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;
139 spin_lock(&inode->i_lock);
140 i_size = i_size_read(inode);
141 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
142 if (i_size > 0 && page->index < end_index)
144 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
147 i_size_write(inode, end);
148 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
150 spin_unlock(&inode->i_lock);
153 /* A writeback failed: mark the page as bad, and invalidate the page cache */
154 static void nfs_set_pageerror(struct page *page)
157 nfs_zap_mapping(page->mapping->host, page->mapping);
160 /* We can set the PG_uptodate flag if we see that a write request
161 * covers the full page.
163 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
165 if (PageUptodate(page))
169 if (count != nfs_page_length(page))
171 SetPageUptodate(page);
174 static int wb_priority(struct writeback_control *wbc)
176 if (wbc->for_reclaim)
177 return FLUSH_HIGHPRI | FLUSH_STABLE;
178 if (wbc->for_kupdate)
184 * NFS congestion control
187 int nfs_congestion_kb;
189 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
190 #define NFS_CONGESTION_OFF_THRESH \
191 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
193 static int nfs_set_page_writeback(struct page *page)
195 int ret = test_set_page_writeback(page);
198 struct inode *inode = page->mapping->host;
199 struct nfs_server *nfss = NFS_SERVER(inode);
201 if (atomic_long_inc_return(&nfss->writeback) >
202 NFS_CONGESTION_ON_THRESH)
203 set_bdi_congested(&nfss->backing_dev_info, WRITE);
208 static void nfs_end_page_writeback(struct page *page)
210 struct inode *inode = page->mapping->host;
211 struct nfs_server *nfss = NFS_SERVER(inode);
213 end_page_writeback(page);
214 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
215 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
219 * Find an associated nfs write request, and prepare to flush it out
220 * May return an error if the user signalled nfs_wait_on_request().
222 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
225 struct inode *inode = page->mapping->host;
226 struct nfs_page *req;
229 spin_lock(&inode->i_lock);
231 req = nfs_page_find_request_locked(page);
233 spin_unlock(&inode->i_lock);
236 if (nfs_set_page_tag_locked(req))
238 /* Note: If we hold the page lock, as is the case in nfs_writepage,
239 * then the call to nfs_set_page_tag_locked() will always
240 * succeed provided that someone hasn't already marked the
241 * request as dirty (in which case we don't care).
243 spin_unlock(&inode->i_lock);
244 ret = nfs_wait_on_request(req);
245 nfs_release_request(req);
248 spin_lock(&inode->i_lock);
250 if (test_bit(PG_CLEAN, &req->wb_flags)) {
251 spin_unlock(&inode->i_lock);
254 if (nfs_set_page_writeback(page) != 0) {
255 spin_unlock(&inode->i_lock);
258 spin_unlock(&inode->i_lock);
259 if (!nfs_pageio_add_request(pgio, req)) {
260 nfs_redirty_request(req);
261 return pgio->pg_error;
266 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
268 struct inode *inode = page->mapping->host;
270 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
271 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
273 nfs_pageio_cond_complete(pgio, page->index);
274 return nfs_page_async_flush(pgio, page);
278 * Write an mmapped page to the server.
280 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
282 struct nfs_pageio_descriptor pgio;
285 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
286 err = nfs_do_writepage(page, wbc, &pgio);
287 nfs_pageio_complete(&pgio);
290 if (pgio.pg_error < 0)
291 return pgio.pg_error;
295 int nfs_writepage(struct page *page, struct writeback_control *wbc)
299 ret = nfs_writepage_locked(page, wbc);
304 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
308 ret = nfs_do_writepage(page, wbc, data);
313 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
315 struct inode *inode = mapping->host;
316 struct nfs_pageio_descriptor pgio;
319 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
321 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
322 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
323 nfs_pageio_complete(&pgio);
326 if (pgio.pg_error < 0)
327 return pgio.pg_error;
332 * Insert a write request into an inode
334 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
336 struct nfs_inode *nfsi = NFS_I(inode);
339 error = radix_tree_preload(GFP_NOFS);
343 /* Lock the request! */
344 nfs_lock_request_dontget(req);
346 spin_lock(&inode->i_lock);
347 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
351 if (nfs_have_delegation(inode, FMODE_WRITE))
354 SetPagePrivate(req->wb_page);
355 set_page_private(req->wb_page, (unsigned long)req);
357 kref_get(&req->wb_kref);
358 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
359 NFS_PAGE_TAG_LOCKED);
360 spin_unlock(&inode->i_lock);
361 radix_tree_preload_end();
367 * Remove a write request from an inode
369 static void nfs_inode_remove_request(struct nfs_page *req)
371 struct inode *inode = req->wb_context->path.dentry->d_inode;
372 struct nfs_inode *nfsi = NFS_I(inode);
374 BUG_ON (!NFS_WBACK_BUSY(req));
376 spin_lock(&inode->i_lock);
377 set_page_private(req->wb_page, 0);
378 ClearPagePrivate(req->wb_page);
379 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
382 spin_unlock(&inode->i_lock);
385 spin_unlock(&inode->i_lock);
386 nfs_clear_request(req);
387 nfs_release_request(req);
391 nfs_mark_request_dirty(struct nfs_page *req)
393 __set_page_dirty_nobuffers(req->wb_page);
396 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
398 * Add a request to the inode's commit list.
401 nfs_mark_request_commit(struct nfs_page *req)
403 struct inode *inode = req->wb_context->path.dentry->d_inode;
404 struct nfs_inode *nfsi = NFS_I(inode);
406 spin_lock(&inode->i_lock);
407 set_bit(PG_CLEAN, &(req)->wb_flags);
408 radix_tree_tag_set(&nfsi->nfs_page_tree,
410 NFS_PAGE_TAG_COMMIT);
411 spin_unlock(&inode->i_lock);
412 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
413 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
414 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
418 nfs_clear_request_commit(struct nfs_page *req)
420 struct page *page = req->wb_page;
422 if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
423 dec_zone_page_state(page, NR_UNSTABLE_NFS);
424 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
431 int nfs_write_need_commit(struct nfs_write_data *data)
433 return data->verf.committed != NFS_FILE_SYNC;
437 int nfs_reschedule_unstable_write(struct nfs_page *req)
439 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
440 nfs_mark_request_commit(req);
443 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
444 nfs_mark_request_dirty(req);
451 nfs_mark_request_commit(struct nfs_page *req)
456 nfs_clear_request_commit(struct nfs_page *req)
462 int nfs_write_need_commit(struct nfs_write_data *data)
468 int nfs_reschedule_unstable_write(struct nfs_page *req)
475 * Wait for a request to complete.
477 * Interruptible by fatal signals only.
479 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
481 struct nfs_inode *nfsi = NFS_I(inode);
482 struct nfs_page *req;
483 pgoff_t idx_end, next;
484 unsigned int res = 0;
490 idx_end = idx_start + npages - 1;
493 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
494 if (req->wb_index > idx_end)
497 next = req->wb_index + 1;
498 BUG_ON(!NFS_WBACK_BUSY(req));
500 kref_get(&req->wb_kref);
501 spin_unlock(&inode->i_lock);
502 error = nfs_wait_on_request(req);
503 nfs_release_request(req);
504 spin_lock(&inode->i_lock);
512 static void nfs_cancel_commit_list(struct list_head *head)
514 struct nfs_page *req;
516 while(!list_empty(head)) {
517 req = nfs_list_entry(head->next);
518 nfs_list_remove_request(req);
519 nfs_clear_request_commit(req);
520 nfs_inode_remove_request(req);
521 nfs_unlock_request(req);
526 nfs_need_commit(struct nfs_inode *nfsi)
528 return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
531 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
533 * nfs_scan_commit - Scan an inode for commit requests
534 * @inode: NFS inode to scan
535 * @dst: destination list
536 * @idx_start: lower bound of page->index to scan.
537 * @npages: idx_start + npages sets the upper bound to scan.
539 * Moves requests from the inode's 'commit' request list.
540 * The requests are *not* checked to ensure that they form a contiguous set.
543 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
545 struct nfs_inode *nfsi = NFS_I(inode);
547 if (!nfs_need_commit(nfsi))
550 return nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
553 static inline int nfs_need_commit(struct nfs_inode *nfsi)
558 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
565 * Search for an existing write request, and attempt to update
566 * it to reflect a new dirty region on a given page.
568 * If the attempt fails, then the existing request is flushed out
571 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
576 struct nfs_page *req;
581 if (!PagePrivate(page))
584 end = offset + bytes;
585 spin_lock(&inode->i_lock);
588 req = nfs_page_find_request_locked(page);
592 rqend = req->wb_offset + req->wb_bytes;
594 * Tell the caller to flush out the request if
595 * the offsets are non-contiguous.
596 * Note: nfs_flush_incompatible() will already
597 * have flushed out requests having wrong owners.
600 || end < req->wb_offset)
603 if (nfs_set_page_tag_locked(req))
606 /* The request is locked, so wait and then retry */
607 spin_unlock(&inode->i_lock);
608 error = nfs_wait_on_request(req);
609 nfs_release_request(req);
612 spin_lock(&inode->i_lock);
615 if (nfs_clear_request_commit(req))
616 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
617 req->wb_index, NFS_PAGE_TAG_COMMIT);
619 /* Okay, the request matches. Update the region */
620 if (offset < req->wb_offset) {
621 req->wb_offset = offset;
622 req->wb_pgbase = offset;
625 req->wb_bytes = end - req->wb_offset;
627 req->wb_bytes = rqend - req->wb_offset;
629 spin_unlock(&inode->i_lock);
632 spin_unlock(&inode->i_lock);
633 nfs_release_request(req);
634 error = nfs_wb_page(inode, page);
636 return ERR_PTR(error);
640 * Try to update an existing write request, or create one if there is none.
642 * Note: Should always be called with the Page Lock held to prevent races
643 * if we have to add a new request. Also assumes that the caller has
644 * already called nfs_flush_incompatible() if necessary.
646 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
647 struct page *page, unsigned int offset, unsigned int bytes)
649 struct inode *inode = page->mapping->host;
650 struct nfs_page *req;
653 req = nfs_try_to_update_request(inode, page, offset, bytes);
656 req = nfs_create_request(ctx, inode, page, offset, bytes);
659 error = nfs_inode_add_request(inode, req);
661 nfs_release_request(req);
662 req = ERR_PTR(error);
668 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
669 unsigned int offset, unsigned int count)
671 struct nfs_page *req;
673 req = nfs_setup_write_request(ctx, page, offset, count);
676 /* Update file length */
677 nfs_grow_file(page, offset, count);
678 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
679 nfs_clear_page_tag_locked(req);
683 int nfs_flush_incompatible(struct file *file, struct page *page)
685 struct nfs_open_context *ctx = nfs_file_open_context(file);
686 struct nfs_page *req;
687 int do_flush, status;
689 * Look for a request corresponding to this page. If there
690 * is one, and it belongs to another file, we flush it out
691 * before we try to copy anything into the page. Do this
692 * due to the lack of an ACCESS-type call in NFSv2.
693 * Also do the same if we find a request from an existing
697 req = nfs_page_find_request(page);
700 do_flush = req->wb_page != page || req->wb_context != ctx;
701 nfs_release_request(req);
704 status = nfs_wb_page(page->mapping->host, page);
705 } while (status == 0);
710 * If the page cache is marked as unsafe or invalid, then we can't rely on
711 * the PageUptodate() flag. In this case, we will need to turn off
712 * write optimisations that depend on the page contents being correct.
714 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
716 return PageUptodate(page) &&
717 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
721 * Update and possibly write a cached page of an NFS file.
723 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
724 * things with a page scheduled for an RPC call (e.g. invalidate it).
726 int nfs_updatepage(struct file *file, struct page *page,
727 unsigned int offset, unsigned int count)
729 struct nfs_open_context *ctx = nfs_file_open_context(file);
730 struct inode *inode = page->mapping->host;
733 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
735 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
736 file->f_path.dentry->d_parent->d_name.name,
737 file->f_path.dentry->d_name.name, count,
738 (long long)(page_offset(page) + offset));
740 /* If we're not using byte range locks, and we know the page
741 * is up to date, it may be more efficient to extend the write
742 * to cover the entire page in order to avoid fragmentation
745 if (nfs_write_pageuptodate(page, inode) &&
746 inode->i_flock == NULL &&
747 !(file->f_flags & O_SYNC)) {
748 count = max(count + offset, nfs_page_length(page));
752 status = nfs_writepage_setup(ctx, page, offset, count);
754 nfs_set_pageerror(page);
756 __set_page_dirty_nobuffers(page);
758 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
759 status, (long long)i_size_read(inode));
763 static void nfs_writepage_release(struct nfs_page *req)
766 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
767 nfs_end_page_writeback(req->wb_page);
768 nfs_inode_remove_request(req);
770 nfs_end_page_writeback(req->wb_page);
771 nfs_clear_page_tag_locked(req);
774 static int flush_task_priority(int how)
776 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
778 return RPC_PRIORITY_HIGH;
780 return RPC_PRIORITY_LOW;
782 return RPC_PRIORITY_NORMAL;
786 * Set up the argument/result storage required for the RPC call.
788 static int nfs_write_rpcsetup(struct nfs_page *req,
789 struct nfs_write_data *data,
790 const struct rpc_call_ops *call_ops,
791 unsigned int count, unsigned int offset,
794 struct inode *inode = req->wb_context->path.dentry->d_inode;
795 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
796 int priority = flush_task_priority(how);
797 struct rpc_task *task;
798 struct rpc_message msg = {
799 .rpc_argp = &data->args,
800 .rpc_resp = &data->res,
801 .rpc_cred = req->wb_context->cred,
803 struct rpc_task_setup task_setup_data = {
804 .rpc_client = NFS_CLIENT(inode),
807 .callback_ops = call_ops,
808 .callback_data = data,
809 .workqueue = nfsiod_workqueue,
811 .priority = priority,
814 /* Set up the RPC argument and reply structs
815 * NB: take care not to mess about with data->commit et al. */
818 data->inode = inode = req->wb_context->path.dentry->d_inode;
819 data->cred = msg.rpc_cred;
821 data->args.fh = NFS_FH(inode);
822 data->args.offset = req_offset(req) + offset;
823 data->args.pgbase = req->wb_pgbase + offset;
824 data->args.pages = data->pagevec;
825 data->args.count = count;
826 data->args.context = get_nfs_open_context(req->wb_context);
827 data->args.stable = NFS_UNSTABLE;
828 if (how & FLUSH_STABLE) {
829 data->args.stable = NFS_DATA_SYNC;
830 if (!nfs_need_commit(NFS_I(inode)))
831 data->args.stable = NFS_FILE_SYNC;
834 data->res.fattr = &data->fattr;
835 data->res.count = count;
836 data->res.verf = &data->verf;
837 nfs_fattr_init(&data->fattr);
839 /* Set up the initial task struct. */
840 NFS_PROTO(inode)->write_setup(data, &msg);
842 dprintk("NFS: %5u initiated write call "
843 "(req %s/%lld, %u bytes @ offset %llu)\n",
846 (long long)NFS_FILEID(inode),
848 (unsigned long long)data->args.offset);
850 task = rpc_run_task(&task_setup_data);
852 return PTR_ERR(task);
857 /* If a nfs_flush_* function fails, it should remove reqs from @head and
858 * call this on each, which will prepare them to be retried on next
859 * writeback using standard nfs.
861 static void nfs_redirty_request(struct nfs_page *req)
863 nfs_mark_request_dirty(req);
864 nfs_end_page_writeback(req->wb_page);
865 nfs_clear_page_tag_locked(req);
869 * Generate multiple small requests to write out a single
870 * contiguous dirty area on one page.
872 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
874 struct nfs_page *req = nfs_list_entry(head->next);
875 struct page *page = req->wb_page;
876 struct nfs_write_data *data;
877 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
883 nfs_list_remove_request(req);
887 size_t len = min(nbytes, wsize);
889 data = nfs_writedata_alloc(1);
892 list_add(&data->pages, &list);
895 } while (nbytes != 0);
896 atomic_set(&req->wb_complete, requests);
898 ClearPageError(page);
904 data = list_entry(list.next, struct nfs_write_data, pages);
905 list_del_init(&data->pages);
907 data->pagevec[0] = page;
911 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
917 } while (nbytes != 0);
922 while (!list_empty(&list)) {
923 data = list_entry(list.next, struct nfs_write_data, pages);
924 list_del(&data->pages);
925 nfs_writedata_release(data);
927 nfs_redirty_request(req);
932 * Create an RPC task for the given write request and kick it.
933 * The page must have been locked by the caller.
935 * It may happen that the page we're passed is not marked dirty.
936 * This is the case if nfs_updatepage detects a conflicting request
937 * that has been written but not committed.
939 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
941 struct nfs_page *req;
943 struct nfs_write_data *data;
945 data = nfs_writedata_alloc(npages);
949 pages = data->pagevec;
950 while (!list_empty(head)) {
951 req = nfs_list_entry(head->next);
952 nfs_list_remove_request(req);
953 nfs_list_add_request(req, &data->pages);
954 ClearPageError(req->wb_page);
955 *pages++ = req->wb_page;
957 req = nfs_list_entry(data->pages.next);
959 /* Set up the argument struct */
960 return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
962 while (!list_empty(head)) {
963 req = nfs_list_entry(head->next);
964 nfs_list_remove_request(req);
965 nfs_redirty_request(req);
970 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
971 struct inode *inode, int ioflags)
973 size_t wsize = NFS_SERVER(inode)->wsize;
975 if (wsize < PAGE_CACHE_SIZE)
976 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
978 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
982 * Handle a write reply that flushed part of a page.
984 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
986 struct nfs_write_data *data = calldata;
988 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
990 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
992 NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
993 data->req->wb_bytes, (long long)req_offset(data->req));
995 nfs_writeback_done(task, data);
998 static void nfs_writeback_release_partial(void *calldata)
1000 struct nfs_write_data *data = calldata;
1001 struct nfs_page *req = data->req;
1002 struct page *page = req->wb_page;
1003 int status = data->task.tk_status;
1006 nfs_set_pageerror(page);
1007 nfs_context_set_write_error(req->wb_context, status);
1008 dprintk(", error = %d\n", status);
1012 if (nfs_write_need_commit(data)) {
1013 struct inode *inode = page->mapping->host;
1015 spin_lock(&inode->i_lock);
1016 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1017 /* Do nothing we need to resend the writes */
1018 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1019 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1020 dprintk(" defer commit\n");
1021 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1022 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1023 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1024 dprintk(" server reboot detected\n");
1026 spin_unlock(&inode->i_lock);
1031 if (atomic_dec_and_test(&req->wb_complete))
1032 nfs_writepage_release(req);
1033 nfs_writedata_release(calldata);
1036 static const struct rpc_call_ops nfs_write_partial_ops = {
1037 .rpc_call_done = nfs_writeback_done_partial,
1038 .rpc_release = nfs_writeback_release_partial,
1042 * Handle a write reply that flushes a whole page.
1044 * FIXME: There is an inherent race with invalidate_inode_pages and
1045 * writebacks since the page->count is kept > 1 for as long
1046 * as the page has a write request pending.
1048 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1050 struct nfs_write_data *data = calldata;
1052 nfs_writeback_done(task, data);
1055 static void nfs_writeback_release_full(void *calldata)
1057 struct nfs_write_data *data = calldata;
1058 int status = data->task.tk_status;
1060 /* Update attributes as result of writeback. */
1061 while (!list_empty(&data->pages)) {
1062 struct nfs_page *req = nfs_list_entry(data->pages.next);
1063 struct page *page = req->wb_page;
1065 nfs_list_remove_request(req);
1067 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1069 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1070 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1072 (long long)req_offset(req));
1075 nfs_set_pageerror(page);
1076 nfs_context_set_write_error(req->wb_context, status);
1077 dprintk(", error = %d\n", status);
1078 goto remove_request;
1081 if (nfs_write_need_commit(data)) {
1082 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1083 nfs_mark_request_commit(req);
1084 nfs_end_page_writeback(page);
1085 dprintk(" marked for commit\n");
1090 nfs_end_page_writeback(page);
1091 nfs_inode_remove_request(req);
1093 nfs_clear_page_tag_locked(req);
1095 nfs_writedata_release(calldata);
1098 static const struct rpc_call_ops nfs_write_full_ops = {
1099 .rpc_call_done = nfs_writeback_done_full,
1100 .rpc_release = nfs_writeback_release_full,
1105 * This function is called when the WRITE call is complete.
1107 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1109 struct nfs_writeargs *argp = &data->args;
1110 struct nfs_writeres *resp = &data->res;
1113 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1114 task->tk_pid, task->tk_status);
1117 * ->write_done will attempt to use post-op attributes to detect
1118 * conflicting writes by other clients. A strict interpretation
1119 * of close-to-open would allow us to continue caching even if
1120 * another writer had changed the file, but some applications
1121 * depend on tighter cache coherency when writing.
1123 status = NFS_PROTO(data->inode)->write_done(task, data);
1126 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1128 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1129 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1130 /* We tried a write call, but the server did not
1131 * commit data to stable storage even though we
1133 * Note: There is a known bug in Tru64 < 5.0 in which
1134 * the server reports NFS_DATA_SYNC, but performs
1135 * NFS_FILE_SYNC. We therefore implement this checking
1136 * as a dprintk() in order to avoid filling syslog.
1138 static unsigned long complain;
1140 if (time_before(complain, jiffies)) {
1141 dprintk("NFS: faulty NFS server %s:"
1142 " (committed = %d) != (stable = %d)\n",
1143 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1144 resp->verf->committed, argp->stable);
1145 complain = jiffies + 300 * HZ;
1149 /* Is this a short write? */
1150 if (task->tk_status >= 0 && resp->count < argp->count) {
1151 static unsigned long complain;
1153 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1155 /* Has the server at least made some progress? */
1156 if (resp->count != 0) {
1157 /* Was this an NFSv2 write or an NFSv3 stable write? */
1158 if (resp->verf->committed != NFS_UNSTABLE) {
1159 /* Resend from where the server left off */
1160 argp->offset += resp->count;
1161 argp->pgbase += resp->count;
1162 argp->count -= resp->count;
1164 /* Resend as a stable write in order to avoid
1165 * headaches in the case of a server crash.
1167 argp->stable = NFS_FILE_SYNC;
1169 rpc_restart_call(task);
1172 if (time_before(complain, jiffies)) {
1174 "NFS: Server wrote zero bytes, expected %u.\n",
1176 complain = jiffies + 300 * HZ;
1178 /* Can't do anything about it except throw an error. */
1179 task->tk_status = -EIO;
1185 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1186 void nfs_commitdata_release(void *data)
1188 struct nfs_write_data *wdata = data;
1190 put_nfs_open_context(wdata->args.context);
1191 nfs_commit_free(wdata);
1195 * Set up the argument/result storage required for the RPC call.
1197 static int nfs_commit_rpcsetup(struct list_head *head,
1198 struct nfs_write_data *data,
1201 struct nfs_page *first = nfs_list_entry(head->next);
1202 struct inode *inode = first->wb_context->path.dentry->d_inode;
1203 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1204 int priority = flush_task_priority(how);
1205 struct rpc_task *task;
1206 struct rpc_message msg = {
1207 .rpc_argp = &data->args,
1208 .rpc_resp = &data->res,
1209 .rpc_cred = first->wb_context->cred,
1211 struct rpc_task_setup task_setup_data = {
1212 .task = &data->task,
1213 .rpc_client = NFS_CLIENT(inode),
1214 .rpc_message = &msg,
1215 .callback_ops = &nfs_commit_ops,
1216 .callback_data = data,
1217 .workqueue = nfsiod_workqueue,
1219 .priority = priority,
1222 /* Set up the RPC argument and reply structs
1223 * NB: take care not to mess about with data->commit et al. */
1225 list_splice_init(head, &data->pages);
1227 data->inode = inode;
1228 data->cred = msg.rpc_cred;
1230 data->args.fh = NFS_FH(data->inode);
1231 /* Note: we always request a commit of the entire inode */
1232 data->args.offset = 0;
1233 data->args.count = 0;
1234 data->args.context = get_nfs_open_context(first->wb_context);
1235 data->res.count = 0;
1236 data->res.fattr = &data->fattr;
1237 data->res.verf = &data->verf;
1238 nfs_fattr_init(&data->fattr);
1240 /* Set up the initial task struct. */
1241 NFS_PROTO(inode)->commit_setup(data, &msg);
1243 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1245 task = rpc_run_task(&task_setup_data);
1247 return PTR_ERR(task);
1253 * Commit dirty pages
1256 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1258 struct nfs_write_data *data;
1259 struct nfs_page *req;
1261 data = nfs_commitdata_alloc();
1266 /* Set up the argument struct */
1267 return nfs_commit_rpcsetup(head, data, how);
1269 while (!list_empty(head)) {
1270 req = nfs_list_entry(head->next);
1271 nfs_list_remove_request(req);
1272 nfs_mark_request_commit(req);
1273 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1274 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1276 nfs_clear_page_tag_locked(req);
1282 * COMMIT call returned
1284 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1286 struct nfs_write_data *data = calldata;
1288 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1289 task->tk_pid, task->tk_status);
1291 /* Call the NFS version-specific code */
1292 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1296 static void nfs_commit_release(void *calldata)
1298 struct nfs_write_data *data = calldata;
1299 struct nfs_page *req;
1300 int status = data->task.tk_status;
1302 while (!list_empty(&data->pages)) {
1303 req = nfs_list_entry(data->pages.next);
1304 nfs_list_remove_request(req);
1305 nfs_clear_request_commit(req);
1307 dprintk("NFS: commit (%s/%lld %d@%lld)",
1308 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1309 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1311 (long long)req_offset(req));
1313 nfs_context_set_write_error(req->wb_context, status);
1314 nfs_inode_remove_request(req);
1315 dprintk(", error = %d\n", status);
1319 /* Okay, COMMIT succeeded, apparently. Check the verifier
1320 * returned by the server against all stored verfs. */
1321 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1322 /* We have a match */
1323 nfs_inode_remove_request(req);
1327 /* We have a mismatch. Write the page again */
1328 dprintk(" mismatch\n");
1329 nfs_mark_request_dirty(req);
1331 nfs_clear_page_tag_locked(req);
1333 nfs_commitdata_release(calldata);
1336 static const struct rpc_call_ops nfs_commit_ops = {
1337 .rpc_call_done = nfs_commit_done,
1338 .rpc_release = nfs_commit_release,
1341 int nfs_commit_inode(struct inode *inode, int how)
1346 spin_lock(&inode->i_lock);
1347 res = nfs_scan_commit(inode, &head, 0, 0);
1348 spin_unlock(&inode->i_lock);
1350 int error = nfs_commit_list(inode, &head, how);
1357 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1363 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1365 struct inode *inode = mapping->host;
1366 pgoff_t idx_start, idx_end;
1367 unsigned int npages = 0;
1369 int nocommit = how & FLUSH_NOCOMMIT;
1373 if (wbc->range_cyclic)
1376 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1377 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1378 if (idx_end > idx_start) {
1379 pgoff_t l_npages = 1 + idx_end - idx_start;
1381 if (sizeof(npages) != sizeof(l_npages) &&
1382 (pgoff_t)npages != l_npages)
1386 how &= ~FLUSH_NOCOMMIT;
1387 spin_lock(&inode->i_lock);
1389 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1394 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1397 if (how & FLUSH_INVALIDATE) {
1398 spin_unlock(&inode->i_lock);
1399 nfs_cancel_commit_list(&head);
1401 spin_lock(&inode->i_lock);
1404 pages += nfs_scan_commit(inode, &head, 0, 0);
1405 spin_unlock(&inode->i_lock);
1406 ret = nfs_commit_list(inode, &head, how);
1407 spin_lock(&inode->i_lock);
1410 spin_unlock(&inode->i_lock);
1414 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1418 ret = nfs_writepages(mapping, wbc);
1421 ret = nfs_sync_mapping_wait(mapping, wbc, how);
1426 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1430 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1431 static int nfs_write_mapping(struct address_space *mapping, int how)
1433 struct writeback_control wbc = {
1434 .bdi = mapping->backing_dev_info,
1435 .sync_mode = WB_SYNC_NONE,
1436 .nr_to_write = LONG_MAX,
1438 .range_end = LLONG_MAX,
1439 .for_writepages = 1,
1443 ret = __nfs_write_mapping(mapping, &wbc, how);
1446 wbc.sync_mode = WB_SYNC_ALL;
1447 return __nfs_write_mapping(mapping, &wbc, how);
1451 * flush the inode to disk.
1453 int nfs_wb_all(struct inode *inode)
1455 return nfs_write_mapping(inode->i_mapping, 0);
1458 int nfs_wb_nocommit(struct inode *inode)
1460 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1463 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1465 struct nfs_page *req;
1466 loff_t range_start = page_offset(page);
1467 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1468 struct writeback_control wbc = {
1469 .bdi = page->mapping->backing_dev_info,
1470 .sync_mode = WB_SYNC_ALL,
1471 .nr_to_write = LONG_MAX,
1472 .range_start = range_start,
1473 .range_end = range_end,
1477 BUG_ON(!PageLocked(page));
1479 req = nfs_page_find_request(page);
1482 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1483 nfs_release_request(req);
1486 if (nfs_lock_request_dontget(req)) {
1487 nfs_inode_remove_request(req);
1489 * In case nfs_inode_remove_request has marked the
1490 * page as being dirty
1492 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1493 nfs_unlock_request(req);
1496 ret = nfs_wait_on_request(req);
1500 if (!PagePrivate(page))
1502 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1507 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1510 loff_t range_start = page_offset(page);
1511 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1512 struct writeback_control wbc = {
1513 .bdi = page->mapping->backing_dev_info,
1514 .sync_mode = WB_SYNC_ALL,
1515 .nr_to_write = LONG_MAX,
1516 .range_start = range_start,
1517 .range_end = range_end,
1522 if (clear_page_dirty_for_io(page)) {
1523 ret = nfs_writepage_locked(page, &wbc);
1526 } else if (!PagePrivate(page))
1528 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1531 } while (PagePrivate(page));
1534 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1539 * Write back all requests on one page - we do this before reading it.
1541 int nfs_wb_page(struct inode *inode, struct page* page)
1543 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1546 int __init nfs_init_writepagecache(void)
1548 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1549 sizeof(struct nfs_write_data),
1550 0, SLAB_HWCACHE_ALIGN,
1552 if (nfs_wdata_cachep == NULL)
1555 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1557 if (nfs_wdata_mempool == NULL)
1560 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1562 if (nfs_commit_mempool == NULL)
1566 * NFS congestion size, scale with available memory.
1578 * This allows larger machines to have larger/more transfers.
1579 * Limit the default to 256M
1581 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1582 if (nfs_congestion_kb > 256*1024)
1583 nfs_congestion_kb = 256*1024;
1588 void nfs_destroy_writepagecache(void)
1590 mempool_destroy(nfs_commit_mempool);
1591 mempool_destroy(nfs_wdata_mempool);
1592 kmem_cache_destroy(nfs_wdata_cachep);