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_NEED_COMMIT, &req->wb_flags)) {
246 /* This request is marked for commit */
247 spin_unlock(&inode->i_lock);
248 nfs_clear_page_tag_locked(req);
249 nfs_pageio_complete(pgio);
252 if (nfs_set_page_writeback(page) != 0) {
253 spin_unlock(&inode->i_lock);
256 spin_unlock(&inode->i_lock);
257 if (!nfs_pageio_add_request(pgio, req)) {
258 nfs_redirty_request(req);
259 return pgio->pg_error;
264 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
266 struct inode *inode = page->mapping->host;
268 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
269 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
271 nfs_pageio_cond_complete(pgio, page->index);
272 return nfs_page_async_flush(pgio, page);
276 * Write an mmapped page to the server.
278 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
280 struct nfs_pageio_descriptor pgio;
283 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
284 err = nfs_do_writepage(page, wbc, &pgio);
285 nfs_pageio_complete(&pgio);
288 if (pgio.pg_error < 0)
289 return pgio.pg_error;
293 int nfs_writepage(struct page *page, struct writeback_control *wbc)
297 ret = nfs_writepage_locked(page, wbc);
302 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
306 ret = nfs_do_writepage(page, wbc, data);
311 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
313 struct inode *inode = mapping->host;
314 struct nfs_pageio_descriptor pgio;
317 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
319 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
320 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
321 nfs_pageio_complete(&pgio);
324 if (pgio.pg_error < 0)
325 return pgio.pg_error;
330 * Insert a write request into an inode
332 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
334 struct nfs_inode *nfsi = NFS_I(inode);
337 error = radix_tree_preload(GFP_NOFS);
341 /* Lock the request! */
342 nfs_lock_request_dontget(req);
344 spin_lock(&inode->i_lock);
345 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
349 if (nfs_have_delegation(inode, FMODE_WRITE))
352 SetPagePrivate(req->wb_page);
353 set_page_private(req->wb_page, (unsigned long)req);
355 kref_get(&req->wb_kref);
356 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
357 NFS_PAGE_TAG_LOCKED);
358 spin_unlock(&inode->i_lock);
359 radix_tree_preload_end();
365 * Remove a write request from an inode
367 static void nfs_inode_remove_request(struct nfs_page *req)
369 struct inode *inode = req->wb_context->path.dentry->d_inode;
370 struct nfs_inode *nfsi = NFS_I(inode);
372 BUG_ON (!NFS_WBACK_BUSY(req));
374 spin_lock(&inode->i_lock);
375 set_page_private(req->wb_page, 0);
376 ClearPagePrivate(req->wb_page);
377 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
380 spin_unlock(&inode->i_lock);
383 spin_unlock(&inode->i_lock);
384 nfs_clear_request(req);
385 nfs_release_request(req);
389 nfs_mark_request_dirty(struct nfs_page *req)
391 __set_page_dirty_nobuffers(req->wb_page);
395 * Check if a request is dirty
398 nfs_dirty_request(struct nfs_page *req)
400 struct page *page = req->wb_page;
402 if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
404 return !PageWriteback(page);
407 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
409 * Add a request to the inode's commit list.
412 nfs_mark_request_commit(struct nfs_page *req)
414 struct inode *inode = req->wb_context->path.dentry->d_inode;
415 struct nfs_inode *nfsi = NFS_I(inode);
417 spin_lock(&inode->i_lock);
419 set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
420 radix_tree_tag_set(&nfsi->nfs_page_tree,
422 NFS_PAGE_TAG_COMMIT);
423 spin_unlock(&inode->i_lock);
424 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
425 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
426 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
430 int nfs_write_need_commit(struct nfs_write_data *data)
432 return data->verf.committed != NFS_FILE_SYNC;
436 int nfs_reschedule_unstable_write(struct nfs_page *req)
438 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
439 nfs_mark_request_commit(req);
442 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
443 nfs_mark_request_dirty(req);
450 nfs_mark_request_commit(struct nfs_page *req)
455 int nfs_write_need_commit(struct nfs_write_data *data)
461 int nfs_reschedule_unstable_write(struct nfs_page *req)
468 * Wait for a request to complete.
470 * Interruptible by fatal signals only.
472 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
474 struct nfs_inode *nfsi = NFS_I(inode);
475 struct nfs_page *req;
476 pgoff_t idx_end, next;
477 unsigned int res = 0;
483 idx_end = idx_start + npages - 1;
486 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
487 if (req->wb_index > idx_end)
490 next = req->wb_index + 1;
491 BUG_ON(!NFS_WBACK_BUSY(req));
493 kref_get(&req->wb_kref);
494 spin_unlock(&inode->i_lock);
495 error = nfs_wait_on_request(req);
496 nfs_release_request(req);
497 spin_lock(&inode->i_lock);
505 static void nfs_cancel_commit_list(struct list_head *head)
507 struct nfs_page *req;
509 while(!list_empty(head)) {
510 req = nfs_list_entry(head->next);
511 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
512 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
514 nfs_list_remove_request(req);
515 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
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.
587 if (!nfs_dirty_request(req)
589 || end < req->wb_offset)
592 if (nfs_set_page_tag_locked(req))
595 /* The request is locked, so wait and then retry */
596 spin_unlock(&inode->i_lock);
597 error = nfs_wait_on_request(req);
598 nfs_release_request(req);
601 spin_lock(&inode->i_lock);
604 /* Okay, the request matches. Update the region */
605 if (offset < req->wb_offset) {
606 req->wb_offset = offset;
607 req->wb_pgbase = offset;
610 req->wb_bytes = end - req->wb_offset;
612 req->wb_bytes = rqend - req->wb_offset;
614 spin_unlock(&inode->i_lock);
617 spin_unlock(&inode->i_lock);
618 nfs_release_request(req);
619 error = nfs_wb_page(inode, page);
621 return ERR_PTR(error);
625 * Try to update an existing write request, or create one if there is none.
627 * Note: Should always be called with the Page Lock held to prevent races
628 * if we have to add a new request. Also assumes that the caller has
629 * already called nfs_flush_incompatible() if necessary.
631 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
632 struct page *page, unsigned int offset, unsigned int bytes)
634 struct inode *inode = page->mapping->host;
635 struct nfs_page *req;
638 req = nfs_try_to_update_request(inode, page, offset, bytes);
641 req = nfs_create_request(ctx, inode, page, offset, bytes);
644 error = nfs_inode_add_request(inode, req);
646 nfs_release_request(req);
647 req = ERR_PTR(error);
653 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
654 unsigned int offset, unsigned int count)
656 struct nfs_page *req;
658 req = nfs_setup_write_request(ctx, page, offset, count);
661 /* Update file length */
662 nfs_grow_file(page, offset, count);
663 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
664 nfs_clear_page_tag_locked(req);
668 int nfs_flush_incompatible(struct file *file, struct page *page)
670 struct nfs_open_context *ctx = nfs_file_open_context(file);
671 struct nfs_page *req;
672 int do_flush, status;
674 * Look for a request corresponding to this page. If there
675 * is one, and it belongs to another file, we flush it out
676 * before we try to copy anything into the page. Do this
677 * due to the lack of an ACCESS-type call in NFSv2.
678 * Also do the same if we find a request from an existing
682 req = nfs_page_find_request(page);
685 do_flush = req->wb_page != page || req->wb_context != ctx
686 || !nfs_dirty_request(req);
687 nfs_release_request(req);
690 status = nfs_wb_page(page->mapping->host, page);
691 } while (status == 0);
696 * If the page cache is marked as unsafe or invalid, then we can't rely on
697 * the PageUptodate() flag. In this case, we will need to turn off
698 * write optimisations that depend on the page contents being correct.
700 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
702 return PageUptodate(page) &&
703 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
707 * Update and possibly write a cached page of an NFS file.
709 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
710 * things with a page scheduled for an RPC call (e.g. invalidate it).
712 int nfs_updatepage(struct file *file, struct page *page,
713 unsigned int offset, unsigned int count)
715 struct nfs_open_context *ctx = nfs_file_open_context(file);
716 struct inode *inode = page->mapping->host;
719 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
721 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
722 file->f_path.dentry->d_parent->d_name.name,
723 file->f_path.dentry->d_name.name, count,
724 (long long)(page_offset(page) + offset));
726 /* If we're not using byte range locks, and we know the page
727 * is up to date, it may be more efficient to extend the write
728 * to cover the entire page in order to avoid fragmentation
731 if (nfs_write_pageuptodate(page, inode) &&
732 inode->i_flock == NULL &&
733 !(file->f_flags & O_SYNC)) {
734 count = max(count + offset, nfs_page_length(page));
738 status = nfs_writepage_setup(ctx, page, offset, count);
740 nfs_set_pageerror(page);
742 __set_page_dirty_nobuffers(page);
744 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
745 status, (long long)i_size_read(inode));
749 static void nfs_writepage_release(struct nfs_page *req)
752 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
753 nfs_end_page_writeback(req->wb_page);
754 nfs_inode_remove_request(req);
756 nfs_end_page_writeback(req->wb_page);
757 nfs_clear_page_tag_locked(req);
760 static int flush_task_priority(int how)
762 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
764 return RPC_PRIORITY_HIGH;
766 return RPC_PRIORITY_LOW;
768 return RPC_PRIORITY_NORMAL;
772 * Set up the argument/result storage required for the RPC call.
774 static int nfs_write_rpcsetup(struct nfs_page *req,
775 struct nfs_write_data *data,
776 const struct rpc_call_ops *call_ops,
777 unsigned int count, unsigned int offset,
780 struct inode *inode = req->wb_context->path.dentry->d_inode;
781 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
782 int priority = flush_task_priority(how);
783 struct rpc_task *task;
784 struct rpc_message msg = {
785 .rpc_argp = &data->args,
786 .rpc_resp = &data->res,
787 .rpc_cred = req->wb_context->cred,
789 struct rpc_task_setup task_setup_data = {
790 .rpc_client = NFS_CLIENT(inode),
793 .callback_ops = call_ops,
794 .callback_data = data,
795 .workqueue = nfsiod_workqueue,
797 .priority = priority,
800 /* Set up the RPC argument and reply structs
801 * NB: take care not to mess about with data->commit et al. */
804 data->inode = inode = req->wb_context->path.dentry->d_inode;
805 data->cred = msg.rpc_cred;
807 data->args.fh = NFS_FH(inode);
808 data->args.offset = req_offset(req) + offset;
809 data->args.pgbase = req->wb_pgbase + offset;
810 data->args.pages = data->pagevec;
811 data->args.count = count;
812 data->args.context = get_nfs_open_context(req->wb_context);
813 data->args.stable = NFS_UNSTABLE;
814 if (how & FLUSH_STABLE) {
815 data->args.stable = NFS_DATA_SYNC;
816 if (!NFS_I(inode)->ncommit)
817 data->args.stable = NFS_FILE_SYNC;
820 data->res.fattr = &data->fattr;
821 data->res.count = count;
822 data->res.verf = &data->verf;
823 nfs_fattr_init(&data->fattr);
825 /* Set up the initial task struct. */
826 NFS_PROTO(inode)->write_setup(data, &msg);
828 dprintk("NFS: %5u initiated write call "
829 "(req %s/%lld, %u bytes @ offset %llu)\n",
832 (long long)NFS_FILEID(inode),
834 (unsigned long long)data->args.offset);
836 task = rpc_run_task(&task_setup_data);
838 return PTR_ERR(task);
843 /* If a nfs_flush_* function fails, it should remove reqs from @head and
844 * call this on each, which will prepare them to be retried on next
845 * writeback using standard nfs.
847 static void nfs_redirty_request(struct nfs_page *req)
849 nfs_mark_request_dirty(req);
850 nfs_end_page_writeback(req->wb_page);
851 nfs_clear_page_tag_locked(req);
855 * Generate multiple small requests to write out a single
856 * contiguous dirty area on one page.
858 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
860 struct nfs_page *req = nfs_list_entry(head->next);
861 struct page *page = req->wb_page;
862 struct nfs_write_data *data;
863 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
869 nfs_list_remove_request(req);
873 size_t len = min(nbytes, wsize);
875 data = nfs_writedata_alloc(1);
878 list_add(&data->pages, &list);
881 } while (nbytes != 0);
882 atomic_set(&req->wb_complete, requests);
884 ClearPageError(page);
890 data = list_entry(list.next, struct nfs_write_data, pages);
891 list_del_init(&data->pages);
893 data->pagevec[0] = page;
897 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
903 } while (nbytes != 0);
908 while (!list_empty(&list)) {
909 data = list_entry(list.next, struct nfs_write_data, pages);
910 list_del(&data->pages);
911 nfs_writedata_release(data);
913 nfs_redirty_request(req);
918 * Create an RPC task for the given write request and kick it.
919 * The page must have been locked by the caller.
921 * It may happen that the page we're passed is not marked dirty.
922 * This is the case if nfs_updatepage detects a conflicting request
923 * that has been written but not committed.
925 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
927 struct nfs_page *req;
929 struct nfs_write_data *data;
931 data = nfs_writedata_alloc(npages);
935 pages = data->pagevec;
936 while (!list_empty(head)) {
937 req = nfs_list_entry(head->next);
938 nfs_list_remove_request(req);
939 nfs_list_add_request(req, &data->pages);
940 ClearPageError(req->wb_page);
941 *pages++ = req->wb_page;
943 req = nfs_list_entry(data->pages.next);
945 /* Set up the argument struct */
946 return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
948 while (!list_empty(head)) {
949 req = nfs_list_entry(head->next);
950 nfs_list_remove_request(req);
951 nfs_redirty_request(req);
956 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
957 struct inode *inode, int ioflags)
959 size_t wsize = NFS_SERVER(inode)->wsize;
961 if (wsize < PAGE_CACHE_SIZE)
962 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
964 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
968 * Handle a write reply that flushed part of a page.
970 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
972 struct nfs_write_data *data = calldata;
974 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
976 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
978 NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
979 data->req->wb_bytes, (long long)req_offset(data->req));
981 nfs_writeback_done(task, data);
984 static void nfs_writeback_release_partial(void *calldata)
986 struct nfs_write_data *data = calldata;
987 struct nfs_page *req = data->req;
988 struct page *page = req->wb_page;
989 int status = data->task.tk_status;
992 nfs_set_pageerror(page);
993 nfs_context_set_write_error(req->wb_context, status);
994 dprintk(", error = %d\n", status);
998 if (nfs_write_need_commit(data)) {
999 struct inode *inode = page->mapping->host;
1001 spin_lock(&inode->i_lock);
1002 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1003 /* Do nothing we need to resend the writes */
1004 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1005 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1006 dprintk(" defer commit\n");
1007 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1008 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1009 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1010 dprintk(" server reboot detected\n");
1012 spin_unlock(&inode->i_lock);
1017 if (atomic_dec_and_test(&req->wb_complete))
1018 nfs_writepage_release(req);
1019 nfs_writedata_release(calldata);
1022 static const struct rpc_call_ops nfs_write_partial_ops = {
1023 .rpc_call_done = nfs_writeback_done_partial,
1024 .rpc_release = nfs_writeback_release_partial,
1028 * Handle a write reply that flushes a whole page.
1030 * FIXME: There is an inherent race with invalidate_inode_pages and
1031 * writebacks since the page->count is kept > 1 for as long
1032 * as the page has a write request pending.
1034 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1036 struct nfs_write_data *data = calldata;
1038 nfs_writeback_done(task, data);
1041 static void nfs_writeback_release_full(void *calldata)
1043 struct nfs_write_data *data = calldata;
1044 int status = data->task.tk_status;
1046 /* Update attributes as result of writeback. */
1047 while (!list_empty(&data->pages)) {
1048 struct nfs_page *req = nfs_list_entry(data->pages.next);
1049 struct page *page = req->wb_page;
1051 nfs_list_remove_request(req);
1053 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1055 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1056 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1058 (long long)req_offset(req));
1061 nfs_set_pageerror(page);
1062 nfs_context_set_write_error(req->wb_context, status);
1063 dprintk(", error = %d\n", status);
1064 goto remove_request;
1067 if (nfs_write_need_commit(data)) {
1068 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1069 nfs_mark_request_commit(req);
1070 nfs_end_page_writeback(page);
1071 dprintk(" marked for commit\n");
1076 nfs_end_page_writeback(page);
1077 nfs_inode_remove_request(req);
1079 nfs_clear_page_tag_locked(req);
1081 nfs_writedata_release(calldata);
1084 static const struct rpc_call_ops nfs_write_full_ops = {
1085 .rpc_call_done = nfs_writeback_done_full,
1086 .rpc_release = nfs_writeback_release_full,
1091 * This function is called when the WRITE call is complete.
1093 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1095 struct nfs_writeargs *argp = &data->args;
1096 struct nfs_writeres *resp = &data->res;
1099 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1100 task->tk_pid, task->tk_status);
1103 * ->write_done will attempt to use post-op attributes to detect
1104 * conflicting writes by other clients. A strict interpretation
1105 * of close-to-open would allow us to continue caching even if
1106 * another writer had changed the file, but some applications
1107 * depend on tighter cache coherency when writing.
1109 status = NFS_PROTO(data->inode)->write_done(task, data);
1112 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1114 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1115 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1116 /* We tried a write call, but the server did not
1117 * commit data to stable storage even though we
1119 * Note: There is a known bug in Tru64 < 5.0 in which
1120 * the server reports NFS_DATA_SYNC, but performs
1121 * NFS_FILE_SYNC. We therefore implement this checking
1122 * as a dprintk() in order to avoid filling syslog.
1124 static unsigned long complain;
1126 if (time_before(complain, jiffies)) {
1127 dprintk("NFS: faulty NFS server %s:"
1128 " (committed = %d) != (stable = %d)\n",
1129 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1130 resp->verf->committed, argp->stable);
1131 complain = jiffies + 300 * HZ;
1135 /* Is this a short write? */
1136 if (task->tk_status >= 0 && resp->count < argp->count) {
1137 static unsigned long complain;
1139 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1141 /* Has the server at least made some progress? */
1142 if (resp->count != 0) {
1143 /* Was this an NFSv2 write or an NFSv3 stable write? */
1144 if (resp->verf->committed != NFS_UNSTABLE) {
1145 /* Resend from where the server left off */
1146 argp->offset += resp->count;
1147 argp->pgbase += resp->count;
1148 argp->count -= resp->count;
1150 /* Resend as a stable write in order to avoid
1151 * headaches in the case of a server crash.
1153 argp->stable = NFS_FILE_SYNC;
1155 rpc_restart_call(task);
1158 if (time_before(complain, jiffies)) {
1160 "NFS: Server wrote zero bytes, expected %u.\n",
1162 complain = jiffies + 300 * HZ;
1164 /* Can't do anything about it except throw an error. */
1165 task->tk_status = -EIO;
1171 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1172 void nfs_commitdata_release(void *data)
1174 struct nfs_write_data *wdata = data;
1176 put_nfs_open_context(wdata->args.context);
1177 nfs_commit_free(wdata);
1181 * Set up the argument/result storage required for the RPC call.
1183 static int nfs_commit_rpcsetup(struct list_head *head,
1184 struct nfs_write_data *data,
1187 struct nfs_page *first = nfs_list_entry(head->next);
1188 struct inode *inode = first->wb_context->path.dentry->d_inode;
1189 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1190 int priority = flush_task_priority(how);
1191 struct rpc_task *task;
1192 struct rpc_message msg = {
1193 .rpc_argp = &data->args,
1194 .rpc_resp = &data->res,
1195 .rpc_cred = first->wb_context->cred,
1197 struct rpc_task_setup task_setup_data = {
1198 .task = &data->task,
1199 .rpc_client = NFS_CLIENT(inode),
1200 .rpc_message = &msg,
1201 .callback_ops = &nfs_commit_ops,
1202 .callback_data = data,
1203 .workqueue = nfsiod_workqueue,
1205 .priority = priority,
1208 /* Set up the RPC argument and reply structs
1209 * NB: take care not to mess about with data->commit et al. */
1211 list_splice_init(head, &data->pages);
1213 data->inode = inode;
1214 data->cred = msg.rpc_cred;
1216 data->args.fh = NFS_FH(data->inode);
1217 /* Note: we always request a commit of the entire inode */
1218 data->args.offset = 0;
1219 data->args.count = 0;
1220 data->args.context = get_nfs_open_context(first->wb_context);
1221 data->res.count = 0;
1222 data->res.fattr = &data->fattr;
1223 data->res.verf = &data->verf;
1224 nfs_fattr_init(&data->fattr);
1226 /* Set up the initial task struct. */
1227 NFS_PROTO(inode)->commit_setup(data, &msg);
1229 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1231 task = rpc_run_task(&task_setup_data);
1233 return PTR_ERR(task);
1239 * Commit dirty pages
1242 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1244 struct nfs_write_data *data;
1245 struct nfs_page *req;
1247 data = nfs_commitdata_alloc();
1252 /* Set up the argument struct */
1253 return nfs_commit_rpcsetup(head, data, how);
1255 while (!list_empty(head)) {
1256 req = nfs_list_entry(head->next);
1257 nfs_list_remove_request(req);
1258 nfs_mark_request_commit(req);
1259 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1260 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1262 nfs_clear_page_tag_locked(req);
1268 * COMMIT call returned
1270 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1272 struct nfs_write_data *data = calldata;
1274 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1275 task->tk_pid, task->tk_status);
1277 /* Call the NFS version-specific code */
1278 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1282 static void nfs_commit_release(void *calldata)
1284 struct nfs_write_data *data = calldata;
1285 struct nfs_page *req;
1286 int status = data->task.tk_status;
1288 while (!list_empty(&data->pages)) {
1289 req = nfs_list_entry(data->pages.next);
1290 nfs_list_remove_request(req);
1291 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1292 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1293 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1296 dprintk("NFS: commit (%s/%lld %d@%lld)",
1297 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1298 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1300 (long long)req_offset(req));
1302 nfs_context_set_write_error(req->wb_context, status);
1303 nfs_inode_remove_request(req);
1304 dprintk(", error = %d\n", status);
1308 /* Okay, COMMIT succeeded, apparently. Check the verifier
1309 * returned by the server against all stored verfs. */
1310 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1311 /* We have a match */
1312 nfs_inode_remove_request(req);
1316 /* We have a mismatch. Write the page again */
1317 dprintk(" mismatch\n");
1318 nfs_mark_request_dirty(req);
1320 nfs_clear_page_tag_locked(req);
1322 nfs_commitdata_release(calldata);
1325 static const struct rpc_call_ops nfs_commit_ops = {
1326 .rpc_call_done = nfs_commit_done,
1327 .rpc_release = nfs_commit_release,
1330 int nfs_commit_inode(struct inode *inode, int how)
1335 spin_lock(&inode->i_lock);
1336 res = nfs_scan_commit(inode, &head, 0, 0);
1337 spin_unlock(&inode->i_lock);
1339 int error = nfs_commit_list(inode, &head, how);
1346 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1352 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1354 struct inode *inode = mapping->host;
1355 pgoff_t idx_start, idx_end;
1356 unsigned int npages = 0;
1358 int nocommit = how & FLUSH_NOCOMMIT;
1362 if (wbc->range_cyclic)
1365 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1366 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1367 if (idx_end > idx_start) {
1368 pgoff_t l_npages = 1 + idx_end - idx_start;
1370 if (sizeof(npages) != sizeof(l_npages) &&
1371 (pgoff_t)npages != l_npages)
1375 how &= ~FLUSH_NOCOMMIT;
1376 spin_lock(&inode->i_lock);
1378 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1383 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1386 if (how & FLUSH_INVALIDATE) {
1387 spin_unlock(&inode->i_lock);
1388 nfs_cancel_commit_list(&head);
1390 spin_lock(&inode->i_lock);
1393 pages += nfs_scan_commit(inode, &head, 0, 0);
1394 spin_unlock(&inode->i_lock);
1395 ret = nfs_commit_list(inode, &head, how);
1396 spin_lock(&inode->i_lock);
1399 spin_unlock(&inode->i_lock);
1403 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1407 ret = nfs_writepages(mapping, wbc);
1410 ret = nfs_sync_mapping_wait(mapping, wbc, how);
1415 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1419 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1420 static int nfs_write_mapping(struct address_space *mapping, int how)
1422 struct writeback_control wbc = {
1423 .bdi = mapping->backing_dev_info,
1424 .sync_mode = WB_SYNC_NONE,
1425 .nr_to_write = LONG_MAX,
1426 .for_writepages = 1,
1431 ret = __nfs_write_mapping(mapping, &wbc, how);
1434 wbc.sync_mode = WB_SYNC_ALL;
1435 return __nfs_write_mapping(mapping, &wbc, how);
1439 * flush the inode to disk.
1441 int nfs_wb_all(struct inode *inode)
1443 return nfs_write_mapping(inode->i_mapping, 0);
1446 int nfs_wb_nocommit(struct inode *inode)
1448 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1451 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1453 struct nfs_page *req;
1454 loff_t range_start = page_offset(page);
1455 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1456 struct writeback_control wbc = {
1457 .bdi = page->mapping->backing_dev_info,
1458 .sync_mode = WB_SYNC_ALL,
1459 .nr_to_write = LONG_MAX,
1460 .range_start = range_start,
1461 .range_end = range_end,
1465 BUG_ON(!PageLocked(page));
1467 req = nfs_page_find_request(page);
1470 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1471 nfs_release_request(req);
1474 if (nfs_lock_request_dontget(req)) {
1475 nfs_inode_remove_request(req);
1477 * In case nfs_inode_remove_request has marked the
1478 * page as being dirty
1480 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1481 nfs_unlock_request(req);
1484 ret = nfs_wait_on_request(req);
1488 if (!PagePrivate(page))
1490 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1495 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1498 loff_t range_start = page_offset(page);
1499 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1500 struct writeback_control wbc = {
1501 .bdi = page->mapping->backing_dev_info,
1502 .sync_mode = WB_SYNC_ALL,
1503 .nr_to_write = LONG_MAX,
1504 .range_start = range_start,
1505 .range_end = range_end,
1510 if (clear_page_dirty_for_io(page)) {
1511 ret = nfs_writepage_locked(page, &wbc);
1514 } else if (!PagePrivate(page))
1516 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1519 } while (PagePrivate(page));
1522 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1527 * Write back all requests on one page - we do this before reading it.
1529 int nfs_wb_page(struct inode *inode, struct page* page)
1531 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1534 int __init nfs_init_writepagecache(void)
1536 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1537 sizeof(struct nfs_write_data),
1538 0, SLAB_HWCACHE_ALIGN,
1540 if (nfs_wdata_cachep == NULL)
1543 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1545 if (nfs_wdata_mempool == NULL)
1548 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1550 if (nfs_commit_mempool == NULL)
1554 * NFS congestion size, scale with available memory.
1566 * This allows larger machines to have larger/more transfers.
1567 * Limit the default to 256M
1569 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1570 if (nfs_congestion_kb > 256*1024)
1571 nfs_congestion_kb = 256*1024;
1576 void nfs_destroy_writepagecache(void)
1578 mempool_destroy(nfs_commit_mempool);
1579 mempool_destroy(nfs_wdata_mempool);
1580 kmem_cache_destroy(nfs_wdata_cachep);