NFS: Fix a delegation callback race
[linux-3.10.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 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
35
36 #define MIN_POOL_WRITE          (32)
37 #define MIN_POOL_COMMIT         (4)
38
39 /*
40  * Local function declarations
41  */
42 static void nfs_redirty_request(struct nfs_page *req);
43 static const struct rpc_call_ops nfs_write_common_ops;
44 static const struct rpc_call_ops nfs_commit_ops;
45 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
46 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
47
48 static struct kmem_cache *nfs_wdata_cachep;
49 static mempool_t *nfs_wdata_mempool;
50 static struct kmem_cache *nfs_cdata_cachep;
51 static mempool_t *nfs_commit_mempool;
52
53 struct nfs_commit_data *nfs_commitdata_alloc(void)
54 {
55         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
56
57         if (p) {
58                 memset(p, 0, sizeof(*p));
59                 INIT_LIST_HEAD(&p->pages);
60         }
61         return p;
62 }
63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
64
65 void nfs_commit_free(struct nfs_commit_data *p)
66 {
67         mempool_free(p, nfs_commit_mempool);
68 }
69 EXPORT_SYMBOL_GPL(nfs_commit_free);
70
71 struct nfs_write_header *nfs_writehdr_alloc(void)
72 {
73         struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
74
75         if (p) {
76                 struct nfs_pgio_header *hdr = &p->header;
77
78                 memset(p, 0, sizeof(*p));
79                 INIT_LIST_HEAD(&hdr->pages);
80                 INIT_LIST_HEAD(&hdr->rpc_list);
81                 spin_lock_init(&hdr->lock);
82                 atomic_set(&hdr->refcnt, 0);
83                 hdr->verf = &p->verf;
84         }
85         return p;
86 }
87 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
88
89 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
90                                                   unsigned int pagecount)
91 {
92         struct nfs_write_data *data, *prealloc;
93
94         prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
95         if (prealloc->header == NULL)
96                 data = prealloc;
97         else
98                 data = kzalloc(sizeof(*data), GFP_KERNEL);
99         if (!data)
100                 goto out;
101
102         if (nfs_pgarray_set(&data->pages, pagecount)) {
103                 data->header = hdr;
104                 atomic_inc(&hdr->refcnt);
105         } else {
106                 if (data != prealloc)
107                         kfree(data);
108                 data = NULL;
109         }
110 out:
111         return data;
112 }
113
114 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
115 {
116         struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
117         mempool_free(whdr, nfs_wdata_mempool);
118 }
119 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
120
121 void nfs_writedata_release(struct nfs_write_data *wdata)
122 {
123         struct nfs_pgio_header *hdr = wdata->header;
124         struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
125
126         put_nfs_open_context(wdata->args.context);
127         if (wdata->pages.pagevec != wdata->pages.page_array)
128                 kfree(wdata->pages.pagevec);
129         if (wdata == &write_header->rpc_data) {
130                 wdata->header = NULL;
131                 wdata = NULL;
132         }
133         if (atomic_dec_and_test(&hdr->refcnt))
134                 hdr->completion_ops->completion(hdr);
135         /* Note: we only free the rpc_task after callbacks are done.
136          * See the comment in rpc_free_task() for why
137          */
138         kfree(wdata);
139 }
140 EXPORT_SYMBOL_GPL(nfs_writedata_release);
141
142 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
143 {
144         ctx->error = error;
145         smp_wmb();
146         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
147 }
148
149 static struct nfs_page *
150 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
151 {
152         struct nfs_page *req = NULL;
153
154         if (PagePrivate(page))
155                 req = (struct nfs_page *)page_private(page);
156         else if (unlikely(PageSwapCache(page))) {
157                 struct nfs_page *freq, *t;
158
159                 /* Linearly search the commit list for the correct req */
160                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
161                         if (freq->wb_page == page) {
162                                 req = freq;
163                                 break;
164                         }
165                 }
166         }
167
168         if (req)
169                 kref_get(&req->wb_kref);
170
171         return req;
172 }
173
174 static struct nfs_page *nfs_page_find_request(struct page *page)
175 {
176         struct inode *inode = page_file_mapping(page)->host;
177         struct nfs_page *req = NULL;
178
179         spin_lock(&inode->i_lock);
180         req = nfs_page_find_request_locked(NFS_I(inode), page);
181         spin_unlock(&inode->i_lock);
182         return req;
183 }
184
185 /* Adjust the file length if we're writing beyond the end */
186 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
187 {
188         struct inode *inode = page_file_mapping(page)->host;
189         loff_t end, i_size;
190         pgoff_t end_index;
191
192         spin_lock(&inode->i_lock);
193         i_size = i_size_read(inode);
194         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
195         if (i_size > 0 && page_file_index(page) < end_index)
196                 goto out;
197         end = page_file_offset(page) + ((loff_t)offset+count);
198         if (i_size >= end)
199                 goto out;
200         i_size_write(inode, end);
201         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
202 out:
203         spin_unlock(&inode->i_lock);
204 }
205
206 /* A writeback failed: mark the page as bad, and invalidate the page cache */
207 static void nfs_set_pageerror(struct page *page)
208 {
209         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
210 }
211
212 /* We can set the PG_uptodate flag if we see that a write request
213  * covers the full page.
214  */
215 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
216 {
217         if (PageUptodate(page))
218                 return;
219         if (base != 0)
220                 return;
221         if (count != nfs_page_length(page))
222                 return;
223         SetPageUptodate(page);
224 }
225
226 static int wb_priority(struct writeback_control *wbc)
227 {
228         if (wbc->for_reclaim)
229                 return FLUSH_HIGHPRI | FLUSH_STABLE;
230         if (wbc->for_kupdate || wbc->for_background)
231                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
232         return FLUSH_COND_STABLE;
233 }
234
235 /*
236  * NFS congestion control
237  */
238
239 int nfs_congestion_kb;
240
241 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
242 #define NFS_CONGESTION_OFF_THRESH       \
243         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
244
245 static void nfs_set_page_writeback(struct page *page)
246 {
247         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
248         int ret = test_set_page_writeback(page);
249
250         WARN_ON_ONCE(ret != 0);
251
252         if (atomic_long_inc_return(&nfss->writeback) >
253                         NFS_CONGESTION_ON_THRESH) {
254                 set_bdi_congested(&nfss->backing_dev_info,
255                                         BLK_RW_ASYNC);
256         }
257 }
258
259 static void nfs_end_page_writeback(struct page *page)
260 {
261         struct inode *inode = page_file_mapping(page)->host;
262         struct nfs_server *nfss = NFS_SERVER(inode);
263
264         end_page_writeback(page);
265         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
266                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
267 }
268
269 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
270 {
271         struct inode *inode = page_file_mapping(page)->host;
272         struct nfs_page *req;
273         int ret;
274
275         spin_lock(&inode->i_lock);
276         for (;;) {
277                 req = nfs_page_find_request_locked(NFS_I(inode), page);
278                 if (req == NULL)
279                         break;
280                 if (nfs_lock_request(req))
281                         break;
282                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
283                  *       then the call to nfs_lock_request() will always
284                  *       succeed provided that someone hasn't already marked the
285                  *       request as dirty (in which case we don't care).
286                  */
287                 spin_unlock(&inode->i_lock);
288                 if (!nonblock)
289                         ret = nfs_wait_on_request(req);
290                 else
291                         ret = -EAGAIN;
292                 nfs_release_request(req);
293                 if (ret != 0)
294                         return ERR_PTR(ret);
295                 spin_lock(&inode->i_lock);
296         }
297         spin_unlock(&inode->i_lock);
298         return req;
299 }
300
301 /*
302  * Find an associated nfs write request, and prepare to flush it out
303  * May return an error if the user signalled nfs_wait_on_request().
304  */
305 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
306                                 struct page *page, bool nonblock)
307 {
308         struct nfs_page *req;
309         int ret = 0;
310
311         req = nfs_find_and_lock_request(page, nonblock);
312         if (!req)
313                 goto out;
314         ret = PTR_ERR(req);
315         if (IS_ERR(req))
316                 goto out;
317
318         nfs_set_page_writeback(page);
319         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
320
321         ret = 0;
322         if (!nfs_pageio_add_request(pgio, req)) {
323                 nfs_redirty_request(req);
324                 ret = pgio->pg_error;
325         }
326 out:
327         return ret;
328 }
329
330 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
331 {
332         struct inode *inode = page_file_mapping(page)->host;
333         int ret;
334
335         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
336         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
337
338         nfs_pageio_cond_complete(pgio, page_file_index(page));
339         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
340         if (ret == -EAGAIN) {
341                 redirty_page_for_writepage(wbc, page);
342                 ret = 0;
343         }
344         return ret;
345 }
346
347 /*
348  * Write an mmapped page to the server.
349  */
350 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
351 {
352         struct nfs_pageio_descriptor pgio;
353         int err;
354
355         NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio,
356                                                           page->mapping->host,
357                                                           wb_priority(wbc),
358                                                           &nfs_async_write_completion_ops);
359         err = nfs_do_writepage(page, wbc, &pgio);
360         nfs_pageio_complete(&pgio);
361         if (err < 0)
362                 return err;
363         if (pgio.pg_error < 0)
364                 return pgio.pg_error;
365         return 0;
366 }
367
368 int nfs_writepage(struct page *page, struct writeback_control *wbc)
369 {
370         int ret;
371
372         ret = nfs_writepage_locked(page, wbc);
373         unlock_page(page);
374         return ret;
375 }
376
377 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
378 {
379         int ret;
380
381         ret = nfs_do_writepage(page, wbc, data);
382         unlock_page(page);
383         return ret;
384 }
385
386 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
387 {
388         struct inode *inode = mapping->host;
389         unsigned long *bitlock = &NFS_I(inode)->flags;
390         struct nfs_pageio_descriptor pgio;
391         int err;
392
393         /* Stop dirtying of new pages while we sync */
394         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
395                         nfs_wait_bit_killable, TASK_KILLABLE);
396         if (err)
397                 goto out_err;
398
399         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
400
401         NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
402         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
403         nfs_pageio_complete(&pgio);
404
405         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
406         smp_mb__after_clear_bit();
407         wake_up_bit(bitlock, NFS_INO_FLUSHING);
408
409         if (err < 0)
410                 goto out_err;
411         err = pgio.pg_error;
412         if (err < 0)
413                 goto out_err;
414         return 0;
415 out_err:
416         return err;
417 }
418
419 /*
420  * Insert a write request into an inode
421  */
422 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
423 {
424         struct nfs_inode *nfsi = NFS_I(inode);
425
426         /* Lock the request! */
427         nfs_lock_request(req);
428
429         spin_lock(&inode->i_lock);
430         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
431                 inode->i_version++;
432         /*
433          * Swap-space should not get truncated. Hence no need to plug the race
434          * with invalidate/truncate.
435          */
436         if (likely(!PageSwapCache(req->wb_page))) {
437                 set_bit(PG_MAPPED, &req->wb_flags);
438                 SetPagePrivate(req->wb_page);
439                 set_page_private(req->wb_page, (unsigned long)req);
440         }
441         nfsi->npages++;
442         kref_get(&req->wb_kref);
443         spin_unlock(&inode->i_lock);
444 }
445
446 /*
447  * Remove a write request from an inode
448  */
449 static void nfs_inode_remove_request(struct nfs_page *req)
450 {
451         struct inode *inode = req->wb_context->dentry->d_inode;
452         struct nfs_inode *nfsi = NFS_I(inode);
453
454         spin_lock(&inode->i_lock);
455         if (likely(!PageSwapCache(req->wb_page))) {
456                 set_page_private(req->wb_page, 0);
457                 ClearPagePrivate(req->wb_page);
458                 clear_bit(PG_MAPPED, &req->wb_flags);
459         }
460         nfsi->npages--;
461         spin_unlock(&inode->i_lock);
462         nfs_release_request(req);
463 }
464
465 static void
466 nfs_mark_request_dirty(struct nfs_page *req)
467 {
468         __set_page_dirty_nobuffers(req->wb_page);
469 }
470
471 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
472 /**
473  * nfs_request_add_commit_list - add request to a commit list
474  * @req: pointer to a struct nfs_page
475  * @dst: commit list head
476  * @cinfo: holds list lock and accounting info
477  *
478  * This sets the PG_CLEAN bit, updates the cinfo count of
479  * number of outstanding requests requiring a commit as well as
480  * the MM page stats.
481  *
482  * The caller must _not_ hold the cinfo->lock, but must be
483  * holding the nfs_page lock.
484  */
485 void
486 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
487                             struct nfs_commit_info *cinfo)
488 {
489         set_bit(PG_CLEAN, &(req)->wb_flags);
490         spin_lock(cinfo->lock);
491         nfs_list_add_request(req, dst);
492         cinfo->mds->ncommit++;
493         spin_unlock(cinfo->lock);
494         if (!cinfo->dreq) {
495                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
496                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
497                              BDI_RECLAIMABLE);
498                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
499                                    I_DIRTY_DATASYNC);
500         }
501 }
502 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
503
504 /**
505  * nfs_request_remove_commit_list - Remove request from a commit list
506  * @req: pointer to a nfs_page
507  * @cinfo: holds list lock and accounting info
508  *
509  * This clears the PG_CLEAN bit, and updates the cinfo's count of
510  * number of outstanding requests requiring a commit
511  * It does not update the MM page stats.
512  *
513  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
514  */
515 void
516 nfs_request_remove_commit_list(struct nfs_page *req,
517                                struct nfs_commit_info *cinfo)
518 {
519         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
520                 return;
521         nfs_list_remove_request(req);
522         cinfo->mds->ncommit--;
523 }
524 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
525
526 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
527                                       struct inode *inode)
528 {
529         cinfo->lock = &inode->i_lock;
530         cinfo->mds = &NFS_I(inode)->commit_info;
531         cinfo->ds = pnfs_get_ds_info(inode);
532         cinfo->dreq = NULL;
533         cinfo->completion_ops = &nfs_commit_completion_ops;
534 }
535
536 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
537                     struct inode *inode,
538                     struct nfs_direct_req *dreq)
539 {
540         if (dreq)
541                 nfs_init_cinfo_from_dreq(cinfo, dreq);
542         else
543                 nfs_init_cinfo_from_inode(cinfo, inode);
544 }
545 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
546
547 /*
548  * Add a request to the inode's commit list.
549  */
550 void
551 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
552                         struct nfs_commit_info *cinfo)
553 {
554         if (pnfs_mark_request_commit(req, lseg, cinfo))
555                 return;
556         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
557 }
558
559 static void
560 nfs_clear_page_commit(struct page *page)
561 {
562         dec_zone_page_state(page, NR_UNSTABLE_NFS);
563         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
564 }
565
566 static void
567 nfs_clear_request_commit(struct nfs_page *req)
568 {
569         if (test_bit(PG_CLEAN, &req->wb_flags)) {
570                 struct inode *inode = req->wb_context->dentry->d_inode;
571                 struct nfs_commit_info cinfo;
572
573                 nfs_init_cinfo_from_inode(&cinfo, inode);
574                 if (!pnfs_clear_request_commit(req, &cinfo)) {
575                         spin_lock(cinfo.lock);
576                         nfs_request_remove_commit_list(req, &cinfo);
577                         spin_unlock(cinfo.lock);
578                 }
579                 nfs_clear_page_commit(req->wb_page);
580         }
581 }
582
583 static inline
584 int nfs_write_need_commit(struct nfs_write_data *data)
585 {
586         if (data->verf.committed == NFS_DATA_SYNC)
587                 return data->header->lseg == NULL;
588         return data->verf.committed != NFS_FILE_SYNC;
589 }
590
591 #else
592 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
593                                       struct inode *inode)
594 {
595 }
596
597 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
598                     struct inode *inode,
599                     struct nfs_direct_req *dreq)
600 {
601 }
602
603 void
604 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
605                         struct nfs_commit_info *cinfo)
606 {
607 }
608
609 static void
610 nfs_clear_request_commit(struct nfs_page *req)
611 {
612 }
613
614 static inline
615 int nfs_write_need_commit(struct nfs_write_data *data)
616 {
617         return 0;
618 }
619
620 #endif
621
622 static void nfs_write_completion(struct nfs_pgio_header *hdr)
623 {
624         struct nfs_commit_info cinfo;
625         unsigned long bytes = 0;
626
627         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
628                 goto out;
629         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
630         while (!list_empty(&hdr->pages)) {
631                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
632
633                 bytes += req->wb_bytes;
634                 nfs_list_remove_request(req);
635                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
636                     (hdr->good_bytes < bytes)) {
637                         nfs_set_pageerror(req->wb_page);
638                         nfs_context_set_write_error(req->wb_context, hdr->error);
639                         goto remove_req;
640                 }
641                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
642                         nfs_mark_request_dirty(req);
643                         goto next;
644                 }
645                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
646                         memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
647                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
648                         goto next;
649                 }
650 remove_req:
651                 nfs_inode_remove_request(req);
652 next:
653                 nfs_unlock_request(req);
654                 nfs_end_page_writeback(req->wb_page);
655                 nfs_release_request(req);
656         }
657 out:
658         hdr->release(hdr);
659 }
660
661 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
662 static unsigned long
663 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
664 {
665         return cinfo->mds->ncommit;
666 }
667
668 /* cinfo->lock held by caller */
669 int
670 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
671                      struct nfs_commit_info *cinfo, int max)
672 {
673         struct nfs_page *req, *tmp;
674         int ret = 0;
675
676         list_for_each_entry_safe(req, tmp, src, wb_list) {
677                 if (!nfs_lock_request(req))
678                         continue;
679                 kref_get(&req->wb_kref);
680                 if (cond_resched_lock(cinfo->lock))
681                         list_safe_reset_next(req, tmp, wb_list);
682                 nfs_request_remove_commit_list(req, cinfo);
683                 nfs_list_add_request(req, dst);
684                 ret++;
685                 if ((ret == max) && !cinfo->dreq)
686                         break;
687         }
688         return ret;
689 }
690
691 /*
692  * nfs_scan_commit - Scan an inode for commit requests
693  * @inode: NFS inode to scan
694  * @dst: mds destination list
695  * @cinfo: mds and ds lists of reqs ready to commit
696  *
697  * Moves requests from the inode's 'commit' request list.
698  * The requests are *not* checked to ensure that they form a contiguous set.
699  */
700 int
701 nfs_scan_commit(struct inode *inode, struct list_head *dst,
702                 struct nfs_commit_info *cinfo)
703 {
704         int ret = 0;
705
706         spin_lock(cinfo->lock);
707         if (cinfo->mds->ncommit > 0) {
708                 const int max = INT_MAX;
709
710                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
711                                            cinfo, max);
712                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
713         }
714         spin_unlock(cinfo->lock);
715         return ret;
716 }
717
718 #else
719 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
720 {
721         return 0;
722 }
723
724 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
725                     struct nfs_commit_info *cinfo)
726 {
727         return 0;
728 }
729 #endif
730
731 /*
732  * Search for an existing write request, and attempt to update
733  * it to reflect a new dirty region on a given page.
734  *
735  * If the attempt fails, then the existing request is flushed out
736  * to disk.
737  */
738 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
739                 struct page *page,
740                 unsigned int offset,
741                 unsigned int bytes)
742 {
743         struct nfs_page *req;
744         unsigned int rqend;
745         unsigned int end;
746         int error;
747
748         if (!PagePrivate(page))
749                 return NULL;
750
751         end = offset + bytes;
752         spin_lock(&inode->i_lock);
753
754         for (;;) {
755                 req = nfs_page_find_request_locked(NFS_I(inode), page);
756                 if (req == NULL)
757                         goto out_unlock;
758
759                 rqend = req->wb_offset + req->wb_bytes;
760                 /*
761                  * Tell the caller to flush out the request if
762                  * the offsets are non-contiguous.
763                  * Note: nfs_flush_incompatible() will already
764                  * have flushed out requests having wrong owners.
765                  */
766                 if (offset > rqend
767                     || end < req->wb_offset)
768                         goto out_flushme;
769
770                 if (nfs_lock_request(req))
771                         break;
772
773                 /* The request is locked, so wait and then retry */
774                 spin_unlock(&inode->i_lock);
775                 error = nfs_wait_on_request(req);
776                 nfs_release_request(req);
777                 if (error != 0)
778                         goto out_err;
779                 spin_lock(&inode->i_lock);
780         }
781
782         /* Okay, the request matches. Update the region */
783         if (offset < req->wb_offset) {
784                 req->wb_offset = offset;
785                 req->wb_pgbase = offset;
786         }
787         if (end > rqend)
788                 req->wb_bytes = end - req->wb_offset;
789         else
790                 req->wb_bytes = rqend - req->wb_offset;
791 out_unlock:
792         spin_unlock(&inode->i_lock);
793         if (req)
794                 nfs_clear_request_commit(req);
795         return req;
796 out_flushme:
797         spin_unlock(&inode->i_lock);
798         nfs_release_request(req);
799         error = nfs_wb_page(inode, page);
800 out_err:
801         return ERR_PTR(error);
802 }
803
804 /*
805  * Try to update an existing write request, or create one if there is none.
806  *
807  * Note: Should always be called with the Page Lock held to prevent races
808  * if we have to add a new request. Also assumes that the caller has
809  * already called nfs_flush_incompatible() if necessary.
810  */
811 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
812                 struct page *page, unsigned int offset, unsigned int bytes)
813 {
814         struct inode *inode = page_file_mapping(page)->host;
815         struct nfs_page *req;
816
817         req = nfs_try_to_update_request(inode, page, offset, bytes);
818         if (req != NULL)
819                 goto out;
820         req = nfs_create_request(ctx, inode, page, offset, bytes);
821         if (IS_ERR(req))
822                 goto out;
823         nfs_inode_add_request(inode, req);
824 out:
825         return req;
826 }
827
828 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
829                 unsigned int offset, unsigned int count)
830 {
831         struct nfs_page *req;
832
833         req = nfs_setup_write_request(ctx, page, offset, count);
834         if (IS_ERR(req))
835                 return PTR_ERR(req);
836         /* Update file length */
837         nfs_grow_file(page, offset, count);
838         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
839         nfs_mark_request_dirty(req);
840         nfs_unlock_and_release_request(req);
841         return 0;
842 }
843
844 int nfs_flush_incompatible(struct file *file, struct page *page)
845 {
846         struct nfs_open_context *ctx = nfs_file_open_context(file);
847         struct nfs_lock_context *l_ctx;
848         struct nfs_page *req;
849         int do_flush, status;
850         /*
851          * Look for a request corresponding to this page. If there
852          * is one, and it belongs to another file, we flush it out
853          * before we try to copy anything into the page. Do this
854          * due to the lack of an ACCESS-type call in NFSv2.
855          * Also do the same if we find a request from an existing
856          * dropped page.
857          */
858         do {
859                 req = nfs_page_find_request(page);
860                 if (req == NULL)
861                         return 0;
862                 l_ctx = req->wb_lock_context;
863                 do_flush = req->wb_page != page || req->wb_context != ctx;
864                 if (l_ctx) {
865                         do_flush |= l_ctx->lockowner.l_owner != current->files
866                                 || l_ctx->lockowner.l_pid != current->tgid;
867                 }
868                 nfs_release_request(req);
869                 if (!do_flush)
870                         return 0;
871                 status = nfs_wb_page(page_file_mapping(page)->host, page);
872         } while (status == 0);
873         return status;
874 }
875
876 /*
877  * If the page cache is marked as unsafe or invalid, then we can't rely on
878  * the PageUptodate() flag. In this case, we will need to turn off
879  * write optimisations that depend on the page contents being correct.
880  */
881 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
882 {
883         if (nfs_have_delegated_attributes(inode))
884                 goto out;
885         if (NFS_I(inode)->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
886                 return false;
887 out:
888         return PageUptodate(page) != 0;
889 }
890
891 /*
892  * Update and possibly write a cached page of an NFS file.
893  *
894  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
895  * things with a page scheduled for an RPC call (e.g. invalidate it).
896  */
897 int nfs_updatepage(struct file *file, struct page *page,
898                 unsigned int offset, unsigned int count)
899 {
900         struct nfs_open_context *ctx = nfs_file_open_context(file);
901         struct inode    *inode = page_file_mapping(page)->host;
902         int             status = 0;
903
904         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
905
906         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
907                 file->f_path.dentry->d_parent->d_name.name,
908                 file->f_path.dentry->d_name.name, count,
909                 (long long)(page_file_offset(page) + offset));
910
911         /* If we're not using byte range locks, and we know the page
912          * is up to date, it may be more efficient to extend the write
913          * to cover the entire page in order to avoid fragmentation
914          * inefficiencies.
915          */
916         if (nfs_write_pageuptodate(page, inode) &&
917                         inode->i_flock == NULL &&
918                         !(file->f_flags & O_DSYNC)) {
919                 count = max(count + offset, nfs_page_length(page));
920                 offset = 0;
921         }
922
923         status = nfs_writepage_setup(ctx, page, offset, count);
924         if (status < 0)
925                 nfs_set_pageerror(page);
926         else
927                 __set_page_dirty_nobuffers(page);
928
929         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
930                         status, (long long)i_size_read(inode));
931         return status;
932 }
933
934 static int flush_task_priority(int how)
935 {
936         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
937                 case FLUSH_HIGHPRI:
938                         return RPC_PRIORITY_HIGH;
939                 case FLUSH_LOWPRI:
940                         return RPC_PRIORITY_LOW;
941         }
942         return RPC_PRIORITY_NORMAL;
943 }
944
945 int nfs_initiate_write(struct rpc_clnt *clnt,
946                        struct nfs_write_data *data,
947                        const struct rpc_call_ops *call_ops,
948                        int how, int flags)
949 {
950         struct inode *inode = data->header->inode;
951         int priority = flush_task_priority(how);
952         struct rpc_task *task;
953         struct rpc_message msg = {
954                 .rpc_argp = &data->args,
955                 .rpc_resp = &data->res,
956                 .rpc_cred = data->header->cred,
957         };
958         struct rpc_task_setup task_setup_data = {
959                 .rpc_client = clnt,
960                 .task = &data->task,
961                 .rpc_message = &msg,
962                 .callback_ops = call_ops,
963                 .callback_data = data,
964                 .workqueue = nfsiod_workqueue,
965                 .flags = RPC_TASK_ASYNC | flags,
966                 .priority = priority,
967         };
968         int ret = 0;
969
970         /* Set up the initial task struct.  */
971         NFS_PROTO(inode)->write_setup(data, &msg);
972
973         dprintk("NFS: %5u initiated write call "
974                 "(req %s/%lld, %u bytes @ offset %llu)\n",
975                 data->task.tk_pid,
976                 inode->i_sb->s_id,
977                 (long long)NFS_FILEID(inode),
978                 data->args.count,
979                 (unsigned long long)data->args.offset);
980
981         task = rpc_run_task(&task_setup_data);
982         if (IS_ERR(task)) {
983                 ret = PTR_ERR(task);
984                 goto out;
985         }
986         if (how & FLUSH_SYNC) {
987                 ret = rpc_wait_for_completion_task(task);
988                 if (ret == 0)
989                         ret = task->tk_status;
990         }
991         rpc_put_task(task);
992 out:
993         return ret;
994 }
995 EXPORT_SYMBOL_GPL(nfs_initiate_write);
996
997 /*
998  * Set up the argument/result storage required for the RPC call.
999  */
1000 static void nfs_write_rpcsetup(struct nfs_write_data *data,
1001                 unsigned int count, unsigned int offset,
1002                 int how, struct nfs_commit_info *cinfo)
1003 {
1004         struct nfs_page *req = data->header->req;
1005
1006         /* Set up the RPC argument and reply structs
1007          * NB: take care not to mess about with data->commit et al. */
1008
1009         data->args.fh     = NFS_FH(data->header->inode);
1010         data->args.offset = req_offset(req) + offset;
1011         /* pnfs_set_layoutcommit needs this */
1012         data->mds_offset = data->args.offset;
1013         data->args.pgbase = req->wb_pgbase + offset;
1014         data->args.pages  = data->pages.pagevec;
1015         data->args.count  = count;
1016         data->args.context = get_nfs_open_context(req->wb_context);
1017         data->args.lock_context = req->wb_lock_context;
1018         data->args.stable  = NFS_UNSTABLE;
1019         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1020         case 0:
1021                 break;
1022         case FLUSH_COND_STABLE:
1023                 if (nfs_reqs_to_commit(cinfo))
1024                         break;
1025         default:
1026                 data->args.stable = NFS_FILE_SYNC;
1027         }
1028
1029         data->res.fattr   = &data->fattr;
1030         data->res.count   = count;
1031         data->res.verf    = &data->verf;
1032         nfs_fattr_init(&data->fattr);
1033 }
1034
1035 static int nfs_do_write(struct nfs_write_data *data,
1036                 const struct rpc_call_ops *call_ops,
1037                 int how)
1038 {
1039         struct inode *inode = data->header->inode;
1040
1041         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1042 }
1043
1044 static int nfs_do_multiple_writes(struct list_head *head,
1045                 const struct rpc_call_ops *call_ops,
1046                 int how)
1047 {
1048         struct nfs_write_data *data;
1049         int ret = 0;
1050
1051         while (!list_empty(head)) {
1052                 int ret2;
1053
1054                 data = list_first_entry(head, struct nfs_write_data, list);
1055                 list_del_init(&data->list);
1056                 
1057                 ret2 = nfs_do_write(data, call_ops, how);
1058                  if (ret == 0)
1059                          ret = ret2;
1060         }
1061         return ret;
1062 }
1063
1064 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1065  * call this on each, which will prepare them to be retried on next
1066  * writeback using standard nfs.
1067  */
1068 static void nfs_redirty_request(struct nfs_page *req)
1069 {
1070         nfs_mark_request_dirty(req);
1071         nfs_unlock_request(req);
1072         nfs_end_page_writeback(req->wb_page);
1073         nfs_release_request(req);
1074 }
1075
1076 static void nfs_async_write_error(struct list_head *head)
1077 {
1078         struct nfs_page *req;
1079
1080         while (!list_empty(head)) {
1081                 req = nfs_list_entry(head->next);
1082                 nfs_list_remove_request(req);
1083                 nfs_redirty_request(req);
1084         }
1085 }
1086
1087 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1088         .error_cleanup = nfs_async_write_error,
1089         .completion = nfs_write_completion,
1090 };
1091
1092 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1093                 struct nfs_pgio_header *hdr)
1094 {
1095         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1096         while (!list_empty(&hdr->rpc_list)) {
1097                 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1098                                 struct nfs_write_data, list);
1099                 list_del(&data->list);
1100                 nfs_writedata_release(data);
1101         }
1102         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1103 }
1104
1105 /*
1106  * Generate multiple small requests to write out a single
1107  * contiguous dirty area on one page.
1108  */
1109 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1110                            struct nfs_pgio_header *hdr)
1111 {
1112         struct nfs_page *req = hdr->req;
1113         struct page *page = req->wb_page;
1114         struct nfs_write_data *data;
1115         size_t wsize = desc->pg_bsize, nbytes;
1116         unsigned int offset;
1117         int requests = 0;
1118         struct nfs_commit_info cinfo;
1119
1120         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1121
1122         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1123             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1124              desc->pg_count > wsize))
1125                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1126
1127
1128         offset = 0;
1129         nbytes = desc->pg_count;
1130         do {
1131                 size_t len = min(nbytes, wsize);
1132
1133                 data = nfs_writedata_alloc(hdr, 1);
1134                 if (!data) {
1135                         nfs_flush_error(desc, hdr);
1136                         return -ENOMEM;
1137                 }
1138                 data->pages.pagevec[0] = page;
1139                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1140                 list_add(&data->list, &hdr->rpc_list);
1141                 requests++;
1142                 nbytes -= len;
1143                 offset += len;
1144         } while (nbytes != 0);
1145         nfs_list_remove_request(req);
1146         nfs_list_add_request(req, &hdr->pages);
1147         desc->pg_rpc_callops = &nfs_write_common_ops;
1148         return 0;
1149 }
1150
1151 /*
1152  * Create an RPC task for the given write request and kick it.
1153  * The page must have been locked by the caller.
1154  *
1155  * It may happen that the page we're passed is not marked dirty.
1156  * This is the case if nfs_updatepage detects a conflicting request
1157  * that has been written but not committed.
1158  */
1159 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1160                          struct nfs_pgio_header *hdr)
1161 {
1162         struct nfs_page         *req;
1163         struct page             **pages;
1164         struct nfs_write_data   *data;
1165         struct list_head *head = &desc->pg_list;
1166         struct nfs_commit_info cinfo;
1167
1168         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1169                                                            desc->pg_count));
1170         if (!data) {
1171                 nfs_flush_error(desc, hdr);
1172                 return -ENOMEM;
1173         }
1174
1175         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1176         pages = data->pages.pagevec;
1177         while (!list_empty(head)) {
1178                 req = nfs_list_entry(head->next);
1179                 nfs_list_remove_request(req);
1180                 nfs_list_add_request(req, &hdr->pages);
1181                 *pages++ = req->wb_page;
1182         }
1183
1184         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1185             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1186                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1187
1188         /* Set up the argument struct */
1189         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1190         list_add(&data->list, &hdr->rpc_list);
1191         desc->pg_rpc_callops = &nfs_write_common_ops;
1192         return 0;
1193 }
1194
1195 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1196                       struct nfs_pgio_header *hdr)
1197 {
1198         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1199                 return nfs_flush_multi(desc, hdr);
1200         return nfs_flush_one(desc, hdr);
1201 }
1202 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1203
1204 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1205 {
1206         struct nfs_write_header *whdr;
1207         struct nfs_pgio_header *hdr;
1208         int ret;
1209
1210         whdr = nfs_writehdr_alloc();
1211         if (!whdr) {
1212                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1213                 return -ENOMEM;
1214         }
1215         hdr = &whdr->header;
1216         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1217         atomic_inc(&hdr->refcnt);
1218         ret = nfs_generic_flush(desc, hdr);
1219         if (ret == 0)
1220                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1221                                              desc->pg_rpc_callops,
1222                                              desc->pg_ioflags);
1223         if (atomic_dec_and_test(&hdr->refcnt))
1224                 hdr->completion_ops->completion(hdr);
1225         return ret;
1226 }
1227
1228 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1229         .pg_test = nfs_generic_pg_test,
1230         .pg_doio = nfs_generic_pg_writepages,
1231 };
1232
1233 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1234                                struct inode *inode, int ioflags,
1235                                const struct nfs_pgio_completion_ops *compl_ops)
1236 {
1237         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1238                                 NFS_SERVER(inode)->wsize, ioflags);
1239 }
1240 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1241
1242 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1243 {
1244         pgio->pg_ops = &nfs_pageio_write_ops;
1245         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1246 }
1247 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1248
1249
1250 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1251 {
1252         struct nfs_write_data *data = calldata;
1253         NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1254         if (unlikely(test_bit(NFS_CONTEXT_BAD, &data->args.context->flags)))
1255                 rpc_exit(task, -EIO);
1256 }
1257
1258 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1259 {
1260         struct nfs_commit_data *data = calldata;
1261
1262         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1263 }
1264
1265 /*
1266  * Handle a write reply that flushes a whole page.
1267  *
1268  * FIXME: There is an inherent race with invalidate_inode_pages and
1269  *        writebacks since the page->count is kept > 1 for as long
1270  *        as the page has a write request pending.
1271  */
1272 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1273 {
1274         struct nfs_write_data   *data = calldata;
1275
1276         nfs_writeback_done(task, data);
1277 }
1278
1279 static void nfs_writeback_release_common(void *calldata)
1280 {
1281         struct nfs_write_data   *data = calldata;
1282         struct nfs_pgio_header *hdr = data->header;
1283         int status = data->task.tk_status;
1284
1285         if ((status >= 0) && nfs_write_need_commit(data)) {
1286                 spin_lock(&hdr->lock);
1287                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1288                         ; /* Do nothing */
1289                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1290                         memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1291                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1292                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1293                 spin_unlock(&hdr->lock);
1294         }
1295         nfs_writedata_release(data);
1296 }
1297
1298 static const struct rpc_call_ops nfs_write_common_ops = {
1299         .rpc_call_prepare = nfs_write_prepare,
1300         .rpc_call_done = nfs_writeback_done_common,
1301         .rpc_release = nfs_writeback_release_common,
1302 };
1303
1304
1305 /*
1306  * This function is called when the WRITE call is complete.
1307  */
1308 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1309 {
1310         struct nfs_writeargs    *argp = &data->args;
1311         struct nfs_writeres     *resp = &data->res;
1312         struct inode            *inode = data->header->inode;
1313         int status;
1314
1315         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1316                 task->tk_pid, task->tk_status);
1317
1318         /*
1319          * ->write_done will attempt to use post-op attributes to detect
1320          * conflicting writes by other clients.  A strict interpretation
1321          * of close-to-open would allow us to continue caching even if
1322          * another writer had changed the file, but some applications
1323          * depend on tighter cache coherency when writing.
1324          */
1325         status = NFS_PROTO(inode)->write_done(task, data);
1326         if (status != 0)
1327                 return;
1328         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1329
1330 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1331         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1332                 /* We tried a write call, but the server did not
1333                  * commit data to stable storage even though we
1334                  * requested it.
1335                  * Note: There is a known bug in Tru64 < 5.0 in which
1336                  *       the server reports NFS_DATA_SYNC, but performs
1337                  *       NFS_FILE_SYNC. We therefore implement this checking
1338                  *       as a dprintk() in order to avoid filling syslog.
1339                  */
1340                 static unsigned long    complain;
1341
1342                 /* Note this will print the MDS for a DS write */
1343                 if (time_before(complain, jiffies)) {
1344                         dprintk("NFS:       faulty NFS server %s:"
1345                                 " (committed = %d) != (stable = %d)\n",
1346                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1347                                 resp->verf->committed, argp->stable);
1348                         complain = jiffies + 300 * HZ;
1349                 }
1350         }
1351 #endif
1352         if (task->tk_status < 0)
1353                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1354         else if (resp->count < argp->count) {
1355                 static unsigned long    complain;
1356
1357                 /* This a short write! */
1358                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1359
1360                 /* Has the server at least made some progress? */
1361                 if (resp->count == 0) {
1362                         if (time_before(complain, jiffies)) {
1363                                 printk(KERN_WARNING
1364                                        "NFS: Server wrote zero bytes, expected %u.\n",
1365                                        argp->count);
1366                                 complain = jiffies + 300 * HZ;
1367                         }
1368                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1369                         task->tk_status = -EIO;
1370                         return;
1371                 }
1372                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1373                 if (resp->verf->committed != NFS_UNSTABLE) {
1374                         /* Resend from where the server left off */
1375                         data->mds_offset += resp->count;
1376                         argp->offset += resp->count;
1377                         argp->pgbase += resp->count;
1378                         argp->count -= resp->count;
1379                 } else {
1380                         /* Resend as a stable write in order to avoid
1381                          * headaches in the case of a server crash.
1382                          */
1383                         argp->stable = NFS_FILE_SYNC;
1384                 }
1385                 rpc_restart_call_prepare(task);
1386         }
1387 }
1388
1389
1390 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1391 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1392 {
1393         int ret;
1394
1395         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1396                 return 1;
1397         if (!may_wait)
1398                 return 0;
1399         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1400                                 NFS_INO_COMMIT,
1401                                 nfs_wait_bit_killable,
1402                                 TASK_KILLABLE);
1403         return (ret < 0) ? ret : 1;
1404 }
1405
1406 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1407 {
1408         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1409         smp_mb__after_clear_bit();
1410         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1411 }
1412
1413 void nfs_commitdata_release(struct nfs_commit_data *data)
1414 {
1415         put_nfs_open_context(data->context);
1416         nfs_commit_free(data);
1417 }
1418 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1419
1420 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1421                         const struct rpc_call_ops *call_ops,
1422                         int how, int flags)
1423 {
1424         struct rpc_task *task;
1425         int priority = flush_task_priority(how);
1426         struct rpc_message msg = {
1427                 .rpc_argp = &data->args,
1428                 .rpc_resp = &data->res,
1429                 .rpc_cred = data->cred,
1430         };
1431         struct rpc_task_setup task_setup_data = {
1432                 .task = &data->task,
1433                 .rpc_client = clnt,
1434                 .rpc_message = &msg,
1435                 .callback_ops = call_ops,
1436                 .callback_data = data,
1437                 .workqueue = nfsiod_workqueue,
1438                 .flags = RPC_TASK_ASYNC | flags,
1439                 .priority = priority,
1440         };
1441         /* Set up the initial task struct.  */
1442         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1443
1444         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1445
1446         task = rpc_run_task(&task_setup_data);
1447         if (IS_ERR(task))
1448                 return PTR_ERR(task);
1449         if (how & FLUSH_SYNC)
1450                 rpc_wait_for_completion_task(task);
1451         rpc_put_task(task);
1452         return 0;
1453 }
1454 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1455
1456 /*
1457  * Set up the argument/result storage required for the RPC call.
1458  */
1459 void nfs_init_commit(struct nfs_commit_data *data,
1460                      struct list_head *head,
1461                      struct pnfs_layout_segment *lseg,
1462                      struct nfs_commit_info *cinfo)
1463 {
1464         struct nfs_page *first = nfs_list_entry(head->next);
1465         struct inode *inode = first->wb_context->dentry->d_inode;
1466
1467         /* Set up the RPC argument and reply structs
1468          * NB: take care not to mess about with data->commit et al. */
1469
1470         list_splice_init(head, &data->pages);
1471
1472         data->inode       = inode;
1473         data->cred        = first->wb_context->cred;
1474         data->lseg        = lseg; /* reference transferred */
1475         data->mds_ops     = &nfs_commit_ops;
1476         data->completion_ops = cinfo->completion_ops;
1477         data->dreq        = cinfo->dreq;
1478
1479         data->args.fh     = NFS_FH(data->inode);
1480         /* Note: we always request a commit of the entire inode */
1481         data->args.offset = 0;
1482         data->args.count  = 0;
1483         data->context     = get_nfs_open_context(first->wb_context);
1484         data->res.fattr   = &data->fattr;
1485         data->res.verf    = &data->verf;
1486         nfs_fattr_init(&data->fattr);
1487 }
1488 EXPORT_SYMBOL_GPL(nfs_init_commit);
1489
1490 void nfs_retry_commit(struct list_head *page_list,
1491                       struct pnfs_layout_segment *lseg,
1492                       struct nfs_commit_info *cinfo)
1493 {
1494         struct nfs_page *req;
1495
1496         while (!list_empty(page_list)) {
1497                 req = nfs_list_entry(page_list->next);
1498                 nfs_list_remove_request(req);
1499                 nfs_mark_request_commit(req, lseg, cinfo);
1500                 if (!cinfo->dreq) {
1501                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1502                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1503                                      BDI_RECLAIMABLE);
1504                 }
1505                 nfs_unlock_and_release_request(req);
1506         }
1507 }
1508 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1509
1510 /*
1511  * Commit dirty pages
1512  */
1513 static int
1514 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1515                 struct nfs_commit_info *cinfo)
1516 {
1517         struct nfs_commit_data  *data;
1518
1519         data = nfs_commitdata_alloc();
1520
1521         if (!data)
1522                 goto out_bad;
1523
1524         /* Set up the argument struct */
1525         nfs_init_commit(data, head, NULL, cinfo);
1526         atomic_inc(&cinfo->mds->rpcs_out);
1527         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1528                                    how, 0);
1529  out_bad:
1530         nfs_retry_commit(head, NULL, cinfo);
1531         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1532         return -ENOMEM;
1533 }
1534
1535 /*
1536  * COMMIT call returned
1537  */
1538 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1539 {
1540         struct nfs_commit_data  *data = calldata;
1541
1542         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1543                                 task->tk_pid, task->tk_status);
1544
1545         /* Call the NFS version-specific code */
1546         NFS_PROTO(data->inode)->commit_done(task, data);
1547 }
1548
1549 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1550 {
1551         struct nfs_page *req;
1552         int status = data->task.tk_status;
1553         struct nfs_commit_info cinfo;
1554
1555         while (!list_empty(&data->pages)) {
1556                 req = nfs_list_entry(data->pages.next);
1557                 nfs_list_remove_request(req);
1558                 nfs_clear_page_commit(req->wb_page);
1559
1560                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1561                         req->wb_context->dentry->d_sb->s_id,
1562                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1563                         req->wb_bytes,
1564                         (long long)req_offset(req));
1565                 if (status < 0) {
1566                         nfs_context_set_write_error(req->wb_context, status);
1567                         nfs_inode_remove_request(req);
1568                         dprintk(", error = %d\n", status);
1569                         goto next;
1570                 }
1571
1572                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1573                  * returned by the server against all stored verfs. */
1574                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1575                         /* We have a match */
1576                         nfs_inode_remove_request(req);
1577                         dprintk(" OK\n");
1578                         goto next;
1579                 }
1580                 /* We have a mismatch. Write the page again */
1581                 dprintk(" mismatch\n");
1582                 nfs_mark_request_dirty(req);
1583                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1584         next:
1585                 nfs_unlock_and_release_request(req);
1586         }
1587         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1588         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1589                 nfs_commit_clear_lock(NFS_I(data->inode));
1590 }
1591
1592 static void nfs_commit_release(void *calldata)
1593 {
1594         struct nfs_commit_data *data = calldata;
1595
1596         data->completion_ops->completion(data);
1597         nfs_commitdata_release(calldata);
1598 }
1599
1600 static const struct rpc_call_ops nfs_commit_ops = {
1601         .rpc_call_prepare = nfs_commit_prepare,
1602         .rpc_call_done = nfs_commit_done,
1603         .rpc_release = nfs_commit_release,
1604 };
1605
1606 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1607         .completion = nfs_commit_release_pages,
1608         .error_cleanup = nfs_commit_clear_lock,
1609 };
1610
1611 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1612                             int how, struct nfs_commit_info *cinfo)
1613 {
1614         int status;
1615
1616         status = pnfs_commit_list(inode, head, how, cinfo);
1617         if (status == PNFS_NOT_ATTEMPTED)
1618                 status = nfs_commit_list(inode, head, how, cinfo);
1619         return status;
1620 }
1621
1622 int nfs_commit_inode(struct inode *inode, int how)
1623 {
1624         LIST_HEAD(head);
1625         struct nfs_commit_info cinfo;
1626         int may_wait = how & FLUSH_SYNC;
1627         int res;
1628
1629         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1630         if (res <= 0)
1631                 goto out_mark_dirty;
1632         nfs_init_cinfo_from_inode(&cinfo, inode);
1633         res = nfs_scan_commit(inode, &head, &cinfo);
1634         if (res) {
1635                 int error;
1636
1637                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1638                 if (error < 0)
1639                         return error;
1640                 if (!may_wait)
1641                         goto out_mark_dirty;
1642                 error = wait_on_bit(&NFS_I(inode)->flags,
1643                                 NFS_INO_COMMIT,
1644                                 nfs_wait_bit_killable,
1645                                 TASK_KILLABLE);
1646                 if (error < 0)
1647                         return error;
1648         } else
1649                 nfs_commit_clear_lock(NFS_I(inode));
1650         return res;
1651         /* Note: If we exit without ensuring that the commit is complete,
1652          * we must mark the inode as dirty. Otherwise, future calls to
1653          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1654          * that the data is on the disk.
1655          */
1656 out_mark_dirty:
1657         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1658         return res;
1659 }
1660
1661 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1662 {
1663         struct nfs_inode *nfsi = NFS_I(inode);
1664         int flags = FLUSH_SYNC;
1665         int ret = 0;
1666
1667         /* no commits means nothing needs to be done */
1668         if (!nfsi->commit_info.ncommit)
1669                 return ret;
1670
1671         if (wbc->sync_mode == WB_SYNC_NONE) {
1672                 /* Don't commit yet if this is a non-blocking flush and there
1673                  * are a lot of outstanding writes for this mapping.
1674                  */
1675                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1676                         goto out_mark_dirty;
1677
1678                 /* don't wait for the COMMIT response */
1679                 flags = 0;
1680         }
1681
1682         ret = nfs_commit_inode(inode, flags);
1683         if (ret >= 0) {
1684                 if (wbc->sync_mode == WB_SYNC_NONE) {
1685                         if (ret < wbc->nr_to_write)
1686                                 wbc->nr_to_write -= ret;
1687                         else
1688                                 wbc->nr_to_write = 0;
1689                 }
1690                 return 0;
1691         }
1692 out_mark_dirty:
1693         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1694         return ret;
1695 }
1696 #else
1697 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1698 {
1699         return 0;
1700 }
1701 #endif
1702
1703 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1704 {
1705         return nfs_commit_unstable_pages(inode, wbc);
1706 }
1707 EXPORT_SYMBOL_GPL(nfs_write_inode);
1708
1709 /*
1710  * flush the inode to disk.
1711  */
1712 int nfs_wb_all(struct inode *inode)
1713 {
1714         struct writeback_control wbc = {
1715                 .sync_mode = WB_SYNC_ALL,
1716                 .nr_to_write = LONG_MAX,
1717                 .range_start = 0,
1718                 .range_end = LLONG_MAX,
1719         };
1720
1721         return sync_inode(inode, &wbc);
1722 }
1723 EXPORT_SYMBOL_GPL(nfs_wb_all);
1724
1725 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1726 {
1727         struct nfs_page *req;
1728         int ret = 0;
1729
1730         for (;;) {
1731                 wait_on_page_writeback(page);
1732                 req = nfs_page_find_request(page);
1733                 if (req == NULL)
1734                         break;
1735                 if (nfs_lock_request(req)) {
1736                         nfs_clear_request_commit(req);
1737                         nfs_inode_remove_request(req);
1738                         /*
1739                          * In case nfs_inode_remove_request has marked the
1740                          * page as being dirty
1741                          */
1742                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1743                         nfs_unlock_and_release_request(req);
1744                         break;
1745                 }
1746                 ret = nfs_wait_on_request(req);
1747                 nfs_release_request(req);
1748                 if (ret < 0)
1749                         break;
1750         }
1751         return ret;
1752 }
1753
1754 /*
1755  * Write back all requests on one page - we do this before reading it.
1756  */
1757 int nfs_wb_page(struct inode *inode, struct page *page)
1758 {
1759         loff_t range_start = page_file_offset(page);
1760         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1761         struct writeback_control wbc = {
1762                 .sync_mode = WB_SYNC_ALL,
1763                 .nr_to_write = 0,
1764                 .range_start = range_start,
1765                 .range_end = range_end,
1766         };
1767         int ret;
1768
1769         for (;;) {
1770                 wait_on_page_writeback(page);
1771                 if (clear_page_dirty_for_io(page)) {
1772                         ret = nfs_writepage_locked(page, &wbc);
1773                         if (ret < 0)
1774                                 goto out_error;
1775                         continue;
1776                 }
1777                 if (!PagePrivate(page))
1778                         break;
1779                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1780                 if (ret < 0)
1781                         goto out_error;
1782         }
1783         return 0;
1784 out_error:
1785         return ret;
1786 }
1787
1788 #ifdef CONFIG_MIGRATION
1789 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1790                 struct page *page, enum migrate_mode mode)
1791 {
1792         /*
1793          * If PagePrivate is set, then the page is currently associated with
1794          * an in-progress read or write request. Don't try to migrate it.
1795          *
1796          * FIXME: we could do this in principle, but we'll need a way to ensure
1797          *        that we can safely release the inode reference while holding
1798          *        the page lock.
1799          */
1800         if (PagePrivate(page))
1801                 return -EBUSY;
1802
1803         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1804                 return -EBUSY;
1805
1806         return migrate_page(mapping, newpage, page, mode);
1807 }
1808 #endif
1809
1810 int __init nfs_init_writepagecache(void)
1811 {
1812         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1813                                              sizeof(struct nfs_write_header),
1814                                              0, SLAB_HWCACHE_ALIGN,
1815                                              NULL);
1816         if (nfs_wdata_cachep == NULL)
1817                 return -ENOMEM;
1818
1819         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1820                                                      nfs_wdata_cachep);
1821         if (nfs_wdata_mempool == NULL)
1822                 goto out_destroy_write_cache;
1823
1824         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1825                                              sizeof(struct nfs_commit_data),
1826                                              0, SLAB_HWCACHE_ALIGN,
1827                                              NULL);
1828         if (nfs_cdata_cachep == NULL)
1829                 goto out_destroy_write_mempool;
1830
1831         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1832                                                       nfs_cdata_cachep);
1833         if (nfs_commit_mempool == NULL)
1834                 goto out_destroy_commit_cache;
1835
1836         /*
1837          * NFS congestion size, scale with available memory.
1838          *
1839          *  64MB:    8192k
1840          * 128MB:   11585k
1841          * 256MB:   16384k
1842          * 512MB:   23170k
1843          *   1GB:   32768k
1844          *   2GB:   46340k
1845          *   4GB:   65536k
1846          *   8GB:   92681k
1847          *  16GB:  131072k
1848          *
1849          * This allows larger machines to have larger/more transfers.
1850          * Limit the default to 256M
1851          */
1852         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1853         if (nfs_congestion_kb > 256*1024)
1854                 nfs_congestion_kb = 256*1024;
1855
1856         return 0;
1857
1858 out_destroy_commit_cache:
1859         kmem_cache_destroy(nfs_cdata_cachep);
1860 out_destroy_write_mempool:
1861         mempool_destroy(nfs_wdata_mempool);
1862 out_destroy_write_cache:
1863         kmem_cache_destroy(nfs_wdata_cachep);
1864         return -ENOMEM;
1865 }
1866
1867 void nfs_destroy_writepagecache(void)
1868 {
1869         mempool_destroy(nfs_commit_mempool);
1870         kmem_cache_destroy(nfs_cdata_cachep);
1871         mempool_destroy(nfs_wdata_mempool);
1872         kmem_cache_destroy(nfs_wdata_cachep);
1873 }
1874