4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
9 #include <linux/types.h>
10 #include <linux/slab.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
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>
24 #include <linux/freezer.h>
25 #include <linux/wait.h>
26 #include <linux/iversion.h>
28 #include <linux/uaccess.h>
29 #include <linux/sched/mm.h>
31 #include "delegation.h"
40 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
42 #define MIN_POOL_WRITE (32)
43 #define MIN_POOL_COMMIT (4)
45 struct nfs_io_completion {
46 void (*complete)(void *data);
52 * Local function declarations
54 static void nfs_redirty_request(struct nfs_page *req);
55 static const struct rpc_call_ops nfs_commit_ops;
56 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
57 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
58 static const struct nfs_rw_ops nfs_rw_write_ops;
59 static void nfs_clear_request_commit(struct nfs_page *req);
60 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
62 static struct nfs_page *
63 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
66 static struct kmem_cache *nfs_wdata_cachep;
67 static mempool_t *nfs_wdata_mempool;
68 static struct kmem_cache *nfs_cdata_cachep;
69 static mempool_t *nfs_commit_mempool;
71 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
73 struct nfs_commit_data *p;
76 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
78 /* It is OK to do some reclaim, not no safe to wait
79 * for anything to be returned to the pool.
80 * mempool_alloc() cannot handle that particular combination,
81 * so we need two separate attempts.
83 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
85 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
86 __GFP_NOWARN | __GFP_NORETRY);
91 memset(p, 0, sizeof(*p));
92 INIT_LIST_HEAD(&p->pages);
95 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
97 void nfs_commit_free(struct nfs_commit_data *p)
99 mempool_free(p, nfs_commit_mempool);
101 EXPORT_SYMBOL_GPL(nfs_commit_free);
103 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
105 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
107 memset(p, 0, sizeof(*p));
108 p->rw_mode = FMODE_WRITE;
112 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
114 mempool_free(hdr, nfs_wdata_mempool);
117 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
119 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
122 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
123 void (*complete)(void *), void *data)
125 ioc->complete = complete;
127 kref_init(&ioc->refcount);
130 static void nfs_io_completion_release(struct kref *kref)
132 struct nfs_io_completion *ioc = container_of(kref,
133 struct nfs_io_completion, refcount);
134 ioc->complete(ioc->data);
138 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
141 kref_get(&ioc->refcount);
144 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
147 kref_put(&ioc->refcount, nfs_io_completion_release);
150 static struct nfs_page *
151 nfs_page_private_request(struct page *page)
153 if (!PagePrivate(page))
155 return (struct nfs_page *)page_private(page);
159 * nfs_page_find_head_request_locked - find head request associated with @page
161 * must be called while holding the inode lock.
163 * returns matching head request with reference held, or NULL if not found.
165 static struct nfs_page *
166 nfs_page_find_private_request(struct page *page)
168 struct address_space *mapping = page_file_mapping(page);
169 struct nfs_page *req;
171 if (!PagePrivate(page))
173 spin_lock(&mapping->private_lock);
174 req = nfs_page_private_request(page);
176 WARN_ON_ONCE(req->wb_head != req);
177 kref_get(&req->wb_kref);
179 spin_unlock(&mapping->private_lock);
183 static struct nfs_page *
184 nfs_page_find_swap_request(struct page *page)
186 struct inode *inode = page_file_mapping(page)->host;
187 struct nfs_inode *nfsi = NFS_I(inode);
188 struct nfs_page *req = NULL;
189 if (!PageSwapCache(page))
191 mutex_lock(&nfsi->commit_mutex);
192 if (PageSwapCache(page)) {
193 req = nfs_page_search_commits_for_head_request_locked(nfsi,
196 WARN_ON_ONCE(req->wb_head != req);
197 kref_get(&req->wb_kref);
200 mutex_unlock(&nfsi->commit_mutex);
205 * nfs_page_find_head_request - find head request associated with @page
207 * returns matching head request with reference held, or NULL if not found.
209 static struct nfs_page *nfs_page_find_head_request(struct page *page)
211 struct nfs_page *req;
213 req = nfs_page_find_private_request(page);
215 req = nfs_page_find_swap_request(page);
219 /* Adjust the file length if we're writing beyond the end */
220 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
222 struct inode *inode = page_file_mapping(page)->host;
226 spin_lock(&inode->i_lock);
227 i_size = i_size_read(inode);
228 end_index = (i_size - 1) >> PAGE_SHIFT;
229 if (i_size > 0 && page_index(page) < end_index)
231 end = page_file_offset(page) + ((loff_t)offset+count);
234 i_size_write(inode, end);
235 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
236 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
238 spin_unlock(&inode->i_lock);
241 /* A writeback failed: mark the page as bad, and invalidate the page cache */
242 static void nfs_set_pageerror(struct address_space *mapping)
244 nfs_zap_mapping(mapping->host, mapping);
248 * nfs_page_group_search_locked
249 * @head - head request of page group
250 * @page_offset - offset into page
252 * Search page group with head @head to find a request that contains the
253 * page offset @page_offset.
255 * Returns a pointer to the first matching nfs request, or NULL if no
258 * Must be called with the page group lock held
260 static struct nfs_page *
261 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
263 struct nfs_page *req;
267 if (page_offset >= req->wb_pgbase &&
268 page_offset < (req->wb_pgbase + req->wb_bytes))
271 req = req->wb_this_page;
272 } while (req != head);
278 * nfs_page_group_covers_page
279 * @head - head request of page group
281 * Return true if the page group with head @head covers the whole page,
282 * returns false otherwise
284 static bool nfs_page_group_covers_page(struct nfs_page *req)
286 struct nfs_page *tmp;
287 unsigned int pos = 0;
288 unsigned int len = nfs_page_length(req->wb_page);
290 nfs_page_group_lock(req);
293 tmp = nfs_page_group_search_locked(req->wb_head, pos);
296 pos = tmp->wb_pgbase + tmp->wb_bytes;
299 nfs_page_group_unlock(req);
303 /* We can set the PG_uptodate flag if we see that a write request
304 * covers the full page.
306 static void nfs_mark_uptodate(struct nfs_page *req)
308 if (PageUptodate(req->wb_page))
310 if (!nfs_page_group_covers_page(req))
312 SetPageUptodate(req->wb_page);
315 static int wb_priority(struct writeback_control *wbc)
319 if (wbc->sync_mode == WB_SYNC_ALL)
320 ret = FLUSH_COND_STABLE;
325 * NFS congestion control
328 int nfs_congestion_kb;
330 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
331 #define NFS_CONGESTION_OFF_THRESH \
332 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
334 static void nfs_set_page_writeback(struct page *page)
336 struct inode *inode = page_file_mapping(page)->host;
337 struct nfs_server *nfss = NFS_SERVER(inode);
338 int ret = test_set_page_writeback(page);
340 WARN_ON_ONCE(ret != 0);
342 if (atomic_long_inc_return(&nfss->writeback) >
343 NFS_CONGESTION_ON_THRESH)
344 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
347 static void nfs_end_page_writeback(struct nfs_page *req)
349 struct inode *inode = page_file_mapping(req->wb_page)->host;
350 struct nfs_server *nfss = NFS_SERVER(inode);
353 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
354 nfs_unlock_request(req);
358 end_page_writeback(req->wb_page);
359 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
360 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
364 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
366 * this is a helper function for nfs_lock_and_join_requests
368 * @inode - inode associated with request page group, must be holding inode lock
369 * @head - head request of page group, must be holding head lock
370 * @req - request that couldn't lock and needs to wait on the req bit lock
372 * NOTE: this must be called holding page_group bit lock
373 * which will be released before returning.
375 * returns 0 on success, < 0 on error.
378 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
379 struct nfs_page *req)
381 struct nfs_page *tmp;
383 /* relinquish all the locks successfully grabbed this run */
384 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
385 if (!kref_read(&tmp->wb_kref))
387 nfs_unlock_and_release_request(tmp);
392 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
394 * @destroy_list - request list (using wb_this_page) terminated by @old_head
395 * @old_head - the old head of the list
397 * All subrequests must be locked and removed from all lists, so at this point
398 * they are only "active" in this function, and possibly in nfs_wait_on_request
399 * with a reference held by some other context.
402 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
403 struct nfs_page *old_head,
406 while (destroy_list) {
407 struct nfs_page *subreq = destroy_list;
409 destroy_list = (subreq->wb_this_page == old_head) ?
410 NULL : subreq->wb_this_page;
412 WARN_ON_ONCE(old_head != subreq->wb_head);
414 /* make sure old group is not used */
415 subreq->wb_this_page = subreq;
417 clear_bit(PG_REMOVE, &subreq->wb_flags);
419 /* Note: races with nfs_page_group_destroy() */
420 if (!kref_read(&subreq->wb_kref)) {
421 /* Check if we raced with nfs_page_group_destroy() */
422 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
423 nfs_free_request(subreq);
427 subreq->wb_head = subreq;
429 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
430 nfs_release_request(subreq);
431 atomic_long_dec(&NFS_I(inode)->nrequests);
434 /* subreq is now totally disconnected from page group or any
435 * write / commit lists. last chance to wake any waiters */
436 nfs_unlock_and_release_request(subreq);
441 * nfs_lock_and_join_requests - join all subreqs to the head req and return
442 * a locked reference, cancelling any pending
443 * operations for this page.
445 * @page - the page used to lookup the "page group" of nfs_page structures
447 * This function joins all sub requests to the head request by first
448 * locking all requests in the group, cancelling any pending operations
449 * and finally updating the head request to cover the whole range covered by
450 * the (former) group. All subrequests are removed from any write or commit
451 * lists, unlinked from the group and destroyed.
453 * Returns a locked, referenced pointer to the head request - which after
454 * this call is guaranteed to be the only request associated with the page.
455 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
456 * error was encountered.
458 static struct nfs_page *
459 nfs_lock_and_join_requests(struct page *page)
461 struct inode *inode = page_file_mapping(page)->host;
462 struct nfs_page *head, *subreq;
463 struct nfs_page *destroy_list = NULL;
464 unsigned int total_bytes;
469 * A reference is taken only on the head request which acts as a
470 * reference to the whole page group - the group will not be destroyed
471 * until the head reference is released.
473 head = nfs_page_find_head_request(page);
477 /* lock the page head first in order to avoid an ABBA inefficiency */
478 if (!nfs_lock_request(head)) {
479 ret = nfs_wait_on_request(head);
480 nfs_release_request(head);
486 /* Ensure that nobody removed the request before we locked it */
487 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
488 nfs_unlock_and_release_request(head);
492 ret = nfs_page_group_lock(head);
494 goto release_request;
496 /* lock each request in the page group */
497 total_bytes = head->wb_bytes;
498 for (subreq = head->wb_this_page; subreq != head;
499 subreq = subreq->wb_this_page) {
501 if (!kref_get_unless_zero(&subreq->wb_kref)) {
502 if (subreq->wb_offset == head->wb_offset + total_bytes)
503 total_bytes += subreq->wb_bytes;
507 while (!nfs_lock_request(subreq)) {
509 * Unlock page to allow nfs_page_group_sync_on_bit()
512 nfs_page_group_unlock(head);
513 ret = nfs_wait_on_request(subreq);
515 ret = nfs_page_group_lock(head);
517 nfs_unroll_locks(inode, head, subreq);
518 nfs_release_request(subreq);
519 goto release_request;
523 * Subrequests are always contiguous, non overlapping
524 * and in order - but may be repeated (mirrored writes).
526 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
527 /* keep track of how many bytes this group covers */
528 total_bytes += subreq->wb_bytes;
529 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
530 ((subreq->wb_offset + subreq->wb_bytes) >
531 (head->wb_offset + total_bytes)))) {
532 nfs_page_group_unlock(head);
533 nfs_unroll_locks(inode, head, subreq);
534 nfs_unlock_and_release_request(subreq);
536 goto release_request;
540 /* Now that all requests are locked, make sure they aren't on any list.
541 * Commit list removal accounting is done after locks are dropped */
544 nfs_clear_request_commit(subreq);
545 subreq = subreq->wb_this_page;
546 } while (subreq != head);
548 /* unlink subrequests from head, destroy them later */
549 if (head->wb_this_page != head) {
550 /* destroy list will be terminated by head */
551 destroy_list = head->wb_this_page;
552 head->wb_this_page = head;
554 /* change head request to cover whole range that
555 * the former page group covered */
556 head->wb_bytes = total_bytes;
559 /* Postpone destruction of this request */
560 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
561 set_bit(PG_INODE_REF, &head->wb_flags);
562 kref_get(&head->wb_kref);
563 atomic_long_inc(&NFS_I(inode)->nrequests);
566 nfs_page_group_unlock(head);
568 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
570 /* Did we lose a race with nfs_inode_remove_request()? */
571 if (!(PagePrivate(page) || PageSwapCache(page))) {
572 nfs_unlock_and_release_request(head);
576 /* still holds ref on head from nfs_page_find_head_request
577 * and still has lock on head from lock loop */
581 nfs_unlock_and_release_request(head);
585 static void nfs_write_error_remove_page(struct nfs_page *req)
587 SetPageError(req->wb_page);
588 nfs_end_page_writeback(req);
589 nfs_release_request(req);
593 nfs_error_is_fatal_on_server(int err)
601 return nfs_error_is_fatal(err);
605 * Find an associated nfs write request, and prepare to flush it out
606 * May return an error if the user signalled nfs_wait_on_request().
608 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
611 struct nfs_page *req;
614 req = nfs_lock_and_join_requests(page);
621 nfs_set_page_writeback(page);
622 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
624 ret = req->wb_context->error;
625 /* If there is a fatal error that covers this write, just exit */
626 if (nfs_error_is_fatal_on_server(ret))
630 if (!nfs_pageio_add_request(pgio, req)) {
631 ret = pgio->pg_error;
633 * Remove the problematic req upon fatal errors on the server
635 if (nfs_error_is_fatal(ret)) {
636 nfs_context_set_write_error(req->wb_context, ret);
637 if (nfs_error_is_fatal_on_server(ret))
641 nfs_redirty_request(req);
643 nfs_add_stats(page_file_mapping(page)->host,
644 NFSIOS_WRITEPAGES, 1);
648 nfs_write_error_remove_page(req);
652 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
653 struct nfs_pageio_descriptor *pgio)
657 nfs_pageio_cond_complete(pgio, page_index(page));
658 ret = nfs_page_async_flush(pgio, page);
659 if (ret == -EAGAIN) {
660 redirty_page_for_writepage(wbc, page);
667 * Write an mmapped page to the server.
669 static int nfs_writepage_locked(struct page *page,
670 struct writeback_control *wbc)
672 struct nfs_pageio_descriptor pgio;
673 struct inode *inode = page_file_mapping(page)->host;
676 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
677 nfs_pageio_init_write(&pgio, inode, 0,
678 false, &nfs_async_write_completion_ops);
679 err = nfs_do_writepage(page, wbc, &pgio);
680 nfs_pageio_complete(&pgio);
683 if (pgio.pg_error < 0)
684 return pgio.pg_error;
688 int nfs_writepage(struct page *page, struct writeback_control *wbc)
692 ret = nfs_writepage_locked(page, wbc);
697 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
701 ret = nfs_do_writepage(page, wbc, data);
706 static void nfs_io_completion_commit(void *inode)
708 nfs_commit_inode(inode, 0);
711 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
713 struct inode *inode = mapping->host;
714 struct nfs_pageio_descriptor pgio;
715 struct nfs_io_completion *ioc;
716 unsigned int pflags = memalloc_nofs_save();
719 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
721 ioc = nfs_io_completion_alloc(GFP_NOFS);
723 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
725 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
726 &nfs_async_write_completion_ops);
727 pgio.pg_io_completion = ioc;
728 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
729 nfs_pageio_complete(&pgio);
730 nfs_io_completion_put(ioc);
732 memalloc_nofs_restore(pflags);
745 * Insert a write request into an inode
747 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
749 struct address_space *mapping = page_file_mapping(req->wb_page);
750 struct nfs_inode *nfsi = NFS_I(inode);
752 WARN_ON_ONCE(req->wb_this_page != req);
754 /* Lock the request! */
755 nfs_lock_request(req);
758 * Swap-space should not get truncated. Hence no need to plug the race
759 * with invalidate/truncate.
761 spin_lock(&mapping->private_lock);
762 if (!nfs_have_writebacks(inode) &&
763 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
764 inode_inc_iversion_raw(inode);
765 if (likely(!PageSwapCache(req->wb_page))) {
766 set_bit(PG_MAPPED, &req->wb_flags);
767 SetPagePrivate(req->wb_page);
768 set_page_private(req->wb_page, (unsigned long)req);
770 spin_unlock(&mapping->private_lock);
771 atomic_long_inc(&nfsi->nrequests);
772 /* this a head request for a page group - mark it as having an
773 * extra reference so sub groups can follow suit.
774 * This flag also informs pgio layer when to bump nrequests when
775 * adding subrequests. */
776 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
777 kref_get(&req->wb_kref);
781 * Remove a write request from an inode
783 static void nfs_inode_remove_request(struct nfs_page *req)
785 struct address_space *mapping = page_file_mapping(req->wb_page);
786 struct inode *inode = mapping->host;
787 struct nfs_inode *nfsi = NFS_I(inode);
788 struct nfs_page *head;
790 atomic_long_dec(&nfsi->nrequests);
791 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
794 spin_lock(&mapping->private_lock);
795 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
796 set_page_private(head->wb_page, 0);
797 ClearPagePrivate(head->wb_page);
798 clear_bit(PG_MAPPED, &head->wb_flags);
800 spin_unlock(&mapping->private_lock);
803 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
804 nfs_release_request(req);
808 nfs_mark_request_dirty(struct nfs_page *req)
811 __set_page_dirty_nobuffers(req->wb_page);
815 * nfs_page_search_commits_for_head_request_locked
817 * Search through commit lists on @inode for the head request for @page.
818 * Must be called while holding the inode (which is cinfo) lock.
820 * Returns the head request if found, or NULL if not found.
822 static struct nfs_page *
823 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
826 struct nfs_page *freq, *t;
827 struct nfs_commit_info cinfo;
828 struct inode *inode = &nfsi->vfs_inode;
830 nfs_init_cinfo_from_inode(&cinfo, inode);
832 /* search through pnfs commit lists */
833 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
835 return freq->wb_head;
837 /* Linearly search the commit list for the correct request */
838 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
839 if (freq->wb_page == page)
840 return freq->wb_head;
847 * nfs_request_add_commit_list_locked - add request to a commit list
848 * @req: pointer to a struct nfs_page
849 * @dst: commit list head
850 * @cinfo: holds list lock and accounting info
852 * This sets the PG_CLEAN bit, updates the cinfo count of
853 * number of outstanding requests requiring a commit as well as
856 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
860 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
861 struct nfs_commit_info *cinfo)
863 set_bit(PG_CLEAN, &req->wb_flags);
864 nfs_list_add_request(req, dst);
865 atomic_long_inc(&cinfo->mds->ncommit);
867 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
870 * nfs_request_add_commit_list - add request to a commit list
871 * @req: pointer to a struct nfs_page
872 * @cinfo: holds list lock and accounting info
874 * This sets the PG_CLEAN bit, updates the cinfo count of
875 * number of outstanding requests requiring a commit as well as
878 * The caller must _not_ hold the cinfo->lock, but must be
879 * holding the nfs_page lock.
882 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
884 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
885 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
886 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
888 nfs_mark_page_unstable(req->wb_page, cinfo);
890 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
893 * nfs_request_remove_commit_list - Remove request from a commit list
894 * @req: pointer to a nfs_page
895 * @cinfo: holds list lock and accounting info
897 * This clears the PG_CLEAN bit, and updates the cinfo's count of
898 * number of outstanding requests requiring a commit
899 * It does not update the MM page stats.
901 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
904 nfs_request_remove_commit_list(struct nfs_page *req,
905 struct nfs_commit_info *cinfo)
907 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
909 nfs_list_remove_request(req);
910 atomic_long_dec(&cinfo->mds->ncommit);
912 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
914 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
917 cinfo->inode = inode;
918 cinfo->mds = &NFS_I(inode)->commit_info;
919 cinfo->ds = pnfs_get_ds_info(inode);
921 cinfo->completion_ops = &nfs_commit_completion_ops;
924 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
926 struct nfs_direct_req *dreq)
929 nfs_init_cinfo_from_dreq(cinfo, dreq);
931 nfs_init_cinfo_from_inode(cinfo, inode);
933 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
936 * Add a request to the inode's commit list.
939 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
940 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
942 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
944 nfs_request_add_commit_list(req, cinfo);
948 nfs_clear_page_commit(struct page *page)
950 dec_node_page_state(page, NR_UNSTABLE_NFS);
951 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
955 /* Called holding the request lock on @req */
957 nfs_clear_request_commit(struct nfs_page *req)
959 if (test_bit(PG_CLEAN, &req->wb_flags)) {
960 struct inode *inode = d_inode(req->wb_context->dentry);
961 struct nfs_commit_info cinfo;
963 nfs_init_cinfo_from_inode(&cinfo, inode);
964 mutex_lock(&NFS_I(inode)->commit_mutex);
965 if (!pnfs_clear_request_commit(req, &cinfo)) {
966 nfs_request_remove_commit_list(req, &cinfo);
968 mutex_unlock(&NFS_I(inode)->commit_mutex);
969 nfs_clear_page_commit(req->wb_page);
973 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
975 if (hdr->verf.committed == NFS_DATA_SYNC)
976 return hdr->lseg == NULL;
977 return hdr->verf.committed != NFS_FILE_SYNC;
980 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
982 nfs_io_completion_get(hdr->io_completion);
985 static void nfs_write_completion(struct nfs_pgio_header *hdr)
987 struct nfs_commit_info cinfo;
988 unsigned long bytes = 0;
990 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
992 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
993 while (!list_empty(&hdr->pages)) {
994 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
996 bytes += req->wb_bytes;
997 nfs_list_remove_request(req);
998 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
999 (hdr->good_bytes < bytes)) {
1000 nfs_set_pageerror(page_file_mapping(req->wb_page));
1001 nfs_context_set_write_error(req->wb_context, hdr->error);
1004 if (nfs_write_need_commit(hdr)) {
1005 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1006 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1007 hdr->pgio_mirror_idx);
1011 nfs_inode_remove_request(req);
1013 nfs_end_page_writeback(req);
1014 nfs_release_request(req);
1017 nfs_io_completion_put(hdr->io_completion);
1022 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1024 return atomic_long_read(&cinfo->mds->ncommit);
1027 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1029 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1030 struct nfs_commit_info *cinfo, int max)
1032 struct nfs_page *req, *tmp;
1036 list_for_each_entry_safe(req, tmp, src, wb_list) {
1037 kref_get(&req->wb_kref);
1038 if (!nfs_lock_request(req)) {
1041 /* Prevent deadlock with nfs_lock_and_join_requests */
1042 if (!list_empty(dst)) {
1043 nfs_release_request(req);
1046 /* Ensure we make progress to prevent livelock */
1047 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1048 status = nfs_wait_on_request(req);
1049 nfs_release_request(req);
1050 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1055 nfs_request_remove_commit_list(req, cinfo);
1056 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1057 nfs_list_add_request(req, dst);
1059 if ((ret == max) && !cinfo->dreq)
1065 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1068 * nfs_scan_commit - Scan an inode for commit requests
1069 * @inode: NFS inode to scan
1070 * @dst: mds destination list
1071 * @cinfo: mds and ds lists of reqs ready to commit
1073 * Moves requests from the inode's 'commit' request list.
1074 * The requests are *not* checked to ensure that they form a contiguous set.
1077 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1078 struct nfs_commit_info *cinfo)
1082 if (!atomic_long_read(&cinfo->mds->ncommit))
1084 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1085 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1086 const int max = INT_MAX;
1088 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1090 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1092 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1097 * Search for an existing write request, and attempt to update
1098 * it to reflect a new dirty region on a given page.
1100 * If the attempt fails, then the existing request is flushed out
1103 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1105 unsigned int offset,
1108 struct nfs_page *req;
1113 end = offset + bytes;
1115 req = nfs_lock_and_join_requests(page);
1116 if (IS_ERR_OR_NULL(req))
1119 rqend = req->wb_offset + req->wb_bytes;
1121 * Tell the caller to flush out the request if
1122 * the offsets are non-contiguous.
1123 * Note: nfs_flush_incompatible() will already
1124 * have flushed out requests having wrong owners.
1126 if (offset > rqend || end < req->wb_offset)
1129 /* Okay, the request matches. Update the region */
1130 if (offset < req->wb_offset) {
1131 req->wb_offset = offset;
1132 req->wb_pgbase = offset;
1135 req->wb_bytes = end - req->wb_offset;
1137 req->wb_bytes = rqend - req->wb_offset;
1141 * Note: we mark the request dirty here because
1142 * nfs_lock_and_join_requests() cannot preserve
1143 * commit flags, so we have to replay the write.
1145 nfs_mark_request_dirty(req);
1146 nfs_unlock_and_release_request(req);
1147 error = nfs_wb_page(inode, page);
1148 return (error < 0) ? ERR_PTR(error) : NULL;
1152 * Try to update an existing write request, or create one if there is none.
1154 * Note: Should always be called with the Page Lock held to prevent races
1155 * if we have to add a new request. Also assumes that the caller has
1156 * already called nfs_flush_incompatible() if necessary.
1158 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1159 struct page *page, unsigned int offset, unsigned int bytes)
1161 struct inode *inode = page_file_mapping(page)->host;
1162 struct nfs_page *req;
1164 req = nfs_try_to_update_request(inode, page, offset, bytes);
1167 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1170 nfs_inode_add_request(inode, req);
1175 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1176 unsigned int offset, unsigned int count)
1178 struct nfs_page *req;
1180 req = nfs_setup_write_request(ctx, page, offset, count);
1182 return PTR_ERR(req);
1183 /* Update file length */
1184 nfs_grow_file(page, offset, count);
1185 nfs_mark_uptodate(req);
1186 nfs_mark_request_dirty(req);
1187 nfs_unlock_and_release_request(req);
1191 int nfs_flush_incompatible(struct file *file, struct page *page)
1193 struct nfs_open_context *ctx = nfs_file_open_context(file);
1194 struct nfs_lock_context *l_ctx;
1195 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1196 struct nfs_page *req;
1197 int do_flush, status;
1199 * Look for a request corresponding to this page. If there
1200 * is one, and it belongs to another file, we flush it out
1201 * before we try to copy anything into the page. Do this
1202 * due to the lack of an ACCESS-type call in NFSv2.
1203 * Also do the same if we find a request from an existing
1207 req = nfs_page_find_head_request(page);
1210 l_ctx = req->wb_lock_context;
1211 do_flush = req->wb_page != page ||
1212 !nfs_match_open_context(req->wb_context, ctx);
1213 if (l_ctx && flctx &&
1214 !(list_empty_careful(&flctx->flc_posix) &&
1215 list_empty_careful(&flctx->flc_flock))) {
1216 do_flush |= l_ctx->lockowner != current->files;
1218 nfs_release_request(req);
1221 status = nfs_wb_page(page_file_mapping(page)->host, page);
1222 } while (status == 0);
1227 * Avoid buffered writes when a open context credential's key would
1230 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1232 * Return 0 and set a credential flag which triggers the inode to flush
1233 * and performs NFS_FILE_SYNC writes if the key will expired within
1234 * RPC_KEY_EXPIRE_TIMEO.
1237 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1239 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1241 if (nfs_ctx_key_to_expire(ctx, inode) &&
1243 /* Already expired! */
1249 * Test if the open context credential key is marked to expire soon.
1251 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1253 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1254 struct rpc_cred *cred = ctx->ll_cred;
1255 struct auth_cred acred = {
1259 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1261 ctx->ll_cred = NULL;
1265 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1266 if (!cred || IS_ERR(cred))
1268 ctx->ll_cred = cred;
1269 return !!(cred->cr_ops->crkey_timeout &&
1270 cred->cr_ops->crkey_timeout(cred));
1274 * If the page cache is marked as unsafe or invalid, then we can't rely on
1275 * the PageUptodate() flag. In this case, we will need to turn off
1276 * write optimisations that depend on the page contents being correct.
1278 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1280 struct nfs_inode *nfsi = NFS_I(inode);
1282 if (nfs_have_delegated_attributes(inode))
1284 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1287 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1290 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1292 return PageUptodate(page) != 0;
1296 is_whole_file_wrlock(struct file_lock *fl)
1298 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1299 fl->fl_type == F_WRLCK;
1302 /* If we know the page is up to date, and we're not using byte range locks (or
1303 * if we have the whole file locked for writing), it may be more efficient to
1304 * extend the write to cover the entire page in order to avoid fragmentation
1307 * If the file is opened for synchronous writes then we can just skip the rest
1310 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1313 struct file_lock_context *flctx = inode->i_flctx;
1314 struct file_lock *fl;
1316 if (file->f_flags & O_DSYNC)
1318 if (!nfs_write_pageuptodate(page, inode))
1320 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1322 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1323 list_empty_careful(&flctx->flc_posix)))
1326 /* Check to see if there are whole file write locks */
1328 spin_lock(&flctx->flc_lock);
1329 if (!list_empty(&flctx->flc_posix)) {
1330 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1332 if (is_whole_file_wrlock(fl))
1334 } else if (!list_empty(&flctx->flc_flock)) {
1335 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1337 if (fl->fl_type == F_WRLCK)
1340 spin_unlock(&flctx->flc_lock);
1345 * Update and possibly write a cached page of an NFS file.
1347 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1348 * things with a page scheduled for an RPC call (e.g. invalidate it).
1350 int nfs_updatepage(struct file *file, struct page *page,
1351 unsigned int offset, unsigned int count)
1353 struct nfs_open_context *ctx = nfs_file_open_context(file);
1354 struct address_space *mapping = page_file_mapping(page);
1355 struct inode *inode = mapping->host;
1358 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1360 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1361 file, count, (long long)(page_file_offset(page) + offset));
1366 if (nfs_can_extend_write(file, page, inode)) {
1367 count = max(count + offset, nfs_page_length(page));
1371 status = nfs_writepage_setup(ctx, page, offset, count);
1373 nfs_set_pageerror(mapping);
1375 __set_page_dirty_nobuffers(page);
1377 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1378 status, (long long)i_size_read(inode));
1382 static int flush_task_priority(int how)
1384 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1386 return RPC_PRIORITY_HIGH;
1388 return RPC_PRIORITY_LOW;
1390 return RPC_PRIORITY_NORMAL;
1393 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1394 struct rpc_message *msg,
1395 const struct nfs_rpc_ops *rpc_ops,
1396 struct rpc_task_setup *task_setup_data, int how)
1398 int priority = flush_task_priority(how);
1400 task_setup_data->priority = priority;
1401 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1402 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1406 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1407 * call this on each, which will prepare them to be retried on next
1408 * writeback using standard nfs.
1410 static void nfs_redirty_request(struct nfs_page *req)
1412 nfs_mark_request_dirty(req);
1413 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1414 nfs_end_page_writeback(req);
1415 nfs_release_request(req);
1418 static void nfs_async_write_error(struct list_head *head, int error)
1420 struct nfs_page *req;
1422 while (!list_empty(head)) {
1423 req = nfs_list_entry(head->next);
1424 nfs_list_remove_request(req);
1425 if (nfs_error_is_fatal(error)) {
1426 nfs_context_set_write_error(req->wb_context, error);
1427 if (nfs_error_is_fatal_on_server(error)) {
1428 nfs_write_error_remove_page(req);
1432 nfs_redirty_request(req);
1436 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1438 nfs_async_write_error(&hdr->pages, 0);
1439 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1440 hdr->args.offset + hdr->args.count - 1);
1443 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1444 .init_hdr = nfs_async_write_init,
1445 .error_cleanup = nfs_async_write_error,
1446 .completion = nfs_write_completion,
1447 .reschedule_io = nfs_async_write_reschedule_io,
1450 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1451 struct inode *inode, int ioflags, bool force_mds,
1452 const struct nfs_pgio_completion_ops *compl_ops)
1454 struct nfs_server *server = NFS_SERVER(inode);
1455 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1457 #ifdef CONFIG_NFS_V4_1
1458 if (server->pnfs_curr_ld && !force_mds)
1459 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1461 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1462 server->wsize, ioflags);
1464 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1466 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1468 struct nfs_pgio_mirror *mirror;
1470 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1471 pgio->pg_ops->pg_cleanup(pgio);
1473 pgio->pg_ops = &nfs_pgio_rw_ops;
1475 nfs_pageio_stop_mirroring(pgio);
1477 mirror = &pgio->pg_mirrors[0];
1478 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1480 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1483 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1485 struct nfs_commit_data *data = calldata;
1487 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1491 * Special version of should_remove_suid() that ignores capabilities.
1493 static int nfs_should_remove_suid(const struct inode *inode)
1495 umode_t mode = inode->i_mode;
1498 /* suid always must be killed */
1499 if (unlikely(mode & S_ISUID))
1500 kill = ATTR_KILL_SUID;
1503 * sgid without any exec bits is just a mandatory locking mark; leave
1504 * it alone. If some exec bits are set, it's a real sgid; kill it.
1506 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1507 kill |= ATTR_KILL_SGID;
1509 if (unlikely(kill && S_ISREG(mode)))
1515 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1516 struct nfs_fattr *fattr)
1518 struct nfs_pgio_args *argp = &hdr->args;
1519 struct nfs_pgio_res *resp = &hdr->res;
1520 u64 size = argp->offset + resp->count;
1522 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1524 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1525 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1528 if (size != fattr->size)
1530 /* Set attribute barrier */
1531 nfs_fattr_set_barrier(fattr);
1532 /* ...and update size */
1533 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1536 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1538 struct nfs_fattr *fattr = &hdr->fattr;
1539 struct inode *inode = hdr->inode;
1541 spin_lock(&inode->i_lock);
1542 nfs_writeback_check_extend(hdr, fattr);
1543 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1544 spin_unlock(&inode->i_lock);
1546 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1549 * This function is called when the WRITE call is complete.
1551 static int nfs_writeback_done(struct rpc_task *task,
1552 struct nfs_pgio_header *hdr,
1553 struct inode *inode)
1558 * ->write_done will attempt to use post-op attributes to detect
1559 * conflicting writes by other clients. A strict interpretation
1560 * of close-to-open would allow us to continue caching even if
1561 * another writer had changed the file, but some applications
1562 * depend on tighter cache coherency when writing.
1564 status = NFS_PROTO(inode)->write_done(task, hdr);
1568 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1569 trace_nfs_writeback_done(inode, task->tk_status,
1570 hdr->args.offset, hdr->res.verf);
1572 if (hdr->res.verf->committed < hdr->args.stable &&
1573 task->tk_status >= 0) {
1574 /* We tried a write call, but the server did not
1575 * commit data to stable storage even though we
1577 * Note: There is a known bug in Tru64 < 5.0 in which
1578 * the server reports NFS_DATA_SYNC, but performs
1579 * NFS_FILE_SYNC. We therefore implement this checking
1580 * as a dprintk() in order to avoid filling syslog.
1582 static unsigned long complain;
1584 /* Note this will print the MDS for a DS write */
1585 if (time_before(complain, jiffies)) {
1586 dprintk("NFS: faulty NFS server %s:"
1587 " (committed = %d) != (stable = %d)\n",
1588 NFS_SERVER(inode)->nfs_client->cl_hostname,
1589 hdr->res.verf->committed, hdr->args.stable);
1590 complain = jiffies + 300 * HZ;
1594 /* Deal with the suid/sgid bit corner case */
1595 if (nfs_should_remove_suid(inode)) {
1596 spin_lock(&inode->i_lock);
1597 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1598 spin_unlock(&inode->i_lock);
1604 * This function is called when the WRITE call is complete.
1606 static void nfs_writeback_result(struct rpc_task *task,
1607 struct nfs_pgio_header *hdr)
1609 struct nfs_pgio_args *argp = &hdr->args;
1610 struct nfs_pgio_res *resp = &hdr->res;
1612 if (resp->count < argp->count) {
1613 static unsigned long complain;
1615 /* This a short write! */
1616 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1618 /* Has the server at least made some progress? */
1619 if (resp->count == 0) {
1620 if (time_before(complain, jiffies)) {
1622 "NFS: Server wrote zero bytes, expected %u.\n",
1624 complain = jiffies + 300 * HZ;
1626 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1627 task->tk_status = -EIO;
1631 /* For non rpc-based layout drivers, retry-through-MDS */
1632 if (!task->tk_ops) {
1633 hdr->pnfs_error = -EAGAIN;
1637 /* Was this an NFSv2 write or an NFSv3 stable write? */
1638 if (resp->verf->committed != NFS_UNSTABLE) {
1639 /* Resend from where the server left off */
1640 hdr->mds_offset += resp->count;
1641 argp->offset += resp->count;
1642 argp->pgbase += resp->count;
1643 argp->count -= resp->count;
1645 /* Resend as a stable write in order to avoid
1646 * headaches in the case of a server crash.
1648 argp->stable = NFS_FILE_SYNC;
1650 rpc_restart_call_prepare(task);
1654 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1656 return wait_var_event_killable(&cinfo->rpcs_out,
1657 !atomic_read(&cinfo->rpcs_out));
1660 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1662 atomic_inc(&cinfo->rpcs_out);
1665 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1667 if (atomic_dec_and_test(&cinfo->rpcs_out))
1668 wake_up_var(&cinfo->rpcs_out);
1671 void nfs_commitdata_release(struct nfs_commit_data *data)
1673 put_nfs_open_context(data->context);
1674 nfs_commit_free(data);
1676 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1678 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1679 const struct nfs_rpc_ops *nfs_ops,
1680 const struct rpc_call_ops *call_ops,
1683 struct rpc_task *task;
1684 int priority = flush_task_priority(how);
1685 struct rpc_message msg = {
1686 .rpc_argp = &data->args,
1687 .rpc_resp = &data->res,
1688 .rpc_cred = data->cred,
1690 struct rpc_task_setup task_setup_data = {
1691 .task = &data->task,
1693 .rpc_message = &msg,
1694 .callback_ops = call_ops,
1695 .callback_data = data,
1696 .workqueue = nfsiod_workqueue,
1697 .flags = RPC_TASK_ASYNC | flags,
1698 .priority = priority,
1700 /* Set up the initial task struct. */
1701 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1702 trace_nfs_initiate_commit(data);
1704 dprintk("NFS: initiated commit call\n");
1706 task = rpc_run_task(&task_setup_data);
1708 return PTR_ERR(task);
1709 if (how & FLUSH_SYNC)
1710 rpc_wait_for_completion_task(task);
1714 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1716 static loff_t nfs_get_lwb(struct list_head *head)
1719 struct nfs_page *req;
1721 list_for_each_entry(req, head, wb_list)
1722 if (lwb < (req_offset(req) + req->wb_bytes))
1723 lwb = req_offset(req) + req->wb_bytes;
1729 * Set up the argument/result storage required for the RPC call.
1731 void nfs_init_commit(struct nfs_commit_data *data,
1732 struct list_head *head,
1733 struct pnfs_layout_segment *lseg,
1734 struct nfs_commit_info *cinfo)
1736 struct nfs_page *first = nfs_list_entry(head->next);
1737 struct inode *inode = d_inode(first->wb_context->dentry);
1739 /* Set up the RPC argument and reply structs
1740 * NB: take care not to mess about with data->commit et al. */
1742 list_splice_init(head, &data->pages);
1744 data->inode = inode;
1745 data->cred = first->wb_context->cred;
1746 data->lseg = lseg; /* reference transferred */
1747 /* only set lwb for pnfs commit */
1749 data->lwb = nfs_get_lwb(&data->pages);
1750 data->mds_ops = &nfs_commit_ops;
1751 data->completion_ops = cinfo->completion_ops;
1752 data->dreq = cinfo->dreq;
1754 data->args.fh = NFS_FH(data->inode);
1755 /* Note: we always request a commit of the entire inode */
1756 data->args.offset = 0;
1757 data->args.count = 0;
1758 data->context = get_nfs_open_context(first->wb_context);
1759 data->res.fattr = &data->fattr;
1760 data->res.verf = &data->verf;
1761 nfs_fattr_init(&data->fattr);
1763 EXPORT_SYMBOL_GPL(nfs_init_commit);
1765 void nfs_retry_commit(struct list_head *page_list,
1766 struct pnfs_layout_segment *lseg,
1767 struct nfs_commit_info *cinfo,
1770 struct nfs_page *req;
1772 while (!list_empty(page_list)) {
1773 req = nfs_list_entry(page_list->next);
1774 nfs_list_remove_request(req);
1775 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1777 nfs_clear_page_commit(req->wb_page);
1778 nfs_unlock_and_release_request(req);
1781 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1784 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1785 struct nfs_page *req)
1787 __set_page_dirty_nobuffers(req->wb_page);
1791 * Commit dirty pages
1794 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1795 struct nfs_commit_info *cinfo)
1797 struct nfs_commit_data *data;
1799 /* another commit raced with us */
1800 if (list_empty(head))
1803 data = nfs_commitdata_alloc(true);
1805 /* Set up the argument struct */
1806 nfs_init_commit(data, head, NULL, cinfo);
1807 atomic_inc(&cinfo->mds->rpcs_out);
1808 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1809 data->mds_ops, how, 0);
1813 * COMMIT call returned
1815 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1817 struct nfs_commit_data *data = calldata;
1819 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1820 task->tk_pid, task->tk_status);
1822 /* Call the NFS version-specific code */
1823 NFS_PROTO(data->inode)->commit_done(task, data);
1824 trace_nfs_commit_done(data);
1827 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1829 struct nfs_page *req;
1830 int status = data->task.tk_status;
1831 struct nfs_commit_info cinfo;
1832 struct nfs_server *nfss;
1834 while (!list_empty(&data->pages)) {
1835 req = nfs_list_entry(data->pages.next);
1836 nfs_list_remove_request(req);
1838 nfs_clear_page_commit(req->wb_page);
1840 dprintk("NFS: commit (%s/%llu %d@%lld)",
1841 req->wb_context->dentry->d_sb->s_id,
1842 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1844 (long long)req_offset(req));
1846 nfs_context_set_write_error(req->wb_context, status);
1848 nfs_inode_remove_request(req);
1849 dprintk_cont(", error = %d\n", status);
1853 /* Okay, COMMIT succeeded, apparently. Check the verifier
1854 * returned by the server against all stored verfs. */
1855 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1856 /* We have a match */
1858 nfs_inode_remove_request(req);
1859 dprintk_cont(" OK\n");
1862 /* We have a mismatch. Write the page again */
1863 dprintk_cont(" mismatch\n");
1864 nfs_mark_request_dirty(req);
1865 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1867 nfs_unlock_and_release_request(req);
1868 /* Latency breaker */
1871 nfss = NFS_SERVER(data->inode);
1872 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1873 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1875 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1876 nfs_commit_end(cinfo.mds);
1879 static void nfs_commit_release(void *calldata)
1881 struct nfs_commit_data *data = calldata;
1883 data->completion_ops->completion(data);
1884 nfs_commitdata_release(calldata);
1887 static const struct rpc_call_ops nfs_commit_ops = {
1888 .rpc_call_prepare = nfs_commit_prepare,
1889 .rpc_call_done = nfs_commit_done,
1890 .rpc_release = nfs_commit_release,
1893 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1894 .completion = nfs_commit_release_pages,
1895 .resched_write = nfs_commit_resched_write,
1898 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1899 int how, struct nfs_commit_info *cinfo)
1903 status = pnfs_commit_list(inode, head, how, cinfo);
1904 if (status == PNFS_NOT_ATTEMPTED)
1905 status = nfs_commit_list(inode, head, how, cinfo);
1909 static int __nfs_commit_inode(struct inode *inode, int how,
1910 struct writeback_control *wbc)
1913 struct nfs_commit_info cinfo;
1914 int may_wait = how & FLUSH_SYNC;
1917 nfs_init_cinfo_from_inode(&cinfo, inode);
1918 nfs_commit_begin(cinfo.mds);
1920 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1923 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1927 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1928 if (nscan < wbc->nr_to_write)
1929 wbc->nr_to_write -= nscan;
1931 wbc->nr_to_write = 0;
1933 if (nscan < INT_MAX)
1937 nfs_commit_end(cinfo.mds);
1938 if (ret || !may_wait)
1940 return wait_on_commit(cinfo.mds);
1943 int nfs_commit_inode(struct inode *inode, int how)
1945 return __nfs_commit_inode(inode, how, NULL);
1947 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1949 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1951 struct nfs_inode *nfsi = NFS_I(inode);
1952 int flags = FLUSH_SYNC;
1955 if (wbc->sync_mode == WB_SYNC_NONE) {
1956 /* no commits means nothing needs to be done */
1957 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1958 goto check_requests_outstanding;
1960 /* Don't commit yet if this is a non-blocking flush and there
1961 * are a lot of outstanding writes for this mapping.
1963 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1964 goto out_mark_dirty;
1966 /* don't wait for the COMMIT response */
1970 ret = __nfs_commit_inode(inode, flags, wbc);
1972 if (flags & FLUSH_SYNC)
1974 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1975 goto out_mark_dirty;
1977 check_requests_outstanding:
1978 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1981 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1984 EXPORT_SYMBOL_GPL(nfs_write_inode);
1987 * Wrapper for filemap_write_and_wait_range()
1989 * Needed for pNFS in order to ensure data becomes visible to the
1992 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1993 loff_t lstart, loff_t lend)
1997 ret = filemap_write_and_wait_range(mapping, lstart, lend);
1999 ret = pnfs_sync_inode(mapping->host, true);
2002 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2005 * flush the inode to disk.
2007 int nfs_wb_all(struct inode *inode)
2011 trace_nfs_writeback_inode_enter(inode);
2013 ret = filemap_write_and_wait(inode->i_mapping);
2016 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2019 pnfs_sync_inode(inode, true);
2023 trace_nfs_writeback_inode_exit(inode, ret);
2026 EXPORT_SYMBOL_GPL(nfs_wb_all);
2028 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2030 struct nfs_page *req;
2033 wait_on_page_writeback(page);
2035 /* blocking call to cancel all requests and join to a single (head)
2037 req = nfs_lock_and_join_requests(page);
2042 /* all requests from this page have been cancelled by
2043 * nfs_lock_and_join_requests, so just remove the head
2044 * request from the inode / page_private pointer and
2046 nfs_inode_remove_request(req);
2047 nfs_unlock_and_release_request(req);
2054 * Write back all requests on one page - we do this before reading it.
2056 int nfs_wb_page(struct inode *inode, struct page *page)
2058 loff_t range_start = page_file_offset(page);
2059 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2060 struct writeback_control wbc = {
2061 .sync_mode = WB_SYNC_ALL,
2063 .range_start = range_start,
2064 .range_end = range_end,
2068 trace_nfs_writeback_page_enter(inode);
2071 wait_on_page_writeback(page);
2072 if (clear_page_dirty_for_io(page)) {
2073 ret = nfs_writepage_locked(page, &wbc);
2079 if (!PagePrivate(page))
2081 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2086 trace_nfs_writeback_page_exit(inode, ret);
2090 #ifdef CONFIG_MIGRATION
2091 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2092 struct page *page, enum migrate_mode mode)
2095 * If PagePrivate is set, then the page is currently associated with
2096 * an in-progress read or write request. Don't try to migrate it.
2098 * FIXME: we could do this in principle, but we'll need a way to ensure
2099 * that we can safely release the inode reference while holding
2102 if (PagePrivate(page))
2105 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2108 return migrate_page(mapping, newpage, page, mode);
2112 int __init nfs_init_writepagecache(void)
2114 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2115 sizeof(struct nfs_pgio_header),
2116 0, SLAB_HWCACHE_ALIGN,
2118 if (nfs_wdata_cachep == NULL)
2121 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2123 if (nfs_wdata_mempool == NULL)
2124 goto out_destroy_write_cache;
2126 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2127 sizeof(struct nfs_commit_data),
2128 0, SLAB_HWCACHE_ALIGN,
2130 if (nfs_cdata_cachep == NULL)
2131 goto out_destroy_write_mempool;
2133 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2135 if (nfs_commit_mempool == NULL)
2136 goto out_destroy_commit_cache;
2139 * NFS congestion size, scale with available memory.
2151 * This allows larger machines to have larger/more transfers.
2152 * Limit the default to 256M
2154 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2155 if (nfs_congestion_kb > 256*1024)
2156 nfs_congestion_kb = 256*1024;
2160 out_destroy_commit_cache:
2161 kmem_cache_destroy(nfs_cdata_cachep);
2162 out_destroy_write_mempool:
2163 mempool_destroy(nfs_wdata_mempool);
2164 out_destroy_write_cache:
2165 kmem_cache_destroy(nfs_wdata_cachep);
2169 void nfs_destroy_writepagecache(void)
2171 mempool_destroy(nfs_commit_mempool);
2172 kmem_cache_destroy(nfs_cdata_cachep);
2173 mempool_destroy(nfs_wdata_mempool);
2174 kmem_cache_destroy(nfs_wdata_cachep);
2177 static const struct nfs_rw_ops nfs_rw_write_ops = {
2178 .rw_alloc_header = nfs_writehdr_alloc,
2179 .rw_free_header = nfs_writehdr_free,
2180 .rw_done = nfs_writeback_done,
2181 .rw_result = nfs_writeback_result,
2182 .rw_initiate = nfs_initiate_write,