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_commit_info *cinfo,
62 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 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);
295 /* A writeback failed: mark the page as bad, and invalidate the page cache */
296 static void nfs_set_pageerror(struct address_space *mapping)
298 struct inode *inode = mapping->host;
300 nfs_zap_mapping(mapping->host, mapping);
301 /* Force file size revalidation */
302 spin_lock(&inode->i_lock);
303 NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
304 NFS_INO_REVAL_PAGECACHE |
305 NFS_INO_INVALID_SIZE;
306 spin_unlock(&inode->i_lock);
309 static void nfs_mapping_set_error(struct page *page, int error)
311 struct address_space *mapping = page_file_mapping(page);
314 filemap_set_wb_err(mapping, error);
316 errseq_set(&mapping->host->i_sb->s_wb_err,
317 error == -ENOSPC ? -ENOSPC : -EIO);
318 nfs_set_pageerror(mapping);
322 * nfs_page_group_search_locked
323 * @head - head request of page group
324 * @page_offset - offset into page
326 * Search page group with head @head to find a request that contains the
327 * page offset @page_offset.
329 * Returns a pointer to the first matching nfs request, or NULL if no
332 * Must be called with the page group lock held
334 static struct nfs_page *
335 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
337 struct nfs_page *req;
341 if (page_offset >= req->wb_pgbase &&
342 page_offset < (req->wb_pgbase + req->wb_bytes))
345 req = req->wb_this_page;
346 } while (req != head);
352 * nfs_page_group_covers_page
353 * @head - head request of page group
355 * Return true if the page group with head @head covers the whole page,
356 * returns false otherwise
358 static bool nfs_page_group_covers_page(struct nfs_page *req)
360 struct nfs_page *tmp;
361 unsigned int pos = 0;
362 unsigned int len = nfs_page_length(req->wb_page);
364 nfs_page_group_lock(req);
367 tmp = nfs_page_group_search_locked(req->wb_head, pos);
370 pos = tmp->wb_pgbase + tmp->wb_bytes;
373 nfs_page_group_unlock(req);
377 /* We can set the PG_uptodate flag if we see that a write request
378 * covers the full page.
380 static void nfs_mark_uptodate(struct nfs_page *req)
382 if (PageUptodate(req->wb_page))
384 if (!nfs_page_group_covers_page(req))
386 SetPageUptodate(req->wb_page);
389 static int wb_priority(struct writeback_control *wbc)
393 if (wbc->sync_mode == WB_SYNC_ALL)
394 ret = FLUSH_COND_STABLE;
399 * NFS congestion control
402 int nfs_congestion_kb;
404 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
405 #define NFS_CONGESTION_OFF_THRESH \
406 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
408 static void nfs_set_page_writeback(struct page *page)
410 struct inode *inode = page_file_mapping(page)->host;
411 struct nfs_server *nfss = NFS_SERVER(inode);
412 int ret = test_set_page_writeback(page);
414 WARN_ON_ONCE(ret != 0);
416 if (atomic_long_inc_return(&nfss->writeback) >
417 NFS_CONGESTION_ON_THRESH)
418 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
421 static void nfs_end_page_writeback(struct nfs_page *req)
423 struct inode *inode = page_file_mapping(req->wb_page)->host;
424 struct nfs_server *nfss = NFS_SERVER(inode);
427 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
428 nfs_unlock_request(req);
432 end_page_writeback(req->wb_page);
433 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
434 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
438 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
440 * @destroy_list - request list (using wb_this_page) terminated by @old_head
441 * @old_head - the old head of the list
443 * All subrequests must be locked and removed from all lists, so at this point
444 * they are only "active" in this function, and possibly in nfs_wait_on_request
445 * with a reference held by some other context.
448 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
449 struct nfs_page *old_head,
452 while (destroy_list) {
453 struct nfs_page *subreq = destroy_list;
455 destroy_list = (subreq->wb_this_page == old_head) ?
456 NULL : subreq->wb_this_page;
458 /* Note: lock subreq in order to change subreq->wb_head */
459 nfs_page_set_headlock(subreq);
460 WARN_ON_ONCE(old_head != subreq->wb_head);
462 /* make sure old group is not used */
463 subreq->wb_this_page = subreq;
464 subreq->wb_head = subreq;
466 clear_bit(PG_REMOVE, &subreq->wb_flags);
468 /* Note: races with nfs_page_group_destroy() */
469 if (!kref_read(&subreq->wb_kref)) {
470 /* Check if we raced with nfs_page_group_destroy() */
471 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
472 nfs_page_clear_headlock(subreq);
473 nfs_free_request(subreq);
475 nfs_page_clear_headlock(subreq);
478 nfs_page_clear_headlock(subreq);
480 nfs_release_request(old_head);
482 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
483 nfs_release_request(subreq);
484 atomic_long_dec(&NFS_I(inode)->nrequests);
487 /* subreq is now totally disconnected from page group or any
488 * write / commit lists. last chance to wake any waiters */
489 nfs_unlock_and_release_request(subreq);
494 * nfs_join_page_group - destroy subrequests of the head req
495 * @head: the page used to lookup the "page group" of nfs_page structures
496 * @inode: Inode to which the request belongs.
498 * This function joins all sub requests to the head request by first
499 * locking all requests in the group, cancelling any pending operations
500 * and finally updating the head request to cover the whole range covered by
501 * the (former) group. All subrequests are removed from any write or commit
502 * lists, unlinked from the group and destroyed.
504 void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
507 struct nfs_page *subreq;
508 struct nfs_page *destroy_list = NULL;
509 unsigned int pgbase, off, bytes;
511 pgbase = head->wb_pgbase;
512 bytes = head->wb_bytes;
513 off = head->wb_offset;
514 for (subreq = head->wb_this_page; subreq != head;
515 subreq = subreq->wb_this_page) {
516 /* Subrequests should always form a contiguous range */
517 if (pgbase > subreq->wb_pgbase) {
518 off -= pgbase - subreq->wb_pgbase;
519 bytes += pgbase - subreq->wb_pgbase;
520 pgbase = subreq->wb_pgbase;
522 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
526 /* Set the head request's range to cover the former page group */
527 head->wb_pgbase = pgbase;
528 head->wb_bytes = bytes;
529 head->wb_offset = off;
531 /* Now that all requests are locked, make sure they aren't on any list.
532 * Commit list removal accounting is done after locks are dropped */
535 nfs_clear_request_commit(cinfo, subreq);
536 subreq = subreq->wb_this_page;
537 } while (subreq != head);
539 /* unlink subrequests from head, destroy them later */
540 if (head->wb_this_page != head) {
541 /* destroy list will be terminated by head */
542 destroy_list = head->wb_this_page;
543 head->wb_this_page = head;
546 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
550 * nfs_lock_and_join_requests - join all subreqs to the head req
551 * @page: the page used to lookup the "page group" of nfs_page structures
553 * This function joins all sub requests to the head request by first
554 * locking all requests in the group, cancelling any pending operations
555 * and finally updating the head request to cover the whole range covered by
556 * the (former) group. All subrequests are removed from any write or commit
557 * lists, unlinked from the group and destroyed.
559 * Returns a locked, referenced pointer to the head request - which after
560 * this call is guaranteed to be the only request associated with the page.
561 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
562 * error was encountered.
564 static struct nfs_page *
565 nfs_lock_and_join_requests(struct page *page)
567 struct inode *inode = page_file_mapping(page)->host;
568 struct nfs_page *head;
569 struct nfs_commit_info cinfo;
572 nfs_init_cinfo_from_inode(&cinfo, inode);
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, &cinfo, inode);
594 static void nfs_write_error(struct nfs_page *req, int error)
596 trace_nfs_write_error(req, error);
597 nfs_mapping_set_error(req->wb_page, error);
598 nfs_inode_remove_request(req);
599 nfs_end_page_writeback(req);
600 nfs_release_request(req);
604 * Find an associated nfs write request, and prepare to flush it out
605 * May return an error if the user signalled nfs_wait_on_request().
607 static int nfs_page_async_flush(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(ret)) {
635 if (nfs_error_is_fatal_on_server(ret))
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)
656 nfs_pageio_cond_complete(pgio, page_index(page));
657 ret = nfs_page_async_flush(pgio, page);
658 if (ret == -EAGAIN) {
659 redirty_page_for_writepage(wbc, page);
660 ret = AOP_WRITEPAGE_ACTIVATE;
666 * Write an mmapped page to the server.
668 static int nfs_writepage_locked(struct page *page,
669 struct writeback_control *wbc)
671 struct nfs_pageio_descriptor pgio;
672 struct inode *inode = page_file_mapping(page)->host;
675 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
676 nfs_pageio_init_write(&pgio, inode, 0,
677 false, &nfs_async_write_completion_ops);
678 err = nfs_do_writepage(page, wbc, &pgio);
680 nfs_pageio_complete(&pgio);
684 int nfs_writepage(struct page *page, struct writeback_control *wbc)
688 ret = nfs_writepage_locked(page, wbc);
689 if (ret != AOP_WRITEPAGE_ACTIVATE)
694 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
698 ret = nfs_do_writepage(page, wbc, data);
699 if (ret != AOP_WRITEPAGE_ACTIVATE)
704 static void nfs_io_completion_commit(void *inode)
706 nfs_commit_inode(inode, 0);
709 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
711 struct inode *inode = mapping->host;
712 struct nfs_pageio_descriptor pgio;
713 struct nfs_io_completion *ioc;
716 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
718 ioc = nfs_io_completion_alloc(GFP_KERNEL);
720 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
722 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
723 &nfs_async_write_completion_ops);
724 pgio.pg_io_completion = ioc;
725 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
727 nfs_pageio_complete(&pgio);
728 nfs_io_completion_put(ioc);
738 * Insert a write request into an inode
740 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
742 struct address_space *mapping = page_file_mapping(req->wb_page);
743 struct nfs_inode *nfsi = NFS_I(inode);
745 WARN_ON_ONCE(req->wb_this_page != req);
747 /* Lock the request! */
748 nfs_lock_request(req);
751 * Swap-space should not get truncated. Hence no need to plug the race
752 * with invalidate/truncate.
754 spin_lock(&mapping->private_lock);
755 if (!nfs_have_writebacks(inode) &&
756 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
757 inode_inc_iversion_raw(inode);
758 if (likely(!PageSwapCache(req->wb_page))) {
759 set_bit(PG_MAPPED, &req->wb_flags);
760 SetPagePrivate(req->wb_page);
761 set_page_private(req->wb_page, (unsigned long)req);
763 spin_unlock(&mapping->private_lock);
764 atomic_long_inc(&nfsi->nrequests);
765 /* this a head request for a page group - mark it as having an
766 * extra reference so sub groups can follow suit.
767 * This flag also informs pgio layer when to bump nrequests when
768 * adding subrequests. */
769 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
770 kref_get(&req->wb_kref);
774 * Remove a write request from an inode
776 static void nfs_inode_remove_request(struct nfs_page *req)
778 struct address_space *mapping = page_file_mapping(req->wb_page);
779 struct inode *inode = mapping->host;
780 struct nfs_inode *nfsi = NFS_I(inode);
781 struct nfs_page *head;
783 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
786 spin_lock(&mapping->private_lock);
787 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
788 set_page_private(head->wb_page, 0);
789 ClearPagePrivate(head->wb_page);
790 clear_bit(PG_MAPPED, &head->wb_flags);
792 spin_unlock(&mapping->private_lock);
795 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
796 nfs_release_request(req);
797 atomic_long_dec(&nfsi->nrequests);
802 nfs_mark_request_dirty(struct nfs_page *req)
805 __set_page_dirty_nobuffers(req->wb_page);
809 * nfs_page_search_commits_for_head_request_locked
811 * Search through commit lists on @inode for the head request for @page.
812 * Must be called while holding the inode (which is cinfo) lock.
814 * Returns the head request if found, or NULL if not found.
816 static struct nfs_page *
817 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
820 struct nfs_page *freq, *t;
821 struct nfs_commit_info cinfo;
822 struct inode *inode = &nfsi->vfs_inode;
824 nfs_init_cinfo_from_inode(&cinfo, inode);
826 /* search through pnfs commit lists */
827 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
829 return freq->wb_head;
831 /* Linearly search the commit list for the correct request */
832 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
833 if (freq->wb_page == page)
834 return freq->wb_head;
841 * nfs_request_add_commit_list_locked - add request to a commit list
842 * @req: pointer to a struct nfs_page
843 * @dst: commit list head
844 * @cinfo: holds list lock and accounting info
846 * This sets the PG_CLEAN bit, updates the cinfo count of
847 * number of outstanding requests requiring a commit as well as
850 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
854 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
855 struct nfs_commit_info *cinfo)
857 set_bit(PG_CLEAN, &req->wb_flags);
858 nfs_list_add_request(req, dst);
859 atomic_long_inc(&cinfo->mds->ncommit);
861 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
864 * nfs_request_add_commit_list - add request to a commit list
865 * @req: pointer to a struct nfs_page
866 * @cinfo: holds list lock and accounting info
868 * This sets the PG_CLEAN bit, updates the cinfo count of
869 * number of outstanding requests requiring a commit as well as
872 * The caller must _not_ hold the cinfo->lock, but must be
873 * holding the nfs_page lock.
876 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
878 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
879 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
880 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
882 nfs_mark_page_unstable(req->wb_page, cinfo);
884 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
887 * nfs_request_remove_commit_list - Remove request from a commit list
888 * @req: pointer to a nfs_page
889 * @cinfo: holds list lock and accounting info
891 * This clears the PG_CLEAN bit, and updates the cinfo's count of
892 * number of outstanding requests requiring a commit
893 * It does not update the MM page stats.
895 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
898 nfs_request_remove_commit_list(struct nfs_page *req,
899 struct nfs_commit_info *cinfo)
901 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
903 nfs_list_remove_request(req);
904 atomic_long_dec(&cinfo->mds->ncommit);
906 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
908 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
911 cinfo->inode = inode;
912 cinfo->mds = &NFS_I(inode)->commit_info;
913 cinfo->ds = pnfs_get_ds_info(inode);
915 cinfo->completion_ops = &nfs_commit_completion_ops;
918 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
920 struct nfs_direct_req *dreq)
923 nfs_init_cinfo_from_dreq(cinfo, dreq);
925 nfs_init_cinfo_from_inode(cinfo, inode);
927 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
930 * Add a request to the inode's commit list.
933 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
934 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
936 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
938 nfs_request_add_commit_list(req, cinfo);
942 nfs_clear_page_commit(struct page *page)
944 dec_node_page_state(page, NR_WRITEBACK);
945 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
949 /* Called holding the request lock on @req */
950 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
951 struct nfs_page *req)
953 if (test_bit(PG_CLEAN, &req->wb_flags)) {
954 struct nfs_open_context *ctx = nfs_req_openctx(req);
955 struct inode *inode = d_inode(ctx->dentry);
957 mutex_lock(&NFS_I(inode)->commit_mutex);
958 if (!pnfs_clear_request_commit(req, cinfo)) {
959 nfs_request_remove_commit_list(req, cinfo);
961 mutex_unlock(&NFS_I(inode)->commit_mutex);
962 nfs_clear_page_commit(req->wb_page);
966 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
968 if (hdr->verf.committed == NFS_DATA_SYNC)
969 return hdr->lseg == NULL;
970 return hdr->verf.committed != NFS_FILE_SYNC;
973 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
975 nfs_io_completion_get(hdr->io_completion);
978 static void nfs_write_completion(struct nfs_pgio_header *hdr)
980 struct nfs_commit_info cinfo;
981 unsigned long bytes = 0;
983 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
985 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
986 while (!list_empty(&hdr->pages)) {
987 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
989 bytes += req->wb_bytes;
990 nfs_list_remove_request(req);
991 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
992 (hdr->good_bytes < bytes)) {
993 trace_nfs_comp_error(req, hdr->error);
994 nfs_mapping_set_error(req->wb_page, hdr->error);
997 if (nfs_write_need_commit(hdr)) {
998 /* Reset wb_nio, since the write was successful. */
1000 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1001 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1002 hdr->pgio_mirror_idx);
1006 nfs_inode_remove_request(req);
1008 nfs_end_page_writeback(req);
1009 nfs_release_request(req);
1012 nfs_io_completion_put(hdr->io_completion);
1017 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1019 return atomic_long_read(&cinfo->mds->ncommit);
1022 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1024 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1025 struct nfs_commit_info *cinfo, int max)
1027 struct nfs_page *req, *tmp;
1030 list_for_each_entry_safe(req, tmp, src, wb_list) {
1031 kref_get(&req->wb_kref);
1032 if (!nfs_lock_request(req)) {
1033 nfs_release_request(req);
1036 nfs_request_remove_commit_list(req, cinfo);
1037 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1038 nfs_list_add_request(req, dst);
1040 if ((ret == max) && !cinfo->dreq)
1046 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1049 * nfs_scan_commit - Scan an inode for commit requests
1050 * @inode: NFS inode to scan
1051 * @dst: mds destination list
1052 * @cinfo: mds and ds lists of reqs ready to commit
1054 * Moves requests from the inode's 'commit' request list.
1055 * The requests are *not* checked to ensure that they form a contiguous set.
1058 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1059 struct nfs_commit_info *cinfo)
1063 if (!atomic_long_read(&cinfo->mds->ncommit))
1065 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1066 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1067 const int max = INT_MAX;
1069 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1071 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1073 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1078 * Search for an existing write request, and attempt to update
1079 * it to reflect a new dirty region on a given page.
1081 * If the attempt fails, then the existing request is flushed out
1084 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1086 unsigned int offset,
1089 struct nfs_page *req;
1094 end = offset + bytes;
1096 req = nfs_lock_and_join_requests(page);
1097 if (IS_ERR_OR_NULL(req))
1100 rqend = req->wb_offset + req->wb_bytes;
1102 * Tell the caller to flush out the request if
1103 * the offsets are non-contiguous.
1104 * Note: nfs_flush_incompatible() will already
1105 * have flushed out requests having wrong owners.
1107 if (offset > rqend || end < req->wb_offset)
1110 /* Okay, the request matches. Update the region */
1111 if (offset < req->wb_offset) {
1112 req->wb_offset = offset;
1113 req->wb_pgbase = offset;
1116 req->wb_bytes = end - req->wb_offset;
1118 req->wb_bytes = rqend - req->wb_offset;
1123 * Note: we mark the request dirty here because
1124 * nfs_lock_and_join_requests() cannot preserve
1125 * commit flags, so we have to replay the write.
1127 nfs_mark_request_dirty(req);
1128 nfs_unlock_and_release_request(req);
1129 error = nfs_wb_page(inode, page);
1130 return (error < 0) ? ERR_PTR(error) : NULL;
1134 * Try to update an existing write request, or create one if there is none.
1136 * Note: Should always be called with the Page Lock held to prevent races
1137 * if we have to add a new request. Also assumes that the caller has
1138 * already called nfs_flush_incompatible() if necessary.
1140 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1141 struct page *page, unsigned int offset, unsigned int bytes)
1143 struct inode *inode = page_file_mapping(page)->host;
1144 struct nfs_page *req;
1146 req = nfs_try_to_update_request(inode, page, offset, bytes);
1149 req = nfs_create_request(ctx, page, offset, bytes);
1152 nfs_inode_add_request(inode, req);
1157 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1158 unsigned int offset, unsigned int count)
1160 struct nfs_page *req;
1162 req = nfs_setup_write_request(ctx, page, offset, count);
1164 return PTR_ERR(req);
1165 /* Update file length */
1166 nfs_grow_file(page, offset, count);
1167 nfs_mark_uptodate(req);
1168 nfs_mark_request_dirty(req);
1169 nfs_unlock_and_release_request(req);
1173 int nfs_flush_incompatible(struct file *file, struct page *page)
1175 struct nfs_open_context *ctx = nfs_file_open_context(file);
1176 struct nfs_lock_context *l_ctx;
1177 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1178 struct nfs_page *req;
1179 int do_flush, status;
1181 * Look for a request corresponding to this page. If there
1182 * is one, and it belongs to another file, we flush it out
1183 * before we try to copy anything into the page. Do this
1184 * due to the lack of an ACCESS-type call in NFSv2.
1185 * Also do the same if we find a request from an existing
1189 req = nfs_page_find_head_request(page);
1192 l_ctx = req->wb_lock_context;
1193 do_flush = req->wb_page != page ||
1194 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1195 if (l_ctx && flctx &&
1196 !(list_empty_careful(&flctx->flc_posix) &&
1197 list_empty_careful(&flctx->flc_flock))) {
1198 do_flush |= l_ctx->lockowner != current->files;
1200 nfs_release_request(req);
1203 status = nfs_wb_page(page_file_mapping(page)->host, page);
1204 } while (status == 0);
1209 * Avoid buffered writes when a open context credential's key would
1212 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1214 * Return 0 and set a credential flag which triggers the inode to flush
1215 * and performs NFS_FILE_SYNC writes if the key will expired within
1216 * RPC_KEY_EXPIRE_TIMEO.
1219 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1221 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1223 if (nfs_ctx_key_to_expire(ctx, inode) &&
1225 /* Already expired! */
1231 * Test if the open context credential key is marked to expire soon.
1233 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1235 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1236 struct rpc_cred *cred = ctx->ll_cred;
1237 struct auth_cred acred = {
1241 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1243 ctx->ll_cred = NULL;
1247 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1248 if (!cred || IS_ERR(cred))
1250 ctx->ll_cred = cred;
1251 return !!(cred->cr_ops->crkey_timeout &&
1252 cred->cr_ops->crkey_timeout(cred));
1256 * If the page cache is marked as unsafe or invalid, then we can't rely on
1257 * the PageUptodate() flag. In this case, we will need to turn off
1258 * write optimisations that depend on the page contents being correct.
1260 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1262 struct nfs_inode *nfsi = NFS_I(inode);
1264 if (nfs_have_delegated_attributes(inode))
1266 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1269 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1272 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1274 return PageUptodate(page) != 0;
1278 is_whole_file_wrlock(struct file_lock *fl)
1280 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1281 fl->fl_type == F_WRLCK;
1284 /* If we know the page is up to date, and we're not using byte range locks (or
1285 * if we have the whole file locked for writing), it may be more efficient to
1286 * extend the write to cover the entire page in order to avoid fragmentation
1289 * If the file is opened for synchronous writes then we can just skip the rest
1292 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1295 struct file_lock_context *flctx = inode->i_flctx;
1296 struct file_lock *fl;
1298 if (file->f_flags & O_DSYNC)
1300 if (!nfs_write_pageuptodate(page, inode))
1302 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1304 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1305 list_empty_careful(&flctx->flc_posix)))
1308 /* Check to see if there are whole file write locks */
1310 spin_lock(&flctx->flc_lock);
1311 if (!list_empty(&flctx->flc_posix)) {
1312 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1314 if (is_whole_file_wrlock(fl))
1316 } else if (!list_empty(&flctx->flc_flock)) {
1317 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1319 if (fl->fl_type == F_WRLCK)
1322 spin_unlock(&flctx->flc_lock);
1327 * Update and possibly write a cached page of an NFS file.
1329 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1330 * things with a page scheduled for an RPC call (e.g. invalidate it).
1332 int nfs_updatepage(struct file *file, struct page *page,
1333 unsigned int offset, unsigned int count)
1335 struct nfs_open_context *ctx = nfs_file_open_context(file);
1336 struct address_space *mapping = page_file_mapping(page);
1337 struct inode *inode = mapping->host;
1340 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1342 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1343 file, count, (long long)(page_file_offset(page) + offset));
1348 if (nfs_can_extend_write(file, page, inode)) {
1349 count = max(count + offset, nfs_page_length(page));
1353 status = nfs_writepage_setup(ctx, page, offset, count);
1355 nfs_set_pageerror(mapping);
1357 __set_page_dirty_nobuffers(page);
1359 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1360 status, (long long)i_size_read(inode));
1364 static int flush_task_priority(int how)
1366 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1368 return RPC_PRIORITY_HIGH;
1370 return RPC_PRIORITY_LOW;
1372 return RPC_PRIORITY_NORMAL;
1375 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1376 struct rpc_message *msg,
1377 const struct nfs_rpc_ops *rpc_ops,
1378 struct rpc_task_setup *task_setup_data, int how)
1380 int priority = flush_task_priority(how);
1382 task_setup_data->priority = priority;
1383 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1384 trace_nfs_initiate_write(hdr);
1387 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1388 * call this on each, which will prepare them to be retried on next
1389 * writeback using standard nfs.
1391 static void nfs_redirty_request(struct nfs_page *req)
1393 /* Bump the transmission count */
1395 nfs_mark_request_dirty(req);
1396 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1397 nfs_end_page_writeback(req);
1398 nfs_release_request(req);
1401 static void nfs_async_write_error(struct list_head *head, int error)
1403 struct nfs_page *req;
1405 while (!list_empty(head)) {
1406 req = nfs_list_entry(head->next);
1407 nfs_list_remove_request(req);
1408 if (nfs_error_is_fatal_on_server(error))
1409 nfs_write_error(req, error);
1411 nfs_redirty_request(req);
1415 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1417 nfs_async_write_error(&hdr->pages, 0);
1418 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1419 hdr->args.offset + hdr->args.count - 1);
1422 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1423 .init_hdr = nfs_async_write_init,
1424 .error_cleanup = nfs_async_write_error,
1425 .completion = nfs_write_completion,
1426 .reschedule_io = nfs_async_write_reschedule_io,
1429 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1430 struct inode *inode, int ioflags, bool force_mds,
1431 const struct nfs_pgio_completion_ops *compl_ops)
1433 struct nfs_server *server = NFS_SERVER(inode);
1434 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1436 #ifdef CONFIG_NFS_V4_1
1437 if (server->pnfs_curr_ld && !force_mds)
1438 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1440 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1441 server->wsize, ioflags);
1443 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1445 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1447 struct nfs_pgio_mirror *mirror;
1449 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1450 pgio->pg_ops->pg_cleanup(pgio);
1452 pgio->pg_ops = &nfs_pgio_rw_ops;
1454 nfs_pageio_stop_mirroring(pgio);
1456 mirror = &pgio->pg_mirrors[0];
1457 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1459 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1462 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1464 struct nfs_commit_data *data = calldata;
1466 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1470 * Special version of should_remove_suid() that ignores capabilities.
1472 static int nfs_should_remove_suid(const struct inode *inode)
1474 umode_t mode = inode->i_mode;
1477 /* suid always must be killed */
1478 if (unlikely(mode & S_ISUID))
1479 kill = ATTR_KILL_SUID;
1482 * sgid without any exec bits is just a mandatory locking mark; leave
1483 * it alone. If some exec bits are set, it's a real sgid; kill it.
1485 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1486 kill |= ATTR_KILL_SGID;
1488 if (unlikely(kill && S_ISREG(mode)))
1494 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1495 struct nfs_fattr *fattr)
1497 struct nfs_pgio_args *argp = &hdr->args;
1498 struct nfs_pgio_res *resp = &hdr->res;
1499 u64 size = argp->offset + resp->count;
1501 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1503 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1504 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1507 if (size != fattr->size)
1509 /* Set attribute barrier */
1510 nfs_fattr_set_barrier(fattr);
1511 /* ...and update size */
1512 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1515 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1517 struct nfs_fattr *fattr = &hdr->fattr;
1518 struct inode *inode = hdr->inode;
1520 spin_lock(&inode->i_lock);
1521 nfs_writeback_check_extend(hdr, fattr);
1522 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1523 spin_unlock(&inode->i_lock);
1525 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1528 * This function is called when the WRITE call is complete.
1530 static int nfs_writeback_done(struct rpc_task *task,
1531 struct nfs_pgio_header *hdr,
1532 struct inode *inode)
1537 * ->write_done will attempt to use post-op attributes to detect
1538 * conflicting writes by other clients. A strict interpretation
1539 * of close-to-open would allow us to continue caching even if
1540 * another writer had changed the file, but some applications
1541 * depend on tighter cache coherency when writing.
1543 status = NFS_PROTO(inode)->write_done(task, hdr);
1547 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1548 trace_nfs_writeback_done(task, hdr);
1550 if (hdr->res.verf->committed < hdr->args.stable &&
1551 task->tk_status >= 0) {
1552 /* We tried a write call, but the server did not
1553 * commit data to stable storage even though we
1555 * Note: There is a known bug in Tru64 < 5.0 in which
1556 * the server reports NFS_DATA_SYNC, but performs
1557 * NFS_FILE_SYNC. We therefore implement this checking
1558 * as a dprintk() in order to avoid filling syslog.
1560 static unsigned long complain;
1562 /* Note this will print the MDS for a DS write */
1563 if (time_before(complain, jiffies)) {
1564 dprintk("NFS: faulty NFS server %s:"
1565 " (committed = %d) != (stable = %d)\n",
1566 NFS_SERVER(inode)->nfs_client->cl_hostname,
1567 hdr->res.verf->committed, hdr->args.stable);
1568 complain = jiffies + 300 * HZ;
1572 /* Deal with the suid/sgid bit corner case */
1573 if (nfs_should_remove_suid(inode)) {
1574 spin_lock(&inode->i_lock);
1575 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1576 spin_unlock(&inode->i_lock);
1582 * This function is called when the WRITE call is complete.
1584 static void nfs_writeback_result(struct rpc_task *task,
1585 struct nfs_pgio_header *hdr)
1587 struct nfs_pgio_args *argp = &hdr->args;
1588 struct nfs_pgio_res *resp = &hdr->res;
1590 if (resp->count < argp->count) {
1591 static unsigned long complain;
1593 /* This a short write! */
1594 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1596 /* Has the server at least made some progress? */
1597 if (resp->count == 0) {
1598 if (time_before(complain, jiffies)) {
1600 "NFS: Server wrote zero bytes, expected %u.\n",
1602 complain = jiffies + 300 * HZ;
1604 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1605 task->tk_status = -EIO;
1609 /* For non rpc-based layout drivers, retry-through-MDS */
1610 if (!task->tk_ops) {
1611 hdr->pnfs_error = -EAGAIN;
1615 /* Was this an NFSv2 write or an NFSv3 stable write? */
1616 if (resp->verf->committed != NFS_UNSTABLE) {
1617 /* Resend from where the server left off */
1618 hdr->mds_offset += resp->count;
1619 argp->offset += resp->count;
1620 argp->pgbase += resp->count;
1621 argp->count -= resp->count;
1623 /* Resend as a stable write in order to avoid
1624 * headaches in the case of a server crash.
1626 argp->stable = NFS_FILE_SYNC;
1629 resp->verf->committed = 0;
1630 rpc_restart_call_prepare(task);
1634 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1636 return wait_var_event_killable(&cinfo->rpcs_out,
1637 !atomic_read(&cinfo->rpcs_out));
1640 void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1642 atomic_inc(&cinfo->rpcs_out);
1645 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1647 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1648 wake_up_var(&cinfo->rpcs_out);
1654 void nfs_commitdata_release(struct nfs_commit_data *data)
1656 put_nfs_open_context(data->context);
1657 nfs_commit_free(data);
1659 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1661 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1662 const struct nfs_rpc_ops *nfs_ops,
1663 const struct rpc_call_ops *call_ops,
1666 struct rpc_task *task;
1667 int priority = flush_task_priority(how);
1668 struct rpc_message msg = {
1669 .rpc_argp = &data->args,
1670 .rpc_resp = &data->res,
1671 .rpc_cred = data->cred,
1673 struct rpc_task_setup task_setup_data = {
1674 .task = &data->task,
1676 .rpc_message = &msg,
1677 .callback_ops = call_ops,
1678 .callback_data = data,
1679 .workqueue = nfsiod_workqueue,
1680 .flags = RPC_TASK_ASYNC | flags,
1681 .priority = priority,
1683 /* Set up the initial task struct. */
1684 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1685 trace_nfs_initiate_commit(data);
1687 dprintk("NFS: initiated commit call\n");
1689 task = rpc_run_task(&task_setup_data);
1691 return PTR_ERR(task);
1692 if (how & FLUSH_SYNC)
1693 rpc_wait_for_completion_task(task);
1697 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1699 static loff_t nfs_get_lwb(struct list_head *head)
1702 struct nfs_page *req;
1704 list_for_each_entry(req, head, wb_list)
1705 if (lwb < (req_offset(req) + req->wb_bytes))
1706 lwb = req_offset(req) + req->wb_bytes;
1712 * Set up the argument/result storage required for the RPC call.
1714 void nfs_init_commit(struct nfs_commit_data *data,
1715 struct list_head *head,
1716 struct pnfs_layout_segment *lseg,
1717 struct nfs_commit_info *cinfo)
1719 struct nfs_page *first;
1720 struct nfs_open_context *ctx;
1721 struct inode *inode;
1723 /* Set up the RPC argument and reply structs
1724 * NB: take care not to mess about with data->commit et al. */
1727 list_splice_init(head, &data->pages);
1729 first = nfs_list_entry(data->pages.next);
1730 ctx = nfs_req_openctx(first);
1731 inode = d_inode(ctx->dentry);
1733 data->inode = inode;
1734 data->cred = ctx->cred;
1735 data->lseg = lseg; /* reference transferred */
1736 /* only set lwb for pnfs commit */
1738 data->lwb = nfs_get_lwb(&data->pages);
1739 data->mds_ops = &nfs_commit_ops;
1740 data->completion_ops = cinfo->completion_ops;
1741 data->dreq = cinfo->dreq;
1743 data->args.fh = NFS_FH(data->inode);
1744 /* Note: we always request a commit of the entire inode */
1745 data->args.offset = 0;
1746 data->args.count = 0;
1747 data->context = get_nfs_open_context(ctx);
1748 data->res.fattr = &data->fattr;
1749 data->res.verf = &data->verf;
1750 nfs_fattr_init(&data->fattr);
1751 nfs_commit_begin(cinfo->mds);
1753 EXPORT_SYMBOL_GPL(nfs_init_commit);
1755 void nfs_retry_commit(struct list_head *page_list,
1756 struct pnfs_layout_segment *lseg,
1757 struct nfs_commit_info *cinfo,
1760 struct nfs_page *req;
1762 while (!list_empty(page_list)) {
1763 req = nfs_list_entry(page_list->next);
1764 nfs_list_remove_request(req);
1765 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1767 nfs_clear_page_commit(req->wb_page);
1768 nfs_unlock_and_release_request(req);
1771 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1774 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1775 struct nfs_page *req)
1777 __set_page_dirty_nobuffers(req->wb_page);
1781 * Commit dirty pages
1784 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1785 struct nfs_commit_info *cinfo)
1787 struct nfs_commit_data *data;
1789 /* another commit raced with us */
1790 if (list_empty(head))
1793 data = nfs_commitdata_alloc();
1795 nfs_retry_commit(head, NULL, cinfo, -1);
1799 /* Set up the argument struct */
1800 nfs_init_commit(data, head, NULL, cinfo);
1801 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1802 data->mds_ops, how, RPC_TASK_CRED_NOREF);
1806 * COMMIT call returned
1808 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1810 struct nfs_commit_data *data = calldata;
1812 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1813 task->tk_pid, task->tk_status);
1815 /* Call the NFS version-specific code */
1816 NFS_PROTO(data->inode)->commit_done(task, data);
1817 trace_nfs_commit_done(task, data);
1820 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1822 const struct nfs_writeverf *verf = data->res.verf;
1823 struct nfs_page *req;
1824 int status = data->task.tk_status;
1825 struct nfs_commit_info cinfo;
1826 struct nfs_server *nfss;
1828 while (!list_empty(&data->pages)) {
1829 req = nfs_list_entry(data->pages.next);
1830 nfs_list_remove_request(req);
1832 nfs_clear_page_commit(req->wb_page);
1834 dprintk("NFS: commit (%s/%llu %d@%lld)",
1835 nfs_req_openctx(req)->dentry->d_sb->s_id,
1836 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1838 (long long)req_offset(req));
1841 trace_nfs_commit_error(req, status);
1842 nfs_mapping_set_error(req->wb_page, status);
1843 nfs_inode_remove_request(req);
1845 dprintk_cont(", error = %d\n", status);
1849 /* Okay, COMMIT succeeded, apparently. Check the verifier
1850 * returned by the server against all stored verfs. */
1851 if (nfs_write_match_verf(verf, req)) {
1852 /* We have a match */
1854 nfs_inode_remove_request(req);
1855 dprintk_cont(" OK\n");
1858 /* We have a mismatch. Write the page again */
1859 dprintk_cont(" mismatch\n");
1860 nfs_mark_request_dirty(req);
1861 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1863 nfs_unlock_and_release_request(req);
1864 /* Latency breaker */
1867 nfss = NFS_SERVER(data->inode);
1868 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1869 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1871 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1872 nfs_commit_end(cinfo.mds);
1875 static void nfs_commit_release(void *calldata)
1877 struct nfs_commit_data *data = calldata;
1879 data->completion_ops->completion(data);
1880 nfs_commitdata_release(calldata);
1883 static const struct rpc_call_ops nfs_commit_ops = {
1884 .rpc_call_prepare = nfs_commit_prepare,
1885 .rpc_call_done = nfs_commit_done,
1886 .rpc_release = nfs_commit_release,
1889 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1890 .completion = nfs_commit_release_pages,
1891 .resched_write = nfs_commit_resched_write,
1894 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1895 int how, struct nfs_commit_info *cinfo)
1899 status = pnfs_commit_list(inode, head, how, cinfo);
1900 if (status == PNFS_NOT_ATTEMPTED)
1901 status = nfs_commit_list(inode, head, how, cinfo);
1905 static int __nfs_commit_inode(struct inode *inode, int how,
1906 struct writeback_control *wbc)
1909 struct nfs_commit_info cinfo;
1910 int may_wait = how & FLUSH_SYNC;
1914 nfs_init_cinfo_from_inode(&cinfo, inode);
1915 nfs_commit_begin(cinfo.mds);
1917 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1920 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1924 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1925 if (nscan < wbc->nr_to_write)
1926 wbc->nr_to_write -= nscan;
1928 wbc->nr_to_write = 0;
1930 if (nscan < INT_MAX)
1934 nfs_commit_end(cinfo.mds);
1935 if (ret || !may_wait)
1937 return wait_on_commit(cinfo.mds);
1940 int nfs_commit_inode(struct inode *inode, int how)
1942 return __nfs_commit_inode(inode, how, NULL);
1944 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1946 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1948 struct nfs_inode *nfsi = NFS_I(inode);
1949 int flags = FLUSH_SYNC;
1952 if (wbc->sync_mode == WB_SYNC_NONE) {
1953 /* no commits means nothing needs to be done */
1954 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1955 goto check_requests_outstanding;
1957 /* Don't commit yet if this is a non-blocking flush and there
1958 * are a lot of outstanding writes for this mapping.
1960 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1961 goto out_mark_dirty;
1963 /* don't wait for the COMMIT response */
1967 ret = __nfs_commit_inode(inode, flags, wbc);
1969 if (flags & FLUSH_SYNC)
1971 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1972 goto out_mark_dirty;
1974 check_requests_outstanding:
1975 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1978 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1981 EXPORT_SYMBOL_GPL(nfs_write_inode);
1984 * Wrapper for filemap_write_and_wait_range()
1986 * Needed for pNFS in order to ensure data becomes visible to the
1989 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1990 loff_t lstart, loff_t lend)
1994 ret = filemap_write_and_wait_range(mapping, lstart, lend);
1996 ret = pnfs_sync_inode(mapping->host, true);
1999 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2002 * flush the inode to disk.
2004 int nfs_wb_all(struct inode *inode)
2008 trace_nfs_writeback_inode_enter(inode);
2010 ret = filemap_write_and_wait(inode->i_mapping);
2013 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2016 pnfs_sync_inode(inode, true);
2020 trace_nfs_writeback_inode_exit(inode, ret);
2023 EXPORT_SYMBOL_GPL(nfs_wb_all);
2025 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2027 struct nfs_page *req;
2030 wait_on_page_writeback(page);
2032 /* blocking call to cancel all requests and join to a single (head)
2034 req = nfs_lock_and_join_requests(page);
2039 /* all requests from this page have been cancelled by
2040 * nfs_lock_and_join_requests, so just remove the head
2041 * request from the inode / page_private pointer and
2043 nfs_inode_remove_request(req);
2044 nfs_unlock_and_release_request(req);
2051 * Write back all requests on one page - we do this before reading it.
2053 int nfs_wb_page(struct inode *inode, struct page *page)
2055 loff_t range_start = page_file_offset(page);
2056 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2057 struct writeback_control wbc = {
2058 .sync_mode = WB_SYNC_ALL,
2060 .range_start = range_start,
2061 .range_end = range_end,
2065 trace_nfs_writeback_page_enter(inode);
2068 wait_on_page_writeback(page);
2069 if (clear_page_dirty_for_io(page)) {
2070 ret = nfs_writepage_locked(page, &wbc);
2076 if (!PagePrivate(page))
2078 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2083 trace_nfs_writeback_page_exit(inode, ret);
2087 #ifdef CONFIG_MIGRATION
2088 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2089 struct page *page, enum migrate_mode mode)
2092 * If PagePrivate is set, then the page is currently associated with
2093 * an in-progress read or write request. Don't try to migrate it.
2095 * FIXME: we could do this in principle, but we'll need a way to ensure
2096 * that we can safely release the inode reference while holding
2099 if (PagePrivate(page))
2102 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2105 return migrate_page(mapping, newpage, page, mode);
2109 int __init nfs_init_writepagecache(void)
2111 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2112 sizeof(struct nfs_pgio_header),
2113 0, SLAB_HWCACHE_ALIGN,
2115 if (nfs_wdata_cachep == NULL)
2118 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2120 if (nfs_wdata_mempool == NULL)
2121 goto out_destroy_write_cache;
2123 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2124 sizeof(struct nfs_commit_data),
2125 0, SLAB_HWCACHE_ALIGN,
2127 if (nfs_cdata_cachep == NULL)
2128 goto out_destroy_write_mempool;
2130 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2132 if (nfs_commit_mempool == NULL)
2133 goto out_destroy_commit_cache;
2136 * NFS congestion size, scale with available memory.
2148 * This allows larger machines to have larger/more transfers.
2149 * Limit the default to 256M
2151 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2152 if (nfs_congestion_kb > 256*1024)
2153 nfs_congestion_kb = 256*1024;
2157 out_destroy_commit_cache:
2158 kmem_cache_destroy(nfs_cdata_cachep);
2159 out_destroy_write_mempool:
2160 mempool_destroy(nfs_wdata_mempool);
2161 out_destroy_write_cache:
2162 kmem_cache_destroy(nfs_wdata_cachep);
2166 void nfs_destroy_writepagecache(void)
2168 mempool_destroy(nfs_commit_mempool);
2169 kmem_cache_destroy(nfs_cdata_cachep);
2170 mempool_destroy(nfs_wdata_mempool);
2171 kmem_cache_destroy(nfs_wdata_cachep);
2174 static const struct nfs_rw_ops nfs_rw_write_ops = {
2175 .rw_alloc_header = nfs_writehdr_alloc,
2176 .rw_free_header = nfs_writehdr_free,
2177 .rw_done = nfs_writeback_done,
2178 .rw_result = nfs_writeback_result,
2179 .rw_initiate = nfs_initiate_write,