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>
16 #include <linux/migrate.h>
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
24 #include <asm/uaccess.h>
26 #include "delegation.h"
32 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
34 #define MIN_POOL_WRITE (32)
35 #define MIN_POOL_COMMIT (4)
38 * Local function declarations
40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
41 struct inode *inode, int ioflags);
42 static void nfs_redirty_request(struct nfs_page *req);
43 static const struct rpc_call_ops nfs_write_partial_ops;
44 static const struct rpc_call_ops nfs_write_full_ops;
45 static const struct rpc_call_ops nfs_commit_ops;
47 static struct kmem_cache *nfs_wdata_cachep;
48 static mempool_t *nfs_wdata_mempool;
49 static mempool_t *nfs_commit_mempool;
51 struct nfs_write_data *nfs_commitdata_alloc(void)
53 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
56 memset(p, 0, sizeof(*p));
57 INIT_LIST_HEAD(&p->pages);
62 void nfs_commit_free(struct nfs_write_data *p)
64 if (p && (p->pagevec != &p->page_array[0]))
66 mempool_free(p, nfs_commit_mempool);
69 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
71 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
74 memset(p, 0, sizeof(*p));
75 INIT_LIST_HEAD(&p->pages);
76 p->npages = pagecount;
77 if (pagecount <= ARRAY_SIZE(p->page_array))
78 p->pagevec = p->page_array;
80 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
82 mempool_free(p, nfs_wdata_mempool);
90 void nfs_writedata_free(struct nfs_write_data *p)
92 if (p && (p->pagevec != &p->page_array[0]))
94 mempool_free(p, nfs_wdata_mempool);
97 static void nfs_writedata_release(struct nfs_write_data *wdata)
99 put_nfs_open_context(wdata->args.context);
100 nfs_writedata_free(wdata);
103 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
107 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
110 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
112 struct nfs_page *req = NULL;
114 if (PagePrivate(page)) {
115 req = (struct nfs_page *)page_private(page);
117 kref_get(&req->wb_kref);
122 static struct nfs_page *nfs_page_find_request(struct page *page)
124 struct inode *inode = page->mapping->host;
125 struct nfs_page *req = NULL;
127 spin_lock(&inode->i_lock);
128 req = nfs_page_find_request_locked(page);
129 spin_unlock(&inode->i_lock);
133 /* Adjust the file length if we're writing beyond the end */
134 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
136 struct inode *inode = page->mapping->host;
140 spin_lock(&inode->i_lock);
141 i_size = i_size_read(inode);
142 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
143 if (i_size > 0 && page->index < end_index)
145 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
148 i_size_write(inode, end);
149 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
151 spin_unlock(&inode->i_lock);
154 /* A writeback failed: mark the page as bad, and invalidate the page cache */
155 static void nfs_set_pageerror(struct page *page)
158 nfs_zap_mapping(page->mapping->host, page->mapping);
161 /* We can set the PG_uptodate flag if we see that a write request
162 * covers the full page.
164 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
166 if (PageUptodate(page))
170 if (count != nfs_page_length(page))
172 SetPageUptodate(page);
175 static int wb_priority(struct writeback_control *wbc)
177 if (wbc->for_reclaim)
178 return FLUSH_HIGHPRI | FLUSH_STABLE;
179 if (wbc->for_kupdate || wbc->for_background)
185 * NFS congestion control
188 int nfs_congestion_kb;
190 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
191 #define NFS_CONGESTION_OFF_THRESH \
192 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
194 static int nfs_set_page_writeback(struct page *page)
196 int ret = test_set_page_writeback(page);
199 struct inode *inode = page->mapping->host;
200 struct nfs_server *nfss = NFS_SERVER(inode);
202 page_cache_get(page);
203 if (atomic_long_inc_return(&nfss->writeback) >
204 NFS_CONGESTION_ON_THRESH) {
205 set_bdi_congested(&nfss->backing_dev_info,
212 static void nfs_end_page_writeback(struct page *page)
214 struct inode *inode = page->mapping->host;
215 struct nfs_server *nfss = NFS_SERVER(inode);
217 end_page_writeback(page);
218 page_cache_release(page);
219 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
220 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
223 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
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);
234 if (nfs_set_page_tag_locked(req))
236 /* Note: If we hold the page lock, as is the case in nfs_writepage,
237 * then the call to nfs_set_page_tag_locked() will always
238 * succeed provided that someone hasn't already marked the
239 * request as dirty (in which case we don't care).
241 spin_unlock(&inode->i_lock);
243 ret = nfs_wait_on_request(req);
246 nfs_release_request(req);
249 spin_lock(&inode->i_lock);
251 spin_unlock(&inode->i_lock);
256 * Find an associated nfs write request, and prepare to flush it out
257 * May return an error if the user signalled nfs_wait_on_request().
259 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
260 struct page *page, bool nonblock)
262 struct nfs_page *req;
265 req = nfs_find_and_lock_request(page, nonblock);
272 ret = nfs_set_page_writeback(page);
274 BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
276 if (!nfs_pageio_add_request(pgio, req)) {
277 nfs_redirty_request(req);
278 ret = pgio->pg_error;
284 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
286 struct inode *inode = page->mapping->host;
289 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
290 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
292 nfs_pageio_cond_complete(pgio, page->index);
293 ret = nfs_page_async_flush(pgio, page,
294 wbc->sync_mode == WB_SYNC_NONE ||
295 wbc->nonblocking != 0);
296 if (ret == -EAGAIN) {
297 redirty_page_for_writepage(wbc, page);
304 * Write an mmapped page to the server.
306 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
308 struct nfs_pageio_descriptor pgio;
311 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
312 err = nfs_do_writepage(page, wbc, &pgio);
313 nfs_pageio_complete(&pgio);
316 if (pgio.pg_error < 0)
317 return pgio.pg_error;
321 int nfs_writepage(struct page *page, struct writeback_control *wbc)
325 ret = nfs_writepage_locked(page, wbc);
330 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
334 ret = nfs_do_writepage(page, wbc, data);
339 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
341 struct inode *inode = mapping->host;
342 unsigned long *bitlock = &NFS_I(inode)->flags;
343 struct nfs_pageio_descriptor pgio;
346 /* Stop dirtying of new pages while we sync */
347 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
348 nfs_wait_bit_killable, TASK_KILLABLE);
352 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
354 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
355 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
356 nfs_pageio_complete(&pgio);
358 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
359 smp_mb__after_clear_bit();
360 wake_up_bit(bitlock, NFS_INO_FLUSHING);
373 * Insert a write request into an inode
375 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
377 struct nfs_inode *nfsi = NFS_I(inode);
380 error = radix_tree_preload(GFP_NOFS);
384 /* Lock the request! */
385 nfs_lock_request_dontget(req);
387 spin_lock(&inode->i_lock);
388 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
392 if (nfs_have_delegation(inode, FMODE_WRITE))
395 SetPagePrivate(req->wb_page);
396 set_page_private(req->wb_page, (unsigned long)req);
398 kref_get(&req->wb_kref);
399 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
400 NFS_PAGE_TAG_LOCKED);
401 spin_unlock(&inode->i_lock);
402 radix_tree_preload_end();
408 * Remove a write request from an inode
410 static void nfs_inode_remove_request(struct nfs_page *req)
412 struct inode *inode = req->wb_context->path.dentry->d_inode;
413 struct nfs_inode *nfsi = NFS_I(inode);
415 BUG_ON (!NFS_WBACK_BUSY(req));
417 spin_lock(&inode->i_lock);
418 set_page_private(req->wb_page, 0);
419 ClearPagePrivate(req->wb_page);
420 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
423 spin_unlock(&inode->i_lock);
426 spin_unlock(&inode->i_lock);
427 nfs_clear_request(req);
428 nfs_release_request(req);
432 nfs_mark_request_dirty(struct nfs_page *req)
434 __set_page_dirty_nobuffers(req->wb_page);
435 __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
438 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
440 * Add a request to the inode's commit list.
443 nfs_mark_request_commit(struct nfs_page *req)
445 struct inode *inode = req->wb_context->path.dentry->d_inode;
446 struct nfs_inode *nfsi = NFS_I(inode);
448 spin_lock(&inode->i_lock);
449 set_bit(PG_CLEAN, &(req)->wb_flags);
450 radix_tree_tag_set(&nfsi->nfs_page_tree,
452 NFS_PAGE_TAG_COMMIT);
454 spin_unlock(&inode->i_lock);
455 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
456 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
457 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
461 nfs_clear_request_commit(struct nfs_page *req)
463 struct page *page = req->wb_page;
465 if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
466 dec_zone_page_state(page, NR_UNSTABLE_NFS);
467 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
474 int nfs_write_need_commit(struct nfs_write_data *data)
476 return data->verf.committed != NFS_FILE_SYNC;
480 int nfs_reschedule_unstable_write(struct nfs_page *req)
482 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
483 nfs_mark_request_commit(req);
486 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
487 nfs_mark_request_dirty(req);
494 nfs_mark_request_commit(struct nfs_page *req)
499 nfs_clear_request_commit(struct nfs_page *req)
505 int nfs_write_need_commit(struct nfs_write_data *data)
511 int nfs_reschedule_unstable_write(struct nfs_page *req)
517 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
519 nfs_need_commit(struct nfs_inode *nfsi)
521 return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
525 * nfs_scan_commit - Scan an inode for commit requests
526 * @inode: NFS inode to scan
527 * @dst: destination list
528 * @idx_start: lower bound of page->index to scan.
529 * @npages: idx_start + npages sets the upper bound to scan.
531 * Moves requests from the inode's 'commit' request list.
532 * The requests are *not* checked to ensure that they form a contiguous set.
535 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
537 struct nfs_inode *nfsi = NFS_I(inode);
540 if (!nfs_need_commit(nfsi))
543 ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
545 nfsi->ncommit -= ret;
546 if (nfs_need_commit(NFS_I(inode)))
547 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
551 static inline int nfs_need_commit(struct nfs_inode *nfsi)
556 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
563 * Search for an existing write request, and attempt to update
564 * it to reflect a new dirty region on a given page.
566 * If the attempt fails, then the existing request is flushed out
569 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
574 struct nfs_page *req;
579 if (!PagePrivate(page))
582 end = offset + bytes;
583 spin_lock(&inode->i_lock);
586 req = nfs_page_find_request_locked(page);
590 rqend = req->wb_offset + req->wb_bytes;
592 * Tell the caller to flush out the request if
593 * the offsets are non-contiguous.
594 * Note: nfs_flush_incompatible() will already
595 * have flushed out requests having wrong owners.
598 || end < req->wb_offset)
601 if (nfs_set_page_tag_locked(req))
604 /* The request is locked, so wait and then retry */
605 spin_unlock(&inode->i_lock);
606 error = nfs_wait_on_request(req);
607 nfs_release_request(req);
610 spin_lock(&inode->i_lock);
613 if (nfs_clear_request_commit(req) &&
614 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
615 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
616 NFS_I(inode)->ncommit--;
618 /* Okay, the request matches. Update the region */
619 if (offset < req->wb_offset) {
620 req->wb_offset = offset;
621 req->wb_pgbase = offset;
624 req->wb_bytes = end - req->wb_offset;
626 req->wb_bytes = rqend - req->wb_offset;
628 spin_unlock(&inode->i_lock);
631 spin_unlock(&inode->i_lock);
632 nfs_release_request(req);
633 error = nfs_wb_page(inode, page);
635 return ERR_PTR(error);
639 * Try to update an existing write request, or create one if there is none.
641 * Note: Should always be called with the Page Lock held to prevent races
642 * if we have to add a new request. Also assumes that the caller has
643 * already called nfs_flush_incompatible() if necessary.
645 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
646 struct page *page, unsigned int offset, unsigned int bytes)
648 struct inode *inode = page->mapping->host;
649 struct nfs_page *req;
652 req = nfs_try_to_update_request(inode, page, offset, bytes);
655 req = nfs_create_request(ctx, inode, page, offset, bytes);
658 error = nfs_inode_add_request(inode, req);
660 nfs_release_request(req);
661 req = ERR_PTR(error);
667 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
668 unsigned int offset, unsigned int count)
670 struct nfs_page *req;
672 req = nfs_setup_write_request(ctx, page, offset, count);
675 nfs_mark_request_dirty(req);
676 /* Update file length */
677 nfs_grow_file(page, offset, count);
678 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
679 nfs_mark_request_dirty(req);
680 nfs_clear_page_tag_locked(req);
684 int nfs_flush_incompatible(struct file *file, struct page *page)
686 struct nfs_open_context *ctx = nfs_file_open_context(file);
687 struct nfs_page *req;
688 int do_flush, status;
690 * Look for a request corresponding to this page. If there
691 * is one, and it belongs to another file, we flush it out
692 * before we try to copy anything into the page. Do this
693 * due to the lack of an ACCESS-type call in NFSv2.
694 * Also do the same if we find a request from an existing
698 req = nfs_page_find_request(page);
701 do_flush = req->wb_page != page || req->wb_context != ctx ||
702 req->wb_lock_context->lockowner != current->files ||
703 req->wb_lock_context->pid != current->tgid;
704 nfs_release_request(req);
707 status = nfs_wb_page(page->mapping->host, page);
708 } while (status == 0);
713 * If the page cache is marked as unsafe or invalid, then we can't rely on
714 * the PageUptodate() flag. In this case, we will need to turn off
715 * write optimisations that depend on the page contents being correct.
717 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
719 return PageUptodate(page) &&
720 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
724 * Update and possibly write a cached page of an NFS file.
726 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
727 * things with a page scheduled for an RPC call (e.g. invalidate it).
729 int nfs_updatepage(struct file *file, struct page *page,
730 unsigned int offset, unsigned int count)
732 struct nfs_open_context *ctx = nfs_file_open_context(file);
733 struct inode *inode = page->mapping->host;
736 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
738 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
739 file->f_path.dentry->d_parent->d_name.name,
740 file->f_path.dentry->d_name.name, count,
741 (long long)(page_offset(page) + offset));
743 /* If we're not using byte range locks, and we know the page
744 * is up to date, it may be more efficient to extend the write
745 * to cover the entire page in order to avoid fragmentation
748 if (nfs_write_pageuptodate(page, inode) &&
749 inode->i_flock == NULL &&
750 !(file->f_flags & O_DSYNC)) {
751 count = max(count + offset, nfs_page_length(page));
755 status = nfs_writepage_setup(ctx, page, offset, count);
757 nfs_set_pageerror(page);
759 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
760 status, (long long)i_size_read(inode));
764 static void nfs_writepage_release(struct nfs_page *req)
766 struct page *page = req->wb_page;
768 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req))
769 nfs_inode_remove_request(req);
770 nfs_clear_page_tag_locked(req);
771 nfs_end_page_writeback(page);
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 priority = flush_task_priority(how);
796 struct rpc_task *task;
797 struct rpc_message msg = {
798 .rpc_argp = &data->args,
799 .rpc_resp = &data->res,
800 .rpc_cred = req->wb_context->cred,
802 struct rpc_task_setup task_setup_data = {
803 .rpc_client = NFS_CLIENT(inode),
806 .callback_ops = call_ops,
807 .callback_data = data,
808 .workqueue = nfsiod_workqueue,
809 .flags = RPC_TASK_ASYNC,
810 .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.lock_context = req->wb_lock_context;
828 data->args.stable = NFS_UNSTABLE;
829 if (how & FLUSH_STABLE) {
830 data->args.stable = NFS_DATA_SYNC;
831 if (!nfs_need_commit(NFS_I(inode)))
832 data->args.stable = NFS_FILE_SYNC;
835 data->res.fattr = &data->fattr;
836 data->res.count = count;
837 data->res.verf = &data->verf;
838 nfs_fattr_init(&data->fattr);
840 /* Set up the initial task struct. */
841 NFS_PROTO(inode)->write_setup(data, &msg);
843 dprintk("NFS: %5u initiated write call "
844 "(req %s/%lld, %u bytes @ offset %llu)\n",
847 (long long)NFS_FILEID(inode),
849 (unsigned long long)data->args.offset);
851 task = rpc_run_task(&task_setup_data);
856 if (how & FLUSH_SYNC) {
857 ret = rpc_wait_for_completion_task(task);
859 ret = task->tk_status;
866 /* If a nfs_flush_* function fails, it should remove reqs from @head and
867 * call this on each, which will prepare them to be retried on next
868 * writeback using standard nfs.
870 static void nfs_redirty_request(struct nfs_page *req)
872 struct page *page = req->wb_page;
874 nfs_mark_request_dirty(req);
875 nfs_clear_page_tag_locked(req);
876 nfs_end_page_writeback(page);
880 * Generate multiple small requests to write out a single
881 * contiguous dirty area on one page.
883 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
885 struct nfs_page *req = nfs_list_entry(head->next);
886 struct page *page = req->wb_page;
887 struct nfs_write_data *data;
888 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
894 nfs_list_remove_request(req);
898 size_t len = min(nbytes, wsize);
900 data = nfs_writedata_alloc(1);
903 list_add(&data->pages, &list);
906 } while (nbytes != 0);
907 atomic_set(&req->wb_complete, requests);
909 ClearPageError(page);
915 data = list_entry(list.next, struct nfs_write_data, pages);
916 list_del_init(&data->pages);
918 data->pagevec[0] = page;
922 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
928 } while (nbytes != 0);
933 while (!list_empty(&list)) {
934 data = list_entry(list.next, struct nfs_write_data, pages);
935 list_del(&data->pages);
936 nfs_writedata_release(data);
938 nfs_redirty_request(req);
943 * Create an RPC task for the given write request and kick it.
944 * The page must have been locked by the caller.
946 * It may happen that the page we're passed is not marked dirty.
947 * This is the case if nfs_updatepage detects a conflicting request
948 * that has been written but not committed.
950 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
952 struct nfs_page *req;
954 struct nfs_write_data *data;
956 data = nfs_writedata_alloc(npages);
960 pages = data->pagevec;
961 while (!list_empty(head)) {
962 req = nfs_list_entry(head->next);
963 nfs_list_remove_request(req);
964 nfs_list_add_request(req, &data->pages);
965 ClearPageError(req->wb_page);
966 *pages++ = req->wb_page;
968 req = nfs_list_entry(data->pages.next);
970 /* Set up the argument struct */
971 return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
973 while (!list_empty(head)) {
974 req = nfs_list_entry(head->next);
975 nfs_list_remove_request(req);
976 nfs_redirty_request(req);
981 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
982 struct inode *inode, int ioflags)
984 size_t wsize = NFS_SERVER(inode)->wsize;
986 if (wsize < PAGE_CACHE_SIZE)
987 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
989 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
993 * Handle a write reply that flushed part of a page.
995 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
997 struct nfs_write_data *data = calldata;
999 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1001 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1003 NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1004 data->req->wb_bytes, (long long)req_offset(data->req));
1006 nfs_writeback_done(task, data);
1009 static void nfs_writeback_release_partial(void *calldata)
1011 struct nfs_write_data *data = calldata;
1012 struct nfs_page *req = data->req;
1013 struct page *page = req->wb_page;
1014 int status = data->task.tk_status;
1017 nfs_set_pageerror(page);
1018 nfs_context_set_write_error(req->wb_context, status);
1019 dprintk(", error = %d\n", status);
1023 if (nfs_write_need_commit(data)) {
1024 struct inode *inode = page->mapping->host;
1026 spin_lock(&inode->i_lock);
1027 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1028 /* Do nothing we need to resend the writes */
1029 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1030 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1031 dprintk(" defer commit\n");
1032 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1033 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1034 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1035 dprintk(" server reboot detected\n");
1037 spin_unlock(&inode->i_lock);
1042 if (atomic_dec_and_test(&req->wb_complete))
1043 nfs_writepage_release(req);
1044 nfs_writedata_release(calldata);
1047 #if defined(CONFIG_NFS_V4_1)
1048 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1050 struct nfs_write_data *data = calldata;
1052 if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1053 &data->args.seq_args,
1054 &data->res.seq_res, 1, task))
1056 rpc_call_start(task);
1058 #endif /* CONFIG_NFS_V4_1 */
1060 static const struct rpc_call_ops nfs_write_partial_ops = {
1061 #if defined(CONFIG_NFS_V4_1)
1062 .rpc_call_prepare = nfs_write_prepare,
1063 #endif /* CONFIG_NFS_V4_1 */
1064 .rpc_call_done = nfs_writeback_done_partial,
1065 .rpc_release = nfs_writeback_release_partial,
1069 * Handle a write reply that flushes a whole page.
1071 * FIXME: There is an inherent race with invalidate_inode_pages and
1072 * writebacks since the page->count is kept > 1 for as long
1073 * as the page has a write request pending.
1075 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1077 struct nfs_write_data *data = calldata;
1079 nfs_writeback_done(task, data);
1082 static void nfs_writeback_release_full(void *calldata)
1084 struct nfs_write_data *data = calldata;
1085 int status = data->task.tk_status;
1087 /* Update attributes as result of writeback. */
1088 while (!list_empty(&data->pages)) {
1089 struct nfs_page *req = nfs_list_entry(data->pages.next);
1090 struct page *page = req->wb_page;
1092 nfs_list_remove_request(req);
1094 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1096 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1097 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1099 (long long)req_offset(req));
1102 nfs_set_pageerror(page);
1103 nfs_context_set_write_error(req->wb_context, status);
1104 dprintk(", error = %d\n", status);
1105 goto remove_request;
1108 if (nfs_write_need_commit(data)) {
1109 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1110 nfs_mark_request_commit(req);
1111 dprintk(" marked for commit\n");
1116 nfs_inode_remove_request(req);
1118 nfs_clear_page_tag_locked(req);
1119 nfs_end_page_writeback(page);
1121 nfs_writedata_release(calldata);
1124 static const struct rpc_call_ops nfs_write_full_ops = {
1125 #if defined(CONFIG_NFS_V4_1)
1126 .rpc_call_prepare = nfs_write_prepare,
1127 #endif /* CONFIG_NFS_V4_1 */
1128 .rpc_call_done = nfs_writeback_done_full,
1129 .rpc_release = nfs_writeback_release_full,
1134 * This function is called when the WRITE call is complete.
1136 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1138 struct nfs_writeargs *argp = &data->args;
1139 struct nfs_writeres *resp = &data->res;
1140 struct nfs_server *server = NFS_SERVER(data->inode);
1143 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1144 task->tk_pid, task->tk_status);
1147 * ->write_done will attempt to use post-op attributes to detect
1148 * conflicting writes by other clients. A strict interpretation
1149 * of close-to-open would allow us to continue caching even if
1150 * another writer had changed the file, but some applications
1151 * depend on tighter cache coherency when writing.
1153 status = NFS_PROTO(data->inode)->write_done(task, data);
1156 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1158 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1159 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1160 /* We tried a write call, but the server did not
1161 * commit data to stable storage even though we
1163 * Note: There is a known bug in Tru64 < 5.0 in which
1164 * the server reports NFS_DATA_SYNC, but performs
1165 * NFS_FILE_SYNC. We therefore implement this checking
1166 * as a dprintk() in order to avoid filling syslog.
1168 static unsigned long complain;
1170 if (time_before(complain, jiffies)) {
1171 dprintk("NFS: faulty NFS server %s:"
1172 " (committed = %d) != (stable = %d)\n",
1173 server->nfs_client->cl_hostname,
1174 resp->verf->committed, argp->stable);
1175 complain = jiffies + 300 * HZ;
1179 /* Is this a short write? */
1180 if (task->tk_status >= 0 && resp->count < argp->count) {
1181 static unsigned long complain;
1183 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1185 /* Has the server at least made some progress? */
1186 if (resp->count != 0) {
1187 /* Was this an NFSv2 write or an NFSv3 stable write? */
1188 if (resp->verf->committed != NFS_UNSTABLE) {
1189 /* Resend from where the server left off */
1190 argp->offset += resp->count;
1191 argp->pgbase += resp->count;
1192 argp->count -= resp->count;
1194 /* Resend as a stable write in order to avoid
1195 * headaches in the case of a server crash.
1197 argp->stable = NFS_FILE_SYNC;
1199 nfs_restart_rpc(task, server->nfs_client);
1202 if (time_before(complain, jiffies)) {
1204 "NFS: Server wrote zero bytes, expected %u.\n",
1206 complain = jiffies + 300 * HZ;
1208 /* Can't do anything about it except throw an error. */
1209 task->tk_status = -EIO;
1215 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1216 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1218 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1220 if (may_wait && !out_of_line_wait_on_bit_lock(&nfsi->flags,
1221 NFS_INO_COMMIT, nfs_wait_bit_killable,
1227 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1229 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1230 smp_mb__after_clear_bit();
1231 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1235 static void nfs_commitdata_release(void *data)
1237 struct nfs_write_data *wdata = data;
1239 put_nfs_open_context(wdata->args.context);
1240 nfs_commit_free(wdata);
1244 * Set up the argument/result storage required for the RPC call.
1246 static int nfs_commit_rpcsetup(struct list_head *head,
1247 struct nfs_write_data *data,
1250 struct nfs_page *first = nfs_list_entry(head->next);
1251 struct inode *inode = first->wb_context->path.dentry->d_inode;
1252 int priority = flush_task_priority(how);
1253 struct rpc_task *task;
1254 struct rpc_message msg = {
1255 .rpc_argp = &data->args,
1256 .rpc_resp = &data->res,
1257 .rpc_cred = first->wb_context->cred,
1259 struct rpc_task_setup task_setup_data = {
1260 .task = &data->task,
1261 .rpc_client = NFS_CLIENT(inode),
1262 .rpc_message = &msg,
1263 .callback_ops = &nfs_commit_ops,
1264 .callback_data = data,
1265 .workqueue = nfsiod_workqueue,
1266 .flags = RPC_TASK_ASYNC,
1267 .priority = priority,
1270 /* Set up the RPC argument and reply structs
1271 * NB: take care not to mess about with data->commit et al. */
1273 list_splice_init(head, &data->pages);
1275 data->inode = inode;
1276 data->cred = msg.rpc_cred;
1278 data->args.fh = NFS_FH(data->inode);
1279 /* Note: we always request a commit of the entire inode */
1280 data->args.offset = 0;
1281 data->args.count = 0;
1282 data->args.context = get_nfs_open_context(first->wb_context);
1283 data->res.count = 0;
1284 data->res.fattr = &data->fattr;
1285 data->res.verf = &data->verf;
1286 nfs_fattr_init(&data->fattr);
1288 /* Set up the initial task struct. */
1289 NFS_PROTO(inode)->commit_setup(data, &msg);
1291 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1293 task = rpc_run_task(&task_setup_data);
1295 return PTR_ERR(task);
1301 * Commit dirty pages
1304 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1306 struct nfs_write_data *data;
1307 struct nfs_page *req;
1309 data = nfs_commitdata_alloc();
1314 /* Set up the argument struct */
1315 return nfs_commit_rpcsetup(head, data, how);
1317 while (!list_empty(head)) {
1318 req = nfs_list_entry(head->next);
1319 nfs_list_remove_request(req);
1320 nfs_mark_request_commit(req);
1321 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1322 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1324 nfs_clear_page_tag_locked(req);
1326 nfs_commit_clear_lock(NFS_I(inode));
1331 * COMMIT call returned
1333 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1335 struct nfs_write_data *data = calldata;
1337 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1338 task->tk_pid, task->tk_status);
1340 /* Call the NFS version-specific code */
1341 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1345 static void nfs_commit_release(void *calldata)
1347 struct nfs_write_data *data = calldata;
1348 struct nfs_page *req;
1349 int status = data->task.tk_status;
1351 while (!list_empty(&data->pages)) {
1352 req = nfs_list_entry(data->pages.next);
1353 nfs_list_remove_request(req);
1354 nfs_clear_request_commit(req);
1356 dprintk("NFS: commit (%s/%lld %d@%lld)",
1357 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1358 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1360 (long long)req_offset(req));
1362 nfs_context_set_write_error(req->wb_context, status);
1363 nfs_inode_remove_request(req);
1364 dprintk(", error = %d\n", status);
1368 /* Okay, COMMIT succeeded, apparently. Check the verifier
1369 * returned by the server against all stored verfs. */
1370 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1371 /* We have a match */
1372 nfs_inode_remove_request(req);
1376 /* We have a mismatch. Write the page again */
1377 dprintk(" mismatch\n");
1378 nfs_mark_request_dirty(req);
1380 nfs_clear_page_tag_locked(req);
1382 nfs_commit_clear_lock(NFS_I(data->inode));
1383 nfs_commitdata_release(calldata);
1386 static const struct rpc_call_ops nfs_commit_ops = {
1387 #if defined(CONFIG_NFS_V4_1)
1388 .rpc_call_prepare = nfs_write_prepare,
1389 #endif /* CONFIG_NFS_V4_1 */
1390 .rpc_call_done = nfs_commit_done,
1391 .rpc_release = nfs_commit_release,
1394 int nfs_commit_inode(struct inode *inode, int how)
1397 int may_wait = how & FLUSH_SYNC;
1400 if (!nfs_commit_set_lock(NFS_I(inode), may_wait))
1401 goto out_mark_dirty;
1402 spin_lock(&inode->i_lock);
1403 res = nfs_scan_commit(inode, &head, 0, 0);
1404 spin_unlock(&inode->i_lock);
1406 int error = nfs_commit_list(inode, &head, how);
1410 wait_on_bit(&NFS_I(inode)->flags, NFS_INO_COMMIT,
1411 nfs_wait_bit_killable,
1414 goto out_mark_dirty;
1416 nfs_commit_clear_lock(NFS_I(inode));
1418 /* Note: If we exit without ensuring that the commit is complete,
1419 * we must mark the inode as dirty. Otherwise, future calls to
1420 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1421 * that the data is on the disk.
1424 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1428 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1430 struct nfs_inode *nfsi = NFS_I(inode);
1431 int flags = FLUSH_SYNC;
1434 if (wbc->sync_mode == WB_SYNC_NONE) {
1435 /* Don't commit yet if this is a non-blocking flush and there
1436 * are a lot of outstanding writes for this mapping.
1438 if (nfsi->ncommit <= (nfsi->npages >> 1))
1439 goto out_mark_dirty;
1441 /* don't wait for the COMMIT response */
1445 ret = nfs_commit_inode(inode, flags);
1447 if (wbc->sync_mode == WB_SYNC_NONE) {
1448 if (ret < wbc->nr_to_write)
1449 wbc->nr_to_write -= ret;
1451 wbc->nr_to_write = 0;
1456 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1460 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1466 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1468 return nfs_commit_unstable_pages(inode, wbc);
1472 * flush the inode to disk.
1474 int nfs_wb_all(struct inode *inode)
1476 struct writeback_control wbc = {
1477 .sync_mode = WB_SYNC_ALL,
1478 .nr_to_write = LONG_MAX,
1480 .range_end = LLONG_MAX,
1483 return sync_inode(inode, &wbc);
1486 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1488 struct nfs_page *req;
1491 BUG_ON(!PageLocked(page));
1493 wait_on_page_writeback(page);
1494 req = nfs_page_find_request(page);
1497 if (nfs_lock_request_dontget(req)) {
1498 nfs_inode_remove_request(req);
1500 * In case nfs_inode_remove_request has marked the
1501 * page as being dirty
1503 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1504 nfs_unlock_request(req);
1507 ret = nfs_wait_on_request(req);
1508 nfs_release_request(req);
1516 * Write back all requests on one page - we do this before reading it.
1518 int nfs_wb_page(struct inode *inode, struct page *page)
1520 loff_t range_start = page_offset(page);
1521 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1522 struct writeback_control wbc = {
1523 .sync_mode = WB_SYNC_ALL,
1525 .range_start = range_start,
1526 .range_end = range_end,
1531 wait_on_page_writeback(page);
1532 if (clear_page_dirty_for_io(page)) {
1533 ret = nfs_writepage_locked(page, &wbc);
1538 if (!PagePrivate(page))
1540 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1549 #ifdef CONFIG_MIGRATION
1550 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1553 struct nfs_page *req;
1556 nfs_fscache_release_page(page, GFP_KERNEL);
1558 req = nfs_find_and_lock_request(page, false);
1563 ret = migrate_page(mapping, newpage, page);
1568 page_cache_get(newpage);
1569 spin_lock(&mapping->host->i_lock);
1570 req->wb_page = newpage;
1571 SetPagePrivate(newpage);
1572 set_page_private(newpage, (unsigned long)req);
1573 ClearPagePrivate(page);
1574 set_page_private(page, 0);
1575 spin_unlock(&mapping->host->i_lock);
1576 page_cache_release(page);
1578 nfs_clear_page_tag_locked(req);
1584 int __init nfs_init_writepagecache(void)
1586 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1587 sizeof(struct nfs_write_data),
1588 0, SLAB_HWCACHE_ALIGN,
1590 if (nfs_wdata_cachep == NULL)
1593 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1595 if (nfs_wdata_mempool == NULL)
1598 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1600 if (nfs_commit_mempool == NULL)
1604 * NFS congestion size, scale with available memory.
1616 * This allows larger machines to have larger/more transfers.
1617 * Limit the default to 256M
1619 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1620 if (nfs_congestion_kb > 256*1024)
1621 nfs_congestion_kb = 256*1024;
1626 void nfs_destroy_writepagecache(void)
1628 mempool_destroy(nfs_commit_mempool);
1629 mempool_destroy(nfs_wdata_mempool);
1630 kmem_cache_destroy(nfs_wdata_cachep);