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>
28 #include <linux/filelock.h>
30 #include <linux/uaccess.h>
31 #include <linux/sched/mm.h>
33 #include "delegation.h"
42 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
44 #define MIN_POOL_WRITE (32)
45 #define MIN_POOL_COMMIT (4)
47 struct nfs_io_completion {
48 void (*complete)(void *data);
54 * Local function declarations
56 static void nfs_redirty_request(struct nfs_page *req);
57 static const struct rpc_call_ops nfs_commit_ops;
58 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
59 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
60 static const struct nfs_rw_ops nfs_rw_write_ops;
61 static void nfs_inode_remove_request(struct nfs_page *req);
62 static void nfs_clear_request_commit(struct nfs_page *req);
63 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
65 static struct nfs_page *
66 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
69 static struct kmem_cache *nfs_wdata_cachep;
70 static mempool_t *nfs_wdata_mempool;
71 static struct kmem_cache *nfs_cdata_cachep;
72 static mempool_t *nfs_commit_mempool;
74 struct nfs_commit_data *nfs_commitdata_alloc(void)
76 struct nfs_commit_data *p;
78 p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
80 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
83 memset(p, 0, sizeof(*p));
85 INIT_LIST_HEAD(&p->pages);
88 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
90 void nfs_commit_free(struct nfs_commit_data *p)
92 mempool_free(p, nfs_commit_mempool);
94 EXPORT_SYMBOL_GPL(nfs_commit_free);
96 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
98 struct nfs_pgio_header *p;
100 p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
102 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
105 memset(p, 0, sizeof(*p));
107 p->rw_mode = FMODE_WRITE;
111 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
113 mempool_free(hdr, nfs_wdata_mempool);
116 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
118 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
121 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
122 void (*complete)(void *), void *data)
124 ioc->complete = complete;
126 kref_init(&ioc->refcount);
129 static void nfs_io_completion_release(struct kref *kref)
131 struct nfs_io_completion *ioc = container_of(kref,
132 struct nfs_io_completion, refcount);
133 ioc->complete(ioc->data);
137 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
140 kref_get(&ioc->refcount);
143 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
146 kref_put(&ioc->refcount, nfs_io_completion_release);
150 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
152 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
153 kref_get(&req->wb_kref);
154 atomic_long_inc(&NFS_I(inode)->nrequests);
159 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
163 if (!test_bit(PG_REMOVE, &req->wb_flags))
165 ret = nfs_page_group_lock(req);
168 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
169 nfs_page_set_inode_ref(req, inode);
170 nfs_page_group_unlock(req);
174 static struct nfs_page *
175 nfs_page_private_request(struct page *page)
177 if (!PagePrivate(page))
179 return (struct nfs_page *)page_private(page);
183 * nfs_page_find_head_request_locked - find head request associated with @page
185 * must be called while holding the inode lock.
187 * returns matching head request with reference held, or NULL if not found.
189 static struct nfs_page *
190 nfs_page_find_private_request(struct page *page)
192 struct address_space *mapping = page_file_mapping(page);
193 struct nfs_page *req;
195 if (!PagePrivate(page))
197 spin_lock(&mapping->private_lock);
198 req = nfs_page_private_request(page);
200 WARN_ON_ONCE(req->wb_head != req);
201 kref_get(&req->wb_kref);
203 spin_unlock(&mapping->private_lock);
207 static struct nfs_page *
208 nfs_page_find_swap_request(struct page *page)
210 struct inode *inode = page_file_mapping(page)->host;
211 struct nfs_inode *nfsi = NFS_I(inode);
212 struct nfs_page *req = NULL;
213 if (!PageSwapCache(page))
215 mutex_lock(&nfsi->commit_mutex);
216 if (PageSwapCache(page)) {
217 req = nfs_page_search_commits_for_head_request_locked(nfsi,
220 WARN_ON_ONCE(req->wb_head != req);
221 kref_get(&req->wb_kref);
224 mutex_unlock(&nfsi->commit_mutex);
229 * nfs_page_find_head_request - find head request associated with @page
231 * returns matching head request with reference held, or NULL if not found.
233 static struct nfs_page *nfs_page_find_head_request(struct page *page)
235 struct nfs_page *req;
237 req = nfs_page_find_private_request(page);
239 req = nfs_page_find_swap_request(page);
243 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
245 struct inode *inode = page_file_mapping(page)->host;
246 struct nfs_page *req, *head;
250 req = nfs_page_find_head_request(page);
253 head = nfs_page_group_lock_head(req);
255 nfs_release_request(req);
258 ret = nfs_cancel_remove_inode(head, inode);
260 nfs_unlock_and_release_request(head);
263 /* Ensure that nobody removed the request before we locked it */
264 if (head == nfs_page_private_request(page))
266 if (PageSwapCache(page))
268 nfs_unlock_and_release_request(head);
273 /* Adjust the file length if we're writing beyond the end */
274 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
276 struct inode *inode = page_file_mapping(page)->host;
280 spin_lock(&inode->i_lock);
281 i_size = i_size_read(inode);
282 end_index = (i_size - 1) >> PAGE_SHIFT;
283 if (i_size > 0 && page_index(page) < end_index)
285 end = page_file_offset(page) + ((loff_t)offset+count);
288 trace_nfs_size_grow(inode, end);
289 i_size_write(inode, end);
290 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
291 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
293 spin_unlock(&inode->i_lock);
294 nfs_fscache_invalidate(inode, 0);
297 /* A writeback failed: mark the page as bad, and invalidate the page cache */
298 static void nfs_set_pageerror(struct address_space *mapping)
300 struct inode *inode = mapping->host;
302 nfs_zap_mapping(mapping->host, mapping);
303 /* Force file size revalidation */
304 spin_lock(&inode->i_lock);
305 nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
306 NFS_INO_INVALID_CHANGE |
307 NFS_INO_INVALID_SIZE);
308 spin_unlock(&inode->i_lock);
311 static void nfs_mapping_set_error(struct page *page, int error)
313 struct address_space *mapping = page_file_mapping(page);
316 filemap_set_wb_err(mapping, error);
318 errseq_set(&mapping->host->i_sb->s_wb_err,
319 error == -ENOSPC ? -ENOSPC : -EIO);
320 nfs_set_pageerror(mapping);
324 * nfs_page_group_search_locked
325 * @head - head request of page group
326 * @page_offset - offset into page
328 * Search page group with head @head to find a request that contains the
329 * page offset @page_offset.
331 * Returns a pointer to the first matching nfs request, or NULL if no
334 * Must be called with the page group lock held
336 static struct nfs_page *
337 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
339 struct nfs_page *req;
343 if (page_offset >= req->wb_pgbase &&
344 page_offset < (req->wb_pgbase + req->wb_bytes))
347 req = req->wb_this_page;
348 } while (req != head);
354 * nfs_page_group_covers_page
355 * @head - head request of page group
357 * Return true if the page group with head @head covers the whole page,
358 * returns false otherwise
360 static bool nfs_page_group_covers_page(struct nfs_page *req)
362 struct nfs_page *tmp;
363 unsigned int pos = 0;
364 unsigned int len = nfs_page_length(req->wb_page);
366 nfs_page_group_lock(req);
369 tmp = nfs_page_group_search_locked(req->wb_head, pos);
372 pos = tmp->wb_pgbase + tmp->wb_bytes;
375 nfs_page_group_unlock(req);
379 /* We can set the PG_uptodate flag if we see that a write request
380 * covers the full page.
382 static void nfs_mark_uptodate(struct nfs_page *req)
384 if (PageUptodate(req->wb_page))
386 if (!nfs_page_group_covers_page(req))
388 SetPageUptodate(req->wb_page);
391 static int wb_priority(struct writeback_control *wbc)
395 if (wbc->sync_mode == WB_SYNC_ALL)
396 ret = FLUSH_COND_STABLE;
401 * NFS congestion control
404 int nfs_congestion_kb;
406 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
407 #define NFS_CONGESTION_OFF_THRESH \
408 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
410 static void nfs_set_page_writeback(struct page *page)
412 struct inode *inode = page_file_mapping(page)->host;
413 struct nfs_server *nfss = NFS_SERVER(inode);
414 int ret = test_set_page_writeback(page);
416 WARN_ON_ONCE(ret != 0);
418 if (atomic_long_inc_return(&nfss->writeback) >
419 NFS_CONGESTION_ON_THRESH)
420 nfss->write_congested = 1;
423 static void nfs_end_page_writeback(struct nfs_page *req)
425 struct inode *inode = page_file_mapping(req->wb_page)->host;
426 struct nfs_server *nfss = NFS_SERVER(inode);
429 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
430 nfs_unlock_request(req);
434 end_page_writeback(req->wb_page);
435 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
436 nfss->write_congested = 0;
440 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
442 * @destroy_list - request list (using wb_this_page) terminated by @old_head
443 * @old_head - the old head of the list
445 * All subrequests must be locked and removed from all lists, so at this point
446 * they are only "active" in this function, and possibly in nfs_wait_on_request
447 * with a reference held by some other context.
450 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
451 struct nfs_page *old_head,
454 while (destroy_list) {
455 struct nfs_page *subreq = destroy_list;
457 destroy_list = (subreq->wb_this_page == old_head) ?
458 NULL : subreq->wb_this_page;
460 /* Note: lock subreq in order to change subreq->wb_head */
461 nfs_page_set_headlock(subreq);
462 WARN_ON_ONCE(old_head != subreq->wb_head);
464 /* make sure old group is not used */
465 subreq->wb_this_page = subreq;
466 subreq->wb_head = subreq;
468 clear_bit(PG_REMOVE, &subreq->wb_flags);
470 /* Note: races with nfs_page_group_destroy() */
471 if (!kref_read(&subreq->wb_kref)) {
472 /* Check if we raced with nfs_page_group_destroy() */
473 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
474 nfs_page_clear_headlock(subreq);
475 nfs_free_request(subreq);
477 nfs_page_clear_headlock(subreq);
480 nfs_page_clear_headlock(subreq);
482 nfs_release_request(old_head);
484 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
485 nfs_release_request(subreq);
486 atomic_long_dec(&NFS_I(inode)->nrequests);
489 /* subreq is now totally disconnected from page group or any
490 * write / commit lists. last chance to wake any waiters */
491 nfs_unlock_and_release_request(subreq);
496 * nfs_join_page_group - destroy subrequests of the head req
497 * @head: the page used to lookup the "page group" of nfs_page structures
498 * @inode: Inode to which the request belongs.
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.
507 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
509 struct nfs_page *subreq;
510 struct nfs_page *destroy_list = NULL;
511 unsigned int pgbase, off, bytes;
513 pgbase = head->wb_pgbase;
514 bytes = head->wb_bytes;
515 off = head->wb_offset;
516 for (subreq = head->wb_this_page; subreq != head;
517 subreq = subreq->wb_this_page) {
518 /* Subrequests should always form a contiguous range */
519 if (pgbase > subreq->wb_pgbase) {
520 off -= pgbase - subreq->wb_pgbase;
521 bytes += pgbase - subreq->wb_pgbase;
522 pgbase = subreq->wb_pgbase;
524 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
528 /* Set the head request's range to cover the former page group */
529 head->wb_pgbase = pgbase;
530 head->wb_bytes = bytes;
531 head->wb_offset = off;
533 /* Now that all requests are locked, make sure they aren't on any list.
534 * Commit list removal accounting is done after locks are dropped */
537 nfs_clear_request_commit(subreq);
538 subreq = subreq->wb_this_page;
539 } while (subreq != head);
541 /* unlink subrequests from head, destroy them later */
542 if (head->wb_this_page != head) {
543 /* destroy list will be terminated by head */
544 destroy_list = head->wb_this_page;
545 head->wb_this_page = head;
548 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
552 * nfs_lock_and_join_requests - join all subreqs to the head req
553 * @page: the page used to lookup the "page group" of nfs_page structures
555 * This function joins all sub requests to the head request by first
556 * locking all requests in the group, cancelling any pending operations
557 * and finally updating the head request to cover the whole range covered by
558 * the (former) group. All subrequests are removed from any write or commit
559 * lists, unlinked from the group and destroyed.
561 * Returns a locked, referenced pointer to the head request - which after
562 * this call is guaranteed to be the only request associated with the page.
563 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
564 * error was encountered.
566 static struct nfs_page *
567 nfs_lock_and_join_requests(struct page *page)
569 struct inode *inode = page_file_mapping(page)->host;
570 struct nfs_page *head;
574 * A reference is taken only on the head request which acts as a
575 * reference to the whole page group - the group will not be destroyed
576 * until the head reference is released.
578 head = nfs_find_and_lock_page_request(page);
579 if (IS_ERR_OR_NULL(head))
582 /* lock each request in the page group */
583 ret = nfs_page_group_lock_subrequests(head);
585 nfs_unlock_and_release_request(head);
589 nfs_join_page_group(head, inode);
594 static void nfs_write_error(struct nfs_page *req, int error)
596 trace_nfs_write_error(page_file_mapping(req->wb_page)->host, req,
598 nfs_mapping_set_error(req->wb_page, error);
599 nfs_inode_remove_request(req);
600 nfs_end_page_writeback(req);
601 nfs_release_request(req);
605 * Find an associated nfs write request, and prepare to flush it out
606 * May return an error if the user signalled nfs_wait_on_request().
608 static int nfs_page_async_flush(struct page *page,
609 struct writeback_control *wbc,
610 struct nfs_pageio_descriptor *pgio)
612 struct nfs_page *req;
615 req = nfs_lock_and_join_requests(page);
622 nfs_set_page_writeback(page);
623 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
625 /* If there is a fatal error that covers this write, just exit */
626 ret = pgio->pg_error;
627 if (nfs_error_is_fatal_on_server(ret))
631 if (!nfs_pageio_add_request(pgio, req)) {
632 ret = pgio->pg_error;
634 * Remove the problematic req upon fatal errors on the server
636 if (nfs_error_is_fatal_on_server(ret))
638 if (wbc->sync_mode == WB_SYNC_NONE)
639 ret = AOP_WRITEPAGE_ACTIVATE;
640 redirty_page_for_writepage(wbc, page);
641 nfs_redirty_request(req);
644 nfs_add_stats(page_file_mapping(page)->host,
645 NFSIOS_WRITEPAGES, 1);
649 nfs_write_error(req, ret);
653 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
654 struct nfs_pageio_descriptor *pgio)
656 nfs_pageio_cond_complete(pgio, page_index(page));
657 return nfs_page_async_flush(page, wbc, pgio);
661 * Write an mmapped page to the server.
663 static int nfs_writepage_locked(struct page *page,
664 struct writeback_control *wbc)
666 struct nfs_pageio_descriptor pgio;
667 struct inode *inode = page_file_mapping(page)->host;
670 if (wbc->sync_mode == WB_SYNC_NONE &&
671 NFS_SERVER(inode)->write_congested)
672 return AOP_WRITEPAGE_ACTIVATE;
674 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
675 nfs_pageio_init_write(&pgio, inode, 0,
676 false, &nfs_async_write_completion_ops);
677 err = nfs_do_writepage(page, wbc, &pgio);
679 nfs_pageio_complete(&pgio);
683 int nfs_writepage(struct page *page, struct writeback_control *wbc)
687 ret = nfs_writepage_locked(page, wbc);
688 if (ret != AOP_WRITEPAGE_ACTIVATE)
693 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
697 ret = nfs_do_writepage(page, wbc, data);
698 if (ret != AOP_WRITEPAGE_ACTIVATE)
703 static void nfs_io_completion_commit(void *inode)
705 nfs_commit_inode(inode, 0);
708 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
710 struct inode *inode = mapping->host;
711 struct nfs_pageio_descriptor pgio;
712 struct nfs_io_completion *ioc = NULL;
713 unsigned int mntflags = NFS_SERVER(inode)->flags;
717 if (wbc->sync_mode == WB_SYNC_NONE &&
718 NFS_SERVER(inode)->write_congested)
721 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
723 if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
724 wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
725 ioc = nfs_io_completion_alloc(GFP_KERNEL);
727 nfs_io_completion_init(ioc, nfs_io_completion_commit,
729 priority = wb_priority(wbc);
733 nfs_pageio_init_write(&pgio, inode, priority, false,
734 &nfs_async_write_completion_ops);
735 pgio.pg_io_completion = ioc;
736 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
739 nfs_pageio_complete(&pgio);
740 } while (err < 0 && !nfs_error_is_fatal(err));
741 nfs_io_completion_put(ioc);
751 * Insert a write request into an inode
753 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
755 struct address_space *mapping = page_file_mapping(req->wb_page);
756 struct nfs_inode *nfsi = NFS_I(inode);
758 WARN_ON_ONCE(req->wb_this_page != req);
760 /* Lock the request! */
761 nfs_lock_request(req);
764 * Swap-space should not get truncated. Hence no need to plug the race
765 * with invalidate/truncate.
767 spin_lock(&mapping->private_lock);
768 if (likely(!PageSwapCache(req->wb_page))) {
769 set_bit(PG_MAPPED, &req->wb_flags);
770 SetPagePrivate(req->wb_page);
771 set_page_private(req->wb_page, (unsigned long)req);
773 spin_unlock(&mapping->private_lock);
774 atomic_long_inc(&nfsi->nrequests);
775 /* this a head request for a page group - mark it as having an
776 * extra reference so sub groups can follow suit.
777 * This flag also informs pgio layer when to bump nrequests when
778 * adding subrequests. */
779 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
780 kref_get(&req->wb_kref);
784 * Remove a write request from an inode
786 static void nfs_inode_remove_request(struct nfs_page *req)
788 struct address_space *mapping = page_file_mapping(req->wb_page);
789 struct inode *inode = mapping->host;
790 struct nfs_inode *nfsi = NFS_I(inode);
791 struct nfs_page *head;
793 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
796 spin_lock(&mapping->private_lock);
797 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
798 set_page_private(head->wb_page, 0);
799 ClearPagePrivate(head->wb_page);
800 clear_bit(PG_MAPPED, &head->wb_flags);
802 spin_unlock(&mapping->private_lock);
805 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
806 nfs_release_request(req);
807 atomic_long_dec(&nfsi->nrequests);
812 nfs_mark_request_dirty(struct nfs_page *req)
815 __set_page_dirty_nobuffers(req->wb_page);
819 * nfs_page_search_commits_for_head_request_locked
821 * Search through commit lists on @inode for the head request for @page.
822 * Must be called while holding the inode (which is cinfo) lock.
824 * Returns the head request if found, or NULL if not found.
826 static struct nfs_page *
827 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
830 struct nfs_page *freq, *t;
831 struct nfs_commit_info cinfo;
832 struct inode *inode = &nfsi->vfs_inode;
834 nfs_init_cinfo_from_inode(&cinfo, inode);
836 /* search through pnfs commit lists */
837 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
839 return freq->wb_head;
841 /* Linearly search the commit list for the correct request */
842 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
843 if (freq->wb_page == page)
844 return freq->wb_head;
851 * nfs_request_add_commit_list_locked - add request to a commit list
852 * @req: pointer to a struct nfs_page
853 * @dst: commit list head
854 * @cinfo: holds list lock and accounting info
856 * This sets the PG_CLEAN bit, updates the cinfo count of
857 * number of outstanding requests requiring a commit as well as
860 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
864 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
865 struct nfs_commit_info *cinfo)
867 set_bit(PG_CLEAN, &req->wb_flags);
868 nfs_list_add_request(req, dst);
869 atomic_long_inc(&cinfo->mds->ncommit);
871 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
874 * nfs_request_add_commit_list - add request to a commit list
875 * @req: pointer to a struct nfs_page
876 * @cinfo: holds list lock and accounting info
878 * This sets the PG_CLEAN bit, updates the cinfo count of
879 * number of outstanding requests requiring a commit as well as
882 * The caller must _not_ hold the cinfo->lock, but must be
883 * holding the nfs_page lock.
886 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
888 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
889 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
890 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
892 nfs_mark_page_unstable(req->wb_page, cinfo);
894 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
897 * nfs_request_remove_commit_list - Remove request from a commit list
898 * @req: pointer to a nfs_page
899 * @cinfo: holds list lock and accounting info
901 * This clears the PG_CLEAN bit, and updates the cinfo's count of
902 * number of outstanding requests requiring a commit
903 * It does not update the MM page stats.
905 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
908 nfs_request_remove_commit_list(struct nfs_page *req,
909 struct nfs_commit_info *cinfo)
911 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
913 nfs_list_remove_request(req);
914 atomic_long_dec(&cinfo->mds->ncommit);
916 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
918 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
921 cinfo->inode = inode;
922 cinfo->mds = &NFS_I(inode)->commit_info;
923 cinfo->ds = pnfs_get_ds_info(inode);
925 cinfo->completion_ops = &nfs_commit_completion_ops;
928 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
930 struct nfs_direct_req *dreq)
933 nfs_init_cinfo_from_dreq(cinfo, dreq);
935 nfs_init_cinfo_from_inode(cinfo, inode);
937 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
940 * Add a request to the inode's commit list.
943 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
944 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
946 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
948 nfs_request_add_commit_list(req, cinfo);
952 nfs_clear_page_commit(struct page *page)
954 dec_node_page_state(page, NR_WRITEBACK);
955 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
959 /* Called holding the request lock on @req */
961 nfs_clear_request_commit(struct nfs_page *req)
963 if (test_bit(PG_CLEAN, &req->wb_flags)) {
964 struct nfs_open_context *ctx = nfs_req_openctx(req);
965 struct inode *inode = d_inode(ctx->dentry);
966 struct nfs_commit_info cinfo;
968 nfs_init_cinfo_from_inode(&cinfo, inode);
969 mutex_lock(&NFS_I(inode)->commit_mutex);
970 if (!pnfs_clear_request_commit(req, &cinfo)) {
971 nfs_request_remove_commit_list(req, &cinfo);
973 mutex_unlock(&NFS_I(inode)->commit_mutex);
974 nfs_clear_page_commit(req->wb_page);
978 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
980 if (hdr->verf.committed == NFS_DATA_SYNC)
981 return hdr->lseg == NULL;
982 return hdr->verf.committed != NFS_FILE_SYNC;
985 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
987 nfs_io_completion_get(hdr->io_completion);
990 static void nfs_write_completion(struct nfs_pgio_header *hdr)
992 struct nfs_commit_info cinfo;
993 unsigned long bytes = 0;
995 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
997 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
998 while (!list_empty(&hdr->pages)) {
999 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1001 bytes += req->wb_bytes;
1002 nfs_list_remove_request(req);
1003 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1004 (hdr->good_bytes < bytes)) {
1005 trace_nfs_comp_error(hdr->inode, req, hdr->error);
1006 nfs_mapping_set_error(req->wb_page, hdr->error);
1009 if (nfs_write_need_commit(hdr)) {
1010 /* Reset wb_nio, since the write was successful. */
1012 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1013 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1014 hdr->pgio_mirror_idx);
1018 nfs_inode_remove_request(req);
1020 nfs_end_page_writeback(req);
1021 nfs_release_request(req);
1024 nfs_io_completion_put(hdr->io_completion);
1029 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1031 return atomic_long_read(&cinfo->mds->ncommit);
1034 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1036 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1037 struct nfs_commit_info *cinfo, int max)
1039 struct nfs_page *req, *tmp;
1042 list_for_each_entry_safe(req, tmp, src, wb_list) {
1043 kref_get(&req->wb_kref);
1044 if (!nfs_lock_request(req)) {
1045 nfs_release_request(req);
1048 nfs_request_remove_commit_list(req, cinfo);
1049 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1050 nfs_list_add_request(req, dst);
1052 if ((ret == max) && !cinfo->dreq)
1058 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1061 * nfs_scan_commit - Scan an inode for commit requests
1062 * @inode: NFS inode to scan
1063 * @dst: mds destination list
1064 * @cinfo: mds and ds lists of reqs ready to commit
1066 * Moves requests from the inode's 'commit' request list.
1067 * The requests are *not* checked to ensure that they form a contiguous set.
1070 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1071 struct nfs_commit_info *cinfo)
1075 if (!atomic_long_read(&cinfo->mds->ncommit))
1077 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1078 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1079 const int max = INT_MAX;
1081 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1083 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1085 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1090 * Search for an existing write request, and attempt to update
1091 * it to reflect a new dirty region on a given page.
1093 * If the attempt fails, then the existing request is flushed out
1096 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1098 unsigned int offset,
1101 struct nfs_page *req;
1106 end = offset + bytes;
1108 req = nfs_lock_and_join_requests(page);
1109 if (IS_ERR_OR_NULL(req))
1112 rqend = req->wb_offset + req->wb_bytes;
1114 * Tell the caller to flush out the request if
1115 * the offsets are non-contiguous.
1116 * Note: nfs_flush_incompatible() will already
1117 * have flushed out requests having wrong owners.
1119 if (offset > rqend || end < req->wb_offset)
1122 /* Okay, the request matches. Update the region */
1123 if (offset < req->wb_offset) {
1124 req->wb_offset = offset;
1125 req->wb_pgbase = offset;
1128 req->wb_bytes = end - req->wb_offset;
1130 req->wb_bytes = rqend - req->wb_offset;
1135 * Note: we mark the request dirty here because
1136 * nfs_lock_and_join_requests() cannot preserve
1137 * commit flags, so we have to replay the write.
1139 nfs_mark_request_dirty(req);
1140 nfs_unlock_and_release_request(req);
1141 error = nfs_wb_page(inode, page);
1142 return (error < 0) ? ERR_PTR(error) : NULL;
1146 * Try to update an existing write request, or create one if there is none.
1148 * Note: Should always be called with the Page Lock held to prevent races
1149 * if we have to add a new request. Also assumes that the caller has
1150 * already called nfs_flush_incompatible() if necessary.
1152 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1153 struct page *page, unsigned int offset, unsigned int bytes)
1155 struct inode *inode = page_file_mapping(page)->host;
1156 struct nfs_page *req;
1158 req = nfs_try_to_update_request(inode, page, offset, bytes);
1161 req = nfs_create_request(ctx, page, offset, bytes);
1164 nfs_inode_add_request(inode, req);
1169 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1170 unsigned int offset, unsigned int count)
1172 struct nfs_page *req;
1174 req = nfs_setup_write_request(ctx, page, offset, count);
1176 return PTR_ERR(req);
1177 /* Update file length */
1178 nfs_grow_file(page, offset, count);
1179 nfs_mark_uptodate(req);
1180 nfs_mark_request_dirty(req);
1181 nfs_unlock_and_release_request(req);
1185 int nfs_flush_incompatible(struct file *file, struct page *page)
1187 struct nfs_open_context *ctx = nfs_file_open_context(file);
1188 struct nfs_lock_context *l_ctx;
1189 struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1190 struct nfs_page *req;
1191 int do_flush, status;
1193 * Look for a request corresponding to this page. If there
1194 * is one, and it belongs to another file, we flush it out
1195 * before we try to copy anything into the page. Do this
1196 * due to the lack of an ACCESS-type call in NFSv2.
1197 * Also do the same if we find a request from an existing
1201 req = nfs_page_find_head_request(page);
1204 l_ctx = req->wb_lock_context;
1205 do_flush = req->wb_page != page ||
1206 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1207 if (l_ctx && flctx &&
1208 !(list_empty_careful(&flctx->flc_posix) &&
1209 list_empty_careful(&flctx->flc_flock))) {
1210 do_flush |= l_ctx->lockowner != current->files;
1212 nfs_release_request(req);
1215 status = nfs_wb_page(page_file_mapping(page)->host, page);
1216 } while (status == 0);
1221 * Avoid buffered writes when a open context credential's key would
1224 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1226 * Return 0 and set a credential flag which triggers the inode to flush
1227 * and performs NFS_FILE_SYNC writes if the key will expired within
1228 * RPC_KEY_EXPIRE_TIMEO.
1231 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1233 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1235 if (nfs_ctx_key_to_expire(ctx, inode) &&
1236 !rcu_access_pointer(ctx->ll_cred))
1237 /* Already expired! */
1243 * Test if the open context credential key is marked to expire soon.
1245 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1247 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1248 struct rpc_cred *cred, *new, *old = NULL;
1249 struct auth_cred acred = {
1255 cred = rcu_dereference(ctx->ll_cred);
1256 if (cred && !(cred->cr_ops->crkey_timeout &&
1257 cred->cr_ops->crkey_timeout(cred)))
1261 new = auth->au_ops->lookup_cred(auth, &acred, 0);
1266 if (IS_ERR_OR_NULL(new)) {
1269 } else if (new->cr_ops->crkey_timeout &&
1270 new->cr_ops->crkey_timeout(new))
1274 old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1275 RCU_INITIALIZER(new)), 1);
1283 * If the page cache is marked as unsafe or invalid, then we can't rely on
1284 * the PageUptodate() flag. In this case, we will need to turn off
1285 * write optimisations that depend on the page contents being correct.
1287 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1288 unsigned int pagelen)
1290 struct nfs_inode *nfsi = NFS_I(inode);
1292 if (nfs_have_delegated_attributes(inode))
1294 if (nfsi->cache_validity &
1295 (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1298 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1301 if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1303 return PageUptodate(page) != 0;
1307 is_whole_file_wrlock(struct file_lock *fl)
1309 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1310 fl->fl_type == F_WRLCK;
1313 /* If we know the page is up to date, and we're not using byte range locks (or
1314 * if we have the whole file locked for writing), it may be more efficient to
1315 * extend the write to cover the entire page in order to avoid fragmentation
1318 * If the file is opened for synchronous writes then we can just skip the rest
1321 static int nfs_can_extend_write(struct file *file, struct page *page,
1322 struct inode *inode, unsigned int pagelen)
1325 struct file_lock_context *flctx = locks_inode_context(inode);
1326 struct file_lock *fl;
1328 if (file->f_flags & O_DSYNC)
1330 if (!nfs_write_pageuptodate(page, inode, pagelen))
1332 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1334 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1335 list_empty_careful(&flctx->flc_posix)))
1338 /* Check to see if there are whole file write locks */
1340 spin_lock(&flctx->flc_lock);
1341 if (!list_empty(&flctx->flc_posix)) {
1342 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1344 if (is_whole_file_wrlock(fl))
1346 } else if (!list_empty(&flctx->flc_flock)) {
1347 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1349 if (fl->fl_type == F_WRLCK)
1352 spin_unlock(&flctx->flc_lock);
1357 * Update and possibly write a cached page of an NFS file.
1359 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1360 * things with a page scheduled for an RPC call (e.g. invalidate it).
1362 int nfs_updatepage(struct file *file, struct page *page,
1363 unsigned int offset, unsigned int count)
1365 struct nfs_open_context *ctx = nfs_file_open_context(file);
1366 struct address_space *mapping = page_file_mapping(page);
1367 struct inode *inode = mapping->host;
1368 unsigned int pagelen = nfs_page_length(page);
1371 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1373 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1374 file, count, (long long)(page_file_offset(page) + offset));
1379 if (nfs_can_extend_write(file, page, inode, pagelen)) {
1380 count = max(count + offset, pagelen);
1384 status = nfs_writepage_setup(ctx, page, offset, count);
1386 nfs_set_pageerror(mapping);
1388 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1389 status, (long long)i_size_read(inode));
1393 static int flush_task_priority(int how)
1395 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1397 return RPC_PRIORITY_HIGH;
1399 return RPC_PRIORITY_LOW;
1401 return RPC_PRIORITY_NORMAL;
1404 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1405 struct rpc_message *msg,
1406 const struct nfs_rpc_ops *rpc_ops,
1407 struct rpc_task_setup *task_setup_data, int how)
1409 int priority = flush_task_priority(how);
1411 if (IS_SWAPFILE(hdr->inode))
1412 task_setup_data->flags |= RPC_TASK_SWAPPER;
1413 task_setup_data->priority = priority;
1414 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1415 trace_nfs_initiate_write(hdr);
1418 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1419 * call this on each, which will prepare them to be retried on next
1420 * writeback using standard nfs.
1422 static void nfs_redirty_request(struct nfs_page *req)
1424 struct nfs_inode *nfsi = NFS_I(page_file_mapping(req->wb_page)->host);
1426 /* Bump the transmission count */
1428 nfs_mark_request_dirty(req);
1429 atomic_long_inc(&nfsi->redirtied_pages);
1430 nfs_end_page_writeback(req);
1431 nfs_release_request(req);
1434 static void nfs_async_write_error(struct list_head *head, int error)
1436 struct nfs_page *req;
1438 while (!list_empty(head)) {
1439 req = nfs_list_entry(head->next);
1440 nfs_list_remove_request(req);
1441 if (nfs_error_is_fatal_on_server(error))
1442 nfs_write_error(req, error);
1444 nfs_redirty_request(req);
1448 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1450 nfs_async_write_error(&hdr->pages, 0);
1453 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1454 .init_hdr = nfs_async_write_init,
1455 .error_cleanup = nfs_async_write_error,
1456 .completion = nfs_write_completion,
1457 .reschedule_io = nfs_async_write_reschedule_io,
1460 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1461 struct inode *inode, int ioflags, bool force_mds,
1462 const struct nfs_pgio_completion_ops *compl_ops)
1464 struct nfs_server *server = NFS_SERVER(inode);
1465 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1467 #ifdef CONFIG_NFS_V4_1
1468 if (server->pnfs_curr_ld && !force_mds)
1469 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1471 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1472 server->wsize, ioflags);
1474 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1476 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1478 struct nfs_pgio_mirror *mirror;
1480 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1481 pgio->pg_ops->pg_cleanup(pgio);
1483 pgio->pg_ops = &nfs_pgio_rw_ops;
1485 nfs_pageio_stop_mirroring(pgio);
1487 mirror = &pgio->pg_mirrors[0];
1488 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1490 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1493 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1495 struct nfs_commit_data *data = calldata;
1497 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1500 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1501 struct nfs_fattr *fattr)
1503 struct nfs_pgio_args *argp = &hdr->args;
1504 struct nfs_pgio_res *resp = &hdr->res;
1505 u64 size = argp->offset + resp->count;
1507 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1509 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1510 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1513 if (size != fattr->size)
1515 /* Set attribute barrier */
1516 nfs_fattr_set_barrier(fattr);
1517 /* ...and update size */
1518 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1521 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1523 struct nfs_fattr *fattr = &hdr->fattr;
1524 struct inode *inode = hdr->inode;
1526 spin_lock(&inode->i_lock);
1527 nfs_writeback_check_extend(hdr, fattr);
1528 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1529 spin_unlock(&inode->i_lock);
1531 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1534 * This function is called when the WRITE call is complete.
1536 static int nfs_writeback_done(struct rpc_task *task,
1537 struct nfs_pgio_header *hdr,
1538 struct inode *inode)
1543 * ->write_done will attempt to use post-op attributes to detect
1544 * conflicting writes by other clients. A strict interpretation
1545 * of close-to-open would allow us to continue caching even if
1546 * another writer had changed the file, but some applications
1547 * depend on tighter cache coherency when writing.
1549 status = NFS_PROTO(inode)->write_done(task, hdr);
1553 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1554 trace_nfs_writeback_done(task, hdr);
1556 if (task->tk_status >= 0) {
1557 enum nfs3_stable_how committed = hdr->res.verf->committed;
1559 if (committed == NFS_UNSTABLE) {
1561 * We have some uncommitted data on the server at
1562 * this point, so ensure that we keep track of that
1563 * fact irrespective of what later writes do.
1565 set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1568 if (committed < hdr->args.stable) {
1569 /* We tried a write call, but the server did not
1570 * commit data to stable storage even though we
1572 * Note: There is a known bug in Tru64 < 5.0 in which
1573 * the server reports NFS_DATA_SYNC, but performs
1574 * NFS_FILE_SYNC. We therefore implement this checking
1575 * as a dprintk() in order to avoid filling syslog.
1577 static unsigned long complain;
1579 /* Note this will print the MDS for a DS write */
1580 if (time_before(complain, jiffies)) {
1581 dprintk("NFS: faulty NFS server %s:"
1582 " (committed = %d) != (stable = %d)\n",
1583 NFS_SERVER(inode)->nfs_client->cl_hostname,
1584 committed, hdr->args.stable);
1585 complain = jiffies + 300 * HZ;
1590 /* Deal with the suid/sgid bit corner case */
1591 if (nfs_should_remove_suid(inode)) {
1592 spin_lock(&inode->i_lock);
1593 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1594 spin_unlock(&inode->i_lock);
1600 * This function is called when the WRITE call is complete.
1602 static void nfs_writeback_result(struct rpc_task *task,
1603 struct nfs_pgio_header *hdr)
1605 struct nfs_pgio_args *argp = &hdr->args;
1606 struct nfs_pgio_res *resp = &hdr->res;
1608 if (resp->count < argp->count) {
1609 static unsigned long complain;
1611 /* This a short write! */
1612 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1614 /* Has the server at least made some progress? */
1615 if (resp->count == 0) {
1616 if (time_before(complain, jiffies)) {
1618 "NFS: Server wrote zero bytes, expected %u.\n",
1620 complain = jiffies + 300 * HZ;
1622 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1623 task->tk_status = -EIO;
1627 /* For non rpc-based layout drivers, retry-through-MDS */
1628 if (!task->tk_ops) {
1629 hdr->pnfs_error = -EAGAIN;
1633 /* Was this an NFSv2 write or an NFSv3 stable write? */
1634 if (resp->verf->committed != NFS_UNSTABLE) {
1635 /* Resend from where the server left off */
1636 hdr->mds_offset += resp->count;
1637 argp->offset += resp->count;
1638 argp->pgbase += resp->count;
1639 argp->count -= resp->count;
1641 /* Resend as a stable write in order to avoid
1642 * headaches in the case of a server crash.
1644 argp->stable = NFS_FILE_SYNC;
1647 resp->verf->committed = 0;
1648 rpc_restart_call_prepare(task);
1652 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1654 return wait_var_event_killable(&cinfo->rpcs_out,
1655 !atomic_read(&cinfo->rpcs_out));
1658 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1660 atomic_inc(&cinfo->rpcs_out);
1663 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1665 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1666 wake_up_var(&cinfo->rpcs_out);
1672 void nfs_commitdata_release(struct nfs_commit_data *data)
1674 put_nfs_open_context(data->context);
1675 nfs_commit_free(data);
1677 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1679 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1680 const struct nfs_rpc_ops *nfs_ops,
1681 const struct rpc_call_ops *call_ops,
1684 struct rpc_task *task;
1685 int priority = flush_task_priority(how);
1686 struct rpc_message msg = {
1687 .rpc_argp = &data->args,
1688 .rpc_resp = &data->res,
1689 .rpc_cred = data->cred,
1691 struct rpc_task_setup task_setup_data = {
1692 .task = &data->task,
1694 .rpc_message = &msg,
1695 .callback_ops = call_ops,
1696 .callback_data = data,
1697 .workqueue = nfsiod_workqueue,
1698 .flags = RPC_TASK_ASYNC | flags,
1699 .priority = priority,
1702 if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1703 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1705 /* Set up the initial task struct. */
1706 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1707 trace_nfs_initiate_commit(data);
1709 dprintk("NFS: initiated commit call\n");
1711 task = rpc_run_task(&task_setup_data);
1713 return PTR_ERR(task);
1714 if (how & FLUSH_SYNC)
1715 rpc_wait_for_completion_task(task);
1719 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1721 static loff_t nfs_get_lwb(struct list_head *head)
1724 struct nfs_page *req;
1726 list_for_each_entry(req, head, wb_list)
1727 if (lwb < (req_offset(req) + req->wb_bytes))
1728 lwb = req_offset(req) + req->wb_bytes;
1734 * Set up the argument/result storage required for the RPC call.
1736 void nfs_init_commit(struct nfs_commit_data *data,
1737 struct list_head *head,
1738 struct pnfs_layout_segment *lseg,
1739 struct nfs_commit_info *cinfo)
1741 struct nfs_page *first;
1742 struct nfs_open_context *ctx;
1743 struct inode *inode;
1745 /* Set up the RPC argument and reply structs
1746 * NB: take care not to mess about with data->commit et al. */
1749 list_splice_init(head, &data->pages);
1751 first = nfs_list_entry(data->pages.next);
1752 ctx = nfs_req_openctx(first);
1753 inode = d_inode(ctx->dentry);
1755 data->inode = inode;
1756 data->cred = ctx->cred;
1757 data->lseg = lseg; /* reference transferred */
1758 /* only set lwb for pnfs commit */
1760 data->lwb = nfs_get_lwb(&data->pages);
1761 data->mds_ops = &nfs_commit_ops;
1762 data->completion_ops = cinfo->completion_ops;
1763 data->dreq = cinfo->dreq;
1765 data->args.fh = NFS_FH(data->inode);
1766 /* Note: we always request a commit of the entire inode */
1767 data->args.offset = 0;
1768 data->args.count = 0;
1769 data->context = get_nfs_open_context(ctx);
1770 data->res.fattr = &data->fattr;
1771 data->res.verf = &data->verf;
1772 nfs_fattr_init(&data->fattr);
1773 nfs_commit_begin(cinfo->mds);
1775 EXPORT_SYMBOL_GPL(nfs_init_commit);
1777 void nfs_retry_commit(struct list_head *page_list,
1778 struct pnfs_layout_segment *lseg,
1779 struct nfs_commit_info *cinfo,
1782 struct nfs_page *req;
1784 while (!list_empty(page_list)) {
1785 req = nfs_list_entry(page_list->next);
1786 nfs_list_remove_request(req);
1787 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1789 nfs_clear_page_commit(req->wb_page);
1790 nfs_unlock_and_release_request(req);
1793 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1796 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1797 struct nfs_page *req)
1799 __set_page_dirty_nobuffers(req->wb_page);
1803 * Commit dirty pages
1806 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1807 struct nfs_commit_info *cinfo)
1809 struct nfs_commit_data *data;
1810 unsigned short task_flags = 0;
1812 /* another commit raced with us */
1813 if (list_empty(head))
1816 data = nfs_commitdata_alloc();
1818 nfs_retry_commit(head, NULL, cinfo, -1);
1822 /* Set up the argument struct */
1823 nfs_init_commit(data, head, NULL, cinfo);
1824 if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1825 task_flags = RPC_TASK_MOVEABLE;
1826 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1828 RPC_TASK_CRED_NOREF | task_flags);
1832 * COMMIT call returned
1834 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1836 struct nfs_commit_data *data = calldata;
1838 /* Call the NFS version-specific code */
1839 NFS_PROTO(data->inode)->commit_done(task, data);
1840 trace_nfs_commit_done(task, data);
1843 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1845 const struct nfs_writeverf *verf = data->res.verf;
1846 struct nfs_page *req;
1847 int status = data->task.tk_status;
1848 struct nfs_commit_info cinfo;
1849 struct nfs_server *nfss;
1851 while (!list_empty(&data->pages)) {
1852 req = nfs_list_entry(data->pages.next);
1853 nfs_list_remove_request(req);
1855 nfs_clear_page_commit(req->wb_page);
1857 dprintk("NFS: commit (%s/%llu %d@%lld)",
1858 nfs_req_openctx(req)->dentry->d_sb->s_id,
1859 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1861 (long long)req_offset(req));
1864 trace_nfs_commit_error(data->inode, req,
1866 nfs_mapping_set_error(req->wb_page, status);
1867 nfs_inode_remove_request(req);
1869 dprintk_cont(", error = %d\n", status);
1873 /* Okay, COMMIT succeeded, apparently. Check the verifier
1874 * returned by the server against all stored verfs. */
1875 if (nfs_write_match_verf(verf, req)) {
1876 /* We have a match */
1878 nfs_inode_remove_request(req);
1879 dprintk_cont(" OK\n");
1882 /* We have a mismatch. Write the page again */
1883 dprintk_cont(" mismatch\n");
1884 nfs_mark_request_dirty(req);
1885 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1887 nfs_unlock_and_release_request(req);
1888 /* Latency breaker */
1891 nfss = NFS_SERVER(data->inode);
1892 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1893 nfss->write_congested = 0;
1895 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1896 nfs_commit_end(cinfo.mds);
1899 static void nfs_commit_release(void *calldata)
1901 struct nfs_commit_data *data = calldata;
1903 data->completion_ops->completion(data);
1904 nfs_commitdata_release(calldata);
1907 static const struct rpc_call_ops nfs_commit_ops = {
1908 .rpc_call_prepare = nfs_commit_prepare,
1909 .rpc_call_done = nfs_commit_done,
1910 .rpc_release = nfs_commit_release,
1913 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1914 .completion = nfs_commit_release_pages,
1915 .resched_write = nfs_commit_resched_write,
1918 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1919 int how, struct nfs_commit_info *cinfo)
1923 status = pnfs_commit_list(inode, head, how, cinfo);
1924 if (status == PNFS_NOT_ATTEMPTED)
1925 status = nfs_commit_list(inode, head, how, cinfo);
1929 static int __nfs_commit_inode(struct inode *inode, int how,
1930 struct writeback_control *wbc)
1933 struct nfs_commit_info cinfo;
1934 int may_wait = how & FLUSH_SYNC;
1938 nfs_init_cinfo_from_inode(&cinfo, inode);
1939 nfs_commit_begin(cinfo.mds);
1941 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1944 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1948 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1949 if (nscan < wbc->nr_to_write)
1950 wbc->nr_to_write -= nscan;
1952 wbc->nr_to_write = 0;
1954 if (nscan < INT_MAX)
1958 nfs_commit_end(cinfo.mds);
1959 if (ret || !may_wait)
1961 return wait_on_commit(cinfo.mds);
1964 int nfs_commit_inode(struct inode *inode, int how)
1966 return __nfs_commit_inode(inode, how, NULL);
1968 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1970 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1972 struct nfs_inode *nfsi = NFS_I(inode);
1973 int flags = FLUSH_SYNC;
1976 if (wbc->sync_mode == WB_SYNC_NONE) {
1977 /* no commits means nothing needs to be done */
1978 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1979 goto check_requests_outstanding;
1981 /* Don't commit yet if this is a non-blocking flush and there
1982 * are a lot of outstanding writes for this mapping.
1984 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1985 goto out_mark_dirty;
1987 /* don't wait for the COMMIT response */
1991 ret = __nfs_commit_inode(inode, flags, wbc);
1993 if (flags & FLUSH_SYNC)
1995 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1996 goto out_mark_dirty;
1998 check_requests_outstanding:
1999 if (!atomic_read(&nfsi->commit_info.rpcs_out))
2002 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2005 EXPORT_SYMBOL_GPL(nfs_write_inode);
2008 * Wrapper for filemap_write_and_wait_range()
2010 * Needed for pNFS in order to ensure data becomes visible to the
2013 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2014 loff_t lstart, loff_t lend)
2018 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2020 ret = pnfs_sync_inode(mapping->host, true);
2023 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2026 * flush the inode to disk.
2028 int nfs_wb_all(struct inode *inode)
2032 trace_nfs_writeback_inode_enter(inode);
2034 ret = filemap_write_and_wait(inode->i_mapping);
2037 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2040 pnfs_sync_inode(inode, true);
2044 trace_nfs_writeback_inode_exit(inode, ret);
2047 EXPORT_SYMBOL_GPL(nfs_wb_all);
2049 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2051 struct nfs_page *req;
2054 folio_wait_writeback(folio);
2056 /* blocking call to cancel all requests and join to a single (head)
2058 req = nfs_lock_and_join_requests(&folio->page);
2063 /* all requests from this folio have been cancelled by
2064 * nfs_lock_and_join_requests, so just remove the head
2065 * request from the inode / page_private pointer and
2067 nfs_inode_remove_request(req);
2068 nfs_unlock_and_release_request(req);
2075 * Write back all requests on one page - we do this before reading it.
2077 int nfs_wb_page(struct inode *inode, struct page *page)
2079 loff_t range_start = page_file_offset(page);
2080 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2081 struct writeback_control wbc = {
2082 .sync_mode = WB_SYNC_ALL,
2084 .range_start = range_start,
2085 .range_end = range_end,
2089 trace_nfs_writeback_page_enter(inode);
2092 wait_on_page_writeback(page);
2093 if (clear_page_dirty_for_io(page)) {
2094 ret = nfs_writepage_locked(page, &wbc);
2100 if (!PagePrivate(page))
2102 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2107 trace_nfs_writeback_page_exit(inode, ret);
2111 #ifdef CONFIG_MIGRATION
2112 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2113 struct folio *src, enum migrate_mode mode)
2116 * If the private flag is set, the folio is currently associated with
2117 * an in-progress read or write request. Don't try to migrate it.
2119 * FIXME: we could do this in principle, but we'll need a way to ensure
2120 * that we can safely release the inode reference while holding
2123 if (folio_test_private(src))
2126 if (folio_test_fscache(src)) {
2127 if (mode == MIGRATE_ASYNC)
2129 folio_wait_fscache(src);
2132 return migrate_folio(mapping, dst, src, mode);
2136 int __init nfs_init_writepagecache(void)
2138 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2139 sizeof(struct nfs_pgio_header),
2140 0, SLAB_HWCACHE_ALIGN,
2142 if (nfs_wdata_cachep == NULL)
2145 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2147 if (nfs_wdata_mempool == NULL)
2148 goto out_destroy_write_cache;
2150 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2151 sizeof(struct nfs_commit_data),
2152 0, SLAB_HWCACHE_ALIGN,
2154 if (nfs_cdata_cachep == NULL)
2155 goto out_destroy_write_mempool;
2157 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2159 if (nfs_commit_mempool == NULL)
2160 goto out_destroy_commit_cache;
2163 * NFS congestion size, scale with available memory.
2175 * This allows larger machines to have larger/more transfers.
2176 * Limit the default to 256M
2178 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2179 if (nfs_congestion_kb > 256*1024)
2180 nfs_congestion_kb = 256*1024;
2184 out_destroy_commit_cache:
2185 kmem_cache_destroy(nfs_cdata_cachep);
2186 out_destroy_write_mempool:
2187 mempool_destroy(nfs_wdata_mempool);
2188 out_destroy_write_cache:
2189 kmem_cache_destroy(nfs_wdata_cachep);
2193 void nfs_destroy_writepagecache(void)
2195 mempool_destroy(nfs_commit_mempool);
2196 kmem_cache_destroy(nfs_cdata_cachep);
2197 mempool_destroy(nfs_wdata_mempool);
2198 kmem_cache_destroy(nfs_wdata_cachep);
2201 static const struct nfs_rw_ops nfs_rw_write_ops = {
2202 .rw_alloc_header = nfs_writehdr_alloc,
2203 .rw_free_header = nfs_writehdr_free,
2204 .rw_done = nfs_writeback_done,
2205 .rw_result = nfs_writeback_result,
2206 .rw_initiate = nfs_initiate_write,