projects/linux·blob master · refs

2213 lines · 58.6 KiB

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