Loading...
1/*
2 * linux/fs/nfs/write.c
3 *
4 * Write file data over NFS.
5 *
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7 */
8
9#include <linux/types.h>
10#include <linux/slab.h>
11#include <linux/mm.h>
12#include <linux/pagemap.h>
13#include <linux/file.h>
14#include <linux/writeback.h>
15#include <linux/swap.h>
16#include <linux/migrate.h>
17
18#include <linux/sunrpc/clnt.h>
19#include <linux/nfs_fs.h>
20#include <linux/nfs_mount.h>
21#include <linux/nfs_page.h>
22#include <linux/backing-dev.h>
23#include <linux/export.h>
24
25#include <asm/uaccess.h>
26
27#include "delegation.h"
28#include "internal.h"
29#include "iostat.h"
30#include "nfs4_fs.h"
31#include "fscache.h"
32#include "pnfs.h"
33
34#define NFSDBG_FACILITY NFSDBG_PAGECACHE
35
36#define MIN_POOL_WRITE (32)
37#define MIN_POOL_COMMIT (4)
38
39/*
40 * Local function declarations
41 */
42static void nfs_redirty_request(struct nfs_page *req);
43static const struct rpc_call_ops nfs_write_common_ops;
44static const struct rpc_call_ops nfs_commit_ops;
45static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
46static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
47
48static struct kmem_cache *nfs_wdata_cachep;
49static mempool_t *nfs_wdata_mempool;
50static struct kmem_cache *nfs_cdata_cachep;
51static mempool_t *nfs_commit_mempool;
52
53struct nfs_commit_data *nfs_commitdata_alloc(void)
54{
55 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
56
57 if (p) {
58 memset(p, 0, sizeof(*p));
59 INIT_LIST_HEAD(&p->pages);
60 }
61 return p;
62}
63EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
64
65void nfs_commit_free(struct nfs_commit_data *p)
66{
67 mempool_free(p, nfs_commit_mempool);
68}
69EXPORT_SYMBOL_GPL(nfs_commit_free);
70
71struct nfs_write_header *nfs_writehdr_alloc(void)
72{
73 struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
74
75 if (p) {
76 struct nfs_pgio_header *hdr = &p->header;
77
78 memset(p, 0, sizeof(*p));
79 INIT_LIST_HEAD(&hdr->pages);
80 INIT_LIST_HEAD(&hdr->rpc_list);
81 spin_lock_init(&hdr->lock);
82 atomic_set(&hdr->refcnt, 0);
83 hdr->verf = &p->verf;
84 }
85 return p;
86}
87
88static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
89 unsigned int pagecount)
90{
91 struct nfs_write_data *data, *prealloc;
92
93 prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
94 if (prealloc->header == NULL)
95 data = prealloc;
96 else
97 data = kzalloc(sizeof(*data), GFP_KERNEL);
98 if (!data)
99 goto out;
100
101 if (nfs_pgarray_set(&data->pages, pagecount)) {
102 data->header = hdr;
103 atomic_inc(&hdr->refcnt);
104 } else {
105 if (data != prealloc)
106 kfree(data);
107 data = NULL;
108 }
109out:
110 return data;
111}
112
113void nfs_writehdr_free(struct nfs_pgio_header *hdr)
114{
115 struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
116 mempool_free(whdr, nfs_wdata_mempool);
117}
118
119void nfs_writedata_release(struct nfs_write_data *wdata)
120{
121 struct nfs_pgio_header *hdr = wdata->header;
122 struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
123
124 put_nfs_open_context(wdata->args.context);
125 if (wdata->pages.pagevec != wdata->pages.page_array)
126 kfree(wdata->pages.pagevec);
127 if (wdata != &write_header->rpc_data)
128 kfree(wdata);
129 else
130 wdata->header = NULL;
131 if (atomic_dec_and_test(&hdr->refcnt))
132 hdr->completion_ops->completion(hdr);
133}
134
135static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
136{
137 ctx->error = error;
138 smp_wmb();
139 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
140}
141
142static struct nfs_page *nfs_page_find_request_locked(struct page *page)
143{
144 struct nfs_page *req = NULL;
145
146 if (PagePrivate(page)) {
147 req = (struct nfs_page *)page_private(page);
148 if (req != NULL)
149 kref_get(&req->wb_kref);
150 }
151 return req;
152}
153
154static struct nfs_page *nfs_page_find_request(struct page *page)
155{
156 struct inode *inode = page->mapping->host;
157 struct nfs_page *req = NULL;
158
159 spin_lock(&inode->i_lock);
160 req = nfs_page_find_request_locked(page);
161 spin_unlock(&inode->i_lock);
162 return req;
163}
164
165/* Adjust the file length if we're writing beyond the end */
166static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
167{
168 struct inode *inode = page->mapping->host;
169 loff_t end, i_size;
170 pgoff_t end_index;
171
172 spin_lock(&inode->i_lock);
173 i_size = i_size_read(inode);
174 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
175 if (i_size > 0 && page->index < end_index)
176 goto out;
177 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
178 if (i_size >= end)
179 goto out;
180 i_size_write(inode, end);
181 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
182out:
183 spin_unlock(&inode->i_lock);
184}
185
186/* A writeback failed: mark the page as bad, and invalidate the page cache */
187static void nfs_set_pageerror(struct page *page)
188{
189 SetPageError(page);
190 nfs_zap_mapping(page->mapping->host, page->mapping);
191}
192
193/* We can set the PG_uptodate flag if we see that a write request
194 * covers the full page.
195 */
196static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
197{
198 if (PageUptodate(page))
199 return;
200 if (base != 0)
201 return;
202 if (count != nfs_page_length(page))
203 return;
204 SetPageUptodate(page);
205}
206
207static int wb_priority(struct writeback_control *wbc)
208{
209 if (wbc->for_reclaim)
210 return FLUSH_HIGHPRI | FLUSH_STABLE;
211 if (wbc->for_kupdate || wbc->for_background)
212 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
213 return FLUSH_COND_STABLE;
214}
215
216/*
217 * NFS congestion control
218 */
219
220int nfs_congestion_kb;
221
222#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
223#define NFS_CONGESTION_OFF_THRESH \
224 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
225
226static int nfs_set_page_writeback(struct page *page)
227{
228 int ret = test_set_page_writeback(page);
229
230 if (!ret) {
231 struct inode *inode = page->mapping->host;
232 struct nfs_server *nfss = NFS_SERVER(inode);
233
234 if (atomic_long_inc_return(&nfss->writeback) >
235 NFS_CONGESTION_ON_THRESH) {
236 set_bdi_congested(&nfss->backing_dev_info,
237 BLK_RW_ASYNC);
238 }
239 }
240 return ret;
241}
242
243static void nfs_end_page_writeback(struct page *page)
244{
245 struct inode *inode = page->mapping->host;
246 struct nfs_server *nfss = NFS_SERVER(inode);
247
248 end_page_writeback(page);
249 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
250 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
251}
252
253static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
254{
255 struct inode *inode = page->mapping->host;
256 struct nfs_page *req;
257 int ret;
258
259 spin_lock(&inode->i_lock);
260 for (;;) {
261 req = nfs_page_find_request_locked(page);
262 if (req == NULL)
263 break;
264 if (nfs_lock_request(req))
265 break;
266 /* Note: If we hold the page lock, as is the case in nfs_writepage,
267 * then the call to nfs_lock_request() will always
268 * succeed provided that someone hasn't already marked the
269 * request as dirty (in which case we don't care).
270 */
271 spin_unlock(&inode->i_lock);
272 if (!nonblock)
273 ret = nfs_wait_on_request(req);
274 else
275 ret = -EAGAIN;
276 nfs_release_request(req);
277 if (ret != 0)
278 return ERR_PTR(ret);
279 spin_lock(&inode->i_lock);
280 }
281 spin_unlock(&inode->i_lock);
282 return req;
283}
284
285/*
286 * Find an associated nfs write request, and prepare to flush it out
287 * May return an error if the user signalled nfs_wait_on_request().
288 */
289static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
290 struct page *page, bool nonblock)
291{
292 struct nfs_page *req;
293 int ret = 0;
294
295 req = nfs_find_and_lock_request(page, nonblock);
296 if (!req)
297 goto out;
298 ret = PTR_ERR(req);
299 if (IS_ERR(req))
300 goto out;
301
302 ret = nfs_set_page_writeback(page);
303 BUG_ON(ret != 0);
304 BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
305
306 if (!nfs_pageio_add_request(pgio, req)) {
307 nfs_redirty_request(req);
308 ret = pgio->pg_error;
309 }
310out:
311 return ret;
312}
313
314static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
315{
316 struct inode *inode = page->mapping->host;
317 int ret;
318
319 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
320 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
321
322 nfs_pageio_cond_complete(pgio, page->index);
323 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
324 if (ret == -EAGAIN) {
325 redirty_page_for_writepage(wbc, page);
326 ret = 0;
327 }
328 return ret;
329}
330
331/*
332 * Write an mmapped page to the server.
333 */
334static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
335{
336 struct nfs_pageio_descriptor pgio;
337 int err;
338
339 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
340 &nfs_async_write_completion_ops);
341 err = nfs_do_writepage(page, wbc, &pgio);
342 nfs_pageio_complete(&pgio);
343 if (err < 0)
344 return err;
345 if (pgio.pg_error < 0)
346 return pgio.pg_error;
347 return 0;
348}
349
350int nfs_writepage(struct page *page, struct writeback_control *wbc)
351{
352 int ret;
353
354 ret = nfs_writepage_locked(page, wbc);
355 unlock_page(page);
356 return ret;
357}
358
359static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
360{
361 int ret;
362
363 ret = nfs_do_writepage(page, wbc, data);
364 unlock_page(page);
365 return ret;
366}
367
368int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
369{
370 struct inode *inode = mapping->host;
371 unsigned long *bitlock = &NFS_I(inode)->flags;
372 struct nfs_pageio_descriptor pgio;
373 int err;
374
375 /* Stop dirtying of new pages while we sync */
376 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
377 nfs_wait_bit_killable, TASK_KILLABLE);
378 if (err)
379 goto out_err;
380
381 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
382
383 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc),
384 &nfs_async_write_completion_ops);
385 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
386 nfs_pageio_complete(&pgio);
387
388 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
389 smp_mb__after_clear_bit();
390 wake_up_bit(bitlock, NFS_INO_FLUSHING);
391
392 if (err < 0)
393 goto out_err;
394 err = pgio.pg_error;
395 if (err < 0)
396 goto out_err;
397 return 0;
398out_err:
399 return err;
400}
401
402/*
403 * Insert a write request into an inode
404 */
405static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
406{
407 struct nfs_inode *nfsi = NFS_I(inode);
408
409 /* Lock the request! */
410 nfs_lock_request(req);
411
412 spin_lock(&inode->i_lock);
413 if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
414 inode->i_version++;
415 set_bit(PG_MAPPED, &req->wb_flags);
416 SetPagePrivate(req->wb_page);
417 set_page_private(req->wb_page, (unsigned long)req);
418 nfsi->npages++;
419 kref_get(&req->wb_kref);
420 spin_unlock(&inode->i_lock);
421}
422
423/*
424 * Remove a write request from an inode
425 */
426static void nfs_inode_remove_request(struct nfs_page *req)
427{
428 struct inode *inode = req->wb_context->dentry->d_inode;
429 struct nfs_inode *nfsi = NFS_I(inode);
430
431 BUG_ON (!NFS_WBACK_BUSY(req));
432
433 spin_lock(&inode->i_lock);
434 set_page_private(req->wb_page, 0);
435 ClearPagePrivate(req->wb_page);
436 clear_bit(PG_MAPPED, &req->wb_flags);
437 nfsi->npages--;
438 spin_unlock(&inode->i_lock);
439 nfs_release_request(req);
440}
441
442static void
443nfs_mark_request_dirty(struct nfs_page *req)
444{
445 __set_page_dirty_nobuffers(req->wb_page);
446}
447
448#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
449/**
450 * nfs_request_add_commit_list - add request to a commit list
451 * @req: pointer to a struct nfs_page
452 * @dst: commit list head
453 * @cinfo: holds list lock and accounting info
454 *
455 * This sets the PG_CLEAN bit, updates the cinfo count of
456 * number of outstanding requests requiring a commit as well as
457 * the MM page stats.
458 *
459 * The caller must _not_ hold the cinfo->lock, but must be
460 * holding the nfs_page lock.
461 */
462void
463nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
464 struct nfs_commit_info *cinfo)
465{
466 set_bit(PG_CLEAN, &(req)->wb_flags);
467 spin_lock(cinfo->lock);
468 nfs_list_add_request(req, dst);
469 cinfo->mds->ncommit++;
470 spin_unlock(cinfo->lock);
471 if (!cinfo->dreq) {
472 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
473 inc_bdi_stat(req->wb_page->mapping->backing_dev_info,
474 BDI_RECLAIMABLE);
475 __mark_inode_dirty(req->wb_context->dentry->d_inode,
476 I_DIRTY_DATASYNC);
477 }
478}
479EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
480
481/**
482 * nfs_request_remove_commit_list - Remove request from a commit list
483 * @req: pointer to a nfs_page
484 * @cinfo: holds list lock and accounting info
485 *
486 * This clears the PG_CLEAN bit, and updates the cinfo's count of
487 * number of outstanding requests requiring a commit
488 * It does not update the MM page stats.
489 *
490 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
491 */
492void
493nfs_request_remove_commit_list(struct nfs_page *req,
494 struct nfs_commit_info *cinfo)
495{
496 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
497 return;
498 nfs_list_remove_request(req);
499 cinfo->mds->ncommit--;
500}
501EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
502
503static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
504 struct inode *inode)
505{
506 cinfo->lock = &inode->i_lock;
507 cinfo->mds = &NFS_I(inode)->commit_info;
508 cinfo->ds = pnfs_get_ds_info(inode);
509 cinfo->dreq = NULL;
510 cinfo->completion_ops = &nfs_commit_completion_ops;
511}
512
513void nfs_init_cinfo(struct nfs_commit_info *cinfo,
514 struct inode *inode,
515 struct nfs_direct_req *dreq)
516{
517 if (dreq)
518 nfs_init_cinfo_from_dreq(cinfo, dreq);
519 else
520 nfs_init_cinfo_from_inode(cinfo, inode);
521}
522EXPORT_SYMBOL_GPL(nfs_init_cinfo);
523
524/*
525 * Add a request to the inode's commit list.
526 */
527void
528nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
529 struct nfs_commit_info *cinfo)
530{
531 if (pnfs_mark_request_commit(req, lseg, cinfo))
532 return;
533 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
534}
535
536static void
537nfs_clear_page_commit(struct page *page)
538{
539 dec_zone_page_state(page, NR_UNSTABLE_NFS);
540 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
541}
542
543static void
544nfs_clear_request_commit(struct nfs_page *req)
545{
546 if (test_bit(PG_CLEAN, &req->wb_flags)) {
547 struct inode *inode = req->wb_context->dentry->d_inode;
548 struct nfs_commit_info cinfo;
549
550 nfs_init_cinfo_from_inode(&cinfo, inode);
551 if (!pnfs_clear_request_commit(req, &cinfo)) {
552 spin_lock(cinfo.lock);
553 nfs_request_remove_commit_list(req, &cinfo);
554 spin_unlock(cinfo.lock);
555 }
556 nfs_clear_page_commit(req->wb_page);
557 }
558}
559
560static inline
561int nfs_write_need_commit(struct nfs_write_data *data)
562{
563 if (data->verf.committed == NFS_DATA_SYNC)
564 return data->header->lseg == NULL;
565 return data->verf.committed != NFS_FILE_SYNC;
566}
567
568#else
569static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
570 struct inode *inode)
571{
572}
573
574void nfs_init_cinfo(struct nfs_commit_info *cinfo,
575 struct inode *inode,
576 struct nfs_direct_req *dreq)
577{
578}
579
580void
581nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
582 struct nfs_commit_info *cinfo)
583{
584}
585
586static void
587nfs_clear_request_commit(struct nfs_page *req)
588{
589}
590
591static inline
592int nfs_write_need_commit(struct nfs_write_data *data)
593{
594 return 0;
595}
596
597#endif
598
599static void nfs_write_completion(struct nfs_pgio_header *hdr)
600{
601 struct nfs_commit_info cinfo;
602 unsigned long bytes = 0;
603
604 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
605 goto out;
606 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
607 while (!list_empty(&hdr->pages)) {
608 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
609
610 bytes += req->wb_bytes;
611 nfs_list_remove_request(req);
612 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
613 (hdr->good_bytes < bytes)) {
614 nfs_set_pageerror(req->wb_page);
615 nfs_context_set_write_error(req->wb_context, hdr->error);
616 goto remove_req;
617 }
618 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
619 nfs_mark_request_dirty(req);
620 goto next;
621 }
622 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
623 memcpy(&req->wb_verf, hdr->verf, sizeof(req->wb_verf));
624 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
625 goto next;
626 }
627remove_req:
628 nfs_inode_remove_request(req);
629next:
630 nfs_unlock_request(req);
631 nfs_end_page_writeback(req->wb_page);
632 nfs_release_request(req);
633 }
634out:
635 hdr->release(hdr);
636}
637
638#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
639static unsigned long
640nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
641{
642 return cinfo->mds->ncommit;
643}
644
645/* cinfo->lock held by caller */
646int
647nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
648 struct nfs_commit_info *cinfo, int max)
649{
650 struct nfs_page *req, *tmp;
651 int ret = 0;
652
653 list_for_each_entry_safe(req, tmp, src, wb_list) {
654 if (!nfs_lock_request(req))
655 continue;
656 kref_get(&req->wb_kref);
657 if (cond_resched_lock(cinfo->lock))
658 list_safe_reset_next(req, tmp, wb_list);
659 nfs_request_remove_commit_list(req, cinfo);
660 nfs_list_add_request(req, dst);
661 ret++;
662 if ((ret == max) && !cinfo->dreq)
663 break;
664 }
665 return ret;
666}
667
668/*
669 * nfs_scan_commit - Scan an inode for commit requests
670 * @inode: NFS inode to scan
671 * @dst: mds destination list
672 * @cinfo: mds and ds lists of reqs ready to commit
673 *
674 * Moves requests from the inode's 'commit' request list.
675 * The requests are *not* checked to ensure that they form a contiguous set.
676 */
677int
678nfs_scan_commit(struct inode *inode, struct list_head *dst,
679 struct nfs_commit_info *cinfo)
680{
681 int ret = 0;
682
683 spin_lock(cinfo->lock);
684 if (cinfo->mds->ncommit > 0) {
685 const int max = INT_MAX;
686
687 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
688 cinfo, max);
689 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
690 }
691 spin_unlock(cinfo->lock);
692 return ret;
693}
694
695#else
696static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
697{
698 return 0;
699}
700
701int nfs_scan_commit(struct inode *inode, struct list_head *dst,
702 struct nfs_commit_info *cinfo)
703{
704 return 0;
705}
706#endif
707
708/*
709 * Search for an existing write request, and attempt to update
710 * it to reflect a new dirty region on a given page.
711 *
712 * If the attempt fails, then the existing request is flushed out
713 * to disk.
714 */
715static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
716 struct page *page,
717 unsigned int offset,
718 unsigned int bytes)
719{
720 struct nfs_page *req;
721 unsigned int rqend;
722 unsigned int end;
723 int error;
724
725 if (!PagePrivate(page))
726 return NULL;
727
728 end = offset + bytes;
729 spin_lock(&inode->i_lock);
730
731 for (;;) {
732 req = nfs_page_find_request_locked(page);
733 if (req == NULL)
734 goto out_unlock;
735
736 rqend = req->wb_offset + req->wb_bytes;
737 /*
738 * Tell the caller to flush out the request if
739 * the offsets are non-contiguous.
740 * Note: nfs_flush_incompatible() will already
741 * have flushed out requests having wrong owners.
742 */
743 if (offset > rqend
744 || end < req->wb_offset)
745 goto out_flushme;
746
747 if (nfs_lock_request(req))
748 break;
749
750 /* The request is locked, so wait and then retry */
751 spin_unlock(&inode->i_lock);
752 error = nfs_wait_on_request(req);
753 nfs_release_request(req);
754 if (error != 0)
755 goto out_err;
756 spin_lock(&inode->i_lock);
757 }
758
759 /* Okay, the request matches. Update the region */
760 if (offset < req->wb_offset) {
761 req->wb_offset = offset;
762 req->wb_pgbase = offset;
763 }
764 if (end > rqend)
765 req->wb_bytes = end - req->wb_offset;
766 else
767 req->wb_bytes = rqend - req->wb_offset;
768out_unlock:
769 spin_unlock(&inode->i_lock);
770 if (req)
771 nfs_clear_request_commit(req);
772 return req;
773out_flushme:
774 spin_unlock(&inode->i_lock);
775 nfs_release_request(req);
776 error = nfs_wb_page(inode, page);
777out_err:
778 return ERR_PTR(error);
779}
780
781/*
782 * Try to update an existing write request, or create one if there is none.
783 *
784 * Note: Should always be called with the Page Lock held to prevent races
785 * if we have to add a new request. Also assumes that the caller has
786 * already called nfs_flush_incompatible() if necessary.
787 */
788static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
789 struct page *page, unsigned int offset, unsigned int bytes)
790{
791 struct inode *inode = page->mapping->host;
792 struct nfs_page *req;
793
794 req = nfs_try_to_update_request(inode, page, offset, bytes);
795 if (req != NULL)
796 goto out;
797 req = nfs_create_request(ctx, inode, page, offset, bytes);
798 if (IS_ERR(req))
799 goto out;
800 nfs_inode_add_request(inode, req);
801out:
802 return req;
803}
804
805static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
806 unsigned int offset, unsigned int count)
807{
808 struct nfs_page *req;
809
810 req = nfs_setup_write_request(ctx, page, offset, count);
811 if (IS_ERR(req))
812 return PTR_ERR(req);
813 /* Update file length */
814 nfs_grow_file(page, offset, count);
815 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
816 nfs_mark_request_dirty(req);
817 nfs_unlock_and_release_request(req);
818 return 0;
819}
820
821int nfs_flush_incompatible(struct file *file, struct page *page)
822{
823 struct nfs_open_context *ctx = nfs_file_open_context(file);
824 struct nfs_page *req;
825 int do_flush, status;
826 /*
827 * Look for a request corresponding to this page. If there
828 * is one, and it belongs to another file, we flush it out
829 * before we try to copy anything into the page. Do this
830 * due to the lack of an ACCESS-type call in NFSv2.
831 * Also do the same if we find a request from an existing
832 * dropped page.
833 */
834 do {
835 req = nfs_page_find_request(page);
836 if (req == NULL)
837 return 0;
838 do_flush = req->wb_page != page || req->wb_context != ctx ||
839 req->wb_lock_context->lockowner != current->files ||
840 req->wb_lock_context->pid != current->tgid;
841 nfs_release_request(req);
842 if (!do_flush)
843 return 0;
844 status = nfs_wb_page(page->mapping->host, page);
845 } while (status == 0);
846 return status;
847}
848
849/*
850 * If the page cache is marked as unsafe or invalid, then we can't rely on
851 * the PageUptodate() flag. In this case, we will need to turn off
852 * write optimisations that depend on the page contents being correct.
853 */
854static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
855{
856 if (nfs_have_delegated_attributes(inode))
857 goto out;
858 if (NFS_I(inode)->cache_validity & NFS_INO_REVAL_PAGECACHE)
859 return false;
860out:
861 return PageUptodate(page) != 0;
862}
863
864/*
865 * Update and possibly write a cached page of an NFS file.
866 *
867 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
868 * things with a page scheduled for an RPC call (e.g. invalidate it).
869 */
870int nfs_updatepage(struct file *file, struct page *page,
871 unsigned int offset, unsigned int count)
872{
873 struct nfs_open_context *ctx = nfs_file_open_context(file);
874 struct inode *inode = page->mapping->host;
875 int status = 0;
876
877 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
878
879 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
880 file->f_path.dentry->d_parent->d_name.name,
881 file->f_path.dentry->d_name.name, count,
882 (long long)(page_offset(page) + offset));
883
884 /* If we're not using byte range locks, and we know the page
885 * is up to date, it may be more efficient to extend the write
886 * to cover the entire page in order to avoid fragmentation
887 * inefficiencies.
888 */
889 if (nfs_write_pageuptodate(page, inode) &&
890 inode->i_flock == NULL &&
891 !(file->f_flags & O_DSYNC)) {
892 count = max(count + offset, nfs_page_length(page));
893 offset = 0;
894 }
895
896 status = nfs_writepage_setup(ctx, page, offset, count);
897 if (status < 0)
898 nfs_set_pageerror(page);
899 else
900 __set_page_dirty_nobuffers(page);
901
902 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
903 status, (long long)i_size_read(inode));
904 return status;
905}
906
907static int flush_task_priority(int how)
908{
909 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
910 case FLUSH_HIGHPRI:
911 return RPC_PRIORITY_HIGH;
912 case FLUSH_LOWPRI:
913 return RPC_PRIORITY_LOW;
914 }
915 return RPC_PRIORITY_NORMAL;
916}
917
918int nfs_initiate_write(struct rpc_clnt *clnt,
919 struct nfs_write_data *data,
920 const struct rpc_call_ops *call_ops,
921 int how, int flags)
922{
923 struct inode *inode = data->header->inode;
924 int priority = flush_task_priority(how);
925 struct rpc_task *task;
926 struct rpc_message msg = {
927 .rpc_argp = &data->args,
928 .rpc_resp = &data->res,
929 .rpc_cred = data->header->cred,
930 };
931 struct rpc_task_setup task_setup_data = {
932 .rpc_client = clnt,
933 .task = &data->task,
934 .rpc_message = &msg,
935 .callback_ops = call_ops,
936 .callback_data = data,
937 .workqueue = nfsiod_workqueue,
938 .flags = RPC_TASK_ASYNC | flags,
939 .priority = priority,
940 };
941 int ret = 0;
942
943 /* Set up the initial task struct. */
944 NFS_PROTO(inode)->write_setup(data, &msg);
945
946 dprintk("NFS: %5u initiated write call "
947 "(req %s/%lld, %u bytes @ offset %llu)\n",
948 data->task.tk_pid,
949 inode->i_sb->s_id,
950 (long long)NFS_FILEID(inode),
951 data->args.count,
952 (unsigned long long)data->args.offset);
953
954 task = rpc_run_task(&task_setup_data);
955 if (IS_ERR(task)) {
956 ret = PTR_ERR(task);
957 goto out;
958 }
959 if (how & FLUSH_SYNC) {
960 ret = rpc_wait_for_completion_task(task);
961 if (ret == 0)
962 ret = task->tk_status;
963 }
964 rpc_put_task(task);
965out:
966 return ret;
967}
968EXPORT_SYMBOL_GPL(nfs_initiate_write);
969
970/*
971 * Set up the argument/result storage required for the RPC call.
972 */
973static void nfs_write_rpcsetup(struct nfs_write_data *data,
974 unsigned int count, unsigned int offset,
975 int how, struct nfs_commit_info *cinfo)
976{
977 struct nfs_page *req = data->header->req;
978
979 /* Set up the RPC argument and reply structs
980 * NB: take care not to mess about with data->commit et al. */
981
982 data->args.fh = NFS_FH(data->header->inode);
983 data->args.offset = req_offset(req) + offset;
984 /* pnfs_set_layoutcommit needs this */
985 data->mds_offset = data->args.offset;
986 data->args.pgbase = req->wb_pgbase + offset;
987 data->args.pages = data->pages.pagevec;
988 data->args.count = count;
989 data->args.context = get_nfs_open_context(req->wb_context);
990 data->args.lock_context = req->wb_lock_context;
991 data->args.stable = NFS_UNSTABLE;
992 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
993 case 0:
994 break;
995 case FLUSH_COND_STABLE:
996 if (nfs_reqs_to_commit(cinfo))
997 break;
998 default:
999 data->args.stable = NFS_FILE_SYNC;
1000 }
1001
1002 data->res.fattr = &data->fattr;
1003 data->res.count = count;
1004 data->res.verf = &data->verf;
1005 nfs_fattr_init(&data->fattr);
1006}
1007
1008static int nfs_do_write(struct nfs_write_data *data,
1009 const struct rpc_call_ops *call_ops,
1010 int how)
1011{
1012 struct inode *inode = data->header->inode;
1013
1014 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1015}
1016
1017static int nfs_do_multiple_writes(struct list_head *head,
1018 const struct rpc_call_ops *call_ops,
1019 int how)
1020{
1021 struct nfs_write_data *data;
1022 int ret = 0;
1023
1024 while (!list_empty(head)) {
1025 int ret2;
1026
1027 data = list_first_entry(head, struct nfs_write_data, list);
1028 list_del_init(&data->list);
1029
1030 ret2 = nfs_do_write(data, call_ops, how);
1031 if (ret == 0)
1032 ret = ret2;
1033 }
1034 return ret;
1035}
1036
1037/* If a nfs_flush_* function fails, it should remove reqs from @head and
1038 * call this on each, which will prepare them to be retried on next
1039 * writeback using standard nfs.
1040 */
1041static void nfs_redirty_request(struct nfs_page *req)
1042{
1043 nfs_mark_request_dirty(req);
1044 nfs_unlock_request(req);
1045 nfs_end_page_writeback(req->wb_page);
1046 nfs_release_request(req);
1047}
1048
1049static void nfs_async_write_error(struct list_head *head)
1050{
1051 struct nfs_page *req;
1052
1053 while (!list_empty(head)) {
1054 req = nfs_list_entry(head->next);
1055 nfs_list_remove_request(req);
1056 nfs_redirty_request(req);
1057 }
1058}
1059
1060static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1061 .error_cleanup = nfs_async_write_error,
1062 .completion = nfs_write_completion,
1063};
1064
1065static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1066 struct nfs_pgio_header *hdr)
1067{
1068 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1069 while (!list_empty(&hdr->rpc_list)) {
1070 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1071 struct nfs_write_data, list);
1072 list_del(&data->list);
1073 nfs_writedata_release(data);
1074 }
1075 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1076}
1077
1078/*
1079 * Generate multiple small requests to write out a single
1080 * contiguous dirty area on one page.
1081 */
1082static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1083 struct nfs_pgio_header *hdr)
1084{
1085 struct nfs_page *req = hdr->req;
1086 struct page *page = req->wb_page;
1087 struct nfs_write_data *data;
1088 size_t wsize = desc->pg_bsize, nbytes;
1089 unsigned int offset;
1090 int requests = 0;
1091 struct nfs_commit_info cinfo;
1092
1093 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1094
1095 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1096 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1097 desc->pg_count > wsize))
1098 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1099
1100
1101 offset = 0;
1102 nbytes = desc->pg_count;
1103 do {
1104 size_t len = min(nbytes, wsize);
1105
1106 data = nfs_writedata_alloc(hdr, 1);
1107 if (!data) {
1108 nfs_flush_error(desc, hdr);
1109 return -ENOMEM;
1110 }
1111 data->pages.pagevec[0] = page;
1112 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1113 list_add(&data->list, &hdr->rpc_list);
1114 requests++;
1115 nbytes -= len;
1116 offset += len;
1117 } while (nbytes != 0);
1118 nfs_list_remove_request(req);
1119 nfs_list_add_request(req, &hdr->pages);
1120 desc->pg_rpc_callops = &nfs_write_common_ops;
1121 return 0;
1122}
1123
1124/*
1125 * Create an RPC task for the given write request and kick it.
1126 * The page must have been locked by the caller.
1127 *
1128 * It may happen that the page we're passed is not marked dirty.
1129 * This is the case if nfs_updatepage detects a conflicting request
1130 * that has been written but not committed.
1131 */
1132static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1133 struct nfs_pgio_header *hdr)
1134{
1135 struct nfs_page *req;
1136 struct page **pages;
1137 struct nfs_write_data *data;
1138 struct list_head *head = &desc->pg_list;
1139 struct nfs_commit_info cinfo;
1140
1141 data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1142 desc->pg_count));
1143 if (!data) {
1144 nfs_flush_error(desc, hdr);
1145 return -ENOMEM;
1146 }
1147
1148 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1149 pages = data->pages.pagevec;
1150 while (!list_empty(head)) {
1151 req = nfs_list_entry(head->next);
1152 nfs_list_remove_request(req);
1153 nfs_list_add_request(req, &hdr->pages);
1154 *pages++ = req->wb_page;
1155 }
1156
1157 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1158 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1159 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1160
1161 /* Set up the argument struct */
1162 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1163 list_add(&data->list, &hdr->rpc_list);
1164 desc->pg_rpc_callops = &nfs_write_common_ops;
1165 return 0;
1166}
1167
1168int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1169 struct nfs_pgio_header *hdr)
1170{
1171 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1172 return nfs_flush_multi(desc, hdr);
1173 return nfs_flush_one(desc, hdr);
1174}
1175
1176static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1177{
1178 struct nfs_write_header *whdr;
1179 struct nfs_pgio_header *hdr;
1180 int ret;
1181
1182 whdr = nfs_writehdr_alloc();
1183 if (!whdr) {
1184 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1185 return -ENOMEM;
1186 }
1187 hdr = &whdr->header;
1188 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1189 atomic_inc(&hdr->refcnt);
1190 ret = nfs_generic_flush(desc, hdr);
1191 if (ret == 0)
1192 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1193 desc->pg_rpc_callops,
1194 desc->pg_ioflags);
1195 if (atomic_dec_and_test(&hdr->refcnt))
1196 hdr->completion_ops->completion(hdr);
1197 return ret;
1198}
1199
1200static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1201 .pg_test = nfs_generic_pg_test,
1202 .pg_doio = nfs_generic_pg_writepages,
1203};
1204
1205void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1206 struct inode *inode, int ioflags,
1207 const struct nfs_pgio_completion_ops *compl_ops)
1208{
1209 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1210 NFS_SERVER(inode)->wsize, ioflags);
1211}
1212
1213void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1214{
1215 pgio->pg_ops = &nfs_pageio_write_ops;
1216 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1217}
1218EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1219
1220void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1221 struct inode *inode, int ioflags,
1222 const struct nfs_pgio_completion_ops *compl_ops)
1223{
1224 if (!pnfs_pageio_init_write(pgio, inode, ioflags, compl_ops))
1225 nfs_pageio_init_write_mds(pgio, inode, ioflags, compl_ops);
1226}
1227
1228void nfs_write_prepare(struct rpc_task *task, void *calldata)
1229{
1230 struct nfs_write_data *data = calldata;
1231 NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1232}
1233
1234void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1235{
1236 struct nfs_commit_data *data = calldata;
1237
1238 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1239}
1240
1241/*
1242 * Handle a write reply that flushes a whole page.
1243 *
1244 * FIXME: There is an inherent race with invalidate_inode_pages and
1245 * writebacks since the page->count is kept > 1 for as long
1246 * as the page has a write request pending.
1247 */
1248static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1249{
1250 struct nfs_write_data *data = calldata;
1251
1252 nfs_writeback_done(task, data);
1253}
1254
1255static void nfs_writeback_release_common(void *calldata)
1256{
1257 struct nfs_write_data *data = calldata;
1258 struct nfs_pgio_header *hdr = data->header;
1259 int status = data->task.tk_status;
1260
1261 if ((status >= 0) && nfs_write_need_commit(data)) {
1262 spin_lock(&hdr->lock);
1263 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1264 ; /* Do nothing */
1265 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1266 memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1267 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1268 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1269 spin_unlock(&hdr->lock);
1270 }
1271 nfs_writedata_release(data);
1272}
1273
1274static const struct rpc_call_ops nfs_write_common_ops = {
1275 .rpc_call_prepare = nfs_write_prepare,
1276 .rpc_call_done = nfs_writeback_done_common,
1277 .rpc_release = nfs_writeback_release_common,
1278};
1279
1280
1281/*
1282 * This function is called when the WRITE call is complete.
1283 */
1284void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1285{
1286 struct nfs_writeargs *argp = &data->args;
1287 struct nfs_writeres *resp = &data->res;
1288 struct inode *inode = data->header->inode;
1289 int status;
1290
1291 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1292 task->tk_pid, task->tk_status);
1293
1294 /*
1295 * ->write_done will attempt to use post-op attributes to detect
1296 * conflicting writes by other clients. A strict interpretation
1297 * of close-to-open would allow us to continue caching even if
1298 * another writer had changed the file, but some applications
1299 * depend on tighter cache coherency when writing.
1300 */
1301 status = NFS_PROTO(inode)->write_done(task, data);
1302 if (status != 0)
1303 return;
1304 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1305
1306#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1307 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1308 /* We tried a write call, but the server did not
1309 * commit data to stable storage even though we
1310 * requested it.
1311 * Note: There is a known bug in Tru64 < 5.0 in which
1312 * the server reports NFS_DATA_SYNC, but performs
1313 * NFS_FILE_SYNC. We therefore implement this checking
1314 * as a dprintk() in order to avoid filling syslog.
1315 */
1316 static unsigned long complain;
1317
1318 /* Note this will print the MDS for a DS write */
1319 if (time_before(complain, jiffies)) {
1320 dprintk("NFS: faulty NFS server %s:"
1321 " (committed = %d) != (stable = %d)\n",
1322 NFS_SERVER(inode)->nfs_client->cl_hostname,
1323 resp->verf->committed, argp->stable);
1324 complain = jiffies + 300 * HZ;
1325 }
1326 }
1327#endif
1328 if (task->tk_status < 0)
1329 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1330 else if (resp->count < argp->count) {
1331 static unsigned long complain;
1332
1333 /* This a short write! */
1334 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1335
1336 /* Has the server at least made some progress? */
1337 if (resp->count == 0) {
1338 if (time_before(complain, jiffies)) {
1339 printk(KERN_WARNING
1340 "NFS: Server wrote zero bytes, expected %u.\n",
1341 argp->count);
1342 complain = jiffies + 300 * HZ;
1343 }
1344 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1345 task->tk_status = -EIO;
1346 return;
1347 }
1348 /* Was this an NFSv2 write or an NFSv3 stable write? */
1349 if (resp->verf->committed != NFS_UNSTABLE) {
1350 /* Resend from where the server left off */
1351 data->mds_offset += resp->count;
1352 argp->offset += resp->count;
1353 argp->pgbase += resp->count;
1354 argp->count -= resp->count;
1355 } else {
1356 /* Resend as a stable write in order to avoid
1357 * headaches in the case of a server crash.
1358 */
1359 argp->stable = NFS_FILE_SYNC;
1360 }
1361 rpc_restart_call_prepare(task);
1362 }
1363}
1364
1365
1366#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1367static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1368{
1369 int ret;
1370
1371 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1372 return 1;
1373 if (!may_wait)
1374 return 0;
1375 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1376 NFS_INO_COMMIT,
1377 nfs_wait_bit_killable,
1378 TASK_KILLABLE);
1379 return (ret < 0) ? ret : 1;
1380}
1381
1382static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1383{
1384 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1385 smp_mb__after_clear_bit();
1386 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1387}
1388
1389void nfs_commitdata_release(struct nfs_commit_data *data)
1390{
1391 put_nfs_open_context(data->context);
1392 nfs_commit_free(data);
1393}
1394EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1395
1396int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1397 const struct rpc_call_ops *call_ops,
1398 int how, int flags)
1399{
1400 struct rpc_task *task;
1401 int priority = flush_task_priority(how);
1402 struct rpc_message msg = {
1403 .rpc_argp = &data->args,
1404 .rpc_resp = &data->res,
1405 .rpc_cred = data->cred,
1406 };
1407 struct rpc_task_setup task_setup_data = {
1408 .task = &data->task,
1409 .rpc_client = clnt,
1410 .rpc_message = &msg,
1411 .callback_ops = call_ops,
1412 .callback_data = data,
1413 .workqueue = nfsiod_workqueue,
1414 .flags = RPC_TASK_ASYNC | flags,
1415 .priority = priority,
1416 };
1417 /* Set up the initial task struct. */
1418 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1419
1420 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1421
1422 task = rpc_run_task(&task_setup_data);
1423 if (IS_ERR(task))
1424 return PTR_ERR(task);
1425 if (how & FLUSH_SYNC)
1426 rpc_wait_for_completion_task(task);
1427 rpc_put_task(task);
1428 return 0;
1429}
1430EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1431
1432/*
1433 * Set up the argument/result storage required for the RPC call.
1434 */
1435void nfs_init_commit(struct nfs_commit_data *data,
1436 struct list_head *head,
1437 struct pnfs_layout_segment *lseg,
1438 struct nfs_commit_info *cinfo)
1439{
1440 struct nfs_page *first = nfs_list_entry(head->next);
1441 struct inode *inode = first->wb_context->dentry->d_inode;
1442
1443 /* Set up the RPC argument and reply structs
1444 * NB: take care not to mess about with data->commit et al. */
1445
1446 list_splice_init(head, &data->pages);
1447
1448 data->inode = inode;
1449 data->cred = first->wb_context->cred;
1450 data->lseg = lseg; /* reference transferred */
1451 data->mds_ops = &nfs_commit_ops;
1452 data->completion_ops = cinfo->completion_ops;
1453 data->dreq = cinfo->dreq;
1454
1455 data->args.fh = NFS_FH(data->inode);
1456 /* Note: we always request a commit of the entire inode */
1457 data->args.offset = 0;
1458 data->args.count = 0;
1459 data->context = get_nfs_open_context(first->wb_context);
1460 data->res.fattr = &data->fattr;
1461 data->res.verf = &data->verf;
1462 nfs_fattr_init(&data->fattr);
1463}
1464EXPORT_SYMBOL_GPL(nfs_init_commit);
1465
1466void nfs_retry_commit(struct list_head *page_list,
1467 struct pnfs_layout_segment *lseg,
1468 struct nfs_commit_info *cinfo)
1469{
1470 struct nfs_page *req;
1471
1472 while (!list_empty(page_list)) {
1473 req = nfs_list_entry(page_list->next);
1474 nfs_list_remove_request(req);
1475 nfs_mark_request_commit(req, lseg, cinfo);
1476 if (!cinfo->dreq) {
1477 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1478 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1479 BDI_RECLAIMABLE);
1480 }
1481 nfs_unlock_and_release_request(req);
1482 }
1483}
1484EXPORT_SYMBOL_GPL(nfs_retry_commit);
1485
1486/*
1487 * Commit dirty pages
1488 */
1489static int
1490nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1491 struct nfs_commit_info *cinfo)
1492{
1493 struct nfs_commit_data *data;
1494
1495 data = nfs_commitdata_alloc();
1496
1497 if (!data)
1498 goto out_bad;
1499
1500 /* Set up the argument struct */
1501 nfs_init_commit(data, head, NULL, cinfo);
1502 atomic_inc(&cinfo->mds->rpcs_out);
1503 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1504 how, 0);
1505 out_bad:
1506 nfs_retry_commit(head, NULL, cinfo);
1507 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1508 return -ENOMEM;
1509}
1510
1511/*
1512 * COMMIT call returned
1513 */
1514static void nfs_commit_done(struct rpc_task *task, void *calldata)
1515{
1516 struct nfs_commit_data *data = calldata;
1517
1518 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1519 task->tk_pid, task->tk_status);
1520
1521 /* Call the NFS version-specific code */
1522 NFS_PROTO(data->inode)->commit_done(task, data);
1523}
1524
1525static void nfs_commit_release_pages(struct nfs_commit_data *data)
1526{
1527 struct nfs_page *req;
1528 int status = data->task.tk_status;
1529 struct nfs_commit_info cinfo;
1530
1531 while (!list_empty(&data->pages)) {
1532 req = nfs_list_entry(data->pages.next);
1533 nfs_list_remove_request(req);
1534 nfs_clear_page_commit(req->wb_page);
1535
1536 dprintk("NFS: commit (%s/%lld %d@%lld)",
1537 req->wb_context->dentry->d_sb->s_id,
1538 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1539 req->wb_bytes,
1540 (long long)req_offset(req));
1541 if (status < 0) {
1542 nfs_context_set_write_error(req->wb_context, status);
1543 nfs_inode_remove_request(req);
1544 dprintk(", error = %d\n", status);
1545 goto next;
1546 }
1547
1548 /* Okay, COMMIT succeeded, apparently. Check the verifier
1549 * returned by the server against all stored verfs. */
1550 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1551 /* We have a match */
1552 nfs_inode_remove_request(req);
1553 dprintk(" OK\n");
1554 goto next;
1555 }
1556 /* We have a mismatch. Write the page again */
1557 dprintk(" mismatch\n");
1558 nfs_mark_request_dirty(req);
1559 next:
1560 nfs_unlock_and_release_request(req);
1561 }
1562 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1563 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1564 nfs_commit_clear_lock(NFS_I(data->inode));
1565}
1566
1567static void nfs_commit_release(void *calldata)
1568{
1569 struct nfs_commit_data *data = calldata;
1570
1571 data->completion_ops->completion(data);
1572 nfs_commitdata_release(calldata);
1573}
1574
1575static const struct rpc_call_ops nfs_commit_ops = {
1576 .rpc_call_prepare = nfs_commit_prepare,
1577 .rpc_call_done = nfs_commit_done,
1578 .rpc_release = nfs_commit_release,
1579};
1580
1581static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1582 .completion = nfs_commit_release_pages,
1583 .error_cleanup = nfs_commit_clear_lock,
1584};
1585
1586int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1587 int how, struct nfs_commit_info *cinfo)
1588{
1589 int status;
1590
1591 status = pnfs_commit_list(inode, head, how, cinfo);
1592 if (status == PNFS_NOT_ATTEMPTED)
1593 status = nfs_commit_list(inode, head, how, cinfo);
1594 return status;
1595}
1596
1597int nfs_commit_inode(struct inode *inode, int how)
1598{
1599 LIST_HEAD(head);
1600 struct nfs_commit_info cinfo;
1601 int may_wait = how & FLUSH_SYNC;
1602 int res;
1603
1604 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1605 if (res <= 0)
1606 goto out_mark_dirty;
1607 nfs_init_cinfo_from_inode(&cinfo, inode);
1608 res = nfs_scan_commit(inode, &head, &cinfo);
1609 if (res) {
1610 int error;
1611
1612 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1613 if (error < 0)
1614 return error;
1615 if (!may_wait)
1616 goto out_mark_dirty;
1617 error = wait_on_bit(&NFS_I(inode)->flags,
1618 NFS_INO_COMMIT,
1619 nfs_wait_bit_killable,
1620 TASK_KILLABLE);
1621 if (error < 0)
1622 return error;
1623 } else
1624 nfs_commit_clear_lock(NFS_I(inode));
1625 return res;
1626 /* Note: If we exit without ensuring that the commit is complete,
1627 * we must mark the inode as dirty. Otherwise, future calls to
1628 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1629 * that the data is on the disk.
1630 */
1631out_mark_dirty:
1632 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1633 return res;
1634}
1635
1636static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1637{
1638 struct nfs_inode *nfsi = NFS_I(inode);
1639 int flags = FLUSH_SYNC;
1640 int ret = 0;
1641
1642 /* no commits means nothing needs to be done */
1643 if (!nfsi->commit_info.ncommit)
1644 return ret;
1645
1646 if (wbc->sync_mode == WB_SYNC_NONE) {
1647 /* Don't commit yet if this is a non-blocking flush and there
1648 * are a lot of outstanding writes for this mapping.
1649 */
1650 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1651 goto out_mark_dirty;
1652
1653 /* don't wait for the COMMIT response */
1654 flags = 0;
1655 }
1656
1657 ret = nfs_commit_inode(inode, flags);
1658 if (ret >= 0) {
1659 if (wbc->sync_mode == WB_SYNC_NONE) {
1660 if (ret < wbc->nr_to_write)
1661 wbc->nr_to_write -= ret;
1662 else
1663 wbc->nr_to_write = 0;
1664 }
1665 return 0;
1666 }
1667out_mark_dirty:
1668 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1669 return ret;
1670}
1671#else
1672static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1673{
1674 return 0;
1675}
1676#endif
1677
1678int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1679{
1680 int ret;
1681
1682 ret = nfs_commit_unstable_pages(inode, wbc);
1683 if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1684 int status;
1685 bool sync = true;
1686
1687 if (wbc->sync_mode == WB_SYNC_NONE)
1688 sync = false;
1689
1690 status = pnfs_layoutcommit_inode(inode, sync);
1691 if (status < 0)
1692 return status;
1693 }
1694 return ret;
1695}
1696
1697/*
1698 * flush the inode to disk.
1699 */
1700int nfs_wb_all(struct inode *inode)
1701{
1702 struct writeback_control wbc = {
1703 .sync_mode = WB_SYNC_ALL,
1704 .nr_to_write = LONG_MAX,
1705 .range_start = 0,
1706 .range_end = LLONG_MAX,
1707 };
1708
1709 return sync_inode(inode, &wbc);
1710}
1711
1712int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1713{
1714 struct nfs_page *req;
1715 int ret = 0;
1716
1717 BUG_ON(!PageLocked(page));
1718 for (;;) {
1719 wait_on_page_writeback(page);
1720 req = nfs_page_find_request(page);
1721 if (req == NULL)
1722 break;
1723 if (nfs_lock_request(req)) {
1724 nfs_clear_request_commit(req);
1725 nfs_inode_remove_request(req);
1726 /*
1727 * In case nfs_inode_remove_request has marked the
1728 * page as being dirty
1729 */
1730 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1731 nfs_unlock_and_release_request(req);
1732 break;
1733 }
1734 ret = nfs_wait_on_request(req);
1735 nfs_release_request(req);
1736 if (ret < 0)
1737 break;
1738 }
1739 return ret;
1740}
1741
1742/*
1743 * Write back all requests on one page - we do this before reading it.
1744 */
1745int nfs_wb_page(struct inode *inode, struct page *page)
1746{
1747 loff_t range_start = page_offset(page);
1748 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1749 struct writeback_control wbc = {
1750 .sync_mode = WB_SYNC_ALL,
1751 .nr_to_write = 0,
1752 .range_start = range_start,
1753 .range_end = range_end,
1754 };
1755 int ret;
1756
1757 for (;;) {
1758 wait_on_page_writeback(page);
1759 if (clear_page_dirty_for_io(page)) {
1760 ret = nfs_writepage_locked(page, &wbc);
1761 if (ret < 0)
1762 goto out_error;
1763 continue;
1764 }
1765 if (!PagePrivate(page))
1766 break;
1767 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1768 if (ret < 0)
1769 goto out_error;
1770 }
1771 return 0;
1772out_error:
1773 return ret;
1774}
1775
1776#ifdef CONFIG_MIGRATION
1777int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1778 struct page *page, enum migrate_mode mode)
1779{
1780 /*
1781 * If PagePrivate is set, then the page is currently associated with
1782 * an in-progress read or write request. Don't try to migrate it.
1783 *
1784 * FIXME: we could do this in principle, but we'll need a way to ensure
1785 * that we can safely release the inode reference while holding
1786 * the page lock.
1787 */
1788 if (PagePrivate(page))
1789 return -EBUSY;
1790
1791 nfs_fscache_release_page(page, GFP_KERNEL);
1792
1793 return migrate_page(mapping, newpage, page, mode);
1794}
1795#endif
1796
1797int __init nfs_init_writepagecache(void)
1798{
1799 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1800 sizeof(struct nfs_write_header),
1801 0, SLAB_HWCACHE_ALIGN,
1802 NULL);
1803 if (nfs_wdata_cachep == NULL)
1804 return -ENOMEM;
1805
1806 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1807 nfs_wdata_cachep);
1808 if (nfs_wdata_mempool == NULL)
1809 goto out_destroy_write_cache;
1810
1811 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1812 sizeof(struct nfs_commit_data),
1813 0, SLAB_HWCACHE_ALIGN,
1814 NULL);
1815 if (nfs_cdata_cachep == NULL)
1816 goto out_destroy_write_mempool;
1817
1818 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1819 nfs_wdata_cachep);
1820 if (nfs_commit_mempool == NULL)
1821 goto out_destroy_commit_cache;
1822
1823 /*
1824 * NFS congestion size, scale with available memory.
1825 *
1826 * 64MB: 8192k
1827 * 128MB: 11585k
1828 * 256MB: 16384k
1829 * 512MB: 23170k
1830 * 1GB: 32768k
1831 * 2GB: 46340k
1832 * 4GB: 65536k
1833 * 8GB: 92681k
1834 * 16GB: 131072k
1835 *
1836 * This allows larger machines to have larger/more transfers.
1837 * Limit the default to 256M
1838 */
1839 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1840 if (nfs_congestion_kb > 256*1024)
1841 nfs_congestion_kb = 256*1024;
1842
1843 return 0;
1844
1845out_destroy_commit_cache:
1846 kmem_cache_destroy(nfs_cdata_cachep);
1847out_destroy_write_mempool:
1848 mempool_destroy(nfs_wdata_mempool);
1849out_destroy_write_cache:
1850 kmem_cache_destroy(nfs_wdata_cachep);
1851 return -ENOMEM;
1852}
1853
1854void nfs_destroy_writepagecache(void)
1855{
1856 mempool_destroy(nfs_commit_mempool);
1857 kmem_cache_destroy(nfs_cdata_cachep);
1858 mempool_destroy(nfs_wdata_mempool);
1859 kmem_cache_destroy(nfs_wdata_cachep);
1860}
1861
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/nfs/write.c
4 *
5 * Write file data over NFS.
6 *
7 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
8 */
9
10#include <linux/types.h>
11#include <linux/slab.h>
12#include <linux/mm.h>
13#include <linux/pagemap.h>
14#include <linux/file.h>
15#include <linux/writeback.h>
16#include <linux/swap.h>
17#include <linux/migrate.h>
18
19#include <linux/sunrpc/clnt.h>
20#include <linux/nfs_fs.h>
21#include <linux/nfs_mount.h>
22#include <linux/nfs_page.h>
23#include <linux/backing-dev.h>
24#include <linux/export.h>
25#include <linux/freezer.h>
26#include <linux/wait.h>
27#include <linux/iversion.h>
28#include <linux/filelock.h>
29
30#include <linux/uaccess.h>
31#include <linux/sched/mm.h>
32
33#include "delegation.h"
34#include "internal.h"
35#include "iostat.h"
36#include "nfs4_fs.h"
37#include "fscache.h"
38#include "pnfs.h"
39
40#include "nfstrace.h"
41
42#define NFSDBG_FACILITY NFSDBG_PAGECACHE
43
44#define MIN_POOL_WRITE (32)
45#define MIN_POOL_COMMIT (4)
46
47struct nfs_io_completion {
48 void (*complete)(void *data);
49 void *data;
50 struct kref refcount;
51};
52
53/*
54 * Local function declarations
55 */
56static void nfs_redirty_request(struct nfs_page *req);
57static const struct rpc_call_ops nfs_commit_ops;
58static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
59static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
60static const struct nfs_rw_ops nfs_rw_write_ops;
61static void nfs_inode_remove_request(struct nfs_page *req);
62static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
63 struct nfs_page *req);
64static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
65 struct inode *inode);
66
67static struct kmem_cache *nfs_wdata_cachep;
68static mempool_t *nfs_wdata_mempool;
69static struct kmem_cache *nfs_cdata_cachep;
70static mempool_t *nfs_commit_mempool;
71
72struct nfs_commit_data *nfs_commitdata_alloc(void)
73{
74 struct nfs_commit_data *p;
75
76 p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
77 if (!p) {
78 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
79 if (!p)
80 return NULL;
81 memset(p, 0, sizeof(*p));
82 }
83 INIT_LIST_HEAD(&p->pages);
84 return p;
85}
86EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
87
88void nfs_commit_free(struct nfs_commit_data *p)
89{
90 mempool_free(p, nfs_commit_mempool);
91}
92EXPORT_SYMBOL_GPL(nfs_commit_free);
93
94static struct nfs_pgio_header *nfs_writehdr_alloc(void)
95{
96 struct nfs_pgio_header *p;
97
98 p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
99 if (!p) {
100 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
101 if (!p)
102 return NULL;
103 memset(p, 0, sizeof(*p));
104 }
105 p->rw_mode = FMODE_WRITE;
106 return p;
107}
108
109static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
110{
111 mempool_free(hdr, nfs_wdata_mempool);
112}
113
114static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
115{
116 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
117}
118
119static void nfs_io_completion_init(struct nfs_io_completion *ioc,
120 void (*complete)(void *), void *data)
121{
122 ioc->complete = complete;
123 ioc->data = data;
124 kref_init(&ioc->refcount);
125}
126
127static void nfs_io_completion_release(struct kref *kref)
128{
129 struct nfs_io_completion *ioc = container_of(kref,
130 struct nfs_io_completion, refcount);
131 ioc->complete(ioc->data);
132 kfree(ioc);
133}
134
135static void nfs_io_completion_get(struct nfs_io_completion *ioc)
136{
137 if (ioc != NULL)
138 kref_get(&ioc->refcount);
139}
140
141static void nfs_io_completion_put(struct nfs_io_completion *ioc)
142{
143 if (ioc != NULL)
144 kref_put(&ioc->refcount, nfs_io_completion_release);
145}
146
147static void
148nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
149{
150 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
151 kref_get(&req->wb_kref);
152 atomic_long_inc(&NFS_I(inode)->nrequests);
153 }
154}
155
156static int
157nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
158{
159 int ret;
160
161 if (!test_bit(PG_REMOVE, &req->wb_flags))
162 return 0;
163 ret = nfs_page_group_lock(req);
164 if (ret)
165 return ret;
166 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
167 nfs_page_set_inode_ref(req, inode);
168 nfs_page_group_unlock(req);
169 return 0;
170}
171
172/**
173 * nfs_folio_find_head_request - find head request associated with a folio
174 * @folio: pointer to folio
175 *
176 * must be called while holding the inode lock.
177 *
178 * returns matching head request with reference held, or NULL if not found.
179 */
180static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
181{
182 struct address_space *mapping = folio->mapping;
183 struct nfs_page *req;
184
185 if (!folio_test_private(folio))
186 return NULL;
187 spin_lock(&mapping->i_private_lock);
188 req = folio->private;
189 if (req) {
190 WARN_ON_ONCE(req->wb_head != req);
191 kref_get(&req->wb_kref);
192 }
193 spin_unlock(&mapping->i_private_lock);
194 return req;
195}
196
197/* Adjust the file length if we're writing beyond the end */
198static void nfs_grow_file(struct folio *folio, unsigned int offset,
199 unsigned int count)
200{
201 struct inode *inode = folio->mapping->host;
202 loff_t end, i_size;
203 pgoff_t end_index;
204
205 spin_lock(&inode->i_lock);
206 i_size = i_size_read(inode);
207 end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio);
208 if (i_size > 0 && folio->index < end_index)
209 goto out;
210 end = folio_pos(folio) + (loff_t)offset + (loff_t)count;
211 if (i_size >= end)
212 goto out;
213 trace_nfs_size_grow(inode, end);
214 i_size_write(inode, end);
215 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
216 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
217out:
218 /* Atomically update timestamps if they are delegated to us. */
219 nfs_update_delegated_mtime_locked(inode);
220 spin_unlock(&inode->i_lock);
221 nfs_fscache_invalidate(inode, 0);
222}
223
224/* A writeback failed: mark the page as bad, and invalidate the page cache */
225static void nfs_set_pageerror(struct address_space *mapping)
226{
227 struct inode *inode = mapping->host;
228
229 nfs_zap_mapping(mapping->host, mapping);
230 /* Force file size revalidation */
231 spin_lock(&inode->i_lock);
232 nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
233 NFS_INO_INVALID_CHANGE |
234 NFS_INO_INVALID_SIZE);
235 spin_unlock(&inode->i_lock);
236}
237
238static void nfs_mapping_set_error(struct folio *folio, int error)
239{
240 struct address_space *mapping = folio->mapping;
241
242 filemap_set_wb_err(mapping, error);
243 if (mapping->host)
244 errseq_set(&mapping->host->i_sb->s_wb_err,
245 error == -ENOSPC ? -ENOSPC : -EIO);
246 nfs_set_pageerror(mapping);
247}
248
249/*
250 * nfs_page_group_search_locked
251 * @head - head request of page group
252 * @page_offset - offset into page
253 *
254 * Search page group with head @head to find a request that contains the
255 * page offset @page_offset.
256 *
257 * Returns a pointer to the first matching nfs request, or NULL if no
258 * match is found.
259 *
260 * Must be called with the page group lock held
261 */
262static struct nfs_page *
263nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
264{
265 struct nfs_page *req;
266
267 req = head;
268 do {
269 if (page_offset >= req->wb_pgbase &&
270 page_offset < (req->wb_pgbase + req->wb_bytes))
271 return req;
272
273 req = req->wb_this_page;
274 } while (req != head);
275
276 return NULL;
277}
278
279/*
280 * nfs_page_group_covers_page
281 * @head - head request of page group
282 *
283 * Return true if the page group with head @head covers the whole page,
284 * returns false otherwise
285 */
286static bool nfs_page_group_covers_page(struct nfs_page *req)
287{
288 unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
289 struct nfs_page *tmp;
290 unsigned int pos = 0;
291
292 nfs_page_group_lock(req);
293
294 for (;;) {
295 tmp = nfs_page_group_search_locked(req->wb_head, pos);
296 if (!tmp)
297 break;
298 pos = tmp->wb_pgbase + tmp->wb_bytes;
299 }
300
301 nfs_page_group_unlock(req);
302 return pos >= len;
303}
304
305/* We can set the PG_uptodate flag if we see that a write request
306 * covers the full page.
307 */
308static void nfs_mark_uptodate(struct nfs_page *req)
309{
310 struct folio *folio = nfs_page_to_folio(req);
311
312 if (folio_test_uptodate(folio))
313 return;
314 if (!nfs_page_group_covers_page(req))
315 return;
316 folio_mark_uptodate(folio);
317}
318
319static int wb_priority(struct writeback_control *wbc)
320{
321 int ret = 0;
322
323 if (wbc->sync_mode == WB_SYNC_ALL)
324 ret = FLUSH_COND_STABLE;
325 return ret;
326}
327
328/*
329 * NFS congestion control
330 */
331
332int nfs_congestion_kb;
333
334#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
335#define NFS_CONGESTION_OFF_THRESH \
336 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
337
338static void nfs_folio_set_writeback(struct folio *folio)
339{
340 struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
341
342 folio_start_writeback(folio);
343 if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
344 nfss->write_congested = 1;
345}
346
347static void nfs_folio_end_writeback(struct folio *folio)
348{
349 struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
350
351 folio_end_writeback(folio);
352 if (atomic_long_dec_return(&nfss->writeback) <
353 NFS_CONGESTION_OFF_THRESH) {
354 nfss->write_congested = 0;
355 wake_up_all(&nfss->write_congestion_wait);
356 }
357}
358
359static void nfs_page_end_writeback(struct nfs_page *req)
360{
361 if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
362 nfs_unlock_request(req);
363 nfs_folio_end_writeback(nfs_page_to_folio(req));
364 } else
365 nfs_unlock_request(req);
366}
367
368/*
369 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
370 *
371 * @destroy_list - request list (using wb_this_page) terminated by @old_head
372 * @old_head - the old head of the list
373 *
374 * All subrequests must be locked and removed from all lists, so at this point
375 * they are only "active" in this function, and possibly in nfs_wait_on_request
376 * with a reference held by some other context.
377 */
378static void
379nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
380 struct nfs_page *old_head,
381 struct inode *inode)
382{
383 while (destroy_list) {
384 struct nfs_page *subreq = destroy_list;
385
386 destroy_list = (subreq->wb_this_page == old_head) ?
387 NULL : subreq->wb_this_page;
388
389 /* Note: lock subreq in order to change subreq->wb_head */
390 nfs_page_set_headlock(subreq);
391 WARN_ON_ONCE(old_head != subreq->wb_head);
392
393 /* make sure old group is not used */
394 subreq->wb_this_page = subreq;
395 subreq->wb_head = subreq;
396
397 clear_bit(PG_REMOVE, &subreq->wb_flags);
398
399 /* Note: races with nfs_page_group_destroy() */
400 if (!kref_read(&subreq->wb_kref)) {
401 /* Check if we raced with nfs_page_group_destroy() */
402 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
403 nfs_page_clear_headlock(subreq);
404 nfs_free_request(subreq);
405 } else
406 nfs_page_clear_headlock(subreq);
407 continue;
408 }
409 nfs_page_clear_headlock(subreq);
410
411 nfs_release_request(old_head);
412
413 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
414 nfs_release_request(subreq);
415 atomic_long_dec(&NFS_I(inode)->nrequests);
416 }
417
418 /* subreq is now totally disconnected from page group or any
419 * write / commit lists. last chance to wake any waiters */
420 nfs_unlock_and_release_request(subreq);
421 }
422}
423
424/*
425 * nfs_join_page_group - destroy subrequests of the head req
426 * @head: the page used to lookup the "page group" of nfs_page structures
427 * @inode: Inode to which the request belongs.
428 *
429 * This function joins all sub requests to the head request by first
430 * locking all requests in the group, cancelling any pending operations
431 * and finally updating the head request to cover the whole range covered by
432 * the (former) group. All subrequests are removed from any write or commit
433 * lists, unlinked from the group and destroyed.
434 */
435void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
436 struct inode *inode)
437{
438 struct nfs_page *subreq;
439 struct nfs_page *destroy_list = NULL;
440 unsigned int pgbase, off, bytes;
441
442 pgbase = head->wb_pgbase;
443 bytes = head->wb_bytes;
444 off = head->wb_offset;
445 for (subreq = head->wb_this_page; subreq != head;
446 subreq = subreq->wb_this_page) {
447 /* Subrequests should always form a contiguous range */
448 if (pgbase > subreq->wb_pgbase) {
449 off -= pgbase - subreq->wb_pgbase;
450 bytes += pgbase - subreq->wb_pgbase;
451 pgbase = subreq->wb_pgbase;
452 }
453 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
454 - pgbase, bytes);
455 }
456
457 /* Set the head request's range to cover the former page group */
458 head->wb_pgbase = pgbase;
459 head->wb_bytes = bytes;
460 head->wb_offset = off;
461
462 /* Now that all requests are locked, make sure they aren't on any list.
463 * Commit list removal accounting is done after locks are dropped */
464 subreq = head;
465 do {
466 nfs_clear_request_commit(cinfo, subreq);
467 subreq = subreq->wb_this_page;
468 } while (subreq != head);
469
470 /* unlink subrequests from head, destroy them later */
471 if (head->wb_this_page != head) {
472 /* destroy list will be terminated by head */
473 destroy_list = head->wb_this_page;
474 head->wb_this_page = head;
475 }
476
477 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
478}
479
480/**
481 * nfs_wait_on_request - Wait for a request to complete.
482 * @req: request to wait upon.
483 *
484 * Interruptible by fatal signals only.
485 * The user is responsible for holding a count on the request.
486 */
487static int nfs_wait_on_request(struct nfs_page *req)
488{
489 if (!test_bit(PG_BUSY, &req->wb_flags))
490 return 0;
491 set_bit(PG_CONTENDED2, &req->wb_flags);
492 smp_mb__after_atomic();
493 return wait_on_bit_io(&req->wb_flags, PG_BUSY,
494 TASK_UNINTERRUPTIBLE);
495}
496
497/*
498 * nfs_unroll_locks - unlock all newly locked reqs and wait on @req
499 * @head: head request of page group, must be holding head lock
500 * @req: request that couldn't lock and needs to wait on the req bit lock
501 *
502 * This is a helper function for nfs_lock_and_join_requests
503 * returns 0 on success, < 0 on error.
504 */
505static void
506nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
507{
508 struct nfs_page *tmp;
509
510 /* relinquish all the locks successfully grabbed this run */
511 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
512 if (!kref_read(&tmp->wb_kref))
513 continue;
514 nfs_unlock_and_release_request(tmp);
515 }
516}
517
518/*
519 * nfs_page_group_lock_subreq - try to lock a subrequest
520 * @head: head request of page group
521 * @subreq: request to lock
522 *
523 * This is a helper function for nfs_lock_and_join_requests which
524 * must be called with the head request and page group both locked.
525 * On error, it returns with the page group unlocked.
526 */
527static int
528nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
529{
530 int ret;
531
532 if (!kref_get_unless_zero(&subreq->wb_kref))
533 return 0;
534 while (!nfs_lock_request(subreq)) {
535 nfs_page_group_unlock(head);
536 ret = nfs_wait_on_request(subreq);
537 if (!ret)
538 ret = nfs_page_group_lock(head);
539 if (ret < 0) {
540 nfs_unroll_locks(head, subreq);
541 nfs_release_request(subreq);
542 return ret;
543 }
544 }
545 return 0;
546}
547
548/*
549 * nfs_lock_and_join_requests - join all subreqs to the head req
550 * @folio: the folio used to lookup the "page group" of nfs_page structures
551 *
552 * This function joins all sub requests to the head request by first
553 * locking all requests in the group, cancelling any pending operations
554 * and finally updating the head request to cover the whole range covered by
555 * the (former) group. All subrequests are removed from any write or commit
556 * lists, unlinked from the group and destroyed.
557 *
558 * Returns a locked, referenced pointer to the head request - which after
559 * this call is guaranteed to be the only request associated with the page.
560 * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
561 * error was encountered.
562 */
563static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
564{
565 struct inode *inode = folio->mapping->host;
566 struct nfs_page *head, *subreq;
567 struct nfs_commit_info cinfo;
568 int ret;
569
570 /*
571 * A reference is taken only on the head request which acts as a
572 * reference to the whole page group - the group will not be destroyed
573 * until the head reference is released.
574 */
575retry:
576 head = nfs_folio_find_head_request(folio);
577 if (!head)
578 return NULL;
579
580 while (!nfs_lock_request(head)) {
581 ret = nfs_wait_on_request(head);
582 if (ret < 0)
583 return ERR_PTR(ret);
584 }
585
586 /* Ensure that nobody removed the request before we locked it */
587 if (head != folio->private) {
588 nfs_unlock_and_release_request(head);
589 goto retry;
590 }
591
592 ret = nfs_cancel_remove_inode(head, inode);
593 if (ret < 0)
594 goto out_unlock;
595
596 ret = nfs_page_group_lock(head);
597 if (ret < 0)
598 goto out_unlock;
599
600 /* lock each request in the page group */
601 for (subreq = head->wb_this_page;
602 subreq != head;
603 subreq = subreq->wb_this_page) {
604 ret = nfs_page_group_lock_subreq(head, subreq);
605 if (ret < 0)
606 goto out_unlock;
607 }
608
609 nfs_page_group_unlock(head);
610
611 nfs_init_cinfo_from_inode(&cinfo, inode);
612 nfs_join_page_group(head, &cinfo, inode);
613 return head;
614
615out_unlock:
616 nfs_unlock_and_release_request(head);
617 return ERR_PTR(ret);
618}
619
620static void nfs_write_error(struct nfs_page *req, int error)
621{
622 trace_nfs_write_error(nfs_page_to_inode(req), req, error);
623 nfs_mapping_set_error(nfs_page_to_folio(req), error);
624 nfs_inode_remove_request(req);
625 nfs_page_end_writeback(req);
626 nfs_release_request(req);
627}
628
629/*
630 * Find an associated nfs write request, and prepare to flush it out
631 * May return an error if the user signalled nfs_wait_on_request().
632 */
633static int nfs_page_async_flush(struct folio *folio,
634 struct writeback_control *wbc,
635 struct nfs_pageio_descriptor *pgio)
636{
637 struct nfs_page *req;
638 int ret = 0;
639
640 req = nfs_lock_and_join_requests(folio);
641 if (!req)
642 goto out;
643 ret = PTR_ERR(req);
644 if (IS_ERR(req))
645 goto out;
646
647 nfs_folio_set_writeback(folio);
648 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
649
650 /* If there is a fatal error that covers this write, just exit */
651 ret = pgio->pg_error;
652 if (nfs_error_is_fatal_on_server(ret))
653 goto out_launder;
654
655 ret = 0;
656 if (!nfs_pageio_add_request(pgio, req)) {
657 ret = pgio->pg_error;
658 /*
659 * Remove the problematic req upon fatal errors on the server
660 */
661 if (nfs_error_is_fatal_on_server(ret))
662 goto out_launder;
663 if (wbc->sync_mode == WB_SYNC_NONE)
664 ret = AOP_WRITEPAGE_ACTIVATE;
665 folio_redirty_for_writepage(wbc, folio);
666 nfs_redirty_request(req);
667 pgio->pg_error = 0;
668 } else
669 nfs_add_stats(folio->mapping->host,
670 NFSIOS_WRITEPAGES, 1);
671out:
672 return ret;
673out_launder:
674 nfs_write_error(req, ret);
675 return 0;
676}
677
678static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
679 struct nfs_pageio_descriptor *pgio)
680{
681 nfs_pageio_cond_complete(pgio, folio->index);
682 return nfs_page_async_flush(folio, wbc, pgio);
683}
684
685/*
686 * Write an mmapped page to the server.
687 */
688static int nfs_writepage_locked(struct folio *folio,
689 struct writeback_control *wbc)
690{
691 struct nfs_pageio_descriptor pgio;
692 struct inode *inode = folio->mapping->host;
693 int err;
694
695 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
696 nfs_pageio_init_write(&pgio, inode, 0, false,
697 &nfs_async_write_completion_ops);
698 err = nfs_do_writepage(folio, wbc, &pgio);
699 pgio.pg_error = 0;
700 nfs_pageio_complete(&pgio);
701 return err;
702}
703
704static int nfs_writepages_callback(struct folio *folio,
705 struct writeback_control *wbc, void *data)
706{
707 int ret;
708
709 ret = nfs_do_writepage(folio, wbc, data);
710 if (ret != AOP_WRITEPAGE_ACTIVATE)
711 folio_unlock(folio);
712 return ret;
713}
714
715static void nfs_io_completion_commit(void *inode)
716{
717 nfs_commit_inode(inode, 0);
718}
719
720int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
721{
722 struct inode *inode = mapping->host;
723 struct nfs_pageio_descriptor pgio;
724 struct nfs_io_completion *ioc = NULL;
725 unsigned int mntflags = NFS_SERVER(inode)->flags;
726 struct nfs_server *nfss = NFS_SERVER(inode);
727 int priority = 0;
728 int err;
729
730 /* Wait with writeback until write congestion eases */
731 if (wbc->sync_mode == WB_SYNC_NONE && nfss->write_congested) {
732 err = wait_event_killable(nfss->write_congestion_wait,
733 nfss->write_congested == 0);
734 if (err)
735 return err;
736 }
737
738 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
739
740 if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
741 wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
742 ioc = nfs_io_completion_alloc(GFP_KERNEL);
743 if (ioc)
744 nfs_io_completion_init(ioc, nfs_io_completion_commit,
745 inode);
746 priority = wb_priority(wbc);
747 }
748
749 do {
750 nfs_pageio_init_write(&pgio, inode, priority, false,
751 &nfs_async_write_completion_ops);
752 pgio.pg_io_completion = ioc;
753 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
754 &pgio);
755 pgio.pg_error = 0;
756 nfs_pageio_complete(&pgio);
757 if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
758 break;
759 } while (err < 0 && !nfs_error_is_fatal(err));
760 nfs_io_completion_put(ioc);
761
762 if (err < 0)
763 goto out_err;
764 return 0;
765out_err:
766 return err;
767}
768
769/*
770 * Insert a write request into an inode
771 */
772static void nfs_inode_add_request(struct nfs_page *req)
773{
774 struct folio *folio = nfs_page_to_folio(req);
775 struct address_space *mapping = folio->mapping;
776 struct nfs_inode *nfsi = NFS_I(mapping->host);
777
778 WARN_ON_ONCE(req->wb_this_page != req);
779
780 /* Lock the request! */
781 nfs_lock_request(req);
782 spin_lock(&mapping->i_private_lock);
783 set_bit(PG_MAPPED, &req->wb_flags);
784 folio_set_private(folio);
785 folio->private = req;
786 spin_unlock(&mapping->i_private_lock);
787 atomic_long_inc(&nfsi->nrequests);
788 /* this a head request for a page group - mark it as having an
789 * extra reference so sub groups can follow suit.
790 * This flag also informs pgio layer when to bump nrequests when
791 * adding subrequests. */
792 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
793 kref_get(&req->wb_kref);
794}
795
796/*
797 * Remove a write request from an inode
798 */
799static void nfs_inode_remove_request(struct nfs_page *req)
800{
801 struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
802
803 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
804 struct folio *folio = nfs_page_to_folio(req->wb_head);
805 struct address_space *mapping = folio->mapping;
806
807 spin_lock(&mapping->i_private_lock);
808 if (likely(folio)) {
809 folio->private = NULL;
810 folio_clear_private(folio);
811 clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
812 }
813 spin_unlock(&mapping->i_private_lock);
814 }
815
816 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
817 atomic_long_dec(&nfsi->nrequests);
818 nfs_release_request(req);
819 }
820}
821
822static void nfs_mark_request_dirty(struct nfs_page *req)
823{
824 struct folio *folio = nfs_page_to_folio(req);
825 if (folio)
826 filemap_dirty_folio(folio_mapping(folio), folio);
827}
828
829/**
830 * nfs_request_add_commit_list_locked - add request to a commit list
831 * @req: pointer to a struct nfs_page
832 * @dst: commit list head
833 * @cinfo: holds list lock and accounting info
834 *
835 * This sets the PG_CLEAN bit, updates the cinfo count of
836 * number of outstanding requests requiring a commit as well as
837 * the MM page stats.
838 *
839 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
840 * nfs_page lock.
841 */
842void
843nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
844 struct nfs_commit_info *cinfo)
845{
846 set_bit(PG_CLEAN, &req->wb_flags);
847 nfs_list_add_request(req, dst);
848 atomic_long_inc(&cinfo->mds->ncommit);
849}
850EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
851
852/**
853 * nfs_request_add_commit_list - add request to a commit list
854 * @req: pointer to a struct nfs_page
855 * @cinfo: holds list lock and accounting info
856 *
857 * This sets the PG_CLEAN bit, updates the cinfo count of
858 * number of outstanding requests requiring a commit as well as
859 * the MM page stats.
860 *
861 * The caller must _not_ hold the cinfo->lock, but must be
862 * holding the nfs_page lock.
863 */
864void
865nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
866{
867 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
868 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
869 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
870 nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
871}
872EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
873
874/**
875 * nfs_request_remove_commit_list - Remove request from a commit list
876 * @req: pointer to a nfs_page
877 * @cinfo: holds list lock and accounting info
878 *
879 * This clears the PG_CLEAN bit, and updates the cinfo's count of
880 * number of outstanding requests requiring a commit
881 * It does not update the MM page stats.
882 *
883 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
884 */
885void
886nfs_request_remove_commit_list(struct nfs_page *req,
887 struct nfs_commit_info *cinfo)
888{
889 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
890 return;
891 nfs_list_remove_request(req);
892 atomic_long_dec(&cinfo->mds->ncommit);
893}
894EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
895
896static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
897 struct inode *inode)
898{
899 cinfo->inode = inode;
900 cinfo->mds = &NFS_I(inode)->commit_info;
901 cinfo->ds = pnfs_get_ds_info(inode);
902 cinfo->dreq = NULL;
903 cinfo->completion_ops = &nfs_commit_completion_ops;
904}
905
906void nfs_init_cinfo(struct nfs_commit_info *cinfo,
907 struct inode *inode,
908 struct nfs_direct_req *dreq)
909{
910 if (dreq)
911 nfs_init_cinfo_from_dreq(cinfo, dreq);
912 else
913 nfs_init_cinfo_from_inode(cinfo, inode);
914}
915EXPORT_SYMBOL_GPL(nfs_init_cinfo);
916
917/*
918 * Add a request to the inode's commit list.
919 */
920void
921nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
922 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
923{
924 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
925 return;
926 nfs_request_add_commit_list(req, cinfo);
927}
928
929static void nfs_folio_clear_commit(struct folio *folio)
930{
931 if (folio) {
932 long nr = folio_nr_pages(folio);
933
934 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
935 wb_stat_mod(&inode_to_bdi(folio->mapping->host)->wb,
936 WB_WRITEBACK, -nr);
937 }
938}
939
940/* Called holding the request lock on @req */
941static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
942 struct nfs_page *req)
943{
944 if (test_bit(PG_CLEAN, &req->wb_flags)) {
945 struct nfs_open_context *ctx = nfs_req_openctx(req);
946 struct inode *inode = d_inode(ctx->dentry);
947
948 mutex_lock(&NFS_I(inode)->commit_mutex);
949 if (!pnfs_clear_request_commit(req, cinfo)) {
950 nfs_request_remove_commit_list(req, cinfo);
951 }
952 mutex_unlock(&NFS_I(inode)->commit_mutex);
953 nfs_folio_clear_commit(nfs_page_to_folio(req));
954 }
955}
956
957int nfs_write_need_commit(struct nfs_pgio_header *hdr)
958{
959 if (hdr->verf.committed == NFS_DATA_SYNC)
960 return hdr->lseg == NULL;
961 return hdr->verf.committed != NFS_FILE_SYNC;
962}
963
964static void nfs_async_write_init(struct nfs_pgio_header *hdr)
965{
966 nfs_io_completion_get(hdr->io_completion);
967}
968
969static void nfs_write_completion(struct nfs_pgio_header *hdr)
970{
971 struct nfs_commit_info cinfo;
972 unsigned long bytes = 0;
973
974 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
975 goto out;
976 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
977 while (!list_empty(&hdr->pages)) {
978 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
979
980 bytes += req->wb_bytes;
981 nfs_list_remove_request(req);
982 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
983 (hdr->good_bytes < bytes)) {
984 trace_nfs_comp_error(hdr->inode, req, hdr->error);
985 nfs_mapping_set_error(nfs_page_to_folio(req),
986 hdr->error);
987 goto remove_req;
988 }
989 if (nfs_write_need_commit(hdr)) {
990 /* Reset wb_nio, since the write was successful. */
991 req->wb_nio = 0;
992 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
993 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
994 hdr->pgio_mirror_idx);
995 goto next;
996 }
997remove_req:
998 nfs_inode_remove_request(req);
999next:
1000 nfs_page_end_writeback(req);
1001 nfs_release_request(req);
1002 }
1003out:
1004 nfs_io_completion_put(hdr->io_completion);
1005 hdr->release(hdr);
1006}
1007
1008unsigned long
1009nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1010{
1011 return atomic_long_read(&cinfo->mds->ncommit);
1012}
1013
1014/* NFS_I(cinfo->inode)->commit_mutex held by caller */
1015int
1016nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1017 struct nfs_commit_info *cinfo, int max)
1018{
1019 struct nfs_page *req, *tmp;
1020 int ret = 0;
1021
1022 list_for_each_entry_safe(req, tmp, src, wb_list) {
1023 kref_get(&req->wb_kref);
1024 if (!nfs_lock_request(req)) {
1025 nfs_release_request(req);
1026 continue;
1027 }
1028 nfs_request_remove_commit_list(req, cinfo);
1029 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1030 nfs_list_add_request(req, dst);
1031 ret++;
1032 if ((ret == max) && !cinfo->dreq)
1033 break;
1034 cond_resched();
1035 }
1036 return ret;
1037}
1038EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1039
1040/*
1041 * nfs_scan_commit - Scan an inode for commit requests
1042 * @inode: NFS inode to scan
1043 * @dst: mds destination list
1044 * @cinfo: mds and ds lists of reqs ready to commit
1045 *
1046 * Moves requests from the inode's 'commit' request list.
1047 * The requests are *not* checked to ensure that they form a contiguous set.
1048 */
1049int
1050nfs_scan_commit(struct inode *inode, struct list_head *dst,
1051 struct nfs_commit_info *cinfo)
1052{
1053 int ret = 0;
1054
1055 if (!atomic_long_read(&cinfo->mds->ncommit))
1056 return 0;
1057 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1058 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1059 const int max = INT_MAX;
1060
1061 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1062 cinfo, max);
1063 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1064 }
1065 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1066 return ret;
1067}
1068
1069/*
1070 * Search for an existing write request, and attempt to update
1071 * it to reflect a new dirty region on a given page.
1072 *
1073 * If the attempt fails, then the existing request is flushed out
1074 * to disk.
1075 */
1076static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1077 unsigned int offset,
1078 unsigned int bytes)
1079{
1080 struct nfs_page *req;
1081 unsigned int rqend;
1082 unsigned int end;
1083 int error;
1084
1085 end = offset + bytes;
1086
1087 req = nfs_lock_and_join_requests(folio);
1088 if (IS_ERR_OR_NULL(req))
1089 return req;
1090
1091 rqend = req->wb_offset + req->wb_bytes;
1092 /*
1093 * Tell the caller to flush out the request if
1094 * the offsets are non-contiguous.
1095 * Note: nfs_flush_incompatible() will already
1096 * have flushed out requests having wrong owners.
1097 */
1098 if (offset > rqend || end < req->wb_offset)
1099 goto out_flushme;
1100
1101 /* Okay, the request matches. Update the region */
1102 if (offset < req->wb_offset) {
1103 req->wb_offset = offset;
1104 req->wb_pgbase = offset;
1105 }
1106 if (end > rqend)
1107 req->wb_bytes = end - req->wb_offset;
1108 else
1109 req->wb_bytes = rqend - req->wb_offset;
1110 req->wb_nio = 0;
1111 return req;
1112out_flushme:
1113 /*
1114 * Note: we mark the request dirty here because
1115 * nfs_lock_and_join_requests() cannot preserve
1116 * commit flags, so we have to replay the write.
1117 */
1118 nfs_mark_request_dirty(req);
1119 nfs_unlock_and_release_request(req);
1120 error = nfs_wb_folio(folio->mapping->host, folio);
1121 return (error < 0) ? ERR_PTR(error) : NULL;
1122}
1123
1124/*
1125 * Try to update an existing write request, or create one if there is none.
1126 *
1127 * Note: Should always be called with the Page Lock held to prevent races
1128 * if we have to add a new request. Also assumes that the caller has
1129 * already called nfs_flush_incompatible() if necessary.
1130 */
1131static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1132 struct folio *folio,
1133 unsigned int offset,
1134 unsigned int bytes)
1135{
1136 struct nfs_page *req;
1137
1138 req = nfs_try_to_update_request(folio, offset, bytes);
1139 if (req != NULL)
1140 goto out;
1141 req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1142 if (IS_ERR(req))
1143 goto out;
1144 nfs_inode_add_request(req);
1145out:
1146 return req;
1147}
1148
1149static int nfs_writepage_setup(struct nfs_open_context *ctx,
1150 struct folio *folio, unsigned int offset,
1151 unsigned int count)
1152{
1153 struct nfs_page *req;
1154
1155 req = nfs_setup_write_request(ctx, folio, offset, count);
1156 if (IS_ERR(req))
1157 return PTR_ERR(req);
1158 /* Update file length */
1159 nfs_grow_file(folio, offset, count);
1160 nfs_mark_uptodate(req);
1161 nfs_mark_request_dirty(req);
1162 nfs_unlock_and_release_request(req);
1163 return 0;
1164}
1165
1166int nfs_flush_incompatible(struct file *file, struct folio *folio)
1167{
1168 struct nfs_open_context *ctx = nfs_file_open_context(file);
1169 struct nfs_lock_context *l_ctx;
1170 struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1171 struct nfs_page *req;
1172 int do_flush, status;
1173 /*
1174 * Look for a request corresponding to this page. If there
1175 * is one, and it belongs to another file, we flush it out
1176 * before we try to copy anything into the page. Do this
1177 * due to the lack of an ACCESS-type call in NFSv2.
1178 * Also do the same if we find a request from an existing
1179 * dropped page.
1180 */
1181 do {
1182 req = nfs_folio_find_head_request(folio);
1183 if (req == NULL)
1184 return 0;
1185 l_ctx = req->wb_lock_context;
1186 do_flush = nfs_page_to_folio(req) != folio ||
1187 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1188 if (l_ctx && flctx &&
1189 !(list_empty_careful(&flctx->flc_posix) &&
1190 list_empty_careful(&flctx->flc_flock))) {
1191 do_flush |= l_ctx->lockowner != current->files;
1192 }
1193 nfs_release_request(req);
1194 if (!do_flush)
1195 return 0;
1196 status = nfs_wb_folio(folio->mapping->host, folio);
1197 } while (status == 0);
1198 return status;
1199}
1200
1201/*
1202 * Avoid buffered writes when a open context credential's key would
1203 * expire soon.
1204 *
1205 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1206 *
1207 * Return 0 and set a credential flag which triggers the inode to flush
1208 * and performs NFS_FILE_SYNC writes if the key will expired within
1209 * RPC_KEY_EXPIRE_TIMEO.
1210 */
1211int
1212nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1213{
1214 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1215
1216 if (nfs_ctx_key_to_expire(ctx, inode) &&
1217 !rcu_access_pointer(ctx->ll_cred))
1218 /* Already expired! */
1219 return -EACCES;
1220 return 0;
1221}
1222
1223/*
1224 * Test if the open context credential key is marked to expire soon.
1225 */
1226bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1227{
1228 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1229 struct rpc_cred *cred, *new, *old = NULL;
1230 struct auth_cred acred = {
1231 .cred = ctx->cred,
1232 };
1233 bool ret = false;
1234
1235 rcu_read_lock();
1236 cred = rcu_dereference(ctx->ll_cred);
1237 if (cred && !(cred->cr_ops->crkey_timeout &&
1238 cred->cr_ops->crkey_timeout(cred)))
1239 goto out;
1240 rcu_read_unlock();
1241
1242 new = auth->au_ops->lookup_cred(auth, &acred, 0);
1243 if (new == cred) {
1244 put_rpccred(new);
1245 return true;
1246 }
1247 if (IS_ERR_OR_NULL(new)) {
1248 new = NULL;
1249 ret = true;
1250 } else if (new->cr_ops->crkey_timeout &&
1251 new->cr_ops->crkey_timeout(new))
1252 ret = true;
1253
1254 rcu_read_lock();
1255 old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1256 RCU_INITIALIZER(new)), 1);
1257out:
1258 rcu_read_unlock();
1259 put_rpccred(old);
1260 return ret;
1261}
1262
1263/*
1264 * If the page cache is marked as unsafe or invalid, then we can't rely on
1265 * the PageUptodate() flag. In this case, we will need to turn off
1266 * write optimisations that depend on the page contents being correct.
1267 */
1268static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1269{
1270 struct inode *inode = folio->mapping->host;
1271 struct nfs_inode *nfsi = NFS_I(inode);
1272
1273 if (nfs_have_delegated_attributes(inode))
1274 goto out;
1275 if (nfsi->cache_validity &
1276 (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1277 return false;
1278 smp_rmb();
1279 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1280 return false;
1281out:
1282 if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1283 return false;
1284 return folio_test_uptodate(folio) != 0;
1285}
1286
1287static bool
1288is_whole_file_wrlock(struct file_lock *fl)
1289{
1290 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1291 lock_is_write(fl);
1292}
1293
1294/* If we know the page is up to date, and we're not using byte range locks (or
1295 * if we have the whole file locked for writing), it may be more efficient to
1296 * extend the write to cover the entire page in order to avoid fragmentation
1297 * inefficiencies.
1298 *
1299 * If the file is opened for synchronous writes then we can just skip the rest
1300 * of the checks.
1301 */
1302static int nfs_can_extend_write(struct file *file, struct folio *folio,
1303 unsigned int pagelen)
1304{
1305 struct inode *inode = file_inode(file);
1306 struct file_lock_context *flctx = locks_inode_context(inode);
1307 struct file_lock *fl;
1308 int ret;
1309 unsigned int mntflags = NFS_SERVER(inode)->flags;
1310
1311 if (mntflags & NFS_MOUNT_NO_ALIGNWRITE)
1312 return 0;
1313 if (file->f_flags & O_DSYNC)
1314 return 0;
1315 if (!nfs_folio_write_uptodate(folio, pagelen))
1316 return 0;
1317 if (nfs_have_write_delegation(inode))
1318 return 1;
1319 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1320 list_empty_careful(&flctx->flc_posix)))
1321 return 1;
1322
1323 /* Check to see if there are whole file write locks */
1324 ret = 0;
1325 spin_lock(&flctx->flc_lock);
1326 if (!list_empty(&flctx->flc_posix)) {
1327 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1328 c.flc_list);
1329 if (is_whole_file_wrlock(fl))
1330 ret = 1;
1331 } else if (!list_empty(&flctx->flc_flock)) {
1332 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1333 c.flc_list);
1334 if (lock_is_write(fl))
1335 ret = 1;
1336 }
1337 spin_unlock(&flctx->flc_lock);
1338 return ret;
1339}
1340
1341/*
1342 * Update and possibly write a cached page of an NFS file.
1343 *
1344 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1345 * things with a page scheduled for an RPC call (e.g. invalidate it).
1346 */
1347int nfs_update_folio(struct file *file, struct folio *folio,
1348 unsigned int offset, unsigned int count)
1349{
1350 struct nfs_open_context *ctx = nfs_file_open_context(file);
1351 struct address_space *mapping = folio->mapping;
1352 struct inode *inode = mapping->host;
1353 unsigned int pagelen = nfs_folio_length(folio);
1354 int status = 0;
1355
1356 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1357
1358 dprintk("NFS: nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1359 (long long)(folio_pos(folio) + offset));
1360
1361 if (!count)
1362 goto out;
1363
1364 if (nfs_can_extend_write(file, folio, pagelen)) {
1365 unsigned int end = count + offset;
1366
1367 offset = round_down(offset, PAGE_SIZE);
1368 if (end < pagelen)
1369 end = min(round_up(end, PAGE_SIZE), pagelen);
1370 count = end - offset;
1371 }
1372
1373 status = nfs_writepage_setup(ctx, folio, offset, count);
1374 if (status < 0)
1375 nfs_set_pageerror(mapping);
1376out:
1377 dprintk("NFS: nfs_update_folio returns %d (isize %lld)\n",
1378 status, (long long)i_size_read(inode));
1379 return status;
1380}
1381
1382static int flush_task_priority(int how)
1383{
1384 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1385 case FLUSH_HIGHPRI:
1386 return RPC_PRIORITY_HIGH;
1387 case FLUSH_LOWPRI:
1388 return RPC_PRIORITY_LOW;
1389 }
1390 return RPC_PRIORITY_NORMAL;
1391}
1392
1393static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1394 struct rpc_message *msg,
1395 const struct nfs_rpc_ops *rpc_ops,
1396 struct rpc_task_setup *task_setup_data, int how)
1397{
1398 int priority = flush_task_priority(how);
1399
1400 if (IS_SWAPFILE(hdr->inode))
1401 task_setup_data->flags |= RPC_TASK_SWAPPER;
1402 task_setup_data->priority = priority;
1403 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1404 trace_nfs_initiate_write(hdr);
1405}
1406
1407/* If a nfs_flush_* function fails, it should remove reqs from @head and
1408 * call this on each, which will prepare them to be retried on next
1409 * writeback using standard nfs.
1410 */
1411static void nfs_redirty_request(struct nfs_page *req)
1412{
1413 struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1414
1415 /* Bump the transmission count */
1416 req->wb_nio++;
1417 nfs_mark_request_dirty(req);
1418 atomic_long_inc(&nfsi->redirtied_pages);
1419 nfs_page_end_writeback(req);
1420 nfs_release_request(req);
1421}
1422
1423static void nfs_async_write_error(struct list_head *head, int error)
1424{
1425 struct nfs_page *req;
1426
1427 while (!list_empty(head)) {
1428 req = nfs_list_entry(head->next);
1429 nfs_list_remove_request(req);
1430 if (nfs_error_is_fatal_on_server(error))
1431 nfs_write_error(req, error);
1432 else
1433 nfs_redirty_request(req);
1434 }
1435}
1436
1437static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1438{
1439 nfs_async_write_error(&hdr->pages, 0);
1440}
1441
1442static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1443 .init_hdr = nfs_async_write_init,
1444 .error_cleanup = nfs_async_write_error,
1445 .completion = nfs_write_completion,
1446 .reschedule_io = nfs_async_write_reschedule_io,
1447};
1448
1449void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1450 struct inode *inode, int ioflags, bool force_mds,
1451 const struct nfs_pgio_completion_ops *compl_ops)
1452{
1453 struct nfs_server *server = NFS_SERVER(inode);
1454 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1455
1456#ifdef CONFIG_NFS_V4_1
1457 if (server->pnfs_curr_ld && !force_mds)
1458 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1459#endif
1460 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1461 server->wsize, ioflags);
1462}
1463EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1464
1465void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1466{
1467 struct nfs_pgio_mirror *mirror;
1468
1469 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1470 pgio->pg_ops->pg_cleanup(pgio);
1471
1472 pgio->pg_ops = &nfs_pgio_rw_ops;
1473
1474 nfs_pageio_stop_mirroring(pgio);
1475
1476 mirror = &pgio->pg_mirrors[0];
1477 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1478}
1479EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1480
1481
1482void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1483{
1484 struct nfs_commit_data *data = calldata;
1485
1486 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1487}
1488
1489static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1490 struct nfs_fattr *fattr)
1491{
1492 struct nfs_pgio_args *argp = &hdr->args;
1493 struct nfs_pgio_res *resp = &hdr->res;
1494 u64 size = argp->offset + resp->count;
1495
1496 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1497 fattr->size = size;
1498 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1499 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1500 return;
1501 }
1502 if (size != fattr->size)
1503 return;
1504 /* Set attribute barrier */
1505 nfs_fattr_set_barrier(fattr);
1506 /* ...and update size */
1507 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1508}
1509
1510void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1511{
1512 struct nfs_fattr *fattr = &hdr->fattr;
1513 struct inode *inode = hdr->inode;
1514
1515 if (nfs_have_delegated_mtime(inode)) {
1516 spin_lock(&inode->i_lock);
1517 nfs_set_cache_invalid(inode, NFS_INO_INVALID_BLOCKS);
1518 spin_unlock(&inode->i_lock);
1519 return;
1520 }
1521
1522 spin_lock(&inode->i_lock);
1523 nfs_writeback_check_extend(hdr, fattr);
1524 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1525 spin_unlock(&inode->i_lock);
1526}
1527EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1528
1529/*
1530 * This function is called when the WRITE call is complete.
1531 */
1532static int nfs_writeback_done(struct rpc_task *task,
1533 struct nfs_pgio_header *hdr,
1534 struct inode *inode)
1535{
1536 int status;
1537
1538 /*
1539 * ->write_done will attempt to use post-op attributes to detect
1540 * conflicting writes by other clients. A strict interpretation
1541 * of close-to-open would allow us to continue caching even if
1542 * another writer had changed the file, but some applications
1543 * depend on tighter cache coherency when writing.
1544 */
1545 status = NFS_PROTO(inode)->write_done(task, hdr);
1546 if (status != 0)
1547 return status;
1548
1549 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1550 trace_nfs_writeback_done(task, hdr);
1551
1552 if (task->tk_status >= 0) {
1553 enum nfs3_stable_how committed = hdr->res.verf->committed;
1554
1555 if (committed == NFS_UNSTABLE) {
1556 /*
1557 * We have some uncommitted data on the server at
1558 * this point, so ensure that we keep track of that
1559 * fact irrespective of what later writes do.
1560 */
1561 set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1562 }
1563
1564 if (committed < hdr->args.stable) {
1565 /* We tried a write call, but the server did not
1566 * commit data to stable storage even though we
1567 * requested it.
1568 * Note: There is a known bug in Tru64 < 5.0 in which
1569 * the server reports NFS_DATA_SYNC, but performs
1570 * NFS_FILE_SYNC. We therefore implement this checking
1571 * as a dprintk() in order to avoid filling syslog.
1572 */
1573 static unsigned long complain;
1574
1575 /* Note this will print the MDS for a DS write */
1576 if (time_before(complain, jiffies)) {
1577 dprintk("NFS: faulty NFS server %s:"
1578 " (committed = %d) != (stable = %d)\n",
1579 NFS_SERVER(inode)->nfs_client->cl_hostname,
1580 committed, hdr->args.stable);
1581 complain = jiffies + 300 * HZ;
1582 }
1583 }
1584 }
1585
1586 /* Deal with the suid/sgid bit corner case */
1587 if (nfs_should_remove_suid(inode)) {
1588 spin_lock(&inode->i_lock);
1589 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1590 spin_unlock(&inode->i_lock);
1591 }
1592 return 0;
1593}
1594
1595/*
1596 * This function is called when the WRITE call is complete.
1597 */
1598static void nfs_writeback_result(struct rpc_task *task,
1599 struct nfs_pgio_header *hdr)
1600{
1601 struct nfs_pgio_args *argp = &hdr->args;
1602 struct nfs_pgio_res *resp = &hdr->res;
1603
1604 if (resp->count < argp->count) {
1605 static unsigned long complain;
1606
1607 /* This a short write! */
1608 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1609
1610 /* Has the server at least made some progress? */
1611 if (resp->count == 0) {
1612 if (time_before(complain, jiffies)) {
1613 printk(KERN_WARNING
1614 "NFS: Server wrote zero bytes, expected %u.\n",
1615 argp->count);
1616 complain = jiffies + 300 * HZ;
1617 }
1618 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1619 task->tk_status = -EIO;
1620 return;
1621 }
1622
1623 /* For non rpc-based layout drivers, retry-through-MDS */
1624 if (!task->tk_ops) {
1625 hdr->pnfs_error = -EAGAIN;
1626 return;
1627 }
1628
1629 /* Was this an NFSv2 write or an NFSv3 stable write? */
1630 if (resp->verf->committed != NFS_UNSTABLE) {
1631 /* Resend from where the server left off */
1632 hdr->mds_offset += resp->count;
1633 argp->offset += resp->count;
1634 argp->pgbase += resp->count;
1635 argp->count -= resp->count;
1636 } else {
1637 /* Resend as a stable write in order to avoid
1638 * headaches in the case of a server crash.
1639 */
1640 argp->stable = NFS_FILE_SYNC;
1641 }
1642 resp->count = 0;
1643 resp->verf->committed = 0;
1644 rpc_restart_call_prepare(task);
1645 }
1646}
1647
1648static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1649{
1650 return wait_var_event_killable(&cinfo->rpcs_out,
1651 !atomic_read(&cinfo->rpcs_out));
1652}
1653
1654void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1655{
1656 atomic_inc(&cinfo->rpcs_out);
1657}
1658
1659bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1660{
1661 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1662 wake_up_var(&cinfo->rpcs_out);
1663 return true;
1664 }
1665 return false;
1666}
1667
1668void nfs_commitdata_release(struct nfs_commit_data *data)
1669{
1670 put_nfs_open_context(data->context);
1671 nfs_commit_free(data);
1672}
1673EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1674
1675int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1676 const struct nfs_rpc_ops *nfs_ops,
1677 const struct rpc_call_ops *call_ops,
1678 int how, int flags,
1679 struct nfsd_file *localio)
1680{
1681 struct rpc_task *task;
1682 int priority = flush_task_priority(how);
1683 struct rpc_message msg = {
1684 .rpc_argp = &data->args,
1685 .rpc_resp = &data->res,
1686 .rpc_cred = data->cred,
1687 };
1688 struct rpc_task_setup task_setup_data = {
1689 .task = &data->task,
1690 .rpc_client = clnt,
1691 .rpc_message = &msg,
1692 .callback_ops = call_ops,
1693 .callback_data = data,
1694 .workqueue = nfsiod_workqueue,
1695 .flags = RPC_TASK_ASYNC | flags,
1696 .priority = priority,
1697 };
1698
1699 if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1700 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1701
1702 /* Set up the initial task struct. */
1703 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1704 trace_nfs_initiate_commit(data);
1705
1706 dprintk("NFS: initiated commit call\n");
1707
1708 if (localio)
1709 return nfs_local_commit(localio, data, call_ops, how);
1710
1711 task = rpc_run_task(&task_setup_data);
1712 if (IS_ERR(task))
1713 return PTR_ERR(task);
1714 if (how & FLUSH_SYNC)
1715 rpc_wait_for_completion_task(task);
1716 rpc_put_task(task);
1717 return 0;
1718}
1719EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1720
1721static loff_t nfs_get_lwb(struct list_head *head)
1722{
1723 loff_t lwb = 0;
1724 struct nfs_page *req;
1725
1726 list_for_each_entry(req, head, wb_list)
1727 if (lwb < (req_offset(req) + req->wb_bytes))
1728 lwb = req_offset(req) + req->wb_bytes;
1729
1730 return lwb;
1731}
1732
1733/*
1734 * Set up the argument/result storage required for the RPC call.
1735 */
1736void nfs_init_commit(struct nfs_commit_data *data,
1737 struct list_head *head,
1738 struct pnfs_layout_segment *lseg,
1739 struct nfs_commit_info *cinfo)
1740{
1741 struct nfs_page *first;
1742 struct nfs_open_context *ctx;
1743 struct inode *inode;
1744
1745 /* Set up the RPC argument and reply structs
1746 * NB: take care not to mess about with data->commit et al. */
1747
1748 if (head)
1749 list_splice_init(head, &data->pages);
1750
1751 first = nfs_list_entry(data->pages.next);
1752 ctx = nfs_req_openctx(first);
1753 inode = d_inode(ctx->dentry);
1754
1755 data->inode = inode;
1756 data->cred = ctx->cred;
1757 data->lseg = lseg; /* reference transferred */
1758 /* only set lwb for pnfs commit */
1759 if (lseg)
1760 data->lwb = nfs_get_lwb(&data->pages);
1761 data->mds_ops = &nfs_commit_ops;
1762 data->completion_ops = cinfo->completion_ops;
1763 data->dreq = cinfo->dreq;
1764
1765 data->args.fh = NFS_FH(data->inode);
1766 /* Note: we always request a commit of the entire inode */
1767 data->args.offset = 0;
1768 data->args.count = 0;
1769 data->context = get_nfs_open_context(ctx);
1770 data->res.fattr = &data->fattr;
1771 data->res.verf = &data->verf;
1772 nfs_fattr_init(&data->fattr);
1773 nfs_commit_begin(cinfo->mds);
1774}
1775EXPORT_SYMBOL_GPL(nfs_init_commit);
1776
1777void nfs_retry_commit(struct list_head *page_list,
1778 struct pnfs_layout_segment *lseg,
1779 struct nfs_commit_info *cinfo,
1780 u32 ds_commit_idx)
1781{
1782 struct nfs_page *req;
1783
1784 while (!list_empty(page_list)) {
1785 req = nfs_list_entry(page_list->next);
1786 nfs_list_remove_request(req);
1787 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1788 nfs_folio_clear_commit(nfs_page_to_folio(req));
1789 nfs_unlock_and_release_request(req);
1790 }
1791}
1792EXPORT_SYMBOL_GPL(nfs_retry_commit);
1793
1794static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1795 struct nfs_page *req)
1796{
1797 struct folio *folio = nfs_page_to_folio(req);
1798
1799 filemap_dirty_folio(folio_mapping(folio), folio);
1800}
1801
1802/*
1803 * Commit dirty pages
1804 */
1805static int
1806nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1807 struct nfs_commit_info *cinfo)
1808{
1809 struct nfs_commit_data *data;
1810 struct nfsd_file *localio;
1811 unsigned short task_flags = 0;
1812
1813 /* another commit raced with us */
1814 if (list_empty(head))
1815 return 0;
1816
1817 data = nfs_commitdata_alloc();
1818 if (!data) {
1819 nfs_retry_commit(head, NULL, cinfo, -1);
1820 return -ENOMEM;
1821 }
1822
1823 /* Set up the argument struct */
1824 nfs_init_commit(data, head, NULL, cinfo);
1825 if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1826 task_flags = RPC_TASK_MOVEABLE;
1827
1828 localio = nfs_local_open_fh(NFS_SERVER(inode)->nfs_client, data->cred,
1829 data->args.fh, data->context->mode);
1830 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1831 data->mds_ops, how,
1832 RPC_TASK_CRED_NOREF | task_flags, localio);
1833}
1834
1835/*
1836 * COMMIT call returned
1837 */
1838static void nfs_commit_done(struct rpc_task *task, void *calldata)
1839{
1840 struct nfs_commit_data *data = calldata;
1841
1842 /* Call the NFS version-specific code */
1843 NFS_PROTO(data->inode)->commit_done(task, data);
1844 trace_nfs_commit_done(task, data);
1845}
1846
1847static void nfs_commit_release_pages(struct nfs_commit_data *data)
1848{
1849 const struct nfs_writeverf *verf = data->res.verf;
1850 struct nfs_page *req;
1851 int status = data->task.tk_status;
1852 struct nfs_commit_info cinfo;
1853 struct folio *folio;
1854
1855 while (!list_empty(&data->pages)) {
1856 req = nfs_list_entry(data->pages.next);
1857 nfs_list_remove_request(req);
1858 folio = nfs_page_to_folio(req);
1859 nfs_folio_clear_commit(folio);
1860
1861 dprintk("NFS: commit (%s/%llu %d@%lld)",
1862 nfs_req_openctx(req)->dentry->d_sb->s_id,
1863 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1864 req->wb_bytes,
1865 (long long)req_offset(req));
1866 if (status < 0) {
1867 if (folio) {
1868 trace_nfs_commit_error(data->inode, req,
1869 status);
1870 nfs_mapping_set_error(folio, status);
1871 nfs_inode_remove_request(req);
1872 }
1873 dprintk_cont(", error = %d\n", status);
1874 goto next;
1875 }
1876
1877 /* Okay, COMMIT succeeded, apparently. Check the verifier
1878 * returned by the server against all stored verfs. */
1879 if (nfs_write_match_verf(verf, req)) {
1880 /* We have a match */
1881 if (folio)
1882 nfs_inode_remove_request(req);
1883 dprintk_cont(" OK\n");
1884 goto next;
1885 }
1886 /* We have a mismatch. Write the page again */
1887 dprintk_cont(" mismatch\n");
1888 nfs_mark_request_dirty(req);
1889 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1890 next:
1891 nfs_unlock_and_release_request(req);
1892 /* Latency breaker */
1893 cond_resched();
1894 }
1895
1896 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1897 nfs_commit_end(cinfo.mds);
1898}
1899
1900static void nfs_commit_release(void *calldata)
1901{
1902 struct nfs_commit_data *data = calldata;
1903
1904 data->completion_ops->completion(data);
1905 nfs_commitdata_release(calldata);
1906}
1907
1908static const struct rpc_call_ops nfs_commit_ops = {
1909 .rpc_call_prepare = nfs_commit_prepare,
1910 .rpc_call_done = nfs_commit_done,
1911 .rpc_release = nfs_commit_release,
1912};
1913
1914static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1915 .completion = nfs_commit_release_pages,
1916 .resched_write = nfs_commit_resched_write,
1917};
1918
1919int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1920 int how, struct nfs_commit_info *cinfo)
1921{
1922 int status;
1923
1924 status = pnfs_commit_list(inode, head, how, cinfo);
1925 if (status == PNFS_NOT_ATTEMPTED)
1926 status = nfs_commit_list(inode, head, how, cinfo);
1927 return status;
1928}
1929
1930static int __nfs_commit_inode(struct inode *inode, int how,
1931 struct writeback_control *wbc)
1932{
1933 LIST_HEAD(head);
1934 struct nfs_commit_info cinfo;
1935 int may_wait = how & FLUSH_SYNC;
1936 int ret, nscan;
1937
1938 how &= ~FLUSH_SYNC;
1939 nfs_init_cinfo_from_inode(&cinfo, inode);
1940 nfs_commit_begin(cinfo.mds);
1941 for (;;) {
1942 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1943 if (ret <= 0)
1944 break;
1945 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1946 if (ret < 0)
1947 break;
1948 ret = 0;
1949 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1950 if (nscan < wbc->nr_to_write)
1951 wbc->nr_to_write -= nscan;
1952 else
1953 wbc->nr_to_write = 0;
1954 }
1955 if (nscan < INT_MAX)
1956 break;
1957 cond_resched();
1958 }
1959 nfs_commit_end(cinfo.mds);
1960 if (ret || !may_wait)
1961 return ret;
1962 return wait_on_commit(cinfo.mds);
1963}
1964
1965int nfs_commit_inode(struct inode *inode, int how)
1966{
1967 return __nfs_commit_inode(inode, how, NULL);
1968}
1969EXPORT_SYMBOL_GPL(nfs_commit_inode);
1970
1971int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1972{
1973 struct nfs_inode *nfsi = NFS_I(inode);
1974 int flags = FLUSH_SYNC;
1975 int ret = 0;
1976
1977 if (wbc->sync_mode == WB_SYNC_NONE) {
1978 /* no commits means nothing needs to be done */
1979 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1980 goto check_requests_outstanding;
1981
1982 /* Don't commit yet if this is a non-blocking flush and there
1983 * are a lot of outstanding writes for this mapping.
1984 */
1985 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1986 goto out_mark_dirty;
1987
1988 /* don't wait for the COMMIT response */
1989 flags = 0;
1990 }
1991
1992 ret = __nfs_commit_inode(inode, flags, wbc);
1993 if (!ret) {
1994 if (flags & FLUSH_SYNC)
1995 return 0;
1996 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1997 goto out_mark_dirty;
1998
1999check_requests_outstanding:
2000 if (!atomic_read(&nfsi->commit_info.rpcs_out))
2001 return ret;
2002out_mark_dirty:
2003 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2004 return ret;
2005}
2006EXPORT_SYMBOL_GPL(nfs_write_inode);
2007
2008/*
2009 * Wrapper for filemap_write_and_wait_range()
2010 *
2011 * Needed for pNFS in order to ensure data becomes visible to the
2012 * client.
2013 */
2014int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2015 loff_t lstart, loff_t lend)
2016{
2017 int ret;
2018
2019 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2020 if (ret == 0)
2021 ret = pnfs_sync_inode(mapping->host, true);
2022 return ret;
2023}
2024EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2025
2026/*
2027 * flush the inode to disk.
2028 */
2029int nfs_wb_all(struct inode *inode)
2030{
2031 int ret;
2032
2033 trace_nfs_writeback_inode_enter(inode);
2034
2035 ret = filemap_write_and_wait(inode->i_mapping);
2036 if (ret)
2037 goto out;
2038 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2039 if (ret < 0)
2040 goto out;
2041 pnfs_sync_inode(inode, true);
2042 ret = 0;
2043
2044out:
2045 trace_nfs_writeback_inode_exit(inode, ret);
2046 return ret;
2047}
2048EXPORT_SYMBOL_GPL(nfs_wb_all);
2049
2050int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2051{
2052 struct nfs_page *req;
2053 int ret = 0;
2054
2055 folio_wait_writeback(folio);
2056
2057 /* blocking call to cancel all requests and join to a single (head)
2058 * request */
2059 req = nfs_lock_and_join_requests(folio);
2060
2061 if (IS_ERR(req)) {
2062 ret = PTR_ERR(req);
2063 } else if (req) {
2064 /* all requests from this folio have been cancelled by
2065 * nfs_lock_and_join_requests, so just remove the head
2066 * request from the inode / page_private pointer and
2067 * release it */
2068 nfs_inode_remove_request(req);
2069 nfs_unlock_and_release_request(req);
2070 }
2071
2072 return ret;
2073}
2074
2075/**
2076 * nfs_wb_folio - Write back all requests on one page
2077 * @inode: pointer to page
2078 * @folio: pointer to folio
2079 *
2080 * Assumes that the folio has been locked by the caller, and will
2081 * not unlock it.
2082 */
2083int nfs_wb_folio(struct inode *inode, struct folio *folio)
2084{
2085 loff_t range_start = folio_pos(folio);
2086 size_t len = folio_size(folio);
2087 struct writeback_control wbc = {
2088 .sync_mode = WB_SYNC_ALL,
2089 .nr_to_write = 0,
2090 .range_start = range_start,
2091 .range_end = range_start + len - 1,
2092 };
2093 int ret;
2094
2095 trace_nfs_writeback_folio(inode, range_start, len);
2096
2097 for (;;) {
2098 folio_wait_writeback(folio);
2099 if (folio_clear_dirty_for_io(folio)) {
2100 ret = nfs_writepage_locked(folio, &wbc);
2101 if (ret < 0)
2102 goto out_error;
2103 continue;
2104 }
2105 ret = 0;
2106 if (!folio_test_private(folio))
2107 break;
2108 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2109 if (ret < 0)
2110 goto out_error;
2111 }
2112out_error:
2113 trace_nfs_writeback_folio_done(inode, range_start, len, ret);
2114 return ret;
2115}
2116
2117#ifdef CONFIG_MIGRATION
2118int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2119 struct folio *src, enum migrate_mode mode)
2120{
2121 /*
2122 * If the private flag is set, the folio is currently associated with
2123 * an in-progress read or write request. Don't try to migrate it.
2124 *
2125 * FIXME: we could do this in principle, but we'll need a way to ensure
2126 * that we can safely release the inode reference while holding
2127 * the folio lock.
2128 */
2129 if (folio_test_private(src))
2130 return -EBUSY;
2131
2132 if (folio_test_private_2(src)) { /* [DEPRECATED] */
2133 if (mode == MIGRATE_ASYNC)
2134 return -EBUSY;
2135 folio_wait_private_2(src);
2136 }
2137
2138 return migrate_folio(mapping, dst, src, mode);
2139}
2140#endif
2141
2142int __init nfs_init_writepagecache(void)
2143{
2144 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2145 sizeof(struct nfs_pgio_header),
2146 0, SLAB_HWCACHE_ALIGN,
2147 NULL);
2148 if (nfs_wdata_cachep == NULL)
2149 return -ENOMEM;
2150
2151 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2152 nfs_wdata_cachep);
2153 if (nfs_wdata_mempool == NULL)
2154 goto out_destroy_write_cache;
2155
2156 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2157 sizeof(struct nfs_commit_data),
2158 0, SLAB_HWCACHE_ALIGN,
2159 NULL);
2160 if (nfs_cdata_cachep == NULL)
2161 goto out_destroy_write_mempool;
2162
2163 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2164 nfs_cdata_cachep);
2165 if (nfs_commit_mempool == NULL)
2166 goto out_destroy_commit_cache;
2167
2168 /*
2169 * NFS congestion size, scale with available memory.
2170 *
2171 * 64MB: 8192k
2172 * 128MB: 11585k
2173 * 256MB: 16384k
2174 * 512MB: 23170k
2175 * 1GB: 32768k
2176 * 2GB: 46340k
2177 * 4GB: 65536k
2178 * 8GB: 92681k
2179 * 16GB: 131072k
2180 *
2181 * This allows larger machines to have larger/more transfers.
2182 * Limit the default to 256M
2183 */
2184 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2185 if (nfs_congestion_kb > 256*1024)
2186 nfs_congestion_kb = 256*1024;
2187
2188 return 0;
2189
2190out_destroy_commit_cache:
2191 kmem_cache_destroy(nfs_cdata_cachep);
2192out_destroy_write_mempool:
2193 mempool_destroy(nfs_wdata_mempool);
2194out_destroy_write_cache:
2195 kmem_cache_destroy(nfs_wdata_cachep);
2196 return -ENOMEM;
2197}
2198
2199void nfs_destroy_writepagecache(void)
2200{
2201 mempool_destroy(nfs_commit_mempool);
2202 kmem_cache_destroy(nfs_cdata_cachep);
2203 mempool_destroy(nfs_wdata_mempool);
2204 kmem_cache_destroy(nfs_wdata_cachep);
2205}
2206
2207static const struct nfs_rw_ops nfs_rw_write_ops = {
2208 .rw_alloc_header = nfs_writehdr_alloc,
2209 .rw_free_header = nfs_writehdr_free,
2210 .rw_done = nfs_writeback_done,
2211 .rw_result = nfs_writeback_result,
2212 .rw_initiate = nfs_initiate_write,
2213};