4 * Writing file data over NFS.
6 * We do it like this: When a (user) process wishes to write data to an
7 * NFS file, a write request is allocated that contains the RPC task data
8 * plus some info on the page to be written, and added to the inode's
9 * write chain. If the process writes past the end of the page, an async
10 * RPC call to write the page is scheduled immediately; otherwise, the call
11 * is delayed for a few seconds.
13 * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
15 * Write requests are kept on the inode's writeback list. Each entry in
16 * that list references the page (portion) to be written. When the
17 * cache timeout has expired, the RPC task is woken up, and tries to
18 * lock the page. As soon as it manages to do so, the request is moved
19 * from the writeback list to the writelock list.
21 * Note: we must make sure never to confuse the inode passed in the
22 * write_page request with the one in page->inode. As far as I understand
23 * it, these are different when doing a swap-out.
25 * To understand everything that goes on here and in the NFS read code,
26 * one should be aware that a page is locked in exactly one of the following
29 * - A write request is in progress.
30 * - A user process is in generic_file_write/nfs_update_page
31 * - A user process is in generic_file_read
33 * Also note that because of the way pages are invalidated in
34 * nfs_revalidate_inode, the following assertions hold:
36 * - If a page is dirty, there will be no read requests (a page will
37 * not be re-read unless invalidated by nfs_revalidate_inode).
38 * - If the page is not uptodate, there will be no pending write
39 * requests, and no process will be in nfs_update_page.
41 * FIXME: Interaction with the vmscan routines is not optimal yet.
42 * Either vmscan must be made nfs-savvy, or we need a different page
43 * reclaim concept that supports something like FS-independent
44 * buffer_heads with a b_ops-> field.
46 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
49 #include <linux/types.h>
50 #include <linux/slab.h>
52 #include <linux/pagemap.h>
53 #include <linux/file.h>
54 #include <linux/writeback.h>
56 #include <linux/sunrpc/clnt.h>
57 #include <linux/nfs_fs.h>
58 #include <linux/nfs_mount.h>
59 #include <linux/nfs_page.h>
60 #include <linux/backing-dev.h>
62 #include <asm/uaccess.h>
63 #include <linux/smp_lock.h>
65 #include "delegation.h"
69 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
71 #define MIN_POOL_WRITE (32)
72 #define MIN_POOL_COMMIT (4)
75 * Local function declarations
77 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
79 unsigned int, unsigned int);
80 static int nfs_wait_on_write_congestion(struct address_space *, int);
81 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
82 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
83 static int nfs_wb_page_priority(struct inode *inode, struct page *page, int how);
84 static const struct rpc_call_ops nfs_write_partial_ops;
85 static const struct rpc_call_ops nfs_write_full_ops;
86 static const struct rpc_call_ops nfs_commit_ops;
88 static kmem_cache_t *nfs_wdata_cachep;
89 static mempool_t *nfs_wdata_mempool;
90 static mempool_t *nfs_commit_mempool;
92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
94 struct nfs_write_data *nfs_commit_alloc(void)
96 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
99 memset(p, 0, sizeof(*p));
100 INIT_LIST_HEAD(&p->pages);
105 void nfs_commit_rcu_free(struct rcu_head *head)
107 struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
108 if (p && (p->pagevec != &p->page_array[0]))
110 mempool_free(p, nfs_commit_mempool);
113 void nfs_commit_free(struct nfs_write_data *wdata)
115 call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
118 struct nfs_write_data *nfs_writedata_alloc(size_t len)
120 unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
121 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
124 memset(p, 0, sizeof(*p));
125 INIT_LIST_HEAD(&p->pages);
126 p->npages = pagecount;
127 if (pagecount <= ARRAY_SIZE(p->page_array))
128 p->pagevec = p->page_array;
130 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
132 mempool_free(p, nfs_wdata_mempool);
140 static void nfs_writedata_rcu_free(struct rcu_head *head)
142 struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
143 if (p && (p->pagevec != &p->page_array[0]))
145 mempool_free(p, nfs_wdata_mempool);
148 static void nfs_writedata_free(struct nfs_write_data *wdata)
150 call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
153 void nfs_writedata_release(void *wdata)
155 nfs_writedata_free(wdata);
158 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
160 struct nfs_page *req = NULL;
162 if (PagePrivate(page)) {
163 req = (struct nfs_page *)page_private(page);
165 atomic_inc(&req->wb_count);
170 static struct nfs_page *nfs_page_find_request(struct page *page)
172 struct nfs_page *req = NULL;
173 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
176 req = nfs_page_find_request_locked(page);
177 spin_unlock(req_lock);
181 /* Adjust the file length if we're writing beyond the end */
182 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
184 struct inode *inode = page->mapping->host;
185 loff_t end, i_size = i_size_read(inode);
186 unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
188 if (i_size > 0 && page->index < end_index)
190 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
193 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
194 i_size_write(inode, end);
197 /* We can set the PG_uptodate flag if we see that a write request
198 * covers the full page.
200 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
202 if (PageUptodate(page))
206 if (count != nfs_page_length(page))
208 if (count != PAGE_CACHE_SIZE)
209 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
210 SetPageUptodate(page);
213 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
214 unsigned int offset, unsigned int count)
216 struct nfs_page *req;
220 req = nfs_update_request(ctx, page, offset, count);
226 ret = nfs_wb_page(page->mapping->host, page);
230 /* Update file length */
231 nfs_grow_file(page, offset, count);
232 /* Set the PG_uptodate flag? */
233 nfs_mark_uptodate(page, offset, count);
234 nfs_unlock_request(req);
238 static int wb_priority(struct writeback_control *wbc)
240 if (wbc->for_reclaim)
241 return FLUSH_HIGHPRI;
242 if (wbc->for_kupdate)
248 * Write an mmapped page to the server.
250 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
252 struct nfs_open_context *ctx;
253 struct inode *inode = page->mapping->host;
254 struct nfs_page *req;
258 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
259 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
261 req = nfs_page_find_request(page);
263 int flushme = test_bit(PG_NEED_FLUSH, &req->wb_flags);
264 nfs_release_request(req);
267 /* Ensure we've flushed out the invalid write */
268 nfs_wb_page_priority(inode, page, wb_priority(wbc) | FLUSH_STABLE | FLUSH_NOWRITEPAGE);
271 offset = nfs_page_length(page);
275 ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
280 err = nfs_writepage_setup(ctx, page, 0, offset);
281 put_nfs_open_context(ctx);
284 if (!wbc->for_writepages)
285 nfs_flush_mapping(page->mapping, wbc, wb_priority(wbc));
289 int nfs_writepage(struct page *page, struct writeback_control *wbc)
293 err = nfs_writepage_locked(page, wbc);
299 * Note: causes nfs_update_request() to block on the assumption
300 * that the writeback is generated due to memory pressure.
302 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
304 struct backing_dev_info *bdi = mapping->backing_dev_info;
305 struct inode *inode = mapping->host;
308 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
310 err = generic_writepages(mapping, wbc);
313 while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
314 if (wbc->nonblocking)
316 nfs_wait_on_write_congestion(mapping, 0);
318 err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
321 nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
322 if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
323 err = nfs_wait_on_requests(inode, 0, 0);
327 err = nfs_commit_inode(inode, wb_priority(wbc));
331 clear_bit(BDI_write_congested, &bdi->state);
332 wake_up_all(&nfs_write_congestion);
333 congestion_end(WRITE);
338 * Insert a write request into an inode
340 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
342 struct nfs_inode *nfsi = NFS_I(inode);
345 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
346 BUG_ON(error == -EEXIST);
351 nfs_begin_data_update(inode);
352 if (nfs_have_delegation(inode, FMODE_WRITE))
355 SetPagePrivate(req->wb_page);
356 set_page_private(req->wb_page, (unsigned long)req);
358 atomic_inc(&req->wb_count);
363 * Insert a write request into an inode
365 static void nfs_inode_remove_request(struct nfs_page *req)
367 struct inode *inode = req->wb_context->dentry->d_inode;
368 struct nfs_inode *nfsi = NFS_I(inode);
370 BUG_ON (!NFS_WBACK_BUSY(req));
372 spin_lock(&nfsi->req_lock);
373 set_page_private(req->wb_page, 0);
374 ClearPagePrivate(req->wb_page);
375 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
378 spin_unlock(&nfsi->req_lock);
379 nfs_end_data_update(inode);
382 spin_unlock(&nfsi->req_lock);
383 nfs_clear_request(req);
384 nfs_release_request(req);
388 * Add a request to the inode's dirty list.
391 nfs_mark_request_dirty(struct nfs_page *req)
393 struct inode *inode = req->wb_context->dentry->d_inode;
394 struct nfs_inode *nfsi = NFS_I(inode);
396 spin_lock(&nfsi->req_lock);
397 radix_tree_tag_set(&nfsi->nfs_page_tree,
398 req->wb_index, NFS_PAGE_TAG_DIRTY);
399 nfs_list_add_request(req, &nfsi->dirty);
401 spin_unlock(&nfsi->req_lock);
402 inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
403 mark_inode_dirty(inode);
407 * Check if a request is dirty
410 nfs_dirty_request(struct nfs_page *req)
412 struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
413 return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
416 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
418 * Add a request to the inode's commit list.
421 nfs_mark_request_commit(struct nfs_page *req)
423 struct inode *inode = req->wb_context->dentry->d_inode;
424 struct nfs_inode *nfsi = NFS_I(inode);
426 spin_lock(&nfsi->req_lock);
427 nfs_list_add_request(req, &nfsi->commit);
429 spin_unlock(&nfsi->req_lock);
430 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
431 mark_inode_dirty(inode);
436 * Wait for a request to complete.
438 * Interruptible by signals only if mounted with intr flag.
440 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
442 struct nfs_inode *nfsi = NFS_I(inode);
443 struct nfs_page *req;
444 unsigned long idx_end, next;
445 unsigned int res = 0;
451 idx_end = idx_start + npages - 1;
454 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
455 if (req->wb_index > idx_end)
458 next = req->wb_index + 1;
459 BUG_ON(!NFS_WBACK_BUSY(req));
461 atomic_inc(&req->wb_count);
462 spin_unlock(&nfsi->req_lock);
463 error = nfs_wait_on_request(req);
464 nfs_release_request(req);
465 spin_lock(&nfsi->req_lock);
473 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
475 struct nfs_inode *nfsi = NFS_I(inode);
478 spin_lock(&nfsi->req_lock);
479 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
480 spin_unlock(&nfsi->req_lock);
484 static void nfs_cancel_dirty_list(struct list_head *head)
486 struct nfs_page *req;
487 while(!list_empty(head)) {
488 req = nfs_list_entry(head->next);
489 nfs_list_remove_request(req);
490 nfs_inode_remove_request(req);
491 nfs_clear_page_writeback(req);
495 static void nfs_cancel_commit_list(struct list_head *head)
497 struct nfs_page *req;
499 while(!list_empty(head)) {
500 req = nfs_list_entry(head->next);
501 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
502 nfs_list_remove_request(req);
503 nfs_inode_remove_request(req);
504 nfs_unlock_request(req);
508 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
510 * nfs_scan_commit - Scan an inode for commit requests
511 * @inode: NFS inode to scan
512 * @dst: destination list
513 * @idx_start: lower bound of page->index to scan.
514 * @npages: idx_start + npages sets the upper bound to scan.
516 * Moves requests from the inode's 'commit' request list.
517 * The requests are *not* checked to ensure that they form a contiguous set.
520 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
522 struct nfs_inode *nfsi = NFS_I(inode);
525 if (nfsi->ncommit != 0) {
526 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
527 nfsi->ncommit -= res;
528 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
529 printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
534 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
540 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
542 struct backing_dev_info *bdi = mapping->backing_dev_info;
548 if (!bdi_write_congested(bdi))
551 nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
554 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
557 rpc_clnt_sigmask(clnt, &oldset);
558 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
559 if (bdi_write_congested(bdi)) {
565 rpc_clnt_sigunmask(clnt, &oldset);
567 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
568 if (bdi_write_congested(bdi))
571 finish_wait(&nfs_write_congestion, &wait);
577 * Try to update any existing write request, or create one if there is none.
578 * In order to match, the request's credentials must match those of
579 * the calling process.
581 * Note: Should always be called with the Page Lock held!
583 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
584 struct page *page, unsigned int offset, unsigned int bytes)
586 struct inode *inode = page->mapping->host;
587 struct nfs_inode *nfsi = NFS_I(inode);
588 struct nfs_page *req, *new = NULL;
589 unsigned long rqend, end;
591 end = offset + bytes;
593 if (nfs_wait_on_write_congestion(page->mapping, NFS_SERVER(inode)->flags & NFS_MOUNT_INTR))
594 return ERR_PTR(-ERESTARTSYS);
596 /* Loop over all inode entries and see if we find
597 * A request for the page we wish to update
599 spin_lock(&nfsi->req_lock);
600 req = nfs_page_find_request_locked(page);
602 if (!nfs_lock_request_dontget(req)) {
605 spin_unlock(&nfsi->req_lock);
606 error = nfs_wait_on_request(req);
607 nfs_release_request(req);
610 nfs_release_request(new);
611 return ERR_PTR(error);
615 spin_unlock(&nfsi->req_lock);
617 nfs_release_request(new);
623 nfs_lock_request_dontget(new);
624 error = nfs_inode_add_request(inode, new);
626 spin_unlock(&nfsi->req_lock);
627 nfs_unlock_request(new);
628 return ERR_PTR(error);
630 spin_unlock(&nfsi->req_lock);
631 nfs_mark_request_dirty(new);
634 spin_unlock(&nfsi->req_lock);
636 new = nfs_create_request(ctx, inode, page, offset, bytes);
641 /* We have a request for our page.
642 * If the creds don't match, or the
643 * page addresses don't match,
644 * tell the caller to wait on the conflicting
647 rqend = req->wb_offset + req->wb_bytes;
648 if (req->wb_context != ctx
649 || req->wb_page != page
650 || !nfs_dirty_request(req)
651 || offset > rqend || end < req->wb_offset) {
652 nfs_unlock_request(req);
653 return ERR_PTR(-EBUSY);
656 /* Okay, the request matches. Update the region */
657 if (offset < req->wb_offset) {
658 req->wb_offset = offset;
659 req->wb_pgbase = offset;
660 req->wb_bytes = rqend - req->wb_offset;
664 req->wb_bytes = end - req->wb_offset;
669 int nfs_flush_incompatible(struct file *file, struct page *page)
671 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
672 struct nfs_page *req;
673 int do_flush, status;
675 * Look for a request corresponding to this page. If there
676 * is one, and it belongs to another file, we flush it out
677 * before we try to copy anything into the page. Do this
678 * due to the lack of an ACCESS-type call in NFSv2.
679 * Also do the same if we find a request from an existing
683 req = nfs_page_find_request(page);
686 do_flush = req->wb_page != page || req->wb_context != ctx
687 || test_bit(PG_NEED_FLUSH, &req->wb_flags);
688 nfs_release_request(req);
691 status = nfs_wb_page(page->mapping->host, page);
692 } while (status == 0);
697 * Update and possibly write a cached page of an NFS file.
699 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
700 * things with a page scheduled for an RPC call (e.g. invalidate it).
702 int nfs_updatepage(struct file *file, struct page *page,
703 unsigned int offset, unsigned int count)
705 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
706 struct inode *inode = page->mapping->host;
709 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
711 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
712 file->f_dentry->d_parent->d_name.name,
713 file->f_dentry->d_name.name, count,
714 (long long)(page_offset(page) +offset));
716 /* If we're not using byte range locks, and we know the page
717 * is entirely in cache, it may be more efficient to avoid
718 * fragmenting write requests.
720 if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
721 count = max(count + offset, nfs_page_length(page));
725 status = nfs_writepage_setup(ctx, page, offset, count);
727 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
728 status, (long long)i_size_read(inode));
730 ClearPageUptodate(page);
734 static void nfs_writepage_release(struct nfs_page *req)
736 end_page_writeback(req->wb_page);
738 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
739 if (!PageError(req->wb_page)) {
740 if (NFS_NEED_RESCHED(req)) {
741 nfs_mark_request_dirty(req);
743 } else if (NFS_NEED_COMMIT(req)) {
744 nfs_mark_request_commit(req);
748 nfs_inode_remove_request(req);
751 nfs_clear_commit(req);
752 nfs_clear_reschedule(req);
754 nfs_inode_remove_request(req);
756 nfs_clear_page_writeback(req);
759 static inline int flush_task_priority(int how)
761 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
763 return RPC_PRIORITY_HIGH;
765 return RPC_PRIORITY_LOW;
767 return RPC_PRIORITY_NORMAL;
771 * Set up the argument/result storage required for the RPC call.
773 static void nfs_write_rpcsetup(struct nfs_page *req,
774 struct nfs_write_data *data,
775 const struct rpc_call_ops *call_ops,
776 unsigned int count, unsigned int offset,
782 /* Set up the RPC argument and reply structs
783 * NB: take care not to mess about with data->commit et al. */
786 data->inode = inode = req->wb_context->dentry->d_inode;
787 data->cred = req->wb_context->cred;
789 data->args.fh = NFS_FH(inode);
790 data->args.offset = req_offset(req) + offset;
791 data->args.pgbase = req->wb_pgbase + offset;
792 data->args.pages = data->pagevec;
793 data->args.count = count;
794 data->args.context = req->wb_context;
796 data->res.fattr = &data->fattr;
797 data->res.count = count;
798 data->res.verf = &data->verf;
799 nfs_fattr_init(&data->fattr);
801 /* Set up the initial task struct. */
802 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
803 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
804 NFS_PROTO(inode)->write_setup(data, how);
806 data->task.tk_priority = flush_task_priority(how);
807 data->task.tk_cookie = (unsigned long)inode;
809 dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
812 (long long)NFS_FILEID(inode),
814 (unsigned long long)data->args.offset);
817 static void nfs_execute_write(struct nfs_write_data *data)
819 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
822 rpc_clnt_sigmask(clnt, &oldset);
823 rpc_execute(&data->task);
824 rpc_clnt_sigunmask(clnt, &oldset);
828 * Generate multiple small requests to write out a single
829 * contiguous dirty area on one page.
831 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
833 struct nfs_page *req = nfs_list_entry(head->next);
834 struct page *page = req->wb_page;
835 struct nfs_write_data *data;
836 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
841 nfs_list_remove_request(req);
843 nbytes = req->wb_bytes;
845 size_t len = min(nbytes, wsize);
847 data = nfs_writedata_alloc(len);
850 list_add(&data->pages, &list);
853 } while (nbytes != 0);
854 atomic_set(&req->wb_complete, requests);
856 ClearPageError(page);
857 set_page_writeback(page);
859 nbytes = req->wb_bytes;
861 data = list_entry(list.next, struct nfs_write_data, pages);
862 list_del_init(&data->pages);
864 data->pagevec[0] = page;
866 if (nbytes > wsize) {
867 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
872 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
873 nbytes, offset, how);
876 nfs_execute_write(data);
877 } while (nbytes != 0);
882 while (!list_empty(&list)) {
883 data = list_entry(list.next, struct nfs_write_data, pages);
884 list_del(&data->pages);
885 nfs_writedata_release(data);
887 nfs_mark_request_dirty(req);
888 nfs_clear_page_writeback(req);
893 * Create an RPC task for the given write request and kick it.
894 * The page must have been locked by the caller.
896 * It may happen that the page we're passed is not marked dirty.
897 * This is the case if nfs_updatepage detects a conflicting request
898 * that has been written but not committed.
900 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
902 struct nfs_page *req;
904 struct nfs_write_data *data;
907 data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
911 pages = data->pagevec;
913 while (!list_empty(head)) {
914 req = nfs_list_entry(head->next);
915 nfs_list_remove_request(req);
916 nfs_list_add_request(req, &data->pages);
917 ClearPageError(req->wb_page);
918 set_page_writeback(req->wb_page);
919 *pages++ = req->wb_page;
920 count += req->wb_bytes;
922 req = nfs_list_entry(data->pages.next);
924 /* Set up the argument struct */
925 nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
927 nfs_execute_write(data);
930 while (!list_empty(head)) {
931 struct nfs_page *req = nfs_list_entry(head->next);
932 nfs_list_remove_request(req);
933 nfs_mark_request_dirty(req);
934 nfs_clear_page_writeback(req);
939 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
941 LIST_HEAD(one_request);
942 int (*flush_one)(struct inode *, struct list_head *, int);
943 struct nfs_page *req;
944 int wpages = NFS_SERVER(inode)->wpages;
945 int wsize = NFS_SERVER(inode)->wsize;
948 flush_one = nfs_flush_one;
949 if (wsize < PAGE_CACHE_SIZE)
950 flush_one = nfs_flush_multi;
951 /* For single writes, FLUSH_STABLE is more efficient */
952 if (npages <= wpages && npages == NFS_I(inode)->npages
953 && nfs_list_entry(head->next)->wb_bytes <= wsize)
957 nfs_coalesce_requests(head, &one_request, wpages);
958 req = nfs_list_entry(one_request.next);
959 error = flush_one(inode, &one_request, how);
962 } while (!list_empty(head));
965 while (!list_empty(head)) {
966 req = nfs_list_entry(head->next);
967 nfs_list_remove_request(req);
968 nfs_mark_request_dirty(req);
969 nfs_clear_page_writeback(req);
975 * Handle a write reply that flushed part of a page.
977 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
979 struct nfs_write_data *data = calldata;
980 struct nfs_page *req = data->req;
981 struct page *page = req->wb_page;
983 dprintk("NFS: write (%s/%Ld %d@%Ld)",
984 req->wb_context->dentry->d_inode->i_sb->s_id,
985 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
987 (long long)req_offset(req));
989 if (nfs_writeback_done(task, data) != 0)
992 if (task->tk_status < 0) {
993 ClearPageUptodate(page);
995 req->wb_context->error = task->tk_status;
996 dprintk(", error = %d\n", task->tk_status);
998 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
999 if (data->verf.committed < NFS_FILE_SYNC) {
1000 if (!NFS_NEED_COMMIT(req)) {
1001 nfs_defer_commit(req);
1002 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1003 dprintk(" defer commit\n");
1004 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1005 nfs_defer_reschedule(req);
1006 dprintk(" server reboot detected\n");
1013 if (atomic_dec_and_test(&req->wb_complete))
1014 nfs_writepage_release(req);
1017 static const struct rpc_call_ops nfs_write_partial_ops = {
1018 .rpc_call_done = nfs_writeback_done_partial,
1019 .rpc_release = nfs_writedata_release,
1023 * Handle a write reply that flushes a whole page.
1025 * FIXME: There is an inherent race with invalidate_inode_pages and
1026 * writebacks since the page->count is kept > 1 for as long
1027 * as the page has a write request pending.
1029 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1031 struct nfs_write_data *data = calldata;
1032 struct nfs_page *req;
1035 if (nfs_writeback_done(task, data) != 0)
1038 /* Update attributes as result of writeback. */
1039 while (!list_empty(&data->pages)) {
1040 req = nfs_list_entry(data->pages.next);
1041 nfs_list_remove_request(req);
1042 page = req->wb_page;
1044 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1045 req->wb_context->dentry->d_inode->i_sb->s_id,
1046 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1048 (long long)req_offset(req));
1050 if (task->tk_status < 0) {
1051 ClearPageUptodate(page);
1053 req->wb_context->error = task->tk_status;
1054 end_page_writeback(page);
1055 nfs_inode_remove_request(req);
1056 dprintk(", error = %d\n", task->tk_status);
1059 end_page_writeback(page);
1061 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1062 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1063 nfs_inode_remove_request(req);
1067 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1068 nfs_mark_request_commit(req);
1069 dprintk(" marked for commit\n");
1071 nfs_inode_remove_request(req);
1074 nfs_clear_page_writeback(req);
1078 static const struct rpc_call_ops nfs_write_full_ops = {
1079 .rpc_call_done = nfs_writeback_done_full,
1080 .rpc_release = nfs_writedata_release,
1085 * This function is called when the WRITE call is complete.
1087 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1089 struct nfs_writeargs *argp = &data->args;
1090 struct nfs_writeres *resp = &data->res;
1093 dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1094 task->tk_pid, task->tk_status);
1097 * ->write_done will attempt to use post-op attributes to detect
1098 * conflicting writes by other clients. A strict interpretation
1099 * of close-to-open would allow us to continue caching even if
1100 * another writer had changed the file, but some applications
1101 * depend on tighter cache coherency when writing.
1103 status = NFS_PROTO(data->inode)->write_done(task, data);
1106 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1108 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1109 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1110 /* We tried a write call, but the server did not
1111 * commit data to stable storage even though we
1113 * Note: There is a known bug in Tru64 < 5.0 in which
1114 * the server reports NFS_DATA_SYNC, but performs
1115 * NFS_FILE_SYNC. We therefore implement this checking
1116 * as a dprintk() in order to avoid filling syslog.
1118 static unsigned long complain;
1120 if (time_before(complain, jiffies)) {
1121 dprintk("NFS: faulty NFS server %s:"
1122 " (committed = %d) != (stable = %d)\n",
1123 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1124 resp->verf->committed, argp->stable);
1125 complain = jiffies + 300 * HZ;
1129 /* Is this a short write? */
1130 if (task->tk_status >= 0 && resp->count < argp->count) {
1131 static unsigned long complain;
1133 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1135 /* Has the server at least made some progress? */
1136 if (resp->count != 0) {
1137 /* Was this an NFSv2 write or an NFSv3 stable write? */
1138 if (resp->verf->committed != NFS_UNSTABLE) {
1139 /* Resend from where the server left off */
1140 argp->offset += resp->count;
1141 argp->pgbase += resp->count;
1142 argp->count -= resp->count;
1144 /* Resend as a stable write in order to avoid
1145 * headaches in the case of a server crash.
1147 argp->stable = NFS_FILE_SYNC;
1149 rpc_restart_call(task);
1152 if (time_before(complain, jiffies)) {
1154 "NFS: Server wrote zero bytes, expected %u.\n",
1156 complain = jiffies + 300 * HZ;
1158 /* Can't do anything about it except throw an error. */
1159 task->tk_status = -EIO;
1165 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1166 void nfs_commit_release(void *wdata)
1168 nfs_commit_free(wdata);
1172 * Set up the argument/result storage required for the RPC call.
1174 static void nfs_commit_rpcsetup(struct list_head *head,
1175 struct nfs_write_data *data,
1178 struct nfs_page *first;
1179 struct inode *inode;
1182 /* Set up the RPC argument and reply structs
1183 * NB: take care not to mess about with data->commit et al. */
1185 list_splice_init(head, &data->pages);
1186 first = nfs_list_entry(data->pages.next);
1187 inode = first->wb_context->dentry->d_inode;
1189 data->inode = inode;
1190 data->cred = first->wb_context->cred;
1192 data->args.fh = NFS_FH(data->inode);
1193 /* Note: we always request a commit of the entire inode */
1194 data->args.offset = 0;
1195 data->args.count = 0;
1196 data->res.count = 0;
1197 data->res.fattr = &data->fattr;
1198 data->res.verf = &data->verf;
1199 nfs_fattr_init(&data->fattr);
1201 /* Set up the initial task struct. */
1202 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1203 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1204 NFS_PROTO(inode)->commit_setup(data, how);
1206 data->task.tk_priority = flush_task_priority(how);
1207 data->task.tk_cookie = (unsigned long)inode;
1209 dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1213 * Commit dirty pages
1216 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1218 struct nfs_write_data *data;
1219 struct nfs_page *req;
1221 data = nfs_commit_alloc();
1226 /* Set up the argument struct */
1227 nfs_commit_rpcsetup(head, data, how);
1229 nfs_execute_write(data);
1232 while (!list_empty(head)) {
1233 req = nfs_list_entry(head->next);
1234 nfs_list_remove_request(req);
1235 nfs_mark_request_commit(req);
1236 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1237 nfs_clear_page_writeback(req);
1243 * COMMIT call returned
1245 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1247 struct nfs_write_data *data = calldata;
1248 struct nfs_page *req;
1250 dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1251 task->tk_pid, task->tk_status);
1253 /* Call the NFS version-specific code */
1254 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1257 while (!list_empty(&data->pages)) {
1258 req = nfs_list_entry(data->pages.next);
1259 nfs_list_remove_request(req);
1260 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1262 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1263 req->wb_context->dentry->d_inode->i_sb->s_id,
1264 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1266 (long long)req_offset(req));
1267 if (task->tk_status < 0) {
1268 req->wb_context->error = task->tk_status;
1269 nfs_inode_remove_request(req);
1270 dprintk(", error = %d\n", task->tk_status);
1274 /* Okay, COMMIT succeeded, apparently. Check the verifier
1275 * returned by the server against all stored verfs. */
1276 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1277 /* We have a match */
1278 nfs_inode_remove_request(req);
1282 /* We have a mismatch. Write the page again */
1283 dprintk(" mismatch\n");
1284 nfs_mark_request_dirty(req);
1286 nfs_clear_page_writeback(req);
1290 static const struct rpc_call_ops nfs_commit_ops = {
1291 .rpc_call_done = nfs_commit_done,
1292 .rpc_release = nfs_commit_release,
1295 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1301 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1303 struct nfs_inode *nfsi = NFS_I(mapping->host);
1307 spin_lock(&nfsi->req_lock);
1308 res = nfs_scan_dirty(mapping, wbc, &head);
1309 spin_unlock(&nfsi->req_lock);
1311 int error = nfs_flush_list(mapping->host, &head, res, how);
1318 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1319 int nfs_commit_inode(struct inode *inode, int how)
1321 struct nfs_inode *nfsi = NFS_I(inode);
1325 spin_lock(&nfsi->req_lock);
1326 res = nfs_scan_commit(inode, &head, 0, 0);
1327 spin_unlock(&nfsi->req_lock);
1329 int error = nfs_commit_list(inode, &head, how);
1337 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1339 struct inode *inode = mapping->host;
1340 struct nfs_inode *nfsi = NFS_I(inode);
1341 unsigned long idx_start, idx_end;
1342 unsigned int npages = 0;
1344 int nocommit = how & FLUSH_NOCOMMIT;
1348 if (wbc->range_cyclic)
1351 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1352 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1353 if (idx_end > idx_start) {
1354 unsigned long l_npages = 1 + idx_end - idx_start;
1356 if (sizeof(npages) != sizeof(l_npages) &&
1357 (unsigned long)npages != l_npages)
1361 how &= ~FLUSH_NOCOMMIT;
1362 spin_lock(&nfsi->req_lock);
1364 wbc->pages_skipped = 0;
1365 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1368 pages = nfs_scan_dirty(mapping, wbc, &head);
1370 spin_unlock(&nfsi->req_lock);
1371 if (how & FLUSH_INVALIDATE) {
1372 nfs_cancel_dirty_list(&head);
1375 ret = nfs_flush_list(inode, &head, pages, how);
1376 spin_lock(&nfsi->req_lock);
1379 if (wbc->pages_skipped != 0)
1383 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1385 if (wbc->pages_skipped != 0)
1389 if (how & FLUSH_INVALIDATE) {
1390 spin_unlock(&nfsi->req_lock);
1391 nfs_cancel_commit_list(&head);
1393 spin_lock(&nfsi->req_lock);
1396 pages += nfs_scan_commit(inode, &head, 0, 0);
1397 spin_unlock(&nfsi->req_lock);
1398 ret = nfs_commit_list(inode, &head, how);
1399 spin_lock(&nfsi->req_lock);
1401 spin_unlock(&nfsi->req_lock);
1406 * flush the inode to disk.
1408 int nfs_wb_all(struct inode *inode)
1410 struct address_space *mapping = inode->i_mapping;
1411 struct writeback_control wbc = {
1412 .bdi = mapping->backing_dev_info,
1413 .sync_mode = WB_SYNC_ALL,
1414 .nr_to_write = LONG_MAX,
1419 ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1425 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1427 struct writeback_control wbc = {
1428 .bdi = mapping->backing_dev_info,
1429 .sync_mode = WB_SYNC_ALL,
1430 .nr_to_write = LONG_MAX,
1431 .range_start = range_start,
1432 .range_end = range_end,
1436 ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1442 static int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1444 loff_t range_start = page_offset(page);
1445 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1446 struct writeback_control wbc = {
1447 .bdi = page->mapping->backing_dev_info,
1448 .sync_mode = WB_SYNC_ALL,
1449 .nr_to_write = LONG_MAX,
1450 .range_start = range_start,
1451 .range_end = range_end,
1455 BUG_ON(!PageLocked(page));
1456 if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1457 ret = nfs_writepage_locked(page, &wbc);
1461 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1469 * Write back all requests on one page - we do this before reading it.
1471 int nfs_wb_page(struct inode *inode, struct page* page)
1473 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1476 int nfs_set_page_dirty(struct page *page)
1478 struct nfs_page *req;
1480 req = nfs_page_find_request(page);
1482 /* Mark any existing write requests for flushing */
1483 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1484 nfs_release_request(req);
1486 return __set_page_dirty_nobuffers(page);
1490 int __init nfs_init_writepagecache(void)
1492 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1493 sizeof(struct nfs_write_data),
1494 0, SLAB_HWCACHE_ALIGN,
1496 if (nfs_wdata_cachep == NULL)
1499 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1501 if (nfs_wdata_mempool == NULL)
1504 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1506 if (nfs_commit_mempool == NULL)
1512 void nfs_destroy_writepagecache(void)
1514 mempool_destroy(nfs_commit_mempool);
1515 mempool_destroy(nfs_wdata_mempool);
1516 kmem_cache_destroy(nfs_wdata_cachep);