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 struct nfs_page *nfs_page_find_request_locked(struct page *page)
115 struct nfs_page *req = NULL;
117 if (PagePrivate(page)) {
118 req = (struct nfs_page *)page_private(page);
120 kref_get(&req->wb_kref);
125 static struct nfs_page *nfs_page_find_request(struct page *page)
127 struct nfs_page *req = NULL;
128 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
131 req = nfs_page_find_request_locked(page);
132 spin_unlock(req_lock);
136 /* Adjust the file length if we're writing beyond the end */
137 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
139 struct inode *inode = page->mapping->host;
140 loff_t end, i_size = i_size_read(inode);
141 pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
143 if (i_size > 0 && page->index < end_index)
145 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
148 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
149 i_size_write(inode, end);
152 /* A writeback failed: mark the page as bad, and invalidate the page cache */
153 static void nfs_set_pageerror(struct page *page)
156 nfs_zap_mapping(page->mapping->host, page->mapping);
159 /* We can set the PG_uptodate flag if we see that a write request
160 * covers the full page.
162 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
164 if (PageUptodate(page))
168 if (count != nfs_page_length(page))
170 if (count != PAGE_CACHE_SIZE)
171 zero_user_page(page, count, PAGE_CACHE_SIZE - count, KM_USER0);
172 SetPageUptodate(page);
175 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
176 unsigned int offset, unsigned int count)
178 struct nfs_page *req;
182 req = nfs_update_request(ctx, page, offset, count);
188 ret = nfs_wb_page(page->mapping->host, page);
192 /* Update file length */
193 nfs_grow_file(page, offset, count);
194 nfs_unlock_request(req);
198 static int wb_priority(struct writeback_control *wbc)
200 if (wbc->for_reclaim)
201 return FLUSH_HIGHPRI | FLUSH_STABLE;
202 if (wbc->for_kupdate)
208 * NFS congestion control
211 int nfs_congestion_kb;
213 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
214 #define NFS_CONGESTION_OFF_THRESH \
215 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
217 static int nfs_set_page_writeback(struct page *page)
219 int ret = test_set_page_writeback(page);
222 struct inode *inode = page->mapping->host;
223 struct nfs_server *nfss = NFS_SERVER(inode);
225 if (atomic_long_inc_return(&nfss->writeback) >
226 NFS_CONGESTION_ON_THRESH)
227 set_bdi_congested(&nfss->backing_dev_info, WRITE);
232 static void nfs_end_page_writeback(struct page *page)
234 struct inode *inode = page->mapping->host;
235 struct nfs_server *nfss = NFS_SERVER(inode);
237 end_page_writeback(page);
238 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
239 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
240 congestion_end(WRITE);
245 * Find an associated nfs write request, and prepare to flush it out
246 * Returns 1 if there was no write request, or if the request was
247 * already tagged by nfs_set_page_dirty.Returns 0 if the request
249 * May also 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 nfs_page *req;
255 struct nfs_inode *nfsi = NFS_I(page->mapping->host);
256 spinlock_t *req_lock = &nfsi->req_lock;
261 req = nfs_page_find_request_locked(page);
263 spin_unlock(req_lock);
266 if (nfs_lock_request_dontget(req))
268 /* Note: If we hold the page lock, as is the case in nfs_writepage,
269 * then the call to nfs_lock_request_dontget() will always
270 * succeed provided that someone hasn't already marked the
271 * request as dirty (in which case we don't care).
273 spin_unlock(req_lock);
274 ret = nfs_wait_on_request(req);
275 nfs_release_request(req);
280 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
281 /* This request is marked for commit */
282 spin_unlock(req_lock);
283 nfs_unlock_request(req);
284 nfs_pageio_complete(pgio);
287 if (nfs_set_page_writeback(page) != 0) {
288 spin_unlock(req_lock);
291 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
292 NFS_PAGE_TAG_LOCKED);
293 ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
294 spin_unlock(req_lock);
295 nfs_pageio_add_request(pgio, req);
300 * Write an mmapped page to the server.
302 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
304 struct nfs_pageio_descriptor mypgio, *pgio;
305 struct nfs_open_context *ctx;
306 struct inode *inode = page->mapping->host;
310 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
311 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
313 if (wbc->for_writepages)
314 pgio = wbc->fs_private;
316 nfs_pageio_init_write(&mypgio, inode, wb_priority(wbc));
320 nfs_pageio_cond_complete(pgio, page->index);
322 err = nfs_page_async_flush(pgio, page);
326 offset = nfs_page_length(page);
330 nfs_pageio_cond_complete(pgio, page->index);
332 ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
337 err = nfs_writepage_setup(ctx, page, 0, offset);
338 put_nfs_open_context(ctx);
341 err = nfs_page_async_flush(pgio, page);
345 if (!wbc->for_writepages)
346 nfs_pageio_complete(pgio);
350 int nfs_writepage(struct page *page, struct writeback_control *wbc)
354 err = nfs_writepage_locked(page, wbc);
359 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
361 struct inode *inode = mapping->host;
362 struct nfs_pageio_descriptor pgio;
365 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
367 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
368 wbc->fs_private = &pgio;
369 err = generic_writepages(mapping, wbc);
370 nfs_pageio_complete(&pgio);
374 return pgio.pg_error;
379 * Insert a write request into an inode
381 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
383 struct nfs_inode *nfsi = NFS_I(inode);
386 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
387 BUG_ON(error == -EEXIST);
392 nfs_begin_data_update(inode);
393 if (nfs_have_delegation(inode, FMODE_WRITE))
396 SetPagePrivate(req->wb_page);
397 set_page_private(req->wb_page, (unsigned long)req);
398 if (PageDirty(req->wb_page))
399 set_bit(PG_NEED_FLUSH, &req->wb_flags);
401 kref_get(&req->wb_kref);
406 * Remove a write request from an inode
408 static void nfs_inode_remove_request(struct nfs_page *req)
410 struct inode *inode = req->wb_context->path.dentry->d_inode;
411 struct nfs_inode *nfsi = NFS_I(inode);
413 BUG_ON (!NFS_WBACK_BUSY(req));
415 spin_lock(&nfsi->req_lock);
416 set_page_private(req->wb_page, 0);
417 ClearPagePrivate(req->wb_page);
418 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
419 if (test_and_clear_bit(PG_NEED_FLUSH, &req->wb_flags))
420 __set_page_dirty_nobuffers(req->wb_page);
423 spin_unlock(&nfsi->req_lock);
424 nfs_end_data_update(inode);
427 spin_unlock(&nfsi->req_lock);
428 nfs_clear_request(req);
429 nfs_release_request(req);
433 nfs_redirty_request(struct nfs_page *req)
435 __set_page_dirty_nobuffers(req->wb_page);
439 * Check if a request is dirty
442 nfs_dirty_request(struct nfs_page *req)
444 struct page *page = req->wb_page;
446 if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
448 return !PageWriteback(req->wb_page);
451 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
453 * Add a request to the inode's commit list.
456 nfs_mark_request_commit(struct nfs_page *req)
458 struct inode *inode = req->wb_context->path.dentry->d_inode;
459 struct nfs_inode *nfsi = NFS_I(inode);
461 spin_lock(&nfsi->req_lock);
462 nfs_list_add_request(req, &nfsi->commit);
464 set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
465 radix_tree_tag_set(&nfsi->nfs_page_tree,
467 NFS_PAGE_TAG_COMMIT);
468 spin_unlock(&nfsi->req_lock);
469 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
470 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
474 int nfs_write_need_commit(struct nfs_write_data *data)
476 return data->verf.committed != NFS_FILE_SYNC;
480 int nfs_reschedule_unstable_write(struct nfs_page *req)
482 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
483 nfs_mark_request_commit(req);
486 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
487 nfs_redirty_request(req);
494 nfs_mark_request_commit(struct nfs_page *req)
499 int nfs_write_need_commit(struct nfs_write_data *data)
505 int nfs_reschedule_unstable_write(struct nfs_page *req)
512 * Wait for a request to complete.
514 * Interruptible by signals only if mounted with intr flag.
516 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
518 struct nfs_inode *nfsi = NFS_I(inode);
519 struct nfs_page *req;
520 pgoff_t idx_end, next;
521 unsigned int res = 0;
527 idx_end = idx_start + npages - 1;
530 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
531 if (req->wb_index > idx_end)
534 next = req->wb_index + 1;
535 BUG_ON(!NFS_WBACK_BUSY(req));
537 kref_get(&req->wb_kref);
538 spin_unlock(&nfsi->req_lock);
539 error = nfs_wait_on_request(req);
540 nfs_release_request(req);
541 spin_lock(&nfsi->req_lock);
549 static void nfs_cancel_commit_list(struct list_head *head)
551 struct nfs_page *req;
553 while(!list_empty(head)) {
554 req = nfs_list_entry(head->next);
555 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
556 nfs_list_remove_request(req);
557 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
558 nfs_inode_remove_request(req);
559 nfs_unlock_request(req);
563 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
565 * nfs_scan_commit - Scan an inode for commit requests
566 * @inode: NFS inode to scan
567 * @dst: destination list
568 * @idx_start: lower bound of page->index to scan.
569 * @npages: idx_start + npages sets the upper bound to scan.
571 * Moves requests from the inode's 'commit' request list.
572 * The requests are *not* checked to ensure that they form a contiguous set.
575 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
577 struct nfs_inode *nfsi = NFS_I(inode);
580 if (nfsi->ncommit != 0) {
581 res = nfs_scan_list(nfsi, dst, idx_start, npages,
582 NFS_PAGE_TAG_COMMIT);
583 nfsi->ncommit -= res;
584 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
585 printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
590 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
597 * Try to update any existing write request, or create one if there is none.
598 * In order to match, the request's credentials must match those of
599 * the calling process.
601 * Note: Should always be called with the Page Lock held!
603 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
604 struct page *page, unsigned int offset, unsigned int bytes)
606 struct address_space *mapping = page->mapping;
607 struct inode *inode = mapping->host;
608 struct nfs_inode *nfsi = NFS_I(inode);
609 struct nfs_page *req, *new = NULL;
612 end = offset + bytes;
615 /* Loop over all inode entries and see if we find
616 * A request for the page we wish to update
618 spin_lock(&nfsi->req_lock);
619 req = nfs_page_find_request_locked(page);
621 if (!nfs_lock_request_dontget(req)) {
624 spin_unlock(&nfsi->req_lock);
625 error = nfs_wait_on_request(req);
626 nfs_release_request(req);
629 nfs_release_request(new);
630 return ERR_PTR(error);
634 spin_unlock(&nfsi->req_lock);
636 nfs_release_request(new);
642 nfs_lock_request_dontget(new);
643 error = nfs_inode_add_request(inode, new);
645 spin_unlock(&nfsi->req_lock);
646 nfs_unlock_request(new);
647 return ERR_PTR(error);
649 spin_unlock(&nfsi->req_lock);
652 spin_unlock(&nfsi->req_lock);
654 new = nfs_create_request(ctx, inode, page, offset, bytes);
659 /* We have a request for our page.
660 * If the creds don't match, or the
661 * page addresses don't match,
662 * tell the caller to wait on the conflicting
665 rqend = req->wb_offset + req->wb_bytes;
666 if (req->wb_context != ctx
667 || req->wb_page != page
668 || !nfs_dirty_request(req)
669 || offset > rqend || end < req->wb_offset) {
670 nfs_unlock_request(req);
671 return ERR_PTR(-EBUSY);
674 /* Okay, the request matches. Update the region */
675 if (offset < req->wb_offset) {
676 req->wb_offset = offset;
677 req->wb_pgbase = offset;
678 req->wb_bytes = rqend - req->wb_offset;
682 req->wb_bytes = end - req->wb_offset;
687 int nfs_flush_incompatible(struct file *file, struct page *page)
689 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
690 struct nfs_page *req;
691 int do_flush, status;
693 * Look for a request corresponding to this page. If there
694 * is one, and it belongs to another file, we flush it out
695 * before we try to copy anything into the page. Do this
696 * due to the lack of an ACCESS-type call in NFSv2.
697 * Also do the same if we find a request from an existing
701 req = nfs_page_find_request(page);
704 do_flush = req->wb_page != page || req->wb_context != ctx
705 || !nfs_dirty_request(req);
706 nfs_release_request(req);
709 status = nfs_wb_page(page->mapping->host, page);
710 } while (status == 0);
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 = (struct nfs_open_context *)file->private_data;
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 entirely in cache, it may be more efficient to avoid
736 * fragmenting write requests.
738 if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
739 count = max(count + offset, nfs_page_length(page));
743 status = nfs_writepage_setup(ctx, page, offset, count);
744 __set_page_dirty_nobuffers(page);
746 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
747 status, (long long)i_size_read(inode));
749 nfs_set_pageerror(page);
753 static void nfs_writepage_release(struct nfs_page *req)
756 if (PageError(req->wb_page)) {
757 nfs_end_page_writeback(req->wb_page);
758 nfs_inode_remove_request(req);
759 } else if (!nfs_reschedule_unstable_write(req)) {
760 /* Set the PG_uptodate flag */
761 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
762 nfs_end_page_writeback(req->wb_page);
763 nfs_inode_remove_request(req);
765 nfs_end_page_writeback(req->wb_page);
766 nfs_clear_page_tag_locked(req);
769 static inline int flush_task_priority(int how)
771 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
773 return RPC_PRIORITY_HIGH;
775 return RPC_PRIORITY_LOW;
777 return RPC_PRIORITY_NORMAL;
781 * Set up the argument/result storage required for the RPC call.
783 static void nfs_write_rpcsetup(struct nfs_page *req,
784 struct nfs_write_data *data,
785 const struct rpc_call_ops *call_ops,
786 unsigned int count, unsigned int offset,
792 /* Set up the RPC argument and reply structs
793 * NB: take care not to mess about with data->commit et al. */
796 data->inode = inode = req->wb_context->path.dentry->d_inode;
797 data->cred = req->wb_context->cred;
799 data->args.fh = NFS_FH(inode);
800 data->args.offset = req_offset(req) + offset;
801 data->args.pgbase = req->wb_pgbase + offset;
802 data->args.pages = data->pagevec;
803 data->args.count = count;
804 data->args.context = req->wb_context;
806 data->res.fattr = &data->fattr;
807 data->res.count = count;
808 data->res.verf = &data->verf;
809 nfs_fattr_init(&data->fattr);
811 /* Set up the initial task struct. */
812 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
813 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
814 NFS_PROTO(inode)->write_setup(data, how);
816 data->task.tk_priority = flush_task_priority(how);
817 data->task.tk_cookie = (unsigned long)inode;
819 dprintk("NFS: %5u initiated write call "
820 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
823 (long long)NFS_FILEID(inode),
825 (unsigned long long)data->args.offset);
828 static void nfs_execute_write(struct nfs_write_data *data)
830 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
833 rpc_clnt_sigmask(clnt, &oldset);
834 rpc_execute(&data->task);
835 rpc_clnt_sigunmask(clnt, &oldset);
839 * Generate multiple small requests to write out a single
840 * contiguous dirty area on one page.
842 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
844 struct nfs_page *req = nfs_list_entry(head->next);
845 struct page *page = req->wb_page;
846 struct nfs_write_data *data;
847 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
852 nfs_list_remove_request(req);
856 size_t len = min(nbytes, wsize);
858 data = nfs_writedata_alloc(1);
861 list_add(&data->pages, &list);
864 } while (nbytes != 0);
865 atomic_set(&req->wb_complete, requests);
867 ClearPageError(page);
871 data = list_entry(list.next, struct nfs_write_data, pages);
872 list_del_init(&data->pages);
874 data->pagevec[0] = page;
878 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
882 nfs_execute_write(data);
883 } while (nbytes != 0);
888 while (!list_empty(&list)) {
889 data = list_entry(list.next, struct nfs_write_data, pages);
890 list_del(&data->pages);
891 nfs_writedata_release(data);
893 nfs_redirty_request(req);
894 nfs_end_page_writeback(req->wb_page);
895 nfs_clear_page_tag_locked(req);
900 * Create an RPC task for the given write request and kick it.
901 * The page must have been locked by the caller.
903 * It may happen that the page we're passed is not marked dirty.
904 * This is the case if nfs_updatepage detects a conflicting request
905 * that has been written but not committed.
907 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
909 struct nfs_page *req;
911 struct nfs_write_data *data;
913 data = nfs_writedata_alloc(npages);
917 pages = data->pagevec;
918 while (!list_empty(head)) {
919 req = nfs_list_entry(head->next);
920 nfs_list_remove_request(req);
921 nfs_list_add_request(req, &data->pages);
922 ClearPageError(req->wb_page);
923 *pages++ = req->wb_page;
925 req = nfs_list_entry(data->pages.next);
927 /* Set up the argument struct */
928 nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
930 nfs_execute_write(data);
933 while (!list_empty(head)) {
934 req = nfs_list_entry(head->next);
935 nfs_list_remove_request(req);
936 nfs_redirty_request(req);
937 nfs_end_page_writeback(req->wb_page);
938 nfs_clear_page_tag_locked(req);
943 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
944 struct inode *inode, int ioflags)
946 int wsize = NFS_SERVER(inode)->wsize;
948 if (wsize < PAGE_CACHE_SIZE)
949 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
951 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
955 * Handle a write reply that flushed part of a page.
957 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
959 struct nfs_write_data *data = calldata;
960 struct nfs_page *req = data->req;
961 struct page *page = req->wb_page;
963 dprintk("NFS: write (%s/%Ld %d@%Ld)",
964 req->wb_context->path.dentry->d_inode->i_sb->s_id,
965 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
967 (long long)req_offset(req));
969 if (nfs_writeback_done(task, data) != 0)
972 if (task->tk_status < 0) {
973 nfs_set_pageerror(page);
974 req->wb_context->error = task->tk_status;
975 dprintk(", error = %d\n", task->tk_status);
979 if (nfs_write_need_commit(data)) {
980 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
983 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
984 /* Do nothing we need to resend the writes */
985 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
986 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
987 dprintk(" defer commit\n");
988 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
989 set_bit(PG_NEED_RESCHED, &req->wb_flags);
990 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
991 dprintk(" server reboot detected\n");
993 spin_unlock(req_lock);
998 if (atomic_dec_and_test(&req->wb_complete))
999 nfs_writepage_release(req);
1002 static const struct rpc_call_ops nfs_write_partial_ops = {
1003 .rpc_call_done = nfs_writeback_done_partial,
1004 .rpc_release = nfs_writedata_release,
1008 * Handle a write reply that flushes a whole page.
1010 * FIXME: There is an inherent race with invalidate_inode_pages and
1011 * writebacks since the page->count is kept > 1 for as long
1012 * as the page has a write request pending.
1014 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1016 struct nfs_write_data *data = calldata;
1017 struct nfs_page *req;
1020 if (nfs_writeback_done(task, data) != 0)
1023 /* Update attributes as result of writeback. */
1024 while (!list_empty(&data->pages)) {
1025 req = nfs_list_entry(data->pages.next);
1026 nfs_list_remove_request(req);
1027 page = req->wb_page;
1029 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1030 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1031 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1033 (long long)req_offset(req));
1035 if (task->tk_status < 0) {
1036 nfs_set_pageerror(page);
1037 req->wb_context->error = task->tk_status;
1038 dprintk(", error = %d\n", task->tk_status);
1039 goto remove_request;
1042 if (nfs_write_need_commit(data)) {
1043 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1044 nfs_mark_request_commit(req);
1045 nfs_end_page_writeback(page);
1046 dprintk(" marked for commit\n");
1049 /* Set the PG_uptodate flag? */
1050 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1053 nfs_end_page_writeback(page);
1054 nfs_inode_remove_request(req);
1056 nfs_clear_page_tag_locked(req);
1060 static const struct rpc_call_ops nfs_write_full_ops = {
1061 .rpc_call_done = nfs_writeback_done_full,
1062 .rpc_release = nfs_writedata_release,
1067 * This function is called when the WRITE call is complete.
1069 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1071 struct nfs_writeargs *argp = &data->args;
1072 struct nfs_writeres *resp = &data->res;
1075 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1076 task->tk_pid, task->tk_status);
1079 * ->write_done will attempt to use post-op attributes to detect
1080 * conflicting writes by other clients. A strict interpretation
1081 * of close-to-open would allow us to continue caching even if
1082 * another writer had changed the file, but some applications
1083 * depend on tighter cache coherency when writing.
1085 status = NFS_PROTO(data->inode)->write_done(task, data);
1088 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1090 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1091 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1092 /* We tried a write call, but the server did not
1093 * commit data to stable storage even though we
1095 * Note: There is a known bug in Tru64 < 5.0 in which
1096 * the server reports NFS_DATA_SYNC, but performs
1097 * NFS_FILE_SYNC. We therefore implement this checking
1098 * as a dprintk() in order to avoid filling syslog.
1100 static unsigned long complain;
1102 if (time_before(complain, jiffies)) {
1103 dprintk("NFS: faulty NFS server %s:"
1104 " (committed = %d) != (stable = %d)\n",
1105 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1106 resp->verf->committed, argp->stable);
1107 complain = jiffies + 300 * HZ;
1111 /* Is this a short write? */
1112 if (task->tk_status >= 0 && resp->count < argp->count) {
1113 static unsigned long complain;
1115 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1117 /* Has the server at least made some progress? */
1118 if (resp->count != 0) {
1119 /* Was this an NFSv2 write or an NFSv3 stable write? */
1120 if (resp->verf->committed != NFS_UNSTABLE) {
1121 /* Resend from where the server left off */
1122 argp->offset += resp->count;
1123 argp->pgbase += resp->count;
1124 argp->count -= resp->count;
1126 /* Resend as a stable write in order to avoid
1127 * headaches in the case of a server crash.
1129 argp->stable = NFS_FILE_SYNC;
1131 rpc_restart_call(task);
1134 if (time_before(complain, jiffies)) {
1136 "NFS: Server wrote zero bytes, expected %u.\n",
1138 complain = jiffies + 300 * HZ;
1140 /* Can't do anything about it except throw an error. */
1141 task->tk_status = -EIO;
1147 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1148 void nfs_commit_release(void *wdata)
1150 nfs_commit_free(wdata);
1154 * Set up the argument/result storage required for the RPC call.
1156 static void nfs_commit_rpcsetup(struct list_head *head,
1157 struct nfs_write_data *data,
1160 struct nfs_page *first;
1161 struct inode *inode;
1164 /* Set up the RPC argument and reply structs
1165 * NB: take care not to mess about with data->commit et al. */
1167 list_splice_init(head, &data->pages);
1168 first = nfs_list_entry(data->pages.next);
1169 inode = first->wb_context->path.dentry->d_inode;
1171 data->inode = inode;
1172 data->cred = first->wb_context->cred;
1174 data->args.fh = NFS_FH(data->inode);
1175 /* Note: we always request a commit of the entire inode */
1176 data->args.offset = 0;
1177 data->args.count = 0;
1178 data->res.count = 0;
1179 data->res.fattr = &data->fattr;
1180 data->res.verf = &data->verf;
1181 nfs_fattr_init(&data->fattr);
1183 /* Set up the initial task struct. */
1184 flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1185 rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1186 NFS_PROTO(inode)->commit_setup(data, how);
1188 data->task.tk_priority = flush_task_priority(how);
1189 data->task.tk_cookie = (unsigned long)inode;
1191 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1195 * Commit dirty pages
1198 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1200 struct nfs_write_data *data;
1201 struct nfs_page *req;
1203 data = nfs_commit_alloc();
1208 /* Set up the argument struct */
1209 nfs_commit_rpcsetup(head, data, how);
1211 nfs_execute_write(data);
1214 while (!list_empty(head)) {
1215 req = nfs_list_entry(head->next);
1216 nfs_list_remove_request(req);
1217 nfs_mark_request_commit(req);
1218 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1219 nfs_clear_page_tag_locked(req);
1225 * COMMIT call returned
1227 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1229 struct nfs_write_data *data = calldata;
1230 struct nfs_page *req;
1232 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1233 task->tk_pid, task->tk_status);
1235 /* Call the NFS version-specific code */
1236 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1239 while (!list_empty(&data->pages)) {
1240 req = nfs_list_entry(data->pages.next);
1241 nfs_list_remove_request(req);
1242 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1243 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1245 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1246 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1247 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1249 (long long)req_offset(req));
1250 if (task->tk_status < 0) {
1251 req->wb_context->error = task->tk_status;
1252 nfs_inode_remove_request(req);
1253 dprintk(", error = %d\n", task->tk_status);
1257 /* Okay, COMMIT succeeded, apparently. Check the verifier
1258 * returned by the server against all stored verfs. */
1259 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1260 /* We have a match */
1261 /* Set the PG_uptodate flag */
1262 nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1264 nfs_inode_remove_request(req);
1268 /* We have a mismatch. Write the page again */
1269 dprintk(" mismatch\n");
1270 nfs_redirty_request(req);
1272 nfs_clear_page_tag_locked(req);
1276 static const struct rpc_call_ops nfs_commit_ops = {
1277 .rpc_call_done = nfs_commit_done,
1278 .rpc_release = nfs_commit_release,
1281 int nfs_commit_inode(struct inode *inode, int how)
1283 struct nfs_inode *nfsi = NFS_I(inode);
1287 spin_lock(&nfsi->req_lock);
1288 res = nfs_scan_commit(inode, &head, 0, 0);
1289 spin_unlock(&nfsi->req_lock);
1291 int error = nfs_commit_list(inode, &head, how);
1298 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1304 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1306 struct inode *inode = mapping->host;
1307 struct nfs_inode *nfsi = NFS_I(inode);
1308 pgoff_t idx_start, idx_end;
1309 unsigned int npages = 0;
1311 int nocommit = how & FLUSH_NOCOMMIT;
1315 if (wbc->range_cyclic)
1318 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1319 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1320 if (idx_end > idx_start) {
1321 pgoff_t l_npages = 1 + idx_end - idx_start;
1323 if (sizeof(npages) != sizeof(l_npages) &&
1324 (pgoff_t)npages != l_npages)
1328 how &= ~FLUSH_NOCOMMIT;
1329 spin_lock(&nfsi->req_lock);
1331 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1336 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1339 if (how & FLUSH_INVALIDATE) {
1340 spin_unlock(&nfsi->req_lock);
1341 nfs_cancel_commit_list(&head);
1343 spin_lock(&nfsi->req_lock);
1346 pages += nfs_scan_commit(inode, &head, 0, 0);
1347 spin_unlock(&nfsi->req_lock);
1348 ret = nfs_commit_list(inode, &head, how);
1349 spin_lock(&nfsi->req_lock);
1351 spin_unlock(&nfsi->req_lock);
1356 * flush the inode to disk.
1358 int nfs_wb_all(struct inode *inode)
1360 struct address_space *mapping = inode->i_mapping;
1361 struct writeback_control wbc = {
1362 .bdi = mapping->backing_dev_info,
1363 .sync_mode = WB_SYNC_ALL,
1364 .nr_to_write = LONG_MAX,
1365 .for_writepages = 1,
1370 ret = nfs_writepages(mapping, &wbc);
1373 ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1377 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1381 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1383 struct writeback_control wbc = {
1384 .bdi = mapping->backing_dev_info,
1385 .sync_mode = WB_SYNC_ALL,
1386 .nr_to_write = LONG_MAX,
1387 .range_start = range_start,
1388 .range_end = range_end,
1389 .for_writepages = 1,
1393 ret = nfs_writepages(mapping, &wbc);
1396 ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1400 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1404 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1406 loff_t range_start = page_offset(page);
1407 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1408 struct writeback_control wbc = {
1409 .bdi = page->mapping->backing_dev_info,
1410 .sync_mode = WB_SYNC_ALL,
1411 .nr_to_write = LONG_MAX,
1412 .range_start = range_start,
1413 .range_end = range_end,
1417 BUG_ON(!PageLocked(page));
1418 if (clear_page_dirty_for_io(page)) {
1419 ret = nfs_writepage_locked(page, &wbc);
1423 if (!PagePrivate(page))
1425 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1429 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1434 * Write back all requests on one page - we do this before reading it.
1436 int nfs_wb_page(struct inode *inode, struct page* page)
1438 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1441 int nfs_set_page_dirty(struct page *page)
1443 struct address_space *mapping = page->mapping;
1444 struct inode *inode;
1445 spinlock_t *req_lock;
1446 struct nfs_page *req;
1451 inode = mapping->host;
1454 req_lock = &NFS_I(inode)->req_lock;
1455 spin_lock(req_lock);
1456 req = nfs_page_find_request_locked(page);
1458 /* Mark any existing write requests for flushing */
1459 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1460 spin_unlock(req_lock);
1461 nfs_release_request(req);
1464 ret = __set_page_dirty_nobuffers(page);
1465 spin_unlock(req_lock);
1468 return !TestSetPageDirty(page);
1472 int __init nfs_init_writepagecache(void)
1474 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1475 sizeof(struct nfs_write_data),
1476 0, SLAB_HWCACHE_ALIGN,
1478 if (nfs_wdata_cachep == NULL)
1481 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1483 if (nfs_wdata_mempool == NULL)
1486 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1488 if (nfs_commit_mempool == NULL)
1492 * NFS congestion size, scale with available memory.
1504 * This allows larger machines to have larger/more transfers.
1505 * Limit the default to 256M
1507 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1508 if (nfs_congestion_kb > 256*1024)
1509 nfs_congestion_kb = 256*1024;
1514 void nfs_destroy_writepagecache(void)
1516 mempool_destroy(nfs_commit_mempool);
1517 mempool_destroy(nfs_wdata_mempool);
1518 kmem_cache_destroy(nfs_wdata_cachep);