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#include <linux/export.h>
24
25#include <asm/uaccess.h>
26
27#include "delegation.h"
28#include "internal.h"
29#include "iostat.h"
30#include "nfs4_fs.h"
31#include "fscache.h"
32#include "pnfs.h"
33
34#define NFSDBG_FACILITY NFSDBG_PAGECACHE
35
36#define MIN_POOL_WRITE (32)
37#define MIN_POOL_COMMIT (4)
38
39/*
40 * Local function declarations
41 */
42static void nfs_redirty_request(struct nfs_page *req);
43static const struct rpc_call_ops nfs_write_common_ops;
44static const struct rpc_call_ops nfs_commit_ops;
45static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
46static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
47
48static struct kmem_cache *nfs_wdata_cachep;
49static mempool_t *nfs_wdata_mempool;
50static struct kmem_cache *nfs_cdata_cachep;
51static mempool_t *nfs_commit_mempool;
52
53struct nfs_commit_data *nfs_commitdata_alloc(void)
54{
55 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
56
57 if (p) {
58 memset(p, 0, sizeof(*p));
59 INIT_LIST_HEAD(&p->pages);
60 }
61 return p;
62}
63EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
64
65void nfs_commit_free(struct nfs_commit_data *p)
66{
67 mempool_free(p, nfs_commit_mempool);
68}
69EXPORT_SYMBOL_GPL(nfs_commit_free);
70
71struct nfs_write_header *nfs_writehdr_alloc(void)
72{
73 struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
74
75 if (p) {
76 struct nfs_pgio_header *hdr = &p->header;
77
78 memset(p, 0, sizeof(*p));
79 INIT_LIST_HEAD(&hdr->pages);
80 INIT_LIST_HEAD(&hdr->rpc_list);
81 spin_lock_init(&hdr->lock);
82 atomic_set(&hdr->refcnt, 0);
83 hdr->verf = &p->verf;
84 }
85 return p;
86}
87
88static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
89 unsigned int pagecount)
90{
91 struct nfs_write_data *data, *prealloc;
92
93 prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
94 if (prealloc->header == NULL)
95 data = prealloc;
96 else
97 data = kzalloc(sizeof(*data), GFP_KERNEL);
98 if (!data)
99 goto out;
100
101 if (nfs_pgarray_set(&data->pages, pagecount)) {
102 data->header = hdr;
103 atomic_inc(&hdr->refcnt);
104 } else {
105 if (data != prealloc)
106 kfree(data);
107 data = NULL;
108 }
109out:
110 return data;
111}
112
113void nfs_writehdr_free(struct nfs_pgio_header *hdr)
114{
115 struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
116 mempool_free(whdr, nfs_wdata_mempool);
117}
118
119void nfs_writedata_release(struct nfs_write_data *wdata)
120{
121 struct nfs_pgio_header *hdr = wdata->header;
122 struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
123
124 put_nfs_open_context(wdata->args.context);
125 if (wdata->pages.pagevec != wdata->pages.page_array)
126 kfree(wdata->pages.pagevec);
127 if (wdata != &write_header->rpc_data)
128 kfree(wdata);
129 else
130 wdata->header = NULL;
131 if (atomic_dec_and_test(&hdr->refcnt))
132 hdr->completion_ops->completion(hdr);
133}
134
135static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
136{
137 ctx->error = error;
138 smp_wmb();
139 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
140}
141
142static struct nfs_page *nfs_page_find_request_locked(struct page *page)
143{
144 struct nfs_page *req = NULL;
145
146 if (PagePrivate(page)) {
147 req = (struct nfs_page *)page_private(page);
148 if (req != NULL)
149 kref_get(&req->wb_kref);
150 }
151 return req;
152}
153
154static struct nfs_page *nfs_page_find_request(struct page *page)
155{
156 struct inode *inode = page->mapping->host;
157 struct nfs_page *req = NULL;
158
159 spin_lock(&inode->i_lock);
160 req = nfs_page_find_request_locked(page);
161 spin_unlock(&inode->i_lock);
162 return req;
163}
164
165/* Adjust the file length if we're writing beyond the end */
166static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
167{
168 struct inode *inode = page->mapping->host;
169 loff_t end, i_size;
170 pgoff_t end_index;
171
172 spin_lock(&inode->i_lock);
173 i_size = i_size_read(inode);
174 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
175 if (i_size > 0 && page->index < end_index)
176 goto out;
177 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
178 if (i_size >= end)
179 goto out;
180 i_size_write(inode, end);
181 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
182out:
183 spin_unlock(&inode->i_lock);
184}
185
186/* A writeback failed: mark the page as bad, and invalidate the page cache */
187static void nfs_set_pageerror(struct page *page)
188{
189 SetPageError(page);
190 nfs_zap_mapping(page->mapping->host, page->mapping);
191}
192
193/* We can set the PG_uptodate flag if we see that a write request
194 * covers the full page.
195 */
196static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
197{
198 if (PageUptodate(page))
199 return;
200 if (base != 0)
201 return;
202 if (count != nfs_page_length(page))
203 return;
204 SetPageUptodate(page);
205}
206
207static int wb_priority(struct writeback_control *wbc)
208{
209 if (wbc->for_reclaim)
210 return FLUSH_HIGHPRI | FLUSH_STABLE;
211 if (wbc->for_kupdate || wbc->for_background)
212 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
213 return FLUSH_COND_STABLE;
214}
215
216/*
217 * NFS congestion control
218 */
219
220int nfs_congestion_kb;
221
222#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
223#define NFS_CONGESTION_OFF_THRESH \
224 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
225
226static int nfs_set_page_writeback(struct page *page)
227{
228 int ret = test_set_page_writeback(page);
229
230 if (!ret) {
231 struct inode *inode = page->mapping->host;
232 struct nfs_server *nfss = NFS_SERVER(inode);
233
234 if (atomic_long_inc_return(&nfss->writeback) >
235 NFS_CONGESTION_ON_THRESH) {
236 set_bdi_congested(&nfss->backing_dev_info,
237 BLK_RW_ASYNC);
238 }
239 }
240 return ret;
241}
242
243static void nfs_end_page_writeback(struct page *page)
244{
245 struct inode *inode = page->mapping->host;
246 struct nfs_server *nfss = NFS_SERVER(inode);
247
248 end_page_writeback(page);
249 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
250 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
251}
252
253static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
254{
255 struct inode *inode = page->mapping->host;
256 struct nfs_page *req;
257 int ret;
258
259 spin_lock(&inode->i_lock);
260 for (;;) {
261 req = nfs_page_find_request_locked(page);
262 if (req == NULL)
263 break;
264 if (nfs_lock_request(req))
265 break;
266 /* Note: If we hold the page lock, as is the case in nfs_writepage,
267 * then the call to nfs_lock_request() will always
268 * succeed provided that someone hasn't already marked the
269 * request as dirty (in which case we don't care).
270 */
271 spin_unlock(&inode->i_lock);
272 if (!nonblock)
273 ret = nfs_wait_on_request(req);
274 else
275 ret = -EAGAIN;
276 nfs_release_request(req);
277 if (ret != 0)
278 return ERR_PTR(ret);
279 spin_lock(&inode->i_lock);
280 }
281 spin_unlock(&inode->i_lock);
282 return req;
283}
284
285/*
286 * Find an associated nfs write request, and prepare to flush it out
287 * May return an error if the user signalled nfs_wait_on_request().
288 */
289static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
290 struct page *page, bool nonblock)
291{
292 struct nfs_page *req;
293 int ret = 0;
294
295 req = nfs_find_and_lock_request(page, nonblock);
296 if (!req)
297 goto out;
298 ret = PTR_ERR(req);
299 if (IS_ERR(req))
300 goto out;
301
302 ret = nfs_set_page_writeback(page);
303 BUG_ON(ret != 0);
304 BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
305
306 if (!nfs_pageio_add_request(pgio, req)) {
307 nfs_redirty_request(req);
308 ret = pgio->pg_error;
309 }
310out:
311 return ret;
312}
313
314static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
315{
316 struct inode *inode = page->mapping->host;
317 int ret;
318
319 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
320 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
321
322 nfs_pageio_cond_complete(pgio, page->index);
323 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
324 if (ret == -EAGAIN) {
325 redirty_page_for_writepage(wbc, page);
326 ret = 0;
327 }
328 return ret;
329}
330
331/*
332 * Write an mmapped page to the server.
333 */
334static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
335{
336 struct nfs_pageio_descriptor pgio;
337 int err;
338
339 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
340 &nfs_async_write_completion_ops);
341 err = nfs_do_writepage(page, wbc, &pgio);
342 nfs_pageio_complete(&pgio);
343 if (err < 0)
344 return err;
345 if (pgio.pg_error < 0)
346 return pgio.pg_error;
347 return 0;
348}
349
350int nfs_writepage(struct page *page, struct writeback_control *wbc)
351{
352 int ret;
353
354 ret = nfs_writepage_locked(page, wbc);
355 unlock_page(page);
356 return ret;
357}
358
359static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
360{
361 int ret;
362
363 ret = nfs_do_writepage(page, wbc, data);
364 unlock_page(page);
365 return ret;
366}
367
368int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
369{
370 struct inode *inode = mapping->host;
371 unsigned long *bitlock = &NFS_I(inode)->flags;
372 struct nfs_pageio_descriptor pgio;
373 int err;
374
375 /* Stop dirtying of new pages while we sync */
376 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
377 nfs_wait_bit_killable, TASK_KILLABLE);
378 if (err)
379 goto out_err;
380
381 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
382
383 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc),
384 &nfs_async_write_completion_ops);
385 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
386 nfs_pageio_complete(&pgio);
387
388 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
389 smp_mb__after_clear_bit();
390 wake_up_bit(bitlock, NFS_INO_FLUSHING);
391
392 if (err < 0)
393 goto out_err;
394 err = pgio.pg_error;
395 if (err < 0)
396 goto out_err;
397 return 0;
398out_err:
399 return err;
400}
401
402/*
403 * Insert a write request into an inode
404 */
405static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
406{
407 struct nfs_inode *nfsi = NFS_I(inode);
408
409 /* Lock the request! */
410 nfs_lock_request(req);
411
412 spin_lock(&inode->i_lock);
413 if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
414 inode->i_version++;
415 set_bit(PG_MAPPED, &req->wb_flags);
416 SetPagePrivate(req->wb_page);
417 set_page_private(req->wb_page, (unsigned long)req);
418 nfsi->npages++;
419 kref_get(&req->wb_kref);
420 spin_unlock(&inode->i_lock);
421}
422
423/*
424 * Remove a write request from an inode
425 */
426static void nfs_inode_remove_request(struct nfs_page *req)
427{
428 struct inode *inode = req->wb_context->dentry->d_inode;
429 struct nfs_inode *nfsi = NFS_I(inode);
430
431 BUG_ON (!NFS_WBACK_BUSY(req));
432
433 spin_lock(&inode->i_lock);
434 set_page_private(req->wb_page, 0);
435 ClearPagePrivate(req->wb_page);
436 clear_bit(PG_MAPPED, &req->wb_flags);
437 nfsi->npages--;
438 spin_unlock(&inode->i_lock);
439 nfs_release_request(req);
440}
441
442static void
443nfs_mark_request_dirty(struct nfs_page *req)
444{
445 __set_page_dirty_nobuffers(req->wb_page);
446}
447
448#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
449/**
450 * nfs_request_add_commit_list - add request to a commit list
451 * @req: pointer to a struct nfs_page
452 * @dst: commit list head
453 * @cinfo: holds list lock and accounting info
454 *
455 * This sets the PG_CLEAN bit, updates the cinfo count of
456 * number of outstanding requests requiring a commit as well as
457 * the MM page stats.
458 *
459 * The caller must _not_ hold the cinfo->lock, but must be
460 * holding the nfs_page lock.
461 */
462void
463nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
464 struct nfs_commit_info *cinfo)
465{
466 set_bit(PG_CLEAN, &(req)->wb_flags);
467 spin_lock(cinfo->lock);
468 nfs_list_add_request(req, dst);
469 cinfo->mds->ncommit++;
470 spin_unlock(cinfo->lock);
471 if (!cinfo->dreq) {
472 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
473 inc_bdi_stat(req->wb_page->mapping->backing_dev_info,
474 BDI_RECLAIMABLE);
475 __mark_inode_dirty(req->wb_context->dentry->d_inode,
476 I_DIRTY_DATASYNC);
477 }
478}
479EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
480
481/**
482 * nfs_request_remove_commit_list - Remove request from a commit list
483 * @req: pointer to a nfs_page
484 * @cinfo: holds list lock and accounting info
485 *
486 * This clears the PG_CLEAN bit, and updates the cinfo's count of
487 * number of outstanding requests requiring a commit
488 * It does not update the MM page stats.
489 *
490 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
491 */
492void
493nfs_request_remove_commit_list(struct nfs_page *req,
494 struct nfs_commit_info *cinfo)
495{
496 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
497 return;
498 nfs_list_remove_request(req);
499 cinfo->mds->ncommit--;
500}
501EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
502
503static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
504 struct inode *inode)
505{
506 cinfo->lock = &inode->i_lock;
507 cinfo->mds = &NFS_I(inode)->commit_info;
508 cinfo->ds = pnfs_get_ds_info(inode);
509 cinfo->dreq = NULL;
510 cinfo->completion_ops = &nfs_commit_completion_ops;
511}
512
513void nfs_init_cinfo(struct nfs_commit_info *cinfo,
514 struct inode *inode,
515 struct nfs_direct_req *dreq)
516{
517 if (dreq)
518 nfs_init_cinfo_from_dreq(cinfo, dreq);
519 else
520 nfs_init_cinfo_from_inode(cinfo, inode);
521}
522EXPORT_SYMBOL_GPL(nfs_init_cinfo);
523
524/*
525 * Add a request to the inode's commit list.
526 */
527void
528nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
529 struct nfs_commit_info *cinfo)
530{
531 if (pnfs_mark_request_commit(req, lseg, cinfo))
532 return;
533 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
534}
535
536static void
537nfs_clear_page_commit(struct page *page)
538{
539 dec_zone_page_state(page, NR_UNSTABLE_NFS);
540 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
541}
542
543static void
544nfs_clear_request_commit(struct nfs_page *req)
545{
546 if (test_bit(PG_CLEAN, &req->wb_flags)) {
547 struct inode *inode = req->wb_context->dentry->d_inode;
548 struct nfs_commit_info cinfo;
549
550 nfs_init_cinfo_from_inode(&cinfo, inode);
551 if (!pnfs_clear_request_commit(req, &cinfo)) {
552 spin_lock(cinfo.lock);
553 nfs_request_remove_commit_list(req, &cinfo);
554 spin_unlock(cinfo.lock);
555 }
556 nfs_clear_page_commit(req->wb_page);
557 }
558}
559
560static inline
561int nfs_write_need_commit(struct nfs_write_data *data)
562{
563 if (data->verf.committed == NFS_DATA_SYNC)
564 return data->header->lseg == NULL;
565 return data->verf.committed != NFS_FILE_SYNC;
566}
567
568#else
569static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
570 struct inode *inode)
571{
572}
573
574void nfs_init_cinfo(struct nfs_commit_info *cinfo,
575 struct inode *inode,
576 struct nfs_direct_req *dreq)
577{
578}
579
580void
581nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
582 struct nfs_commit_info *cinfo)
583{
584}
585
586static void
587nfs_clear_request_commit(struct nfs_page *req)
588{
589}
590
591static inline
592int nfs_write_need_commit(struct nfs_write_data *data)
593{
594 return 0;
595}
596
597#endif
598
599static void nfs_write_completion(struct nfs_pgio_header *hdr)
600{
601 struct nfs_commit_info cinfo;
602 unsigned long bytes = 0;
603
604 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
605 goto out;
606 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
607 while (!list_empty(&hdr->pages)) {
608 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
609
610 bytes += req->wb_bytes;
611 nfs_list_remove_request(req);
612 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
613 (hdr->good_bytes < bytes)) {
614 nfs_set_pageerror(req->wb_page);
615 nfs_context_set_write_error(req->wb_context, hdr->error);
616 goto remove_req;
617 }
618 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
619 nfs_mark_request_dirty(req);
620 goto next;
621 }
622 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
623 memcpy(&req->wb_verf, hdr->verf, sizeof(req->wb_verf));
624 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
625 goto next;
626 }
627remove_req:
628 nfs_inode_remove_request(req);
629next:
630 nfs_unlock_request(req);
631 nfs_end_page_writeback(req->wb_page);
632 nfs_release_request(req);
633 }
634out:
635 hdr->release(hdr);
636}
637
638#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
639static unsigned long
640nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
641{
642 return cinfo->mds->ncommit;
643}
644
645/* cinfo->lock held by caller */
646int
647nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
648 struct nfs_commit_info *cinfo, int max)
649{
650 struct nfs_page *req, *tmp;
651 int ret = 0;
652
653 list_for_each_entry_safe(req, tmp, src, wb_list) {
654 if (!nfs_lock_request(req))
655 continue;
656 kref_get(&req->wb_kref);
657 if (cond_resched_lock(cinfo->lock))
658 list_safe_reset_next(req, tmp, wb_list);
659 nfs_request_remove_commit_list(req, cinfo);
660 nfs_list_add_request(req, dst);
661 ret++;
662 if ((ret == max) && !cinfo->dreq)
663 break;
664 }
665 return ret;
666}
667
668/*
669 * nfs_scan_commit - Scan an inode for commit requests
670 * @inode: NFS inode to scan
671 * @dst: mds destination list
672 * @cinfo: mds and ds lists of reqs ready to commit
673 *
674 * Moves requests from the inode's 'commit' request list.
675 * The requests are *not* checked to ensure that they form a contiguous set.
676 */
677int
678nfs_scan_commit(struct inode *inode, struct list_head *dst,
679 struct nfs_commit_info *cinfo)
680{
681 int ret = 0;
682
683 spin_lock(cinfo->lock);
684 if (cinfo->mds->ncommit > 0) {
685 const int max = INT_MAX;
686
687 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
688 cinfo, max);
689 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
690 }
691 spin_unlock(cinfo->lock);
692 return ret;
693}
694
695#else
696static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
697{
698 return 0;
699}
700
701int nfs_scan_commit(struct inode *inode, struct list_head *dst,
702 struct nfs_commit_info *cinfo)
703{
704 return 0;
705}
706#endif
707
708/*
709 * Search for an existing write request, and attempt to update
710 * it to reflect a new dirty region on a given page.
711 *
712 * If the attempt fails, then the existing request is flushed out
713 * to disk.
714 */
715static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
716 struct page *page,
717 unsigned int offset,
718 unsigned int bytes)
719{
720 struct nfs_page *req;
721 unsigned int rqend;
722 unsigned int end;
723 int error;
724
725 if (!PagePrivate(page))
726 return NULL;
727
728 end = offset + bytes;
729 spin_lock(&inode->i_lock);
730
731 for (;;) {
732 req = nfs_page_find_request_locked(page);
733 if (req == NULL)
734 goto out_unlock;
735
736 rqend = req->wb_offset + req->wb_bytes;
737 /*
738 * Tell the caller to flush out the request if
739 * the offsets are non-contiguous.
740 * Note: nfs_flush_incompatible() will already
741 * have flushed out requests having wrong owners.
742 */
743 if (offset > rqend
744 || end < req->wb_offset)
745 goto out_flushme;
746
747 if (nfs_lock_request(req))
748 break;
749
750 /* The request is locked, so wait and then retry */
751 spin_unlock(&inode->i_lock);
752 error = nfs_wait_on_request(req);
753 nfs_release_request(req);
754 if (error != 0)
755 goto out_err;
756 spin_lock(&inode->i_lock);
757 }
758
759 /* Okay, the request matches. Update the region */
760 if (offset < req->wb_offset) {
761 req->wb_offset = offset;
762 req->wb_pgbase = offset;
763 }
764 if (end > rqend)
765 req->wb_bytes = end - req->wb_offset;
766 else
767 req->wb_bytes = rqend - req->wb_offset;
768out_unlock:
769 spin_unlock(&inode->i_lock);
770 if (req)
771 nfs_clear_request_commit(req);
772 return req;
773out_flushme:
774 spin_unlock(&inode->i_lock);
775 nfs_release_request(req);
776 error = nfs_wb_page(inode, page);
777out_err:
778 return ERR_PTR(error);
779}
780
781/*
782 * Try to update an existing write request, or create one if there is none.
783 *
784 * Note: Should always be called with the Page Lock held to prevent races
785 * if we have to add a new request. Also assumes that the caller has
786 * already called nfs_flush_incompatible() if necessary.
787 */
788static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
789 struct page *page, unsigned int offset, unsigned int bytes)
790{
791 struct inode *inode = page->mapping->host;
792 struct nfs_page *req;
793
794 req = nfs_try_to_update_request(inode, page, offset, bytes);
795 if (req != NULL)
796 goto out;
797 req = nfs_create_request(ctx, inode, page, offset, bytes);
798 if (IS_ERR(req))
799 goto out;
800 nfs_inode_add_request(inode, req);
801out:
802 return req;
803}
804
805static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
806 unsigned int offset, unsigned int count)
807{
808 struct nfs_page *req;
809
810 req = nfs_setup_write_request(ctx, page, offset, count);
811 if (IS_ERR(req))
812 return PTR_ERR(req);
813 /* Update file length */
814 nfs_grow_file(page, offset, count);
815 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
816 nfs_mark_request_dirty(req);
817 nfs_unlock_and_release_request(req);
818 return 0;
819}
820
821int nfs_flush_incompatible(struct file *file, struct page *page)
822{
823 struct nfs_open_context *ctx = nfs_file_open_context(file);
824 struct nfs_page *req;
825 int do_flush, status;
826 /*
827 * Look for a request corresponding to this page. If there
828 * is one, and it belongs to another file, we flush it out
829 * before we try to copy anything into the page. Do this
830 * due to the lack of an ACCESS-type call in NFSv2.
831 * Also do the same if we find a request from an existing
832 * dropped page.
833 */
834 do {
835 req = nfs_page_find_request(page);
836 if (req == NULL)
837 return 0;
838 do_flush = req->wb_page != page || req->wb_context != ctx ||
839 req->wb_lock_context->lockowner != current->files ||
840 req->wb_lock_context->pid != current->tgid;
841 nfs_release_request(req);
842 if (!do_flush)
843 return 0;
844 status = nfs_wb_page(page->mapping->host, page);
845 } while (status == 0);
846 return status;
847}
848
849/*
850 * If the page cache is marked as unsafe or invalid, then we can't rely on
851 * the PageUptodate() flag. In this case, we will need to turn off
852 * write optimisations that depend on the page contents being correct.
853 */
854static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
855{
856 if (nfs_have_delegated_attributes(inode))
857 goto out;
858 if (NFS_I(inode)->cache_validity & NFS_INO_REVAL_PAGECACHE)
859 return false;
860out:
861 return PageUptodate(page) != 0;
862}
863
864/*
865 * Update and possibly write a cached page of an NFS file.
866 *
867 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
868 * things with a page scheduled for an RPC call (e.g. invalidate it).
869 */
870int nfs_updatepage(struct file *file, struct page *page,
871 unsigned int offset, unsigned int count)
872{
873 struct nfs_open_context *ctx = nfs_file_open_context(file);
874 struct inode *inode = page->mapping->host;
875 int status = 0;
876
877 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
878
879 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
880 file->f_path.dentry->d_parent->d_name.name,
881 file->f_path.dentry->d_name.name, count,
882 (long long)(page_offset(page) + offset));
883
884 /* If we're not using byte range locks, and we know the page
885 * is up to date, it may be more efficient to extend the write
886 * to cover the entire page in order to avoid fragmentation
887 * inefficiencies.
888 */
889 if (nfs_write_pageuptodate(page, inode) &&
890 inode->i_flock == NULL &&
891 !(file->f_flags & O_DSYNC)) {
892 count = max(count + offset, nfs_page_length(page));
893 offset = 0;
894 }
895
896 status = nfs_writepage_setup(ctx, page, offset, count);
897 if (status < 0)
898 nfs_set_pageerror(page);
899 else
900 __set_page_dirty_nobuffers(page);
901
902 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
903 status, (long long)i_size_read(inode));
904 return status;
905}
906
907static int flush_task_priority(int how)
908{
909 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
910 case FLUSH_HIGHPRI:
911 return RPC_PRIORITY_HIGH;
912 case FLUSH_LOWPRI:
913 return RPC_PRIORITY_LOW;
914 }
915 return RPC_PRIORITY_NORMAL;
916}
917
918int nfs_initiate_write(struct rpc_clnt *clnt,
919 struct nfs_write_data *data,
920 const struct rpc_call_ops *call_ops,
921 int how, int flags)
922{
923 struct inode *inode = data->header->inode;
924 int priority = flush_task_priority(how);
925 struct rpc_task *task;
926 struct rpc_message msg = {
927 .rpc_argp = &data->args,
928 .rpc_resp = &data->res,
929 .rpc_cred = data->header->cred,
930 };
931 struct rpc_task_setup task_setup_data = {
932 .rpc_client = clnt,
933 .task = &data->task,
934 .rpc_message = &msg,
935 .callback_ops = call_ops,
936 .callback_data = data,
937 .workqueue = nfsiod_workqueue,
938 .flags = RPC_TASK_ASYNC | flags,
939 .priority = priority,
940 };
941 int ret = 0;
942
943 /* Set up the initial task struct. */
944 NFS_PROTO(inode)->write_setup(data, &msg);
945
946 dprintk("NFS: %5u initiated write call "
947 "(req %s/%lld, %u bytes @ offset %llu)\n",
948 data->task.tk_pid,
949 inode->i_sb->s_id,
950 (long long)NFS_FILEID(inode),
951 data->args.count,
952 (unsigned long long)data->args.offset);
953
954 task = rpc_run_task(&task_setup_data);
955 if (IS_ERR(task)) {
956 ret = PTR_ERR(task);
957 goto out;
958 }
959 if (how & FLUSH_SYNC) {
960 ret = rpc_wait_for_completion_task(task);
961 if (ret == 0)
962 ret = task->tk_status;
963 }
964 rpc_put_task(task);
965out:
966 return ret;
967}
968EXPORT_SYMBOL_GPL(nfs_initiate_write);
969
970/*
971 * Set up the argument/result storage required for the RPC call.
972 */
973static void nfs_write_rpcsetup(struct nfs_write_data *data,
974 unsigned int count, unsigned int offset,
975 int how, struct nfs_commit_info *cinfo)
976{
977 struct nfs_page *req = data->header->req;
978
979 /* Set up the RPC argument and reply structs
980 * NB: take care not to mess about with data->commit et al. */
981
982 data->args.fh = NFS_FH(data->header->inode);
983 data->args.offset = req_offset(req) + offset;
984 /* pnfs_set_layoutcommit needs this */
985 data->mds_offset = data->args.offset;
986 data->args.pgbase = req->wb_pgbase + offset;
987 data->args.pages = data->pages.pagevec;
988 data->args.count = count;
989 data->args.context = get_nfs_open_context(req->wb_context);
990 data->args.lock_context = req->wb_lock_context;
991 data->args.stable = NFS_UNSTABLE;
992 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
993 case 0:
994 break;
995 case FLUSH_COND_STABLE:
996 if (nfs_reqs_to_commit(cinfo))
997 break;
998 default:
999 data->args.stable = NFS_FILE_SYNC;
1000 }
1001
1002 data->res.fattr = &data->fattr;
1003 data->res.count = count;
1004 data->res.verf = &data->verf;
1005 nfs_fattr_init(&data->fattr);
1006}
1007
1008static int nfs_do_write(struct nfs_write_data *data,
1009 const struct rpc_call_ops *call_ops,
1010 int how)
1011{
1012 struct inode *inode = data->header->inode;
1013
1014 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1015}
1016
1017static int nfs_do_multiple_writes(struct list_head *head,
1018 const struct rpc_call_ops *call_ops,
1019 int how)
1020{
1021 struct nfs_write_data *data;
1022 int ret = 0;
1023
1024 while (!list_empty(head)) {
1025 int ret2;
1026
1027 data = list_first_entry(head, struct nfs_write_data, list);
1028 list_del_init(&data->list);
1029
1030 ret2 = nfs_do_write(data, call_ops, how);
1031 if (ret == 0)
1032 ret = ret2;
1033 }
1034 return ret;
1035}
1036
1037/* If a nfs_flush_* function fails, it should remove reqs from @head and
1038 * call this on each, which will prepare them to be retried on next
1039 * writeback using standard nfs.
1040 */
1041static void nfs_redirty_request(struct nfs_page *req)
1042{
1043 nfs_mark_request_dirty(req);
1044 nfs_unlock_request(req);
1045 nfs_end_page_writeback(req->wb_page);
1046 nfs_release_request(req);
1047}
1048
1049static void nfs_async_write_error(struct list_head *head)
1050{
1051 struct nfs_page *req;
1052
1053 while (!list_empty(head)) {
1054 req = nfs_list_entry(head->next);
1055 nfs_list_remove_request(req);
1056 nfs_redirty_request(req);
1057 }
1058}
1059
1060static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1061 .error_cleanup = nfs_async_write_error,
1062 .completion = nfs_write_completion,
1063};
1064
1065static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1066 struct nfs_pgio_header *hdr)
1067{
1068 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1069 while (!list_empty(&hdr->rpc_list)) {
1070 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1071 struct nfs_write_data, list);
1072 list_del(&data->list);
1073 nfs_writedata_release(data);
1074 }
1075 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1076}
1077
1078/*
1079 * Generate multiple small requests to write out a single
1080 * contiguous dirty area on one page.
1081 */
1082static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1083 struct nfs_pgio_header *hdr)
1084{
1085 struct nfs_page *req = hdr->req;
1086 struct page *page = req->wb_page;
1087 struct nfs_write_data *data;
1088 size_t wsize = desc->pg_bsize, nbytes;
1089 unsigned int offset;
1090 int requests = 0;
1091 struct nfs_commit_info cinfo;
1092
1093 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1094
1095 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1096 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1097 desc->pg_count > wsize))
1098 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1099
1100
1101 offset = 0;
1102 nbytes = desc->pg_count;
1103 do {
1104 size_t len = min(nbytes, wsize);
1105
1106 data = nfs_writedata_alloc(hdr, 1);
1107 if (!data) {
1108 nfs_flush_error(desc, hdr);
1109 return -ENOMEM;
1110 }
1111 data->pages.pagevec[0] = page;
1112 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1113 list_add(&data->list, &hdr->rpc_list);
1114 requests++;
1115 nbytes -= len;
1116 offset += len;
1117 } while (nbytes != 0);
1118 nfs_list_remove_request(req);
1119 nfs_list_add_request(req, &hdr->pages);
1120 desc->pg_rpc_callops = &nfs_write_common_ops;
1121 return 0;
1122}
1123
1124/*
1125 * Create an RPC task for the given write request and kick it.
1126 * The page must have been locked by the caller.
1127 *
1128 * It may happen that the page we're passed is not marked dirty.
1129 * This is the case if nfs_updatepage detects a conflicting request
1130 * that has been written but not committed.
1131 */
1132static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1133 struct nfs_pgio_header *hdr)
1134{
1135 struct nfs_page *req;
1136 struct page **pages;
1137 struct nfs_write_data *data;
1138 struct list_head *head = &desc->pg_list;
1139 struct nfs_commit_info cinfo;
1140
1141 data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1142 desc->pg_count));
1143 if (!data) {
1144 nfs_flush_error(desc, hdr);
1145 return -ENOMEM;
1146 }
1147
1148 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1149 pages = data->pages.pagevec;
1150 while (!list_empty(head)) {
1151 req = nfs_list_entry(head->next);
1152 nfs_list_remove_request(req);
1153 nfs_list_add_request(req, &hdr->pages);
1154 *pages++ = req->wb_page;
1155 }
1156
1157 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1158 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1159 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1160
1161 /* Set up the argument struct */
1162 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1163 list_add(&data->list, &hdr->rpc_list);
1164 desc->pg_rpc_callops = &nfs_write_common_ops;
1165 return 0;
1166}
1167
1168int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1169 struct nfs_pgio_header *hdr)
1170{
1171 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1172 return nfs_flush_multi(desc, hdr);
1173 return nfs_flush_one(desc, hdr);
1174}
1175
1176static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1177{
1178 struct nfs_write_header *whdr;
1179 struct nfs_pgio_header *hdr;
1180 int ret;
1181
1182 whdr = nfs_writehdr_alloc();
1183 if (!whdr) {
1184 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1185 return -ENOMEM;
1186 }
1187 hdr = &whdr->header;
1188 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1189 atomic_inc(&hdr->refcnt);
1190 ret = nfs_generic_flush(desc, hdr);
1191 if (ret == 0)
1192 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1193 desc->pg_rpc_callops,
1194 desc->pg_ioflags);
1195 if (atomic_dec_and_test(&hdr->refcnt))
1196 hdr->completion_ops->completion(hdr);
1197 return ret;
1198}
1199
1200static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1201 .pg_test = nfs_generic_pg_test,
1202 .pg_doio = nfs_generic_pg_writepages,
1203};
1204
1205void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1206 struct inode *inode, int ioflags,
1207 const struct nfs_pgio_completion_ops *compl_ops)
1208{
1209 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1210 NFS_SERVER(inode)->wsize, ioflags);
1211}
1212
1213void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1214{
1215 pgio->pg_ops = &nfs_pageio_write_ops;
1216 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1217}
1218EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1219
1220void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1221 struct inode *inode, int ioflags,
1222 const struct nfs_pgio_completion_ops *compl_ops)
1223{
1224 if (!pnfs_pageio_init_write(pgio, inode, ioflags, compl_ops))
1225 nfs_pageio_init_write_mds(pgio, inode, ioflags, compl_ops);
1226}
1227
1228void nfs_write_prepare(struct rpc_task *task, void *calldata)
1229{
1230 struct nfs_write_data *data = calldata;
1231 NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1232}
1233
1234void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1235{
1236 struct nfs_commit_data *data = calldata;
1237
1238 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1239}
1240
1241/*
1242 * Handle a write reply that flushes a whole page.
1243 *
1244 * FIXME: There is an inherent race with invalidate_inode_pages and
1245 * writebacks since the page->count is kept > 1 for as long
1246 * as the page has a write request pending.
1247 */
1248static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1249{
1250 struct nfs_write_data *data = calldata;
1251
1252 nfs_writeback_done(task, data);
1253}
1254
1255static void nfs_writeback_release_common(void *calldata)
1256{
1257 struct nfs_write_data *data = calldata;
1258 struct nfs_pgio_header *hdr = data->header;
1259 int status = data->task.tk_status;
1260
1261 if ((status >= 0) && nfs_write_need_commit(data)) {
1262 spin_lock(&hdr->lock);
1263 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1264 ; /* Do nothing */
1265 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1266 memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1267 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1268 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1269 spin_unlock(&hdr->lock);
1270 }
1271 nfs_writedata_release(data);
1272}
1273
1274static const struct rpc_call_ops nfs_write_common_ops = {
1275 .rpc_call_prepare = nfs_write_prepare,
1276 .rpc_call_done = nfs_writeback_done_common,
1277 .rpc_release = nfs_writeback_release_common,
1278};
1279
1280
1281/*
1282 * This function is called when the WRITE call is complete.
1283 */
1284void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1285{
1286 struct nfs_writeargs *argp = &data->args;
1287 struct nfs_writeres *resp = &data->res;
1288 struct inode *inode = data->header->inode;
1289 int status;
1290
1291 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1292 task->tk_pid, task->tk_status);
1293
1294 /*
1295 * ->write_done will attempt to use post-op attributes to detect
1296 * conflicting writes by other clients. A strict interpretation
1297 * of close-to-open would allow us to continue caching even if
1298 * another writer had changed the file, but some applications
1299 * depend on tighter cache coherency when writing.
1300 */
1301 status = NFS_PROTO(inode)->write_done(task, data);
1302 if (status != 0)
1303 return;
1304 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1305
1306#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1307 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1308 /* We tried a write call, but the server did not
1309 * commit data to stable storage even though we
1310 * requested it.
1311 * Note: There is a known bug in Tru64 < 5.0 in which
1312 * the server reports NFS_DATA_SYNC, but performs
1313 * NFS_FILE_SYNC. We therefore implement this checking
1314 * as a dprintk() in order to avoid filling syslog.
1315 */
1316 static unsigned long complain;
1317
1318 /* Note this will print the MDS for a DS write */
1319 if (time_before(complain, jiffies)) {
1320 dprintk("NFS: faulty NFS server %s:"
1321 " (committed = %d) != (stable = %d)\n",
1322 NFS_SERVER(inode)->nfs_client->cl_hostname,
1323 resp->verf->committed, argp->stable);
1324 complain = jiffies + 300 * HZ;
1325 }
1326 }
1327#endif
1328 if (task->tk_status < 0)
1329 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1330 else if (resp->count < argp->count) {
1331 static unsigned long complain;
1332
1333 /* This a short write! */
1334 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1335
1336 /* Has the server at least made some progress? */
1337 if (resp->count == 0) {
1338 if (time_before(complain, jiffies)) {
1339 printk(KERN_WARNING
1340 "NFS: Server wrote zero bytes, expected %u.\n",
1341 argp->count);
1342 complain = jiffies + 300 * HZ;
1343 }
1344 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1345 task->tk_status = -EIO;
1346 return;
1347 }
1348 /* Was this an NFSv2 write or an NFSv3 stable write? */
1349 if (resp->verf->committed != NFS_UNSTABLE) {
1350 /* Resend from where the server left off */
1351 data->mds_offset += resp->count;
1352 argp->offset += resp->count;
1353 argp->pgbase += resp->count;
1354 argp->count -= resp->count;
1355 } else {
1356 /* Resend as a stable write in order to avoid
1357 * headaches in the case of a server crash.
1358 */
1359 argp->stable = NFS_FILE_SYNC;
1360 }
1361 rpc_restart_call_prepare(task);
1362 }
1363}
1364
1365
1366#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1367static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1368{
1369 int ret;
1370
1371 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1372 return 1;
1373 if (!may_wait)
1374 return 0;
1375 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1376 NFS_INO_COMMIT,
1377 nfs_wait_bit_killable,
1378 TASK_KILLABLE);
1379 return (ret < 0) ? ret : 1;
1380}
1381
1382static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1383{
1384 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1385 smp_mb__after_clear_bit();
1386 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1387}
1388
1389void nfs_commitdata_release(struct nfs_commit_data *data)
1390{
1391 put_nfs_open_context(data->context);
1392 nfs_commit_free(data);
1393}
1394EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1395
1396int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1397 const struct rpc_call_ops *call_ops,
1398 int how, int flags)
1399{
1400 struct rpc_task *task;
1401 int priority = flush_task_priority(how);
1402 struct rpc_message msg = {
1403 .rpc_argp = &data->args,
1404 .rpc_resp = &data->res,
1405 .rpc_cred = data->cred,
1406 };
1407 struct rpc_task_setup task_setup_data = {
1408 .task = &data->task,
1409 .rpc_client = clnt,
1410 .rpc_message = &msg,
1411 .callback_ops = call_ops,
1412 .callback_data = data,
1413 .workqueue = nfsiod_workqueue,
1414 .flags = RPC_TASK_ASYNC | flags,
1415 .priority = priority,
1416 };
1417 /* Set up the initial task struct. */
1418 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1419
1420 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1421
1422 task = rpc_run_task(&task_setup_data);
1423 if (IS_ERR(task))
1424 return PTR_ERR(task);
1425 if (how & FLUSH_SYNC)
1426 rpc_wait_for_completion_task(task);
1427 rpc_put_task(task);
1428 return 0;
1429}
1430EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1431
1432/*
1433 * Set up the argument/result storage required for the RPC call.
1434 */
1435void nfs_init_commit(struct nfs_commit_data *data,
1436 struct list_head *head,
1437 struct pnfs_layout_segment *lseg,
1438 struct nfs_commit_info *cinfo)
1439{
1440 struct nfs_page *first = nfs_list_entry(head->next);
1441 struct inode *inode = first->wb_context->dentry->d_inode;
1442
1443 /* Set up the RPC argument and reply structs
1444 * NB: take care not to mess about with data->commit et al. */
1445
1446 list_splice_init(head, &data->pages);
1447
1448 data->inode = inode;
1449 data->cred = first->wb_context->cred;
1450 data->lseg = lseg; /* reference transferred */
1451 data->mds_ops = &nfs_commit_ops;
1452 data->completion_ops = cinfo->completion_ops;
1453 data->dreq = cinfo->dreq;
1454
1455 data->args.fh = NFS_FH(data->inode);
1456 /* Note: we always request a commit of the entire inode */
1457 data->args.offset = 0;
1458 data->args.count = 0;
1459 data->context = get_nfs_open_context(first->wb_context);
1460 data->res.fattr = &data->fattr;
1461 data->res.verf = &data->verf;
1462 nfs_fattr_init(&data->fattr);
1463}
1464EXPORT_SYMBOL_GPL(nfs_init_commit);
1465
1466void nfs_retry_commit(struct list_head *page_list,
1467 struct pnfs_layout_segment *lseg,
1468 struct nfs_commit_info *cinfo)
1469{
1470 struct nfs_page *req;
1471
1472 while (!list_empty(page_list)) {
1473 req = nfs_list_entry(page_list->next);
1474 nfs_list_remove_request(req);
1475 nfs_mark_request_commit(req, lseg, cinfo);
1476 if (!cinfo->dreq) {
1477 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1478 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1479 BDI_RECLAIMABLE);
1480 }
1481 nfs_unlock_and_release_request(req);
1482 }
1483}
1484EXPORT_SYMBOL_GPL(nfs_retry_commit);
1485
1486/*
1487 * Commit dirty pages
1488 */
1489static int
1490nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1491 struct nfs_commit_info *cinfo)
1492{
1493 struct nfs_commit_data *data;
1494
1495 data = nfs_commitdata_alloc();
1496
1497 if (!data)
1498 goto out_bad;
1499
1500 /* Set up the argument struct */
1501 nfs_init_commit(data, head, NULL, cinfo);
1502 atomic_inc(&cinfo->mds->rpcs_out);
1503 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1504 how, 0);
1505 out_bad:
1506 nfs_retry_commit(head, NULL, cinfo);
1507 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1508 return -ENOMEM;
1509}
1510
1511/*
1512 * COMMIT call returned
1513 */
1514static void nfs_commit_done(struct rpc_task *task, void *calldata)
1515{
1516 struct nfs_commit_data *data = calldata;
1517
1518 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1519 task->tk_pid, task->tk_status);
1520
1521 /* Call the NFS version-specific code */
1522 NFS_PROTO(data->inode)->commit_done(task, data);
1523}
1524
1525static void nfs_commit_release_pages(struct nfs_commit_data *data)
1526{
1527 struct nfs_page *req;
1528 int status = data->task.tk_status;
1529 struct nfs_commit_info cinfo;
1530
1531 while (!list_empty(&data->pages)) {
1532 req = nfs_list_entry(data->pages.next);
1533 nfs_list_remove_request(req);
1534 nfs_clear_page_commit(req->wb_page);
1535
1536 dprintk("NFS: commit (%s/%lld %d@%lld)",
1537 req->wb_context->dentry->d_sb->s_id,
1538 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1539 req->wb_bytes,
1540 (long long)req_offset(req));
1541 if (status < 0) {
1542 nfs_context_set_write_error(req->wb_context, status);
1543 nfs_inode_remove_request(req);
1544 dprintk(", error = %d\n", status);
1545 goto next;
1546 }
1547
1548 /* Okay, COMMIT succeeded, apparently. Check the verifier
1549 * returned by the server against all stored verfs. */
1550 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1551 /* We have a match */
1552 nfs_inode_remove_request(req);
1553 dprintk(" OK\n");
1554 goto next;
1555 }
1556 /* We have a mismatch. Write the page again */
1557 dprintk(" mismatch\n");
1558 nfs_mark_request_dirty(req);
1559 next:
1560 nfs_unlock_and_release_request(req);
1561 }
1562 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1563 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1564 nfs_commit_clear_lock(NFS_I(data->inode));
1565}
1566
1567static void nfs_commit_release(void *calldata)
1568{
1569 struct nfs_commit_data *data = calldata;
1570
1571 data->completion_ops->completion(data);
1572 nfs_commitdata_release(calldata);
1573}
1574
1575static const struct rpc_call_ops nfs_commit_ops = {
1576 .rpc_call_prepare = nfs_commit_prepare,
1577 .rpc_call_done = nfs_commit_done,
1578 .rpc_release = nfs_commit_release,
1579};
1580
1581static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1582 .completion = nfs_commit_release_pages,
1583 .error_cleanup = nfs_commit_clear_lock,
1584};
1585
1586int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1587 int how, struct nfs_commit_info *cinfo)
1588{
1589 int status;
1590
1591 status = pnfs_commit_list(inode, head, how, cinfo);
1592 if (status == PNFS_NOT_ATTEMPTED)
1593 status = nfs_commit_list(inode, head, how, cinfo);
1594 return status;
1595}
1596
1597int nfs_commit_inode(struct inode *inode, int how)
1598{
1599 LIST_HEAD(head);
1600 struct nfs_commit_info cinfo;
1601 int may_wait = how & FLUSH_SYNC;
1602 int res;
1603
1604 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1605 if (res <= 0)
1606 goto out_mark_dirty;
1607 nfs_init_cinfo_from_inode(&cinfo, inode);
1608 res = nfs_scan_commit(inode, &head, &cinfo);
1609 if (res) {
1610 int error;
1611
1612 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1613 if (error < 0)
1614 return error;
1615 if (!may_wait)
1616 goto out_mark_dirty;
1617 error = wait_on_bit(&NFS_I(inode)->flags,
1618 NFS_INO_COMMIT,
1619 nfs_wait_bit_killable,
1620 TASK_KILLABLE);
1621 if (error < 0)
1622 return error;
1623 } else
1624 nfs_commit_clear_lock(NFS_I(inode));
1625 return res;
1626 /* Note: If we exit without ensuring that the commit is complete,
1627 * we must mark the inode as dirty. Otherwise, future calls to
1628 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1629 * that the data is on the disk.
1630 */
1631out_mark_dirty:
1632 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1633 return res;
1634}
1635
1636static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1637{
1638 struct nfs_inode *nfsi = NFS_I(inode);
1639 int flags = FLUSH_SYNC;
1640 int ret = 0;
1641
1642 /* no commits means nothing needs to be done */
1643 if (!nfsi->commit_info.ncommit)
1644 return ret;
1645
1646 if (wbc->sync_mode == WB_SYNC_NONE) {
1647 /* Don't commit yet if this is a non-blocking flush and there
1648 * are a lot of outstanding writes for this mapping.
1649 */
1650 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1651 goto out_mark_dirty;
1652
1653 /* don't wait for the COMMIT response */
1654 flags = 0;
1655 }
1656
1657 ret = nfs_commit_inode(inode, flags);
1658 if (ret >= 0) {
1659 if (wbc->sync_mode == WB_SYNC_NONE) {
1660 if (ret < wbc->nr_to_write)
1661 wbc->nr_to_write -= ret;
1662 else
1663 wbc->nr_to_write = 0;
1664 }
1665 return 0;
1666 }
1667out_mark_dirty:
1668 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1669 return ret;
1670}
1671#else
1672static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1673{
1674 return 0;
1675}
1676#endif
1677
1678int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1679{
1680 int ret;
1681
1682 ret = nfs_commit_unstable_pages(inode, wbc);
1683 if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1684 int status;
1685 bool sync = true;
1686
1687 if (wbc->sync_mode == WB_SYNC_NONE)
1688 sync = false;
1689
1690 status = pnfs_layoutcommit_inode(inode, sync);
1691 if (status < 0)
1692 return status;
1693 }
1694 return ret;
1695}
1696
1697/*
1698 * flush the inode to disk.
1699 */
1700int nfs_wb_all(struct inode *inode)
1701{
1702 struct writeback_control wbc = {
1703 .sync_mode = WB_SYNC_ALL,
1704 .nr_to_write = LONG_MAX,
1705 .range_start = 0,
1706 .range_end = LLONG_MAX,
1707 };
1708
1709 return sync_inode(inode, &wbc);
1710}
1711
1712int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1713{
1714 struct nfs_page *req;
1715 int ret = 0;
1716
1717 BUG_ON(!PageLocked(page));
1718 for (;;) {
1719 wait_on_page_writeback(page);
1720 req = nfs_page_find_request(page);
1721 if (req == NULL)
1722 break;
1723 if (nfs_lock_request(req)) {
1724 nfs_clear_request_commit(req);
1725 nfs_inode_remove_request(req);
1726 /*
1727 * In case nfs_inode_remove_request has marked the
1728 * page as being dirty
1729 */
1730 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1731 nfs_unlock_and_release_request(req);
1732 break;
1733 }
1734 ret = nfs_wait_on_request(req);
1735 nfs_release_request(req);
1736 if (ret < 0)
1737 break;
1738 }
1739 return ret;
1740}
1741
1742/*
1743 * Write back all requests on one page - we do this before reading it.
1744 */
1745int nfs_wb_page(struct inode *inode, struct page *page)
1746{
1747 loff_t range_start = page_offset(page);
1748 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1749 struct writeback_control wbc = {
1750 .sync_mode = WB_SYNC_ALL,
1751 .nr_to_write = 0,
1752 .range_start = range_start,
1753 .range_end = range_end,
1754 };
1755 int ret;
1756
1757 for (;;) {
1758 wait_on_page_writeback(page);
1759 if (clear_page_dirty_for_io(page)) {
1760 ret = nfs_writepage_locked(page, &wbc);
1761 if (ret < 0)
1762 goto out_error;
1763 continue;
1764 }
1765 if (!PagePrivate(page))
1766 break;
1767 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1768 if (ret < 0)
1769 goto out_error;
1770 }
1771 return 0;
1772out_error:
1773 return ret;
1774}
1775
1776#ifdef CONFIG_MIGRATION
1777int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1778 struct page *page, enum migrate_mode mode)
1779{
1780 /*
1781 * If PagePrivate is set, then the page is currently associated with
1782 * an in-progress read or write request. Don't try to migrate it.
1783 *
1784 * FIXME: we could do this in principle, but we'll need a way to ensure
1785 * that we can safely release the inode reference while holding
1786 * the page lock.
1787 */
1788 if (PagePrivate(page))
1789 return -EBUSY;
1790
1791 nfs_fscache_release_page(page, GFP_KERNEL);
1792
1793 return migrate_page(mapping, newpage, page, mode);
1794}
1795#endif
1796
1797int __init nfs_init_writepagecache(void)
1798{
1799 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1800 sizeof(struct nfs_write_header),
1801 0, SLAB_HWCACHE_ALIGN,
1802 NULL);
1803 if (nfs_wdata_cachep == NULL)
1804 return -ENOMEM;
1805
1806 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1807 nfs_wdata_cachep);
1808 if (nfs_wdata_mempool == NULL)
1809 goto out_destroy_write_cache;
1810
1811 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1812 sizeof(struct nfs_commit_data),
1813 0, SLAB_HWCACHE_ALIGN,
1814 NULL);
1815 if (nfs_cdata_cachep == NULL)
1816 goto out_destroy_write_mempool;
1817
1818 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1819 nfs_wdata_cachep);
1820 if (nfs_commit_mempool == NULL)
1821 goto out_destroy_commit_cache;
1822
1823 /*
1824 * NFS congestion size, scale with available memory.
1825 *
1826 * 64MB: 8192k
1827 * 128MB: 11585k
1828 * 256MB: 16384k
1829 * 512MB: 23170k
1830 * 1GB: 32768k
1831 * 2GB: 46340k
1832 * 4GB: 65536k
1833 * 8GB: 92681k
1834 * 16GB: 131072k
1835 *
1836 * This allows larger machines to have larger/more transfers.
1837 * Limit the default to 256M
1838 */
1839 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1840 if (nfs_congestion_kb > 256*1024)
1841 nfs_congestion_kb = 256*1024;
1842
1843 return 0;
1844
1845out_destroy_commit_cache:
1846 kmem_cache_destroy(nfs_cdata_cachep);
1847out_destroy_write_mempool:
1848 mempool_destroy(nfs_wdata_mempool);
1849out_destroy_write_cache:
1850 kmem_cache_destroy(nfs_wdata_cachep);
1851 return -ENOMEM;
1852}
1853
1854void nfs_destroy_writepagecache(void)
1855{
1856 mempool_destroy(nfs_commit_mempool);
1857 kmem_cache_destroy(nfs_cdata_cachep);
1858 mempool_destroy(nfs_wdata_mempool);
1859 kmem_cache_destroy(nfs_wdata_cachep);
1860}
1861
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
29#include <linux/uaccess.h>
30#include <linux/sched/mm.h>
31
32#include "delegation.h"
33#include "internal.h"
34#include "iostat.h"
35#include "nfs4_fs.h"
36#include "fscache.h"
37#include "pnfs.h"
38
39#include "nfstrace.h"
40
41#define NFSDBG_FACILITY NFSDBG_PAGECACHE
42
43#define MIN_POOL_WRITE (32)
44#define MIN_POOL_COMMIT (4)
45
46struct nfs_io_completion {
47 void (*complete)(void *data);
48 void *data;
49 struct kref refcount;
50};
51
52/*
53 * Local function declarations
54 */
55static void nfs_redirty_request(struct nfs_page *req);
56static const struct rpc_call_ops nfs_commit_ops;
57static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
58static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
59static const struct nfs_rw_ops nfs_rw_write_ops;
60static void nfs_inode_remove_request(struct nfs_page *req);
61static void nfs_clear_request_commit(struct nfs_page *req);
62static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
63 struct inode *inode);
64static struct nfs_page *
65nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
66 struct page *page);
67
68static struct kmem_cache *nfs_wdata_cachep;
69static mempool_t *nfs_wdata_mempool;
70static struct kmem_cache *nfs_cdata_cachep;
71static mempool_t *nfs_commit_mempool;
72
73struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
74{
75 struct nfs_commit_data *p;
76
77 if (never_fail)
78 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
79 else {
80 /* It is OK to do some reclaim, not no safe to wait
81 * for anything to be returned to the pool.
82 * mempool_alloc() cannot handle that particular combination,
83 * so we need two separate attempts.
84 */
85 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
86 if (!p)
87 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
88 __GFP_NOWARN | __GFP_NORETRY);
89 if (!p)
90 return NULL;
91 }
92
93 memset(p, 0, sizeof(*p));
94 INIT_LIST_HEAD(&p->pages);
95 return p;
96}
97EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
98
99void nfs_commit_free(struct nfs_commit_data *p)
100{
101 mempool_free(p, nfs_commit_mempool);
102}
103EXPORT_SYMBOL_GPL(nfs_commit_free);
104
105static struct nfs_pgio_header *nfs_writehdr_alloc(void)
106{
107 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_KERNEL);
108
109 memset(p, 0, sizeof(*p));
110 p->rw_mode = FMODE_WRITE;
111 return p;
112}
113
114static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
115{
116 mempool_free(hdr, nfs_wdata_mempool);
117}
118
119static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
120{
121 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
122}
123
124static void nfs_io_completion_init(struct nfs_io_completion *ioc,
125 void (*complete)(void *), void *data)
126{
127 ioc->complete = complete;
128 ioc->data = data;
129 kref_init(&ioc->refcount);
130}
131
132static void nfs_io_completion_release(struct kref *kref)
133{
134 struct nfs_io_completion *ioc = container_of(kref,
135 struct nfs_io_completion, refcount);
136 ioc->complete(ioc->data);
137 kfree(ioc);
138}
139
140static void nfs_io_completion_get(struct nfs_io_completion *ioc)
141{
142 if (ioc != NULL)
143 kref_get(&ioc->refcount);
144}
145
146static void nfs_io_completion_put(struct nfs_io_completion *ioc)
147{
148 if (ioc != NULL)
149 kref_put(&ioc->refcount, nfs_io_completion_release);
150}
151
152static void
153nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
154{
155 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
156 kref_get(&req->wb_kref);
157 atomic_long_inc(&NFS_I(inode)->nrequests);
158 }
159}
160
161static int
162nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
163{
164 int ret;
165
166 if (!test_bit(PG_REMOVE, &req->wb_flags))
167 return 0;
168 ret = nfs_page_group_lock(req);
169 if (ret)
170 return ret;
171 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
172 nfs_page_set_inode_ref(req, inode);
173 nfs_page_group_unlock(req);
174 return 0;
175}
176
177static struct nfs_page *
178nfs_page_private_request(struct page *page)
179{
180 if (!PagePrivate(page))
181 return NULL;
182 return (struct nfs_page *)page_private(page);
183}
184
185/*
186 * nfs_page_find_head_request_locked - find head request associated with @page
187 *
188 * must be called while holding the inode lock.
189 *
190 * returns matching head request with reference held, or NULL if not found.
191 */
192static struct nfs_page *
193nfs_page_find_private_request(struct page *page)
194{
195 struct address_space *mapping = page_file_mapping(page);
196 struct nfs_page *req;
197
198 if (!PagePrivate(page))
199 return NULL;
200 spin_lock(&mapping->private_lock);
201 req = nfs_page_private_request(page);
202 if (req) {
203 WARN_ON_ONCE(req->wb_head != req);
204 kref_get(&req->wb_kref);
205 }
206 spin_unlock(&mapping->private_lock);
207 return req;
208}
209
210static struct nfs_page *
211nfs_page_find_swap_request(struct page *page)
212{
213 struct inode *inode = page_file_mapping(page)->host;
214 struct nfs_inode *nfsi = NFS_I(inode);
215 struct nfs_page *req = NULL;
216 if (!PageSwapCache(page))
217 return NULL;
218 mutex_lock(&nfsi->commit_mutex);
219 if (PageSwapCache(page)) {
220 req = nfs_page_search_commits_for_head_request_locked(nfsi,
221 page);
222 if (req) {
223 WARN_ON_ONCE(req->wb_head != req);
224 kref_get(&req->wb_kref);
225 }
226 }
227 mutex_unlock(&nfsi->commit_mutex);
228 return req;
229}
230
231/*
232 * nfs_page_find_head_request - find head request associated with @page
233 *
234 * returns matching head request with reference held, or NULL if not found.
235 */
236static struct nfs_page *nfs_page_find_head_request(struct page *page)
237{
238 struct nfs_page *req;
239
240 req = nfs_page_find_private_request(page);
241 if (!req)
242 req = nfs_page_find_swap_request(page);
243 return req;
244}
245
246static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
247{
248 struct inode *inode = page_file_mapping(page)->host;
249 struct nfs_page *req, *head;
250 int ret;
251
252 for (;;) {
253 req = nfs_page_find_head_request(page);
254 if (!req)
255 return req;
256 head = nfs_page_group_lock_head(req);
257 if (head != req)
258 nfs_release_request(req);
259 if (IS_ERR(head))
260 return head;
261 ret = nfs_cancel_remove_inode(head, inode);
262 if (ret < 0) {
263 nfs_unlock_and_release_request(head);
264 return ERR_PTR(ret);
265 }
266 /* Ensure that nobody removed the request before we locked it */
267 if (head == nfs_page_private_request(page))
268 break;
269 if (PageSwapCache(page))
270 break;
271 nfs_unlock_and_release_request(head);
272 }
273 return head;
274}
275
276/* Adjust the file length if we're writing beyond the end */
277static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
278{
279 struct inode *inode = page_file_mapping(page)->host;
280 loff_t end, i_size;
281 pgoff_t end_index;
282
283 spin_lock(&inode->i_lock);
284 i_size = i_size_read(inode);
285 end_index = (i_size - 1) >> PAGE_SHIFT;
286 if (i_size > 0 && page_index(page) < end_index)
287 goto out;
288 end = page_file_offset(page) + ((loff_t)offset+count);
289 if (i_size >= end)
290 goto out;
291 i_size_write(inode, end);
292 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
293 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
294out:
295 spin_unlock(&inode->i_lock);
296}
297
298/* A writeback failed: mark the page as bad, and invalidate the page cache */
299static void nfs_set_pageerror(struct address_space *mapping)
300{
301 struct inode *inode = mapping->host;
302
303 nfs_zap_mapping(mapping->host, mapping);
304 /* Force file size revalidation */
305 spin_lock(&inode->i_lock);
306 nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
307 NFS_INO_REVAL_PAGECACHE |
308 NFS_INO_INVALID_SIZE);
309 spin_unlock(&inode->i_lock);
310}
311
312static void nfs_mapping_set_error(struct page *page, int error)
313{
314 struct address_space *mapping = page_file_mapping(page);
315
316 SetPageError(page);
317 mapping_set_error(mapping, error);
318 nfs_set_pageerror(mapping);
319}
320
321/*
322 * nfs_page_group_search_locked
323 * @head - head request of page group
324 * @page_offset - offset into page
325 *
326 * Search page group with head @head to find a request that contains the
327 * page offset @page_offset.
328 *
329 * Returns a pointer to the first matching nfs request, or NULL if no
330 * match is found.
331 *
332 * Must be called with the page group lock held
333 */
334static struct nfs_page *
335nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
336{
337 struct nfs_page *req;
338
339 req = head;
340 do {
341 if (page_offset >= req->wb_pgbase &&
342 page_offset < (req->wb_pgbase + req->wb_bytes))
343 return req;
344
345 req = req->wb_this_page;
346 } while (req != head);
347
348 return NULL;
349}
350
351/*
352 * nfs_page_group_covers_page
353 * @head - head request of page group
354 *
355 * Return true if the page group with head @head covers the whole page,
356 * returns false otherwise
357 */
358static bool nfs_page_group_covers_page(struct nfs_page *req)
359{
360 struct nfs_page *tmp;
361 unsigned int pos = 0;
362 unsigned int len = nfs_page_length(req->wb_page);
363
364 nfs_page_group_lock(req);
365
366 for (;;) {
367 tmp = nfs_page_group_search_locked(req->wb_head, pos);
368 if (!tmp)
369 break;
370 pos = tmp->wb_pgbase + tmp->wb_bytes;
371 }
372
373 nfs_page_group_unlock(req);
374 return pos >= len;
375}
376
377/* We can set the PG_uptodate flag if we see that a write request
378 * covers the full page.
379 */
380static void nfs_mark_uptodate(struct nfs_page *req)
381{
382 if (PageUptodate(req->wb_page))
383 return;
384 if (!nfs_page_group_covers_page(req))
385 return;
386 SetPageUptodate(req->wb_page);
387}
388
389static int wb_priority(struct writeback_control *wbc)
390{
391 int ret = 0;
392
393 if (wbc->sync_mode == WB_SYNC_ALL)
394 ret = FLUSH_COND_STABLE;
395 return ret;
396}
397
398/*
399 * NFS congestion control
400 */
401
402int nfs_congestion_kb;
403
404#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
405#define NFS_CONGESTION_OFF_THRESH \
406 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
407
408static void nfs_set_page_writeback(struct page *page)
409{
410 struct inode *inode = page_file_mapping(page)->host;
411 struct nfs_server *nfss = NFS_SERVER(inode);
412 int ret = test_set_page_writeback(page);
413
414 WARN_ON_ONCE(ret != 0);
415
416 if (atomic_long_inc_return(&nfss->writeback) >
417 NFS_CONGESTION_ON_THRESH)
418 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
419}
420
421static void nfs_end_page_writeback(struct nfs_page *req)
422{
423 struct inode *inode = page_file_mapping(req->wb_page)->host;
424 struct nfs_server *nfss = NFS_SERVER(inode);
425 bool is_done;
426
427 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
428 nfs_unlock_request(req);
429 if (!is_done)
430 return;
431
432 end_page_writeback(req->wb_page);
433 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
434 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
435}
436
437/*
438 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
439 *
440 * @destroy_list - request list (using wb_this_page) terminated by @old_head
441 * @old_head - the old head of the list
442 *
443 * All subrequests must be locked and removed from all lists, so at this point
444 * they are only "active" in this function, and possibly in nfs_wait_on_request
445 * with a reference held by some other context.
446 */
447static void
448nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
449 struct nfs_page *old_head,
450 struct inode *inode)
451{
452 while (destroy_list) {
453 struct nfs_page *subreq = destroy_list;
454
455 destroy_list = (subreq->wb_this_page == old_head) ?
456 NULL : subreq->wb_this_page;
457
458 /* Note: lock subreq in order to change subreq->wb_head */
459 nfs_page_set_headlock(subreq);
460 WARN_ON_ONCE(old_head != subreq->wb_head);
461
462 /* make sure old group is not used */
463 subreq->wb_this_page = subreq;
464 subreq->wb_head = subreq;
465
466 clear_bit(PG_REMOVE, &subreq->wb_flags);
467
468 /* Note: races with nfs_page_group_destroy() */
469 if (!kref_read(&subreq->wb_kref)) {
470 /* Check if we raced with nfs_page_group_destroy() */
471 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
472 nfs_page_clear_headlock(subreq);
473 nfs_free_request(subreq);
474 } else
475 nfs_page_clear_headlock(subreq);
476 continue;
477 }
478 nfs_page_clear_headlock(subreq);
479
480 nfs_release_request(old_head);
481
482 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
483 nfs_release_request(subreq);
484 atomic_long_dec(&NFS_I(inode)->nrequests);
485 }
486
487 /* subreq is now totally disconnected from page group or any
488 * write / commit lists. last chance to wake any waiters */
489 nfs_unlock_and_release_request(subreq);
490 }
491}
492
493/*
494 * nfs_join_page_group - destroy subrequests of the head req
495 * @head: the page used to lookup the "page group" of nfs_page structures
496 * @inode: Inode to which the request belongs.
497 *
498 * This function joins all sub requests to the head request by first
499 * locking all requests in the group, cancelling any pending operations
500 * and finally updating the head request to cover the whole range covered by
501 * the (former) group. All subrequests are removed from any write or commit
502 * lists, unlinked from the group and destroyed.
503 */
504void
505nfs_join_page_group(struct nfs_page *head, struct inode *inode)
506{
507 struct nfs_page *subreq;
508 struct nfs_page *destroy_list = NULL;
509 unsigned int pgbase, off, bytes;
510
511 pgbase = head->wb_pgbase;
512 bytes = head->wb_bytes;
513 off = head->wb_offset;
514 for (subreq = head->wb_this_page; subreq != head;
515 subreq = subreq->wb_this_page) {
516 /* Subrequests should always form a contiguous range */
517 if (pgbase > subreq->wb_pgbase) {
518 off -= pgbase - subreq->wb_pgbase;
519 bytes += pgbase - subreq->wb_pgbase;
520 pgbase = subreq->wb_pgbase;
521 }
522 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
523 - pgbase, bytes);
524 }
525
526 /* Set the head request's range to cover the former page group */
527 head->wb_pgbase = pgbase;
528 head->wb_bytes = bytes;
529 head->wb_offset = off;
530
531 /* Now that all requests are locked, make sure they aren't on any list.
532 * Commit list removal accounting is done after locks are dropped */
533 subreq = head;
534 do {
535 nfs_clear_request_commit(subreq);
536 subreq = subreq->wb_this_page;
537 } while (subreq != head);
538
539 /* unlink subrequests from head, destroy them later */
540 if (head->wb_this_page != head) {
541 /* destroy list will be terminated by head */
542 destroy_list = head->wb_this_page;
543 head->wb_this_page = head;
544 }
545
546 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
547}
548
549/*
550 * nfs_lock_and_join_requests - join all subreqs to the head req
551 * @page: the page used to lookup the "page group" of nfs_page structures
552 *
553 * This function joins all sub requests to the head request by first
554 * locking all requests in the group, cancelling any pending operations
555 * and finally updating the head request to cover the whole range covered by
556 * the (former) group. All subrequests are removed from any write or commit
557 * lists, unlinked from the group and destroyed.
558 *
559 * Returns a locked, referenced pointer to the head request - which after
560 * this call is guaranteed to be the only request associated with the page.
561 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
562 * error was encountered.
563 */
564static struct nfs_page *
565nfs_lock_and_join_requests(struct page *page)
566{
567 struct inode *inode = page_file_mapping(page)->host;
568 struct nfs_page *head;
569 int ret;
570
571 /*
572 * A reference is taken only on the head request which acts as a
573 * reference to the whole page group - the group will not be destroyed
574 * until the head reference is released.
575 */
576 head = nfs_find_and_lock_page_request(page);
577 if (IS_ERR_OR_NULL(head))
578 return head;
579
580 /* lock each request in the page group */
581 ret = nfs_page_group_lock_subrequests(head);
582 if (ret < 0) {
583 nfs_unlock_and_release_request(head);
584 return ERR_PTR(ret);
585 }
586
587 nfs_join_page_group(head, inode);
588
589 return head;
590}
591
592static void nfs_write_error(struct nfs_page *req, int error)
593{
594 trace_nfs_write_error(req, error);
595 nfs_mapping_set_error(req->wb_page, error);
596 nfs_inode_remove_request(req);
597 nfs_end_page_writeback(req);
598 nfs_release_request(req);
599}
600
601/*
602 * Find an associated nfs write request, and prepare to flush it out
603 * May return an error if the user signalled nfs_wait_on_request().
604 */
605static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
606 struct page *page)
607{
608 struct nfs_page *req;
609 int ret = 0;
610
611 req = nfs_lock_and_join_requests(page);
612 if (!req)
613 goto out;
614 ret = PTR_ERR(req);
615 if (IS_ERR(req))
616 goto out;
617
618 nfs_set_page_writeback(page);
619 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
620
621 /* If there is a fatal error that covers this write, just exit */
622 ret = pgio->pg_error;
623 if (nfs_error_is_fatal_on_server(ret))
624 goto out_launder;
625
626 ret = 0;
627 if (!nfs_pageio_add_request(pgio, req)) {
628 ret = pgio->pg_error;
629 /*
630 * Remove the problematic req upon fatal errors on the server
631 */
632 if (nfs_error_is_fatal(ret)) {
633 if (nfs_error_is_fatal_on_server(ret))
634 goto out_launder;
635 } else
636 ret = -EAGAIN;
637 nfs_redirty_request(req);
638 pgio->pg_error = 0;
639 } else
640 nfs_add_stats(page_file_mapping(page)->host,
641 NFSIOS_WRITEPAGES, 1);
642out:
643 return ret;
644out_launder:
645 nfs_write_error(req, ret);
646 return 0;
647}
648
649static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
650 struct nfs_pageio_descriptor *pgio)
651{
652 int ret;
653
654 nfs_pageio_cond_complete(pgio, page_index(page));
655 ret = nfs_page_async_flush(pgio, page);
656 if (ret == -EAGAIN) {
657 redirty_page_for_writepage(wbc, page);
658 ret = AOP_WRITEPAGE_ACTIVATE;
659 }
660 return ret;
661}
662
663/*
664 * Write an mmapped page to the server.
665 */
666static int nfs_writepage_locked(struct page *page,
667 struct writeback_control *wbc)
668{
669 struct nfs_pageio_descriptor pgio;
670 struct inode *inode = page_file_mapping(page)->host;
671 int err;
672
673 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
674 nfs_pageio_init_write(&pgio, inode, 0,
675 false, &nfs_async_write_completion_ops);
676 err = nfs_do_writepage(page, wbc, &pgio);
677 pgio.pg_error = 0;
678 nfs_pageio_complete(&pgio);
679 if (err < 0)
680 return err;
681 if (nfs_error_is_fatal(pgio.pg_error))
682 return pgio.pg_error;
683 return 0;
684}
685
686int nfs_writepage(struct page *page, struct writeback_control *wbc)
687{
688 int ret;
689
690 ret = nfs_writepage_locked(page, wbc);
691 if (ret != AOP_WRITEPAGE_ACTIVATE)
692 unlock_page(page);
693 return ret;
694}
695
696static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
697{
698 int ret;
699
700 ret = nfs_do_writepage(page, wbc, data);
701 if (ret != AOP_WRITEPAGE_ACTIVATE)
702 unlock_page(page);
703 return ret;
704}
705
706static void nfs_io_completion_commit(void *inode)
707{
708 nfs_commit_inode(inode, 0);
709}
710
711int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
712{
713 struct inode *inode = mapping->host;
714 struct nfs_pageio_descriptor pgio;
715 struct nfs_io_completion *ioc = NULL;
716 unsigned int mntflags = NFS_SERVER(inode)->flags;
717 int priority = 0;
718 int err;
719
720 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
721
722 if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
723 wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
724 ioc = nfs_io_completion_alloc(GFP_KERNEL);
725 if (ioc)
726 nfs_io_completion_init(ioc, nfs_io_completion_commit,
727 inode);
728 priority = wb_priority(wbc);
729 }
730
731 nfs_pageio_init_write(&pgio, inode, priority, false,
732 &nfs_async_write_completion_ops);
733 pgio.pg_io_completion = ioc;
734 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
735 pgio.pg_error = 0;
736 nfs_pageio_complete(&pgio);
737 nfs_io_completion_put(ioc);
738
739 if (err < 0)
740 goto out_err;
741 err = pgio.pg_error;
742 if (nfs_error_is_fatal(err))
743 goto out_err;
744 return 0;
745out_err:
746 return err;
747}
748
749/*
750 * Insert a write request into an inode
751 */
752static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
753{
754 struct address_space *mapping = page_file_mapping(req->wb_page);
755 struct nfs_inode *nfsi = NFS_I(inode);
756
757 WARN_ON_ONCE(req->wb_this_page != req);
758
759 /* Lock the request! */
760 nfs_lock_request(req);
761
762 /*
763 * Swap-space should not get truncated. Hence no need to plug the race
764 * with invalidate/truncate.
765 */
766 spin_lock(&mapping->private_lock);
767 if (likely(!PageSwapCache(req->wb_page))) {
768 set_bit(PG_MAPPED, &req->wb_flags);
769 SetPagePrivate(req->wb_page);
770 set_page_private(req->wb_page, (unsigned long)req);
771 }
772 spin_unlock(&mapping->private_lock);
773 atomic_long_inc(&nfsi->nrequests);
774 /* this a head request for a page group - mark it as having an
775 * extra reference so sub groups can follow suit.
776 * This flag also informs pgio layer when to bump nrequests when
777 * adding subrequests. */
778 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
779 kref_get(&req->wb_kref);
780}
781
782/*
783 * Remove a write request from an inode
784 */
785static void nfs_inode_remove_request(struct nfs_page *req)
786{
787 struct address_space *mapping = page_file_mapping(req->wb_page);
788 struct inode *inode = mapping->host;
789 struct nfs_inode *nfsi = NFS_I(inode);
790 struct nfs_page *head;
791
792 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
793 head = req->wb_head;
794
795 spin_lock(&mapping->private_lock);
796 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
797 set_page_private(head->wb_page, 0);
798 ClearPagePrivate(head->wb_page);
799 clear_bit(PG_MAPPED, &head->wb_flags);
800 }
801 spin_unlock(&mapping->private_lock);
802 }
803
804 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
805 nfs_release_request(req);
806 atomic_long_dec(&nfsi->nrequests);
807 }
808}
809
810static void
811nfs_mark_request_dirty(struct nfs_page *req)
812{
813 if (req->wb_page)
814 __set_page_dirty_nobuffers(req->wb_page);
815}
816
817/*
818 * nfs_page_search_commits_for_head_request_locked
819 *
820 * Search through commit lists on @inode for the head request for @page.
821 * Must be called while holding the inode (which is cinfo) lock.
822 *
823 * Returns the head request if found, or NULL if not found.
824 */
825static struct nfs_page *
826nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
827 struct page *page)
828{
829 struct nfs_page *freq, *t;
830 struct nfs_commit_info cinfo;
831 struct inode *inode = &nfsi->vfs_inode;
832
833 nfs_init_cinfo_from_inode(&cinfo, inode);
834
835 /* search through pnfs commit lists */
836 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
837 if (freq)
838 return freq->wb_head;
839
840 /* Linearly search the commit list for the correct request */
841 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
842 if (freq->wb_page == page)
843 return freq->wb_head;
844 }
845
846 return NULL;
847}
848
849/**
850 * nfs_request_add_commit_list_locked - add request to a commit list
851 * @req: pointer to a struct nfs_page
852 * @dst: commit list head
853 * @cinfo: holds list lock and accounting info
854 *
855 * This sets the PG_CLEAN bit, updates the cinfo count of
856 * number of outstanding requests requiring a commit as well as
857 * the MM page stats.
858 *
859 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
860 * nfs_page lock.
861 */
862void
863nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
864 struct nfs_commit_info *cinfo)
865{
866 set_bit(PG_CLEAN, &req->wb_flags);
867 nfs_list_add_request(req, dst);
868 atomic_long_inc(&cinfo->mds->ncommit);
869}
870EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
871
872/**
873 * nfs_request_add_commit_list - add request to a commit list
874 * @req: pointer to a struct nfs_page
875 * @cinfo: holds list lock and accounting info
876 *
877 * This sets the PG_CLEAN bit, updates the cinfo count of
878 * number of outstanding requests requiring a commit as well as
879 * the MM page stats.
880 *
881 * The caller must _not_ hold the cinfo->lock, but must be
882 * holding the nfs_page lock.
883 */
884void
885nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
886{
887 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
888 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
889 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
890 if (req->wb_page)
891 nfs_mark_page_unstable(req->wb_page, cinfo);
892}
893EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
894
895/**
896 * nfs_request_remove_commit_list - Remove request from a commit list
897 * @req: pointer to a nfs_page
898 * @cinfo: holds list lock and accounting info
899 *
900 * This clears the PG_CLEAN bit, and updates the cinfo's count of
901 * number of outstanding requests requiring a commit
902 * It does not update the MM page stats.
903 *
904 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
905 */
906void
907nfs_request_remove_commit_list(struct nfs_page *req,
908 struct nfs_commit_info *cinfo)
909{
910 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
911 return;
912 nfs_list_remove_request(req);
913 atomic_long_dec(&cinfo->mds->ncommit);
914}
915EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
916
917static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
918 struct inode *inode)
919{
920 cinfo->inode = inode;
921 cinfo->mds = &NFS_I(inode)->commit_info;
922 cinfo->ds = pnfs_get_ds_info(inode);
923 cinfo->dreq = NULL;
924 cinfo->completion_ops = &nfs_commit_completion_ops;
925}
926
927void nfs_init_cinfo(struct nfs_commit_info *cinfo,
928 struct inode *inode,
929 struct nfs_direct_req *dreq)
930{
931 if (dreq)
932 nfs_init_cinfo_from_dreq(cinfo, dreq);
933 else
934 nfs_init_cinfo_from_inode(cinfo, inode);
935}
936EXPORT_SYMBOL_GPL(nfs_init_cinfo);
937
938/*
939 * Add a request to the inode's commit list.
940 */
941void
942nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
943 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
944{
945 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
946 return;
947 nfs_request_add_commit_list(req, cinfo);
948}
949
950static void
951nfs_clear_page_commit(struct page *page)
952{
953 dec_node_page_state(page, NR_WRITEBACK);
954 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
955 WB_WRITEBACK);
956}
957
958/* Called holding the request lock on @req */
959static void
960nfs_clear_request_commit(struct nfs_page *req)
961{
962 if (test_bit(PG_CLEAN, &req->wb_flags)) {
963 struct nfs_open_context *ctx = nfs_req_openctx(req);
964 struct inode *inode = d_inode(ctx->dentry);
965 struct nfs_commit_info cinfo;
966
967 nfs_init_cinfo_from_inode(&cinfo, inode);
968 mutex_lock(&NFS_I(inode)->commit_mutex);
969 if (!pnfs_clear_request_commit(req, &cinfo)) {
970 nfs_request_remove_commit_list(req, &cinfo);
971 }
972 mutex_unlock(&NFS_I(inode)->commit_mutex);
973 nfs_clear_page_commit(req->wb_page);
974 }
975}
976
977int nfs_write_need_commit(struct nfs_pgio_header *hdr)
978{
979 if (hdr->verf.committed == NFS_DATA_SYNC)
980 return hdr->lseg == NULL;
981 return hdr->verf.committed != NFS_FILE_SYNC;
982}
983
984static void nfs_async_write_init(struct nfs_pgio_header *hdr)
985{
986 nfs_io_completion_get(hdr->io_completion);
987}
988
989static void nfs_write_completion(struct nfs_pgio_header *hdr)
990{
991 struct nfs_commit_info cinfo;
992 unsigned long bytes = 0;
993
994 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
995 goto out;
996 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
997 while (!list_empty(&hdr->pages)) {
998 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
999
1000 bytes += req->wb_bytes;
1001 nfs_list_remove_request(req);
1002 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1003 (hdr->good_bytes < bytes)) {
1004 trace_nfs_comp_error(req, hdr->error);
1005 nfs_mapping_set_error(req->wb_page, hdr->error);
1006 goto remove_req;
1007 }
1008 if (nfs_write_need_commit(hdr)) {
1009 /* Reset wb_nio, since the write was successful. */
1010 req->wb_nio = 0;
1011 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1012 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1013 hdr->pgio_mirror_idx);
1014 goto next;
1015 }
1016remove_req:
1017 nfs_inode_remove_request(req);
1018next:
1019 nfs_end_page_writeback(req);
1020 nfs_release_request(req);
1021 }
1022out:
1023 nfs_io_completion_put(hdr->io_completion);
1024 hdr->release(hdr);
1025}
1026
1027unsigned long
1028nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1029{
1030 return atomic_long_read(&cinfo->mds->ncommit);
1031}
1032
1033/* NFS_I(cinfo->inode)->commit_mutex held by caller */
1034int
1035nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1036 struct nfs_commit_info *cinfo, int max)
1037{
1038 struct nfs_page *req, *tmp;
1039 int ret = 0;
1040
1041restart:
1042 list_for_each_entry_safe(req, tmp, src, wb_list) {
1043 kref_get(&req->wb_kref);
1044 if (!nfs_lock_request(req)) {
1045 int status;
1046
1047 /* Prevent deadlock with nfs_lock_and_join_requests */
1048 if (!list_empty(dst)) {
1049 nfs_release_request(req);
1050 continue;
1051 }
1052 /* Ensure we make progress to prevent livelock */
1053 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1054 status = nfs_wait_on_request(req);
1055 nfs_release_request(req);
1056 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1057 if (status < 0)
1058 break;
1059 goto restart;
1060 }
1061 nfs_request_remove_commit_list(req, cinfo);
1062 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1063 nfs_list_add_request(req, dst);
1064 ret++;
1065 if ((ret == max) && !cinfo->dreq)
1066 break;
1067 cond_resched();
1068 }
1069 return ret;
1070}
1071EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1072
1073/*
1074 * nfs_scan_commit - Scan an inode for commit requests
1075 * @inode: NFS inode to scan
1076 * @dst: mds destination list
1077 * @cinfo: mds and ds lists of reqs ready to commit
1078 *
1079 * Moves requests from the inode's 'commit' request list.
1080 * The requests are *not* checked to ensure that they form a contiguous set.
1081 */
1082int
1083nfs_scan_commit(struct inode *inode, struct list_head *dst,
1084 struct nfs_commit_info *cinfo)
1085{
1086 int ret = 0;
1087
1088 if (!atomic_long_read(&cinfo->mds->ncommit))
1089 return 0;
1090 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1091 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1092 const int max = INT_MAX;
1093
1094 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1095 cinfo, max);
1096 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1097 }
1098 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1099 return ret;
1100}
1101
1102/*
1103 * Search for an existing write request, and attempt to update
1104 * it to reflect a new dirty region on a given page.
1105 *
1106 * If the attempt fails, then the existing request is flushed out
1107 * to disk.
1108 */
1109static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1110 struct page *page,
1111 unsigned int offset,
1112 unsigned int bytes)
1113{
1114 struct nfs_page *req;
1115 unsigned int rqend;
1116 unsigned int end;
1117 int error;
1118
1119 end = offset + bytes;
1120
1121 req = nfs_lock_and_join_requests(page);
1122 if (IS_ERR_OR_NULL(req))
1123 return req;
1124
1125 rqend = req->wb_offset + req->wb_bytes;
1126 /*
1127 * Tell the caller to flush out the request if
1128 * the offsets are non-contiguous.
1129 * Note: nfs_flush_incompatible() will already
1130 * have flushed out requests having wrong owners.
1131 */
1132 if (offset > rqend || end < req->wb_offset)
1133 goto out_flushme;
1134
1135 /* Okay, the request matches. Update the region */
1136 if (offset < req->wb_offset) {
1137 req->wb_offset = offset;
1138 req->wb_pgbase = offset;
1139 }
1140 if (end > rqend)
1141 req->wb_bytes = end - req->wb_offset;
1142 else
1143 req->wb_bytes = rqend - req->wb_offset;
1144 req->wb_nio = 0;
1145 return req;
1146out_flushme:
1147 /*
1148 * Note: we mark the request dirty here because
1149 * nfs_lock_and_join_requests() cannot preserve
1150 * commit flags, so we have to replay the write.
1151 */
1152 nfs_mark_request_dirty(req);
1153 nfs_unlock_and_release_request(req);
1154 error = nfs_wb_page(inode, page);
1155 return (error < 0) ? ERR_PTR(error) : NULL;
1156}
1157
1158/*
1159 * Try to update an existing write request, or create one if there is none.
1160 *
1161 * Note: Should always be called with the Page Lock held to prevent races
1162 * if we have to add a new request. Also assumes that the caller has
1163 * already called nfs_flush_incompatible() if necessary.
1164 */
1165static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1166 struct page *page, unsigned int offset, unsigned int bytes)
1167{
1168 struct inode *inode = page_file_mapping(page)->host;
1169 struct nfs_page *req;
1170
1171 req = nfs_try_to_update_request(inode, page, offset, bytes);
1172 if (req != NULL)
1173 goto out;
1174 req = nfs_create_request(ctx, page, offset, bytes);
1175 if (IS_ERR(req))
1176 goto out;
1177 nfs_inode_add_request(inode, req);
1178out:
1179 return req;
1180}
1181
1182static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1183 unsigned int offset, unsigned int count)
1184{
1185 struct nfs_page *req;
1186
1187 req = nfs_setup_write_request(ctx, page, offset, count);
1188 if (IS_ERR(req))
1189 return PTR_ERR(req);
1190 /* Update file length */
1191 nfs_grow_file(page, offset, count);
1192 nfs_mark_uptodate(req);
1193 nfs_mark_request_dirty(req);
1194 nfs_unlock_and_release_request(req);
1195 return 0;
1196}
1197
1198int nfs_flush_incompatible(struct file *file, struct page *page)
1199{
1200 struct nfs_open_context *ctx = nfs_file_open_context(file);
1201 struct nfs_lock_context *l_ctx;
1202 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1203 struct nfs_page *req;
1204 int do_flush, status;
1205 /*
1206 * Look for a request corresponding to this page. If there
1207 * is one, and it belongs to another file, we flush it out
1208 * before we try to copy anything into the page. Do this
1209 * due to the lack of an ACCESS-type call in NFSv2.
1210 * Also do the same if we find a request from an existing
1211 * dropped page.
1212 */
1213 do {
1214 req = nfs_page_find_head_request(page);
1215 if (req == NULL)
1216 return 0;
1217 l_ctx = req->wb_lock_context;
1218 do_flush = req->wb_page != page ||
1219 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1220 if (l_ctx && flctx &&
1221 !(list_empty_careful(&flctx->flc_posix) &&
1222 list_empty_careful(&flctx->flc_flock))) {
1223 do_flush |= l_ctx->lockowner != current->files;
1224 }
1225 nfs_release_request(req);
1226 if (!do_flush)
1227 return 0;
1228 status = nfs_wb_page(page_file_mapping(page)->host, page);
1229 } while (status == 0);
1230 return status;
1231}
1232
1233/*
1234 * Avoid buffered writes when a open context credential's key would
1235 * expire soon.
1236 *
1237 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1238 *
1239 * Return 0 and set a credential flag which triggers the inode to flush
1240 * and performs NFS_FILE_SYNC writes if the key will expired within
1241 * RPC_KEY_EXPIRE_TIMEO.
1242 */
1243int
1244nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1245{
1246 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1247
1248 if (nfs_ctx_key_to_expire(ctx, inode) &&
1249 !ctx->ll_cred)
1250 /* Already expired! */
1251 return -EACCES;
1252 return 0;
1253}
1254
1255/*
1256 * Test if the open context credential key is marked to expire soon.
1257 */
1258bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1259{
1260 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1261 struct rpc_cred *cred = ctx->ll_cred;
1262 struct auth_cred acred = {
1263 .cred = ctx->cred,
1264 };
1265
1266 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1267 put_rpccred(cred);
1268 ctx->ll_cred = NULL;
1269 cred = NULL;
1270 }
1271 if (!cred)
1272 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1273 if (!cred || IS_ERR(cred))
1274 return true;
1275 ctx->ll_cred = cred;
1276 return !!(cred->cr_ops->crkey_timeout &&
1277 cred->cr_ops->crkey_timeout(cred));
1278}
1279
1280/*
1281 * If the page cache is marked as unsafe or invalid, then we can't rely on
1282 * the PageUptodate() flag. In this case, we will need to turn off
1283 * write optimisations that depend on the page contents being correct.
1284 */
1285static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1286 unsigned int pagelen)
1287{
1288 struct nfs_inode *nfsi = NFS_I(inode);
1289
1290 if (nfs_have_delegated_attributes(inode))
1291 goto out;
1292 if (nfsi->cache_validity &
1293 (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1294 return false;
1295 smp_rmb();
1296 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1297 return false;
1298out:
1299 if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1300 return false;
1301 return PageUptodate(page) != 0;
1302}
1303
1304static bool
1305is_whole_file_wrlock(struct file_lock *fl)
1306{
1307 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1308 fl->fl_type == F_WRLCK;
1309}
1310
1311/* If we know the page is up to date, and we're not using byte range locks (or
1312 * if we have the whole file locked for writing), it may be more efficient to
1313 * extend the write to cover the entire page in order to avoid fragmentation
1314 * inefficiencies.
1315 *
1316 * If the file is opened for synchronous writes then we can just skip the rest
1317 * of the checks.
1318 */
1319static int nfs_can_extend_write(struct file *file, struct page *page,
1320 struct inode *inode, unsigned int pagelen)
1321{
1322 int ret;
1323 struct file_lock_context *flctx = inode->i_flctx;
1324 struct file_lock *fl;
1325
1326 if (file->f_flags & O_DSYNC)
1327 return 0;
1328 if (!nfs_write_pageuptodate(page, inode, pagelen))
1329 return 0;
1330 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1331 return 1;
1332 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1333 list_empty_careful(&flctx->flc_posix)))
1334 return 1;
1335
1336 /* Check to see if there are whole file write locks */
1337 ret = 0;
1338 spin_lock(&flctx->flc_lock);
1339 if (!list_empty(&flctx->flc_posix)) {
1340 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1341 fl_list);
1342 if (is_whole_file_wrlock(fl))
1343 ret = 1;
1344 } else if (!list_empty(&flctx->flc_flock)) {
1345 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1346 fl_list);
1347 if (fl->fl_type == F_WRLCK)
1348 ret = 1;
1349 }
1350 spin_unlock(&flctx->flc_lock);
1351 return ret;
1352}
1353
1354/*
1355 * Update and possibly write a cached page of an NFS file.
1356 *
1357 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1358 * things with a page scheduled for an RPC call (e.g. invalidate it).
1359 */
1360int nfs_updatepage(struct file *file, struct page *page,
1361 unsigned int offset, unsigned int count)
1362{
1363 struct nfs_open_context *ctx = nfs_file_open_context(file);
1364 struct address_space *mapping = page_file_mapping(page);
1365 struct inode *inode = mapping->host;
1366 unsigned int pagelen = nfs_page_length(page);
1367 int status = 0;
1368
1369 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1370
1371 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1372 file, count, (long long)(page_file_offset(page) + offset));
1373
1374 if (!count)
1375 goto out;
1376
1377 if (nfs_can_extend_write(file, page, inode, pagelen)) {
1378 count = max(count + offset, pagelen);
1379 offset = 0;
1380 }
1381
1382 status = nfs_writepage_setup(ctx, page, offset, count);
1383 if (status < 0)
1384 nfs_set_pageerror(mapping);
1385 else
1386 __set_page_dirty_nobuffers(page);
1387out:
1388 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1389 status, (long long)i_size_read(inode));
1390 return status;
1391}
1392
1393static int flush_task_priority(int how)
1394{
1395 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1396 case FLUSH_HIGHPRI:
1397 return RPC_PRIORITY_HIGH;
1398 case FLUSH_LOWPRI:
1399 return RPC_PRIORITY_LOW;
1400 }
1401 return RPC_PRIORITY_NORMAL;
1402}
1403
1404static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1405 struct rpc_message *msg,
1406 const struct nfs_rpc_ops *rpc_ops,
1407 struct rpc_task_setup *task_setup_data, int how)
1408{
1409 int priority = flush_task_priority(how);
1410
1411 task_setup_data->priority = priority;
1412 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1413 trace_nfs_initiate_write(hdr);
1414}
1415
1416/* If a nfs_flush_* function fails, it should remove reqs from @head and
1417 * call this on each, which will prepare them to be retried on next
1418 * writeback using standard nfs.
1419 */
1420static void nfs_redirty_request(struct nfs_page *req)
1421{
1422 /* Bump the transmission count */
1423 req->wb_nio++;
1424 nfs_mark_request_dirty(req);
1425 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1426 nfs_end_page_writeback(req);
1427 nfs_release_request(req);
1428}
1429
1430static void nfs_async_write_error(struct list_head *head, int error)
1431{
1432 struct nfs_page *req;
1433
1434 while (!list_empty(head)) {
1435 req = nfs_list_entry(head->next);
1436 nfs_list_remove_request(req);
1437 if (nfs_error_is_fatal(error))
1438 nfs_write_error(req, error);
1439 else
1440 nfs_redirty_request(req);
1441 }
1442}
1443
1444static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1445{
1446 nfs_async_write_error(&hdr->pages, 0);
1447 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1448 hdr->args.offset + hdr->args.count - 1);
1449}
1450
1451static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1452 .init_hdr = nfs_async_write_init,
1453 .error_cleanup = nfs_async_write_error,
1454 .completion = nfs_write_completion,
1455 .reschedule_io = nfs_async_write_reschedule_io,
1456};
1457
1458void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1459 struct inode *inode, int ioflags, bool force_mds,
1460 const struct nfs_pgio_completion_ops *compl_ops)
1461{
1462 struct nfs_server *server = NFS_SERVER(inode);
1463 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1464
1465#ifdef CONFIG_NFS_V4_1
1466 if (server->pnfs_curr_ld && !force_mds)
1467 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1468#endif
1469 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1470 server->wsize, ioflags);
1471}
1472EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1473
1474void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1475{
1476 struct nfs_pgio_mirror *mirror;
1477
1478 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1479 pgio->pg_ops->pg_cleanup(pgio);
1480
1481 pgio->pg_ops = &nfs_pgio_rw_ops;
1482
1483 nfs_pageio_stop_mirroring(pgio);
1484
1485 mirror = &pgio->pg_mirrors[0];
1486 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1487}
1488EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1489
1490
1491void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1492{
1493 struct nfs_commit_data *data = calldata;
1494
1495 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1496}
1497
1498/*
1499 * Special version of should_remove_suid() that ignores capabilities.
1500 */
1501static int nfs_should_remove_suid(const struct inode *inode)
1502{
1503 umode_t mode = inode->i_mode;
1504 int kill = 0;
1505
1506 /* suid always must be killed */
1507 if (unlikely(mode & S_ISUID))
1508 kill = ATTR_KILL_SUID;
1509
1510 /*
1511 * sgid without any exec bits is just a mandatory locking mark; leave
1512 * it alone. If some exec bits are set, it's a real sgid; kill it.
1513 */
1514 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1515 kill |= ATTR_KILL_SGID;
1516
1517 if (unlikely(kill && S_ISREG(mode)))
1518 return kill;
1519
1520 return 0;
1521}
1522
1523static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1524 struct nfs_fattr *fattr)
1525{
1526 struct nfs_pgio_args *argp = &hdr->args;
1527 struct nfs_pgio_res *resp = &hdr->res;
1528 u64 size = argp->offset + resp->count;
1529
1530 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1531 fattr->size = size;
1532 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1533 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1534 return;
1535 }
1536 if (size != fattr->size)
1537 return;
1538 /* Set attribute barrier */
1539 nfs_fattr_set_barrier(fattr);
1540 /* ...and update size */
1541 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1542}
1543
1544void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1545{
1546 struct nfs_fattr *fattr = &hdr->fattr;
1547 struct inode *inode = hdr->inode;
1548
1549 spin_lock(&inode->i_lock);
1550 nfs_writeback_check_extend(hdr, fattr);
1551 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1552 spin_unlock(&inode->i_lock);
1553}
1554EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1555
1556/*
1557 * This function is called when the WRITE call is complete.
1558 */
1559static int nfs_writeback_done(struct rpc_task *task,
1560 struct nfs_pgio_header *hdr,
1561 struct inode *inode)
1562{
1563 int status;
1564
1565 /*
1566 * ->write_done will attempt to use post-op attributes to detect
1567 * conflicting writes by other clients. A strict interpretation
1568 * of close-to-open would allow us to continue caching even if
1569 * another writer had changed the file, but some applications
1570 * depend on tighter cache coherency when writing.
1571 */
1572 status = NFS_PROTO(inode)->write_done(task, hdr);
1573 if (status != 0)
1574 return status;
1575
1576 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1577 trace_nfs_writeback_done(task, hdr);
1578
1579 if (hdr->res.verf->committed < hdr->args.stable &&
1580 task->tk_status >= 0) {
1581 /* We tried a write call, but the server did not
1582 * commit data to stable storage even though we
1583 * requested it.
1584 * Note: There is a known bug in Tru64 < 5.0 in which
1585 * the server reports NFS_DATA_SYNC, but performs
1586 * NFS_FILE_SYNC. We therefore implement this checking
1587 * as a dprintk() in order to avoid filling syslog.
1588 */
1589 static unsigned long complain;
1590
1591 /* Note this will print the MDS for a DS write */
1592 if (time_before(complain, jiffies)) {
1593 dprintk("NFS: faulty NFS server %s:"
1594 " (committed = %d) != (stable = %d)\n",
1595 NFS_SERVER(inode)->nfs_client->cl_hostname,
1596 hdr->res.verf->committed, hdr->args.stable);
1597 complain = jiffies + 300 * HZ;
1598 }
1599 }
1600
1601 /* Deal with the suid/sgid bit corner case */
1602 if (nfs_should_remove_suid(inode)) {
1603 spin_lock(&inode->i_lock);
1604 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1605 spin_unlock(&inode->i_lock);
1606 }
1607 return 0;
1608}
1609
1610/*
1611 * This function is called when the WRITE call is complete.
1612 */
1613static void nfs_writeback_result(struct rpc_task *task,
1614 struct nfs_pgio_header *hdr)
1615{
1616 struct nfs_pgio_args *argp = &hdr->args;
1617 struct nfs_pgio_res *resp = &hdr->res;
1618
1619 if (resp->count < argp->count) {
1620 static unsigned long complain;
1621
1622 /* This a short write! */
1623 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1624
1625 /* Has the server at least made some progress? */
1626 if (resp->count == 0) {
1627 if (time_before(complain, jiffies)) {
1628 printk(KERN_WARNING
1629 "NFS: Server wrote zero bytes, expected %u.\n",
1630 argp->count);
1631 complain = jiffies + 300 * HZ;
1632 }
1633 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1634 task->tk_status = -EIO;
1635 return;
1636 }
1637
1638 /* For non rpc-based layout drivers, retry-through-MDS */
1639 if (!task->tk_ops) {
1640 hdr->pnfs_error = -EAGAIN;
1641 return;
1642 }
1643
1644 /* Was this an NFSv2 write or an NFSv3 stable write? */
1645 if (resp->verf->committed != NFS_UNSTABLE) {
1646 /* Resend from where the server left off */
1647 hdr->mds_offset += resp->count;
1648 argp->offset += resp->count;
1649 argp->pgbase += resp->count;
1650 argp->count -= resp->count;
1651 } else {
1652 /* Resend as a stable write in order to avoid
1653 * headaches in the case of a server crash.
1654 */
1655 argp->stable = NFS_FILE_SYNC;
1656 }
1657 resp->count = 0;
1658 resp->verf->committed = 0;
1659 rpc_restart_call_prepare(task);
1660 }
1661}
1662
1663static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1664{
1665 return wait_var_event_killable(&cinfo->rpcs_out,
1666 !atomic_read(&cinfo->rpcs_out));
1667}
1668
1669static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1670{
1671 atomic_inc(&cinfo->rpcs_out);
1672}
1673
1674static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1675{
1676 if (atomic_dec_and_test(&cinfo->rpcs_out))
1677 wake_up_var(&cinfo->rpcs_out);
1678}
1679
1680void nfs_commitdata_release(struct nfs_commit_data *data)
1681{
1682 put_nfs_open_context(data->context);
1683 nfs_commit_free(data);
1684}
1685EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1686
1687int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1688 const struct nfs_rpc_ops *nfs_ops,
1689 const struct rpc_call_ops *call_ops,
1690 int how, int flags)
1691{
1692 struct rpc_task *task;
1693 int priority = flush_task_priority(how);
1694 struct rpc_message msg = {
1695 .rpc_argp = &data->args,
1696 .rpc_resp = &data->res,
1697 .rpc_cred = data->cred,
1698 };
1699 struct rpc_task_setup task_setup_data = {
1700 .task = &data->task,
1701 .rpc_client = clnt,
1702 .rpc_message = &msg,
1703 .callback_ops = call_ops,
1704 .callback_data = data,
1705 .workqueue = nfsiod_workqueue,
1706 .flags = RPC_TASK_ASYNC | flags,
1707 .priority = priority,
1708 };
1709 /* Set up the initial task struct. */
1710 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1711 trace_nfs_initiate_commit(data);
1712
1713 dprintk("NFS: initiated commit call\n");
1714
1715 task = rpc_run_task(&task_setup_data);
1716 if (IS_ERR(task))
1717 return PTR_ERR(task);
1718 if (how & FLUSH_SYNC)
1719 rpc_wait_for_completion_task(task);
1720 rpc_put_task(task);
1721 return 0;
1722}
1723EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1724
1725static loff_t nfs_get_lwb(struct list_head *head)
1726{
1727 loff_t lwb = 0;
1728 struct nfs_page *req;
1729
1730 list_for_each_entry(req, head, wb_list)
1731 if (lwb < (req_offset(req) + req->wb_bytes))
1732 lwb = req_offset(req) + req->wb_bytes;
1733
1734 return lwb;
1735}
1736
1737/*
1738 * Set up the argument/result storage required for the RPC call.
1739 */
1740void nfs_init_commit(struct nfs_commit_data *data,
1741 struct list_head *head,
1742 struct pnfs_layout_segment *lseg,
1743 struct nfs_commit_info *cinfo)
1744{
1745 struct nfs_page *first;
1746 struct nfs_open_context *ctx;
1747 struct inode *inode;
1748
1749 /* Set up the RPC argument and reply structs
1750 * NB: take care not to mess about with data->commit et al. */
1751
1752 if (head)
1753 list_splice_init(head, &data->pages);
1754
1755 first = nfs_list_entry(data->pages.next);
1756 ctx = nfs_req_openctx(first);
1757 inode = d_inode(ctx->dentry);
1758
1759 data->inode = inode;
1760 data->cred = ctx->cred;
1761 data->lseg = lseg; /* reference transferred */
1762 /* only set lwb for pnfs commit */
1763 if (lseg)
1764 data->lwb = nfs_get_lwb(&data->pages);
1765 data->mds_ops = &nfs_commit_ops;
1766 data->completion_ops = cinfo->completion_ops;
1767 data->dreq = cinfo->dreq;
1768
1769 data->args.fh = NFS_FH(data->inode);
1770 /* Note: we always request a commit of the entire inode */
1771 data->args.offset = 0;
1772 data->args.count = 0;
1773 data->context = get_nfs_open_context(ctx);
1774 data->res.fattr = &data->fattr;
1775 data->res.verf = &data->verf;
1776 nfs_fattr_init(&data->fattr);
1777}
1778EXPORT_SYMBOL_GPL(nfs_init_commit);
1779
1780void nfs_retry_commit(struct list_head *page_list,
1781 struct pnfs_layout_segment *lseg,
1782 struct nfs_commit_info *cinfo,
1783 u32 ds_commit_idx)
1784{
1785 struct nfs_page *req;
1786
1787 while (!list_empty(page_list)) {
1788 req = nfs_list_entry(page_list->next);
1789 nfs_list_remove_request(req);
1790 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1791 if (!cinfo->dreq)
1792 nfs_clear_page_commit(req->wb_page);
1793 nfs_unlock_and_release_request(req);
1794 }
1795}
1796EXPORT_SYMBOL_GPL(nfs_retry_commit);
1797
1798static void
1799nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1800 struct nfs_page *req)
1801{
1802 __set_page_dirty_nobuffers(req->wb_page);
1803}
1804
1805/*
1806 * Commit dirty pages
1807 */
1808static int
1809nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1810 struct nfs_commit_info *cinfo)
1811{
1812 struct nfs_commit_data *data;
1813 unsigned short task_flags = 0;
1814
1815 /* another commit raced with us */
1816 if (list_empty(head))
1817 return 0;
1818
1819 data = nfs_commitdata_alloc(true);
1820
1821 /* Set up the argument struct */
1822 nfs_init_commit(data, head, NULL, cinfo);
1823 atomic_inc(&cinfo->mds->rpcs_out);
1824 if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1825 task_flags = RPC_TASK_MOVEABLE;
1826 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1827 data->mds_ops, how,
1828 RPC_TASK_CRED_NOREF | task_flags);
1829}
1830
1831/*
1832 * COMMIT call returned
1833 */
1834static void nfs_commit_done(struct rpc_task *task, void *calldata)
1835{
1836 struct nfs_commit_data *data = calldata;
1837
1838 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1839 task->tk_pid, task->tk_status);
1840
1841 /* Call the NFS version-specific code */
1842 NFS_PROTO(data->inode)->commit_done(task, data);
1843 trace_nfs_commit_done(task, data);
1844}
1845
1846static void nfs_commit_release_pages(struct nfs_commit_data *data)
1847{
1848 const struct nfs_writeverf *verf = data->res.verf;
1849 struct nfs_page *req;
1850 int status = data->task.tk_status;
1851 struct nfs_commit_info cinfo;
1852 struct nfs_server *nfss;
1853
1854 while (!list_empty(&data->pages)) {
1855 req = nfs_list_entry(data->pages.next);
1856 nfs_list_remove_request(req);
1857 if (req->wb_page)
1858 nfs_clear_page_commit(req->wb_page);
1859
1860 dprintk("NFS: commit (%s/%llu %d@%lld)",
1861 nfs_req_openctx(req)->dentry->d_sb->s_id,
1862 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1863 req->wb_bytes,
1864 (long long)req_offset(req));
1865 if (status < 0) {
1866 if (req->wb_page) {
1867 trace_nfs_commit_error(req, status);
1868 nfs_mapping_set_error(req->wb_page, status);
1869 nfs_inode_remove_request(req);
1870 }
1871 dprintk_cont(", error = %d\n", status);
1872 goto next;
1873 }
1874
1875 /* Okay, COMMIT succeeded, apparently. Check the verifier
1876 * returned by the server against all stored verfs. */
1877 if (nfs_write_match_verf(verf, req)) {
1878 /* We have a match */
1879 if (req->wb_page)
1880 nfs_inode_remove_request(req);
1881 dprintk_cont(" OK\n");
1882 goto next;
1883 }
1884 /* We have a mismatch. Write the page again */
1885 dprintk_cont(" mismatch\n");
1886 nfs_mark_request_dirty(req);
1887 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1888 next:
1889 nfs_unlock_and_release_request(req);
1890 /* Latency breaker */
1891 cond_resched();
1892 }
1893 nfss = NFS_SERVER(data->inode);
1894 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1895 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1896
1897 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1898 nfs_commit_end(cinfo.mds);
1899}
1900
1901static void nfs_commit_release(void *calldata)
1902{
1903 struct nfs_commit_data *data = calldata;
1904
1905 data->completion_ops->completion(data);
1906 nfs_commitdata_release(calldata);
1907}
1908
1909static const struct rpc_call_ops nfs_commit_ops = {
1910 .rpc_call_prepare = nfs_commit_prepare,
1911 .rpc_call_done = nfs_commit_done,
1912 .rpc_release = nfs_commit_release,
1913};
1914
1915static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1916 .completion = nfs_commit_release_pages,
1917 .resched_write = nfs_commit_resched_write,
1918};
1919
1920int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1921 int how, struct nfs_commit_info *cinfo)
1922{
1923 int status;
1924
1925 status = pnfs_commit_list(inode, head, how, cinfo);
1926 if (status == PNFS_NOT_ATTEMPTED)
1927 status = nfs_commit_list(inode, head, how, cinfo);
1928 return status;
1929}
1930
1931static int __nfs_commit_inode(struct inode *inode, int how,
1932 struct writeback_control *wbc)
1933{
1934 LIST_HEAD(head);
1935 struct nfs_commit_info cinfo;
1936 int may_wait = how & FLUSH_SYNC;
1937 int ret, nscan;
1938
1939 nfs_init_cinfo_from_inode(&cinfo, inode);
1940 nfs_commit_begin(cinfo.mds);
1941 for (;;) {
1942 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1943 if (ret <= 0)
1944 break;
1945 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1946 if (ret < 0)
1947 break;
1948 ret = 0;
1949 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1950 if (nscan < wbc->nr_to_write)
1951 wbc->nr_to_write -= nscan;
1952 else
1953 wbc->nr_to_write = 0;
1954 }
1955 if (nscan < INT_MAX)
1956 break;
1957 cond_resched();
1958 }
1959 nfs_commit_end(cinfo.mds);
1960 if (ret || !may_wait)
1961 return ret;
1962 return wait_on_commit(cinfo.mds);
1963}
1964
1965int nfs_commit_inode(struct inode *inode, int how)
1966{
1967 return __nfs_commit_inode(inode, how, NULL);
1968}
1969EXPORT_SYMBOL_GPL(nfs_commit_inode);
1970
1971int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1972{
1973 struct nfs_inode *nfsi = NFS_I(inode);
1974 int flags = FLUSH_SYNC;
1975 int ret = 0;
1976
1977 if (wbc->sync_mode == WB_SYNC_NONE) {
1978 /* no commits means nothing needs to be done */
1979 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1980 goto check_requests_outstanding;
1981
1982 /* Don't commit yet if this is a non-blocking flush and there
1983 * are a lot of outstanding writes for this mapping.
1984 */
1985 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1986 goto out_mark_dirty;
1987
1988 /* don't wait for the COMMIT response */
1989 flags = 0;
1990 }
1991
1992 ret = __nfs_commit_inode(inode, flags, wbc);
1993 if (!ret) {
1994 if (flags & FLUSH_SYNC)
1995 return 0;
1996 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1997 goto out_mark_dirty;
1998
1999check_requests_outstanding:
2000 if (!atomic_read(&nfsi->commit_info.rpcs_out))
2001 return ret;
2002out_mark_dirty:
2003 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2004 return ret;
2005}
2006EXPORT_SYMBOL_GPL(nfs_write_inode);
2007
2008/*
2009 * Wrapper for filemap_write_and_wait_range()
2010 *
2011 * Needed for pNFS in order to ensure data becomes visible to the
2012 * client.
2013 */
2014int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2015 loff_t lstart, loff_t lend)
2016{
2017 int ret;
2018
2019 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2020 if (ret == 0)
2021 ret = pnfs_sync_inode(mapping->host, true);
2022 return ret;
2023}
2024EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2025
2026/*
2027 * flush the inode to disk.
2028 */
2029int nfs_wb_all(struct inode *inode)
2030{
2031 int ret;
2032
2033 trace_nfs_writeback_inode_enter(inode);
2034
2035 ret = filemap_write_and_wait(inode->i_mapping);
2036 if (ret)
2037 goto out;
2038 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2039 if (ret < 0)
2040 goto out;
2041 pnfs_sync_inode(inode, true);
2042 ret = 0;
2043
2044out:
2045 trace_nfs_writeback_inode_exit(inode, ret);
2046 return ret;
2047}
2048EXPORT_SYMBOL_GPL(nfs_wb_all);
2049
2050int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2051{
2052 struct nfs_page *req;
2053 int ret = 0;
2054
2055 wait_on_page_writeback(page);
2056
2057 /* blocking call to cancel all requests and join to a single (head)
2058 * request */
2059 req = nfs_lock_and_join_requests(page);
2060
2061 if (IS_ERR(req)) {
2062 ret = PTR_ERR(req);
2063 } else if (req) {
2064 /* all requests from this page have been cancelled by
2065 * nfs_lock_and_join_requests, so just remove the head
2066 * request from the inode / page_private pointer and
2067 * release it */
2068 nfs_inode_remove_request(req);
2069 nfs_unlock_and_release_request(req);
2070 }
2071
2072 return ret;
2073}
2074
2075/*
2076 * Write back all requests on one page - we do this before reading it.
2077 */
2078int nfs_wb_page(struct inode *inode, struct page *page)
2079{
2080 loff_t range_start = page_file_offset(page);
2081 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2082 struct writeback_control wbc = {
2083 .sync_mode = WB_SYNC_ALL,
2084 .nr_to_write = 0,
2085 .range_start = range_start,
2086 .range_end = range_end,
2087 };
2088 int ret;
2089
2090 trace_nfs_writeback_page_enter(inode);
2091
2092 for (;;) {
2093 wait_on_page_writeback(page);
2094 if (clear_page_dirty_for_io(page)) {
2095 ret = nfs_writepage_locked(page, &wbc);
2096 if (ret < 0)
2097 goto out_error;
2098 continue;
2099 }
2100 ret = 0;
2101 if (!PagePrivate(page))
2102 break;
2103 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2104 if (ret < 0)
2105 goto out_error;
2106 }
2107out_error:
2108 trace_nfs_writeback_page_exit(inode, ret);
2109 return ret;
2110}
2111
2112#ifdef CONFIG_MIGRATION
2113int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2114 struct page *page, enum migrate_mode mode)
2115{
2116 /*
2117 * If PagePrivate is set, then the page is currently associated with
2118 * an in-progress read or write request. Don't try to migrate it.
2119 *
2120 * FIXME: we could do this in principle, but we'll need a way to ensure
2121 * that we can safely release the inode reference while holding
2122 * the page lock.
2123 */
2124 if (PagePrivate(page))
2125 return -EBUSY;
2126
2127 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2128 return -EBUSY;
2129
2130 return migrate_page(mapping, newpage, page, mode);
2131}
2132#endif
2133
2134int __init nfs_init_writepagecache(void)
2135{
2136 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2137 sizeof(struct nfs_pgio_header),
2138 0, SLAB_HWCACHE_ALIGN,
2139 NULL);
2140 if (nfs_wdata_cachep == NULL)
2141 return -ENOMEM;
2142
2143 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2144 nfs_wdata_cachep);
2145 if (nfs_wdata_mempool == NULL)
2146 goto out_destroy_write_cache;
2147
2148 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2149 sizeof(struct nfs_commit_data),
2150 0, SLAB_HWCACHE_ALIGN,
2151 NULL);
2152 if (nfs_cdata_cachep == NULL)
2153 goto out_destroy_write_mempool;
2154
2155 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2156 nfs_cdata_cachep);
2157 if (nfs_commit_mempool == NULL)
2158 goto out_destroy_commit_cache;
2159
2160 /*
2161 * NFS congestion size, scale with available memory.
2162 *
2163 * 64MB: 8192k
2164 * 128MB: 11585k
2165 * 256MB: 16384k
2166 * 512MB: 23170k
2167 * 1GB: 32768k
2168 * 2GB: 46340k
2169 * 4GB: 65536k
2170 * 8GB: 92681k
2171 * 16GB: 131072k
2172 *
2173 * This allows larger machines to have larger/more transfers.
2174 * Limit the default to 256M
2175 */
2176 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2177 if (nfs_congestion_kb > 256*1024)
2178 nfs_congestion_kb = 256*1024;
2179
2180 return 0;
2181
2182out_destroy_commit_cache:
2183 kmem_cache_destroy(nfs_cdata_cachep);
2184out_destroy_write_mempool:
2185 mempool_destroy(nfs_wdata_mempool);
2186out_destroy_write_cache:
2187 kmem_cache_destroy(nfs_wdata_cachep);
2188 return -ENOMEM;
2189}
2190
2191void nfs_destroy_writepagecache(void)
2192{
2193 mempool_destroy(nfs_commit_mempool);
2194 kmem_cache_destroy(nfs_cdata_cachep);
2195 mempool_destroy(nfs_wdata_mempool);
2196 kmem_cache_destroy(nfs_wdata_cachep);
2197}
2198
2199static const struct nfs_rw_ops nfs_rw_write_ops = {
2200 .rw_alloc_header = nfs_writehdr_alloc,
2201 .rw_free_header = nfs_writehdr_free,
2202 .rw_done = nfs_writeback_done,
2203 .rw_result = nfs_writeback_result,
2204 .rw_initiate = nfs_initiate_write,
2205};