HID: hiddev: Fix slab-out-of-bounds write in hiddev_ioctl_usage()
[linux/fpc-iii.git] / fs / nfs / write.c
blob6e81a5b5858ec9a9e34a42fbcbd8a3a25af6017b
1 /*
2 * linux/fs/nfs/write.c
4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7 */
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>
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>
25 #include <asm/uaccess.h>
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"
34 #include "nfstrace.h"
36 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
38 #define MIN_POOL_WRITE (32)
39 #define MIN_POOL_COMMIT (4)
42 * Local function declarations
44 static void nfs_redirty_request(struct nfs_page *req);
45 static const struct rpc_call_ops nfs_commit_ops;
46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
48 static const struct nfs_rw_ops nfs_rw_write_ops;
49 static void nfs_clear_request_commit(struct nfs_page *req);
50 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
51 struct inode *inode);
52 static struct nfs_page *
53 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
54 struct page *page);
56 static struct kmem_cache *nfs_wdata_cachep;
57 static mempool_t *nfs_wdata_mempool;
58 static struct kmem_cache *nfs_cdata_cachep;
59 static mempool_t *nfs_commit_mempool;
61 struct nfs_commit_data *nfs_commitdata_alloc(void)
63 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
65 if (p) {
66 memset(p, 0, sizeof(*p));
67 INIT_LIST_HEAD(&p->pages);
69 return p;
71 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
73 void nfs_commit_free(struct nfs_commit_data *p)
75 mempool_free(p, nfs_commit_mempool);
77 EXPORT_SYMBOL_GPL(nfs_commit_free);
79 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
81 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
83 if (p)
84 memset(p, 0, sizeof(*p));
85 return p;
88 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
90 mempool_free(hdr, nfs_wdata_mempool);
93 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
95 ctx->error = error;
96 smp_wmb();
97 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
101 * nfs_page_find_head_request_locked - find head request associated with @page
103 * must be called while holding the inode lock.
105 * returns matching head request with reference held, or NULL if not found.
107 static struct nfs_page *
108 nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page)
110 struct nfs_page *req = NULL;
112 if (PagePrivate(page))
113 req = (struct nfs_page *)page_private(page);
114 else if (unlikely(PageSwapCache(page)))
115 req = nfs_page_search_commits_for_head_request_locked(nfsi,
116 page);
118 if (req) {
119 WARN_ON_ONCE(req->wb_head != req);
120 kref_get(&req->wb_kref);
123 return req;
127 * nfs_page_find_head_request - find head request associated with @page
129 * returns matching head request with reference held, or NULL if not found.
131 static struct nfs_page *nfs_page_find_head_request(struct page *page)
133 struct inode *inode = page_file_mapping(page)->host;
134 struct nfs_page *req = NULL;
136 spin_lock(&inode->i_lock);
137 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
138 spin_unlock(&inode->i_lock);
139 return req;
142 /* Adjust the file length if we're writing beyond the end */
143 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
145 struct inode *inode = page_file_mapping(page)->host;
146 loff_t end, i_size;
147 pgoff_t end_index;
149 spin_lock(&inode->i_lock);
150 i_size = i_size_read(inode);
151 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
152 if (i_size > 0 && page_file_index(page) < end_index)
153 goto out;
154 end = page_file_offset(page) + ((loff_t)offset+count);
155 if (i_size >= end)
156 goto out;
157 i_size_write(inode, end);
158 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
159 out:
160 spin_unlock(&inode->i_lock);
163 /* A writeback failed: mark the page as bad, and invalidate the page cache */
164 static void nfs_set_pageerror(struct page *page)
166 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
170 * nfs_page_group_search_locked
171 * @head - head request of page group
172 * @page_offset - offset into page
174 * Search page group with head @head to find a request that contains the
175 * page offset @page_offset.
177 * Returns a pointer to the first matching nfs request, or NULL if no
178 * match is found.
180 * Must be called with the page group lock held
182 static struct nfs_page *
183 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
185 struct nfs_page *req;
187 WARN_ON_ONCE(head != head->wb_head);
188 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
190 req = head;
191 do {
192 if (page_offset >= req->wb_pgbase &&
193 page_offset < (req->wb_pgbase + req->wb_bytes))
194 return req;
196 req = req->wb_this_page;
197 } while (req != head);
199 return NULL;
203 * nfs_page_group_covers_page
204 * @head - head request of page group
206 * Return true if the page group with head @head covers the whole page,
207 * returns false otherwise
209 static bool nfs_page_group_covers_page(struct nfs_page *req)
211 struct nfs_page *tmp;
212 unsigned int pos = 0;
213 unsigned int len = nfs_page_length(req->wb_page);
215 nfs_page_group_lock(req, false);
217 do {
218 tmp = nfs_page_group_search_locked(req->wb_head, pos);
219 if (tmp) {
220 /* no way this should happen */
221 WARN_ON_ONCE(tmp->wb_pgbase != pos);
222 pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
224 } while (tmp && pos < len);
226 nfs_page_group_unlock(req);
227 WARN_ON_ONCE(pos > len);
228 return pos == len;
231 /* We can set the PG_uptodate flag if we see that a write request
232 * covers the full page.
234 static void nfs_mark_uptodate(struct nfs_page *req)
236 if (PageUptodate(req->wb_page))
237 return;
238 if (!nfs_page_group_covers_page(req))
239 return;
240 SetPageUptodate(req->wb_page);
243 static int wb_priority(struct writeback_control *wbc)
245 int ret = 0;
246 if (wbc->for_reclaim)
247 return FLUSH_HIGHPRI | FLUSH_STABLE;
248 if (wbc->sync_mode == WB_SYNC_ALL)
249 ret = FLUSH_COND_STABLE;
250 if (wbc->for_kupdate || wbc->for_background)
251 ret |= FLUSH_LOWPRI;
252 return ret;
256 * NFS congestion control
259 int nfs_congestion_kb;
261 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
262 #define NFS_CONGESTION_OFF_THRESH \
263 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
265 static void nfs_set_page_writeback(struct page *page)
267 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
268 int ret = test_set_page_writeback(page);
270 WARN_ON_ONCE(ret != 0);
272 if (atomic_long_inc_return(&nfss->writeback) >
273 NFS_CONGESTION_ON_THRESH) {
274 set_bdi_congested(&nfss->backing_dev_info,
275 BLK_RW_ASYNC);
279 static void nfs_end_page_writeback(struct nfs_page *req)
281 struct inode *inode = page_file_mapping(req->wb_page)->host;
282 struct nfs_server *nfss = NFS_SERVER(inode);
284 if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
285 return;
287 end_page_writeback(req->wb_page);
288 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
289 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
293 /* nfs_page_group_clear_bits
294 * @req - an nfs request
295 * clears all page group related bits from @req
297 static void
298 nfs_page_group_clear_bits(struct nfs_page *req)
300 clear_bit(PG_TEARDOWN, &req->wb_flags);
301 clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
302 clear_bit(PG_UPTODATE, &req->wb_flags);
303 clear_bit(PG_WB_END, &req->wb_flags);
304 clear_bit(PG_REMOVE, &req->wb_flags);
309 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
311 * this is a helper function for nfs_lock_and_join_requests
313 * @inode - inode associated with request page group, must be holding inode lock
314 * @head - head request of page group, must be holding head lock
315 * @req - request that couldn't lock and needs to wait on the req bit lock
316 * @nonblock - if true, don't actually wait
318 * NOTE: this must be called holding page_group bit lock and inode spin lock
319 * and BOTH will be released before returning.
321 * returns 0 on success, < 0 on error.
323 static int
324 nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
325 struct nfs_page *req, bool nonblock)
326 __releases(&inode->i_lock)
328 struct nfs_page *tmp;
329 int ret;
331 /* relinquish all the locks successfully grabbed this run */
332 for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
333 nfs_unlock_request(tmp);
335 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
337 /* grab a ref on the request that will be waited on */
338 kref_get(&req->wb_kref);
340 nfs_page_group_unlock(head);
341 spin_unlock(&inode->i_lock);
343 /* release ref from nfs_page_find_head_request_locked */
344 nfs_release_request(head);
346 if (!nonblock)
347 ret = nfs_wait_on_request(req);
348 else
349 ret = -EAGAIN;
350 nfs_release_request(req);
352 return ret;
356 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
358 * @destroy_list - request list (using wb_this_page) terminated by @old_head
359 * @old_head - the old head of the list
361 * All subrequests must be locked and removed from all lists, so at this point
362 * they are only "active" in this function, and possibly in nfs_wait_on_request
363 * with a reference held by some other context.
365 static void
366 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
367 struct nfs_page *old_head)
369 while (destroy_list) {
370 struct nfs_page *subreq = destroy_list;
372 destroy_list = (subreq->wb_this_page == old_head) ?
373 NULL : subreq->wb_this_page;
375 WARN_ON_ONCE(old_head != subreq->wb_head);
377 /* make sure old group is not used */
378 subreq->wb_head = subreq;
379 subreq->wb_this_page = subreq;
381 /* subreq is now totally disconnected from page group or any
382 * write / commit lists. last chance to wake any waiters */
383 nfs_unlock_request(subreq);
385 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
386 /* release ref on old head request */
387 nfs_release_request(old_head);
389 nfs_page_group_clear_bits(subreq);
391 /* release the PG_INODE_REF reference */
392 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
393 nfs_release_request(subreq);
394 else
395 WARN_ON_ONCE(1);
396 } else {
397 WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
398 /* zombie requests have already released the last
399 * reference and were waiting on the rest of the
400 * group to complete. Since it's no longer part of a
401 * group, simply free the request */
402 nfs_page_group_clear_bits(subreq);
403 nfs_free_request(subreq);
409 * nfs_lock_and_join_requests - join all subreqs to the head req and return
410 * a locked reference, cancelling any pending
411 * operations for this page.
413 * @page - the page used to lookup the "page group" of nfs_page structures
414 * @nonblock - if true, don't block waiting for request locks
416 * This function joins all sub requests to the head request by first
417 * locking all requests in the group, cancelling any pending operations
418 * and finally updating the head request to cover the whole range covered by
419 * the (former) group. All subrequests are removed from any write or commit
420 * lists, unlinked from the group and destroyed.
422 * Returns a locked, referenced pointer to the head request - which after
423 * this call is guaranteed to be the only request associated with the page.
424 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
425 * error was encountered.
427 static struct nfs_page *
428 nfs_lock_and_join_requests(struct page *page, bool nonblock)
430 struct inode *inode = page_file_mapping(page)->host;
431 struct nfs_page *head, *subreq;
432 struct nfs_page *destroy_list = NULL;
433 unsigned int total_bytes;
434 int ret;
436 try_again:
437 total_bytes = 0;
439 WARN_ON_ONCE(destroy_list);
441 spin_lock(&inode->i_lock);
444 * A reference is taken only on the head request which acts as a
445 * reference to the whole page group - the group will not be destroyed
446 * until the head reference is released.
448 head = nfs_page_find_head_request_locked(NFS_I(inode), page);
450 if (!head) {
451 spin_unlock(&inode->i_lock);
452 return NULL;
455 /* holding inode lock, so always make a non-blocking call to try the
456 * page group lock */
457 ret = nfs_page_group_lock(head, true);
458 if (ret < 0) {
459 spin_unlock(&inode->i_lock);
461 if (!nonblock && ret == -EAGAIN) {
462 nfs_page_group_lock_wait(head);
463 nfs_release_request(head);
464 goto try_again;
467 nfs_release_request(head);
468 return ERR_PTR(ret);
471 /* lock each request in the page group */
472 subreq = head;
473 do {
475 * Subrequests are always contiguous, non overlapping
476 * and in order - but may be repeated (mirrored writes).
478 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
479 /* keep track of how many bytes this group covers */
480 total_bytes += subreq->wb_bytes;
481 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
482 ((subreq->wb_offset + subreq->wb_bytes) >
483 (head->wb_offset + total_bytes)))) {
484 nfs_page_group_unlock(head);
485 spin_unlock(&inode->i_lock);
486 return ERR_PTR(-EIO);
489 if (!nfs_lock_request(subreq)) {
490 /* releases page group bit lock and
491 * inode spin lock and all references */
492 ret = nfs_unroll_locks_and_wait(inode, head,
493 subreq, nonblock);
495 if (ret == 0)
496 goto try_again;
498 return ERR_PTR(ret);
501 subreq = subreq->wb_this_page;
502 } while (subreq != head);
504 /* Now that all requests are locked, make sure they aren't on any list.
505 * Commit list removal accounting is done after locks are dropped */
506 subreq = head;
507 do {
508 nfs_clear_request_commit(subreq);
509 subreq = subreq->wb_this_page;
510 } while (subreq != head);
512 /* unlink subrequests from head, destroy them later */
513 if (head->wb_this_page != head) {
514 /* destroy list will be terminated by head */
515 destroy_list = head->wb_this_page;
516 head->wb_this_page = head;
518 /* change head request to cover whole range that
519 * the former page group covered */
520 head->wb_bytes = total_bytes;
524 * prepare head request to be added to new pgio descriptor
526 nfs_page_group_clear_bits(head);
529 * some part of the group was still on the inode list - otherwise
530 * the group wouldn't be involved in async write.
531 * grab a reference for the head request, iff it needs one.
533 if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
534 kref_get(&head->wb_kref);
536 nfs_page_group_unlock(head);
538 /* drop lock to clean uprequests on destroy list */
539 spin_unlock(&inode->i_lock);
541 nfs_destroy_unlinked_subrequests(destroy_list, head);
543 /* still holds ref on head from nfs_page_find_head_request_locked
544 * and still has lock on head from lock loop */
545 return head;
549 * Find an associated nfs write request, and prepare to flush it out
550 * May return an error if the user signalled nfs_wait_on_request().
552 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
553 struct page *page, bool nonblock)
555 struct nfs_page *req;
556 int ret = 0;
558 req = nfs_lock_and_join_requests(page, nonblock);
559 if (!req)
560 goto out;
561 ret = PTR_ERR(req);
562 if (IS_ERR(req))
563 goto out;
565 nfs_set_page_writeback(page);
566 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
568 ret = 0;
569 if (!nfs_pageio_add_request(pgio, req)) {
570 nfs_redirty_request(req);
571 ret = pgio->pg_error;
572 } else
573 nfs_add_stats(page_file_mapping(page)->host,
574 NFSIOS_WRITEPAGES, 1);
575 out:
576 return ret;
579 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
581 int ret;
583 nfs_pageio_cond_complete(pgio, page_file_index(page));
584 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
585 if (ret == -EAGAIN) {
586 redirty_page_for_writepage(wbc, page);
587 ret = 0;
589 return ret;
593 * Write an mmapped page to the server.
595 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
597 struct nfs_pageio_descriptor pgio;
598 struct inode *inode = page_file_mapping(page)->host;
599 int err;
601 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
602 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc),
603 false, &nfs_async_write_completion_ops);
604 err = nfs_do_writepage(page, wbc, &pgio);
605 nfs_pageio_complete(&pgio);
606 if (err < 0)
607 return err;
608 if (pgio.pg_error < 0)
609 return pgio.pg_error;
610 return 0;
613 int nfs_writepage(struct page *page, struct writeback_control *wbc)
615 int ret;
617 ret = nfs_writepage_locked(page, wbc);
618 unlock_page(page);
619 return ret;
622 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
624 int ret;
626 ret = nfs_do_writepage(page, wbc, data);
627 unlock_page(page);
628 return ret;
631 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
633 struct inode *inode = mapping->host;
634 unsigned long *bitlock = &NFS_I(inode)->flags;
635 struct nfs_pageio_descriptor pgio;
636 int err;
638 /* Stop dirtying of new pages while we sync */
639 err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING,
640 nfs_wait_bit_killable, TASK_KILLABLE);
641 if (err)
642 goto out_err;
644 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
646 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
647 &nfs_async_write_completion_ops);
648 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
649 nfs_pageio_complete(&pgio);
651 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
652 smp_mb__after_atomic();
653 wake_up_bit(bitlock, NFS_INO_FLUSHING);
655 if (err < 0)
656 goto out_err;
657 err = pgio.pg_error;
658 if (err < 0)
659 goto out_err;
660 return 0;
661 out_err:
662 return err;
666 * Insert a write request into an inode
668 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
670 struct nfs_inode *nfsi = NFS_I(inode);
672 WARN_ON_ONCE(req->wb_this_page != req);
674 /* Lock the request! */
675 nfs_lock_request(req);
677 spin_lock(&inode->i_lock);
678 if (!nfsi->nrequests &&
679 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
680 inode->i_version++;
682 * Swap-space should not get truncated. Hence no need to plug the race
683 * with invalidate/truncate.
685 if (likely(!PageSwapCache(req->wb_page))) {
686 set_bit(PG_MAPPED, &req->wb_flags);
687 SetPagePrivate(req->wb_page);
688 set_page_private(req->wb_page, (unsigned long)req);
690 nfsi->nrequests++;
691 /* this a head request for a page group - mark it as having an
692 * extra reference so sub groups can follow suit.
693 * This flag also informs pgio layer when to bump nrequests when
694 * adding subrequests. */
695 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
696 kref_get(&req->wb_kref);
697 spin_unlock(&inode->i_lock);
701 * Remove a write request from an inode
703 static void nfs_inode_remove_request(struct nfs_page *req)
705 struct inode *inode = d_inode(req->wb_context->dentry);
706 struct nfs_inode *nfsi = NFS_I(inode);
707 struct nfs_page *head;
709 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
710 head = req->wb_head;
712 spin_lock(&inode->i_lock);
713 if (likely(!PageSwapCache(head->wb_page))) {
714 set_page_private(head->wb_page, 0);
715 ClearPagePrivate(head->wb_page);
716 smp_mb__after_atomic();
717 wake_up_page(head->wb_page, PG_private);
718 clear_bit(PG_MAPPED, &head->wb_flags);
720 nfsi->nrequests--;
721 spin_unlock(&inode->i_lock);
722 } else {
723 spin_lock(&inode->i_lock);
724 nfsi->nrequests--;
725 spin_unlock(&inode->i_lock);
728 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
729 nfs_release_request(req);
732 static void
733 nfs_mark_request_dirty(struct nfs_page *req)
735 __set_page_dirty_nobuffers(req->wb_page);
739 * nfs_page_search_commits_for_head_request_locked
741 * Search through commit lists on @inode for the head request for @page.
742 * Must be called while holding the inode (which is cinfo) lock.
744 * Returns the head request if found, or NULL if not found.
746 static struct nfs_page *
747 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
748 struct page *page)
750 struct nfs_page *freq, *t;
751 struct nfs_commit_info cinfo;
752 struct inode *inode = &nfsi->vfs_inode;
754 nfs_init_cinfo_from_inode(&cinfo, inode);
756 /* search through pnfs commit lists */
757 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
758 if (freq)
759 return freq->wb_head;
761 /* Linearly search the commit list for the correct request */
762 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
763 if (freq->wb_page == page)
764 return freq->wb_head;
767 return NULL;
771 * nfs_request_add_commit_list_locked - add request to a commit list
772 * @req: pointer to a struct nfs_page
773 * @dst: commit list head
774 * @cinfo: holds list lock and accounting info
776 * This sets the PG_CLEAN bit, updates the cinfo count of
777 * number of outstanding requests requiring a commit as well as
778 * the MM page stats.
780 * The caller must hold the cinfo->lock, and the nfs_page lock.
782 void
783 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
784 struct nfs_commit_info *cinfo)
786 set_bit(PG_CLEAN, &req->wb_flags);
787 nfs_list_add_request(req, dst);
788 cinfo->mds->ncommit++;
790 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
793 * nfs_request_add_commit_list - add request to a commit list
794 * @req: pointer to a struct nfs_page
795 * @dst: commit list head
796 * @cinfo: holds list lock and accounting info
798 * This sets the PG_CLEAN bit, updates the cinfo count of
799 * number of outstanding requests requiring a commit as well as
800 * the MM page stats.
802 * The caller must _not_ hold the cinfo->lock, but must be
803 * holding the nfs_page lock.
805 void
806 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
807 struct nfs_commit_info *cinfo)
809 spin_lock(cinfo->lock);
810 nfs_request_add_commit_list_locked(req, dst, cinfo);
811 spin_unlock(cinfo->lock);
812 nfs_mark_page_unstable(req->wb_page, cinfo);
814 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
817 * nfs_request_remove_commit_list - Remove request from a commit list
818 * @req: pointer to a nfs_page
819 * @cinfo: holds list lock and accounting info
821 * This clears the PG_CLEAN bit, and updates the cinfo's count of
822 * number of outstanding requests requiring a commit
823 * It does not update the MM page stats.
825 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
827 void
828 nfs_request_remove_commit_list(struct nfs_page *req,
829 struct nfs_commit_info *cinfo)
831 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
832 return;
833 nfs_list_remove_request(req);
834 cinfo->mds->ncommit--;
836 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
838 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
839 struct inode *inode)
841 cinfo->lock = &inode->i_lock;
842 cinfo->mds = &NFS_I(inode)->commit_info;
843 cinfo->ds = pnfs_get_ds_info(inode);
844 cinfo->dreq = NULL;
845 cinfo->completion_ops = &nfs_commit_completion_ops;
848 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
849 struct inode *inode,
850 struct nfs_direct_req *dreq)
852 if (dreq)
853 nfs_init_cinfo_from_dreq(cinfo, dreq);
854 else
855 nfs_init_cinfo_from_inode(cinfo, inode);
857 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
860 * Add a request to the inode's commit list.
862 void
863 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
864 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
866 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
867 return;
868 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
871 static void
872 nfs_clear_page_commit(struct page *page)
874 dec_zone_page_state(page, NR_UNSTABLE_NFS);
875 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
876 WB_RECLAIMABLE);
879 /* Called holding inode (/cinfo) lock */
880 static void
881 nfs_clear_request_commit(struct nfs_page *req)
883 if (test_bit(PG_CLEAN, &req->wb_flags)) {
884 struct inode *inode = d_inode(req->wb_context->dentry);
885 struct nfs_commit_info cinfo;
887 nfs_init_cinfo_from_inode(&cinfo, inode);
888 if (!pnfs_clear_request_commit(req, &cinfo)) {
889 nfs_request_remove_commit_list(req, &cinfo);
891 nfs_clear_page_commit(req->wb_page);
895 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
897 if (hdr->verf.committed == NFS_DATA_SYNC)
898 return hdr->lseg == NULL;
899 return hdr->verf.committed != NFS_FILE_SYNC;
902 static void nfs_write_completion(struct nfs_pgio_header *hdr)
904 struct nfs_commit_info cinfo;
905 unsigned long bytes = 0;
907 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
908 goto out;
909 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
910 while (!list_empty(&hdr->pages)) {
911 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
913 bytes += req->wb_bytes;
914 nfs_list_remove_request(req);
915 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
916 (hdr->good_bytes < bytes)) {
917 nfs_set_pageerror(req->wb_page);
918 nfs_context_set_write_error(req->wb_context, hdr->error);
919 goto remove_req;
921 if (nfs_write_need_commit(hdr)) {
922 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
923 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
924 hdr->pgio_mirror_idx);
925 goto next;
927 remove_req:
928 nfs_inode_remove_request(req);
929 next:
930 nfs_unlock_request(req);
931 nfs_end_page_writeback(req);
932 nfs_release_request(req);
934 out:
935 hdr->release(hdr);
938 unsigned long
939 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
941 return cinfo->mds->ncommit;
944 /* cinfo->lock held by caller */
946 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
947 struct nfs_commit_info *cinfo, int max)
949 struct nfs_page *req, *tmp;
950 int ret = 0;
952 list_for_each_entry_safe(req, tmp, src, wb_list) {
953 if (!nfs_lock_request(req))
954 continue;
955 kref_get(&req->wb_kref);
956 if (cond_resched_lock(cinfo->lock))
957 list_safe_reset_next(req, tmp, wb_list);
958 nfs_request_remove_commit_list(req, cinfo);
959 nfs_list_add_request(req, dst);
960 ret++;
961 if ((ret == max) && !cinfo->dreq)
962 break;
964 return ret;
968 * nfs_scan_commit - Scan an inode for commit requests
969 * @inode: NFS inode to scan
970 * @dst: mds destination list
971 * @cinfo: mds and ds lists of reqs ready to commit
973 * Moves requests from the inode's 'commit' request list.
974 * The requests are *not* checked to ensure that they form a contiguous set.
977 nfs_scan_commit(struct inode *inode, struct list_head *dst,
978 struct nfs_commit_info *cinfo)
980 int ret = 0;
982 spin_lock(cinfo->lock);
983 if (cinfo->mds->ncommit > 0) {
984 const int max = INT_MAX;
986 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
987 cinfo, max);
988 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
990 spin_unlock(cinfo->lock);
991 return ret;
995 * Search for an existing write request, and attempt to update
996 * it to reflect a new dirty region on a given page.
998 * If the attempt fails, then the existing request is flushed out
999 * to disk.
1001 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1002 struct page *page,
1003 unsigned int offset,
1004 unsigned int bytes)
1006 struct nfs_page *req;
1007 unsigned int rqend;
1008 unsigned int end;
1009 int error;
1011 if (!PagePrivate(page))
1012 return NULL;
1014 end = offset + bytes;
1015 spin_lock(&inode->i_lock);
1017 for (;;) {
1018 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
1019 if (req == NULL)
1020 goto out_unlock;
1022 /* should be handled by nfs_flush_incompatible */
1023 WARN_ON_ONCE(req->wb_head != req);
1024 WARN_ON_ONCE(req->wb_this_page != req);
1026 rqend = req->wb_offset + req->wb_bytes;
1028 * Tell the caller to flush out the request if
1029 * the offsets are non-contiguous.
1030 * Note: nfs_flush_incompatible() will already
1031 * have flushed out requests having wrong owners.
1033 if (offset > rqend
1034 || end < req->wb_offset)
1035 goto out_flushme;
1037 if (nfs_lock_request(req))
1038 break;
1040 /* The request is locked, so wait and then retry */
1041 spin_unlock(&inode->i_lock);
1042 error = nfs_wait_on_request(req);
1043 nfs_release_request(req);
1044 if (error != 0)
1045 goto out_err;
1046 spin_lock(&inode->i_lock);
1049 /* Okay, the request matches. Update the region */
1050 if (offset < req->wb_offset) {
1051 req->wb_offset = offset;
1052 req->wb_pgbase = offset;
1054 if (end > rqend)
1055 req->wb_bytes = end - req->wb_offset;
1056 else
1057 req->wb_bytes = rqend - req->wb_offset;
1058 out_unlock:
1059 if (req)
1060 nfs_clear_request_commit(req);
1061 spin_unlock(&inode->i_lock);
1062 return req;
1063 out_flushme:
1064 spin_unlock(&inode->i_lock);
1065 nfs_release_request(req);
1066 error = nfs_wb_page(inode, page);
1067 out_err:
1068 return ERR_PTR(error);
1072 * Try to update an existing write request, or create one if there is none.
1074 * Note: Should always be called with the Page Lock held to prevent races
1075 * if we have to add a new request. Also assumes that the caller has
1076 * already called nfs_flush_incompatible() if necessary.
1078 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1079 struct page *page, unsigned int offset, unsigned int bytes)
1081 struct inode *inode = page_file_mapping(page)->host;
1082 struct nfs_page *req;
1084 req = nfs_try_to_update_request(inode, page, offset, bytes);
1085 if (req != NULL)
1086 goto out;
1087 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1088 if (IS_ERR(req))
1089 goto out;
1090 nfs_inode_add_request(inode, req);
1091 out:
1092 return req;
1095 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1096 unsigned int offset, unsigned int count)
1098 struct nfs_page *req;
1100 req = nfs_setup_write_request(ctx, page, offset, count);
1101 if (IS_ERR(req))
1102 return PTR_ERR(req);
1103 /* Update file length */
1104 nfs_grow_file(page, offset, count);
1105 nfs_mark_uptodate(req);
1106 nfs_mark_request_dirty(req);
1107 nfs_unlock_and_release_request(req);
1108 return 0;
1111 int nfs_flush_incompatible(struct file *file, struct page *page)
1113 struct nfs_open_context *ctx = nfs_file_open_context(file);
1114 struct nfs_lock_context *l_ctx;
1115 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1116 struct nfs_page *req;
1117 int do_flush, status;
1119 * Look for a request corresponding to this page. If there
1120 * is one, and it belongs to another file, we flush it out
1121 * before we try to copy anything into the page. Do this
1122 * due to the lack of an ACCESS-type call in NFSv2.
1123 * Also do the same if we find a request from an existing
1124 * dropped page.
1126 do {
1127 req = nfs_page_find_head_request(page);
1128 if (req == NULL)
1129 return 0;
1130 l_ctx = req->wb_lock_context;
1131 do_flush = req->wb_page != page || req->wb_context != ctx;
1132 /* for now, flush if more than 1 request in page_group */
1133 do_flush |= req->wb_this_page != req;
1134 if (l_ctx && flctx &&
1135 !(list_empty_careful(&flctx->flc_posix) &&
1136 list_empty_careful(&flctx->flc_flock))) {
1137 do_flush |= l_ctx->lockowner.l_owner != current->files
1138 || l_ctx->lockowner.l_pid != current->tgid;
1140 nfs_release_request(req);
1141 if (!do_flush)
1142 return 0;
1143 status = nfs_wb_page(page_file_mapping(page)->host, page);
1144 } while (status == 0);
1145 return status;
1149 * Avoid buffered writes when a open context credential's key would
1150 * expire soon.
1152 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1154 * Return 0 and set a credential flag which triggers the inode to flush
1155 * and performs NFS_FILE_SYNC writes if the key will expired within
1156 * RPC_KEY_EXPIRE_TIMEO.
1159 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1161 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1162 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1164 return rpcauth_key_timeout_notify(auth, ctx->cred);
1168 * Test if the open context credential key is marked to expire soon.
1170 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1172 return rpcauth_cred_key_to_expire(ctx->cred);
1176 * If the page cache is marked as unsafe or invalid, then we can't rely on
1177 * the PageUptodate() flag. In this case, we will need to turn off
1178 * write optimisations that depend on the page contents being correct.
1180 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1182 struct nfs_inode *nfsi = NFS_I(inode);
1184 if (nfs_have_delegated_attributes(inode))
1185 goto out;
1186 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1187 return false;
1188 smp_rmb();
1189 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1190 return false;
1191 out:
1192 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1193 return false;
1194 return PageUptodate(page) != 0;
1197 static bool
1198 is_whole_file_wrlock(struct file_lock *fl)
1200 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1201 fl->fl_type == F_WRLCK;
1204 /* If we know the page is up to date, and we're not using byte range locks (or
1205 * if we have the whole file locked for writing), it may be more efficient to
1206 * extend the write to cover the entire page in order to avoid fragmentation
1207 * inefficiencies.
1209 * If the file is opened for synchronous writes then we can just skip the rest
1210 * of the checks.
1212 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1214 int ret;
1215 struct file_lock_context *flctx = inode->i_flctx;
1216 struct file_lock *fl;
1218 if (file->f_flags & O_DSYNC)
1219 return 0;
1220 if (!nfs_write_pageuptodate(page, inode))
1221 return 0;
1222 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1223 return 1;
1224 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1225 list_empty_careful(&flctx->flc_posix)))
1226 return 1;
1228 /* Check to see if there are whole file write locks */
1229 ret = 0;
1230 spin_lock(&flctx->flc_lock);
1231 if (!list_empty(&flctx->flc_posix)) {
1232 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1233 fl_list);
1234 if (is_whole_file_wrlock(fl))
1235 ret = 1;
1236 } else if (!list_empty(&flctx->flc_flock)) {
1237 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1238 fl_list);
1239 if (fl->fl_type == F_WRLCK)
1240 ret = 1;
1242 spin_unlock(&flctx->flc_lock);
1243 return ret;
1247 * Update and possibly write a cached page of an NFS file.
1249 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1250 * things with a page scheduled for an RPC call (e.g. invalidate it).
1252 int nfs_updatepage(struct file *file, struct page *page,
1253 unsigned int offset, unsigned int count)
1255 struct nfs_open_context *ctx = nfs_file_open_context(file);
1256 struct inode *inode = page_file_mapping(page)->host;
1257 int status = 0;
1259 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1261 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1262 file, count, (long long)(page_file_offset(page) + offset));
1264 if (!count)
1265 goto out;
1267 if (nfs_can_extend_write(file, page, inode)) {
1268 count = max(count + offset, nfs_page_length(page));
1269 offset = 0;
1272 status = nfs_writepage_setup(ctx, page, offset, count);
1273 if (status < 0)
1274 nfs_set_pageerror(page);
1275 else
1276 __set_page_dirty_nobuffers(page);
1277 out:
1278 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1279 status, (long long)i_size_read(inode));
1280 return status;
1283 static int flush_task_priority(int how)
1285 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1286 case FLUSH_HIGHPRI:
1287 return RPC_PRIORITY_HIGH;
1288 case FLUSH_LOWPRI:
1289 return RPC_PRIORITY_LOW;
1291 return RPC_PRIORITY_NORMAL;
1294 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1295 struct rpc_message *msg,
1296 const struct nfs_rpc_ops *rpc_ops,
1297 struct rpc_task_setup *task_setup_data, int how)
1299 int priority = flush_task_priority(how);
1301 task_setup_data->priority = priority;
1302 rpc_ops->write_setup(hdr, msg);
1304 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1305 &task_setup_data->rpc_client, msg, hdr);
1308 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1309 * call this on each, which will prepare them to be retried on next
1310 * writeback using standard nfs.
1312 static void nfs_redirty_request(struct nfs_page *req)
1314 nfs_mark_request_dirty(req);
1315 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1316 nfs_unlock_request(req);
1317 nfs_end_page_writeback(req);
1318 nfs_release_request(req);
1321 static void nfs_async_write_error(struct list_head *head)
1323 struct nfs_page *req;
1325 while (!list_empty(head)) {
1326 req = nfs_list_entry(head->next);
1327 nfs_list_remove_request(req);
1328 nfs_redirty_request(req);
1332 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1333 .error_cleanup = nfs_async_write_error,
1334 .completion = nfs_write_completion,
1337 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1338 struct inode *inode, int ioflags, bool force_mds,
1339 const struct nfs_pgio_completion_ops *compl_ops)
1341 struct nfs_server *server = NFS_SERVER(inode);
1342 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1344 #ifdef CONFIG_NFS_V4_1
1345 if (server->pnfs_curr_ld && !force_mds)
1346 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1347 #endif
1348 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1349 server->wsize, ioflags);
1351 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1353 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1355 struct nfs_pgio_mirror *mirror;
1357 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1358 pgio->pg_ops->pg_cleanup(pgio);
1360 pgio->pg_ops = &nfs_pgio_rw_ops;
1362 nfs_pageio_stop_mirroring(pgio);
1364 mirror = &pgio->pg_mirrors[0];
1365 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1367 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1370 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1372 struct nfs_commit_data *data = calldata;
1374 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1378 * Special version of should_remove_suid() that ignores capabilities.
1380 static int nfs_should_remove_suid(const struct inode *inode)
1382 umode_t mode = inode->i_mode;
1383 int kill = 0;
1385 /* suid always must be killed */
1386 if (unlikely(mode & S_ISUID))
1387 kill = ATTR_KILL_SUID;
1390 * sgid without any exec bits is just a mandatory locking mark; leave
1391 * it alone. If some exec bits are set, it's a real sgid; kill it.
1393 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1394 kill |= ATTR_KILL_SGID;
1396 if (unlikely(kill && S_ISREG(mode)))
1397 return kill;
1399 return 0;
1402 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1403 struct nfs_fattr *fattr)
1405 struct nfs_pgio_args *argp = &hdr->args;
1406 struct nfs_pgio_res *resp = &hdr->res;
1407 u64 size = argp->offset + resp->count;
1409 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1410 fattr->size = size;
1411 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1412 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1413 return;
1415 if (size != fattr->size)
1416 return;
1417 /* Set attribute barrier */
1418 nfs_fattr_set_barrier(fattr);
1419 /* ...and update size */
1420 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1423 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1425 struct nfs_fattr *fattr = &hdr->fattr;
1426 struct inode *inode = hdr->inode;
1428 spin_lock(&inode->i_lock);
1429 nfs_writeback_check_extend(hdr, fattr);
1430 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1431 spin_unlock(&inode->i_lock);
1433 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1436 * This function is called when the WRITE call is complete.
1438 static int nfs_writeback_done(struct rpc_task *task,
1439 struct nfs_pgio_header *hdr,
1440 struct inode *inode)
1442 int status;
1445 * ->write_done will attempt to use post-op attributes to detect
1446 * conflicting writes by other clients. A strict interpretation
1447 * of close-to-open would allow us to continue caching even if
1448 * another writer had changed the file, but some applications
1449 * depend on tighter cache coherency when writing.
1451 status = NFS_PROTO(inode)->write_done(task, hdr);
1452 if (status != 0)
1453 return status;
1454 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1456 if (hdr->res.verf->committed < hdr->args.stable &&
1457 task->tk_status >= 0) {
1458 /* We tried a write call, but the server did not
1459 * commit data to stable storage even though we
1460 * requested it.
1461 * Note: There is a known bug in Tru64 < 5.0 in which
1462 * the server reports NFS_DATA_SYNC, but performs
1463 * NFS_FILE_SYNC. We therefore implement this checking
1464 * as a dprintk() in order to avoid filling syslog.
1466 static unsigned long complain;
1468 /* Note this will print the MDS for a DS write */
1469 if (time_before(complain, jiffies)) {
1470 dprintk("NFS: faulty NFS server %s:"
1471 " (committed = %d) != (stable = %d)\n",
1472 NFS_SERVER(inode)->nfs_client->cl_hostname,
1473 hdr->res.verf->committed, hdr->args.stable);
1474 complain = jiffies + 300 * HZ;
1478 /* Deal with the suid/sgid bit corner case */
1479 if (nfs_should_remove_suid(inode))
1480 nfs_mark_for_revalidate(inode);
1481 return 0;
1485 * This function is called when the WRITE call is complete.
1487 static void nfs_writeback_result(struct rpc_task *task,
1488 struct nfs_pgio_header *hdr)
1490 struct nfs_pgio_args *argp = &hdr->args;
1491 struct nfs_pgio_res *resp = &hdr->res;
1493 if (resp->count < argp->count) {
1494 static unsigned long complain;
1496 /* This a short write! */
1497 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1499 /* Has the server at least made some progress? */
1500 if (resp->count == 0) {
1501 if (time_before(complain, jiffies)) {
1502 printk(KERN_WARNING
1503 "NFS: Server wrote zero bytes, expected %u.\n",
1504 argp->count);
1505 complain = jiffies + 300 * HZ;
1507 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1508 task->tk_status = -EIO;
1509 return;
1512 /* For non rpc-based layout drivers, retry-through-MDS */
1513 if (!task->tk_ops) {
1514 hdr->pnfs_error = -EAGAIN;
1515 return;
1518 /* Was this an NFSv2 write or an NFSv3 stable write? */
1519 if (resp->verf->committed != NFS_UNSTABLE) {
1520 /* Resend from where the server left off */
1521 hdr->mds_offset += resp->count;
1522 argp->offset += resp->count;
1523 argp->pgbase += resp->count;
1524 argp->count -= resp->count;
1525 } else {
1526 /* Resend as a stable write in order to avoid
1527 * headaches in the case of a server crash.
1529 argp->stable = NFS_FILE_SYNC;
1531 rpc_restart_call_prepare(task);
1536 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1538 int ret;
1540 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1541 return 1;
1542 if (!may_wait)
1543 return 0;
1544 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1545 NFS_INO_COMMIT,
1546 nfs_wait_bit_killable,
1547 TASK_KILLABLE);
1548 return (ret < 0) ? ret : 1;
1551 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1553 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1554 smp_mb__after_atomic();
1555 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1558 void nfs_commitdata_release(struct nfs_commit_data *data)
1560 put_nfs_open_context(data->context);
1561 nfs_commit_free(data);
1563 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1565 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1566 const struct nfs_rpc_ops *nfs_ops,
1567 const struct rpc_call_ops *call_ops,
1568 int how, int flags)
1570 struct rpc_task *task;
1571 int priority = flush_task_priority(how);
1572 struct rpc_message msg = {
1573 .rpc_argp = &data->args,
1574 .rpc_resp = &data->res,
1575 .rpc_cred = data->cred,
1577 struct rpc_task_setup task_setup_data = {
1578 .task = &data->task,
1579 .rpc_client = clnt,
1580 .rpc_message = &msg,
1581 .callback_ops = call_ops,
1582 .callback_data = data,
1583 .workqueue = nfsiod_workqueue,
1584 .flags = RPC_TASK_ASYNC | flags,
1585 .priority = priority,
1587 /* Set up the initial task struct. */
1588 nfs_ops->commit_setup(data, &msg);
1590 dprintk("NFS: initiated commit call\n");
1592 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1593 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1595 task = rpc_run_task(&task_setup_data);
1596 if (IS_ERR(task))
1597 return PTR_ERR(task);
1598 if (how & FLUSH_SYNC)
1599 rpc_wait_for_completion_task(task);
1600 rpc_put_task(task);
1601 return 0;
1603 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1605 static loff_t nfs_get_lwb(struct list_head *head)
1607 loff_t lwb = 0;
1608 struct nfs_page *req;
1610 list_for_each_entry(req, head, wb_list)
1611 if (lwb < (req_offset(req) + req->wb_bytes))
1612 lwb = req_offset(req) + req->wb_bytes;
1614 return lwb;
1618 * Set up the argument/result storage required for the RPC call.
1620 void nfs_init_commit(struct nfs_commit_data *data,
1621 struct list_head *head,
1622 struct pnfs_layout_segment *lseg,
1623 struct nfs_commit_info *cinfo)
1625 struct nfs_page *first = nfs_list_entry(head->next);
1626 struct inode *inode = d_inode(first->wb_context->dentry);
1628 /* Set up the RPC argument and reply structs
1629 * NB: take care not to mess about with data->commit et al. */
1631 list_splice_init(head, &data->pages);
1633 data->inode = inode;
1634 data->cred = first->wb_context->cred;
1635 data->lseg = lseg; /* reference transferred */
1636 /* only set lwb for pnfs commit */
1637 if (lseg)
1638 data->lwb = nfs_get_lwb(&data->pages);
1639 data->mds_ops = &nfs_commit_ops;
1640 data->completion_ops = cinfo->completion_ops;
1641 data->dreq = cinfo->dreq;
1643 data->args.fh = NFS_FH(data->inode);
1644 /* Note: we always request a commit of the entire inode */
1645 data->args.offset = 0;
1646 data->args.count = 0;
1647 data->context = get_nfs_open_context(first->wb_context);
1648 data->res.fattr = &data->fattr;
1649 data->res.verf = &data->verf;
1650 nfs_fattr_init(&data->fattr);
1652 EXPORT_SYMBOL_GPL(nfs_init_commit);
1654 void nfs_retry_commit(struct list_head *page_list,
1655 struct pnfs_layout_segment *lseg,
1656 struct nfs_commit_info *cinfo,
1657 u32 ds_commit_idx)
1659 struct nfs_page *req;
1661 while (!list_empty(page_list)) {
1662 req = nfs_list_entry(page_list->next);
1663 nfs_list_remove_request(req);
1664 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1665 if (!cinfo->dreq)
1666 nfs_clear_page_commit(req->wb_page);
1667 nfs_unlock_and_release_request(req);
1670 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1673 * Commit dirty pages
1675 static int
1676 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1677 struct nfs_commit_info *cinfo)
1679 struct nfs_commit_data *data;
1681 data = nfs_commitdata_alloc();
1683 if (!data)
1684 goto out_bad;
1686 /* Set up the argument struct */
1687 nfs_init_commit(data, head, NULL, cinfo);
1688 atomic_inc(&cinfo->mds->rpcs_out);
1689 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1690 data->mds_ops, how, 0);
1691 out_bad:
1692 nfs_retry_commit(head, NULL, cinfo, 0);
1693 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1694 return -ENOMEM;
1698 * COMMIT call returned
1700 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1702 struct nfs_commit_data *data = calldata;
1704 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1705 task->tk_pid, task->tk_status);
1707 /* Call the NFS version-specific code */
1708 NFS_PROTO(data->inode)->commit_done(task, data);
1711 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1713 struct nfs_page *req;
1714 int status = data->task.tk_status;
1715 struct nfs_commit_info cinfo;
1716 struct nfs_server *nfss;
1718 while (!list_empty(&data->pages)) {
1719 req = nfs_list_entry(data->pages.next);
1720 nfs_list_remove_request(req);
1721 nfs_clear_page_commit(req->wb_page);
1723 dprintk("NFS: commit (%s/%llu %d@%lld)",
1724 req->wb_context->dentry->d_sb->s_id,
1725 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1726 req->wb_bytes,
1727 (long long)req_offset(req));
1728 if (status < 0) {
1729 nfs_context_set_write_error(req->wb_context, status);
1730 nfs_inode_remove_request(req);
1731 dprintk(", error = %d\n", status);
1732 goto next;
1735 /* Okay, COMMIT succeeded, apparently. Check the verifier
1736 * returned by the server against all stored verfs. */
1737 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1738 /* We have a match */
1739 nfs_inode_remove_request(req);
1740 dprintk(" OK\n");
1741 goto next;
1743 /* We have a mismatch. Write the page again */
1744 dprintk(" mismatch\n");
1745 nfs_mark_request_dirty(req);
1746 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1747 next:
1748 nfs_unlock_and_release_request(req);
1749 /* Latency breaker */
1750 cond_resched();
1752 nfss = NFS_SERVER(data->inode);
1753 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1754 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1756 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1757 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1758 nfs_commit_clear_lock(NFS_I(data->inode));
1761 static void nfs_commit_release(void *calldata)
1763 struct nfs_commit_data *data = calldata;
1765 data->completion_ops->completion(data);
1766 nfs_commitdata_release(calldata);
1769 static const struct rpc_call_ops nfs_commit_ops = {
1770 .rpc_call_prepare = nfs_commit_prepare,
1771 .rpc_call_done = nfs_commit_done,
1772 .rpc_release = nfs_commit_release,
1775 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1776 .completion = nfs_commit_release_pages,
1777 .error_cleanup = nfs_commit_clear_lock,
1780 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1781 int how, struct nfs_commit_info *cinfo)
1783 int status;
1785 status = pnfs_commit_list(inode, head, how, cinfo);
1786 if (status == PNFS_NOT_ATTEMPTED)
1787 status = nfs_commit_list(inode, head, how, cinfo);
1788 return status;
1791 int nfs_commit_inode(struct inode *inode, int how)
1793 LIST_HEAD(head);
1794 struct nfs_commit_info cinfo;
1795 int may_wait = how & FLUSH_SYNC;
1796 int res;
1798 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1799 if (res <= 0)
1800 goto out_mark_dirty;
1801 nfs_init_cinfo_from_inode(&cinfo, inode);
1802 res = nfs_scan_commit(inode, &head, &cinfo);
1803 if (res) {
1804 int error;
1806 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1807 if (error < 0)
1808 return error;
1809 if (!may_wait)
1810 goto out_mark_dirty;
1811 error = wait_on_bit_action(&NFS_I(inode)->flags,
1812 NFS_INO_COMMIT,
1813 nfs_wait_bit_killable,
1814 TASK_KILLABLE);
1815 if (error < 0)
1816 return error;
1817 } else
1818 nfs_commit_clear_lock(NFS_I(inode));
1819 return res;
1820 /* Note: If we exit without ensuring that the commit is complete,
1821 * we must mark the inode as dirty. Otherwise, future calls to
1822 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1823 * that the data is on the disk.
1825 out_mark_dirty:
1826 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1827 return res;
1830 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1832 struct nfs_inode *nfsi = NFS_I(inode);
1833 int flags = FLUSH_SYNC;
1834 int ret = 0;
1836 /* no commits means nothing needs to be done */
1837 if (!nfsi->commit_info.ncommit)
1838 return ret;
1840 if (wbc->sync_mode == WB_SYNC_NONE) {
1841 /* Don't commit yet if this is a non-blocking flush and there
1842 * are a lot of outstanding writes for this mapping.
1844 if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1))
1845 goto out_mark_dirty;
1847 /* don't wait for the COMMIT response */
1848 flags = 0;
1851 ret = nfs_commit_inode(inode, flags);
1852 if (ret >= 0) {
1853 if (wbc->sync_mode == WB_SYNC_NONE) {
1854 if (ret < wbc->nr_to_write)
1855 wbc->nr_to_write -= ret;
1856 else
1857 wbc->nr_to_write = 0;
1859 return 0;
1861 out_mark_dirty:
1862 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1863 return ret;
1865 EXPORT_SYMBOL_GPL(nfs_write_inode);
1868 * flush the inode to disk.
1870 int nfs_wb_all(struct inode *inode)
1872 int ret;
1874 trace_nfs_writeback_inode_enter(inode);
1876 ret = filemap_write_and_wait(inode->i_mapping);
1877 if (ret)
1878 goto out;
1879 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1880 if (ret < 0)
1881 goto out;
1882 pnfs_sync_inode(inode, true);
1883 ret = 0;
1885 out:
1886 trace_nfs_writeback_inode_exit(inode, ret);
1887 return ret;
1889 EXPORT_SYMBOL_GPL(nfs_wb_all);
1891 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1893 struct nfs_page *req;
1894 int ret = 0;
1896 wait_on_page_writeback(page);
1898 /* blocking call to cancel all requests and join to a single (head)
1899 * request */
1900 req = nfs_lock_and_join_requests(page, false);
1902 if (IS_ERR(req)) {
1903 ret = PTR_ERR(req);
1904 } else if (req) {
1905 /* all requests from this page have been cancelled by
1906 * nfs_lock_and_join_requests, so just remove the head
1907 * request from the inode / page_private pointer and
1908 * release it */
1909 nfs_inode_remove_request(req);
1910 nfs_unlock_and_release_request(req);
1913 return ret;
1917 * Write back all requests on one page - we do this before reading it.
1919 int nfs_wb_page(struct inode *inode, struct page *page)
1921 loff_t range_start = page_file_offset(page);
1922 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1923 struct writeback_control wbc = {
1924 .sync_mode = WB_SYNC_ALL,
1925 .nr_to_write = 0,
1926 .range_start = range_start,
1927 .range_end = range_end,
1929 int ret;
1931 trace_nfs_writeback_page_enter(inode);
1933 for (;;) {
1934 wait_on_page_writeback(page);
1935 if (clear_page_dirty_for_io(page)) {
1936 ret = nfs_writepage_locked(page, &wbc);
1937 if (ret < 0)
1938 goto out_error;
1939 continue;
1941 ret = 0;
1942 if (!PagePrivate(page))
1943 break;
1944 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1945 if (ret < 0)
1946 goto out_error;
1948 out_error:
1949 trace_nfs_writeback_page_exit(inode, ret);
1950 return ret;
1953 #ifdef CONFIG_MIGRATION
1954 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1955 struct page *page, enum migrate_mode mode)
1958 * If PagePrivate is set, then the page is currently associated with
1959 * an in-progress read or write request. Don't try to migrate it.
1961 * FIXME: we could do this in principle, but we'll need a way to ensure
1962 * that we can safely release the inode reference while holding
1963 * the page lock.
1965 if (PagePrivate(page))
1966 return -EBUSY;
1968 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1969 return -EBUSY;
1971 return migrate_page(mapping, newpage, page, mode);
1973 #endif
1975 int __init nfs_init_writepagecache(void)
1977 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1978 sizeof(struct nfs_pgio_header),
1979 0, SLAB_HWCACHE_ALIGN,
1980 NULL);
1981 if (nfs_wdata_cachep == NULL)
1982 return -ENOMEM;
1984 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1985 nfs_wdata_cachep);
1986 if (nfs_wdata_mempool == NULL)
1987 goto out_destroy_write_cache;
1989 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1990 sizeof(struct nfs_commit_data),
1991 0, SLAB_HWCACHE_ALIGN,
1992 NULL);
1993 if (nfs_cdata_cachep == NULL)
1994 goto out_destroy_write_mempool;
1996 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1997 nfs_cdata_cachep);
1998 if (nfs_commit_mempool == NULL)
1999 goto out_destroy_commit_cache;
2002 * NFS congestion size, scale with available memory.
2004 * 64MB: 8192k
2005 * 128MB: 11585k
2006 * 256MB: 16384k
2007 * 512MB: 23170k
2008 * 1GB: 32768k
2009 * 2GB: 46340k
2010 * 4GB: 65536k
2011 * 8GB: 92681k
2012 * 16GB: 131072k
2014 * This allows larger machines to have larger/more transfers.
2015 * Limit the default to 256M
2017 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2018 if (nfs_congestion_kb > 256*1024)
2019 nfs_congestion_kb = 256*1024;
2021 return 0;
2023 out_destroy_commit_cache:
2024 kmem_cache_destroy(nfs_cdata_cachep);
2025 out_destroy_write_mempool:
2026 mempool_destroy(nfs_wdata_mempool);
2027 out_destroy_write_cache:
2028 kmem_cache_destroy(nfs_wdata_cachep);
2029 return -ENOMEM;
2032 void nfs_destroy_writepagecache(void)
2034 mempool_destroy(nfs_commit_mempool);
2035 kmem_cache_destroy(nfs_cdata_cachep);
2036 mempool_destroy(nfs_wdata_mempool);
2037 kmem_cache_destroy(nfs_wdata_cachep);
2040 static const struct nfs_rw_ops nfs_rw_write_ops = {
2041 .rw_mode = FMODE_WRITE,
2042 .rw_alloc_header = nfs_writehdr_alloc,
2043 .rw_free_header = nfs_writehdr_free,
2044 .rw_done = nfs_writeback_done,
2045 .rw_result = nfs_writeback_result,
2046 .rw_initiate = nfs_initiate_write,