nfs: Fix cache_validity check in nfs_write_pageuptodate()
[platform/adaptation/renesas_rcar/renesas_kernel.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_write_common_ops;
46 static const struct rpc_call_ops nfs_commit_ops;
47 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
48 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
49
50 static struct kmem_cache *nfs_wdata_cachep;
51 static mempool_t *nfs_wdata_mempool;
52 static struct kmem_cache *nfs_cdata_cachep;
53 static mempool_t *nfs_commit_mempool;
54
55 struct nfs_commit_data *nfs_commitdata_alloc(void)
56 {
57         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
58
59         if (p) {
60                 memset(p, 0, sizeof(*p));
61                 INIT_LIST_HEAD(&p->pages);
62         }
63         return p;
64 }
65 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
66
67 void nfs_commit_free(struct nfs_commit_data *p)
68 {
69         mempool_free(p, nfs_commit_mempool);
70 }
71 EXPORT_SYMBOL_GPL(nfs_commit_free);
72
73 struct nfs_write_header *nfs_writehdr_alloc(void)
74 {
75         struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
76
77         if (p) {
78                 struct nfs_pgio_header *hdr = &p->header;
79
80                 memset(p, 0, sizeof(*p));
81                 INIT_LIST_HEAD(&hdr->pages);
82                 INIT_LIST_HEAD(&hdr->rpc_list);
83                 spin_lock_init(&hdr->lock);
84                 atomic_set(&hdr->refcnt, 0);
85                 hdr->verf = &p->verf;
86         }
87         return p;
88 }
89 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
90
91 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
92                                                   unsigned int pagecount)
93 {
94         struct nfs_write_data *data, *prealloc;
95
96         prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
97         if (prealloc->header == NULL)
98                 data = prealloc;
99         else
100                 data = kzalloc(sizeof(*data), GFP_KERNEL);
101         if (!data)
102                 goto out;
103
104         if (nfs_pgarray_set(&data->pages, pagecount)) {
105                 data->header = hdr;
106                 atomic_inc(&hdr->refcnt);
107         } else {
108                 if (data != prealloc)
109                         kfree(data);
110                 data = NULL;
111         }
112 out:
113         return data;
114 }
115
116 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
117 {
118         struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
119         mempool_free(whdr, nfs_wdata_mempool);
120 }
121 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
122
123 void nfs_writedata_release(struct nfs_write_data *wdata)
124 {
125         struct nfs_pgio_header *hdr = wdata->header;
126         struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
127
128         put_nfs_open_context(wdata->args.context);
129         if (wdata->pages.pagevec != wdata->pages.page_array)
130                 kfree(wdata->pages.pagevec);
131         if (wdata == &write_header->rpc_data) {
132                 wdata->header = NULL;
133                 wdata = NULL;
134         }
135         if (atomic_dec_and_test(&hdr->refcnt))
136                 hdr->completion_ops->completion(hdr);
137         /* Note: we only free the rpc_task after callbacks are done.
138          * See the comment in rpc_free_task() for why
139          */
140         kfree(wdata);
141 }
142 EXPORT_SYMBOL_GPL(nfs_writedata_release);
143
144 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
145 {
146         ctx->error = error;
147         smp_wmb();
148         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
149 }
150
151 static struct nfs_page *
152 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
153 {
154         struct nfs_page *req = NULL;
155
156         if (PagePrivate(page))
157                 req = (struct nfs_page *)page_private(page);
158         else if (unlikely(PageSwapCache(page))) {
159                 struct nfs_page *freq, *t;
160
161                 /* Linearly search the commit list for the correct req */
162                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
163                         if (freq->wb_page == page) {
164                                 req = freq;
165                                 break;
166                         }
167                 }
168         }
169
170         if (req)
171                 kref_get(&req->wb_kref);
172
173         return req;
174 }
175
176 static struct nfs_page *nfs_page_find_request(struct page *page)
177 {
178         struct inode *inode = page_file_mapping(page)->host;
179         struct nfs_page *req = NULL;
180
181         spin_lock(&inode->i_lock);
182         req = nfs_page_find_request_locked(NFS_I(inode), page);
183         spin_unlock(&inode->i_lock);
184         return req;
185 }
186
187 /* Adjust the file length if we're writing beyond the end */
188 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
189 {
190         struct inode *inode = page_file_mapping(page)->host;
191         loff_t end, i_size;
192         pgoff_t end_index;
193
194         spin_lock(&inode->i_lock);
195         i_size = i_size_read(inode);
196         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
197         if (i_size > 0 && page_file_index(page) < end_index)
198                 goto out;
199         end = page_file_offset(page) + ((loff_t)offset+count);
200         if (i_size >= end)
201                 goto out;
202         i_size_write(inode, end);
203         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
204 out:
205         spin_unlock(&inode->i_lock);
206 }
207
208 /* A writeback failed: mark the page as bad, and invalidate the page cache */
209 static void nfs_set_pageerror(struct page *page)
210 {
211         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
212 }
213
214 /* We can set the PG_uptodate flag if we see that a write request
215  * covers the full page.
216  */
217 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
218 {
219         if (PageUptodate(page))
220                 return;
221         if (base != 0)
222                 return;
223         if (count != nfs_page_length(page))
224                 return;
225         SetPageUptodate(page);
226 }
227
228 static int wb_priority(struct writeback_control *wbc)
229 {
230         if (wbc->for_reclaim)
231                 return FLUSH_HIGHPRI | FLUSH_STABLE;
232         if (wbc->for_kupdate || wbc->for_background)
233                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
234         return FLUSH_COND_STABLE;
235 }
236
237 /*
238  * NFS congestion control
239  */
240
241 int nfs_congestion_kb;
242
243 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
244 #define NFS_CONGESTION_OFF_THRESH       \
245         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
246
247 static void nfs_set_page_writeback(struct page *page)
248 {
249         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
250         int ret = test_set_page_writeback(page);
251
252         WARN_ON_ONCE(ret != 0);
253
254         if (atomic_long_inc_return(&nfss->writeback) >
255                         NFS_CONGESTION_ON_THRESH) {
256                 set_bdi_congested(&nfss->backing_dev_info,
257                                         BLK_RW_ASYNC);
258         }
259 }
260
261 static void nfs_end_page_writeback(struct page *page)
262 {
263         struct inode *inode = page_file_mapping(page)->host;
264         struct nfs_server *nfss = NFS_SERVER(inode);
265
266         end_page_writeback(page);
267         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
268                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
269 }
270
271 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
272 {
273         struct inode *inode = page_file_mapping(page)->host;
274         struct nfs_page *req;
275         int ret;
276
277         spin_lock(&inode->i_lock);
278         for (;;) {
279                 req = nfs_page_find_request_locked(NFS_I(inode), page);
280                 if (req == NULL)
281                         break;
282                 if (nfs_lock_request(req))
283                         break;
284                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
285                  *       then the call to nfs_lock_request() will always
286                  *       succeed provided that someone hasn't already marked the
287                  *       request as dirty (in which case we don't care).
288                  */
289                 spin_unlock(&inode->i_lock);
290                 if (!nonblock)
291                         ret = nfs_wait_on_request(req);
292                 else
293                         ret = -EAGAIN;
294                 nfs_release_request(req);
295                 if (ret != 0)
296                         return ERR_PTR(ret);
297                 spin_lock(&inode->i_lock);
298         }
299         spin_unlock(&inode->i_lock);
300         return req;
301 }
302
303 /*
304  * Find an associated nfs write request, and prepare to flush it out
305  * May return an error if the user signalled nfs_wait_on_request().
306  */
307 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
308                                 struct page *page, bool nonblock)
309 {
310         struct nfs_page *req;
311         int ret = 0;
312
313         req = nfs_find_and_lock_request(page, nonblock);
314         if (!req)
315                 goto out;
316         ret = PTR_ERR(req);
317         if (IS_ERR(req))
318                 goto out;
319
320         nfs_set_page_writeback(page);
321         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
322
323         ret = 0;
324         if (!nfs_pageio_add_request(pgio, req)) {
325                 nfs_redirty_request(req);
326                 ret = pgio->pg_error;
327         }
328 out:
329         return ret;
330 }
331
332 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
333 {
334         struct inode *inode = page_file_mapping(page)->host;
335         int ret;
336
337         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
338         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
339
340         nfs_pageio_cond_complete(pgio, page_file_index(page));
341         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
342         if (ret == -EAGAIN) {
343                 redirty_page_for_writepage(wbc, page);
344                 ret = 0;
345         }
346         return ret;
347 }
348
349 /*
350  * Write an mmapped page to the server.
351  */
352 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
353 {
354         struct nfs_pageio_descriptor pgio;
355         int err;
356
357         NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio,
358                                                           page->mapping->host,
359                                                           wb_priority(wbc),
360                                                           &nfs_async_write_completion_ops);
361         err = nfs_do_writepage(page, wbc, &pgio);
362         nfs_pageio_complete(&pgio);
363         if (err < 0)
364                 return err;
365         if (pgio.pg_error < 0)
366                 return pgio.pg_error;
367         return 0;
368 }
369
370 int nfs_writepage(struct page *page, struct writeback_control *wbc)
371 {
372         int ret;
373
374         ret = nfs_writepage_locked(page, wbc);
375         unlock_page(page);
376         return ret;
377 }
378
379 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
380 {
381         int ret;
382
383         ret = nfs_do_writepage(page, wbc, data);
384         unlock_page(page);
385         return ret;
386 }
387
388 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
389 {
390         struct inode *inode = mapping->host;
391         unsigned long *bitlock = &NFS_I(inode)->flags;
392         struct nfs_pageio_descriptor pgio;
393         int err;
394
395         /* Stop dirtying of new pages while we sync */
396         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
397                         nfs_wait_bit_killable, TASK_KILLABLE);
398         if (err)
399                 goto out_err;
400
401         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
402
403         NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
404         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
405         nfs_pageio_complete(&pgio);
406
407         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
408         smp_mb__after_clear_bit();
409         wake_up_bit(bitlock, NFS_INO_FLUSHING);
410
411         if (err < 0)
412                 goto out_err;
413         err = pgio.pg_error;
414         if (err < 0)
415                 goto out_err;
416         return 0;
417 out_err:
418         return err;
419 }
420
421 /*
422  * Insert a write request into an inode
423  */
424 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
425 {
426         struct nfs_inode *nfsi = NFS_I(inode);
427
428         /* Lock the request! */
429         nfs_lock_request(req);
430
431         spin_lock(&inode->i_lock);
432         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
433                 inode->i_version++;
434         /*
435          * Swap-space should not get truncated. Hence no need to plug the race
436          * with invalidate/truncate.
437          */
438         if (likely(!PageSwapCache(req->wb_page))) {
439                 set_bit(PG_MAPPED, &req->wb_flags);
440                 SetPagePrivate(req->wb_page);
441                 set_page_private(req->wb_page, (unsigned long)req);
442         }
443         nfsi->npages++;
444         kref_get(&req->wb_kref);
445         spin_unlock(&inode->i_lock);
446 }
447
448 /*
449  * Remove a write request from an inode
450  */
451 static void nfs_inode_remove_request(struct nfs_page *req)
452 {
453         struct inode *inode = req->wb_context->dentry->d_inode;
454         struct nfs_inode *nfsi = NFS_I(inode);
455
456         spin_lock(&inode->i_lock);
457         if (likely(!PageSwapCache(req->wb_page))) {
458                 set_page_private(req->wb_page, 0);
459                 ClearPagePrivate(req->wb_page);
460                 clear_bit(PG_MAPPED, &req->wb_flags);
461         }
462         nfsi->npages--;
463         spin_unlock(&inode->i_lock);
464         nfs_release_request(req);
465 }
466
467 static void
468 nfs_mark_request_dirty(struct nfs_page *req)
469 {
470         __set_page_dirty_nobuffers(req->wb_page);
471 }
472
473 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
474 /**
475  * nfs_request_add_commit_list - add request to a commit list
476  * @req: pointer to a struct nfs_page
477  * @dst: commit list head
478  * @cinfo: holds list lock and accounting info
479  *
480  * This sets the PG_CLEAN bit, updates the cinfo count of
481  * number of outstanding requests requiring a commit as well as
482  * the MM page stats.
483  *
484  * The caller must _not_ hold the cinfo->lock, but must be
485  * holding the nfs_page lock.
486  */
487 void
488 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
489                             struct nfs_commit_info *cinfo)
490 {
491         set_bit(PG_CLEAN, &(req)->wb_flags);
492         spin_lock(cinfo->lock);
493         nfs_list_add_request(req, dst);
494         cinfo->mds->ncommit++;
495         spin_unlock(cinfo->lock);
496         if (!cinfo->dreq) {
497                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
498                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
499                              BDI_RECLAIMABLE);
500                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
501                                    I_DIRTY_DATASYNC);
502         }
503 }
504 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
505
506 /**
507  * nfs_request_remove_commit_list - Remove request from a commit list
508  * @req: pointer to a nfs_page
509  * @cinfo: holds list lock and accounting info
510  *
511  * This clears the PG_CLEAN bit, and updates the cinfo's count of
512  * number of outstanding requests requiring a commit
513  * It does not update the MM page stats.
514  *
515  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
516  */
517 void
518 nfs_request_remove_commit_list(struct nfs_page *req,
519                                struct nfs_commit_info *cinfo)
520 {
521         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
522                 return;
523         nfs_list_remove_request(req);
524         cinfo->mds->ncommit--;
525 }
526 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
527
528 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
529                                       struct inode *inode)
530 {
531         cinfo->lock = &inode->i_lock;
532         cinfo->mds = &NFS_I(inode)->commit_info;
533         cinfo->ds = pnfs_get_ds_info(inode);
534         cinfo->dreq = NULL;
535         cinfo->completion_ops = &nfs_commit_completion_ops;
536 }
537
538 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
539                     struct inode *inode,
540                     struct nfs_direct_req *dreq)
541 {
542         if (dreq)
543                 nfs_init_cinfo_from_dreq(cinfo, dreq);
544         else
545                 nfs_init_cinfo_from_inode(cinfo, inode);
546 }
547 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
548
549 /*
550  * Add a request to the inode's commit list.
551  */
552 void
553 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
554                         struct nfs_commit_info *cinfo)
555 {
556         if (pnfs_mark_request_commit(req, lseg, cinfo))
557                 return;
558         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
559 }
560
561 static void
562 nfs_clear_page_commit(struct page *page)
563 {
564         dec_zone_page_state(page, NR_UNSTABLE_NFS);
565         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
566 }
567
568 static void
569 nfs_clear_request_commit(struct nfs_page *req)
570 {
571         if (test_bit(PG_CLEAN, &req->wb_flags)) {
572                 struct inode *inode = req->wb_context->dentry->d_inode;
573                 struct nfs_commit_info cinfo;
574
575                 nfs_init_cinfo_from_inode(&cinfo, inode);
576                 if (!pnfs_clear_request_commit(req, &cinfo)) {
577                         spin_lock(cinfo.lock);
578                         nfs_request_remove_commit_list(req, &cinfo);
579                         spin_unlock(cinfo.lock);
580                 }
581                 nfs_clear_page_commit(req->wb_page);
582         }
583 }
584
585 static inline
586 int nfs_write_need_commit(struct nfs_write_data *data)
587 {
588         if (data->verf.committed == NFS_DATA_SYNC)
589                 return data->header->lseg == NULL;
590         return data->verf.committed != NFS_FILE_SYNC;
591 }
592
593 #else
594 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
595                                       struct inode *inode)
596 {
597 }
598
599 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
600                     struct inode *inode,
601                     struct nfs_direct_req *dreq)
602 {
603 }
604
605 void
606 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
607                         struct nfs_commit_info *cinfo)
608 {
609 }
610
611 static void
612 nfs_clear_request_commit(struct nfs_page *req)
613 {
614 }
615
616 static inline
617 int nfs_write_need_commit(struct nfs_write_data *data)
618 {
619         return 0;
620 }
621
622 #endif
623
624 static void nfs_write_completion(struct nfs_pgio_header *hdr)
625 {
626         struct nfs_commit_info cinfo;
627         unsigned long bytes = 0;
628
629         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
630                 goto out;
631         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
632         while (!list_empty(&hdr->pages)) {
633                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
634
635                 bytes += req->wb_bytes;
636                 nfs_list_remove_request(req);
637                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
638                     (hdr->good_bytes < bytes)) {
639                         nfs_set_pageerror(req->wb_page);
640                         nfs_context_set_write_error(req->wb_context, hdr->error);
641                         goto remove_req;
642                 }
643                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
644                         nfs_mark_request_dirty(req);
645                         goto next;
646                 }
647                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
648                         memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
649                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
650                         goto next;
651                 }
652 remove_req:
653                 nfs_inode_remove_request(req);
654 next:
655                 nfs_unlock_request(req);
656                 nfs_end_page_writeback(req->wb_page);
657                 nfs_release_request(req);
658         }
659 out:
660         hdr->release(hdr);
661 }
662
663 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
664 static unsigned long
665 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
666 {
667         return cinfo->mds->ncommit;
668 }
669
670 /* cinfo->lock held by caller */
671 int
672 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
673                      struct nfs_commit_info *cinfo, int max)
674 {
675         struct nfs_page *req, *tmp;
676         int ret = 0;
677
678         list_for_each_entry_safe(req, tmp, src, wb_list) {
679                 if (!nfs_lock_request(req))
680                         continue;
681                 kref_get(&req->wb_kref);
682                 if (cond_resched_lock(cinfo->lock))
683                         list_safe_reset_next(req, tmp, wb_list);
684                 nfs_request_remove_commit_list(req, cinfo);
685                 nfs_list_add_request(req, dst);
686                 ret++;
687                 if ((ret == max) && !cinfo->dreq)
688                         break;
689         }
690         return ret;
691 }
692
693 /*
694  * nfs_scan_commit - Scan an inode for commit requests
695  * @inode: NFS inode to scan
696  * @dst: mds destination list
697  * @cinfo: mds and ds lists of reqs ready to commit
698  *
699  * Moves requests from the inode's 'commit' request list.
700  * The requests are *not* checked to ensure that they form a contiguous set.
701  */
702 int
703 nfs_scan_commit(struct inode *inode, struct list_head *dst,
704                 struct nfs_commit_info *cinfo)
705 {
706         int ret = 0;
707
708         spin_lock(cinfo->lock);
709         if (cinfo->mds->ncommit > 0) {
710                 const int max = INT_MAX;
711
712                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
713                                            cinfo, max);
714                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
715         }
716         spin_unlock(cinfo->lock);
717         return ret;
718 }
719
720 #else
721 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
722 {
723         return 0;
724 }
725
726 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
727                     struct nfs_commit_info *cinfo)
728 {
729         return 0;
730 }
731 #endif
732
733 /*
734  * Search for an existing write request, and attempt to update
735  * it to reflect a new dirty region on a given page.
736  *
737  * If the attempt fails, then the existing request is flushed out
738  * to disk.
739  */
740 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
741                 struct page *page,
742                 unsigned int offset,
743                 unsigned int bytes)
744 {
745         struct nfs_page *req;
746         unsigned int rqend;
747         unsigned int end;
748         int error;
749
750         if (!PagePrivate(page))
751                 return NULL;
752
753         end = offset + bytes;
754         spin_lock(&inode->i_lock);
755
756         for (;;) {
757                 req = nfs_page_find_request_locked(NFS_I(inode), page);
758                 if (req == NULL)
759                         goto out_unlock;
760
761                 rqend = req->wb_offset + req->wb_bytes;
762                 /*
763                  * Tell the caller to flush out the request if
764                  * the offsets are non-contiguous.
765                  * Note: nfs_flush_incompatible() will already
766                  * have flushed out requests having wrong owners.
767                  */
768                 if (offset > rqend
769                     || end < req->wb_offset)
770                         goto out_flushme;
771
772                 if (nfs_lock_request(req))
773                         break;
774
775                 /* The request is locked, so wait and then retry */
776                 spin_unlock(&inode->i_lock);
777                 error = nfs_wait_on_request(req);
778                 nfs_release_request(req);
779                 if (error != 0)
780                         goto out_err;
781                 spin_lock(&inode->i_lock);
782         }
783
784         /* Okay, the request matches. Update the region */
785         if (offset < req->wb_offset) {
786                 req->wb_offset = offset;
787                 req->wb_pgbase = offset;
788         }
789         if (end > rqend)
790                 req->wb_bytes = end - req->wb_offset;
791         else
792                 req->wb_bytes = rqend - req->wb_offset;
793 out_unlock:
794         spin_unlock(&inode->i_lock);
795         if (req)
796                 nfs_clear_request_commit(req);
797         return req;
798 out_flushme:
799         spin_unlock(&inode->i_lock);
800         nfs_release_request(req);
801         error = nfs_wb_page(inode, page);
802 out_err:
803         return ERR_PTR(error);
804 }
805
806 /*
807  * Try to update an existing write request, or create one if there is none.
808  *
809  * Note: Should always be called with the Page Lock held to prevent races
810  * if we have to add a new request. Also assumes that the caller has
811  * already called nfs_flush_incompatible() if necessary.
812  */
813 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
814                 struct page *page, unsigned int offset, unsigned int bytes)
815 {
816         struct inode *inode = page_file_mapping(page)->host;
817         struct nfs_page *req;
818
819         req = nfs_try_to_update_request(inode, page, offset, bytes);
820         if (req != NULL)
821                 goto out;
822         req = nfs_create_request(ctx, inode, page, offset, bytes);
823         if (IS_ERR(req))
824                 goto out;
825         nfs_inode_add_request(inode, req);
826 out:
827         return req;
828 }
829
830 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
831                 unsigned int offset, unsigned int count)
832 {
833         struct nfs_page *req;
834
835         req = nfs_setup_write_request(ctx, page, offset, count);
836         if (IS_ERR(req))
837                 return PTR_ERR(req);
838         /* Update file length */
839         nfs_grow_file(page, offset, count);
840         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
841         nfs_mark_request_dirty(req);
842         nfs_unlock_and_release_request(req);
843         return 0;
844 }
845
846 int nfs_flush_incompatible(struct file *file, struct page *page)
847 {
848         struct nfs_open_context *ctx = nfs_file_open_context(file);
849         struct nfs_lock_context *l_ctx;
850         struct nfs_page *req;
851         int do_flush, status;
852         /*
853          * Look for a request corresponding to this page. If there
854          * is one, and it belongs to another file, we flush it out
855          * before we try to copy anything into the page. Do this
856          * due to the lack of an ACCESS-type call in NFSv2.
857          * Also do the same if we find a request from an existing
858          * dropped page.
859          */
860         do {
861                 req = nfs_page_find_request(page);
862                 if (req == NULL)
863                         return 0;
864                 l_ctx = req->wb_lock_context;
865                 do_flush = req->wb_page != page || req->wb_context != ctx;
866                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
867                         do_flush |= l_ctx->lockowner.l_owner != current->files
868                                 || l_ctx->lockowner.l_pid != current->tgid;
869                 }
870                 nfs_release_request(req);
871                 if (!do_flush)
872                         return 0;
873                 status = nfs_wb_page(page_file_mapping(page)->host, page);
874         } while (status == 0);
875         return status;
876 }
877
878 /*
879  * Avoid buffered writes when a open context credential's key would
880  * expire soon.
881  *
882  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
883  *
884  * Return 0 and set a credential flag which triggers the inode to flush
885  * and performs  NFS_FILE_SYNC writes if the key will expired within
886  * RPC_KEY_EXPIRE_TIMEO.
887  */
888 int
889 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
890 {
891         struct nfs_open_context *ctx = nfs_file_open_context(filp);
892         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
893
894         return rpcauth_key_timeout_notify(auth, ctx->cred);
895 }
896
897 /*
898  * Test if the open context credential key is marked to expire soon.
899  */
900 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
901 {
902         return rpcauth_cred_key_to_expire(ctx->cred);
903 }
904
905 /*
906  * If the page cache is marked as unsafe or invalid, then we can't rely on
907  * the PageUptodate() flag. In this case, we will need to turn off
908  * write optimisations that depend on the page contents being correct.
909  */
910 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
911 {
912         struct nfs_inode *nfsi = NFS_I(inode);
913
914         if (nfs_have_delegated_attributes(inode))
915                 goto out;
916         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
917                 return false;
918         smp_rmb();
919         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
920                 return false;
921 out:
922         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
923                 return false;
924         return PageUptodate(page) != 0;
925 }
926
927 /* If we know the page is up to date, and we're not using byte range locks (or
928  * if we have the whole file locked for writing), it may be more efficient to
929  * extend the write to cover the entire page in order to avoid fragmentation
930  * inefficiencies.
931  *
932  * If the file is opened for synchronous writes then we can just skip the rest
933  * of the checks.
934  */
935 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
936 {
937         if (file->f_flags & O_DSYNC)
938                 return 0;
939         if (!nfs_write_pageuptodate(page, inode))
940                 return 0;
941         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
942                 return 1;
943         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
944                         inode->i_flock->fl_end == OFFSET_MAX &&
945                         inode->i_flock->fl_type != F_RDLCK))
946                 return 1;
947         return 0;
948 }
949
950 /*
951  * Update and possibly write a cached page of an NFS file.
952  *
953  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
954  * things with a page scheduled for an RPC call (e.g. invalidate it).
955  */
956 int nfs_updatepage(struct file *file, struct page *page,
957                 unsigned int offset, unsigned int count)
958 {
959         struct nfs_open_context *ctx = nfs_file_open_context(file);
960         struct inode    *inode = page_file_mapping(page)->host;
961         int             status = 0;
962
963         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
964
965         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
966                 file, count, (long long)(page_file_offset(page) + offset));
967
968         if (nfs_can_extend_write(file, page, inode)) {
969                 count = max(count + offset, nfs_page_length(page));
970                 offset = 0;
971         }
972
973         status = nfs_writepage_setup(ctx, page, offset, count);
974         if (status < 0)
975                 nfs_set_pageerror(page);
976         else
977                 __set_page_dirty_nobuffers(page);
978
979         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
980                         status, (long long)i_size_read(inode));
981         return status;
982 }
983
984 static int flush_task_priority(int how)
985 {
986         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
987                 case FLUSH_HIGHPRI:
988                         return RPC_PRIORITY_HIGH;
989                 case FLUSH_LOWPRI:
990                         return RPC_PRIORITY_LOW;
991         }
992         return RPC_PRIORITY_NORMAL;
993 }
994
995 int nfs_initiate_write(struct rpc_clnt *clnt,
996                        struct nfs_write_data *data,
997                        const struct rpc_call_ops *call_ops,
998                        int how, int flags)
999 {
1000         struct inode *inode = data->header->inode;
1001         int priority = flush_task_priority(how);
1002         struct rpc_task *task;
1003         struct rpc_message msg = {
1004                 .rpc_argp = &data->args,
1005                 .rpc_resp = &data->res,
1006                 .rpc_cred = data->header->cred,
1007         };
1008         struct rpc_task_setup task_setup_data = {
1009                 .rpc_client = clnt,
1010                 .task = &data->task,
1011                 .rpc_message = &msg,
1012                 .callback_ops = call_ops,
1013                 .callback_data = data,
1014                 .workqueue = nfsiod_workqueue,
1015                 .flags = RPC_TASK_ASYNC | flags,
1016                 .priority = priority,
1017         };
1018         int ret = 0;
1019
1020         /* Set up the initial task struct.  */
1021         NFS_PROTO(inode)->write_setup(data, &msg);
1022
1023         dprintk("NFS: %5u initiated write call "
1024                 "(req %s/%llu, %u bytes @ offset %llu)\n",
1025                 data->task.tk_pid,
1026                 inode->i_sb->s_id,
1027                 (unsigned long long)NFS_FILEID(inode),
1028                 data->args.count,
1029                 (unsigned long long)data->args.offset);
1030
1031         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1032                                  &task_setup_data.rpc_client, &msg, data);
1033
1034         task = rpc_run_task(&task_setup_data);
1035         if (IS_ERR(task)) {
1036                 ret = PTR_ERR(task);
1037                 goto out;
1038         }
1039         if (how & FLUSH_SYNC) {
1040                 ret = rpc_wait_for_completion_task(task);
1041                 if (ret == 0)
1042                         ret = task->tk_status;
1043         }
1044         rpc_put_task(task);
1045 out:
1046         return ret;
1047 }
1048 EXPORT_SYMBOL_GPL(nfs_initiate_write);
1049
1050 /*
1051  * Set up the argument/result storage required for the RPC call.
1052  */
1053 static void nfs_write_rpcsetup(struct nfs_write_data *data,
1054                 unsigned int count, unsigned int offset,
1055                 int how, struct nfs_commit_info *cinfo)
1056 {
1057         struct nfs_page *req = data->header->req;
1058
1059         /* Set up the RPC argument and reply structs
1060          * NB: take care not to mess about with data->commit et al. */
1061
1062         data->args.fh     = NFS_FH(data->header->inode);
1063         data->args.offset = req_offset(req) + offset;
1064         /* pnfs_set_layoutcommit needs this */
1065         data->mds_offset = data->args.offset;
1066         data->args.pgbase = req->wb_pgbase + offset;
1067         data->args.pages  = data->pages.pagevec;
1068         data->args.count  = count;
1069         data->args.context = get_nfs_open_context(req->wb_context);
1070         data->args.lock_context = req->wb_lock_context;
1071         data->args.stable  = NFS_UNSTABLE;
1072         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1073         case 0:
1074                 break;
1075         case FLUSH_COND_STABLE:
1076                 if (nfs_reqs_to_commit(cinfo))
1077                         break;
1078         default:
1079                 data->args.stable = NFS_FILE_SYNC;
1080         }
1081
1082         data->res.fattr   = &data->fattr;
1083         data->res.count   = count;
1084         data->res.verf    = &data->verf;
1085         nfs_fattr_init(&data->fattr);
1086 }
1087
1088 static int nfs_do_write(struct nfs_write_data *data,
1089                 const struct rpc_call_ops *call_ops,
1090                 int how)
1091 {
1092         struct inode *inode = data->header->inode;
1093
1094         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1095 }
1096
1097 static int nfs_do_multiple_writes(struct list_head *head,
1098                 const struct rpc_call_ops *call_ops,
1099                 int how)
1100 {
1101         struct nfs_write_data *data;
1102         int ret = 0;
1103
1104         while (!list_empty(head)) {
1105                 int ret2;
1106
1107                 data = list_first_entry(head, struct nfs_write_data, list);
1108                 list_del_init(&data->list);
1109                 
1110                 ret2 = nfs_do_write(data, call_ops, how);
1111                  if (ret == 0)
1112                          ret = ret2;
1113         }
1114         return ret;
1115 }
1116
1117 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1118  * call this on each, which will prepare them to be retried on next
1119  * writeback using standard nfs.
1120  */
1121 static void nfs_redirty_request(struct nfs_page *req)
1122 {
1123         nfs_mark_request_dirty(req);
1124         nfs_unlock_request(req);
1125         nfs_end_page_writeback(req->wb_page);
1126         nfs_release_request(req);
1127 }
1128
1129 static void nfs_async_write_error(struct list_head *head)
1130 {
1131         struct nfs_page *req;
1132
1133         while (!list_empty(head)) {
1134                 req = nfs_list_entry(head->next);
1135                 nfs_list_remove_request(req);
1136                 nfs_redirty_request(req);
1137         }
1138 }
1139
1140 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1141         .error_cleanup = nfs_async_write_error,
1142         .completion = nfs_write_completion,
1143 };
1144
1145 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1146                 struct nfs_pgio_header *hdr)
1147 {
1148         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1149         while (!list_empty(&hdr->rpc_list)) {
1150                 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1151                                 struct nfs_write_data, list);
1152                 list_del(&data->list);
1153                 nfs_writedata_release(data);
1154         }
1155         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1156 }
1157
1158 /*
1159  * Generate multiple small requests to write out a single
1160  * contiguous dirty area on one page.
1161  */
1162 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1163                            struct nfs_pgio_header *hdr)
1164 {
1165         struct nfs_page *req = hdr->req;
1166         struct page *page = req->wb_page;
1167         struct nfs_write_data *data;
1168         size_t wsize = desc->pg_bsize, nbytes;
1169         unsigned int offset;
1170         int requests = 0;
1171         struct nfs_commit_info cinfo;
1172
1173         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1174
1175         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1176             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1177              desc->pg_count > wsize))
1178                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1179
1180
1181         offset = 0;
1182         nbytes = desc->pg_count;
1183         do {
1184                 size_t len = min(nbytes, wsize);
1185
1186                 data = nfs_writedata_alloc(hdr, 1);
1187                 if (!data) {
1188                         nfs_flush_error(desc, hdr);
1189                         return -ENOMEM;
1190                 }
1191                 data->pages.pagevec[0] = page;
1192                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1193                 list_add(&data->list, &hdr->rpc_list);
1194                 requests++;
1195                 nbytes -= len;
1196                 offset += len;
1197         } while (nbytes != 0);
1198         nfs_list_remove_request(req);
1199         nfs_list_add_request(req, &hdr->pages);
1200         desc->pg_rpc_callops = &nfs_write_common_ops;
1201         return 0;
1202 }
1203
1204 /*
1205  * Create an RPC task for the given write request and kick it.
1206  * The page must have been locked by the caller.
1207  *
1208  * It may happen that the page we're passed is not marked dirty.
1209  * This is the case if nfs_updatepage detects a conflicting request
1210  * that has been written but not committed.
1211  */
1212 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1213                          struct nfs_pgio_header *hdr)
1214 {
1215         struct nfs_page         *req;
1216         struct page             **pages;
1217         struct nfs_write_data   *data;
1218         struct list_head *head = &desc->pg_list;
1219         struct nfs_commit_info cinfo;
1220
1221         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1222                                                            desc->pg_count));
1223         if (!data) {
1224                 nfs_flush_error(desc, hdr);
1225                 return -ENOMEM;
1226         }
1227
1228         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1229         pages = data->pages.pagevec;
1230         while (!list_empty(head)) {
1231                 req = nfs_list_entry(head->next);
1232                 nfs_list_remove_request(req);
1233                 nfs_list_add_request(req, &hdr->pages);
1234                 *pages++ = req->wb_page;
1235         }
1236
1237         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1238             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1239                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1240
1241         /* Set up the argument struct */
1242         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1243         list_add(&data->list, &hdr->rpc_list);
1244         desc->pg_rpc_callops = &nfs_write_common_ops;
1245         return 0;
1246 }
1247
1248 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1249                       struct nfs_pgio_header *hdr)
1250 {
1251         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1252                 return nfs_flush_multi(desc, hdr);
1253         return nfs_flush_one(desc, hdr);
1254 }
1255 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1256
1257 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1258 {
1259         struct nfs_write_header *whdr;
1260         struct nfs_pgio_header *hdr;
1261         int ret;
1262
1263         whdr = nfs_writehdr_alloc();
1264         if (!whdr) {
1265                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1266                 return -ENOMEM;
1267         }
1268         hdr = &whdr->header;
1269         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1270         atomic_inc(&hdr->refcnt);
1271         ret = nfs_generic_flush(desc, hdr);
1272         if (ret == 0)
1273                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1274                                              desc->pg_rpc_callops,
1275                                              desc->pg_ioflags);
1276         if (atomic_dec_and_test(&hdr->refcnt))
1277                 hdr->completion_ops->completion(hdr);
1278         return ret;
1279 }
1280
1281 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1282         .pg_test = nfs_generic_pg_test,
1283         .pg_doio = nfs_generic_pg_writepages,
1284 };
1285
1286 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1287                                struct inode *inode, int ioflags,
1288                                const struct nfs_pgio_completion_ops *compl_ops)
1289 {
1290         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1291                                 NFS_SERVER(inode)->wsize, ioflags);
1292 }
1293 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1294
1295 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1296 {
1297         pgio->pg_ops = &nfs_pageio_write_ops;
1298         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1299 }
1300 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1301
1302
1303 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1304 {
1305         struct nfs_write_data *data = calldata;
1306         int err;
1307         err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1308         if (err)
1309                 rpc_exit(task, err);
1310 }
1311
1312 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1313 {
1314         struct nfs_commit_data *data = calldata;
1315
1316         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1317 }
1318
1319 /*
1320  * Handle a write reply that flushes a whole page.
1321  *
1322  * FIXME: There is an inherent race with invalidate_inode_pages and
1323  *        writebacks since the page->count is kept > 1 for as long
1324  *        as the page has a write request pending.
1325  */
1326 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1327 {
1328         struct nfs_write_data   *data = calldata;
1329
1330         nfs_writeback_done(task, data);
1331 }
1332
1333 static void nfs_writeback_release_common(void *calldata)
1334 {
1335         struct nfs_write_data   *data = calldata;
1336         struct nfs_pgio_header *hdr = data->header;
1337         int status = data->task.tk_status;
1338
1339         if ((status >= 0) && nfs_write_need_commit(data)) {
1340                 spin_lock(&hdr->lock);
1341                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1342                         ; /* Do nothing */
1343                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1344                         memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1345                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1346                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1347                 spin_unlock(&hdr->lock);
1348         }
1349         nfs_writedata_release(data);
1350 }
1351
1352 static const struct rpc_call_ops nfs_write_common_ops = {
1353         .rpc_call_prepare = nfs_write_prepare,
1354         .rpc_call_done = nfs_writeback_done_common,
1355         .rpc_release = nfs_writeback_release_common,
1356 };
1357
1358
1359 /*
1360  * This function is called when the WRITE call is complete.
1361  */
1362 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1363 {
1364         struct nfs_writeargs    *argp = &data->args;
1365         struct nfs_writeres     *resp = &data->res;
1366         struct inode            *inode = data->header->inode;
1367         int status;
1368
1369         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1370                 task->tk_pid, task->tk_status);
1371
1372         /*
1373          * ->write_done will attempt to use post-op attributes to detect
1374          * conflicting writes by other clients.  A strict interpretation
1375          * of close-to-open would allow us to continue caching even if
1376          * another writer had changed the file, but some applications
1377          * depend on tighter cache coherency when writing.
1378          */
1379         status = NFS_PROTO(inode)->write_done(task, data);
1380         if (status != 0)
1381                 return;
1382         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1383
1384 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1385         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1386                 /* We tried a write call, but the server did not
1387                  * commit data to stable storage even though we
1388                  * requested it.
1389                  * Note: There is a known bug in Tru64 < 5.0 in which
1390                  *       the server reports NFS_DATA_SYNC, but performs
1391                  *       NFS_FILE_SYNC. We therefore implement this checking
1392                  *       as a dprintk() in order to avoid filling syslog.
1393                  */
1394                 static unsigned long    complain;
1395
1396                 /* Note this will print the MDS for a DS write */
1397                 if (time_before(complain, jiffies)) {
1398                         dprintk("NFS:       faulty NFS server %s:"
1399                                 " (committed = %d) != (stable = %d)\n",
1400                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1401                                 resp->verf->committed, argp->stable);
1402                         complain = jiffies + 300 * HZ;
1403                 }
1404         }
1405 #endif
1406         if (task->tk_status < 0)
1407                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1408         else if (resp->count < argp->count) {
1409                 static unsigned long    complain;
1410
1411                 /* This a short write! */
1412                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1413
1414                 /* Has the server at least made some progress? */
1415                 if (resp->count == 0) {
1416                         if (time_before(complain, jiffies)) {
1417                                 printk(KERN_WARNING
1418                                        "NFS: Server wrote zero bytes, expected %u.\n",
1419                                        argp->count);
1420                                 complain = jiffies + 300 * HZ;
1421                         }
1422                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1423                         task->tk_status = -EIO;
1424                         return;
1425                 }
1426                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1427                 if (resp->verf->committed != NFS_UNSTABLE) {
1428                         /* Resend from where the server left off */
1429                         data->mds_offset += resp->count;
1430                         argp->offset += resp->count;
1431                         argp->pgbase += resp->count;
1432                         argp->count -= resp->count;
1433                 } else {
1434                         /* Resend as a stable write in order to avoid
1435                          * headaches in the case of a server crash.
1436                          */
1437                         argp->stable = NFS_FILE_SYNC;
1438                 }
1439                 rpc_restart_call_prepare(task);
1440         }
1441 }
1442
1443
1444 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1445 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1446 {
1447         int ret;
1448
1449         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1450                 return 1;
1451         if (!may_wait)
1452                 return 0;
1453         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1454                                 NFS_INO_COMMIT,
1455                                 nfs_wait_bit_killable,
1456                                 TASK_KILLABLE);
1457         return (ret < 0) ? ret : 1;
1458 }
1459
1460 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1461 {
1462         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1463         smp_mb__after_clear_bit();
1464         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1465 }
1466
1467 void nfs_commitdata_release(struct nfs_commit_data *data)
1468 {
1469         put_nfs_open_context(data->context);
1470         nfs_commit_free(data);
1471 }
1472 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1473
1474 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1475                         const struct rpc_call_ops *call_ops,
1476                         int how, int flags)
1477 {
1478         struct rpc_task *task;
1479         int priority = flush_task_priority(how);
1480         struct rpc_message msg = {
1481                 .rpc_argp = &data->args,
1482                 .rpc_resp = &data->res,
1483                 .rpc_cred = data->cred,
1484         };
1485         struct rpc_task_setup task_setup_data = {
1486                 .task = &data->task,
1487                 .rpc_client = clnt,
1488                 .rpc_message = &msg,
1489                 .callback_ops = call_ops,
1490                 .callback_data = data,
1491                 .workqueue = nfsiod_workqueue,
1492                 .flags = RPC_TASK_ASYNC | flags,
1493                 .priority = priority,
1494         };
1495         /* Set up the initial task struct.  */
1496         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1497
1498         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1499
1500         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1501                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1502
1503         task = rpc_run_task(&task_setup_data);
1504         if (IS_ERR(task))
1505                 return PTR_ERR(task);
1506         if (how & FLUSH_SYNC)
1507                 rpc_wait_for_completion_task(task);
1508         rpc_put_task(task);
1509         return 0;
1510 }
1511 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1512
1513 /*
1514  * Set up the argument/result storage required for the RPC call.
1515  */
1516 void nfs_init_commit(struct nfs_commit_data *data,
1517                      struct list_head *head,
1518                      struct pnfs_layout_segment *lseg,
1519                      struct nfs_commit_info *cinfo)
1520 {
1521         struct nfs_page *first = nfs_list_entry(head->next);
1522         struct inode *inode = first->wb_context->dentry->d_inode;
1523
1524         /* Set up the RPC argument and reply structs
1525          * NB: take care not to mess about with data->commit et al. */
1526
1527         list_splice_init(head, &data->pages);
1528
1529         data->inode       = inode;
1530         data->cred        = first->wb_context->cred;
1531         data->lseg        = lseg; /* reference transferred */
1532         data->mds_ops     = &nfs_commit_ops;
1533         data->completion_ops = cinfo->completion_ops;
1534         data->dreq        = cinfo->dreq;
1535
1536         data->args.fh     = NFS_FH(data->inode);
1537         /* Note: we always request a commit of the entire inode */
1538         data->args.offset = 0;
1539         data->args.count  = 0;
1540         data->context     = get_nfs_open_context(first->wb_context);
1541         data->res.fattr   = &data->fattr;
1542         data->res.verf    = &data->verf;
1543         nfs_fattr_init(&data->fattr);
1544 }
1545 EXPORT_SYMBOL_GPL(nfs_init_commit);
1546
1547 void nfs_retry_commit(struct list_head *page_list,
1548                       struct pnfs_layout_segment *lseg,
1549                       struct nfs_commit_info *cinfo)
1550 {
1551         struct nfs_page *req;
1552
1553         while (!list_empty(page_list)) {
1554                 req = nfs_list_entry(page_list->next);
1555                 nfs_list_remove_request(req);
1556                 nfs_mark_request_commit(req, lseg, cinfo);
1557                 if (!cinfo->dreq) {
1558                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1559                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1560                                      BDI_RECLAIMABLE);
1561                 }
1562                 nfs_unlock_and_release_request(req);
1563         }
1564 }
1565 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1566
1567 /*
1568  * Commit dirty pages
1569  */
1570 static int
1571 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1572                 struct nfs_commit_info *cinfo)
1573 {
1574         struct nfs_commit_data  *data;
1575
1576         data = nfs_commitdata_alloc();
1577
1578         if (!data)
1579                 goto out_bad;
1580
1581         /* Set up the argument struct */
1582         nfs_init_commit(data, head, NULL, cinfo);
1583         atomic_inc(&cinfo->mds->rpcs_out);
1584         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1585                                    how, 0);
1586  out_bad:
1587         nfs_retry_commit(head, NULL, cinfo);
1588         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1589         return -ENOMEM;
1590 }
1591
1592 /*
1593  * COMMIT call returned
1594  */
1595 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1596 {
1597         struct nfs_commit_data  *data = calldata;
1598
1599         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1600                                 task->tk_pid, task->tk_status);
1601
1602         /* Call the NFS version-specific code */
1603         NFS_PROTO(data->inode)->commit_done(task, data);
1604 }
1605
1606 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1607 {
1608         struct nfs_page *req;
1609         int status = data->task.tk_status;
1610         struct nfs_commit_info cinfo;
1611
1612         while (!list_empty(&data->pages)) {
1613                 req = nfs_list_entry(data->pages.next);
1614                 nfs_list_remove_request(req);
1615                 nfs_clear_page_commit(req->wb_page);
1616
1617                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1618                         req->wb_context->dentry->d_sb->s_id,
1619                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1620                         req->wb_bytes,
1621                         (long long)req_offset(req));
1622                 if (status < 0) {
1623                         nfs_context_set_write_error(req->wb_context, status);
1624                         nfs_inode_remove_request(req);
1625                         dprintk(", error = %d\n", status);
1626                         goto next;
1627                 }
1628
1629                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1630                  * returned by the server against all stored verfs. */
1631                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1632                         /* We have a match */
1633                         nfs_inode_remove_request(req);
1634                         dprintk(" OK\n");
1635                         goto next;
1636                 }
1637                 /* We have a mismatch. Write the page again */
1638                 dprintk(" mismatch\n");
1639                 nfs_mark_request_dirty(req);
1640                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1641         next:
1642                 nfs_unlock_and_release_request(req);
1643         }
1644         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1645         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1646                 nfs_commit_clear_lock(NFS_I(data->inode));
1647 }
1648
1649 static void nfs_commit_release(void *calldata)
1650 {
1651         struct nfs_commit_data *data = calldata;
1652
1653         data->completion_ops->completion(data);
1654         nfs_commitdata_release(calldata);
1655 }
1656
1657 static const struct rpc_call_ops nfs_commit_ops = {
1658         .rpc_call_prepare = nfs_commit_prepare,
1659         .rpc_call_done = nfs_commit_done,
1660         .rpc_release = nfs_commit_release,
1661 };
1662
1663 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1664         .completion = nfs_commit_release_pages,
1665         .error_cleanup = nfs_commit_clear_lock,
1666 };
1667
1668 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1669                             int how, struct nfs_commit_info *cinfo)
1670 {
1671         int status;
1672
1673         status = pnfs_commit_list(inode, head, how, cinfo);
1674         if (status == PNFS_NOT_ATTEMPTED)
1675                 status = nfs_commit_list(inode, head, how, cinfo);
1676         return status;
1677 }
1678
1679 int nfs_commit_inode(struct inode *inode, int how)
1680 {
1681         LIST_HEAD(head);
1682         struct nfs_commit_info cinfo;
1683         int may_wait = how & FLUSH_SYNC;
1684         int res;
1685
1686         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1687         if (res <= 0)
1688                 goto out_mark_dirty;
1689         nfs_init_cinfo_from_inode(&cinfo, inode);
1690         res = nfs_scan_commit(inode, &head, &cinfo);
1691         if (res) {
1692                 int error;
1693
1694                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1695                 if (error < 0)
1696                         return error;
1697                 if (!may_wait)
1698                         goto out_mark_dirty;
1699                 error = wait_on_bit(&NFS_I(inode)->flags,
1700                                 NFS_INO_COMMIT,
1701                                 nfs_wait_bit_killable,
1702                                 TASK_KILLABLE);
1703                 if (error < 0)
1704                         return error;
1705         } else
1706                 nfs_commit_clear_lock(NFS_I(inode));
1707         return res;
1708         /* Note: If we exit without ensuring that the commit is complete,
1709          * we must mark the inode as dirty. Otherwise, future calls to
1710          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1711          * that the data is on the disk.
1712          */
1713 out_mark_dirty:
1714         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1715         return res;
1716 }
1717
1718 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1719 {
1720         struct nfs_inode *nfsi = NFS_I(inode);
1721         int flags = FLUSH_SYNC;
1722         int ret = 0;
1723
1724         /* no commits means nothing needs to be done */
1725         if (!nfsi->commit_info.ncommit)
1726                 return ret;
1727
1728         if (wbc->sync_mode == WB_SYNC_NONE) {
1729                 /* Don't commit yet if this is a non-blocking flush and there
1730                  * are a lot of outstanding writes for this mapping.
1731                  */
1732                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1733                         goto out_mark_dirty;
1734
1735                 /* don't wait for the COMMIT response */
1736                 flags = 0;
1737         }
1738
1739         ret = nfs_commit_inode(inode, flags);
1740         if (ret >= 0) {
1741                 if (wbc->sync_mode == WB_SYNC_NONE) {
1742                         if (ret < wbc->nr_to_write)
1743                                 wbc->nr_to_write -= ret;
1744                         else
1745                                 wbc->nr_to_write = 0;
1746                 }
1747                 return 0;
1748         }
1749 out_mark_dirty:
1750         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1751         return ret;
1752 }
1753 #else
1754 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1755 {
1756         return 0;
1757 }
1758 #endif
1759
1760 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1761 {
1762         return nfs_commit_unstable_pages(inode, wbc);
1763 }
1764 EXPORT_SYMBOL_GPL(nfs_write_inode);
1765
1766 /*
1767  * flush the inode to disk.
1768  */
1769 int nfs_wb_all(struct inode *inode)
1770 {
1771         struct writeback_control wbc = {
1772                 .sync_mode = WB_SYNC_ALL,
1773                 .nr_to_write = LONG_MAX,
1774                 .range_start = 0,
1775                 .range_end = LLONG_MAX,
1776         };
1777         int ret;
1778
1779         trace_nfs_writeback_inode_enter(inode);
1780
1781         ret = sync_inode(inode, &wbc);
1782
1783         trace_nfs_writeback_inode_exit(inode, ret);
1784         return ret;
1785 }
1786 EXPORT_SYMBOL_GPL(nfs_wb_all);
1787
1788 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1789 {
1790         struct nfs_page *req;
1791         int ret = 0;
1792
1793         for (;;) {
1794                 wait_on_page_writeback(page);
1795                 req = nfs_page_find_request(page);
1796                 if (req == NULL)
1797                         break;
1798                 if (nfs_lock_request(req)) {
1799                         nfs_clear_request_commit(req);
1800                         nfs_inode_remove_request(req);
1801                         /*
1802                          * In case nfs_inode_remove_request has marked the
1803                          * page as being dirty
1804                          */
1805                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1806                         nfs_unlock_and_release_request(req);
1807                         break;
1808                 }
1809                 ret = nfs_wait_on_request(req);
1810                 nfs_release_request(req);
1811                 if (ret < 0)
1812                         break;
1813         }
1814         return ret;
1815 }
1816
1817 /*
1818  * Write back all requests on one page - we do this before reading it.
1819  */
1820 int nfs_wb_page(struct inode *inode, struct page *page)
1821 {
1822         loff_t range_start = page_file_offset(page);
1823         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1824         struct writeback_control wbc = {
1825                 .sync_mode = WB_SYNC_ALL,
1826                 .nr_to_write = 0,
1827                 .range_start = range_start,
1828                 .range_end = range_end,
1829         };
1830         int ret;
1831
1832         trace_nfs_writeback_page_enter(inode);
1833
1834         for (;;) {
1835                 wait_on_page_writeback(page);
1836                 if (clear_page_dirty_for_io(page)) {
1837                         ret = nfs_writepage_locked(page, &wbc);
1838                         if (ret < 0)
1839                                 goto out_error;
1840                         continue;
1841                 }
1842                 ret = 0;
1843                 if (!PagePrivate(page))
1844                         break;
1845                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1846                 if (ret < 0)
1847                         goto out_error;
1848         }
1849 out_error:
1850         trace_nfs_writeback_page_exit(inode, ret);
1851         return ret;
1852 }
1853
1854 #ifdef CONFIG_MIGRATION
1855 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1856                 struct page *page, enum migrate_mode mode)
1857 {
1858         /*
1859          * If PagePrivate is set, then the page is currently associated with
1860          * an in-progress read or write request. Don't try to migrate it.
1861          *
1862          * FIXME: we could do this in principle, but we'll need a way to ensure
1863          *        that we can safely release the inode reference while holding
1864          *        the page lock.
1865          */
1866         if (PagePrivate(page))
1867                 return -EBUSY;
1868
1869         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1870                 return -EBUSY;
1871
1872         return migrate_page(mapping, newpage, page, mode);
1873 }
1874 #endif
1875
1876 int __init nfs_init_writepagecache(void)
1877 {
1878         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1879                                              sizeof(struct nfs_write_header),
1880                                              0, SLAB_HWCACHE_ALIGN,
1881                                              NULL);
1882         if (nfs_wdata_cachep == NULL)
1883                 return -ENOMEM;
1884
1885         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1886                                                      nfs_wdata_cachep);
1887         if (nfs_wdata_mempool == NULL)
1888                 goto out_destroy_write_cache;
1889
1890         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1891                                              sizeof(struct nfs_commit_data),
1892                                              0, SLAB_HWCACHE_ALIGN,
1893                                              NULL);
1894         if (nfs_cdata_cachep == NULL)
1895                 goto out_destroy_write_mempool;
1896
1897         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1898                                                       nfs_cdata_cachep);
1899         if (nfs_commit_mempool == NULL)
1900                 goto out_destroy_commit_cache;
1901
1902         /*
1903          * NFS congestion size, scale with available memory.
1904          *
1905          *  64MB:    8192k
1906          * 128MB:   11585k
1907          * 256MB:   16384k
1908          * 512MB:   23170k
1909          *   1GB:   32768k
1910          *   2GB:   46340k
1911          *   4GB:   65536k
1912          *   8GB:   92681k
1913          *  16GB:  131072k
1914          *
1915          * This allows larger machines to have larger/more transfers.
1916          * Limit the default to 256M
1917          */
1918         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1919         if (nfs_congestion_kb > 256*1024)
1920                 nfs_congestion_kb = 256*1024;
1921
1922         return 0;
1923
1924 out_destroy_commit_cache:
1925         kmem_cache_destroy(nfs_cdata_cachep);
1926 out_destroy_write_mempool:
1927         mempool_destroy(nfs_wdata_mempool);
1928 out_destroy_write_cache:
1929         kmem_cache_destroy(nfs_wdata_cachep);
1930         return -ENOMEM;
1931 }
1932
1933 void nfs_destroy_writepagecache(void)
1934 {
1935         mempool_destroy(nfs_commit_mempool);
1936         kmem_cache_destroy(nfs_cdata_cachep);
1937         mempool_destroy(nfs_wdata_mempool);
1938         kmem_cache_destroy(nfs_wdata_cachep);
1939 }
1940