode->fid.vnode, first, last);
581
582	ASSERT(wb != NULL);
583
584	pagevec_init(&pv, 0);
585
586	do {
587		_debug("done %lx-%lx", first, last);
588
589		count = last - first + 1;
590		if (count > PAGEVEC_SIZE)
591			count = PAGEVEC_SIZE;
592		pv.nr = find_get_pages_contig(call->mapping, first, count,
593					      pv.pages);
594		ASSERTCMP(pv.nr, ==, count);
595
596		spin_lock(&vnode->writeback_lock);
597		for (loop = 0; loop < count; loop++) {
598			struct page *page = pv.pages[loop];
599			end_page_writeback(page);
600			if (page_private(page) == (unsigned long) wb) {
601				set_page_private(page, 0);
602				ClearPagePrivate(page);
603				wb->usage--;
604			}
605		}
606		free_wb = false;
607		if (wb->usage == 0) {
608			afs_unlink_writeback(wb);
609			free_wb = true;
610		}
611		spin_unlock(&vnode->writeback_lock);
612		first += count;
613		if (free_wb) {
614			afs_free_writeback(wb);
615			wb = NULL;
616		}
617
618		__pagevec_release(&pv);
619	} while (first <= last);
620
621	_leave("");
622}
623
624/*
625 * write to an AFS file
626 */
627ssize_t afs_file_write(struct kiocb *iocb, struct iov_iter *from)
628{
629	struct afs_vnode *vnode = AFS_FS_I(file_inode(iocb->ki_filp));
630	ssize_t result;
631	size_t count = iov_iter_count(from);
632
633	_enter("{%x.%u},{%zu},",
634	       vnode->fid.vid, vnode->fid.vnode, count);
635
636	if (IS_SWAPFILE(&vnode->vfs_inode)) {
637		printk(KERN_INFO
638		       "AFS: Attempt to write to active swap file!\n");
639		return -EBUSY;
640	}
641
642	if (!count)
643		return 0;
644
645	result = generic_file_write_iter(iocb, from);
646	if (IS_ERR_VALUE(result)) {
647		_leave(" = %zd", result);
648		return result;
649	}
650
651	_leave(" = %zd", result);
652	return result;
653}
654
655/*
656 * flush the vnode to the fileserver
657 */
658int afs_writeback_all(struct afs_vnode *vnode)
659{
660	struct address_space *mapping = vnode->vfs_inode.i_mapping;
661	struct writeback_control wbc = {
662		.sync_mode	= WB_SYNC_ALL,
663		.nr_to_write	= LONG_MAX,
664		.range_cyclic	= 1,
665	};
666	int ret;
667
668	_enter("");
669
670	ret = mapping->a_ops->writepages(mapping, &wbc);
671	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
672
673	_leave(" = %d", ret);
674	return ret;
675}
676
677/*
678 * flush any dirty pages for this process, and check for write errors.
679 * - the return status from this call provides a reliable indication of
680 *   whether any write errors occurred for this process.
681 */
682int afs_fsync(struct file *file, loff_t start, loff_t end, int datasync)
683{
684	struct inode *inode = file_inode(file);
685	struct afs_writeback *wb, *xwb;
686	struct afs_vnode *vnode = AFS_FS_I(inode);
687	int ret;
688
689	_enter("{%x:%u},{n=%pD},%d",
690	       vnode->fid.vid, vnode->fid.vnode, file,
691	       datasync);
692
693	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
694	if (ret)
695		return ret;
696	mutex_lock(&inode->i_mutex);
697
698	/* use a writeback record as a marker in the queue - when this reaches
699	 * the front of the queue, all the outstanding writes are either
700	 * completed or rejected */
701	wb = kzalloc(sizeof(*wb), GFP_KERNEL);
702	if (!wb) {
703		ret = -ENOMEM;
704		goto out;
705	}
706	wb->vnode = vnode;
707	wb->first = 0;
708	wb->last = -1;
709	wb->offset_first = 0;
710	wb->to_last = PAGE_SIZE;
711	wb->usage = 1;
712	wb->state = AFS_WBACK_SYNCING;
713	init_waitqueue_head(&wb->waitq);
714
715	spin_lock(&vnode->writeback_lock);
716	list_for_each_entry(xwb, &vnode->writebacks, link) {
717		if (xwb->state == AFS_WBACK_PENDING)
718			xwb->state = AFS_WBACK_CONFLICTING;
719	}
720	list_add_tail(&wb->link, &vnode->writebacks);
721	spin_unlock(&vnode->writeback_lock);
722
723	/* push all the outstanding writebacks to the server */
724	ret = afs_writeback_all(vnode);
725	if (ret < 0) {
726		afs_put_writeback(wb);
727		_leave(" = %d [wb]", ret);
728		goto out;
729	}
730
731	/* wait for the preceding writes to actually complete */
732	ret = wait_event_interruptible(wb->waitq,
733				       wb->state == AFS_WBACK_COMPLETE ||
734				       vnode->writebacks.next == &wb->link);
735	afs_put_writeback(wb);
736	_leave(" = %d", ret);
737out:
738	mutex_unlock(&inode->i_mutex);
739	return ret;
740}
741
742/*
743 * notification that a previously read-only page is about to become writable
744 * - if it returns an error, the caller will deliver a bus error signal
745 */
746int afs_page_mkwrite(struct vm_area_struct *vma, struct page *page)
747{
748	struct afs_vnode *vnode = AFS_FS_I(vma->vm_file->f_mapping->host);
749
750	_enter("{{%x:%u}},{%lx}",
751	       vnode->fid.vid, vnode->fid.vnode, page->index);
752
753	/* wait for the page to be written to the cache before we allow it to
754	 * be modified */
755#ifdef CONFIG_AFS_FSCACHE
756	fscache_wait_on_page_write(vnode->cache, page);
757#endif
758
759	_leave(" = 0");
760	return 0;
761}
762