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/*
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#include <linux/export.h>
24#include <linux/freezer.h>
25#include <linux/wait.h>
26
27#include <asm/uaccess.h>
28
29#include "delegation.h"
30#include "internal.h"
31#include "iostat.h"
32#include "nfs4_fs.h"
33#include "fscache.h"
34#include "pnfs.h"
35
36#include "nfstrace.h"
37
38#define NFSDBG_FACILITY NFSDBG_PAGECACHE
39
40#define MIN_POOL_WRITE (32)
41#define MIN_POOL_COMMIT (4)
42
43/*
44 * Local function declarations
45 */
46static void nfs_redirty_request(struct nfs_page *req);
47static const struct rpc_call_ops nfs_commit_ops;
48static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
49static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
50static const struct nfs_rw_ops nfs_rw_write_ops;
51static void nfs_clear_request_commit(struct nfs_page *req);
52static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
53 struct inode *inode);
54static struct nfs_page *
55nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
56 struct page *page);
57
58static struct kmem_cache *nfs_wdata_cachep;
59static mempool_t *nfs_wdata_mempool;
60static struct kmem_cache *nfs_cdata_cachep;
61static mempool_t *nfs_commit_mempool;
62
63struct nfs_commit_data *nfs_commitdata_alloc(void)
64{
65 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
66
67 if (p) {
68 memset(p, 0, sizeof(*p));
69 INIT_LIST_HEAD(&p->pages);
70 }
71 return p;
72}
73EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
74
75void nfs_commit_free(struct nfs_commit_data *p)
76{
77 mempool_free(p, nfs_commit_mempool);
78}
79EXPORT_SYMBOL_GPL(nfs_commit_free);
80
81static struct nfs_pgio_header *nfs_writehdr_alloc(void)
82{
83 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
84
85 if (p)
86 memset(p, 0, sizeof(*p));
87 return p;
88}
89
90static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
91{
92 mempool_free(hdr, nfs_wdata_mempool);
93}
94
95static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
96{
97 ctx->error = error;
98 smp_wmb();
99 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
100}
101
102/*
103 * nfs_page_find_head_request_locked - find head request associated with @page
104 *
105 * must be called while holding the inode lock.
106 *
107 * returns matching head request with reference held, or NULL if not found.
108 */
109static struct nfs_page *
110nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page)
111{
112 struct nfs_page *req = NULL;
113
114 if (PagePrivate(page))
115 req = (struct nfs_page *)page_private(page);
116 else if (unlikely(PageSwapCache(page)))
117 req = nfs_page_search_commits_for_head_request_locked(nfsi,
118 page);
119
120 if (req) {
121 WARN_ON_ONCE(req->wb_head != req);
122 kref_get(&req->wb_kref);
123 }
124
125 return req;
126}
127
128/*
129 * nfs_page_find_head_request - find head request associated with @page
130 *
131 * returns matching head request with reference held, or NULL if not found.
132 */
133static struct nfs_page *nfs_page_find_head_request(struct page *page)
134{
135 struct inode *inode = page_file_mapping(page)->host;
136 struct nfs_page *req = NULL;
137
138 spin_lock(&inode->i_lock);
139 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
140 spin_unlock(&inode->i_lock);
141 return req;
142}
143
144/* Adjust the file length if we're writing beyond the end */
145static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
146{
147 struct inode *inode = page_file_mapping(page)->host;
148 loff_t end, i_size;
149 pgoff_t end_index;
150
151 spin_lock(&inode->i_lock);
152 i_size = i_size_read(inode);
153 end_index = (i_size - 1) >> PAGE_SHIFT;
154 if (i_size > 0 && page_file_index(page) < end_index)
155 goto out;
156 end = page_file_offset(page) + ((loff_t)offset+count);
157 if (i_size >= end)
158 goto out;
159 i_size_write(inode, end);
160 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
161out:
162 spin_unlock(&inode->i_lock);
163}
164
165/* A writeback failed: mark the page as bad, and invalidate the page cache */
166static void nfs_set_pageerror(struct page *page)
167{
168 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
169}
170
171/*
172 * nfs_page_group_search_locked
173 * @head - head request of page group
174 * @page_offset - offset into page
175 *
176 * Search page group with head @head to find a request that contains the
177 * page offset @page_offset.
178 *
179 * Returns a pointer to the first matching nfs request, or NULL if no
180 * match is found.
181 *
182 * Must be called with the page group lock held
183 */
184static struct nfs_page *
185nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
186{
187 struct nfs_page *req;
188
189 WARN_ON_ONCE(head != head->wb_head);
190 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
191
192 req = head;
193 do {
194 if (page_offset >= req->wb_pgbase &&
195 page_offset < (req->wb_pgbase + req->wb_bytes))
196 return req;
197
198 req = req->wb_this_page;
199 } while (req != head);
200
201 return NULL;
202}
203
204/*
205 * nfs_page_group_covers_page
206 * @head - head request of page group
207 *
208 * Return true if the page group with head @head covers the whole page,
209 * returns false otherwise
210 */
211static bool nfs_page_group_covers_page(struct nfs_page *req)
212{
213 struct nfs_page *tmp;
214 unsigned int pos = 0;
215 unsigned int len = nfs_page_length(req->wb_page);
216
217 nfs_page_group_lock(req, false);
218
219 do {
220 tmp = nfs_page_group_search_locked(req->wb_head, pos);
221 if (tmp) {
222 /* no way this should happen */
223 WARN_ON_ONCE(tmp->wb_pgbase != pos);
224 pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
225 }
226 } while (tmp && pos < len);
227
228 nfs_page_group_unlock(req);
229 WARN_ON_ONCE(pos > len);
230 return pos == len;
231}
232
233/* We can set the PG_uptodate flag if we see that a write request
234 * covers the full page.
235 */
236static void nfs_mark_uptodate(struct nfs_page *req)
237{
238 if (PageUptodate(req->wb_page))
239 return;
240 if (!nfs_page_group_covers_page(req))
241 return;
242 SetPageUptodate(req->wb_page);
243}
244
245static int wb_priority(struct writeback_control *wbc)
246{
247 int ret = 0;
248 if (wbc->for_reclaim)
249 return FLUSH_HIGHPRI | FLUSH_COND_STABLE;
250 if (wbc->sync_mode == WB_SYNC_ALL)
251 ret = FLUSH_COND_STABLE;
252 return ret;
253}
254
255/*
256 * NFS congestion control
257 */
258
259int nfs_congestion_kb;
260
261#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
262#define NFS_CONGESTION_OFF_THRESH \
263 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
264
265static void nfs_set_page_writeback(struct page *page)
266{
267 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
268 int ret = test_set_page_writeback(page);
269
270 WARN_ON_ONCE(ret != 0);
271
272 if (atomic_long_inc_return(&nfss->writeback) >
273 NFS_CONGESTION_ON_THRESH) {
274 set_bdi_congested(&nfss->backing_dev_info,
275 BLK_RW_ASYNC);
276 }
277}
278
279static void nfs_end_page_writeback(struct nfs_page *req)
280{
281 struct inode *inode = page_file_mapping(req->wb_page)->host;
282 struct nfs_server *nfss = NFS_SERVER(inode);
283
284 if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
285 return;
286
287 end_page_writeback(req->wb_page);
288 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
289 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
290}
291
292
293/* nfs_page_group_clear_bits
294 * @req - an nfs request
295 * clears all page group related bits from @req
296 */
297static void
298nfs_page_group_clear_bits(struct nfs_page *req)
299{
300 clear_bit(PG_TEARDOWN, &req->wb_flags);
301 clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
302 clear_bit(PG_UPTODATE, &req->wb_flags);
303 clear_bit(PG_WB_END, &req->wb_flags);
304 clear_bit(PG_REMOVE, &req->wb_flags);
305}
306
307
308/*
309 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
310 *
311 * this is a helper function for nfs_lock_and_join_requests
312 *
313 * @inode - inode associated with request page group, must be holding inode lock
314 * @head - head request of page group, must be holding head lock
315 * @req - request that couldn't lock and needs to wait on the req bit lock
316 * @nonblock - if true, don't actually wait
317 *
318 * NOTE: this must be called holding page_group bit lock and inode spin lock
319 * and BOTH will be released before returning.
320 *
321 * returns 0 on success, < 0 on error.
322 */
323static int
324nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
325 struct nfs_page *req, bool nonblock)
326 __releases(&inode->i_lock)
327{
328 struct nfs_page *tmp;
329 int ret;
330
331 /* relinquish all the locks successfully grabbed this run */
332 for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
333 nfs_unlock_request(tmp);
334
335 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
336
337 /* grab a ref on the request that will be waited on */
338 kref_get(&req->wb_kref);
339
340 nfs_page_group_unlock(head);
341 spin_unlock(&inode->i_lock);
342
343 /* release ref from nfs_page_find_head_request_locked */
344 nfs_release_request(head);
345
346 if (!nonblock)
347 ret = nfs_wait_on_request(req);
348 else
349 ret = -EAGAIN;
350 nfs_release_request(req);
351
352 return ret;
353}
354
355/*
356 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
357 *
358 * @destroy_list - request list (using wb_this_page) terminated by @old_head
359 * @old_head - the old head of the list
360 *
361 * All subrequests must be locked and removed from all lists, so at this point
362 * they are only "active" in this function, and possibly in nfs_wait_on_request
363 * with a reference held by some other context.
364 */
365static void
366nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
367 struct nfs_page *old_head)
368{
369 while (destroy_list) {
370 struct nfs_page *subreq = destroy_list;
371
372 destroy_list = (subreq->wb_this_page == old_head) ?
373 NULL : subreq->wb_this_page;
374
375 WARN_ON_ONCE(old_head != subreq->wb_head);
376
377 /* make sure old group is not used */
378 subreq->wb_head = subreq;
379 subreq->wb_this_page = subreq;
380
381 /* subreq is now totally disconnected from page group or any
382 * write / commit lists. last chance to wake any waiters */
383 nfs_unlock_request(subreq);
384
385 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
386 /* release ref on old head request */
387 nfs_release_request(old_head);
388
389 nfs_page_group_clear_bits(subreq);
390
391 /* release the PG_INODE_REF reference */
392 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
393 nfs_release_request(subreq);
394 else
395 WARN_ON_ONCE(1);
396 } else {
397 WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
398 /* zombie requests have already released the last
399 * reference and were waiting on the rest of the
400 * group to complete. Since it's no longer part of a
401 * group, simply free the request */
402 nfs_page_group_clear_bits(subreq);
403 nfs_free_request(subreq);
404 }
405 }
406}
407
408/*
409 * nfs_lock_and_join_requests - join all subreqs to the head req and return
410 * a locked reference, cancelling any pending
411 * operations for this page.
412 *
413 * @page - the page used to lookup the "page group" of nfs_page structures
414 * @nonblock - if true, don't block waiting for request locks
415 *
416 * This function joins all sub requests to the head request by first
417 * locking all requests in the group, cancelling any pending operations
418 * and finally updating the head request to cover the whole range covered by
419 * the (former) group. All subrequests are removed from any write or commit
420 * lists, unlinked from the group and destroyed.
421 *
422 * Returns a locked, referenced pointer to the head request - which after
423 * this call is guaranteed to be the only request associated with the page.
424 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
425 * error was encountered.
426 */
427static struct nfs_page *
428nfs_lock_and_join_requests(struct page *page, bool nonblock)
429{
430 struct inode *inode = page_file_mapping(page)->host;
431 struct nfs_page *head, *subreq;
432 struct nfs_page *destroy_list = NULL;
433 unsigned int total_bytes;
434 int ret;
435
436try_again:
437 total_bytes = 0;
438
439 WARN_ON_ONCE(destroy_list);
440
441 spin_lock(&inode->i_lock);
442
443 /*
444 * A reference is taken only on the head request which acts as a
445 * reference to the whole page group - the group will not be destroyed
446 * until the head reference is released.
447 */
448 head = nfs_page_find_head_request_locked(NFS_I(inode), page);
449
450 if (!head) {
451 spin_unlock(&inode->i_lock);
452 return NULL;
453 }
454
455 /* holding inode lock, so always make a non-blocking call to try the
456 * page group lock */
457 ret = nfs_page_group_lock(head, true);
458 if (ret < 0) {
459 spin_unlock(&inode->i_lock);
460
461 if (!nonblock && ret == -EAGAIN) {
462 nfs_page_group_lock_wait(head);
463 nfs_release_request(head);
464 goto try_again;
465 }
466
467 nfs_release_request(head);
468 return ERR_PTR(ret);
469 }
470
471 /* lock each request in the page group */
472 subreq = head;
473 do {
474 /*
475 * Subrequests are always contiguous, non overlapping
476 * and in order - but may be repeated (mirrored writes).
477 */
478 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
479 /* keep track of how many bytes this group covers */
480 total_bytes += subreq->wb_bytes;
481 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
482 ((subreq->wb_offset + subreq->wb_bytes) >
483 (head->wb_offset + total_bytes)))) {
484 nfs_page_group_unlock(head);
485 spin_unlock(&inode->i_lock);
486 return ERR_PTR(-EIO);
487 }
488
489 if (!nfs_lock_request(subreq)) {
490 /* releases page group bit lock and
491 * inode spin lock and all references */
492 ret = nfs_unroll_locks_and_wait(inode, head,
493 subreq, nonblock);
494
495 if (ret == 0)
496 goto try_again;
497
498 return ERR_PTR(ret);
499 }
500
501 subreq = subreq->wb_this_page;
502 } while (subreq != head);
503
504 /* Now that all requests are locked, make sure they aren't on any list.
505 * Commit list removal accounting is done after locks are dropped */
506 subreq = head;
507 do {
508 nfs_clear_request_commit(subreq);
509 subreq = subreq->wb_this_page;
510 } while (subreq != head);
511
512 /* unlink subrequests from head, destroy them later */
513 if (head->wb_this_page != head) {
514 /* destroy list will be terminated by head */
515 destroy_list = head->wb_this_page;
516 head->wb_this_page = head;
517
518 /* change head request to cover whole range that
519 * the former page group covered */
520 head->wb_bytes = total_bytes;
521 }
522
523 /*
524 * prepare head request to be added to new pgio descriptor
525 */
526 nfs_page_group_clear_bits(head);
527
528 /*
529 * some part of the group was still on the inode list - otherwise
530 * the group wouldn't be involved in async write.
531 * grab a reference for the head request, iff it needs one.
532 */
533 if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
534 kref_get(&head->wb_kref);
535
536 nfs_page_group_unlock(head);
537
538 /* drop lock to clean uprequests on destroy list */
539 spin_unlock(&inode->i_lock);
540
541 nfs_destroy_unlinked_subrequests(destroy_list, head);
542
543 /* still holds ref on head from nfs_page_find_head_request_locked
544 * and still has lock on head from lock loop */
545 return head;
546}
547
548static void nfs_write_error_remove_page(struct nfs_page *req)
549{
550 nfs_unlock_request(req);
551 nfs_end_page_writeback(req);
552 nfs_release_request(req);
553 generic_error_remove_page(page_file_mapping(req->wb_page),
554 req->wb_page);
555}
556
557/*
558 * Find an associated nfs write request, and prepare to flush it out
559 * May return an error if the user signalled nfs_wait_on_request().
560 */
561static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
562 struct page *page, bool nonblock,
563 bool launder)
564{
565 struct nfs_page *req;
566 int ret = 0;
567
568 req = nfs_lock_and_join_requests(page, nonblock);
569 if (!req)
570 goto out;
571 ret = PTR_ERR(req);
572 if (IS_ERR(req))
573 goto out;
574
575 nfs_set_page_writeback(page);
576 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
577
578 ret = 0;
579 if (!nfs_pageio_add_request(pgio, req)) {
580 ret = pgio->pg_error;
581 /*
582 * Remove the problematic req upon fatal errors
583 * in launder case, while other dirty pages can
584 * still be around until they get flushed.
585 */
586 if (nfs_error_is_fatal(ret)) {
587 nfs_context_set_write_error(req->wb_context, ret);
588 if (launder) {
589 nfs_write_error_remove_page(req);
590 goto out;
591 }
592 }
593 nfs_redirty_request(req);
594 ret = -EAGAIN;
595 } else
596 nfs_add_stats(page_file_mapping(page)->host,
597 NFSIOS_WRITEPAGES, 1);
598out:
599 return ret;
600}
601
602static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
603 struct nfs_pageio_descriptor *pgio, bool launder)
604{
605 int ret;
606
607 nfs_pageio_cond_complete(pgio, page_file_index(page));
608 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE,
609 launder);
610 if (ret == -EAGAIN) {
611 redirty_page_for_writepage(wbc, page);
612 ret = 0;
613 }
614 return ret;
615}
616
617/*
618 * Write an mmapped page to the server.
619 */
620static int nfs_writepage_locked(struct page *page,
621 struct writeback_control *wbc,
622 bool launder)
623{
624 struct nfs_pageio_descriptor pgio;
625 struct inode *inode = page_file_mapping(page)->host;
626 int err;
627
628 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
629 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc),
630 false, &nfs_async_write_completion_ops);
631 err = nfs_do_writepage(page, wbc, &pgio, launder);
632 nfs_pageio_complete(&pgio);
633 if (err < 0)
634 return err;
635 if (pgio.pg_error < 0)
636 return pgio.pg_error;
637 return 0;
638}
639
640int nfs_writepage(struct page *page, struct writeback_control *wbc)
641{
642 int ret;
643
644 ret = nfs_writepage_locked(page, wbc, false);
645 unlock_page(page);
646 return ret;
647}
648
649static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
650{
651 int ret;
652
653 ret = nfs_do_writepage(page, wbc, data, false);
654 unlock_page(page);
655 return ret;
656}
657
658int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
659{
660 struct inode *inode = mapping->host;
661 unsigned long *bitlock = &NFS_I(inode)->flags;
662 struct nfs_pageio_descriptor pgio;
663 int err;
664
665 /* Stop dirtying of new pages while we sync */
666 err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING,
667 nfs_wait_bit_killable, TASK_KILLABLE);
668 if (err)
669 goto out_err;
670
671 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
672
673 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
674 &nfs_async_write_completion_ops);
675 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
676 nfs_pageio_complete(&pgio);
677
678 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
679 smp_mb__after_atomic();
680 wake_up_bit(bitlock, NFS_INO_FLUSHING);
681
682 if (err < 0)
683 goto out_err;
684 err = pgio.pg_error;
685 if (err < 0)
686 goto out_err;
687 return 0;
688out_err:
689 return err;
690}
691
692/*
693 * Insert a write request into an inode
694 */
695static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
696{
697 struct nfs_inode *nfsi = NFS_I(inode);
698
699 WARN_ON_ONCE(req->wb_this_page != req);
700
701 /* Lock the request! */
702 nfs_lock_request(req);
703
704 spin_lock(&inode->i_lock);
705 if (!nfsi->nrequests &&
706 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
707 inode->i_version++;
708 /*
709 * Swap-space should not get truncated. Hence no need to plug the race
710 * with invalidate/truncate.
711 */
712 if (likely(!PageSwapCache(req->wb_page))) {
713 set_bit(PG_MAPPED, &req->wb_flags);
714 SetPagePrivate(req->wb_page);
715 set_page_private(req->wb_page, (unsigned long)req);
716 }
717 nfsi->nrequests++;
718 /* this a head request for a page group - mark it as having an
719 * extra reference so sub groups can follow suit.
720 * This flag also informs pgio layer when to bump nrequests when
721 * adding subrequests. */
722 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
723 kref_get(&req->wb_kref);
724 spin_unlock(&inode->i_lock);
725}
726
727/*
728 * Remove a write request from an inode
729 */
730static void nfs_inode_remove_request(struct nfs_page *req)
731{
732 struct inode *inode = d_inode(req->wb_context->dentry);
733 struct nfs_inode *nfsi = NFS_I(inode);
734 struct nfs_page *head;
735
736 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
737 head = req->wb_head;
738
739 spin_lock(&inode->i_lock);
740 if (likely(!PageSwapCache(head->wb_page))) {
741 set_page_private(head->wb_page, 0);
742 ClearPagePrivate(head->wb_page);
743 smp_mb__after_atomic();
744 wake_up_page(head->wb_page, PG_private);
745 clear_bit(PG_MAPPED, &head->wb_flags);
746 }
747 nfsi->nrequests--;
748 spin_unlock(&inode->i_lock);
749 } else {
750 spin_lock(&inode->i_lock);
751 nfsi->nrequests--;
752 spin_unlock(&inode->i_lock);
753 }
754
755 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
756 nfs_release_request(req);
757}
758
759static void
760nfs_mark_request_dirty(struct nfs_page *req)
761{
762 __set_page_dirty_nobuffers(req->wb_page);
763}
764
765/*
766 * nfs_page_search_commits_for_head_request_locked
767 *
768 * Search through commit lists on @inode for the head request for @page.
769 * Must be called while holding the inode (which is cinfo) lock.
770 *
771 * Returns the head request if found, or NULL if not found.
772 */
773static struct nfs_page *
774nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
775 struct page *page)
776{
777 struct nfs_page *freq, *t;
778 struct nfs_commit_info cinfo;
779 struct inode *inode = &nfsi->vfs_inode;
780
781 nfs_init_cinfo_from_inode(&cinfo, inode);
782
783 /* search through pnfs commit lists */
784 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
785 if (freq)
786 return freq->wb_head;
787
788 /* Linearly search the commit list for the correct request */
789 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
790 if (freq->wb_page == page)
791 return freq->wb_head;
792 }
793
794 return NULL;
795}
796
797/**
798 * nfs_request_add_commit_list_locked - add request to a commit list
799 * @req: pointer to a struct nfs_page
800 * @dst: commit list head
801 * @cinfo: holds list lock and accounting info
802 *
803 * This sets the PG_CLEAN bit, updates the cinfo count of
804 * number of outstanding requests requiring a commit as well as
805 * the MM page stats.
806 *
807 * The caller must hold the cinfo->lock, and the nfs_page lock.
808 */
809void
810nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
811 struct nfs_commit_info *cinfo)
812{
813 set_bit(PG_CLEAN, &req->wb_flags);
814 nfs_list_add_request(req, dst);
815 cinfo->mds->ncommit++;
816}
817EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
818
819/**
820 * nfs_request_add_commit_list - add request to a commit list
821 * @req: pointer to a struct nfs_page
822 * @dst: commit list head
823 * @cinfo: holds list lock and accounting info
824 *
825 * This sets the PG_CLEAN bit, updates the cinfo count of
826 * number of outstanding requests requiring a commit as well as
827 * the MM page stats.
828 *
829 * The caller must _not_ hold the cinfo->lock, but must be
830 * holding the nfs_page lock.
831 */
832void
833nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
834{
835 spin_lock(cinfo->lock);
836 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
837 spin_unlock(cinfo->lock);
838 nfs_mark_page_unstable(req->wb_page, cinfo);
839}
840EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
841
842/**
843 * nfs_request_remove_commit_list - Remove request from a commit list
844 * @req: pointer to a nfs_page
845 * @cinfo: holds list lock and accounting info
846 *
847 * This clears the PG_CLEAN bit, and updates the cinfo's count of
848 * number of outstanding requests requiring a commit
849 * It does not update the MM page stats.
850 *
851 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
852 */
853void
854nfs_request_remove_commit_list(struct nfs_page *req,
855 struct nfs_commit_info *cinfo)
856{
857 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
858 return;
859 nfs_list_remove_request(req);
860 cinfo->mds->ncommit--;
861}
862EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
863
864static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
865 struct inode *inode)
866{
867 cinfo->lock = &inode->i_lock;
868 cinfo->mds = &NFS_I(inode)->commit_info;
869 cinfo->ds = pnfs_get_ds_info(inode);
870 cinfo->dreq = NULL;
871 cinfo->completion_ops = &nfs_commit_completion_ops;
872}
873
874void nfs_init_cinfo(struct nfs_commit_info *cinfo,
875 struct inode *inode,
876 struct nfs_direct_req *dreq)
877{
878 if (dreq)
879 nfs_init_cinfo_from_dreq(cinfo, dreq);
880 else
881 nfs_init_cinfo_from_inode(cinfo, inode);
882}
883EXPORT_SYMBOL_GPL(nfs_init_cinfo);
884
885/*
886 * Add a request to the inode's commit list.
887 */
888void
889nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
890 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
891{
892 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
893 return;
894 nfs_request_add_commit_list(req, cinfo);
895}
896
897static void
898nfs_clear_page_commit(struct page *page)
899{
900 dec_zone_page_state(page, NR_UNSTABLE_NFS);
901 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
902 WB_RECLAIMABLE);
903}
904
905/* Called holding inode (/cinfo) lock */
906static void
907nfs_clear_request_commit(struct nfs_page *req)
908{
909 if (test_bit(PG_CLEAN, &req->wb_flags)) {
910 struct inode *inode = d_inode(req->wb_context->dentry);
911 struct nfs_commit_info cinfo;
912
913 nfs_init_cinfo_from_inode(&cinfo, inode);
914 if (!pnfs_clear_request_commit(req, &cinfo)) {
915 nfs_request_remove_commit_list(req, &cinfo);
916 }
917 nfs_clear_page_commit(req->wb_page);
918 }
919}
920
921int nfs_write_need_commit(struct nfs_pgio_header *hdr)
922{
923 if (hdr->verf.committed == NFS_DATA_SYNC)
924 return hdr->lseg == NULL;
925 return hdr->verf.committed != NFS_FILE_SYNC;
926}
927
928static void nfs_write_completion(struct nfs_pgio_header *hdr)
929{
930 struct nfs_commit_info cinfo;
931 unsigned long bytes = 0;
932
933 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
934 goto out;
935 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
936 while (!list_empty(&hdr->pages)) {
937 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
938
939 bytes += req->wb_bytes;
940 nfs_list_remove_request(req);
941 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
942 (hdr->good_bytes < bytes)) {
943 nfs_set_pageerror(req->wb_page);
944 nfs_context_set_write_error(req->wb_context, hdr->error);
945 goto remove_req;
946 }
947 if (nfs_write_need_commit(hdr)) {
948 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
949 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
950 hdr->pgio_mirror_idx);
951 goto next;
952 }
953remove_req:
954 nfs_inode_remove_request(req);
955next:
956 nfs_unlock_request(req);
957 nfs_end_page_writeback(req);
958 nfs_release_request(req);
959 }
960out:
961 hdr->release(hdr);
962}
963
964unsigned long
965nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
966{
967 return cinfo->mds->ncommit;
968}
969
970/* cinfo->lock held by caller */
971int
972nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
973 struct nfs_commit_info *cinfo, int max)
974{
975 struct nfs_page *req, *tmp;
976 int ret = 0;
977
978 list_for_each_entry_safe(req, tmp, src, wb_list) {
979 if (!nfs_lock_request(req))
980 continue;
981 kref_get(&req->wb_kref);
982 if (cond_resched_lock(cinfo->lock))
983 list_safe_reset_next(req, tmp, wb_list);
984 nfs_request_remove_commit_list(req, cinfo);
985 nfs_list_add_request(req, dst);
986 ret++;
987 if ((ret == max) && !cinfo->dreq)
988 break;
989 }
990 return ret;
991}
992
993/*
994 * nfs_scan_commit - Scan an inode for commit requests
995 * @inode: NFS inode to scan
996 * @dst: mds destination list
997 * @cinfo: mds and ds lists of reqs ready to commit
998 *
999 * Moves requests from the inode's 'commit' request list.
1000 * The requests are *not* checked to ensure that they form a contiguous set.
1001 */
1002int
1003nfs_scan_commit(struct inode *inode, struct list_head *dst,
1004 struct nfs_commit_info *cinfo)
1005{
1006 int ret = 0;
1007
1008 spin_lock(cinfo->lock);
1009 if (cinfo->mds->ncommit > 0) {
1010 const int max = INT_MAX;
1011
1012 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1013 cinfo, max);
1014 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1015 }
1016 spin_unlock(cinfo->lock);
1017 return ret;
1018}
1019
1020/*
1021 * Search for an existing write request, and attempt to update
1022 * it to reflect a new dirty region on a given page.
1023 *
1024 * If the attempt fails, then the existing request is flushed out
1025 * to disk.
1026 */
1027static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1028 struct page *page,
1029 unsigned int offset,
1030 unsigned int bytes)
1031{
1032 struct nfs_page *req;
1033 unsigned int rqend;
1034 unsigned int end;
1035 int error;
1036
1037 if (!PagePrivate(page))
1038 return NULL;
1039
1040 end = offset + bytes;
1041 spin_lock(&inode->i_lock);
1042
1043 for (;;) {
1044 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
1045 if (req == NULL)
1046 goto out_unlock;
1047
1048 /* should be handled by nfs_flush_incompatible */
1049 WARN_ON_ONCE(req->wb_head != req);
1050 WARN_ON_ONCE(req->wb_this_page != req);
1051
1052 rqend = req->wb_offset + req->wb_bytes;
1053 /*
1054 * Tell the caller to flush out the request if
1055 * the offsets are non-contiguous.
1056 * Note: nfs_flush_incompatible() will already
1057 * have flushed out requests having wrong owners.
1058 */
1059 if (offset > rqend
1060 || end < req->wb_offset)
1061 goto out_flushme;
1062
1063 if (nfs_lock_request(req))
1064 break;
1065
1066 /* The request is locked, so wait and then retry */
1067 spin_unlock(&inode->i_lock);
1068 error = nfs_wait_on_request(req);
1069 nfs_release_request(req);
1070 if (error != 0)
1071 goto out_err;
1072 spin_lock(&inode->i_lock);
1073 }
1074
1075 /* Okay, the request matches. Update the region */
1076 if (offset < req->wb_offset) {
1077 req->wb_offset = offset;
1078 req->wb_pgbase = offset;
1079 }
1080 if (end > rqend)
1081 req->wb_bytes = end - req->wb_offset;
1082 else
1083 req->wb_bytes = rqend - req->wb_offset;
1084out_unlock:
1085 if (req)
1086 nfs_clear_request_commit(req);
1087 spin_unlock(&inode->i_lock);
1088 return req;
1089out_flushme:
1090 spin_unlock(&inode->i_lock);
1091 nfs_release_request(req);
1092 error = nfs_wb_page(inode, page);
1093out_err:
1094 return ERR_PTR(error);
1095}
1096
1097/*
1098 * Try to update an existing write request, or create one if there is none.
1099 *
1100 * Note: Should always be called with the Page Lock held to prevent races
1101 * if we have to add a new request. Also assumes that the caller has
1102 * already called nfs_flush_incompatible() if necessary.
1103 */
1104static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1105 struct page *page, unsigned int offset, unsigned int bytes)
1106{
1107 struct inode *inode = page_file_mapping(page)->host;
1108 struct nfs_page *req;
1109
1110 req = nfs_try_to_update_request(inode, page, offset, bytes);
1111 if (req != NULL)
1112 goto out;
1113 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1114 if (IS_ERR(req))
1115 goto out;
1116 nfs_inode_add_request(inode, req);
1117out:
1118 return req;
1119}
1120
1121static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1122 unsigned int offset, unsigned int count)
1123{
1124 struct nfs_page *req;
1125
1126 req = nfs_setup_write_request(ctx, page, offset, count);
1127 if (IS_ERR(req))
1128 return PTR_ERR(req);
1129 /* Update file length */
1130 nfs_grow_file(page, offset, count);
1131 nfs_mark_uptodate(req);
1132 nfs_mark_request_dirty(req);
1133 nfs_unlock_and_release_request(req);
1134 return 0;
1135}
1136
1137int nfs_flush_incompatible(struct file *file, struct page *page)
1138{
1139 struct nfs_open_context *ctx = nfs_file_open_context(file);
1140 struct nfs_lock_context *l_ctx;
1141 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1142 struct nfs_page *req;
1143 int do_flush, status;
1144 /*
1145 * Look for a request corresponding to this page. If there
1146 * is one, and it belongs to another file, we flush it out
1147 * before we try to copy anything into the page. Do this
1148 * due to the lack of an ACCESS-type call in NFSv2.
1149 * Also do the same if we find a request from an existing
1150 * dropped page.
1151 */
1152 do {
1153 req = nfs_page_find_head_request(page);
1154 if (req == NULL)
1155 return 0;
1156 l_ctx = req->wb_lock_context;
1157 do_flush = req->wb_page != page ||
1158 !nfs_match_open_context(req->wb_context, ctx);
1159 /* for now, flush if more than 1 request in page_group */
1160 do_flush |= req->wb_this_page != req;
1161 if (l_ctx && flctx &&
1162 !(list_empty_careful(&flctx->flc_posix) &&
1163 list_empty_careful(&flctx->flc_flock))) {
1164 do_flush |= l_ctx->lockowner.l_owner != current->files
1165 || l_ctx->lockowner.l_pid != current->tgid;
1166 }
1167 nfs_release_request(req);
1168 if (!do_flush)
1169 return 0;
1170 status = nfs_wb_page(page_file_mapping(page)->host, page);
1171 } while (status == 0);
1172 return status;
1173}
1174
1175/*
1176 * Avoid buffered writes when a open context credential's key would
1177 * expire soon.
1178 *
1179 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1180 *
1181 * Return 0 and set a credential flag which triggers the inode to flush
1182 * and performs NFS_FILE_SYNC writes if the key will expired within
1183 * RPC_KEY_EXPIRE_TIMEO.
1184 */
1185int
1186nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1187{
1188 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1189 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1190
1191 return rpcauth_key_timeout_notify(auth, ctx->cred);
1192}
1193
1194/*
1195 * Test if the open context credential key is marked to expire soon.
1196 */
1197bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1198{
1199 return rpcauth_cred_key_to_expire(ctx->cred);
1200}
1201
1202/*
1203 * If the page cache is marked as unsafe or invalid, then we can't rely on
1204 * the PageUptodate() flag. In this case, we will need to turn off
1205 * write optimisations that depend on the page contents being correct.
1206 */
1207static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1208{
1209 struct nfs_inode *nfsi = NFS_I(inode);
1210
1211 if (nfs_have_delegated_attributes(inode))
1212 goto out;
1213 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1214 return false;
1215 smp_rmb();
1216 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1217 return false;
1218out:
1219 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1220 return false;
1221 return PageUptodate(page) != 0;
1222}
1223
1224static bool
1225is_whole_file_wrlock(struct file_lock *fl)
1226{
1227 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1228 fl->fl_type == F_WRLCK;
1229}
1230
1231/* If we know the page is up to date, and we're not using byte range locks (or
1232 * if we have the whole file locked for writing), it may be more efficient to
1233 * extend the write to cover the entire page in order to avoid fragmentation
1234 * inefficiencies.
1235 *
1236 * If the file is opened for synchronous writes then we can just skip the rest
1237 * of the checks.
1238 */
1239static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1240{
1241 int ret;
1242 struct file_lock_context *flctx = inode->i_flctx;
1243 struct file_lock *fl;
1244
1245 if (file->f_flags & O_DSYNC)
1246 return 0;
1247 if (!nfs_write_pageuptodate(page, inode))
1248 return 0;
1249 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1250 return 1;
1251 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1252 list_empty_careful(&flctx->flc_posix)))
1253 return 1;
1254
1255 /* Check to see if there are whole file write locks */
1256 ret = 0;
1257 spin_lock(&flctx->flc_lock);
1258 if (!list_empty(&flctx->flc_posix)) {
1259 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1260 fl_list);
1261 if (is_whole_file_wrlock(fl))
1262 ret = 1;
1263 } else if (!list_empty(&flctx->flc_flock)) {
1264 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1265 fl_list);
1266 if (fl->fl_type == F_WRLCK)
1267 ret = 1;
1268 }
1269 spin_unlock(&flctx->flc_lock);
1270 return ret;
1271}
1272
1273/*
1274 * Update and possibly write a cached page of an NFS file.
1275 *
1276 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1277 * things with a page scheduled for an RPC call (e.g. invalidate it).
1278 */
1279int nfs_updatepage(struct file *file, struct page *page,
1280 unsigned int offset, unsigned int count)
1281{
1282 struct nfs_open_context *ctx = nfs_file_open_context(file);
1283 struct inode *inode = page_file_mapping(page)->host;
1284 int status = 0;
1285
1286 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1287
1288 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1289 file, count, (long long)(page_file_offset(page) + offset));
1290
1291 if (nfs_can_extend_write(file, page, inode)) {
1292 count = max(count + offset, nfs_page_length(page));
1293 offset = 0;
1294 }
1295
1296 status = nfs_writepage_setup(ctx, page, offset, count);
1297 if (status < 0)
1298 nfs_set_pageerror(page);
1299 else
1300 __set_page_dirty_nobuffers(page);
1301
1302 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1303 status, (long long)i_size_read(inode));
1304 return status;
1305}
1306
1307static int flush_task_priority(int how)
1308{
1309 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1310 case FLUSH_HIGHPRI:
1311 return RPC_PRIORITY_HIGH;
1312 case FLUSH_LOWPRI:
1313 return RPC_PRIORITY_LOW;
1314 }
1315 return RPC_PRIORITY_NORMAL;
1316}
1317
1318static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1319 struct rpc_message *msg,
1320 const struct nfs_rpc_ops *rpc_ops,
1321 struct rpc_task_setup *task_setup_data, int how)
1322{
1323 int priority = flush_task_priority(how);
1324
1325 task_setup_data->priority = priority;
1326 rpc_ops->write_setup(hdr, msg);
1327
1328 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1329 &task_setup_data->rpc_client, msg, hdr);
1330}
1331
1332/* If a nfs_flush_* function fails, it should remove reqs from @head and
1333 * call this on each, which will prepare them to be retried on next
1334 * writeback using standard nfs.
1335 */
1336static void nfs_redirty_request(struct nfs_page *req)
1337{
1338 nfs_mark_request_dirty(req);
1339 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1340 nfs_unlock_request(req);
1341 nfs_end_page_writeback(req);
1342 nfs_release_request(req);
1343}
1344
1345static void nfs_async_write_error(struct list_head *head)
1346{
1347 struct nfs_page *req;
1348
1349 while (!list_empty(head)) {
1350 req = nfs_list_entry(head->next);
1351 nfs_list_remove_request(req);
1352 nfs_redirty_request(req);
1353 }
1354}
1355
1356static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1357{
1358 nfs_async_write_error(&hdr->pages);
1359}
1360
1361static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1362 .error_cleanup = nfs_async_write_error,
1363 .completion = nfs_write_completion,
1364 .reschedule_io = nfs_async_write_reschedule_io,
1365};
1366
1367void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1368 struct inode *inode, int ioflags, bool force_mds,
1369 const struct nfs_pgio_completion_ops *compl_ops)
1370{
1371 struct nfs_server *server = NFS_SERVER(inode);
1372 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1373
1374#ifdef CONFIG_NFS_V4_1
1375 if (server->pnfs_curr_ld && !force_mds)
1376 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1377#endif
1378 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1379 server->wsize, ioflags);
1380}
1381EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1382
1383void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1384{
1385 struct nfs_pgio_mirror *mirror;
1386
1387 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1388 pgio->pg_ops->pg_cleanup(pgio);
1389
1390 pgio->pg_ops = &nfs_pgio_rw_ops;
1391
1392 nfs_pageio_stop_mirroring(pgio);
1393
1394 mirror = &pgio->pg_mirrors[0];
1395 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1396}
1397EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1398
1399
1400void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1401{
1402 struct nfs_commit_data *data = calldata;
1403
1404 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1405}
1406
1407/*
1408 * Special version of should_remove_suid() that ignores capabilities.
1409 */
1410static int nfs_should_remove_suid(const struct inode *inode)
1411{
1412 umode_t mode = inode->i_mode;
1413 int kill = 0;
1414
1415 /* suid always must be killed */
1416 if (unlikely(mode & S_ISUID))
1417 kill = ATTR_KILL_SUID;
1418
1419 /*
1420 * sgid without any exec bits is just a mandatory locking mark; leave
1421 * it alone. If some exec bits are set, it's a real sgid; kill it.
1422 */
1423 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1424 kill |= ATTR_KILL_SGID;
1425
1426 if (unlikely(kill && S_ISREG(mode)))
1427 return kill;
1428
1429 return 0;
1430}
1431
1432static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1433 struct nfs_fattr *fattr)
1434{
1435 struct nfs_pgio_args *argp = &hdr->args;
1436 struct nfs_pgio_res *resp = &hdr->res;
1437 u64 size = argp->offset + resp->count;
1438
1439 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1440 fattr->size = size;
1441 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1442 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1443 return;
1444 }
1445 if (size != fattr->size)
1446 return;
1447 /* Set attribute barrier */
1448 nfs_fattr_set_barrier(fattr);
1449 /* ...and update size */
1450 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1451}
1452
1453void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1454{
1455 struct nfs_fattr *fattr = &hdr->fattr;
1456 struct inode *inode = hdr->inode;
1457
1458 spin_lock(&inode->i_lock);
1459 nfs_writeback_check_extend(hdr, fattr);
1460 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1461 spin_unlock(&inode->i_lock);
1462}
1463EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1464
1465/*
1466 * This function is called when the WRITE call is complete.
1467 */
1468static int nfs_writeback_done(struct rpc_task *task,
1469 struct nfs_pgio_header *hdr,
1470 struct inode *inode)
1471{
1472 int status;
1473
1474 /*
1475 * ->write_done will attempt to use post-op attributes to detect
1476 * conflicting writes by other clients. A strict interpretation
1477 * of close-to-open would allow us to continue caching even if
1478 * another writer had changed the file, but some applications
1479 * depend on tighter cache coherency when writing.
1480 */
1481 status = NFS_PROTO(inode)->write_done(task, hdr);
1482 if (status != 0)
1483 return status;
1484 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1485
1486 if (hdr->res.verf->committed < hdr->args.stable &&
1487 task->tk_status >= 0) {
1488 /* We tried a write call, but the server did not
1489 * commit data to stable storage even though we
1490 * requested it.
1491 * Note: There is a known bug in Tru64 < 5.0 in which
1492 * the server reports NFS_DATA_SYNC, but performs
1493 * NFS_FILE_SYNC. We therefore implement this checking
1494 * as a dprintk() in order to avoid filling syslog.
1495 */
1496 static unsigned long complain;
1497
1498 /* Note this will print the MDS for a DS write */
1499 if (time_before(complain, jiffies)) {
1500 dprintk("NFS: faulty NFS server %s:"
1501 " (committed = %d) != (stable = %d)\n",
1502 NFS_SERVER(inode)->nfs_client->cl_hostname,
1503 hdr->res.verf->committed, hdr->args.stable);
1504 complain = jiffies + 300 * HZ;
1505 }
1506 }
1507
1508 /* Deal with the suid/sgid bit corner case */
1509 if (nfs_should_remove_suid(inode))
1510 nfs_mark_for_revalidate(inode);
1511 return 0;
1512}
1513
1514/*
1515 * This function is called when the WRITE call is complete.
1516 */
1517static void nfs_writeback_result(struct rpc_task *task,
1518 struct nfs_pgio_header *hdr)
1519{
1520 struct nfs_pgio_args *argp = &hdr->args;
1521 struct nfs_pgio_res *resp = &hdr->res;
1522
1523 if (resp->count < argp->count) {
1524 static unsigned long complain;
1525
1526 /* This a short write! */
1527 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1528
1529 /* Has the server at least made some progress? */
1530 if (resp->count == 0) {
1531 if (time_before(complain, jiffies)) {
1532 printk(KERN_WARNING
1533 "NFS: Server wrote zero bytes, expected %u.\n",
1534 argp->count);
1535 complain = jiffies + 300 * HZ;
1536 }
1537 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1538 task->tk_status = -EIO;
1539 return;
1540 }
1541
1542 /* For non rpc-based layout drivers, retry-through-MDS */
1543 if (!task->tk_ops) {
1544 hdr->pnfs_error = -EAGAIN;
1545 return;
1546 }
1547
1548 /* Was this an NFSv2 write or an NFSv3 stable write? */
1549 if (resp->verf->committed != NFS_UNSTABLE) {
1550 /* Resend from where the server left off */
1551 hdr->mds_offset += resp->count;
1552 argp->offset += resp->count;
1553 argp->pgbase += resp->count;
1554 argp->count -= resp->count;
1555 } else {
1556 /* Resend as a stable write in order to avoid
1557 * headaches in the case of a server crash.
1558 */
1559 argp->stable = NFS_FILE_SYNC;
1560 }
1561 rpc_restart_call_prepare(task);
1562 }
1563}
1564
1565static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1566{
1567 return wait_on_atomic_t(&cinfo->rpcs_out,
1568 nfs_wait_atomic_killable, TASK_KILLABLE);
1569}
1570
1571static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1572{
1573 atomic_inc(&cinfo->rpcs_out);
1574}
1575
1576static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1577{
1578 if (atomic_dec_and_test(&cinfo->rpcs_out))
1579 wake_up_atomic_t(&cinfo->rpcs_out);
1580}
1581
1582void nfs_commitdata_release(struct nfs_commit_data *data)
1583{
1584 put_nfs_open_context(data->context);
1585 nfs_commit_free(data);
1586}
1587EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1588
1589int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1590 const struct nfs_rpc_ops *nfs_ops,
1591 const struct rpc_call_ops *call_ops,
1592 int how, int flags)
1593{
1594 struct rpc_task *task;
1595 int priority = flush_task_priority(how);
1596 struct rpc_message msg = {
1597 .rpc_argp = &data->args,
1598 .rpc_resp = &data->res,
1599 .rpc_cred = data->cred,
1600 };
1601 struct rpc_task_setup task_setup_data = {
1602 .task = &data->task,
1603 .rpc_client = clnt,
1604 .rpc_message = &msg,
1605 .callback_ops = call_ops,
1606 .callback_data = data,
1607 .workqueue = nfsiod_workqueue,
1608 .flags = RPC_TASK_ASYNC | flags,
1609 .priority = priority,
1610 };
1611 /* Set up the initial task struct. */
1612 nfs_ops->commit_setup(data, &msg);
1613
1614 dprintk("NFS: initiated commit call\n");
1615
1616 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1617 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1618
1619 task = rpc_run_task(&task_setup_data);
1620 if (IS_ERR(task))
1621 return PTR_ERR(task);
1622 if (how & FLUSH_SYNC)
1623 rpc_wait_for_completion_task(task);
1624 rpc_put_task(task);
1625 return 0;
1626}
1627EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1628
1629static loff_t nfs_get_lwb(struct list_head *head)
1630{
1631 loff_t lwb = 0;
1632 struct nfs_page *req;
1633
1634 list_for_each_entry(req, head, wb_list)
1635 if (lwb < (req_offset(req) + req->wb_bytes))
1636 lwb = req_offset(req) + req->wb_bytes;
1637
1638 return lwb;
1639}
1640
1641/*
1642 * Set up the argument/result storage required for the RPC call.
1643 */
1644void nfs_init_commit(struct nfs_commit_data *data,
1645 struct list_head *head,
1646 struct pnfs_layout_segment *lseg,
1647 struct nfs_commit_info *cinfo)
1648{
1649 struct nfs_page *first = nfs_list_entry(head->next);
1650 struct inode *inode = d_inode(first->wb_context->dentry);
1651
1652 /* Set up the RPC argument and reply structs
1653 * NB: take care not to mess about with data->commit et al. */
1654
1655 list_splice_init(head, &data->pages);
1656
1657 data->inode = inode;
1658 data->cred = first->wb_context->cred;
1659 data->lseg = lseg; /* reference transferred */
1660 /* only set lwb for pnfs commit */
1661 if (lseg)
1662 data->lwb = nfs_get_lwb(&data->pages);
1663 data->mds_ops = &nfs_commit_ops;
1664 data->completion_ops = cinfo->completion_ops;
1665 data->dreq = cinfo->dreq;
1666
1667 data->args.fh = NFS_FH(data->inode);
1668 /* Note: we always request a commit of the entire inode */
1669 data->args.offset = 0;
1670 data->args.count = 0;
1671 data->context = get_nfs_open_context(first->wb_context);
1672 data->res.fattr = &data->fattr;
1673 data->res.verf = &data->verf;
1674 nfs_fattr_init(&data->fattr);
1675}
1676EXPORT_SYMBOL_GPL(nfs_init_commit);
1677
1678void nfs_retry_commit(struct list_head *page_list,
1679 struct pnfs_layout_segment *lseg,
1680 struct nfs_commit_info *cinfo,
1681 u32 ds_commit_idx)
1682{
1683 struct nfs_page *req;
1684
1685 while (!list_empty(page_list)) {
1686 req = nfs_list_entry(page_list->next);
1687 nfs_list_remove_request(req);
1688 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1689 if (!cinfo->dreq)
1690 nfs_clear_page_commit(req->wb_page);
1691 nfs_unlock_and_release_request(req);
1692 }
1693}
1694EXPORT_SYMBOL_GPL(nfs_retry_commit);
1695
1696static void
1697nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1698 struct nfs_page *req)
1699{
1700 __set_page_dirty_nobuffers(req->wb_page);
1701}
1702
1703/*
1704 * Commit dirty pages
1705 */
1706static int
1707nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1708 struct nfs_commit_info *cinfo)
1709{
1710 struct nfs_commit_data *data;
1711
1712 data = nfs_commitdata_alloc();
1713
1714 if (!data)
1715 goto out_bad;
1716
1717 /* Set up the argument struct */
1718 nfs_init_commit(data, head, NULL, cinfo);
1719 atomic_inc(&cinfo->mds->rpcs_out);
1720 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1721 data->mds_ops, how, 0);
1722 out_bad:
1723 nfs_retry_commit(head, NULL, cinfo, 0);
1724 return -ENOMEM;
1725}
1726
1727/*
1728 * COMMIT call returned
1729 */
1730static void nfs_commit_done(struct rpc_task *task, void *calldata)
1731{
1732 struct nfs_commit_data *data = calldata;
1733
1734 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1735 task->tk_pid, task->tk_status);
1736
1737 /* Call the NFS version-specific code */
1738 NFS_PROTO(data->inode)->commit_done(task, data);
1739}
1740
1741static void nfs_commit_release_pages(struct nfs_commit_data *data)
1742{
1743 struct nfs_page *req;
1744 int status = data->task.tk_status;
1745 struct nfs_commit_info cinfo;
1746 struct nfs_server *nfss;
1747
1748 while (!list_empty(&data->pages)) {
1749 req = nfs_list_entry(data->pages.next);
1750 nfs_list_remove_request(req);
1751 nfs_clear_page_commit(req->wb_page);
1752
1753 dprintk("NFS: commit (%s/%llu %d@%lld)",
1754 req->wb_context->dentry->d_sb->s_id,
1755 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1756 req->wb_bytes,
1757 (long long)req_offset(req));
1758 if (status < 0) {
1759 nfs_context_set_write_error(req->wb_context, status);
1760 nfs_inode_remove_request(req);
1761 dprintk(", error = %d\n", status);
1762 goto next;
1763 }
1764
1765 /* Okay, COMMIT succeeded, apparently. Check the verifier
1766 * returned by the server against all stored verfs. */
1767 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1768 /* We have a match */
1769 nfs_inode_remove_request(req);
1770 dprintk(" OK\n");
1771 goto next;
1772 }
1773 /* We have a mismatch. Write the page again */
1774 dprintk(" mismatch\n");
1775 nfs_mark_request_dirty(req);
1776 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1777 next:
1778 nfs_unlock_and_release_request(req);
1779 }
1780 nfss = NFS_SERVER(data->inode);
1781 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1782 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1783
1784 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1785 nfs_commit_end(cinfo.mds);
1786}
1787
1788static void nfs_commit_release(void *calldata)
1789{
1790 struct nfs_commit_data *data = calldata;
1791
1792 data->completion_ops->completion(data);
1793 nfs_commitdata_release(calldata);
1794}
1795
1796static const struct rpc_call_ops nfs_commit_ops = {
1797 .rpc_call_prepare = nfs_commit_prepare,
1798 .rpc_call_done = nfs_commit_done,
1799 .rpc_release = nfs_commit_release,
1800};
1801
1802static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1803 .completion = nfs_commit_release_pages,
1804 .resched_write = nfs_commit_resched_write,
1805};
1806
1807int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1808 int how, struct nfs_commit_info *cinfo)
1809{
1810 int status;
1811
1812 status = pnfs_commit_list(inode, head, how, cinfo);
1813 if (status == PNFS_NOT_ATTEMPTED)
1814 status = nfs_commit_list(inode, head, how, cinfo);
1815 return status;
1816}
1817
1818int nfs_commit_inode(struct inode *inode, int how)
1819{
1820 LIST_HEAD(head);
1821 struct nfs_commit_info cinfo;
1822 int may_wait = how & FLUSH_SYNC;
1823 int error = 0;
1824 int res;
1825
1826 nfs_init_cinfo_from_inode(&cinfo, inode);
1827 nfs_commit_begin(cinfo.mds);
1828 res = nfs_scan_commit(inode, &head, &cinfo);
1829 if (res)
1830 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1831 nfs_commit_end(cinfo.mds);
1832 if (error < 0)
1833 goto out_error;
1834 if (!may_wait)
1835 goto out_mark_dirty;
1836 error = wait_on_commit(cinfo.mds);
1837 if (error < 0)
1838 return error;
1839 return res;
1840out_error:
1841 res = error;
1842 /* Note: If we exit without ensuring that the commit is complete,
1843 * we must mark the inode as dirty. Otherwise, future calls to
1844 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1845 * that the data is on the disk.
1846 */
1847out_mark_dirty:
1848 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1849 return res;
1850}
1851EXPORT_SYMBOL_GPL(nfs_commit_inode);
1852
1853int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1854{
1855 struct nfs_inode *nfsi = NFS_I(inode);
1856 int flags = FLUSH_SYNC;
1857 int ret = 0;
1858
1859 /* no commits means nothing needs to be done */
1860 if (!nfsi->commit_info.ncommit)
1861 return ret;
1862
1863 if (wbc->sync_mode == WB_SYNC_NONE) {
1864 /* Don't commit yet if this is a non-blocking flush and there
1865 * are a lot of outstanding writes for this mapping.
1866 */
1867 if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1))
1868 goto out_mark_dirty;
1869
1870 /* don't wait for the COMMIT response */
1871 flags = 0;
1872 }
1873
1874 ret = nfs_commit_inode(inode, flags);
1875 if (ret >= 0) {
1876 if (wbc->sync_mode == WB_SYNC_NONE) {
1877 if (ret < wbc->nr_to_write)
1878 wbc->nr_to_write -= ret;
1879 else
1880 wbc->nr_to_write = 0;
1881 }
1882 return 0;
1883 }
1884out_mark_dirty:
1885 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1886 return ret;
1887}
1888EXPORT_SYMBOL_GPL(nfs_write_inode);
1889
1890/*
1891 * flush the inode to disk.
1892 */
1893int nfs_wb_all(struct inode *inode)
1894{
1895 int ret;
1896
1897 trace_nfs_writeback_inode_enter(inode);
1898
1899 ret = filemap_write_and_wait(inode->i_mapping);
1900 if (ret)
1901 goto out;
1902 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1903 if (ret < 0)
1904 goto out;
1905 pnfs_sync_inode(inode, true);
1906 ret = 0;
1907
1908out:
1909 trace_nfs_writeback_inode_exit(inode, ret);
1910 return ret;
1911}
1912EXPORT_SYMBOL_GPL(nfs_wb_all);
1913
1914int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1915{
1916 struct nfs_page *req;
1917 int ret = 0;
1918
1919 wait_on_page_writeback(page);
1920
1921 /* blocking call to cancel all requests and join to a single (head)
1922 * request */
1923 req = nfs_lock_and_join_requests(page, false);
1924
1925 if (IS_ERR(req)) {
1926 ret = PTR_ERR(req);
1927 } else if (req) {
1928 /* all requests from this page have been cancelled by
1929 * nfs_lock_and_join_requests, so just remove the head
1930 * request from the inode / page_private pointer and
1931 * release it */
1932 nfs_inode_remove_request(req);
1933 nfs_unlock_and_release_request(req);
1934 }
1935
1936 return ret;
1937}
1938
1939/*
1940 * Write back all requests on one page - we do this before reading it.
1941 */
1942int nfs_wb_single_page(struct inode *inode, struct page *page, bool launder)
1943{
1944 loff_t range_start = page_file_offset(page);
1945 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
1946 struct writeback_control wbc = {
1947 .sync_mode = WB_SYNC_ALL,
1948 .nr_to_write = 0,
1949 .range_start = range_start,
1950 .range_end = range_end,
1951 };
1952 int ret;
1953
1954 trace_nfs_writeback_page_enter(inode);
1955
1956 for (;;) {
1957 wait_on_page_writeback(page);
1958 if (clear_page_dirty_for_io(page)) {
1959 ret = nfs_writepage_locked(page, &wbc, launder);
1960 if (ret < 0)
1961 goto out_error;
1962 continue;
1963 }
1964 ret = 0;
1965 if (!PagePrivate(page))
1966 break;
1967 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1968 if (ret < 0)
1969 goto out_error;
1970 }
1971out_error:
1972 trace_nfs_writeback_page_exit(inode, ret);
1973 return ret;
1974}
1975
1976#ifdef CONFIG_MIGRATION
1977int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1978 struct page *page, enum migrate_mode mode)
1979{
1980 /*
1981 * If PagePrivate is set, then the page is currently associated with
1982 * an in-progress read or write request. Don't try to migrate it.
1983 *
1984 * FIXME: we could do this in principle, but we'll need a way to ensure
1985 * that we can safely release the inode reference while holding
1986 * the page lock.
1987 */
1988 if (PagePrivate(page))
1989 return -EBUSY;
1990
1991 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1992 return -EBUSY;
1993
1994 return migrate_page(mapping, newpage, page, mode);
1995}
1996#endif
1997
1998int __init nfs_init_writepagecache(void)
1999{
2000 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2001 sizeof(struct nfs_pgio_header),
2002 0, SLAB_HWCACHE_ALIGN,
2003 NULL);
2004 if (nfs_wdata_cachep == NULL)
2005 return -ENOMEM;
2006
2007 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2008 nfs_wdata_cachep);
2009 if (nfs_wdata_mempool == NULL)
2010 goto out_destroy_write_cache;
2011
2012 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2013 sizeof(struct nfs_commit_data),
2014 0, SLAB_HWCACHE_ALIGN,
2015 NULL);
2016 if (nfs_cdata_cachep == NULL)
2017 goto out_destroy_write_mempool;
2018
2019 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2020 nfs_cdata_cachep);
2021 if (nfs_commit_mempool == NULL)
2022 goto out_destroy_commit_cache;
2023
2024 /*
2025 * NFS congestion size, scale with available memory.
2026 *
2027 * 64MB: 8192k
2028 * 128MB: 11585k
2029 * 256MB: 16384k
2030 * 512MB: 23170k
2031 * 1GB: 32768k
2032 * 2GB: 46340k
2033 * 4GB: 65536k
2034 * 8GB: 92681k
2035 * 16GB: 131072k
2036 *
2037 * This allows larger machines to have larger/more transfers.
2038 * Limit the default to 256M
2039 */
2040 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2041 if (nfs_congestion_kb > 256*1024)
2042 nfs_congestion_kb = 256*1024;
2043
2044 return 0;
2045
2046out_destroy_commit_cache:
2047 kmem_cache_destroy(nfs_cdata_cachep);
2048out_destroy_write_mempool:
2049 mempool_destroy(nfs_wdata_mempool);
2050out_destroy_write_cache:
2051 kmem_cache_destroy(nfs_wdata_cachep);
2052 return -ENOMEM;
2053}
2054
2055void nfs_destroy_writepagecache(void)
2056{
2057 mempool_destroy(nfs_commit_mempool);
2058 kmem_cache_destroy(nfs_cdata_cachep);
2059 mempool_destroy(nfs_wdata_mempool);
2060 kmem_cache_destroy(nfs_wdata_cachep);
2061}
2062
2063static const struct nfs_rw_ops nfs_rw_write_ops = {
2064 .rw_mode = FMODE_WRITE,
2065 .rw_alloc_header = nfs_writehdr_alloc,
2066 .rw_free_header = nfs_writehdr_free,
2067 .rw_done = nfs_writeback_done,
2068 .rw_result = nfs_writeback_result,
2069 .rw_initiate = nfs_initiate_write,
2070};