4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
9 #include <linux/types.h>
10 #include <linux/slab.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
17 #include <linux/sunrpc/clnt.h>
18 #include <linux/nfs_fs.h>
19 #include <linux/nfs_mount.h>
20 #include <linux/nfs_page.h>
21 #include <linux/backing-dev.h>
23 #include <asm/uaccess.h>
24 #include <linux/smp_lock.h>
26 #include "delegation.h"
30 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
32 #define MIN_POOL_WRITE (32)
33 #define MIN_POOL_COMMIT (4)
36 * Local function declarations
38 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
40 unsigned int, unsigned int);
41 static void nfs_mark_request_dirty(struct nfs_page *req);
42 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
43 static const struct rpc_call_ops nfs_write_partial_ops;
44 static const struct rpc_call_ops nfs_write_full_ops;
45 static const struct rpc_call_ops nfs_commit_ops;
47 static struct kmem_cache *nfs_wdata_cachep;
48 static mempool_t *nfs_wdata_mempool;
49 static mempool_t *nfs_commit_mempool;
51 struct nfs_write_data *nfs_commit_alloc(void)
53 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
56 memset(p, 0, sizeof(*p));
57 INIT_LIST_HEAD(&p->pages);
62 void nfs_commit_rcu_free(struct rcu_head *head)
64 struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
65 if (p && (p->pagevec != &p->page_array[0]))
67 mempool_free(p, nfs_commit_mempool);
70 void nfs_commit_free(struct nfs_write_data *wdata)
72 call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
75 struct nfs_write_data *nfs_writedata_alloc(size_t len)
77 unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
78 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
81 memset(p, 0, sizeof(*p));
82 INIT_LIST_HEAD(&p->pages);
83 p->npages = pagecount;
84 if (pagecount <= ARRAY_SIZE(p->page_array))
85 p->pagevec = p->page_array;
87 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
89 mempool_free(p, nfs_wdata_mempool);
97 static void nfs_writedata_rcu_free(struct rcu_head *head)
99 struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
100 if (p && (p->pagevec != &p->page_array[0]))
102 mempool_free(p, nfs_wdata_mempool);
105 static void nfs_writedata_free(struct nfs_write_data *wdata)
107 call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
110 void nfs_writedata_release(void *wdata)
112 nfs_writedata_free(wdata);
115 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
117 struct nfs_page *req = NULL;
119 if (PagePrivate(page)) {
120 req = (struct nfs_page *)page_private(page);
122 atomic_inc(&req->wb_count);
127 static struct nfs_page *nfs_page_find_request(struct page *page)
129 struct nfs_page *req = NULL;
130 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
133 req = nfs_page_find_request_locked(page);
134 spin_unlock(req_lock);
138 /* Adjust the file length if we're writing beyond the end */
139 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
141 struct inode *inode = page->mapping->host;
142 loff_t end, i_size = i_size_read(inode);
143 unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
145 if (i_size > 0 && page->index < end_index)
147 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
150 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
151 i_size_write(inode, end);
154 /* A writeback failed: mark the page as bad, and invalidate the page cache */
155 static void nfs_set_pageerror(struct page *page)
158 nfs_zap_mapping(page->mapping->host, page->mapping);
161 /* We can set the PG_uptodate flag if we see that a write request
162 * covers the full page.
164 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
166 if (PageUptodate(page))
170 if (count != nfs_page_length(page))
172 if (count != PAGE_CACHE_SIZE)
173 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
174 SetPageUptodate(page);
177 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
178 unsigned int offset, unsigned int count)
180 struct nfs_page *req;
184 req = nfs_update_request(ctx, page, offset, count);
190 ret = nfs_wb_page(page->mapping->host, page);
194 /* Update file length */
195 nfs_grow_file(page, offset, count);
196 /* Set the PG_uptodate flag? */
197 nfs_mark_uptodate(page, offset, count);
198 nfs_unlock_request(req);
202 static int wb_priority(struct writeback_control *wbc)
204 if (wbc->for_reclaim)
205 return FLUSH_HIGHPRI;
206 if (wbc->for_kupdate)
212 * NFS congestion control
215 int nfs_congestion_kb;
217 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
218 #define NFS_CONGESTION_OFF_THRESH \
219 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
221 static int nfs_set_page_writeback(struct page *page)
223 int ret = test_set_page_writeback(page);
226 struct inode *inode = page->mapping->host;
227 struct nfs_server *nfss = NFS_SERVER(inode);
229 if (atomic_inc_return(&nfss->writeback) >
230 NFS_CONGESTION_ON_THRESH)
231 set_bdi_congested(&nfss->backing_dev_info, WRITE);
236 static void nfs_end_page_writeback(struct page *page)
238 struct inode *inode = page->mapping->host;
239 struct nfs_server *nfss = NFS_SERVER(inode);
241 end_page_writeback(page);
242 if (atomic_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
243 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
244 congestion_end(WRITE);
249 * Find an associated nfs write request, and prepare to flush it out
250 * Returns 1 if there was no write request, or if the request was
251 * already tagged by nfs_set_page_dirty.Returns 0 if the request
253 * May also return an error if the user signalled nfs_wait_on_request().
255 static int nfs_page_mark_flush(struct page *page)
257 struct nfs_page *req;
258 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
263 req = nfs_page_find_request_locked(page);
265 spin_unlock(req_lock);
268 if (nfs_lock_request_dontget(req))
270 /* Note: If we hold the page lock, as is the case in nfs_writepage,
271 * then the call to nfs_lock_request_dontget() will always
272 * succeed provided that someone hasn't already marked the
273 * request as dirty (in which case we don't care).
275 spin_unlock(req_lock);
276 ret = nfs_wait_on_request(req);
277 nfs_release_request(req);
282 spin_unlock(req_lock);
283 if (nfs_set_page_writeback(page) == 0) {
284 nfs_list_remove_request(req);
285 nfs_mark_request_dirty(req);
287 ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
288 nfs_unlock_request(req);
293 * Write an mmapped page to the server.
295 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
297 struct nfs_open_context *ctx;
298 struct inode *inode = page->mapping->host;
302 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
303 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
305 err = nfs_page_mark_flush(page);
309 offset = nfs_page_length(page);
313 ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
318 err = nfs_writepage_setup(ctx, page, 0, offset);
319 put_nfs_open_context(ctx);
322 err = nfs_page_mark_flush(page);
326 if (!wbc->for_writepages)
327 nfs_flush_mapping(page->mapping, wbc, FLUSH_STABLE|wb_priority(wbc));
331 int nfs_writepage(struct page *page, struct writeback_control *wbc)
335 err = nfs_writepage_locked(page, wbc);
340 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
342 struct inode *inode = mapping->host;
345 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
347 err = generic_writepages(mapping, wbc);
350 err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
353 nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
360 * Insert a write request into an inode
362 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
364 struct nfs_inode *nfsi = NFS_I(inode);
367 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
368 BUG_ON(error == -EEXIST);
373 nfs_begin_data_update(inode);
374 if (nfs_have_delegation(inode, FMODE_WRITE))
377 SetPagePrivate(req->wb_page);
378 set_page_private(req->wb_page, (unsigned long)req);
380 atomic_inc(&req->wb_count);
385 * Remove a write request from an inode
387 static void nfs_inode_remove_request(struct nfs_page *req)
389 struct inode *inode = req->wb_context->dentry->d_inode;
390 struct nfs_inode *nfsi = NFS_I(inode);
392 BUG_ON (!NFS_WBACK_BUSY(req));
394 spin_lock(&nfsi->req_lock);
395 set_page_private(req->wb_page, 0);
396 ClearPagePrivate(req->wb_page);
397 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
400 spin_unlock(&nfsi->req_lock);
401 nfs_end_data_update(inode);
404 spin_unlock(&nfsi->req_lock);
405 nfs_clear_request(req);
406 nfs_release_request(req);
410 * Add a request to the inode's dirty list.
413 nfs_mark_request_dirty(struct nfs_page *req)
415 struct inode *inode = req->wb_context->dentry->d_inode;
416 struct nfs_inode *nfsi = NFS_I(inode);
418 spin_lock(&nfsi->req_lock);
419 radix_tree_tag_set(&nfsi->nfs_page_tree,
420 req->wb_index, NFS_PAGE_TAG_DIRTY);
421 nfs_list_add_request(req, &nfsi->dirty);
423 spin_unlock(&nfsi->req_lock);
424 __mark_inode_dirty(inode, I_DIRTY_PAGES);
428 nfs_redirty_request(struct nfs_page *req)
430 __set_page_dirty_nobuffers(req->wb_page);
434 * Check if a request is dirty
437 nfs_dirty_request(struct nfs_page *req)
439 struct page *page = req->wb_page;
443 return !PageWriteback(req->wb_page);
446 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
448 * Add a request to the inode's commit list.
451 nfs_mark_request_commit(struct nfs_page *req)
453 struct inode *inode = req->wb_context->dentry->d_inode;
454 struct nfs_inode *nfsi = NFS_I(inode);
456 spin_lock(&nfsi->req_lock);
457 nfs_list_add_request(req, &nfsi->commit);
459 spin_unlock(&nfsi->req_lock);
460 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
461 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
465 int nfs_write_need_commit(struct nfs_write_data *data)
467 return data->verf.committed != NFS_FILE_SYNC;
471 int nfs_reschedule_unstable_write(struct nfs_page *req)
473 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
474 nfs_mark_request_commit(req);
477 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
478 nfs_redirty_request(req);
485 nfs_mark_request_commit(struct nfs_page *req)
490 int nfs_write_need_commit(struct nfs_write_data *data)
496 int nfs_reschedule_unstable_write(struct nfs_page *req)
503 * Wait for a request to complete.
505 * Interruptible by signals only if mounted with intr flag.
507 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
509 struct nfs_inode *nfsi = NFS_I(inode);
510 struct nfs_page *req;
511 unsigned long idx_end, next;
512 unsigned int res = 0;
518 idx_end = idx_start + npages - 1;
521 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
522 if (req->wb_index > idx_end)
525 next = req->wb_index + 1;
526 BUG_ON(!NFS_WBACK_BUSY(req));
528 atomic_inc(&req->wb_count);
529 spin_unlock(&nfsi->req_lock);
530 error = nfs_wait_on_request(req);
531 nfs_release_request(req);
532 spin_lock(&nfsi->req_lock);
540 static void nfs_cancel_dirty_list(struct list_head *head)
542 struct nfs_page *req;
543 while(!list_empty(head)) {
544 req = nfs_list_entry(head->next);
545 nfs_list_remove_request(req);
546 nfs_end_page_writeback(req->wb_page);
547 nfs_inode_remove_request(req);
548 nfs_clear_page_writeback(req);
552 static void nfs_cancel_commit_list(struct list_head *head)
554 struct nfs_page *req;
556 while(!list_empty(head)) {
557 req = nfs_list_entry(head->next);
558 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
559 nfs_list_remove_request(req);
560 nfs_inode_remove_request(req);
561 nfs_unlock_request(req);
565 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
567 * nfs_scan_commit - Scan an inode for commit requests
568 * @inode: NFS inode to scan
569 * @dst: destination list
570 * @idx_start: lower bound of page->index to scan.
571 * @npages: idx_start + npages sets the upper bound to scan.
573 * Moves requests from the inode's 'commit' request list.
574 * The requests are *not* checked to ensure that they form a contiguous set.
577 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
579 struct nfs_inode *nfsi = NFS_I(inode);
582 if (nfsi->ncommit != 0) {
583 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
584 nfsi->ncommit -= res;
585 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
586 printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
591 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
597 static int nfs_wait_on_write_congestion(struct address_space *mapping)
599 struct inode *inode = mapping->host;
600 struct backing_dev_info *bdi = mapping->backing_dev_info;
605 if (!bdi_write_congested(bdi))
608 nfs_inc_stats(inode, NFSIOS_CONGESTIONWAIT);
611 struct rpc_clnt *clnt = NFS_CLIENT(inode);
614 rpc_clnt_sigmask(clnt, &oldset);
615 ret = congestion_wait_interruptible(WRITE, HZ/10);
616 rpc_clnt_sigunmask(clnt, &oldset);
617 if (ret == -ERESTARTSYS)
620 } while (bdi_write_congested(bdi));
626 * Try to update any existing write request, or create one if there is none.
627 * In order to match, the request's credentials must match those of
628 * the calling process.
630 * Note: Should always be called with the Page Lock held!
632 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
633 struct page *page, unsigned int offset, unsigned int bytes)
635 struct address_space *mapping = page->mapping;
636 struct inode *inode = mapping->host;
637 struct nfs_inode *nfsi = NFS_I(inode);
638 struct nfs_page *req, *new = NULL;
639 unsigned long rqend, end;
641 end = offset + bytes;
643 if (nfs_wait_on_write_congestion(mapping))
644 return ERR_PTR(-ERESTARTSYS);
646 /* Loop over all inode entries and see if we find
647 * A request for the page we wish to update
649 spin_lock(&nfsi->req_lock);
650 req = nfs_page_find_request_locked(page);
652 if (!nfs_lock_request_dontget(req)) {
655 spin_unlock(&nfsi->req_lock);
656 error = nfs_wait_on_request(req);
657 nfs_release_request(req);
660 nfs_release_request(new);
661 return ERR_PTR(error);
665 spin_unlock(&nfsi->req_lock);
667 nfs_release_request(new);
673 nfs_lock_request_dontget(new);
674 error = nfs_inode_add_request(inode, new);
676 spin_unlock(&nfsi->req_lock);
677 nfs_unlock_request(new);
678 return ERR_PTR(error);
680 spin_unlock(&nfsi->req_lock);
683 spin_unlock(&nfsi->req_lock);
685 new = nfs_create_request(ctx, inode, page, offset, bytes);
690 /* We have a request for our page.
691 * If the creds don't match, or the
692 * page addresses don't match,
693 * tell the caller to wait on the conflicting
696 rqend = req->wb_offset + req->wb_bytes;
697 if (req->wb_context != ctx
698 || req->wb_page != page
699 || !nfs_dirty_request(req)
700 || offset > rqend || end < req->wb_offset) {
701 nfs_unlock_request(req);
702 return ERR_PTR(-EBUSY);
705 /* Okay, the request matches. Update the region */
706 if (offset < req->wb_offset) {
707 req->wb_offset = offset;
708 req->wb_pgbase = offset;
709 req->wb_bytes = rqend - req->wb_offset;
713 req->wb_bytes = end - req->wb_offset;
718 int nfs_flush_incompatible(struct file *file, struct page *page)
720 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
721 struct nfs_page *req;
722 int do_flush, status;
724 * Look for a request corresponding to this page. If there
725 * is one, and it belongs to another file, we flush it out
726 * before we try to copy anything into the page. Do this
727 * due to the lack of an ACCESS-type call in NFSv2.
728 * Also do the same if we find a request from an existing
732 req = nfs_page_find_request(page);
735 do_flush = req->wb_page != page || req->wb_context != ctx
736 || !nfs_dirty_request(req);
737 nfs_release_request(req);
740 status = nfs_wb_page(page->mapping->host, page);
741 } while (status == 0);
746 * Update and possibly write a cached page of an NFS file.
748 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
749 * things with a page scheduled for an RPC call (e.g. invalidate it).
751 int nfs_updatepage(struct file *file, struct page *page,
752 unsigned int offset, unsigned int count)
754 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
755 struct inode *inode = page->mapping->host;
758 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
760 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
761 file->f_path.dentry->d_parent->d_name.name,
762 file->f_path.dentry->d_name.name, count,
763 (long long)(page_offset(page) +offset));
765 /* If we're not using byte range locks, and we know the page
766 * is entirely in cache, it may be more efficient to avoid
767 * fragmenting write requests.
769 if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
770 count = max(count + offset, nfs_page_length(page));
774 status = nfs_writepage_setup(ctx, page, offset, count);
775 __set_page_dirty_nobuffers(page);
777 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
778 status, (long long)i_size_read(inode));
780 nfs_set_pageerror(page);
784 static void nfs_writepage_release(struct nfs_page *req)
787 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
788 nfs_end_page_writeback(req->wb_page);
789 nfs_inode_remove_request(req);
791 nfs_end_page_writeback(req->wb_page);
792 nfs_clear_page_writeback(req);
795 static inline int flush_task_priority(int how)
797 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
799 return RPC_PRIORITY_HIGH;
801 return RPC_PRIORITY_LOW;
803 return RPC_PRIORITY_NORMAL;
807 * Set up the argument/result storage required for the RPC call.
809 static void nfs_write_rpcsetup(struct nfs_page *req,
810 struct nfs_write_data *data,
811 const struct rpc_call_ops *call_ops,
812 unsigned int count, unsigned int offset,
818 /* Set up the RPC argument and reply structs
819 * NB: take care not to mess about with data->commit et al. */
822 data->inode = inode = req->wb_context->dentry->d_inode;
823 data->cred = req->wb_context->cred;
825 data->args.fh = NFS_FH(inode);
826 data->args.offset = req_offset(req) + offset;
827 data->args.pgbase = req->wb_pgbase + offset;
828 data->args.pages = data->pagevec;
829 data->args.count = count;
830 data->args.context = req->wb_context;
832 data->res.fattr = &data->fattr;
833 data->res.count = count;
834 data->res.verf = &data->verf;
835 nfs_fattr_init(&data->fattr);
837 /* Set up the initial task struct. */
838 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
839 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
840 NFS_PROTO(inode)->write_setup(data, how);
842 data->task.tk_priority = flush_task_priority(how);
843 data->task.tk_cookie = (unsigned long)inode;
845 dprintk("NFS: %5u initiated write call "
846 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
849 (long long)NFS_FILEID(inode),
851 (unsigned long long)data->args.offset);
854 static void nfs_execute_write(struct nfs_write_data *data)
856 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
859 rpc_clnt_sigmask(clnt, &oldset);
860 rpc_execute(&data->task);
861 rpc_clnt_sigunmask(clnt, &oldset);
865 * Generate multiple small requests to write out a single
866 * contiguous dirty area on one page.
868 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
870 struct nfs_page *req = nfs_list_entry(head->next);
871 struct page *page = req->wb_page;
872 struct nfs_write_data *data;
873 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
878 nfs_list_remove_request(req);
880 nbytes = req->wb_bytes;
882 size_t len = min(nbytes, wsize);
884 data = nfs_writedata_alloc(len);
887 list_add(&data->pages, &list);
890 } while (nbytes != 0);
891 atomic_set(&req->wb_complete, requests);
893 ClearPageError(page);
895 nbytes = req->wb_bytes;
897 data = list_entry(list.next, struct nfs_write_data, pages);
898 list_del_init(&data->pages);
900 data->pagevec[0] = page;
902 if (nbytes > wsize) {
903 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
908 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
909 nbytes, offset, how);
912 nfs_execute_write(data);
913 } while (nbytes != 0);
918 while (!list_empty(&list)) {
919 data = list_entry(list.next, struct nfs_write_data, pages);
920 list_del(&data->pages);
921 nfs_writedata_release(data);
923 nfs_redirty_request(req);
924 nfs_end_page_writeback(req->wb_page);
925 nfs_clear_page_writeback(req);
930 * Create an RPC task for the given write request and kick it.
931 * The page must have been locked by the caller.
933 * It may happen that the page we're passed is not marked dirty.
934 * This is the case if nfs_updatepage detects a conflicting request
935 * that has been written but not committed.
937 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
939 struct nfs_page *req;
941 struct nfs_write_data *data;
944 data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
948 pages = data->pagevec;
950 while (!list_empty(head)) {
951 req = nfs_list_entry(head->next);
952 nfs_list_remove_request(req);
953 nfs_list_add_request(req, &data->pages);
954 ClearPageError(req->wb_page);
955 *pages++ = req->wb_page;
956 count += req->wb_bytes;
958 req = nfs_list_entry(data->pages.next);
960 /* Set up the argument struct */
961 nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
963 nfs_execute_write(data);
966 while (!list_empty(head)) {
967 struct nfs_page *req = nfs_list_entry(head->next);
968 nfs_list_remove_request(req);
969 nfs_redirty_request(req);
970 nfs_end_page_writeback(req->wb_page);
971 nfs_clear_page_writeback(req);
976 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
978 LIST_HEAD(one_request);
979 int (*flush_one)(struct inode *, struct list_head *, int);
980 struct nfs_page *req;
981 int wpages = NFS_SERVER(inode)->wpages;
982 int wsize = NFS_SERVER(inode)->wsize;
985 flush_one = nfs_flush_one;
986 if (wsize < PAGE_CACHE_SIZE)
987 flush_one = nfs_flush_multi;
988 /* For single writes, FLUSH_STABLE is more efficient */
989 if (npages <= wpages && npages == NFS_I(inode)->npages
990 && nfs_list_entry(head->next)->wb_bytes <= wsize)
994 nfs_coalesce_requests(head, &one_request, wpages);
995 req = nfs_list_entry(one_request.next);
996 error = flush_one(inode, &one_request, how);
999 } while (!list_empty(head));
1002 while (!list_empty(head)) {
1003 req = nfs_list_entry(head->next);
1004 nfs_list_remove_request(req);
1005 nfs_redirty_request(req);
1006 nfs_end_page_writeback(req->wb_page);
1007 nfs_clear_page_writeback(req);
1013 * Handle a write reply that flushed part of a page.
1015 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1017 struct nfs_write_data *data = calldata;
1018 struct nfs_page *req = data->req;
1019 struct page *page = req->wb_page;
1021 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1022 req->wb_context->dentry->d_inode->i_sb->s_id,
1023 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1025 (long long)req_offset(req));
1027 if (nfs_writeback_done(task, data) != 0)
1030 if (task->tk_status < 0) {
1031 nfs_set_pageerror(page);
1032 req->wb_context->error = task->tk_status;
1033 dprintk(", error = %d\n", task->tk_status);
1037 if (nfs_write_need_commit(data)) {
1038 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
1040 spin_lock(req_lock);
1041 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1042 /* Do nothing we need to resend the writes */
1043 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1044 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1045 dprintk(" defer commit\n");
1046 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1047 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1048 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1049 dprintk(" server reboot detected\n");
1051 spin_unlock(req_lock);
1056 if (atomic_dec_and_test(&req->wb_complete))
1057 nfs_writepage_release(req);
1060 static const struct rpc_call_ops nfs_write_partial_ops = {
1061 .rpc_call_done = nfs_writeback_done_partial,
1062 .rpc_release = nfs_writedata_release,
1066 * Handle a write reply that flushes a whole page.
1068 * FIXME: There is an inherent race with invalidate_inode_pages and
1069 * writebacks since the page->count is kept > 1 for as long
1070 * as the page has a write request pending.
1072 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1074 struct nfs_write_data *data = calldata;
1075 struct nfs_page *req;
1078 if (nfs_writeback_done(task, data) != 0)
1081 /* Update attributes as result of writeback. */
1082 while (!list_empty(&data->pages)) {
1083 req = nfs_list_entry(data->pages.next);
1084 nfs_list_remove_request(req);
1085 page = req->wb_page;
1087 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1088 req->wb_context->dentry->d_inode->i_sb->s_id,
1089 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1091 (long long)req_offset(req));
1093 if (task->tk_status < 0) {
1094 nfs_set_pageerror(page);
1095 req->wb_context->error = task->tk_status;
1096 dprintk(", error = %d\n", task->tk_status);
1097 goto remove_request;
1100 if (nfs_write_need_commit(data)) {
1101 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1102 nfs_mark_request_commit(req);
1103 nfs_end_page_writeback(page);
1104 dprintk(" marked for commit\n");
1109 nfs_end_page_writeback(page);
1110 nfs_inode_remove_request(req);
1112 nfs_clear_page_writeback(req);
1116 static const struct rpc_call_ops nfs_write_full_ops = {
1117 .rpc_call_done = nfs_writeback_done_full,
1118 .rpc_release = nfs_writedata_release,
1123 * This function is called when the WRITE call is complete.
1125 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1127 struct nfs_writeargs *argp = &data->args;
1128 struct nfs_writeres *resp = &data->res;
1131 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1132 task->tk_pid, task->tk_status);
1135 * ->write_done will attempt to use post-op attributes to detect
1136 * conflicting writes by other clients. A strict interpretation
1137 * of close-to-open would allow us to continue caching even if
1138 * another writer had changed the file, but some applications
1139 * depend on tighter cache coherency when writing.
1141 status = NFS_PROTO(data->inode)->write_done(task, data);
1144 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1146 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1147 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1148 /* We tried a write call, but the server did not
1149 * commit data to stable storage even though we
1151 * Note: There is a known bug in Tru64 < 5.0 in which
1152 * the server reports NFS_DATA_SYNC, but performs
1153 * NFS_FILE_SYNC. We therefore implement this checking
1154 * as a dprintk() in order to avoid filling syslog.
1156 static unsigned long complain;
1158 if (time_before(complain, jiffies)) {
1159 dprintk("NFS: faulty NFS server %s:"
1160 " (committed = %d) != (stable = %d)\n",
1161 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1162 resp->verf->committed, argp->stable);
1163 complain = jiffies + 300 * HZ;
1167 /* Is this a short write? */
1168 if (task->tk_status >= 0 && resp->count < argp->count) {
1169 static unsigned long complain;
1171 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1173 /* Has the server at least made some progress? */
1174 if (resp->count != 0) {
1175 /* Was this an NFSv2 write or an NFSv3 stable write? */
1176 if (resp->verf->committed != NFS_UNSTABLE) {
1177 /* Resend from where the server left off */
1178 argp->offset += resp->count;
1179 argp->pgbase += resp->count;
1180 argp->count -= resp->count;
1182 /* Resend as a stable write in order to avoid
1183 * headaches in the case of a server crash.
1185 argp->stable = NFS_FILE_SYNC;
1187 rpc_restart_call(task);
1190 if (time_before(complain, jiffies)) {
1192 "NFS: Server wrote zero bytes, expected %u.\n",
1194 complain = jiffies + 300 * HZ;
1196 /* Can't do anything about it except throw an error. */
1197 task->tk_status = -EIO;
1203 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1204 void nfs_commit_release(void *wdata)
1206 nfs_commit_free(wdata);
1210 * Set up the argument/result storage required for the RPC call.
1212 static void nfs_commit_rpcsetup(struct list_head *head,
1213 struct nfs_write_data *data,
1216 struct nfs_page *first;
1217 struct inode *inode;
1220 /* Set up the RPC argument and reply structs
1221 * NB: take care not to mess about with data->commit et al. */
1223 list_splice_init(head, &data->pages);
1224 first = nfs_list_entry(data->pages.next);
1225 inode = first->wb_context->dentry->d_inode;
1227 data->inode = inode;
1228 data->cred = first->wb_context->cred;
1230 data->args.fh = NFS_FH(data->inode);
1231 /* Note: we always request a commit of the entire inode */
1232 data->args.offset = 0;
1233 data->args.count = 0;
1234 data->res.count = 0;
1235 data->res.fattr = &data->fattr;
1236 data->res.verf = &data->verf;
1237 nfs_fattr_init(&data->fattr);
1239 /* Set up the initial task struct. */
1240 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1241 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1242 NFS_PROTO(inode)->commit_setup(data, how);
1244 data->task.tk_priority = flush_task_priority(how);
1245 data->task.tk_cookie = (unsigned long)inode;
1247 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1251 * Commit dirty pages
1254 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1256 struct nfs_write_data *data;
1257 struct nfs_page *req;
1259 data = nfs_commit_alloc();
1264 /* Set up the argument struct */
1265 nfs_commit_rpcsetup(head, data, how);
1267 nfs_execute_write(data);
1270 while (!list_empty(head)) {
1271 req = nfs_list_entry(head->next);
1272 nfs_list_remove_request(req);
1273 nfs_mark_request_commit(req);
1274 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1275 nfs_clear_page_writeback(req);
1281 * COMMIT call returned
1283 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1285 struct nfs_write_data *data = calldata;
1286 struct nfs_page *req;
1288 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1289 task->tk_pid, task->tk_status);
1291 /* Call the NFS version-specific code */
1292 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1295 while (!list_empty(&data->pages)) {
1296 req = nfs_list_entry(data->pages.next);
1297 nfs_list_remove_request(req);
1298 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1300 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1301 req->wb_context->dentry->d_inode->i_sb->s_id,
1302 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1304 (long long)req_offset(req));
1305 if (task->tk_status < 0) {
1306 req->wb_context->error = task->tk_status;
1307 nfs_inode_remove_request(req);
1308 dprintk(", error = %d\n", task->tk_status);
1312 /* Okay, COMMIT succeeded, apparently. Check the verifier
1313 * returned by the server against all stored verfs. */
1314 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1315 /* We have a match */
1316 nfs_inode_remove_request(req);
1320 /* We have a mismatch. Write the page again */
1321 dprintk(" mismatch\n");
1322 nfs_redirty_request(req);
1324 nfs_clear_page_writeback(req);
1328 static const struct rpc_call_ops nfs_commit_ops = {
1329 .rpc_call_done = nfs_commit_done,
1330 .rpc_release = nfs_commit_release,
1333 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1339 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1341 struct nfs_inode *nfsi = NFS_I(mapping->host);
1345 spin_lock(&nfsi->req_lock);
1346 res = nfs_scan_dirty(mapping, wbc, &head);
1347 spin_unlock(&nfsi->req_lock);
1349 int error = nfs_flush_list(mapping->host, &head, res, how);
1356 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1357 int nfs_commit_inode(struct inode *inode, int how)
1359 struct nfs_inode *nfsi = NFS_I(inode);
1363 spin_lock(&nfsi->req_lock);
1364 res = nfs_scan_commit(inode, &head, 0, 0);
1365 spin_unlock(&nfsi->req_lock);
1367 int error = nfs_commit_list(inode, &head, how);
1375 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1377 struct inode *inode = mapping->host;
1378 struct nfs_inode *nfsi = NFS_I(inode);
1379 unsigned long idx_start, idx_end;
1380 unsigned int npages = 0;
1382 int nocommit = how & FLUSH_NOCOMMIT;
1386 if (wbc->range_cyclic)
1389 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1390 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1391 if (idx_end > idx_start) {
1392 unsigned long l_npages = 1 + idx_end - idx_start;
1394 if (sizeof(npages) != sizeof(l_npages) &&
1395 (unsigned long)npages != l_npages)
1399 how &= ~FLUSH_NOCOMMIT;
1400 spin_lock(&nfsi->req_lock);
1402 wbc->pages_skipped = 0;
1403 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1406 pages = nfs_scan_dirty(mapping, wbc, &head);
1408 spin_unlock(&nfsi->req_lock);
1409 if (how & FLUSH_INVALIDATE) {
1410 nfs_cancel_dirty_list(&head);
1413 ret = nfs_flush_list(inode, &head, pages, how);
1414 spin_lock(&nfsi->req_lock);
1417 if (wbc->pages_skipped != 0)
1421 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1423 if (wbc->pages_skipped != 0)
1427 if (how & FLUSH_INVALIDATE) {
1428 spin_unlock(&nfsi->req_lock);
1429 nfs_cancel_commit_list(&head);
1431 spin_lock(&nfsi->req_lock);
1434 pages += nfs_scan_commit(inode, &head, 0, 0);
1435 spin_unlock(&nfsi->req_lock);
1436 ret = nfs_commit_list(inode, &head, how);
1437 spin_lock(&nfsi->req_lock);
1439 spin_unlock(&nfsi->req_lock);
1444 * flush the inode to disk.
1446 int nfs_wb_all(struct inode *inode)
1448 struct address_space *mapping = inode->i_mapping;
1449 struct writeback_control wbc = {
1450 .bdi = mapping->backing_dev_info,
1451 .sync_mode = WB_SYNC_ALL,
1452 .nr_to_write = LONG_MAX,
1453 .for_writepages = 1,
1458 ret = generic_writepages(mapping, &wbc);
1461 ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1465 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1469 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1471 struct writeback_control wbc = {
1472 .bdi = mapping->backing_dev_info,
1473 .sync_mode = WB_SYNC_ALL,
1474 .nr_to_write = LONG_MAX,
1475 .range_start = range_start,
1476 .range_end = range_end,
1477 .for_writepages = 1,
1481 if (!(how & FLUSH_NOWRITEPAGE)) {
1482 ret = generic_writepages(mapping, &wbc);
1486 ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1490 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1494 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1496 loff_t range_start = page_offset(page);
1497 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1498 struct writeback_control wbc = {
1499 .bdi = page->mapping->backing_dev_info,
1500 .sync_mode = WB_SYNC_ALL,
1501 .nr_to_write = LONG_MAX,
1502 .range_start = range_start,
1503 .range_end = range_end,
1507 BUG_ON(!PageLocked(page));
1508 if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1509 ret = nfs_writepage_locked(page, &wbc);
1513 if (!PagePrivate(page))
1515 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1519 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1524 * Write back all requests on one page - we do this before reading it.
1526 int nfs_wb_page(struct inode *inode, struct page* page)
1528 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1531 int nfs_set_page_dirty(struct page *page)
1533 struct nfs_page *req;
1535 req = nfs_page_find_request(page);
1537 /* Mark any existing write requests for flushing */
1538 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1539 nfs_release_request(req);
1541 return __set_page_dirty_nobuffers(page);
1545 int __init nfs_init_writepagecache(void)
1547 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1548 sizeof(struct nfs_write_data),
1549 0, SLAB_HWCACHE_ALIGN,
1551 if (nfs_wdata_cachep == NULL)
1554 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1556 if (nfs_wdata_mempool == NULL)
1559 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1561 if (nfs_commit_mempool == NULL)
1565 * NFS congestion size, scale with available memory.
1577 * This allows larger machines to have larger/more transfers.
1578 * Limit the default to 256M
1580 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1581 if (nfs_congestion_kb > 256*1024)
1582 nfs_congestion_kb = 256*1024;
1587 void nfs_destroy_writepagecache(void)
1589 mempool_destroy(nfs_commit_mempool);
1590 mempool_destroy(nfs_wdata_mempool);
1591 kmem_cache_destroy(nfs_wdata_cachep);