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/*
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#include <linux/freezer.h>
25#include <linux/wait.h>
26
27#include <asm/uaccess.h>
28
29#include "delegation.h"
30#include "internal.h"
31#include "iostat.h"
32#include "nfs4_fs.h"
33#include "fscache.h"
34#include "pnfs.h"
35
36#include "nfstrace.h"
37
38#define NFSDBG_FACILITY NFSDBG_PAGECACHE
39
40#define MIN_POOL_WRITE (32)
41#define MIN_POOL_COMMIT (4)
42
43/*
44 * Local function declarations
45 */
46static void nfs_redirty_request(struct nfs_page *req);
47static const struct rpc_call_ops nfs_commit_ops;
48static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
49static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
50static const struct nfs_rw_ops nfs_rw_write_ops;
51static void nfs_clear_request_commit(struct nfs_page *req);
52static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
53 struct inode *inode);
54static struct nfs_page *
55nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
56 struct page *page);
57
58static struct kmem_cache *nfs_wdata_cachep;
59static mempool_t *nfs_wdata_mempool;
60static struct kmem_cache *nfs_cdata_cachep;
61static mempool_t *nfs_commit_mempool;
62
63struct nfs_commit_data *nfs_commitdata_alloc(void)
64{
65 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
66
67 if (p) {
68 memset(p, 0, sizeof(*p));
69 INIT_LIST_HEAD(&p->pages);
70 }
71 return p;
72}
73EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
74
75void nfs_commit_free(struct nfs_commit_data *p)
76{
77 mempool_free(p, nfs_commit_mempool);
78}
79EXPORT_SYMBOL_GPL(nfs_commit_free);
80
81static struct nfs_pgio_header *nfs_writehdr_alloc(void)
82{
83 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
84
85 if (p)
86 memset(p, 0, sizeof(*p));
87 return p;
88}
89
90static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
91{
92 mempool_free(hdr, nfs_wdata_mempool);
93}
94
95static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
96{
97 ctx->error = error;
98 smp_wmb();
99 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
100}
101
102/*
103 * nfs_page_find_head_request_locked - find head request associated with @page
104 *
105 * must be called while holding the inode lock.
106 *
107 * returns matching head request with reference held, or NULL if not found.
108 */
109static struct nfs_page *
110nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page)
111{
112 struct nfs_page *req = NULL;
113
114 if (PagePrivate(page))
115 req = (struct nfs_page *)page_private(page);
116 else if (unlikely(PageSwapCache(page)))
117 req = nfs_page_search_commits_for_head_request_locked(nfsi,
118 page);
119
120 if (req) {
121 WARN_ON_ONCE(req->wb_head != req);
122 kref_get(&req->wb_kref);
123 }
124
125 return req;
126}
127
128/*
129 * nfs_page_find_head_request - find head request associated with @page
130 *
131 * returns matching head request with reference held, or NULL if not found.
132 */
133static struct nfs_page *nfs_page_find_head_request(struct page *page)
134{
135 struct inode *inode = page_file_mapping(page)->host;
136 struct nfs_page *req = NULL;
137
138 spin_lock(&inode->i_lock);
139 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
140 spin_unlock(&inode->i_lock);
141 return req;
142}
143
144/* Adjust the file length if we're writing beyond the end */
145static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
146{
147 struct inode *inode = page_file_mapping(page)->host;
148 loff_t end, i_size;
149 pgoff_t end_index;
150
151 spin_lock(&inode->i_lock);
152 i_size = i_size_read(inode);
153 end_index = (i_size - 1) >> PAGE_SHIFT;
154 if (i_size > 0 && page_file_index(page) < end_index)
155 goto out;
156 end = page_file_offset(page) + ((loff_t)offset+count);
157 if (i_size >= end)
158 goto out;
159 i_size_write(inode, end);
160 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
161out:
162 spin_unlock(&inode->i_lock);
163}
164
165/* A writeback failed: mark the page as bad, and invalidate the page cache */
166static void nfs_set_pageerror(struct page *page)
167{
168 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
169}
170
171/*
172 * nfs_page_group_search_locked
173 * @head - head request of page group
174 * @page_offset - offset into page
175 *
176 * Search page group with head @head to find a request that contains the
177 * page offset @page_offset.
178 *
179 * Returns a pointer to the first matching nfs request, or NULL if no
180 * match is found.
181 *
182 * Must be called with the page group lock held
183 */
184static struct nfs_page *
185nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
186{
187 struct nfs_page *req;
188
189 WARN_ON_ONCE(head != head->wb_head);
190 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
191
192 req = head;
193 do {
194 if (page_offset >= req->wb_pgbase &&
195 page_offset < (req->wb_pgbase + req->wb_bytes))
196 return req;
197
198 req = req->wb_this_page;
199 } while (req != head);
200
201 return NULL;
202}
203
204/*
205 * nfs_page_group_covers_page
206 * @head - head request of page group
207 *
208 * Return true if the page group with head @head covers the whole page,
209 * returns false otherwise
210 */
211static bool nfs_page_group_covers_page(struct nfs_page *req)
212{
213 struct nfs_page *tmp;
214 unsigned int pos = 0;
215 unsigned int len = nfs_page_length(req->wb_page);
216
217 nfs_page_group_lock(req, false);
218
219 do {
220 tmp = nfs_page_group_search_locked(req->wb_head, pos);
221 if (tmp) {
222 /* no way this should happen */
223 WARN_ON_ONCE(tmp->wb_pgbase != pos);
224 pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
225 }
226 } while (tmp && pos < len);
227
228 nfs_page_group_unlock(req);
229 WARN_ON_ONCE(pos > len);
230 return pos == len;
231}
232
233/* We can set the PG_uptodate flag if we see that a write request
234 * covers the full page.
235 */
236static void nfs_mark_uptodate(struct nfs_page *req)
237{
238 if (PageUptodate(req->wb_page))
239 return;
240 if (!nfs_page_group_covers_page(req))
241 return;
242 SetPageUptodate(req->wb_page);
243}
244
245static int wb_priority(struct writeback_control *wbc)
246{
247 int ret = 0;
248 if (wbc->for_reclaim)
249 return FLUSH_HIGHPRI | FLUSH_COND_STABLE;
250 if (wbc->sync_mode == WB_SYNC_ALL)
251 ret = FLUSH_COND_STABLE;
252 return ret;
253}
254
255/*
256 * NFS congestion control
257 */
258
259int nfs_congestion_kb;
260
261#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
262#define NFS_CONGESTION_OFF_THRESH \
263 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
264
265static void nfs_set_page_writeback(struct page *page)
266{
267 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
268 int ret = test_set_page_writeback(page);
269
270 WARN_ON_ONCE(ret != 0);
271
272 if (atomic_long_inc_return(&nfss->writeback) >
273 NFS_CONGESTION_ON_THRESH) {
274 set_bdi_congested(&nfss->backing_dev_info,
275 BLK_RW_ASYNC);
276 }
277}
278
279static void nfs_end_page_writeback(struct nfs_page *req)
280{
281 struct inode *inode = page_file_mapping(req->wb_page)->host;
282 struct nfs_server *nfss = NFS_SERVER(inode);
283
284 if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
285 return;
286
287 end_page_writeback(req->wb_page);
288 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
289 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
290}
291
292
293/* nfs_page_group_clear_bits
294 * @req - an nfs request
295 * clears all page group related bits from @req
296 */
297static void
298nfs_page_group_clear_bits(struct nfs_page *req)
299{
300 clear_bit(PG_TEARDOWN, &req->wb_flags);
301 clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
302 clear_bit(PG_UPTODATE, &req->wb_flags);
303 clear_bit(PG_WB_END, &req->wb_flags);
304 clear_bit(PG_REMOVE, &req->wb_flags);
305}
306
307
308/*
309 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
310 *
311 * this is a helper function for nfs_lock_and_join_requests
312 *
313 * @inode - inode associated with request page group, must be holding inode lock
314 * @head - head request of page group, must be holding head lock
315 * @req - request that couldn't lock and needs to wait on the req bit lock
316 * @nonblock - if true, don't actually wait
317 *
318 * NOTE: this must be called holding page_group bit lock and inode spin lock
319 * and BOTH will be released before returning.
320 *
321 * returns 0 on success, < 0 on error.
322 */
323static int
324nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
325 struct nfs_page *req, bool nonblock)
326 __releases(&inode->i_lock)
327{
328 struct nfs_page *tmp;
329 int ret;
330
331 /* relinquish all the locks successfully grabbed this run */
332 for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
333 nfs_unlock_request(tmp);
334
335 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
336
337 /* grab a ref on the request that will be waited on */
338 kref_get(&req->wb_kref);
339
340 nfs_page_group_unlock(head);
341 spin_unlock(&inode->i_lock);
342
343 /* release ref from nfs_page_find_head_request_locked */
344 nfs_release_request(head);
345
346 if (!nonblock)
347 ret = nfs_wait_on_request(req);
348 else
349 ret = -EAGAIN;
350 nfs_release_request(req);
351
352 return ret;
353}
354
355/*
356 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
357 *
358 * @destroy_list - request list (using wb_this_page) terminated by @old_head
359 * @old_head - the old head of the list
360 *
361 * All subrequests must be locked and removed from all lists, so at this point
362 * they are only "active" in this function, and possibly in nfs_wait_on_request
363 * with a reference held by some other context.
364 */
365static void
366nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
367 struct nfs_page *old_head)
368{
369 while (destroy_list) {
370 struct nfs_page *subreq = destroy_list;
371
372 destroy_list = (subreq->wb_this_page == old_head) ?
373 NULL : subreq->wb_this_page;
374
375 WARN_ON_ONCE(old_head != subreq->wb_head);
376
377 /* make sure old group is not used */
378 subreq->wb_head = subreq;
379 subreq->wb_this_page = subreq;
380
381 /* subreq is now totally disconnected from page group or any
382 * write / commit lists. last chance to wake any waiters */
383 nfs_unlock_request(subreq);
384
385 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
386 /* release ref on old head request */
387 nfs_release_request(old_head);
388
389 nfs_page_group_clear_bits(subreq);
390
391 /* release the PG_INODE_REF reference */
392 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
393 nfs_release_request(subreq);
394 else
395 WARN_ON_ONCE(1);
396 } else {
397 WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
398 /* zombie requests have already released the last
399 * reference and were waiting on the rest of the
400 * group to complete. Since it's no longer part of a
401 * group, simply free the request */
402 nfs_page_group_clear_bits(subreq);
403 nfs_free_request(subreq);
404 }
405 }
406}
407
408/*
409 * nfs_lock_and_join_requests - join all subreqs to the head req and return
410 * a locked reference, cancelling any pending
411 * operations for this page.
412 *
413 * @page - the page used to lookup the "page group" of nfs_page structures
414 * @nonblock - if true, don't block waiting for request locks
415 *
416 * This function joins all sub requests to the head request by first
417 * locking all requests in the group, cancelling any pending operations
418 * and finally updating the head request to cover the whole range covered by
419 * the (former) group. All subrequests are removed from any write or commit
420 * lists, unlinked from the group and destroyed.
421 *
422 * Returns a locked, referenced pointer to the head request - which after
423 * this call is guaranteed to be the only request associated with the page.
424 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
425 * error was encountered.
426 */
427static struct nfs_page *
428nfs_lock_and_join_requests(struct page *page, bool nonblock)
429{
430 struct inode *inode = page_file_mapping(page)->host;
431 struct nfs_page *head, *subreq;
432 struct nfs_page *destroy_list = NULL;
433 unsigned int total_bytes;
434 int ret;
435
436try_again:
437 total_bytes = 0;
438
439 WARN_ON_ONCE(destroy_list);
440
441 spin_lock(&inode->i_lock);
442
443 /*
444 * A reference is taken only on the head request which acts as a
445 * reference to the whole page group - the group will not be destroyed
446 * until the head reference is released.
447 */
448 head = nfs_page_find_head_request_locked(NFS_I(inode), page);
449
450 if (!head) {
451 spin_unlock(&inode->i_lock);
452 return NULL;
453 }
454
455 /* holding inode lock, so always make a non-blocking call to try the
456 * page group lock */
457 ret = nfs_page_group_lock(head, true);
458 if (ret < 0) {
459 spin_unlock(&inode->i_lock);
460
461 if (!nonblock && ret == -EAGAIN) {
462 nfs_page_group_lock_wait(head);
463 nfs_release_request(head);
464 goto try_again;
465 }
466
467 nfs_release_request(head);
468 return ERR_PTR(ret);
469 }
470
471 /* lock each request in the page group */
472 subreq = head;
473 do {
474 /*
475 * Subrequests are always contiguous, non overlapping
476 * and in order - but may be repeated (mirrored writes).
477 */
478 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
479 /* keep track of how many bytes this group covers */
480 total_bytes += subreq->wb_bytes;
481 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
482 ((subreq->wb_offset + subreq->wb_bytes) >
483 (head->wb_offset + total_bytes)))) {
484 nfs_page_group_unlock(head);
485 spin_unlock(&inode->i_lock);
486 return ERR_PTR(-EIO);
487 }
488
489 if (!nfs_lock_request(subreq)) {
490 /* releases page group bit lock and
491 * inode spin lock and all references */
492 ret = nfs_unroll_locks_and_wait(inode, head,
493 subreq, nonblock);
494
495 if (ret == 0)
496 goto try_again;
497
498 return ERR_PTR(ret);
499 }
500
501 subreq = subreq->wb_this_page;
502 } while (subreq != head);
503
504 /* Now that all requests are locked, make sure they aren't on any list.
505 * Commit list removal accounting is done after locks are dropped */
506 subreq = head;
507 do {
508 nfs_clear_request_commit(subreq);
509 subreq = subreq->wb_this_page;
510 } while (subreq != head);
511
512 /* unlink subrequests from head, destroy them later */
513 if (head->wb_this_page != head) {
514 /* destroy list will be terminated by head */
515 destroy_list = head->wb_this_page;
516 head->wb_this_page = head;
517
518 /* change head request to cover whole range that
519 * the former page group covered */
520 head->wb_bytes = total_bytes;
521 }
522
523 /*
524 * prepare head request to be added to new pgio descriptor
525 */
526 nfs_page_group_clear_bits(head);
527
528 /*
529 * some part of the group was still on the inode list - otherwise
530 * the group wouldn't be involved in async write.
531 * grab a reference for the head request, iff it needs one.
532 */
533 if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
534 kref_get(&head->wb_kref);
535
536 nfs_page_group_unlock(head);
537
538 /* drop lock to clean uprequests on destroy list */
539 spin_unlock(&inode->i_lock);
540
541 nfs_destroy_unlinked_subrequests(destroy_list, head);
542
543 /* still holds ref on head from nfs_page_find_head_request_locked
544 * and still has lock on head from lock loop */
545 return head;
546}
547
548static void nfs_write_error_remove_page(struct nfs_page *req)
549{
550 nfs_unlock_request(req);
551 nfs_end_page_writeback(req);
552 nfs_release_request(req);
553 generic_error_remove_page(page_file_mapping(req->wb_page),
554 req->wb_page);
555}
556
557/*
558 * Find an associated nfs write request, and prepare to flush it out
559 * May return an error if the user signalled nfs_wait_on_request().
560 */
561static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
562 struct page *page, bool nonblock,
563 bool launder)
564{
565 struct nfs_page *req;
566 int ret = 0;
567
568 req = nfs_lock_and_join_requests(page, nonblock);
569 if (!req)
570 goto out;
571 ret = PTR_ERR(req);
572 if (IS_ERR(req))
573 goto out;
574
575 nfs_set_page_writeback(page);
576 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
577
578 ret = 0;
579 if (!nfs_pageio_add_request(pgio, req)) {
580 ret = pgio->pg_error;
581 /*
582 * Remove the problematic req upon fatal errors
583 * in launder case, while other dirty pages can
584 * still be around until they get flushed.
585 */
586 if (nfs_error_is_fatal(ret)) {
587 nfs_context_set_write_error(req->wb_context, ret);
588 if (launder) {
589 nfs_write_error_remove_page(req);
590 goto out;
591 }
592 }
593 nfs_redirty_request(req);
594 ret = -EAGAIN;
595 } else
596 nfs_add_stats(page_file_mapping(page)->host,
597 NFSIOS_WRITEPAGES, 1);
598out:
599 return ret;
600}
601
602static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
603 struct nfs_pageio_descriptor *pgio, bool launder)
604{
605 int ret;
606
607 nfs_pageio_cond_complete(pgio, page_file_index(page));
608 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE,
609 launder);
610 if (ret == -EAGAIN) {
611 redirty_page_for_writepage(wbc, page);
612 ret = 0;
613 }
614 return ret;
615}
616
617/*
618 * Write an mmapped page to the server.
619 */
620static int nfs_writepage_locked(struct page *page,
621 struct writeback_control *wbc,
622 bool launder)
623{
624 struct nfs_pageio_descriptor pgio;
625 struct inode *inode = page_file_mapping(page)->host;
626 int err;
627
628 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
629 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc),
630 false, &nfs_async_write_completion_ops);
631 err = nfs_do_writepage(page, wbc, &pgio, launder);
632 nfs_pageio_complete(&pgio);
633 if (err < 0)
634 return err;
635 if (pgio.pg_error < 0)
636 return pgio.pg_error;
637 return 0;
638}
639
640int nfs_writepage(struct page *page, struct writeback_control *wbc)
641{
642 int ret;
643
644 ret = nfs_writepage_locked(page, wbc, false);
645 unlock_page(page);
646 return ret;
647}
648
649static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
650{
651 int ret;
652
653 ret = nfs_do_writepage(page, wbc, data, false);
654 unlock_page(page);
655 return ret;
656}
657
658int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
659{
660 struct inode *inode = mapping->host;
661 unsigned long *bitlock = &NFS_I(inode)->flags;
662 struct nfs_pageio_descriptor pgio;
663 int err;
664
665 /* Stop dirtying of new pages while we sync */
666 err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING,
667 nfs_wait_bit_killable, TASK_KILLABLE);
668 if (err)
669 goto out_err;
670
671 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
672
673 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
674 &nfs_async_write_completion_ops);
675 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
676 nfs_pageio_complete(&pgio);
677
678 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
679 smp_mb__after_atomic();
680 wake_up_bit(bitlock, NFS_INO_FLUSHING);
681
682 if (err < 0)
683 goto out_err;
684 err = pgio.pg_error;
685 if (err < 0)
686 goto out_err;
687 return 0;
688out_err:
689 return err;
690}
691
692/*
693 * Insert a write request into an inode
694 */
695static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
696{
697 struct nfs_inode *nfsi = NFS_I(inode);
698
699 WARN_ON_ONCE(req->wb_this_page != req);
700
701 /* Lock the request! */
702 nfs_lock_request(req);
703
704 spin_lock(&inode->i_lock);
705 if (!nfsi->nrequests &&
706 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
707 inode->i_version++;
708 /*
709 * Swap-space should not get truncated. Hence no need to plug the race
710 * with invalidate/truncate.
711 */
712 if (likely(!PageSwapCache(req->wb_page))) {
713 set_bit(PG_MAPPED, &req->wb_flags);
714 SetPagePrivate(req->wb_page);
715 set_page_private(req->wb_page, (unsigned long)req);
716 }
717 nfsi->nrequests++;
718 /* this a head request for a page group - mark it as having an
719 * extra reference so sub groups can follow suit.
720 * This flag also informs pgio layer when to bump nrequests when
721 * adding subrequests. */
722 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
723 kref_get(&req->wb_kref);
724 spin_unlock(&inode->i_lock);
725}
726
727/*
728 * Remove a write request from an inode
729 */
730static void nfs_inode_remove_request(struct nfs_page *req)
731{
732 struct inode *inode = d_inode(req->wb_context->dentry);
733 struct nfs_inode *nfsi = NFS_I(inode);
734 struct nfs_page *head;
735
736 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
737 head = req->wb_head;
738
739 spin_lock(&inode->i_lock);
740 if (likely(!PageSwapCache(head->wb_page))) {
741 set_page_private(head->wb_page, 0);
742 ClearPagePrivate(head->wb_page);
743 smp_mb__after_atomic();
744 wake_up_page(head->wb_page, PG_private);
745 clear_bit(PG_MAPPED, &head->wb_flags);
746 }
747 nfsi->nrequests--;
748 spin_unlock(&inode->i_lock);
749 } else {
750 spin_lock(&inode->i_lock);
751 nfsi->nrequests--;
752 spin_unlock(&inode->i_lock);
753 }
754
755 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
756 nfs_release_request(req);
757}
758
759static void
760nfs_mark_request_dirty(struct nfs_page *req)
761{
762 __set_page_dirty_nobuffers(req->wb_page);
763}
764
765/*
766 * nfs_page_search_commits_for_head_request_locked
767 *
768 * Search through commit lists on @inode for the head request for @page.
769 * Must be called while holding the inode (which is cinfo) lock.
770 *
771 * Returns the head request if found, or NULL if not found.
772 */
773static struct nfs_page *
774nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
775 struct page *page)
776{
777 struct nfs_page *freq, *t;
778 struct nfs_commit_info cinfo;
779 struct inode *inode = &nfsi->vfs_inode;
780
781 nfs_init_cinfo_from_inode(&cinfo, inode);
782
783 /* search through pnfs commit lists */
784 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
785 if (freq)
786 return freq->wb_head;
787
788 /* Linearly search the commit list for the correct request */
789 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
790 if (freq->wb_page == page)
791 return freq->wb_head;
792 }
793
794 return NULL;
795}
796
797/**
798 * nfs_request_add_commit_list_locked - add request to a commit list
799 * @req: pointer to a struct nfs_page
800 * @dst: commit list head
801 * @cinfo: holds list lock and accounting info
802 *
803 * This sets the PG_CLEAN bit, updates the cinfo count of
804 * number of outstanding requests requiring a commit as well as
805 * the MM page stats.
806 *
807 * The caller must hold the cinfo->lock, and the nfs_page lock.
808 */
809void
810nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
811 struct nfs_commit_info *cinfo)
812{
813 set_bit(PG_CLEAN, &req->wb_flags);
814 nfs_list_add_request(req, dst);
815 cinfo->mds->ncommit++;
816}
817EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
818
819/**
820 * nfs_request_add_commit_list - add request to a commit list
821 * @req: pointer to a struct nfs_page
822 * @dst: commit list head
823 * @cinfo: holds list lock and accounting info
824 *
825 * This sets the PG_CLEAN bit, updates the cinfo count of
826 * number of outstanding requests requiring a commit as well as
827 * the MM page stats.
828 *
829 * The caller must _not_ hold the cinfo->lock, but must be
830 * holding the nfs_page lock.
831 */
832void
833nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
834{
835 spin_lock(cinfo->lock);
836 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
837 spin_unlock(cinfo->lock);
838 nfs_mark_page_unstable(req->wb_page, cinfo);
839}
840EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
841
842/**
843 * nfs_request_remove_commit_list - Remove request from a commit list
844 * @req: pointer to a nfs_page
845 * @cinfo: holds list lock and accounting info
846 *
847 * This clears the PG_CLEAN bit, and updates the cinfo's count of
848 * number of outstanding requests requiring a commit
849 * It does not update the MM page stats.
850 *
851 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
852 */
853void
854nfs_request_remove_commit_list(struct nfs_page *req,
855 struct nfs_commit_info *cinfo)
856{
857 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
858 return;
859 nfs_list_remove_request(req);
860 cinfo->mds->ncommit--;
861}
862EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
863
864static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
865 struct inode *inode)
866{
867 cinfo->lock = &inode->i_lock;
868 cinfo->mds = &NFS_I(inode)->commit_info;
869 cinfo->ds = pnfs_get_ds_info(inode);
870 cinfo->dreq = NULL;
871 cinfo->completion_ops = &nfs_commit_completion_ops;
872}
873
874void nfs_init_cinfo(struct nfs_commit_info *cinfo,
875 struct inode *inode,
876 struct nfs_direct_req *dreq)
877{
878 if (dreq)
879 nfs_init_cinfo_from_dreq(cinfo, dreq);
880 else
881 nfs_init_cinfo_from_inode(cinfo, inode);
882}
883EXPORT_SYMBOL_GPL(nfs_init_cinfo);
884
885/*
886 * Add a request to the inode's commit list.
887 */
888void
889nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
890 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
891{
892 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
893 return;
894 nfs_request_add_commit_list(req, cinfo);
895}
896
897static void
898nfs_clear_page_commit(struct page *page)
899{
900 dec_zone_page_state(page, NR_UNSTABLE_NFS);
901 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
902 WB_RECLAIMABLE);
903}
904
905/* Called holding inode (/cinfo) lock */
906static void
907nfs_clear_request_commit(struct nfs_page *req)
908{
909 if (test_bit(PG_CLEAN, &req->wb_flags)) {
910 struct inode *inode = d_inode(req->wb_context->dentry);
911 struct nfs_commit_info cinfo;
912
913 nfs_init_cinfo_from_inode(&cinfo, inode);
914 if (!pnfs_clear_request_commit(req, &cinfo)) {
915 nfs_request_remove_commit_list(req, &cinfo);
916 }
917 nfs_clear_page_commit(req->wb_page);
918 }
919}
920
921int nfs_write_need_commit(struct nfs_pgio_header *hdr)
922{
923 if (hdr->verf.committed == NFS_DATA_SYNC)
924 return hdr->lseg == NULL;
925 return hdr->verf.committed != NFS_FILE_SYNC;
926}
927
928static void nfs_write_completion(struct nfs_pgio_header *hdr)
929{
930 struct nfs_commit_info cinfo;
931 unsigned long bytes = 0;
932
933 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
934 goto out;
935 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
936 while (!list_empty(&hdr->pages)) {
937 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
938
939 bytes += req->wb_bytes;
940 nfs_list_remove_request(req);
941 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
942 (hdr->good_bytes < bytes)) {
943 nfs_set_pageerror(req->wb_page);
944 nfs_context_set_write_error(req->wb_context, hdr->error);
945 goto remove_req;
946 }
947 if (nfs_write_need_commit(hdr)) {
948 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
949 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
950 hdr->pgio_mirror_idx);
951 goto next;
952 }
953remove_req:
954 nfs_inode_remove_request(req);
955next:
956 nfs_unlock_request(req);
957 nfs_end_page_writeback(req);
958 nfs_release_request(req);
959 }
960out:
961 hdr->release(hdr);
962}
963
964unsigned long
965nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
966{
967 return cinfo->mds->ncommit;
968}
969
970/* cinfo->lock held by caller */
971int
972nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
973 struct nfs_commit_info *cinfo, int max)
974{
975 struct nfs_page *req, *tmp;
976 int ret = 0;
977
978 list_for_each_entry_safe(req, tmp, src, wb_list) {
979 if (!nfs_lock_request(req))
980 continue;
981 kref_get(&req->wb_kref);
982 if (cond_resched_lock(cinfo->lock))
983 list_safe_reset_next(req, tmp, wb_list);
984 nfs_request_remove_commit_list(req, cinfo);
985 nfs_list_add_request(req, dst);
986 ret++;
987 if ((ret == max) && !cinfo->dreq)
988 break;
989 }
990 return ret;
991}
992
993/*
994 * nfs_scan_commit - Scan an inode for commit requests
995 * @inode: NFS inode to scan
996 * @dst: mds destination list
997 * @cinfo: mds and ds lists of reqs ready to commit
998 *
999 * Moves requests from the inode's 'commit' request list.
1000 * The requests are *not* checked to ensure that they form a contiguous set.
1001 */
1002int
1003nfs_scan_commit(struct inode *inode, struct list_head *dst,
1004 struct nfs_commit_info *cinfo)
1005{
1006 int ret = 0;
1007
1008 spin_lock(cinfo->lock);
1009 if (cinfo->mds->ncommit > 0) {
1010 const int max = INT_MAX;
1011
1012 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1013 cinfo, max);
1014 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1015 }
1016 spin_unlock(cinfo->lock);
1017 return ret;
1018}
1019
1020/*
1021 * Search for an existing write request, and attempt to update
1022 * it to reflect a new dirty region on a given page.
1023 *
1024 * If the attempt fails, then the existing request is flushed out
1025 * to disk.
1026 */
1027static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1028 struct page *page,
1029 unsigned int offset,
1030 unsigned int bytes)
1031{
1032 struct nfs_page *req;
1033 unsigned int rqend;
1034 unsigned int end;
1035 int error;
1036
1037 if (!PagePrivate(page))
1038 return NULL;
1039
1040 end = offset + bytes;
1041 spin_lock(&inode->i_lock);
1042
1043 for (;;) {
1044 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
1045 if (req == NULL)
1046 goto out_unlock;
1047
1048 /* should be handled by nfs_flush_incompatible */
1049 WARN_ON_ONCE(req->wb_head != req);
1050 WARN_ON_ONCE(req->wb_this_page != req);
1051
1052 rqend = req->wb_offset + req->wb_bytes;
1053 /*
1054 * Tell the caller to flush out the request if
1055 * the offsets are non-contiguous.
1056 * Note: nfs_flush_incompatible() will already
1057 * have flushed out requests having wrong owners.
1058 */
1059 if (offset > rqend
1060 || end < req->wb_offset)
1061 goto out_flushme;
1062
1063 if (nfs_lock_request(req))
1064 break;
1065
1066 /* The request is locked, so wait and then retry */
1067 spin_unlock(&inode->i_lock);
1068 error = nfs_wait_on_request(req);
1069 nfs_release_request(req);
1070 if (error != 0)
1071 goto out_err;
1072 spin_lock(&inode->i_lock);
1073 }
1074
1075 /* Okay, the request matches. Update the region */
1076 if (offset < req->wb_offset) {
1077 req->wb_offset = offset;
1078 req->wb_pgbase = offset;
1079 }
1080 if (end > rqend)
1081 req->wb_bytes = end - req->wb_offset;
1082 else
1083 req->wb_bytes = rqend - req->wb_offset;
1084out_unlock:
1085 if (req)
1086 nfs_clear_request_commit(req);
1087 spin_unlock(&inode->i_lock);
1088 return req;
1089out_flushme:
1090 spin_unlock(&inode->i_lock);
1091 nfs_release_request(req);
1092 error = nfs_wb_page(inode, page);
1093out_err:
1094 return ERR_PTR(error);
1095}
1096
1097/*
1098 * Try to update an existing write request, or create one if there is none.
1099 *
1100 * Note: Should always be called with the Page Lock held to prevent races
1101 * if we have to add a new request. Also assumes that the caller has
1102 * already called nfs_flush_incompatible() if necessary.
1103 */
1104static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1105 struct page *page, unsigned int offset, unsigned int bytes)
1106{
1107 struct inode *inode = page_file_mapping(page)->host;
1108 struct nfs_page *req;
1109
1110 req = nfs_try_to_update_request(inode, page, offset, bytes);
1111 if (req != NULL)
1112 goto out;
1113 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1114 if (IS_ERR(req))
1115 goto out;
1116 nfs_inode_add_request(inode, req);
1117out:
1118 return req;
1119}
1120
1121static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1122 unsigned int offset, unsigned int count)
1123{
1124 struct nfs_page *req;
1125
1126 req = nfs_setup_write_request(ctx, page, offset, count);
1127 if (IS_ERR(req))
1128 return PTR_ERR(req);
1129 /* Update file length */
1130 nfs_grow_file(page, offset, count);
1131 nfs_mark_uptodate(req);
1132 nfs_mark_request_dirty(req);
1133 nfs_unlock_and_release_request(req);
1134 return 0;
1135}
1136
1137int nfs_flush_incompatible(struct file *file, struct page *page)
1138{
1139 struct nfs_open_context *ctx = nfs_file_open_context(file);
1140 struct nfs_lock_context *l_ctx;
1141 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1142 struct nfs_page *req;
1143 int do_flush, status;
1144 /*
1145 * Look for a request corresponding to this page. If there
1146 * is one, and it belongs to another file, we flush it out
1147 * before we try to copy anything into the page. Do this
1148 * due to the lack of an ACCESS-type call in NFSv2.
1149 * Also do the same if we find a request from an existing
1150 * dropped page.
1151 */
1152 do {
1153 req = nfs_page_find_head_request(page);
1154 if (req == NULL)
1155 return 0;
1156 l_ctx = req->wb_lock_context;
1157 do_flush = req->wb_page != page ||
1158 !nfs_match_open_context(req->wb_context, ctx);
1159 /* for now, flush if more than 1 request in page_group */
1160 do_flush |= req->wb_this_page != req;
1161 if (l_ctx && flctx &&
1162 !(list_empty_careful(&flctx->flc_posix) &&
1163 list_empty_careful(&flctx->flc_flock))) {
1164 do_flush |= l_ctx->lockowner.l_owner != current->files
1165 || l_ctx->lockowner.l_pid != current->tgid;
1166 }
1167 nfs_release_request(req);
1168 if (!do_flush)
1169 return 0;
1170 status = nfs_wb_page(page_file_mapping(page)->host, page);
1171 } while (status == 0);
1172 return status;
1173}
1174
1175/*
1176 * Avoid buffered writes when a open context credential's key would
1177 * expire soon.
1178 *
1179 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1180 *
1181 * Return 0 and set a credential flag which triggers the inode to flush
1182 * and performs NFS_FILE_SYNC writes if the key will expired within
1183 * RPC_KEY_EXPIRE_TIMEO.
1184 */
1185int
1186nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1187{
1188 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1189 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1190
1191 return rpcauth_key_timeout_notify(auth, ctx->cred);
1192}
1193
1194/*
1195 * Test if the open context credential key is marked to expire soon.
1196 */
1197bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1198{
1199 return rpcauth_cred_key_to_expire(ctx->cred);
1200}
1201
1202/*
1203 * If the page cache is marked as unsafe or invalid, then we can't rely on
1204 * the PageUptodate() flag. In this case, we will need to turn off
1205 * write optimisations that depend on the page contents being correct.
1206 */
1207static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1208{
1209 struct nfs_inode *nfsi = NFS_I(inode);
1210
1211 if (nfs_have_delegated_attributes(inode))
1212 goto out;
1213 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1214 return false;
1215 smp_rmb();
1216 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1217 return false;
1218out:
1219 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1220 return false;
1221 return PageUptodate(page) != 0;
1222}
1223
1224static bool
1225is_whole_file_wrlock(struct file_lock *fl)
1226{
1227 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1228 fl->fl_type == F_WRLCK;
1229}
1230
1231/* If we know the page is up to date, and we're not using byte range locks (or
1232 * if we have the whole file locked for writing), it may be more efficient to
1233 * extend the write to cover the entire page in order to avoid fragmentation
1234 * inefficiencies.
1235 *
1236 * If the file is opened for synchronous writes then we can just skip the rest
1237 * of the checks.
1238 */
1239static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1240{
1241 int ret;
1242 struct file_lock_context *flctx = inode->i_flctx;
1243 struct file_lock *fl;
1244
1245 if (file->f_flags & O_DSYNC)
1246 return 0;
1247 if (!nfs_write_pageuptodate(page, inode))
1248 return 0;
1249 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1250 return 1;
1251 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1252 list_empty_careful(&flctx->flc_posix)))
1253 return 1;
1254
1255 /* Check to see if there are whole file write locks */
1256 ret = 0;
1257 spin_lock(&flctx->flc_lock);
1258 if (!list_empty(&flctx->flc_posix)) {
1259 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1260 fl_list);
1261 if (is_whole_file_wrlock(fl))
1262 ret = 1;
1263 } else if (!list_empty(&flctx->flc_flock)) {
1264 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1265 fl_list);
1266 if (fl->fl_type == F_WRLCK)
1267 ret = 1;
1268 }
1269 spin_unlock(&flctx->flc_lock);
1270 return ret;
1271}
1272
1273/*
1274 * Update and possibly write a cached page of an NFS file.
1275 *
1276 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1277 * things with a page scheduled for an RPC call (e.g. invalidate it).
1278 */
1279int nfs_updatepage(struct file *file, struct page *page,
1280 unsigned int offset, unsigned int count)
1281{
1282 struct nfs_open_context *ctx = nfs_file_open_context(file);
1283 struct inode *inode = page_file_mapping(page)->host;
1284 int status = 0;
1285
1286 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1287
1288 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1289 file, count, (long long)(page_file_offset(page) + offset));
1290
1291 if (nfs_can_extend_write(file, page, inode)) {
1292 count = max(count + offset, nfs_page_length(page));
1293 offset = 0;
1294 }
1295
1296 status = nfs_writepage_setup(ctx, page, offset, count);
1297 if (status < 0)
1298 nfs_set_pageerror(page);
1299 else
1300 __set_page_dirty_nobuffers(page);
1301
1302 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1303 status, (long long)i_size_read(inode));
1304 return status;
1305}
1306
1307static int flush_task_priority(int how)
1308{
1309 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1310 case FLUSH_HIGHPRI:
1311 return RPC_PRIORITY_HIGH;
1312 case FLUSH_LOWPRI:
1313 return RPC_PRIORITY_LOW;
1314 }
1315 return RPC_PRIORITY_NORMAL;
1316}
1317
1318static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1319 struct rpc_message *msg,
1320 const struct nfs_rpc_ops *rpc_ops,
1321 struct rpc_task_setup *task_setup_data, int how)
1322{
1323 int priority = flush_task_priority(how);
1324
1325 task_setup_data->priority = priority;
1326 rpc_ops->write_setup(hdr, msg);
1327
1328 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1329 &task_setup_data->rpc_client, msg, hdr);
1330}
1331
1332/* If a nfs_flush_* function fails, it should remove reqs from @head and
1333 * call this on each, which will prepare them to be retried on next
1334 * writeback using standard nfs.
1335 */
1336static void nfs_redirty_request(struct nfs_page *req)
1337{
1338 nfs_mark_request_dirty(req);
1339 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1340 nfs_unlock_request(req);
1341 nfs_end_page_writeback(req);
1342 nfs_release_request(req);
1343}
1344
1345static void nfs_async_write_error(struct list_head *head)
1346{
1347 struct nfs_page *req;
1348
1349 while (!list_empty(head)) {
1350 req = nfs_list_entry(head->next);
1351 nfs_list_remove_request(req);
1352 nfs_redirty_request(req);
1353 }
1354}
1355
1356static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1357{
1358 nfs_async_write_error(&hdr->pages);
1359}
1360
1361static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1362 .error_cleanup = nfs_async_write_error,
1363 .completion = nfs_write_completion,
1364 .reschedule_io = nfs_async_write_reschedule_io,
1365};
1366
1367void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1368 struct inode *inode, int ioflags, bool force_mds,
1369 const struct nfs_pgio_completion_ops *compl_ops)
1370{
1371 struct nfs_server *server = NFS_SERVER(inode);
1372 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1373
1374#ifdef CONFIG_NFS_V4_1
1375 if (server->pnfs_curr_ld && !force_mds)
1376 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1377#endif
1378 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1379 server->wsize, ioflags);
1380}
1381EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1382
1383void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1384{
1385 struct nfs_pgio_mirror *mirror;
1386
1387 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1388 pgio->pg_ops->pg_cleanup(pgio);
1389
1390 pgio->pg_ops = &nfs_pgio_rw_ops;
1391
1392 nfs_pageio_stop_mirroring(pgio);
1393
1394 mirror = &pgio->pg_mirrors[0];
1395 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1396}
1397EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1398
1399
1400void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1401{
1402 struct nfs_commit_data *data = calldata;
1403
1404 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1405}
1406
1407/*
1408 * Special version of should_remove_suid() that ignores capabilities.
1409 */
1410static int nfs_should_remove_suid(const struct inode *inode)
1411{
1412 umode_t mode = inode->i_mode;
1413 int kill = 0;
1414
1415 /* suid always must be killed */
1416 if (unlikely(mode & S_ISUID))
1417 kill = ATTR_KILL_SUID;
1418
1419 /*
1420 * sgid without any exec bits is just a mandatory locking mark; leave
1421 * it alone. If some exec bits are set, it's a real sgid; kill it.
1422 */
1423 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1424 kill |= ATTR_KILL_SGID;
1425
1426 if (unlikely(kill && S_ISREG(mode)))
1427 return kill;
1428
1429 return 0;
1430}
1431
1432static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1433 struct nfs_fattr *fattr)
1434{
1435 struct nfs_pgio_args *argp = &hdr->args;
1436 struct nfs_pgio_res *resp = &hdr->res;
1437 u64 size = argp->offset + resp->count;
1438
1439 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1440 fattr->size = size;
1441 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1442 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1443 return;
1444 }
1445 if (size != fattr->size)
1446 return;
1447 /* Set attribute barrier */
1448 nfs_fattr_set_barrier(fattr);
1449 /* ...and update size */
1450 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1451}
1452
1453void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1454{
1455 struct nfs_fattr *fattr = &hdr->fattr;
1456 struct inode *inode = hdr->inode;
1457
1458 spin_lock(&inode->i_lock);
1459 nfs_writeback_check_extend(hdr, fattr);
1460 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1461 spin_unlock(&inode->i_lock);
1462}
1463EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1464
1465/*
1466 * This function is called when the WRITE call is complete.
1467 */
1468static int nfs_writeback_done(struct rpc_task *task,
1469 struct nfs_pgio_header *hdr,
1470 struct inode *inode)
1471{
1472 int status;
1473
1474 /*
1475 * ->write_done will attempt to use post-op attributes to detect
1476 * conflicting writes by other clients. A strict interpretation
1477 * of close-to-open would allow us to continue caching even if
1478 * another writer had changed the file, but some applications
1479 * depend on tighter cache coherency when writing.
1480 */
1481 status = NFS_PROTO(inode)->write_done(task, hdr);
1482 if (status != 0)
1483 return status;
1484 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1485
1486 if (hdr->res.verf->committed < hdr->args.stable &&
1487 task->tk_status >= 0) {
1488 /* We tried a write call, but the server did not
1489 * commit data to stable storage even though we
1490 * requested it.
1491 * Note: There is a known bug in Tru64 < 5.0 in which
1492 * the server reports NFS_DATA_SYNC, but performs
1493 * NFS_FILE_SYNC. We therefore implement this checking
1494 * as a dprintk() in order to avoid filling syslog.
1495 */
1496 static unsigned long complain;
1497
1498 /* Note this will print the MDS for a DS write */
1499 if (time_before(complain, jiffies)) {
1500 dprintk("NFS: faulty NFS server %s:"
1501 " (committed = %d) != (stable = %d)\n",
1502 NFS_SERVER(inode)->nfs_client->cl_hostname,
1503 hdr->res.verf->committed, hdr->args.stable);
1504 complain = jiffies + 300 * HZ;
1505 }
1506 }
1507
1508 /* Deal with the suid/sgid bit corner case */
1509 if (nfs_should_remove_suid(inode))
1510 nfs_mark_for_revalidate(inode);
1511 return 0;
1512}
1513
1514/*
1515 * This function is called when the WRITE call is complete.
1516 */
1517static void nfs_writeback_result(struct rpc_task *task,
1518 struct nfs_pgio_header *hdr)
1519{
1520 struct nfs_pgio_args *argp = &hdr->args;
1521 struct nfs_pgio_res *resp = &hdr->res;
1522
1523 if (resp->count < argp->count) {
1524 static unsigned long complain;
1525
1526 /* This a short write! */
1527 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1528
1529 /* Has the server at least made some progress? */
1530 if (resp->count == 0) {
1531 if (time_before(complain, jiffies)) {
1532 printk(KERN_WARNING
1533 "NFS: Server wrote zero bytes, expected %u.\n",
1534 argp->count);
1535 complain = jiffies + 300 * HZ;
1536 }
1537 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1538 task->tk_status = -EIO;
1539 return;
1540 }
1541
1542 /* For non rpc-based layout drivers, retry-through-MDS */
1543 if (!task->tk_ops) {
1544 hdr->pnfs_error = -EAGAIN;
1545 return;
1546 }
1547
1548 /* Was this an NFSv2 write or an NFSv3 stable write? */
1549 if (resp->verf->committed != NFS_UNSTABLE) {
1550 /* Resend from where the server left off */
1551 hdr->mds_offset += resp->count;
1552 argp->offset += resp->count;
1553 argp->pgbase += resp->count;
1554 argp->count -= resp->count;
1555 } else {
1556 /* Resend as a stable write in order to avoid
1557 * headaches in the case of a server crash.
1558 */
1559 argp->stable = NFS_FILE_SYNC;
1560 }
1561 rpc_restart_call_prepare(task);
1562 }
1563}
1564
1565static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1566{
1567 return wait_on_atomic_t(&cinfo->rpcs_out,
1568 nfs_wait_atomic_killable, TASK_KILLABLE);
1569}
1570
1571static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1572{
1573 atomic_inc(&cinfo->rpcs_out);
1574}
1575
1576static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1577{
1578 if (atomic_dec_and_test(&cinfo->rpcs_out))
1579 wake_up_atomic_t(&cinfo->rpcs_out);
1580}
1581
1582void nfs_commitdata_release(struct nfs_commit_data *data)
1583{
1584 put_nfs_open_context(data->context);
1585 nfs_commit_free(data);
1586}
1587EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1588
1589int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1590 const struct nfs_rpc_ops *nfs_ops,
1591 const struct rpc_call_ops *call_ops,
1592 int how, int flags)
1593{
1594 struct rpc_task *task;
1595 int priority = flush_task_priority(how);
1596 struct rpc_message msg = {
1597 .rpc_argp = &data->args,
1598 .rpc_resp = &data->res,
1599 .rpc_cred = data->cred,
1600 };
1601 struct rpc_task_setup task_setup_data = {
1602 .task = &data->task,
1603 .rpc_client = clnt,
1604 .rpc_message = &msg,
1605 .callback_ops = call_ops,
1606 .callback_data = data,
1607 .workqueue = nfsiod_workqueue,
1608 .flags = RPC_TASK_ASYNC | flags,
1609 .priority = priority,
1610 };
1611 /* Set up the initial task struct. */
1612 nfs_ops->commit_setup(data, &msg);
1613
1614 dprintk("NFS: initiated commit call\n");
1615
1616 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1617 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1618
1619 task = rpc_run_task(&task_setup_data);
1620 if (IS_ERR(task))
1621 return PTR_ERR(task);
1622 if (how & FLUSH_SYNC)
1623 rpc_wait_for_completion_task(task);
1624 rpc_put_task(task);
1625 return 0;
1626}
1627EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1628
1629static loff_t nfs_get_lwb(struct list_head *head)
1630{
1631 loff_t lwb = 0;
1632 struct nfs_page *req;
1633
1634 list_for_each_entry(req, head, wb_list)
1635 if (lwb < (req_offset(req) + req->wb_bytes))
1636 lwb = req_offset(req) + req->wb_bytes;
1637
1638 return lwb;
1639}
1640
1641/*
1642 * Set up the argument/result storage required for the RPC call.
1643 */
1644void nfs_init_commit(struct nfs_commit_data *data,
1645 struct list_head *head,
1646 struct pnfs_layout_segment *lseg,
1647 struct nfs_commit_info *cinfo)
1648{
1649 struct nfs_page *first = nfs_list_entry(head->next);
1650 struct inode *inode = d_inode(first->wb_context->dentry);
1651
1652 /* Set up the RPC argument and reply structs
1653 * NB: take care not to mess about with data->commit et al. */
1654
1655 list_splice_init(head, &data->pages);
1656
1657 data->inode = inode;
1658 data->cred = first->wb_context->cred;
1659 data->lseg = lseg; /* reference transferred */
1660 /* only set lwb for pnfs commit */
1661 if (lseg)
1662 data->lwb = nfs_get_lwb(&data->pages);
1663 data->mds_ops = &nfs_commit_ops;
1664 data->completion_ops = cinfo->completion_ops;
1665 data->dreq = cinfo->dreq;
1666
1667 data->args.fh = NFS_FH(data->inode);
1668 /* Note: we always request a commit of the entire inode */
1669 data->args.offset = 0;
1670 data->args.count = 0;
1671 data->context = get_nfs_open_context(first->wb_context);
1672 data->res.fattr = &data->fattr;
1673 data->res.verf = &data->verf;
1674 nfs_fattr_init(&data->fattr);
1675}
1676EXPORT_SYMBOL_GPL(nfs_init_commit);
1677
1678void nfs_retry_commit(struct list_head *page_list,
1679 struct pnfs_layout_segment *lseg,
1680 struct nfs_commit_info *cinfo,
1681 u32 ds_commit_idx)
1682{
1683 struct nfs_page *req;
1684
1685 while (!list_empty(page_list)) {
1686 req = nfs_list_entry(page_list->next);
1687 nfs_list_remove_request(req);
1688 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1689 if (!cinfo->dreq)
1690 nfs_clear_page_commit(req->wb_page);
1691 nfs_unlock_and_release_request(req);
1692 }
1693}
1694EXPORT_SYMBOL_GPL(nfs_retry_commit);
1695
1696static void
1697nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1698 struct nfs_page *req)
1699{
1700 __set_page_dirty_nobuffers(req->wb_page);
1701}
1702
1703/*
1704 * Commit dirty pages
1705 */
1706static int
1707nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1708 struct nfs_commit_info *cinfo)
1709{
1710 struct nfs_commit_data *data;
1711
1712 data = nfs_commitdata_alloc();
1713
1714 if (!data)
1715 goto out_bad;
1716
1717 /* Set up the argument struct */
1718 nfs_init_commit(data, head, NULL, cinfo);
1719 atomic_inc(&cinfo->mds->rpcs_out);
1720 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1721 data->mds_ops, how, 0);
1722 out_bad:
1723 nfs_retry_commit(head, NULL, cinfo, 0);
1724 return -ENOMEM;
1725}
1726
1727/*
1728 * COMMIT call returned
1729 */
1730static void nfs_commit_done(struct rpc_task *task, void *calldata)
1731{
1732 struct nfs_commit_data *data = calldata;
1733
1734 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1735 task->tk_pid, task->tk_status);
1736
1737 /* Call the NFS version-specific code */
1738 NFS_PROTO(data->inode)->commit_done(task, data);
1739}
1740
1741static void nfs_commit_release_pages(struct nfs_commit_data *data)
1742{
1743 struct nfs_page *req;
1744 int status = data->task.tk_status;
1745 struct nfs_commit_info cinfo;
1746 struct nfs_server *nfss;
1747
1748 while (!list_empty(&data->pages)) {
1749 req = nfs_list_entry(data->pages.next);
1750 nfs_list_remove_request(req);
1751 nfs_clear_page_commit(req->wb_page);
1752
1753 dprintk("NFS: commit (%s/%llu %d@%lld)",
1754 req->wb_context->dentry->d_sb->s_id,
1755 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1756 req->wb_bytes,
1757 (long long)req_offset(req));
1758 if (status < 0) {
1759 nfs_context_set_write_error(req->wb_context, status);
1760 nfs_inode_remove_request(req);
1761 dprintk(", error = %d\n", status);
1762 goto next;
1763 }
1764
1765 /* Okay, COMMIT succeeded, apparently. Check the verifier
1766 * returned by the server against all stored verfs. */
1767 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1768 /* We have a match */
1769 nfs_inode_remove_request(req);
1770 dprintk(" OK\n");
1771 goto next;
1772 }
1773 /* We have a mismatch. Write the page again */
1774 dprintk(" mismatch\n");
1775 nfs_mark_request_dirty(req);
1776 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1777 next:
1778 nfs_unlock_and_release_request(req);
1779 }
1780 nfss = NFS_SERVER(data->inode);
1781 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1782 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1783
1784 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1785 nfs_commit_end(cinfo.mds);
1786}
1787
1788static void nfs_commit_release(void *calldata)
1789{
1790 struct nfs_commit_data *data = calldata;
1791
1792 data->completion_ops->completion(data);
1793 nfs_commitdata_release(calldata);
1794}
1795
1796static const struct rpc_call_ops nfs_commit_ops = {
1797 .rpc_call_prepare = nfs_commit_prepare,
1798 .rpc_call_done = nfs_commit_done,
1799 .rpc_release = nfs_commit_release,
1800};
1801
1802static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1803 .completion = nfs_commit_release_pages,
1804 .resched_write = nfs_commit_resched_write,
1805};
1806
1807int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1808 int how, struct nfs_commit_info *cinfo)
1809{
1810 int status;
1811
1812 status = pnfs_commit_list(inode, head, how, cinfo);
1813 if (status == PNFS_NOT_ATTEMPTED)
1814 status = nfs_commit_list(inode, head, how, cinfo);
1815 return status;
1816}
1817
1818int nfs_commit_inode(struct inode *inode, int how)
1819{
1820 LIST_HEAD(head);
1821 struct nfs_commit_info cinfo;
1822 int may_wait = how & FLUSH_SYNC;
1823 int error = 0;
1824 int res;
1825
1826 nfs_init_cinfo_from_inode(&cinfo, inode);
1827 nfs_commit_begin(cinfo.mds);
1828 res = nfs_scan_commit(inode, &head, &cinfo);
1829 if (res)
1830 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1831 nfs_commit_end(cinfo.mds);
1832 if (error < 0)
1833 goto out_error;
1834 if (!may_wait)
1835 goto out_mark_dirty;
1836 error = wait_on_commit(cinfo.mds);
1837 if (error < 0)
1838 return error;
1839 return res;
1840out_error:
1841 res = error;
1842 /* Note: If we exit without ensuring that the commit is complete,
1843 * we must mark the inode as dirty. Otherwise, future calls to
1844 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1845 * that the data is on the disk.
1846 */
1847out_mark_dirty:
1848 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1849 return res;
1850}
1851EXPORT_SYMBOL_GPL(nfs_commit_inode);
1852
1853int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1854{
1855 struct nfs_inode *nfsi = NFS_I(inode);
1856 int flags = FLUSH_SYNC;
1857 int ret = 0;
1858
1859 /* no commits means nothing needs to be done */
1860 if (!nfsi->commit_info.ncommit)
1861 return ret;
1862
1863 if (wbc->sync_mode == WB_SYNC_NONE) {
1864 /* Don't commit yet if this is a non-blocking flush and there
1865 * are a lot of outstanding writes for this mapping.
1866 */
1867 if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1))
1868 goto out_mark_dirty;
1869
1870 /* don't wait for the COMMIT response */
1871 flags = 0;
1872 }
1873
1874 ret = nfs_commit_inode(inode, flags);
1875 if (ret >= 0) {
1876 if (wbc->sync_mode == WB_SYNC_NONE) {
1877 if (ret < wbc->nr_to_write)
1878 wbc->nr_to_write -= ret;
1879 else
1880 wbc->nr_to_write = 0;
1881 }
1882 return 0;
1883 }
1884out_mark_dirty:
1885 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1886 return ret;
1887}
1888EXPORT_SYMBOL_GPL(nfs_write_inode);
1889
1890/*
1891 * flush the inode to disk.
1892 */
1893int nfs_wb_all(struct inode *inode)
1894{
1895 int ret;
1896
1897 trace_nfs_writeback_inode_enter(inode);
1898
1899 ret = filemap_write_and_wait(inode->i_mapping);
1900 if (ret)
1901 goto out;
1902 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1903 if (ret < 0)
1904 goto out;
1905 pnfs_sync_inode(inode, true);
1906 ret = 0;
1907
1908out:
1909 trace_nfs_writeback_inode_exit(inode, ret);
1910 return ret;
1911}
1912EXPORT_SYMBOL_GPL(nfs_wb_all);
1913
1914int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1915{
1916 struct nfs_page *req;
1917 int ret = 0;
1918
1919 wait_on_page_writeback(page);
1920
1921 /* blocking call to cancel all requests and join to a single (head)
1922 * request */
1923 req = nfs_lock_and_join_requests(page, false);
1924
1925 if (IS_ERR(req)) {
1926 ret = PTR_ERR(req);
1927 } else if (req) {
1928 /* all requests from this page have been cancelled by
1929 * nfs_lock_and_join_requests, so just remove the head
1930 * request from the inode / page_private pointer and
1931 * release it */
1932 nfs_inode_remove_request(req);
1933 nfs_unlock_and_release_request(req);
1934 }
1935
1936 return ret;
1937}
1938
1939/*
1940 * Write back all requests on one page - we do this before reading it.
1941 */
1942int nfs_wb_single_page(struct inode *inode, struct page *page, bool launder)
1943{
1944 loff_t range_start = page_file_offset(page);
1945 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
1946 struct writeback_control wbc = {
1947 .sync_mode = WB_SYNC_ALL,
1948 .nr_to_write = 0,
1949 .range_start = range_start,
1950 .range_end = range_end,
1951 };
1952 int ret;
1953
1954 trace_nfs_writeback_page_enter(inode);
1955
1956 for (;;) {
1957 wait_on_page_writeback(page);
1958 if (clear_page_dirty_for_io(page)) {
1959 ret = nfs_writepage_locked(page, &wbc, launder);
1960 if (ret < 0)
1961 goto out_error;
1962 continue;
1963 }
1964 ret = 0;
1965 if (!PagePrivate(page))
1966 break;
1967 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1968 if (ret < 0)
1969 goto out_error;
1970 }
1971out_error:
1972 trace_nfs_writeback_page_exit(inode, ret);
1973 return ret;
1974}
1975
1976#ifdef CONFIG_MIGRATION
1977int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1978 struct page *page, enum migrate_mode mode)
1979{
1980 /*
1981 * If PagePrivate is set, then the page is currently associated with
1982 * an in-progress read or write request. Don't try to migrate it.
1983 *
1984 * FIXME: we could do this in principle, but we'll need a way to ensure
1985 * that we can safely release the inode reference while holding
1986 * the page lock.
1987 */
1988 if (PagePrivate(page))
1989 return -EBUSY;
1990
1991 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1992 return -EBUSY;
1993
1994 return migrate_page(mapping, newpage, page, mode);
1995}
1996#endif
1997
1998int __init nfs_init_writepagecache(void)
1999{
2000 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2001 sizeof(struct nfs_pgio_header),
2002 0, SLAB_HWCACHE_ALIGN,
2003 NULL);
2004 if (nfs_wdata_cachep == NULL)
2005 return -ENOMEM;
2006
2007 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2008 nfs_wdata_cachep);
2009 if (nfs_wdata_mempool == NULL)
2010 goto out_destroy_write_cache;
2011
2012 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2013 sizeof(struct nfs_commit_data),
2014 0, SLAB_HWCACHE_ALIGN,
2015 NULL);
2016 if (nfs_cdata_cachep == NULL)
2017 goto out_destroy_write_mempool;
2018
2019 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2020 nfs_cdata_cachep);
2021 if (nfs_commit_mempool == NULL)
2022 goto out_destroy_commit_cache;
2023
2024 /*
2025 * NFS congestion size, scale with available memory.
2026 *
2027 * 64MB: 8192k
2028 * 128MB: 11585k
2029 * 256MB: 16384k
2030 * 512MB: 23170k
2031 * 1GB: 32768k
2032 * 2GB: 46340k
2033 * 4GB: 65536k
2034 * 8GB: 92681k
2035 * 16GB: 131072k
2036 *
2037 * This allows larger machines to have larger/more transfers.
2038 * Limit the default to 256M
2039 */
2040 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2041 if (nfs_congestion_kb > 256*1024)
2042 nfs_congestion_kb = 256*1024;
2043
2044 return 0;
2045
2046out_destroy_commit_cache:
2047 kmem_cache_destroy(nfs_cdata_cachep);
2048out_destroy_write_mempool:
2049 mempool_destroy(nfs_wdata_mempool);
2050out_destroy_write_cache:
2051 kmem_cache_destroy(nfs_wdata_cachep);
2052 return -ENOMEM;
2053}
2054
2055void nfs_destroy_writepagecache(void)
2056{
2057 mempool_destroy(nfs_commit_mempool);
2058 kmem_cache_destroy(nfs_cdata_cachep);
2059 mempool_destroy(nfs_wdata_mempool);
2060 kmem_cache_destroy(nfs_wdata_cachep);
2061}
2062
2063static const struct nfs_rw_ops nfs_rw_write_ops = {
2064 .rw_mode = FMODE_WRITE,
2065 .rw_alloc_header = nfs_writehdr_alloc,
2066 .rw_free_header = nfs_writehdr_free,
2067 .rw_done = nfs_writeback_done,
2068 .rw_result = nfs_writeback_result,
2069 .rw_initiate = nfs_initiate_write,
2070};