]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - fs/nfs/write.c
NFS: rewrite directio write to use async coalesce code
[mirror_ubuntu-bionic-kernel.git] / fs / nfs / write.c
1 /*
2 * linux/fs/nfs/write.c
3 *
4 * Write file data over NFS.
5 *
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7 */
8
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/mm.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
17
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
23 #include <linux/export.h>
24
25 #include <asm/uaccess.h>
26
27 #include "delegation.h"
28 #include "internal.h"
29 #include "iostat.h"
30 #include "nfs4_fs.h"
31 #include "fscache.h"
32 #include "pnfs.h"
33
34 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
35
36 #define MIN_POOL_WRITE (32)
37 #define MIN_POOL_COMMIT (4)
38
39 /*
40 * Local function declarations
41 */
42 static void nfs_redirty_request(struct nfs_page *req);
43 static const struct rpc_call_ops nfs_write_common_ops;
44 static const struct rpc_call_ops nfs_commit_ops;
45 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
46 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
47
48 static struct kmem_cache *nfs_wdata_cachep;
49 static mempool_t *nfs_wdata_mempool;
50 static struct kmem_cache *nfs_cdata_cachep;
51 static mempool_t *nfs_commit_mempool;
52
53 struct nfs_commit_data *nfs_commitdata_alloc(void)
54 {
55 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
56
57 if (p) {
58 memset(p, 0, sizeof(*p));
59 INIT_LIST_HEAD(&p->pages);
60 }
61 return p;
62 }
63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
64
65 void nfs_commit_free(struct nfs_commit_data *p)
66 {
67 mempool_free(p, nfs_commit_mempool);
68 }
69 EXPORT_SYMBOL_GPL(nfs_commit_free);
70
71 struct nfs_write_header *nfs_writehdr_alloc(void)
72 {
73 struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
74
75 if (p) {
76 struct nfs_pgio_header *hdr = &p->header;
77
78 memset(p, 0, sizeof(*p));
79 INIT_LIST_HEAD(&hdr->pages);
80 INIT_LIST_HEAD(&hdr->rpc_list);
81 spin_lock_init(&hdr->lock);
82 atomic_set(&hdr->refcnt, 0);
83 }
84 return p;
85 }
86
87 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
88 unsigned int pagecount)
89 {
90 struct nfs_write_data *data, *prealloc;
91
92 prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
93 if (prealloc->header == NULL)
94 data = prealloc;
95 else
96 data = kzalloc(sizeof(*data), GFP_KERNEL);
97 if (!data)
98 goto out;
99
100 if (nfs_pgarray_set(&data->pages, pagecount)) {
101 data->header = hdr;
102 atomic_inc(&hdr->refcnt);
103 } else {
104 if (data != prealloc)
105 kfree(data);
106 data = NULL;
107 }
108 out:
109 return data;
110 }
111
112 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
113 {
114 struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
115 mempool_free(whdr, nfs_wdata_mempool);
116 }
117
118 void nfs_writedata_release(struct nfs_write_data *wdata)
119 {
120 struct nfs_pgio_header *hdr = wdata->header;
121 struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
122
123 put_nfs_open_context(wdata->args.context);
124 if (wdata->pages.pagevec != wdata->pages.page_array)
125 kfree(wdata->pages.pagevec);
126 if (wdata != &write_header->rpc_data)
127 kfree(wdata);
128 else
129 wdata->header = NULL;
130 if (atomic_dec_and_test(&hdr->refcnt))
131 hdr->completion_ops->completion(hdr);
132 }
133
134 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
135 {
136 ctx->error = error;
137 smp_wmb();
138 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
139 }
140
141 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
142 {
143 struct nfs_page *req = NULL;
144
145 if (PagePrivate(page)) {
146 req = (struct nfs_page *)page_private(page);
147 if (req != NULL)
148 kref_get(&req->wb_kref);
149 }
150 return req;
151 }
152
153 static struct nfs_page *nfs_page_find_request(struct page *page)
154 {
155 struct inode *inode = page->mapping->host;
156 struct nfs_page *req = NULL;
157
158 spin_lock(&inode->i_lock);
159 req = nfs_page_find_request_locked(page);
160 spin_unlock(&inode->i_lock);
161 return req;
162 }
163
164 /* Adjust the file length if we're writing beyond the end */
165 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
166 {
167 struct inode *inode = page->mapping->host;
168 loff_t end, i_size;
169 pgoff_t end_index;
170
171 spin_lock(&inode->i_lock);
172 i_size = i_size_read(inode);
173 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
174 if (i_size > 0 && page->index < end_index)
175 goto out;
176 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
177 if (i_size >= end)
178 goto out;
179 i_size_write(inode, end);
180 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
181 out:
182 spin_unlock(&inode->i_lock);
183 }
184
185 /* A writeback failed: mark the page as bad, and invalidate the page cache */
186 static void nfs_set_pageerror(struct page *page)
187 {
188 SetPageError(page);
189 nfs_zap_mapping(page->mapping->host, page->mapping);
190 }
191
192 /* We can set the PG_uptodate flag if we see that a write request
193 * covers the full page.
194 */
195 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
196 {
197 if (PageUptodate(page))
198 return;
199 if (base != 0)
200 return;
201 if (count != nfs_page_length(page))
202 return;
203 SetPageUptodate(page);
204 }
205
206 static int wb_priority(struct writeback_control *wbc)
207 {
208 if (wbc->for_reclaim)
209 return FLUSH_HIGHPRI | FLUSH_STABLE;
210 if (wbc->for_kupdate || wbc->for_background)
211 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
212 return FLUSH_COND_STABLE;
213 }
214
215 /*
216 * NFS congestion control
217 */
218
219 int nfs_congestion_kb;
220
221 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
222 #define NFS_CONGESTION_OFF_THRESH \
223 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
224
225 static int nfs_set_page_writeback(struct page *page)
226 {
227 int ret = test_set_page_writeback(page);
228
229 if (!ret) {
230 struct inode *inode = page->mapping->host;
231 struct nfs_server *nfss = NFS_SERVER(inode);
232
233 page_cache_get(page);
234 if (atomic_long_inc_return(&nfss->writeback) >
235 NFS_CONGESTION_ON_THRESH) {
236 set_bdi_congested(&nfss->backing_dev_info,
237 BLK_RW_ASYNC);
238 }
239 }
240 return ret;
241 }
242
243 static void nfs_end_page_writeback(struct page *page)
244 {
245 struct inode *inode = page->mapping->host;
246 struct nfs_server *nfss = NFS_SERVER(inode);
247
248 end_page_writeback(page);
249 page_cache_release(page);
250 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
251 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
252 }
253
254 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
255 {
256 struct inode *inode = page->mapping->host;
257 struct nfs_page *req;
258 int ret;
259
260 spin_lock(&inode->i_lock);
261 for (;;) {
262 req = nfs_page_find_request_locked(page);
263 if (req == NULL)
264 break;
265 if (nfs_lock_request_dontget(req))
266 break;
267 /* Note: If we hold the page lock, as is the case in nfs_writepage,
268 * then the call to nfs_lock_request_dontget() will always
269 * succeed provided that someone hasn't already marked the
270 * request as dirty (in which case we don't care).
271 */
272 spin_unlock(&inode->i_lock);
273 if (!nonblock)
274 ret = nfs_wait_on_request(req);
275 else
276 ret = -EAGAIN;
277 nfs_release_request(req);
278 if (ret != 0)
279 return ERR_PTR(ret);
280 spin_lock(&inode->i_lock);
281 }
282 spin_unlock(&inode->i_lock);
283 return req;
284 }
285
286 /*
287 * Find an associated nfs write request, and prepare to flush it out
288 * May return an error if the user signalled nfs_wait_on_request().
289 */
290 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
291 struct page *page, bool nonblock)
292 {
293 struct nfs_page *req;
294 int ret = 0;
295
296 req = nfs_find_and_lock_request(page, nonblock);
297 if (!req)
298 goto out;
299 ret = PTR_ERR(req);
300 if (IS_ERR(req))
301 goto out;
302
303 ret = nfs_set_page_writeback(page);
304 BUG_ON(ret != 0);
305 BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
306
307 if (!nfs_pageio_add_request(pgio, req)) {
308 nfs_redirty_request(req);
309 ret = pgio->pg_error;
310 }
311 out:
312 return ret;
313 }
314
315 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
316 {
317 struct inode *inode = page->mapping->host;
318 int ret;
319
320 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
321 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
322
323 nfs_pageio_cond_complete(pgio, page->index);
324 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
325 if (ret == -EAGAIN) {
326 redirty_page_for_writepage(wbc, page);
327 ret = 0;
328 }
329 return ret;
330 }
331
332 /*
333 * Write an mmapped page to the server.
334 */
335 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
336 {
337 struct nfs_pageio_descriptor pgio;
338 int err;
339
340 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
341 &nfs_async_write_completion_ops);
342 err = nfs_do_writepage(page, wbc, &pgio);
343 nfs_pageio_complete(&pgio);
344 if (err < 0)
345 return err;
346 if (pgio.pg_error < 0)
347 return pgio.pg_error;
348 return 0;
349 }
350
351 int nfs_writepage(struct page *page, struct writeback_control *wbc)
352 {
353 int ret;
354
355 ret = nfs_writepage_locked(page, wbc);
356 unlock_page(page);
357 return ret;
358 }
359
360 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
361 {
362 int ret;
363
364 ret = nfs_do_writepage(page, wbc, data);
365 unlock_page(page);
366 return ret;
367 }
368
369 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
370 {
371 struct inode *inode = mapping->host;
372 unsigned long *bitlock = &NFS_I(inode)->flags;
373 struct nfs_pageio_descriptor pgio;
374 int err;
375
376 /* Stop dirtying of new pages while we sync */
377 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
378 nfs_wait_bit_killable, TASK_KILLABLE);
379 if (err)
380 goto out_err;
381
382 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
383
384 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc),
385 &nfs_async_write_completion_ops);
386 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
387 nfs_pageio_complete(&pgio);
388
389 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
390 smp_mb__after_clear_bit();
391 wake_up_bit(bitlock, NFS_INO_FLUSHING);
392
393 if (err < 0)
394 goto out_err;
395 err = pgio.pg_error;
396 if (err < 0)
397 goto out_err;
398 return 0;
399 out_err:
400 return err;
401 }
402
403 /*
404 * Insert a write request into an inode
405 */
406 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
407 {
408 struct nfs_inode *nfsi = NFS_I(inode);
409
410 /* Lock the request! */
411 nfs_lock_request_dontget(req);
412
413 spin_lock(&inode->i_lock);
414 if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
415 inode->i_version++;
416 set_bit(PG_MAPPED, &req->wb_flags);
417 SetPagePrivate(req->wb_page);
418 set_page_private(req->wb_page, (unsigned long)req);
419 nfsi->npages++;
420 kref_get(&req->wb_kref);
421 spin_unlock(&inode->i_lock);
422 }
423
424 /*
425 * Remove a write request from an inode
426 */
427 static void nfs_inode_remove_request(struct nfs_page *req)
428 {
429 struct inode *inode = req->wb_context->dentry->d_inode;
430 struct nfs_inode *nfsi = NFS_I(inode);
431
432 BUG_ON (!NFS_WBACK_BUSY(req));
433
434 spin_lock(&inode->i_lock);
435 set_page_private(req->wb_page, 0);
436 ClearPagePrivate(req->wb_page);
437 clear_bit(PG_MAPPED, &req->wb_flags);
438 nfsi->npages--;
439 spin_unlock(&inode->i_lock);
440 nfs_release_request(req);
441 }
442
443 static void
444 nfs_mark_request_dirty(struct nfs_page *req)
445 {
446 __set_page_dirty_nobuffers(req->wb_page);
447 }
448
449 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
450 /**
451 * nfs_request_add_commit_list - add request to a commit list
452 * @req: pointer to a struct nfs_page
453 * @dst: commit list head
454 * @cinfo: holds list lock and accounting info
455 *
456 * This sets the PG_CLEAN bit, updates the cinfo count of
457 * number of outstanding requests requiring a commit as well as
458 * the MM page stats.
459 *
460 * The caller must _not_ hold the cinfo->lock, but must be
461 * holding the nfs_page lock.
462 */
463 void
464 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
465 struct nfs_commit_info *cinfo)
466 {
467 set_bit(PG_CLEAN, &(req)->wb_flags);
468 spin_lock(cinfo->lock);
469 nfs_list_add_request(req, dst);
470 cinfo->mds->ncommit++;
471 spin_unlock(cinfo->lock);
472 if (!cinfo->dreq) {
473 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
474 inc_bdi_stat(req->wb_page->mapping->backing_dev_info,
475 BDI_RECLAIMABLE);
476 __mark_inode_dirty(req->wb_context->dentry->d_inode,
477 I_DIRTY_DATASYNC);
478 }
479 }
480 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
481
482 /**
483 * nfs_request_remove_commit_list - Remove request from a commit list
484 * @req: pointer to a nfs_page
485 * @cinfo: holds list lock and accounting info
486 *
487 * This clears the PG_CLEAN bit, and updates the cinfo's count of
488 * number of outstanding requests requiring a commit
489 * It does not update the MM page stats.
490 *
491 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
492 */
493 void
494 nfs_request_remove_commit_list(struct nfs_page *req,
495 struct nfs_commit_info *cinfo)
496 {
497 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
498 return;
499 nfs_list_remove_request(req);
500 cinfo->mds->ncommit--;
501 }
502 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
503
504 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
505 struct inode *inode)
506 {
507 cinfo->lock = &inode->i_lock;
508 cinfo->mds = &NFS_I(inode)->commit_info;
509 cinfo->ds = pnfs_get_ds_info(inode);
510 cinfo->dreq = NULL;
511 cinfo->completion_ops = &nfs_commit_completion_ops;
512 }
513
514 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
515 struct inode *inode,
516 struct nfs_direct_req *dreq)
517 {
518 if (dreq)
519 nfs_init_cinfo_from_dreq(cinfo, dreq);
520 else
521 nfs_init_cinfo_from_inode(cinfo, inode);
522 }
523 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
524
525 /*
526 * Add a request to the inode's commit list.
527 */
528 void
529 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
530 struct nfs_commit_info *cinfo)
531 {
532 if (pnfs_mark_request_commit(req, lseg, cinfo))
533 return;
534 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
535 }
536
537 static void
538 nfs_clear_page_commit(struct page *page)
539 {
540 dec_zone_page_state(page, NR_UNSTABLE_NFS);
541 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
542 }
543
544 static void
545 nfs_clear_request_commit(struct nfs_page *req)
546 {
547 if (test_bit(PG_CLEAN, &req->wb_flags)) {
548 struct inode *inode = req->wb_context->dentry->d_inode;
549 struct nfs_commit_info cinfo;
550
551 nfs_init_cinfo_from_inode(&cinfo, inode);
552 if (!pnfs_clear_request_commit(req, &cinfo)) {
553 spin_lock(cinfo.lock);
554 nfs_request_remove_commit_list(req, &cinfo);
555 spin_unlock(cinfo.lock);
556 }
557 nfs_clear_page_commit(req->wb_page);
558 }
559 }
560
561 static inline
562 int nfs_write_need_commit(struct nfs_write_data *data)
563 {
564 if (data->verf.committed == NFS_DATA_SYNC)
565 return data->header->lseg == NULL;
566 return data->verf.committed != NFS_FILE_SYNC;
567 }
568
569 #else
570 void
571 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
572 struct nfs_commit_info *cinfo)
573 {
574 }
575
576 static void
577 nfs_clear_request_commit(struct nfs_page *req)
578 {
579 }
580
581 static inline
582 int nfs_write_need_commit(struct nfs_write_data *data)
583 {
584 return 0;
585 }
586
587 #endif
588
589 static void nfs_write_completion(struct nfs_pgio_header *hdr)
590 {
591 struct nfs_commit_info cinfo;
592 unsigned long bytes = 0;
593
594 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
595 goto out;
596 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
597 while (!list_empty(&hdr->pages)) {
598 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
599 struct page *page = req->wb_page;
600
601 bytes += req->wb_bytes;
602 nfs_list_remove_request(req);
603 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
604 (hdr->good_bytes < bytes)) {
605 nfs_set_pageerror(page);
606 nfs_context_set_write_error(req->wb_context, hdr->error);
607 goto remove_req;
608 }
609 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
610 nfs_mark_request_dirty(req);
611 goto next;
612 }
613 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
614 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
615 goto next;
616 }
617 remove_req:
618 nfs_inode_remove_request(req);
619 next:
620 nfs_unlock_request(req);
621 nfs_end_page_writeback(page);
622 }
623 out:
624 hdr->release(hdr);
625 }
626
627 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
628 static unsigned long
629 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
630 {
631 return cinfo->mds->ncommit;
632 }
633
634 /* cinfo->lock held by caller */
635 int
636 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
637 struct nfs_commit_info *cinfo, int max)
638 {
639 struct nfs_page *req, *tmp;
640 int ret = 0;
641
642 list_for_each_entry_safe(req, tmp, src, wb_list) {
643 if (!nfs_lock_request(req))
644 continue;
645 if (cond_resched_lock(cinfo->lock))
646 list_safe_reset_next(req, tmp, wb_list);
647 nfs_request_remove_commit_list(req, cinfo);
648 nfs_list_add_request(req, dst);
649 ret++;
650 if ((ret == max) && !cinfo->dreq)
651 break;
652 }
653 return ret;
654 }
655
656 /*
657 * nfs_scan_commit - Scan an inode for commit requests
658 * @inode: NFS inode to scan
659 * @dst: mds destination list
660 * @cinfo: mds and ds lists of reqs ready to commit
661 *
662 * Moves requests from the inode's 'commit' request list.
663 * The requests are *not* checked to ensure that they form a contiguous set.
664 */
665 int
666 nfs_scan_commit(struct inode *inode, struct list_head *dst,
667 struct nfs_commit_info *cinfo)
668 {
669 int ret = 0;
670
671 spin_lock(cinfo->lock);
672 if (cinfo->mds->ncommit > 0) {
673 const int max = INT_MAX;
674
675 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
676 cinfo, max);
677 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
678 }
679 spin_unlock(cinfo->lock);
680 return ret;
681 }
682
683 #else
684 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
685 {
686 return 0;
687 }
688
689 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
690 struct nfs_commit_info *cinfo)
691 {
692 return 0;
693 }
694 #endif
695
696 /*
697 * Search for an existing write request, and attempt to update
698 * it to reflect a new dirty region on a given page.
699 *
700 * If the attempt fails, then the existing request is flushed out
701 * to disk.
702 */
703 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
704 struct page *page,
705 unsigned int offset,
706 unsigned int bytes)
707 {
708 struct nfs_page *req;
709 unsigned int rqend;
710 unsigned int end;
711 int error;
712
713 if (!PagePrivate(page))
714 return NULL;
715
716 end = offset + bytes;
717 spin_lock(&inode->i_lock);
718
719 for (;;) {
720 req = nfs_page_find_request_locked(page);
721 if (req == NULL)
722 goto out_unlock;
723
724 rqend = req->wb_offset + req->wb_bytes;
725 /*
726 * Tell the caller to flush out the request if
727 * the offsets are non-contiguous.
728 * Note: nfs_flush_incompatible() will already
729 * have flushed out requests having wrong owners.
730 */
731 if (offset > rqend
732 || end < req->wb_offset)
733 goto out_flushme;
734
735 if (nfs_lock_request_dontget(req))
736 break;
737
738 /* The request is locked, so wait and then retry */
739 spin_unlock(&inode->i_lock);
740 error = nfs_wait_on_request(req);
741 nfs_release_request(req);
742 if (error != 0)
743 goto out_err;
744 spin_lock(&inode->i_lock);
745 }
746
747 /* Okay, the request matches. Update the region */
748 if (offset < req->wb_offset) {
749 req->wb_offset = offset;
750 req->wb_pgbase = offset;
751 }
752 if (end > rqend)
753 req->wb_bytes = end - req->wb_offset;
754 else
755 req->wb_bytes = rqend - req->wb_offset;
756 out_unlock:
757 spin_unlock(&inode->i_lock);
758 if (req)
759 nfs_clear_request_commit(req);
760 return req;
761 out_flushme:
762 spin_unlock(&inode->i_lock);
763 nfs_release_request(req);
764 error = nfs_wb_page(inode, page);
765 out_err:
766 return ERR_PTR(error);
767 }
768
769 /*
770 * Try to update an existing write request, or create one if there is none.
771 *
772 * Note: Should always be called with the Page Lock held to prevent races
773 * if we have to add a new request. Also assumes that the caller has
774 * already called nfs_flush_incompatible() if necessary.
775 */
776 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
777 struct page *page, unsigned int offset, unsigned int bytes)
778 {
779 struct inode *inode = page->mapping->host;
780 struct nfs_page *req;
781
782 req = nfs_try_to_update_request(inode, page, offset, bytes);
783 if (req != NULL)
784 goto out;
785 req = nfs_create_request(ctx, inode, page, offset, bytes);
786 if (IS_ERR(req))
787 goto out;
788 nfs_inode_add_request(inode, req);
789 out:
790 return req;
791 }
792
793 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
794 unsigned int offset, unsigned int count)
795 {
796 struct nfs_page *req;
797
798 req = nfs_setup_write_request(ctx, page, offset, count);
799 if (IS_ERR(req))
800 return PTR_ERR(req);
801 /* Update file length */
802 nfs_grow_file(page, offset, count);
803 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
804 nfs_mark_request_dirty(req);
805 nfs_unlock_request(req);
806 return 0;
807 }
808
809 int nfs_flush_incompatible(struct file *file, struct page *page)
810 {
811 struct nfs_open_context *ctx = nfs_file_open_context(file);
812 struct nfs_page *req;
813 int do_flush, status;
814 /*
815 * Look for a request corresponding to this page. If there
816 * is one, and it belongs to another file, we flush it out
817 * before we try to copy anything into the page. Do this
818 * due to the lack of an ACCESS-type call in NFSv2.
819 * Also do the same if we find a request from an existing
820 * dropped page.
821 */
822 do {
823 req = nfs_page_find_request(page);
824 if (req == NULL)
825 return 0;
826 do_flush = req->wb_page != page || req->wb_context != ctx ||
827 req->wb_lock_context->lockowner != current->files ||
828 req->wb_lock_context->pid != current->tgid;
829 nfs_release_request(req);
830 if (!do_flush)
831 return 0;
832 status = nfs_wb_page(page->mapping->host, page);
833 } while (status == 0);
834 return status;
835 }
836
837 /*
838 * If the page cache is marked as unsafe or invalid, then we can't rely on
839 * the PageUptodate() flag. In this case, we will need to turn off
840 * write optimisations that depend on the page contents being correct.
841 */
842 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
843 {
844 return PageUptodate(page) &&
845 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
846 }
847
848 /*
849 * Update and possibly write a cached page of an NFS file.
850 *
851 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
852 * things with a page scheduled for an RPC call (e.g. invalidate it).
853 */
854 int nfs_updatepage(struct file *file, struct page *page,
855 unsigned int offset, unsigned int count)
856 {
857 struct nfs_open_context *ctx = nfs_file_open_context(file);
858 struct inode *inode = page->mapping->host;
859 int status = 0;
860
861 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
862
863 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
864 file->f_path.dentry->d_parent->d_name.name,
865 file->f_path.dentry->d_name.name, count,
866 (long long)(page_offset(page) + offset));
867
868 /* If we're not using byte range locks, and we know the page
869 * is up to date, it may be more efficient to extend the write
870 * to cover the entire page in order to avoid fragmentation
871 * inefficiencies.
872 */
873 if (nfs_write_pageuptodate(page, inode) &&
874 inode->i_flock == NULL &&
875 !(file->f_flags & O_DSYNC)) {
876 count = max(count + offset, nfs_page_length(page));
877 offset = 0;
878 }
879
880 status = nfs_writepage_setup(ctx, page, offset, count);
881 if (status < 0)
882 nfs_set_pageerror(page);
883 else
884 __set_page_dirty_nobuffers(page);
885
886 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
887 status, (long long)i_size_read(inode));
888 return status;
889 }
890
891 static int flush_task_priority(int how)
892 {
893 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
894 case FLUSH_HIGHPRI:
895 return RPC_PRIORITY_HIGH;
896 case FLUSH_LOWPRI:
897 return RPC_PRIORITY_LOW;
898 }
899 return RPC_PRIORITY_NORMAL;
900 }
901
902 int nfs_initiate_write(struct rpc_clnt *clnt,
903 struct nfs_write_data *data,
904 const struct rpc_call_ops *call_ops,
905 int how)
906 {
907 struct inode *inode = data->header->inode;
908 int priority = flush_task_priority(how);
909 struct rpc_task *task;
910 struct rpc_message msg = {
911 .rpc_argp = &data->args,
912 .rpc_resp = &data->res,
913 .rpc_cred = data->header->cred,
914 };
915 struct rpc_task_setup task_setup_data = {
916 .rpc_client = clnt,
917 .task = &data->task,
918 .rpc_message = &msg,
919 .callback_ops = call_ops,
920 .callback_data = data,
921 .workqueue = nfsiod_workqueue,
922 .flags = RPC_TASK_ASYNC,
923 .priority = priority,
924 };
925 int ret = 0;
926
927 /* Set up the initial task struct. */
928 NFS_PROTO(inode)->write_setup(data, &msg);
929
930 dprintk("NFS: %5u initiated write call "
931 "(req %s/%lld, %u bytes @ offset %llu)\n",
932 data->task.tk_pid,
933 inode->i_sb->s_id,
934 (long long)NFS_FILEID(inode),
935 data->args.count,
936 (unsigned long long)data->args.offset);
937
938 task = rpc_run_task(&task_setup_data);
939 if (IS_ERR(task)) {
940 ret = PTR_ERR(task);
941 goto out;
942 }
943 if (how & FLUSH_SYNC) {
944 ret = rpc_wait_for_completion_task(task);
945 if (ret == 0)
946 ret = task->tk_status;
947 }
948 rpc_put_task(task);
949 out:
950 return ret;
951 }
952 EXPORT_SYMBOL_GPL(nfs_initiate_write);
953
954 /*
955 * Set up the argument/result storage required for the RPC call.
956 */
957 static void nfs_write_rpcsetup(struct nfs_write_data *data,
958 unsigned int count, unsigned int offset,
959 int how, struct nfs_commit_info *cinfo)
960 {
961 struct nfs_page *req = data->header->req;
962
963 /* Set up the RPC argument and reply structs
964 * NB: take care not to mess about with data->commit et al. */
965
966 data->args.fh = NFS_FH(data->header->inode);
967 data->args.offset = req_offset(req) + offset;
968 /* pnfs_set_layoutcommit needs this */
969 data->mds_offset = data->args.offset;
970 data->args.pgbase = req->wb_pgbase + offset;
971 data->args.pages = data->pages.pagevec;
972 data->args.count = count;
973 data->args.context = get_nfs_open_context(req->wb_context);
974 data->args.lock_context = req->wb_lock_context;
975 data->args.stable = NFS_UNSTABLE;
976 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
977 case 0:
978 break;
979 case FLUSH_COND_STABLE:
980 if (nfs_reqs_to_commit(cinfo))
981 break;
982 default:
983 data->args.stable = NFS_FILE_SYNC;
984 }
985
986 data->res.fattr = &data->fattr;
987 data->res.count = count;
988 data->res.verf = &data->verf;
989 nfs_fattr_init(&data->fattr);
990 }
991
992 static int nfs_do_write(struct nfs_write_data *data,
993 const struct rpc_call_ops *call_ops,
994 int how)
995 {
996 struct inode *inode = data->header->inode;
997
998 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how);
999 }
1000
1001 static int nfs_do_multiple_writes(struct list_head *head,
1002 const struct rpc_call_ops *call_ops,
1003 int how)
1004 {
1005 struct nfs_write_data *data;
1006 int ret = 0;
1007
1008 while (!list_empty(head)) {
1009 int ret2;
1010
1011 data = list_first_entry(head, struct nfs_write_data, list);
1012 list_del_init(&data->list);
1013
1014 ret2 = nfs_do_write(data, call_ops, how);
1015 if (ret == 0)
1016 ret = ret2;
1017 }
1018 return ret;
1019 }
1020
1021 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1022 * call this on each, which will prepare them to be retried on next
1023 * writeback using standard nfs.
1024 */
1025 static void nfs_redirty_request(struct nfs_page *req)
1026 {
1027 struct page *page = req->wb_page;
1028
1029 nfs_mark_request_dirty(req);
1030 nfs_unlock_request(req);
1031 nfs_end_page_writeback(page);
1032 }
1033
1034 static void nfs_async_write_error(struct list_head *head)
1035 {
1036 struct nfs_page *req;
1037
1038 while (!list_empty(head)) {
1039 req = nfs_list_entry(head->next);
1040 nfs_list_remove_request(req);
1041 nfs_redirty_request(req);
1042 }
1043 }
1044
1045 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1046 .error_cleanup = nfs_async_write_error,
1047 .completion = nfs_write_completion,
1048 };
1049
1050 /*
1051 * Generate multiple small requests to write out a single
1052 * contiguous dirty area on one page.
1053 */
1054 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1055 struct nfs_pgio_header *hdr)
1056 {
1057 struct nfs_page *req = hdr->req;
1058 struct page *page = req->wb_page;
1059 struct nfs_write_data *data;
1060 size_t wsize = desc->pg_bsize, nbytes;
1061 unsigned int offset;
1062 int requests = 0;
1063 int ret = 0;
1064 struct nfs_commit_info cinfo;
1065
1066 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1067 nfs_list_remove_request(req);
1068 nfs_list_add_request(req, &hdr->pages);
1069
1070 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1071 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1072 desc->pg_count > wsize))
1073 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1074
1075
1076 offset = 0;
1077 nbytes = desc->pg_count;
1078 do {
1079 size_t len = min(nbytes, wsize);
1080
1081 data = nfs_writedata_alloc(hdr, 1);
1082 if (!data)
1083 goto out_bad;
1084 data->pages.pagevec[0] = page;
1085 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1086 list_add(&data->list, &hdr->rpc_list);
1087 requests++;
1088 nbytes -= len;
1089 offset += len;
1090 } while (nbytes != 0);
1091 desc->pg_rpc_callops = &nfs_write_common_ops;
1092 return ret;
1093
1094 out_bad:
1095 while (!list_empty(&hdr->rpc_list)) {
1096 data = list_first_entry(&hdr->rpc_list, struct nfs_write_data, list);
1097 list_del(&data->list);
1098 nfs_writedata_release(data);
1099 }
1100 desc->pg_completion_ops->error_cleanup(&hdr->pages);
1101 return -ENOMEM;
1102 }
1103
1104 /*
1105 * Create an RPC task for the given write request and kick it.
1106 * The page must have been locked by the caller.
1107 *
1108 * It may happen that the page we're passed is not marked dirty.
1109 * This is the case if nfs_updatepage detects a conflicting request
1110 * that has been written but not committed.
1111 */
1112 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1113 struct nfs_pgio_header *hdr)
1114 {
1115 struct nfs_page *req;
1116 struct page **pages;
1117 struct nfs_write_data *data;
1118 struct list_head *head = &desc->pg_list;
1119 int ret = 0;
1120 struct nfs_commit_info cinfo;
1121
1122 data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1123 desc->pg_count));
1124 if (!data) {
1125 desc->pg_completion_ops->error_cleanup(head);
1126 ret = -ENOMEM;
1127 goto out;
1128 }
1129
1130 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1131 pages = data->pages.pagevec;
1132 while (!list_empty(head)) {
1133 req = nfs_list_entry(head->next);
1134 nfs_list_remove_request(req);
1135 nfs_list_add_request(req, &hdr->pages);
1136 *pages++ = req->wb_page;
1137 }
1138
1139 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1140 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1141 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1142
1143 /* Set up the argument struct */
1144 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1145 list_add(&data->list, &hdr->rpc_list);
1146 desc->pg_rpc_callops = &nfs_write_common_ops;
1147 out:
1148 return ret;
1149 }
1150
1151 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1152 struct nfs_pgio_header *hdr)
1153 {
1154 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1155 return nfs_flush_multi(desc, hdr);
1156 return nfs_flush_one(desc, hdr);
1157 }
1158
1159 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1160 {
1161 struct nfs_write_header *whdr;
1162 struct nfs_pgio_header *hdr;
1163 int ret;
1164
1165 whdr = nfs_writehdr_alloc();
1166 if (!whdr) {
1167 desc->pg_completion_ops->error_cleanup(&hdr->pages);
1168 return -ENOMEM;
1169 }
1170 hdr = &whdr->header;
1171 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1172 atomic_inc(&hdr->refcnt);
1173 ret = nfs_generic_flush(desc, hdr);
1174 if (ret == 0)
1175 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1176 desc->pg_rpc_callops,
1177 desc->pg_ioflags);
1178 else
1179 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1180 if (atomic_dec_and_test(&hdr->refcnt))
1181 hdr->completion_ops->completion(hdr);
1182 return ret;
1183 }
1184
1185 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1186 .pg_test = nfs_generic_pg_test,
1187 .pg_doio = nfs_generic_pg_writepages,
1188 };
1189
1190 void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1191 struct inode *inode, int ioflags,
1192 const struct nfs_pgio_completion_ops *compl_ops)
1193 {
1194 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1195 NFS_SERVER(inode)->wsize, ioflags);
1196 }
1197
1198 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1199 {
1200 pgio->pg_ops = &nfs_pageio_write_ops;
1201 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1202 }
1203 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1204
1205 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1206 struct inode *inode, int ioflags,
1207 const struct nfs_pgio_completion_ops *compl_ops)
1208 {
1209 if (!pnfs_pageio_init_write(pgio, inode, ioflags, compl_ops))
1210 nfs_pageio_init_write_mds(pgio, inode, ioflags, compl_ops);
1211 }
1212
1213 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1214 {
1215 struct nfs_write_data *data = calldata;
1216 NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1217 }
1218
1219 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1220 {
1221 struct nfs_commit_data *data = calldata;
1222
1223 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1224 }
1225
1226 /*
1227 * Handle a write reply that flushes a whole page.
1228 *
1229 * FIXME: There is an inherent race with invalidate_inode_pages and
1230 * writebacks since the page->count is kept > 1 for as long
1231 * as the page has a write request pending.
1232 */
1233 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1234 {
1235 struct nfs_write_data *data = calldata;
1236
1237 nfs_writeback_done(task, data);
1238 }
1239
1240 static void nfs_writeback_release_common(void *calldata)
1241 {
1242 struct nfs_write_data *data = calldata;
1243 struct nfs_pgio_header *hdr = data->header;
1244 int status = data->task.tk_status;
1245 struct nfs_page *req = hdr->req;
1246
1247 if ((status >= 0) && nfs_write_need_commit(data)) {
1248 spin_lock(&hdr->lock);
1249 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1250 ; /* Do nothing */
1251 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1252 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1253 else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf)))
1254 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1255 spin_unlock(&hdr->lock);
1256 }
1257 nfs_writedata_release(data);
1258 }
1259
1260 static const struct rpc_call_ops nfs_write_common_ops = {
1261 .rpc_call_prepare = nfs_write_prepare,
1262 .rpc_call_done = nfs_writeback_done_common,
1263 .rpc_release = nfs_writeback_release_common,
1264 };
1265
1266
1267 /*
1268 * This function is called when the WRITE call is complete.
1269 */
1270 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1271 {
1272 struct nfs_writeargs *argp = &data->args;
1273 struct nfs_writeres *resp = &data->res;
1274 struct inode *inode = data->header->inode;
1275 int status;
1276
1277 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1278 task->tk_pid, task->tk_status);
1279
1280 /*
1281 * ->write_done will attempt to use post-op attributes to detect
1282 * conflicting writes by other clients. A strict interpretation
1283 * of close-to-open would allow us to continue caching even if
1284 * another writer had changed the file, but some applications
1285 * depend on tighter cache coherency when writing.
1286 */
1287 status = NFS_PROTO(inode)->write_done(task, data);
1288 if (status != 0)
1289 return;
1290 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1291
1292 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1293 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1294 /* We tried a write call, but the server did not
1295 * commit data to stable storage even though we
1296 * requested it.
1297 * Note: There is a known bug in Tru64 < 5.0 in which
1298 * the server reports NFS_DATA_SYNC, but performs
1299 * NFS_FILE_SYNC. We therefore implement this checking
1300 * as a dprintk() in order to avoid filling syslog.
1301 */
1302 static unsigned long complain;
1303
1304 /* Note this will print the MDS for a DS write */
1305 if (time_before(complain, jiffies)) {
1306 dprintk("NFS: faulty NFS server %s:"
1307 " (committed = %d) != (stable = %d)\n",
1308 NFS_SERVER(inode)->nfs_client->cl_hostname,
1309 resp->verf->committed, argp->stable);
1310 complain = jiffies + 300 * HZ;
1311 }
1312 }
1313 #endif
1314 if (task->tk_status < 0)
1315 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1316 else if (resp->count < argp->count) {
1317 static unsigned long complain;
1318
1319 /* This a short write! */
1320 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1321
1322 /* Has the server at least made some progress? */
1323 if (resp->count == 0) {
1324 if (time_before(complain, jiffies)) {
1325 printk(KERN_WARNING
1326 "NFS: Server wrote zero bytes, expected %u.\n",
1327 argp->count);
1328 complain = jiffies + 300 * HZ;
1329 }
1330 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1331 task->tk_status = -EIO;
1332 return;
1333 }
1334 /* Was this an NFSv2 write or an NFSv3 stable write? */
1335 if (resp->verf->committed != NFS_UNSTABLE) {
1336 /* Resend from where the server left off */
1337 data->mds_offset += resp->count;
1338 argp->offset += resp->count;
1339 argp->pgbase += resp->count;
1340 argp->count -= resp->count;
1341 } else {
1342 /* Resend as a stable write in order to avoid
1343 * headaches in the case of a server crash.
1344 */
1345 argp->stable = NFS_FILE_SYNC;
1346 }
1347 rpc_restart_call_prepare(task);
1348 }
1349 }
1350
1351
1352 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1353 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1354 {
1355 int ret;
1356
1357 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1358 return 1;
1359 if (!may_wait)
1360 return 0;
1361 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1362 NFS_INO_COMMIT,
1363 nfs_wait_bit_killable,
1364 TASK_KILLABLE);
1365 return (ret < 0) ? ret : 1;
1366 }
1367
1368 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1369 {
1370 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1371 smp_mb__after_clear_bit();
1372 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1373 }
1374
1375 void nfs_commitdata_release(struct nfs_commit_data *data)
1376 {
1377 put_nfs_open_context(data->context);
1378 nfs_commit_free(data);
1379 }
1380 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1381
1382 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1383 const struct rpc_call_ops *call_ops,
1384 int how)
1385 {
1386 struct rpc_task *task;
1387 int priority = flush_task_priority(how);
1388 struct rpc_message msg = {
1389 .rpc_argp = &data->args,
1390 .rpc_resp = &data->res,
1391 .rpc_cred = data->cred,
1392 };
1393 struct rpc_task_setup task_setup_data = {
1394 .task = &data->task,
1395 .rpc_client = clnt,
1396 .rpc_message = &msg,
1397 .callback_ops = call_ops,
1398 .callback_data = data,
1399 .workqueue = nfsiod_workqueue,
1400 .flags = RPC_TASK_ASYNC,
1401 .priority = priority,
1402 };
1403 /* Set up the initial task struct. */
1404 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1405
1406 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1407
1408 task = rpc_run_task(&task_setup_data);
1409 if (IS_ERR(task))
1410 return PTR_ERR(task);
1411 if (how & FLUSH_SYNC)
1412 rpc_wait_for_completion_task(task);
1413 rpc_put_task(task);
1414 return 0;
1415 }
1416 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1417
1418 /*
1419 * Set up the argument/result storage required for the RPC call.
1420 */
1421 void nfs_init_commit(struct nfs_commit_data *data,
1422 struct list_head *head,
1423 struct pnfs_layout_segment *lseg,
1424 struct nfs_commit_info *cinfo)
1425 {
1426 struct nfs_page *first = nfs_list_entry(head->next);
1427 struct inode *inode = first->wb_context->dentry->d_inode;
1428
1429 /* Set up the RPC argument and reply structs
1430 * NB: take care not to mess about with data->commit et al. */
1431
1432 list_splice_init(head, &data->pages);
1433
1434 data->inode = inode;
1435 data->cred = first->wb_context->cred;
1436 data->lseg = lseg; /* reference transferred */
1437 data->mds_ops = &nfs_commit_ops;
1438 data->completion_ops = cinfo->completion_ops;
1439 data->dreq = cinfo->dreq;
1440
1441 data->args.fh = NFS_FH(data->inode);
1442 /* Note: we always request a commit of the entire inode */
1443 data->args.offset = 0;
1444 data->args.count = 0;
1445 data->context = get_nfs_open_context(first->wb_context);
1446 data->res.fattr = &data->fattr;
1447 data->res.verf = &data->verf;
1448 nfs_fattr_init(&data->fattr);
1449 }
1450 EXPORT_SYMBOL_GPL(nfs_init_commit);
1451
1452 void nfs_retry_commit(struct list_head *page_list,
1453 struct pnfs_layout_segment *lseg,
1454 struct nfs_commit_info *cinfo)
1455 {
1456 struct nfs_page *req;
1457
1458 while (!list_empty(page_list)) {
1459 req = nfs_list_entry(page_list->next);
1460 nfs_list_remove_request(req);
1461 nfs_mark_request_commit(req, lseg, cinfo);
1462 if (!cinfo->dreq) {
1463 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1464 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1465 BDI_RECLAIMABLE);
1466 }
1467 nfs_unlock_request(req);
1468 }
1469 }
1470 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1471
1472 /*
1473 * Commit dirty pages
1474 */
1475 static int
1476 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1477 struct nfs_commit_info *cinfo)
1478 {
1479 struct nfs_commit_data *data;
1480
1481 data = nfs_commitdata_alloc();
1482
1483 if (!data)
1484 goto out_bad;
1485
1486 /* Set up the argument struct */
1487 nfs_init_commit(data, head, NULL, cinfo);
1488 atomic_inc(&cinfo->mds->rpcs_out);
1489 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops, how);
1490 out_bad:
1491 nfs_retry_commit(head, NULL, cinfo);
1492 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1493 return -ENOMEM;
1494 }
1495
1496 /*
1497 * COMMIT call returned
1498 */
1499 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1500 {
1501 struct nfs_commit_data *data = calldata;
1502
1503 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1504 task->tk_pid, task->tk_status);
1505
1506 /* Call the NFS version-specific code */
1507 NFS_PROTO(data->inode)->commit_done(task, data);
1508 }
1509
1510 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1511 {
1512 struct nfs_page *req;
1513 int status = data->task.tk_status;
1514 struct nfs_commit_info cinfo;
1515
1516 while (!list_empty(&data->pages)) {
1517 req = nfs_list_entry(data->pages.next);
1518 nfs_list_remove_request(req);
1519 nfs_clear_page_commit(req->wb_page);
1520
1521 dprintk("NFS: commit (%s/%lld %d@%lld)",
1522 req->wb_context->dentry->d_sb->s_id,
1523 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1524 req->wb_bytes,
1525 (long long)req_offset(req));
1526 if (status < 0) {
1527 nfs_context_set_write_error(req->wb_context, status);
1528 nfs_inode_remove_request(req);
1529 dprintk(", error = %d\n", status);
1530 goto next;
1531 }
1532
1533 /* Okay, COMMIT succeeded, apparently. Check the verifier
1534 * returned by the server against all stored verfs. */
1535 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1536 /* We have a match */
1537 nfs_inode_remove_request(req);
1538 dprintk(" OK\n");
1539 goto next;
1540 }
1541 /* We have a mismatch. Write the page again */
1542 dprintk(" mismatch\n");
1543 nfs_mark_request_dirty(req);
1544 next:
1545 nfs_unlock_request(req);
1546 }
1547 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1548 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1549 nfs_commit_clear_lock(NFS_I(data->inode));
1550 }
1551
1552 static void nfs_commit_release(void *calldata)
1553 {
1554 struct nfs_commit_data *data = calldata;
1555
1556 data->completion_ops->completion(data);
1557 nfs_commitdata_release(calldata);
1558 }
1559
1560 static const struct rpc_call_ops nfs_commit_ops = {
1561 .rpc_call_prepare = nfs_commit_prepare,
1562 .rpc_call_done = nfs_commit_done,
1563 .rpc_release = nfs_commit_release,
1564 };
1565
1566 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1567 .completion = nfs_commit_release_pages,
1568 .error_cleanup = nfs_commit_clear_lock,
1569 };
1570
1571 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1572 int how, struct nfs_commit_info *cinfo)
1573 {
1574 int status;
1575
1576 status = pnfs_commit_list(inode, head, how, cinfo);
1577 if (status == PNFS_NOT_ATTEMPTED)
1578 status = nfs_commit_list(inode, head, how, cinfo);
1579 return status;
1580 }
1581
1582 int nfs_commit_inode(struct inode *inode, int how)
1583 {
1584 LIST_HEAD(head);
1585 struct nfs_commit_info cinfo;
1586 int may_wait = how & FLUSH_SYNC;
1587 int res;
1588
1589 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1590 if (res <= 0)
1591 goto out_mark_dirty;
1592 nfs_init_cinfo_from_inode(&cinfo, inode);
1593 res = nfs_scan_commit(inode, &head, &cinfo);
1594 if (res) {
1595 int error;
1596
1597 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1598 if (error < 0)
1599 return error;
1600 if (!may_wait)
1601 goto out_mark_dirty;
1602 error = wait_on_bit(&NFS_I(inode)->flags,
1603 NFS_INO_COMMIT,
1604 nfs_wait_bit_killable,
1605 TASK_KILLABLE);
1606 if (error < 0)
1607 return error;
1608 } else
1609 nfs_commit_clear_lock(NFS_I(inode));
1610 return res;
1611 /* Note: If we exit without ensuring that the commit is complete,
1612 * we must mark the inode as dirty. Otherwise, future calls to
1613 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1614 * that the data is on the disk.
1615 */
1616 out_mark_dirty:
1617 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1618 return res;
1619 }
1620
1621 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1622 {
1623 struct nfs_inode *nfsi = NFS_I(inode);
1624 int flags = FLUSH_SYNC;
1625 int ret = 0;
1626
1627 /* no commits means nothing needs to be done */
1628 if (!nfsi->commit_info.ncommit)
1629 return ret;
1630
1631 if (wbc->sync_mode == WB_SYNC_NONE) {
1632 /* Don't commit yet if this is a non-blocking flush and there
1633 * are a lot of outstanding writes for this mapping.
1634 */
1635 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1636 goto out_mark_dirty;
1637
1638 /* don't wait for the COMMIT response */
1639 flags = 0;
1640 }
1641
1642 ret = nfs_commit_inode(inode, flags);
1643 if (ret >= 0) {
1644 if (wbc->sync_mode == WB_SYNC_NONE) {
1645 if (ret < wbc->nr_to_write)
1646 wbc->nr_to_write -= ret;
1647 else
1648 wbc->nr_to_write = 0;
1649 }
1650 return 0;
1651 }
1652 out_mark_dirty:
1653 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1654 return ret;
1655 }
1656 #else
1657 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1658 {
1659 return 0;
1660 }
1661 #endif
1662
1663 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1664 {
1665 int ret;
1666
1667 ret = nfs_commit_unstable_pages(inode, wbc);
1668 if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1669 int status;
1670 bool sync = true;
1671
1672 if (wbc->sync_mode == WB_SYNC_NONE)
1673 sync = false;
1674
1675 status = pnfs_layoutcommit_inode(inode, sync);
1676 if (status < 0)
1677 return status;
1678 }
1679 return ret;
1680 }
1681
1682 /*
1683 * flush the inode to disk.
1684 */
1685 int nfs_wb_all(struct inode *inode)
1686 {
1687 struct writeback_control wbc = {
1688 .sync_mode = WB_SYNC_ALL,
1689 .nr_to_write = LONG_MAX,
1690 .range_start = 0,
1691 .range_end = LLONG_MAX,
1692 };
1693
1694 return sync_inode(inode, &wbc);
1695 }
1696
1697 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1698 {
1699 struct nfs_page *req;
1700 int ret = 0;
1701
1702 BUG_ON(!PageLocked(page));
1703 for (;;) {
1704 wait_on_page_writeback(page);
1705 req = nfs_page_find_request(page);
1706 if (req == NULL)
1707 break;
1708 if (nfs_lock_request_dontget(req)) {
1709 nfs_clear_request_commit(req);
1710 nfs_inode_remove_request(req);
1711 /*
1712 * In case nfs_inode_remove_request has marked the
1713 * page as being dirty
1714 */
1715 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1716 nfs_unlock_request(req);
1717 break;
1718 }
1719 ret = nfs_wait_on_request(req);
1720 nfs_release_request(req);
1721 if (ret < 0)
1722 break;
1723 }
1724 return ret;
1725 }
1726
1727 /*
1728 * Write back all requests on one page - we do this before reading it.
1729 */
1730 int nfs_wb_page(struct inode *inode, struct page *page)
1731 {
1732 loff_t range_start = page_offset(page);
1733 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1734 struct writeback_control wbc = {
1735 .sync_mode = WB_SYNC_ALL,
1736 .nr_to_write = 0,
1737 .range_start = range_start,
1738 .range_end = range_end,
1739 };
1740 int ret;
1741
1742 for (;;) {
1743 wait_on_page_writeback(page);
1744 if (clear_page_dirty_for_io(page)) {
1745 ret = nfs_writepage_locked(page, &wbc);
1746 if (ret < 0)
1747 goto out_error;
1748 continue;
1749 }
1750 if (!PagePrivate(page))
1751 break;
1752 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1753 if (ret < 0)
1754 goto out_error;
1755 }
1756 return 0;
1757 out_error:
1758 return ret;
1759 }
1760
1761 #ifdef CONFIG_MIGRATION
1762 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1763 struct page *page, enum migrate_mode mode)
1764 {
1765 /*
1766 * If PagePrivate is set, then the page is currently associated with
1767 * an in-progress read or write request. Don't try to migrate it.
1768 *
1769 * FIXME: we could do this in principle, but we'll need a way to ensure
1770 * that we can safely release the inode reference while holding
1771 * the page lock.
1772 */
1773 if (PagePrivate(page))
1774 return -EBUSY;
1775
1776 nfs_fscache_release_page(page, GFP_KERNEL);
1777
1778 return migrate_page(mapping, newpage, page, mode);
1779 }
1780 #endif
1781
1782 int __init nfs_init_writepagecache(void)
1783 {
1784 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1785 sizeof(struct nfs_write_header),
1786 0, SLAB_HWCACHE_ALIGN,
1787 NULL);
1788 if (nfs_wdata_cachep == NULL)
1789 return -ENOMEM;
1790
1791 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1792 nfs_wdata_cachep);
1793 if (nfs_wdata_mempool == NULL)
1794 return -ENOMEM;
1795
1796 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1797 sizeof(struct nfs_commit_data),
1798 0, SLAB_HWCACHE_ALIGN,
1799 NULL);
1800 if (nfs_cdata_cachep == NULL)
1801 return -ENOMEM;
1802
1803 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1804 nfs_wdata_cachep);
1805 if (nfs_commit_mempool == NULL)
1806 return -ENOMEM;
1807
1808 /*
1809 * NFS congestion size, scale with available memory.
1810 *
1811 * 64MB: 8192k
1812 * 128MB: 11585k
1813 * 256MB: 16384k
1814 * 512MB: 23170k
1815 * 1GB: 32768k
1816 * 2GB: 46340k
1817 * 4GB: 65536k
1818 * 8GB: 92681k
1819 * 16GB: 131072k
1820 *
1821 * This allows larger machines to have larger/more transfers.
1822 * Limit the default to 256M
1823 */
1824 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1825 if (nfs_congestion_kb > 256*1024)
1826 nfs_congestion_kb = 256*1024;
1827
1828 return 0;
1829 }
1830
1831 void nfs_destroy_writepagecache(void)
1832 {
1833 mempool_destroy(nfs_commit_mempool);
1834 mempool_destroy(nfs_wdata_mempool);
1835 kmem_cache_destroy(nfs_wdata_cachep);
1836 }
1837