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 void 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);
372 if (nfs_have_delegation(inode, FMODE_WRITE))
375 SetPagePrivate(req->wb_page);
376 set_page_private(req->wb_page, (unsigned long)req);
378 kref_get(&req->wb_kref);
379 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
380 NFS_PAGE_TAG_LOCKED);
384 * Remove a write request from an inode
386 static void nfs_inode_remove_request(struct nfs_page *req)
388 struct inode *inode = req->wb_context->path.dentry->d_inode;
389 struct nfs_inode *nfsi = NFS_I(inode);
391 BUG_ON (!NFS_WBACK_BUSY(req));
393 spin_lock(&inode->i_lock);
394 set_page_private(req->wb_page, 0);
395 ClearPagePrivate(req->wb_page);
396 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
399 spin_unlock(&inode->i_lock);
402 spin_unlock(&inode->i_lock);
403 nfs_clear_request(req);
404 nfs_release_request(req);
408 nfs_redirty_request(struct nfs_page *req)
410 __set_page_dirty_nobuffers(req->wb_page);
414 * Check if a request is dirty
417 nfs_dirty_request(struct nfs_page *req)
419 struct page *page = req->wb_page;
421 if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
423 return !PageWriteback(req->wb_page);
426 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
428 * Add a request to the inode's commit list.
431 nfs_mark_request_commit(struct nfs_page *req)
433 struct inode *inode = req->wb_context->path.dentry->d_inode;
434 struct nfs_inode *nfsi = NFS_I(inode);
436 spin_lock(&inode->i_lock);
438 set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
439 radix_tree_tag_set(&nfsi->nfs_page_tree,
441 NFS_PAGE_TAG_COMMIT);
442 spin_unlock(&inode->i_lock);
443 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
444 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
445 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
449 int nfs_write_need_commit(struct nfs_write_data *data)
451 return data->verf.committed != NFS_FILE_SYNC;
455 int nfs_reschedule_unstable_write(struct nfs_page *req)
457 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
458 nfs_mark_request_commit(req);
461 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
462 nfs_redirty_request(req);
469 nfs_mark_request_commit(struct nfs_page *req)
474 int nfs_write_need_commit(struct nfs_write_data *data)
480 int nfs_reschedule_unstable_write(struct nfs_page *req)
487 * Wait for a request to complete.
489 * Interruptible by fatal signals only.
491 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
493 struct nfs_inode *nfsi = NFS_I(inode);
494 struct nfs_page *req;
495 pgoff_t idx_end, next;
496 unsigned int res = 0;
502 idx_end = idx_start + npages - 1;
505 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
506 if (req->wb_index > idx_end)
509 next = req->wb_index + 1;
510 BUG_ON(!NFS_WBACK_BUSY(req));
512 kref_get(&req->wb_kref);
513 spin_unlock(&inode->i_lock);
514 error = nfs_wait_on_request(req);
515 nfs_release_request(req);
516 spin_lock(&inode->i_lock);
524 static void nfs_cancel_commit_list(struct list_head *head)
526 struct nfs_page *req;
528 while(!list_empty(head)) {
529 req = nfs_list_entry(head->next);
530 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
531 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
533 nfs_list_remove_request(req);
534 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
535 nfs_inode_remove_request(req);
536 nfs_unlock_request(req);
540 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
542 * nfs_scan_commit - Scan an inode for commit requests
543 * @inode: NFS inode to scan
544 * @dst: destination list
545 * @idx_start: lower bound of page->index to scan.
546 * @npages: idx_start + npages sets the upper bound to scan.
548 * Moves requests from the inode's 'commit' request list.
549 * The requests are *not* checked to ensure that they form a contiguous set.
552 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
554 struct nfs_inode *nfsi = NFS_I(inode);
557 if (nfsi->ncommit != 0) {
558 res = nfs_scan_list(nfsi, dst, idx_start, npages,
559 NFS_PAGE_TAG_COMMIT);
560 nfsi->ncommit -= res;
565 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
572 * Try to update any existing write request, or create one if there is none.
573 * In order to match, the request's credentials must match those of
574 * the calling process.
576 * Note: Should always be called with the Page Lock held!
578 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
579 struct page *page, unsigned int offset, unsigned int bytes)
581 struct address_space *mapping = page->mapping;
582 struct inode *inode = mapping->host;
583 struct nfs_page *req, *new = NULL;
586 end = offset + bytes;
589 /* Loop over all inode entries and see if we find
590 * A request for the page we wish to update
593 if (radix_tree_preload(GFP_NOFS)) {
594 nfs_release_request(new);
595 return ERR_PTR(-ENOMEM);
599 spin_lock(&inode->i_lock);
600 req = nfs_page_find_request_locked(page);
602 if (!nfs_set_page_tag_locked(req)) {
605 spin_unlock(&inode->i_lock);
606 error = nfs_wait_on_request(req);
607 nfs_release_request(req);
610 radix_tree_preload_end();
611 nfs_release_request(new);
613 return ERR_PTR(error);
617 spin_unlock(&inode->i_lock);
619 radix_tree_preload_end();
620 nfs_release_request(new);
626 nfs_lock_request_dontget(new);
627 nfs_inode_add_request(inode, new);
628 spin_unlock(&inode->i_lock);
629 radix_tree_preload_end();
633 spin_unlock(&inode->i_lock);
635 new = nfs_create_request(ctx, inode, page, offset, bytes);
640 /* We have a request for our page.
641 * If the creds don't match, or the
642 * page addresses don't match,
643 * tell the caller to wait on the conflicting
646 rqend = req->wb_offset + req->wb_bytes;
647 if (req->wb_context != ctx
648 || req->wb_page != page
649 || !nfs_dirty_request(req)
650 || offset > rqend || end < req->wb_offset) {
651 nfs_clear_page_tag_locked(req);
652 return ERR_PTR(-EBUSY);
655 /* Okay, the request matches. Update the region */
656 if (offset < req->wb_offset) {
657 req->wb_offset = offset;
658 req->wb_pgbase = offset;
659 req->wb_bytes = max(end, rqend) - req->wb_offset;
664 req->wb_bytes = end - req->wb_offset;
668 /* If this page might potentially be marked as up to date,
669 * then we need to zero any uninitalised data. */
670 if (req->wb_pgbase == 0 && req->wb_bytes != PAGE_CACHE_SIZE
671 && !PageUptodate(req->wb_page))
672 zero_user_segment(req->wb_page, req->wb_bytes, PAGE_CACHE_SIZE);
676 int nfs_flush_incompatible(struct file *file, struct page *page)
678 struct nfs_open_context *ctx = nfs_file_open_context(file);
679 struct nfs_page *req;
680 int do_flush, status;
682 * Look for a request corresponding to this page. If there
683 * is one, and it belongs to another file, we flush it out
684 * before we try to copy anything into the page. Do this
685 * due to the lack of an ACCESS-type call in NFSv2.
686 * Also do the same if we find a request from an existing
690 req = nfs_page_find_request(page);
693 do_flush = req->wb_page != page || req->wb_context != ctx
694 || !nfs_dirty_request(req);
695 nfs_release_request(req);
698 status = nfs_wb_page(page->mapping->host, page);
699 } while (status == 0);
704 * If the page cache is marked as unsafe or invalid, then we can't rely on
705 * the PageUptodate() flag. In this case, we will need to turn off
706 * write optimisations that depend on the page contents being correct.
708 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
710 return PageUptodate(page) &&
711 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
715 * Update and possibly write a cached page of an NFS file.
717 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
718 * things with a page scheduled for an RPC call (e.g. invalidate it).
720 int nfs_updatepage(struct file *file, struct page *page,
721 unsigned int offset, unsigned int count)
723 struct nfs_open_context *ctx = nfs_file_open_context(file);
724 struct inode *inode = page->mapping->host;
727 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
729 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
730 file->f_path.dentry->d_parent->d_name.name,
731 file->f_path.dentry->d_name.name, count,
732 (long long)(page_offset(page) +offset));
734 /* If we're not using byte range locks, and we know the page
735 * is up to date, it may be more efficient to extend the write
736 * to cover the entire page in order to avoid fragmentation
739 if (nfs_write_pageuptodate(page, inode) &&
740 inode->i_flock == NULL &&
741 !(file->f_mode & O_SYNC)) {
742 count = max(count + offset, nfs_page_length(page));
746 status = nfs_writepage_setup(ctx, page, offset, count);
747 __set_page_dirty_nobuffers(page);
749 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
750 status, (long long)i_size_read(inode));
752 nfs_set_pageerror(page);
756 static void nfs_writepage_release(struct nfs_page *req)
759 if (PageError(req->wb_page)) {
760 nfs_end_page_writeback(req->wb_page);
761 nfs_inode_remove_request(req);
762 } else if (!nfs_reschedule_unstable_write(req)) {
763 /* Set the PG_uptodate flag */
764 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
765 nfs_end_page_writeback(req->wb_page);
766 nfs_inode_remove_request(req);
768 nfs_end_page_writeback(req->wb_page);
769 nfs_clear_page_tag_locked(req);
772 static int flush_task_priority(int how)
774 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
776 return RPC_PRIORITY_HIGH;
778 return RPC_PRIORITY_LOW;
780 return RPC_PRIORITY_NORMAL;
784 * Set up the argument/result storage required for the RPC call.
786 static void nfs_write_rpcsetup(struct nfs_page *req,
787 struct nfs_write_data *data,
788 const struct rpc_call_ops *call_ops,
789 unsigned int count, unsigned int offset,
792 struct inode *inode = req->wb_context->path.dentry->d_inode;
793 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
794 int priority = flush_task_priority(how);
795 struct rpc_task *task;
796 struct rpc_message msg = {
797 .rpc_argp = &data->args,
798 .rpc_resp = &data->res,
799 .rpc_cred = req->wb_context->cred,
801 struct rpc_task_setup task_setup_data = {
802 .rpc_client = NFS_CLIENT(inode),
805 .callback_ops = call_ops,
806 .callback_data = data,
808 .priority = priority,
811 /* Set up the RPC argument and reply structs
812 * NB: take care not to mess about with data->commit et al. */
815 data->inode = inode = req->wb_context->path.dentry->d_inode;
816 data->cred = msg.rpc_cred;
818 data->args.fh = NFS_FH(inode);
819 data->args.offset = req_offset(req) + offset;
820 data->args.pgbase = req->wb_pgbase + offset;
821 data->args.pages = data->pagevec;
822 data->args.count = count;
823 data->args.context = req->wb_context;
824 data->args.stable = NFS_UNSTABLE;
825 if (how & FLUSH_STABLE) {
826 data->args.stable = NFS_DATA_SYNC;
827 if (!NFS_I(inode)->ncommit)
828 data->args.stable = NFS_FILE_SYNC;
831 data->res.fattr = &data->fattr;
832 data->res.count = count;
833 data->res.verf = &data->verf;
834 nfs_fattr_init(&data->fattr);
836 /* Set up the initial task struct. */
837 NFS_PROTO(inode)->write_setup(data, &msg);
839 dprintk("NFS: %5u initiated write call "
840 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
843 (long long)NFS_FILEID(inode),
845 (unsigned long long)data->args.offset);
847 task = rpc_run_task(&task_setup_data);
853 * Generate multiple small requests to write out a single
854 * contiguous dirty area on one page.
856 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
858 struct nfs_page *req = nfs_list_entry(head->next);
859 struct page *page = req->wb_page;
860 struct nfs_write_data *data;
861 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
866 nfs_list_remove_request(req);
870 size_t len = min(nbytes, wsize);
872 data = nfs_writedata_alloc(1);
875 list_add(&data->pages, &list);
878 } while (nbytes != 0);
879 atomic_set(&req->wb_complete, requests);
881 ClearPageError(page);
885 data = list_entry(list.next, struct nfs_write_data, pages);
886 list_del_init(&data->pages);
888 data->pagevec[0] = page;
892 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
896 } while (nbytes != 0);
901 while (!list_empty(&list)) {
902 data = list_entry(list.next, struct nfs_write_data, pages);
903 list_del(&data->pages);
904 nfs_writedata_release(data);
906 nfs_redirty_request(req);
907 nfs_end_page_writeback(req->wb_page);
908 nfs_clear_page_tag_locked(req);
913 * Create an RPC task for the given write request and kick it.
914 * The page must have been locked by the caller.
916 * It may happen that the page we're passed is not marked dirty.
917 * This is the case if nfs_updatepage detects a conflicting request
918 * that has been written but not committed.
920 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
922 struct nfs_page *req;
924 struct nfs_write_data *data;
926 data = nfs_writedata_alloc(npages);
930 pages = data->pagevec;
931 while (!list_empty(head)) {
932 req = nfs_list_entry(head->next);
933 nfs_list_remove_request(req);
934 nfs_list_add_request(req, &data->pages);
935 ClearPageError(req->wb_page);
936 *pages++ = req->wb_page;
938 req = nfs_list_entry(data->pages.next);
940 /* Set up the argument struct */
941 nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
945 while (!list_empty(head)) {
946 req = nfs_list_entry(head->next);
947 nfs_list_remove_request(req);
948 nfs_redirty_request(req);
949 nfs_end_page_writeback(req->wb_page);
950 nfs_clear_page_tag_locked(req);
955 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
956 struct inode *inode, int ioflags)
958 size_t wsize = NFS_SERVER(inode)->wsize;
960 if (wsize < PAGE_CACHE_SIZE)
961 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
963 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
967 * Handle a write reply that flushed part of a page.
969 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
971 struct nfs_write_data *data = calldata;
972 struct nfs_page *req = data->req;
973 struct page *page = req->wb_page;
975 dprintk("NFS: write (%s/%Ld %d@%Ld)",
976 req->wb_context->path.dentry->d_inode->i_sb->s_id,
977 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
979 (long long)req_offset(req));
981 if (nfs_writeback_done(task, data) != 0)
984 if (task->tk_status < 0) {
985 nfs_set_pageerror(page);
986 nfs_context_set_write_error(req->wb_context, task->tk_status);
987 dprintk(", error = %d\n", task->tk_status);
991 if (nfs_write_need_commit(data)) {
992 struct inode *inode = page->mapping->host;
994 spin_lock(&inode->i_lock);
995 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
996 /* Do nothing we need to resend the writes */
997 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
998 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
999 dprintk(" defer commit\n");
1000 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1001 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1002 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1003 dprintk(" server reboot detected\n");
1005 spin_unlock(&inode->i_lock);
1010 if (atomic_dec_and_test(&req->wb_complete))
1011 nfs_writepage_release(req);
1014 static const struct rpc_call_ops nfs_write_partial_ops = {
1015 .rpc_call_done = nfs_writeback_done_partial,
1016 .rpc_release = nfs_writedata_release,
1020 * Handle a write reply that flushes a whole page.
1022 * FIXME: There is an inherent race with invalidate_inode_pages and
1023 * writebacks since the page->count is kept > 1 for as long
1024 * as the page has a write request pending.
1026 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1028 struct nfs_write_data *data = calldata;
1029 struct nfs_page *req;
1032 if (nfs_writeback_done(task, data) != 0)
1035 /* Update attributes as result of writeback. */
1036 while (!list_empty(&data->pages)) {
1037 req = nfs_list_entry(data->pages.next);
1038 nfs_list_remove_request(req);
1039 page = req->wb_page;
1041 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1042 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1043 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1045 (long long)req_offset(req));
1047 if (task->tk_status < 0) {
1048 nfs_set_pageerror(page);
1049 nfs_context_set_write_error(req->wb_context, task->tk_status);
1050 dprintk(", error = %d\n", task->tk_status);
1051 goto remove_request;
1054 if (nfs_write_need_commit(data)) {
1055 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1056 nfs_mark_request_commit(req);
1057 nfs_end_page_writeback(page);
1058 dprintk(" marked for commit\n");
1061 /* Set the PG_uptodate flag? */
1062 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1065 nfs_end_page_writeback(page);
1066 nfs_inode_remove_request(req);
1068 nfs_clear_page_tag_locked(req);
1072 static const struct rpc_call_ops nfs_write_full_ops = {
1073 .rpc_call_done = nfs_writeback_done_full,
1074 .rpc_release = nfs_writedata_release,
1079 * This function is called when the WRITE call is complete.
1081 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1083 struct nfs_writeargs *argp = &data->args;
1084 struct nfs_writeres *resp = &data->res;
1087 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1088 task->tk_pid, task->tk_status);
1091 * ->write_done will attempt to use post-op attributes to detect
1092 * conflicting writes by other clients. A strict interpretation
1093 * of close-to-open would allow us to continue caching even if
1094 * another writer had changed the file, but some applications
1095 * depend on tighter cache coherency when writing.
1097 status = NFS_PROTO(data->inode)->write_done(task, data);
1100 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1102 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1103 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1104 /* We tried a write call, but the server did not
1105 * commit data to stable storage even though we
1107 * Note: There is a known bug in Tru64 < 5.0 in which
1108 * the server reports NFS_DATA_SYNC, but performs
1109 * NFS_FILE_SYNC. We therefore implement this checking
1110 * as a dprintk() in order to avoid filling syslog.
1112 static unsigned long complain;
1114 if (time_before(complain, jiffies)) {
1115 dprintk("NFS: faulty NFS server %s:"
1116 " (committed = %d) != (stable = %d)\n",
1117 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1118 resp->verf->committed, argp->stable);
1119 complain = jiffies + 300 * HZ;
1123 /* Is this a short write? */
1124 if (task->tk_status >= 0 && resp->count < argp->count) {
1125 static unsigned long complain;
1127 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1129 /* Has the server at least made some progress? */
1130 if (resp->count != 0) {
1131 /* Was this an NFSv2 write or an NFSv3 stable write? */
1132 if (resp->verf->committed != NFS_UNSTABLE) {
1133 /* Resend from where the server left off */
1134 argp->offset += resp->count;
1135 argp->pgbase += resp->count;
1136 argp->count -= resp->count;
1138 /* Resend as a stable write in order to avoid
1139 * headaches in the case of a server crash.
1141 argp->stable = NFS_FILE_SYNC;
1143 rpc_restart_call(task);
1146 if (time_before(complain, jiffies)) {
1148 "NFS: Server wrote zero bytes, expected %u.\n",
1150 complain = jiffies + 300 * HZ;
1152 /* Can't do anything about it except throw an error. */
1153 task->tk_status = -EIO;
1159 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1160 void nfs_commit_release(void *wdata)
1162 nfs_commit_free(wdata);
1166 * Set up the argument/result storage required for the RPC call.
1168 static void nfs_commit_rpcsetup(struct list_head *head,
1169 struct nfs_write_data *data,
1172 struct nfs_page *first = nfs_list_entry(head->next);
1173 struct inode *inode = first->wb_context->path.dentry->d_inode;
1174 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1175 int priority = flush_task_priority(how);
1176 struct rpc_task *task;
1177 struct rpc_message msg = {
1178 .rpc_argp = &data->args,
1179 .rpc_resp = &data->res,
1180 .rpc_cred = first->wb_context->cred,
1182 struct rpc_task_setup task_setup_data = {
1183 .task = &data->task,
1184 .rpc_client = NFS_CLIENT(inode),
1185 .rpc_message = &msg,
1186 .callback_ops = &nfs_commit_ops,
1187 .callback_data = data,
1189 .priority = priority,
1192 /* Set up the RPC argument and reply structs
1193 * NB: take care not to mess about with data->commit et al. */
1195 list_splice_init(head, &data->pages);
1197 data->inode = inode;
1198 data->cred = msg.rpc_cred;
1200 data->args.fh = NFS_FH(data->inode);
1201 /* Note: we always request a commit of the entire inode */
1202 data->args.offset = 0;
1203 data->args.count = 0;
1204 data->res.count = 0;
1205 data->res.fattr = &data->fattr;
1206 data->res.verf = &data->verf;
1207 nfs_fattr_init(&data->fattr);
1209 /* Set up the initial task struct. */
1210 NFS_PROTO(inode)->commit_setup(data, &msg);
1212 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1214 task = rpc_run_task(&task_setup_data);
1220 * Commit dirty pages
1223 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1225 struct nfs_write_data *data;
1226 struct nfs_page *req;
1228 data = nfs_commit_alloc();
1233 /* Set up the argument struct */
1234 nfs_commit_rpcsetup(head, data, how);
1238 while (!list_empty(head)) {
1239 req = nfs_list_entry(head->next);
1240 nfs_list_remove_request(req);
1241 nfs_mark_request_commit(req);
1242 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1243 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1245 nfs_clear_page_tag_locked(req);
1251 * COMMIT call returned
1253 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1255 struct nfs_write_data *data = calldata;
1256 struct nfs_page *req;
1258 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1259 task->tk_pid, task->tk_status);
1261 /* Call the NFS version-specific code */
1262 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1265 while (!list_empty(&data->pages)) {
1266 req = nfs_list_entry(data->pages.next);
1267 nfs_list_remove_request(req);
1268 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1269 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1270 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1273 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1274 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1275 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1277 (long long)req_offset(req));
1278 if (task->tk_status < 0) {
1279 nfs_context_set_write_error(req->wb_context, task->tk_status);
1280 nfs_inode_remove_request(req);
1281 dprintk(", error = %d\n", task->tk_status);
1285 /* Okay, COMMIT succeeded, apparently. Check the verifier
1286 * returned by the server against all stored verfs. */
1287 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1288 /* We have a match */
1289 /* Set the PG_uptodate flag */
1290 nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1292 nfs_inode_remove_request(req);
1296 /* We have a mismatch. Write the page again */
1297 dprintk(" mismatch\n");
1298 nfs_redirty_request(req);
1300 nfs_clear_page_tag_locked(req);
1304 static const struct rpc_call_ops nfs_commit_ops = {
1305 .rpc_call_done = nfs_commit_done,
1306 .rpc_release = nfs_commit_release,
1309 int nfs_commit_inode(struct inode *inode, int how)
1314 spin_lock(&inode->i_lock);
1315 res = nfs_scan_commit(inode, &head, 0, 0);
1316 spin_unlock(&inode->i_lock);
1318 int error = nfs_commit_list(inode, &head, how);
1325 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1331 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1333 struct inode *inode = mapping->host;
1334 pgoff_t idx_start, idx_end;
1335 unsigned int npages = 0;
1337 int nocommit = how & FLUSH_NOCOMMIT;
1341 if (wbc->range_cyclic)
1344 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1345 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1346 if (idx_end > idx_start) {
1347 pgoff_t l_npages = 1 + idx_end - idx_start;
1349 if (sizeof(npages) != sizeof(l_npages) &&
1350 (pgoff_t)npages != l_npages)
1354 how &= ~FLUSH_NOCOMMIT;
1355 spin_lock(&inode->i_lock);
1357 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1362 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1365 if (how & FLUSH_INVALIDATE) {
1366 spin_unlock(&inode->i_lock);
1367 nfs_cancel_commit_list(&head);
1369 spin_lock(&inode->i_lock);
1372 pages += nfs_scan_commit(inode, &head, 0, 0);
1373 spin_unlock(&inode->i_lock);
1374 ret = nfs_commit_list(inode, &head, how);
1375 spin_lock(&inode->i_lock);
1378 spin_unlock(&inode->i_lock);
1382 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1386 ret = nfs_writepages(mapping, wbc);
1389 ret = nfs_sync_mapping_wait(mapping, wbc, how);
1394 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1398 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1399 static int nfs_write_mapping(struct address_space *mapping, int how)
1401 struct writeback_control wbc = {
1402 .bdi = mapping->backing_dev_info,
1403 .sync_mode = WB_SYNC_NONE,
1404 .nr_to_write = LONG_MAX,
1405 .for_writepages = 1,
1410 ret = __nfs_write_mapping(mapping, &wbc, how);
1413 wbc.sync_mode = WB_SYNC_ALL;
1414 return __nfs_write_mapping(mapping, &wbc, how);
1418 * flush the inode to disk.
1420 int nfs_wb_all(struct inode *inode)
1422 return nfs_write_mapping(inode->i_mapping, 0);
1425 int nfs_wb_nocommit(struct inode *inode)
1427 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1430 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1432 struct nfs_page *req;
1433 loff_t range_start = page_offset(page);
1434 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1435 struct writeback_control wbc = {
1436 .bdi = page->mapping->backing_dev_info,
1437 .sync_mode = WB_SYNC_ALL,
1438 .nr_to_write = LONG_MAX,
1439 .range_start = range_start,
1440 .range_end = range_end,
1444 BUG_ON(!PageLocked(page));
1446 req = nfs_page_find_request(page);
1449 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1450 nfs_release_request(req);
1453 if (nfs_lock_request_dontget(req)) {
1454 nfs_inode_remove_request(req);
1456 * In case nfs_inode_remove_request has marked the
1457 * page as being dirty
1459 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1460 nfs_unlock_request(req);
1463 ret = nfs_wait_on_request(req);
1467 if (!PagePrivate(page))
1469 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1474 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1477 loff_t range_start = page_offset(page);
1478 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1479 struct writeback_control wbc = {
1480 .bdi = page->mapping->backing_dev_info,
1481 .sync_mode = WB_SYNC_ALL,
1482 .nr_to_write = LONG_MAX,
1483 .range_start = range_start,
1484 .range_end = range_end,
1488 BUG_ON(!PageLocked(page));
1489 if (clear_page_dirty_for_io(page)) {
1490 ret = nfs_writepage_locked(page, &wbc);
1494 if (!PagePrivate(page))
1496 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1500 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1505 * Write back all requests on one page - we do this before reading it.
1507 int nfs_wb_page(struct inode *inode, struct page* page)
1509 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1512 int __init nfs_init_writepagecache(void)
1514 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1515 sizeof(struct nfs_write_data),
1516 0, SLAB_HWCACHE_ALIGN,
1518 if (nfs_wdata_cachep == NULL)
1521 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1523 if (nfs_wdata_mempool == NULL)
1526 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1528 if (nfs_commit_mempool == NULL)
1532 * NFS congestion size, scale with available memory.
1544 * This allows larger machines to have larger/more transfers.
1545 * Limit the default to 256M
1547 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1548 if (nfs_congestion_kb > 256*1024)
1549 nfs_congestion_kb = 256*1024;
1554 void nfs_destroy_writepagecache(void)
1556 mempool_destroy(nfs_commit_mempool);
1557 mempool_destroy(nfs_wdata_mempool);
1558 kmem_cache_destroy(nfs_wdata_cachep);