]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/nfs/write.c
Merge tag 'arc-v3.16-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/vgupta/arc
[karo-tx-linux.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_atomic();
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_INVALID_DATA|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         return PageUptodate(page) != 0;
923 }
924
925 /* If we know the page is up to date, and we're not using byte range locks (or
926  * if we have the whole file locked for writing), it may be more efficient to
927  * extend the write to cover the entire page in order to avoid fragmentation
928  * inefficiencies.
929  *
930  * If the file is opened for synchronous writes then we can just skip the rest
931  * of the checks.
932  */
933 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
934 {
935         if (file->f_flags & O_DSYNC)
936                 return 0;
937         if (!nfs_write_pageuptodate(page, inode))
938                 return 0;
939         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
940                 return 1;
941         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
942                         inode->i_flock->fl_end == OFFSET_MAX &&
943                         inode->i_flock->fl_type != F_RDLCK))
944                 return 1;
945         return 0;
946 }
947
948 /*
949  * Update and possibly write a cached page of an NFS file.
950  *
951  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
952  * things with a page scheduled for an RPC call (e.g. invalidate it).
953  */
954 int nfs_updatepage(struct file *file, struct page *page,
955                 unsigned int offset, unsigned int count)
956 {
957         struct nfs_open_context *ctx = nfs_file_open_context(file);
958         struct inode    *inode = page_file_mapping(page)->host;
959         int             status = 0;
960
961         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
962
963         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
964                 file, count, (long long)(page_file_offset(page) + offset));
965
966         if (nfs_can_extend_write(file, page, inode)) {
967                 count = max(count + offset, nfs_page_length(page));
968                 offset = 0;
969         }
970
971         status = nfs_writepage_setup(ctx, page, offset, count);
972         if (status < 0)
973                 nfs_set_pageerror(page);
974         else
975                 __set_page_dirty_nobuffers(page);
976
977         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
978                         status, (long long)i_size_read(inode));
979         return status;
980 }
981
982 static int flush_task_priority(int how)
983 {
984         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
985                 case FLUSH_HIGHPRI:
986                         return RPC_PRIORITY_HIGH;
987                 case FLUSH_LOWPRI:
988                         return RPC_PRIORITY_LOW;
989         }
990         return RPC_PRIORITY_NORMAL;
991 }
992
993 int nfs_initiate_write(struct rpc_clnt *clnt,
994                        struct nfs_write_data *data,
995                        const struct rpc_call_ops *call_ops,
996                        int how, int flags)
997 {
998         struct inode *inode = data->header->inode;
999         int priority = flush_task_priority(how);
1000         struct rpc_task *task;
1001         struct rpc_message msg = {
1002                 .rpc_argp = &data->args,
1003                 .rpc_resp = &data->res,
1004                 .rpc_cred = data->header->cred,
1005         };
1006         struct rpc_task_setup task_setup_data = {
1007                 .rpc_client = clnt,
1008                 .task = &data->task,
1009                 .rpc_message = &msg,
1010                 .callback_ops = call_ops,
1011                 .callback_data = data,
1012                 .workqueue = nfsiod_workqueue,
1013                 .flags = RPC_TASK_ASYNC | flags,
1014                 .priority = priority,
1015         };
1016         int ret = 0;
1017
1018         /* Set up the initial task struct.  */
1019         NFS_PROTO(inode)->write_setup(data, &msg);
1020
1021         dprintk("NFS: %5u initiated write call "
1022                 "(req %s/%llu, %u bytes @ offset %llu)\n",
1023                 data->task.tk_pid,
1024                 inode->i_sb->s_id,
1025                 (unsigned long long)NFS_FILEID(inode),
1026                 data->args.count,
1027                 (unsigned long long)data->args.offset);
1028
1029         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1030                                  &task_setup_data.rpc_client, &msg, data);
1031
1032         task = rpc_run_task(&task_setup_data);
1033         if (IS_ERR(task)) {
1034                 ret = PTR_ERR(task);
1035                 goto out;
1036         }
1037         if (how & FLUSH_SYNC) {
1038                 ret = rpc_wait_for_completion_task(task);
1039                 if (ret == 0)
1040                         ret = task->tk_status;
1041         }
1042         rpc_put_task(task);
1043 out:
1044         return ret;
1045 }
1046 EXPORT_SYMBOL_GPL(nfs_initiate_write);
1047
1048 /*
1049  * Set up the argument/result storage required for the RPC call.
1050  */
1051 static void nfs_write_rpcsetup(struct nfs_write_data *data,
1052                 unsigned int count, unsigned int offset,
1053                 int how, struct nfs_commit_info *cinfo)
1054 {
1055         struct nfs_page *req = data->header->req;
1056
1057         /* Set up the RPC argument and reply structs
1058          * NB: take care not to mess about with data->commit et al. */
1059
1060         data->args.fh     = NFS_FH(data->header->inode);
1061         data->args.offset = req_offset(req) + offset;
1062         /* pnfs_set_layoutcommit needs this */
1063         data->mds_offset = data->args.offset;
1064         data->args.pgbase = req->wb_pgbase + offset;
1065         data->args.pages  = data->pages.pagevec;
1066         data->args.count  = count;
1067         data->args.context = get_nfs_open_context(req->wb_context);
1068         data->args.lock_context = req->wb_lock_context;
1069         data->args.stable  = NFS_UNSTABLE;
1070         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1071         case 0:
1072                 break;
1073         case FLUSH_COND_STABLE:
1074                 if (nfs_reqs_to_commit(cinfo))
1075                         break;
1076         default:
1077                 data->args.stable = NFS_FILE_SYNC;
1078         }
1079
1080         data->res.fattr   = &data->fattr;
1081         data->res.count   = count;
1082         data->res.verf    = &data->verf;
1083         nfs_fattr_init(&data->fattr);
1084 }
1085
1086 static int nfs_do_write(struct nfs_write_data *data,
1087                 const struct rpc_call_ops *call_ops,
1088                 int how)
1089 {
1090         struct inode *inode = data->header->inode;
1091
1092         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1093 }
1094
1095 static int nfs_do_multiple_writes(struct list_head *head,
1096                 const struct rpc_call_ops *call_ops,
1097                 int how)
1098 {
1099         struct nfs_write_data *data;
1100         int ret = 0;
1101
1102         while (!list_empty(head)) {
1103                 int ret2;
1104
1105                 data = list_first_entry(head, struct nfs_write_data, list);
1106                 list_del_init(&data->list);
1107                 
1108                 ret2 = nfs_do_write(data, call_ops, how);
1109                  if (ret == 0)
1110                          ret = ret2;
1111         }
1112         return ret;
1113 }
1114
1115 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1116  * call this on each, which will prepare them to be retried on next
1117  * writeback using standard nfs.
1118  */
1119 static void nfs_redirty_request(struct nfs_page *req)
1120 {
1121         nfs_mark_request_dirty(req);
1122         nfs_unlock_request(req);
1123         nfs_end_page_writeback(req->wb_page);
1124         nfs_release_request(req);
1125 }
1126
1127 static void nfs_async_write_error(struct list_head *head)
1128 {
1129         struct nfs_page *req;
1130
1131         while (!list_empty(head)) {
1132                 req = nfs_list_entry(head->next);
1133                 nfs_list_remove_request(req);
1134                 nfs_redirty_request(req);
1135         }
1136 }
1137
1138 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1139         .error_cleanup = nfs_async_write_error,
1140         .completion = nfs_write_completion,
1141 };
1142
1143 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1144                 struct nfs_pgio_header *hdr)
1145 {
1146         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1147         while (!list_empty(&hdr->rpc_list)) {
1148                 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1149                                 struct nfs_write_data, list);
1150                 list_del(&data->list);
1151                 nfs_writedata_release(data);
1152         }
1153         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1154 }
1155
1156 /*
1157  * Generate multiple small requests to write out a single
1158  * contiguous dirty area on one page.
1159  */
1160 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1161                            struct nfs_pgio_header *hdr)
1162 {
1163         struct nfs_page *req = hdr->req;
1164         struct page *page = req->wb_page;
1165         struct nfs_write_data *data;
1166         size_t wsize = desc->pg_bsize, nbytes;
1167         unsigned int offset;
1168         int requests = 0;
1169         struct nfs_commit_info cinfo;
1170
1171         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1172
1173         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1174             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1175              desc->pg_count > wsize))
1176                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1177
1178
1179         offset = 0;
1180         nbytes = desc->pg_count;
1181         do {
1182                 size_t len = min(nbytes, wsize);
1183
1184                 data = nfs_writedata_alloc(hdr, 1);
1185                 if (!data) {
1186                         nfs_flush_error(desc, hdr);
1187                         return -ENOMEM;
1188                 }
1189                 data->pages.pagevec[0] = page;
1190                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1191                 list_add(&data->list, &hdr->rpc_list);
1192                 requests++;
1193                 nbytes -= len;
1194                 offset += len;
1195         } while (nbytes != 0);
1196         nfs_list_remove_request(req);
1197         nfs_list_add_request(req, &hdr->pages);
1198         desc->pg_rpc_callops = &nfs_write_common_ops;
1199         return 0;
1200 }
1201
1202 /*
1203  * Create an RPC task for the given write request and kick it.
1204  * The page must have been locked by the caller.
1205  *
1206  * It may happen that the page we're passed is not marked dirty.
1207  * This is the case if nfs_updatepage detects a conflicting request
1208  * that has been written but not committed.
1209  */
1210 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1211                          struct nfs_pgio_header *hdr)
1212 {
1213         struct nfs_page         *req;
1214         struct page             **pages;
1215         struct nfs_write_data   *data;
1216         struct list_head *head = &desc->pg_list;
1217         struct nfs_commit_info cinfo;
1218
1219         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1220                                                            desc->pg_count));
1221         if (!data) {
1222                 nfs_flush_error(desc, hdr);
1223                 return -ENOMEM;
1224         }
1225
1226         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1227         pages = data->pages.pagevec;
1228         while (!list_empty(head)) {
1229                 req = nfs_list_entry(head->next);
1230                 nfs_list_remove_request(req);
1231                 nfs_list_add_request(req, &hdr->pages);
1232                 *pages++ = req->wb_page;
1233         }
1234
1235         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1236             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1237                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1238
1239         /* Set up the argument struct */
1240         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1241         list_add(&data->list, &hdr->rpc_list);
1242         desc->pg_rpc_callops = &nfs_write_common_ops;
1243         return 0;
1244 }
1245
1246 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1247                       struct nfs_pgio_header *hdr)
1248 {
1249         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1250                 return nfs_flush_multi(desc, hdr);
1251         return nfs_flush_one(desc, hdr);
1252 }
1253 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1254
1255 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1256 {
1257         struct nfs_write_header *whdr;
1258         struct nfs_pgio_header *hdr;
1259         int ret;
1260
1261         whdr = nfs_writehdr_alloc();
1262         if (!whdr) {
1263                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1264                 return -ENOMEM;
1265         }
1266         hdr = &whdr->header;
1267         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1268         atomic_inc(&hdr->refcnt);
1269         ret = nfs_generic_flush(desc, hdr);
1270         if (ret == 0)
1271                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1272                                              desc->pg_rpc_callops,
1273                                              desc->pg_ioflags);
1274         if (atomic_dec_and_test(&hdr->refcnt))
1275                 hdr->completion_ops->completion(hdr);
1276         return ret;
1277 }
1278
1279 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1280         .pg_test = nfs_generic_pg_test,
1281         .pg_doio = nfs_generic_pg_writepages,
1282 };
1283
1284 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1285                                struct inode *inode, int ioflags,
1286                                const struct nfs_pgio_completion_ops *compl_ops)
1287 {
1288         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1289                                 NFS_SERVER(inode)->wsize, ioflags);
1290 }
1291 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1292
1293 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1294 {
1295         pgio->pg_ops = &nfs_pageio_write_ops;
1296         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1297 }
1298 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1299
1300
1301 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1302 {
1303         struct nfs_write_data *data = calldata;
1304         int err;
1305         err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1306         if (err)
1307                 rpc_exit(task, err);
1308 }
1309
1310 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1311 {
1312         struct nfs_commit_data *data = calldata;
1313
1314         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1315 }
1316
1317 /*
1318  * Handle a write reply that flushes a whole page.
1319  *
1320  * FIXME: There is an inherent race with invalidate_inode_pages and
1321  *        writebacks since the page->count is kept > 1 for as long
1322  *        as the page has a write request pending.
1323  */
1324 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1325 {
1326         struct nfs_write_data   *data = calldata;
1327
1328         nfs_writeback_done(task, data);
1329 }
1330
1331 static void nfs_writeback_release_common(void *calldata)
1332 {
1333         struct nfs_write_data   *data = calldata;
1334         struct nfs_pgio_header *hdr = data->header;
1335         int status = data->task.tk_status;
1336
1337         if ((status >= 0) && nfs_write_need_commit(data)) {
1338                 spin_lock(&hdr->lock);
1339                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1340                         ; /* Do nothing */
1341                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1342                         memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1343                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1344                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1345                 spin_unlock(&hdr->lock);
1346         }
1347         nfs_writedata_release(data);
1348 }
1349
1350 static const struct rpc_call_ops nfs_write_common_ops = {
1351         .rpc_call_prepare = nfs_write_prepare,
1352         .rpc_call_done = nfs_writeback_done_common,
1353         .rpc_release = nfs_writeback_release_common,
1354 };
1355
1356
1357 /*
1358  * This function is called when the WRITE call is complete.
1359  */
1360 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1361 {
1362         struct nfs_writeargs    *argp = &data->args;
1363         struct nfs_writeres     *resp = &data->res;
1364         struct inode            *inode = data->header->inode;
1365         int status;
1366
1367         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1368                 task->tk_pid, task->tk_status);
1369
1370         /*
1371          * ->write_done will attempt to use post-op attributes to detect
1372          * conflicting writes by other clients.  A strict interpretation
1373          * of close-to-open would allow us to continue caching even if
1374          * another writer had changed the file, but some applications
1375          * depend on tighter cache coherency when writing.
1376          */
1377         status = NFS_PROTO(inode)->write_done(task, data);
1378         if (status != 0)
1379                 return;
1380         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1381
1382 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1383         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1384                 /* We tried a write call, but the server did not
1385                  * commit data to stable storage even though we
1386                  * requested it.
1387                  * Note: There is a known bug in Tru64 < 5.0 in which
1388                  *       the server reports NFS_DATA_SYNC, but performs
1389                  *       NFS_FILE_SYNC. We therefore implement this checking
1390                  *       as a dprintk() in order to avoid filling syslog.
1391                  */
1392                 static unsigned long    complain;
1393
1394                 /* Note this will print the MDS for a DS write */
1395                 if (time_before(complain, jiffies)) {
1396                         dprintk("NFS:       faulty NFS server %s:"
1397                                 " (committed = %d) != (stable = %d)\n",
1398                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1399                                 resp->verf->committed, argp->stable);
1400                         complain = jiffies + 300 * HZ;
1401                 }
1402         }
1403 #endif
1404         if (task->tk_status < 0)
1405                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1406         else if (resp->count < argp->count) {
1407                 static unsigned long    complain;
1408
1409                 /* This a short write! */
1410                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1411
1412                 /* Has the server at least made some progress? */
1413                 if (resp->count == 0) {
1414                         if (time_before(complain, jiffies)) {
1415                                 printk(KERN_WARNING
1416                                        "NFS: Server wrote zero bytes, expected %u.\n",
1417                                        argp->count);
1418                                 complain = jiffies + 300 * HZ;
1419                         }
1420                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1421                         task->tk_status = -EIO;
1422                         return;
1423                 }
1424                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1425                 if (resp->verf->committed != NFS_UNSTABLE) {
1426                         /* Resend from where the server left off */
1427                         data->mds_offset += resp->count;
1428                         argp->offset += resp->count;
1429                         argp->pgbase += resp->count;
1430                         argp->count -= resp->count;
1431                 } else {
1432                         /* Resend as a stable write in order to avoid
1433                          * headaches in the case of a server crash.
1434                          */
1435                         argp->stable = NFS_FILE_SYNC;
1436                 }
1437                 rpc_restart_call_prepare(task);
1438         }
1439 }
1440
1441
1442 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1443 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1444 {
1445         int ret;
1446
1447         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1448                 return 1;
1449         if (!may_wait)
1450                 return 0;
1451         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1452                                 NFS_INO_COMMIT,
1453                                 nfs_wait_bit_killable,
1454                                 TASK_KILLABLE);
1455         return (ret < 0) ? ret : 1;
1456 }
1457
1458 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1459 {
1460         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1461         smp_mb__after_atomic();
1462         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1463 }
1464
1465 void nfs_commitdata_release(struct nfs_commit_data *data)
1466 {
1467         put_nfs_open_context(data->context);
1468         nfs_commit_free(data);
1469 }
1470 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1471
1472 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1473                         const struct rpc_call_ops *call_ops,
1474                         int how, int flags)
1475 {
1476         struct rpc_task *task;
1477         int priority = flush_task_priority(how);
1478         struct rpc_message msg = {
1479                 .rpc_argp = &data->args,
1480                 .rpc_resp = &data->res,
1481                 .rpc_cred = data->cred,
1482         };
1483         struct rpc_task_setup task_setup_data = {
1484                 .task = &data->task,
1485                 .rpc_client = clnt,
1486                 .rpc_message = &msg,
1487                 .callback_ops = call_ops,
1488                 .callback_data = data,
1489                 .workqueue = nfsiod_workqueue,
1490                 .flags = RPC_TASK_ASYNC | flags,
1491                 .priority = priority,
1492         };
1493         /* Set up the initial task struct.  */
1494         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1495
1496         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1497
1498         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1499                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1500
1501         task = rpc_run_task(&task_setup_data);
1502         if (IS_ERR(task))
1503                 return PTR_ERR(task);
1504         if (how & FLUSH_SYNC)
1505                 rpc_wait_for_completion_task(task);
1506         rpc_put_task(task);
1507         return 0;
1508 }
1509 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1510
1511 /*
1512  * Set up the argument/result storage required for the RPC call.
1513  */
1514 void nfs_init_commit(struct nfs_commit_data *data,
1515                      struct list_head *head,
1516                      struct pnfs_layout_segment *lseg,
1517                      struct nfs_commit_info *cinfo)
1518 {
1519         struct nfs_page *first = nfs_list_entry(head->next);
1520         struct inode *inode = first->wb_context->dentry->d_inode;
1521
1522         /* Set up the RPC argument and reply structs
1523          * NB: take care not to mess about with data->commit et al. */
1524
1525         list_splice_init(head, &data->pages);
1526
1527         data->inode       = inode;
1528         data->cred        = first->wb_context->cred;
1529         data->lseg        = lseg; /* reference transferred */
1530         data->mds_ops     = &nfs_commit_ops;
1531         data->completion_ops = cinfo->completion_ops;
1532         data->dreq        = cinfo->dreq;
1533
1534         data->args.fh     = NFS_FH(data->inode);
1535         /* Note: we always request a commit of the entire inode */
1536         data->args.offset = 0;
1537         data->args.count  = 0;
1538         data->context     = get_nfs_open_context(first->wb_context);
1539         data->res.fattr   = &data->fattr;
1540         data->res.verf    = &data->verf;
1541         nfs_fattr_init(&data->fattr);
1542 }
1543 EXPORT_SYMBOL_GPL(nfs_init_commit);
1544
1545 void nfs_retry_commit(struct list_head *page_list,
1546                       struct pnfs_layout_segment *lseg,
1547                       struct nfs_commit_info *cinfo)
1548 {
1549         struct nfs_page *req;
1550
1551         while (!list_empty(page_list)) {
1552                 req = nfs_list_entry(page_list->next);
1553                 nfs_list_remove_request(req);
1554                 nfs_mark_request_commit(req, lseg, cinfo);
1555                 if (!cinfo->dreq) {
1556                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1557                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1558                                      BDI_RECLAIMABLE);
1559                 }
1560                 nfs_unlock_and_release_request(req);
1561         }
1562 }
1563 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1564
1565 /*
1566  * Commit dirty pages
1567  */
1568 static int
1569 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1570                 struct nfs_commit_info *cinfo)
1571 {
1572         struct nfs_commit_data  *data;
1573
1574         data = nfs_commitdata_alloc();
1575
1576         if (!data)
1577                 goto out_bad;
1578
1579         /* Set up the argument struct */
1580         nfs_init_commit(data, head, NULL, cinfo);
1581         atomic_inc(&cinfo->mds->rpcs_out);
1582         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1583                                    how, 0);
1584  out_bad:
1585         nfs_retry_commit(head, NULL, cinfo);
1586         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1587         return -ENOMEM;
1588 }
1589
1590 /*
1591  * COMMIT call returned
1592  */
1593 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1594 {
1595         struct nfs_commit_data  *data = calldata;
1596
1597         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1598                                 task->tk_pid, task->tk_status);
1599
1600         /* Call the NFS version-specific code */
1601         NFS_PROTO(data->inode)->commit_done(task, data);
1602 }
1603
1604 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1605 {
1606         struct nfs_page *req;
1607         int status = data->task.tk_status;
1608         struct nfs_commit_info cinfo;
1609
1610         while (!list_empty(&data->pages)) {
1611                 req = nfs_list_entry(data->pages.next);
1612                 nfs_list_remove_request(req);
1613                 nfs_clear_page_commit(req->wb_page);
1614
1615                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1616                         req->wb_context->dentry->d_sb->s_id,
1617                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1618                         req->wb_bytes,
1619                         (long long)req_offset(req));
1620                 if (status < 0) {
1621                         nfs_context_set_write_error(req->wb_context, status);
1622                         nfs_inode_remove_request(req);
1623                         dprintk(", error = %d\n", status);
1624                         goto next;
1625                 }
1626
1627                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1628                  * returned by the server against all stored verfs. */
1629                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1630                         /* We have a match */
1631                         nfs_inode_remove_request(req);
1632                         dprintk(" OK\n");
1633                         goto next;
1634                 }
1635                 /* We have a mismatch. Write the page again */
1636                 dprintk(" mismatch\n");
1637                 nfs_mark_request_dirty(req);
1638                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1639         next:
1640                 nfs_unlock_and_release_request(req);
1641         }
1642         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1643         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1644                 nfs_commit_clear_lock(NFS_I(data->inode));
1645 }
1646
1647 static void nfs_commit_release(void *calldata)
1648 {
1649         struct nfs_commit_data *data = calldata;
1650
1651         data->completion_ops->completion(data);
1652         nfs_commitdata_release(calldata);
1653 }
1654
1655 static const struct rpc_call_ops nfs_commit_ops = {
1656         .rpc_call_prepare = nfs_commit_prepare,
1657         .rpc_call_done = nfs_commit_done,
1658         .rpc_release = nfs_commit_release,
1659 };
1660
1661 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1662         .completion = nfs_commit_release_pages,
1663         .error_cleanup = nfs_commit_clear_lock,
1664 };
1665
1666 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1667                             int how, struct nfs_commit_info *cinfo)
1668 {
1669         int status;
1670
1671         status = pnfs_commit_list(inode, head, how, cinfo);
1672         if (status == PNFS_NOT_ATTEMPTED)
1673                 status = nfs_commit_list(inode, head, how, cinfo);
1674         return status;
1675 }
1676
1677 int nfs_commit_inode(struct inode *inode, int how)
1678 {
1679         LIST_HEAD(head);
1680         struct nfs_commit_info cinfo;
1681         int may_wait = how & FLUSH_SYNC;
1682         int res;
1683
1684         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1685         if (res <= 0)
1686                 goto out_mark_dirty;
1687         nfs_init_cinfo_from_inode(&cinfo, inode);
1688         res = nfs_scan_commit(inode, &head, &cinfo);
1689         if (res) {
1690                 int error;
1691
1692                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1693                 if (error < 0)
1694                         return error;
1695                 if (!may_wait)
1696                         goto out_mark_dirty;
1697                 error = wait_on_bit(&NFS_I(inode)->flags,
1698                                 NFS_INO_COMMIT,
1699                                 nfs_wait_bit_killable,
1700                                 TASK_KILLABLE);
1701                 if (error < 0)
1702                         return error;
1703         } else
1704                 nfs_commit_clear_lock(NFS_I(inode));
1705         return res;
1706         /* Note: If we exit without ensuring that the commit is complete,
1707          * we must mark the inode as dirty. Otherwise, future calls to
1708          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1709          * that the data is on the disk.
1710          */
1711 out_mark_dirty:
1712         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1713         return res;
1714 }
1715
1716 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1717 {
1718         struct nfs_inode *nfsi = NFS_I(inode);
1719         int flags = FLUSH_SYNC;
1720         int ret = 0;
1721
1722         /* no commits means nothing needs to be done */
1723         if (!nfsi->commit_info.ncommit)
1724                 return ret;
1725
1726         if (wbc->sync_mode == WB_SYNC_NONE) {
1727                 /* Don't commit yet if this is a non-blocking flush and there
1728                  * are a lot of outstanding writes for this mapping.
1729                  */
1730                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1731                         goto out_mark_dirty;
1732
1733                 /* don't wait for the COMMIT response */
1734                 flags = 0;
1735         }
1736
1737         ret = nfs_commit_inode(inode, flags);
1738         if (ret >= 0) {
1739                 if (wbc->sync_mode == WB_SYNC_NONE) {
1740                         if (ret < wbc->nr_to_write)
1741                                 wbc->nr_to_write -= ret;
1742                         else
1743                                 wbc->nr_to_write = 0;
1744                 }
1745                 return 0;
1746         }
1747 out_mark_dirty:
1748         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1749         return ret;
1750 }
1751 #else
1752 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1753 {
1754         return 0;
1755 }
1756 #endif
1757
1758 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1759 {
1760         return nfs_commit_unstable_pages(inode, wbc);
1761 }
1762 EXPORT_SYMBOL_GPL(nfs_write_inode);
1763
1764 /*
1765  * flush the inode to disk.
1766  */
1767 int nfs_wb_all(struct inode *inode)
1768 {
1769         struct writeback_control wbc = {
1770                 .sync_mode = WB_SYNC_ALL,
1771                 .nr_to_write = LONG_MAX,
1772                 .range_start = 0,
1773                 .range_end = LLONG_MAX,
1774         };
1775         int ret;
1776
1777         trace_nfs_writeback_inode_enter(inode);
1778
1779         ret = sync_inode(inode, &wbc);
1780
1781         trace_nfs_writeback_inode_exit(inode, ret);
1782         return ret;
1783 }
1784 EXPORT_SYMBOL_GPL(nfs_wb_all);
1785
1786 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1787 {
1788         struct nfs_page *req;
1789         int ret = 0;
1790
1791         for (;;) {
1792                 wait_on_page_writeback(page);
1793                 req = nfs_page_find_request(page);
1794                 if (req == NULL)
1795                         break;
1796                 if (nfs_lock_request(req)) {
1797                         nfs_clear_request_commit(req);
1798                         nfs_inode_remove_request(req);
1799                         /*
1800                          * In case nfs_inode_remove_request has marked the
1801                          * page as being dirty
1802                          */
1803                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1804                         nfs_unlock_and_release_request(req);
1805                         break;
1806                 }
1807                 ret = nfs_wait_on_request(req);
1808                 nfs_release_request(req);
1809                 if (ret < 0)
1810                         break;
1811         }
1812         return ret;
1813 }
1814
1815 /*
1816  * Write back all requests on one page - we do this before reading it.
1817  */
1818 int nfs_wb_page(struct inode *inode, struct page *page)
1819 {
1820         loff_t range_start = page_file_offset(page);
1821         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1822         struct writeback_control wbc = {
1823                 .sync_mode = WB_SYNC_ALL,
1824                 .nr_to_write = 0,
1825                 .range_start = range_start,
1826                 .range_end = range_end,
1827         };
1828         int ret;
1829
1830         trace_nfs_writeback_page_enter(inode);
1831
1832         for (;;) {
1833                 wait_on_page_writeback(page);
1834                 if (clear_page_dirty_for_io(page)) {
1835                         ret = nfs_writepage_locked(page, &wbc);
1836                         if (ret < 0)
1837                                 goto out_error;
1838                         continue;
1839                 }
1840                 ret = 0;
1841                 if (!PagePrivate(page))
1842                         break;
1843                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1844                 if (ret < 0)
1845                         goto out_error;
1846         }
1847 out_error:
1848         trace_nfs_writeback_page_exit(inode, ret);
1849         return ret;
1850 }
1851
1852 #ifdef CONFIG_MIGRATION
1853 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1854                 struct page *page, enum migrate_mode mode)
1855 {
1856         /*
1857          * If PagePrivate is set, then the page is currently associated with
1858          * an in-progress read or write request. Don't try to migrate it.
1859          *
1860          * FIXME: we could do this in principle, but we'll need a way to ensure
1861          *        that we can safely release the inode reference while holding
1862          *        the page lock.
1863          */
1864         if (PagePrivate(page))
1865                 return -EBUSY;
1866
1867         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1868                 return -EBUSY;
1869
1870         return migrate_page(mapping, newpage, page, mode);
1871 }
1872 #endif
1873
1874 int __init nfs_init_writepagecache(void)
1875 {
1876         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1877                                              sizeof(struct nfs_write_header),
1878                                              0, SLAB_HWCACHE_ALIGN,
1879                                              NULL);
1880         if (nfs_wdata_cachep == NULL)
1881                 return -ENOMEM;
1882
1883         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1884                                                      nfs_wdata_cachep);
1885         if (nfs_wdata_mempool == NULL)
1886                 goto out_destroy_write_cache;
1887
1888         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1889                                              sizeof(struct nfs_commit_data),
1890                                              0, SLAB_HWCACHE_ALIGN,
1891                                              NULL);
1892         if (nfs_cdata_cachep == NULL)
1893                 goto out_destroy_write_mempool;
1894
1895         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1896                                                       nfs_cdata_cachep);
1897         if (nfs_commit_mempool == NULL)
1898                 goto out_destroy_commit_cache;
1899
1900         /*
1901          * NFS congestion size, scale with available memory.
1902          *
1903          *  64MB:    8192k
1904          * 128MB:   11585k
1905          * 256MB:   16384k
1906          * 512MB:   23170k
1907          *   1GB:   32768k
1908          *   2GB:   46340k
1909          *   4GB:   65536k
1910          *   8GB:   92681k
1911          *  16GB:  131072k
1912          *
1913          * This allows larger machines to have larger/more transfers.
1914          * Limit the default to 256M
1915          */
1916         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1917         if (nfs_congestion_kb > 256*1024)
1918                 nfs_congestion_kb = 256*1024;
1919
1920         return 0;
1921
1922 out_destroy_commit_cache:
1923         kmem_cache_destroy(nfs_cdata_cachep);
1924 out_destroy_write_mempool:
1925         mempool_destroy(nfs_wdata_mempool);
1926 out_destroy_write_cache:
1927         kmem_cache_destroy(nfs_wdata_cachep);
1928         return -ENOMEM;
1929 }
1930
1931 void nfs_destroy_writepagecache(void)
1932 {
1933         mempool_destroy(nfs_commit_mempool);
1934         kmem_cache_destroy(nfs_cdata_cachep);
1935         mempool_destroy(nfs_wdata_mempool);
1936         kmem_cache_destroy(nfs_wdata_cachep);
1937 }
1938