Loading...
1/*
2 * linux/fs/nfs/write.c
3 *
4 * Write file data over NFS.
5 *
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7 */
8
9#include <linux/types.h>
10#include <linux/slab.h>
11#include <linux/mm.h>
12#include <linux/pagemap.h>
13#include <linux/file.h>
14#include <linux/writeback.h>
15#include <linux/swap.h>
16#include <linux/migrate.h>
17
18#include <linux/sunrpc/clnt.h>
19#include <linux/nfs_fs.h>
20#include <linux/nfs_mount.h>
21#include <linux/nfs_page.h>
22#include <linux/backing-dev.h>
23
24#include <asm/uaccess.h>
25
26#include "delegation.h"
27#include "internal.h"
28#include "iostat.h"
29#include "nfs4_fs.h"
30#include "fscache.h"
31#include "pnfs.h"
32
33#define NFSDBG_FACILITY NFSDBG_PAGECACHE
34
35#define MIN_POOL_WRITE (32)
36#define MIN_POOL_COMMIT (4)
37
38/*
39 * Local function declarations
40 */
41static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
42 struct inode *inode, int ioflags);
43static void nfs_redirty_request(struct nfs_page *req);
44static const struct rpc_call_ops nfs_write_partial_ops;
45static const struct rpc_call_ops nfs_write_full_ops;
46static const struct rpc_call_ops nfs_commit_ops;
47
48static struct kmem_cache *nfs_wdata_cachep;
49static mempool_t *nfs_wdata_mempool;
50static mempool_t *nfs_commit_mempool;
51
52struct nfs_write_data *nfs_commitdata_alloc(void)
53{
54 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
55
56 if (p) {
57 memset(p, 0, sizeof(*p));
58 INIT_LIST_HEAD(&p->pages);
59 }
60 return p;
61}
62EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
63
64void nfs_commit_free(struct nfs_write_data *p)
65{
66 if (p && (p->pagevec != &p->page_array[0]))
67 kfree(p->pagevec);
68 mempool_free(p, nfs_commit_mempool);
69}
70EXPORT_SYMBOL_GPL(nfs_commit_free);
71
72struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
73{
74 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
75
76 if (p) {
77 memset(p, 0, sizeof(*p));
78 INIT_LIST_HEAD(&p->pages);
79 p->npages = pagecount;
80 if (pagecount <= ARRAY_SIZE(p->page_array))
81 p->pagevec = p->page_array;
82 else {
83 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
84 if (!p->pagevec) {
85 mempool_free(p, nfs_wdata_mempool);
86 p = NULL;
87 }
88 }
89 }
90 return p;
91}
92
93void nfs_writedata_free(struct nfs_write_data *p)
94{
95 if (p && (p->pagevec != &p->page_array[0]))
96 kfree(p->pagevec);
97 mempool_free(p, nfs_wdata_mempool);
98}
99
100void nfs_writedata_release(struct nfs_write_data *wdata)
101{
102 put_lseg(wdata->lseg);
103 put_nfs_open_context(wdata->args.context);
104 nfs_writedata_free(wdata);
105}
106
107static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
108{
109 ctx->error = error;
110 smp_wmb();
111 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
112}
113
114static struct nfs_page *nfs_page_find_request_locked(struct page *page)
115{
116 struct nfs_page *req = NULL;
117
118 if (PagePrivate(page)) {
119 req = (struct nfs_page *)page_private(page);
120 if (req != NULL)
121 kref_get(&req->wb_kref);
122 }
123 return req;
124}
125
126static struct nfs_page *nfs_page_find_request(struct page *page)
127{
128 struct inode *inode = page->mapping->host;
129 struct nfs_page *req = NULL;
130
131 spin_lock(&inode->i_lock);
132 req = nfs_page_find_request_locked(page);
133 spin_unlock(&inode->i_lock);
134 return req;
135}
136
137/* Adjust the file length if we're writing beyond the end */
138static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
139{
140 struct inode *inode = page->mapping->host;
141 loff_t end, i_size;
142 pgoff_t end_index;
143
144 spin_lock(&inode->i_lock);
145 i_size = i_size_read(inode);
146 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
147 if (i_size > 0 && page->index < end_index)
148 goto out;
149 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
150 if (i_size >= end)
151 goto out;
152 i_size_write(inode, end);
153 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
154out:
155 spin_unlock(&inode->i_lock);
156}
157
158/* A writeback failed: mark the page as bad, and invalidate the page cache */
159static void nfs_set_pageerror(struct page *page)
160{
161 SetPageError(page);
162 nfs_zap_mapping(page->mapping->host, page->mapping);
163}
164
165/* We can set the PG_uptodate flag if we see that a write request
166 * covers the full page.
167 */
168static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
169{
170 if (PageUptodate(page))
171 return;
172 if (base != 0)
173 return;
174 if (count != nfs_page_length(page))
175 return;
176 SetPageUptodate(page);
177}
178
179static int wb_priority(struct writeback_control *wbc)
180{
181 if (wbc->for_reclaim)
182 return FLUSH_HIGHPRI | FLUSH_STABLE;
183 if (wbc->for_kupdate || wbc->for_background)
184 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
185 return FLUSH_COND_STABLE;
186}
187
188/*
189 * NFS congestion control
190 */
191
192int nfs_congestion_kb;
193
194#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
195#define NFS_CONGESTION_OFF_THRESH \
196 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
197
198static int nfs_set_page_writeback(struct page *page)
199{
200 int ret = test_set_page_writeback(page);
201
202 if (!ret) {
203 struct inode *inode = page->mapping->host;
204 struct nfs_server *nfss = NFS_SERVER(inode);
205
206 page_cache_get(page);
207 if (atomic_long_inc_return(&nfss->writeback) >
208 NFS_CONGESTION_ON_THRESH) {
209 set_bdi_congested(&nfss->backing_dev_info,
210 BLK_RW_ASYNC);
211 }
212 }
213 return ret;
214}
215
216static void nfs_end_page_writeback(struct page *page)
217{
218 struct inode *inode = page->mapping->host;
219 struct nfs_server *nfss = NFS_SERVER(inode);
220
221 end_page_writeback(page);
222 page_cache_release(page);
223 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
224 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
225}
226
227static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
228{
229 struct inode *inode = page->mapping->host;
230 struct nfs_page *req;
231 int ret;
232
233 spin_lock(&inode->i_lock);
234 for (;;) {
235 req = nfs_page_find_request_locked(page);
236 if (req == NULL)
237 break;
238 if (nfs_set_page_tag_locked(req))
239 break;
240 /* Note: If we hold the page lock, as is the case in nfs_writepage,
241 * then the call to nfs_set_page_tag_locked() will always
242 * succeed provided that someone hasn't already marked the
243 * request as dirty (in which case we don't care).
244 */
245 spin_unlock(&inode->i_lock);
246 if (!nonblock)
247 ret = nfs_wait_on_request(req);
248 else
249 ret = -EAGAIN;
250 nfs_release_request(req);
251 if (ret != 0)
252 return ERR_PTR(ret);
253 spin_lock(&inode->i_lock);
254 }
255 spin_unlock(&inode->i_lock);
256 return req;
257}
258
259/*
260 * Find an associated nfs write request, and prepare to flush it out
261 * May return an error if the user signalled nfs_wait_on_request().
262 */
263static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
264 struct page *page, bool nonblock)
265{
266 struct nfs_page *req;
267 int ret = 0;
268
269 req = nfs_find_and_lock_request(page, nonblock);
270 if (!req)
271 goto out;
272 ret = PTR_ERR(req);
273 if (IS_ERR(req))
274 goto out;
275
276 ret = nfs_set_page_writeback(page);
277 BUG_ON(ret != 0);
278 BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
279
280 if (!nfs_pageio_add_request(pgio, req)) {
281 nfs_redirty_request(req);
282 ret = pgio->pg_error;
283 }
284out:
285 return ret;
286}
287
288static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
289{
290 struct inode *inode = page->mapping->host;
291 int ret;
292
293 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
294 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
295
296 nfs_pageio_cond_complete(pgio, page->index);
297 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
298 if (ret == -EAGAIN) {
299 redirty_page_for_writepage(wbc, page);
300 ret = 0;
301 }
302 return ret;
303}
304
305/*
306 * Write an mmapped page to the server.
307 */
308static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
309{
310 struct nfs_pageio_descriptor pgio;
311 int err;
312
313 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
314 err = nfs_do_writepage(page, wbc, &pgio);
315 nfs_pageio_complete(&pgio);
316 if (err < 0)
317 return err;
318 if (pgio.pg_error < 0)
319 return pgio.pg_error;
320 return 0;
321}
322
323int nfs_writepage(struct page *page, struct writeback_control *wbc)
324{
325 int ret;
326
327 ret = nfs_writepage_locked(page, wbc);
328 unlock_page(page);
329 return ret;
330}
331
332static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
333{
334 int ret;
335
336 ret = nfs_do_writepage(page, wbc, data);
337 unlock_page(page);
338 return ret;
339}
340
341int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
342{
343 struct inode *inode = mapping->host;
344 unsigned long *bitlock = &NFS_I(inode)->flags;
345 struct nfs_pageio_descriptor pgio;
346 int err;
347
348 /* Stop dirtying of new pages while we sync */
349 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
350 nfs_wait_bit_killable, TASK_KILLABLE);
351 if (err)
352 goto out_err;
353
354 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
355
356 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
357 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
358 nfs_pageio_complete(&pgio);
359
360 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
361 smp_mb__after_clear_bit();
362 wake_up_bit(bitlock, NFS_INO_FLUSHING);
363
364 if (err < 0)
365 goto out_err;
366 err = pgio.pg_error;
367 if (err < 0)
368 goto out_err;
369 return 0;
370out_err:
371 return err;
372}
373
374/*
375 * Insert a write request into an inode
376 */
377static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
378{
379 struct nfs_inode *nfsi = NFS_I(inode);
380 int error;
381
382 error = radix_tree_preload(GFP_NOFS);
383 if (error != 0)
384 goto out;
385
386 /* Lock the request! */
387 nfs_lock_request_dontget(req);
388
389 spin_lock(&inode->i_lock);
390 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
391 BUG_ON(error);
392 if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
393 nfsi->change_attr++;
394 set_bit(PG_MAPPED, &req->wb_flags);
395 SetPagePrivate(req->wb_page);
396 set_page_private(req->wb_page, (unsigned long)req);
397 nfsi->npages++;
398 kref_get(&req->wb_kref);
399 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
400 NFS_PAGE_TAG_LOCKED);
401 spin_unlock(&inode->i_lock);
402 radix_tree_preload_end();
403out:
404 return error;
405}
406
407/*
408 * Remove a write request from an inode
409 */
410static void nfs_inode_remove_request(struct nfs_page *req)
411{
412 struct inode *inode = req->wb_context->dentry->d_inode;
413 struct nfs_inode *nfsi = NFS_I(inode);
414
415 BUG_ON (!NFS_WBACK_BUSY(req));
416
417 spin_lock(&inode->i_lock);
418 set_page_private(req->wb_page, 0);
419 ClearPagePrivate(req->wb_page);
420 clear_bit(PG_MAPPED, &req->wb_flags);
421 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
422 nfsi->npages--;
423 spin_unlock(&inode->i_lock);
424 nfs_release_request(req);
425}
426
427static void
428nfs_mark_request_dirty(struct nfs_page *req)
429{
430 __set_page_dirty_nobuffers(req->wb_page);
431 __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
432}
433
434#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
435/*
436 * Add a request to the inode's commit list.
437 */
438static void
439nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
440{
441 struct inode *inode = req->wb_context->dentry->d_inode;
442 struct nfs_inode *nfsi = NFS_I(inode);
443
444 spin_lock(&inode->i_lock);
445 set_bit(PG_CLEAN, &(req)->wb_flags);
446 radix_tree_tag_set(&nfsi->nfs_page_tree,
447 req->wb_index,
448 NFS_PAGE_TAG_COMMIT);
449 nfsi->ncommit++;
450 spin_unlock(&inode->i_lock);
451 pnfs_mark_request_commit(req, lseg);
452 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
453 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
454 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
455}
456
457static int
458nfs_clear_request_commit(struct nfs_page *req)
459{
460 struct page *page = req->wb_page;
461
462 if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
463 dec_zone_page_state(page, NR_UNSTABLE_NFS);
464 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
465 return 1;
466 }
467 return 0;
468}
469
470static inline
471int nfs_write_need_commit(struct nfs_write_data *data)
472{
473 if (data->verf.committed == NFS_DATA_SYNC)
474 return data->lseg == NULL;
475 else
476 return data->verf.committed != NFS_FILE_SYNC;
477}
478
479static inline
480int nfs_reschedule_unstable_write(struct nfs_page *req,
481 struct nfs_write_data *data)
482{
483 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
484 nfs_mark_request_commit(req, data->lseg);
485 return 1;
486 }
487 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
488 nfs_mark_request_dirty(req);
489 return 1;
490 }
491 return 0;
492}
493#else
494static inline void
495nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
496{
497}
498
499static inline int
500nfs_clear_request_commit(struct nfs_page *req)
501{
502 return 0;
503}
504
505static inline
506int nfs_write_need_commit(struct nfs_write_data *data)
507{
508 return 0;
509}
510
511static inline
512int nfs_reschedule_unstable_write(struct nfs_page *req,
513 struct nfs_write_data *data)
514{
515 return 0;
516}
517#endif
518
519#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
520static int
521nfs_need_commit(struct nfs_inode *nfsi)
522{
523 return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
524}
525
526/*
527 * nfs_scan_commit - Scan an inode for commit requests
528 * @inode: NFS inode to scan
529 * @dst: destination list
530 * @idx_start: lower bound of page->index to scan.
531 * @npages: idx_start + npages sets the upper bound to scan.
532 *
533 * Moves requests from the inode's 'commit' request list.
534 * The requests are *not* checked to ensure that they form a contiguous set.
535 */
536static int
537nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
538{
539 struct nfs_inode *nfsi = NFS_I(inode);
540 int ret;
541
542 if (!nfs_need_commit(nfsi))
543 return 0;
544
545 spin_lock(&inode->i_lock);
546 ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
547 if (ret > 0)
548 nfsi->ncommit -= ret;
549 spin_unlock(&inode->i_lock);
550
551 if (nfs_need_commit(NFS_I(inode)))
552 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
553
554 return ret;
555}
556#else
557static inline int nfs_need_commit(struct nfs_inode *nfsi)
558{
559 return 0;
560}
561
562static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
563{
564 return 0;
565}
566#endif
567
568/*
569 * Search for an existing write request, and attempt to update
570 * it to reflect a new dirty region on a given page.
571 *
572 * If the attempt fails, then the existing request is flushed out
573 * to disk.
574 */
575static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
576 struct page *page,
577 unsigned int offset,
578 unsigned int bytes)
579{
580 struct nfs_page *req;
581 unsigned int rqend;
582 unsigned int end;
583 int error;
584
585 if (!PagePrivate(page))
586 return NULL;
587
588 end = offset + bytes;
589 spin_lock(&inode->i_lock);
590
591 for (;;) {
592 req = nfs_page_find_request_locked(page);
593 if (req == NULL)
594 goto out_unlock;
595
596 rqend = req->wb_offset + req->wb_bytes;
597 /*
598 * Tell the caller to flush out the request if
599 * the offsets are non-contiguous.
600 * Note: nfs_flush_incompatible() will already
601 * have flushed out requests having wrong owners.
602 */
603 if (offset > rqend
604 || end < req->wb_offset)
605 goto out_flushme;
606
607 if (nfs_set_page_tag_locked(req))
608 break;
609
610 /* The request is locked, so wait and then retry */
611 spin_unlock(&inode->i_lock);
612 error = nfs_wait_on_request(req);
613 nfs_release_request(req);
614 if (error != 0)
615 goto out_err;
616 spin_lock(&inode->i_lock);
617 }
618
619 if (nfs_clear_request_commit(req) &&
620 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
621 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL) {
622 NFS_I(inode)->ncommit--;
623 pnfs_clear_request_commit(req);
624 }
625
626 /* Okay, the request matches. Update the region */
627 if (offset < req->wb_offset) {
628 req->wb_offset = offset;
629 req->wb_pgbase = offset;
630 }
631 if (end > rqend)
632 req->wb_bytes = end - req->wb_offset;
633 else
634 req->wb_bytes = rqend - req->wb_offset;
635out_unlock:
636 spin_unlock(&inode->i_lock);
637 return req;
638out_flushme:
639 spin_unlock(&inode->i_lock);
640 nfs_release_request(req);
641 error = nfs_wb_page(inode, page);
642out_err:
643 return ERR_PTR(error);
644}
645
646/*
647 * Try to update an existing write request, or create one if there is none.
648 *
649 * Note: Should always be called with the Page Lock held to prevent races
650 * if we have to add a new request. Also assumes that the caller has
651 * already called nfs_flush_incompatible() if necessary.
652 */
653static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
654 struct page *page, unsigned int offset, unsigned int bytes)
655{
656 struct inode *inode = page->mapping->host;
657 struct nfs_page *req;
658 int error;
659
660 req = nfs_try_to_update_request(inode, page, offset, bytes);
661 if (req != NULL)
662 goto out;
663 req = nfs_create_request(ctx, inode, page, offset, bytes);
664 if (IS_ERR(req))
665 goto out;
666 error = nfs_inode_add_request(inode, req);
667 if (error != 0) {
668 nfs_release_request(req);
669 req = ERR_PTR(error);
670 }
671out:
672 return req;
673}
674
675static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
676 unsigned int offset, unsigned int count)
677{
678 struct nfs_page *req;
679
680 req = nfs_setup_write_request(ctx, page, offset, count);
681 if (IS_ERR(req))
682 return PTR_ERR(req);
683 /* Update file length */
684 nfs_grow_file(page, offset, count);
685 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
686 nfs_mark_request_dirty(req);
687 nfs_clear_page_tag_locked(req);
688 return 0;
689}
690
691int nfs_flush_incompatible(struct file *file, struct page *page)
692{
693 struct nfs_open_context *ctx = nfs_file_open_context(file);
694 struct nfs_page *req;
695 int do_flush, status;
696 /*
697 * Look for a request corresponding to this page. If there
698 * is one, and it belongs to another file, we flush it out
699 * before we try to copy anything into the page. Do this
700 * due to the lack of an ACCESS-type call in NFSv2.
701 * Also do the same if we find a request from an existing
702 * dropped page.
703 */
704 do {
705 req = nfs_page_find_request(page);
706 if (req == NULL)
707 return 0;
708 do_flush = req->wb_page != page || req->wb_context != ctx ||
709 req->wb_lock_context->lockowner != current->files ||
710 req->wb_lock_context->pid != current->tgid;
711 nfs_release_request(req);
712 if (!do_flush)
713 return 0;
714 status = nfs_wb_page(page->mapping->host, page);
715 } while (status == 0);
716 return status;
717}
718
719/*
720 * If the page cache is marked as unsafe or invalid, then we can't rely on
721 * the PageUptodate() flag. In this case, we will need to turn off
722 * write optimisations that depend on the page contents being correct.
723 */
724static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
725{
726 return PageUptodate(page) &&
727 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
728}
729
730/*
731 * Update and possibly write a cached page of an NFS file.
732 *
733 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
734 * things with a page scheduled for an RPC call (e.g. invalidate it).
735 */
736int nfs_updatepage(struct file *file, struct page *page,
737 unsigned int offset, unsigned int count)
738{
739 struct nfs_open_context *ctx = nfs_file_open_context(file);
740 struct inode *inode = page->mapping->host;
741 int status = 0;
742
743 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
744
745 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
746 file->f_path.dentry->d_parent->d_name.name,
747 file->f_path.dentry->d_name.name, count,
748 (long long)(page_offset(page) + offset));
749
750 /* If we're not using byte range locks, and we know the page
751 * is up to date, it may be more efficient to extend the write
752 * to cover the entire page in order to avoid fragmentation
753 * inefficiencies.
754 */
755 if (nfs_write_pageuptodate(page, inode) &&
756 inode->i_flock == NULL &&
757 !(file->f_flags & O_DSYNC)) {
758 count = max(count + offset, nfs_page_length(page));
759 offset = 0;
760 }
761
762 status = nfs_writepage_setup(ctx, page, offset, count);
763 if (status < 0)
764 nfs_set_pageerror(page);
765
766 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
767 status, (long long)i_size_read(inode));
768 return status;
769}
770
771static void nfs_writepage_release(struct nfs_page *req,
772 struct nfs_write_data *data)
773{
774 struct page *page = req->wb_page;
775
776 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
777 nfs_inode_remove_request(req);
778 nfs_clear_page_tag_locked(req);
779 nfs_end_page_writeback(page);
780}
781
782static int flush_task_priority(int how)
783{
784 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
785 case FLUSH_HIGHPRI:
786 return RPC_PRIORITY_HIGH;
787 case FLUSH_LOWPRI:
788 return RPC_PRIORITY_LOW;
789 }
790 return RPC_PRIORITY_NORMAL;
791}
792
793int nfs_initiate_write(struct nfs_write_data *data,
794 struct rpc_clnt *clnt,
795 const struct rpc_call_ops *call_ops,
796 int how)
797{
798 struct inode *inode = data->inode;
799 int priority = flush_task_priority(how);
800 struct rpc_task *task;
801 struct rpc_message msg = {
802 .rpc_argp = &data->args,
803 .rpc_resp = &data->res,
804 .rpc_cred = data->cred,
805 };
806 struct rpc_task_setup task_setup_data = {
807 .rpc_client = clnt,
808 .task = &data->task,
809 .rpc_message = &msg,
810 .callback_ops = call_ops,
811 .callback_data = data,
812 .workqueue = nfsiod_workqueue,
813 .flags = RPC_TASK_ASYNC,
814 .priority = priority,
815 };
816 int ret = 0;
817
818 /* Set up the initial task struct. */
819 NFS_PROTO(inode)->write_setup(data, &msg);
820
821 dprintk("NFS: %5u initiated write call "
822 "(req %s/%lld, %u bytes @ offset %llu)\n",
823 data->task.tk_pid,
824 inode->i_sb->s_id,
825 (long long)NFS_FILEID(inode),
826 data->args.count,
827 (unsigned long long)data->args.offset);
828
829 task = rpc_run_task(&task_setup_data);
830 if (IS_ERR(task)) {
831 ret = PTR_ERR(task);
832 goto out;
833 }
834 if (how & FLUSH_SYNC) {
835 ret = rpc_wait_for_completion_task(task);
836 if (ret == 0)
837 ret = task->tk_status;
838 }
839 rpc_put_task(task);
840out:
841 return ret;
842}
843EXPORT_SYMBOL_GPL(nfs_initiate_write);
844
845/*
846 * Set up the argument/result storage required for the RPC call.
847 */
848static void nfs_write_rpcsetup(struct nfs_page *req,
849 struct nfs_write_data *data,
850 unsigned int count, unsigned int offset,
851 int how)
852{
853 struct inode *inode = req->wb_context->dentry->d_inode;
854
855 /* Set up the RPC argument and reply structs
856 * NB: take care not to mess about with data->commit et al. */
857
858 data->req = req;
859 data->inode = inode = req->wb_context->dentry->d_inode;
860 data->cred = req->wb_context->cred;
861
862 data->args.fh = NFS_FH(inode);
863 data->args.offset = req_offset(req) + offset;
864 /* pnfs_set_layoutcommit needs this */
865 data->mds_offset = data->args.offset;
866 data->args.pgbase = req->wb_pgbase + offset;
867 data->args.pages = data->pagevec;
868 data->args.count = count;
869 data->args.context = get_nfs_open_context(req->wb_context);
870 data->args.lock_context = req->wb_lock_context;
871 data->args.stable = NFS_UNSTABLE;
872 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
873 case 0:
874 break;
875 case FLUSH_COND_STABLE:
876 if (nfs_need_commit(NFS_I(inode)))
877 break;
878 default:
879 data->args.stable = NFS_FILE_SYNC;
880 }
881
882 data->res.fattr = &data->fattr;
883 data->res.count = count;
884 data->res.verf = &data->verf;
885 nfs_fattr_init(&data->fattr);
886}
887
888static int nfs_do_write(struct nfs_write_data *data,
889 const struct rpc_call_ops *call_ops,
890 int how)
891{
892 struct inode *inode = data->args.context->dentry->d_inode;
893
894 return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
895}
896
897static int nfs_do_multiple_writes(struct list_head *head,
898 const struct rpc_call_ops *call_ops,
899 int how)
900{
901 struct nfs_write_data *data;
902 int ret = 0;
903
904 while (!list_empty(head)) {
905 int ret2;
906
907 data = list_entry(head->next, struct nfs_write_data, list);
908 list_del_init(&data->list);
909
910 ret2 = nfs_do_write(data, call_ops, how);
911 if (ret == 0)
912 ret = ret2;
913 }
914 return ret;
915}
916
917/* If a nfs_flush_* function fails, it should remove reqs from @head and
918 * call this on each, which will prepare them to be retried on next
919 * writeback using standard nfs.
920 */
921static void nfs_redirty_request(struct nfs_page *req)
922{
923 struct page *page = req->wb_page;
924
925 nfs_mark_request_dirty(req);
926 nfs_clear_page_tag_locked(req);
927 nfs_end_page_writeback(page);
928}
929
930/*
931 * Generate multiple small requests to write out a single
932 * contiguous dirty area on one page.
933 */
934static int nfs_flush_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
935{
936 struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
937 struct page *page = req->wb_page;
938 struct nfs_write_data *data;
939 size_t wsize = desc->pg_bsize, nbytes;
940 unsigned int offset;
941 int requests = 0;
942 int ret = 0;
943
944 nfs_list_remove_request(req);
945
946 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
947 (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
948 desc->pg_count > wsize))
949 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
950
951
952 offset = 0;
953 nbytes = desc->pg_count;
954 do {
955 size_t len = min(nbytes, wsize);
956
957 data = nfs_writedata_alloc(1);
958 if (!data)
959 goto out_bad;
960 data->pagevec[0] = page;
961 nfs_write_rpcsetup(req, data, len, offset, desc->pg_ioflags);
962 list_add(&data->list, res);
963 requests++;
964 nbytes -= len;
965 offset += len;
966 } while (nbytes != 0);
967 atomic_set(&req->wb_complete, requests);
968 desc->pg_rpc_callops = &nfs_write_partial_ops;
969 return ret;
970
971out_bad:
972 while (!list_empty(res)) {
973 data = list_entry(res->next, struct nfs_write_data, list);
974 list_del(&data->list);
975 nfs_writedata_free(data);
976 }
977 nfs_redirty_request(req);
978 return -ENOMEM;
979}
980
981/*
982 * Create an RPC task for the given write request and kick it.
983 * The page must have been locked by the caller.
984 *
985 * It may happen that the page we're passed is not marked dirty.
986 * This is the case if nfs_updatepage detects a conflicting request
987 * that has been written but not committed.
988 */
989static int nfs_flush_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
990{
991 struct nfs_page *req;
992 struct page **pages;
993 struct nfs_write_data *data;
994 struct list_head *head = &desc->pg_list;
995 int ret = 0;
996
997 data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
998 desc->pg_count));
999 if (!data) {
1000 while (!list_empty(head)) {
1001 req = nfs_list_entry(head->next);
1002 nfs_list_remove_request(req);
1003 nfs_redirty_request(req);
1004 }
1005 ret = -ENOMEM;
1006 goto out;
1007 }
1008 pages = data->pagevec;
1009 while (!list_empty(head)) {
1010 req = nfs_list_entry(head->next);
1011 nfs_list_remove_request(req);
1012 nfs_list_add_request(req, &data->pages);
1013 ClearPageError(req->wb_page);
1014 *pages++ = req->wb_page;
1015 }
1016 req = nfs_list_entry(data->pages.next);
1017
1018 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1019 (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1020 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1021
1022 /* Set up the argument struct */
1023 nfs_write_rpcsetup(req, data, desc->pg_count, 0, desc->pg_ioflags);
1024 list_add(&data->list, res);
1025 desc->pg_rpc_callops = &nfs_write_full_ops;
1026out:
1027 return ret;
1028}
1029
1030int nfs_generic_flush(struct nfs_pageio_descriptor *desc, struct list_head *head)
1031{
1032 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1033 return nfs_flush_multi(desc, head);
1034 return nfs_flush_one(desc, head);
1035}
1036
1037static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1038{
1039 LIST_HEAD(head);
1040 int ret;
1041
1042 ret = nfs_generic_flush(desc, &head);
1043 if (ret == 0)
1044 ret = nfs_do_multiple_writes(&head, desc->pg_rpc_callops,
1045 desc->pg_ioflags);
1046 return ret;
1047}
1048
1049static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1050 .pg_test = nfs_generic_pg_test,
1051 .pg_doio = nfs_generic_pg_writepages,
1052};
1053
1054static void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1055 struct inode *inode, int ioflags)
1056{
1057 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops,
1058 NFS_SERVER(inode)->wsize, ioflags);
1059}
1060
1061void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1062{
1063 pgio->pg_ops = &nfs_pageio_write_ops;
1064 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1065}
1066EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1067
1068static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1069 struct inode *inode, int ioflags)
1070{
1071 if (!pnfs_pageio_init_write(pgio, inode, ioflags))
1072 nfs_pageio_init_write_mds(pgio, inode, ioflags);
1073}
1074
1075/*
1076 * Handle a write reply that flushed part of a page.
1077 */
1078static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1079{
1080 struct nfs_write_data *data = calldata;
1081
1082 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1083 task->tk_pid,
1084 data->req->wb_context->dentry->d_inode->i_sb->s_id,
1085 (long long)
1086 NFS_FILEID(data->req->wb_context->dentry->d_inode),
1087 data->req->wb_bytes, (long long)req_offset(data->req));
1088
1089 nfs_writeback_done(task, data);
1090}
1091
1092static void nfs_writeback_release_partial(void *calldata)
1093{
1094 struct nfs_write_data *data = calldata;
1095 struct nfs_page *req = data->req;
1096 struct page *page = req->wb_page;
1097 int status = data->task.tk_status;
1098
1099 if (status < 0) {
1100 nfs_set_pageerror(page);
1101 nfs_context_set_write_error(req->wb_context, status);
1102 dprintk(", error = %d\n", status);
1103 goto out;
1104 }
1105
1106 if (nfs_write_need_commit(data)) {
1107 struct inode *inode = page->mapping->host;
1108
1109 spin_lock(&inode->i_lock);
1110 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1111 /* Do nothing we need to resend the writes */
1112 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1113 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1114 dprintk(" defer commit\n");
1115 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1116 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1117 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1118 dprintk(" server reboot detected\n");
1119 }
1120 spin_unlock(&inode->i_lock);
1121 } else
1122 dprintk(" OK\n");
1123
1124out:
1125 if (atomic_dec_and_test(&req->wb_complete))
1126 nfs_writepage_release(req, data);
1127 nfs_writedata_release(calldata);
1128}
1129
1130#if defined(CONFIG_NFS_V4_1)
1131void nfs_write_prepare(struct rpc_task *task, void *calldata)
1132{
1133 struct nfs_write_data *data = calldata;
1134
1135 if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1136 &data->args.seq_args,
1137 &data->res.seq_res, 1, task))
1138 return;
1139 rpc_call_start(task);
1140}
1141#endif /* CONFIG_NFS_V4_1 */
1142
1143static const struct rpc_call_ops nfs_write_partial_ops = {
1144#if defined(CONFIG_NFS_V4_1)
1145 .rpc_call_prepare = nfs_write_prepare,
1146#endif /* CONFIG_NFS_V4_1 */
1147 .rpc_call_done = nfs_writeback_done_partial,
1148 .rpc_release = nfs_writeback_release_partial,
1149};
1150
1151/*
1152 * Handle a write reply that flushes a whole page.
1153 *
1154 * FIXME: There is an inherent race with invalidate_inode_pages and
1155 * writebacks since the page->count is kept > 1 for as long
1156 * as the page has a write request pending.
1157 */
1158static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1159{
1160 struct nfs_write_data *data = calldata;
1161
1162 nfs_writeback_done(task, data);
1163}
1164
1165static void nfs_writeback_release_full(void *calldata)
1166{
1167 struct nfs_write_data *data = calldata;
1168 int status = data->task.tk_status;
1169
1170 /* Update attributes as result of writeback. */
1171 while (!list_empty(&data->pages)) {
1172 struct nfs_page *req = nfs_list_entry(data->pages.next);
1173 struct page *page = req->wb_page;
1174
1175 nfs_list_remove_request(req);
1176
1177 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1178 data->task.tk_pid,
1179 req->wb_context->dentry->d_inode->i_sb->s_id,
1180 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1181 req->wb_bytes,
1182 (long long)req_offset(req));
1183
1184 if (status < 0) {
1185 nfs_set_pageerror(page);
1186 nfs_context_set_write_error(req->wb_context, status);
1187 dprintk(", error = %d\n", status);
1188 goto remove_request;
1189 }
1190
1191 if (nfs_write_need_commit(data)) {
1192 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1193 nfs_mark_request_commit(req, data->lseg);
1194 dprintk(" marked for commit\n");
1195 goto next;
1196 }
1197 dprintk(" OK\n");
1198remove_request:
1199 nfs_inode_remove_request(req);
1200 next:
1201 nfs_clear_page_tag_locked(req);
1202 nfs_end_page_writeback(page);
1203 }
1204 nfs_writedata_release(calldata);
1205}
1206
1207static const struct rpc_call_ops nfs_write_full_ops = {
1208#if defined(CONFIG_NFS_V4_1)
1209 .rpc_call_prepare = nfs_write_prepare,
1210#endif /* CONFIG_NFS_V4_1 */
1211 .rpc_call_done = nfs_writeback_done_full,
1212 .rpc_release = nfs_writeback_release_full,
1213};
1214
1215
1216/*
1217 * This function is called when the WRITE call is complete.
1218 */
1219void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1220{
1221 struct nfs_writeargs *argp = &data->args;
1222 struct nfs_writeres *resp = &data->res;
1223 struct nfs_server *server = NFS_SERVER(data->inode);
1224 int status;
1225
1226 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1227 task->tk_pid, task->tk_status);
1228
1229 /*
1230 * ->write_done will attempt to use post-op attributes to detect
1231 * conflicting writes by other clients. A strict interpretation
1232 * of close-to-open would allow us to continue caching even if
1233 * another writer had changed the file, but some applications
1234 * depend on tighter cache coherency when writing.
1235 */
1236 status = NFS_PROTO(data->inode)->write_done(task, data);
1237 if (status != 0)
1238 return;
1239 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1240
1241#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1242 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1243 /* We tried a write call, but the server did not
1244 * commit data to stable storage even though we
1245 * requested it.
1246 * Note: There is a known bug in Tru64 < 5.0 in which
1247 * the server reports NFS_DATA_SYNC, but performs
1248 * NFS_FILE_SYNC. We therefore implement this checking
1249 * as a dprintk() in order to avoid filling syslog.
1250 */
1251 static unsigned long complain;
1252
1253 /* Note this will print the MDS for a DS write */
1254 if (time_before(complain, jiffies)) {
1255 dprintk("NFS: faulty NFS server %s:"
1256 " (committed = %d) != (stable = %d)\n",
1257 server->nfs_client->cl_hostname,
1258 resp->verf->committed, argp->stable);
1259 complain = jiffies + 300 * HZ;
1260 }
1261 }
1262#endif
1263 /* Is this a short write? */
1264 if (task->tk_status >= 0 && resp->count < argp->count) {
1265 static unsigned long complain;
1266
1267 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1268
1269 /* Has the server at least made some progress? */
1270 if (resp->count != 0) {
1271 /* Was this an NFSv2 write or an NFSv3 stable write? */
1272 if (resp->verf->committed != NFS_UNSTABLE) {
1273 /* Resend from where the server left off */
1274 data->mds_offset += resp->count;
1275 argp->offset += resp->count;
1276 argp->pgbase += resp->count;
1277 argp->count -= resp->count;
1278 } else {
1279 /* Resend as a stable write in order to avoid
1280 * headaches in the case of a server crash.
1281 */
1282 argp->stable = NFS_FILE_SYNC;
1283 }
1284 nfs_restart_rpc(task, server->nfs_client);
1285 return;
1286 }
1287 if (time_before(complain, jiffies)) {
1288 printk(KERN_WARNING
1289 "NFS: Server wrote zero bytes, expected %u.\n",
1290 argp->count);
1291 complain = jiffies + 300 * HZ;
1292 }
1293 /* Can't do anything about it except throw an error. */
1294 task->tk_status = -EIO;
1295 }
1296 return;
1297}
1298
1299
1300#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1301static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1302{
1303 int ret;
1304
1305 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1306 return 1;
1307 if (!may_wait)
1308 return 0;
1309 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1310 NFS_INO_COMMIT,
1311 nfs_wait_bit_killable,
1312 TASK_KILLABLE);
1313 return (ret < 0) ? ret : 1;
1314}
1315
1316void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1317{
1318 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1319 smp_mb__after_clear_bit();
1320 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1321}
1322EXPORT_SYMBOL_GPL(nfs_commit_clear_lock);
1323
1324void nfs_commitdata_release(void *data)
1325{
1326 struct nfs_write_data *wdata = data;
1327
1328 put_lseg(wdata->lseg);
1329 put_nfs_open_context(wdata->args.context);
1330 nfs_commit_free(wdata);
1331}
1332EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1333
1334int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1335 const struct rpc_call_ops *call_ops,
1336 int how)
1337{
1338 struct rpc_task *task;
1339 int priority = flush_task_priority(how);
1340 struct rpc_message msg = {
1341 .rpc_argp = &data->args,
1342 .rpc_resp = &data->res,
1343 .rpc_cred = data->cred,
1344 };
1345 struct rpc_task_setup task_setup_data = {
1346 .task = &data->task,
1347 .rpc_client = clnt,
1348 .rpc_message = &msg,
1349 .callback_ops = call_ops,
1350 .callback_data = data,
1351 .workqueue = nfsiod_workqueue,
1352 .flags = RPC_TASK_ASYNC,
1353 .priority = priority,
1354 };
1355 /* Set up the initial task struct. */
1356 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1357
1358 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1359
1360 task = rpc_run_task(&task_setup_data);
1361 if (IS_ERR(task))
1362 return PTR_ERR(task);
1363 if (how & FLUSH_SYNC)
1364 rpc_wait_for_completion_task(task);
1365 rpc_put_task(task);
1366 return 0;
1367}
1368EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1369
1370/*
1371 * Set up the argument/result storage required for the RPC call.
1372 */
1373void nfs_init_commit(struct nfs_write_data *data,
1374 struct list_head *head,
1375 struct pnfs_layout_segment *lseg)
1376{
1377 struct nfs_page *first = nfs_list_entry(head->next);
1378 struct inode *inode = first->wb_context->dentry->d_inode;
1379
1380 /* Set up the RPC argument and reply structs
1381 * NB: take care not to mess about with data->commit et al. */
1382
1383 list_splice_init(head, &data->pages);
1384
1385 data->inode = inode;
1386 data->cred = first->wb_context->cred;
1387 data->lseg = lseg; /* reference transferred */
1388 data->mds_ops = &nfs_commit_ops;
1389
1390 data->args.fh = NFS_FH(data->inode);
1391 /* Note: we always request a commit of the entire inode */
1392 data->args.offset = 0;
1393 data->args.count = 0;
1394 data->args.context = get_nfs_open_context(first->wb_context);
1395 data->res.count = 0;
1396 data->res.fattr = &data->fattr;
1397 data->res.verf = &data->verf;
1398 nfs_fattr_init(&data->fattr);
1399}
1400EXPORT_SYMBOL_GPL(nfs_init_commit);
1401
1402void nfs_retry_commit(struct list_head *page_list,
1403 struct pnfs_layout_segment *lseg)
1404{
1405 struct nfs_page *req;
1406
1407 while (!list_empty(page_list)) {
1408 req = nfs_list_entry(page_list->next);
1409 nfs_list_remove_request(req);
1410 nfs_mark_request_commit(req, lseg);
1411 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1412 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1413 BDI_RECLAIMABLE);
1414 nfs_clear_page_tag_locked(req);
1415 }
1416}
1417EXPORT_SYMBOL_GPL(nfs_retry_commit);
1418
1419/*
1420 * Commit dirty pages
1421 */
1422static int
1423nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1424{
1425 struct nfs_write_data *data;
1426
1427 data = nfs_commitdata_alloc();
1428
1429 if (!data)
1430 goto out_bad;
1431
1432 /* Set up the argument struct */
1433 nfs_init_commit(data, head, NULL);
1434 return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1435 out_bad:
1436 nfs_retry_commit(head, NULL);
1437 nfs_commit_clear_lock(NFS_I(inode));
1438 return -ENOMEM;
1439}
1440
1441/*
1442 * COMMIT call returned
1443 */
1444static void nfs_commit_done(struct rpc_task *task, void *calldata)
1445{
1446 struct nfs_write_data *data = calldata;
1447
1448 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1449 task->tk_pid, task->tk_status);
1450
1451 /* Call the NFS version-specific code */
1452 NFS_PROTO(data->inode)->commit_done(task, data);
1453}
1454
1455void nfs_commit_release_pages(struct nfs_write_data *data)
1456{
1457 struct nfs_page *req;
1458 int status = data->task.tk_status;
1459
1460 while (!list_empty(&data->pages)) {
1461 req = nfs_list_entry(data->pages.next);
1462 nfs_list_remove_request(req);
1463 nfs_clear_request_commit(req);
1464
1465 dprintk("NFS: commit (%s/%lld %d@%lld)",
1466 req->wb_context->dentry->d_sb->s_id,
1467 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1468 req->wb_bytes,
1469 (long long)req_offset(req));
1470 if (status < 0) {
1471 nfs_context_set_write_error(req->wb_context, status);
1472 nfs_inode_remove_request(req);
1473 dprintk(", error = %d\n", status);
1474 goto next;
1475 }
1476
1477 /* Okay, COMMIT succeeded, apparently. Check the verifier
1478 * returned by the server against all stored verfs. */
1479 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1480 /* We have a match */
1481 nfs_inode_remove_request(req);
1482 dprintk(" OK\n");
1483 goto next;
1484 }
1485 /* We have a mismatch. Write the page again */
1486 dprintk(" mismatch\n");
1487 nfs_mark_request_dirty(req);
1488 next:
1489 nfs_clear_page_tag_locked(req);
1490 }
1491}
1492EXPORT_SYMBOL_GPL(nfs_commit_release_pages);
1493
1494static void nfs_commit_release(void *calldata)
1495{
1496 struct nfs_write_data *data = calldata;
1497
1498 nfs_commit_release_pages(data);
1499 nfs_commit_clear_lock(NFS_I(data->inode));
1500 nfs_commitdata_release(calldata);
1501}
1502
1503static const struct rpc_call_ops nfs_commit_ops = {
1504#if defined(CONFIG_NFS_V4_1)
1505 .rpc_call_prepare = nfs_write_prepare,
1506#endif /* CONFIG_NFS_V4_1 */
1507 .rpc_call_done = nfs_commit_done,
1508 .rpc_release = nfs_commit_release,
1509};
1510
1511int nfs_commit_inode(struct inode *inode, int how)
1512{
1513 LIST_HEAD(head);
1514 int may_wait = how & FLUSH_SYNC;
1515 int res;
1516
1517 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1518 if (res <= 0)
1519 goto out_mark_dirty;
1520 res = nfs_scan_commit(inode, &head, 0, 0);
1521 if (res) {
1522 int error;
1523
1524 error = pnfs_commit_list(inode, &head, how);
1525 if (error == PNFS_NOT_ATTEMPTED)
1526 error = nfs_commit_list(inode, &head, how);
1527 if (error < 0)
1528 return error;
1529 if (!may_wait)
1530 goto out_mark_dirty;
1531 error = wait_on_bit(&NFS_I(inode)->flags,
1532 NFS_INO_COMMIT,
1533 nfs_wait_bit_killable,
1534 TASK_KILLABLE);
1535 if (error < 0)
1536 return error;
1537 } else
1538 nfs_commit_clear_lock(NFS_I(inode));
1539 return res;
1540 /* Note: If we exit without ensuring that the commit is complete,
1541 * we must mark the inode as dirty. Otherwise, future calls to
1542 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1543 * that the data is on the disk.
1544 */
1545out_mark_dirty:
1546 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1547 return res;
1548}
1549
1550static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1551{
1552 struct nfs_inode *nfsi = NFS_I(inode);
1553 int flags = FLUSH_SYNC;
1554 int ret = 0;
1555
1556 if (wbc->sync_mode == WB_SYNC_NONE) {
1557 /* Don't commit yet if this is a non-blocking flush and there
1558 * are a lot of outstanding writes for this mapping.
1559 */
1560 if (nfsi->ncommit <= (nfsi->npages >> 1))
1561 goto out_mark_dirty;
1562
1563 /* don't wait for the COMMIT response */
1564 flags = 0;
1565 }
1566
1567 ret = nfs_commit_inode(inode, flags);
1568 if (ret >= 0) {
1569 if (wbc->sync_mode == WB_SYNC_NONE) {
1570 if (ret < wbc->nr_to_write)
1571 wbc->nr_to_write -= ret;
1572 else
1573 wbc->nr_to_write = 0;
1574 }
1575 return 0;
1576 }
1577out_mark_dirty:
1578 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1579 return ret;
1580}
1581#else
1582static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1583{
1584 return 0;
1585}
1586#endif
1587
1588int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1589{
1590 int ret;
1591
1592 ret = nfs_commit_unstable_pages(inode, wbc);
1593 if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1594 int status;
1595 bool sync = true;
1596
1597 if (wbc->sync_mode == WB_SYNC_NONE)
1598 sync = false;
1599
1600 status = pnfs_layoutcommit_inode(inode, sync);
1601 if (status < 0)
1602 return status;
1603 }
1604 return ret;
1605}
1606
1607/*
1608 * flush the inode to disk.
1609 */
1610int nfs_wb_all(struct inode *inode)
1611{
1612 struct writeback_control wbc = {
1613 .sync_mode = WB_SYNC_ALL,
1614 .nr_to_write = LONG_MAX,
1615 .range_start = 0,
1616 .range_end = LLONG_MAX,
1617 };
1618
1619 return sync_inode(inode, &wbc);
1620}
1621
1622int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1623{
1624 struct nfs_page *req;
1625 int ret = 0;
1626
1627 BUG_ON(!PageLocked(page));
1628 for (;;) {
1629 wait_on_page_writeback(page);
1630 req = nfs_page_find_request(page);
1631 if (req == NULL)
1632 break;
1633 if (nfs_lock_request_dontget(req)) {
1634 nfs_inode_remove_request(req);
1635 /*
1636 * In case nfs_inode_remove_request has marked the
1637 * page as being dirty
1638 */
1639 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1640 nfs_unlock_request(req);
1641 break;
1642 }
1643 ret = nfs_wait_on_request(req);
1644 nfs_release_request(req);
1645 if (ret < 0)
1646 break;
1647 }
1648 return ret;
1649}
1650
1651/*
1652 * Write back all requests on one page - we do this before reading it.
1653 */
1654int nfs_wb_page(struct inode *inode, struct page *page)
1655{
1656 loff_t range_start = page_offset(page);
1657 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1658 struct writeback_control wbc = {
1659 .sync_mode = WB_SYNC_ALL,
1660 .nr_to_write = 0,
1661 .range_start = range_start,
1662 .range_end = range_end,
1663 };
1664 int ret;
1665
1666 for (;;) {
1667 wait_on_page_writeback(page);
1668 if (clear_page_dirty_for_io(page)) {
1669 ret = nfs_writepage_locked(page, &wbc);
1670 if (ret < 0)
1671 goto out_error;
1672 continue;
1673 }
1674 if (!PagePrivate(page))
1675 break;
1676 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1677 if (ret < 0)
1678 goto out_error;
1679 }
1680 return 0;
1681out_error:
1682 return ret;
1683}
1684
1685#ifdef CONFIG_MIGRATION
1686int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1687 struct page *page)
1688{
1689 struct nfs_page *req;
1690 int ret;
1691
1692 nfs_fscache_release_page(page, GFP_KERNEL);
1693
1694 req = nfs_find_and_lock_request(page, false);
1695 ret = PTR_ERR(req);
1696 if (IS_ERR(req))
1697 goto out;
1698
1699 ret = migrate_page(mapping, newpage, page);
1700 if (!req)
1701 goto out;
1702 if (ret)
1703 goto out_unlock;
1704 page_cache_get(newpage);
1705 spin_lock(&mapping->host->i_lock);
1706 req->wb_page = newpage;
1707 SetPagePrivate(newpage);
1708 set_page_private(newpage, (unsigned long)req);
1709 ClearPagePrivate(page);
1710 set_page_private(page, 0);
1711 spin_unlock(&mapping->host->i_lock);
1712 page_cache_release(page);
1713out_unlock:
1714 nfs_clear_page_tag_locked(req);
1715out:
1716 return ret;
1717}
1718#endif
1719
1720int __init nfs_init_writepagecache(void)
1721{
1722 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1723 sizeof(struct nfs_write_data),
1724 0, SLAB_HWCACHE_ALIGN,
1725 NULL);
1726 if (nfs_wdata_cachep == NULL)
1727 return -ENOMEM;
1728
1729 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1730 nfs_wdata_cachep);
1731 if (nfs_wdata_mempool == NULL)
1732 return -ENOMEM;
1733
1734 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1735 nfs_wdata_cachep);
1736 if (nfs_commit_mempool == NULL)
1737 return -ENOMEM;
1738
1739 /*
1740 * NFS congestion size, scale with available memory.
1741 *
1742 * 64MB: 8192k
1743 * 128MB: 11585k
1744 * 256MB: 16384k
1745 * 512MB: 23170k
1746 * 1GB: 32768k
1747 * 2GB: 46340k
1748 * 4GB: 65536k
1749 * 8GB: 92681k
1750 * 16GB: 131072k
1751 *
1752 * This allows larger machines to have larger/more transfers.
1753 * Limit the default to 256M
1754 */
1755 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1756 if (nfs_congestion_kb > 256*1024)
1757 nfs_congestion_kb = 256*1024;
1758
1759 return 0;
1760}
1761
1762void nfs_destroy_writepagecache(void)
1763{
1764 mempool_destroy(nfs_commit_mempool);
1765 mempool_destroy(nfs_wdata_mempool);
1766 kmem_cache_destroy(nfs_wdata_cachep);
1767}
1768
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/nfs/write.c
4 *
5 * Write file data over NFS.
6 *
7 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
8 */
9
10#include <linux/types.h>
11#include <linux/slab.h>
12#include <linux/mm.h>
13#include <linux/pagemap.h>
14#include <linux/file.h>
15#include <linux/writeback.h>
16#include <linux/swap.h>
17#include <linux/migrate.h>
18
19#include <linux/sunrpc/clnt.h>
20#include <linux/nfs_fs.h>
21#include <linux/nfs_mount.h>
22#include <linux/nfs_page.h>
23#include <linux/backing-dev.h>
24#include <linux/export.h>
25#include <linux/freezer.h>
26#include <linux/wait.h>
27#include <linux/iversion.h>
28
29#include <linux/uaccess.h>
30#include <linux/sched/mm.h>
31
32#include "delegation.h"
33#include "internal.h"
34#include "iostat.h"
35#include "nfs4_fs.h"
36#include "fscache.h"
37#include "pnfs.h"
38
39#include "nfstrace.h"
40
41#define NFSDBG_FACILITY NFSDBG_PAGECACHE
42
43#define MIN_POOL_WRITE (32)
44#define MIN_POOL_COMMIT (4)
45
46struct nfs_io_completion {
47 void (*complete)(void *data);
48 void *data;
49 struct kref refcount;
50};
51
52/*
53 * Local function declarations
54 */
55static void nfs_redirty_request(struct nfs_page *req);
56static const struct rpc_call_ops nfs_commit_ops;
57static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
58static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
59static const struct nfs_rw_ops nfs_rw_write_ops;
60static void nfs_inode_remove_request(struct nfs_page *req);
61static void nfs_clear_request_commit(struct nfs_page *req);
62static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
63 struct inode *inode);
64static struct nfs_page *
65nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
66 struct page *page);
67
68static struct kmem_cache *nfs_wdata_cachep;
69static mempool_t *nfs_wdata_mempool;
70static struct kmem_cache *nfs_cdata_cachep;
71static mempool_t *nfs_commit_mempool;
72
73struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
74{
75 struct nfs_commit_data *p;
76
77 if (never_fail)
78 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
79 else {
80 /* It is OK to do some reclaim, not no safe to wait
81 * for anything to be returned to the pool.
82 * mempool_alloc() cannot handle that particular combination,
83 * so we need two separate attempts.
84 */
85 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
86 if (!p)
87 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
88 __GFP_NOWARN | __GFP_NORETRY);
89 if (!p)
90 return NULL;
91 }
92
93 memset(p, 0, sizeof(*p));
94 INIT_LIST_HEAD(&p->pages);
95 return p;
96}
97EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
98
99void nfs_commit_free(struct nfs_commit_data *p)
100{
101 mempool_free(p, nfs_commit_mempool);
102}
103EXPORT_SYMBOL_GPL(nfs_commit_free);
104
105static struct nfs_pgio_header *nfs_writehdr_alloc(void)
106{
107 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_KERNEL);
108
109 memset(p, 0, sizeof(*p));
110 p->rw_mode = FMODE_WRITE;
111 return p;
112}
113
114static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
115{
116 mempool_free(hdr, nfs_wdata_mempool);
117}
118
119static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
120{
121 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
122}
123
124static void nfs_io_completion_init(struct nfs_io_completion *ioc,
125 void (*complete)(void *), void *data)
126{
127 ioc->complete = complete;
128 ioc->data = data;
129 kref_init(&ioc->refcount);
130}
131
132static void nfs_io_completion_release(struct kref *kref)
133{
134 struct nfs_io_completion *ioc = container_of(kref,
135 struct nfs_io_completion, refcount);
136 ioc->complete(ioc->data);
137 kfree(ioc);
138}
139
140static void nfs_io_completion_get(struct nfs_io_completion *ioc)
141{
142 if (ioc != NULL)
143 kref_get(&ioc->refcount);
144}
145
146static void nfs_io_completion_put(struct nfs_io_completion *ioc)
147{
148 if (ioc != NULL)
149 kref_put(&ioc->refcount, nfs_io_completion_release);
150}
151
152static struct nfs_page *
153nfs_page_private_request(struct page *page)
154{
155 if (!PagePrivate(page))
156 return NULL;
157 return (struct nfs_page *)page_private(page);
158}
159
160/*
161 * nfs_page_find_head_request_locked - find head request associated with @page
162 *
163 * must be called while holding the inode lock.
164 *
165 * returns matching head request with reference held, or NULL if not found.
166 */
167static struct nfs_page *
168nfs_page_find_private_request(struct page *page)
169{
170 struct address_space *mapping = page_file_mapping(page);
171 struct nfs_page *req;
172
173 if (!PagePrivate(page))
174 return NULL;
175 spin_lock(&mapping->private_lock);
176 req = nfs_page_private_request(page);
177 if (req) {
178 WARN_ON_ONCE(req->wb_head != req);
179 kref_get(&req->wb_kref);
180 }
181 spin_unlock(&mapping->private_lock);
182 return req;
183}
184
185static struct nfs_page *
186nfs_page_find_swap_request(struct page *page)
187{
188 struct inode *inode = page_file_mapping(page)->host;
189 struct nfs_inode *nfsi = NFS_I(inode);
190 struct nfs_page *req = NULL;
191 if (!PageSwapCache(page))
192 return NULL;
193 mutex_lock(&nfsi->commit_mutex);
194 if (PageSwapCache(page)) {
195 req = nfs_page_search_commits_for_head_request_locked(nfsi,
196 page);
197 if (req) {
198 WARN_ON_ONCE(req->wb_head != req);
199 kref_get(&req->wb_kref);
200 }
201 }
202 mutex_unlock(&nfsi->commit_mutex);
203 return req;
204}
205
206/*
207 * nfs_page_find_head_request - find head request associated with @page
208 *
209 * returns matching head request with reference held, or NULL if not found.
210 */
211static struct nfs_page *nfs_page_find_head_request(struct page *page)
212{
213 struct nfs_page *req;
214
215 req = nfs_page_find_private_request(page);
216 if (!req)
217 req = nfs_page_find_swap_request(page);
218 return req;
219}
220
221/* Adjust the file length if we're writing beyond the end */
222static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
223{
224 struct inode *inode = page_file_mapping(page)->host;
225 loff_t end, i_size;
226 pgoff_t end_index;
227
228 spin_lock(&inode->i_lock);
229 i_size = i_size_read(inode);
230 end_index = (i_size - 1) >> PAGE_SHIFT;
231 if (i_size > 0 && page_index(page) < end_index)
232 goto out;
233 end = page_file_offset(page) + ((loff_t)offset+count);
234 if (i_size >= end)
235 goto out;
236 i_size_write(inode, end);
237 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
238 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
239out:
240 spin_unlock(&inode->i_lock);
241}
242
243/* A writeback failed: mark the page as bad, and invalidate the page cache */
244static void nfs_set_pageerror(struct address_space *mapping)
245{
246 nfs_zap_mapping(mapping->host, mapping);
247}
248
249static void nfs_mapping_set_error(struct page *page, int error)
250{
251 SetPageError(page);
252 mapping_set_error(page_file_mapping(page), error);
253}
254
255/*
256 * nfs_page_group_search_locked
257 * @head - head request of page group
258 * @page_offset - offset into page
259 *
260 * Search page group with head @head to find a request that contains the
261 * page offset @page_offset.
262 *
263 * Returns a pointer to the first matching nfs request, or NULL if no
264 * match is found.
265 *
266 * Must be called with the page group lock held
267 */
268static struct nfs_page *
269nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
270{
271 struct nfs_page *req;
272
273 req = head;
274 do {
275 if (page_offset >= req->wb_pgbase &&
276 page_offset < (req->wb_pgbase + req->wb_bytes))
277 return req;
278
279 req = req->wb_this_page;
280 } while (req != head);
281
282 return NULL;
283}
284
285/*
286 * nfs_page_group_covers_page
287 * @head - head request of page group
288 *
289 * Return true if the page group with head @head covers the whole page,
290 * returns false otherwise
291 */
292static bool nfs_page_group_covers_page(struct nfs_page *req)
293{
294 struct nfs_page *tmp;
295 unsigned int pos = 0;
296 unsigned int len = nfs_page_length(req->wb_page);
297
298 nfs_page_group_lock(req);
299
300 for (;;) {
301 tmp = nfs_page_group_search_locked(req->wb_head, pos);
302 if (!tmp)
303 break;
304 pos = tmp->wb_pgbase + tmp->wb_bytes;
305 }
306
307 nfs_page_group_unlock(req);
308 return pos >= len;
309}
310
311/* We can set the PG_uptodate flag if we see that a write request
312 * covers the full page.
313 */
314static void nfs_mark_uptodate(struct nfs_page *req)
315{
316 if (PageUptodate(req->wb_page))
317 return;
318 if (!nfs_page_group_covers_page(req))
319 return;
320 SetPageUptodate(req->wb_page);
321}
322
323static int wb_priority(struct writeback_control *wbc)
324{
325 int ret = 0;
326
327 if (wbc->sync_mode == WB_SYNC_ALL)
328 ret = FLUSH_COND_STABLE;
329 return ret;
330}
331
332/*
333 * NFS congestion control
334 */
335
336int nfs_congestion_kb;
337
338#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
339#define NFS_CONGESTION_OFF_THRESH \
340 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
341
342static void nfs_set_page_writeback(struct page *page)
343{
344 struct inode *inode = page_file_mapping(page)->host;
345 struct nfs_server *nfss = NFS_SERVER(inode);
346 int ret = test_set_page_writeback(page);
347
348 WARN_ON_ONCE(ret != 0);
349
350 if (atomic_long_inc_return(&nfss->writeback) >
351 NFS_CONGESTION_ON_THRESH)
352 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
353}
354
355static void nfs_end_page_writeback(struct nfs_page *req)
356{
357 struct inode *inode = page_file_mapping(req->wb_page)->host;
358 struct nfs_server *nfss = NFS_SERVER(inode);
359 bool is_done;
360
361 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
362 nfs_unlock_request(req);
363 if (!is_done)
364 return;
365
366 end_page_writeback(req->wb_page);
367 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
368 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
369}
370
371/*
372 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
373 *
374 * this is a helper function for nfs_lock_and_join_requests
375 *
376 * @inode - inode associated with request page group, must be holding inode lock
377 * @head - head request of page group, must be holding head lock
378 * @req - request that couldn't lock and needs to wait on the req bit lock
379 *
380 * NOTE: this must be called holding page_group bit lock
381 * which will be released before returning.
382 *
383 * returns 0 on success, < 0 on error.
384 */
385static void
386nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
387 struct nfs_page *req)
388{
389 struct nfs_page *tmp;
390
391 /* relinquish all the locks successfully grabbed this run */
392 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
393 if (!kref_read(&tmp->wb_kref))
394 continue;
395 nfs_unlock_and_release_request(tmp);
396 }
397}
398
399/*
400 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
401 *
402 * @destroy_list - request list (using wb_this_page) terminated by @old_head
403 * @old_head - the old head of the list
404 *
405 * All subrequests must be locked and removed from all lists, so at this point
406 * they are only "active" in this function, and possibly in nfs_wait_on_request
407 * with a reference held by some other context.
408 */
409static void
410nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
411 struct nfs_page *old_head,
412 struct inode *inode)
413{
414 while (destroy_list) {
415 struct nfs_page *subreq = destroy_list;
416
417 destroy_list = (subreq->wb_this_page == old_head) ?
418 NULL : subreq->wb_this_page;
419
420 WARN_ON_ONCE(old_head != subreq->wb_head);
421
422 /* make sure old group is not used */
423 subreq->wb_this_page = subreq;
424
425 clear_bit(PG_REMOVE, &subreq->wb_flags);
426
427 /* Note: races with nfs_page_group_destroy() */
428 if (!kref_read(&subreq->wb_kref)) {
429 /* Check if we raced with nfs_page_group_destroy() */
430 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
431 nfs_free_request(subreq);
432 continue;
433 }
434
435 subreq->wb_head = subreq;
436
437 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
438 nfs_release_request(subreq);
439 atomic_long_dec(&NFS_I(inode)->nrequests);
440 }
441
442 /* subreq is now totally disconnected from page group or any
443 * write / commit lists. last chance to wake any waiters */
444 nfs_unlock_and_release_request(subreq);
445 }
446}
447
448/*
449 * nfs_lock_and_join_requests - join all subreqs to the head req and return
450 * a locked reference, cancelling any pending
451 * operations for this page.
452 *
453 * @page - the page used to lookup the "page group" of nfs_page structures
454 *
455 * This function joins all sub requests to the head request by first
456 * locking all requests in the group, cancelling any pending operations
457 * and finally updating the head request to cover the whole range covered by
458 * the (former) group. All subrequests are removed from any write or commit
459 * lists, unlinked from the group and destroyed.
460 *
461 * Returns a locked, referenced pointer to the head request - which after
462 * this call is guaranteed to be the only request associated with the page.
463 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
464 * error was encountered.
465 */
466static struct nfs_page *
467nfs_lock_and_join_requests(struct page *page)
468{
469 struct inode *inode = page_file_mapping(page)->host;
470 struct nfs_page *head, *subreq;
471 struct nfs_page *destroy_list = NULL;
472 unsigned int total_bytes;
473 int ret;
474
475try_again:
476 /*
477 * A reference is taken only on the head request which acts as a
478 * reference to the whole page group - the group will not be destroyed
479 * until the head reference is released.
480 */
481 head = nfs_page_find_head_request(page);
482 if (!head)
483 return NULL;
484
485 /* lock the page head first in order to avoid an ABBA inefficiency */
486 if (!nfs_lock_request(head)) {
487 ret = nfs_wait_on_request(head);
488 nfs_release_request(head);
489 if (ret < 0)
490 return ERR_PTR(ret);
491 goto try_again;
492 }
493
494 /* Ensure that nobody removed the request before we locked it */
495 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
496 nfs_unlock_and_release_request(head);
497 goto try_again;
498 }
499
500 ret = nfs_page_group_lock(head);
501 if (ret < 0)
502 goto release_request;
503
504 /* lock each request in the page group */
505 total_bytes = head->wb_bytes;
506 for (subreq = head->wb_this_page; subreq != head;
507 subreq = subreq->wb_this_page) {
508
509 if (!kref_get_unless_zero(&subreq->wb_kref)) {
510 if (subreq->wb_offset == head->wb_offset + total_bytes)
511 total_bytes += subreq->wb_bytes;
512 continue;
513 }
514
515 while (!nfs_lock_request(subreq)) {
516 /*
517 * Unlock page to allow nfs_page_group_sync_on_bit()
518 * to succeed
519 */
520 nfs_page_group_unlock(head);
521 ret = nfs_wait_on_request(subreq);
522 if (!ret)
523 ret = nfs_page_group_lock(head);
524 if (ret < 0) {
525 nfs_unroll_locks(inode, head, subreq);
526 nfs_release_request(subreq);
527 goto release_request;
528 }
529 }
530 /*
531 * Subrequests are always contiguous, non overlapping
532 * and in order - but may be repeated (mirrored writes).
533 */
534 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
535 /* keep track of how many bytes this group covers */
536 total_bytes += subreq->wb_bytes;
537 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
538 ((subreq->wb_offset + subreq->wb_bytes) >
539 (head->wb_offset + total_bytes)))) {
540 nfs_page_group_unlock(head);
541 nfs_unroll_locks(inode, head, subreq);
542 nfs_unlock_and_release_request(subreq);
543 ret = -EIO;
544 goto release_request;
545 }
546 }
547
548 /* Now that all requests are locked, make sure they aren't on any list.
549 * Commit list removal accounting is done after locks are dropped */
550 subreq = head;
551 do {
552 nfs_clear_request_commit(subreq);
553 subreq = subreq->wb_this_page;
554 } while (subreq != head);
555
556 /* unlink subrequests from head, destroy them later */
557 if (head->wb_this_page != head) {
558 /* destroy list will be terminated by head */
559 destroy_list = head->wb_this_page;
560 head->wb_this_page = head;
561
562 /* change head request to cover whole range that
563 * the former page group covered */
564 head->wb_bytes = total_bytes;
565 }
566
567 /* Postpone destruction of this request */
568 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
569 set_bit(PG_INODE_REF, &head->wb_flags);
570 kref_get(&head->wb_kref);
571 atomic_long_inc(&NFS_I(inode)->nrequests);
572 }
573
574 nfs_page_group_unlock(head);
575
576 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
577
578 /* Did we lose a race with nfs_inode_remove_request()? */
579 if (!(PagePrivate(page) || PageSwapCache(page))) {
580 nfs_unlock_and_release_request(head);
581 return NULL;
582 }
583
584 /* still holds ref on head from nfs_page_find_head_request
585 * and still has lock on head from lock loop */
586 return head;
587
588release_request:
589 nfs_unlock_and_release_request(head);
590 return ERR_PTR(ret);
591}
592
593static void nfs_write_error(struct nfs_page *req, int error)
594{
595 nfs_set_pageerror(page_file_mapping(req->wb_page));
596 nfs_mapping_set_error(req->wb_page, error);
597 nfs_inode_remove_request(req);
598 nfs_end_page_writeback(req);
599 nfs_release_request(req);
600}
601
602/*
603 * Find an associated nfs write request, and prepare to flush it out
604 * May return an error if the user signalled nfs_wait_on_request().
605 */
606static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
607 struct page *page)
608{
609 struct nfs_page *req;
610 int ret = 0;
611
612 req = nfs_lock_and_join_requests(page);
613 if (!req)
614 goto out;
615 ret = PTR_ERR(req);
616 if (IS_ERR(req))
617 goto out;
618
619 nfs_set_page_writeback(page);
620 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
621
622 /* If there is a fatal error that covers this write, just exit */
623 ret = pgio->pg_error;
624 if (nfs_error_is_fatal_on_server(ret))
625 goto out_launder;
626
627 ret = 0;
628 if (!nfs_pageio_add_request(pgio, req)) {
629 ret = pgio->pg_error;
630 /*
631 * Remove the problematic req upon fatal errors on the server
632 */
633 if (nfs_error_is_fatal(ret)) {
634 if (nfs_error_is_fatal_on_server(ret))
635 goto out_launder;
636 } else
637 ret = -EAGAIN;
638 nfs_redirty_request(req);
639 pgio->pg_error = 0;
640 } else
641 nfs_add_stats(page_file_mapping(page)->host,
642 NFSIOS_WRITEPAGES, 1);
643out:
644 return ret;
645out_launder:
646 nfs_write_error(req, ret);
647 return 0;
648}
649
650static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
651 struct nfs_pageio_descriptor *pgio)
652{
653 int ret;
654
655 nfs_pageio_cond_complete(pgio, page_index(page));
656 ret = nfs_page_async_flush(pgio, page);
657 if (ret == -EAGAIN) {
658 redirty_page_for_writepage(wbc, page);
659 ret = AOP_WRITEPAGE_ACTIVATE;
660 }
661 return ret;
662}
663
664/*
665 * Write an mmapped page to the server.
666 */
667static int nfs_writepage_locked(struct page *page,
668 struct writeback_control *wbc)
669{
670 struct nfs_pageio_descriptor pgio;
671 struct inode *inode = page_file_mapping(page)->host;
672 int err;
673
674 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
675 nfs_pageio_init_write(&pgio, inode, 0,
676 false, &nfs_async_write_completion_ops);
677 err = nfs_do_writepage(page, wbc, &pgio);
678 pgio.pg_error = 0;
679 nfs_pageio_complete(&pgio);
680 if (err < 0)
681 return err;
682 if (nfs_error_is_fatal(pgio.pg_error))
683 return pgio.pg_error;
684 return 0;
685}
686
687int nfs_writepage(struct page *page, struct writeback_control *wbc)
688{
689 int ret;
690
691 ret = nfs_writepage_locked(page, wbc);
692 if (ret != AOP_WRITEPAGE_ACTIVATE)
693 unlock_page(page);
694 return ret;
695}
696
697static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
698{
699 int ret;
700
701 ret = nfs_do_writepage(page, wbc, data);
702 if (ret != AOP_WRITEPAGE_ACTIVATE)
703 unlock_page(page);
704 return ret;
705}
706
707static void nfs_io_completion_commit(void *inode)
708{
709 nfs_commit_inode(inode, 0);
710}
711
712int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
713{
714 struct inode *inode = mapping->host;
715 struct nfs_pageio_descriptor pgio;
716 struct nfs_io_completion *ioc;
717 int err;
718
719 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
720
721 ioc = nfs_io_completion_alloc(GFP_KERNEL);
722 if (ioc)
723 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
724
725 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
726 &nfs_async_write_completion_ops);
727 pgio.pg_io_completion = ioc;
728 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
729 pgio.pg_error = 0;
730 nfs_pageio_complete(&pgio);
731 nfs_io_completion_put(ioc);
732
733 if (err < 0)
734 goto out_err;
735 err = pgio.pg_error;
736 if (nfs_error_is_fatal(err))
737 goto out_err;
738 return 0;
739out_err:
740 return err;
741}
742
743/*
744 * Insert a write request into an inode
745 */
746static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
747{
748 struct address_space *mapping = page_file_mapping(req->wb_page);
749 struct nfs_inode *nfsi = NFS_I(inode);
750
751 WARN_ON_ONCE(req->wb_this_page != req);
752
753 /* Lock the request! */
754 nfs_lock_request(req);
755
756 /*
757 * Swap-space should not get truncated. Hence no need to plug the race
758 * with invalidate/truncate.
759 */
760 spin_lock(&mapping->private_lock);
761 if (!nfs_have_writebacks(inode) &&
762 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
763 inode_inc_iversion_raw(inode);
764 if (likely(!PageSwapCache(req->wb_page))) {
765 set_bit(PG_MAPPED, &req->wb_flags);
766 SetPagePrivate(req->wb_page);
767 set_page_private(req->wb_page, (unsigned long)req);
768 }
769 spin_unlock(&mapping->private_lock);
770 atomic_long_inc(&nfsi->nrequests);
771 /* this a head request for a page group - mark it as having an
772 * extra reference so sub groups can follow suit.
773 * This flag also informs pgio layer when to bump nrequests when
774 * adding subrequests. */
775 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
776 kref_get(&req->wb_kref);
777}
778
779/*
780 * Remove a write request from an inode
781 */
782static void nfs_inode_remove_request(struct nfs_page *req)
783{
784 struct address_space *mapping = page_file_mapping(req->wb_page);
785 struct inode *inode = mapping->host;
786 struct nfs_inode *nfsi = NFS_I(inode);
787 struct nfs_page *head;
788
789 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
790 head = req->wb_head;
791
792 spin_lock(&mapping->private_lock);
793 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
794 set_page_private(head->wb_page, 0);
795 ClearPagePrivate(head->wb_page);
796 clear_bit(PG_MAPPED, &head->wb_flags);
797 }
798 spin_unlock(&mapping->private_lock);
799 }
800
801 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
802 nfs_release_request(req);
803 atomic_long_dec(&nfsi->nrequests);
804 }
805}
806
807static void
808nfs_mark_request_dirty(struct nfs_page *req)
809{
810 if (req->wb_page)
811 __set_page_dirty_nobuffers(req->wb_page);
812}
813
814/*
815 * nfs_page_search_commits_for_head_request_locked
816 *
817 * Search through commit lists on @inode for the head request for @page.
818 * Must be called while holding the inode (which is cinfo) lock.
819 *
820 * Returns the head request if found, or NULL if not found.
821 */
822static struct nfs_page *
823nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
824 struct page *page)
825{
826 struct nfs_page *freq, *t;
827 struct nfs_commit_info cinfo;
828 struct inode *inode = &nfsi->vfs_inode;
829
830 nfs_init_cinfo_from_inode(&cinfo, inode);
831
832 /* search through pnfs commit lists */
833 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
834 if (freq)
835 return freq->wb_head;
836
837 /* Linearly search the commit list for the correct request */
838 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
839 if (freq->wb_page == page)
840 return freq->wb_head;
841 }
842
843 return NULL;
844}
845
846/**
847 * nfs_request_add_commit_list_locked - add request to a commit list
848 * @req: pointer to a struct nfs_page
849 * @dst: commit list head
850 * @cinfo: holds list lock and accounting info
851 *
852 * This sets the PG_CLEAN bit, updates the cinfo count of
853 * number of outstanding requests requiring a commit as well as
854 * the MM page stats.
855 *
856 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
857 * nfs_page lock.
858 */
859void
860nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
861 struct nfs_commit_info *cinfo)
862{
863 set_bit(PG_CLEAN, &req->wb_flags);
864 nfs_list_add_request(req, dst);
865 atomic_long_inc(&cinfo->mds->ncommit);
866}
867EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
868
869/**
870 * nfs_request_add_commit_list - add request to a commit list
871 * @req: pointer to a struct nfs_page
872 * @cinfo: holds list lock and accounting info
873 *
874 * This sets the PG_CLEAN bit, updates the cinfo count of
875 * number of outstanding requests requiring a commit as well as
876 * the MM page stats.
877 *
878 * The caller must _not_ hold the cinfo->lock, but must be
879 * holding the nfs_page lock.
880 */
881void
882nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
883{
884 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
885 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
886 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
887 if (req->wb_page)
888 nfs_mark_page_unstable(req->wb_page, cinfo);
889}
890EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
891
892/**
893 * nfs_request_remove_commit_list - Remove request from a commit list
894 * @req: pointer to a nfs_page
895 * @cinfo: holds list lock and accounting info
896 *
897 * This clears the PG_CLEAN bit, and updates the cinfo's count of
898 * number of outstanding requests requiring a commit
899 * It does not update the MM page stats.
900 *
901 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
902 */
903void
904nfs_request_remove_commit_list(struct nfs_page *req,
905 struct nfs_commit_info *cinfo)
906{
907 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
908 return;
909 nfs_list_remove_request(req);
910 atomic_long_dec(&cinfo->mds->ncommit);
911}
912EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
913
914static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
915 struct inode *inode)
916{
917 cinfo->inode = inode;
918 cinfo->mds = &NFS_I(inode)->commit_info;
919 cinfo->ds = pnfs_get_ds_info(inode);
920 cinfo->dreq = NULL;
921 cinfo->completion_ops = &nfs_commit_completion_ops;
922}
923
924void nfs_init_cinfo(struct nfs_commit_info *cinfo,
925 struct inode *inode,
926 struct nfs_direct_req *dreq)
927{
928 if (dreq)
929 nfs_init_cinfo_from_dreq(cinfo, dreq);
930 else
931 nfs_init_cinfo_from_inode(cinfo, inode);
932}
933EXPORT_SYMBOL_GPL(nfs_init_cinfo);
934
935/*
936 * Add a request to the inode's commit list.
937 */
938void
939nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
940 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
941{
942 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
943 return;
944 nfs_request_add_commit_list(req, cinfo);
945}
946
947static void
948nfs_clear_page_commit(struct page *page)
949{
950 dec_node_page_state(page, NR_UNSTABLE_NFS);
951 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
952 WB_RECLAIMABLE);
953}
954
955/* Called holding the request lock on @req */
956static void
957nfs_clear_request_commit(struct nfs_page *req)
958{
959 if (test_bit(PG_CLEAN, &req->wb_flags)) {
960 struct nfs_open_context *ctx = nfs_req_openctx(req);
961 struct inode *inode = d_inode(ctx->dentry);
962 struct nfs_commit_info cinfo;
963
964 nfs_init_cinfo_from_inode(&cinfo, inode);
965 mutex_lock(&NFS_I(inode)->commit_mutex);
966 if (!pnfs_clear_request_commit(req, &cinfo)) {
967 nfs_request_remove_commit_list(req, &cinfo);
968 }
969 mutex_unlock(&NFS_I(inode)->commit_mutex);
970 nfs_clear_page_commit(req->wb_page);
971 }
972}
973
974int nfs_write_need_commit(struct nfs_pgio_header *hdr)
975{
976 if (hdr->verf.committed == NFS_DATA_SYNC)
977 return hdr->lseg == NULL;
978 return hdr->verf.committed != NFS_FILE_SYNC;
979}
980
981static void nfs_async_write_init(struct nfs_pgio_header *hdr)
982{
983 nfs_io_completion_get(hdr->io_completion);
984}
985
986static void nfs_write_completion(struct nfs_pgio_header *hdr)
987{
988 struct nfs_commit_info cinfo;
989 unsigned long bytes = 0;
990
991 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
992 goto out;
993 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
994 while (!list_empty(&hdr->pages)) {
995 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
996
997 bytes += req->wb_bytes;
998 nfs_list_remove_request(req);
999 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1000 (hdr->good_bytes < bytes)) {
1001 nfs_set_pageerror(page_file_mapping(req->wb_page));
1002 nfs_mapping_set_error(req->wb_page, hdr->error);
1003 goto remove_req;
1004 }
1005 if (nfs_write_need_commit(hdr)) {
1006 /* Reset wb_nio, since the write was successful. */
1007 req->wb_nio = 0;
1008 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1009 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1010 hdr->pgio_mirror_idx);
1011 goto next;
1012 }
1013remove_req:
1014 nfs_inode_remove_request(req);
1015next:
1016 nfs_end_page_writeback(req);
1017 nfs_release_request(req);
1018 }
1019out:
1020 nfs_io_completion_put(hdr->io_completion);
1021 hdr->release(hdr);
1022}
1023
1024unsigned long
1025nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1026{
1027 return atomic_long_read(&cinfo->mds->ncommit);
1028}
1029
1030/* NFS_I(cinfo->inode)->commit_mutex held by caller */
1031int
1032nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1033 struct nfs_commit_info *cinfo, int max)
1034{
1035 struct nfs_page *req, *tmp;
1036 int ret = 0;
1037
1038restart:
1039 list_for_each_entry_safe(req, tmp, src, wb_list) {
1040 kref_get(&req->wb_kref);
1041 if (!nfs_lock_request(req)) {
1042 int status;
1043
1044 /* Prevent deadlock with nfs_lock_and_join_requests */
1045 if (!list_empty(dst)) {
1046 nfs_release_request(req);
1047 continue;
1048 }
1049 /* Ensure we make progress to prevent livelock */
1050 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1051 status = nfs_wait_on_request(req);
1052 nfs_release_request(req);
1053 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1054 if (status < 0)
1055 break;
1056 goto restart;
1057 }
1058 nfs_request_remove_commit_list(req, cinfo);
1059 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1060 nfs_list_add_request(req, dst);
1061 ret++;
1062 if ((ret == max) && !cinfo->dreq)
1063 break;
1064 cond_resched();
1065 }
1066 return ret;
1067}
1068EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1069
1070/*
1071 * nfs_scan_commit - Scan an inode for commit requests
1072 * @inode: NFS inode to scan
1073 * @dst: mds destination list
1074 * @cinfo: mds and ds lists of reqs ready to commit
1075 *
1076 * Moves requests from the inode's 'commit' request list.
1077 * The requests are *not* checked to ensure that they form a contiguous set.
1078 */
1079int
1080nfs_scan_commit(struct inode *inode, struct list_head *dst,
1081 struct nfs_commit_info *cinfo)
1082{
1083 int ret = 0;
1084
1085 if (!atomic_long_read(&cinfo->mds->ncommit))
1086 return 0;
1087 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1088 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1089 const int max = INT_MAX;
1090
1091 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1092 cinfo, max);
1093 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1094 }
1095 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1096 return ret;
1097}
1098
1099/*
1100 * Search for an existing write request, and attempt to update
1101 * it to reflect a new dirty region on a given page.
1102 *
1103 * If the attempt fails, then the existing request is flushed out
1104 * to disk.
1105 */
1106static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1107 struct page *page,
1108 unsigned int offset,
1109 unsigned int bytes)
1110{
1111 struct nfs_page *req;
1112 unsigned int rqend;
1113 unsigned int end;
1114 int error;
1115
1116 end = offset + bytes;
1117
1118 req = nfs_lock_and_join_requests(page);
1119 if (IS_ERR_OR_NULL(req))
1120 return req;
1121
1122 rqend = req->wb_offset + req->wb_bytes;
1123 /*
1124 * Tell the caller to flush out the request if
1125 * the offsets are non-contiguous.
1126 * Note: nfs_flush_incompatible() will already
1127 * have flushed out requests having wrong owners.
1128 */
1129 if (offset > rqend || end < req->wb_offset)
1130 goto out_flushme;
1131
1132 /* Okay, the request matches. Update the region */
1133 if (offset < req->wb_offset) {
1134 req->wb_offset = offset;
1135 req->wb_pgbase = offset;
1136 }
1137 if (end > rqend)
1138 req->wb_bytes = end - req->wb_offset;
1139 else
1140 req->wb_bytes = rqend - req->wb_offset;
1141 req->wb_nio = 0;
1142 return req;
1143out_flushme:
1144 /*
1145 * Note: we mark the request dirty here because
1146 * nfs_lock_and_join_requests() cannot preserve
1147 * commit flags, so we have to replay the write.
1148 */
1149 nfs_mark_request_dirty(req);
1150 nfs_unlock_and_release_request(req);
1151 error = nfs_wb_page(inode, page);
1152 return (error < 0) ? ERR_PTR(error) : NULL;
1153}
1154
1155/*
1156 * Try to update an existing write request, or create one if there is none.
1157 *
1158 * Note: Should always be called with the Page Lock held to prevent races
1159 * if we have to add a new request. Also assumes that the caller has
1160 * already called nfs_flush_incompatible() if necessary.
1161 */
1162static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1163 struct page *page, unsigned int offset, unsigned int bytes)
1164{
1165 struct inode *inode = page_file_mapping(page)->host;
1166 struct nfs_page *req;
1167
1168 req = nfs_try_to_update_request(inode, page, offset, bytes);
1169 if (req != NULL)
1170 goto out;
1171 req = nfs_create_request(ctx, page, offset, bytes);
1172 if (IS_ERR(req))
1173 goto out;
1174 nfs_inode_add_request(inode, req);
1175out:
1176 return req;
1177}
1178
1179static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1180 unsigned int offset, unsigned int count)
1181{
1182 struct nfs_page *req;
1183
1184 req = nfs_setup_write_request(ctx, page, offset, count);
1185 if (IS_ERR(req))
1186 return PTR_ERR(req);
1187 /* Update file length */
1188 nfs_grow_file(page, offset, count);
1189 nfs_mark_uptodate(req);
1190 nfs_mark_request_dirty(req);
1191 nfs_unlock_and_release_request(req);
1192 return 0;
1193}
1194
1195int nfs_flush_incompatible(struct file *file, struct page *page)
1196{
1197 struct nfs_open_context *ctx = nfs_file_open_context(file);
1198 struct nfs_lock_context *l_ctx;
1199 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1200 struct nfs_page *req;
1201 int do_flush, status;
1202 /*
1203 * Look for a request corresponding to this page. If there
1204 * is one, and it belongs to another file, we flush it out
1205 * before we try to copy anything into the page. Do this
1206 * due to the lack of an ACCESS-type call in NFSv2.
1207 * Also do the same if we find a request from an existing
1208 * dropped page.
1209 */
1210 do {
1211 req = nfs_page_find_head_request(page);
1212 if (req == NULL)
1213 return 0;
1214 l_ctx = req->wb_lock_context;
1215 do_flush = req->wb_page != page ||
1216 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1217 if (l_ctx && flctx &&
1218 !(list_empty_careful(&flctx->flc_posix) &&
1219 list_empty_careful(&flctx->flc_flock))) {
1220 do_flush |= l_ctx->lockowner != current->files;
1221 }
1222 nfs_release_request(req);
1223 if (!do_flush)
1224 return 0;
1225 status = nfs_wb_page(page_file_mapping(page)->host, page);
1226 } while (status == 0);
1227 return status;
1228}
1229
1230/*
1231 * Avoid buffered writes when a open context credential's key would
1232 * expire soon.
1233 *
1234 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1235 *
1236 * Return 0 and set a credential flag which triggers the inode to flush
1237 * and performs NFS_FILE_SYNC writes if the key will expired within
1238 * RPC_KEY_EXPIRE_TIMEO.
1239 */
1240int
1241nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1242{
1243 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1244
1245 if (nfs_ctx_key_to_expire(ctx, inode) &&
1246 !ctx->ll_cred)
1247 /* Already expired! */
1248 return -EACCES;
1249 return 0;
1250}
1251
1252/*
1253 * Test if the open context credential key is marked to expire soon.
1254 */
1255bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1256{
1257 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1258 struct rpc_cred *cred = ctx->ll_cred;
1259 struct auth_cred acred = {
1260 .cred = ctx->cred,
1261 };
1262
1263 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1264 put_rpccred(cred);
1265 ctx->ll_cred = NULL;
1266 cred = NULL;
1267 }
1268 if (!cred)
1269 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1270 if (!cred || IS_ERR(cred))
1271 return true;
1272 ctx->ll_cred = cred;
1273 return !!(cred->cr_ops->crkey_timeout &&
1274 cred->cr_ops->crkey_timeout(cred));
1275}
1276
1277/*
1278 * If the page cache is marked as unsafe or invalid, then we can't rely on
1279 * the PageUptodate() flag. In this case, we will need to turn off
1280 * write optimisations that depend on the page contents being correct.
1281 */
1282static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1283{
1284 struct nfs_inode *nfsi = NFS_I(inode);
1285
1286 if (nfs_have_delegated_attributes(inode))
1287 goto out;
1288 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1289 return false;
1290 smp_rmb();
1291 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1292 return false;
1293out:
1294 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1295 return false;
1296 return PageUptodate(page) != 0;
1297}
1298
1299static bool
1300is_whole_file_wrlock(struct file_lock *fl)
1301{
1302 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1303 fl->fl_type == F_WRLCK;
1304}
1305
1306/* If we know the page is up to date, and we're not using byte range locks (or
1307 * if we have the whole file locked for writing), it may be more efficient to
1308 * extend the write to cover the entire page in order to avoid fragmentation
1309 * inefficiencies.
1310 *
1311 * If the file is opened for synchronous writes then we can just skip the rest
1312 * of the checks.
1313 */
1314static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1315{
1316 int ret;
1317 struct file_lock_context *flctx = inode->i_flctx;
1318 struct file_lock *fl;
1319
1320 if (file->f_flags & O_DSYNC)
1321 return 0;
1322 if (!nfs_write_pageuptodate(page, inode))
1323 return 0;
1324 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1325 return 1;
1326 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1327 list_empty_careful(&flctx->flc_posix)))
1328 return 1;
1329
1330 /* Check to see if there are whole file write locks */
1331 ret = 0;
1332 spin_lock(&flctx->flc_lock);
1333 if (!list_empty(&flctx->flc_posix)) {
1334 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1335 fl_list);
1336 if (is_whole_file_wrlock(fl))
1337 ret = 1;
1338 } else if (!list_empty(&flctx->flc_flock)) {
1339 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1340 fl_list);
1341 if (fl->fl_type == F_WRLCK)
1342 ret = 1;
1343 }
1344 spin_unlock(&flctx->flc_lock);
1345 return ret;
1346}
1347
1348/*
1349 * Update and possibly write a cached page of an NFS file.
1350 *
1351 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1352 * things with a page scheduled for an RPC call (e.g. invalidate it).
1353 */
1354int nfs_updatepage(struct file *file, struct page *page,
1355 unsigned int offset, unsigned int count)
1356{
1357 struct nfs_open_context *ctx = nfs_file_open_context(file);
1358 struct address_space *mapping = page_file_mapping(page);
1359 struct inode *inode = mapping->host;
1360 int status = 0;
1361
1362 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1363
1364 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1365 file, count, (long long)(page_file_offset(page) + offset));
1366
1367 if (!count)
1368 goto out;
1369
1370 if (nfs_can_extend_write(file, page, inode)) {
1371 count = max(count + offset, nfs_page_length(page));
1372 offset = 0;
1373 }
1374
1375 status = nfs_writepage_setup(ctx, page, offset, count);
1376 if (status < 0)
1377 nfs_set_pageerror(mapping);
1378 else
1379 __set_page_dirty_nobuffers(page);
1380out:
1381 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1382 status, (long long)i_size_read(inode));
1383 return status;
1384}
1385
1386static int flush_task_priority(int how)
1387{
1388 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1389 case FLUSH_HIGHPRI:
1390 return RPC_PRIORITY_HIGH;
1391 case FLUSH_LOWPRI:
1392 return RPC_PRIORITY_LOW;
1393 }
1394 return RPC_PRIORITY_NORMAL;
1395}
1396
1397static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1398 struct rpc_message *msg,
1399 const struct nfs_rpc_ops *rpc_ops,
1400 struct rpc_task_setup *task_setup_data, int how)
1401{
1402 int priority = flush_task_priority(how);
1403
1404 task_setup_data->priority = priority;
1405 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1406 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1407 hdr->args.stable);
1408}
1409
1410/* If a nfs_flush_* function fails, it should remove reqs from @head and
1411 * call this on each, which will prepare them to be retried on next
1412 * writeback using standard nfs.
1413 */
1414static void nfs_redirty_request(struct nfs_page *req)
1415{
1416 /* Bump the transmission count */
1417 req->wb_nio++;
1418 nfs_mark_request_dirty(req);
1419 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1420 nfs_end_page_writeback(req);
1421 nfs_release_request(req);
1422}
1423
1424static void nfs_async_write_error(struct list_head *head, int error)
1425{
1426 struct nfs_page *req;
1427
1428 while (!list_empty(head)) {
1429 req = nfs_list_entry(head->next);
1430 nfs_list_remove_request(req);
1431 if (nfs_error_is_fatal(error))
1432 nfs_write_error(req, error);
1433 else
1434 nfs_redirty_request(req);
1435 }
1436}
1437
1438static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1439{
1440 nfs_async_write_error(&hdr->pages, 0);
1441 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1442 hdr->args.offset + hdr->args.count - 1);
1443}
1444
1445static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1446 .init_hdr = nfs_async_write_init,
1447 .error_cleanup = nfs_async_write_error,
1448 .completion = nfs_write_completion,
1449 .reschedule_io = nfs_async_write_reschedule_io,
1450};
1451
1452void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1453 struct inode *inode, int ioflags, bool force_mds,
1454 const struct nfs_pgio_completion_ops *compl_ops)
1455{
1456 struct nfs_server *server = NFS_SERVER(inode);
1457 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1458
1459#ifdef CONFIG_NFS_V4_1
1460 if (server->pnfs_curr_ld && !force_mds)
1461 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1462#endif
1463 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1464 server->wsize, ioflags);
1465}
1466EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1467
1468void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1469{
1470 struct nfs_pgio_mirror *mirror;
1471
1472 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1473 pgio->pg_ops->pg_cleanup(pgio);
1474
1475 pgio->pg_ops = &nfs_pgio_rw_ops;
1476
1477 nfs_pageio_stop_mirroring(pgio);
1478
1479 mirror = &pgio->pg_mirrors[0];
1480 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1481}
1482EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1483
1484
1485void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1486{
1487 struct nfs_commit_data *data = calldata;
1488
1489 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1490}
1491
1492/*
1493 * Special version of should_remove_suid() that ignores capabilities.
1494 */
1495static int nfs_should_remove_suid(const struct inode *inode)
1496{
1497 umode_t mode = inode->i_mode;
1498 int kill = 0;
1499
1500 /* suid always must be killed */
1501 if (unlikely(mode & S_ISUID))
1502 kill = ATTR_KILL_SUID;
1503
1504 /*
1505 * sgid without any exec bits is just a mandatory locking mark; leave
1506 * it alone. If some exec bits are set, it's a real sgid; kill it.
1507 */
1508 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1509 kill |= ATTR_KILL_SGID;
1510
1511 if (unlikely(kill && S_ISREG(mode)))
1512 return kill;
1513
1514 return 0;
1515}
1516
1517static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1518 struct nfs_fattr *fattr)
1519{
1520 struct nfs_pgio_args *argp = &hdr->args;
1521 struct nfs_pgio_res *resp = &hdr->res;
1522 u64 size = argp->offset + resp->count;
1523
1524 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1525 fattr->size = size;
1526 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1527 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1528 return;
1529 }
1530 if (size != fattr->size)
1531 return;
1532 /* Set attribute barrier */
1533 nfs_fattr_set_barrier(fattr);
1534 /* ...and update size */
1535 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1536}
1537
1538void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1539{
1540 struct nfs_fattr *fattr = &hdr->fattr;
1541 struct inode *inode = hdr->inode;
1542
1543 spin_lock(&inode->i_lock);
1544 nfs_writeback_check_extend(hdr, fattr);
1545 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1546 spin_unlock(&inode->i_lock);
1547}
1548EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1549
1550/*
1551 * This function is called when the WRITE call is complete.
1552 */
1553static int nfs_writeback_done(struct rpc_task *task,
1554 struct nfs_pgio_header *hdr,
1555 struct inode *inode)
1556{
1557 int status;
1558
1559 /*
1560 * ->write_done will attempt to use post-op attributes to detect
1561 * conflicting writes by other clients. A strict interpretation
1562 * of close-to-open would allow us to continue caching even if
1563 * another writer had changed the file, but some applications
1564 * depend on tighter cache coherency when writing.
1565 */
1566 status = NFS_PROTO(inode)->write_done(task, hdr);
1567 if (status != 0)
1568 return status;
1569
1570 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1571 trace_nfs_writeback_done(inode, task->tk_status,
1572 hdr->args.offset, hdr->res.verf);
1573
1574 if (hdr->res.verf->committed < hdr->args.stable &&
1575 task->tk_status >= 0) {
1576 /* We tried a write call, but the server did not
1577 * commit data to stable storage even though we
1578 * requested it.
1579 * Note: There is a known bug in Tru64 < 5.0 in which
1580 * the server reports NFS_DATA_SYNC, but performs
1581 * NFS_FILE_SYNC. We therefore implement this checking
1582 * as a dprintk() in order to avoid filling syslog.
1583 */
1584 static unsigned long complain;
1585
1586 /* Note this will print the MDS for a DS write */
1587 if (time_before(complain, jiffies)) {
1588 dprintk("NFS: faulty NFS server %s:"
1589 " (committed = %d) != (stable = %d)\n",
1590 NFS_SERVER(inode)->nfs_client->cl_hostname,
1591 hdr->res.verf->committed, hdr->args.stable);
1592 complain = jiffies + 300 * HZ;
1593 }
1594 }
1595
1596 /* Deal with the suid/sgid bit corner case */
1597 if (nfs_should_remove_suid(inode)) {
1598 spin_lock(&inode->i_lock);
1599 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1600 spin_unlock(&inode->i_lock);
1601 }
1602 return 0;
1603}
1604
1605/*
1606 * This function is called when the WRITE call is complete.
1607 */
1608static void nfs_writeback_result(struct rpc_task *task,
1609 struct nfs_pgio_header *hdr)
1610{
1611 struct nfs_pgio_args *argp = &hdr->args;
1612 struct nfs_pgio_res *resp = &hdr->res;
1613
1614 if (resp->count < argp->count) {
1615 static unsigned long complain;
1616
1617 /* This a short write! */
1618 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1619
1620 /* Has the server at least made some progress? */
1621 if (resp->count == 0) {
1622 if (time_before(complain, jiffies)) {
1623 printk(KERN_WARNING
1624 "NFS: Server wrote zero bytes, expected %u.\n",
1625 argp->count);
1626 complain = jiffies + 300 * HZ;
1627 }
1628 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1629 task->tk_status = -EIO;
1630 return;
1631 }
1632
1633 /* For non rpc-based layout drivers, retry-through-MDS */
1634 if (!task->tk_ops) {
1635 hdr->pnfs_error = -EAGAIN;
1636 return;
1637 }
1638
1639 /* Was this an NFSv2 write or an NFSv3 stable write? */
1640 if (resp->verf->committed != NFS_UNSTABLE) {
1641 /* Resend from where the server left off */
1642 hdr->mds_offset += resp->count;
1643 argp->offset += resp->count;
1644 argp->pgbase += resp->count;
1645 argp->count -= resp->count;
1646 } else {
1647 /* Resend as a stable write in order to avoid
1648 * headaches in the case of a server crash.
1649 */
1650 argp->stable = NFS_FILE_SYNC;
1651 }
1652 rpc_restart_call_prepare(task);
1653 }
1654}
1655
1656static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1657{
1658 return wait_var_event_killable(&cinfo->rpcs_out,
1659 !atomic_read(&cinfo->rpcs_out));
1660}
1661
1662static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1663{
1664 atomic_inc(&cinfo->rpcs_out);
1665}
1666
1667static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1668{
1669 if (atomic_dec_and_test(&cinfo->rpcs_out))
1670 wake_up_var(&cinfo->rpcs_out);
1671}
1672
1673void nfs_commitdata_release(struct nfs_commit_data *data)
1674{
1675 put_nfs_open_context(data->context);
1676 nfs_commit_free(data);
1677}
1678EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1679
1680int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1681 const struct nfs_rpc_ops *nfs_ops,
1682 const struct rpc_call_ops *call_ops,
1683 int how, int flags)
1684{
1685 struct rpc_task *task;
1686 int priority = flush_task_priority(how);
1687 struct rpc_message msg = {
1688 .rpc_argp = &data->args,
1689 .rpc_resp = &data->res,
1690 .rpc_cred = data->cred,
1691 };
1692 struct rpc_task_setup task_setup_data = {
1693 .task = &data->task,
1694 .rpc_client = clnt,
1695 .rpc_message = &msg,
1696 .callback_ops = call_ops,
1697 .callback_data = data,
1698 .workqueue = nfsiod_workqueue,
1699 .flags = RPC_TASK_ASYNC | flags,
1700 .priority = priority,
1701 };
1702 /* Set up the initial task struct. */
1703 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1704 trace_nfs_initiate_commit(data);
1705
1706 dprintk("NFS: initiated commit call\n");
1707
1708 task = rpc_run_task(&task_setup_data);
1709 if (IS_ERR(task))
1710 return PTR_ERR(task);
1711 if (how & FLUSH_SYNC)
1712 rpc_wait_for_completion_task(task);
1713 rpc_put_task(task);
1714 return 0;
1715}
1716EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1717
1718static loff_t nfs_get_lwb(struct list_head *head)
1719{
1720 loff_t lwb = 0;
1721 struct nfs_page *req;
1722
1723 list_for_each_entry(req, head, wb_list)
1724 if (lwb < (req_offset(req) + req->wb_bytes))
1725 lwb = req_offset(req) + req->wb_bytes;
1726
1727 return lwb;
1728}
1729
1730/*
1731 * Set up the argument/result storage required for the RPC call.
1732 */
1733void nfs_init_commit(struct nfs_commit_data *data,
1734 struct list_head *head,
1735 struct pnfs_layout_segment *lseg,
1736 struct nfs_commit_info *cinfo)
1737{
1738 struct nfs_page *first = nfs_list_entry(head->next);
1739 struct nfs_open_context *ctx = nfs_req_openctx(first);
1740 struct inode *inode = d_inode(ctx->dentry);
1741
1742 /* Set up the RPC argument and reply structs
1743 * NB: take care not to mess about with data->commit et al. */
1744
1745 list_splice_init(head, &data->pages);
1746
1747 data->inode = inode;
1748 data->cred = ctx->cred;
1749 data->lseg = lseg; /* reference transferred */
1750 /* only set lwb for pnfs commit */
1751 if (lseg)
1752 data->lwb = nfs_get_lwb(&data->pages);
1753 data->mds_ops = &nfs_commit_ops;
1754 data->completion_ops = cinfo->completion_ops;
1755 data->dreq = cinfo->dreq;
1756
1757 data->args.fh = NFS_FH(data->inode);
1758 /* Note: we always request a commit of the entire inode */
1759 data->args.offset = 0;
1760 data->args.count = 0;
1761 data->context = get_nfs_open_context(ctx);
1762 data->res.fattr = &data->fattr;
1763 data->res.verf = &data->verf;
1764 nfs_fattr_init(&data->fattr);
1765}
1766EXPORT_SYMBOL_GPL(nfs_init_commit);
1767
1768void nfs_retry_commit(struct list_head *page_list,
1769 struct pnfs_layout_segment *lseg,
1770 struct nfs_commit_info *cinfo,
1771 u32 ds_commit_idx)
1772{
1773 struct nfs_page *req;
1774
1775 while (!list_empty(page_list)) {
1776 req = nfs_list_entry(page_list->next);
1777 nfs_list_remove_request(req);
1778 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1779 if (!cinfo->dreq)
1780 nfs_clear_page_commit(req->wb_page);
1781 nfs_unlock_and_release_request(req);
1782 }
1783}
1784EXPORT_SYMBOL_GPL(nfs_retry_commit);
1785
1786static void
1787nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1788 struct nfs_page *req)
1789{
1790 __set_page_dirty_nobuffers(req->wb_page);
1791}
1792
1793/*
1794 * Commit dirty pages
1795 */
1796static int
1797nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1798 struct nfs_commit_info *cinfo)
1799{
1800 struct nfs_commit_data *data;
1801
1802 /* another commit raced with us */
1803 if (list_empty(head))
1804 return 0;
1805
1806 data = nfs_commitdata_alloc(true);
1807
1808 /* Set up the argument struct */
1809 nfs_init_commit(data, head, NULL, cinfo);
1810 atomic_inc(&cinfo->mds->rpcs_out);
1811 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1812 data->mds_ops, how, 0);
1813}
1814
1815/*
1816 * COMMIT call returned
1817 */
1818static void nfs_commit_done(struct rpc_task *task, void *calldata)
1819{
1820 struct nfs_commit_data *data = calldata;
1821
1822 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1823 task->tk_pid, task->tk_status);
1824
1825 /* Call the NFS version-specific code */
1826 NFS_PROTO(data->inode)->commit_done(task, data);
1827 trace_nfs_commit_done(data);
1828}
1829
1830static void nfs_commit_release_pages(struct nfs_commit_data *data)
1831{
1832 struct nfs_page *req;
1833 int status = data->task.tk_status;
1834 struct nfs_commit_info cinfo;
1835 struct nfs_server *nfss;
1836
1837 while (!list_empty(&data->pages)) {
1838 req = nfs_list_entry(data->pages.next);
1839 nfs_list_remove_request(req);
1840 if (req->wb_page)
1841 nfs_clear_page_commit(req->wb_page);
1842
1843 dprintk("NFS: commit (%s/%llu %d@%lld)",
1844 nfs_req_openctx(req)->dentry->d_sb->s_id,
1845 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1846 req->wb_bytes,
1847 (long long)req_offset(req));
1848 if (status < 0) {
1849 if (req->wb_page) {
1850 nfs_mapping_set_error(req->wb_page, status);
1851 nfs_inode_remove_request(req);
1852 }
1853 dprintk_cont(", error = %d\n", status);
1854 goto next;
1855 }
1856
1857 /* Okay, COMMIT succeeded, apparently. Check the verifier
1858 * returned by the server against all stored verfs. */
1859 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1860 /* We have a match */
1861 if (req->wb_page)
1862 nfs_inode_remove_request(req);
1863 dprintk_cont(" OK\n");
1864 goto next;
1865 }
1866 /* We have a mismatch. Write the page again */
1867 dprintk_cont(" mismatch\n");
1868 nfs_mark_request_dirty(req);
1869 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1870 next:
1871 nfs_unlock_and_release_request(req);
1872 /* Latency breaker */
1873 cond_resched();
1874 }
1875 nfss = NFS_SERVER(data->inode);
1876 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1877 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1878
1879 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1880 nfs_commit_end(cinfo.mds);
1881}
1882
1883static void nfs_commit_release(void *calldata)
1884{
1885 struct nfs_commit_data *data = calldata;
1886
1887 data->completion_ops->completion(data);
1888 nfs_commitdata_release(calldata);
1889}
1890
1891static const struct rpc_call_ops nfs_commit_ops = {
1892 .rpc_call_prepare = nfs_commit_prepare,
1893 .rpc_call_done = nfs_commit_done,
1894 .rpc_release = nfs_commit_release,
1895};
1896
1897static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1898 .completion = nfs_commit_release_pages,
1899 .resched_write = nfs_commit_resched_write,
1900};
1901
1902int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1903 int how, struct nfs_commit_info *cinfo)
1904{
1905 int status;
1906
1907 status = pnfs_commit_list(inode, head, how, cinfo);
1908 if (status == PNFS_NOT_ATTEMPTED)
1909 status = nfs_commit_list(inode, head, how, cinfo);
1910 return status;
1911}
1912
1913static int __nfs_commit_inode(struct inode *inode, int how,
1914 struct writeback_control *wbc)
1915{
1916 LIST_HEAD(head);
1917 struct nfs_commit_info cinfo;
1918 int may_wait = how & FLUSH_SYNC;
1919 int ret, nscan;
1920
1921 nfs_init_cinfo_from_inode(&cinfo, inode);
1922 nfs_commit_begin(cinfo.mds);
1923 for (;;) {
1924 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1925 if (ret <= 0)
1926 break;
1927 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1928 if (ret < 0)
1929 break;
1930 ret = 0;
1931 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1932 if (nscan < wbc->nr_to_write)
1933 wbc->nr_to_write -= nscan;
1934 else
1935 wbc->nr_to_write = 0;
1936 }
1937 if (nscan < INT_MAX)
1938 break;
1939 cond_resched();
1940 }
1941 nfs_commit_end(cinfo.mds);
1942 if (ret || !may_wait)
1943 return ret;
1944 return wait_on_commit(cinfo.mds);
1945}
1946
1947int nfs_commit_inode(struct inode *inode, int how)
1948{
1949 return __nfs_commit_inode(inode, how, NULL);
1950}
1951EXPORT_SYMBOL_GPL(nfs_commit_inode);
1952
1953int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1954{
1955 struct nfs_inode *nfsi = NFS_I(inode);
1956 int flags = FLUSH_SYNC;
1957 int ret = 0;
1958
1959 if (wbc->sync_mode == WB_SYNC_NONE) {
1960 /* no commits means nothing needs to be done */
1961 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1962 goto check_requests_outstanding;
1963
1964 /* Don't commit yet if this is a non-blocking flush and there
1965 * are a lot of outstanding writes for this mapping.
1966 */
1967 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1968 goto out_mark_dirty;
1969
1970 /* don't wait for the COMMIT response */
1971 flags = 0;
1972 }
1973
1974 ret = __nfs_commit_inode(inode, flags, wbc);
1975 if (!ret) {
1976 if (flags & FLUSH_SYNC)
1977 return 0;
1978 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1979 goto out_mark_dirty;
1980
1981check_requests_outstanding:
1982 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1983 return ret;
1984out_mark_dirty:
1985 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1986 return ret;
1987}
1988EXPORT_SYMBOL_GPL(nfs_write_inode);
1989
1990/*
1991 * Wrapper for filemap_write_and_wait_range()
1992 *
1993 * Needed for pNFS in order to ensure data becomes visible to the
1994 * client.
1995 */
1996int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1997 loff_t lstart, loff_t lend)
1998{
1999 int ret;
2000
2001 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2002 if (ret == 0)
2003 ret = pnfs_sync_inode(mapping->host, true);
2004 return ret;
2005}
2006EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2007
2008/*
2009 * flush the inode to disk.
2010 */
2011int nfs_wb_all(struct inode *inode)
2012{
2013 int ret;
2014
2015 trace_nfs_writeback_inode_enter(inode);
2016
2017 ret = filemap_write_and_wait(inode->i_mapping);
2018 if (ret)
2019 goto out;
2020 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2021 if (ret < 0)
2022 goto out;
2023 pnfs_sync_inode(inode, true);
2024 ret = 0;
2025
2026out:
2027 trace_nfs_writeback_inode_exit(inode, ret);
2028 return ret;
2029}
2030EXPORT_SYMBOL_GPL(nfs_wb_all);
2031
2032int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2033{
2034 struct nfs_page *req;
2035 int ret = 0;
2036
2037 wait_on_page_writeback(page);
2038
2039 /* blocking call to cancel all requests and join to a single (head)
2040 * request */
2041 req = nfs_lock_and_join_requests(page);
2042
2043 if (IS_ERR(req)) {
2044 ret = PTR_ERR(req);
2045 } else if (req) {
2046 /* all requests from this page have been cancelled by
2047 * nfs_lock_and_join_requests, so just remove the head
2048 * request from the inode / page_private pointer and
2049 * release it */
2050 nfs_inode_remove_request(req);
2051 nfs_unlock_and_release_request(req);
2052 }
2053
2054 return ret;
2055}
2056
2057/*
2058 * Write back all requests on one page - we do this before reading it.
2059 */
2060int nfs_wb_page(struct inode *inode, struct page *page)
2061{
2062 loff_t range_start = page_file_offset(page);
2063 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2064 struct writeback_control wbc = {
2065 .sync_mode = WB_SYNC_ALL,
2066 .nr_to_write = 0,
2067 .range_start = range_start,
2068 .range_end = range_end,
2069 };
2070 int ret;
2071
2072 trace_nfs_writeback_page_enter(inode);
2073
2074 for (;;) {
2075 wait_on_page_writeback(page);
2076 if (clear_page_dirty_for_io(page)) {
2077 ret = nfs_writepage_locked(page, &wbc);
2078 if (ret < 0)
2079 goto out_error;
2080 continue;
2081 }
2082 ret = 0;
2083 if (!PagePrivate(page))
2084 break;
2085 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2086 if (ret < 0)
2087 goto out_error;
2088 }
2089out_error:
2090 trace_nfs_writeback_page_exit(inode, ret);
2091 return ret;
2092}
2093
2094#ifdef CONFIG_MIGRATION
2095int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2096 struct page *page, enum migrate_mode mode)
2097{
2098 /*
2099 * If PagePrivate is set, then the page is currently associated with
2100 * an in-progress read or write request. Don't try to migrate it.
2101 *
2102 * FIXME: we could do this in principle, but we'll need a way to ensure
2103 * that we can safely release the inode reference while holding
2104 * the page lock.
2105 */
2106 if (PagePrivate(page))
2107 return -EBUSY;
2108
2109 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2110 return -EBUSY;
2111
2112 return migrate_page(mapping, newpage, page, mode);
2113}
2114#endif
2115
2116int __init nfs_init_writepagecache(void)
2117{
2118 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2119 sizeof(struct nfs_pgio_header),
2120 0, SLAB_HWCACHE_ALIGN,
2121 NULL);
2122 if (nfs_wdata_cachep == NULL)
2123 return -ENOMEM;
2124
2125 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2126 nfs_wdata_cachep);
2127 if (nfs_wdata_mempool == NULL)
2128 goto out_destroy_write_cache;
2129
2130 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2131 sizeof(struct nfs_commit_data),
2132 0, SLAB_HWCACHE_ALIGN,
2133 NULL);
2134 if (nfs_cdata_cachep == NULL)
2135 goto out_destroy_write_mempool;
2136
2137 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2138 nfs_cdata_cachep);
2139 if (nfs_commit_mempool == NULL)
2140 goto out_destroy_commit_cache;
2141
2142 /*
2143 * NFS congestion size, scale with available memory.
2144 *
2145 * 64MB: 8192k
2146 * 128MB: 11585k
2147 * 256MB: 16384k
2148 * 512MB: 23170k
2149 * 1GB: 32768k
2150 * 2GB: 46340k
2151 * 4GB: 65536k
2152 * 8GB: 92681k
2153 * 16GB: 131072k
2154 *
2155 * This allows larger machines to have larger/more transfers.
2156 * Limit the default to 256M
2157 */
2158 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2159 if (nfs_congestion_kb > 256*1024)
2160 nfs_congestion_kb = 256*1024;
2161
2162 return 0;
2163
2164out_destroy_commit_cache:
2165 kmem_cache_destroy(nfs_cdata_cachep);
2166out_destroy_write_mempool:
2167 mempool_destroy(nfs_wdata_mempool);
2168out_destroy_write_cache:
2169 kmem_cache_destroy(nfs_wdata_cachep);
2170 return -ENOMEM;
2171}
2172
2173void nfs_destroy_writepagecache(void)
2174{
2175 mempool_destroy(nfs_commit_mempool);
2176 kmem_cache_destroy(nfs_cdata_cachep);
2177 mempool_destroy(nfs_wdata_mempool);
2178 kmem_cache_destroy(nfs_wdata_cachep);
2179}
2180
2181static const struct nfs_rw_ops nfs_rw_write_ops = {
2182 .rw_alloc_header = nfs_writehdr_alloc,
2183 .rw_free_header = nfs_writehdr_free,
2184 .rw_done = nfs_writeback_done,
2185 .rw_result = nfs_writeback_result,
2186 .rw_initiate = nfs_initiate_write,
2187};