xref: /openbmc/linux/fs/nfs/write.c (revision 62ed948cb1405fe95d61d8c6445c102e0c9da0a6)
1 /*
2  * linux/fs/nfs/write.c
3  *
4  * Writing file data over NFS.
5  *
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.
12  *
13  * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
14  *
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.
20  *
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.
24  *
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
27  * cases:
28  *
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
32  *
33  * Also note that because of the way pages are invalidated in
34  * nfs_revalidate_inode, the following assertions hold:
35  *
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.
40  *
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.
45  *
46  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
47  */
48 
49 #include <linux/config.h>
50 #include <linux/types.h>
51 #include <linux/slab.h>
52 #include <linux/mm.h>
53 #include <linux/pagemap.h>
54 #include <linux/file.h>
55 #include <linux/mpage.h>
56 #include <linux/writeback.h>
57 
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>
64 
65 #include "delegation.h"
66 #include "iostat.h"
67 
68 #define NFSDBG_FACILITY		NFSDBG_PAGECACHE
69 
70 #define MIN_POOL_WRITE		(32)
71 #define MIN_POOL_COMMIT		(4)
72 
73 /*
74  * Local function declarations
75  */
76 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
77 					    struct inode *,
78 					    struct page *,
79 					    unsigned int, unsigned int);
80 static int nfs_wait_on_write_congestion(struct address_space *, int);
81 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
82 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
83 			   unsigned int npages, int how);
84 static const struct rpc_call_ops nfs_write_partial_ops;
85 static const struct rpc_call_ops nfs_write_full_ops;
86 static const struct rpc_call_ops nfs_commit_ops;
87 
88 static kmem_cache_t *nfs_wdata_cachep;
89 static mempool_t *nfs_wdata_mempool;
90 static mempool_t *nfs_commit_mempool;
91 
92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
93 
94 struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
95 {
96 	struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
97 
98 	if (p) {
99 		memset(p, 0, sizeof(*p));
100 		INIT_LIST_HEAD(&p->pages);
101 		if (pagecount <= ARRAY_SIZE(p->page_array))
102 			p->pagevec = p->page_array;
103 		else {
104 			p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
105 			if (!p->pagevec) {
106 				mempool_free(p, nfs_commit_mempool);
107 				p = NULL;
108 			}
109 		}
110 	}
111 	return p;
112 }
113 
114 void nfs_commit_free(struct nfs_write_data *p)
115 {
116 	if (p && (p->pagevec != &p->page_array[0]))
117 		kfree(p->pagevec);
118 	mempool_free(p, nfs_commit_mempool);
119 }
120 
121 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
122 {
123 	struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
124 
125 	if (p) {
126 		memset(p, 0, sizeof(*p));
127 		INIT_LIST_HEAD(&p->pages);
128 		if (pagecount <= ARRAY_SIZE(p->page_array))
129 			p->pagevec = p->page_array;
130 		else {
131 			p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
132 			if (!p->pagevec) {
133 				mempool_free(p, nfs_wdata_mempool);
134 				p = NULL;
135 			}
136 		}
137 	}
138 	return p;
139 }
140 
141 void nfs_writedata_free(struct nfs_write_data *p)
142 {
143 	if (p && (p->pagevec != &p->page_array[0]))
144 		kfree(p->pagevec);
145 	mempool_free(p, nfs_wdata_mempool);
146 }
147 
148 void nfs_writedata_release(void *wdata)
149 {
150 	nfs_writedata_free(wdata);
151 }
152 
153 /* Adjust the file length if we're writing beyond the end */
154 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
155 {
156 	struct inode *inode = page->mapping->host;
157 	loff_t end, i_size = i_size_read(inode);
158 	unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
159 
160 	if (i_size > 0 && page->index < end_index)
161 		return;
162 	end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
163 	if (i_size >= end)
164 		return;
165 	nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
166 	i_size_write(inode, end);
167 }
168 
169 /* We can set the PG_uptodate flag if we see that a write request
170  * covers the full page.
171  */
172 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
173 {
174 	loff_t end_offs;
175 
176 	if (PageUptodate(page))
177 		return;
178 	if (base != 0)
179 		return;
180 	if (count == PAGE_CACHE_SIZE) {
181 		SetPageUptodate(page);
182 		return;
183 	}
184 
185 	end_offs = i_size_read(page->mapping->host) - 1;
186 	if (end_offs < 0)
187 		return;
188 	/* Is this the last page? */
189 	if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
190 		return;
191 	/* This is the last page: set PG_uptodate if we cover the entire
192 	 * extent of the data, then zero the rest of the page.
193 	 */
194 	if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
195 		memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
196 		SetPageUptodate(page);
197 	}
198 }
199 
200 /*
201  * Write a page synchronously.
202  * Offset is the data offset within the page.
203  */
204 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
205 		struct page *page, unsigned int offset, unsigned int count,
206 		int how)
207 {
208 	unsigned int	wsize = NFS_SERVER(inode)->wsize;
209 	int		result, written = 0;
210 	struct nfs_write_data *wdata;
211 
212 	wdata = nfs_writedata_alloc(1);
213 	if (!wdata)
214 		return -ENOMEM;
215 
216 	wdata->flags = how;
217 	wdata->cred = ctx->cred;
218 	wdata->inode = inode;
219 	wdata->args.fh = NFS_FH(inode);
220 	wdata->args.context = ctx;
221 	wdata->args.pages = &page;
222 	wdata->args.stable = NFS_FILE_SYNC;
223 	wdata->args.pgbase = offset;
224 	wdata->args.count = wsize;
225 	wdata->res.fattr = &wdata->fattr;
226 	wdata->res.verf = &wdata->verf;
227 
228 	dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
229 		inode->i_sb->s_id,
230 		(long long)NFS_FILEID(inode),
231 		count, (long long)(page_offset(page) + offset));
232 
233 	set_page_writeback(page);
234 	nfs_begin_data_update(inode);
235 	do {
236 		if (count < wsize)
237 			wdata->args.count = count;
238 		wdata->args.offset = page_offset(page) + wdata->args.pgbase;
239 
240 		result = NFS_PROTO(inode)->write(wdata);
241 
242 		if (result < 0) {
243 			/* Must mark the page invalid after I/O error */
244 			ClearPageUptodate(page);
245 			goto io_error;
246 		}
247 		if (result < wdata->args.count)
248 			printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
249 					wdata->args.count, result);
250 
251 		wdata->args.offset += result;
252 	        wdata->args.pgbase += result;
253 		written += result;
254 		count -= result;
255 		nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
256 	} while (count);
257 	/* Update file length */
258 	nfs_grow_file(page, offset, written);
259 	/* Set the PG_uptodate flag? */
260 	nfs_mark_uptodate(page, offset, written);
261 
262 	if (PageError(page))
263 		ClearPageError(page);
264 
265 io_error:
266 	nfs_end_data_update(inode);
267 	end_page_writeback(page);
268 	nfs_writedata_free(wdata);
269 	return written ? written : result;
270 }
271 
272 static int nfs_writepage_async(struct nfs_open_context *ctx,
273 		struct inode *inode, struct page *page,
274 		unsigned int offset, unsigned int count)
275 {
276 	struct nfs_page	*req;
277 
278 	req = nfs_update_request(ctx, inode, page, offset, count);
279 	if (IS_ERR(req))
280 		return PTR_ERR(req);
281 	/* Update file length */
282 	nfs_grow_file(page, offset, count);
283 	/* Set the PG_uptodate flag? */
284 	nfs_mark_uptodate(page, offset, count);
285 	nfs_unlock_request(req);
286 	return 0;
287 }
288 
289 static int wb_priority(struct writeback_control *wbc)
290 {
291 	if (wbc->for_reclaim)
292 		return FLUSH_HIGHPRI;
293 	if (wbc->for_kupdate)
294 		return FLUSH_LOWPRI;
295 	return 0;
296 }
297 
298 /*
299  * Write an mmapped page to the server.
300  */
301 int nfs_writepage(struct page *page, struct writeback_control *wbc)
302 {
303 	struct nfs_open_context *ctx;
304 	struct inode *inode = page->mapping->host;
305 	unsigned long end_index;
306 	unsigned offset = PAGE_CACHE_SIZE;
307 	loff_t i_size = i_size_read(inode);
308 	int inode_referenced = 0;
309 	int priority = wb_priority(wbc);
310 	int err;
311 
312 	nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
313 	nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
314 
315 	/*
316 	 * Note: We need to ensure that we have a reference to the inode
317 	 *       if we are to do asynchronous writes. If not, waiting
318 	 *       in nfs_wait_on_request() may deadlock with clear_inode().
319 	 *
320 	 *       If igrab() fails here, then it is in any case safe to
321 	 *       call nfs_wb_page(), since there will be no pending writes.
322 	 */
323 	if (igrab(inode) != 0)
324 		inode_referenced = 1;
325 	end_index = i_size >> PAGE_CACHE_SHIFT;
326 
327 	/* Ensure we've flushed out any previous writes */
328 	nfs_wb_page_priority(inode, page, priority);
329 
330 	/* easy case */
331 	if (page->index < end_index)
332 		goto do_it;
333 	/* things got complicated... */
334 	offset = i_size & (PAGE_CACHE_SIZE-1);
335 
336 	/* OK, are we completely out? */
337 	err = 0; /* potential race with truncate - ignore */
338 	if (page->index >= end_index+1 || !offset)
339 		goto out;
340 do_it:
341 	ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
342 	if (ctx == NULL) {
343 		err = -EBADF;
344 		goto out;
345 	}
346 	lock_kernel();
347 	if (!IS_SYNC(inode) && inode_referenced) {
348 		err = nfs_writepage_async(ctx, inode, page, 0, offset);
349 		if (!wbc->for_writepages)
350 			nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
351 	} else {
352 		err = nfs_writepage_sync(ctx, inode, page, 0,
353 						offset, priority);
354 		if (err >= 0) {
355 			if (err != offset)
356 				redirty_page_for_writepage(wbc, page);
357 			err = 0;
358 		}
359 	}
360 	unlock_kernel();
361 	put_nfs_open_context(ctx);
362 out:
363 	unlock_page(page);
364 	if (inode_referenced)
365 		iput(inode);
366 	return err;
367 }
368 
369 /*
370  * Note: causes nfs_update_request() to block on the assumption
371  * 	 that the writeback is generated due to memory pressure.
372  */
373 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
374 {
375 	struct backing_dev_info *bdi = mapping->backing_dev_info;
376 	struct inode *inode = mapping->host;
377 	int err;
378 
379 	nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
380 
381 	err = generic_writepages(mapping, wbc);
382 	if (err)
383 		return err;
384 	while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
385 		if (wbc->nonblocking)
386 			return 0;
387 		nfs_wait_on_write_congestion(mapping, 0);
388 	}
389 	err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
390 	if (err < 0)
391 		goto out;
392 	nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
393 	wbc->nr_to_write -= err;
394 	if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
395 		err = nfs_wait_on_requests(inode, 0, 0);
396 		if (err < 0)
397 			goto out;
398 	}
399 	err = nfs_commit_inode(inode, wb_priority(wbc));
400 	if (err > 0) {
401 		wbc->nr_to_write -= err;
402 		err = 0;
403 	}
404 out:
405 	clear_bit(BDI_write_congested, &bdi->state);
406 	wake_up_all(&nfs_write_congestion);
407 	return err;
408 }
409 
410 /*
411  * Insert a write request into an inode
412  */
413 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
414 {
415 	struct nfs_inode *nfsi = NFS_I(inode);
416 	int error;
417 
418 	error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
419 	BUG_ON(error == -EEXIST);
420 	if (error)
421 		return error;
422 	if (!nfsi->npages) {
423 		igrab(inode);
424 		nfs_begin_data_update(inode);
425 		if (nfs_have_delegation(inode, FMODE_WRITE))
426 			nfsi->change_attr++;
427 	}
428 	SetPagePrivate(req->wb_page);
429 	nfsi->npages++;
430 	atomic_inc(&req->wb_count);
431 	return 0;
432 }
433 
434 /*
435  * Insert a write request into an inode
436  */
437 static void nfs_inode_remove_request(struct nfs_page *req)
438 {
439 	struct inode *inode = req->wb_context->dentry->d_inode;
440 	struct nfs_inode *nfsi = NFS_I(inode);
441 
442 	BUG_ON (!NFS_WBACK_BUSY(req));
443 
444 	spin_lock(&nfsi->req_lock);
445 	ClearPagePrivate(req->wb_page);
446 	radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
447 	nfsi->npages--;
448 	if (!nfsi->npages) {
449 		spin_unlock(&nfsi->req_lock);
450 		nfs_end_data_update(inode);
451 		iput(inode);
452 	} else
453 		spin_unlock(&nfsi->req_lock);
454 	nfs_clear_request(req);
455 	nfs_release_request(req);
456 }
457 
458 /*
459  * Find a request
460  */
461 static inline struct nfs_page *
462 _nfs_find_request(struct inode *inode, unsigned long index)
463 {
464 	struct nfs_inode *nfsi = NFS_I(inode);
465 	struct nfs_page *req;
466 
467 	req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
468 	if (req)
469 		atomic_inc(&req->wb_count);
470 	return req;
471 }
472 
473 static struct nfs_page *
474 nfs_find_request(struct inode *inode, unsigned long index)
475 {
476 	struct nfs_page		*req;
477 	struct nfs_inode	*nfsi = NFS_I(inode);
478 
479 	spin_lock(&nfsi->req_lock);
480 	req = _nfs_find_request(inode, index);
481 	spin_unlock(&nfsi->req_lock);
482 	return req;
483 }
484 
485 /*
486  * Add a request to the inode's dirty list.
487  */
488 static void
489 nfs_mark_request_dirty(struct nfs_page *req)
490 {
491 	struct inode *inode = req->wb_context->dentry->d_inode;
492 	struct nfs_inode *nfsi = NFS_I(inode);
493 
494 	spin_lock(&nfsi->req_lock);
495 	radix_tree_tag_set(&nfsi->nfs_page_tree,
496 			req->wb_index, NFS_PAGE_TAG_DIRTY);
497 	nfs_list_add_request(req, &nfsi->dirty);
498 	nfsi->ndirty++;
499 	spin_unlock(&nfsi->req_lock);
500 	inc_page_state(nr_dirty);
501 	mark_inode_dirty(inode);
502 }
503 
504 /*
505  * Check if a request is dirty
506  */
507 static inline int
508 nfs_dirty_request(struct nfs_page *req)
509 {
510 	struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
511 	return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
512 }
513 
514 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
515 /*
516  * Add a request to the inode's commit list.
517  */
518 static void
519 nfs_mark_request_commit(struct nfs_page *req)
520 {
521 	struct inode *inode = req->wb_context->dentry->d_inode;
522 	struct nfs_inode *nfsi = NFS_I(inode);
523 
524 	spin_lock(&nfsi->req_lock);
525 	nfs_list_add_request(req, &nfsi->commit);
526 	nfsi->ncommit++;
527 	spin_unlock(&nfsi->req_lock);
528 	inc_page_state(nr_unstable);
529 	mark_inode_dirty(inode);
530 }
531 #endif
532 
533 /*
534  * Wait for a request to complete.
535  *
536  * Interruptible by signals only if mounted with intr flag.
537  */
538 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
539 {
540 	struct nfs_inode *nfsi = NFS_I(inode);
541 	struct nfs_page *req;
542 	unsigned long		idx_end, next;
543 	unsigned int		res = 0;
544 	int			error;
545 
546 	if (npages == 0)
547 		idx_end = ~0;
548 	else
549 		idx_end = idx_start + npages - 1;
550 
551 	next = idx_start;
552 	while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
553 		if (req->wb_index > idx_end)
554 			break;
555 
556 		next = req->wb_index + 1;
557 		BUG_ON(!NFS_WBACK_BUSY(req));
558 
559 		atomic_inc(&req->wb_count);
560 		spin_unlock(&nfsi->req_lock);
561 		error = nfs_wait_on_request(req);
562 		nfs_release_request(req);
563 		spin_lock(&nfsi->req_lock);
564 		if (error < 0)
565 			return error;
566 		res++;
567 	}
568 	return res;
569 }
570 
571 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
572 {
573 	struct nfs_inode *nfsi = NFS_I(inode);
574 	int ret;
575 
576 	spin_lock(&nfsi->req_lock);
577 	ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
578 	spin_unlock(&nfsi->req_lock);
579 	return ret;
580 }
581 
582 static void nfs_cancel_requests(struct list_head *head)
583 {
584 	struct nfs_page *req;
585 	while(!list_empty(head)) {
586 		req = nfs_list_entry(head->next);
587 		nfs_list_remove_request(req);
588 		nfs_inode_remove_request(req);
589 		nfs_clear_page_writeback(req);
590 	}
591 }
592 
593 /*
594  * nfs_scan_dirty - Scan an inode for dirty requests
595  * @inode: NFS inode to scan
596  * @dst: destination list
597  * @idx_start: lower bound of page->index to scan.
598  * @npages: idx_start + npages sets the upper bound to scan.
599  *
600  * Moves requests from the inode's dirty page list.
601  * The requests are *not* checked to ensure that they form a contiguous set.
602  */
603 static int
604 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
605 {
606 	struct nfs_inode *nfsi = NFS_I(inode);
607 	int res = 0;
608 
609 	if (nfsi->ndirty != 0) {
610 		res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
611 		nfsi->ndirty -= res;
612 		sub_page_state(nr_dirty,res);
613 		if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
614 			printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
615 	}
616 	return res;
617 }
618 
619 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
620 /*
621  * nfs_scan_commit - Scan an inode for commit requests
622  * @inode: NFS inode to scan
623  * @dst: destination list
624  * @idx_start: lower bound of page->index to scan.
625  * @npages: idx_start + npages sets the upper bound to scan.
626  *
627  * Moves requests from the inode's 'commit' request list.
628  * The requests are *not* checked to ensure that they form a contiguous set.
629  */
630 static int
631 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
632 {
633 	struct nfs_inode *nfsi = NFS_I(inode);
634 	int res = 0;
635 
636 	if (nfsi->ncommit != 0) {
637 		res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
638 		nfsi->ncommit -= res;
639 		if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
640 			printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
641 	}
642 	return res;
643 }
644 #else
645 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
646 {
647 	return 0;
648 }
649 #endif
650 
651 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
652 {
653 	struct backing_dev_info *bdi = mapping->backing_dev_info;
654 	DEFINE_WAIT(wait);
655 	int ret = 0;
656 
657 	might_sleep();
658 
659 	if (!bdi_write_congested(bdi))
660 		return 0;
661 
662 	nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
663 
664 	if (intr) {
665 		struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
666 		sigset_t oldset;
667 
668 		rpc_clnt_sigmask(clnt, &oldset);
669 		prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
670 		if (bdi_write_congested(bdi)) {
671 			if (signalled())
672 				ret = -ERESTARTSYS;
673 			else
674 				schedule();
675 		}
676 		rpc_clnt_sigunmask(clnt, &oldset);
677 	} else {
678 		prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
679 		if (bdi_write_congested(bdi))
680 			schedule();
681 	}
682 	finish_wait(&nfs_write_congestion, &wait);
683 	return ret;
684 }
685 
686 
687 /*
688  * Try to update any existing write request, or create one if there is none.
689  * In order to match, the request's credentials must match those of
690  * the calling process.
691  *
692  * Note: Should always be called with the Page Lock held!
693  */
694 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
695 		struct inode *inode, struct page *page,
696 		unsigned int offset, unsigned int bytes)
697 {
698 	struct nfs_server *server = NFS_SERVER(inode);
699 	struct nfs_inode *nfsi = NFS_I(inode);
700 	struct nfs_page		*req, *new = NULL;
701 	unsigned long		rqend, end;
702 
703 	end = offset + bytes;
704 
705 	if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
706 		return ERR_PTR(-ERESTARTSYS);
707 	for (;;) {
708 		/* Loop over all inode entries and see if we find
709 		 * A request for the page we wish to update
710 		 */
711 		spin_lock(&nfsi->req_lock);
712 		req = _nfs_find_request(inode, page->index);
713 		if (req) {
714 			if (!nfs_lock_request_dontget(req)) {
715 				int error;
716 				spin_unlock(&nfsi->req_lock);
717 				error = nfs_wait_on_request(req);
718 				nfs_release_request(req);
719 				if (error < 0) {
720 					if (new)
721 						nfs_release_request(new);
722 					return ERR_PTR(error);
723 				}
724 				continue;
725 			}
726 			spin_unlock(&nfsi->req_lock);
727 			if (new)
728 				nfs_release_request(new);
729 			break;
730 		}
731 
732 		if (new) {
733 			int error;
734 			nfs_lock_request_dontget(new);
735 			error = nfs_inode_add_request(inode, new);
736 			if (error) {
737 				spin_unlock(&nfsi->req_lock);
738 				nfs_unlock_request(new);
739 				return ERR_PTR(error);
740 			}
741 			spin_unlock(&nfsi->req_lock);
742 			nfs_mark_request_dirty(new);
743 			return new;
744 		}
745 		spin_unlock(&nfsi->req_lock);
746 
747 		new = nfs_create_request(ctx, inode, page, offset, bytes);
748 		if (IS_ERR(new))
749 			return new;
750 	}
751 
752 	/* We have a request for our page.
753 	 * If the creds don't match, or the
754 	 * page addresses don't match,
755 	 * tell the caller to wait on the conflicting
756 	 * request.
757 	 */
758 	rqend = req->wb_offset + req->wb_bytes;
759 	if (req->wb_context != ctx
760 	    || req->wb_page != page
761 	    || !nfs_dirty_request(req)
762 	    || offset > rqend || end < req->wb_offset) {
763 		nfs_unlock_request(req);
764 		return ERR_PTR(-EBUSY);
765 	}
766 
767 	/* Okay, the request matches. Update the region */
768 	if (offset < req->wb_offset) {
769 		req->wb_offset = offset;
770 		req->wb_pgbase = offset;
771 		req->wb_bytes = rqend - req->wb_offset;
772 	}
773 
774 	if (end > rqend)
775 		req->wb_bytes = end - req->wb_offset;
776 
777 	return req;
778 }
779 
780 int nfs_flush_incompatible(struct file *file, struct page *page)
781 {
782 	struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
783 	struct inode	*inode = page->mapping->host;
784 	struct nfs_page	*req;
785 	int		status = 0;
786 	/*
787 	 * Look for a request corresponding to this page. If there
788 	 * is one, and it belongs to another file, we flush it out
789 	 * before we try to copy anything into the page. Do this
790 	 * due to the lack of an ACCESS-type call in NFSv2.
791 	 * Also do the same if we find a request from an existing
792 	 * dropped page.
793 	 */
794 	req = nfs_find_request(inode, page->index);
795 	if (req) {
796 		if (req->wb_page != page || ctx != req->wb_context)
797 			status = nfs_wb_page(inode, page);
798 		nfs_release_request(req);
799 	}
800 	return (status < 0) ? status : 0;
801 }
802 
803 /*
804  * Update and possibly write a cached page of an NFS file.
805  *
806  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
807  * things with a page scheduled for an RPC call (e.g. invalidate it).
808  */
809 int nfs_updatepage(struct file *file, struct page *page,
810 		unsigned int offset, unsigned int count)
811 {
812 	struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
813 	struct inode	*inode = page->mapping->host;
814 	struct nfs_page	*req;
815 	int		status = 0;
816 
817 	nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
818 
819 	dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
820 		file->f_dentry->d_parent->d_name.name,
821 		file->f_dentry->d_name.name, count,
822 		(long long)(page_offset(page) +offset));
823 
824 	if (IS_SYNC(inode)) {
825 		status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
826 		if (status > 0) {
827 			if (offset == 0 && status == PAGE_CACHE_SIZE)
828 				SetPageUptodate(page);
829 			return 0;
830 		}
831 		return status;
832 	}
833 
834 	/* If we're not using byte range locks, and we know the page
835 	 * is entirely in cache, it may be more efficient to avoid
836 	 * fragmenting write requests.
837 	 */
838 	if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
839 		loff_t end_offs = i_size_read(inode) - 1;
840 		unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
841 
842 		count += offset;
843 		offset = 0;
844 		if (unlikely(end_offs < 0)) {
845 			/* Do nothing */
846 		} else if (page->index == end_index) {
847 			unsigned int pglen;
848 			pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
849 			if (count < pglen)
850 				count = pglen;
851 		} else if (page->index < end_index)
852 			count = PAGE_CACHE_SIZE;
853 	}
854 
855 	/*
856 	 * Try to find an NFS request corresponding to this page
857 	 * and update it.
858 	 * If the existing request cannot be updated, we must flush
859 	 * it out now.
860 	 */
861 	do {
862 		req = nfs_update_request(ctx, inode, page, offset, count);
863 		status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
864 		if (status != -EBUSY)
865 			break;
866 		/* Request could not be updated. Flush it out and try again */
867 		status = nfs_wb_page(inode, page);
868 	} while (status >= 0);
869 	if (status < 0)
870 		goto done;
871 
872 	status = 0;
873 
874 	/* Update file length */
875 	nfs_grow_file(page, offset, count);
876 	/* Set the PG_uptodate flag? */
877 	nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
878 	nfs_unlock_request(req);
879 done:
880         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
881 			status, (long long)i_size_read(inode));
882 	if (status < 0)
883 		ClearPageUptodate(page);
884 	return status;
885 }
886 
887 static void nfs_writepage_release(struct nfs_page *req)
888 {
889 	end_page_writeback(req->wb_page);
890 
891 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
892 	if (!PageError(req->wb_page)) {
893 		if (NFS_NEED_RESCHED(req)) {
894 			nfs_mark_request_dirty(req);
895 			goto out;
896 		} else if (NFS_NEED_COMMIT(req)) {
897 			nfs_mark_request_commit(req);
898 			goto out;
899 		}
900 	}
901 	nfs_inode_remove_request(req);
902 
903 out:
904 	nfs_clear_commit(req);
905 	nfs_clear_reschedule(req);
906 #else
907 	nfs_inode_remove_request(req);
908 #endif
909 	nfs_clear_page_writeback(req);
910 }
911 
912 static inline int flush_task_priority(int how)
913 {
914 	switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
915 		case FLUSH_HIGHPRI:
916 			return RPC_PRIORITY_HIGH;
917 		case FLUSH_LOWPRI:
918 			return RPC_PRIORITY_LOW;
919 	}
920 	return RPC_PRIORITY_NORMAL;
921 }
922 
923 /*
924  * Set up the argument/result storage required for the RPC call.
925  */
926 static void nfs_write_rpcsetup(struct nfs_page *req,
927 		struct nfs_write_data *data,
928 		const struct rpc_call_ops *call_ops,
929 		unsigned int count, unsigned int offset,
930 		int how)
931 {
932 	struct inode		*inode;
933 	int flags;
934 
935 	/* Set up the RPC argument and reply structs
936 	 * NB: take care not to mess about with data->commit et al. */
937 
938 	data->req = req;
939 	data->inode = inode = req->wb_context->dentry->d_inode;
940 	data->cred = req->wb_context->cred;
941 
942 	data->args.fh     = NFS_FH(inode);
943 	data->args.offset = req_offset(req) + offset;
944 	data->args.pgbase = req->wb_pgbase + offset;
945 	data->args.pages  = data->pagevec;
946 	data->args.count  = count;
947 	data->args.context = req->wb_context;
948 
949 	data->res.fattr   = &data->fattr;
950 	data->res.count   = count;
951 	data->res.verf    = &data->verf;
952 	nfs_fattr_init(&data->fattr);
953 
954 	/* Set up the initial task struct.  */
955 	flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
956 	rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
957 	NFS_PROTO(inode)->write_setup(data, how);
958 
959 	data->task.tk_priority = flush_task_priority(how);
960 	data->task.tk_cookie = (unsigned long)inode;
961 
962 	dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
963 		data->task.tk_pid,
964 		inode->i_sb->s_id,
965 		(long long)NFS_FILEID(inode),
966 		count,
967 		(unsigned long long)data->args.offset);
968 }
969 
970 static void nfs_execute_write(struct nfs_write_data *data)
971 {
972 	struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
973 	sigset_t oldset;
974 
975 	rpc_clnt_sigmask(clnt, &oldset);
976 	lock_kernel();
977 	rpc_execute(&data->task);
978 	unlock_kernel();
979 	rpc_clnt_sigunmask(clnt, &oldset);
980 }
981 
982 /*
983  * Generate multiple small requests to write out a single
984  * contiguous dirty area on one page.
985  */
986 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
987 {
988 	struct nfs_page *req = nfs_list_entry(head->next);
989 	struct page *page = req->wb_page;
990 	struct nfs_write_data *data;
991 	unsigned int wsize = NFS_SERVER(inode)->wsize;
992 	unsigned int nbytes, offset;
993 	int requests = 0;
994 	LIST_HEAD(list);
995 
996 	nfs_list_remove_request(req);
997 
998 	nbytes = req->wb_bytes;
999 	for (;;) {
1000 		data = nfs_writedata_alloc(1);
1001 		if (!data)
1002 			goto out_bad;
1003 		list_add(&data->pages, &list);
1004 		requests++;
1005 		if (nbytes <= wsize)
1006 			break;
1007 		nbytes -= wsize;
1008 	}
1009 	atomic_set(&req->wb_complete, requests);
1010 
1011 	ClearPageError(page);
1012 	set_page_writeback(page);
1013 	offset = 0;
1014 	nbytes = req->wb_bytes;
1015 	do {
1016 		data = list_entry(list.next, struct nfs_write_data, pages);
1017 		list_del_init(&data->pages);
1018 
1019 		data->pagevec[0] = page;
1020 
1021 		if (nbytes > wsize) {
1022 			nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1023 					wsize, offset, how);
1024 			offset += wsize;
1025 			nbytes -= wsize;
1026 		} else {
1027 			nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1028 					nbytes, offset, how);
1029 			nbytes = 0;
1030 		}
1031 		nfs_execute_write(data);
1032 	} while (nbytes != 0);
1033 
1034 	return 0;
1035 
1036 out_bad:
1037 	while (!list_empty(&list)) {
1038 		data = list_entry(list.next, struct nfs_write_data, pages);
1039 		list_del(&data->pages);
1040 		nfs_writedata_free(data);
1041 	}
1042 	nfs_mark_request_dirty(req);
1043 	nfs_clear_page_writeback(req);
1044 	return -ENOMEM;
1045 }
1046 
1047 /*
1048  * Create an RPC task for the given write request and kick it.
1049  * The page must have been locked by the caller.
1050  *
1051  * It may happen that the page we're passed is not marked dirty.
1052  * This is the case if nfs_updatepage detects a conflicting request
1053  * that has been written but not committed.
1054  */
1055 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
1056 {
1057 	struct nfs_page		*req;
1058 	struct page		**pages;
1059 	struct nfs_write_data	*data;
1060 	unsigned int		count;
1061 
1062 	data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
1063 	if (!data)
1064 		goto out_bad;
1065 
1066 	pages = data->pagevec;
1067 	count = 0;
1068 	while (!list_empty(head)) {
1069 		req = nfs_list_entry(head->next);
1070 		nfs_list_remove_request(req);
1071 		nfs_list_add_request(req, &data->pages);
1072 		ClearPageError(req->wb_page);
1073 		set_page_writeback(req->wb_page);
1074 		*pages++ = req->wb_page;
1075 		count += req->wb_bytes;
1076 	}
1077 	req = nfs_list_entry(data->pages.next);
1078 
1079 	/* Set up the argument struct */
1080 	nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1081 
1082 	nfs_execute_write(data);
1083 	return 0;
1084  out_bad:
1085 	while (!list_empty(head)) {
1086 		struct nfs_page *req = nfs_list_entry(head->next);
1087 		nfs_list_remove_request(req);
1088 		nfs_mark_request_dirty(req);
1089 		nfs_clear_page_writeback(req);
1090 	}
1091 	return -ENOMEM;
1092 }
1093 
1094 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1095 {
1096 	LIST_HEAD(one_request);
1097 	int (*flush_one)(struct inode *, struct list_head *, int);
1098 	struct nfs_page	*req;
1099 	int wpages = NFS_SERVER(inode)->wpages;
1100 	int wsize = NFS_SERVER(inode)->wsize;
1101 	int error;
1102 
1103 	flush_one = nfs_flush_one;
1104 	if (wsize < PAGE_CACHE_SIZE)
1105 		flush_one = nfs_flush_multi;
1106 	/* For single writes, FLUSH_STABLE is more efficient */
1107 	if (npages <= wpages && npages == NFS_I(inode)->npages
1108 			&& nfs_list_entry(head->next)->wb_bytes <= wsize)
1109 		how |= FLUSH_STABLE;
1110 
1111 	do {
1112 		nfs_coalesce_requests(head, &one_request, wpages);
1113 		req = nfs_list_entry(one_request.next);
1114 		error = flush_one(inode, &one_request, how);
1115 		if (error < 0)
1116 			goto out_err;
1117 	} while (!list_empty(head));
1118 	return 0;
1119 out_err:
1120 	while (!list_empty(head)) {
1121 		req = nfs_list_entry(head->next);
1122 		nfs_list_remove_request(req);
1123 		nfs_mark_request_dirty(req);
1124 		nfs_clear_page_writeback(req);
1125 	}
1126 	return error;
1127 }
1128 
1129 /*
1130  * Handle a write reply that flushed part of a page.
1131  */
1132 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1133 {
1134 	struct nfs_write_data	*data = calldata;
1135 	struct nfs_page		*req = data->req;
1136 	struct page		*page = req->wb_page;
1137 
1138 	dprintk("NFS: write (%s/%Ld %d@%Ld)",
1139 		req->wb_context->dentry->d_inode->i_sb->s_id,
1140 		(long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1141 		req->wb_bytes,
1142 		(long long)req_offset(req));
1143 
1144 	if (nfs_writeback_done(task, data) != 0)
1145 		return;
1146 
1147 	if (task->tk_status < 0) {
1148 		ClearPageUptodate(page);
1149 		SetPageError(page);
1150 		req->wb_context->error = task->tk_status;
1151 		dprintk(", error = %d\n", task->tk_status);
1152 	} else {
1153 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1154 		if (data->verf.committed < NFS_FILE_SYNC) {
1155 			if (!NFS_NEED_COMMIT(req)) {
1156 				nfs_defer_commit(req);
1157 				memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1158 				dprintk(" defer commit\n");
1159 			} else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1160 				nfs_defer_reschedule(req);
1161 				dprintk(" server reboot detected\n");
1162 			}
1163 		} else
1164 #endif
1165 			dprintk(" OK\n");
1166 	}
1167 
1168 	if (atomic_dec_and_test(&req->wb_complete))
1169 		nfs_writepage_release(req);
1170 }
1171 
1172 static const struct rpc_call_ops nfs_write_partial_ops = {
1173 	.rpc_call_done = nfs_writeback_done_partial,
1174 	.rpc_release = nfs_writedata_release,
1175 };
1176 
1177 /*
1178  * Handle a write reply that flushes a whole page.
1179  *
1180  * FIXME: There is an inherent race with invalidate_inode_pages and
1181  *	  writebacks since the page->count is kept > 1 for as long
1182  *	  as the page has a write request pending.
1183  */
1184 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1185 {
1186 	struct nfs_write_data	*data = calldata;
1187 	struct nfs_page		*req;
1188 	struct page		*page;
1189 
1190 	if (nfs_writeback_done(task, data) != 0)
1191 		return;
1192 
1193 	/* Update attributes as result of writeback. */
1194 	while (!list_empty(&data->pages)) {
1195 		req = nfs_list_entry(data->pages.next);
1196 		nfs_list_remove_request(req);
1197 		page = req->wb_page;
1198 
1199 		dprintk("NFS: write (%s/%Ld %d@%Ld)",
1200 			req->wb_context->dentry->d_inode->i_sb->s_id,
1201 			(long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1202 			req->wb_bytes,
1203 			(long long)req_offset(req));
1204 
1205 		if (task->tk_status < 0) {
1206 			ClearPageUptodate(page);
1207 			SetPageError(page);
1208 			req->wb_context->error = task->tk_status;
1209 			end_page_writeback(page);
1210 			nfs_inode_remove_request(req);
1211 			dprintk(", error = %d\n", task->tk_status);
1212 			goto next;
1213 		}
1214 		end_page_writeback(page);
1215 
1216 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1217 		if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1218 			nfs_inode_remove_request(req);
1219 			dprintk(" OK\n");
1220 			goto next;
1221 		}
1222 		memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1223 		nfs_mark_request_commit(req);
1224 		dprintk(" marked for commit\n");
1225 #else
1226 		nfs_inode_remove_request(req);
1227 #endif
1228 	next:
1229 		nfs_clear_page_writeback(req);
1230 	}
1231 }
1232 
1233 static const struct rpc_call_ops nfs_write_full_ops = {
1234 	.rpc_call_done = nfs_writeback_done_full,
1235 	.rpc_release = nfs_writedata_release,
1236 };
1237 
1238 
1239 /*
1240  * This function is called when the WRITE call is complete.
1241  */
1242 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1243 {
1244 	struct nfs_writeargs	*argp = &data->args;
1245 	struct nfs_writeres	*resp = &data->res;
1246 	int status;
1247 
1248 	dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1249 		task->tk_pid, task->tk_status);
1250 
1251 	/* Call the NFS version-specific code */
1252 	status = NFS_PROTO(data->inode)->write_done(task, data);
1253 	if (status != 0)
1254 		return status;
1255 	nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1256 
1257 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1258 	if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1259 		/* We tried a write call, but the server did not
1260 		 * commit data to stable storage even though we
1261 		 * requested it.
1262 		 * Note: There is a known bug in Tru64 < 5.0 in which
1263 		 *	 the server reports NFS_DATA_SYNC, but performs
1264 		 *	 NFS_FILE_SYNC. We therefore implement this checking
1265 		 *	 as a dprintk() in order to avoid filling syslog.
1266 		 */
1267 		static unsigned long    complain;
1268 
1269 		if (time_before(complain, jiffies)) {
1270 			dprintk("NFS: faulty NFS server %s:"
1271 				" (committed = %d) != (stable = %d)\n",
1272 				NFS_SERVER(data->inode)->hostname,
1273 				resp->verf->committed, argp->stable);
1274 			complain = jiffies + 300 * HZ;
1275 		}
1276 	}
1277 #endif
1278 	/* Is this a short write? */
1279 	if (task->tk_status >= 0 && resp->count < argp->count) {
1280 		static unsigned long    complain;
1281 
1282 		nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1283 
1284 		/* Has the server at least made some progress? */
1285 		if (resp->count != 0) {
1286 			/* Was this an NFSv2 write or an NFSv3 stable write? */
1287 			if (resp->verf->committed != NFS_UNSTABLE) {
1288 				/* Resend from where the server left off */
1289 				argp->offset += resp->count;
1290 				argp->pgbase += resp->count;
1291 				argp->count -= resp->count;
1292 			} else {
1293 				/* Resend as a stable write in order to avoid
1294 				 * headaches in the case of a server crash.
1295 				 */
1296 				argp->stable = NFS_FILE_SYNC;
1297 			}
1298 			rpc_restart_call(task);
1299 			return -EAGAIN;
1300 		}
1301 		if (time_before(complain, jiffies)) {
1302 			printk(KERN_WARNING
1303 			       "NFS: Server wrote zero bytes, expected %u.\n",
1304 					argp->count);
1305 			complain = jiffies + 300 * HZ;
1306 		}
1307 		/* Can't do anything about it except throw an error. */
1308 		task->tk_status = -EIO;
1309 	}
1310 	return 0;
1311 }
1312 
1313 
1314 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1315 void nfs_commit_release(void *wdata)
1316 {
1317 	nfs_commit_free(wdata);
1318 }
1319 
1320 /*
1321  * Set up the argument/result storage required for the RPC call.
1322  */
1323 static void nfs_commit_rpcsetup(struct list_head *head,
1324 		struct nfs_write_data *data,
1325 		int how)
1326 {
1327 	struct nfs_page		*first;
1328 	struct inode		*inode;
1329 	int flags;
1330 
1331 	/* Set up the RPC argument and reply structs
1332 	 * NB: take care not to mess about with data->commit et al. */
1333 
1334 	list_splice_init(head, &data->pages);
1335 	first = nfs_list_entry(data->pages.next);
1336 	inode = first->wb_context->dentry->d_inode;
1337 
1338 	data->inode	  = inode;
1339 	data->cred	  = first->wb_context->cred;
1340 
1341 	data->args.fh     = NFS_FH(data->inode);
1342 	/* Note: we always request a commit of the entire inode */
1343 	data->args.offset = 0;
1344 	data->args.count  = 0;
1345 	data->res.count   = 0;
1346 	data->res.fattr   = &data->fattr;
1347 	data->res.verf    = &data->verf;
1348 	nfs_fattr_init(&data->fattr);
1349 
1350 	/* Set up the initial task struct.  */
1351 	flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1352 	rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1353 	NFS_PROTO(inode)->commit_setup(data, how);
1354 
1355 	data->task.tk_priority = flush_task_priority(how);
1356 	data->task.tk_cookie = (unsigned long)inode;
1357 
1358 	dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1359 }
1360 
1361 /*
1362  * Commit dirty pages
1363  */
1364 static int
1365 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1366 {
1367 	struct nfs_write_data	*data;
1368 	struct nfs_page         *req;
1369 
1370 	data = nfs_commit_alloc(NFS_SERVER(inode)->wpages);
1371 
1372 	if (!data)
1373 		goto out_bad;
1374 
1375 	/* Set up the argument struct */
1376 	nfs_commit_rpcsetup(head, data, how);
1377 
1378 	nfs_execute_write(data);
1379 	return 0;
1380  out_bad:
1381 	while (!list_empty(head)) {
1382 		req = nfs_list_entry(head->next);
1383 		nfs_list_remove_request(req);
1384 		nfs_mark_request_commit(req);
1385 		nfs_clear_page_writeback(req);
1386 	}
1387 	return -ENOMEM;
1388 }
1389 
1390 /*
1391  * COMMIT call returned
1392  */
1393 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1394 {
1395 	struct nfs_write_data	*data = calldata;
1396 	struct nfs_page		*req;
1397 	int res = 0;
1398 
1399         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1400                                 task->tk_pid, task->tk_status);
1401 
1402 	/* Call the NFS version-specific code */
1403 	if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1404 		return;
1405 
1406 	while (!list_empty(&data->pages)) {
1407 		req = nfs_list_entry(data->pages.next);
1408 		nfs_list_remove_request(req);
1409 
1410 		dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1411 			req->wb_context->dentry->d_inode->i_sb->s_id,
1412 			(long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1413 			req->wb_bytes,
1414 			(long long)req_offset(req));
1415 		if (task->tk_status < 0) {
1416 			req->wb_context->error = task->tk_status;
1417 			nfs_inode_remove_request(req);
1418 			dprintk(", error = %d\n", task->tk_status);
1419 			goto next;
1420 		}
1421 
1422 		/* Okay, COMMIT succeeded, apparently. Check the verifier
1423 		 * returned by the server against all stored verfs. */
1424 		if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1425 			/* We have a match */
1426 			nfs_inode_remove_request(req);
1427 			dprintk(" OK\n");
1428 			goto next;
1429 		}
1430 		/* We have a mismatch. Write the page again */
1431 		dprintk(" mismatch\n");
1432 		nfs_mark_request_dirty(req);
1433 	next:
1434 		nfs_clear_page_writeback(req);
1435 		res++;
1436 	}
1437 	sub_page_state(nr_unstable,res);
1438 }
1439 
1440 static const struct rpc_call_ops nfs_commit_ops = {
1441 	.rpc_call_done = nfs_commit_done,
1442 	.rpc_release = nfs_commit_release,
1443 };
1444 #else
1445 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1446 {
1447 	return 0;
1448 }
1449 #endif
1450 
1451 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1452 			   unsigned int npages, int how)
1453 {
1454 	struct nfs_inode *nfsi = NFS_I(inode);
1455 	LIST_HEAD(head);
1456 	int res;
1457 
1458 	spin_lock(&nfsi->req_lock);
1459 	res = nfs_scan_dirty(inode, &head, idx_start, npages);
1460 	spin_unlock(&nfsi->req_lock);
1461 	if (res) {
1462 		int error = nfs_flush_list(inode, &head, res, how);
1463 		if (error < 0)
1464 			return error;
1465 	}
1466 	return res;
1467 }
1468 
1469 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1470 int nfs_commit_inode(struct inode *inode, int how)
1471 {
1472 	struct nfs_inode *nfsi = NFS_I(inode);
1473 	LIST_HEAD(head);
1474 	int res;
1475 
1476 	spin_lock(&nfsi->req_lock);
1477 	res = nfs_scan_commit(inode, &head, 0, 0);
1478 	spin_unlock(&nfsi->req_lock);
1479 	if (res) {
1480 		int error = nfs_commit_list(inode, &head, how);
1481 		if (error < 0)
1482 			return error;
1483 	}
1484 	return res;
1485 }
1486 #endif
1487 
1488 int nfs_sync_inode_wait(struct inode *inode, unsigned long idx_start,
1489 		unsigned int npages, int how)
1490 {
1491 	struct nfs_inode *nfsi = NFS_I(inode);
1492 	LIST_HEAD(head);
1493 	int nocommit = how & FLUSH_NOCOMMIT;
1494 	int pages, ret;
1495 
1496 	how &= ~FLUSH_NOCOMMIT;
1497 	spin_lock(&nfsi->req_lock);
1498 	do {
1499 		ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1500 		if (ret != 0)
1501 			continue;
1502 		pages = nfs_scan_dirty(inode, &head, idx_start, npages);
1503 		if (pages != 0) {
1504 			spin_unlock(&nfsi->req_lock);
1505 			if (how & FLUSH_INVALIDATE)
1506 				nfs_cancel_requests(&head);
1507 			else
1508 				ret = nfs_flush_list(inode, &head, pages, how);
1509 			spin_lock(&nfsi->req_lock);
1510 			continue;
1511 		}
1512 		if (nocommit)
1513 			break;
1514 		pages = nfs_scan_commit(inode, &head, idx_start, npages);
1515 		if (pages == 0)
1516 			break;
1517 		if (how & FLUSH_INVALIDATE) {
1518 			spin_unlock(&nfsi->req_lock);
1519 			nfs_cancel_requests(&head);
1520 			spin_lock(&nfsi->req_lock);
1521 			continue;
1522 		}
1523 		pages += nfs_scan_commit(inode, &head, 0, 0);
1524 		spin_unlock(&nfsi->req_lock);
1525 		ret = nfs_commit_list(inode, &head, how);
1526 		spin_lock(&nfsi->req_lock);
1527 	} while (ret >= 0);
1528 	spin_unlock(&nfsi->req_lock);
1529 	return ret;
1530 }
1531 
1532 int __init nfs_init_writepagecache(void)
1533 {
1534 	nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1535 					     sizeof(struct nfs_write_data),
1536 					     0, SLAB_HWCACHE_ALIGN,
1537 					     NULL, NULL);
1538 	if (nfs_wdata_cachep == NULL)
1539 		return -ENOMEM;
1540 
1541 	nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1542 						     nfs_wdata_cachep);
1543 	if (nfs_wdata_mempool == NULL)
1544 		return -ENOMEM;
1545 
1546 	nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1547 						      nfs_wdata_cachep);
1548 	if (nfs_commit_mempool == NULL)
1549 		return -ENOMEM;
1550 
1551 	return 0;
1552 }
1553 
1554 void __exit nfs_destroy_writepagecache(void)
1555 {
1556 	mempool_destroy(nfs_commit_mempool);
1557 	mempool_destroy(nfs_wdata_mempool);
1558 	if (kmem_cache_destroy(nfs_wdata_cachep))
1559 		printk(KERN_INFO "nfs_write_data: not all structures were freed\n");
1560 }
1561 
1562