4 * Writing file data over NFS.
6 * We do it like this: When a (user) process wishes to write data to an
7 * NFS file, a write request is allocated that contains the RPC task data
8 * plus some info on the page to be written, and added to the inode's
9 * write chain. If the process writes past the end of the page, an async
10 * RPC call to write the page is scheduled immediately; otherwise, the call
11 * is delayed for a few seconds.
13 * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
15 * Write requests are kept on the inode's writeback list. Each entry in
16 * that list references the page (portion) to be written. When the
17 * cache timeout has expired, the RPC task is woken up, and tries to
18 * lock the page. As soon as it manages to do so, the request is moved
19 * from the writeback list to the writelock list.
21 * Note: we must make sure never to confuse the inode passed in the
22 * write_page request with the one in page->inode. As far as I understand
23 * it, these are different when doing a swap-out.
25 * To understand everything that goes on here and in the NFS read code,
26 * one should be aware that a page is locked in exactly one of the following
29 * - A write request is in progress.
30 * - A user process is in generic_file_write/nfs_update_page
31 * - A user process is in generic_file_read
33 * Also note that because of the way pages are invalidated in
34 * nfs_revalidate_inode, the following assertions hold:
36 * - If a page is dirty, there will be no read requests (a page will
37 * not be re-read unless invalidated by nfs_revalidate_inode).
38 * - If the page is not uptodate, there will be no pending write
39 * requests, and no process will be in nfs_update_page.
41 * FIXME: Interaction with the vmscan routines is not optimal yet.
42 * Either vmscan must be made nfs-savvy, or we need a different page
43 * reclaim concept that supports something like FS-independent
44 * buffer_heads with a b_ops-> field.
46 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
49 #include <linux/config.h>
50 #include <linux/types.h>
51 #include <linux/slab.h>
53 #include <linux/pagemap.h>
54 #include <linux/file.h>
55 #include <linux/mpage.h>
56 #include <linux/writeback.h>
58 #include <linux/sunrpc/clnt.h>
59 #include <linux/nfs_fs.h>
60 #include <linux/nfs_mount.h>
61 #include <linux/nfs_page.h>
62 #include <asm/uaccess.h>
63 #include <linux/smp_lock.h>
65 #include "delegation.h"
67 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
69 #define MIN_POOL_WRITE (32)
70 #define MIN_POOL_COMMIT (4)
73 * Local function declarations
75 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
78 unsigned int, unsigned int);
79 static void nfs_writeback_done_partial(struct nfs_write_data *, int);
80 static void nfs_writeback_done_full(struct nfs_write_data *, int);
81 static int nfs_wait_on_write_congestion(struct address_space *, int);
82 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
83 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
84 unsigned int npages, int how);
86 static kmem_cache_t *nfs_wdata_cachep;
87 mempool_t *nfs_wdata_mempool;
88 static mempool_t *nfs_commit_mempool;
90 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
92 static inline struct nfs_write_data *nfs_commit_alloc(void)
94 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
96 memset(p, 0, sizeof(*p));
97 INIT_LIST_HEAD(&p->pages);
102 static inline void nfs_commit_free(struct nfs_write_data *p)
104 mempool_free(p, nfs_commit_mempool);
107 void nfs_writedata_release(void *wdata)
109 nfs_writedata_free(wdata);
112 /* Adjust the file length if we're writing beyond the end */
113 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
115 struct inode *inode = page->mapping->host;
116 loff_t end, i_size = i_size_read(inode);
117 unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
119 if (i_size > 0 && page->index < end_index)
121 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
124 i_size_write(inode, end);
127 /* We can set the PG_uptodate flag if we see that a write request
128 * covers the full page.
130 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
134 if (PageUptodate(page))
138 if (count == PAGE_CACHE_SIZE) {
139 SetPageUptodate(page);
143 end_offs = i_size_read(page->mapping->host) - 1;
146 /* Is this the last page? */
147 if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
149 /* This is the last page: set PG_uptodate if we cover the entire
150 * extent of the data, then zero the rest of the page.
152 if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
153 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
154 SetPageUptodate(page);
159 * Write a page synchronously.
160 * Offset is the data offset within the page.
162 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
163 struct page *page, unsigned int offset, unsigned int count,
166 unsigned int wsize = NFS_SERVER(inode)->wsize;
167 int result, written = 0;
168 struct nfs_write_data *wdata;
170 wdata = nfs_writedata_alloc();
175 wdata->cred = ctx->cred;
176 wdata->inode = inode;
177 wdata->args.fh = NFS_FH(inode);
178 wdata->args.context = ctx;
179 wdata->args.pages = &page;
180 wdata->args.stable = NFS_FILE_SYNC;
181 wdata->args.pgbase = offset;
182 wdata->args.count = wsize;
183 wdata->res.fattr = &wdata->fattr;
184 wdata->res.verf = &wdata->verf;
186 dprintk("NFS: nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
188 (long long)NFS_FILEID(inode),
189 count, (long long)(page_offset(page) + offset));
191 set_page_writeback(page);
192 nfs_begin_data_update(inode);
195 wdata->args.count = count;
196 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
198 result = NFS_PROTO(inode)->write(wdata);
201 /* Must mark the page invalid after I/O error */
202 ClearPageUptodate(page);
205 if (result < wdata->args.count)
206 printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
207 wdata->args.count, result);
209 wdata->args.offset += result;
210 wdata->args.pgbase += result;
214 /* Update file length */
215 nfs_grow_file(page, offset, written);
216 /* Set the PG_uptodate flag? */
217 nfs_mark_uptodate(page, offset, written);
220 ClearPageError(page);
223 nfs_end_data_update(inode);
224 end_page_writeback(page);
225 nfs_writedata_free(wdata);
226 return written ? written : result;
229 static int nfs_writepage_async(struct nfs_open_context *ctx,
230 struct inode *inode, struct page *page,
231 unsigned int offset, unsigned int count)
233 struct nfs_page *req;
235 req = nfs_update_request(ctx, inode, page, offset, count);
238 /* Update file length */
239 nfs_grow_file(page, offset, count);
240 /* Set the PG_uptodate flag? */
241 nfs_mark_uptodate(page, offset, count);
242 nfs_unlock_request(req);
246 static int wb_priority(struct writeback_control *wbc)
248 if (wbc->for_reclaim)
249 return FLUSH_HIGHPRI;
250 if (wbc->for_kupdate)
256 * Write an mmapped page to the server.
258 int nfs_writepage(struct page *page, struct writeback_control *wbc)
260 struct nfs_open_context *ctx;
261 struct inode *inode = page->mapping->host;
262 unsigned long end_index;
263 unsigned offset = PAGE_CACHE_SIZE;
264 loff_t i_size = i_size_read(inode);
265 int inode_referenced = 0;
266 int priority = wb_priority(wbc);
270 * Note: We need to ensure that we have a reference to the inode
271 * if we are to do asynchronous writes. If not, waiting
272 * in nfs_wait_on_request() may deadlock with clear_inode().
274 * If igrab() fails here, then it is in any case safe to
275 * call nfs_wb_page(), since there will be no pending writes.
277 if (igrab(inode) != 0)
278 inode_referenced = 1;
279 end_index = i_size >> PAGE_CACHE_SHIFT;
281 /* Ensure we've flushed out any previous writes */
282 nfs_wb_page_priority(inode, page, priority);
285 if (page->index < end_index)
287 /* things got complicated... */
288 offset = i_size & (PAGE_CACHE_SIZE-1);
290 /* OK, are we completely out? */
291 err = 0; /* potential race with truncate - ignore */
292 if (page->index >= end_index+1 || !offset)
295 ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
301 if (!IS_SYNC(inode) && inode_referenced) {
302 err = nfs_writepage_async(ctx, inode, page, 0, offset);
303 if (!wbc->for_writepages)
304 nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
306 err = nfs_writepage_sync(ctx, inode, page, 0,
310 redirty_page_for_writepage(wbc, page);
315 put_nfs_open_context(ctx);
318 if (inode_referenced)
324 * Note: causes nfs_update_request() to block on the assumption
325 * that the writeback is generated due to memory pressure.
327 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
329 struct backing_dev_info *bdi = mapping->backing_dev_info;
330 struct inode *inode = mapping->host;
333 err = generic_writepages(mapping, wbc);
336 while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
337 if (wbc->nonblocking)
339 nfs_wait_on_write_congestion(mapping, 0);
341 err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
344 wbc->nr_to_write -= err;
345 if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
346 err = nfs_wait_on_requests(inode, 0, 0);
350 err = nfs_commit_inode(inode, wb_priority(wbc));
352 wbc->nr_to_write -= err;
356 clear_bit(BDI_write_congested, &bdi->state);
357 wake_up_all(&nfs_write_congestion);
362 * Insert a write request into an inode
364 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
366 struct nfs_inode *nfsi = NFS_I(inode);
369 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
370 BUG_ON(error == -EEXIST);
375 nfs_begin_data_update(inode);
376 if (nfs_have_delegation(inode, FMODE_WRITE))
380 atomic_inc(&req->wb_count);
385 * Insert a write request into an inode
387 static void nfs_inode_remove_request(struct nfs_page *req)
389 struct inode *inode = req->wb_context->dentry->d_inode;
390 struct nfs_inode *nfsi = NFS_I(inode);
392 BUG_ON (!NFS_WBACK_BUSY(req));
394 spin_lock(&nfsi->req_lock);
395 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
398 spin_unlock(&nfsi->req_lock);
399 nfs_end_data_update(inode);
402 spin_unlock(&nfsi->req_lock);
403 nfs_clear_request(req);
404 nfs_release_request(req);
410 static inline struct nfs_page *
411 _nfs_find_request(struct inode *inode, unsigned long index)
413 struct nfs_inode *nfsi = NFS_I(inode);
414 struct nfs_page *req;
416 req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
418 atomic_inc(&req->wb_count);
422 static struct nfs_page *
423 nfs_find_request(struct inode *inode, unsigned long index)
425 struct nfs_page *req;
426 struct nfs_inode *nfsi = NFS_I(inode);
428 spin_lock(&nfsi->req_lock);
429 req = _nfs_find_request(inode, index);
430 spin_unlock(&nfsi->req_lock);
435 * Add a request to the inode's dirty list.
438 nfs_mark_request_dirty(struct nfs_page *req)
440 struct inode *inode = req->wb_context->dentry->d_inode;
441 struct nfs_inode *nfsi = NFS_I(inode);
443 spin_lock(&nfsi->req_lock);
444 radix_tree_tag_set(&nfsi->nfs_page_tree,
445 req->wb_index, NFS_PAGE_TAG_DIRTY);
446 nfs_list_add_request(req, &nfsi->dirty);
448 spin_unlock(&nfsi->req_lock);
449 inc_page_state(nr_dirty);
450 mark_inode_dirty(inode);
454 * Check if a request is dirty
457 nfs_dirty_request(struct nfs_page *req)
459 struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
460 return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
463 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
465 * Add a request to the inode's commit list.
468 nfs_mark_request_commit(struct nfs_page *req)
470 struct inode *inode = req->wb_context->dentry->d_inode;
471 struct nfs_inode *nfsi = NFS_I(inode);
473 spin_lock(&nfsi->req_lock);
474 nfs_list_add_request(req, &nfsi->commit);
476 spin_unlock(&nfsi->req_lock);
477 inc_page_state(nr_unstable);
478 mark_inode_dirty(inode);
483 * Wait for a request to complete.
485 * Interruptible by signals only if mounted with intr flag.
488 nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
490 struct nfs_inode *nfsi = NFS_I(inode);
491 struct nfs_page *req;
492 unsigned long idx_end, next;
493 unsigned int res = 0;
499 idx_end = idx_start + npages - 1;
501 spin_lock(&nfsi->req_lock);
503 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
504 if (req->wb_index > idx_end)
507 next = req->wb_index + 1;
508 BUG_ON(!NFS_WBACK_BUSY(req));
510 atomic_inc(&req->wb_count);
511 spin_unlock(&nfsi->req_lock);
512 error = nfs_wait_on_request(req);
513 nfs_release_request(req);
516 spin_lock(&nfsi->req_lock);
519 spin_unlock(&nfsi->req_lock);
524 * nfs_scan_dirty - Scan an inode for dirty requests
525 * @inode: NFS inode to scan
526 * @dst: destination list
527 * @idx_start: lower bound of page->index to scan.
528 * @npages: idx_start + npages sets the upper bound to scan.
530 * Moves requests from the inode's dirty page list.
531 * The requests are *not* checked to ensure that they form a contiguous set.
534 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
536 struct nfs_inode *nfsi = NFS_I(inode);
539 if (nfsi->ndirty != 0) {
540 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
542 sub_page_state(nr_dirty,res);
543 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
544 printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
549 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
551 * nfs_scan_commit - Scan an inode for commit requests
552 * @inode: NFS inode to scan
553 * @dst: destination list
554 * @idx_start: lower bound of page->index to scan.
555 * @npages: idx_start + npages sets the upper bound to scan.
557 * Moves requests from the inode's 'commit' request list.
558 * The requests are *not* checked to ensure that they form a contiguous set.
561 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
563 struct nfs_inode *nfsi = NFS_I(inode);
566 if (nfsi->ncommit != 0) {
567 res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);
568 nfsi->ncommit -= res;
569 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
570 printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
576 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
578 struct backing_dev_info *bdi = mapping->backing_dev_info;
584 if (!bdi_write_congested(bdi))
587 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
590 rpc_clnt_sigmask(clnt, &oldset);
591 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
592 if (bdi_write_congested(bdi)) {
598 rpc_clnt_sigunmask(clnt, &oldset);
600 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
601 if (bdi_write_congested(bdi))
604 finish_wait(&nfs_write_congestion, &wait);
610 * Try to update any existing write request, or create one if there is none.
611 * In order to match, the request's credentials must match those of
612 * the calling process.
614 * Note: Should always be called with the Page Lock held!
616 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
617 struct inode *inode, struct page *page,
618 unsigned int offset, unsigned int bytes)
620 struct nfs_server *server = NFS_SERVER(inode);
621 struct nfs_inode *nfsi = NFS_I(inode);
622 struct nfs_page *req, *new = NULL;
623 unsigned long rqend, end;
625 end = offset + bytes;
627 if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
628 return ERR_PTR(-ERESTARTSYS);
630 /* Loop over all inode entries and see if we find
631 * A request for the page we wish to update
633 spin_lock(&nfsi->req_lock);
634 req = _nfs_find_request(inode, page->index);
636 if (!nfs_lock_request_dontget(req)) {
638 spin_unlock(&nfsi->req_lock);
639 error = nfs_wait_on_request(req);
640 nfs_release_request(req);
642 return ERR_PTR(error);
645 spin_unlock(&nfsi->req_lock);
647 nfs_release_request(new);
653 nfs_lock_request_dontget(new);
654 error = nfs_inode_add_request(inode, new);
656 spin_unlock(&nfsi->req_lock);
657 nfs_unlock_request(new);
658 return ERR_PTR(error);
660 spin_unlock(&nfsi->req_lock);
661 nfs_mark_request_dirty(new);
664 spin_unlock(&nfsi->req_lock);
666 new = nfs_create_request(ctx, inode, page, offset, bytes);
671 /* We have a request for our page.
672 * If the creds don't match, or the
673 * page addresses don't match,
674 * tell the caller to wait on the conflicting
677 rqend = req->wb_offset + req->wb_bytes;
678 if (req->wb_context != ctx
679 || req->wb_page != page
680 || !nfs_dirty_request(req)
681 || offset > rqend || end < req->wb_offset) {
682 nfs_unlock_request(req);
683 return ERR_PTR(-EBUSY);
686 /* Okay, the request matches. Update the region */
687 if (offset < req->wb_offset) {
688 req->wb_offset = offset;
689 req->wb_pgbase = offset;
690 req->wb_bytes = rqend - req->wb_offset;
694 req->wb_bytes = end - req->wb_offset;
699 int nfs_flush_incompatible(struct file *file, struct page *page)
701 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
702 struct inode *inode = page->mapping->host;
703 struct nfs_page *req;
706 * Look for a request corresponding to this page. If there
707 * is one, and it belongs to another file, we flush it out
708 * before we try to copy anything into the page. Do this
709 * due to the lack of an ACCESS-type call in NFSv2.
710 * Also do the same if we find a request from an existing
713 req = nfs_find_request(inode, page->index);
715 if (req->wb_page != page || ctx != req->wb_context)
716 status = nfs_wb_page(inode, page);
717 nfs_release_request(req);
719 return (status < 0) ? status : 0;
723 * Update and possibly write a cached page of an NFS file.
725 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
726 * things with a page scheduled for an RPC call (e.g. invalidate it).
728 int nfs_updatepage(struct file *file, struct page *page,
729 unsigned int offset, unsigned int count)
731 struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
732 struct inode *inode = page->mapping->host;
733 struct nfs_page *req;
736 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
737 file->f_dentry->d_parent->d_name.name,
738 file->f_dentry->d_name.name, count,
739 (long long)(page_offset(page) +offset));
741 if (IS_SYNC(inode)) {
742 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
744 if (offset == 0 && status == PAGE_CACHE_SIZE)
745 SetPageUptodate(page);
751 /* If we're not using byte range locks, and we know the page
752 * is entirely in cache, it may be more efficient to avoid
753 * fragmenting write requests.
755 if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
756 loff_t end_offs = i_size_read(inode) - 1;
757 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
761 if (unlikely(end_offs < 0)) {
763 } else if (page->index == end_index) {
765 pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
768 } else if (page->index < end_index)
769 count = PAGE_CACHE_SIZE;
773 * Try to find an NFS request corresponding to this page
775 * If the existing request cannot be updated, we must flush
779 req = nfs_update_request(ctx, inode, page, offset, count);
780 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
781 if (status != -EBUSY)
783 /* Request could not be updated. Flush it out and try again */
784 status = nfs_wb_page(inode, page);
785 } while (status >= 0);
791 /* Update file length */
792 nfs_grow_file(page, offset, count);
793 /* Set the PG_uptodate flag? */
794 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
795 nfs_unlock_request(req);
797 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
798 status, (long long)i_size_read(inode));
800 ClearPageUptodate(page);
804 static void nfs_writepage_release(struct nfs_page *req)
806 end_page_writeback(req->wb_page);
808 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
809 if (!PageError(req->wb_page)) {
810 if (NFS_NEED_RESCHED(req)) {
811 nfs_mark_request_dirty(req);
813 } else if (NFS_NEED_COMMIT(req)) {
814 nfs_mark_request_commit(req);
818 nfs_inode_remove_request(req);
821 nfs_clear_commit(req);
822 nfs_clear_reschedule(req);
824 nfs_inode_remove_request(req);
826 nfs_clear_page_writeback(req);
829 static inline int flush_task_priority(int how)
831 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
833 return RPC_PRIORITY_HIGH;
835 return RPC_PRIORITY_LOW;
837 return RPC_PRIORITY_NORMAL;
841 * Set up the argument/result storage required for the RPC call.
843 static void nfs_write_rpcsetup(struct nfs_page *req,
844 struct nfs_write_data *data,
845 unsigned int count, unsigned int offset,
850 /* Set up the RPC argument and reply structs
851 * NB: take care not to mess about with data->commit et al. */
854 data->inode = inode = req->wb_context->dentry->d_inode;
855 data->cred = req->wb_context->cred;
857 data->args.fh = NFS_FH(inode);
858 data->args.offset = req_offset(req) + offset;
859 data->args.pgbase = req->wb_pgbase + offset;
860 data->args.pages = data->pagevec;
861 data->args.count = count;
862 data->args.context = req->wb_context;
864 data->res.fattr = &data->fattr;
865 data->res.count = count;
866 data->res.verf = &data->verf;
867 nfs_fattr_init(&data->fattr);
869 NFS_PROTO(inode)->write_setup(data, how);
871 data->task.tk_priority = flush_task_priority(how);
872 data->task.tk_cookie = (unsigned long)inode;
874 dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
877 (long long)NFS_FILEID(inode),
879 (unsigned long long)data->args.offset);
882 static void nfs_execute_write(struct nfs_write_data *data)
884 struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
887 rpc_clnt_sigmask(clnt, &oldset);
889 rpc_execute(&data->task);
891 rpc_clnt_sigunmask(clnt, &oldset);
895 * Generate multiple small requests to write out a single
896 * contiguous dirty area on one page.
898 static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how)
900 struct nfs_page *req = nfs_list_entry(head->next);
901 struct page *page = req->wb_page;
902 struct nfs_write_data *data;
903 unsigned int wsize = NFS_SERVER(inode)->wsize;
904 unsigned int nbytes, offset;
908 nfs_list_remove_request(req);
910 nbytes = req->wb_bytes;
912 data = nfs_writedata_alloc();
915 list_add(&data->pages, &list);
921 atomic_set(&req->wb_complete, requests);
923 ClearPageError(page);
924 set_page_writeback(page);
926 nbytes = req->wb_bytes;
928 data = list_entry(list.next, struct nfs_write_data, pages);
929 list_del_init(&data->pages);
931 data->pagevec[0] = page;
932 data->complete = nfs_writeback_done_partial;
934 if (nbytes > wsize) {
935 nfs_write_rpcsetup(req, data, wsize, offset, how);
939 nfs_write_rpcsetup(req, data, nbytes, offset, how);
942 nfs_execute_write(data);
943 } while (nbytes != 0);
948 while (!list_empty(&list)) {
949 data = list_entry(list.next, struct nfs_write_data, pages);
950 list_del(&data->pages);
951 nfs_writedata_free(data);
953 nfs_mark_request_dirty(req);
954 nfs_clear_page_writeback(req);
959 * Create an RPC task for the given write request and kick it.
960 * The page must have been locked by the caller.
962 * It may happen that the page we're passed is not marked dirty.
963 * This is the case if nfs_updatepage detects a conflicting request
964 * that has been written but not committed.
966 static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
968 struct nfs_page *req;
970 struct nfs_write_data *data;
973 if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE)
974 return nfs_flush_multi(head, inode, how);
976 data = nfs_writedata_alloc();
980 pages = data->pagevec;
982 while (!list_empty(head)) {
983 req = nfs_list_entry(head->next);
984 nfs_list_remove_request(req);
985 nfs_list_add_request(req, &data->pages);
986 ClearPageError(req->wb_page);
987 set_page_writeback(req->wb_page);
988 *pages++ = req->wb_page;
989 count += req->wb_bytes;
991 req = nfs_list_entry(data->pages.next);
993 data->complete = nfs_writeback_done_full;
994 /* Set up the argument struct */
995 nfs_write_rpcsetup(req, data, count, 0, how);
997 nfs_execute_write(data);
1000 while (!list_empty(head)) {
1001 struct nfs_page *req = nfs_list_entry(head->next);
1002 nfs_list_remove_request(req);
1003 nfs_mark_request_dirty(req);
1004 nfs_clear_page_writeback(req);
1010 nfs_flush_list(struct list_head *head, int wpages, int how)
1012 LIST_HEAD(one_request);
1013 struct nfs_page *req;
1015 unsigned int pages = 0;
1017 while (!list_empty(head)) {
1018 pages += nfs_coalesce_requests(head, &one_request, wpages);
1019 req = nfs_list_entry(one_request.next);
1020 error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how);
1027 while (!list_empty(head)) {
1028 req = nfs_list_entry(head->next);
1029 nfs_list_remove_request(req);
1030 nfs_mark_request_dirty(req);
1031 nfs_clear_page_writeback(req);
1037 * Handle a write reply that flushed part of a page.
1039 static void nfs_writeback_done_partial(struct nfs_write_data *data, int status)
1041 struct nfs_page *req = data->req;
1042 struct page *page = req->wb_page;
1044 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1045 req->wb_context->dentry->d_inode->i_sb->s_id,
1046 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1048 (long long)req_offset(req));
1051 ClearPageUptodate(page);
1053 req->wb_context->error = status;
1054 dprintk(", error = %d\n", status);
1056 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1057 if (data->verf.committed < NFS_FILE_SYNC) {
1058 if (!NFS_NEED_COMMIT(req)) {
1059 nfs_defer_commit(req);
1060 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1061 dprintk(" defer commit\n");
1062 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1063 nfs_defer_reschedule(req);
1064 dprintk(" server reboot detected\n");
1071 if (atomic_dec_and_test(&req->wb_complete))
1072 nfs_writepage_release(req);
1076 * Handle a write reply that flushes a whole page.
1078 * FIXME: There is an inherent race with invalidate_inode_pages and
1079 * writebacks since the page->count is kept > 1 for as long
1080 * as the page has a write request pending.
1082 static void nfs_writeback_done_full(struct nfs_write_data *data, int status)
1084 struct nfs_page *req;
1087 /* Update attributes as result of writeback. */
1088 while (!list_empty(&data->pages)) {
1089 req = nfs_list_entry(data->pages.next);
1090 nfs_list_remove_request(req);
1091 page = req->wb_page;
1093 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1094 req->wb_context->dentry->d_inode->i_sb->s_id,
1095 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1097 (long long)req_offset(req));
1100 ClearPageUptodate(page);
1102 req->wb_context->error = status;
1103 end_page_writeback(page);
1104 nfs_inode_remove_request(req);
1105 dprintk(", error = %d\n", status);
1108 end_page_writeback(page);
1110 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1111 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1112 nfs_inode_remove_request(req);
1116 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1117 nfs_mark_request_commit(req);
1118 dprintk(" marked for commit\n");
1120 nfs_inode_remove_request(req);
1123 nfs_clear_page_writeback(req);
1128 * This function is called when the WRITE call is complete.
1130 void nfs_writeback_done(struct rpc_task *task, void *calldata)
1132 struct nfs_write_data *data = calldata;
1133 struct nfs_writeargs *argp = &data->args;
1134 struct nfs_writeres *resp = &data->res;
1136 dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1137 task->tk_pid, task->tk_status);
1139 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1140 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1141 /* We tried a write call, but the server did not
1142 * commit data to stable storage even though we
1144 * Note: There is a known bug in Tru64 < 5.0 in which
1145 * the server reports NFS_DATA_SYNC, but performs
1146 * NFS_FILE_SYNC. We therefore implement this checking
1147 * as a dprintk() in order to avoid filling syslog.
1149 static unsigned long complain;
1151 if (time_before(complain, jiffies)) {
1152 dprintk("NFS: faulty NFS server %s:"
1153 " (committed = %d) != (stable = %d)\n",
1154 NFS_SERVER(data->inode)->hostname,
1155 resp->verf->committed, argp->stable);
1156 complain = jiffies + 300 * HZ;
1160 /* Is this a short write? */
1161 if (task->tk_status >= 0 && resp->count < argp->count) {
1162 static unsigned long complain;
1164 /* Has the server at least made some progress? */
1165 if (resp->count != 0) {
1166 /* Was this an NFSv2 write or an NFSv3 stable write? */
1167 if (resp->verf->committed != NFS_UNSTABLE) {
1168 /* Resend from where the server left off */
1169 argp->offset += resp->count;
1170 argp->pgbase += resp->count;
1171 argp->count -= resp->count;
1173 /* Resend as a stable write in order to avoid
1174 * headaches in the case of a server crash.
1176 argp->stable = NFS_FILE_SYNC;
1178 rpc_restart_call(task);
1181 if (time_before(complain, jiffies)) {
1183 "NFS: Server wrote zero bytes, expected %u.\n",
1185 complain = jiffies + 300 * HZ;
1187 /* Can't do anything about it except throw an error. */
1188 task->tk_status = -EIO;
1192 * Process the nfs_page list
1194 data->complete(data, task->tk_status);
1198 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1199 void nfs_commit_release(void *wdata)
1201 nfs_commit_free(wdata);
1205 * Set up the argument/result storage required for the RPC call.
1207 static void nfs_commit_rpcsetup(struct list_head *head,
1208 struct nfs_write_data *data, int how)
1210 struct nfs_page *first;
1211 struct inode *inode;
1213 /* Set up the RPC argument and reply structs
1214 * NB: take care not to mess about with data->commit et al. */
1216 list_splice_init(head, &data->pages);
1217 first = nfs_list_entry(data->pages.next);
1218 inode = first->wb_context->dentry->d_inode;
1220 data->inode = inode;
1221 data->cred = first->wb_context->cred;
1223 data->args.fh = NFS_FH(data->inode);
1224 /* Note: we always request a commit of the entire inode */
1225 data->args.offset = 0;
1226 data->args.count = 0;
1227 data->res.count = 0;
1228 data->res.fattr = &data->fattr;
1229 data->res.verf = &data->verf;
1230 nfs_fattr_init(&data->fattr);
1232 NFS_PROTO(inode)->commit_setup(data, how);
1234 data->task.tk_priority = flush_task_priority(how);
1235 data->task.tk_cookie = (unsigned long)inode;
1237 dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1241 * Commit dirty pages
1244 nfs_commit_list(struct list_head *head, int how)
1246 struct nfs_write_data *data;
1247 struct nfs_page *req;
1249 data = nfs_commit_alloc();
1254 /* Set up the argument struct */
1255 nfs_commit_rpcsetup(head, data, how);
1257 nfs_execute_write(data);
1260 while (!list_empty(head)) {
1261 req = nfs_list_entry(head->next);
1262 nfs_list_remove_request(req);
1263 nfs_mark_request_commit(req);
1264 nfs_clear_page_writeback(req);
1270 * COMMIT call returned
1272 void nfs_commit_done(struct rpc_task *task, void *calldata)
1274 struct nfs_write_data *data = calldata;
1275 struct nfs_page *req;
1278 dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1279 task->tk_pid, task->tk_status);
1281 while (!list_empty(&data->pages)) {
1282 req = nfs_list_entry(data->pages.next);
1283 nfs_list_remove_request(req);
1285 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1286 req->wb_context->dentry->d_inode->i_sb->s_id,
1287 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1289 (long long)req_offset(req));
1290 if (task->tk_status < 0) {
1291 req->wb_context->error = task->tk_status;
1292 nfs_inode_remove_request(req);
1293 dprintk(", error = %d\n", task->tk_status);
1297 /* Okay, COMMIT succeeded, apparently. Check the verifier
1298 * returned by the server against all stored verfs. */
1299 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1300 /* We have a match */
1301 nfs_inode_remove_request(req);
1305 /* We have a mismatch. Write the page again */
1306 dprintk(" mismatch\n");
1307 nfs_mark_request_dirty(req);
1309 nfs_clear_page_writeback(req);
1312 sub_page_state(nr_unstable,res);
1316 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1317 unsigned int npages, int how)
1319 struct nfs_inode *nfsi = NFS_I(inode);
1324 spin_lock(&nfsi->req_lock);
1325 res = nfs_scan_dirty(inode, &head, idx_start, npages);
1326 spin_unlock(&nfsi->req_lock);
1328 struct nfs_server *server = NFS_SERVER(inode);
1330 /* For single writes, FLUSH_STABLE is more efficient */
1331 if (res == nfsi->npages && nfsi->npages <= server->wpages) {
1332 if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize)
1333 how |= FLUSH_STABLE;
1335 error = nfs_flush_list(&head, server->wpages, how);
1342 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1343 int nfs_commit_inode(struct inode *inode, int how)
1345 struct nfs_inode *nfsi = NFS_I(inode);
1350 spin_lock(&nfsi->req_lock);
1351 res = nfs_scan_commit(inode, &head, 0, 0);
1352 spin_unlock(&nfsi->req_lock);
1354 error = nfs_commit_list(&head, how);
1362 int nfs_sync_inode(struct inode *inode, unsigned long idx_start,
1363 unsigned int npages, int how)
1368 wait = how & FLUSH_WAIT;
1374 error = nfs_wait_on_requests(inode, idx_start, npages);
1376 error = nfs_flush_inode(inode, idx_start, npages, how);
1377 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1379 error = nfs_commit_inode(inode, how);
1381 } while (error > 0);
1385 int nfs_init_writepagecache(void)
1387 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1388 sizeof(struct nfs_write_data),
1389 0, SLAB_HWCACHE_ALIGN,
1391 if (nfs_wdata_cachep == NULL)
1394 nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE,
1398 if (nfs_wdata_mempool == NULL)
1401 nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT,
1405 if (nfs_commit_mempool == NULL)
1411 void nfs_destroy_writepagecache(void)
1413 mempool_destroy(nfs_commit_mempool);
1414 mempool_destroy(nfs_wdata_mempool);
1415 if (kmem_cache_destroy(nfs_wdata_cachep))
1416 printk(KERN_INFO "nfs_write_data: not all structures were freed\n");