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 void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
38 struct inode *inode, int ioflags);
39 static void nfs_redirty_request(struct nfs_page *req);
40 static const struct rpc_call_ops nfs_write_partial_ops;
41 static const struct rpc_call_ops nfs_write_full_ops;
42 static const struct rpc_call_ops nfs_commit_ops;
44 static struct kmem_cache *nfs_wdata_cachep;
45 static mempool_t *nfs_wdata_mempool;
46 static mempool_t *nfs_commit_mempool;
48 struct nfs_write_data *nfs_commitdata_alloc(void)
50 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
53 memset(p, 0, sizeof(*p));
54 INIT_LIST_HEAD(&p->pages);
59 void nfs_commit_free(struct nfs_write_data *p)
61 if (p && (p->pagevec != &p->page_array[0]))
63 mempool_free(p, nfs_commit_mempool);
66 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
68 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
71 memset(p, 0, sizeof(*p));
72 INIT_LIST_HEAD(&p->pages);
73 p->npages = pagecount;
74 if (pagecount <= ARRAY_SIZE(p->page_array))
75 p->pagevec = p->page_array;
77 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
79 mempool_free(p, nfs_wdata_mempool);
87 static void nfs_writedata_free(struct nfs_write_data *p)
89 if (p && (p->pagevec != &p->page_array[0]))
91 mempool_free(p, nfs_wdata_mempool);
94 void nfs_writedata_release(void *data)
96 struct nfs_write_data *wdata = data;
98 put_nfs_open_context(wdata->args.context);
99 nfs_writedata_free(wdata);
102 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
106 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
109 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
111 struct nfs_page *req = NULL;
113 if (PagePrivate(page)) {
114 req = (struct nfs_page *)page_private(page);
116 kref_get(&req->wb_kref);
121 static struct nfs_page *nfs_page_find_request(struct page *page)
123 struct inode *inode = page->mapping->host;
124 struct nfs_page *req = NULL;
126 spin_lock(&inode->i_lock);
127 req = nfs_page_find_request_locked(page);
128 spin_unlock(&inode->i_lock);
132 /* Adjust the file length if we're writing beyond the end */
133 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
135 struct inode *inode = page->mapping->host;
136 loff_t end, i_size = i_size_read(inode);
137 pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
139 if (i_size > 0 && page->index < end_index)
141 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
144 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
145 i_size_write(inode, end);
148 /* A writeback failed: mark the page as bad, and invalidate the page cache */
149 static void nfs_set_pageerror(struct page *page)
152 nfs_zap_mapping(page->mapping->host, page->mapping);
155 /* We can set the PG_uptodate flag if we see that a write request
156 * covers the full page.
158 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
160 if (PageUptodate(page))
164 if (count != nfs_page_length(page))
166 SetPageUptodate(page);
169 static int wb_priority(struct writeback_control *wbc)
171 if (wbc->for_reclaim)
172 return FLUSH_HIGHPRI | FLUSH_STABLE;
173 if (wbc->for_kupdate)
179 * NFS congestion control
182 int nfs_congestion_kb;
184 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
185 #define NFS_CONGESTION_OFF_THRESH \
186 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
188 static int nfs_set_page_writeback(struct page *page)
190 int ret = test_set_page_writeback(page);
193 struct inode *inode = page->mapping->host;
194 struct nfs_server *nfss = NFS_SERVER(inode);
196 if (atomic_long_inc_return(&nfss->writeback) >
197 NFS_CONGESTION_ON_THRESH)
198 set_bdi_congested(&nfss->backing_dev_info, WRITE);
203 static void nfs_end_page_writeback(struct page *page)
205 struct inode *inode = page->mapping->host;
206 struct nfs_server *nfss = NFS_SERVER(inode);
208 end_page_writeback(page);
209 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
210 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
214 * Find an associated nfs write request, and prepare to flush it out
215 * May return an error if the user signalled nfs_wait_on_request().
217 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
220 struct inode *inode = page->mapping->host;
221 struct nfs_page *req;
224 spin_lock(&inode->i_lock);
226 req = nfs_page_find_request_locked(page);
228 spin_unlock(&inode->i_lock);
231 if (nfs_set_page_tag_locked(req))
233 /* Note: If we hold the page lock, as is the case in nfs_writepage,
234 * then the call to nfs_set_page_tag_locked() will always
235 * succeed provided that someone hasn't already marked the
236 * request as dirty (in which case we don't care).
238 spin_unlock(&inode->i_lock);
239 ret = nfs_wait_on_request(req);
240 nfs_release_request(req);
243 spin_lock(&inode->i_lock);
245 if (test_bit(PG_CLEAN, &req->wb_flags)) {
246 spin_unlock(&inode->i_lock);
249 if (nfs_set_page_writeback(page) != 0) {
250 spin_unlock(&inode->i_lock);
253 spin_unlock(&inode->i_lock);
254 if (!nfs_pageio_add_request(pgio, req)) {
255 nfs_redirty_request(req);
256 return pgio->pg_error;
261 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
263 struct inode *inode = page->mapping->host;
265 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
266 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
268 nfs_pageio_cond_complete(pgio, page->index);
269 return nfs_page_async_flush(pgio, page);
273 * Write an mmapped page to the server.
275 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
277 struct nfs_pageio_descriptor pgio;
280 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
281 err = nfs_do_writepage(page, wbc, &pgio);
282 nfs_pageio_complete(&pgio);
285 if (pgio.pg_error < 0)
286 return pgio.pg_error;
290 int nfs_writepage(struct page *page, struct writeback_control *wbc)
294 ret = nfs_writepage_locked(page, wbc);
299 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
303 ret = nfs_do_writepage(page, wbc, data);
308 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
310 struct inode *inode = mapping->host;
311 struct nfs_pageio_descriptor pgio;
314 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
316 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
317 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
318 nfs_pageio_complete(&pgio);
321 if (pgio.pg_error < 0)
322 return pgio.pg_error;
327 * Insert a write request into an inode
329 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
331 struct nfs_inode *nfsi = NFS_I(inode);
334 error = radix_tree_preload(GFP_NOFS);
338 /* Lock the request! */
339 nfs_lock_request_dontget(req);
341 spin_lock(&inode->i_lock);
342 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
346 if (nfs_have_delegation(inode, FMODE_WRITE))
349 SetPagePrivate(req->wb_page);
350 set_page_private(req->wb_page, (unsigned long)req);
352 kref_get(&req->wb_kref);
353 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
354 NFS_PAGE_TAG_LOCKED);
355 spin_unlock(&inode->i_lock);
356 radix_tree_preload_end();
362 * Remove a write request from an inode
364 static void nfs_inode_remove_request(struct nfs_page *req)
366 struct inode *inode = req->wb_context->path.dentry->d_inode;
367 struct nfs_inode *nfsi = NFS_I(inode);
369 BUG_ON (!NFS_WBACK_BUSY(req));
371 spin_lock(&inode->i_lock);
372 set_page_private(req->wb_page, 0);
373 ClearPagePrivate(req->wb_page);
374 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
377 spin_unlock(&inode->i_lock);
380 spin_unlock(&inode->i_lock);
381 nfs_clear_request(req);
382 nfs_release_request(req);
386 nfs_mark_request_dirty(struct nfs_page *req)
388 __set_page_dirty_nobuffers(req->wb_page);
391 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
393 * Add a request to the inode's commit list.
396 nfs_mark_request_commit(struct nfs_page *req)
398 struct inode *inode = req->wb_context->path.dentry->d_inode;
399 struct nfs_inode *nfsi = NFS_I(inode);
401 spin_lock(&inode->i_lock);
403 set_bit(PG_CLEAN, &(req)->wb_flags);
404 radix_tree_tag_set(&nfsi->nfs_page_tree,
406 NFS_PAGE_TAG_COMMIT);
407 spin_unlock(&inode->i_lock);
408 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
409 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
410 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
414 nfs_clear_request_commit(struct nfs_page *req)
416 struct page *page = req->wb_page;
418 if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
419 dec_zone_page_state(page, NR_UNSTABLE_NFS);
420 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
427 int nfs_write_need_commit(struct nfs_write_data *data)
429 return data->verf.committed != NFS_FILE_SYNC;
433 int nfs_reschedule_unstable_write(struct nfs_page *req)
435 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
436 nfs_mark_request_commit(req);
439 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
440 nfs_mark_request_dirty(req);
447 nfs_mark_request_commit(struct nfs_page *req)
452 nfs_clear_request_commit(struct nfs_page *req)
458 int nfs_write_need_commit(struct nfs_write_data *data)
464 int nfs_reschedule_unstable_write(struct nfs_page *req)
471 * Wait for a request to complete.
473 * Interruptible by fatal signals only.
475 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
477 struct nfs_inode *nfsi = NFS_I(inode);
478 struct nfs_page *req;
479 pgoff_t idx_end, next;
480 unsigned int res = 0;
486 idx_end = idx_start + npages - 1;
489 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
490 if (req->wb_index > idx_end)
493 next = req->wb_index + 1;
494 BUG_ON(!NFS_WBACK_BUSY(req));
496 kref_get(&req->wb_kref);
497 spin_unlock(&inode->i_lock);
498 error = nfs_wait_on_request(req);
499 nfs_release_request(req);
500 spin_lock(&inode->i_lock);
508 static void nfs_cancel_commit_list(struct list_head *head)
510 struct nfs_page *req;
512 while(!list_empty(head)) {
513 req = nfs_list_entry(head->next);
514 nfs_list_remove_request(req);
515 nfs_clear_request_commit(req);
516 nfs_inode_remove_request(req);
517 nfs_unlock_request(req);
521 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
523 * nfs_scan_commit - Scan an inode for commit requests
524 * @inode: NFS inode to scan
525 * @dst: destination list
526 * @idx_start: lower bound of page->index to scan.
527 * @npages: idx_start + npages sets the upper bound to scan.
529 * Moves requests from the inode's 'commit' request list.
530 * The requests are *not* checked to ensure that they form a contiguous set.
533 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
535 struct nfs_inode *nfsi = NFS_I(inode);
538 if (nfsi->ncommit != 0) {
539 res = nfs_scan_list(nfsi, dst, idx_start, npages,
540 NFS_PAGE_TAG_COMMIT);
541 nfsi->ncommit -= res;
546 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
553 * Search for an existing write request, and attempt to update
554 * it to reflect a new dirty region on a given page.
556 * If the attempt fails, then the existing request is flushed out
559 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
564 struct nfs_page *req;
569 if (!PagePrivate(page))
572 end = offset + bytes;
573 spin_lock(&inode->i_lock);
576 req = nfs_page_find_request_locked(page);
580 rqend = req->wb_offset + req->wb_bytes;
582 * Tell the caller to flush out the request if
583 * the offsets are non-contiguous.
584 * Note: nfs_flush_incompatible() will already
585 * have flushed out requests having wrong owners.
588 || end < req->wb_offset)
591 if (nfs_set_page_tag_locked(req))
594 /* The request is locked, so wait and then retry */
595 spin_unlock(&inode->i_lock);
596 error = nfs_wait_on_request(req);
597 nfs_release_request(req);
600 spin_lock(&inode->i_lock);
603 if (nfs_clear_request_commit(req))
604 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
605 req->wb_index, NFS_PAGE_TAG_COMMIT);
607 /* Okay, the request matches. Update the region */
608 if (offset < req->wb_offset) {
609 req->wb_offset = offset;
610 req->wb_pgbase = offset;
613 req->wb_bytes = end - req->wb_offset;
615 req->wb_bytes = rqend - req->wb_offset;
617 spin_unlock(&inode->i_lock);
620 spin_unlock(&inode->i_lock);
621 nfs_release_request(req);
622 error = nfs_wb_page(inode, page);
624 return ERR_PTR(error);
628 * Try to update an existing write request, or create one if there is none.
630 * Note: Should always be called with the Page Lock held to prevent races
631 * if we have to add a new request. Also assumes that the caller has
632 * already called nfs_flush_incompatible() if necessary.
634 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
635 struct page *page, unsigned int offset, unsigned int bytes)
637 struct inode *inode = page->mapping->host;
638 struct nfs_page *req;
641 req = nfs_try_to_update_request(inode, page, offset, bytes);
644 req = nfs_create_request(ctx, inode, page, offset, bytes);
647 error = nfs_inode_add_request(inode, req);
649 nfs_release_request(req);
650 req = ERR_PTR(error);
656 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
657 unsigned int offset, unsigned int count)
659 struct nfs_page *req;
661 req = nfs_setup_write_request(ctx, page, offset, count);
664 /* Update file length */
665 nfs_grow_file(page, offset, count);
666 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
667 nfs_clear_page_tag_locked(req);
671 int nfs_flush_incompatible(struct file *file, struct page *page)
673 struct nfs_open_context *ctx = nfs_file_open_context(file);
674 struct nfs_page *req;
675 int do_flush, status;
677 * Look for a request corresponding to this page. If there
678 * is one, and it belongs to another file, we flush it out
679 * before we try to copy anything into the page. Do this
680 * due to the lack of an ACCESS-type call in NFSv2.
681 * Also do the same if we find a request from an existing
685 req = nfs_page_find_request(page);
688 do_flush = req->wb_page != page || req->wb_context != ctx;
689 nfs_release_request(req);
692 status = nfs_wb_page(page->mapping->host, page);
693 } while (status == 0);
698 * If the page cache is marked as unsafe or invalid, then we can't rely on
699 * the PageUptodate() flag. In this case, we will need to turn off
700 * write optimisations that depend on the page contents being correct.
702 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
704 return PageUptodate(page) &&
705 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
709 * Update and possibly write a cached page of an NFS file.
711 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
712 * things with a page scheduled for an RPC call (e.g. invalidate it).
714 int nfs_updatepage(struct file *file, struct page *page,
715 unsigned int offset, unsigned int count)
717 struct nfs_open_context *ctx = nfs_file_open_context(file);
718 struct inode *inode = page->mapping->host;
721 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
723 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
724 file->f_path.dentry->d_parent->d_name.name,
725 file->f_path.dentry->d_name.name, count,
726 (long long)(page_offset(page) + offset));
728 /* If we're not using byte range locks, and we know the page
729 * is up to date, it may be more efficient to extend the write
730 * to cover the entire page in order to avoid fragmentation
733 if (nfs_write_pageuptodate(page, inode) &&
734 inode->i_flock == NULL &&
735 !(file->f_flags & O_SYNC)) {
736 count = max(count + offset, nfs_page_length(page));
740 status = nfs_writepage_setup(ctx, page, offset, count);
742 nfs_set_pageerror(page);
744 __set_page_dirty_nobuffers(page);
746 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
747 status, (long long)i_size_read(inode));
751 static void nfs_writepage_release(struct nfs_page *req)
754 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
755 nfs_end_page_writeback(req->wb_page);
756 nfs_inode_remove_request(req);
758 nfs_end_page_writeback(req->wb_page);
759 nfs_clear_page_tag_locked(req);
762 static int flush_task_priority(int how)
764 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
766 return RPC_PRIORITY_HIGH;
768 return RPC_PRIORITY_LOW;
770 return RPC_PRIORITY_NORMAL;
774 * Set up the argument/result storage required for the RPC call.
776 static int nfs_write_rpcsetup(struct nfs_page *req,
777 struct nfs_write_data *data,
778 const struct rpc_call_ops *call_ops,
779 unsigned int count, unsigned int offset,
782 struct inode *inode = req->wb_context->path.dentry->d_inode;
783 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
784 int priority = flush_task_priority(how);
785 struct rpc_task *task;
786 struct rpc_message msg = {
787 .rpc_argp = &data->args,
788 .rpc_resp = &data->res,
789 .rpc_cred = req->wb_context->cred,
791 struct rpc_task_setup task_setup_data = {
792 .rpc_client = NFS_CLIENT(inode),
795 .callback_ops = call_ops,
796 .callback_data = data,
797 .workqueue = nfsiod_workqueue,
799 .priority = priority,
802 /* Set up the RPC argument and reply structs
803 * NB: take care not to mess about with data->commit et al. */
806 data->inode = inode = req->wb_context->path.dentry->d_inode;
807 data->cred = msg.rpc_cred;
809 data->args.fh = NFS_FH(inode);
810 data->args.offset = req_offset(req) + offset;
811 data->args.pgbase = req->wb_pgbase + offset;
812 data->args.pages = data->pagevec;
813 data->args.count = count;
814 data->args.context = get_nfs_open_context(req->wb_context);
815 data->args.stable = NFS_UNSTABLE;
816 if (how & FLUSH_STABLE) {
817 data->args.stable = NFS_DATA_SYNC;
818 if (!NFS_I(inode)->ncommit)
819 data->args.stable = NFS_FILE_SYNC;
822 data->res.fattr = &data->fattr;
823 data->res.count = count;
824 data->res.verf = &data->verf;
825 nfs_fattr_init(&data->fattr);
827 /* Set up the initial task struct. */
828 NFS_PROTO(inode)->write_setup(data, &msg);
830 dprintk("NFS: %5u initiated write call "
831 "(req %s/%lld, %u bytes @ offset %llu)\n",
834 (long long)NFS_FILEID(inode),
836 (unsigned long long)data->args.offset);
838 task = rpc_run_task(&task_setup_data);
840 return PTR_ERR(task);
845 /* If a nfs_flush_* function fails, it should remove reqs from @head and
846 * call this on each, which will prepare them to be retried on next
847 * writeback using standard nfs.
849 static void nfs_redirty_request(struct nfs_page *req)
851 nfs_mark_request_dirty(req);
852 nfs_end_page_writeback(req->wb_page);
853 nfs_clear_page_tag_locked(req);
857 * Generate multiple small requests to write out a single
858 * contiguous dirty area on one page.
860 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
862 struct nfs_page *req = nfs_list_entry(head->next);
863 struct page *page = req->wb_page;
864 struct nfs_write_data *data;
865 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
871 nfs_list_remove_request(req);
875 size_t len = min(nbytes, wsize);
877 data = nfs_writedata_alloc(1);
880 list_add(&data->pages, &list);
883 } while (nbytes != 0);
884 atomic_set(&req->wb_complete, requests);
886 ClearPageError(page);
892 data = list_entry(list.next, struct nfs_write_data, pages);
893 list_del_init(&data->pages);
895 data->pagevec[0] = page;
899 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
905 } while (nbytes != 0);
910 while (!list_empty(&list)) {
911 data = list_entry(list.next, struct nfs_write_data, pages);
912 list_del(&data->pages);
913 nfs_writedata_release(data);
915 nfs_redirty_request(req);
920 * Create an RPC task for the given write request and kick it.
921 * The page must have been locked by the caller.
923 * It may happen that the page we're passed is not marked dirty.
924 * This is the case if nfs_updatepage detects a conflicting request
925 * that has been written but not committed.
927 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
929 struct nfs_page *req;
931 struct nfs_write_data *data;
933 data = nfs_writedata_alloc(npages);
937 pages = data->pagevec;
938 while (!list_empty(head)) {
939 req = nfs_list_entry(head->next);
940 nfs_list_remove_request(req);
941 nfs_list_add_request(req, &data->pages);
942 ClearPageError(req->wb_page);
943 *pages++ = req->wb_page;
945 req = nfs_list_entry(data->pages.next);
947 /* Set up the argument struct */
948 return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
950 while (!list_empty(head)) {
951 req = nfs_list_entry(head->next);
952 nfs_list_remove_request(req);
953 nfs_redirty_request(req);
958 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
959 struct inode *inode, int ioflags)
961 size_t wsize = NFS_SERVER(inode)->wsize;
963 if (wsize < PAGE_CACHE_SIZE)
964 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
966 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
970 * Handle a write reply that flushed part of a page.
972 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
974 struct nfs_write_data *data = calldata;
976 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
978 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
980 NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
981 data->req->wb_bytes, (long long)req_offset(data->req));
983 nfs_writeback_done(task, data);
986 static void nfs_writeback_release_partial(void *calldata)
988 struct nfs_write_data *data = calldata;
989 struct nfs_page *req = data->req;
990 struct page *page = req->wb_page;
991 int status = data->task.tk_status;
994 nfs_set_pageerror(page);
995 nfs_context_set_write_error(req->wb_context, status);
996 dprintk(", error = %d\n", status);
1000 if (nfs_write_need_commit(data)) {
1001 struct inode *inode = page->mapping->host;
1003 spin_lock(&inode->i_lock);
1004 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1005 /* Do nothing we need to resend the writes */
1006 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1007 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1008 dprintk(" defer commit\n");
1009 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1010 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1011 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1012 dprintk(" server reboot detected\n");
1014 spin_unlock(&inode->i_lock);
1019 if (atomic_dec_and_test(&req->wb_complete))
1020 nfs_writepage_release(req);
1021 nfs_writedata_release(calldata);
1024 static const struct rpc_call_ops nfs_write_partial_ops = {
1025 .rpc_call_done = nfs_writeback_done_partial,
1026 .rpc_release = nfs_writeback_release_partial,
1030 * Handle a write reply that flushes a whole page.
1032 * FIXME: There is an inherent race with invalidate_inode_pages and
1033 * writebacks since the page->count is kept > 1 for as long
1034 * as the page has a write request pending.
1036 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1038 struct nfs_write_data *data = calldata;
1040 nfs_writeback_done(task, data);
1043 static void nfs_writeback_release_full(void *calldata)
1045 struct nfs_write_data *data = calldata;
1046 int status = data->task.tk_status;
1048 /* Update attributes as result of writeback. */
1049 while (!list_empty(&data->pages)) {
1050 struct nfs_page *req = nfs_list_entry(data->pages.next);
1051 struct page *page = req->wb_page;
1053 nfs_list_remove_request(req);
1055 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1057 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1058 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1060 (long long)req_offset(req));
1063 nfs_set_pageerror(page);
1064 nfs_context_set_write_error(req->wb_context, status);
1065 dprintk(", error = %d\n", status);
1066 goto remove_request;
1069 if (nfs_write_need_commit(data)) {
1070 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1071 nfs_mark_request_commit(req);
1072 nfs_end_page_writeback(page);
1073 dprintk(" marked for commit\n");
1078 nfs_end_page_writeback(page);
1079 nfs_inode_remove_request(req);
1081 nfs_clear_page_tag_locked(req);
1083 nfs_writedata_release(calldata);
1086 static const struct rpc_call_ops nfs_write_full_ops = {
1087 .rpc_call_done = nfs_writeback_done_full,
1088 .rpc_release = nfs_writeback_release_full,
1093 * This function is called when the WRITE call is complete.
1095 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1097 struct nfs_writeargs *argp = &data->args;
1098 struct nfs_writeres *resp = &data->res;
1101 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1102 task->tk_pid, task->tk_status);
1105 * ->write_done will attempt to use post-op attributes to detect
1106 * conflicting writes by other clients. A strict interpretation
1107 * of close-to-open would allow us to continue caching even if
1108 * another writer had changed the file, but some applications
1109 * depend on tighter cache coherency when writing.
1111 status = NFS_PROTO(data->inode)->write_done(task, data);
1114 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1116 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1117 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1118 /* We tried a write call, but the server did not
1119 * commit data to stable storage even though we
1121 * Note: There is a known bug in Tru64 < 5.0 in which
1122 * the server reports NFS_DATA_SYNC, but performs
1123 * NFS_FILE_SYNC. We therefore implement this checking
1124 * as a dprintk() in order to avoid filling syslog.
1126 static unsigned long complain;
1128 if (time_before(complain, jiffies)) {
1129 dprintk("NFS: faulty NFS server %s:"
1130 " (committed = %d) != (stable = %d)\n",
1131 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1132 resp->verf->committed, argp->stable);
1133 complain = jiffies + 300 * HZ;
1137 /* Is this a short write? */
1138 if (task->tk_status >= 0 && resp->count < argp->count) {
1139 static unsigned long complain;
1141 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1143 /* Has the server at least made some progress? */
1144 if (resp->count != 0) {
1145 /* Was this an NFSv2 write or an NFSv3 stable write? */
1146 if (resp->verf->committed != NFS_UNSTABLE) {
1147 /* Resend from where the server left off */
1148 argp->offset += resp->count;
1149 argp->pgbase += resp->count;
1150 argp->count -= resp->count;
1152 /* Resend as a stable write in order to avoid
1153 * headaches in the case of a server crash.
1155 argp->stable = NFS_FILE_SYNC;
1157 rpc_restart_call(task);
1160 if (time_before(complain, jiffies)) {
1162 "NFS: Server wrote zero bytes, expected %u.\n",
1164 complain = jiffies + 300 * HZ;
1166 /* Can't do anything about it except throw an error. */
1167 task->tk_status = -EIO;
1173 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1174 void nfs_commitdata_release(void *data)
1176 struct nfs_write_data *wdata = data;
1178 put_nfs_open_context(wdata->args.context);
1179 nfs_commit_free(wdata);
1183 * Set up the argument/result storage required for the RPC call.
1185 static int nfs_commit_rpcsetup(struct list_head *head,
1186 struct nfs_write_data *data,
1189 struct nfs_page *first = nfs_list_entry(head->next);
1190 struct inode *inode = first->wb_context->path.dentry->d_inode;
1191 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1192 int priority = flush_task_priority(how);
1193 struct rpc_task *task;
1194 struct rpc_message msg = {
1195 .rpc_argp = &data->args,
1196 .rpc_resp = &data->res,
1197 .rpc_cred = first->wb_context->cred,
1199 struct rpc_task_setup task_setup_data = {
1200 .task = &data->task,
1201 .rpc_client = NFS_CLIENT(inode),
1202 .rpc_message = &msg,
1203 .callback_ops = &nfs_commit_ops,
1204 .callback_data = data,
1205 .workqueue = nfsiod_workqueue,
1207 .priority = priority,
1210 /* Set up the RPC argument and reply structs
1211 * NB: take care not to mess about with data->commit et al. */
1213 list_splice_init(head, &data->pages);
1215 data->inode = inode;
1216 data->cred = msg.rpc_cred;
1218 data->args.fh = NFS_FH(data->inode);
1219 /* Note: we always request a commit of the entire inode */
1220 data->args.offset = 0;
1221 data->args.count = 0;
1222 data->args.context = get_nfs_open_context(first->wb_context);
1223 data->res.count = 0;
1224 data->res.fattr = &data->fattr;
1225 data->res.verf = &data->verf;
1226 nfs_fattr_init(&data->fattr);
1228 /* Set up the initial task struct. */
1229 NFS_PROTO(inode)->commit_setup(data, &msg);
1231 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1233 task = rpc_run_task(&task_setup_data);
1235 return PTR_ERR(task);
1241 * Commit dirty pages
1244 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1246 struct nfs_write_data *data;
1247 struct nfs_page *req;
1249 data = nfs_commitdata_alloc();
1254 /* Set up the argument struct */
1255 return nfs_commit_rpcsetup(head, data, how);
1257 while (!list_empty(head)) {
1258 req = nfs_list_entry(head->next);
1259 nfs_list_remove_request(req);
1260 nfs_mark_request_commit(req);
1261 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1262 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1264 nfs_clear_page_tag_locked(req);
1270 * COMMIT call returned
1272 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1274 struct nfs_write_data *data = calldata;
1276 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1277 task->tk_pid, task->tk_status);
1279 /* Call the NFS version-specific code */
1280 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1284 static void nfs_commit_release(void *calldata)
1286 struct nfs_write_data *data = calldata;
1287 struct nfs_page *req;
1288 int status = data->task.tk_status;
1290 while (!list_empty(&data->pages)) {
1291 req = nfs_list_entry(data->pages.next);
1292 nfs_list_remove_request(req);
1293 nfs_clear_request_commit(req);
1295 dprintk("NFS: commit (%s/%lld %d@%lld)",
1296 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1297 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1299 (long long)req_offset(req));
1301 nfs_context_set_write_error(req->wb_context, status);
1302 nfs_inode_remove_request(req);
1303 dprintk(", error = %d\n", status);
1307 /* Okay, COMMIT succeeded, apparently. Check the verifier
1308 * returned by the server against all stored verfs. */
1309 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1310 /* We have a match */
1311 nfs_inode_remove_request(req);
1315 /* We have a mismatch. Write the page again */
1316 dprintk(" mismatch\n");
1317 nfs_mark_request_dirty(req);
1319 nfs_clear_page_tag_locked(req);
1321 nfs_commitdata_release(calldata);
1324 static const struct rpc_call_ops nfs_commit_ops = {
1325 .rpc_call_done = nfs_commit_done,
1326 .rpc_release = nfs_commit_release,
1329 int nfs_commit_inode(struct inode *inode, int how)
1334 spin_lock(&inode->i_lock);
1335 res = nfs_scan_commit(inode, &head, 0, 0);
1336 spin_unlock(&inode->i_lock);
1338 int error = nfs_commit_list(inode, &head, how);
1345 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1351 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1353 struct inode *inode = mapping->host;
1354 pgoff_t idx_start, idx_end;
1355 unsigned int npages = 0;
1357 int nocommit = how & FLUSH_NOCOMMIT;
1361 if (wbc->range_cyclic)
1364 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1365 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1366 if (idx_end > idx_start) {
1367 pgoff_t l_npages = 1 + idx_end - idx_start;
1369 if (sizeof(npages) != sizeof(l_npages) &&
1370 (pgoff_t)npages != l_npages)
1374 how &= ~FLUSH_NOCOMMIT;
1375 spin_lock(&inode->i_lock);
1377 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1382 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1385 if (how & FLUSH_INVALIDATE) {
1386 spin_unlock(&inode->i_lock);
1387 nfs_cancel_commit_list(&head);
1389 spin_lock(&inode->i_lock);
1392 pages += nfs_scan_commit(inode, &head, 0, 0);
1393 spin_unlock(&inode->i_lock);
1394 ret = nfs_commit_list(inode, &head, how);
1395 spin_lock(&inode->i_lock);
1398 spin_unlock(&inode->i_lock);
1402 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1406 ret = nfs_writepages(mapping, wbc);
1409 ret = nfs_sync_mapping_wait(mapping, wbc, how);
1414 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1418 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1419 static int nfs_write_mapping(struct address_space *mapping, int how)
1421 struct writeback_control wbc = {
1422 .bdi = mapping->backing_dev_info,
1423 .sync_mode = WB_SYNC_NONE,
1424 .nr_to_write = LONG_MAX,
1425 .for_writepages = 1,
1430 ret = __nfs_write_mapping(mapping, &wbc, how);
1433 wbc.sync_mode = WB_SYNC_ALL;
1434 return __nfs_write_mapping(mapping, &wbc, how);
1438 * flush the inode to disk.
1440 int nfs_wb_all(struct inode *inode)
1442 return nfs_write_mapping(inode->i_mapping, 0);
1445 int nfs_wb_nocommit(struct inode *inode)
1447 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1450 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1452 struct nfs_page *req;
1453 loff_t range_start = page_offset(page);
1454 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1455 struct writeback_control wbc = {
1456 .bdi = page->mapping->backing_dev_info,
1457 .sync_mode = WB_SYNC_ALL,
1458 .nr_to_write = LONG_MAX,
1459 .range_start = range_start,
1460 .range_end = range_end,
1464 BUG_ON(!PageLocked(page));
1466 req = nfs_page_find_request(page);
1469 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1470 nfs_release_request(req);
1473 if (nfs_lock_request_dontget(req)) {
1474 nfs_inode_remove_request(req);
1476 * In case nfs_inode_remove_request has marked the
1477 * page as being dirty
1479 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1480 nfs_unlock_request(req);
1483 ret = nfs_wait_on_request(req);
1487 if (!PagePrivate(page))
1489 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1494 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1497 loff_t range_start = page_offset(page);
1498 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1499 struct writeback_control wbc = {
1500 .bdi = page->mapping->backing_dev_info,
1501 .sync_mode = WB_SYNC_ALL,
1502 .nr_to_write = LONG_MAX,
1503 .range_start = range_start,
1504 .range_end = range_end,
1509 if (clear_page_dirty_for_io(page)) {
1510 ret = nfs_writepage_locked(page, &wbc);
1513 } else if (!PagePrivate(page))
1515 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1518 } while (PagePrivate(page));
1521 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1526 * Write back all requests on one page - we do this before reading it.
1528 int nfs_wb_page(struct inode *inode, struct page* page)
1530 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1533 int __init nfs_init_writepagecache(void)
1535 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1536 sizeof(struct nfs_write_data),
1537 0, SLAB_HWCACHE_ALIGN,
1539 if (nfs_wdata_cachep == NULL)
1542 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1544 if (nfs_wdata_mempool == NULL)
1547 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1549 if (nfs_commit_mempool == NULL)
1553 * NFS congestion size, scale with available memory.
1565 * This allows larger machines to have larger/more transfers.
1566 * Limit the default to 256M
1568 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1569 if (nfs_congestion_kb > 256*1024)
1570 nfs_congestion_kb = 256*1024;
1575 void nfs_destroy_writepagecache(void)
1577 mempool_destroy(nfs_commit_mempool);
1578 mempool_destroy(nfs_wdata_mempool);
1579 kmem_cache_destroy(nfs_wdata_cachep);