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>
25 #include "delegation.h"
29 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
31 #define MIN_POOL_WRITE (32)
32 #define MIN_POOL_COMMIT (4)
35 * Local function declarations
37 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
39 unsigned int, unsigned int);
40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
41 struct inode *inode, int ioflags);
42 static const struct rpc_call_ops nfs_write_partial_ops;
43 static const struct rpc_call_ops nfs_write_full_ops;
44 static const struct rpc_call_ops nfs_commit_ops;
46 static struct kmem_cache *nfs_wdata_cachep;
47 static mempool_t *nfs_wdata_mempool;
48 static mempool_t *nfs_commit_mempool;
50 struct nfs_write_data *nfs_commit_alloc(void)
52 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
55 memset(p, 0, sizeof(*p));
56 INIT_LIST_HEAD(&p->pages);
61 static void nfs_commit_rcu_free(struct rcu_head *head)
63 struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
64 if (p && (p->pagevec != &p->page_array[0]))
66 mempool_free(p, nfs_commit_mempool);
69 void nfs_commit_free(struct nfs_write_data *wdata)
71 call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
74 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
76 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
79 memset(p, 0, sizeof(*p));
80 INIT_LIST_HEAD(&p->pages);
81 p->npages = pagecount;
82 if (pagecount <= ARRAY_SIZE(p->page_array))
83 p->pagevec = p->page_array;
85 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
87 mempool_free(p, nfs_wdata_mempool);
95 static void nfs_writedata_rcu_free(struct rcu_head *head)
97 struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
98 if (p && (p->pagevec != &p->page_array[0]))
100 mempool_free(p, nfs_wdata_mempool);
103 static void nfs_writedata_free(struct nfs_write_data *wdata)
105 call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
108 void nfs_writedata_release(void *wdata)
110 nfs_writedata_free(wdata);
113 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
117 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
120 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
122 struct nfs_page *req = NULL;
124 if (PagePrivate(page)) {
125 req = (struct nfs_page *)page_private(page);
127 kref_get(&req->wb_kref);
132 static struct nfs_page *nfs_page_find_request(struct page *page)
134 struct inode *inode = page->mapping->host;
135 struct nfs_page *req = NULL;
137 spin_lock(&inode->i_lock);
138 req = nfs_page_find_request_locked(page);
139 spin_unlock(&inode->i_lock);
143 /* Adjust the file length if we're writing beyond the end */
144 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
146 struct inode *inode = page->mapping->host;
147 loff_t end, i_size = i_size_read(inode);
148 pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
150 if (i_size > 0 && page->index < end_index)
152 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
155 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
156 i_size_write(inode, end);
159 /* A writeback failed: mark the page as bad, and invalidate the page cache */
160 static void nfs_set_pageerror(struct page *page)
163 nfs_zap_mapping(page->mapping->host, page->mapping);
166 /* We can set the PG_uptodate flag if we see that a write request
167 * covers the full page.
169 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
171 if (PageUptodate(page))
175 if (count != nfs_page_length(page))
177 SetPageUptodate(page);
180 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
181 unsigned int offset, unsigned int count)
183 struct nfs_page *req;
187 req = nfs_update_request(ctx, page, offset, count);
193 ret = nfs_wb_page(page->mapping->host, page);
197 /* Update file length */
198 nfs_grow_file(page, offset, count);
199 nfs_clear_page_tag_locked(req);
203 static int wb_priority(struct writeback_control *wbc)
205 if (wbc->for_reclaim)
206 return FLUSH_HIGHPRI | FLUSH_STABLE;
207 if (wbc->for_kupdate)
213 * NFS congestion control
216 int nfs_congestion_kb;
218 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
219 #define NFS_CONGESTION_OFF_THRESH \
220 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
222 static int nfs_set_page_writeback(struct page *page)
224 int ret = test_set_page_writeback(page);
227 struct inode *inode = page->mapping->host;
228 struct nfs_server *nfss = NFS_SERVER(inode);
230 if (atomic_long_inc_return(&nfss->writeback) >
231 NFS_CONGESTION_ON_THRESH)
232 set_bdi_congested(&nfss->backing_dev_info, WRITE);
237 static void nfs_end_page_writeback(struct page *page)
239 struct inode *inode = page->mapping->host;
240 struct nfs_server *nfss = NFS_SERVER(inode);
242 end_page_writeback(page);
243 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
244 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
248 * Find an associated nfs write request, and prepare to flush it out
249 * May return an error if the user signalled nfs_wait_on_request().
251 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
254 struct inode *inode = page->mapping->host;
255 struct nfs_page *req;
258 spin_lock(&inode->i_lock);
260 req = nfs_page_find_request_locked(page);
262 spin_unlock(&inode->i_lock);
265 if (nfs_set_page_tag_locked(req))
267 /* Note: If we hold the page lock, as is the case in nfs_writepage,
268 * then the call to nfs_set_page_tag_locked() will always
269 * succeed provided that someone hasn't already marked the
270 * request as dirty (in which case we don't care).
272 spin_unlock(&inode->i_lock);
273 ret = nfs_wait_on_request(req);
274 nfs_release_request(req);
277 spin_lock(&inode->i_lock);
279 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
280 /* This request is marked for commit */
281 spin_unlock(&inode->i_lock);
282 nfs_clear_page_tag_locked(req);
283 nfs_pageio_complete(pgio);
286 if (nfs_set_page_writeback(page) != 0) {
287 spin_unlock(&inode->i_lock);
290 spin_unlock(&inode->i_lock);
291 nfs_pageio_add_request(pgio, req);
295 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
297 struct inode *inode = page->mapping->host;
299 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
300 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
302 nfs_pageio_cond_complete(pgio, page->index);
303 return nfs_page_async_flush(pgio, page);
307 * Write an mmapped page to the server.
309 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
311 struct nfs_pageio_descriptor pgio;
314 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
315 err = nfs_do_writepage(page, wbc, &pgio);
316 nfs_pageio_complete(&pgio);
319 if (pgio.pg_error < 0)
320 return pgio.pg_error;
324 int nfs_writepage(struct page *page, struct writeback_control *wbc)
328 ret = nfs_writepage_locked(page, wbc);
333 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
337 ret = nfs_do_writepage(page, wbc, data);
342 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
344 struct inode *inode = mapping->host;
345 struct nfs_pageio_descriptor pgio;
348 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
350 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
351 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
352 nfs_pageio_complete(&pgio);
355 if (pgio.pg_error < 0)
356 return pgio.pg_error;
361 * Insert a write request into an inode
363 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
365 struct nfs_inode *nfsi = NFS_I(inode);
368 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
369 BUG_ON(error == -EEXIST);
374 if (nfs_have_delegation(inode, FMODE_WRITE))
377 SetPagePrivate(req->wb_page);
378 set_page_private(req->wb_page, (unsigned long)req);
380 kref_get(&req->wb_kref);
381 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
386 * Remove a write request from an inode
388 static void nfs_inode_remove_request(struct nfs_page *req)
390 struct inode *inode = req->wb_context->path.dentry->d_inode;
391 struct nfs_inode *nfsi = NFS_I(inode);
393 BUG_ON (!NFS_WBACK_BUSY(req));
395 spin_lock(&inode->i_lock);
396 set_page_private(req->wb_page, 0);
397 ClearPagePrivate(req->wb_page);
398 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
401 spin_unlock(&inode->i_lock);
404 spin_unlock(&inode->i_lock);
405 nfs_clear_request(req);
406 nfs_release_request(req);
410 nfs_redirty_request(struct nfs_page *req)
412 __set_page_dirty_nobuffers(req->wb_page);
416 * Check if a request is dirty
419 nfs_dirty_request(struct nfs_page *req)
421 struct page *page = req->wb_page;
423 if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
425 return !PageWriteback(req->wb_page);
428 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
430 * Add a request to the inode's commit list.
433 nfs_mark_request_commit(struct nfs_page *req)
435 struct inode *inode = req->wb_context->path.dentry->d_inode;
436 struct nfs_inode *nfsi = NFS_I(inode);
438 spin_lock(&inode->i_lock);
440 set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
441 radix_tree_tag_set(&nfsi->nfs_page_tree,
443 NFS_PAGE_TAG_COMMIT);
444 spin_unlock(&inode->i_lock);
445 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
446 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
447 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
451 int nfs_write_need_commit(struct nfs_write_data *data)
453 return data->verf.committed != NFS_FILE_SYNC;
457 int nfs_reschedule_unstable_write(struct nfs_page *req)
459 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
460 nfs_mark_request_commit(req);
463 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
464 nfs_redirty_request(req);
471 nfs_mark_request_commit(struct nfs_page *req)
476 int nfs_write_need_commit(struct nfs_write_data *data)
482 int nfs_reschedule_unstable_write(struct nfs_page *req)
489 * Wait for a request to complete.
491 * Interruptible by signals only if mounted with intr flag.
493 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
495 struct nfs_inode *nfsi = NFS_I(inode);
496 struct nfs_page *req;
497 pgoff_t idx_end, next;
498 unsigned int res = 0;
504 idx_end = idx_start + npages - 1;
507 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
508 if (req->wb_index > idx_end)
511 next = req->wb_index + 1;
512 BUG_ON(!NFS_WBACK_BUSY(req));
514 kref_get(&req->wb_kref);
515 spin_unlock(&inode->i_lock);
516 error = nfs_wait_on_request(req);
517 nfs_release_request(req);
518 spin_lock(&inode->i_lock);
526 static void nfs_cancel_commit_list(struct list_head *head)
528 struct nfs_page *req;
530 while(!list_empty(head)) {
531 req = nfs_list_entry(head->next);
532 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
533 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
535 nfs_list_remove_request(req);
536 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
537 nfs_inode_remove_request(req);
538 nfs_unlock_request(req);
542 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
544 * nfs_scan_commit - Scan an inode for commit requests
545 * @inode: NFS inode to scan
546 * @dst: destination list
547 * @idx_start: lower bound of page->index to scan.
548 * @npages: idx_start + npages sets the upper bound to scan.
550 * Moves requests from the inode's 'commit' request list.
551 * The requests are *not* checked to ensure that they form a contiguous set.
554 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
556 struct nfs_inode *nfsi = NFS_I(inode);
559 if (nfsi->ncommit != 0) {
560 res = nfs_scan_list(nfsi, dst, idx_start, npages,
561 NFS_PAGE_TAG_COMMIT);
562 nfsi->ncommit -= res;
567 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
574 * Try to update any existing write request, or create one if there is none.
575 * In order to match, the request's credentials must match those of
576 * the calling process.
578 * Note: Should always be called with the Page Lock held!
580 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
581 struct page *page, unsigned int offset, unsigned int bytes)
583 struct address_space *mapping = page->mapping;
584 struct inode *inode = mapping->host;
585 struct nfs_page *req, *new = NULL;
588 end = offset + bytes;
591 /* Loop over all inode entries and see if we find
592 * A request for the page we wish to update
594 spin_lock(&inode->i_lock);
595 req = nfs_page_find_request_locked(page);
597 if (!nfs_set_page_tag_locked(req)) {
600 spin_unlock(&inode->i_lock);
601 error = nfs_wait_on_request(req);
602 nfs_release_request(req);
605 nfs_release_request(new);
606 return ERR_PTR(error);
610 spin_unlock(&inode->i_lock);
612 nfs_release_request(new);
618 nfs_lock_request_dontget(new);
619 error = nfs_inode_add_request(inode, new);
621 spin_unlock(&inode->i_lock);
622 nfs_unlock_request(new);
623 return ERR_PTR(error);
625 spin_unlock(&inode->i_lock);
629 spin_unlock(&inode->i_lock);
631 new = nfs_create_request(ctx, inode, page, offset, bytes);
636 /* We have a request for our page.
637 * If the creds don't match, or the
638 * page addresses don't match,
639 * tell the caller to wait on the conflicting
642 rqend = req->wb_offset + req->wb_bytes;
643 if (req->wb_context != ctx
644 || req->wb_page != page
645 || !nfs_dirty_request(req)
646 || offset > rqend || end < req->wb_offset) {
647 nfs_clear_page_tag_locked(req);
648 return ERR_PTR(-EBUSY);
651 /* Okay, the request matches. Update the region */
652 if (offset < req->wb_offset) {
653 req->wb_offset = offset;
654 req->wb_pgbase = offset;
655 req->wb_bytes = max(end, rqend) - req->wb_offset;
660 req->wb_bytes = end - req->wb_offset;
664 /* If this page might potentially be marked as up to date,
665 * then we need to zero any uninitalised data. */
666 if (req->wb_pgbase == 0 && req->wb_bytes != PAGE_CACHE_SIZE
667 && !PageUptodate(req->wb_page))
668 zero_user_page(req->wb_page, req->wb_bytes,
669 PAGE_CACHE_SIZE - req->wb_bytes,
674 int nfs_flush_incompatible(struct file *file, struct page *page)
676 struct nfs_open_context *ctx = nfs_file_open_context(file);
677 struct nfs_page *req;
678 int do_flush, status;
680 * Look for a request corresponding to this page. If there
681 * is one, and it belongs to another file, we flush it out
682 * before we try to copy anything into the page. Do this
683 * due to the lack of an ACCESS-type call in NFSv2.
684 * Also do the same if we find a request from an existing
688 req = nfs_page_find_request(page);
691 do_flush = req->wb_page != page || req->wb_context != ctx
692 || !nfs_dirty_request(req);
693 nfs_release_request(req);
696 status = nfs_wb_page(page->mapping->host, page);
697 } while (status == 0);
702 * Update and possibly write a cached page of an NFS file.
704 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
705 * things with a page scheduled for an RPC call (e.g. invalidate it).
707 int nfs_updatepage(struct file *file, struct page *page,
708 unsigned int offset, unsigned int count)
710 struct nfs_open_context *ctx = nfs_file_open_context(file);
711 struct inode *inode = page->mapping->host;
714 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
716 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
717 file->f_path.dentry->d_parent->d_name.name,
718 file->f_path.dentry->d_name.name, count,
719 (long long)(page_offset(page) +offset));
721 /* If we're not using byte range locks, and we know the page
722 * is entirely in cache, it may be more efficient to avoid
723 * fragmenting write requests.
725 if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
726 count = max(count + offset, nfs_page_length(page));
730 status = nfs_writepage_setup(ctx, page, offset, count);
731 __set_page_dirty_nobuffers(page);
733 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
734 status, (long long)i_size_read(inode));
736 nfs_set_pageerror(page);
740 static void nfs_writepage_release(struct nfs_page *req)
743 if (PageError(req->wb_page)) {
744 nfs_end_page_writeback(req->wb_page);
745 nfs_inode_remove_request(req);
746 } else if (!nfs_reschedule_unstable_write(req)) {
747 /* Set the PG_uptodate flag */
748 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
749 nfs_end_page_writeback(req->wb_page);
750 nfs_inode_remove_request(req);
752 nfs_end_page_writeback(req->wb_page);
753 nfs_clear_page_tag_locked(req);
756 static inline int flush_task_priority(int how)
758 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
760 return RPC_PRIORITY_HIGH;
762 return RPC_PRIORITY_LOW;
764 return RPC_PRIORITY_NORMAL;
768 * Set up the argument/result storage required for the RPC call.
770 static void nfs_write_rpcsetup(struct nfs_page *req,
771 struct nfs_write_data *data,
772 const struct rpc_call_ops *call_ops,
773 unsigned int count, unsigned int offset,
779 /* Set up the RPC argument and reply structs
780 * NB: take care not to mess about with data->commit et al. */
783 data->inode = inode = req->wb_context->path.dentry->d_inode;
784 data->cred = req->wb_context->cred;
786 data->args.fh = NFS_FH(inode);
787 data->args.offset = req_offset(req) + offset;
788 data->args.pgbase = req->wb_pgbase + offset;
789 data->args.pages = data->pagevec;
790 data->args.count = count;
791 data->args.context = req->wb_context;
793 data->res.fattr = &data->fattr;
794 data->res.count = count;
795 data->res.verf = &data->verf;
796 nfs_fattr_init(&data->fattr);
798 /* Set up the initial task struct. */
799 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
800 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
801 NFS_PROTO(inode)->write_setup(data, how);
803 data->task.tk_priority = flush_task_priority(how);
804 data->task.tk_cookie = (unsigned long)inode;
806 dprintk("NFS: %5u initiated write call "
807 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
810 (long long)NFS_FILEID(inode),
812 (unsigned long long)data->args.offset);
815 static void nfs_execute_write(struct nfs_write_data *data)
817 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
820 rpc_clnt_sigmask(clnt, &oldset);
821 rpc_execute(&data->task);
822 rpc_clnt_sigunmask(clnt, &oldset);
826 * Generate multiple small requests to write out a single
827 * contiguous dirty area on one page.
829 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
831 struct nfs_page *req = nfs_list_entry(head->next);
832 struct page *page = req->wb_page;
833 struct nfs_write_data *data;
834 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
839 nfs_list_remove_request(req);
843 size_t len = min(nbytes, wsize);
845 data = nfs_writedata_alloc(1);
848 list_add(&data->pages, &list);
851 } while (nbytes != 0);
852 atomic_set(&req->wb_complete, requests);
854 ClearPageError(page);
858 data = list_entry(list.next, struct nfs_write_data, pages);
859 list_del_init(&data->pages);
861 data->pagevec[0] = page;
865 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
869 nfs_execute_write(data);
870 } while (nbytes != 0);
875 while (!list_empty(&list)) {
876 data = list_entry(list.next, struct nfs_write_data, pages);
877 list_del(&data->pages);
878 nfs_writedata_release(data);
880 nfs_redirty_request(req);
881 nfs_end_page_writeback(req->wb_page);
882 nfs_clear_page_tag_locked(req);
887 * Create an RPC task for the given write request and kick it.
888 * The page must have been locked by the caller.
890 * It may happen that the page we're passed is not marked dirty.
891 * This is the case if nfs_updatepage detects a conflicting request
892 * that has been written but not committed.
894 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
896 struct nfs_page *req;
898 struct nfs_write_data *data;
900 data = nfs_writedata_alloc(npages);
904 pages = data->pagevec;
905 while (!list_empty(head)) {
906 req = nfs_list_entry(head->next);
907 nfs_list_remove_request(req);
908 nfs_list_add_request(req, &data->pages);
909 ClearPageError(req->wb_page);
910 *pages++ = req->wb_page;
912 req = nfs_list_entry(data->pages.next);
914 /* Set up the argument struct */
915 nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
917 nfs_execute_write(data);
920 while (!list_empty(head)) {
921 req = nfs_list_entry(head->next);
922 nfs_list_remove_request(req);
923 nfs_redirty_request(req);
924 nfs_end_page_writeback(req->wb_page);
925 nfs_clear_page_tag_locked(req);
930 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
931 struct inode *inode, int ioflags)
933 int wsize = NFS_SERVER(inode)->wsize;
935 if (wsize < PAGE_CACHE_SIZE)
936 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
938 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
942 * Handle a write reply that flushed part of a page.
944 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
946 struct nfs_write_data *data = calldata;
947 struct nfs_page *req = data->req;
948 struct page *page = req->wb_page;
950 dprintk("NFS: write (%s/%Ld %d@%Ld)",
951 req->wb_context->path.dentry->d_inode->i_sb->s_id,
952 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
954 (long long)req_offset(req));
956 if (nfs_writeback_done(task, data) != 0)
959 if (task->tk_status < 0) {
960 nfs_set_pageerror(page);
961 nfs_context_set_write_error(req->wb_context, task->tk_status);
962 dprintk(", error = %d\n", task->tk_status);
966 if (nfs_write_need_commit(data)) {
967 struct inode *inode = page->mapping->host;
969 spin_lock(&inode->i_lock);
970 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
971 /* Do nothing we need to resend the writes */
972 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
973 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
974 dprintk(" defer commit\n");
975 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
976 set_bit(PG_NEED_RESCHED, &req->wb_flags);
977 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
978 dprintk(" server reboot detected\n");
980 spin_unlock(&inode->i_lock);
985 if (atomic_dec_and_test(&req->wb_complete))
986 nfs_writepage_release(req);
989 static const struct rpc_call_ops nfs_write_partial_ops = {
990 .rpc_call_done = nfs_writeback_done_partial,
991 .rpc_release = nfs_writedata_release,
995 * Handle a write reply that flushes a whole page.
997 * FIXME: There is an inherent race with invalidate_inode_pages and
998 * writebacks since the page->count is kept > 1 for as long
999 * as the page has a write request pending.
1001 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1003 struct nfs_write_data *data = calldata;
1004 struct nfs_page *req;
1007 if (nfs_writeback_done(task, data) != 0)
1010 /* Update attributes as result of writeback. */
1011 while (!list_empty(&data->pages)) {
1012 req = nfs_list_entry(data->pages.next);
1013 nfs_list_remove_request(req);
1014 page = req->wb_page;
1016 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1017 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1018 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1020 (long long)req_offset(req));
1022 if (task->tk_status < 0) {
1023 nfs_set_pageerror(page);
1024 nfs_context_set_write_error(req->wb_context, task->tk_status);
1025 dprintk(", error = %d\n", task->tk_status);
1026 goto remove_request;
1029 if (nfs_write_need_commit(data)) {
1030 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1031 nfs_mark_request_commit(req);
1032 nfs_end_page_writeback(page);
1033 dprintk(" marked for commit\n");
1036 /* Set the PG_uptodate flag? */
1037 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1040 nfs_end_page_writeback(page);
1041 nfs_inode_remove_request(req);
1043 nfs_clear_page_tag_locked(req);
1047 static const struct rpc_call_ops nfs_write_full_ops = {
1048 .rpc_call_done = nfs_writeback_done_full,
1049 .rpc_release = nfs_writedata_release,
1054 * This function is called when the WRITE call is complete.
1056 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1058 struct nfs_writeargs *argp = &data->args;
1059 struct nfs_writeres *resp = &data->res;
1062 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1063 task->tk_pid, task->tk_status);
1066 * ->write_done will attempt to use post-op attributes to detect
1067 * conflicting writes by other clients. A strict interpretation
1068 * of close-to-open would allow us to continue caching even if
1069 * another writer had changed the file, but some applications
1070 * depend on tighter cache coherency when writing.
1072 status = NFS_PROTO(data->inode)->write_done(task, data);
1075 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1077 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1078 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1079 /* We tried a write call, but the server did not
1080 * commit data to stable storage even though we
1082 * Note: There is a known bug in Tru64 < 5.0 in which
1083 * the server reports NFS_DATA_SYNC, but performs
1084 * NFS_FILE_SYNC. We therefore implement this checking
1085 * as a dprintk() in order to avoid filling syslog.
1087 static unsigned long complain;
1089 if (time_before(complain, jiffies)) {
1090 dprintk("NFS: faulty NFS server %s:"
1091 " (committed = %d) != (stable = %d)\n",
1092 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1093 resp->verf->committed, argp->stable);
1094 complain = jiffies + 300 * HZ;
1098 /* Is this a short write? */
1099 if (task->tk_status >= 0 && resp->count < argp->count) {
1100 static unsigned long complain;
1102 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1104 /* Has the server at least made some progress? */
1105 if (resp->count != 0) {
1106 /* Was this an NFSv2 write or an NFSv3 stable write? */
1107 if (resp->verf->committed != NFS_UNSTABLE) {
1108 /* Resend from where the server left off */
1109 argp->offset += resp->count;
1110 argp->pgbase += resp->count;
1111 argp->count -= resp->count;
1113 /* Resend as a stable write in order to avoid
1114 * headaches in the case of a server crash.
1116 argp->stable = NFS_FILE_SYNC;
1118 rpc_restart_call(task);
1121 if (time_before(complain, jiffies)) {
1123 "NFS: Server wrote zero bytes, expected %u.\n",
1125 complain = jiffies + 300 * HZ;
1127 /* Can't do anything about it except throw an error. */
1128 task->tk_status = -EIO;
1134 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1135 void nfs_commit_release(void *wdata)
1137 nfs_commit_free(wdata);
1141 * Set up the argument/result storage required for the RPC call.
1143 static void nfs_commit_rpcsetup(struct list_head *head,
1144 struct nfs_write_data *data,
1147 struct nfs_page *first;
1148 struct inode *inode;
1151 /* Set up the RPC argument and reply structs
1152 * NB: take care not to mess about with data->commit et al. */
1154 list_splice_init(head, &data->pages);
1155 first = nfs_list_entry(data->pages.next);
1156 inode = first->wb_context->path.dentry->d_inode;
1158 data->inode = inode;
1159 data->cred = first->wb_context->cred;
1161 data->args.fh = NFS_FH(data->inode);
1162 /* Note: we always request a commit of the entire inode */
1163 data->args.offset = 0;
1164 data->args.count = 0;
1165 data->res.count = 0;
1166 data->res.fattr = &data->fattr;
1167 data->res.verf = &data->verf;
1168 nfs_fattr_init(&data->fattr);
1170 /* Set up the initial task struct. */
1171 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1172 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1173 NFS_PROTO(inode)->commit_setup(data, how);
1175 data->task.tk_priority = flush_task_priority(how);
1176 data->task.tk_cookie = (unsigned long)inode;
1178 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1182 * Commit dirty pages
1185 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1187 struct nfs_write_data *data;
1188 struct nfs_page *req;
1190 data = nfs_commit_alloc();
1195 /* Set up the argument struct */
1196 nfs_commit_rpcsetup(head, data, how);
1198 nfs_execute_write(data);
1201 while (!list_empty(head)) {
1202 req = nfs_list_entry(head->next);
1203 nfs_list_remove_request(req);
1204 nfs_mark_request_commit(req);
1205 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1206 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1208 nfs_clear_page_tag_locked(req);
1214 * COMMIT call returned
1216 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1218 struct nfs_write_data *data = calldata;
1219 struct nfs_page *req;
1221 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1222 task->tk_pid, task->tk_status);
1224 /* Call the NFS version-specific code */
1225 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1228 while (!list_empty(&data->pages)) {
1229 req = nfs_list_entry(data->pages.next);
1230 nfs_list_remove_request(req);
1231 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1232 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1233 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1236 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1237 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1238 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1240 (long long)req_offset(req));
1241 if (task->tk_status < 0) {
1242 nfs_context_set_write_error(req->wb_context, task->tk_status);
1243 nfs_inode_remove_request(req);
1244 dprintk(", error = %d\n", task->tk_status);
1248 /* Okay, COMMIT succeeded, apparently. Check the verifier
1249 * returned by the server against all stored verfs. */
1250 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1251 /* We have a match */
1252 /* Set the PG_uptodate flag */
1253 nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1255 nfs_inode_remove_request(req);
1259 /* We have a mismatch. Write the page again */
1260 dprintk(" mismatch\n");
1261 nfs_redirty_request(req);
1263 nfs_clear_page_tag_locked(req);
1267 static const struct rpc_call_ops nfs_commit_ops = {
1268 .rpc_call_done = nfs_commit_done,
1269 .rpc_release = nfs_commit_release,
1272 int nfs_commit_inode(struct inode *inode, int how)
1277 spin_lock(&inode->i_lock);
1278 res = nfs_scan_commit(inode, &head, 0, 0);
1279 spin_unlock(&inode->i_lock);
1281 int error = nfs_commit_list(inode, &head, how);
1288 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1294 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1296 struct inode *inode = mapping->host;
1297 pgoff_t idx_start, idx_end;
1298 unsigned int npages = 0;
1300 int nocommit = how & FLUSH_NOCOMMIT;
1304 if (wbc->range_cyclic)
1307 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1308 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1309 if (idx_end > idx_start) {
1310 pgoff_t l_npages = 1 + idx_end - idx_start;
1312 if (sizeof(npages) != sizeof(l_npages) &&
1313 (pgoff_t)npages != l_npages)
1317 how &= ~FLUSH_NOCOMMIT;
1318 spin_lock(&inode->i_lock);
1320 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1325 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1328 if (how & FLUSH_INVALIDATE) {
1329 spin_unlock(&inode->i_lock);
1330 nfs_cancel_commit_list(&head);
1332 spin_lock(&inode->i_lock);
1335 pages += nfs_scan_commit(inode, &head, 0, 0);
1336 spin_unlock(&inode->i_lock);
1337 ret = nfs_commit_list(inode, &head, how);
1338 spin_lock(&inode->i_lock);
1341 spin_unlock(&inode->i_lock);
1345 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1349 ret = nfs_writepages(mapping, wbc);
1352 ret = nfs_sync_mapping_wait(mapping, wbc, how);
1357 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1361 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1362 static int nfs_write_mapping(struct address_space *mapping, int how)
1364 struct writeback_control wbc = {
1365 .bdi = mapping->backing_dev_info,
1366 .sync_mode = WB_SYNC_NONE,
1367 .nr_to_write = LONG_MAX,
1368 .for_writepages = 1,
1373 ret = __nfs_write_mapping(mapping, &wbc, how);
1376 wbc.sync_mode = WB_SYNC_ALL;
1377 return __nfs_write_mapping(mapping, &wbc, how);
1381 * flush the inode to disk.
1383 int nfs_wb_all(struct inode *inode)
1385 return nfs_write_mapping(inode->i_mapping, 0);
1388 int nfs_wb_nocommit(struct inode *inode)
1390 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1393 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1395 struct nfs_page *req;
1396 loff_t range_start = page_offset(page);
1397 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1398 struct writeback_control wbc = {
1399 .bdi = page->mapping->backing_dev_info,
1400 .sync_mode = WB_SYNC_ALL,
1401 .nr_to_write = LONG_MAX,
1402 .range_start = range_start,
1403 .range_end = range_end,
1407 BUG_ON(!PageLocked(page));
1409 req = nfs_page_find_request(page);
1412 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1413 nfs_release_request(req);
1416 if (nfs_lock_request_dontget(req)) {
1417 nfs_inode_remove_request(req);
1419 * In case nfs_inode_remove_request has marked the
1420 * page as being dirty
1422 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1423 nfs_unlock_request(req);
1426 ret = nfs_wait_on_request(req);
1430 if (!PagePrivate(page))
1432 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1437 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1440 loff_t range_start = page_offset(page);
1441 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1442 struct writeback_control wbc = {
1443 .bdi = page->mapping->backing_dev_info,
1444 .sync_mode = WB_SYNC_ALL,
1445 .nr_to_write = LONG_MAX,
1446 .range_start = range_start,
1447 .range_end = range_end,
1451 BUG_ON(!PageLocked(page));
1452 if (clear_page_dirty_for_io(page)) {
1453 ret = nfs_writepage_locked(page, &wbc);
1457 if (!PagePrivate(page))
1459 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1463 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1468 * Write back all requests on one page - we do this before reading it.
1470 int nfs_wb_page(struct inode *inode, struct page* page)
1472 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1475 int __init nfs_init_writepagecache(void)
1477 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1478 sizeof(struct nfs_write_data),
1479 0, SLAB_HWCACHE_ALIGN,
1481 if (nfs_wdata_cachep == NULL)
1484 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1486 if (nfs_wdata_mempool == NULL)
1489 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1491 if (nfs_commit_mempool == NULL)
1495 * NFS congestion size, scale with available memory.
1507 * This allows larger machines to have larger/more transfers.
1508 * Limit the default to 256M
1510 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1511 if (nfs_congestion_kb > 256*1024)
1512 nfs_congestion_kb = 256*1024;
1517 void nfs_destroy_writepagecache(void)
1519 mempool_destroy(nfs_commit_mempool);
1520 mempool_destroy(nfs_wdata_mempool);
1521 kmem_cache_destroy(nfs_wdata_cachep);