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>
22 #include <linux/blkdev.h>
24 #include <asm/uaccess.h>
26 #include "delegation.h"
31 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
33 #define MIN_POOL_WRITE (32)
34 #define MIN_POOL_COMMIT (4)
37 * Local function declarations
39 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
40 struct inode *inode, int ioflags);
41 static void nfs_redirty_request(struct nfs_page *req);
42 static const struct rpc_call_ops nfs_write_partial_ops;
43 static const struct rpc_call_ops nfs_write_full_ops;
44 static const struct rpc_call_ops nfs_commit_ops;
46 static struct kmem_cache *nfs_wdata_cachep;
47 static mempool_t *nfs_wdata_mempool;
48 static mempool_t *nfs_commit_mempool;
50 struct nfs_write_data *nfs_commitdata_alloc(void)
52 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
55 memset(p, 0, sizeof(*p));
56 INIT_LIST_HEAD(&p->pages);
57 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
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 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
78 if (pagecount <= ARRAY_SIZE(p->page_array))
79 p->pagevec = p->page_array;
81 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
83 mempool_free(p, nfs_wdata_mempool);
91 static void nfs_writedata_free(struct nfs_write_data *p)
93 if (p && (p->pagevec != &p->page_array[0]))
95 mempool_free(p, nfs_wdata_mempool);
98 void nfs_writedata_release(void *data)
100 struct nfs_write_data *wdata = data;
102 put_nfs_open_context(wdata->args.context);
103 nfs_writedata_free(wdata);
106 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
110 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
113 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
115 struct nfs_page *req = NULL;
117 if (PagePrivate(page)) {
118 req = (struct nfs_page *)page_private(page);
120 kref_get(&req->wb_kref);
125 static struct nfs_page *nfs_page_find_request(struct page *page)
127 struct inode *inode = page->mapping->host;
128 struct nfs_page *req = NULL;
130 spin_lock(&inode->i_lock);
131 req = nfs_page_find_request_locked(page);
132 spin_unlock(&inode->i_lock);
136 /* Adjust the file length if we're writing beyond the end */
137 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
139 struct inode *inode = page->mapping->host;
143 spin_lock(&inode->i_lock);
144 i_size = i_size_read(inode);
145 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
146 if (i_size > 0 && page->index < end_index)
148 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
151 i_size_write(inode, end);
152 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
154 spin_unlock(&inode->i_lock);
157 /* A writeback failed: mark the page as bad, and invalidate the page cache */
158 static void nfs_set_pageerror(struct page *page)
161 nfs_zap_mapping(page->mapping->host, page->mapping);
164 /* We can set the PG_uptodate flag if we see that a write request
165 * covers the full page.
167 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
169 if (PageUptodate(page))
173 if (count != nfs_page_length(page))
175 SetPageUptodate(page);
178 static int wb_priority(struct writeback_control *wbc)
180 if (wbc->for_reclaim)
181 return FLUSH_HIGHPRI | FLUSH_STABLE;
182 if (wbc->for_kupdate)
188 * NFS congestion control
191 int nfs_congestion_kb;
193 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
194 #define NFS_CONGESTION_OFF_THRESH \
195 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
197 static int nfs_set_page_writeback(struct page *page)
199 int ret = test_set_page_writeback(page);
202 struct inode *inode = page->mapping->host;
203 struct nfs_server *nfss = NFS_SERVER(inode);
205 if (atomic_long_inc_return(&nfss->writeback) >
206 NFS_CONGESTION_ON_THRESH) {
207 set_bdi_congested(&nfss->backing_dev_info,
214 static void nfs_end_page_writeback(struct page *page)
216 struct inode *inode = page->mapping->host;
217 struct nfs_server *nfss = NFS_SERVER(inode);
219 end_page_writeback(page);
220 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
221 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
225 * Find an associated nfs write request, and prepare to flush it out
226 * May return an error if the user signalled nfs_wait_on_request().
228 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
231 struct inode *inode = page->mapping->host;
232 struct nfs_page *req;
235 spin_lock(&inode->i_lock);
237 req = nfs_page_find_request_locked(page);
239 spin_unlock(&inode->i_lock);
242 if (nfs_set_page_tag_locked(req))
244 /* Note: If we hold the page lock, as is the case in nfs_writepage,
245 * then the call to nfs_set_page_tag_locked() will always
246 * succeed provided that someone hasn't already marked the
247 * request as dirty (in which case we don't care).
249 spin_unlock(&inode->i_lock);
250 ret = nfs_wait_on_request(req);
251 nfs_release_request(req);
254 spin_lock(&inode->i_lock);
256 if (test_bit(PG_CLEAN, &req->wb_flags)) {
257 spin_unlock(&inode->i_lock);
260 if (nfs_set_page_writeback(page) != 0) {
261 spin_unlock(&inode->i_lock);
264 spin_unlock(&inode->i_lock);
265 if (!nfs_pageio_add_request(pgio, req)) {
266 nfs_redirty_request(req);
267 return pgio->pg_error;
272 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
274 struct inode *inode = page->mapping->host;
276 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
277 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
279 nfs_pageio_cond_complete(pgio, page->index);
280 return nfs_page_async_flush(pgio, page);
284 * Write an mmapped page to the server.
286 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
288 struct nfs_pageio_descriptor pgio;
291 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
292 err = nfs_do_writepage(page, wbc, &pgio);
293 nfs_pageio_complete(&pgio);
296 if (pgio.pg_error < 0)
297 return pgio.pg_error;
301 int nfs_writepage(struct page *page, struct writeback_control *wbc)
305 ret = nfs_writepage_locked(page, wbc);
310 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
314 ret = nfs_do_writepage(page, wbc, data);
319 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
321 struct inode *inode = mapping->host;
322 unsigned long *bitlock = &NFS_I(inode)->flags;
323 struct nfs_pageio_descriptor pgio;
326 /* Stop dirtying of new pages while we sync */
327 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
328 nfs_wait_bit_killable, TASK_KILLABLE);
332 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
334 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
335 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
336 nfs_pageio_complete(&pgio);
338 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
339 smp_mb__after_clear_bit();
340 wake_up_bit(bitlock, NFS_INO_FLUSHING);
353 * Insert a write request into an inode
355 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
357 struct nfs_inode *nfsi = NFS_I(inode);
360 error = radix_tree_preload(GFP_NOFS);
364 /* Lock the request! */
365 nfs_lock_request_dontget(req);
367 spin_lock(&inode->i_lock);
368 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
372 if (nfs_have_delegation(inode, FMODE_WRITE))
375 SetPagePrivate(req->wb_page);
376 set_page_private(req->wb_page, (unsigned long)req);
378 kref_get(&req->wb_kref);
379 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
380 NFS_PAGE_TAG_LOCKED);
381 spin_unlock(&inode->i_lock);
382 radix_tree_preload_end();
388 * Remove a write request from an inode
390 static void nfs_inode_remove_request(struct nfs_page *req)
392 struct inode *inode = req->wb_context->path.dentry->d_inode;
393 struct nfs_inode *nfsi = NFS_I(inode);
395 BUG_ON (!NFS_WBACK_BUSY(req));
397 spin_lock(&inode->i_lock);
398 set_page_private(req->wb_page, 0);
399 ClearPagePrivate(req->wb_page);
400 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
403 spin_unlock(&inode->i_lock);
406 spin_unlock(&inode->i_lock);
407 nfs_clear_request(req);
408 nfs_release_request(req);
412 nfs_mark_request_dirty(struct nfs_page *req)
414 __set_page_dirty_nobuffers(req->wb_page);
417 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
419 * Add a request to the inode's commit list.
422 nfs_mark_request_commit(struct nfs_page *req)
424 struct inode *inode = req->wb_context->path.dentry->d_inode;
425 struct nfs_inode *nfsi = NFS_I(inode);
427 spin_lock(&inode->i_lock);
428 set_bit(PG_CLEAN, &(req)->wb_flags);
429 radix_tree_tag_set(&nfsi->nfs_page_tree,
431 NFS_PAGE_TAG_COMMIT);
432 spin_unlock(&inode->i_lock);
433 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
434 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
435 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
439 nfs_clear_request_commit(struct nfs_page *req)
441 struct page *page = req->wb_page;
443 if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
444 dec_zone_page_state(page, NR_UNSTABLE_NFS);
445 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
452 int nfs_write_need_commit(struct nfs_write_data *data)
454 return data->verf.committed != NFS_FILE_SYNC;
458 int nfs_reschedule_unstable_write(struct nfs_page *req)
460 if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
461 nfs_mark_request_commit(req);
464 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
465 nfs_mark_request_dirty(req);
472 nfs_mark_request_commit(struct nfs_page *req)
477 nfs_clear_request_commit(struct nfs_page *req)
483 int nfs_write_need_commit(struct nfs_write_data *data)
489 int nfs_reschedule_unstable_write(struct nfs_page *req)
496 * Wait for a request to complete.
498 * Interruptible by fatal signals only.
500 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
502 struct nfs_inode *nfsi = NFS_I(inode);
503 struct nfs_page *req;
504 pgoff_t idx_end, next;
505 unsigned int res = 0;
511 idx_end = idx_start + npages - 1;
514 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
515 if (req->wb_index > idx_end)
518 next = req->wb_index + 1;
519 BUG_ON(!NFS_WBACK_BUSY(req));
521 kref_get(&req->wb_kref);
522 spin_unlock(&inode->i_lock);
523 error = nfs_wait_on_request(req);
524 nfs_release_request(req);
525 spin_lock(&inode->i_lock);
533 static void nfs_cancel_commit_list(struct list_head *head)
535 struct nfs_page *req;
537 while(!list_empty(head)) {
538 req = nfs_list_entry(head->next);
539 nfs_list_remove_request(req);
540 nfs_clear_request_commit(req);
541 nfs_inode_remove_request(req);
542 nfs_unlock_request(req);
546 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
548 nfs_need_commit(struct nfs_inode *nfsi)
550 return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
554 * nfs_scan_commit - Scan an inode for commit requests
555 * @inode: NFS inode to scan
556 * @dst: destination list
557 * @idx_start: lower bound of page->index to scan.
558 * @npages: idx_start + npages sets the upper bound to scan.
560 * Moves requests from the inode's 'commit' request list.
561 * The requests are *not* checked to ensure that they form a contiguous set.
564 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
566 struct nfs_inode *nfsi = NFS_I(inode);
568 if (!nfs_need_commit(nfsi))
571 return nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
574 static inline int nfs_need_commit(struct nfs_inode *nfsi)
579 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
586 * Search for an existing write request, and attempt to update
587 * it to reflect a new dirty region on a given page.
589 * If the attempt fails, then the existing request is flushed out
592 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
597 struct nfs_page *req;
602 if (!PagePrivate(page))
605 end = offset + bytes;
606 spin_lock(&inode->i_lock);
609 req = nfs_page_find_request_locked(page);
613 rqend = req->wb_offset + req->wb_bytes;
615 * Tell the caller to flush out the request if
616 * the offsets are non-contiguous.
617 * Note: nfs_flush_incompatible() will already
618 * have flushed out requests having wrong owners.
621 || end < req->wb_offset)
624 if (nfs_set_page_tag_locked(req))
627 /* The request is locked, so wait and then retry */
628 spin_unlock(&inode->i_lock);
629 error = nfs_wait_on_request(req);
630 nfs_release_request(req);
633 spin_lock(&inode->i_lock);
636 if (nfs_clear_request_commit(req))
637 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
638 req->wb_index, NFS_PAGE_TAG_COMMIT);
640 /* Okay, the request matches. Update the region */
641 if (offset < req->wb_offset) {
642 req->wb_offset = offset;
643 req->wb_pgbase = offset;
646 req->wb_bytes = end - req->wb_offset;
648 req->wb_bytes = rqend - req->wb_offset;
650 spin_unlock(&inode->i_lock);
653 spin_unlock(&inode->i_lock);
654 nfs_release_request(req);
655 error = nfs_wb_page(inode, page);
657 return ERR_PTR(error);
661 * Try to update an existing write request, or create one if there is none.
663 * Note: Should always be called with the Page Lock held to prevent races
664 * if we have to add a new request. Also assumes that the caller has
665 * already called nfs_flush_incompatible() if necessary.
667 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
668 struct page *page, unsigned int offset, unsigned int bytes)
670 struct inode *inode = page->mapping->host;
671 struct nfs_page *req;
674 req = nfs_try_to_update_request(inode, page, offset, bytes);
677 req = nfs_create_request(ctx, inode, page, offset, bytes);
680 error = nfs_inode_add_request(inode, req);
682 nfs_release_request(req);
683 req = ERR_PTR(error);
689 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
690 unsigned int offset, unsigned int count)
692 struct nfs_page *req;
694 req = nfs_setup_write_request(ctx, page, offset, count);
697 /* Update file length */
698 nfs_grow_file(page, offset, count);
699 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
700 nfs_clear_page_tag_locked(req);
704 int nfs_flush_incompatible(struct file *file, struct page *page)
706 struct nfs_open_context *ctx = nfs_file_open_context(file);
707 struct nfs_page *req;
708 int do_flush, status;
710 * Look for a request corresponding to this page. If there
711 * is one, and it belongs to another file, we flush it out
712 * before we try to copy anything into the page. Do this
713 * due to the lack of an ACCESS-type call in NFSv2.
714 * Also do the same if we find a request from an existing
718 req = nfs_page_find_request(page);
721 do_flush = req->wb_page != page || req->wb_context != ctx;
722 nfs_release_request(req);
725 status = nfs_wb_page(page->mapping->host, page);
726 } while (status == 0);
731 * If the page cache is marked as unsafe or invalid, then we can't rely on
732 * the PageUptodate() flag. In this case, we will need to turn off
733 * write optimisations that depend on the page contents being correct.
735 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
737 return PageUptodate(page) &&
738 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
742 * Update and possibly write a cached page of an NFS file.
744 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
745 * things with a page scheduled for an RPC call (e.g. invalidate it).
747 int nfs_updatepage(struct file *file, struct page *page,
748 unsigned int offset, unsigned int count)
750 struct nfs_open_context *ctx = nfs_file_open_context(file);
751 struct inode *inode = page->mapping->host;
754 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
756 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
757 file->f_path.dentry->d_parent->d_name.name,
758 file->f_path.dentry->d_name.name, count,
759 (long long)(page_offset(page) + offset));
761 /* If we're not using byte range locks, and we know the page
762 * is up to date, it may be more efficient to extend the write
763 * to cover the entire page in order to avoid fragmentation
766 if (nfs_write_pageuptodate(page, inode) &&
767 inode->i_flock == NULL &&
768 !(file->f_flags & O_SYNC)) {
769 count = max(count + offset, nfs_page_length(page));
773 status = nfs_writepage_setup(ctx, page, offset, count);
775 nfs_set_pageerror(page);
777 __set_page_dirty_nobuffers(page);
779 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
780 status, (long long)i_size_read(inode));
784 static void nfs_writepage_release(struct nfs_page *req)
787 if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
788 nfs_end_page_writeback(req->wb_page);
789 nfs_inode_remove_request(req);
791 nfs_end_page_writeback(req->wb_page);
792 nfs_clear_page_tag_locked(req);
795 static int flush_task_priority(int how)
797 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
799 return RPC_PRIORITY_HIGH;
801 return RPC_PRIORITY_LOW;
803 return RPC_PRIORITY_NORMAL;
807 * Set up the argument/result storage required for the RPC call.
809 static int nfs_write_rpcsetup(struct nfs_page *req,
810 struct nfs_write_data *data,
811 const struct rpc_call_ops *call_ops,
812 unsigned int count, unsigned int offset,
815 struct inode *inode = req->wb_context->path.dentry->d_inode;
816 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
817 int priority = flush_task_priority(how);
818 struct rpc_task *task;
819 struct rpc_message msg = {
820 .rpc_argp = &data->args,
821 .rpc_resp = &data->res,
822 .rpc_cred = req->wb_context->cred,
824 struct rpc_task_setup task_setup_data = {
825 .rpc_client = NFS_CLIENT(inode),
828 .callback_ops = call_ops,
829 .callback_data = data,
830 .workqueue = nfsiod_workqueue,
832 .priority = priority,
835 /* Set up the RPC argument and reply structs
836 * NB: take care not to mess about with data->commit et al. */
839 data->inode = inode = req->wb_context->path.dentry->d_inode;
840 data->cred = msg.rpc_cred;
842 data->args.fh = NFS_FH(inode);
843 data->args.offset = req_offset(req) + offset;
844 data->args.pgbase = req->wb_pgbase + offset;
845 data->args.pages = data->pagevec;
846 data->args.count = count;
847 data->args.context = get_nfs_open_context(req->wb_context);
848 data->args.stable = NFS_UNSTABLE;
849 if (how & FLUSH_STABLE) {
850 data->args.stable = NFS_DATA_SYNC;
851 if (!nfs_need_commit(NFS_I(inode)))
852 data->args.stable = NFS_FILE_SYNC;
855 data->res.fattr = &data->fattr;
856 data->res.count = count;
857 data->res.verf = &data->verf;
858 nfs_fattr_init(&data->fattr);
860 /* Set up the initial task struct. */
861 NFS_PROTO(inode)->write_setup(data, &msg);
863 dprintk("NFS: %5u initiated write call "
864 "(req %s/%lld, %u bytes @ offset %llu)\n",
867 (long long)NFS_FILEID(inode),
869 (unsigned long long)data->args.offset);
871 task = rpc_run_task(&task_setup_data);
873 return PTR_ERR(task);
878 /* If a nfs_flush_* function fails, it should remove reqs from @head and
879 * call this on each, which will prepare them to be retried on next
880 * writeback using standard nfs.
882 static void nfs_redirty_request(struct nfs_page *req)
884 nfs_mark_request_dirty(req);
885 nfs_end_page_writeback(req->wb_page);
886 nfs_clear_page_tag_locked(req);
890 * Generate multiple small requests to write out a single
891 * contiguous dirty area on one page.
893 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
895 struct nfs_page *req = nfs_list_entry(head->next);
896 struct page *page = req->wb_page;
897 struct nfs_write_data *data;
898 size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
904 nfs_list_remove_request(req);
908 size_t len = min(nbytes, wsize);
910 data = nfs_writedata_alloc(1);
913 list_add(&data->pages, &list);
916 } while (nbytes != 0);
917 atomic_set(&req->wb_complete, requests);
919 ClearPageError(page);
925 data = list_entry(list.next, struct nfs_write_data, pages);
926 list_del_init(&data->pages);
928 data->pagevec[0] = page;
932 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
938 } while (nbytes != 0);
943 while (!list_empty(&list)) {
944 data = list_entry(list.next, struct nfs_write_data, pages);
945 list_del(&data->pages);
946 nfs_writedata_release(data);
948 nfs_redirty_request(req);
953 * Create an RPC task for the given write request and kick it.
954 * The page must have been locked by the caller.
956 * It may happen that the page we're passed is not marked dirty.
957 * This is the case if nfs_updatepage detects a conflicting request
958 * that has been written but not committed.
960 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
962 struct nfs_page *req;
964 struct nfs_write_data *data;
966 data = nfs_writedata_alloc(npages);
970 pages = data->pagevec;
971 while (!list_empty(head)) {
972 req = nfs_list_entry(head->next);
973 nfs_list_remove_request(req);
974 nfs_list_add_request(req, &data->pages);
975 ClearPageError(req->wb_page);
976 *pages++ = req->wb_page;
978 req = nfs_list_entry(data->pages.next);
980 /* Set up the argument struct */
981 return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
983 while (!list_empty(head)) {
984 req = nfs_list_entry(head->next);
985 nfs_list_remove_request(req);
986 nfs_redirty_request(req);
991 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
992 struct inode *inode, int ioflags)
994 size_t wsize = NFS_SERVER(inode)->wsize;
996 if (wsize < PAGE_CACHE_SIZE)
997 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
999 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1003 * Handle a write reply that flushed part of a page.
1005 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1007 struct nfs_write_data *data = calldata;
1009 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1011 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1013 NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1014 data->req->wb_bytes, (long long)req_offset(data->req));
1016 nfs_writeback_done(task, data);
1019 static void nfs_writeback_release_partial(void *calldata)
1021 struct nfs_write_data *data = calldata;
1022 struct nfs_page *req = data->req;
1023 struct page *page = req->wb_page;
1024 int status = data->task.tk_status;
1027 nfs_set_pageerror(page);
1028 nfs_context_set_write_error(req->wb_context, status);
1029 dprintk(", error = %d\n", status);
1033 if (nfs_write_need_commit(data)) {
1034 struct inode *inode = page->mapping->host;
1036 spin_lock(&inode->i_lock);
1037 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1038 /* Do nothing we need to resend the writes */
1039 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1040 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1041 dprintk(" defer commit\n");
1042 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1043 set_bit(PG_NEED_RESCHED, &req->wb_flags);
1044 clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1045 dprintk(" server reboot detected\n");
1047 spin_unlock(&inode->i_lock);
1052 if (atomic_dec_and_test(&req->wb_complete))
1053 nfs_writepage_release(req);
1054 nfs_writedata_release(calldata);
1057 #if defined(CONFIG_NFS_V4_1)
1058 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1060 struct nfs_write_data *data = calldata;
1061 struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1063 if (nfs4_setup_sequence(clp, &data->args.seq_args,
1064 &data->res.seq_res, 1, task))
1066 rpc_call_start(task);
1068 #endif /* CONFIG_NFS_V4_1 */
1070 static const struct rpc_call_ops nfs_write_partial_ops = {
1071 #if defined(CONFIG_NFS_V4_1)
1072 .rpc_call_prepare = nfs_write_prepare,
1073 #endif /* CONFIG_NFS_V4_1 */
1074 .rpc_call_done = nfs_writeback_done_partial,
1075 .rpc_release = nfs_writeback_release_partial,
1079 * Handle a write reply that flushes a whole page.
1081 * FIXME: There is an inherent race with invalidate_inode_pages and
1082 * writebacks since the page->count is kept > 1 for as long
1083 * as the page has a write request pending.
1085 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1087 struct nfs_write_data *data = calldata;
1089 nfs_writeback_done(task, data);
1092 static void nfs_writeback_release_full(void *calldata)
1094 struct nfs_write_data *data = calldata;
1095 int status = data->task.tk_status;
1097 /* Update attributes as result of writeback. */
1098 while (!list_empty(&data->pages)) {
1099 struct nfs_page *req = nfs_list_entry(data->pages.next);
1100 struct page *page = req->wb_page;
1102 nfs_list_remove_request(req);
1104 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1106 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1107 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1109 (long long)req_offset(req));
1112 nfs_set_pageerror(page);
1113 nfs_context_set_write_error(req->wb_context, status);
1114 dprintk(", error = %d\n", status);
1115 goto remove_request;
1118 if (nfs_write_need_commit(data)) {
1119 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1120 nfs_mark_request_commit(req);
1121 nfs_end_page_writeback(page);
1122 dprintk(" marked for commit\n");
1127 nfs_end_page_writeback(page);
1128 nfs_inode_remove_request(req);
1130 nfs_clear_page_tag_locked(req);
1132 nfs_writedata_release(calldata);
1135 static const struct rpc_call_ops nfs_write_full_ops = {
1136 #if defined(CONFIG_NFS_V4_1)
1137 .rpc_call_prepare = nfs_write_prepare,
1138 #endif /* CONFIG_NFS_V4_1 */
1139 .rpc_call_done = nfs_writeback_done_full,
1140 .rpc_release = nfs_writeback_release_full,
1145 * This function is called when the WRITE call is complete.
1147 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1149 struct nfs_writeargs *argp = &data->args;
1150 struct nfs_writeres *resp = &data->res;
1151 struct nfs_server *server = NFS_SERVER(data->inode);
1154 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1155 task->tk_pid, task->tk_status);
1158 * ->write_done will attempt to use post-op attributes to detect
1159 * conflicting writes by other clients. A strict interpretation
1160 * of close-to-open would allow us to continue caching even if
1161 * another writer had changed the file, but some applications
1162 * depend on tighter cache coherency when writing.
1164 status = NFS_PROTO(data->inode)->write_done(task, data);
1167 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1169 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1170 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1171 /* We tried a write call, but the server did not
1172 * commit data to stable storage even though we
1174 * Note: There is a known bug in Tru64 < 5.0 in which
1175 * the server reports NFS_DATA_SYNC, but performs
1176 * NFS_FILE_SYNC. We therefore implement this checking
1177 * as a dprintk() in order to avoid filling syslog.
1179 static unsigned long complain;
1181 if (time_before(complain, jiffies)) {
1182 dprintk("NFS: faulty NFS server %s:"
1183 " (committed = %d) != (stable = %d)\n",
1184 server->nfs_client->cl_hostname,
1185 resp->verf->committed, argp->stable);
1186 complain = jiffies + 300 * HZ;
1190 /* Is this a short write? */
1191 if (task->tk_status >= 0 && resp->count < argp->count) {
1192 static unsigned long complain;
1194 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1196 /* Has the server at least made some progress? */
1197 if (resp->count != 0) {
1198 /* Was this an NFSv2 write or an NFSv3 stable write? */
1199 if (resp->verf->committed != NFS_UNSTABLE) {
1200 /* Resend from where the server left off */
1201 argp->offset += resp->count;
1202 argp->pgbase += resp->count;
1203 argp->count -= resp->count;
1205 /* Resend as a stable write in order to avoid
1206 * headaches in the case of a server crash.
1208 argp->stable = NFS_FILE_SYNC;
1210 nfs4_restart_rpc(task, server->nfs_client);
1213 if (time_before(complain, jiffies)) {
1215 "NFS: Server wrote zero bytes, expected %u.\n",
1217 complain = jiffies + 300 * HZ;
1219 /* Can't do anything about it except throw an error. */
1220 task->tk_status = -EIO;
1222 nfs4_sequence_free_slot(server->nfs_client, &data->res.seq_res);
1227 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1228 void nfs_commitdata_release(void *data)
1230 struct nfs_write_data *wdata = data;
1232 put_nfs_open_context(wdata->args.context);
1233 nfs_commit_free(wdata);
1237 * Set up the argument/result storage required for the RPC call.
1239 static int nfs_commit_rpcsetup(struct list_head *head,
1240 struct nfs_write_data *data,
1243 struct nfs_page *first = nfs_list_entry(head->next);
1244 struct inode *inode = first->wb_context->path.dentry->d_inode;
1245 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1246 int priority = flush_task_priority(how);
1247 struct rpc_task *task;
1248 struct rpc_message msg = {
1249 .rpc_argp = &data->args,
1250 .rpc_resp = &data->res,
1251 .rpc_cred = first->wb_context->cred,
1253 struct rpc_task_setup task_setup_data = {
1254 .task = &data->task,
1255 .rpc_client = NFS_CLIENT(inode),
1256 .rpc_message = &msg,
1257 .callback_ops = &nfs_commit_ops,
1258 .callback_data = data,
1259 .workqueue = nfsiod_workqueue,
1261 .priority = priority,
1264 /* Set up the RPC argument and reply structs
1265 * NB: take care not to mess about with data->commit et al. */
1267 list_splice_init(head, &data->pages);
1269 data->inode = inode;
1270 data->cred = msg.rpc_cred;
1272 data->args.fh = NFS_FH(data->inode);
1273 /* Note: we always request a commit of the entire inode */
1274 data->args.offset = 0;
1275 data->args.count = 0;
1276 data->args.context = get_nfs_open_context(first->wb_context);
1277 data->res.count = 0;
1278 data->res.fattr = &data->fattr;
1279 data->res.verf = &data->verf;
1280 nfs_fattr_init(&data->fattr);
1282 /* Set up the initial task struct. */
1283 NFS_PROTO(inode)->commit_setup(data, &msg);
1285 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1287 task = rpc_run_task(&task_setup_data);
1289 return PTR_ERR(task);
1295 * Commit dirty pages
1298 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1300 struct nfs_write_data *data;
1301 struct nfs_page *req;
1303 data = nfs_commitdata_alloc();
1308 /* Set up the argument struct */
1309 return nfs_commit_rpcsetup(head, data, how);
1311 while (!list_empty(head)) {
1312 req = nfs_list_entry(head->next);
1313 nfs_list_remove_request(req);
1314 nfs_mark_request_commit(req);
1315 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1316 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1318 nfs_clear_page_tag_locked(req);
1324 * COMMIT call returned
1326 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1328 struct nfs_write_data *data = calldata;
1330 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1331 task->tk_pid, task->tk_status);
1333 /* Call the NFS version-specific code */
1334 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1338 static void nfs_commit_release(void *calldata)
1340 struct nfs_write_data *data = calldata;
1341 struct nfs_page *req;
1342 int status = data->task.tk_status;
1344 while (!list_empty(&data->pages)) {
1345 req = nfs_list_entry(data->pages.next);
1346 nfs_list_remove_request(req);
1347 nfs_clear_request_commit(req);
1349 dprintk("NFS: commit (%s/%lld %d@%lld)",
1350 req->wb_context->path.dentry->d_inode->i_sb->s_id,
1351 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1353 (long long)req_offset(req));
1355 nfs_context_set_write_error(req->wb_context, status);
1356 nfs_inode_remove_request(req);
1357 dprintk(", error = %d\n", status);
1361 /* Okay, COMMIT succeeded, apparently. Check the verifier
1362 * returned by the server against all stored verfs. */
1363 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1364 /* We have a match */
1365 nfs_inode_remove_request(req);
1369 /* We have a mismatch. Write the page again */
1370 dprintk(" mismatch\n");
1371 nfs_mark_request_dirty(req);
1373 nfs_clear_page_tag_locked(req);
1375 nfs_commitdata_release(calldata);
1378 static const struct rpc_call_ops nfs_commit_ops = {
1379 #if defined(CONFIG_NFS_V4_1)
1380 .rpc_call_prepare = nfs_write_prepare,
1381 #endif /* CONFIG_NFS_V4_1 */
1382 .rpc_call_done = nfs_commit_done,
1383 .rpc_release = nfs_commit_release,
1386 int nfs_commit_inode(struct inode *inode, int how)
1391 spin_lock(&inode->i_lock);
1392 res = nfs_scan_commit(inode, &head, 0, 0);
1393 spin_unlock(&inode->i_lock);
1395 int error = nfs_commit_list(inode, &head, how);
1402 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1408 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1410 struct inode *inode = mapping->host;
1411 pgoff_t idx_start, idx_end;
1412 unsigned int npages = 0;
1414 int nocommit = how & FLUSH_NOCOMMIT;
1418 if (wbc->range_cyclic)
1421 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1422 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1423 if (idx_end > idx_start) {
1424 pgoff_t l_npages = 1 + idx_end - idx_start;
1426 if (sizeof(npages) != sizeof(l_npages) &&
1427 (pgoff_t)npages != l_npages)
1431 how &= ~FLUSH_NOCOMMIT;
1432 spin_lock(&inode->i_lock);
1434 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1439 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1442 if (how & FLUSH_INVALIDATE) {
1443 spin_unlock(&inode->i_lock);
1444 nfs_cancel_commit_list(&head);
1446 spin_lock(&inode->i_lock);
1449 pages += nfs_scan_commit(inode, &head, 0, 0);
1450 spin_unlock(&inode->i_lock);
1451 ret = nfs_commit_list(inode, &head, how);
1452 spin_lock(&inode->i_lock);
1455 spin_unlock(&inode->i_lock);
1459 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1463 ret = nfs_writepages(mapping, wbc);
1466 ret = nfs_sync_mapping_wait(mapping, wbc, how);
1471 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1475 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1476 static int nfs_write_mapping(struct address_space *mapping, int how)
1478 struct writeback_control wbc = {
1479 .bdi = mapping->backing_dev_info,
1480 .sync_mode = WB_SYNC_ALL,
1481 .nr_to_write = LONG_MAX,
1483 .range_end = LLONG_MAX,
1484 .for_writepages = 1,
1487 return __nfs_write_mapping(mapping, &wbc, how);
1491 * flush the inode to disk.
1493 int nfs_wb_all(struct inode *inode)
1495 return nfs_write_mapping(inode->i_mapping, 0);
1498 int nfs_wb_nocommit(struct inode *inode)
1500 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1503 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1505 struct nfs_page *req;
1506 loff_t range_start = page_offset(page);
1507 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1508 struct writeback_control wbc = {
1509 .bdi = page->mapping->backing_dev_info,
1510 .sync_mode = WB_SYNC_ALL,
1511 .nr_to_write = LONG_MAX,
1512 .range_start = range_start,
1513 .range_end = range_end,
1517 BUG_ON(!PageLocked(page));
1519 req = nfs_page_find_request(page);
1522 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1523 nfs_release_request(req);
1526 if (nfs_lock_request_dontget(req)) {
1527 nfs_inode_remove_request(req);
1529 * In case nfs_inode_remove_request has marked the
1530 * page as being dirty
1532 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1533 nfs_unlock_request(req);
1536 ret = nfs_wait_on_request(req);
1540 if (!PagePrivate(page))
1542 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1547 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1550 loff_t range_start = page_offset(page);
1551 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1552 struct writeback_control wbc = {
1553 .bdi = page->mapping->backing_dev_info,
1554 .sync_mode = WB_SYNC_ALL,
1555 .nr_to_write = LONG_MAX,
1556 .range_start = range_start,
1557 .range_end = range_end,
1562 if (clear_page_dirty_for_io(page)) {
1563 ret = nfs_writepage_locked(page, &wbc);
1566 } else if (!PagePrivate(page))
1568 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1571 } while (PagePrivate(page));
1574 __mark_inode_dirty(inode, I_DIRTY_PAGES);
1579 * Write back all requests on one page - we do this before reading it.
1581 int nfs_wb_page(struct inode *inode, struct page* page)
1583 return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1586 int __init nfs_init_writepagecache(void)
1588 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1589 sizeof(struct nfs_write_data),
1590 0, SLAB_HWCACHE_ALIGN,
1592 if (nfs_wdata_cachep == NULL)
1595 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1597 if (nfs_wdata_mempool == NULL)
1600 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1602 if (nfs_commit_mempool == NULL)
1606 * NFS congestion size, scale with available memory.
1618 * This allows larger machines to have larger/more transfers.
1619 * Limit the default to 256M
1621 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1622 if (nfs_congestion_kb > 256*1024)
1623 nfs_congestion_kb = 256*1024;
1628 void nfs_destroy_writepagecache(void)
1630 mempool_destroy(nfs_commit_mempool);
1631 mempool_destroy(nfs_wdata_mempool);
1632 kmem_cache_destroy(nfs_wdata_cachep);