Merge tag 'locks-v3.20-1' of git://git.samba.org/jlayton/linux
[firefly-linux-kernel-4.4.55.git] / fs / nfs / write.c
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
25 #include <asm/uaccess.h>
26
27 #include "delegation.h"
28 #include "internal.h"
29 #include "iostat.h"
30 #include "nfs4_fs.h"
31 #include "fscache.h"
32 #include "pnfs.h"
33
34 #include "nfstrace.h"
35
36 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
37
38 #define MIN_POOL_WRITE          (32)
39 #define MIN_POOL_COMMIT         (4)
40
41 /*
42  * Local function declarations
43  */
44 static void nfs_redirty_request(struct nfs_page *req);
45 static const struct rpc_call_ops nfs_commit_ops;
46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
48 static const struct nfs_rw_ops nfs_rw_write_ops;
49 static void nfs_clear_request_commit(struct nfs_page *req);
50 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
51                                       struct inode *inode);
52 static struct nfs_page *
53 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
54                                                 struct page *page);
55
56 static struct kmem_cache *nfs_wdata_cachep;
57 static mempool_t *nfs_wdata_mempool;
58 static struct kmem_cache *nfs_cdata_cachep;
59 static mempool_t *nfs_commit_mempool;
60
61 struct nfs_commit_data *nfs_commitdata_alloc(void)
62 {
63         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
64
65         if (p) {
66                 memset(p, 0, sizeof(*p));
67                 INIT_LIST_HEAD(&p->pages);
68         }
69         return p;
70 }
71 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
72
73 void nfs_commit_free(struct nfs_commit_data *p)
74 {
75         mempool_free(p, nfs_commit_mempool);
76 }
77 EXPORT_SYMBOL_GPL(nfs_commit_free);
78
79 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
80 {
81         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
82
83         if (p)
84                 memset(p, 0, sizeof(*p));
85         return p;
86 }
87
88 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
89 {
90         mempool_free(hdr, nfs_wdata_mempool);
91 }
92
93 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
94 {
95         ctx->error = error;
96         smp_wmb();
97         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
98 }
99
100 /*
101  * nfs_page_find_head_request_locked - find head request associated with @page
102  *
103  * must be called while holding the inode lock.
104  *
105  * returns matching head request with reference held, or NULL if not found.
106  */
107 static struct nfs_page *
108 nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page)
109 {
110         struct nfs_page *req = NULL;
111
112         if (PagePrivate(page))
113                 req = (struct nfs_page *)page_private(page);
114         else if (unlikely(PageSwapCache(page)))
115                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
116                         page);
117
118         if (req) {
119                 WARN_ON_ONCE(req->wb_head != req);
120                 kref_get(&req->wb_kref);
121         }
122
123         return req;
124 }
125
126 /*
127  * nfs_page_find_head_request - find head request associated with @page
128  *
129  * returns matching head request with reference held, or NULL if not found.
130  */
131 static struct nfs_page *nfs_page_find_head_request(struct page *page)
132 {
133         struct inode *inode = page_file_mapping(page)->host;
134         struct nfs_page *req = NULL;
135
136         spin_lock(&inode->i_lock);
137         req = nfs_page_find_head_request_locked(NFS_I(inode), page);
138         spin_unlock(&inode->i_lock);
139         return req;
140 }
141
142 /* Adjust the file length if we're writing beyond the end */
143 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
144 {
145         struct inode *inode = page_file_mapping(page)->host;
146         loff_t end, i_size;
147         pgoff_t end_index;
148
149         spin_lock(&inode->i_lock);
150         i_size = i_size_read(inode);
151         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
152         if (i_size > 0 && page_file_index(page) < end_index)
153                 goto out;
154         end = page_file_offset(page) + ((loff_t)offset+count);
155         if (i_size >= end)
156                 goto out;
157         i_size_write(inode, end);
158         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
159 out:
160         spin_unlock(&inode->i_lock);
161 }
162
163 /* A writeback failed: mark the page as bad, and invalidate the page cache */
164 static void nfs_set_pageerror(struct page *page)
165 {
166         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
167 }
168
169 /*
170  * nfs_page_group_search_locked
171  * @head - head request of page group
172  * @page_offset - offset into page
173  *
174  * Search page group with head @head to find a request that contains the
175  * page offset @page_offset.
176  *
177  * Returns a pointer to the first matching nfs request, or NULL if no
178  * match is found.
179  *
180  * Must be called with the page group lock held
181  */
182 static struct nfs_page *
183 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
184 {
185         struct nfs_page *req;
186
187         WARN_ON_ONCE(head != head->wb_head);
188         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
189
190         req = head;
191         do {
192                 if (page_offset >= req->wb_pgbase &&
193                     page_offset < (req->wb_pgbase + req->wb_bytes))
194                         return req;
195
196                 req = req->wb_this_page;
197         } while (req != head);
198
199         return NULL;
200 }
201
202 /*
203  * nfs_page_group_covers_page
204  * @head - head request of page group
205  *
206  * Return true if the page group with head @head covers the whole page,
207  * returns false otherwise
208  */
209 static bool nfs_page_group_covers_page(struct nfs_page *req)
210 {
211         struct nfs_page *tmp;
212         unsigned int pos = 0;
213         unsigned int len = nfs_page_length(req->wb_page);
214
215         nfs_page_group_lock(req, false);
216
217         do {
218                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
219                 if (tmp) {
220                         /* no way this should happen */
221                         WARN_ON_ONCE(tmp->wb_pgbase != pos);
222                         pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
223                 }
224         } while (tmp && pos < len);
225
226         nfs_page_group_unlock(req);
227         WARN_ON_ONCE(pos > len);
228         return pos == len;
229 }
230
231 /* We can set the PG_uptodate flag if we see that a write request
232  * covers the full page.
233  */
234 static void nfs_mark_uptodate(struct nfs_page *req)
235 {
236         if (PageUptodate(req->wb_page))
237                 return;
238         if (!nfs_page_group_covers_page(req))
239                 return;
240         SetPageUptodate(req->wb_page);
241 }
242
243 static int wb_priority(struct writeback_control *wbc)
244 {
245         int ret = 0;
246         if (wbc->for_reclaim)
247                 return FLUSH_HIGHPRI | FLUSH_STABLE;
248         if (wbc->sync_mode == WB_SYNC_ALL)
249                 ret = FLUSH_COND_STABLE;
250         if (wbc->for_kupdate || wbc->for_background)
251                 ret |= FLUSH_LOWPRI;
252         return ret;
253 }
254
255 /*
256  * NFS congestion control
257  */
258
259 int nfs_congestion_kb;
260
261 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
262 #define NFS_CONGESTION_OFF_THRESH       \
263         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
264
265 static void nfs_set_page_writeback(struct page *page)
266 {
267         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
268         int ret = test_set_page_writeback(page);
269
270         WARN_ON_ONCE(ret != 0);
271
272         if (atomic_long_inc_return(&nfss->writeback) >
273                         NFS_CONGESTION_ON_THRESH) {
274                 set_bdi_congested(&nfss->backing_dev_info,
275                                         BLK_RW_ASYNC);
276         }
277 }
278
279 static void nfs_end_page_writeback(struct nfs_page *req)
280 {
281         struct inode *inode = page_file_mapping(req->wb_page)->host;
282         struct nfs_server *nfss = NFS_SERVER(inode);
283
284         if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
285                 return;
286
287         end_page_writeback(req->wb_page);
288         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
289                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
290 }
291
292
293 /* nfs_page_group_clear_bits
294  *   @req - an nfs request
295  * clears all page group related bits from @req
296  */
297 static void
298 nfs_page_group_clear_bits(struct nfs_page *req)
299 {
300         clear_bit(PG_TEARDOWN, &req->wb_flags);
301         clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
302         clear_bit(PG_UPTODATE, &req->wb_flags);
303         clear_bit(PG_WB_END, &req->wb_flags);
304         clear_bit(PG_REMOVE, &req->wb_flags);
305 }
306
307
308 /*
309  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
310  *
311  * this is a helper function for nfs_lock_and_join_requests
312  *
313  * @inode - inode associated with request page group, must be holding inode lock
314  * @head  - head request of page group, must be holding head lock
315  * @req   - request that couldn't lock and needs to wait on the req bit lock
316  * @nonblock - if true, don't actually wait
317  *
318  * NOTE: this must be called holding page_group bit lock and inode spin lock
319  *       and BOTH will be released before returning.
320  *
321  * returns 0 on success, < 0 on error.
322  */
323 static int
324 nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
325                           struct nfs_page *req, bool nonblock)
326         __releases(&inode->i_lock)
327 {
328         struct nfs_page *tmp;
329         int ret;
330
331         /* relinquish all the locks successfully grabbed this run */
332         for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
333                 nfs_unlock_request(tmp);
334
335         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
336
337         /* grab a ref on the request that will be waited on */
338         kref_get(&req->wb_kref);
339
340         nfs_page_group_unlock(head);
341         spin_unlock(&inode->i_lock);
342
343         /* release ref from nfs_page_find_head_request_locked */
344         nfs_release_request(head);
345
346         if (!nonblock)
347                 ret = nfs_wait_on_request(req);
348         else
349                 ret = -EAGAIN;
350         nfs_release_request(req);
351
352         return ret;
353 }
354
355 /*
356  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
357  *
358  * @destroy_list - request list (using wb_this_page) terminated by @old_head
359  * @old_head - the old head of the list
360  *
361  * All subrequests must be locked and removed from all lists, so at this point
362  * they are only "active" in this function, and possibly in nfs_wait_on_request
363  * with a reference held by some other context.
364  */
365 static void
366 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
367                                  struct nfs_page *old_head)
368 {
369         while (destroy_list) {
370                 struct nfs_page *subreq = destroy_list;
371
372                 destroy_list = (subreq->wb_this_page == old_head) ?
373                                    NULL : subreq->wb_this_page;
374
375                 WARN_ON_ONCE(old_head != subreq->wb_head);
376
377                 /* make sure old group is not used */
378                 subreq->wb_head = subreq;
379                 subreq->wb_this_page = subreq;
380
381                 /* subreq is now totally disconnected from page group or any
382                  * write / commit lists. last chance to wake any waiters */
383                 nfs_unlock_request(subreq);
384
385                 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
386                         /* release ref on old head request */
387                         nfs_release_request(old_head);
388
389                         nfs_page_group_clear_bits(subreq);
390
391                         /* release the PG_INODE_REF reference */
392                         if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
393                                 nfs_release_request(subreq);
394                         else
395                                 WARN_ON_ONCE(1);
396                 } else {
397                         WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
398                         /* zombie requests have already released the last
399                          * reference and were waiting on the rest of the
400                          * group to complete. Since it's no longer part of a
401                          * group, simply free the request */
402                         nfs_page_group_clear_bits(subreq);
403                         nfs_free_request(subreq);
404                 }
405         }
406 }
407
408 /*
409  * nfs_lock_and_join_requests - join all subreqs to the head req and return
410  *                              a locked reference, cancelling any pending
411  *                              operations for this page.
412  *
413  * @page - the page used to lookup the "page group" of nfs_page structures
414  * @nonblock - if true, don't block waiting for request locks
415  *
416  * This function joins all sub requests to the head request by first
417  * locking all requests in the group, cancelling any pending operations
418  * and finally updating the head request to cover the whole range covered by
419  * the (former) group.  All subrequests are removed from any write or commit
420  * lists, unlinked from the group and destroyed.
421  *
422  * Returns a locked, referenced pointer to the head request - which after
423  * this call is guaranteed to be the only request associated with the page.
424  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
425  * error was encountered.
426  */
427 static struct nfs_page *
428 nfs_lock_and_join_requests(struct page *page, bool nonblock)
429 {
430         struct inode *inode = page_file_mapping(page)->host;
431         struct nfs_page *head, *subreq;
432         struct nfs_page *destroy_list = NULL;
433         unsigned int total_bytes;
434         int ret;
435
436 try_again:
437         total_bytes = 0;
438
439         WARN_ON_ONCE(destroy_list);
440
441         spin_lock(&inode->i_lock);
442
443         /*
444          * A reference is taken only on the head request which acts as a
445          * reference to the whole page group - the group will not be destroyed
446          * until the head reference is released.
447          */
448         head = nfs_page_find_head_request_locked(NFS_I(inode), page);
449
450         if (!head) {
451                 spin_unlock(&inode->i_lock);
452                 return NULL;
453         }
454
455         /* holding inode lock, so always make a non-blocking call to try the
456          * page group lock */
457         ret = nfs_page_group_lock(head, true);
458         if (ret < 0) {
459                 spin_unlock(&inode->i_lock);
460
461                 if (!nonblock && ret == -EAGAIN) {
462                         nfs_page_group_lock_wait(head);
463                         nfs_release_request(head);
464                         goto try_again;
465                 }
466
467                 nfs_release_request(head);
468                 return ERR_PTR(ret);
469         }
470
471         /* lock each request in the page group */
472         subreq = head;
473         do {
474                 /*
475                  * Subrequests are always contiguous, non overlapping
476                  * and in order. If not, it's a programming error.
477                  */
478                 WARN_ON_ONCE(subreq->wb_offset !=
479                      (head->wb_offset + total_bytes));
480
481                 /* keep track of how many bytes this group covers */
482                 total_bytes += subreq->wb_bytes;
483
484                 if (!nfs_lock_request(subreq)) {
485                         /* releases page group bit lock and
486                          * inode spin lock and all references */
487                         ret = nfs_unroll_locks_and_wait(inode, head,
488                                 subreq, nonblock);
489
490                         if (ret == 0)
491                                 goto try_again;
492
493                         return ERR_PTR(ret);
494                 }
495
496                 subreq = subreq->wb_this_page;
497         } while (subreq != head);
498
499         /* Now that all requests are locked, make sure they aren't on any list.
500          * Commit list removal accounting is done after locks are dropped */
501         subreq = head;
502         do {
503                 nfs_clear_request_commit(subreq);
504                 subreq = subreq->wb_this_page;
505         } while (subreq != head);
506
507         /* unlink subrequests from head, destroy them later */
508         if (head->wb_this_page != head) {
509                 /* destroy list will be terminated by head */
510                 destroy_list = head->wb_this_page;
511                 head->wb_this_page = head;
512
513                 /* change head request to cover whole range that
514                  * the former page group covered */
515                 head->wb_bytes = total_bytes;
516         }
517
518         /*
519          * prepare head request to be added to new pgio descriptor
520          */
521         nfs_page_group_clear_bits(head);
522
523         /*
524          * some part of the group was still on the inode list - otherwise
525          * the group wouldn't be involved in async write.
526          * grab a reference for the head request, iff it needs one.
527          */
528         if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
529                 kref_get(&head->wb_kref);
530
531         nfs_page_group_unlock(head);
532
533         /* drop lock to clean uprequests on destroy list */
534         spin_unlock(&inode->i_lock);
535
536         nfs_destroy_unlinked_subrequests(destroy_list, head);
537
538         /* still holds ref on head from nfs_page_find_head_request_locked
539          * and still has lock on head from lock loop */
540         return head;
541 }
542
543 /*
544  * Find an associated nfs write request, and prepare to flush it out
545  * May return an error if the user signalled nfs_wait_on_request().
546  */
547 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
548                                 struct page *page, bool nonblock)
549 {
550         struct nfs_page *req;
551         int ret = 0;
552
553         req = nfs_lock_and_join_requests(page, nonblock);
554         if (!req)
555                 goto out;
556         ret = PTR_ERR(req);
557         if (IS_ERR(req))
558                 goto out;
559
560         nfs_set_page_writeback(page);
561         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
562
563         ret = 0;
564         if (!nfs_pageio_add_request(pgio, req)) {
565                 nfs_redirty_request(req);
566                 ret = pgio->pg_error;
567         }
568 out:
569         return ret;
570 }
571
572 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
573 {
574         struct inode *inode = page_file_mapping(page)->host;
575         int ret;
576
577         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
578         nfs_inc_stats(inode, NFSIOS_WRITEPAGES);
579
580         nfs_pageio_cond_complete(pgio, page_file_index(page));
581         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
582         if (ret == -EAGAIN) {
583                 redirty_page_for_writepage(wbc, page);
584                 ret = 0;
585         }
586         return ret;
587 }
588
589 /*
590  * Write an mmapped page to the server.
591  */
592 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
593 {
594         struct nfs_pageio_descriptor pgio;
595         int err;
596
597         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
598                                 false, &nfs_async_write_completion_ops);
599         err = nfs_do_writepage(page, wbc, &pgio);
600         nfs_pageio_complete(&pgio);
601         if (err < 0)
602                 return err;
603         if (pgio.pg_error < 0)
604                 return pgio.pg_error;
605         return 0;
606 }
607
608 int nfs_writepage(struct page *page, struct writeback_control *wbc)
609 {
610         int ret;
611
612         ret = nfs_writepage_locked(page, wbc);
613         unlock_page(page);
614         return ret;
615 }
616
617 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
618 {
619         int ret;
620
621         ret = nfs_do_writepage(page, wbc, data);
622         unlock_page(page);
623         return ret;
624 }
625
626 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
627 {
628         struct inode *inode = mapping->host;
629         unsigned long *bitlock = &NFS_I(inode)->flags;
630         struct nfs_pageio_descriptor pgio;
631         int err;
632
633         /* Stop dirtying of new pages while we sync */
634         err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING,
635                         nfs_wait_bit_killable, TASK_KILLABLE);
636         if (err)
637                 goto out_err;
638
639         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
640
641         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
642                                 &nfs_async_write_completion_ops);
643         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
644         nfs_pageio_complete(&pgio);
645
646         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
647         smp_mb__after_atomic();
648         wake_up_bit(bitlock, NFS_INO_FLUSHING);
649
650         if (err < 0)
651                 goto out_err;
652         err = pgio.pg_error;
653         if (err < 0)
654                 goto out_err;
655         return 0;
656 out_err:
657         return err;
658 }
659
660 /*
661  * Insert a write request into an inode
662  */
663 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
664 {
665         struct nfs_inode *nfsi = NFS_I(inode);
666
667         WARN_ON_ONCE(req->wb_this_page != req);
668
669         /* Lock the request! */
670         nfs_lock_request(req);
671
672         spin_lock(&inode->i_lock);
673         if (!nfsi->nrequests &&
674             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
675                 inode->i_version++;
676         /*
677          * Swap-space should not get truncated. Hence no need to plug the race
678          * with invalidate/truncate.
679          */
680         if (likely(!PageSwapCache(req->wb_page))) {
681                 set_bit(PG_MAPPED, &req->wb_flags);
682                 SetPagePrivate(req->wb_page);
683                 set_page_private(req->wb_page, (unsigned long)req);
684         }
685         nfsi->nrequests++;
686         /* this a head request for a page group - mark it as having an
687          * extra reference so sub groups can follow suit.
688          * This flag also informs pgio layer when to bump nrequests when
689          * adding subrequests. */
690         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
691         kref_get(&req->wb_kref);
692         spin_unlock(&inode->i_lock);
693 }
694
695 /*
696  * Remove a write request from an inode
697  */
698 static void nfs_inode_remove_request(struct nfs_page *req)
699 {
700         struct inode *inode = req->wb_context->dentry->d_inode;
701         struct nfs_inode *nfsi = NFS_I(inode);
702         struct nfs_page *head;
703
704         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
705                 head = req->wb_head;
706
707                 spin_lock(&inode->i_lock);
708                 if (likely(!PageSwapCache(head->wb_page))) {
709                         set_page_private(head->wb_page, 0);
710                         ClearPagePrivate(head->wb_page);
711                         smp_mb__after_atomic();
712                         wake_up_page(head->wb_page, PG_private);
713                         clear_bit(PG_MAPPED, &head->wb_flags);
714                 }
715                 nfsi->nrequests--;
716                 spin_unlock(&inode->i_lock);
717         } else {
718                 spin_lock(&inode->i_lock);
719                 nfsi->nrequests--;
720                 spin_unlock(&inode->i_lock);
721         }
722
723         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
724                 nfs_release_request(req);
725 }
726
727 static void
728 nfs_mark_request_dirty(struct nfs_page *req)
729 {
730         __set_page_dirty_nobuffers(req->wb_page);
731 }
732
733 /*
734  * nfs_page_search_commits_for_head_request_locked
735  *
736  * Search through commit lists on @inode for the head request for @page.
737  * Must be called while holding the inode (which is cinfo) lock.
738  *
739  * Returns the head request if found, or NULL if not found.
740  */
741 static struct nfs_page *
742 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
743                                                 struct page *page)
744 {
745         struct nfs_page *freq, *t;
746         struct nfs_commit_info cinfo;
747         struct inode *inode = &nfsi->vfs_inode;
748
749         nfs_init_cinfo_from_inode(&cinfo, inode);
750
751         /* search through pnfs commit lists */
752         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
753         if (freq)
754                 return freq->wb_head;
755
756         /* Linearly search the commit list for the correct request */
757         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
758                 if (freq->wb_page == page)
759                         return freq->wb_head;
760         }
761
762         return NULL;
763 }
764
765 /**
766  * nfs_request_add_commit_list - add request to a commit list
767  * @req: pointer to a struct nfs_page
768  * @dst: commit list head
769  * @cinfo: holds list lock and accounting info
770  *
771  * This sets the PG_CLEAN bit, updates the cinfo count of
772  * number of outstanding requests requiring a commit as well as
773  * the MM page stats.
774  *
775  * The caller must _not_ hold the cinfo->lock, but must be
776  * holding the nfs_page lock.
777  */
778 void
779 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
780                             struct nfs_commit_info *cinfo)
781 {
782         set_bit(PG_CLEAN, &(req)->wb_flags);
783         spin_lock(cinfo->lock);
784         nfs_list_add_request(req, dst);
785         cinfo->mds->ncommit++;
786         spin_unlock(cinfo->lock);
787         if (!cinfo->dreq) {
788                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
789                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
790                              BDI_RECLAIMABLE);
791                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
792                                    I_DIRTY_DATASYNC);
793         }
794 }
795 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
796
797 /**
798  * nfs_request_remove_commit_list - Remove request from a commit list
799  * @req: pointer to a nfs_page
800  * @cinfo: holds list lock and accounting info
801  *
802  * This clears the PG_CLEAN bit, and updates the cinfo's count of
803  * number of outstanding requests requiring a commit
804  * It does not update the MM page stats.
805  *
806  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
807  */
808 void
809 nfs_request_remove_commit_list(struct nfs_page *req,
810                                struct nfs_commit_info *cinfo)
811 {
812         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
813                 return;
814         nfs_list_remove_request(req);
815         cinfo->mds->ncommit--;
816 }
817 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
818
819 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
820                                       struct inode *inode)
821 {
822         cinfo->lock = &inode->i_lock;
823         cinfo->mds = &NFS_I(inode)->commit_info;
824         cinfo->ds = pnfs_get_ds_info(inode);
825         cinfo->dreq = NULL;
826         cinfo->completion_ops = &nfs_commit_completion_ops;
827 }
828
829 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
830                     struct inode *inode,
831                     struct nfs_direct_req *dreq)
832 {
833         if (dreq)
834                 nfs_init_cinfo_from_dreq(cinfo, dreq);
835         else
836                 nfs_init_cinfo_from_inode(cinfo, inode);
837 }
838 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
839
840 /*
841  * Add a request to the inode's commit list.
842  */
843 void
844 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
845                         struct nfs_commit_info *cinfo)
846 {
847         if (pnfs_mark_request_commit(req, lseg, cinfo))
848                 return;
849         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
850 }
851
852 static void
853 nfs_clear_page_commit(struct page *page)
854 {
855         dec_zone_page_state(page, NR_UNSTABLE_NFS);
856         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
857 }
858
859 /* Called holding inode (/cinfo) lock */
860 static void
861 nfs_clear_request_commit(struct nfs_page *req)
862 {
863         if (test_bit(PG_CLEAN, &req->wb_flags)) {
864                 struct inode *inode = req->wb_context->dentry->d_inode;
865                 struct nfs_commit_info cinfo;
866
867                 nfs_init_cinfo_from_inode(&cinfo, inode);
868                 if (!pnfs_clear_request_commit(req, &cinfo)) {
869                         nfs_request_remove_commit_list(req, &cinfo);
870                 }
871                 nfs_clear_page_commit(req->wb_page);
872         }
873 }
874
875 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
876 {
877         if (hdr->verf.committed == NFS_DATA_SYNC)
878                 return hdr->lseg == NULL;
879         return hdr->verf.committed != NFS_FILE_SYNC;
880 }
881
882 static void nfs_write_completion(struct nfs_pgio_header *hdr)
883 {
884         struct nfs_commit_info cinfo;
885         unsigned long bytes = 0;
886
887         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
888                 goto out;
889         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
890         while (!list_empty(&hdr->pages)) {
891                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
892
893                 bytes += req->wb_bytes;
894                 nfs_list_remove_request(req);
895                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
896                     (hdr->good_bytes < bytes)) {
897                         nfs_set_pageerror(req->wb_page);
898                         nfs_context_set_write_error(req->wb_context, hdr->error);
899                         goto remove_req;
900                 }
901                 if (nfs_write_need_commit(hdr)) {
902                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
903                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
904                         goto next;
905                 }
906 remove_req:
907                 nfs_inode_remove_request(req);
908 next:
909                 nfs_unlock_request(req);
910                 nfs_end_page_writeback(req);
911                 nfs_release_request(req);
912         }
913 out:
914         hdr->release(hdr);
915 }
916
917 unsigned long
918 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
919 {
920         return cinfo->mds->ncommit;
921 }
922
923 /* cinfo->lock held by caller */
924 int
925 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
926                      struct nfs_commit_info *cinfo, int max)
927 {
928         struct nfs_page *req, *tmp;
929         int ret = 0;
930
931         list_for_each_entry_safe(req, tmp, src, wb_list) {
932                 if (!nfs_lock_request(req))
933                         continue;
934                 kref_get(&req->wb_kref);
935                 if (cond_resched_lock(cinfo->lock))
936                         list_safe_reset_next(req, tmp, wb_list);
937                 nfs_request_remove_commit_list(req, cinfo);
938                 nfs_list_add_request(req, dst);
939                 ret++;
940                 if ((ret == max) && !cinfo->dreq)
941                         break;
942         }
943         return ret;
944 }
945
946 /*
947  * nfs_scan_commit - Scan an inode for commit requests
948  * @inode: NFS inode to scan
949  * @dst: mds destination list
950  * @cinfo: mds and ds lists of reqs ready to commit
951  *
952  * Moves requests from the inode's 'commit' request list.
953  * The requests are *not* checked to ensure that they form a contiguous set.
954  */
955 int
956 nfs_scan_commit(struct inode *inode, struct list_head *dst,
957                 struct nfs_commit_info *cinfo)
958 {
959         int ret = 0;
960
961         spin_lock(cinfo->lock);
962         if (cinfo->mds->ncommit > 0) {
963                 const int max = INT_MAX;
964
965                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
966                                            cinfo, max);
967                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
968         }
969         spin_unlock(cinfo->lock);
970         return ret;
971 }
972
973 /*
974  * Search for an existing write request, and attempt to update
975  * it to reflect a new dirty region on a given page.
976  *
977  * If the attempt fails, then the existing request is flushed out
978  * to disk.
979  */
980 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
981                 struct page *page,
982                 unsigned int offset,
983                 unsigned int bytes)
984 {
985         struct nfs_page *req;
986         unsigned int rqend;
987         unsigned int end;
988         int error;
989
990         if (!PagePrivate(page))
991                 return NULL;
992
993         end = offset + bytes;
994         spin_lock(&inode->i_lock);
995
996         for (;;) {
997                 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
998                 if (req == NULL)
999                         goto out_unlock;
1000
1001                 /* should be handled by nfs_flush_incompatible */
1002                 WARN_ON_ONCE(req->wb_head != req);
1003                 WARN_ON_ONCE(req->wb_this_page != req);
1004
1005                 rqend = req->wb_offset + req->wb_bytes;
1006                 /*
1007                  * Tell the caller to flush out the request if
1008                  * the offsets are non-contiguous.
1009                  * Note: nfs_flush_incompatible() will already
1010                  * have flushed out requests having wrong owners.
1011                  */
1012                 if (offset > rqend
1013                     || end < req->wb_offset)
1014                         goto out_flushme;
1015
1016                 if (nfs_lock_request(req))
1017                         break;
1018
1019                 /* The request is locked, so wait and then retry */
1020                 spin_unlock(&inode->i_lock);
1021                 error = nfs_wait_on_request(req);
1022                 nfs_release_request(req);
1023                 if (error != 0)
1024                         goto out_err;
1025                 spin_lock(&inode->i_lock);
1026         }
1027
1028         /* Okay, the request matches. Update the region */
1029         if (offset < req->wb_offset) {
1030                 req->wb_offset = offset;
1031                 req->wb_pgbase = offset;
1032         }
1033         if (end > rqend)
1034                 req->wb_bytes = end - req->wb_offset;
1035         else
1036                 req->wb_bytes = rqend - req->wb_offset;
1037 out_unlock:
1038         if (req)
1039                 nfs_clear_request_commit(req);
1040         spin_unlock(&inode->i_lock);
1041         return req;
1042 out_flushme:
1043         spin_unlock(&inode->i_lock);
1044         nfs_release_request(req);
1045         error = nfs_wb_page(inode, page);
1046 out_err:
1047         return ERR_PTR(error);
1048 }
1049
1050 /*
1051  * Try to update an existing write request, or create one if there is none.
1052  *
1053  * Note: Should always be called with the Page Lock held to prevent races
1054  * if we have to add a new request. Also assumes that the caller has
1055  * already called nfs_flush_incompatible() if necessary.
1056  */
1057 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1058                 struct page *page, unsigned int offset, unsigned int bytes)
1059 {
1060         struct inode *inode = page_file_mapping(page)->host;
1061         struct nfs_page *req;
1062
1063         req = nfs_try_to_update_request(inode, page, offset, bytes);
1064         if (req != NULL)
1065                 goto out;
1066         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1067         if (IS_ERR(req))
1068                 goto out;
1069         nfs_inode_add_request(inode, req);
1070 out:
1071         return req;
1072 }
1073
1074 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1075                 unsigned int offset, unsigned int count)
1076 {
1077         struct nfs_page *req;
1078
1079         req = nfs_setup_write_request(ctx, page, offset, count);
1080         if (IS_ERR(req))
1081                 return PTR_ERR(req);
1082         /* Update file length */
1083         nfs_grow_file(page, offset, count);
1084         nfs_mark_uptodate(req);
1085         nfs_mark_request_dirty(req);
1086         nfs_unlock_and_release_request(req);
1087         return 0;
1088 }
1089
1090 int nfs_flush_incompatible(struct file *file, struct page *page)
1091 {
1092         struct nfs_open_context *ctx = nfs_file_open_context(file);
1093         struct nfs_lock_context *l_ctx;
1094         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1095         struct nfs_page *req;
1096         int do_flush, status;
1097         /*
1098          * Look for a request corresponding to this page. If there
1099          * is one, and it belongs to another file, we flush it out
1100          * before we try to copy anything into the page. Do this
1101          * due to the lack of an ACCESS-type call in NFSv2.
1102          * Also do the same if we find a request from an existing
1103          * dropped page.
1104          */
1105         do {
1106                 req = nfs_page_find_head_request(page);
1107                 if (req == NULL)
1108                         return 0;
1109                 l_ctx = req->wb_lock_context;
1110                 do_flush = req->wb_page != page || req->wb_context != ctx;
1111                 /* for now, flush if more than 1 request in page_group */
1112                 do_flush |= req->wb_this_page != req;
1113                 if (l_ctx && flctx &&
1114                     !(list_empty_careful(&flctx->flc_posix) &&
1115                       list_empty_careful(&flctx->flc_flock))) {
1116                         do_flush |= l_ctx->lockowner.l_owner != current->files
1117                                 || l_ctx->lockowner.l_pid != current->tgid;
1118                 }
1119                 nfs_release_request(req);
1120                 if (!do_flush)
1121                         return 0;
1122                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1123         } while (status == 0);
1124         return status;
1125 }
1126
1127 /*
1128  * Avoid buffered writes when a open context credential's key would
1129  * expire soon.
1130  *
1131  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1132  *
1133  * Return 0 and set a credential flag which triggers the inode to flush
1134  * and performs  NFS_FILE_SYNC writes if the key will expired within
1135  * RPC_KEY_EXPIRE_TIMEO.
1136  */
1137 int
1138 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1139 {
1140         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1141         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1142
1143         return rpcauth_key_timeout_notify(auth, ctx->cred);
1144 }
1145
1146 /*
1147  * Test if the open context credential key is marked to expire soon.
1148  */
1149 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1150 {
1151         return rpcauth_cred_key_to_expire(ctx->cred);
1152 }
1153
1154 /*
1155  * If the page cache is marked as unsafe or invalid, then we can't rely on
1156  * the PageUptodate() flag. In this case, we will need to turn off
1157  * write optimisations that depend on the page contents being correct.
1158  */
1159 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1160 {
1161         struct nfs_inode *nfsi = NFS_I(inode);
1162
1163         if (nfs_have_delegated_attributes(inode))
1164                 goto out;
1165         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1166                 return false;
1167         smp_rmb();
1168         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1169                 return false;
1170 out:
1171         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1172                 return false;
1173         return PageUptodate(page) != 0;
1174 }
1175
1176 static bool
1177 is_whole_file_wrlock(struct file_lock *fl)
1178 {
1179         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1180                         fl->fl_type == F_WRLCK;
1181 }
1182
1183 /* If we know the page is up to date, and we're not using byte range locks (or
1184  * if we have the whole file locked for writing), it may be more efficient to
1185  * extend the write to cover the entire page in order to avoid fragmentation
1186  * inefficiencies.
1187  *
1188  * If the file is opened for synchronous writes then we can just skip the rest
1189  * of the checks.
1190  */
1191 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1192 {
1193         int ret;
1194         struct file_lock_context *flctx = inode->i_flctx;
1195         struct file_lock *fl;
1196
1197         if (file->f_flags & O_DSYNC)
1198                 return 0;
1199         if (!nfs_write_pageuptodate(page, inode))
1200                 return 0;
1201         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1202                 return 1;
1203         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1204                        list_empty_careful(&flctx->flc_posix)))
1205                 return 0;
1206
1207         /* Check to see if there are whole file write locks */
1208         ret = 0;
1209         spin_lock(&flctx->flc_lock);
1210         if (!list_empty(&flctx->flc_posix)) {
1211                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1212                                         fl_list);
1213                 if (is_whole_file_wrlock(fl))
1214                         ret = 1;
1215         } else if (!list_empty(&flctx->flc_flock)) {
1216                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1217                                         fl_list);
1218                 if (fl->fl_type == F_WRLCK)
1219                         ret = 1;
1220         }
1221         spin_unlock(&flctx->flc_lock);
1222         return ret;
1223 }
1224
1225 /*
1226  * Update and possibly write a cached page of an NFS file.
1227  *
1228  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1229  * things with a page scheduled for an RPC call (e.g. invalidate it).
1230  */
1231 int nfs_updatepage(struct file *file, struct page *page,
1232                 unsigned int offset, unsigned int count)
1233 {
1234         struct nfs_open_context *ctx = nfs_file_open_context(file);
1235         struct inode    *inode = page_file_mapping(page)->host;
1236         int             status = 0;
1237
1238         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1239
1240         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1241                 file, count, (long long)(page_file_offset(page) + offset));
1242
1243         if (nfs_can_extend_write(file, page, inode)) {
1244                 count = max(count + offset, nfs_page_length(page));
1245                 offset = 0;
1246         }
1247
1248         status = nfs_writepage_setup(ctx, page, offset, count);
1249         if (status < 0)
1250                 nfs_set_pageerror(page);
1251         else
1252                 __set_page_dirty_nobuffers(page);
1253
1254         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1255                         status, (long long)i_size_read(inode));
1256         return status;
1257 }
1258
1259 static int flush_task_priority(int how)
1260 {
1261         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1262                 case FLUSH_HIGHPRI:
1263                         return RPC_PRIORITY_HIGH;
1264                 case FLUSH_LOWPRI:
1265                         return RPC_PRIORITY_LOW;
1266         }
1267         return RPC_PRIORITY_NORMAL;
1268 }
1269
1270 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1271                                struct rpc_message *msg,
1272                                struct rpc_task_setup *task_setup_data, int how)
1273 {
1274         struct inode *inode = hdr->inode;
1275         int priority = flush_task_priority(how);
1276
1277         task_setup_data->priority = priority;
1278         NFS_PROTO(inode)->write_setup(hdr, msg);
1279
1280         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1281                                  &task_setup_data->rpc_client, msg, hdr);
1282 }
1283
1284 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1285  * call this on each, which will prepare them to be retried on next
1286  * writeback using standard nfs.
1287  */
1288 static void nfs_redirty_request(struct nfs_page *req)
1289 {
1290         nfs_mark_request_dirty(req);
1291         nfs_unlock_request(req);
1292         nfs_end_page_writeback(req);
1293         nfs_release_request(req);
1294 }
1295
1296 static void nfs_async_write_error(struct list_head *head)
1297 {
1298         struct nfs_page *req;
1299
1300         while (!list_empty(head)) {
1301                 req = nfs_list_entry(head->next);
1302                 nfs_list_remove_request(req);
1303                 nfs_redirty_request(req);
1304         }
1305 }
1306
1307 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1308         .error_cleanup = nfs_async_write_error,
1309         .completion = nfs_write_completion,
1310 };
1311
1312 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1313                                struct inode *inode, int ioflags, bool force_mds,
1314                                const struct nfs_pgio_completion_ops *compl_ops)
1315 {
1316         struct nfs_server *server = NFS_SERVER(inode);
1317         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1318
1319 #ifdef CONFIG_NFS_V4_1
1320         if (server->pnfs_curr_ld && !force_mds)
1321                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1322 #endif
1323         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1324                         server->wsize, ioflags);
1325 }
1326 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1327
1328 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1329 {
1330         pgio->pg_ops = &nfs_pgio_rw_ops;
1331         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1332 }
1333 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1334
1335
1336 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1337 {
1338         struct nfs_commit_data *data = calldata;
1339
1340         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1341 }
1342
1343 static void nfs_writeback_release_common(struct nfs_pgio_header *hdr)
1344 {
1345         /* do nothing! */
1346 }
1347
1348 /*
1349  * Special version of should_remove_suid() that ignores capabilities.
1350  */
1351 static int nfs_should_remove_suid(const struct inode *inode)
1352 {
1353         umode_t mode = inode->i_mode;
1354         int kill = 0;
1355
1356         /* suid always must be killed */
1357         if (unlikely(mode & S_ISUID))
1358                 kill = ATTR_KILL_SUID;
1359
1360         /*
1361          * sgid without any exec bits is just a mandatory locking mark; leave
1362          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1363          */
1364         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1365                 kill |= ATTR_KILL_SGID;
1366
1367         if (unlikely(kill && S_ISREG(mode)))
1368                 return kill;
1369
1370         return 0;
1371 }
1372
1373 /*
1374  * This function is called when the WRITE call is complete.
1375  */
1376 static int nfs_writeback_done(struct rpc_task *task,
1377                               struct nfs_pgio_header *hdr,
1378                               struct inode *inode)
1379 {
1380         int status;
1381
1382         /*
1383          * ->write_done will attempt to use post-op attributes to detect
1384          * conflicting writes by other clients.  A strict interpretation
1385          * of close-to-open would allow us to continue caching even if
1386          * another writer had changed the file, but some applications
1387          * depend on tighter cache coherency when writing.
1388          */
1389         status = NFS_PROTO(inode)->write_done(task, hdr);
1390         if (status != 0)
1391                 return status;
1392         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1393
1394         if (hdr->res.verf->committed < hdr->args.stable &&
1395             task->tk_status >= 0) {
1396                 /* We tried a write call, but the server did not
1397                  * commit data to stable storage even though we
1398                  * requested it.
1399                  * Note: There is a known bug in Tru64 < 5.0 in which
1400                  *       the server reports NFS_DATA_SYNC, but performs
1401                  *       NFS_FILE_SYNC. We therefore implement this checking
1402                  *       as a dprintk() in order to avoid filling syslog.
1403                  */
1404                 static unsigned long    complain;
1405
1406                 /* Note this will print the MDS for a DS write */
1407                 if (time_before(complain, jiffies)) {
1408                         dprintk("NFS:       faulty NFS server %s:"
1409                                 " (committed = %d) != (stable = %d)\n",
1410                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1411                                 hdr->res.verf->committed, hdr->args.stable);
1412                         complain = jiffies + 300 * HZ;
1413                 }
1414         }
1415
1416         /* Deal with the suid/sgid bit corner case */
1417         if (nfs_should_remove_suid(inode))
1418                 nfs_mark_for_revalidate(inode);
1419         return 0;
1420 }
1421
1422 /*
1423  * This function is called when the WRITE call is complete.
1424  */
1425 static void nfs_writeback_result(struct rpc_task *task,
1426                                  struct nfs_pgio_header *hdr)
1427 {
1428         struct nfs_pgio_args    *argp = &hdr->args;
1429         struct nfs_pgio_res     *resp = &hdr->res;
1430
1431         if (resp->count < argp->count) {
1432                 static unsigned long    complain;
1433
1434                 /* This a short write! */
1435                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1436
1437                 /* Has the server at least made some progress? */
1438                 if (resp->count == 0) {
1439                         if (time_before(complain, jiffies)) {
1440                                 printk(KERN_WARNING
1441                                        "NFS: Server wrote zero bytes, expected %u.\n",
1442                                        argp->count);
1443                                 complain = jiffies + 300 * HZ;
1444                         }
1445                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1446                         task->tk_status = -EIO;
1447                         return;
1448                 }
1449                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1450                 if (resp->verf->committed != NFS_UNSTABLE) {
1451                         /* Resend from where the server left off */
1452                         hdr->mds_offset += resp->count;
1453                         argp->offset += resp->count;
1454                         argp->pgbase += resp->count;
1455                         argp->count -= resp->count;
1456                 } else {
1457                         /* Resend as a stable write in order to avoid
1458                          * headaches in the case of a server crash.
1459                          */
1460                         argp->stable = NFS_FILE_SYNC;
1461                 }
1462                 rpc_restart_call_prepare(task);
1463         }
1464 }
1465
1466
1467 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1468 {
1469         int ret;
1470
1471         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1472                 return 1;
1473         if (!may_wait)
1474                 return 0;
1475         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1476                                 NFS_INO_COMMIT,
1477                                 nfs_wait_bit_killable,
1478                                 TASK_KILLABLE);
1479         return (ret < 0) ? ret : 1;
1480 }
1481
1482 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1483 {
1484         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1485         smp_mb__after_atomic();
1486         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1487 }
1488
1489 void nfs_commitdata_release(struct nfs_commit_data *data)
1490 {
1491         put_nfs_open_context(data->context);
1492         nfs_commit_free(data);
1493 }
1494 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1495
1496 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1497                         const struct rpc_call_ops *call_ops,
1498                         int how, int flags)
1499 {
1500         struct rpc_task *task;
1501         int priority = flush_task_priority(how);
1502         struct rpc_message msg = {
1503                 .rpc_argp = &data->args,
1504                 .rpc_resp = &data->res,
1505                 .rpc_cred = data->cred,
1506         };
1507         struct rpc_task_setup task_setup_data = {
1508                 .task = &data->task,
1509                 .rpc_client = clnt,
1510                 .rpc_message = &msg,
1511                 .callback_ops = call_ops,
1512                 .callback_data = data,
1513                 .workqueue = nfsiod_workqueue,
1514                 .flags = RPC_TASK_ASYNC | flags,
1515                 .priority = priority,
1516         };
1517         /* Set up the initial task struct.  */
1518         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1519
1520         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1521
1522         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1523                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1524
1525         task = rpc_run_task(&task_setup_data);
1526         if (IS_ERR(task))
1527                 return PTR_ERR(task);
1528         if (how & FLUSH_SYNC)
1529                 rpc_wait_for_completion_task(task);
1530         rpc_put_task(task);
1531         return 0;
1532 }
1533 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1534
1535 static loff_t nfs_get_lwb(struct list_head *head)
1536 {
1537         loff_t lwb = 0;
1538         struct nfs_page *req;
1539
1540         list_for_each_entry(req, head, wb_list)
1541                 if (lwb < (req_offset(req) + req->wb_bytes))
1542                         lwb = req_offset(req) + req->wb_bytes;
1543
1544         return lwb;
1545 }
1546
1547 /*
1548  * Set up the argument/result storage required for the RPC call.
1549  */
1550 void nfs_init_commit(struct nfs_commit_data *data,
1551                      struct list_head *head,
1552                      struct pnfs_layout_segment *lseg,
1553                      struct nfs_commit_info *cinfo)
1554 {
1555         struct nfs_page *first = nfs_list_entry(head->next);
1556         struct inode *inode = first->wb_context->dentry->d_inode;
1557
1558         /* Set up the RPC argument and reply structs
1559          * NB: take care not to mess about with data->commit et al. */
1560
1561         list_splice_init(head, &data->pages);
1562
1563         data->inode       = inode;
1564         data->cred        = first->wb_context->cred;
1565         data->lseg        = lseg; /* reference transferred */
1566         /* only set lwb for pnfs commit */
1567         if (lseg)
1568                 data->lwb = nfs_get_lwb(&data->pages);
1569         data->mds_ops     = &nfs_commit_ops;
1570         data->completion_ops = cinfo->completion_ops;
1571         data->dreq        = cinfo->dreq;
1572
1573         data->args.fh     = NFS_FH(data->inode);
1574         /* Note: we always request a commit of the entire inode */
1575         data->args.offset = 0;
1576         data->args.count  = 0;
1577         data->context     = get_nfs_open_context(first->wb_context);
1578         data->res.fattr   = &data->fattr;
1579         data->res.verf    = &data->verf;
1580         nfs_fattr_init(&data->fattr);
1581 }
1582 EXPORT_SYMBOL_GPL(nfs_init_commit);
1583
1584 void nfs_retry_commit(struct list_head *page_list,
1585                       struct pnfs_layout_segment *lseg,
1586                       struct nfs_commit_info *cinfo)
1587 {
1588         struct nfs_page *req;
1589
1590         while (!list_empty(page_list)) {
1591                 req = nfs_list_entry(page_list->next);
1592                 nfs_list_remove_request(req);
1593                 nfs_mark_request_commit(req, lseg, cinfo);
1594                 if (!cinfo->dreq) {
1595                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1596                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1597                                      BDI_RECLAIMABLE);
1598                 }
1599                 nfs_unlock_and_release_request(req);
1600         }
1601 }
1602 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1603
1604 /*
1605  * Commit dirty pages
1606  */
1607 static int
1608 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1609                 struct nfs_commit_info *cinfo)
1610 {
1611         struct nfs_commit_data  *data;
1612
1613         data = nfs_commitdata_alloc();
1614
1615         if (!data)
1616                 goto out_bad;
1617
1618         /* Set up the argument struct */
1619         nfs_init_commit(data, head, NULL, cinfo);
1620         atomic_inc(&cinfo->mds->rpcs_out);
1621         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1622                                    how, 0);
1623  out_bad:
1624         nfs_retry_commit(head, NULL, cinfo);
1625         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1626         return -ENOMEM;
1627 }
1628
1629 /*
1630  * COMMIT call returned
1631  */
1632 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1633 {
1634         struct nfs_commit_data  *data = calldata;
1635
1636         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1637                                 task->tk_pid, task->tk_status);
1638
1639         /* Call the NFS version-specific code */
1640         NFS_PROTO(data->inode)->commit_done(task, data);
1641 }
1642
1643 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1644 {
1645         struct nfs_page *req;
1646         int status = data->task.tk_status;
1647         struct nfs_commit_info cinfo;
1648         struct nfs_server *nfss;
1649
1650         while (!list_empty(&data->pages)) {
1651                 req = nfs_list_entry(data->pages.next);
1652                 nfs_list_remove_request(req);
1653                 nfs_clear_page_commit(req->wb_page);
1654
1655                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1656                         req->wb_context->dentry->d_sb->s_id,
1657                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1658                         req->wb_bytes,
1659                         (long long)req_offset(req));
1660                 if (status < 0) {
1661                         nfs_context_set_write_error(req->wb_context, status);
1662                         nfs_inode_remove_request(req);
1663                         dprintk(", error = %d\n", status);
1664                         goto next;
1665                 }
1666
1667                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1668                  * returned by the server against all stored verfs. */
1669                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1670                         /* We have a match */
1671                         nfs_inode_remove_request(req);
1672                         dprintk(" OK\n");
1673                         goto next;
1674                 }
1675                 /* We have a mismatch. Write the page again */
1676                 dprintk(" mismatch\n");
1677                 nfs_mark_request_dirty(req);
1678                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1679         next:
1680                 nfs_unlock_and_release_request(req);
1681         }
1682         nfss = NFS_SERVER(data->inode);
1683         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1684                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1685
1686         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1687         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1688                 nfs_commit_clear_lock(NFS_I(data->inode));
1689 }
1690
1691 static void nfs_commit_release(void *calldata)
1692 {
1693         struct nfs_commit_data *data = calldata;
1694
1695         data->completion_ops->completion(data);
1696         nfs_commitdata_release(calldata);
1697 }
1698
1699 static const struct rpc_call_ops nfs_commit_ops = {
1700         .rpc_call_prepare = nfs_commit_prepare,
1701         .rpc_call_done = nfs_commit_done,
1702         .rpc_release = nfs_commit_release,
1703 };
1704
1705 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1706         .completion = nfs_commit_release_pages,
1707         .error_cleanup = nfs_commit_clear_lock,
1708 };
1709
1710 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1711                             int how, struct nfs_commit_info *cinfo)
1712 {
1713         int status;
1714
1715         status = pnfs_commit_list(inode, head, how, cinfo);
1716         if (status == PNFS_NOT_ATTEMPTED)
1717                 status = nfs_commit_list(inode, head, how, cinfo);
1718         return status;
1719 }
1720
1721 int nfs_commit_inode(struct inode *inode, int how)
1722 {
1723         LIST_HEAD(head);
1724         struct nfs_commit_info cinfo;
1725         int may_wait = how & FLUSH_SYNC;
1726         int res;
1727
1728         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1729         if (res <= 0)
1730                 goto out_mark_dirty;
1731         nfs_init_cinfo_from_inode(&cinfo, inode);
1732         res = nfs_scan_commit(inode, &head, &cinfo);
1733         if (res) {
1734                 int error;
1735
1736                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1737                 if (error < 0)
1738                         return error;
1739                 if (!may_wait)
1740                         goto out_mark_dirty;
1741                 error = wait_on_bit_action(&NFS_I(inode)->flags,
1742                                 NFS_INO_COMMIT,
1743                                 nfs_wait_bit_killable,
1744                                 TASK_KILLABLE);
1745                 if (error < 0)
1746                         return error;
1747         } else
1748                 nfs_commit_clear_lock(NFS_I(inode));
1749         return res;
1750         /* Note: If we exit without ensuring that the commit is complete,
1751          * we must mark the inode as dirty. Otherwise, future calls to
1752          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1753          * that the data is on the disk.
1754          */
1755 out_mark_dirty:
1756         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1757         return res;
1758 }
1759
1760 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1761 {
1762         struct nfs_inode *nfsi = NFS_I(inode);
1763         int flags = FLUSH_SYNC;
1764         int ret = 0;
1765
1766         /* no commits means nothing needs to be done */
1767         if (!nfsi->commit_info.ncommit)
1768                 return ret;
1769
1770         if (wbc->sync_mode == WB_SYNC_NONE) {
1771                 /* Don't commit yet if this is a non-blocking flush and there
1772                  * are a lot of outstanding writes for this mapping.
1773                  */
1774                 if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1))
1775                         goto out_mark_dirty;
1776
1777                 /* don't wait for the COMMIT response */
1778                 flags = 0;
1779         }
1780
1781         ret = nfs_commit_inode(inode, flags);
1782         if (ret >= 0) {
1783                 if (wbc->sync_mode == WB_SYNC_NONE) {
1784                         if (ret < wbc->nr_to_write)
1785                                 wbc->nr_to_write -= ret;
1786                         else
1787                                 wbc->nr_to_write = 0;
1788                 }
1789                 return 0;
1790         }
1791 out_mark_dirty:
1792         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1793         return ret;
1794 }
1795
1796 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1797 {
1798         return nfs_commit_unstable_pages(inode, wbc);
1799 }
1800 EXPORT_SYMBOL_GPL(nfs_write_inode);
1801
1802 /*
1803  * flush the inode to disk.
1804  */
1805 int nfs_wb_all(struct inode *inode)
1806 {
1807         struct writeback_control wbc = {
1808                 .sync_mode = WB_SYNC_ALL,
1809                 .nr_to_write = LONG_MAX,
1810                 .range_start = 0,
1811                 .range_end = LLONG_MAX,
1812         };
1813         int ret;
1814
1815         trace_nfs_writeback_inode_enter(inode);
1816
1817         ret = sync_inode(inode, &wbc);
1818
1819         trace_nfs_writeback_inode_exit(inode, ret);
1820         return ret;
1821 }
1822 EXPORT_SYMBOL_GPL(nfs_wb_all);
1823
1824 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1825 {
1826         struct nfs_page *req;
1827         int ret = 0;
1828
1829         wait_on_page_writeback(page);
1830
1831         /* blocking call to cancel all requests and join to a single (head)
1832          * request */
1833         req = nfs_lock_and_join_requests(page, false);
1834
1835         if (IS_ERR(req)) {
1836                 ret = PTR_ERR(req);
1837         } else if (req) {
1838                 /* all requests from this page have been cancelled by
1839                  * nfs_lock_and_join_requests, so just remove the head
1840                  * request from the inode / page_private pointer and
1841                  * release it */
1842                 nfs_inode_remove_request(req);
1843                 /*
1844                  * In case nfs_inode_remove_request has marked the
1845                  * page as being dirty
1846                  */
1847                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1848                 nfs_unlock_and_release_request(req);
1849         }
1850
1851         return ret;
1852 }
1853
1854 /*
1855  * Write back all requests on one page - we do this before reading it.
1856  */
1857 int nfs_wb_page(struct inode *inode, struct page *page)
1858 {
1859         loff_t range_start = page_file_offset(page);
1860         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1861         struct writeback_control wbc = {
1862                 .sync_mode = WB_SYNC_ALL,
1863                 .nr_to_write = 0,
1864                 .range_start = range_start,
1865                 .range_end = range_end,
1866         };
1867         int ret;
1868
1869         trace_nfs_writeback_page_enter(inode);
1870
1871         for (;;) {
1872                 wait_on_page_writeback(page);
1873                 if (clear_page_dirty_for_io(page)) {
1874                         ret = nfs_writepage_locked(page, &wbc);
1875                         if (ret < 0)
1876                                 goto out_error;
1877                         continue;
1878                 }
1879                 ret = 0;
1880                 if (!PagePrivate(page))
1881                         break;
1882                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1883                 if (ret < 0)
1884                         goto out_error;
1885         }
1886 out_error:
1887         trace_nfs_writeback_page_exit(inode, ret);
1888         return ret;
1889 }
1890
1891 #ifdef CONFIG_MIGRATION
1892 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1893                 struct page *page, enum migrate_mode mode)
1894 {
1895         /*
1896          * If PagePrivate is set, then the page is currently associated with
1897          * an in-progress read or write request. Don't try to migrate it.
1898          *
1899          * FIXME: we could do this in principle, but we'll need a way to ensure
1900          *        that we can safely release the inode reference while holding
1901          *        the page lock.
1902          */
1903         if (PagePrivate(page))
1904                 return -EBUSY;
1905
1906         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1907                 return -EBUSY;
1908
1909         return migrate_page(mapping, newpage, page, mode);
1910 }
1911 #endif
1912
1913 int __init nfs_init_writepagecache(void)
1914 {
1915         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1916                                              sizeof(struct nfs_pgio_header),
1917                                              0, SLAB_HWCACHE_ALIGN,
1918                                              NULL);
1919         if (nfs_wdata_cachep == NULL)
1920                 return -ENOMEM;
1921
1922         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1923                                                      nfs_wdata_cachep);
1924         if (nfs_wdata_mempool == NULL)
1925                 goto out_destroy_write_cache;
1926
1927         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1928                                              sizeof(struct nfs_commit_data),
1929                                              0, SLAB_HWCACHE_ALIGN,
1930                                              NULL);
1931         if (nfs_cdata_cachep == NULL)
1932                 goto out_destroy_write_mempool;
1933
1934         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1935                                                       nfs_cdata_cachep);
1936         if (nfs_commit_mempool == NULL)
1937                 goto out_destroy_commit_cache;
1938
1939         /*
1940          * NFS congestion size, scale with available memory.
1941          *
1942          *  64MB:    8192k
1943          * 128MB:   11585k
1944          * 256MB:   16384k
1945          * 512MB:   23170k
1946          *   1GB:   32768k
1947          *   2GB:   46340k
1948          *   4GB:   65536k
1949          *   8GB:   92681k
1950          *  16GB:  131072k
1951          *
1952          * This allows larger machines to have larger/more transfers.
1953          * Limit the default to 256M
1954          */
1955         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1956         if (nfs_congestion_kb > 256*1024)
1957                 nfs_congestion_kb = 256*1024;
1958
1959         return 0;
1960
1961 out_destroy_commit_cache:
1962         kmem_cache_destroy(nfs_cdata_cachep);
1963 out_destroy_write_mempool:
1964         mempool_destroy(nfs_wdata_mempool);
1965 out_destroy_write_cache:
1966         kmem_cache_destroy(nfs_wdata_cachep);
1967         return -ENOMEM;
1968 }
1969
1970 void nfs_destroy_writepagecache(void)
1971 {
1972         mempool_destroy(nfs_commit_mempool);
1973         kmem_cache_destroy(nfs_cdata_cachep);
1974         mempool_destroy(nfs_wdata_mempool);
1975         kmem_cache_destroy(nfs_wdata_cachep);
1976 }
1977
1978 static const struct nfs_rw_ops nfs_rw_write_ops = {
1979         .rw_mode                = FMODE_WRITE,
1980         .rw_alloc_header        = nfs_writehdr_alloc,
1981         .rw_free_header         = nfs_writehdr_free,
1982         .rw_release             = nfs_writeback_release_common,
1983         .rw_done                = nfs_writeback_done,
1984         .rw_result              = nfs_writeback_result,
1985         .rw_initiate            = nfs_initiate_write,
1986 };