Loading...
1/*
2 * linux/fs/nfs/write.c
3 *
4 * Write file data over NFS.
5 *
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7 */
8
9#include <linux/types.h>
10#include <linux/slab.h>
11#include <linux/mm.h>
12#include <linux/pagemap.h>
13#include <linux/file.h>
14#include <linux/writeback.h>
15#include <linux/swap.h>
16#include <linux/migrate.h>
17
18#include <linux/sunrpc/clnt.h>
19#include <linux/nfs_fs.h>
20#include <linux/nfs_mount.h>
21#include <linux/nfs_page.h>
22#include <linux/backing-dev.h>
23
24#include <asm/uaccess.h>
25
26#include "delegation.h"
27#include "internal.h"
28#include "iostat.h"
29#include "nfs4_fs.h"
30#include "fscache.h"
31#include "pnfs.h"
32
33#define NFSDBG_FACILITY NFSDBG_PAGECACHE
34
35#define MIN_POOL_WRITE (32)
36#define MIN_POOL_COMMIT (4)
37
38/*
39 * Local function declarations
40 */
41static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
42 struct inode *inode, int ioflags);
43static void nfs_redirty_request(struct nfs_page *req);
44static const struct rpc_call_ops nfs_write_partial_ops;
45static const struct rpc_call_ops nfs_write_full_ops;
46static const struct rpc_call_ops nfs_commit_ops;
47
48static struct kmem_cache *nfs_wdata_cachep;
49static mempool_t *nfs_wdata_mempool;
50static mempool_t *nfs_commit_mempool;
51
52struct nfs_write_data *nfs_commitdata_alloc(void)
53{
54 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
55
56 if (p) {
57 memset(p, 0, sizeof(*p));
58 INIT_LIST_HEAD(&p->pages);
59 }
60 return p;
61}
62EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
63
64void nfs_commit_free(struct nfs_write_data *p)
65{
66 if (p && (p->pagevec != &p->page_array[0]))
67 kfree(p->pagevec);
68 mempool_free(p, nfs_commit_mempool);
69}
70EXPORT_SYMBOL_GPL(nfs_commit_free);
71
72struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
73{
74 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
75
76 if (p) {
77 memset(p, 0, sizeof(*p));
78 INIT_LIST_HEAD(&p->pages);
79 p->npages = pagecount;
80 if (pagecount <= ARRAY_SIZE(p->page_array))
81 p->pagevec = p->page_array;
82 else {
83 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
84 if (!p->pagevec) {
85 mempool_free(p, nfs_wdata_mempool);
86 p = NULL;
87 }
88 }
89 }
90 return p;
91}
92
93void nfs_writedata_free(struct nfs_write_data *p)
94{
95 if (p && (p->pagevec != &p->page_array[0]))
96 kfree(p->pagevec);
97 mempool_free(p, nfs_wdata_mempool);
98}
99
100void nfs_writedata_release(struct nfs_write_data *wdata)
101{
102 put_lseg(wdata->lseg);
103 put_nfs_open_context(wdata->args.context);
104 nfs_writedata_free(wdata);
105}
106
107static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
108{
109 ctx->error = error;
110 smp_wmb();
111 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
112}
113
114static struct nfs_page *nfs_page_find_request_locked(struct page *page)
115{
116 struct nfs_page *req = NULL;
117
118 if (PagePrivate(page)) {
119 req = (struct nfs_page *)page_private(page);
120 if (req != NULL)
121 kref_get(&req->wb_kref);
122 }
123 return req;
124}
125
126static struct nfs_page *nfs_page_find_request(struct page *page)
127{
128 struct inode *inode = page->mapping->host;
129 struct nfs_page *req = NULL;
130
131 spin_lock(&inode->i_lock);
132 req = nfs_page_find_request_locked(page);
133 spin_unlock(&inode->i_lock);
134 return req;
135}
136
137/* Adjust the file length if we're writing beyond the end */
138static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
139{
140 struct inode *inode = page->mapping->host;
141 loff_t end, i_size;
142 pgoff_t end_index;
143
144 spin_lock(&inode->i_lock);
145 i_size = i_size_read(inode);
146 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
147 if (i_size > 0 && page->index < end_index)
148 goto out;
149 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
150 if (i_size >= end)
151 goto out;
152 i_size_write(inode, end);
153 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
154out:
155 spin_unlock(&inode->i_lock);
156}
157
158/* A writeback failed: mark the page as bad, and invalidate the page cache */
159static void nfs_set_pageerror(struct page *page)
160{
161 SetPageError(page);
162 nfs_zap_mapping(page->mapping->host, page->mapping);
163}
164
165/* We can set the PG_uptodate flag if we see that a write request
166 * covers the full page.
167 */
168static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
169{
170 if (PageUptodate(page))
171 return;
172 if (base != 0)
173 return;
174 if (count != nfs_page_length(page))
175 return;
176 SetPageUptodate(page);
177}
178
179static int wb_priority(struct writeback_control *wbc)
180{
181 if (wbc->for_reclaim)
182 return FLUSH_HIGHPRI | FLUSH_STABLE;
183 if (wbc->for_kupdate || wbc->for_background)
184 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
185 return FLUSH_COND_STABLE;
186}
187
188/*
189 * NFS congestion control
190 */
191
192int nfs_congestion_kb;
193
194#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
195#define NFS_CONGESTION_OFF_THRESH \
196 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
197
198static int nfs_set_page_writeback(struct page *page)
199{
200 int ret = test_set_page_writeback(page);
201
202 if (!ret) {
203 struct inode *inode = page->mapping->host;
204 struct nfs_server *nfss = NFS_SERVER(inode);
205
206 page_cache_get(page);
207 if (atomic_long_inc_return(&nfss->writeback) >
208 NFS_CONGESTION_ON_THRESH) {
209 set_bdi_congested(&nfss->backing_dev_info,
210 BLK_RW_ASYNC);
211 }
212 }
213 return ret;
214}
215
216static void nfs_end_page_writeback(struct page *page)
217{
218 struct inode *inode = page->mapping->host;
219 struct nfs_server *nfss = NFS_SERVER(inode);
220
221 end_page_writeback(page);
222 page_cache_release(page);
223 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
224 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
225}
226
227static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
228{
229 struct inode *inode = page->mapping->host;
230 struct nfs_page *req;
231 int ret;
232
233 spin_lock(&inode->i_lock);
234 for (;;) {
235 req = nfs_page_find_request_locked(page);
236 if (req == NULL)
237 break;
238 if (nfs_set_page_tag_locked(req))
239 break;
240 /* Note: If we hold the page lock, as is the case in nfs_writepage,
241 * then the call to nfs_set_page_tag_locked() will always
242 * succeed provided that someone hasn't already marked the
243 * request as dirty (in which case we don't care).
244 */
245 spin_unlock(&inode->i_lock);
246 if (!nonblock)
247 ret = nfs_wait_on_request(req);
248 else
249 ret = -EAGAIN;
250 nfs_release_request(req);
251 if (ret != 0)
252 return ERR_PTR(ret);
253 spin_lock(&inode->i_lock);
254 }
255 spin_unlock(&inode->i_lock);
256 return req;
257}
258
259/*
260 * Find an associated nfs write request, and prepare to flush it out
261 * May return an error if the user signalled nfs_wait_on_request().
262 */
263static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
264 struct page *page, bool nonblock)
265{
266 struct nfs_page *req;
267 int ret = 0;
268
269 req = nfs_find_and_lock_request(page, nonblock);
270 if (!req)
271 goto out;
272 ret = PTR_ERR(req);
273 if (IS_ERR(req))
274 goto out;
275
276 ret = nfs_set_page_writeback(page);
277 BUG_ON(ret != 0);
278 BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
279
280 if (!nfs_pageio_add_request(pgio, req)) {
281 nfs_redirty_request(req);
282 ret = pgio->pg_error;
283 }
284out:
285 return ret;
286}
287
288static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
289{
290 struct inode *inode = page->mapping->host;
291 int ret;
292
293 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
294 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
295
296 nfs_pageio_cond_complete(pgio, page->index);
297 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
298 if (ret == -EAGAIN) {
299 redirty_page_for_writepage(wbc, page);
300 ret = 0;
301 }
302 return ret;
303}
304
305/*
306 * Write an mmapped page to the server.
307 */
308static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
309{
310 struct nfs_pageio_descriptor pgio;
311 int err;
312
313 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
314 err = nfs_do_writepage(page, wbc, &pgio);
315 nfs_pageio_complete(&pgio);
316 if (err < 0)
317 return err;
318 if (pgio.pg_error < 0)
319 return pgio.pg_error;
320 return 0;
321}
322
323int nfs_writepage(struct page *page, struct writeback_control *wbc)
324{
325 int ret;
326
327 ret = nfs_writepage_locked(page, wbc);
328 unlock_page(page);
329 return ret;
330}
331
332static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
333{
334 int ret;
335
336 ret = nfs_do_writepage(page, wbc, data);
337 unlock_page(page);
338 return ret;
339}
340
341int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
342{
343 struct inode *inode = mapping->host;
344 unsigned long *bitlock = &NFS_I(inode)->flags;
345 struct nfs_pageio_descriptor pgio;
346 int err;
347
348 /* Stop dirtying of new pages while we sync */
349 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
350 nfs_wait_bit_killable, TASK_KILLABLE);
351 if (err)
352 goto out_err;
353
354 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
355
356 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
357 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
358 nfs_pageio_complete(&pgio);
359
360 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
361 smp_mb__after_clear_bit();
362 wake_up_bit(bitlock, NFS_INO_FLUSHING);
363
364 if (err < 0)
365 goto out_err;
366 err = pgio.pg_error;
367 if (err < 0)
368 goto out_err;
369 return 0;
370out_err:
371 return err;
372}
373
374/*
375 * Insert a write request into an inode
376 */
377static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
378{
379 struct nfs_inode *nfsi = NFS_I(inode);
380 int error;
381
382 error = radix_tree_preload(GFP_NOFS);
383 if (error != 0)
384 goto out;
385
386 /* Lock the request! */
387 nfs_lock_request_dontget(req);
388
389 spin_lock(&inode->i_lock);
390 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
391 BUG_ON(error);
392 if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
393 nfsi->change_attr++;
394 set_bit(PG_MAPPED, &req->wb_flags);
395 SetPagePrivate(req->wb_page);
396 set_page_private(req->wb_page, (unsigned long)req);
397 nfsi->npages++;
398 kref_get(&req->wb_kref);
399 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
400 NFS_PAGE_TAG_LOCKED);
401 spin_unlock(&inode->i_lock);
402 radix_tree_preload_end();
403out:
404 return error;
405}
406
407/*
408 * Remove a write request from an inode
409 */
410static void nfs_inode_remove_request(struct nfs_page *req)
411{
412 struct inode *inode = req->wb_context->dentry->d_inode;
413 struct nfs_inode *nfsi = NFS_I(inode);
414
415 BUG_ON (!NFS_WBACK_BUSY(req));
416
417 spin_lock(&inode->i_lock);
418 set_page_private(req->wb_page, 0);
419 ClearPagePrivate(req->wb_page);
420 clear_bit(PG_MAPPED, &req->wb_flags);
421 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
422 nfsi->npages--;
423 spin_unlock(&inode->i_lock);
424 nfs_release_request(req);
425}
426
427static void
428nfs_mark_request_dirty(struct nfs_page *req)
429{
430 __set_page_dirty_nobuffers(req->wb_page);
431 __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
432}
433
434#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
435/*
436 * Add a request to the inode's commit list.
437 */
438static void
439nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
440{
441 struct inode *inode = req->wb_context->dentry->d_inode;
442 struct nfs_inode *nfsi = NFS_I(inode);
443
444 spin_lock(&inode->i_lock);
445 set_bit(PG_CLEAN, &(req)->wb_flags);
446 radix_tree_tag_set(&nfsi->nfs_page_tree,
447 req->wb_index,
448 NFS_PAGE_TAG_COMMIT);
449 nfsi->ncommit++;
450 spin_unlock(&inode->i_lock);
451 pnfs_mark_request_commit(req, lseg);
452 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
453 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
454 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
455}
456
457static int
458nfs_clear_request_commit(struct nfs_page *req)
459{
460 struct page *page = req->wb_page;
461
462 if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
463 dec_zone_page_state(page, NR_UNSTABLE_NFS);
464 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
465 return 1;
466 }
467 return 0;
468}
469
470static inline
471int nfs_write_need_commit(struct nfs_write_data *data)
472{
473 if (data->verf.committed == NFS_DATA_SYNC)
474 return data->lseg == NULL;
475 else
476 return data->verf.committed != NFS_FILE_SYNC;
477}
478
479static inline
480int nfs_reschedule_unstable_write(struct nfs_page *req,
481 struct nfs_write_data *data)
482{
483 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
484 nfs_mark_request_commit(req, data->lseg);
485 return 1;
486 }
487 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
488 nfs_mark_request_dirty(req);
489 return 1;
490 }
491 return 0;
492}
493#else
494static inline void
495nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
496{
497}
498
499static inline int
500nfs_clear_request_commit(struct nfs_page *req)
501{
502 return 0;
503}
504
505static inline
506int nfs_write_need_commit(struct nfs_write_data *data)
507{
508 return 0;
509}
510
511static inline
512int nfs_reschedule_unstable_write(struct nfs_page *req,
513 struct nfs_write_data *data)
514{
515 return 0;
516}
517#endif
518
519#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
520static int
521nfs_need_commit(struct nfs_inode *nfsi)
522{
523 return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
524}
525
526/*
527 * nfs_scan_commit - Scan an inode for commit requests
528 * @inode: NFS inode to scan
529 * @dst: destination list
530 * @idx_start: lower bound of page->index to scan.
531 * @npages: idx_start + npages sets the upper bound to scan.
532 *
533 * Moves requests from the inode's 'commit' request list.
534 * The requests are *not* checked to ensure that they form a contiguous set.
535 */
536static int
537nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
538{
539 struct nfs_inode *nfsi = NFS_I(inode);
540 int ret;
541
542 if (!nfs_need_commit(nfsi))
543 return 0;
544
545 spin_lock(&inode->i_lock);
546 ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
547 if (ret > 0)
548 nfsi->ncommit -= ret;
549 spin_unlock(&inode->i_lock);
550
551 if (nfs_need_commit(NFS_I(inode)))
552 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
553
554 return ret;
555}
556#else
557static inline int nfs_need_commit(struct nfs_inode *nfsi)
558{
559 return 0;
560}
561
562static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
563{
564 return 0;
565}
566#endif
567
568/*
569 * Search for an existing write request, and attempt to update
570 * it to reflect a new dirty region on a given page.
571 *
572 * If the attempt fails, then the existing request is flushed out
573 * to disk.
574 */
575static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
576 struct page *page,
577 unsigned int offset,
578 unsigned int bytes)
579{
580 struct nfs_page *req;
581 unsigned int rqend;
582 unsigned int end;
583 int error;
584
585 if (!PagePrivate(page))
586 return NULL;
587
588 end = offset + bytes;
589 spin_lock(&inode->i_lock);
590
591 for (;;) {
592 req = nfs_page_find_request_locked(page);
593 if (req == NULL)
594 goto out_unlock;
595
596 rqend = req->wb_offset + req->wb_bytes;
597 /*
598 * Tell the caller to flush out the request if
599 * the offsets are non-contiguous.
600 * Note: nfs_flush_incompatible() will already
601 * have flushed out requests having wrong owners.
602 */
603 if (offset > rqend
604 || end < req->wb_offset)
605 goto out_flushme;
606
607 if (nfs_set_page_tag_locked(req))
608 break;
609
610 /* The request is locked, so wait and then retry */
611 spin_unlock(&inode->i_lock);
612 error = nfs_wait_on_request(req);
613 nfs_release_request(req);
614 if (error != 0)
615 goto out_err;
616 spin_lock(&inode->i_lock);
617 }
618
619 if (nfs_clear_request_commit(req) &&
620 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
621 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL) {
622 NFS_I(inode)->ncommit--;
623 pnfs_clear_request_commit(req);
624 }
625
626 /* Okay, the request matches. Update the region */
627 if (offset < req->wb_offset) {
628 req->wb_offset = offset;
629 req->wb_pgbase = offset;
630 }
631 if (end > rqend)
632 req->wb_bytes = end - req->wb_offset;
633 else
634 req->wb_bytes = rqend - req->wb_offset;
635out_unlock:
636 spin_unlock(&inode->i_lock);
637 return req;
638out_flushme:
639 spin_unlock(&inode->i_lock);
640 nfs_release_request(req);
641 error = nfs_wb_page(inode, page);
642out_err:
643 return ERR_PTR(error);
644}
645
646/*
647 * Try to update an existing write request, or create one if there is none.
648 *
649 * Note: Should always be called with the Page Lock held to prevent races
650 * if we have to add a new request. Also assumes that the caller has
651 * already called nfs_flush_incompatible() if necessary.
652 */
653static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
654 struct page *page, unsigned int offset, unsigned int bytes)
655{
656 struct inode *inode = page->mapping->host;
657 struct nfs_page *req;
658 int error;
659
660 req = nfs_try_to_update_request(inode, page, offset, bytes);
661 if (req != NULL)
662 goto out;
663 req = nfs_create_request(ctx, inode, page, offset, bytes);
664 if (IS_ERR(req))
665 goto out;
666 error = nfs_inode_add_request(inode, req);
667 if (error != 0) {
668 nfs_release_request(req);
669 req = ERR_PTR(error);
670 }
671out:
672 return req;
673}
674
675static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
676 unsigned int offset, unsigned int count)
677{
678 struct nfs_page *req;
679
680 req = nfs_setup_write_request(ctx, page, offset, count);
681 if (IS_ERR(req))
682 return PTR_ERR(req);
683 /* Update file length */
684 nfs_grow_file(page, offset, count);
685 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
686 nfs_mark_request_dirty(req);
687 nfs_clear_page_tag_locked(req);
688 return 0;
689}
690
691int nfs_flush_incompatible(struct file *file, struct page *page)
692{
693 struct nfs_open_context *ctx = nfs_file_open_context(file);
694 struct nfs_page *req;
695 int do_flush, status;
696 /*
697 * Look for a request corresponding to this page. If there
698 * is one, and it belongs to another file, we flush it out
699 * before we try to copy anything into the page. Do this
700 * due to the lack of an ACCESS-type call in NFSv2.
701 * Also do the same if we find a request from an existing
702 * dropped page.
703 */
704 do {
705 req = nfs_page_find_request(page);
706 if (req == NULL)
707 return 0;
708 do_flush = req->wb_page != page || req->wb_context != ctx ||
709 req->wb_lock_context->lockowner != current->files ||
710 req->wb_lock_context->pid != current->tgid;
711 nfs_release_request(req);
712 if (!do_flush)
713 return 0;
714 status = nfs_wb_page(page->mapping->host, page);
715 } while (status == 0);
716 return status;
717}
718
719/*
720 * If the page cache is marked as unsafe or invalid, then we can't rely on
721 * the PageUptodate() flag. In this case, we will need to turn off
722 * write optimisations that depend on the page contents being correct.
723 */
724static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
725{
726 return PageUptodate(page) &&
727 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
728}
729
730/*
731 * Update and possibly write a cached page of an NFS file.
732 *
733 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
734 * things with a page scheduled for an RPC call (e.g. invalidate it).
735 */
736int nfs_updatepage(struct file *file, struct page *page,
737 unsigned int offset, unsigned int count)
738{
739 struct nfs_open_context *ctx = nfs_file_open_context(file);
740 struct inode *inode = page->mapping->host;
741 int status = 0;
742
743 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
744
745 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
746 file->f_path.dentry->d_parent->d_name.name,
747 file->f_path.dentry->d_name.name, count,
748 (long long)(page_offset(page) + offset));
749
750 /* If we're not using byte range locks, and we know the page
751 * is up to date, it may be more efficient to extend the write
752 * to cover the entire page in order to avoid fragmentation
753 * inefficiencies.
754 */
755 if (nfs_write_pageuptodate(page, inode) &&
756 inode->i_flock == NULL &&
757 !(file->f_flags & O_DSYNC)) {
758 count = max(count + offset, nfs_page_length(page));
759 offset = 0;
760 }
761
762 status = nfs_writepage_setup(ctx, page, offset, count);
763 if (status < 0)
764 nfs_set_pageerror(page);
765
766 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
767 status, (long long)i_size_read(inode));
768 return status;
769}
770
771static void nfs_writepage_release(struct nfs_page *req,
772 struct nfs_write_data *data)
773{
774 struct page *page = req->wb_page;
775
776 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
777 nfs_inode_remove_request(req);
778 nfs_clear_page_tag_locked(req);
779 nfs_end_page_writeback(page);
780}
781
782static int flush_task_priority(int how)
783{
784 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
785 case FLUSH_HIGHPRI:
786 return RPC_PRIORITY_HIGH;
787 case FLUSH_LOWPRI:
788 return RPC_PRIORITY_LOW;
789 }
790 return RPC_PRIORITY_NORMAL;
791}
792
793int nfs_initiate_write(struct nfs_write_data *data,
794 struct rpc_clnt *clnt,
795 const struct rpc_call_ops *call_ops,
796 int how)
797{
798 struct inode *inode = data->inode;
799 int priority = flush_task_priority(how);
800 struct rpc_task *task;
801 struct rpc_message msg = {
802 .rpc_argp = &data->args,
803 .rpc_resp = &data->res,
804 .rpc_cred = data->cred,
805 };
806 struct rpc_task_setup task_setup_data = {
807 .rpc_client = clnt,
808 .task = &data->task,
809 .rpc_message = &msg,
810 .callback_ops = call_ops,
811 .callback_data = data,
812 .workqueue = nfsiod_workqueue,
813 .flags = RPC_TASK_ASYNC,
814 .priority = priority,
815 };
816 int ret = 0;
817
818 /* Set up the initial task struct. */
819 NFS_PROTO(inode)->write_setup(data, &msg);
820
821 dprintk("NFS: %5u initiated write call "
822 "(req %s/%lld, %u bytes @ offset %llu)\n",
823 data->task.tk_pid,
824 inode->i_sb->s_id,
825 (long long)NFS_FILEID(inode),
826 data->args.count,
827 (unsigned long long)data->args.offset);
828
829 task = rpc_run_task(&task_setup_data);
830 if (IS_ERR(task)) {
831 ret = PTR_ERR(task);
832 goto out;
833 }
834 if (how & FLUSH_SYNC) {
835 ret = rpc_wait_for_completion_task(task);
836 if (ret == 0)
837 ret = task->tk_status;
838 }
839 rpc_put_task(task);
840out:
841 return ret;
842}
843EXPORT_SYMBOL_GPL(nfs_initiate_write);
844
845/*
846 * Set up the argument/result storage required for the RPC call.
847 */
848static void nfs_write_rpcsetup(struct nfs_page *req,
849 struct nfs_write_data *data,
850 unsigned int count, unsigned int offset,
851 int how)
852{
853 struct inode *inode = req->wb_context->dentry->d_inode;
854
855 /* Set up the RPC argument and reply structs
856 * NB: take care not to mess about with data->commit et al. */
857
858 data->req = req;
859 data->inode = inode = req->wb_context->dentry->d_inode;
860 data->cred = req->wb_context->cred;
861
862 data->args.fh = NFS_FH(inode);
863 data->args.offset = req_offset(req) + offset;
864 /* pnfs_set_layoutcommit needs this */
865 data->mds_offset = data->args.offset;
866 data->args.pgbase = req->wb_pgbase + offset;
867 data->args.pages = data->pagevec;
868 data->args.count = count;
869 data->args.context = get_nfs_open_context(req->wb_context);
870 data->args.lock_context = req->wb_lock_context;
871 data->args.stable = NFS_UNSTABLE;
872 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
873 case 0:
874 break;
875 case FLUSH_COND_STABLE:
876 if (nfs_need_commit(NFS_I(inode)))
877 break;
878 default:
879 data->args.stable = NFS_FILE_SYNC;
880 }
881
882 data->res.fattr = &data->fattr;
883 data->res.count = count;
884 data->res.verf = &data->verf;
885 nfs_fattr_init(&data->fattr);
886}
887
888static int nfs_do_write(struct nfs_write_data *data,
889 const struct rpc_call_ops *call_ops,
890 int how)
891{
892 struct inode *inode = data->args.context->dentry->d_inode;
893
894 return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
895}
896
897static int nfs_do_multiple_writes(struct list_head *head,
898 const struct rpc_call_ops *call_ops,
899 int how)
900{
901 struct nfs_write_data *data;
902 int ret = 0;
903
904 while (!list_empty(head)) {
905 int ret2;
906
907 data = list_entry(head->next, struct nfs_write_data, list);
908 list_del_init(&data->list);
909
910 ret2 = nfs_do_write(data, call_ops, how);
911 if (ret == 0)
912 ret = ret2;
913 }
914 return ret;
915}
916
917/* If a nfs_flush_* function fails, it should remove reqs from @head and
918 * call this on each, which will prepare them to be retried on next
919 * writeback using standard nfs.
920 */
921static void nfs_redirty_request(struct nfs_page *req)
922{
923 struct page *page = req->wb_page;
924
925 nfs_mark_request_dirty(req);
926 nfs_clear_page_tag_locked(req);
927 nfs_end_page_writeback(page);
928}
929
930/*
931 * Generate multiple small requests to write out a single
932 * contiguous dirty area on one page.
933 */
934static int nfs_flush_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
935{
936 struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
937 struct page *page = req->wb_page;
938 struct nfs_write_data *data;
939 size_t wsize = desc->pg_bsize, nbytes;
940 unsigned int offset;
941 int requests = 0;
942 int ret = 0;
943
944 nfs_list_remove_request(req);
945
946 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
947 (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
948 desc->pg_count > wsize))
949 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
950
951
952 offset = 0;
953 nbytes = desc->pg_count;
954 do {
955 size_t len = min(nbytes, wsize);
956
957 data = nfs_writedata_alloc(1);
958 if (!data)
959 goto out_bad;
960 data->pagevec[0] = page;
961 nfs_write_rpcsetup(req, data, len, offset, desc->pg_ioflags);
962 list_add(&data->list, res);
963 requests++;
964 nbytes -= len;
965 offset += len;
966 } while (nbytes != 0);
967 atomic_set(&req->wb_complete, requests);
968 desc->pg_rpc_callops = &nfs_write_partial_ops;
969 return ret;
970
971out_bad:
972 while (!list_empty(res)) {
973 data = list_entry(res->next, struct nfs_write_data, list);
974 list_del(&data->list);
975 nfs_writedata_free(data);
976 }
977 nfs_redirty_request(req);
978 return -ENOMEM;
979}
980
981/*
982 * Create an RPC task for the given write request and kick it.
983 * The page must have been locked by the caller.
984 *
985 * It may happen that the page we're passed is not marked dirty.
986 * This is the case if nfs_updatepage detects a conflicting request
987 * that has been written but not committed.
988 */
989static int nfs_flush_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
990{
991 struct nfs_page *req;
992 struct page **pages;
993 struct nfs_write_data *data;
994 struct list_head *head = &desc->pg_list;
995 int ret = 0;
996
997 data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
998 desc->pg_count));
999 if (!data) {
1000 while (!list_empty(head)) {
1001 req = nfs_list_entry(head->next);
1002 nfs_list_remove_request(req);
1003 nfs_redirty_request(req);
1004 }
1005 ret = -ENOMEM;
1006 goto out;
1007 }
1008 pages = data->pagevec;
1009 while (!list_empty(head)) {
1010 req = nfs_list_entry(head->next);
1011 nfs_list_remove_request(req);
1012 nfs_list_add_request(req, &data->pages);
1013 ClearPageError(req->wb_page);
1014 *pages++ = req->wb_page;
1015 }
1016 req = nfs_list_entry(data->pages.next);
1017
1018 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1019 (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1020 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1021
1022 /* Set up the argument struct */
1023 nfs_write_rpcsetup(req, data, desc->pg_count, 0, desc->pg_ioflags);
1024 list_add(&data->list, res);
1025 desc->pg_rpc_callops = &nfs_write_full_ops;
1026out:
1027 return ret;
1028}
1029
1030int nfs_generic_flush(struct nfs_pageio_descriptor *desc, struct list_head *head)
1031{
1032 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1033 return nfs_flush_multi(desc, head);
1034 return nfs_flush_one(desc, head);
1035}
1036
1037static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1038{
1039 LIST_HEAD(head);
1040 int ret;
1041
1042 ret = nfs_generic_flush(desc, &head);
1043 if (ret == 0)
1044 ret = nfs_do_multiple_writes(&head, desc->pg_rpc_callops,
1045 desc->pg_ioflags);
1046 return ret;
1047}
1048
1049static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1050 .pg_test = nfs_generic_pg_test,
1051 .pg_doio = nfs_generic_pg_writepages,
1052};
1053
1054static void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1055 struct inode *inode, int ioflags)
1056{
1057 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops,
1058 NFS_SERVER(inode)->wsize, ioflags);
1059}
1060
1061void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1062{
1063 pgio->pg_ops = &nfs_pageio_write_ops;
1064 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1065}
1066EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1067
1068static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1069 struct inode *inode, int ioflags)
1070{
1071 if (!pnfs_pageio_init_write(pgio, inode, ioflags))
1072 nfs_pageio_init_write_mds(pgio, inode, ioflags);
1073}
1074
1075/*
1076 * Handle a write reply that flushed part of a page.
1077 */
1078static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1079{
1080 struct nfs_write_data *data = calldata;
1081
1082 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1083 task->tk_pid,
1084 data->req->wb_context->dentry->d_inode->i_sb->s_id,
1085 (long long)
1086 NFS_FILEID(data->req->wb_context->dentry->d_inode),
1087 data->req->wb_bytes, (long long)req_offset(data->req));
1088
1089 nfs_writeback_done(task, data);
1090}
1091
1092static void nfs_writeback_release_partial(void *calldata)
1093{
1094 struct nfs_write_data *data = calldata;
1095 struct nfs_page *req = data->req;
1096 struct page *page = req->wb_page;
1097 int status = data->task.tk_status;
1098
1099 if (status < 0) {
1100 nfs_set_pageerror(page);
1101 nfs_context_set_write_error(req->wb_context, status);
1102 dprintk(", error = %d\n", status);
1103 goto out;
1104 }
1105
1106 if (nfs_write_need_commit(data)) {
1107 struct inode *inode = page->mapping->host;
1108
1109 spin_lock(&inode->i_lock);
1110 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1111 /* Do nothing we need to resend the writes */
1112 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1113 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1114 dprintk(" defer commit\n");
1115 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1116 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1117 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1118 dprintk(" server reboot detected\n");
1119 }
1120 spin_unlock(&inode->i_lock);
1121 } else
1122 dprintk(" OK\n");
1123
1124out:
1125 if (atomic_dec_and_test(&req->wb_complete))
1126 nfs_writepage_release(req, data);
1127 nfs_writedata_release(calldata);
1128}
1129
1130#if defined(CONFIG_NFS_V4_1)
1131void nfs_write_prepare(struct rpc_task *task, void *calldata)
1132{
1133 struct nfs_write_data *data = calldata;
1134
1135 if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1136 &data->args.seq_args,
1137 &data->res.seq_res, 1, task))
1138 return;
1139 rpc_call_start(task);
1140}
1141#endif /* CONFIG_NFS_V4_1 */
1142
1143static const struct rpc_call_ops nfs_write_partial_ops = {
1144#if defined(CONFIG_NFS_V4_1)
1145 .rpc_call_prepare = nfs_write_prepare,
1146#endif /* CONFIG_NFS_V4_1 */
1147 .rpc_call_done = nfs_writeback_done_partial,
1148 .rpc_release = nfs_writeback_release_partial,
1149};
1150
1151/*
1152 * Handle a write reply that flushes a whole page.
1153 *
1154 * FIXME: There is an inherent race with invalidate_inode_pages and
1155 * writebacks since the page->count is kept > 1 for as long
1156 * as the page has a write request pending.
1157 */
1158static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1159{
1160 struct nfs_write_data *data = calldata;
1161
1162 nfs_writeback_done(task, data);
1163}
1164
1165static void nfs_writeback_release_full(void *calldata)
1166{
1167 struct nfs_write_data *data = calldata;
1168 int status = data->task.tk_status;
1169
1170 /* Update attributes as result of writeback. */
1171 while (!list_empty(&data->pages)) {
1172 struct nfs_page *req = nfs_list_entry(data->pages.next);
1173 struct page *page = req->wb_page;
1174
1175 nfs_list_remove_request(req);
1176
1177 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1178 data->task.tk_pid,
1179 req->wb_context->dentry->d_inode->i_sb->s_id,
1180 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1181 req->wb_bytes,
1182 (long long)req_offset(req));
1183
1184 if (status < 0) {
1185 nfs_set_pageerror(page);
1186 nfs_context_set_write_error(req->wb_context, status);
1187 dprintk(", error = %d\n", status);
1188 goto remove_request;
1189 }
1190
1191 if (nfs_write_need_commit(data)) {
1192 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1193 nfs_mark_request_commit(req, data->lseg);
1194 dprintk(" marked for commit\n");
1195 goto next;
1196 }
1197 dprintk(" OK\n");
1198remove_request:
1199 nfs_inode_remove_request(req);
1200 next:
1201 nfs_clear_page_tag_locked(req);
1202 nfs_end_page_writeback(page);
1203 }
1204 nfs_writedata_release(calldata);
1205}
1206
1207static const struct rpc_call_ops nfs_write_full_ops = {
1208#if defined(CONFIG_NFS_V4_1)
1209 .rpc_call_prepare = nfs_write_prepare,
1210#endif /* CONFIG_NFS_V4_1 */
1211 .rpc_call_done = nfs_writeback_done_full,
1212 .rpc_release = nfs_writeback_release_full,
1213};
1214
1215
1216/*
1217 * This function is called when the WRITE call is complete.
1218 */
1219void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1220{
1221 struct nfs_writeargs *argp = &data->args;
1222 struct nfs_writeres *resp = &data->res;
1223 struct nfs_server *server = NFS_SERVER(data->inode);
1224 int status;
1225
1226 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1227 task->tk_pid, task->tk_status);
1228
1229 /*
1230 * ->write_done will attempt to use post-op attributes to detect
1231 * conflicting writes by other clients. A strict interpretation
1232 * of close-to-open would allow us to continue caching even if
1233 * another writer had changed the file, but some applications
1234 * depend on tighter cache coherency when writing.
1235 */
1236 status = NFS_PROTO(data->inode)->write_done(task, data);
1237 if (status != 0)
1238 return;
1239 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1240
1241#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1242 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1243 /* We tried a write call, but the server did not
1244 * commit data to stable storage even though we
1245 * requested it.
1246 * Note: There is a known bug in Tru64 < 5.0 in which
1247 * the server reports NFS_DATA_SYNC, but performs
1248 * NFS_FILE_SYNC. We therefore implement this checking
1249 * as a dprintk() in order to avoid filling syslog.
1250 */
1251 static unsigned long complain;
1252
1253 /* Note this will print the MDS for a DS write */
1254 if (time_before(complain, jiffies)) {
1255 dprintk("NFS: faulty NFS server %s:"
1256 " (committed = %d) != (stable = %d)\n",
1257 server->nfs_client->cl_hostname,
1258 resp->verf->committed, argp->stable);
1259 complain = jiffies + 300 * HZ;
1260 }
1261 }
1262#endif
1263 /* Is this a short write? */
1264 if (task->tk_status >= 0 && resp->count < argp->count) {
1265 static unsigned long complain;
1266
1267 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1268
1269 /* Has the server at least made some progress? */
1270 if (resp->count != 0) {
1271 /* Was this an NFSv2 write or an NFSv3 stable write? */
1272 if (resp->verf->committed != NFS_UNSTABLE) {
1273 /* Resend from where the server left off */
1274 data->mds_offset += resp->count;
1275 argp->offset += resp->count;
1276 argp->pgbase += resp->count;
1277 argp->count -= resp->count;
1278 } else {
1279 /* Resend as a stable write in order to avoid
1280 * headaches in the case of a server crash.
1281 */
1282 argp->stable = NFS_FILE_SYNC;
1283 }
1284 nfs_restart_rpc(task, server->nfs_client);
1285 return;
1286 }
1287 if (time_before(complain, jiffies)) {
1288 printk(KERN_WARNING
1289 "NFS: Server wrote zero bytes, expected %u.\n",
1290 argp->count);
1291 complain = jiffies + 300 * HZ;
1292 }
1293 /* Can't do anything about it except throw an error. */
1294 task->tk_status = -EIO;
1295 }
1296 return;
1297}
1298
1299
1300#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1301static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1302{
1303 int ret;
1304
1305 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1306 return 1;
1307 if (!may_wait)
1308 return 0;
1309 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1310 NFS_INO_COMMIT,
1311 nfs_wait_bit_killable,
1312 TASK_KILLABLE);
1313 return (ret < 0) ? ret : 1;
1314}
1315
1316void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1317{
1318 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1319 smp_mb__after_clear_bit();
1320 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1321}
1322EXPORT_SYMBOL_GPL(nfs_commit_clear_lock);
1323
1324void nfs_commitdata_release(void *data)
1325{
1326 struct nfs_write_data *wdata = data;
1327
1328 put_lseg(wdata->lseg);
1329 put_nfs_open_context(wdata->args.context);
1330 nfs_commit_free(wdata);
1331}
1332EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1333
1334int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1335 const struct rpc_call_ops *call_ops,
1336 int how)
1337{
1338 struct rpc_task *task;
1339 int priority = flush_task_priority(how);
1340 struct rpc_message msg = {
1341 .rpc_argp = &data->args,
1342 .rpc_resp = &data->res,
1343 .rpc_cred = data->cred,
1344 };
1345 struct rpc_task_setup task_setup_data = {
1346 .task = &data->task,
1347 .rpc_client = clnt,
1348 .rpc_message = &msg,
1349 .callback_ops = call_ops,
1350 .callback_data = data,
1351 .workqueue = nfsiod_workqueue,
1352 .flags = RPC_TASK_ASYNC,
1353 .priority = priority,
1354 };
1355 /* Set up the initial task struct. */
1356 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1357
1358 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1359
1360 task = rpc_run_task(&task_setup_data);
1361 if (IS_ERR(task))
1362 return PTR_ERR(task);
1363 if (how & FLUSH_SYNC)
1364 rpc_wait_for_completion_task(task);
1365 rpc_put_task(task);
1366 return 0;
1367}
1368EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1369
1370/*
1371 * Set up the argument/result storage required for the RPC call.
1372 */
1373void nfs_init_commit(struct nfs_write_data *data,
1374 struct list_head *head,
1375 struct pnfs_layout_segment *lseg)
1376{
1377 struct nfs_page *first = nfs_list_entry(head->next);
1378 struct inode *inode = first->wb_context->dentry->d_inode;
1379
1380 /* Set up the RPC argument and reply structs
1381 * NB: take care not to mess about with data->commit et al. */
1382
1383 list_splice_init(head, &data->pages);
1384
1385 data->inode = inode;
1386 data->cred = first->wb_context->cred;
1387 data->lseg = lseg; /* reference transferred */
1388 data->mds_ops = &nfs_commit_ops;
1389
1390 data->args.fh = NFS_FH(data->inode);
1391 /* Note: we always request a commit of the entire inode */
1392 data->args.offset = 0;
1393 data->args.count = 0;
1394 data->args.context = get_nfs_open_context(first->wb_context);
1395 data->res.count = 0;
1396 data->res.fattr = &data->fattr;
1397 data->res.verf = &data->verf;
1398 nfs_fattr_init(&data->fattr);
1399}
1400EXPORT_SYMBOL_GPL(nfs_init_commit);
1401
1402void nfs_retry_commit(struct list_head *page_list,
1403 struct pnfs_layout_segment *lseg)
1404{
1405 struct nfs_page *req;
1406
1407 while (!list_empty(page_list)) {
1408 req = nfs_list_entry(page_list->next);
1409 nfs_list_remove_request(req);
1410 nfs_mark_request_commit(req, lseg);
1411 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1412 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1413 BDI_RECLAIMABLE);
1414 nfs_clear_page_tag_locked(req);
1415 }
1416}
1417EXPORT_SYMBOL_GPL(nfs_retry_commit);
1418
1419/*
1420 * Commit dirty pages
1421 */
1422static int
1423nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1424{
1425 struct nfs_write_data *data;
1426
1427 data = nfs_commitdata_alloc();
1428
1429 if (!data)
1430 goto out_bad;
1431
1432 /* Set up the argument struct */
1433 nfs_init_commit(data, head, NULL);
1434 return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1435 out_bad:
1436 nfs_retry_commit(head, NULL);
1437 nfs_commit_clear_lock(NFS_I(inode));
1438 return -ENOMEM;
1439}
1440
1441/*
1442 * COMMIT call returned
1443 */
1444static void nfs_commit_done(struct rpc_task *task, void *calldata)
1445{
1446 struct nfs_write_data *data = calldata;
1447
1448 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1449 task->tk_pid, task->tk_status);
1450
1451 /* Call the NFS version-specific code */
1452 NFS_PROTO(data->inode)->commit_done(task, data);
1453}
1454
1455void nfs_commit_release_pages(struct nfs_write_data *data)
1456{
1457 struct nfs_page *req;
1458 int status = data->task.tk_status;
1459
1460 while (!list_empty(&data->pages)) {
1461 req = nfs_list_entry(data->pages.next);
1462 nfs_list_remove_request(req);
1463 nfs_clear_request_commit(req);
1464
1465 dprintk("NFS: commit (%s/%lld %d@%lld)",
1466 req->wb_context->dentry->d_sb->s_id,
1467 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1468 req->wb_bytes,
1469 (long long)req_offset(req));
1470 if (status < 0) {
1471 nfs_context_set_write_error(req->wb_context, status);
1472 nfs_inode_remove_request(req);
1473 dprintk(", error = %d\n", status);
1474 goto next;
1475 }
1476
1477 /* Okay, COMMIT succeeded, apparently. Check the verifier
1478 * returned by the server against all stored verfs. */
1479 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1480 /* We have a match */
1481 nfs_inode_remove_request(req);
1482 dprintk(" OK\n");
1483 goto next;
1484 }
1485 /* We have a mismatch. Write the page again */
1486 dprintk(" mismatch\n");
1487 nfs_mark_request_dirty(req);
1488 next:
1489 nfs_clear_page_tag_locked(req);
1490 }
1491}
1492EXPORT_SYMBOL_GPL(nfs_commit_release_pages);
1493
1494static void nfs_commit_release(void *calldata)
1495{
1496 struct nfs_write_data *data = calldata;
1497
1498 nfs_commit_release_pages(data);
1499 nfs_commit_clear_lock(NFS_I(data->inode));
1500 nfs_commitdata_release(calldata);
1501}
1502
1503static const struct rpc_call_ops nfs_commit_ops = {
1504#if defined(CONFIG_NFS_V4_1)
1505 .rpc_call_prepare = nfs_write_prepare,
1506#endif /* CONFIG_NFS_V4_1 */
1507 .rpc_call_done = nfs_commit_done,
1508 .rpc_release = nfs_commit_release,
1509};
1510
1511int nfs_commit_inode(struct inode *inode, int how)
1512{
1513 LIST_HEAD(head);
1514 int may_wait = how & FLUSH_SYNC;
1515 int res;
1516
1517 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1518 if (res <= 0)
1519 goto out_mark_dirty;
1520 res = nfs_scan_commit(inode, &head, 0, 0);
1521 if (res) {
1522 int error;
1523
1524 error = pnfs_commit_list(inode, &head, how);
1525 if (error == PNFS_NOT_ATTEMPTED)
1526 error = nfs_commit_list(inode, &head, how);
1527 if (error < 0)
1528 return error;
1529 if (!may_wait)
1530 goto out_mark_dirty;
1531 error = wait_on_bit(&NFS_I(inode)->flags,
1532 NFS_INO_COMMIT,
1533 nfs_wait_bit_killable,
1534 TASK_KILLABLE);
1535 if (error < 0)
1536 return error;
1537 } else
1538 nfs_commit_clear_lock(NFS_I(inode));
1539 return res;
1540 /* Note: If we exit without ensuring that the commit is complete,
1541 * we must mark the inode as dirty. Otherwise, future calls to
1542 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1543 * that the data is on the disk.
1544 */
1545out_mark_dirty:
1546 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1547 return res;
1548}
1549
1550static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1551{
1552 struct nfs_inode *nfsi = NFS_I(inode);
1553 int flags = FLUSH_SYNC;
1554 int ret = 0;
1555
1556 if (wbc->sync_mode == WB_SYNC_NONE) {
1557 /* Don't commit yet if this is a non-blocking flush and there
1558 * are a lot of outstanding writes for this mapping.
1559 */
1560 if (nfsi->ncommit <= (nfsi->npages >> 1))
1561 goto out_mark_dirty;
1562
1563 /* don't wait for the COMMIT response */
1564 flags = 0;
1565 }
1566
1567 ret = nfs_commit_inode(inode, flags);
1568 if (ret >= 0) {
1569 if (wbc->sync_mode == WB_SYNC_NONE) {
1570 if (ret < wbc->nr_to_write)
1571 wbc->nr_to_write -= ret;
1572 else
1573 wbc->nr_to_write = 0;
1574 }
1575 return 0;
1576 }
1577out_mark_dirty:
1578 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1579 return ret;
1580}
1581#else
1582static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1583{
1584 return 0;
1585}
1586#endif
1587
1588int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1589{
1590 int ret;
1591
1592 ret = nfs_commit_unstable_pages(inode, wbc);
1593 if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1594 int status;
1595 bool sync = true;
1596
1597 if (wbc->sync_mode == WB_SYNC_NONE)
1598 sync = false;
1599
1600 status = pnfs_layoutcommit_inode(inode, sync);
1601 if (status < 0)
1602 return status;
1603 }
1604 return ret;
1605}
1606
1607/*
1608 * flush the inode to disk.
1609 */
1610int nfs_wb_all(struct inode *inode)
1611{
1612 struct writeback_control wbc = {
1613 .sync_mode = WB_SYNC_ALL,
1614 .nr_to_write = LONG_MAX,
1615 .range_start = 0,
1616 .range_end = LLONG_MAX,
1617 };
1618
1619 return sync_inode(inode, &wbc);
1620}
1621
1622int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1623{
1624 struct nfs_page *req;
1625 int ret = 0;
1626
1627 BUG_ON(!PageLocked(page));
1628 for (;;) {
1629 wait_on_page_writeback(page);
1630 req = nfs_page_find_request(page);
1631 if (req == NULL)
1632 break;
1633 if (nfs_lock_request_dontget(req)) {
1634 nfs_inode_remove_request(req);
1635 /*
1636 * In case nfs_inode_remove_request has marked the
1637 * page as being dirty
1638 */
1639 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1640 nfs_unlock_request(req);
1641 break;
1642 }
1643 ret = nfs_wait_on_request(req);
1644 nfs_release_request(req);
1645 if (ret < 0)
1646 break;
1647 }
1648 return ret;
1649}
1650
1651/*
1652 * Write back all requests on one page - we do this before reading it.
1653 */
1654int nfs_wb_page(struct inode *inode, struct page *page)
1655{
1656 loff_t range_start = page_offset(page);
1657 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1658 struct writeback_control wbc = {
1659 .sync_mode = WB_SYNC_ALL,
1660 .nr_to_write = 0,
1661 .range_start = range_start,
1662 .range_end = range_end,
1663 };
1664 int ret;
1665
1666 for (;;) {
1667 wait_on_page_writeback(page);
1668 if (clear_page_dirty_for_io(page)) {
1669 ret = nfs_writepage_locked(page, &wbc);
1670 if (ret < 0)
1671 goto out_error;
1672 continue;
1673 }
1674 if (!PagePrivate(page))
1675 break;
1676 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1677 if (ret < 0)
1678 goto out_error;
1679 }
1680 return 0;
1681out_error:
1682 return ret;
1683}
1684
1685#ifdef CONFIG_MIGRATION
1686int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1687 struct page *page)
1688{
1689 struct nfs_page *req;
1690 int ret;
1691
1692 nfs_fscache_release_page(page, GFP_KERNEL);
1693
1694 req = nfs_find_and_lock_request(page, false);
1695 ret = PTR_ERR(req);
1696 if (IS_ERR(req))
1697 goto out;
1698
1699 ret = migrate_page(mapping, newpage, page);
1700 if (!req)
1701 goto out;
1702 if (ret)
1703 goto out_unlock;
1704 page_cache_get(newpage);
1705 spin_lock(&mapping->host->i_lock);
1706 req->wb_page = newpage;
1707 SetPagePrivate(newpage);
1708 set_page_private(newpage, (unsigned long)req);
1709 ClearPagePrivate(page);
1710 set_page_private(page, 0);
1711 spin_unlock(&mapping->host->i_lock);
1712 page_cache_release(page);
1713out_unlock:
1714 nfs_clear_page_tag_locked(req);
1715out:
1716 return ret;
1717}
1718#endif
1719
1720int __init nfs_init_writepagecache(void)
1721{
1722 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1723 sizeof(struct nfs_write_data),
1724 0, SLAB_HWCACHE_ALIGN,
1725 NULL);
1726 if (nfs_wdata_cachep == NULL)
1727 return -ENOMEM;
1728
1729 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1730 nfs_wdata_cachep);
1731 if (nfs_wdata_mempool == NULL)
1732 return -ENOMEM;
1733
1734 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1735 nfs_wdata_cachep);
1736 if (nfs_commit_mempool == NULL)
1737 return -ENOMEM;
1738
1739 /*
1740 * NFS congestion size, scale with available memory.
1741 *
1742 * 64MB: 8192k
1743 * 128MB: 11585k
1744 * 256MB: 16384k
1745 * 512MB: 23170k
1746 * 1GB: 32768k
1747 * 2GB: 46340k
1748 * 4GB: 65536k
1749 * 8GB: 92681k
1750 * 16GB: 131072k
1751 *
1752 * This allows larger machines to have larger/more transfers.
1753 * Limit the default to 256M
1754 */
1755 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1756 if (nfs_congestion_kb > 256*1024)
1757 nfs_congestion_kb = 256*1024;
1758
1759 return 0;
1760}
1761
1762void nfs_destroy_writepagecache(void)
1763{
1764 mempool_destroy(nfs_commit_mempool);
1765 mempool_destroy(nfs_wdata_mempool);
1766 kmem_cache_destroy(nfs_wdata_cachep);
1767}
1768
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/nfs/write.c
4 *
5 * Write file data over NFS.
6 *
7 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
8 */
9
10#include <linux/types.h>
11#include <linux/slab.h>
12#include <linux/mm.h>
13#include <linux/pagemap.h>
14#include <linux/file.h>
15#include <linux/writeback.h>
16#include <linux/swap.h>
17#include <linux/migrate.h>
18
19#include <linux/sunrpc/clnt.h>
20#include <linux/nfs_fs.h>
21#include <linux/nfs_mount.h>
22#include <linux/nfs_page.h>
23#include <linux/backing-dev.h>
24#include <linux/export.h>
25#include <linux/freezer.h>
26#include <linux/wait.h>
27#include <linux/iversion.h>
28#include <linux/filelock.h>
29
30#include <linux/uaccess.h>
31#include <linux/sched/mm.h>
32
33#include "delegation.h"
34#include "internal.h"
35#include "iostat.h"
36#include "nfs4_fs.h"
37#include "fscache.h"
38#include "pnfs.h"
39
40#include "nfstrace.h"
41
42#define NFSDBG_FACILITY NFSDBG_PAGECACHE
43
44#define MIN_POOL_WRITE (32)
45#define MIN_POOL_COMMIT (4)
46
47struct nfs_io_completion {
48 void (*complete)(void *data);
49 void *data;
50 struct kref refcount;
51};
52
53/*
54 * Local function declarations
55 */
56static void nfs_redirty_request(struct nfs_page *req);
57static const struct rpc_call_ops nfs_commit_ops;
58static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
59static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
60static const struct nfs_rw_ops nfs_rw_write_ops;
61static void nfs_inode_remove_request(struct nfs_page *req);
62static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
63 struct nfs_page *req);
64static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
65 struct inode *inode);
66static struct nfs_page *
67nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
68 struct folio *folio);
69
70static struct kmem_cache *nfs_wdata_cachep;
71static mempool_t *nfs_wdata_mempool;
72static struct kmem_cache *nfs_cdata_cachep;
73static mempool_t *nfs_commit_mempool;
74
75struct nfs_commit_data *nfs_commitdata_alloc(void)
76{
77 struct nfs_commit_data *p;
78
79 p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
80 if (!p) {
81 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
82 if (!p)
83 return NULL;
84 memset(p, 0, sizeof(*p));
85 }
86 INIT_LIST_HEAD(&p->pages);
87 return p;
88}
89EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
90
91void nfs_commit_free(struct nfs_commit_data *p)
92{
93 mempool_free(p, nfs_commit_mempool);
94}
95EXPORT_SYMBOL_GPL(nfs_commit_free);
96
97static struct nfs_pgio_header *nfs_writehdr_alloc(void)
98{
99 struct nfs_pgio_header *p;
100
101 p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
102 if (!p) {
103 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
104 if (!p)
105 return NULL;
106 memset(p, 0, sizeof(*p));
107 }
108 p->rw_mode = FMODE_WRITE;
109 return p;
110}
111
112static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
113{
114 mempool_free(hdr, nfs_wdata_mempool);
115}
116
117static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
118{
119 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
120}
121
122static void nfs_io_completion_init(struct nfs_io_completion *ioc,
123 void (*complete)(void *), void *data)
124{
125 ioc->complete = complete;
126 ioc->data = data;
127 kref_init(&ioc->refcount);
128}
129
130static void nfs_io_completion_release(struct kref *kref)
131{
132 struct nfs_io_completion *ioc = container_of(kref,
133 struct nfs_io_completion, refcount);
134 ioc->complete(ioc->data);
135 kfree(ioc);
136}
137
138static void nfs_io_completion_get(struct nfs_io_completion *ioc)
139{
140 if (ioc != NULL)
141 kref_get(&ioc->refcount);
142}
143
144static void nfs_io_completion_put(struct nfs_io_completion *ioc)
145{
146 if (ioc != NULL)
147 kref_put(&ioc->refcount, nfs_io_completion_release);
148}
149
150static void
151nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
152{
153 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
154 kref_get(&req->wb_kref);
155 atomic_long_inc(&NFS_I(inode)->nrequests);
156 }
157}
158
159static int
160nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
161{
162 int ret;
163
164 if (!test_bit(PG_REMOVE, &req->wb_flags))
165 return 0;
166 ret = nfs_page_group_lock(req);
167 if (ret)
168 return ret;
169 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
170 nfs_page_set_inode_ref(req, inode);
171 nfs_page_group_unlock(req);
172 return 0;
173}
174
175static struct nfs_page *nfs_folio_private_request(struct folio *folio)
176{
177 return folio_get_private(folio);
178}
179
180/**
181 * nfs_folio_find_private_request - find head request associated with a folio
182 * @folio: pointer to folio
183 *
184 * must be called while holding the inode lock.
185 *
186 * returns matching head request with reference held, or NULL if not found.
187 */
188static struct nfs_page *nfs_folio_find_private_request(struct folio *folio)
189{
190 struct address_space *mapping = folio_file_mapping(folio);
191 struct nfs_page *req;
192
193 if (!folio_test_private(folio))
194 return NULL;
195 spin_lock(&mapping->i_private_lock);
196 req = nfs_folio_private_request(folio);
197 if (req) {
198 WARN_ON_ONCE(req->wb_head != req);
199 kref_get(&req->wb_kref);
200 }
201 spin_unlock(&mapping->i_private_lock);
202 return req;
203}
204
205static struct nfs_page *nfs_folio_find_swap_request(struct folio *folio)
206{
207 struct inode *inode = folio_file_mapping(folio)->host;
208 struct nfs_inode *nfsi = NFS_I(inode);
209 struct nfs_page *req = NULL;
210 if (!folio_test_swapcache(folio))
211 return NULL;
212 mutex_lock(&nfsi->commit_mutex);
213 if (folio_test_swapcache(folio)) {
214 req = nfs_page_search_commits_for_head_request_locked(nfsi,
215 folio);
216 if (req) {
217 WARN_ON_ONCE(req->wb_head != req);
218 kref_get(&req->wb_kref);
219 }
220 }
221 mutex_unlock(&nfsi->commit_mutex);
222 return req;
223}
224
225/**
226 * nfs_folio_find_head_request - find head request associated with a folio
227 * @folio: pointer to folio
228 *
229 * returns matching head request with reference held, or NULL if not found.
230 */
231static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
232{
233 struct nfs_page *req;
234
235 req = nfs_folio_find_private_request(folio);
236 if (!req)
237 req = nfs_folio_find_swap_request(folio);
238 return req;
239}
240
241static struct nfs_page *nfs_folio_find_and_lock_request(struct folio *folio)
242{
243 struct inode *inode = folio_file_mapping(folio)->host;
244 struct nfs_page *req, *head;
245 int ret;
246
247 for (;;) {
248 req = nfs_folio_find_head_request(folio);
249 if (!req)
250 return req;
251 head = nfs_page_group_lock_head(req);
252 if (head != req)
253 nfs_release_request(req);
254 if (IS_ERR(head))
255 return head;
256 ret = nfs_cancel_remove_inode(head, inode);
257 if (ret < 0) {
258 nfs_unlock_and_release_request(head);
259 return ERR_PTR(ret);
260 }
261 /* Ensure that nobody removed the request before we locked it */
262 if (head == nfs_folio_private_request(folio))
263 break;
264 if (folio_test_swapcache(folio))
265 break;
266 nfs_unlock_and_release_request(head);
267 }
268 return head;
269}
270
271/* Adjust the file length if we're writing beyond the end */
272static void nfs_grow_file(struct folio *folio, unsigned int offset,
273 unsigned int count)
274{
275 struct inode *inode = folio_file_mapping(folio)->host;
276 loff_t end, i_size;
277 pgoff_t end_index;
278
279 spin_lock(&inode->i_lock);
280 i_size = i_size_read(inode);
281 end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio);
282 if (i_size > 0 && folio_index(folio) < end_index)
283 goto out;
284 end = folio_file_pos(folio) + (loff_t)offset + (loff_t)count;
285 if (i_size >= end)
286 goto out;
287 trace_nfs_size_grow(inode, end);
288 i_size_write(inode, end);
289 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
290 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
291out:
292 spin_unlock(&inode->i_lock);
293 nfs_fscache_invalidate(inode, 0);
294}
295
296/* A writeback failed: mark the page as bad, and invalidate the page cache */
297static void nfs_set_pageerror(struct address_space *mapping)
298{
299 struct inode *inode = mapping->host;
300
301 nfs_zap_mapping(mapping->host, mapping);
302 /* Force file size revalidation */
303 spin_lock(&inode->i_lock);
304 nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
305 NFS_INO_INVALID_CHANGE |
306 NFS_INO_INVALID_SIZE);
307 spin_unlock(&inode->i_lock);
308}
309
310static void nfs_mapping_set_error(struct folio *folio, int error)
311{
312 struct address_space *mapping = folio_file_mapping(folio);
313
314 folio_set_error(folio);
315 filemap_set_wb_err(mapping, error);
316 if (mapping->host)
317 errseq_set(&mapping->host->i_sb->s_wb_err,
318 error == -ENOSPC ? -ENOSPC : -EIO);
319 nfs_set_pageerror(mapping);
320}
321
322/*
323 * nfs_page_group_search_locked
324 * @head - head request of page group
325 * @page_offset - offset into page
326 *
327 * Search page group with head @head to find a request that contains the
328 * page offset @page_offset.
329 *
330 * Returns a pointer to the first matching nfs request, or NULL if no
331 * match is found.
332 *
333 * Must be called with the page group lock held
334 */
335static struct nfs_page *
336nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
337{
338 struct nfs_page *req;
339
340 req = head;
341 do {
342 if (page_offset >= req->wb_pgbase &&
343 page_offset < (req->wb_pgbase + req->wb_bytes))
344 return req;
345
346 req = req->wb_this_page;
347 } while (req != head);
348
349 return NULL;
350}
351
352/*
353 * nfs_page_group_covers_page
354 * @head - head request of page group
355 *
356 * Return true if the page group with head @head covers the whole page,
357 * returns false otherwise
358 */
359static bool nfs_page_group_covers_page(struct nfs_page *req)
360{
361 unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
362 struct nfs_page *tmp;
363 unsigned int pos = 0;
364
365 nfs_page_group_lock(req);
366
367 for (;;) {
368 tmp = nfs_page_group_search_locked(req->wb_head, pos);
369 if (!tmp)
370 break;
371 pos = tmp->wb_pgbase + tmp->wb_bytes;
372 }
373
374 nfs_page_group_unlock(req);
375 return pos >= len;
376}
377
378/* We can set the PG_uptodate flag if we see that a write request
379 * covers the full page.
380 */
381static void nfs_mark_uptodate(struct nfs_page *req)
382{
383 struct folio *folio = nfs_page_to_folio(req);
384
385 if (folio_test_uptodate(folio))
386 return;
387 if (!nfs_page_group_covers_page(req))
388 return;
389 folio_mark_uptodate(folio);
390}
391
392static int wb_priority(struct writeback_control *wbc)
393{
394 int ret = 0;
395
396 if (wbc->sync_mode == WB_SYNC_ALL)
397 ret = FLUSH_COND_STABLE;
398 return ret;
399}
400
401/*
402 * NFS congestion control
403 */
404
405int nfs_congestion_kb;
406
407#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
408#define NFS_CONGESTION_OFF_THRESH \
409 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
410
411static void nfs_folio_set_writeback(struct folio *folio)
412{
413 struct nfs_server *nfss = NFS_SERVER(folio_file_mapping(folio)->host);
414
415 folio_start_writeback(folio);
416 if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
417 nfss->write_congested = 1;
418}
419
420static void nfs_folio_end_writeback(struct folio *folio)
421{
422 struct nfs_server *nfss = NFS_SERVER(folio_file_mapping(folio)->host);
423
424 folio_end_writeback(folio);
425 if (atomic_long_dec_return(&nfss->writeback) <
426 NFS_CONGESTION_OFF_THRESH)
427 nfss->write_congested = 0;
428}
429
430static void nfs_page_end_writeback(struct nfs_page *req)
431{
432 if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
433 nfs_unlock_request(req);
434 nfs_folio_end_writeback(nfs_page_to_folio(req));
435 } else
436 nfs_unlock_request(req);
437}
438
439/*
440 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
441 *
442 * @destroy_list - request list (using wb_this_page) terminated by @old_head
443 * @old_head - the old head of the list
444 *
445 * All subrequests must be locked and removed from all lists, so at this point
446 * they are only "active" in this function, and possibly in nfs_wait_on_request
447 * with a reference held by some other context.
448 */
449static void
450nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
451 struct nfs_page *old_head,
452 struct inode *inode)
453{
454 while (destroy_list) {
455 struct nfs_page *subreq = destroy_list;
456
457 destroy_list = (subreq->wb_this_page == old_head) ?
458 NULL : subreq->wb_this_page;
459
460 /* Note: lock subreq in order to change subreq->wb_head */
461 nfs_page_set_headlock(subreq);
462 WARN_ON_ONCE(old_head != subreq->wb_head);
463
464 /* make sure old group is not used */
465 subreq->wb_this_page = subreq;
466 subreq->wb_head = subreq;
467
468 clear_bit(PG_REMOVE, &subreq->wb_flags);
469
470 /* Note: races with nfs_page_group_destroy() */
471 if (!kref_read(&subreq->wb_kref)) {
472 /* Check if we raced with nfs_page_group_destroy() */
473 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
474 nfs_page_clear_headlock(subreq);
475 nfs_free_request(subreq);
476 } else
477 nfs_page_clear_headlock(subreq);
478 continue;
479 }
480 nfs_page_clear_headlock(subreq);
481
482 nfs_release_request(old_head);
483
484 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
485 nfs_release_request(subreq);
486 atomic_long_dec(&NFS_I(inode)->nrequests);
487 }
488
489 /* subreq is now totally disconnected from page group or any
490 * write / commit lists. last chance to wake any waiters */
491 nfs_unlock_and_release_request(subreq);
492 }
493}
494
495/*
496 * nfs_join_page_group - destroy subrequests of the head req
497 * @head: the page used to lookup the "page group" of nfs_page structures
498 * @inode: Inode to which the request belongs.
499 *
500 * This function joins all sub requests to the head request by first
501 * locking all requests in the group, cancelling any pending operations
502 * and finally updating the head request to cover the whole range covered by
503 * the (former) group. All subrequests are removed from any write or commit
504 * lists, unlinked from the group and destroyed.
505 */
506void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
507 struct inode *inode)
508{
509 struct nfs_page *subreq;
510 struct nfs_page *destroy_list = NULL;
511 unsigned int pgbase, off, bytes;
512
513 pgbase = head->wb_pgbase;
514 bytes = head->wb_bytes;
515 off = head->wb_offset;
516 for (subreq = head->wb_this_page; subreq != head;
517 subreq = subreq->wb_this_page) {
518 /* Subrequests should always form a contiguous range */
519 if (pgbase > subreq->wb_pgbase) {
520 off -= pgbase - subreq->wb_pgbase;
521 bytes += pgbase - subreq->wb_pgbase;
522 pgbase = subreq->wb_pgbase;
523 }
524 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
525 - pgbase, bytes);
526 }
527
528 /* Set the head request's range to cover the former page group */
529 head->wb_pgbase = pgbase;
530 head->wb_bytes = bytes;
531 head->wb_offset = off;
532
533 /* Now that all requests are locked, make sure they aren't on any list.
534 * Commit list removal accounting is done after locks are dropped */
535 subreq = head;
536 do {
537 nfs_clear_request_commit(cinfo, subreq);
538 subreq = subreq->wb_this_page;
539 } while (subreq != head);
540
541 /* unlink subrequests from head, destroy them later */
542 if (head->wb_this_page != head) {
543 /* destroy list will be terminated by head */
544 destroy_list = head->wb_this_page;
545 head->wb_this_page = head;
546 }
547
548 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
549}
550
551/*
552 * nfs_lock_and_join_requests - join all subreqs to the head req
553 * @folio: the folio used to lookup the "page group" of nfs_page structures
554 *
555 * This function joins all sub requests to the head request by first
556 * locking all requests in the group, cancelling any pending operations
557 * and finally updating the head request to cover the whole range covered by
558 * the (former) group. All subrequests are removed from any write or commit
559 * lists, unlinked from the group and destroyed.
560 *
561 * Returns a locked, referenced pointer to the head request - which after
562 * this call is guaranteed to be the only request associated with the page.
563 * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
564 * error was encountered.
565 */
566static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
567{
568 struct inode *inode = folio_file_mapping(folio)->host;
569 struct nfs_page *head;
570 struct nfs_commit_info cinfo;
571 int ret;
572
573 nfs_init_cinfo_from_inode(&cinfo, inode);
574 /*
575 * A reference is taken only on the head request which acts as a
576 * reference to the whole page group - the group will not be destroyed
577 * until the head reference is released.
578 */
579 head = nfs_folio_find_and_lock_request(folio);
580 if (IS_ERR_OR_NULL(head))
581 return head;
582
583 /* lock each request in the page group */
584 ret = nfs_page_group_lock_subrequests(head);
585 if (ret < 0) {
586 nfs_unlock_and_release_request(head);
587 return ERR_PTR(ret);
588 }
589
590 nfs_join_page_group(head, &cinfo, inode);
591
592 return head;
593}
594
595static void nfs_write_error(struct nfs_page *req, int error)
596{
597 trace_nfs_write_error(nfs_page_to_inode(req), req, error);
598 nfs_mapping_set_error(nfs_page_to_folio(req), error);
599 nfs_inode_remove_request(req);
600 nfs_page_end_writeback(req);
601 nfs_release_request(req);
602}
603
604/*
605 * Find an associated nfs write request, and prepare to flush it out
606 * May return an error if the user signalled nfs_wait_on_request().
607 */
608static int nfs_page_async_flush(struct folio *folio,
609 struct writeback_control *wbc,
610 struct nfs_pageio_descriptor *pgio)
611{
612 struct nfs_page *req;
613 int ret = 0;
614
615 req = nfs_lock_and_join_requests(folio);
616 if (!req)
617 goto out;
618 ret = PTR_ERR(req);
619 if (IS_ERR(req))
620 goto out;
621
622 nfs_folio_set_writeback(folio);
623 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
624
625 /* If there is a fatal error that covers this write, just exit */
626 ret = pgio->pg_error;
627 if (nfs_error_is_fatal_on_server(ret))
628 goto out_launder;
629
630 ret = 0;
631 if (!nfs_pageio_add_request(pgio, req)) {
632 ret = pgio->pg_error;
633 /*
634 * Remove the problematic req upon fatal errors on the server
635 */
636 if (nfs_error_is_fatal_on_server(ret))
637 goto out_launder;
638 if (wbc->sync_mode == WB_SYNC_NONE)
639 ret = AOP_WRITEPAGE_ACTIVATE;
640 folio_redirty_for_writepage(wbc, folio);
641 nfs_redirty_request(req);
642 pgio->pg_error = 0;
643 } else
644 nfs_add_stats(folio_file_mapping(folio)->host,
645 NFSIOS_WRITEPAGES, 1);
646out:
647 return ret;
648out_launder:
649 nfs_write_error(req, ret);
650 return 0;
651}
652
653static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
654 struct nfs_pageio_descriptor *pgio)
655{
656 nfs_pageio_cond_complete(pgio, folio_index(folio));
657 return nfs_page_async_flush(folio, wbc, pgio);
658}
659
660/*
661 * Write an mmapped page to the server.
662 */
663static int nfs_writepage_locked(struct folio *folio,
664 struct writeback_control *wbc)
665{
666 struct nfs_pageio_descriptor pgio;
667 struct inode *inode = folio_file_mapping(folio)->host;
668 int err;
669
670 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
671 nfs_pageio_init_write(&pgio, inode, 0, false,
672 &nfs_async_write_completion_ops);
673 err = nfs_do_writepage(folio, wbc, &pgio);
674 pgio.pg_error = 0;
675 nfs_pageio_complete(&pgio);
676 return err;
677}
678
679static int nfs_writepages_callback(struct folio *folio,
680 struct writeback_control *wbc, void *data)
681{
682 int ret;
683
684 ret = nfs_do_writepage(folio, wbc, data);
685 if (ret != AOP_WRITEPAGE_ACTIVATE)
686 folio_unlock(folio);
687 return ret;
688}
689
690static void nfs_io_completion_commit(void *inode)
691{
692 nfs_commit_inode(inode, 0);
693}
694
695int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
696{
697 struct inode *inode = mapping->host;
698 struct nfs_pageio_descriptor pgio;
699 struct nfs_io_completion *ioc = NULL;
700 unsigned int mntflags = NFS_SERVER(inode)->flags;
701 int priority = 0;
702 int err;
703
704 if (wbc->sync_mode == WB_SYNC_NONE &&
705 NFS_SERVER(inode)->write_congested)
706 return 0;
707
708 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
709
710 if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
711 wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
712 ioc = nfs_io_completion_alloc(GFP_KERNEL);
713 if (ioc)
714 nfs_io_completion_init(ioc, nfs_io_completion_commit,
715 inode);
716 priority = wb_priority(wbc);
717 }
718
719 do {
720 nfs_pageio_init_write(&pgio, inode, priority, false,
721 &nfs_async_write_completion_ops);
722 pgio.pg_io_completion = ioc;
723 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
724 &pgio);
725 pgio.pg_error = 0;
726 nfs_pageio_complete(&pgio);
727 if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
728 break;
729 } while (err < 0 && !nfs_error_is_fatal(err));
730 nfs_io_completion_put(ioc);
731
732 if (err < 0)
733 goto out_err;
734 return 0;
735out_err:
736 return err;
737}
738
739/*
740 * Insert a write request into an inode
741 */
742static void nfs_inode_add_request(struct nfs_page *req)
743{
744 struct folio *folio = nfs_page_to_folio(req);
745 struct address_space *mapping = folio_file_mapping(folio);
746 struct nfs_inode *nfsi = NFS_I(mapping->host);
747
748 WARN_ON_ONCE(req->wb_this_page != req);
749
750 /* Lock the request! */
751 nfs_lock_request(req);
752
753 /*
754 * Swap-space should not get truncated. Hence no need to plug the race
755 * with invalidate/truncate.
756 */
757 spin_lock(&mapping->i_private_lock);
758 if (likely(!folio_test_swapcache(folio))) {
759 set_bit(PG_MAPPED, &req->wb_flags);
760 folio_set_private(folio);
761 folio->private = req;
762 }
763 spin_unlock(&mapping->i_private_lock);
764 atomic_long_inc(&nfsi->nrequests);
765 /* this a head request for a page group - mark it as having an
766 * extra reference so sub groups can follow suit.
767 * This flag also informs pgio layer when to bump nrequests when
768 * adding subrequests. */
769 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
770 kref_get(&req->wb_kref);
771}
772
773/*
774 * Remove a write request from an inode
775 */
776static void nfs_inode_remove_request(struct nfs_page *req)
777{
778 struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
779
780 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
781 struct folio *folio = nfs_page_to_folio(req->wb_head);
782 struct address_space *mapping = folio_file_mapping(folio);
783
784 spin_lock(&mapping->i_private_lock);
785 if (likely(folio && !folio_test_swapcache(folio))) {
786 folio->private = NULL;
787 folio_clear_private(folio);
788 clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
789 }
790 spin_unlock(&mapping->i_private_lock);
791 }
792
793 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
794 atomic_long_dec(&nfsi->nrequests);
795 nfs_release_request(req);
796 }
797}
798
799static void nfs_mark_request_dirty(struct nfs_page *req)
800{
801 struct folio *folio = nfs_page_to_folio(req);
802 if (folio)
803 filemap_dirty_folio(folio_mapping(folio), folio);
804}
805
806/*
807 * nfs_page_search_commits_for_head_request_locked
808 *
809 * Search through commit lists on @inode for the head request for @folio.
810 * Must be called while holding the inode (which is cinfo) lock.
811 *
812 * Returns the head request if found, or NULL if not found.
813 */
814static struct nfs_page *
815nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
816 struct folio *folio)
817{
818 struct nfs_page *freq, *t;
819 struct nfs_commit_info cinfo;
820 struct inode *inode = &nfsi->vfs_inode;
821
822 nfs_init_cinfo_from_inode(&cinfo, inode);
823
824 /* search through pnfs commit lists */
825 freq = pnfs_search_commit_reqs(inode, &cinfo, folio);
826 if (freq)
827 return freq->wb_head;
828
829 /* Linearly search the commit list for the correct request */
830 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
831 if (nfs_page_to_folio(freq) == folio)
832 return freq->wb_head;
833 }
834
835 return NULL;
836}
837
838/**
839 * nfs_request_add_commit_list_locked - add request to a commit list
840 * @req: pointer to a struct nfs_page
841 * @dst: commit list head
842 * @cinfo: holds list lock and accounting info
843 *
844 * This sets the PG_CLEAN bit, updates the cinfo count of
845 * number of outstanding requests requiring a commit as well as
846 * the MM page stats.
847 *
848 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
849 * nfs_page lock.
850 */
851void
852nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
853 struct nfs_commit_info *cinfo)
854{
855 set_bit(PG_CLEAN, &req->wb_flags);
856 nfs_list_add_request(req, dst);
857 atomic_long_inc(&cinfo->mds->ncommit);
858}
859EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
860
861/**
862 * nfs_request_add_commit_list - add request to a commit list
863 * @req: pointer to a struct nfs_page
864 * @cinfo: holds list lock and accounting info
865 *
866 * This sets the PG_CLEAN bit, updates the cinfo count of
867 * number of outstanding requests requiring a commit as well as
868 * the MM page stats.
869 *
870 * The caller must _not_ hold the cinfo->lock, but must be
871 * holding the nfs_page lock.
872 */
873void
874nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
875{
876 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
877 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
878 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
879 nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
880}
881EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
882
883/**
884 * nfs_request_remove_commit_list - Remove request from a commit list
885 * @req: pointer to a nfs_page
886 * @cinfo: holds list lock and accounting info
887 *
888 * This clears the PG_CLEAN bit, and updates the cinfo's count of
889 * number of outstanding requests requiring a commit
890 * It does not update the MM page stats.
891 *
892 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
893 */
894void
895nfs_request_remove_commit_list(struct nfs_page *req,
896 struct nfs_commit_info *cinfo)
897{
898 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
899 return;
900 nfs_list_remove_request(req);
901 atomic_long_dec(&cinfo->mds->ncommit);
902}
903EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
904
905static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
906 struct inode *inode)
907{
908 cinfo->inode = inode;
909 cinfo->mds = &NFS_I(inode)->commit_info;
910 cinfo->ds = pnfs_get_ds_info(inode);
911 cinfo->dreq = NULL;
912 cinfo->completion_ops = &nfs_commit_completion_ops;
913}
914
915void nfs_init_cinfo(struct nfs_commit_info *cinfo,
916 struct inode *inode,
917 struct nfs_direct_req *dreq)
918{
919 if (dreq)
920 nfs_init_cinfo_from_dreq(cinfo, dreq);
921 else
922 nfs_init_cinfo_from_inode(cinfo, inode);
923}
924EXPORT_SYMBOL_GPL(nfs_init_cinfo);
925
926/*
927 * Add a request to the inode's commit list.
928 */
929void
930nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
931 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
932{
933 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
934 return;
935 nfs_request_add_commit_list(req, cinfo);
936}
937
938static void nfs_folio_clear_commit(struct folio *folio)
939{
940 if (folio) {
941 long nr = folio_nr_pages(folio);
942
943 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
944 wb_stat_mod(&inode_to_bdi(folio_file_mapping(folio)->host)->wb,
945 WB_WRITEBACK, -nr);
946 }
947}
948
949/* Called holding the request lock on @req */
950static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
951 struct nfs_page *req)
952{
953 if (test_bit(PG_CLEAN, &req->wb_flags)) {
954 struct nfs_open_context *ctx = nfs_req_openctx(req);
955 struct inode *inode = d_inode(ctx->dentry);
956
957 mutex_lock(&NFS_I(inode)->commit_mutex);
958 if (!pnfs_clear_request_commit(req, cinfo)) {
959 nfs_request_remove_commit_list(req, cinfo);
960 }
961 mutex_unlock(&NFS_I(inode)->commit_mutex);
962 nfs_folio_clear_commit(nfs_page_to_folio(req));
963 }
964}
965
966int nfs_write_need_commit(struct nfs_pgio_header *hdr)
967{
968 if (hdr->verf.committed == NFS_DATA_SYNC)
969 return hdr->lseg == NULL;
970 return hdr->verf.committed != NFS_FILE_SYNC;
971}
972
973static void nfs_async_write_init(struct nfs_pgio_header *hdr)
974{
975 nfs_io_completion_get(hdr->io_completion);
976}
977
978static void nfs_write_completion(struct nfs_pgio_header *hdr)
979{
980 struct nfs_commit_info cinfo;
981 unsigned long bytes = 0;
982
983 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
984 goto out;
985 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
986 while (!list_empty(&hdr->pages)) {
987 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
988
989 bytes += req->wb_bytes;
990 nfs_list_remove_request(req);
991 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
992 (hdr->good_bytes < bytes)) {
993 trace_nfs_comp_error(hdr->inode, req, hdr->error);
994 nfs_mapping_set_error(nfs_page_to_folio(req),
995 hdr->error);
996 goto remove_req;
997 }
998 if (nfs_write_need_commit(hdr)) {
999 /* Reset wb_nio, since the write was successful. */
1000 req->wb_nio = 0;
1001 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1002 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1003 hdr->pgio_mirror_idx);
1004 goto next;
1005 }
1006remove_req:
1007 nfs_inode_remove_request(req);
1008next:
1009 nfs_page_end_writeback(req);
1010 nfs_release_request(req);
1011 }
1012out:
1013 nfs_io_completion_put(hdr->io_completion);
1014 hdr->release(hdr);
1015}
1016
1017unsigned long
1018nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1019{
1020 return atomic_long_read(&cinfo->mds->ncommit);
1021}
1022
1023/* NFS_I(cinfo->inode)->commit_mutex held by caller */
1024int
1025nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1026 struct nfs_commit_info *cinfo, int max)
1027{
1028 struct nfs_page *req, *tmp;
1029 int ret = 0;
1030
1031 list_for_each_entry_safe(req, tmp, src, wb_list) {
1032 kref_get(&req->wb_kref);
1033 if (!nfs_lock_request(req)) {
1034 nfs_release_request(req);
1035 continue;
1036 }
1037 nfs_request_remove_commit_list(req, cinfo);
1038 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1039 nfs_list_add_request(req, dst);
1040 ret++;
1041 if ((ret == max) && !cinfo->dreq)
1042 break;
1043 cond_resched();
1044 }
1045 return ret;
1046}
1047EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1048
1049/*
1050 * nfs_scan_commit - Scan an inode for commit requests
1051 * @inode: NFS inode to scan
1052 * @dst: mds destination list
1053 * @cinfo: mds and ds lists of reqs ready to commit
1054 *
1055 * Moves requests from the inode's 'commit' request list.
1056 * The requests are *not* checked to ensure that they form a contiguous set.
1057 */
1058int
1059nfs_scan_commit(struct inode *inode, struct list_head *dst,
1060 struct nfs_commit_info *cinfo)
1061{
1062 int ret = 0;
1063
1064 if (!atomic_long_read(&cinfo->mds->ncommit))
1065 return 0;
1066 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1067 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1068 const int max = INT_MAX;
1069
1070 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1071 cinfo, max);
1072 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1073 }
1074 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1075 return ret;
1076}
1077
1078/*
1079 * Search for an existing write request, and attempt to update
1080 * it to reflect a new dirty region on a given page.
1081 *
1082 * If the attempt fails, then the existing request is flushed out
1083 * to disk.
1084 */
1085static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1086 unsigned int offset,
1087 unsigned int bytes)
1088{
1089 struct nfs_page *req;
1090 unsigned int rqend;
1091 unsigned int end;
1092 int error;
1093
1094 end = offset + bytes;
1095
1096 req = nfs_lock_and_join_requests(folio);
1097 if (IS_ERR_OR_NULL(req))
1098 return req;
1099
1100 rqend = req->wb_offset + req->wb_bytes;
1101 /*
1102 * Tell the caller to flush out the request if
1103 * the offsets are non-contiguous.
1104 * Note: nfs_flush_incompatible() will already
1105 * have flushed out requests having wrong owners.
1106 */
1107 if (offset > rqend || end < req->wb_offset)
1108 goto out_flushme;
1109
1110 /* Okay, the request matches. Update the region */
1111 if (offset < req->wb_offset) {
1112 req->wb_offset = offset;
1113 req->wb_pgbase = offset;
1114 }
1115 if (end > rqend)
1116 req->wb_bytes = end - req->wb_offset;
1117 else
1118 req->wb_bytes = rqend - req->wb_offset;
1119 req->wb_nio = 0;
1120 return req;
1121out_flushme:
1122 /*
1123 * Note: we mark the request dirty here because
1124 * nfs_lock_and_join_requests() cannot preserve
1125 * commit flags, so we have to replay the write.
1126 */
1127 nfs_mark_request_dirty(req);
1128 nfs_unlock_and_release_request(req);
1129 error = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1130 return (error < 0) ? ERR_PTR(error) : NULL;
1131}
1132
1133/*
1134 * Try to update an existing write request, or create one if there is none.
1135 *
1136 * Note: Should always be called with the Page Lock held to prevent races
1137 * if we have to add a new request. Also assumes that the caller has
1138 * already called nfs_flush_incompatible() if necessary.
1139 */
1140static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1141 struct folio *folio,
1142 unsigned int offset,
1143 unsigned int bytes)
1144{
1145 struct nfs_page *req;
1146
1147 req = nfs_try_to_update_request(folio, offset, bytes);
1148 if (req != NULL)
1149 goto out;
1150 req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1151 if (IS_ERR(req))
1152 goto out;
1153 nfs_inode_add_request(req);
1154out:
1155 return req;
1156}
1157
1158static int nfs_writepage_setup(struct nfs_open_context *ctx,
1159 struct folio *folio, unsigned int offset,
1160 unsigned int count)
1161{
1162 struct nfs_page *req;
1163
1164 req = nfs_setup_write_request(ctx, folio, offset, count);
1165 if (IS_ERR(req))
1166 return PTR_ERR(req);
1167 /* Update file length */
1168 nfs_grow_file(folio, offset, count);
1169 nfs_mark_uptodate(req);
1170 nfs_mark_request_dirty(req);
1171 nfs_unlock_and_release_request(req);
1172 return 0;
1173}
1174
1175int nfs_flush_incompatible(struct file *file, struct folio *folio)
1176{
1177 struct nfs_open_context *ctx = nfs_file_open_context(file);
1178 struct nfs_lock_context *l_ctx;
1179 struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1180 struct nfs_page *req;
1181 int do_flush, status;
1182 /*
1183 * Look for a request corresponding to this page. If there
1184 * is one, and it belongs to another file, we flush it out
1185 * before we try to copy anything into the page. Do this
1186 * due to the lack of an ACCESS-type call in NFSv2.
1187 * Also do the same if we find a request from an existing
1188 * dropped page.
1189 */
1190 do {
1191 req = nfs_folio_find_head_request(folio);
1192 if (req == NULL)
1193 return 0;
1194 l_ctx = req->wb_lock_context;
1195 do_flush = nfs_page_to_folio(req) != folio ||
1196 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1197 if (l_ctx && flctx &&
1198 !(list_empty_careful(&flctx->flc_posix) &&
1199 list_empty_careful(&flctx->flc_flock))) {
1200 do_flush |= l_ctx->lockowner != current->files;
1201 }
1202 nfs_release_request(req);
1203 if (!do_flush)
1204 return 0;
1205 status = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1206 } while (status == 0);
1207 return status;
1208}
1209
1210/*
1211 * Avoid buffered writes when a open context credential's key would
1212 * expire soon.
1213 *
1214 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1215 *
1216 * Return 0 and set a credential flag which triggers the inode to flush
1217 * and performs NFS_FILE_SYNC writes if the key will expired within
1218 * RPC_KEY_EXPIRE_TIMEO.
1219 */
1220int
1221nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1222{
1223 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1224
1225 if (nfs_ctx_key_to_expire(ctx, inode) &&
1226 !rcu_access_pointer(ctx->ll_cred))
1227 /* Already expired! */
1228 return -EACCES;
1229 return 0;
1230}
1231
1232/*
1233 * Test if the open context credential key is marked to expire soon.
1234 */
1235bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1236{
1237 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1238 struct rpc_cred *cred, *new, *old = NULL;
1239 struct auth_cred acred = {
1240 .cred = ctx->cred,
1241 };
1242 bool ret = false;
1243
1244 rcu_read_lock();
1245 cred = rcu_dereference(ctx->ll_cred);
1246 if (cred && !(cred->cr_ops->crkey_timeout &&
1247 cred->cr_ops->crkey_timeout(cred)))
1248 goto out;
1249 rcu_read_unlock();
1250
1251 new = auth->au_ops->lookup_cred(auth, &acred, 0);
1252 if (new == cred) {
1253 put_rpccred(new);
1254 return true;
1255 }
1256 if (IS_ERR_OR_NULL(new)) {
1257 new = NULL;
1258 ret = true;
1259 } else if (new->cr_ops->crkey_timeout &&
1260 new->cr_ops->crkey_timeout(new))
1261 ret = true;
1262
1263 rcu_read_lock();
1264 old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1265 RCU_INITIALIZER(new)), 1);
1266out:
1267 rcu_read_unlock();
1268 put_rpccred(old);
1269 return ret;
1270}
1271
1272/*
1273 * If the page cache is marked as unsafe or invalid, then we can't rely on
1274 * the PageUptodate() flag. In this case, we will need to turn off
1275 * write optimisations that depend on the page contents being correct.
1276 */
1277static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1278{
1279 struct inode *inode = folio_file_mapping(folio)->host;
1280 struct nfs_inode *nfsi = NFS_I(inode);
1281
1282 if (nfs_have_delegated_attributes(inode))
1283 goto out;
1284 if (nfsi->cache_validity &
1285 (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1286 return false;
1287 smp_rmb();
1288 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1289 return false;
1290out:
1291 if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1292 return false;
1293 return folio_test_uptodate(folio) != 0;
1294}
1295
1296static bool
1297is_whole_file_wrlock(struct file_lock *fl)
1298{
1299 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1300 lock_is_write(fl);
1301}
1302
1303/* If we know the page is up to date, and we're not using byte range locks (or
1304 * if we have the whole file locked for writing), it may be more efficient to
1305 * extend the write to cover the entire page in order to avoid fragmentation
1306 * inefficiencies.
1307 *
1308 * If the file is opened for synchronous writes then we can just skip the rest
1309 * of the checks.
1310 */
1311static int nfs_can_extend_write(struct file *file, struct folio *folio,
1312 unsigned int pagelen)
1313{
1314 struct inode *inode = file_inode(file);
1315 struct file_lock_context *flctx = locks_inode_context(inode);
1316 struct file_lock *fl;
1317 int ret;
1318
1319 if (file->f_flags & O_DSYNC)
1320 return 0;
1321 if (!nfs_folio_write_uptodate(folio, pagelen))
1322 return 0;
1323 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1324 return 1;
1325 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1326 list_empty_careful(&flctx->flc_posix)))
1327 return 1;
1328
1329 /* Check to see if there are whole file write locks */
1330 ret = 0;
1331 spin_lock(&flctx->flc_lock);
1332 if (!list_empty(&flctx->flc_posix)) {
1333 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1334 c.flc_list);
1335 if (is_whole_file_wrlock(fl))
1336 ret = 1;
1337 } else if (!list_empty(&flctx->flc_flock)) {
1338 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1339 c.flc_list);
1340 if (lock_is_write(fl))
1341 ret = 1;
1342 }
1343 spin_unlock(&flctx->flc_lock);
1344 return ret;
1345}
1346
1347/*
1348 * Update and possibly write a cached page of an NFS file.
1349 *
1350 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1351 * things with a page scheduled for an RPC call (e.g. invalidate it).
1352 */
1353int nfs_update_folio(struct file *file, struct folio *folio,
1354 unsigned int offset, unsigned int count)
1355{
1356 struct nfs_open_context *ctx = nfs_file_open_context(file);
1357 struct address_space *mapping = folio_file_mapping(folio);
1358 struct inode *inode = mapping->host;
1359 unsigned int pagelen = nfs_folio_length(folio);
1360 int status = 0;
1361
1362 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1363
1364 dprintk("NFS: nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1365 (long long)(folio_file_pos(folio) + offset));
1366
1367 if (!count)
1368 goto out;
1369
1370 if (nfs_can_extend_write(file, folio, pagelen)) {
1371 count = max(count + offset, pagelen);
1372 offset = 0;
1373 }
1374
1375 status = nfs_writepage_setup(ctx, folio, offset, count);
1376 if (status < 0)
1377 nfs_set_pageerror(mapping);
1378out:
1379 dprintk("NFS: nfs_update_folio returns %d (isize %lld)\n",
1380 status, (long long)i_size_read(inode));
1381 return status;
1382}
1383
1384static int flush_task_priority(int how)
1385{
1386 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1387 case FLUSH_HIGHPRI:
1388 return RPC_PRIORITY_HIGH;
1389 case FLUSH_LOWPRI:
1390 return RPC_PRIORITY_LOW;
1391 }
1392 return RPC_PRIORITY_NORMAL;
1393}
1394
1395static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1396 struct rpc_message *msg,
1397 const struct nfs_rpc_ops *rpc_ops,
1398 struct rpc_task_setup *task_setup_data, int how)
1399{
1400 int priority = flush_task_priority(how);
1401
1402 if (IS_SWAPFILE(hdr->inode))
1403 task_setup_data->flags |= RPC_TASK_SWAPPER;
1404 task_setup_data->priority = priority;
1405 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1406 trace_nfs_initiate_write(hdr);
1407}
1408
1409/* If a nfs_flush_* function fails, it should remove reqs from @head and
1410 * call this on each, which will prepare them to be retried on next
1411 * writeback using standard nfs.
1412 */
1413static void nfs_redirty_request(struct nfs_page *req)
1414{
1415 struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1416
1417 /* Bump the transmission count */
1418 req->wb_nio++;
1419 nfs_mark_request_dirty(req);
1420 atomic_long_inc(&nfsi->redirtied_pages);
1421 nfs_page_end_writeback(req);
1422 nfs_release_request(req);
1423}
1424
1425static void nfs_async_write_error(struct list_head *head, int error)
1426{
1427 struct nfs_page *req;
1428
1429 while (!list_empty(head)) {
1430 req = nfs_list_entry(head->next);
1431 nfs_list_remove_request(req);
1432 if (nfs_error_is_fatal_on_server(error))
1433 nfs_write_error(req, error);
1434 else
1435 nfs_redirty_request(req);
1436 }
1437}
1438
1439static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1440{
1441 nfs_async_write_error(&hdr->pages, 0);
1442}
1443
1444static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1445 .init_hdr = nfs_async_write_init,
1446 .error_cleanup = nfs_async_write_error,
1447 .completion = nfs_write_completion,
1448 .reschedule_io = nfs_async_write_reschedule_io,
1449};
1450
1451void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1452 struct inode *inode, int ioflags, bool force_mds,
1453 const struct nfs_pgio_completion_ops *compl_ops)
1454{
1455 struct nfs_server *server = NFS_SERVER(inode);
1456 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1457
1458#ifdef CONFIG_NFS_V4_1
1459 if (server->pnfs_curr_ld && !force_mds)
1460 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1461#endif
1462 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1463 server->wsize, ioflags);
1464}
1465EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1466
1467void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1468{
1469 struct nfs_pgio_mirror *mirror;
1470
1471 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1472 pgio->pg_ops->pg_cleanup(pgio);
1473
1474 pgio->pg_ops = &nfs_pgio_rw_ops;
1475
1476 nfs_pageio_stop_mirroring(pgio);
1477
1478 mirror = &pgio->pg_mirrors[0];
1479 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1480}
1481EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1482
1483
1484void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1485{
1486 struct nfs_commit_data *data = calldata;
1487
1488 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1489}
1490
1491static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1492 struct nfs_fattr *fattr)
1493{
1494 struct nfs_pgio_args *argp = &hdr->args;
1495 struct nfs_pgio_res *resp = &hdr->res;
1496 u64 size = argp->offset + resp->count;
1497
1498 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1499 fattr->size = size;
1500 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1501 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1502 return;
1503 }
1504 if (size != fattr->size)
1505 return;
1506 /* Set attribute barrier */
1507 nfs_fattr_set_barrier(fattr);
1508 /* ...and update size */
1509 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1510}
1511
1512void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1513{
1514 struct nfs_fattr *fattr = &hdr->fattr;
1515 struct inode *inode = hdr->inode;
1516
1517 spin_lock(&inode->i_lock);
1518 nfs_writeback_check_extend(hdr, fattr);
1519 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1520 spin_unlock(&inode->i_lock);
1521}
1522EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1523
1524/*
1525 * This function is called when the WRITE call is complete.
1526 */
1527static int nfs_writeback_done(struct rpc_task *task,
1528 struct nfs_pgio_header *hdr,
1529 struct inode *inode)
1530{
1531 int status;
1532
1533 /*
1534 * ->write_done will attempt to use post-op attributes to detect
1535 * conflicting writes by other clients. A strict interpretation
1536 * of close-to-open would allow us to continue caching even if
1537 * another writer had changed the file, but some applications
1538 * depend on tighter cache coherency when writing.
1539 */
1540 status = NFS_PROTO(inode)->write_done(task, hdr);
1541 if (status != 0)
1542 return status;
1543
1544 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1545 trace_nfs_writeback_done(task, hdr);
1546
1547 if (task->tk_status >= 0) {
1548 enum nfs3_stable_how committed = hdr->res.verf->committed;
1549
1550 if (committed == NFS_UNSTABLE) {
1551 /*
1552 * We have some uncommitted data on the server at
1553 * this point, so ensure that we keep track of that
1554 * fact irrespective of what later writes do.
1555 */
1556 set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1557 }
1558
1559 if (committed < hdr->args.stable) {
1560 /* We tried a write call, but the server did not
1561 * commit data to stable storage even though we
1562 * requested it.
1563 * Note: There is a known bug in Tru64 < 5.0 in which
1564 * the server reports NFS_DATA_SYNC, but performs
1565 * NFS_FILE_SYNC. We therefore implement this checking
1566 * as a dprintk() in order to avoid filling syslog.
1567 */
1568 static unsigned long complain;
1569
1570 /* Note this will print the MDS for a DS write */
1571 if (time_before(complain, jiffies)) {
1572 dprintk("NFS: faulty NFS server %s:"
1573 " (committed = %d) != (stable = %d)\n",
1574 NFS_SERVER(inode)->nfs_client->cl_hostname,
1575 committed, hdr->args.stable);
1576 complain = jiffies + 300 * HZ;
1577 }
1578 }
1579 }
1580
1581 /* Deal with the suid/sgid bit corner case */
1582 if (nfs_should_remove_suid(inode)) {
1583 spin_lock(&inode->i_lock);
1584 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1585 spin_unlock(&inode->i_lock);
1586 }
1587 return 0;
1588}
1589
1590/*
1591 * This function is called when the WRITE call is complete.
1592 */
1593static void nfs_writeback_result(struct rpc_task *task,
1594 struct nfs_pgio_header *hdr)
1595{
1596 struct nfs_pgio_args *argp = &hdr->args;
1597 struct nfs_pgio_res *resp = &hdr->res;
1598
1599 if (resp->count < argp->count) {
1600 static unsigned long complain;
1601
1602 /* This a short write! */
1603 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1604
1605 /* Has the server at least made some progress? */
1606 if (resp->count == 0) {
1607 if (time_before(complain, jiffies)) {
1608 printk(KERN_WARNING
1609 "NFS: Server wrote zero bytes, expected %u.\n",
1610 argp->count);
1611 complain = jiffies + 300 * HZ;
1612 }
1613 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1614 task->tk_status = -EIO;
1615 return;
1616 }
1617
1618 /* For non rpc-based layout drivers, retry-through-MDS */
1619 if (!task->tk_ops) {
1620 hdr->pnfs_error = -EAGAIN;
1621 return;
1622 }
1623
1624 /* Was this an NFSv2 write or an NFSv3 stable write? */
1625 if (resp->verf->committed != NFS_UNSTABLE) {
1626 /* Resend from where the server left off */
1627 hdr->mds_offset += resp->count;
1628 argp->offset += resp->count;
1629 argp->pgbase += resp->count;
1630 argp->count -= resp->count;
1631 } else {
1632 /* Resend as a stable write in order to avoid
1633 * headaches in the case of a server crash.
1634 */
1635 argp->stable = NFS_FILE_SYNC;
1636 }
1637 resp->count = 0;
1638 resp->verf->committed = 0;
1639 rpc_restart_call_prepare(task);
1640 }
1641}
1642
1643static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1644{
1645 return wait_var_event_killable(&cinfo->rpcs_out,
1646 !atomic_read(&cinfo->rpcs_out));
1647}
1648
1649void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1650{
1651 atomic_inc(&cinfo->rpcs_out);
1652}
1653
1654bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1655{
1656 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1657 wake_up_var(&cinfo->rpcs_out);
1658 return true;
1659 }
1660 return false;
1661}
1662
1663void nfs_commitdata_release(struct nfs_commit_data *data)
1664{
1665 put_nfs_open_context(data->context);
1666 nfs_commit_free(data);
1667}
1668EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1669
1670int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1671 const struct nfs_rpc_ops *nfs_ops,
1672 const struct rpc_call_ops *call_ops,
1673 int how, int flags)
1674{
1675 struct rpc_task *task;
1676 int priority = flush_task_priority(how);
1677 struct rpc_message msg = {
1678 .rpc_argp = &data->args,
1679 .rpc_resp = &data->res,
1680 .rpc_cred = data->cred,
1681 };
1682 struct rpc_task_setup task_setup_data = {
1683 .task = &data->task,
1684 .rpc_client = clnt,
1685 .rpc_message = &msg,
1686 .callback_ops = call_ops,
1687 .callback_data = data,
1688 .workqueue = nfsiod_workqueue,
1689 .flags = RPC_TASK_ASYNC | flags,
1690 .priority = priority,
1691 };
1692
1693 if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1694 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1695
1696 /* Set up the initial task struct. */
1697 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1698 trace_nfs_initiate_commit(data);
1699
1700 dprintk("NFS: initiated commit call\n");
1701
1702 task = rpc_run_task(&task_setup_data);
1703 if (IS_ERR(task))
1704 return PTR_ERR(task);
1705 if (how & FLUSH_SYNC)
1706 rpc_wait_for_completion_task(task);
1707 rpc_put_task(task);
1708 return 0;
1709}
1710EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1711
1712static loff_t nfs_get_lwb(struct list_head *head)
1713{
1714 loff_t lwb = 0;
1715 struct nfs_page *req;
1716
1717 list_for_each_entry(req, head, wb_list)
1718 if (lwb < (req_offset(req) + req->wb_bytes))
1719 lwb = req_offset(req) + req->wb_bytes;
1720
1721 return lwb;
1722}
1723
1724/*
1725 * Set up the argument/result storage required for the RPC call.
1726 */
1727void nfs_init_commit(struct nfs_commit_data *data,
1728 struct list_head *head,
1729 struct pnfs_layout_segment *lseg,
1730 struct nfs_commit_info *cinfo)
1731{
1732 struct nfs_page *first;
1733 struct nfs_open_context *ctx;
1734 struct inode *inode;
1735
1736 /* Set up the RPC argument and reply structs
1737 * NB: take care not to mess about with data->commit et al. */
1738
1739 if (head)
1740 list_splice_init(head, &data->pages);
1741
1742 first = nfs_list_entry(data->pages.next);
1743 ctx = nfs_req_openctx(first);
1744 inode = d_inode(ctx->dentry);
1745
1746 data->inode = inode;
1747 data->cred = ctx->cred;
1748 data->lseg = lseg; /* reference transferred */
1749 /* only set lwb for pnfs commit */
1750 if (lseg)
1751 data->lwb = nfs_get_lwb(&data->pages);
1752 data->mds_ops = &nfs_commit_ops;
1753 data->completion_ops = cinfo->completion_ops;
1754 data->dreq = cinfo->dreq;
1755
1756 data->args.fh = NFS_FH(data->inode);
1757 /* Note: we always request a commit of the entire inode */
1758 data->args.offset = 0;
1759 data->args.count = 0;
1760 data->context = get_nfs_open_context(ctx);
1761 data->res.fattr = &data->fattr;
1762 data->res.verf = &data->verf;
1763 nfs_fattr_init(&data->fattr);
1764 nfs_commit_begin(cinfo->mds);
1765}
1766EXPORT_SYMBOL_GPL(nfs_init_commit);
1767
1768void nfs_retry_commit(struct list_head *page_list,
1769 struct pnfs_layout_segment *lseg,
1770 struct nfs_commit_info *cinfo,
1771 u32 ds_commit_idx)
1772{
1773 struct nfs_page *req;
1774
1775 while (!list_empty(page_list)) {
1776 req = nfs_list_entry(page_list->next);
1777 nfs_list_remove_request(req);
1778 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1779 nfs_folio_clear_commit(nfs_page_to_folio(req));
1780 nfs_unlock_and_release_request(req);
1781 }
1782}
1783EXPORT_SYMBOL_GPL(nfs_retry_commit);
1784
1785static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1786 struct nfs_page *req)
1787{
1788 struct folio *folio = nfs_page_to_folio(req);
1789
1790 filemap_dirty_folio(folio_mapping(folio), folio);
1791}
1792
1793/*
1794 * Commit dirty pages
1795 */
1796static int
1797nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1798 struct nfs_commit_info *cinfo)
1799{
1800 struct nfs_commit_data *data;
1801 unsigned short task_flags = 0;
1802
1803 /* another commit raced with us */
1804 if (list_empty(head))
1805 return 0;
1806
1807 data = nfs_commitdata_alloc();
1808 if (!data) {
1809 nfs_retry_commit(head, NULL, cinfo, -1);
1810 return -ENOMEM;
1811 }
1812
1813 /* Set up the argument struct */
1814 nfs_init_commit(data, head, NULL, cinfo);
1815 if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1816 task_flags = RPC_TASK_MOVEABLE;
1817 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1818 data->mds_ops, how,
1819 RPC_TASK_CRED_NOREF | task_flags);
1820}
1821
1822/*
1823 * COMMIT call returned
1824 */
1825static void nfs_commit_done(struct rpc_task *task, void *calldata)
1826{
1827 struct nfs_commit_data *data = calldata;
1828
1829 /* Call the NFS version-specific code */
1830 NFS_PROTO(data->inode)->commit_done(task, data);
1831 trace_nfs_commit_done(task, data);
1832}
1833
1834static void nfs_commit_release_pages(struct nfs_commit_data *data)
1835{
1836 const struct nfs_writeverf *verf = data->res.verf;
1837 struct nfs_page *req;
1838 int status = data->task.tk_status;
1839 struct nfs_commit_info cinfo;
1840 struct nfs_server *nfss;
1841 struct folio *folio;
1842
1843 while (!list_empty(&data->pages)) {
1844 req = nfs_list_entry(data->pages.next);
1845 nfs_list_remove_request(req);
1846 folio = nfs_page_to_folio(req);
1847 nfs_folio_clear_commit(folio);
1848
1849 dprintk("NFS: commit (%s/%llu %d@%lld)",
1850 nfs_req_openctx(req)->dentry->d_sb->s_id,
1851 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1852 req->wb_bytes,
1853 (long long)req_offset(req));
1854 if (status < 0) {
1855 if (folio) {
1856 trace_nfs_commit_error(data->inode, req,
1857 status);
1858 nfs_mapping_set_error(folio, status);
1859 nfs_inode_remove_request(req);
1860 }
1861 dprintk_cont(", error = %d\n", status);
1862 goto next;
1863 }
1864
1865 /* Okay, COMMIT succeeded, apparently. Check the verifier
1866 * returned by the server against all stored verfs. */
1867 if (nfs_write_match_verf(verf, req)) {
1868 /* We have a match */
1869 if (folio)
1870 nfs_inode_remove_request(req);
1871 dprintk_cont(" OK\n");
1872 goto next;
1873 }
1874 /* We have a mismatch. Write the page again */
1875 dprintk_cont(" mismatch\n");
1876 nfs_mark_request_dirty(req);
1877 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1878 next:
1879 nfs_unlock_and_release_request(req);
1880 /* Latency breaker */
1881 cond_resched();
1882 }
1883 nfss = NFS_SERVER(data->inode);
1884 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1885 nfss->write_congested = 0;
1886
1887 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1888 nfs_commit_end(cinfo.mds);
1889}
1890
1891static void nfs_commit_release(void *calldata)
1892{
1893 struct nfs_commit_data *data = calldata;
1894
1895 data->completion_ops->completion(data);
1896 nfs_commitdata_release(calldata);
1897}
1898
1899static const struct rpc_call_ops nfs_commit_ops = {
1900 .rpc_call_prepare = nfs_commit_prepare,
1901 .rpc_call_done = nfs_commit_done,
1902 .rpc_release = nfs_commit_release,
1903};
1904
1905static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1906 .completion = nfs_commit_release_pages,
1907 .resched_write = nfs_commit_resched_write,
1908};
1909
1910int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1911 int how, struct nfs_commit_info *cinfo)
1912{
1913 int status;
1914
1915 status = pnfs_commit_list(inode, head, how, cinfo);
1916 if (status == PNFS_NOT_ATTEMPTED)
1917 status = nfs_commit_list(inode, head, how, cinfo);
1918 return status;
1919}
1920
1921static int __nfs_commit_inode(struct inode *inode, int how,
1922 struct writeback_control *wbc)
1923{
1924 LIST_HEAD(head);
1925 struct nfs_commit_info cinfo;
1926 int may_wait = how & FLUSH_SYNC;
1927 int ret, nscan;
1928
1929 how &= ~FLUSH_SYNC;
1930 nfs_init_cinfo_from_inode(&cinfo, inode);
1931 nfs_commit_begin(cinfo.mds);
1932 for (;;) {
1933 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1934 if (ret <= 0)
1935 break;
1936 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1937 if (ret < 0)
1938 break;
1939 ret = 0;
1940 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1941 if (nscan < wbc->nr_to_write)
1942 wbc->nr_to_write -= nscan;
1943 else
1944 wbc->nr_to_write = 0;
1945 }
1946 if (nscan < INT_MAX)
1947 break;
1948 cond_resched();
1949 }
1950 nfs_commit_end(cinfo.mds);
1951 if (ret || !may_wait)
1952 return ret;
1953 return wait_on_commit(cinfo.mds);
1954}
1955
1956int nfs_commit_inode(struct inode *inode, int how)
1957{
1958 return __nfs_commit_inode(inode, how, NULL);
1959}
1960EXPORT_SYMBOL_GPL(nfs_commit_inode);
1961
1962int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1963{
1964 struct nfs_inode *nfsi = NFS_I(inode);
1965 int flags = FLUSH_SYNC;
1966 int ret = 0;
1967
1968 if (wbc->sync_mode == WB_SYNC_NONE) {
1969 /* no commits means nothing needs to be done */
1970 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1971 goto check_requests_outstanding;
1972
1973 /* Don't commit yet if this is a non-blocking flush and there
1974 * are a lot of outstanding writes for this mapping.
1975 */
1976 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1977 goto out_mark_dirty;
1978
1979 /* don't wait for the COMMIT response */
1980 flags = 0;
1981 }
1982
1983 ret = __nfs_commit_inode(inode, flags, wbc);
1984 if (!ret) {
1985 if (flags & FLUSH_SYNC)
1986 return 0;
1987 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1988 goto out_mark_dirty;
1989
1990check_requests_outstanding:
1991 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1992 return ret;
1993out_mark_dirty:
1994 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1995 return ret;
1996}
1997EXPORT_SYMBOL_GPL(nfs_write_inode);
1998
1999/*
2000 * Wrapper for filemap_write_and_wait_range()
2001 *
2002 * Needed for pNFS in order to ensure data becomes visible to the
2003 * client.
2004 */
2005int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2006 loff_t lstart, loff_t lend)
2007{
2008 int ret;
2009
2010 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2011 if (ret == 0)
2012 ret = pnfs_sync_inode(mapping->host, true);
2013 return ret;
2014}
2015EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2016
2017/*
2018 * flush the inode to disk.
2019 */
2020int nfs_wb_all(struct inode *inode)
2021{
2022 int ret;
2023
2024 trace_nfs_writeback_inode_enter(inode);
2025
2026 ret = filemap_write_and_wait(inode->i_mapping);
2027 if (ret)
2028 goto out;
2029 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2030 if (ret < 0)
2031 goto out;
2032 pnfs_sync_inode(inode, true);
2033 ret = 0;
2034
2035out:
2036 trace_nfs_writeback_inode_exit(inode, ret);
2037 return ret;
2038}
2039EXPORT_SYMBOL_GPL(nfs_wb_all);
2040
2041int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2042{
2043 struct nfs_page *req;
2044 int ret = 0;
2045
2046 folio_wait_writeback(folio);
2047
2048 /* blocking call to cancel all requests and join to a single (head)
2049 * request */
2050 req = nfs_lock_and_join_requests(folio);
2051
2052 if (IS_ERR(req)) {
2053 ret = PTR_ERR(req);
2054 } else if (req) {
2055 /* all requests from this folio have been cancelled by
2056 * nfs_lock_and_join_requests, so just remove the head
2057 * request from the inode / page_private pointer and
2058 * release it */
2059 nfs_inode_remove_request(req);
2060 nfs_unlock_and_release_request(req);
2061 }
2062
2063 return ret;
2064}
2065
2066/**
2067 * nfs_wb_folio - Write back all requests on one page
2068 * @inode: pointer to page
2069 * @folio: pointer to folio
2070 *
2071 * Assumes that the folio has been locked by the caller, and will
2072 * not unlock it.
2073 */
2074int nfs_wb_folio(struct inode *inode, struct folio *folio)
2075{
2076 loff_t range_start = folio_file_pos(folio);
2077 loff_t range_end = range_start + (loff_t)folio_size(folio) - 1;
2078 struct writeback_control wbc = {
2079 .sync_mode = WB_SYNC_ALL,
2080 .nr_to_write = 0,
2081 .range_start = range_start,
2082 .range_end = range_end,
2083 };
2084 int ret;
2085
2086 trace_nfs_writeback_folio(inode, folio);
2087
2088 for (;;) {
2089 folio_wait_writeback(folio);
2090 if (folio_clear_dirty_for_io(folio)) {
2091 ret = nfs_writepage_locked(folio, &wbc);
2092 if (ret < 0)
2093 goto out_error;
2094 continue;
2095 }
2096 ret = 0;
2097 if (!folio_test_private(folio))
2098 break;
2099 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2100 if (ret < 0)
2101 goto out_error;
2102 }
2103out_error:
2104 trace_nfs_writeback_folio_done(inode, folio, ret);
2105 return ret;
2106}
2107
2108#ifdef CONFIG_MIGRATION
2109int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2110 struct folio *src, enum migrate_mode mode)
2111{
2112 /*
2113 * If the private flag is set, the folio is currently associated with
2114 * an in-progress read or write request. Don't try to migrate it.
2115 *
2116 * FIXME: we could do this in principle, but we'll need a way to ensure
2117 * that we can safely release the inode reference while holding
2118 * the folio lock.
2119 */
2120 if (folio_test_private(src))
2121 return -EBUSY;
2122
2123 if (folio_test_fscache(src)) {
2124 if (mode == MIGRATE_ASYNC)
2125 return -EBUSY;
2126 folio_wait_fscache(src);
2127 }
2128
2129 return migrate_folio(mapping, dst, src, mode);
2130}
2131#endif
2132
2133int __init nfs_init_writepagecache(void)
2134{
2135 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2136 sizeof(struct nfs_pgio_header),
2137 0, SLAB_HWCACHE_ALIGN,
2138 NULL);
2139 if (nfs_wdata_cachep == NULL)
2140 return -ENOMEM;
2141
2142 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2143 nfs_wdata_cachep);
2144 if (nfs_wdata_mempool == NULL)
2145 goto out_destroy_write_cache;
2146
2147 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2148 sizeof(struct nfs_commit_data),
2149 0, SLAB_HWCACHE_ALIGN,
2150 NULL);
2151 if (nfs_cdata_cachep == NULL)
2152 goto out_destroy_write_mempool;
2153
2154 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2155 nfs_cdata_cachep);
2156 if (nfs_commit_mempool == NULL)
2157 goto out_destroy_commit_cache;
2158
2159 /*
2160 * NFS congestion size, scale with available memory.
2161 *
2162 * 64MB: 8192k
2163 * 128MB: 11585k
2164 * 256MB: 16384k
2165 * 512MB: 23170k
2166 * 1GB: 32768k
2167 * 2GB: 46340k
2168 * 4GB: 65536k
2169 * 8GB: 92681k
2170 * 16GB: 131072k
2171 *
2172 * This allows larger machines to have larger/more transfers.
2173 * Limit the default to 256M
2174 */
2175 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2176 if (nfs_congestion_kb > 256*1024)
2177 nfs_congestion_kb = 256*1024;
2178
2179 return 0;
2180
2181out_destroy_commit_cache:
2182 kmem_cache_destroy(nfs_cdata_cachep);
2183out_destroy_write_mempool:
2184 mempool_destroy(nfs_wdata_mempool);
2185out_destroy_write_cache:
2186 kmem_cache_destroy(nfs_wdata_cachep);
2187 return -ENOMEM;
2188}
2189
2190void nfs_destroy_writepagecache(void)
2191{
2192 mempool_destroy(nfs_commit_mempool);
2193 kmem_cache_destroy(nfs_cdata_cachep);
2194 mempool_destroy(nfs_wdata_mempool);
2195 kmem_cache_destroy(nfs_wdata_cachep);
2196}
2197
2198static const struct nfs_rw_ops nfs_rw_write_ops = {
2199 .rw_alloc_header = nfs_writehdr_alloc,
2200 .rw_free_header = nfs_writehdr_free,
2201 .rw_done = nfs_writeback_done,
2202 .rw_result = nfs_writeback_result,
2203 .rw_initiate = nfs_initiate_write,
2204};