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#include <linux/iversion.h>
27
28#include <linux/uaccess.h>
29
30#include "delegation.h"
31#include "internal.h"
32#include "iostat.h"
33#include "nfs4_fs.h"
34#include "fscache.h"
35#include "pnfs.h"
36
37#include "nfstrace.h"
38
39#define NFSDBG_FACILITY NFSDBG_PAGECACHE
40
41#define MIN_POOL_WRITE (32)
42#define MIN_POOL_COMMIT (4)
43
44struct nfs_io_completion {
45 void (*complete)(void *data);
46 void *data;
47 struct kref refcount;
48};
49
50/*
51 * Local function declarations
52 */
53static void nfs_redirty_request(struct nfs_page *req);
54static const struct rpc_call_ops nfs_commit_ops;
55static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
56static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
57static const struct nfs_rw_ops nfs_rw_write_ops;
58static void nfs_clear_request_commit(struct nfs_page *req);
59static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
60 struct inode *inode);
61static struct nfs_page *
62nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
63 struct page *page);
64
65static struct kmem_cache *nfs_wdata_cachep;
66static mempool_t *nfs_wdata_mempool;
67static struct kmem_cache *nfs_cdata_cachep;
68static mempool_t *nfs_commit_mempool;
69
70struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
71{
72 struct nfs_commit_data *p;
73
74 if (never_fail)
75 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
76 else {
77 /* It is OK to do some reclaim, not no safe to wait
78 * for anything to be returned to the pool.
79 * mempool_alloc() cannot handle that particular combination,
80 * so we need two separate attempts.
81 */
82 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
83 if (!p)
84 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
85 __GFP_NOWARN | __GFP_NORETRY);
86 if (!p)
87 return NULL;
88 }
89
90 memset(p, 0, sizeof(*p));
91 INIT_LIST_HEAD(&p->pages);
92 return p;
93}
94EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
95
96void nfs_commit_free(struct nfs_commit_data *p)
97{
98 mempool_free(p, nfs_commit_mempool);
99}
100EXPORT_SYMBOL_GPL(nfs_commit_free);
101
102static struct nfs_pgio_header *nfs_writehdr_alloc(void)
103{
104 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
105
106 memset(p, 0, sizeof(*p));
107 p->rw_mode = FMODE_WRITE;
108 return p;
109}
110
111static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
112{
113 mempool_free(hdr, nfs_wdata_mempool);
114}
115
116static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
117{
118 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
119}
120
121static void nfs_io_completion_init(struct nfs_io_completion *ioc,
122 void (*complete)(void *), void *data)
123{
124 ioc->complete = complete;
125 ioc->data = data;
126 kref_init(&ioc->refcount);
127}
128
129static void nfs_io_completion_release(struct kref *kref)
130{
131 struct nfs_io_completion *ioc = container_of(kref,
132 struct nfs_io_completion, refcount);
133 ioc->complete(ioc->data);
134 kfree(ioc);
135}
136
137static void nfs_io_completion_get(struct nfs_io_completion *ioc)
138{
139 if (ioc != NULL)
140 kref_get(&ioc->refcount);
141}
142
143static void nfs_io_completion_put(struct nfs_io_completion *ioc)
144{
145 if (ioc != NULL)
146 kref_put(&ioc->refcount, nfs_io_completion_release);
147}
148
149static struct nfs_page *
150nfs_page_private_request(struct page *page)
151{
152 if (!PagePrivate(page))
153 return NULL;
154 return (struct nfs_page *)page_private(page);
155}
156
157/*
158 * nfs_page_find_head_request_locked - find head request associated with @page
159 *
160 * must be called while holding the inode lock.
161 *
162 * returns matching head request with reference held, or NULL if not found.
163 */
164static struct nfs_page *
165nfs_page_find_private_request(struct page *page)
166{
167 struct address_space *mapping = page_file_mapping(page);
168 struct nfs_page *req;
169
170 if (!PagePrivate(page))
171 return NULL;
172 spin_lock(&mapping->private_lock);
173 req = nfs_page_private_request(page);
174 if (req) {
175 WARN_ON_ONCE(req->wb_head != req);
176 kref_get(&req->wb_kref);
177 }
178 spin_unlock(&mapping->private_lock);
179 return req;
180}
181
182static struct nfs_page *
183nfs_page_find_swap_request(struct page *page)
184{
185 struct inode *inode = page_file_mapping(page)->host;
186 struct nfs_inode *nfsi = NFS_I(inode);
187 struct nfs_page *req = NULL;
188 if (!PageSwapCache(page))
189 return NULL;
190 mutex_lock(&nfsi->commit_mutex);
191 if (PageSwapCache(page)) {
192 req = nfs_page_search_commits_for_head_request_locked(nfsi,
193 page);
194 if (req) {
195 WARN_ON_ONCE(req->wb_head != req);
196 kref_get(&req->wb_kref);
197 }
198 }
199 mutex_unlock(&nfsi->commit_mutex);
200 return req;
201}
202
203/*
204 * nfs_page_find_head_request - find head request associated with @page
205 *
206 * returns matching head request with reference held, or NULL if not found.
207 */
208static struct nfs_page *nfs_page_find_head_request(struct page *page)
209{
210 struct nfs_page *req;
211
212 req = nfs_page_find_private_request(page);
213 if (!req)
214 req = nfs_page_find_swap_request(page);
215 return req;
216}
217
218/* Adjust the file length if we're writing beyond the end */
219static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
220{
221 struct inode *inode = page_file_mapping(page)->host;
222 loff_t end, i_size;
223 pgoff_t end_index;
224
225 spin_lock(&inode->i_lock);
226 i_size = i_size_read(inode);
227 end_index = (i_size - 1) >> PAGE_SHIFT;
228 if (i_size > 0 && page_index(page) < end_index)
229 goto out;
230 end = page_file_offset(page) + ((loff_t)offset+count);
231 if (i_size >= end)
232 goto out;
233 i_size_write(inode, end);
234 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
235 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
236out:
237 spin_unlock(&inode->i_lock);
238}
239
240/* A writeback failed: mark the page as bad, and invalidate the page cache */
241static void nfs_set_pageerror(struct page *page)
242{
243 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
244}
245
246/*
247 * nfs_page_group_search_locked
248 * @head - head request of page group
249 * @page_offset - offset into page
250 *
251 * Search page group with head @head to find a request that contains the
252 * page offset @page_offset.
253 *
254 * Returns a pointer to the first matching nfs request, or NULL if no
255 * match is found.
256 *
257 * Must be called with the page group lock held
258 */
259static struct nfs_page *
260nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
261{
262 struct nfs_page *req;
263
264 req = head;
265 do {
266 if (page_offset >= req->wb_pgbase &&
267 page_offset < (req->wb_pgbase + req->wb_bytes))
268 return req;
269
270 req = req->wb_this_page;
271 } while (req != head);
272
273 return NULL;
274}
275
276/*
277 * nfs_page_group_covers_page
278 * @head - head request of page group
279 *
280 * Return true if the page group with head @head covers the whole page,
281 * returns false otherwise
282 */
283static bool nfs_page_group_covers_page(struct nfs_page *req)
284{
285 struct nfs_page *tmp;
286 unsigned int pos = 0;
287 unsigned int len = nfs_page_length(req->wb_page);
288
289 nfs_page_group_lock(req);
290
291 for (;;) {
292 tmp = nfs_page_group_search_locked(req->wb_head, pos);
293 if (!tmp)
294 break;
295 pos = tmp->wb_pgbase + tmp->wb_bytes;
296 }
297
298 nfs_page_group_unlock(req);
299 return pos >= len;
300}
301
302/* We can set the PG_uptodate flag if we see that a write request
303 * covers the full page.
304 */
305static void nfs_mark_uptodate(struct nfs_page *req)
306{
307 if (PageUptodate(req->wb_page))
308 return;
309 if (!nfs_page_group_covers_page(req))
310 return;
311 SetPageUptodate(req->wb_page);
312}
313
314static int wb_priority(struct writeback_control *wbc)
315{
316 int ret = 0;
317
318 if (wbc->sync_mode == WB_SYNC_ALL)
319 ret = FLUSH_COND_STABLE;
320 return ret;
321}
322
323/*
324 * NFS congestion control
325 */
326
327int nfs_congestion_kb;
328
329#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
330#define NFS_CONGESTION_OFF_THRESH \
331 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
332
333static void nfs_set_page_writeback(struct page *page)
334{
335 struct inode *inode = page_file_mapping(page)->host;
336 struct nfs_server *nfss = NFS_SERVER(inode);
337 int ret = test_set_page_writeback(page);
338
339 WARN_ON_ONCE(ret != 0);
340
341 if (atomic_long_inc_return(&nfss->writeback) >
342 NFS_CONGESTION_ON_THRESH)
343 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
344}
345
346static void nfs_end_page_writeback(struct nfs_page *req)
347{
348 struct inode *inode = page_file_mapping(req->wb_page)->host;
349 struct nfs_server *nfss = NFS_SERVER(inode);
350 bool is_done;
351
352 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
353 nfs_unlock_request(req);
354 if (!is_done)
355 return;
356
357 end_page_writeback(req->wb_page);
358 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
359 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
360}
361
362/*
363 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
364 *
365 * this is a helper function for nfs_lock_and_join_requests
366 *
367 * @inode - inode associated with request page group, must be holding inode lock
368 * @head - head request of page group, must be holding head lock
369 * @req - request that couldn't lock and needs to wait on the req bit lock
370 *
371 * NOTE: this must be called holding page_group bit lock
372 * which will be released before returning.
373 *
374 * returns 0 on success, < 0 on error.
375 */
376static void
377nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
378 struct nfs_page *req)
379{
380 struct nfs_page *tmp;
381
382 /* relinquish all the locks successfully grabbed this run */
383 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
384 if (!kref_read(&tmp->wb_kref))
385 continue;
386 nfs_unlock_and_release_request(tmp);
387 }
388}
389
390/*
391 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
392 *
393 * @destroy_list - request list (using wb_this_page) terminated by @old_head
394 * @old_head - the old head of the list
395 *
396 * All subrequests must be locked and removed from all lists, so at this point
397 * they are only "active" in this function, and possibly in nfs_wait_on_request
398 * with a reference held by some other context.
399 */
400static void
401nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
402 struct nfs_page *old_head,
403 struct inode *inode)
404{
405 while (destroy_list) {
406 struct nfs_page *subreq = destroy_list;
407
408 destroy_list = (subreq->wb_this_page == old_head) ?
409 NULL : subreq->wb_this_page;
410
411 WARN_ON_ONCE(old_head != subreq->wb_head);
412
413 /* make sure old group is not used */
414 subreq->wb_this_page = subreq;
415
416 clear_bit(PG_REMOVE, &subreq->wb_flags);
417
418 /* Note: races with nfs_page_group_destroy() */
419 if (!kref_read(&subreq->wb_kref)) {
420 /* Check if we raced with nfs_page_group_destroy() */
421 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
422 nfs_free_request(subreq);
423 continue;
424 }
425
426 subreq->wb_head = subreq;
427
428 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
429 nfs_release_request(subreq);
430 atomic_long_dec(&NFS_I(inode)->nrequests);
431 }
432
433 /* subreq is now totally disconnected from page group or any
434 * write / commit lists. last chance to wake any waiters */
435 nfs_unlock_and_release_request(subreq);
436 }
437}
438
439/*
440 * nfs_lock_and_join_requests - join all subreqs to the head req and return
441 * a locked reference, cancelling any pending
442 * operations for this page.
443 *
444 * @page - the page used to lookup the "page group" of nfs_page structures
445 *
446 * This function joins all sub requests to the head request by first
447 * locking all requests in the group, cancelling any pending operations
448 * and finally updating the head request to cover the whole range covered by
449 * the (former) group. All subrequests are removed from any write or commit
450 * lists, unlinked from the group and destroyed.
451 *
452 * Returns a locked, referenced pointer to the head request - which after
453 * this call is guaranteed to be the only request associated with the page.
454 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
455 * error was encountered.
456 */
457static struct nfs_page *
458nfs_lock_and_join_requests(struct page *page)
459{
460 struct inode *inode = page_file_mapping(page)->host;
461 struct nfs_page *head, *subreq;
462 struct nfs_page *destroy_list = NULL;
463 unsigned int total_bytes;
464 int ret;
465
466try_again:
467 /*
468 * A reference is taken only on the head request which acts as a
469 * reference to the whole page group - the group will not be destroyed
470 * until the head reference is released.
471 */
472 head = nfs_page_find_head_request(page);
473 if (!head)
474 return NULL;
475
476 /* lock the page head first in order to avoid an ABBA inefficiency */
477 if (!nfs_lock_request(head)) {
478 ret = nfs_wait_on_request(head);
479 nfs_release_request(head);
480 if (ret < 0)
481 return ERR_PTR(ret);
482 goto try_again;
483 }
484
485 /* Ensure that nobody removed the request before we locked it */
486 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
487 nfs_unlock_and_release_request(head);
488 goto try_again;
489 }
490
491 ret = nfs_page_group_lock(head);
492 if (ret < 0)
493 goto release_request;
494
495 /* lock each request in the page group */
496 total_bytes = head->wb_bytes;
497 for (subreq = head->wb_this_page; subreq != head;
498 subreq = subreq->wb_this_page) {
499
500 if (!kref_get_unless_zero(&subreq->wb_kref)) {
501 if (subreq->wb_offset == head->wb_offset + total_bytes)
502 total_bytes += subreq->wb_bytes;
503 continue;
504 }
505
506 while (!nfs_lock_request(subreq)) {
507 /*
508 * Unlock page to allow nfs_page_group_sync_on_bit()
509 * to succeed
510 */
511 nfs_page_group_unlock(head);
512 ret = nfs_wait_on_request(subreq);
513 if (!ret)
514 ret = nfs_page_group_lock(head);
515 if (ret < 0) {
516 nfs_unroll_locks(inode, head, subreq);
517 nfs_release_request(subreq);
518 goto release_request;
519 }
520 }
521 /*
522 * Subrequests are always contiguous, non overlapping
523 * and in order - but may be repeated (mirrored writes).
524 */
525 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
526 /* keep track of how many bytes this group covers */
527 total_bytes += subreq->wb_bytes;
528 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
529 ((subreq->wb_offset + subreq->wb_bytes) >
530 (head->wb_offset + total_bytes)))) {
531 nfs_page_group_unlock(head);
532 nfs_unroll_locks(inode, head, subreq);
533 nfs_unlock_and_release_request(subreq);
534 ret = -EIO;
535 goto release_request;
536 }
537 }
538
539 /* Now that all requests are locked, make sure they aren't on any list.
540 * Commit list removal accounting is done after locks are dropped */
541 subreq = head;
542 do {
543 nfs_clear_request_commit(subreq);
544 subreq = subreq->wb_this_page;
545 } while (subreq != head);
546
547 /* unlink subrequests from head, destroy them later */
548 if (head->wb_this_page != head) {
549 /* destroy list will be terminated by head */
550 destroy_list = head->wb_this_page;
551 head->wb_this_page = head;
552
553 /* change head request to cover whole range that
554 * the former page group covered */
555 head->wb_bytes = total_bytes;
556 }
557
558 /* Postpone destruction of this request */
559 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
560 set_bit(PG_INODE_REF, &head->wb_flags);
561 kref_get(&head->wb_kref);
562 atomic_long_inc(&NFS_I(inode)->nrequests);
563 }
564
565 nfs_page_group_unlock(head);
566
567 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
568
569 /* Did we lose a race with nfs_inode_remove_request()? */
570 if (!(PagePrivate(page) || PageSwapCache(page))) {
571 nfs_unlock_and_release_request(head);
572 return NULL;
573 }
574
575 /* still holds ref on head from nfs_page_find_head_request
576 * and still has lock on head from lock loop */
577 return head;
578
579release_request:
580 nfs_unlock_and_release_request(head);
581 return ERR_PTR(ret);
582}
583
584static void nfs_write_error_remove_page(struct nfs_page *req)
585{
586 nfs_end_page_writeback(req);
587 generic_error_remove_page(page_file_mapping(req->wb_page),
588 req->wb_page);
589 nfs_release_request(req);
590}
591
592static bool
593nfs_error_is_fatal_on_server(int err)
594{
595 switch (err) {
596 case 0:
597 case -ERESTARTSYS:
598 case -EINTR:
599 return false;
600 }
601 return nfs_error_is_fatal(err);
602}
603
604/*
605 * Find an associated nfs write request, and prepare to flush it out
606 * May return an error if the user signalled nfs_wait_on_request().
607 */
608static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
609 struct page *page)
610{
611 struct nfs_page *req;
612 int ret = 0;
613
614 req = nfs_lock_and_join_requests(page);
615 if (!req)
616 goto out;
617 ret = PTR_ERR(req);
618 if (IS_ERR(req))
619 goto out;
620
621 nfs_set_page_writeback(page);
622 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
623
624 ret = 0;
625 /* If there is a fatal error that covers this write, just exit */
626 if (nfs_error_is_fatal_on_server(req->wb_context->error))
627 goto out_launder;
628
629 if (!nfs_pageio_add_request(pgio, req)) {
630 ret = pgio->pg_error;
631 /*
632 * Remove the problematic req upon fatal errors on the server
633 */
634 if (nfs_error_is_fatal(ret)) {
635 nfs_context_set_write_error(req->wb_context, ret);
636 if (nfs_error_is_fatal_on_server(ret))
637 goto out_launder;
638 }
639 nfs_redirty_request(req);
640 ret = -EAGAIN;
641 } else
642 nfs_add_stats(page_file_mapping(page)->host,
643 NFSIOS_WRITEPAGES, 1);
644out:
645 return ret;
646out_launder:
647 nfs_write_error_remove_page(req);
648 return ret;
649}
650
651static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
652 struct nfs_pageio_descriptor *pgio)
653{
654 int ret;
655
656 nfs_pageio_cond_complete(pgio, page_index(page));
657 ret = nfs_page_async_flush(pgio, page);
658 if (ret == -EAGAIN) {
659 redirty_page_for_writepage(wbc, page);
660 ret = 0;
661 }
662 return ret;
663}
664
665/*
666 * Write an mmapped page to the server.
667 */
668static int nfs_writepage_locked(struct page *page,
669 struct writeback_control *wbc)
670{
671 struct nfs_pageio_descriptor pgio;
672 struct inode *inode = page_file_mapping(page)->host;
673 int err;
674
675 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
676 nfs_pageio_init_write(&pgio, inode, 0,
677 false, &nfs_async_write_completion_ops);
678 err = nfs_do_writepage(page, wbc, &pgio);
679 nfs_pageio_complete(&pgio);
680 if (err < 0)
681 return err;
682 if (pgio.pg_error < 0)
683 return pgio.pg_error;
684 return 0;
685}
686
687int nfs_writepage(struct page *page, struct writeback_control *wbc)
688{
689 int ret;
690
691 ret = nfs_writepage_locked(page, wbc);
692 unlock_page(page);
693 return ret;
694}
695
696static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
697{
698 int ret;
699
700 ret = nfs_do_writepage(page, wbc, data);
701 unlock_page(page);
702 return ret;
703}
704
705static void nfs_io_completion_commit(void *inode)
706{
707 nfs_commit_inode(inode, 0);
708}
709
710int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
711{
712 struct inode *inode = mapping->host;
713 struct nfs_pageio_descriptor pgio;
714 struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
715 int err;
716
717 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
718
719 if (ioc)
720 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
721
722 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
723 &nfs_async_write_completion_ops);
724 pgio.pg_io_completion = ioc;
725 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
726 nfs_pageio_complete(&pgio);
727 nfs_io_completion_put(ioc);
728
729 if (err < 0)
730 goto out_err;
731 err = pgio.pg_error;
732 if (err < 0)
733 goto out_err;
734 return 0;
735out_err:
736 return err;
737}
738
739/*
740 * Insert a write request into an inode
741 */
742static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
743{
744 struct address_space *mapping = page_file_mapping(req->wb_page);
745 struct nfs_inode *nfsi = NFS_I(inode);
746
747 WARN_ON_ONCE(req->wb_this_page != req);
748
749 /* Lock the request! */
750 nfs_lock_request(req);
751
752 /*
753 * Swap-space should not get truncated. Hence no need to plug the race
754 * with invalidate/truncate.
755 */
756 spin_lock(&mapping->private_lock);
757 if (!nfs_have_writebacks(inode) &&
758 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
759 inode_inc_iversion_raw(inode);
760 if (likely(!PageSwapCache(req->wb_page))) {
761 set_bit(PG_MAPPED, &req->wb_flags);
762 SetPagePrivate(req->wb_page);
763 set_page_private(req->wb_page, (unsigned long)req);
764 }
765 spin_unlock(&mapping->private_lock);
766 atomic_long_inc(&nfsi->nrequests);
767 /* this a head request for a page group - mark it as having an
768 * extra reference so sub groups can follow suit.
769 * This flag also informs pgio layer when to bump nrequests when
770 * adding subrequests. */
771 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
772 kref_get(&req->wb_kref);
773}
774
775/*
776 * Remove a write request from an inode
777 */
778static void nfs_inode_remove_request(struct nfs_page *req)
779{
780 struct address_space *mapping = page_file_mapping(req->wb_page);
781 struct inode *inode = mapping->host;
782 struct nfs_inode *nfsi = NFS_I(inode);
783 struct nfs_page *head;
784
785 atomic_long_dec(&nfsi->nrequests);
786 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
787 head = req->wb_head;
788
789 spin_lock(&mapping->private_lock);
790 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
791 set_page_private(head->wb_page, 0);
792 ClearPagePrivate(head->wb_page);
793 clear_bit(PG_MAPPED, &head->wb_flags);
794 }
795 spin_unlock(&mapping->private_lock);
796 }
797
798 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
799 nfs_release_request(req);
800}
801
802static void
803nfs_mark_request_dirty(struct nfs_page *req)
804{
805 if (req->wb_page)
806 __set_page_dirty_nobuffers(req->wb_page);
807}
808
809/*
810 * nfs_page_search_commits_for_head_request_locked
811 *
812 * Search through commit lists on @inode for the head request for @page.
813 * Must be called while holding the inode (which is cinfo) lock.
814 *
815 * Returns the head request if found, or NULL if not found.
816 */
817static struct nfs_page *
818nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
819 struct page *page)
820{
821 struct nfs_page *freq, *t;
822 struct nfs_commit_info cinfo;
823 struct inode *inode = &nfsi->vfs_inode;
824
825 nfs_init_cinfo_from_inode(&cinfo, inode);
826
827 /* search through pnfs commit lists */
828 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
829 if (freq)
830 return freq->wb_head;
831
832 /* Linearly search the commit list for the correct request */
833 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
834 if (freq->wb_page == page)
835 return freq->wb_head;
836 }
837
838 return NULL;
839}
840
841/**
842 * nfs_request_add_commit_list_locked - add request to a commit list
843 * @req: pointer to a struct nfs_page
844 * @dst: commit list head
845 * @cinfo: holds list lock and accounting info
846 *
847 * This sets the PG_CLEAN bit, updates the cinfo count of
848 * number of outstanding requests requiring a commit as well as
849 * the MM page stats.
850 *
851 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
852 * nfs_page lock.
853 */
854void
855nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
856 struct nfs_commit_info *cinfo)
857{
858 set_bit(PG_CLEAN, &req->wb_flags);
859 nfs_list_add_request(req, dst);
860 atomic_long_inc(&cinfo->mds->ncommit);
861}
862EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
863
864/**
865 * nfs_request_add_commit_list - add request to a commit list
866 * @req: pointer to a struct nfs_page
867 * @dst: commit list head
868 * @cinfo: holds list lock and accounting info
869 *
870 * This sets the PG_CLEAN bit, updates the cinfo count of
871 * number of outstanding requests requiring a commit as well as
872 * the MM page stats.
873 *
874 * The caller must _not_ hold the cinfo->lock, but must be
875 * holding the nfs_page lock.
876 */
877void
878nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
879{
880 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
881 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
882 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
883 if (req->wb_page)
884 nfs_mark_page_unstable(req->wb_page, cinfo);
885}
886EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
887
888/**
889 * nfs_request_remove_commit_list - Remove request from a commit list
890 * @req: pointer to a nfs_page
891 * @cinfo: holds list lock and accounting info
892 *
893 * This clears the PG_CLEAN bit, and updates the cinfo's count of
894 * number of outstanding requests requiring a commit
895 * It does not update the MM page stats.
896 *
897 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
898 */
899void
900nfs_request_remove_commit_list(struct nfs_page *req,
901 struct nfs_commit_info *cinfo)
902{
903 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
904 return;
905 nfs_list_remove_request(req);
906 atomic_long_dec(&cinfo->mds->ncommit);
907}
908EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
909
910static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
911 struct inode *inode)
912{
913 cinfo->inode = inode;
914 cinfo->mds = &NFS_I(inode)->commit_info;
915 cinfo->ds = pnfs_get_ds_info(inode);
916 cinfo->dreq = NULL;
917 cinfo->completion_ops = &nfs_commit_completion_ops;
918}
919
920void nfs_init_cinfo(struct nfs_commit_info *cinfo,
921 struct inode *inode,
922 struct nfs_direct_req *dreq)
923{
924 if (dreq)
925 nfs_init_cinfo_from_dreq(cinfo, dreq);
926 else
927 nfs_init_cinfo_from_inode(cinfo, inode);
928}
929EXPORT_SYMBOL_GPL(nfs_init_cinfo);
930
931/*
932 * Add a request to the inode's commit list.
933 */
934void
935nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
936 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
937{
938 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
939 return;
940 nfs_request_add_commit_list(req, cinfo);
941}
942
943static void
944nfs_clear_page_commit(struct page *page)
945{
946 dec_node_page_state(page, NR_UNSTABLE_NFS);
947 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
948 WB_RECLAIMABLE);
949}
950
951/* Called holding the request lock on @req */
952static void
953nfs_clear_request_commit(struct nfs_page *req)
954{
955 if (test_bit(PG_CLEAN, &req->wb_flags)) {
956 struct inode *inode = d_inode(req->wb_context->dentry);
957 struct nfs_commit_info cinfo;
958
959 nfs_init_cinfo_from_inode(&cinfo, inode);
960 mutex_lock(&NFS_I(inode)->commit_mutex);
961 if (!pnfs_clear_request_commit(req, &cinfo)) {
962 nfs_request_remove_commit_list(req, &cinfo);
963 }
964 mutex_unlock(&NFS_I(inode)->commit_mutex);
965 nfs_clear_page_commit(req->wb_page);
966 }
967}
968
969int nfs_write_need_commit(struct nfs_pgio_header *hdr)
970{
971 if (hdr->verf.committed == NFS_DATA_SYNC)
972 return hdr->lseg == NULL;
973 return hdr->verf.committed != NFS_FILE_SYNC;
974}
975
976static void nfs_async_write_init(struct nfs_pgio_header *hdr)
977{
978 nfs_io_completion_get(hdr->io_completion);
979}
980
981static void nfs_write_completion(struct nfs_pgio_header *hdr)
982{
983 struct nfs_commit_info cinfo;
984 unsigned long bytes = 0;
985
986 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
987 goto out;
988 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
989 while (!list_empty(&hdr->pages)) {
990 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
991
992 bytes += req->wb_bytes;
993 nfs_list_remove_request(req);
994 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
995 (hdr->good_bytes < bytes)) {
996 nfs_set_pageerror(req->wb_page);
997 nfs_context_set_write_error(req->wb_context, hdr->error);
998 goto remove_req;
999 }
1000 if (nfs_write_need_commit(hdr)) {
1001 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1002 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1003 hdr->pgio_mirror_idx);
1004 goto next;
1005 }
1006remove_req:
1007 nfs_inode_remove_request(req);
1008next:
1009 nfs_end_page_writeback(req);
1010 nfs_release_request(req);
1011 }
1012out:
1013 nfs_io_completion_put(hdr->io_completion);
1014 hdr->release(hdr);
1015}
1016
1017unsigned long
1018nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1019{
1020 return atomic_long_read(&cinfo->mds->ncommit);
1021}
1022
1023/* NFS_I(cinfo->inode)->commit_mutex held by caller */
1024int
1025nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1026 struct nfs_commit_info *cinfo, int max)
1027{
1028 struct nfs_page *req, *tmp;
1029 int ret = 0;
1030
1031restart:
1032 list_for_each_entry_safe(req, tmp, src, wb_list) {
1033 kref_get(&req->wb_kref);
1034 if (!nfs_lock_request(req)) {
1035 int status;
1036
1037 /* Prevent deadlock with nfs_lock_and_join_requests */
1038 if (!list_empty(dst)) {
1039 nfs_release_request(req);
1040 continue;
1041 }
1042 /* Ensure we make progress to prevent livelock */
1043 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1044 status = nfs_wait_on_request(req);
1045 nfs_release_request(req);
1046 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1047 if (status < 0)
1048 break;
1049 goto restart;
1050 }
1051 nfs_request_remove_commit_list(req, cinfo);
1052 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1053 nfs_list_add_request(req, dst);
1054 ret++;
1055 if ((ret == max) && !cinfo->dreq)
1056 break;
1057 cond_resched();
1058 }
1059 return ret;
1060}
1061EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1062
1063/*
1064 * nfs_scan_commit - Scan an inode for commit requests
1065 * @inode: NFS inode to scan
1066 * @dst: mds destination list
1067 * @cinfo: mds and ds lists of reqs ready to commit
1068 *
1069 * Moves requests from the inode's 'commit' request list.
1070 * The requests are *not* checked to ensure that they form a contiguous set.
1071 */
1072int
1073nfs_scan_commit(struct inode *inode, struct list_head *dst,
1074 struct nfs_commit_info *cinfo)
1075{
1076 int ret = 0;
1077
1078 if (!atomic_long_read(&cinfo->mds->ncommit))
1079 return 0;
1080 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1081 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1082 const int max = INT_MAX;
1083
1084 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1085 cinfo, max);
1086 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1087 }
1088 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1089 return ret;
1090}
1091
1092/*
1093 * Search for an existing write request, and attempt to update
1094 * it to reflect a new dirty region on a given page.
1095 *
1096 * If the attempt fails, then the existing request is flushed out
1097 * to disk.
1098 */
1099static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1100 struct page *page,
1101 unsigned int offset,
1102 unsigned int bytes)
1103{
1104 struct nfs_page *req;
1105 unsigned int rqend;
1106 unsigned int end;
1107 int error;
1108
1109 end = offset + bytes;
1110
1111 req = nfs_lock_and_join_requests(page);
1112 if (IS_ERR_OR_NULL(req))
1113 return req;
1114
1115 rqend = req->wb_offset + req->wb_bytes;
1116 /*
1117 * Tell the caller to flush out the request if
1118 * the offsets are non-contiguous.
1119 * Note: nfs_flush_incompatible() will already
1120 * have flushed out requests having wrong owners.
1121 */
1122 if (offset > rqend || end < req->wb_offset)
1123 goto out_flushme;
1124
1125 /* Okay, the request matches. Update the region */
1126 if (offset < req->wb_offset) {
1127 req->wb_offset = offset;
1128 req->wb_pgbase = offset;
1129 }
1130 if (end > rqend)
1131 req->wb_bytes = end - req->wb_offset;
1132 else
1133 req->wb_bytes = rqend - req->wb_offset;
1134 return req;
1135out_flushme:
1136 /*
1137 * Note: we mark the request dirty here because
1138 * nfs_lock_and_join_requests() cannot preserve
1139 * commit flags, so we have to replay the write.
1140 */
1141 nfs_mark_request_dirty(req);
1142 nfs_unlock_and_release_request(req);
1143 error = nfs_wb_page(inode, page);
1144 return (error < 0) ? ERR_PTR(error) : NULL;
1145}
1146
1147/*
1148 * Try to update an existing write request, or create one if there is none.
1149 *
1150 * Note: Should always be called with the Page Lock held to prevent races
1151 * if we have to add a new request. Also assumes that the caller has
1152 * already called nfs_flush_incompatible() if necessary.
1153 */
1154static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1155 struct page *page, unsigned int offset, unsigned int bytes)
1156{
1157 struct inode *inode = page_file_mapping(page)->host;
1158 struct nfs_page *req;
1159
1160 req = nfs_try_to_update_request(inode, page, offset, bytes);
1161 if (req != NULL)
1162 goto out;
1163 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1164 if (IS_ERR(req))
1165 goto out;
1166 nfs_inode_add_request(inode, req);
1167out:
1168 return req;
1169}
1170
1171static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1172 unsigned int offset, unsigned int count)
1173{
1174 struct nfs_page *req;
1175
1176 req = nfs_setup_write_request(ctx, page, offset, count);
1177 if (IS_ERR(req))
1178 return PTR_ERR(req);
1179 /* Update file length */
1180 nfs_grow_file(page, offset, count);
1181 nfs_mark_uptodate(req);
1182 nfs_mark_request_dirty(req);
1183 nfs_unlock_and_release_request(req);
1184 return 0;
1185}
1186
1187int nfs_flush_incompatible(struct file *file, struct page *page)
1188{
1189 struct nfs_open_context *ctx = nfs_file_open_context(file);
1190 struct nfs_lock_context *l_ctx;
1191 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1192 struct nfs_page *req;
1193 int do_flush, status;
1194 /*
1195 * Look for a request corresponding to this page. If there
1196 * is one, and it belongs to another file, we flush it out
1197 * before we try to copy anything into the page. Do this
1198 * due to the lack of an ACCESS-type call in NFSv2.
1199 * Also do the same if we find a request from an existing
1200 * dropped page.
1201 */
1202 do {
1203 req = nfs_page_find_head_request(page);
1204 if (req == NULL)
1205 return 0;
1206 l_ctx = req->wb_lock_context;
1207 do_flush = req->wb_page != page ||
1208 !nfs_match_open_context(req->wb_context, ctx);
1209 if (l_ctx && flctx &&
1210 !(list_empty_careful(&flctx->flc_posix) &&
1211 list_empty_careful(&flctx->flc_flock))) {
1212 do_flush |= l_ctx->lockowner != current->files;
1213 }
1214 nfs_release_request(req);
1215 if (!do_flush)
1216 return 0;
1217 status = nfs_wb_page(page_file_mapping(page)->host, page);
1218 } while (status == 0);
1219 return status;
1220}
1221
1222/*
1223 * Avoid buffered writes when a open context credential's key would
1224 * expire soon.
1225 *
1226 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1227 *
1228 * Return 0 and set a credential flag which triggers the inode to flush
1229 * and performs NFS_FILE_SYNC writes if the key will expired within
1230 * RPC_KEY_EXPIRE_TIMEO.
1231 */
1232int
1233nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1234{
1235 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1236 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1237
1238 return rpcauth_key_timeout_notify(auth, ctx->cred);
1239}
1240
1241/*
1242 * Test if the open context credential key is marked to expire soon.
1243 */
1244bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1245{
1246 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1247
1248 return rpcauth_cred_key_to_expire(auth, ctx->cred);
1249}
1250
1251/*
1252 * If the page cache is marked as unsafe or invalid, then we can't rely on
1253 * the PageUptodate() flag. In this case, we will need to turn off
1254 * write optimisations that depend on the page contents being correct.
1255 */
1256static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1257{
1258 struct nfs_inode *nfsi = NFS_I(inode);
1259
1260 if (nfs_have_delegated_attributes(inode))
1261 goto out;
1262 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1263 return false;
1264 smp_rmb();
1265 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1266 return false;
1267out:
1268 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1269 return false;
1270 return PageUptodate(page) != 0;
1271}
1272
1273static bool
1274is_whole_file_wrlock(struct file_lock *fl)
1275{
1276 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1277 fl->fl_type == F_WRLCK;
1278}
1279
1280/* If we know the page is up to date, and we're not using byte range locks (or
1281 * if we have the whole file locked for writing), it may be more efficient to
1282 * extend the write to cover the entire page in order to avoid fragmentation
1283 * inefficiencies.
1284 *
1285 * If the file is opened for synchronous writes then we can just skip the rest
1286 * of the checks.
1287 */
1288static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1289{
1290 int ret;
1291 struct file_lock_context *flctx = inode->i_flctx;
1292 struct file_lock *fl;
1293
1294 if (file->f_flags & O_DSYNC)
1295 return 0;
1296 if (!nfs_write_pageuptodate(page, inode))
1297 return 0;
1298 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1299 return 1;
1300 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1301 list_empty_careful(&flctx->flc_posix)))
1302 return 1;
1303
1304 /* Check to see if there are whole file write locks */
1305 ret = 0;
1306 spin_lock(&flctx->flc_lock);
1307 if (!list_empty(&flctx->flc_posix)) {
1308 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1309 fl_list);
1310 if (is_whole_file_wrlock(fl))
1311 ret = 1;
1312 } else if (!list_empty(&flctx->flc_flock)) {
1313 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1314 fl_list);
1315 if (fl->fl_type == F_WRLCK)
1316 ret = 1;
1317 }
1318 spin_unlock(&flctx->flc_lock);
1319 return ret;
1320}
1321
1322/*
1323 * Update and possibly write a cached page of an NFS file.
1324 *
1325 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1326 * things with a page scheduled for an RPC call (e.g. invalidate it).
1327 */
1328int nfs_updatepage(struct file *file, struct page *page,
1329 unsigned int offset, unsigned int count)
1330{
1331 struct nfs_open_context *ctx = nfs_file_open_context(file);
1332 struct inode *inode = page_file_mapping(page)->host;
1333 int status = 0;
1334
1335 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1336
1337 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1338 file, count, (long long)(page_file_offset(page) + offset));
1339
1340 if (!count)
1341 goto out;
1342
1343 if (nfs_can_extend_write(file, page, inode)) {
1344 count = max(count + offset, nfs_page_length(page));
1345 offset = 0;
1346 }
1347
1348 status = nfs_writepage_setup(ctx, page, offset, count);
1349 if (status < 0)
1350 nfs_set_pageerror(page);
1351 else
1352 __set_page_dirty_nobuffers(page);
1353out:
1354 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1355 status, (long long)i_size_read(inode));
1356 return status;
1357}
1358
1359static int flush_task_priority(int how)
1360{
1361 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1362 case FLUSH_HIGHPRI:
1363 return RPC_PRIORITY_HIGH;
1364 case FLUSH_LOWPRI:
1365 return RPC_PRIORITY_LOW;
1366 }
1367 return RPC_PRIORITY_NORMAL;
1368}
1369
1370static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1371 struct rpc_message *msg,
1372 const struct nfs_rpc_ops *rpc_ops,
1373 struct rpc_task_setup *task_setup_data, int how)
1374{
1375 int priority = flush_task_priority(how);
1376
1377 task_setup_data->priority = priority;
1378 rpc_ops->write_setup(hdr, msg);
1379 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1380 hdr->args.stable);
1381
1382 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1383 &task_setup_data->rpc_client, msg, hdr);
1384}
1385
1386/* If a nfs_flush_* function fails, it should remove reqs from @head and
1387 * call this on each, which will prepare them to be retried on next
1388 * writeback using standard nfs.
1389 */
1390static void nfs_redirty_request(struct nfs_page *req)
1391{
1392 nfs_mark_request_dirty(req);
1393 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1394 nfs_end_page_writeback(req);
1395 nfs_release_request(req);
1396}
1397
1398static void nfs_async_write_error(struct list_head *head)
1399{
1400 struct nfs_page *req;
1401
1402 while (!list_empty(head)) {
1403 req = nfs_list_entry(head->next);
1404 nfs_list_remove_request(req);
1405 nfs_redirty_request(req);
1406 }
1407}
1408
1409static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1410{
1411 nfs_async_write_error(&hdr->pages);
1412}
1413
1414static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1415 .init_hdr = nfs_async_write_init,
1416 .error_cleanup = nfs_async_write_error,
1417 .completion = nfs_write_completion,
1418 .reschedule_io = nfs_async_write_reschedule_io,
1419};
1420
1421void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1422 struct inode *inode, int ioflags, bool force_mds,
1423 const struct nfs_pgio_completion_ops *compl_ops)
1424{
1425 struct nfs_server *server = NFS_SERVER(inode);
1426 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1427
1428#ifdef CONFIG_NFS_V4_1
1429 if (server->pnfs_curr_ld && !force_mds)
1430 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1431#endif
1432 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1433 server->wsize, ioflags);
1434}
1435EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1436
1437void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1438{
1439 struct nfs_pgio_mirror *mirror;
1440
1441 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1442 pgio->pg_ops->pg_cleanup(pgio);
1443
1444 pgio->pg_ops = &nfs_pgio_rw_ops;
1445
1446 nfs_pageio_stop_mirroring(pgio);
1447
1448 mirror = &pgio->pg_mirrors[0];
1449 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1450}
1451EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1452
1453
1454void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1455{
1456 struct nfs_commit_data *data = calldata;
1457
1458 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1459}
1460
1461/*
1462 * Special version of should_remove_suid() that ignores capabilities.
1463 */
1464static int nfs_should_remove_suid(const struct inode *inode)
1465{
1466 umode_t mode = inode->i_mode;
1467 int kill = 0;
1468
1469 /* suid always must be killed */
1470 if (unlikely(mode & S_ISUID))
1471 kill = ATTR_KILL_SUID;
1472
1473 /*
1474 * sgid without any exec bits is just a mandatory locking mark; leave
1475 * it alone. If some exec bits are set, it's a real sgid; kill it.
1476 */
1477 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1478 kill |= ATTR_KILL_SGID;
1479
1480 if (unlikely(kill && S_ISREG(mode)))
1481 return kill;
1482
1483 return 0;
1484}
1485
1486static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1487 struct nfs_fattr *fattr)
1488{
1489 struct nfs_pgio_args *argp = &hdr->args;
1490 struct nfs_pgio_res *resp = &hdr->res;
1491 u64 size = argp->offset + resp->count;
1492
1493 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1494 fattr->size = size;
1495 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1496 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1497 return;
1498 }
1499 if (size != fattr->size)
1500 return;
1501 /* Set attribute barrier */
1502 nfs_fattr_set_barrier(fattr);
1503 /* ...and update size */
1504 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1505}
1506
1507void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1508{
1509 struct nfs_fattr *fattr = &hdr->fattr;
1510 struct inode *inode = hdr->inode;
1511
1512 spin_lock(&inode->i_lock);
1513 nfs_writeback_check_extend(hdr, fattr);
1514 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1515 spin_unlock(&inode->i_lock);
1516}
1517EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1518
1519/*
1520 * This function is called when the WRITE call is complete.
1521 */
1522static int nfs_writeback_done(struct rpc_task *task,
1523 struct nfs_pgio_header *hdr,
1524 struct inode *inode)
1525{
1526 int status;
1527
1528 /*
1529 * ->write_done will attempt to use post-op attributes to detect
1530 * conflicting writes by other clients. A strict interpretation
1531 * of close-to-open would allow us to continue caching even if
1532 * another writer had changed the file, but some applications
1533 * depend on tighter cache coherency when writing.
1534 */
1535 status = NFS_PROTO(inode)->write_done(task, hdr);
1536 if (status != 0)
1537 return status;
1538
1539 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1540 trace_nfs_writeback_done(inode, task->tk_status,
1541 hdr->args.offset, hdr->res.verf);
1542
1543 if (hdr->res.verf->committed < hdr->args.stable &&
1544 task->tk_status >= 0) {
1545 /* We tried a write call, but the server did not
1546 * commit data to stable storage even though we
1547 * requested it.
1548 * Note: There is a known bug in Tru64 < 5.0 in which
1549 * the server reports NFS_DATA_SYNC, but performs
1550 * NFS_FILE_SYNC. We therefore implement this checking
1551 * as a dprintk() in order to avoid filling syslog.
1552 */
1553 static unsigned long complain;
1554
1555 /* Note this will print the MDS for a DS write */
1556 if (time_before(complain, jiffies)) {
1557 dprintk("NFS: faulty NFS server %s:"
1558 " (committed = %d) != (stable = %d)\n",
1559 NFS_SERVER(inode)->nfs_client->cl_hostname,
1560 hdr->res.verf->committed, hdr->args.stable);
1561 complain = jiffies + 300 * HZ;
1562 }
1563 }
1564
1565 /* Deal with the suid/sgid bit corner case */
1566 if (nfs_should_remove_suid(inode)) {
1567 spin_lock(&inode->i_lock);
1568 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1569 spin_unlock(&inode->i_lock);
1570 }
1571 return 0;
1572}
1573
1574/*
1575 * This function is called when the WRITE call is complete.
1576 */
1577static void nfs_writeback_result(struct rpc_task *task,
1578 struct nfs_pgio_header *hdr)
1579{
1580 struct nfs_pgio_args *argp = &hdr->args;
1581 struct nfs_pgio_res *resp = &hdr->res;
1582
1583 if (resp->count < argp->count) {
1584 static unsigned long complain;
1585
1586 /* This a short write! */
1587 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1588
1589 /* Has the server at least made some progress? */
1590 if (resp->count == 0) {
1591 if (time_before(complain, jiffies)) {
1592 printk(KERN_WARNING
1593 "NFS: Server wrote zero bytes, expected %u.\n",
1594 argp->count);
1595 complain = jiffies + 300 * HZ;
1596 }
1597 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1598 task->tk_status = -EIO;
1599 return;
1600 }
1601
1602 /* For non rpc-based layout drivers, retry-through-MDS */
1603 if (!task->tk_ops) {
1604 hdr->pnfs_error = -EAGAIN;
1605 return;
1606 }
1607
1608 /* Was this an NFSv2 write or an NFSv3 stable write? */
1609 if (resp->verf->committed != NFS_UNSTABLE) {
1610 /* Resend from where the server left off */
1611 hdr->mds_offset += resp->count;
1612 argp->offset += resp->count;
1613 argp->pgbase += resp->count;
1614 argp->count -= resp->count;
1615 } else {
1616 /* Resend as a stable write in order to avoid
1617 * headaches in the case of a server crash.
1618 */
1619 argp->stable = NFS_FILE_SYNC;
1620 }
1621 rpc_restart_call_prepare(task);
1622 }
1623}
1624
1625static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1626{
1627 return wait_var_event_killable(&cinfo->rpcs_out,
1628 !atomic_read(&cinfo->rpcs_out));
1629}
1630
1631static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1632{
1633 atomic_inc(&cinfo->rpcs_out);
1634}
1635
1636static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1637{
1638 if (atomic_dec_and_test(&cinfo->rpcs_out))
1639 wake_up_var(&cinfo->rpcs_out);
1640}
1641
1642void nfs_commitdata_release(struct nfs_commit_data *data)
1643{
1644 put_nfs_open_context(data->context);
1645 nfs_commit_free(data);
1646}
1647EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1648
1649int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1650 const struct nfs_rpc_ops *nfs_ops,
1651 const struct rpc_call_ops *call_ops,
1652 int how, int flags)
1653{
1654 struct rpc_task *task;
1655 int priority = flush_task_priority(how);
1656 struct rpc_message msg = {
1657 .rpc_argp = &data->args,
1658 .rpc_resp = &data->res,
1659 .rpc_cred = data->cred,
1660 };
1661 struct rpc_task_setup task_setup_data = {
1662 .task = &data->task,
1663 .rpc_client = clnt,
1664 .rpc_message = &msg,
1665 .callback_ops = call_ops,
1666 .callback_data = data,
1667 .workqueue = nfsiod_workqueue,
1668 .flags = RPC_TASK_ASYNC | flags,
1669 .priority = priority,
1670 };
1671 /* Set up the initial task struct. */
1672 nfs_ops->commit_setup(data, &msg);
1673 trace_nfs_initiate_commit(data);
1674
1675 dprintk("NFS: initiated commit call\n");
1676
1677 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1678 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1679
1680 task = rpc_run_task(&task_setup_data);
1681 if (IS_ERR(task))
1682 return PTR_ERR(task);
1683 if (how & FLUSH_SYNC)
1684 rpc_wait_for_completion_task(task);
1685 rpc_put_task(task);
1686 return 0;
1687}
1688EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1689
1690static loff_t nfs_get_lwb(struct list_head *head)
1691{
1692 loff_t lwb = 0;
1693 struct nfs_page *req;
1694
1695 list_for_each_entry(req, head, wb_list)
1696 if (lwb < (req_offset(req) + req->wb_bytes))
1697 lwb = req_offset(req) + req->wb_bytes;
1698
1699 return lwb;
1700}
1701
1702/*
1703 * Set up the argument/result storage required for the RPC call.
1704 */
1705void nfs_init_commit(struct nfs_commit_data *data,
1706 struct list_head *head,
1707 struct pnfs_layout_segment *lseg,
1708 struct nfs_commit_info *cinfo)
1709{
1710 struct nfs_page *first = nfs_list_entry(head->next);
1711 struct inode *inode = d_inode(first->wb_context->dentry);
1712
1713 /* Set up the RPC argument and reply structs
1714 * NB: take care not to mess about with data->commit et al. */
1715
1716 list_splice_init(head, &data->pages);
1717
1718 data->inode = inode;
1719 data->cred = first->wb_context->cred;
1720 data->lseg = lseg; /* reference transferred */
1721 /* only set lwb for pnfs commit */
1722 if (lseg)
1723 data->lwb = nfs_get_lwb(&data->pages);
1724 data->mds_ops = &nfs_commit_ops;
1725 data->completion_ops = cinfo->completion_ops;
1726 data->dreq = cinfo->dreq;
1727
1728 data->args.fh = NFS_FH(data->inode);
1729 /* Note: we always request a commit of the entire inode */
1730 data->args.offset = 0;
1731 data->args.count = 0;
1732 data->context = get_nfs_open_context(first->wb_context);
1733 data->res.fattr = &data->fattr;
1734 data->res.verf = &data->verf;
1735 nfs_fattr_init(&data->fattr);
1736}
1737EXPORT_SYMBOL_GPL(nfs_init_commit);
1738
1739void nfs_retry_commit(struct list_head *page_list,
1740 struct pnfs_layout_segment *lseg,
1741 struct nfs_commit_info *cinfo,
1742 u32 ds_commit_idx)
1743{
1744 struct nfs_page *req;
1745
1746 while (!list_empty(page_list)) {
1747 req = nfs_list_entry(page_list->next);
1748 nfs_list_remove_request(req);
1749 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1750 if (!cinfo->dreq)
1751 nfs_clear_page_commit(req->wb_page);
1752 nfs_unlock_and_release_request(req);
1753 }
1754}
1755EXPORT_SYMBOL_GPL(nfs_retry_commit);
1756
1757static void
1758nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1759 struct nfs_page *req)
1760{
1761 __set_page_dirty_nobuffers(req->wb_page);
1762}
1763
1764/*
1765 * Commit dirty pages
1766 */
1767static int
1768nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1769 struct nfs_commit_info *cinfo)
1770{
1771 struct nfs_commit_data *data;
1772
1773 /* another commit raced with us */
1774 if (list_empty(head))
1775 return 0;
1776
1777 data = nfs_commitdata_alloc(true);
1778
1779 /* Set up the argument struct */
1780 nfs_init_commit(data, head, NULL, cinfo);
1781 atomic_inc(&cinfo->mds->rpcs_out);
1782 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1783 data->mds_ops, how, 0);
1784}
1785
1786/*
1787 * COMMIT call returned
1788 */
1789static void nfs_commit_done(struct rpc_task *task, void *calldata)
1790{
1791 struct nfs_commit_data *data = calldata;
1792
1793 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1794 task->tk_pid, task->tk_status);
1795
1796 /* Call the NFS version-specific code */
1797 NFS_PROTO(data->inode)->commit_done(task, data);
1798 trace_nfs_commit_done(data);
1799}
1800
1801static void nfs_commit_release_pages(struct nfs_commit_data *data)
1802{
1803 struct nfs_page *req;
1804 int status = data->task.tk_status;
1805 struct nfs_commit_info cinfo;
1806 struct nfs_server *nfss;
1807
1808 while (!list_empty(&data->pages)) {
1809 req = nfs_list_entry(data->pages.next);
1810 nfs_list_remove_request(req);
1811 if (req->wb_page)
1812 nfs_clear_page_commit(req->wb_page);
1813
1814 dprintk("NFS: commit (%s/%llu %d@%lld)",
1815 req->wb_context->dentry->d_sb->s_id,
1816 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1817 req->wb_bytes,
1818 (long long)req_offset(req));
1819 if (status < 0) {
1820 nfs_context_set_write_error(req->wb_context, status);
1821 if (req->wb_page)
1822 nfs_inode_remove_request(req);
1823 dprintk_cont(", error = %d\n", status);
1824 goto next;
1825 }
1826
1827 /* Okay, COMMIT succeeded, apparently. Check the verifier
1828 * returned by the server against all stored verfs. */
1829 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1830 /* We have a match */
1831 if (req->wb_page)
1832 nfs_inode_remove_request(req);
1833 dprintk_cont(" OK\n");
1834 goto next;
1835 }
1836 /* We have a mismatch. Write the page again */
1837 dprintk_cont(" mismatch\n");
1838 nfs_mark_request_dirty(req);
1839 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1840 next:
1841 nfs_unlock_and_release_request(req);
1842 /* Latency breaker */
1843 cond_resched();
1844 }
1845 nfss = NFS_SERVER(data->inode);
1846 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1847 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1848
1849 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1850 nfs_commit_end(cinfo.mds);
1851}
1852
1853static void nfs_commit_release(void *calldata)
1854{
1855 struct nfs_commit_data *data = calldata;
1856
1857 data->completion_ops->completion(data);
1858 nfs_commitdata_release(calldata);
1859}
1860
1861static const struct rpc_call_ops nfs_commit_ops = {
1862 .rpc_call_prepare = nfs_commit_prepare,
1863 .rpc_call_done = nfs_commit_done,
1864 .rpc_release = nfs_commit_release,
1865};
1866
1867static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1868 .completion = nfs_commit_release_pages,
1869 .resched_write = nfs_commit_resched_write,
1870};
1871
1872int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1873 int how, struct nfs_commit_info *cinfo)
1874{
1875 int status;
1876
1877 status = pnfs_commit_list(inode, head, how, cinfo);
1878 if (status == PNFS_NOT_ATTEMPTED)
1879 status = nfs_commit_list(inode, head, how, cinfo);
1880 return status;
1881}
1882
1883static int __nfs_commit_inode(struct inode *inode, int how,
1884 struct writeback_control *wbc)
1885{
1886 LIST_HEAD(head);
1887 struct nfs_commit_info cinfo;
1888 int may_wait = how & FLUSH_SYNC;
1889 int ret, nscan;
1890
1891 nfs_init_cinfo_from_inode(&cinfo, inode);
1892 nfs_commit_begin(cinfo.mds);
1893 for (;;) {
1894 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1895 if (ret <= 0)
1896 break;
1897 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1898 if (ret < 0)
1899 break;
1900 ret = 0;
1901 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1902 if (nscan < wbc->nr_to_write)
1903 wbc->nr_to_write -= nscan;
1904 else
1905 wbc->nr_to_write = 0;
1906 }
1907 if (nscan < INT_MAX)
1908 break;
1909 cond_resched();
1910 }
1911 nfs_commit_end(cinfo.mds);
1912 if (ret || !may_wait)
1913 return ret;
1914 return wait_on_commit(cinfo.mds);
1915}
1916
1917int nfs_commit_inode(struct inode *inode, int how)
1918{
1919 return __nfs_commit_inode(inode, how, NULL);
1920}
1921EXPORT_SYMBOL_GPL(nfs_commit_inode);
1922
1923int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1924{
1925 struct nfs_inode *nfsi = NFS_I(inode);
1926 int flags = FLUSH_SYNC;
1927 int ret = 0;
1928
1929 if (wbc->sync_mode == WB_SYNC_NONE) {
1930 /* no commits means nothing needs to be done */
1931 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1932 goto check_requests_outstanding;
1933
1934 /* Don't commit yet if this is a non-blocking flush and there
1935 * are a lot of outstanding writes for this mapping.
1936 */
1937 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1938 goto out_mark_dirty;
1939
1940 /* don't wait for the COMMIT response */
1941 flags = 0;
1942 }
1943
1944 ret = __nfs_commit_inode(inode, flags, wbc);
1945 if (!ret) {
1946 if (flags & FLUSH_SYNC)
1947 return 0;
1948 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1949 goto out_mark_dirty;
1950
1951check_requests_outstanding:
1952 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1953 return ret;
1954out_mark_dirty:
1955 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1956 return ret;
1957}
1958EXPORT_SYMBOL_GPL(nfs_write_inode);
1959
1960/*
1961 * Wrapper for filemap_write_and_wait_range()
1962 *
1963 * Needed for pNFS in order to ensure data becomes visible to the
1964 * client.
1965 */
1966int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1967 loff_t lstart, loff_t lend)
1968{
1969 int ret;
1970
1971 ret = filemap_write_and_wait_range(mapping, lstart, lend);
1972 if (ret == 0)
1973 ret = pnfs_sync_inode(mapping->host, true);
1974 return ret;
1975}
1976EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1977
1978/*
1979 * flush the inode to disk.
1980 */
1981int nfs_wb_all(struct inode *inode)
1982{
1983 int ret;
1984
1985 trace_nfs_writeback_inode_enter(inode);
1986
1987 ret = filemap_write_and_wait(inode->i_mapping);
1988 if (ret)
1989 goto out;
1990 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1991 if (ret < 0)
1992 goto out;
1993 pnfs_sync_inode(inode, true);
1994 ret = 0;
1995
1996out:
1997 trace_nfs_writeback_inode_exit(inode, ret);
1998 return ret;
1999}
2000EXPORT_SYMBOL_GPL(nfs_wb_all);
2001
2002int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2003{
2004 struct nfs_page *req;
2005 int ret = 0;
2006
2007 wait_on_page_writeback(page);
2008
2009 /* blocking call to cancel all requests and join to a single (head)
2010 * request */
2011 req = nfs_lock_and_join_requests(page);
2012
2013 if (IS_ERR(req)) {
2014 ret = PTR_ERR(req);
2015 } else if (req) {
2016 /* all requests from this page have been cancelled by
2017 * nfs_lock_and_join_requests, so just remove the head
2018 * request from the inode / page_private pointer and
2019 * release it */
2020 nfs_inode_remove_request(req);
2021 nfs_unlock_and_release_request(req);
2022 }
2023
2024 return ret;
2025}
2026
2027/*
2028 * Write back all requests on one page - we do this before reading it.
2029 */
2030int nfs_wb_page(struct inode *inode, struct page *page)
2031{
2032 loff_t range_start = page_file_offset(page);
2033 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2034 struct writeback_control wbc = {
2035 .sync_mode = WB_SYNC_ALL,
2036 .nr_to_write = 0,
2037 .range_start = range_start,
2038 .range_end = range_end,
2039 };
2040 int ret;
2041
2042 trace_nfs_writeback_page_enter(inode);
2043
2044 for (;;) {
2045 wait_on_page_writeback(page);
2046 if (clear_page_dirty_for_io(page)) {
2047 ret = nfs_writepage_locked(page, &wbc);
2048 if (ret < 0)
2049 goto out_error;
2050 continue;
2051 }
2052 ret = 0;
2053 if (!PagePrivate(page))
2054 break;
2055 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2056 if (ret < 0)
2057 goto out_error;
2058 }
2059out_error:
2060 trace_nfs_writeback_page_exit(inode, ret);
2061 return ret;
2062}
2063
2064#ifdef CONFIG_MIGRATION
2065int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2066 struct page *page, enum migrate_mode mode)
2067{
2068 /*
2069 * If PagePrivate is set, then the page is currently associated with
2070 * an in-progress read or write request. Don't try to migrate it.
2071 *
2072 * FIXME: we could do this in principle, but we'll need a way to ensure
2073 * that we can safely release the inode reference while holding
2074 * the page lock.
2075 */
2076 if (PagePrivate(page))
2077 return -EBUSY;
2078
2079 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2080 return -EBUSY;
2081
2082 return migrate_page(mapping, newpage, page, mode);
2083}
2084#endif
2085
2086int __init nfs_init_writepagecache(void)
2087{
2088 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2089 sizeof(struct nfs_pgio_header),
2090 0, SLAB_HWCACHE_ALIGN,
2091 NULL);
2092 if (nfs_wdata_cachep == NULL)
2093 return -ENOMEM;
2094
2095 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2096 nfs_wdata_cachep);
2097 if (nfs_wdata_mempool == NULL)
2098 goto out_destroy_write_cache;
2099
2100 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2101 sizeof(struct nfs_commit_data),
2102 0, SLAB_HWCACHE_ALIGN,
2103 NULL);
2104 if (nfs_cdata_cachep == NULL)
2105 goto out_destroy_write_mempool;
2106
2107 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2108 nfs_cdata_cachep);
2109 if (nfs_commit_mempool == NULL)
2110 goto out_destroy_commit_cache;
2111
2112 /*
2113 * NFS congestion size, scale with available memory.
2114 *
2115 * 64MB: 8192k
2116 * 128MB: 11585k
2117 * 256MB: 16384k
2118 * 512MB: 23170k
2119 * 1GB: 32768k
2120 * 2GB: 46340k
2121 * 4GB: 65536k
2122 * 8GB: 92681k
2123 * 16GB: 131072k
2124 *
2125 * This allows larger machines to have larger/more transfers.
2126 * Limit the default to 256M
2127 */
2128 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2129 if (nfs_congestion_kb > 256*1024)
2130 nfs_congestion_kb = 256*1024;
2131
2132 return 0;
2133
2134out_destroy_commit_cache:
2135 kmem_cache_destroy(nfs_cdata_cachep);
2136out_destroy_write_mempool:
2137 mempool_destroy(nfs_wdata_mempool);
2138out_destroy_write_cache:
2139 kmem_cache_destroy(nfs_wdata_cachep);
2140 return -ENOMEM;
2141}
2142
2143void nfs_destroy_writepagecache(void)
2144{
2145 mempool_destroy(nfs_commit_mempool);
2146 kmem_cache_destroy(nfs_cdata_cachep);
2147 mempool_destroy(nfs_wdata_mempool);
2148 kmem_cache_destroy(nfs_wdata_cachep);
2149}
2150
2151static const struct nfs_rw_ops nfs_rw_write_ops = {
2152 .rw_alloc_header = nfs_writehdr_alloc,
2153 .rw_free_header = nfs_writehdr_free,
2154 .rw_done = nfs_writeback_done,
2155 .rw_result = nfs_writeback_result,
2156 .rw_initiate = nfs_initiate_write,
2157};