1 // SPDX-License-Identifier: GPL-2.0-only
5 * Write file data over NFS.
7 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
10 #include <linux/types.h>
11 #include <linux/slab.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>
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>
29 #include <linux/uaccess.h>
30 #include <linux/sched/mm.h>
32 #include "delegation.h"
41 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
43 #define MIN_POOL_WRITE (32)
44 #define MIN_POOL_COMMIT (4)
46 struct nfs_io_completion {
47 void (*complete)(void *data);
53 * Local function declarations
55 static void nfs_redirty_request(struct nfs_page *req);
56 static const struct rpc_call_ops nfs_commit_ops;
57 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
58 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
59 static const struct nfs_rw_ops nfs_rw_write_ops;
60 static void nfs_inode_remove_request(struct nfs_page *req);
61 static void nfs_clear_request_commit(struct nfs_page *req);
62 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
64 static struct nfs_page *
65 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
68 static struct kmem_cache *nfs_wdata_cachep;
69 static mempool_t *nfs_wdata_mempool;
70 static struct kmem_cache *nfs_cdata_cachep;
71 static mempool_t *nfs_commit_mempool;
73 struct nfs_commit_data *nfs_commitdata_alloc(void)
75 struct nfs_commit_data *p;
77 p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
79 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
82 memset(p, 0, sizeof(*p));
84 INIT_LIST_HEAD(&p->pages);
87 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
89 void nfs_commit_free(struct nfs_commit_data *p)
91 mempool_free(p, nfs_commit_mempool);
93 EXPORT_SYMBOL_GPL(nfs_commit_free);
95 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
97 struct nfs_pgio_header *p;
99 p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
101 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
104 memset(p, 0, sizeof(*p));
106 p->rw_mode = FMODE_WRITE;
110 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
112 mempool_free(hdr, nfs_wdata_mempool);
115 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
117 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
120 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
121 void (*complete)(void *), void *data)
123 ioc->complete = complete;
125 kref_init(&ioc->refcount);
128 static void nfs_io_completion_release(struct kref *kref)
130 struct nfs_io_completion *ioc = container_of(kref,
131 struct nfs_io_completion, refcount);
132 ioc->complete(ioc->data);
136 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
139 kref_get(&ioc->refcount);
142 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
145 kref_put(&ioc->refcount, nfs_io_completion_release);
149 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
151 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
152 kref_get(&req->wb_kref);
153 atomic_long_inc(&NFS_I(inode)->nrequests);
158 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
162 if (!test_bit(PG_REMOVE, &req->wb_flags))
164 ret = nfs_page_group_lock(req);
167 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
168 nfs_page_set_inode_ref(req, inode);
169 nfs_page_group_unlock(req);
173 static struct nfs_page *
174 nfs_page_private_request(struct page *page)
176 if (!PagePrivate(page))
178 return (struct nfs_page *)page_private(page);
182 * nfs_page_find_head_request_locked - find head request associated with @page
184 * must be called while holding the inode lock.
186 * returns matching head request with reference held, or NULL if not found.
188 static struct nfs_page *
189 nfs_page_find_private_request(struct page *page)
191 struct address_space *mapping = page_file_mapping(page);
192 struct nfs_page *req;
194 if (!PagePrivate(page))
196 spin_lock(&mapping->private_lock);
197 req = nfs_page_private_request(page);
199 WARN_ON_ONCE(req->wb_head != req);
200 kref_get(&req->wb_kref);
202 spin_unlock(&mapping->private_lock);
206 static struct nfs_page *
207 nfs_page_find_swap_request(struct page *page)
209 struct inode *inode = page_file_mapping(page)->host;
210 struct nfs_inode *nfsi = NFS_I(inode);
211 struct nfs_page *req = NULL;
212 if (!PageSwapCache(page))
214 mutex_lock(&nfsi->commit_mutex);
215 if (PageSwapCache(page)) {
216 req = nfs_page_search_commits_for_head_request_locked(nfsi,
219 WARN_ON_ONCE(req->wb_head != req);
220 kref_get(&req->wb_kref);
223 mutex_unlock(&nfsi->commit_mutex);
228 * nfs_page_find_head_request - find head request associated with @page
230 * returns matching head request with reference held, or NULL if not found.
232 static struct nfs_page *nfs_page_find_head_request(struct page *page)
234 struct nfs_page *req;
236 req = nfs_page_find_private_request(page);
238 req = nfs_page_find_swap_request(page);
242 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
244 struct inode *inode = page_file_mapping(page)->host;
245 struct nfs_page *req, *head;
249 req = nfs_page_find_head_request(page);
252 head = nfs_page_group_lock_head(req);
254 nfs_release_request(req);
257 ret = nfs_cancel_remove_inode(head, inode);
259 nfs_unlock_and_release_request(head);
262 /* Ensure that nobody removed the request before we locked it */
263 if (head == nfs_page_private_request(page))
265 if (PageSwapCache(page))
267 nfs_unlock_and_release_request(head);
272 /* Adjust the file length if we're writing beyond the end */
273 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
275 struct inode *inode = page_file_mapping(page)->host;
279 spin_lock(&inode->i_lock);
280 i_size = i_size_read(inode);
281 end_index = (i_size - 1) >> PAGE_SHIFT;
282 if (i_size > 0 && page_index(page) < end_index)
284 end = page_file_offset(page) + ((loff_t)offset+count);
287 trace_nfs_size_grow(inode, end);
288 i_size_write(inode, end);
289 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
290 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
292 spin_unlock(&inode->i_lock);
293 nfs_fscache_invalidate(inode, 0);
296 /* A writeback failed: mark the page as bad, and invalidate the page cache */
297 static void nfs_set_pageerror(struct address_space *mapping)
299 struct inode *inode = mapping->host;
301 nfs_zap_mapping(mapping->host, mapping);
302 /* Force file size revalidation */
303 spin_lock(&inode->i_lock);
304 nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
305 NFS_INO_INVALID_CHANGE |
306 NFS_INO_INVALID_SIZE);
307 spin_unlock(&inode->i_lock);
310 static void nfs_mapping_set_error(struct page *page, int error)
312 struct address_space *mapping = page_file_mapping(page);
315 filemap_set_wb_err(mapping, error);
317 errseq_set(&mapping->host->i_sb->s_wb_err,
318 error == -ENOSPC ? -ENOSPC : -EIO);
319 nfs_set_pageerror(mapping);
323 * nfs_page_group_search_locked
324 * @head - head request of page group
325 * @page_offset - offset into page
327 * Search page group with head @head to find a request that contains the
328 * page offset @page_offset.
330 * Returns a pointer to the first matching nfs request, or NULL if no
333 * Must be called with the page group lock held
335 static struct nfs_page *
336 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
338 struct nfs_page *req;
342 if (page_offset >= req->wb_pgbase &&
343 page_offset < (req->wb_pgbase + req->wb_bytes))
346 req = req->wb_this_page;
347 } while (req != head);
353 * nfs_page_group_covers_page
354 * @head - head request of page group
356 * Return true if the page group with head @head covers the whole page,
357 * returns false otherwise
359 static bool nfs_page_group_covers_page(struct nfs_page *req)
361 struct nfs_page *tmp;
362 unsigned int pos = 0;
363 unsigned int len = nfs_page_length(req->wb_page);
365 nfs_page_group_lock(req);
368 tmp = nfs_page_group_search_locked(req->wb_head, pos);
371 pos = tmp->wb_pgbase + tmp->wb_bytes;
374 nfs_page_group_unlock(req);
378 /* We can set the PG_uptodate flag if we see that a write request
379 * covers the full page.
381 static void nfs_mark_uptodate(struct nfs_page *req)
383 if (PageUptodate(req->wb_page))
385 if (!nfs_page_group_covers_page(req))
387 SetPageUptodate(req->wb_page);
390 static int wb_priority(struct writeback_control *wbc)
394 if (wbc->sync_mode == WB_SYNC_ALL)
395 ret = FLUSH_COND_STABLE;
400 * NFS congestion control
403 int nfs_congestion_kb;
405 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
406 #define NFS_CONGESTION_OFF_THRESH \
407 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
409 static void nfs_set_page_writeback(struct page *page)
411 struct inode *inode = page_file_mapping(page)->host;
412 struct nfs_server *nfss = NFS_SERVER(inode);
413 int ret = test_set_page_writeback(page);
415 WARN_ON_ONCE(ret != 0);
417 if (atomic_long_inc_return(&nfss->writeback) >
418 NFS_CONGESTION_ON_THRESH)
419 nfss->write_congested = 1;
422 static void nfs_end_page_writeback(struct nfs_page *req)
424 struct inode *inode = page_file_mapping(req->wb_page)->host;
425 struct nfs_server *nfss = NFS_SERVER(inode);
428 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
429 nfs_unlock_request(req);
433 end_page_writeback(req->wb_page);
434 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
435 nfss->write_congested = 0;
439 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
441 * @destroy_list - request list (using wb_this_page) terminated by @old_head
442 * @old_head - the old head of the list
444 * All subrequests must be locked and removed from all lists, so at this point
445 * they are only "active" in this function, and possibly in nfs_wait_on_request
446 * with a reference held by some other context.
449 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
450 struct nfs_page *old_head,
453 while (destroy_list) {
454 struct nfs_page *subreq = destroy_list;
456 destroy_list = (subreq->wb_this_page == old_head) ?
457 NULL : subreq->wb_this_page;
459 /* Note: lock subreq in order to change subreq->wb_head */
460 nfs_page_set_headlock(subreq);
461 WARN_ON_ONCE(old_head != subreq->wb_head);
463 /* make sure old group is not used */
464 subreq->wb_this_page = subreq;
465 subreq->wb_head = subreq;
467 clear_bit(PG_REMOVE, &subreq->wb_flags);
469 /* Note: races with nfs_page_group_destroy() */
470 if (!kref_read(&subreq->wb_kref)) {
471 /* Check if we raced with nfs_page_group_destroy() */
472 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
473 nfs_page_clear_headlock(subreq);
474 nfs_free_request(subreq);
476 nfs_page_clear_headlock(subreq);
479 nfs_page_clear_headlock(subreq);
481 nfs_release_request(old_head);
483 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
484 nfs_release_request(subreq);
485 atomic_long_dec(&NFS_I(inode)->nrequests);
488 /* subreq is now totally disconnected from page group or any
489 * write / commit lists. last chance to wake any waiters */
490 nfs_unlock_and_release_request(subreq);
495 * nfs_join_page_group - destroy subrequests of the head req
496 * @head: the page used to lookup the "page group" of nfs_page structures
497 * @inode: Inode to which the request belongs.
499 * This function joins all sub requests to the head request by first
500 * locking all requests in the group, cancelling any pending operations
501 * and finally updating the head request to cover the whole range covered by
502 * the (former) group. All subrequests are removed from any write or commit
503 * lists, unlinked from the group and destroyed.
506 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
508 struct nfs_page *subreq;
509 struct nfs_page *destroy_list = NULL;
510 unsigned int pgbase, off, bytes;
512 pgbase = head->wb_pgbase;
513 bytes = head->wb_bytes;
514 off = head->wb_offset;
515 for (subreq = head->wb_this_page; subreq != head;
516 subreq = subreq->wb_this_page) {
517 /* Subrequests should always form a contiguous range */
518 if (pgbase > subreq->wb_pgbase) {
519 off -= pgbase - subreq->wb_pgbase;
520 bytes += pgbase - subreq->wb_pgbase;
521 pgbase = subreq->wb_pgbase;
523 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
527 /* Set the head request's range to cover the former page group */
528 head->wb_pgbase = pgbase;
529 head->wb_bytes = bytes;
530 head->wb_offset = off;
532 /* Now that all requests are locked, make sure they aren't on any list.
533 * Commit list removal accounting is done after locks are dropped */
536 nfs_clear_request_commit(subreq);
537 subreq = subreq->wb_this_page;
538 } while (subreq != head);
540 /* unlink subrequests from head, destroy them later */
541 if (head->wb_this_page != head) {
542 /* destroy list will be terminated by head */
543 destroy_list = head->wb_this_page;
544 head->wb_this_page = head;
547 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
551 * nfs_lock_and_join_requests - join all subreqs to the head req
552 * @page: the page used to lookup the "page group" of nfs_page structures
554 * This function joins all sub requests to the head request by first
555 * locking all requests in the group, cancelling any pending operations
556 * and finally updating the head request to cover the whole range covered by
557 * the (former) group. All subrequests are removed from any write or commit
558 * lists, unlinked from the group and destroyed.
560 * Returns a locked, referenced pointer to the head request - which after
561 * this call is guaranteed to be the only request associated with the page.
562 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
563 * error was encountered.
565 static struct nfs_page *
566 nfs_lock_and_join_requests(struct page *page)
568 struct inode *inode = page_file_mapping(page)->host;
569 struct nfs_page *head;
573 * A reference is taken only on the head request which acts as a
574 * reference to the whole page group - the group will not be destroyed
575 * until the head reference is released.
577 head = nfs_find_and_lock_page_request(page);
578 if (IS_ERR_OR_NULL(head))
581 /* lock each request in the page group */
582 ret = nfs_page_group_lock_subrequests(head);
584 nfs_unlock_and_release_request(head);
588 nfs_join_page_group(head, inode);
593 static void nfs_write_error(struct nfs_page *req, int error)
595 trace_nfs_write_error(req, error);
596 nfs_mapping_set_error(req->wb_page, error);
597 nfs_inode_remove_request(req);
598 nfs_end_page_writeback(req);
599 nfs_release_request(req);
603 * Find an associated nfs write request, and prepare to flush it out
604 * May return an error if the user signalled nfs_wait_on_request().
606 static int nfs_page_async_flush(struct page *page,
607 struct writeback_control *wbc,
608 struct nfs_pageio_descriptor *pgio)
610 struct nfs_page *req;
613 req = nfs_lock_and_join_requests(page);
620 nfs_set_page_writeback(page);
621 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
623 /* If there is a fatal error that covers this write, just exit */
624 ret = pgio->pg_error;
625 if (nfs_error_is_fatal_on_server(ret))
629 if (!nfs_pageio_add_request(pgio, req)) {
630 ret = pgio->pg_error;
632 * Remove the problematic req upon fatal errors on the server
634 if (nfs_error_is_fatal_on_server(ret))
636 if (wbc->sync_mode == WB_SYNC_NONE)
637 ret = AOP_WRITEPAGE_ACTIVATE;
638 redirty_page_for_writepage(wbc, page);
639 nfs_redirty_request(req);
642 nfs_add_stats(page_file_mapping(page)->host,
643 NFSIOS_WRITEPAGES, 1);
647 nfs_write_error(req, ret);
651 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
652 struct nfs_pageio_descriptor *pgio)
654 nfs_pageio_cond_complete(pgio, page_index(page));
655 return nfs_page_async_flush(page, wbc, pgio);
659 * Write an mmapped page to the server.
661 static int nfs_writepage_locked(struct page *page,
662 struct writeback_control *wbc)
664 struct nfs_pageio_descriptor pgio;
665 struct inode *inode = page_file_mapping(page)->host;
668 if (wbc->sync_mode == WB_SYNC_NONE &&
669 NFS_SERVER(inode)->write_congested)
670 return AOP_WRITEPAGE_ACTIVATE;
672 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
673 nfs_pageio_init_write(&pgio, inode, 0,
674 false, &nfs_async_write_completion_ops);
675 err = nfs_do_writepage(page, wbc, &pgio);
677 nfs_pageio_complete(&pgio);
681 int nfs_writepage(struct page *page, struct writeback_control *wbc)
685 ret = nfs_writepage_locked(page, wbc);
686 if (ret != AOP_WRITEPAGE_ACTIVATE)
691 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
695 ret = nfs_do_writepage(page, wbc, data);
696 if (ret != AOP_WRITEPAGE_ACTIVATE)
701 static void nfs_io_completion_commit(void *inode)
703 nfs_commit_inode(inode, 0);
706 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
708 struct inode *inode = mapping->host;
709 struct nfs_pageio_descriptor pgio;
710 struct nfs_io_completion *ioc = NULL;
711 unsigned int mntflags = NFS_SERVER(inode)->flags;
715 if (wbc->sync_mode == WB_SYNC_NONE &&
716 NFS_SERVER(inode)->write_congested)
719 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
721 if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
722 wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
723 ioc = nfs_io_completion_alloc(GFP_KERNEL);
725 nfs_io_completion_init(ioc, nfs_io_completion_commit,
727 priority = wb_priority(wbc);
731 nfs_pageio_init_write(&pgio, inode, priority, false,
732 &nfs_async_write_completion_ops);
733 pgio.pg_io_completion = ioc;
734 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
737 nfs_pageio_complete(&pgio);
738 } while (err < 0 && !nfs_error_is_fatal(err));
739 nfs_io_completion_put(ioc);
749 * Insert a write request into an inode
751 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
753 struct address_space *mapping = page_file_mapping(req->wb_page);
754 struct nfs_inode *nfsi = NFS_I(inode);
756 WARN_ON_ONCE(req->wb_this_page != req);
758 /* Lock the request! */
759 nfs_lock_request(req);
762 * Swap-space should not get truncated. Hence no need to plug the race
763 * with invalidate/truncate.
765 spin_lock(&mapping->private_lock);
766 if (likely(!PageSwapCache(req->wb_page))) {
767 set_bit(PG_MAPPED, &req->wb_flags);
768 SetPagePrivate(req->wb_page);
769 set_page_private(req->wb_page, (unsigned long)req);
771 spin_unlock(&mapping->private_lock);
772 atomic_long_inc(&nfsi->nrequests);
773 /* this a head request for a page group - mark it as having an
774 * extra reference so sub groups can follow suit.
775 * This flag also informs pgio layer when to bump nrequests when
776 * adding subrequests. */
777 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
778 kref_get(&req->wb_kref);
782 * Remove a write request from an inode
784 static void nfs_inode_remove_request(struct nfs_page *req)
786 struct address_space *mapping = page_file_mapping(req->wb_page);
787 struct inode *inode = mapping->host;
788 struct nfs_inode *nfsi = NFS_I(inode);
789 struct nfs_page *head;
791 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
794 spin_lock(&mapping->private_lock);
795 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
796 set_page_private(head->wb_page, 0);
797 ClearPagePrivate(head->wb_page);
798 clear_bit(PG_MAPPED, &head->wb_flags);
800 spin_unlock(&mapping->private_lock);
803 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
804 nfs_release_request(req);
805 atomic_long_dec(&nfsi->nrequests);
810 nfs_mark_request_dirty(struct nfs_page *req)
813 __set_page_dirty_nobuffers(req->wb_page);
817 * nfs_page_search_commits_for_head_request_locked
819 * Search through commit lists on @inode for the head request for @page.
820 * Must be called while holding the inode (which is cinfo) lock.
822 * Returns the head request if found, or NULL if not found.
824 static struct nfs_page *
825 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
828 struct nfs_page *freq, *t;
829 struct nfs_commit_info cinfo;
830 struct inode *inode = &nfsi->vfs_inode;
832 nfs_init_cinfo_from_inode(&cinfo, inode);
834 /* search through pnfs commit lists */
835 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
837 return freq->wb_head;
839 /* Linearly search the commit list for the correct request */
840 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
841 if (freq->wb_page == page)
842 return freq->wb_head;
849 * nfs_request_add_commit_list_locked - add request to a commit list
850 * @req: pointer to a struct nfs_page
851 * @dst: commit list head
852 * @cinfo: holds list lock and accounting info
854 * This sets the PG_CLEAN bit, updates the cinfo count of
855 * number of outstanding requests requiring a commit as well as
858 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
862 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
863 struct nfs_commit_info *cinfo)
865 set_bit(PG_CLEAN, &req->wb_flags);
866 nfs_list_add_request(req, dst);
867 atomic_long_inc(&cinfo->mds->ncommit);
869 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
872 * nfs_request_add_commit_list - add request to a commit list
873 * @req: pointer to a struct nfs_page
874 * @cinfo: holds list lock and accounting info
876 * This sets the PG_CLEAN bit, updates the cinfo count of
877 * number of outstanding requests requiring a commit as well as
880 * The caller must _not_ hold the cinfo->lock, but must be
881 * holding the nfs_page lock.
884 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
886 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
887 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
888 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
890 nfs_mark_page_unstable(req->wb_page, cinfo);
892 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
895 * nfs_request_remove_commit_list - Remove request from a commit list
896 * @req: pointer to a nfs_page
897 * @cinfo: holds list lock and accounting info
899 * This clears the PG_CLEAN bit, and updates the cinfo's count of
900 * number of outstanding requests requiring a commit
901 * It does not update the MM page stats.
903 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
906 nfs_request_remove_commit_list(struct nfs_page *req,
907 struct nfs_commit_info *cinfo)
909 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
911 nfs_list_remove_request(req);
912 atomic_long_dec(&cinfo->mds->ncommit);
914 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
916 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
919 cinfo->inode = inode;
920 cinfo->mds = &NFS_I(inode)->commit_info;
921 cinfo->ds = pnfs_get_ds_info(inode);
923 cinfo->completion_ops = &nfs_commit_completion_ops;
926 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
928 struct nfs_direct_req *dreq)
931 nfs_init_cinfo_from_dreq(cinfo, dreq);
933 nfs_init_cinfo_from_inode(cinfo, inode);
935 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
938 * Add a request to the inode's commit list.
941 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
942 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
944 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
946 nfs_request_add_commit_list(req, cinfo);
950 nfs_clear_page_commit(struct page *page)
952 dec_node_page_state(page, NR_WRITEBACK);
953 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
957 /* Called holding the request lock on @req */
959 nfs_clear_request_commit(struct nfs_page *req)
961 if (test_bit(PG_CLEAN, &req->wb_flags)) {
962 struct nfs_open_context *ctx = nfs_req_openctx(req);
963 struct inode *inode = d_inode(ctx->dentry);
964 struct nfs_commit_info cinfo;
966 nfs_init_cinfo_from_inode(&cinfo, inode);
967 mutex_lock(&NFS_I(inode)->commit_mutex);
968 if (!pnfs_clear_request_commit(req, &cinfo)) {
969 nfs_request_remove_commit_list(req, &cinfo);
971 mutex_unlock(&NFS_I(inode)->commit_mutex);
972 nfs_clear_page_commit(req->wb_page);
976 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
978 if (hdr->verf.committed == NFS_DATA_SYNC)
979 return hdr->lseg == NULL;
980 return hdr->verf.committed != NFS_FILE_SYNC;
983 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
985 nfs_io_completion_get(hdr->io_completion);
988 static void nfs_write_completion(struct nfs_pgio_header *hdr)
990 struct nfs_commit_info cinfo;
991 unsigned long bytes = 0;
993 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
995 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
996 while (!list_empty(&hdr->pages)) {
997 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
999 bytes += req->wb_bytes;
1000 nfs_list_remove_request(req);
1001 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1002 (hdr->good_bytes < bytes)) {
1003 trace_nfs_comp_error(req, hdr->error);
1004 nfs_mapping_set_error(req->wb_page, hdr->error);
1007 if (nfs_write_need_commit(hdr)) {
1008 /* Reset wb_nio, since the write was successful. */
1010 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1011 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1012 hdr->pgio_mirror_idx);
1016 nfs_inode_remove_request(req);
1018 nfs_end_page_writeback(req);
1019 nfs_release_request(req);
1022 nfs_io_completion_put(hdr->io_completion);
1027 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1029 return atomic_long_read(&cinfo->mds->ncommit);
1032 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1034 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1035 struct nfs_commit_info *cinfo, int max)
1037 struct nfs_page *req, *tmp;
1040 list_for_each_entry_safe(req, tmp, src, wb_list) {
1041 kref_get(&req->wb_kref);
1042 if (!nfs_lock_request(req)) {
1043 nfs_release_request(req);
1046 nfs_request_remove_commit_list(req, cinfo);
1047 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1048 nfs_list_add_request(req, dst);
1050 if ((ret == max) && !cinfo->dreq)
1056 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1059 * nfs_scan_commit - Scan an inode for commit requests
1060 * @inode: NFS inode to scan
1061 * @dst: mds destination list
1062 * @cinfo: mds and ds lists of reqs ready to commit
1064 * Moves requests from the inode's 'commit' request list.
1065 * The requests are *not* checked to ensure that they form a contiguous set.
1068 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1069 struct nfs_commit_info *cinfo)
1073 if (!atomic_long_read(&cinfo->mds->ncommit))
1075 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1076 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1077 const int max = INT_MAX;
1079 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1081 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1083 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1088 * Search for an existing write request, and attempt to update
1089 * it to reflect a new dirty region on a given page.
1091 * If the attempt fails, then the existing request is flushed out
1094 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1096 unsigned int offset,
1099 struct nfs_page *req;
1104 end = offset + bytes;
1106 req = nfs_lock_and_join_requests(page);
1107 if (IS_ERR_OR_NULL(req))
1110 rqend = req->wb_offset + req->wb_bytes;
1112 * Tell the caller to flush out the request if
1113 * the offsets are non-contiguous.
1114 * Note: nfs_flush_incompatible() will already
1115 * have flushed out requests having wrong owners.
1117 if (offset > rqend || end < req->wb_offset)
1120 /* Okay, the request matches. Update the region */
1121 if (offset < req->wb_offset) {
1122 req->wb_offset = offset;
1123 req->wb_pgbase = offset;
1126 req->wb_bytes = end - req->wb_offset;
1128 req->wb_bytes = rqend - req->wb_offset;
1133 * Note: we mark the request dirty here because
1134 * nfs_lock_and_join_requests() cannot preserve
1135 * commit flags, so we have to replay the write.
1137 nfs_mark_request_dirty(req);
1138 nfs_unlock_and_release_request(req);
1139 error = nfs_wb_page(inode, page);
1140 return (error < 0) ? ERR_PTR(error) : NULL;
1144 * Try to update an existing write request, or create one if there is none.
1146 * Note: Should always be called with the Page Lock held to prevent races
1147 * if we have to add a new request. Also assumes that the caller has
1148 * already called nfs_flush_incompatible() if necessary.
1150 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1151 struct page *page, unsigned int offset, unsigned int bytes)
1153 struct inode *inode = page_file_mapping(page)->host;
1154 struct nfs_page *req;
1156 req = nfs_try_to_update_request(inode, page, offset, bytes);
1159 req = nfs_create_request(ctx, page, offset, bytes);
1162 nfs_inode_add_request(inode, req);
1167 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1168 unsigned int offset, unsigned int count)
1170 struct nfs_page *req;
1172 req = nfs_setup_write_request(ctx, page, offset, count);
1174 return PTR_ERR(req);
1175 /* Update file length */
1176 nfs_grow_file(page, offset, count);
1177 nfs_mark_uptodate(req);
1178 nfs_mark_request_dirty(req);
1179 nfs_unlock_and_release_request(req);
1183 int nfs_flush_incompatible(struct file *file, struct page *page)
1185 struct nfs_open_context *ctx = nfs_file_open_context(file);
1186 struct nfs_lock_context *l_ctx;
1187 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1188 struct nfs_page *req;
1189 int do_flush, status;
1191 * Look for a request corresponding to this page. If there
1192 * is one, and it belongs to another file, we flush it out
1193 * before we try to copy anything into the page. Do this
1194 * due to the lack of an ACCESS-type call in NFSv2.
1195 * Also do the same if we find a request from an existing
1199 req = nfs_page_find_head_request(page);
1202 l_ctx = req->wb_lock_context;
1203 do_flush = req->wb_page != page ||
1204 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1205 if (l_ctx && flctx &&
1206 !(list_empty_careful(&flctx->flc_posix) &&
1207 list_empty_careful(&flctx->flc_flock))) {
1208 do_flush |= l_ctx->lockowner != current->files;
1210 nfs_release_request(req);
1213 status = nfs_wb_page(page_file_mapping(page)->host, page);
1214 } while (status == 0);
1219 * Avoid buffered writes when a open context credential's key would
1222 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1224 * Return 0 and set a credential flag which triggers the inode to flush
1225 * and performs NFS_FILE_SYNC writes if the key will expired within
1226 * RPC_KEY_EXPIRE_TIMEO.
1229 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1231 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1233 if (nfs_ctx_key_to_expire(ctx, inode) &&
1234 !rcu_access_pointer(ctx->ll_cred))
1235 /* Already expired! */
1241 * Test if the open context credential key is marked to expire soon.
1243 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1245 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1246 struct rpc_cred *cred, *new, *old = NULL;
1247 struct auth_cred acred = {
1253 cred = rcu_dereference(ctx->ll_cred);
1254 if (cred && !(cred->cr_ops->crkey_timeout &&
1255 cred->cr_ops->crkey_timeout(cred)))
1259 new = auth->au_ops->lookup_cred(auth, &acred, 0);
1264 if (IS_ERR_OR_NULL(new)) {
1267 } else if (new->cr_ops->crkey_timeout &&
1268 new->cr_ops->crkey_timeout(new))
1272 old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1273 RCU_INITIALIZER(new)), 1);
1281 * If the page cache is marked as unsafe or invalid, then we can't rely on
1282 * the PageUptodate() flag. In this case, we will need to turn off
1283 * write optimisations that depend on the page contents being correct.
1285 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1286 unsigned int pagelen)
1288 struct nfs_inode *nfsi = NFS_I(inode);
1290 if (nfs_have_delegated_attributes(inode))
1292 if (nfsi->cache_validity &
1293 (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1296 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1299 if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1301 return PageUptodate(page) != 0;
1305 is_whole_file_wrlock(struct file_lock *fl)
1307 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1308 fl->fl_type == F_WRLCK;
1311 /* If we know the page is up to date, and we're not using byte range locks (or
1312 * if we have the whole file locked for writing), it may be more efficient to
1313 * extend the write to cover the entire page in order to avoid fragmentation
1316 * If the file is opened for synchronous writes then we can just skip the rest
1319 static int nfs_can_extend_write(struct file *file, struct page *page,
1320 struct inode *inode, unsigned int pagelen)
1323 struct file_lock_context *flctx = inode->i_flctx;
1324 struct file_lock *fl;
1326 if (file->f_flags & O_DSYNC)
1328 if (!nfs_write_pageuptodate(page, inode, pagelen))
1330 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1332 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1333 list_empty_careful(&flctx->flc_posix)))
1336 /* Check to see if there are whole file write locks */
1338 spin_lock(&flctx->flc_lock);
1339 if (!list_empty(&flctx->flc_posix)) {
1340 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1342 if (is_whole_file_wrlock(fl))
1344 } else if (!list_empty(&flctx->flc_flock)) {
1345 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1347 if (fl->fl_type == F_WRLCK)
1350 spin_unlock(&flctx->flc_lock);
1355 * Update and possibly write a cached page of an NFS file.
1357 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1358 * things with a page scheduled for an RPC call (e.g. invalidate it).
1360 int nfs_updatepage(struct file *file, struct page *page,
1361 unsigned int offset, unsigned int count)
1363 struct nfs_open_context *ctx = nfs_file_open_context(file);
1364 struct address_space *mapping = page_file_mapping(page);
1365 struct inode *inode = mapping->host;
1366 unsigned int pagelen = nfs_page_length(page);
1369 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1371 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1372 file, count, (long long)(page_file_offset(page) + offset));
1377 if (nfs_can_extend_write(file, page, inode, pagelen)) {
1378 count = max(count + offset, pagelen);
1382 status = nfs_writepage_setup(ctx, page, offset, count);
1384 nfs_set_pageerror(mapping);
1386 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1387 status, (long long)i_size_read(inode));
1391 static int flush_task_priority(int how)
1393 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1395 return RPC_PRIORITY_HIGH;
1397 return RPC_PRIORITY_LOW;
1399 return RPC_PRIORITY_NORMAL;
1402 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1403 struct rpc_message *msg,
1404 const struct nfs_rpc_ops *rpc_ops,
1405 struct rpc_task_setup *task_setup_data, int how)
1407 int priority = flush_task_priority(how);
1409 if (IS_SWAPFILE(hdr->inode))
1410 task_setup_data->flags |= RPC_TASK_SWAPPER;
1411 task_setup_data->priority = priority;
1412 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1413 trace_nfs_initiate_write(hdr);
1416 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1417 * call this on each, which will prepare them to be retried on next
1418 * writeback using standard nfs.
1420 static void nfs_redirty_request(struct nfs_page *req)
1422 /* Bump the transmission count */
1424 nfs_mark_request_dirty(req);
1425 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1426 nfs_end_page_writeback(req);
1427 nfs_release_request(req);
1430 static void nfs_async_write_error(struct list_head *head, int error)
1432 struct nfs_page *req;
1434 while (!list_empty(head)) {
1435 req = nfs_list_entry(head->next);
1436 nfs_list_remove_request(req);
1437 if (nfs_error_is_fatal_on_server(error))
1438 nfs_write_error(req, error);
1440 nfs_redirty_request(req);
1444 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1446 nfs_async_write_error(&hdr->pages, 0);
1447 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1448 hdr->args.offset + hdr->args.count - 1);
1451 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1452 .init_hdr = nfs_async_write_init,
1453 .error_cleanup = nfs_async_write_error,
1454 .completion = nfs_write_completion,
1455 .reschedule_io = nfs_async_write_reschedule_io,
1458 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1459 struct inode *inode, int ioflags, bool force_mds,
1460 const struct nfs_pgio_completion_ops *compl_ops)
1462 struct nfs_server *server = NFS_SERVER(inode);
1463 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1465 #ifdef CONFIG_NFS_V4_1
1466 if (server->pnfs_curr_ld && !force_mds)
1467 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1469 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1470 server->wsize, ioflags);
1472 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1474 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1476 struct nfs_pgio_mirror *mirror;
1478 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1479 pgio->pg_ops->pg_cleanup(pgio);
1481 pgio->pg_ops = &nfs_pgio_rw_ops;
1483 nfs_pageio_stop_mirroring(pgio);
1485 mirror = &pgio->pg_mirrors[0];
1486 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1488 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1491 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1493 struct nfs_commit_data *data = calldata;
1495 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1499 * Special version of should_remove_suid() that ignores capabilities.
1501 static int nfs_should_remove_suid(const struct inode *inode)
1503 umode_t mode = inode->i_mode;
1506 /* suid always must be killed */
1507 if (unlikely(mode & S_ISUID))
1508 kill = ATTR_KILL_SUID;
1511 * sgid without any exec bits is just a mandatory locking mark; leave
1512 * it alone. If some exec bits are set, it's a real sgid; kill it.
1514 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1515 kill |= ATTR_KILL_SGID;
1517 if (unlikely(kill && S_ISREG(mode)))
1523 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1524 struct nfs_fattr *fattr)
1526 struct nfs_pgio_args *argp = &hdr->args;
1527 struct nfs_pgio_res *resp = &hdr->res;
1528 u64 size = argp->offset + resp->count;
1530 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1532 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1533 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1536 if (size != fattr->size)
1538 /* Set attribute barrier */
1539 nfs_fattr_set_barrier(fattr);
1540 /* ...and update size */
1541 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1544 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1546 struct nfs_fattr *fattr = &hdr->fattr;
1547 struct inode *inode = hdr->inode;
1549 spin_lock(&inode->i_lock);
1550 nfs_writeback_check_extend(hdr, fattr);
1551 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1552 spin_unlock(&inode->i_lock);
1554 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1557 * This function is called when the WRITE call is complete.
1559 static int nfs_writeback_done(struct rpc_task *task,
1560 struct nfs_pgio_header *hdr,
1561 struct inode *inode)
1566 * ->write_done will attempt to use post-op attributes to detect
1567 * conflicting writes by other clients. A strict interpretation
1568 * of close-to-open would allow us to continue caching even if
1569 * another writer had changed the file, but some applications
1570 * depend on tighter cache coherency when writing.
1572 status = NFS_PROTO(inode)->write_done(task, hdr);
1576 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1577 trace_nfs_writeback_done(task, hdr);
1579 if (hdr->res.verf->committed < hdr->args.stable &&
1580 task->tk_status >= 0) {
1581 /* We tried a write call, but the server did not
1582 * commit data to stable storage even though we
1584 * Note: There is a known bug in Tru64 < 5.0 in which
1585 * the server reports NFS_DATA_SYNC, but performs
1586 * NFS_FILE_SYNC. We therefore implement this checking
1587 * as a dprintk() in order to avoid filling syslog.
1589 static unsigned long complain;
1591 /* Note this will print the MDS for a DS write */
1592 if (time_before(complain, jiffies)) {
1593 dprintk("NFS: faulty NFS server %s:"
1594 " (committed = %d) != (stable = %d)\n",
1595 NFS_SERVER(inode)->nfs_client->cl_hostname,
1596 hdr->res.verf->committed, hdr->args.stable);
1597 complain = jiffies + 300 * HZ;
1601 /* Deal with the suid/sgid bit corner case */
1602 if (nfs_should_remove_suid(inode)) {
1603 spin_lock(&inode->i_lock);
1604 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1605 spin_unlock(&inode->i_lock);
1611 * This function is called when the WRITE call is complete.
1613 static void nfs_writeback_result(struct rpc_task *task,
1614 struct nfs_pgio_header *hdr)
1616 struct nfs_pgio_args *argp = &hdr->args;
1617 struct nfs_pgio_res *resp = &hdr->res;
1619 if (resp->count < argp->count) {
1620 static unsigned long complain;
1622 /* This a short write! */
1623 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1625 /* Has the server at least made some progress? */
1626 if (resp->count == 0) {
1627 if (time_before(complain, jiffies)) {
1629 "NFS: Server wrote zero bytes, expected %u.\n",
1631 complain = jiffies + 300 * HZ;
1633 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1634 task->tk_status = -EIO;
1638 /* For non rpc-based layout drivers, retry-through-MDS */
1639 if (!task->tk_ops) {
1640 hdr->pnfs_error = -EAGAIN;
1644 /* Was this an NFSv2 write or an NFSv3 stable write? */
1645 if (resp->verf->committed != NFS_UNSTABLE) {
1646 /* Resend from where the server left off */
1647 hdr->mds_offset += resp->count;
1648 argp->offset += resp->count;
1649 argp->pgbase += resp->count;
1650 argp->count -= resp->count;
1652 /* Resend as a stable write in order to avoid
1653 * headaches in the case of a server crash.
1655 argp->stable = NFS_FILE_SYNC;
1658 resp->verf->committed = 0;
1659 rpc_restart_call_prepare(task);
1663 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1665 return wait_var_event_killable(&cinfo->rpcs_out,
1666 !atomic_read(&cinfo->rpcs_out));
1669 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1671 atomic_inc(&cinfo->rpcs_out);
1674 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1676 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1677 wake_up_var(&cinfo->rpcs_out);
1683 void nfs_commitdata_release(struct nfs_commit_data *data)
1685 put_nfs_open_context(data->context);
1686 nfs_commit_free(data);
1688 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1690 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1691 const struct nfs_rpc_ops *nfs_ops,
1692 const struct rpc_call_ops *call_ops,
1695 struct rpc_task *task;
1696 int priority = flush_task_priority(how);
1697 struct rpc_message msg = {
1698 .rpc_argp = &data->args,
1699 .rpc_resp = &data->res,
1700 .rpc_cred = data->cred,
1702 struct rpc_task_setup task_setup_data = {
1703 .task = &data->task,
1705 .rpc_message = &msg,
1706 .callback_ops = call_ops,
1707 .callback_data = data,
1708 .workqueue = nfsiod_workqueue,
1709 .flags = RPC_TASK_ASYNC | flags,
1710 .priority = priority,
1713 if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1714 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1716 /* Set up the initial task struct. */
1717 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1718 trace_nfs_initiate_commit(data);
1720 dprintk("NFS: initiated commit call\n");
1722 task = rpc_run_task(&task_setup_data);
1724 return PTR_ERR(task);
1725 if (how & FLUSH_SYNC)
1726 rpc_wait_for_completion_task(task);
1730 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1732 static loff_t nfs_get_lwb(struct list_head *head)
1735 struct nfs_page *req;
1737 list_for_each_entry(req, head, wb_list)
1738 if (lwb < (req_offset(req) + req->wb_bytes))
1739 lwb = req_offset(req) + req->wb_bytes;
1745 * Set up the argument/result storage required for the RPC call.
1747 void nfs_init_commit(struct nfs_commit_data *data,
1748 struct list_head *head,
1749 struct pnfs_layout_segment *lseg,
1750 struct nfs_commit_info *cinfo)
1752 struct nfs_page *first;
1753 struct nfs_open_context *ctx;
1754 struct inode *inode;
1756 /* Set up the RPC argument and reply structs
1757 * NB: take care not to mess about with data->commit et al. */
1760 list_splice_init(head, &data->pages);
1762 first = nfs_list_entry(data->pages.next);
1763 ctx = nfs_req_openctx(first);
1764 inode = d_inode(ctx->dentry);
1766 data->inode = inode;
1767 data->cred = ctx->cred;
1768 data->lseg = lseg; /* reference transferred */
1769 /* only set lwb for pnfs commit */
1771 data->lwb = nfs_get_lwb(&data->pages);
1772 data->mds_ops = &nfs_commit_ops;
1773 data->completion_ops = cinfo->completion_ops;
1774 data->dreq = cinfo->dreq;
1776 data->args.fh = NFS_FH(data->inode);
1777 /* Note: we always request a commit of the entire inode */
1778 data->args.offset = 0;
1779 data->args.count = 0;
1780 data->context = get_nfs_open_context(ctx);
1781 data->res.fattr = &data->fattr;
1782 data->res.verf = &data->verf;
1783 nfs_fattr_init(&data->fattr);
1784 nfs_commit_begin(cinfo->mds);
1786 EXPORT_SYMBOL_GPL(nfs_init_commit);
1788 void nfs_retry_commit(struct list_head *page_list,
1789 struct pnfs_layout_segment *lseg,
1790 struct nfs_commit_info *cinfo,
1793 struct nfs_page *req;
1795 while (!list_empty(page_list)) {
1796 req = nfs_list_entry(page_list->next);
1797 nfs_list_remove_request(req);
1798 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1800 nfs_clear_page_commit(req->wb_page);
1801 nfs_unlock_and_release_request(req);
1804 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1807 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1808 struct nfs_page *req)
1810 __set_page_dirty_nobuffers(req->wb_page);
1814 * Commit dirty pages
1817 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1818 struct nfs_commit_info *cinfo)
1820 struct nfs_commit_data *data;
1821 unsigned short task_flags = 0;
1823 /* another commit raced with us */
1824 if (list_empty(head))
1827 data = nfs_commitdata_alloc();
1829 nfs_retry_commit(head, NULL, cinfo, -1);
1833 /* Set up the argument struct */
1834 nfs_init_commit(data, head, NULL, cinfo);
1835 if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1836 task_flags = RPC_TASK_MOVEABLE;
1837 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1839 RPC_TASK_CRED_NOREF | task_flags);
1843 * COMMIT call returned
1845 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1847 struct nfs_commit_data *data = calldata;
1849 /* Call the NFS version-specific code */
1850 NFS_PROTO(data->inode)->commit_done(task, data);
1851 trace_nfs_commit_done(task, data);
1854 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1856 const struct nfs_writeverf *verf = data->res.verf;
1857 struct nfs_page *req;
1858 int status = data->task.tk_status;
1859 struct nfs_commit_info cinfo;
1860 struct nfs_server *nfss;
1862 while (!list_empty(&data->pages)) {
1863 req = nfs_list_entry(data->pages.next);
1864 nfs_list_remove_request(req);
1866 nfs_clear_page_commit(req->wb_page);
1868 dprintk("NFS: commit (%s/%llu %d@%lld)",
1869 nfs_req_openctx(req)->dentry->d_sb->s_id,
1870 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1872 (long long)req_offset(req));
1875 trace_nfs_commit_error(req, status);
1876 nfs_mapping_set_error(req->wb_page, status);
1877 nfs_inode_remove_request(req);
1879 dprintk_cont(", error = %d\n", status);
1883 /* Okay, COMMIT succeeded, apparently. Check the verifier
1884 * returned by the server against all stored verfs. */
1885 if (nfs_write_match_verf(verf, req)) {
1886 /* We have a match */
1888 nfs_inode_remove_request(req);
1889 dprintk_cont(" OK\n");
1892 /* We have a mismatch. Write the page again */
1893 dprintk_cont(" mismatch\n");
1894 nfs_mark_request_dirty(req);
1895 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1897 nfs_unlock_and_release_request(req);
1898 /* Latency breaker */
1901 nfss = NFS_SERVER(data->inode);
1902 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1903 nfss->write_congested = 0;
1905 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1906 nfs_commit_end(cinfo.mds);
1909 static void nfs_commit_release(void *calldata)
1911 struct nfs_commit_data *data = calldata;
1913 data->completion_ops->completion(data);
1914 nfs_commitdata_release(calldata);
1917 static const struct rpc_call_ops nfs_commit_ops = {
1918 .rpc_call_prepare = nfs_commit_prepare,
1919 .rpc_call_done = nfs_commit_done,
1920 .rpc_release = nfs_commit_release,
1923 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1924 .completion = nfs_commit_release_pages,
1925 .resched_write = nfs_commit_resched_write,
1928 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1929 int how, struct nfs_commit_info *cinfo)
1933 status = pnfs_commit_list(inode, head, how, cinfo);
1934 if (status == PNFS_NOT_ATTEMPTED)
1935 status = nfs_commit_list(inode, head, how, cinfo);
1939 static int __nfs_commit_inode(struct inode *inode, int how,
1940 struct writeback_control *wbc)
1943 struct nfs_commit_info cinfo;
1944 int may_wait = how & FLUSH_SYNC;
1948 nfs_init_cinfo_from_inode(&cinfo, inode);
1949 nfs_commit_begin(cinfo.mds);
1951 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1954 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1958 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1959 if (nscan < wbc->nr_to_write)
1960 wbc->nr_to_write -= nscan;
1962 wbc->nr_to_write = 0;
1964 if (nscan < INT_MAX)
1968 nfs_commit_end(cinfo.mds);
1969 if (ret || !may_wait)
1971 return wait_on_commit(cinfo.mds);
1974 int nfs_commit_inode(struct inode *inode, int how)
1976 return __nfs_commit_inode(inode, how, NULL);
1978 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1980 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1982 struct nfs_inode *nfsi = NFS_I(inode);
1983 int flags = FLUSH_SYNC;
1986 if (wbc->sync_mode == WB_SYNC_NONE) {
1987 /* no commits means nothing needs to be done */
1988 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1989 goto check_requests_outstanding;
1991 /* Don't commit yet if this is a non-blocking flush and there
1992 * are a lot of outstanding writes for this mapping.
1994 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1995 goto out_mark_dirty;
1997 /* don't wait for the COMMIT response */
2001 ret = __nfs_commit_inode(inode, flags, wbc);
2003 if (flags & FLUSH_SYNC)
2005 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
2006 goto out_mark_dirty;
2008 check_requests_outstanding:
2009 if (!atomic_read(&nfsi->commit_info.rpcs_out))
2012 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2015 EXPORT_SYMBOL_GPL(nfs_write_inode);
2018 * Wrapper for filemap_write_and_wait_range()
2020 * Needed for pNFS in order to ensure data becomes visible to the
2023 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2024 loff_t lstart, loff_t lend)
2028 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2030 ret = pnfs_sync_inode(mapping->host, true);
2033 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2036 * flush the inode to disk.
2038 int nfs_wb_all(struct inode *inode)
2042 trace_nfs_writeback_inode_enter(inode);
2044 ret = filemap_write_and_wait(inode->i_mapping);
2047 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2050 pnfs_sync_inode(inode, true);
2054 trace_nfs_writeback_inode_exit(inode, ret);
2057 EXPORT_SYMBOL_GPL(nfs_wb_all);
2059 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2061 struct nfs_page *req;
2064 folio_wait_writeback(folio);
2066 /* blocking call to cancel all requests and join to a single (head)
2068 req = nfs_lock_and_join_requests(&folio->page);
2073 /* all requests from this folio have been cancelled by
2074 * nfs_lock_and_join_requests, so just remove the head
2075 * request from the inode / page_private pointer and
2077 nfs_inode_remove_request(req);
2078 nfs_unlock_and_release_request(req);
2085 * Write back all requests on one page - we do this before reading it.
2087 int nfs_wb_page(struct inode *inode, struct page *page)
2089 loff_t range_start = page_file_offset(page);
2090 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2091 struct writeback_control wbc = {
2092 .sync_mode = WB_SYNC_ALL,
2094 .range_start = range_start,
2095 .range_end = range_end,
2099 trace_nfs_writeback_page_enter(inode);
2102 wait_on_page_writeback(page);
2103 if (clear_page_dirty_for_io(page)) {
2104 ret = nfs_writepage_locked(page, &wbc);
2110 if (!PagePrivate(page))
2112 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2117 trace_nfs_writeback_page_exit(inode, ret);
2121 #ifdef CONFIG_MIGRATION
2122 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2123 struct page *page, enum migrate_mode mode)
2126 * If PagePrivate is set, then the page is currently associated with
2127 * an in-progress read or write request. Don't try to migrate it.
2129 * FIXME: we could do this in principle, but we'll need a way to ensure
2130 * that we can safely release the inode reference while holding
2133 if (PagePrivate(page))
2136 if (PageFsCache(page)) {
2137 if (mode == MIGRATE_ASYNC)
2139 wait_on_page_fscache(page);
2142 return migrate_page(mapping, newpage, page, mode);
2146 int __init nfs_init_writepagecache(void)
2148 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2149 sizeof(struct nfs_pgio_header),
2150 0, SLAB_HWCACHE_ALIGN,
2152 if (nfs_wdata_cachep == NULL)
2155 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2157 if (nfs_wdata_mempool == NULL)
2158 goto out_destroy_write_cache;
2160 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2161 sizeof(struct nfs_commit_data),
2162 0, SLAB_HWCACHE_ALIGN,
2164 if (nfs_cdata_cachep == NULL)
2165 goto out_destroy_write_mempool;
2167 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2169 if (nfs_commit_mempool == NULL)
2170 goto out_destroy_commit_cache;
2173 * NFS congestion size, scale with available memory.
2185 * This allows larger machines to have larger/more transfers.
2186 * Limit the default to 256M
2188 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2189 if (nfs_congestion_kb > 256*1024)
2190 nfs_congestion_kb = 256*1024;
2194 out_destroy_commit_cache:
2195 kmem_cache_destroy(nfs_cdata_cachep);
2196 out_destroy_write_mempool:
2197 mempool_destroy(nfs_wdata_mempool);
2198 out_destroy_write_cache:
2199 kmem_cache_destroy(nfs_wdata_cachep);
2203 void nfs_destroy_writepagecache(void)
2205 mempool_destroy(nfs_commit_mempool);
2206 kmem_cache_destroy(nfs_cdata_cachep);
2207 mempool_destroy(nfs_wdata_mempool);
2208 kmem_cache_destroy(nfs_wdata_cachep);
2211 static const struct nfs_rw_ops nfs_rw_write_ops = {
2212 .rw_alloc_header = nfs_writehdr_alloc,
2213 .rw_free_header = nfs_writehdr_free,
2214 .rw_done = nfs_writeback_done,
2215 .rw_result = nfs_writeback_result,
2216 .rw_initiate = nfs_initiate_write,