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>
30 #include "delegation.h"
39 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
41 #define MIN_POOL_WRITE (32)
42 #define MIN_POOL_COMMIT (4)
44 struct nfs_io_completion {
45 void (*complete)(void *data);
51 * Local function declarations
53 static void nfs_redirty_request(struct nfs_page *req);
54 static const struct rpc_call_ops nfs_commit_ops;
55 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
56 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
57 static const struct nfs_rw_ops nfs_rw_write_ops;
58 static void nfs_clear_request_commit(struct nfs_page *req);
59 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
61 static struct nfs_page *
62 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
65 static struct kmem_cache *nfs_wdata_cachep;
66 static mempool_t *nfs_wdata_mempool;
67 static struct kmem_cache *nfs_cdata_cachep;
68 static mempool_t *nfs_commit_mempool;
70 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
72 struct nfs_commit_data *p;
75 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
77 /* It is OK to do some reclaim, not no safe to wait
78 * for anything to be returned to the pool.
79 * mempool_alloc() cannot handle that particular combination,
80 * so we need two separate attempts.
82 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
84 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
85 __GFP_NOWARN | __GFP_NORETRY);
90 memset(p, 0, sizeof(*p));
91 INIT_LIST_HEAD(&p->pages);
94 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
96 void nfs_commit_free(struct nfs_commit_data *p)
98 mempool_free(p, nfs_commit_mempool);
100 EXPORT_SYMBOL_GPL(nfs_commit_free);
102 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
104 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
106 memset(p, 0, sizeof(*p));
107 p->rw_mode = FMODE_WRITE;
111 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
113 mempool_free(hdr, nfs_wdata_mempool);
116 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
118 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
121 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
122 void (*complete)(void *), void *data)
124 ioc->complete = complete;
126 kref_init(&ioc->refcount);
129 static void nfs_io_completion_release(struct kref *kref)
131 struct nfs_io_completion *ioc = container_of(kref,
132 struct nfs_io_completion, refcount);
133 ioc->complete(ioc->data);
137 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
140 kref_get(&ioc->refcount);
143 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
146 kref_put(&ioc->refcount, nfs_io_completion_release);
149 static struct nfs_page *
150 nfs_page_private_request(struct page *page)
152 if (!PagePrivate(page))
154 return (struct nfs_page *)page_private(page);
158 * nfs_page_find_head_request_locked - find head request associated with @page
160 * must be called while holding the inode lock.
162 * returns matching head request with reference held, or NULL if not found.
164 static struct nfs_page *
165 nfs_page_find_private_request(struct page *page)
167 struct address_space *mapping = page_file_mapping(page);
168 struct nfs_page *req;
170 if (!PagePrivate(page))
172 spin_lock(&mapping->private_lock);
173 req = nfs_page_private_request(page);
175 WARN_ON_ONCE(req->wb_head != req);
176 kref_get(&req->wb_kref);
178 spin_unlock(&mapping->private_lock);
182 static struct nfs_page *
183 nfs_page_find_swap_request(struct page *page)
185 struct inode *inode = page_file_mapping(page)->host;
186 struct nfs_inode *nfsi = NFS_I(inode);
187 struct nfs_page *req = NULL;
188 if (!PageSwapCache(page))
190 mutex_lock(&nfsi->commit_mutex);
191 if (PageSwapCache(page)) {
192 req = nfs_page_search_commits_for_head_request_locked(nfsi,
195 WARN_ON_ONCE(req->wb_head != req);
196 kref_get(&req->wb_kref);
199 mutex_unlock(&nfsi->commit_mutex);
204 * nfs_page_find_head_request - find head request associated with @page
206 * returns matching head request with reference held, or NULL if not found.
208 static struct nfs_page *nfs_page_find_head_request(struct page *page)
210 struct nfs_page *req;
212 req = nfs_page_find_private_request(page);
214 req = nfs_page_find_swap_request(page);
218 /* Adjust the file length if we're writing beyond the end */
219 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
221 struct inode *inode = page_file_mapping(page)->host;
225 spin_lock(&inode->i_lock);
226 i_size = i_size_read(inode);
227 end_index = (i_size - 1) >> PAGE_SHIFT;
228 if (i_size > 0 && page_index(page) < end_index)
230 end = page_file_offset(page) + ((loff_t)offset+count);
233 i_size_write(inode, end);
234 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
235 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
237 spin_unlock(&inode->i_lock);
240 /* A writeback failed: mark the page as bad, and invalidate the page cache */
241 static void nfs_set_pageerror(struct address_space *mapping)
243 struct inode *inode = mapping->host;
245 nfs_zap_mapping(mapping->host, mapping);
246 /* Force file size revalidation */
247 spin_lock(&inode->i_lock);
248 NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
249 NFS_INO_REVAL_PAGECACHE |
250 NFS_INO_INVALID_SIZE;
251 spin_unlock(&inode->i_lock);
255 * nfs_page_group_search_locked
256 * @head - head request of page group
257 * @page_offset - offset into page
259 * Search page group with head @head to find a request that contains the
260 * page offset @page_offset.
262 * Returns a pointer to the first matching nfs request, or NULL if no
265 * Must be called with the page group lock held
267 static struct nfs_page *
268 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
270 struct nfs_page *req;
274 if (page_offset >= req->wb_pgbase &&
275 page_offset < (req->wb_pgbase + req->wb_bytes))
278 req = req->wb_this_page;
279 } while (req != head);
285 * nfs_page_group_covers_page
286 * @head - head request of page group
288 * Return true if the page group with head @head covers the whole page,
289 * returns false otherwise
291 static bool nfs_page_group_covers_page(struct nfs_page *req)
293 struct nfs_page *tmp;
294 unsigned int pos = 0;
295 unsigned int len = nfs_page_length(req->wb_page);
297 nfs_page_group_lock(req);
300 tmp = nfs_page_group_search_locked(req->wb_head, pos);
303 pos = tmp->wb_pgbase + tmp->wb_bytes;
306 nfs_page_group_unlock(req);
310 /* We can set the PG_uptodate flag if we see that a write request
311 * covers the full page.
313 static void nfs_mark_uptodate(struct nfs_page *req)
315 if (PageUptodate(req->wb_page))
317 if (!nfs_page_group_covers_page(req))
319 SetPageUptodate(req->wb_page);
322 static int wb_priority(struct writeback_control *wbc)
326 if (wbc->sync_mode == WB_SYNC_ALL)
327 ret = FLUSH_COND_STABLE;
332 * NFS congestion control
335 int nfs_congestion_kb;
337 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
338 #define NFS_CONGESTION_OFF_THRESH \
339 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
341 static void nfs_set_page_writeback(struct page *page)
343 struct inode *inode = page_file_mapping(page)->host;
344 struct nfs_server *nfss = NFS_SERVER(inode);
345 int ret = test_set_page_writeback(page);
347 WARN_ON_ONCE(ret != 0);
349 if (atomic_long_inc_return(&nfss->writeback) >
350 NFS_CONGESTION_ON_THRESH)
351 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
354 static void nfs_end_page_writeback(struct nfs_page *req)
356 struct inode *inode = page_file_mapping(req->wb_page)->host;
357 struct nfs_server *nfss = NFS_SERVER(inode);
360 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
361 nfs_unlock_request(req);
365 end_page_writeback(req->wb_page);
366 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
367 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
371 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
373 * this is a helper function for nfs_lock_and_join_requests
375 * @inode - inode associated with request page group, must be holding inode lock
376 * @head - head request of page group, must be holding head lock
377 * @req - request that couldn't lock and needs to wait on the req bit lock
379 * NOTE: this must be called holding page_group bit lock
380 * which will be released before returning.
382 * returns 0 on success, < 0 on error.
385 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
386 struct nfs_page *req)
388 struct nfs_page *tmp;
390 /* relinquish all the locks successfully grabbed this run */
391 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
392 if (!kref_read(&tmp->wb_kref))
394 nfs_unlock_and_release_request(tmp);
399 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
401 * @destroy_list - request list (using wb_this_page) terminated by @old_head
402 * @old_head - the old head of the list
404 * All subrequests must be locked and removed from all lists, so at this point
405 * they are only "active" in this function, and possibly in nfs_wait_on_request
406 * with a reference held by some other context.
409 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
410 struct nfs_page *old_head,
413 while (destroy_list) {
414 struct nfs_page *subreq = destroy_list;
416 destroy_list = (subreq->wb_this_page == old_head) ?
417 NULL : subreq->wb_this_page;
419 /* Note: lock subreq in order to change subreq->wb_head */
420 nfs_page_set_headlock(subreq);
421 WARN_ON_ONCE(old_head != subreq->wb_head);
423 /* make sure old group is not used */
424 subreq->wb_this_page = subreq;
425 subreq->wb_head = subreq;
427 clear_bit(PG_REMOVE, &subreq->wb_flags);
429 /* Note: races with nfs_page_group_destroy() */
430 if (!kref_read(&subreq->wb_kref)) {
431 /* Check if we raced with nfs_page_group_destroy() */
432 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
433 nfs_page_clear_headlock(subreq);
434 nfs_free_request(subreq);
436 nfs_page_clear_headlock(subreq);
439 nfs_page_clear_headlock(subreq);
441 nfs_release_request(old_head);
443 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
444 nfs_release_request(subreq);
445 atomic_long_dec(&NFS_I(inode)->nrequests);
448 /* subreq is now totally disconnected from page group or any
449 * write / commit lists. last chance to wake any waiters */
450 nfs_unlock_and_release_request(subreq);
455 * nfs_lock_and_join_requests - join all subreqs to the head req and return
456 * a locked reference, cancelling any pending
457 * operations for this page.
459 * @page - the page used to lookup the "page group" of nfs_page structures
461 * This function joins all sub requests to the head request by first
462 * locking all requests in the group, cancelling any pending operations
463 * and finally updating the head request to cover the whole range covered by
464 * the (former) group. All subrequests are removed from any write or commit
465 * lists, unlinked from the group and destroyed.
467 * Returns a locked, referenced pointer to the head request - which after
468 * this call is guaranteed to be the only request associated with the page.
469 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
470 * error was encountered.
472 static struct nfs_page *
473 nfs_lock_and_join_requests(struct page *page)
475 struct inode *inode = page_file_mapping(page)->host;
476 struct nfs_page *head, *subreq;
477 struct nfs_page *destroy_list = NULL;
478 unsigned int total_bytes;
483 * A reference is taken only on the head request which acts as a
484 * reference to the whole page group - the group will not be destroyed
485 * until the head reference is released.
487 head = nfs_page_find_head_request(page);
491 /* lock the page head first in order to avoid an ABBA inefficiency */
492 if (!nfs_lock_request(head)) {
493 ret = nfs_wait_on_request(head);
494 nfs_release_request(head);
500 /* Ensure that nobody removed the request before we locked it */
501 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
502 nfs_unlock_and_release_request(head);
506 ret = nfs_page_group_lock(head);
508 goto release_request;
510 /* lock each request in the page group */
511 total_bytes = head->wb_bytes;
512 for (subreq = head->wb_this_page; subreq != head;
513 subreq = subreq->wb_this_page) {
515 if (!kref_get_unless_zero(&subreq->wb_kref)) {
516 if (subreq->wb_offset == head->wb_offset + total_bytes)
517 total_bytes += subreq->wb_bytes;
521 while (!nfs_lock_request(subreq)) {
523 * Unlock page to allow nfs_page_group_sync_on_bit()
526 nfs_page_group_unlock(head);
527 ret = nfs_wait_on_request(subreq);
529 ret = nfs_page_group_lock(head);
531 nfs_unroll_locks(inode, head, subreq);
532 nfs_release_request(subreq);
533 goto release_request;
537 * Subrequests are always contiguous, non overlapping
538 * and in order - but may be repeated (mirrored writes).
540 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
541 /* keep track of how many bytes this group covers */
542 total_bytes += subreq->wb_bytes;
543 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
544 ((subreq->wb_offset + subreq->wb_bytes) >
545 (head->wb_offset + total_bytes)))) {
546 nfs_page_group_unlock(head);
547 nfs_unroll_locks(inode, head, subreq);
548 nfs_unlock_and_release_request(subreq);
550 goto release_request;
554 /* Now that all requests are locked, make sure they aren't on any list.
555 * Commit list removal accounting is done after locks are dropped */
558 nfs_clear_request_commit(subreq);
559 subreq = subreq->wb_this_page;
560 } while (subreq != head);
562 /* unlink subrequests from head, destroy them later */
563 if (head->wb_this_page != head) {
564 /* destroy list will be terminated by head */
565 destroy_list = head->wb_this_page;
566 head->wb_this_page = head;
568 /* change head request to cover whole range that
569 * the former page group covered */
570 head->wb_bytes = total_bytes;
573 /* Postpone destruction of this request */
574 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
575 set_bit(PG_INODE_REF, &head->wb_flags);
576 kref_get(&head->wb_kref);
577 atomic_long_inc(&NFS_I(inode)->nrequests);
580 nfs_page_group_unlock(head);
582 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
584 /* Did we lose a race with nfs_inode_remove_request()? */
585 if (!(PagePrivate(page) || PageSwapCache(page))) {
586 nfs_unlock_and_release_request(head);
590 /* still holds ref on head from nfs_page_find_head_request
591 * and still has lock on head from lock loop */
595 nfs_unlock_and_release_request(head);
599 static void nfs_write_error_remove_page(struct nfs_page *req)
601 nfs_end_page_writeback(req);
602 generic_error_remove_page(page_file_mapping(req->wb_page),
604 nfs_release_request(req);
608 nfs_error_is_fatal_on_server(int err)
616 return nfs_error_is_fatal(err);
620 * Find an associated nfs write request, and prepare to flush it out
621 * May return an error if the user signalled nfs_wait_on_request().
623 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
626 struct nfs_page *req;
629 req = nfs_lock_and_join_requests(page);
636 nfs_set_page_writeback(page);
637 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
639 ret = req->wb_context->error;
640 /* If there is a fatal error that covers this write, just exit */
641 if (nfs_error_is_fatal_on_server(ret))
645 if (!nfs_pageio_add_request(pgio, req)) {
646 ret = pgio->pg_error;
648 * Remove the problematic req upon fatal errors on the server
650 if (nfs_error_is_fatal(ret)) {
651 nfs_context_set_write_error(req->wb_context, ret);
652 if (nfs_error_is_fatal_on_server(ret))
656 nfs_redirty_request(req);
658 nfs_add_stats(page_file_mapping(page)->host,
659 NFSIOS_WRITEPAGES, 1);
663 nfs_write_error_remove_page(req);
667 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
668 struct nfs_pageio_descriptor *pgio)
672 nfs_pageio_cond_complete(pgio, page_index(page));
673 ret = nfs_page_async_flush(pgio, page);
674 if (ret == -EAGAIN) {
675 redirty_page_for_writepage(wbc, page);
682 * Write an mmapped page to the server.
684 static int nfs_writepage_locked(struct page *page,
685 struct writeback_control *wbc)
687 struct nfs_pageio_descriptor pgio;
688 struct inode *inode = page_file_mapping(page)->host;
691 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
692 nfs_pageio_init_write(&pgio, inode, 0,
693 false, &nfs_async_write_completion_ops);
694 err = nfs_do_writepage(page, wbc, &pgio);
695 nfs_pageio_complete(&pgio);
698 if (pgio.pg_error < 0)
699 return pgio.pg_error;
703 int nfs_writepage(struct page *page, struct writeback_control *wbc)
707 ret = nfs_writepage_locked(page, wbc);
712 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
716 ret = nfs_do_writepage(page, wbc, data);
721 static void nfs_io_completion_commit(void *inode)
723 nfs_commit_inode(inode, 0);
726 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
728 struct inode *inode = mapping->host;
729 struct nfs_pageio_descriptor pgio;
730 struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
733 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
736 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
738 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
739 &nfs_async_write_completion_ops);
740 pgio.pg_io_completion = ioc;
741 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
742 nfs_pageio_complete(&pgio);
743 nfs_io_completion_put(ioc);
756 * Insert a write request into an inode
758 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
760 struct address_space *mapping = page_file_mapping(req->wb_page);
761 struct nfs_inode *nfsi = NFS_I(inode);
763 WARN_ON_ONCE(req->wb_this_page != req);
765 /* Lock the request! */
766 nfs_lock_request(req);
769 * Swap-space should not get truncated. Hence no need to plug the race
770 * with invalidate/truncate.
772 spin_lock(&mapping->private_lock);
773 if (!nfs_have_writebacks(inode) &&
774 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
775 inode_inc_iversion_raw(inode);
776 if (likely(!PageSwapCache(req->wb_page))) {
777 set_bit(PG_MAPPED, &req->wb_flags);
778 SetPagePrivate(req->wb_page);
779 set_page_private(req->wb_page, (unsigned long)req);
781 spin_unlock(&mapping->private_lock);
782 atomic_long_inc(&nfsi->nrequests);
783 /* this a head request for a page group - mark it as having an
784 * extra reference so sub groups can follow suit.
785 * This flag also informs pgio layer when to bump nrequests when
786 * adding subrequests. */
787 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
788 kref_get(&req->wb_kref);
792 * Remove a write request from an inode
794 static void nfs_inode_remove_request(struct nfs_page *req)
796 struct address_space *mapping = page_file_mapping(req->wb_page);
797 struct inode *inode = mapping->host;
798 struct nfs_inode *nfsi = NFS_I(inode);
799 struct nfs_page *head;
801 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
804 spin_lock(&mapping->private_lock);
805 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
806 set_page_private(head->wb_page, 0);
807 ClearPagePrivate(head->wb_page);
808 clear_bit(PG_MAPPED, &head->wb_flags);
810 spin_unlock(&mapping->private_lock);
813 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
814 nfs_release_request(req);
815 atomic_long_dec(&nfsi->nrequests);
820 nfs_mark_request_dirty(struct nfs_page *req)
823 __set_page_dirty_nobuffers(req->wb_page);
827 * nfs_page_search_commits_for_head_request_locked
829 * Search through commit lists on @inode for the head request for @page.
830 * Must be called while holding the inode (which is cinfo) lock.
832 * Returns the head request if found, or NULL if not found.
834 static struct nfs_page *
835 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
838 struct nfs_page *freq, *t;
839 struct nfs_commit_info cinfo;
840 struct inode *inode = &nfsi->vfs_inode;
842 nfs_init_cinfo_from_inode(&cinfo, inode);
844 /* search through pnfs commit lists */
845 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
847 return freq->wb_head;
849 /* Linearly search the commit list for the correct request */
850 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
851 if (freq->wb_page == page)
852 return freq->wb_head;
859 * nfs_request_add_commit_list_locked - add request to a commit list
860 * @req: pointer to a struct nfs_page
861 * @dst: commit list head
862 * @cinfo: holds list lock and accounting info
864 * This sets the PG_CLEAN bit, updates the cinfo count of
865 * number of outstanding requests requiring a commit as well as
868 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
872 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
873 struct nfs_commit_info *cinfo)
875 set_bit(PG_CLEAN, &req->wb_flags);
876 nfs_list_add_request(req, dst);
877 atomic_long_inc(&cinfo->mds->ncommit);
879 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
882 * nfs_request_add_commit_list - add request to a commit list
883 * @req: pointer to a struct nfs_page
884 * @dst: commit list head
885 * @cinfo: holds list lock and accounting info
887 * This sets the PG_CLEAN bit, updates the cinfo count of
888 * number of outstanding requests requiring a commit as well as
891 * The caller must _not_ hold the cinfo->lock, but must be
892 * holding the nfs_page lock.
895 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
897 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
898 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
899 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
901 nfs_mark_page_unstable(req->wb_page, cinfo);
903 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
906 * nfs_request_remove_commit_list - Remove request from a commit list
907 * @req: pointer to a nfs_page
908 * @cinfo: holds list lock and accounting info
910 * This clears the PG_CLEAN bit, and updates the cinfo's count of
911 * number of outstanding requests requiring a commit
912 * It does not update the MM page stats.
914 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
917 nfs_request_remove_commit_list(struct nfs_page *req,
918 struct nfs_commit_info *cinfo)
920 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
922 nfs_list_remove_request(req);
923 atomic_long_dec(&cinfo->mds->ncommit);
925 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
927 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
930 cinfo->inode = inode;
931 cinfo->mds = &NFS_I(inode)->commit_info;
932 cinfo->ds = pnfs_get_ds_info(inode);
934 cinfo->completion_ops = &nfs_commit_completion_ops;
937 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
939 struct nfs_direct_req *dreq)
942 nfs_init_cinfo_from_dreq(cinfo, dreq);
944 nfs_init_cinfo_from_inode(cinfo, inode);
946 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
949 * Add a request to the inode's commit list.
952 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
953 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
955 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
957 nfs_request_add_commit_list(req, cinfo);
961 nfs_clear_page_commit(struct page *page)
963 dec_node_page_state(page, NR_UNSTABLE_NFS);
964 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
968 /* Called holding the request lock on @req */
970 nfs_clear_request_commit(struct nfs_page *req)
972 if (test_bit(PG_CLEAN, &req->wb_flags)) {
973 struct inode *inode = d_inode(req->wb_context->dentry);
974 struct nfs_commit_info cinfo;
976 nfs_init_cinfo_from_inode(&cinfo, inode);
977 mutex_lock(&NFS_I(inode)->commit_mutex);
978 if (!pnfs_clear_request_commit(req, &cinfo)) {
979 nfs_request_remove_commit_list(req, &cinfo);
981 mutex_unlock(&NFS_I(inode)->commit_mutex);
982 nfs_clear_page_commit(req->wb_page);
986 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
988 if (hdr->verf.committed == NFS_DATA_SYNC)
989 return hdr->lseg == NULL;
990 return hdr->verf.committed != NFS_FILE_SYNC;
993 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
995 nfs_io_completion_get(hdr->io_completion);
998 static void nfs_write_completion(struct nfs_pgio_header *hdr)
1000 struct nfs_commit_info cinfo;
1001 unsigned long bytes = 0;
1003 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1005 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1006 while (!list_empty(&hdr->pages)) {
1007 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1009 bytes += req->wb_bytes;
1010 nfs_list_remove_request(req);
1011 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1012 (hdr->good_bytes < bytes)) {
1013 nfs_set_pageerror(page_file_mapping(req->wb_page));
1014 nfs_context_set_write_error(req->wb_context, hdr->error);
1017 if (nfs_write_need_commit(hdr)) {
1018 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1019 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1020 hdr->pgio_mirror_idx);
1024 nfs_inode_remove_request(req);
1026 nfs_end_page_writeback(req);
1027 nfs_release_request(req);
1030 nfs_io_completion_put(hdr->io_completion);
1035 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1037 return atomic_long_read(&cinfo->mds->ncommit);
1040 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1042 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1043 struct nfs_commit_info *cinfo, int max)
1045 struct nfs_page *req, *tmp;
1048 list_for_each_entry_safe(req, tmp, src, wb_list) {
1049 kref_get(&req->wb_kref);
1050 if (!nfs_lock_request(req)) {
1051 nfs_release_request(req);
1054 nfs_request_remove_commit_list(req, cinfo);
1055 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1056 nfs_list_add_request(req, dst);
1058 if ((ret == max) && !cinfo->dreq)
1064 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1067 * nfs_scan_commit - Scan an inode for commit requests
1068 * @inode: NFS inode to scan
1069 * @dst: mds destination list
1070 * @cinfo: mds and ds lists of reqs ready to commit
1072 * Moves requests from the inode's 'commit' request list.
1073 * The requests are *not* checked to ensure that they form a contiguous set.
1076 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1077 struct nfs_commit_info *cinfo)
1081 if (!atomic_long_read(&cinfo->mds->ncommit))
1083 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1084 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1085 const int max = INT_MAX;
1087 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1089 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1091 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1096 * Search for an existing write request, and attempt to update
1097 * it to reflect a new dirty region on a given page.
1099 * If the attempt fails, then the existing request is flushed out
1102 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1104 unsigned int offset,
1107 struct nfs_page *req;
1112 end = offset + bytes;
1114 req = nfs_lock_and_join_requests(page);
1115 if (IS_ERR_OR_NULL(req))
1118 rqend = req->wb_offset + req->wb_bytes;
1120 * Tell the caller to flush out the request if
1121 * the offsets are non-contiguous.
1122 * Note: nfs_flush_incompatible() will already
1123 * have flushed out requests having wrong owners.
1125 if (offset > rqend || end < req->wb_offset)
1128 /* Okay, the request matches. Update the region */
1129 if (offset < req->wb_offset) {
1130 req->wb_offset = offset;
1131 req->wb_pgbase = offset;
1134 req->wb_bytes = end - req->wb_offset;
1136 req->wb_bytes = rqend - req->wb_offset;
1140 * Note: we mark the request dirty here because
1141 * nfs_lock_and_join_requests() cannot preserve
1142 * commit flags, so we have to replay the write.
1144 nfs_mark_request_dirty(req);
1145 nfs_unlock_and_release_request(req);
1146 error = nfs_wb_page(inode, page);
1147 return (error < 0) ? ERR_PTR(error) : NULL;
1151 * Try to update an existing write request, or create one if there is none.
1153 * Note: Should always be called with the Page Lock held to prevent races
1154 * if we have to add a new request. Also assumes that the caller has
1155 * already called nfs_flush_incompatible() if necessary.
1157 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1158 struct page *page, unsigned int offset, unsigned int bytes)
1160 struct inode *inode = page_file_mapping(page)->host;
1161 struct nfs_page *req;
1163 req = nfs_try_to_update_request(inode, page, offset, bytes);
1166 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1169 nfs_inode_add_request(inode, req);
1174 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1175 unsigned int offset, unsigned int count)
1177 struct nfs_page *req;
1179 req = nfs_setup_write_request(ctx, page, offset, count);
1181 return PTR_ERR(req);
1182 /* Update file length */
1183 nfs_grow_file(page, offset, count);
1184 nfs_mark_uptodate(req);
1185 nfs_mark_request_dirty(req);
1186 nfs_unlock_and_release_request(req);
1190 int nfs_flush_incompatible(struct file *file, struct page *page)
1192 struct nfs_open_context *ctx = nfs_file_open_context(file);
1193 struct nfs_lock_context *l_ctx;
1194 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1195 struct nfs_page *req;
1196 int do_flush, status;
1198 * Look for a request corresponding to this page. If there
1199 * is one, and it belongs to another file, we flush it out
1200 * before we try to copy anything into the page. Do this
1201 * due to the lack of an ACCESS-type call in NFSv2.
1202 * Also do the same if we find a request from an existing
1206 req = nfs_page_find_head_request(page);
1209 l_ctx = req->wb_lock_context;
1210 do_flush = req->wb_page != page ||
1211 !nfs_match_open_context(req->wb_context, ctx);
1212 if (l_ctx && flctx &&
1213 !(list_empty_careful(&flctx->flc_posix) &&
1214 list_empty_careful(&flctx->flc_flock))) {
1215 do_flush |= l_ctx->lockowner != current->files;
1217 nfs_release_request(req);
1220 status = nfs_wb_page(page_file_mapping(page)->host, page);
1221 } while (status == 0);
1226 * Avoid buffered writes when a open context credential's key would
1229 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1231 * Return 0 and set a credential flag which triggers the inode to flush
1232 * and performs NFS_FILE_SYNC writes if the key will expired within
1233 * RPC_KEY_EXPIRE_TIMEO.
1236 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1238 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1239 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1241 return rpcauth_key_timeout_notify(auth, ctx->cred);
1245 * Test if the open context credential key is marked to expire soon.
1247 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1249 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1251 return rpcauth_cred_key_to_expire(auth, ctx->cred);
1255 * If the page cache is marked as unsafe or invalid, then we can't rely on
1256 * the PageUptodate() flag. In this case, we will need to turn off
1257 * write optimisations that depend on the page contents being correct.
1259 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1261 struct nfs_inode *nfsi = NFS_I(inode);
1263 if (nfs_have_delegated_attributes(inode))
1265 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1268 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1271 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1273 return PageUptodate(page) != 0;
1277 is_whole_file_wrlock(struct file_lock *fl)
1279 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1280 fl->fl_type == F_WRLCK;
1283 /* If we know the page is up to date, and we're not using byte range locks (or
1284 * if we have the whole file locked for writing), it may be more efficient to
1285 * extend the write to cover the entire page in order to avoid fragmentation
1288 * If the file is opened for synchronous writes then we can just skip the rest
1291 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1294 struct file_lock_context *flctx = inode->i_flctx;
1295 struct file_lock *fl;
1297 if (file->f_flags & O_DSYNC)
1299 if (!nfs_write_pageuptodate(page, inode))
1301 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1303 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1304 list_empty_careful(&flctx->flc_posix)))
1307 /* Check to see if there are whole file write locks */
1309 spin_lock(&flctx->flc_lock);
1310 if (!list_empty(&flctx->flc_posix)) {
1311 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1313 if (is_whole_file_wrlock(fl))
1315 } else if (!list_empty(&flctx->flc_flock)) {
1316 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1318 if (fl->fl_type == F_WRLCK)
1321 spin_unlock(&flctx->flc_lock);
1326 * Update and possibly write a cached page of an NFS file.
1328 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1329 * things with a page scheduled for an RPC call (e.g. invalidate it).
1331 int nfs_updatepage(struct file *file, struct page *page,
1332 unsigned int offset, unsigned int count)
1334 struct nfs_open_context *ctx = nfs_file_open_context(file);
1335 struct address_space *mapping = page_file_mapping(page);
1336 struct inode *inode = mapping->host;
1339 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1341 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1342 file, count, (long long)(page_file_offset(page) + offset));
1347 if (nfs_can_extend_write(file, page, inode)) {
1348 count = max(count + offset, nfs_page_length(page));
1352 status = nfs_writepage_setup(ctx, page, offset, count);
1354 nfs_set_pageerror(mapping);
1356 __set_page_dirty_nobuffers(page);
1358 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1359 status, (long long)i_size_read(inode));
1363 static int flush_task_priority(int how)
1365 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1367 return RPC_PRIORITY_HIGH;
1369 return RPC_PRIORITY_LOW;
1371 return RPC_PRIORITY_NORMAL;
1374 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1375 struct rpc_message *msg,
1376 const struct nfs_rpc_ops *rpc_ops,
1377 struct rpc_task_setup *task_setup_data, int how)
1379 int priority = flush_task_priority(how);
1381 task_setup_data->priority = priority;
1382 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1383 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1387 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1388 * call this on each, which will prepare them to be retried on next
1389 * writeback using standard nfs.
1391 static void nfs_redirty_request(struct nfs_page *req)
1393 nfs_mark_request_dirty(req);
1394 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1395 nfs_end_page_writeback(req);
1396 nfs_release_request(req);
1399 static void nfs_async_write_error(struct list_head *head, int error)
1401 struct nfs_page *req;
1403 while (!list_empty(head)) {
1404 req = nfs_list_entry(head->next);
1405 nfs_list_remove_request(req);
1406 if (nfs_error_is_fatal(error)) {
1407 nfs_context_set_write_error(req->wb_context, error);
1408 if (nfs_error_is_fatal_on_server(error)) {
1409 nfs_write_error_remove_page(req);
1413 nfs_redirty_request(req);
1417 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1419 nfs_async_write_error(&hdr->pages, 0);
1420 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1421 hdr->args.offset + hdr->args.count - 1);
1424 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1425 .init_hdr = nfs_async_write_init,
1426 .error_cleanup = nfs_async_write_error,
1427 .completion = nfs_write_completion,
1428 .reschedule_io = nfs_async_write_reschedule_io,
1431 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1432 struct inode *inode, int ioflags, bool force_mds,
1433 const struct nfs_pgio_completion_ops *compl_ops)
1435 struct nfs_server *server = NFS_SERVER(inode);
1436 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1438 #ifdef CONFIG_NFS_V4_1
1439 if (server->pnfs_curr_ld && !force_mds)
1440 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1442 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1443 server->wsize, ioflags);
1445 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1447 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1449 struct nfs_pgio_mirror *mirror;
1451 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1452 pgio->pg_ops->pg_cleanup(pgio);
1454 pgio->pg_ops = &nfs_pgio_rw_ops;
1456 nfs_pageio_stop_mirroring(pgio);
1458 mirror = &pgio->pg_mirrors[0];
1459 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1461 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1464 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1466 struct nfs_commit_data *data = calldata;
1468 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1472 * Special version of should_remove_suid() that ignores capabilities.
1474 static int nfs_should_remove_suid(const struct inode *inode)
1476 umode_t mode = inode->i_mode;
1479 /* suid always must be killed */
1480 if (unlikely(mode & S_ISUID))
1481 kill = ATTR_KILL_SUID;
1484 * sgid without any exec bits is just a mandatory locking mark; leave
1485 * it alone. If some exec bits are set, it's a real sgid; kill it.
1487 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1488 kill |= ATTR_KILL_SGID;
1490 if (unlikely(kill && S_ISREG(mode)))
1496 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1497 struct nfs_fattr *fattr)
1499 struct nfs_pgio_args *argp = &hdr->args;
1500 struct nfs_pgio_res *resp = &hdr->res;
1501 u64 size = argp->offset + resp->count;
1503 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1505 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1506 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1509 if (size != fattr->size)
1511 /* Set attribute barrier */
1512 nfs_fattr_set_barrier(fattr);
1513 /* ...and update size */
1514 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1517 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1519 struct nfs_fattr *fattr = &hdr->fattr;
1520 struct inode *inode = hdr->inode;
1522 spin_lock(&inode->i_lock);
1523 nfs_writeback_check_extend(hdr, fattr);
1524 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1525 spin_unlock(&inode->i_lock);
1527 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1530 * This function is called when the WRITE call is complete.
1532 static int nfs_writeback_done(struct rpc_task *task,
1533 struct nfs_pgio_header *hdr,
1534 struct inode *inode)
1539 * ->write_done will attempt to use post-op attributes to detect
1540 * conflicting writes by other clients. A strict interpretation
1541 * of close-to-open would allow us to continue caching even if
1542 * another writer had changed the file, but some applications
1543 * depend on tighter cache coherency when writing.
1545 status = NFS_PROTO(inode)->write_done(task, hdr);
1549 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1550 trace_nfs_writeback_done(inode, task->tk_status,
1551 hdr->args.offset, hdr->res.verf);
1553 if (hdr->res.verf->committed < hdr->args.stable &&
1554 task->tk_status >= 0) {
1555 /* We tried a write call, but the server did not
1556 * commit data to stable storage even though we
1558 * Note: There is a known bug in Tru64 < 5.0 in which
1559 * the server reports NFS_DATA_SYNC, but performs
1560 * NFS_FILE_SYNC. We therefore implement this checking
1561 * as a dprintk() in order to avoid filling syslog.
1563 static unsigned long complain;
1565 /* Note this will print the MDS for a DS write */
1566 if (time_before(complain, jiffies)) {
1567 dprintk("NFS: faulty NFS server %s:"
1568 " (committed = %d) != (stable = %d)\n",
1569 NFS_SERVER(inode)->nfs_client->cl_hostname,
1570 hdr->res.verf->committed, hdr->args.stable);
1571 complain = jiffies + 300 * HZ;
1575 /* Deal with the suid/sgid bit corner case */
1576 if (nfs_should_remove_suid(inode)) {
1577 spin_lock(&inode->i_lock);
1578 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1579 spin_unlock(&inode->i_lock);
1585 * This function is called when the WRITE call is complete.
1587 static void nfs_writeback_result(struct rpc_task *task,
1588 struct nfs_pgio_header *hdr)
1590 struct nfs_pgio_args *argp = &hdr->args;
1591 struct nfs_pgio_res *resp = &hdr->res;
1593 if (resp->count < argp->count) {
1594 static unsigned long complain;
1596 /* This a short write! */
1597 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1599 /* Has the server at least made some progress? */
1600 if (resp->count == 0) {
1601 if (time_before(complain, jiffies)) {
1603 "NFS: Server wrote zero bytes, expected %u.\n",
1605 complain = jiffies + 300 * HZ;
1607 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1608 task->tk_status = -EIO;
1612 /* For non rpc-based layout drivers, retry-through-MDS */
1613 if (!task->tk_ops) {
1614 hdr->pnfs_error = -EAGAIN;
1618 /* Was this an NFSv2 write or an NFSv3 stable write? */
1619 if (resp->verf->committed != NFS_UNSTABLE) {
1620 /* Resend from where the server left off */
1621 hdr->mds_offset += resp->count;
1622 argp->offset += resp->count;
1623 argp->pgbase += resp->count;
1624 argp->count -= resp->count;
1626 /* Resend as a stable write in order to avoid
1627 * headaches in the case of a server crash.
1629 argp->stable = NFS_FILE_SYNC;
1631 rpc_restart_call_prepare(task);
1635 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1637 return wait_var_event_killable(&cinfo->rpcs_out,
1638 !atomic_read(&cinfo->rpcs_out));
1641 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1643 atomic_inc(&cinfo->rpcs_out);
1646 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1648 if (atomic_dec_and_test(&cinfo->rpcs_out))
1649 wake_up_var(&cinfo->rpcs_out);
1652 void nfs_commitdata_release(struct nfs_commit_data *data)
1654 put_nfs_open_context(data->context);
1655 nfs_commit_free(data);
1657 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1659 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1660 const struct nfs_rpc_ops *nfs_ops,
1661 const struct rpc_call_ops *call_ops,
1664 struct rpc_task *task;
1665 int priority = flush_task_priority(how);
1666 struct rpc_message msg = {
1667 .rpc_argp = &data->args,
1668 .rpc_resp = &data->res,
1669 .rpc_cred = data->cred,
1671 struct rpc_task_setup task_setup_data = {
1672 .task = &data->task,
1674 .rpc_message = &msg,
1675 .callback_ops = call_ops,
1676 .callback_data = data,
1677 .workqueue = nfsiod_workqueue,
1678 .flags = RPC_TASK_ASYNC | flags,
1679 .priority = priority,
1681 /* Set up the initial task struct. */
1682 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1683 trace_nfs_initiate_commit(data);
1685 dprintk("NFS: initiated commit call\n");
1687 task = rpc_run_task(&task_setup_data);
1689 return PTR_ERR(task);
1690 if (how & FLUSH_SYNC)
1691 rpc_wait_for_completion_task(task);
1695 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1697 static loff_t nfs_get_lwb(struct list_head *head)
1700 struct nfs_page *req;
1702 list_for_each_entry(req, head, wb_list)
1703 if (lwb < (req_offset(req) + req->wb_bytes))
1704 lwb = req_offset(req) + req->wb_bytes;
1710 * Set up the argument/result storage required for the RPC call.
1712 void nfs_init_commit(struct nfs_commit_data *data,
1713 struct list_head *head,
1714 struct pnfs_layout_segment *lseg,
1715 struct nfs_commit_info *cinfo)
1717 struct nfs_page *first = nfs_list_entry(head->next);
1718 struct inode *inode = d_inode(first->wb_context->dentry);
1720 /* Set up the RPC argument and reply structs
1721 * NB: take care not to mess about with data->commit et al. */
1723 list_splice_init(head, &data->pages);
1725 data->inode = inode;
1726 data->cred = first->wb_context->cred;
1727 data->lseg = lseg; /* reference transferred */
1728 /* only set lwb for pnfs commit */
1730 data->lwb = nfs_get_lwb(&data->pages);
1731 data->mds_ops = &nfs_commit_ops;
1732 data->completion_ops = cinfo->completion_ops;
1733 data->dreq = cinfo->dreq;
1735 data->args.fh = NFS_FH(data->inode);
1736 /* Note: we always request a commit of the entire inode */
1737 data->args.offset = 0;
1738 data->args.count = 0;
1739 data->context = get_nfs_open_context(first->wb_context);
1740 data->res.fattr = &data->fattr;
1741 data->res.verf = &data->verf;
1742 nfs_fattr_init(&data->fattr);
1744 EXPORT_SYMBOL_GPL(nfs_init_commit);
1746 void nfs_retry_commit(struct list_head *page_list,
1747 struct pnfs_layout_segment *lseg,
1748 struct nfs_commit_info *cinfo,
1751 struct nfs_page *req;
1753 while (!list_empty(page_list)) {
1754 req = nfs_list_entry(page_list->next);
1755 nfs_list_remove_request(req);
1756 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1758 nfs_clear_page_commit(req->wb_page);
1759 nfs_unlock_and_release_request(req);
1762 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1765 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1766 struct nfs_page *req)
1768 __set_page_dirty_nobuffers(req->wb_page);
1772 * Commit dirty pages
1775 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1776 struct nfs_commit_info *cinfo)
1778 struct nfs_commit_data *data;
1780 /* another commit raced with us */
1781 if (list_empty(head))
1784 data = nfs_commitdata_alloc(true);
1786 /* Set up the argument struct */
1787 nfs_init_commit(data, head, NULL, cinfo);
1788 atomic_inc(&cinfo->mds->rpcs_out);
1789 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1790 data->mds_ops, how, 0);
1794 * COMMIT call returned
1796 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1798 struct nfs_commit_data *data = calldata;
1800 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1801 task->tk_pid, task->tk_status);
1803 /* Call the NFS version-specific code */
1804 NFS_PROTO(data->inode)->commit_done(task, data);
1805 trace_nfs_commit_done(data);
1808 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1810 const struct nfs_writeverf *verf = data->res.verf;
1811 struct nfs_page *req;
1812 int status = data->task.tk_status;
1813 struct nfs_commit_info cinfo;
1814 struct nfs_server *nfss;
1816 while (!list_empty(&data->pages)) {
1817 req = nfs_list_entry(data->pages.next);
1818 nfs_list_remove_request(req);
1820 nfs_clear_page_commit(req->wb_page);
1822 dprintk("NFS: commit (%s/%llu %d@%lld)",
1823 req->wb_context->dentry->d_sb->s_id,
1824 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1826 (long long)req_offset(req));
1828 nfs_context_set_write_error(req->wb_context, status);
1830 nfs_inode_remove_request(req);
1831 dprintk_cont(", error = %d\n", status);
1835 /* Okay, COMMIT succeeded, apparently. Check the verifier
1836 * returned by the server against all stored verfs. */
1837 if (verf->committed > NFS_UNSTABLE &&
1838 !nfs_write_verifier_cmp(&req->wb_verf, &verf->verifier)) {
1839 /* We have a match */
1841 nfs_inode_remove_request(req);
1842 dprintk_cont(" OK\n");
1845 /* We have a mismatch. Write the page again */
1846 dprintk_cont(" mismatch\n");
1847 nfs_mark_request_dirty(req);
1848 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1850 nfs_unlock_and_release_request(req);
1851 /* Latency breaker */
1854 nfss = NFS_SERVER(data->inode);
1855 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1856 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1858 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1859 nfs_commit_end(cinfo.mds);
1862 static void nfs_commit_release(void *calldata)
1864 struct nfs_commit_data *data = calldata;
1866 data->completion_ops->completion(data);
1867 nfs_commitdata_release(calldata);
1870 static const struct rpc_call_ops nfs_commit_ops = {
1871 .rpc_call_prepare = nfs_commit_prepare,
1872 .rpc_call_done = nfs_commit_done,
1873 .rpc_release = nfs_commit_release,
1876 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1877 .completion = nfs_commit_release_pages,
1878 .resched_write = nfs_commit_resched_write,
1881 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1882 int how, struct nfs_commit_info *cinfo)
1886 status = pnfs_commit_list(inode, head, how, cinfo);
1887 if (status == PNFS_NOT_ATTEMPTED)
1888 status = nfs_commit_list(inode, head, how, cinfo);
1892 static int __nfs_commit_inode(struct inode *inode, int how,
1893 struct writeback_control *wbc)
1896 struct nfs_commit_info cinfo;
1897 int may_wait = how & FLUSH_SYNC;
1901 nfs_init_cinfo_from_inode(&cinfo, inode);
1902 nfs_commit_begin(cinfo.mds);
1904 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1907 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1911 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1912 if (nscan < wbc->nr_to_write)
1913 wbc->nr_to_write -= nscan;
1915 wbc->nr_to_write = 0;
1917 if (nscan < INT_MAX)
1921 nfs_commit_end(cinfo.mds);
1922 if (ret || !may_wait)
1924 return wait_on_commit(cinfo.mds);
1927 int nfs_commit_inode(struct inode *inode, int how)
1929 return __nfs_commit_inode(inode, how, NULL);
1931 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1933 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1935 struct nfs_inode *nfsi = NFS_I(inode);
1936 int flags = FLUSH_SYNC;
1939 if (wbc->sync_mode == WB_SYNC_NONE) {
1940 /* no commits means nothing needs to be done */
1941 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1942 goto check_requests_outstanding;
1944 /* Don't commit yet if this is a non-blocking flush and there
1945 * are a lot of outstanding writes for this mapping.
1947 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1948 goto out_mark_dirty;
1950 /* don't wait for the COMMIT response */
1954 ret = __nfs_commit_inode(inode, flags, wbc);
1956 if (flags & FLUSH_SYNC)
1958 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1959 goto out_mark_dirty;
1961 check_requests_outstanding:
1962 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1965 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1968 EXPORT_SYMBOL_GPL(nfs_write_inode);
1971 * Wrapper for filemap_write_and_wait_range()
1973 * Needed for pNFS in order to ensure data becomes visible to the
1976 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1977 loff_t lstart, loff_t lend)
1981 ret = filemap_write_and_wait_range(mapping, lstart, lend);
1983 ret = pnfs_sync_inode(mapping->host, true);
1986 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1989 * flush the inode to disk.
1991 int nfs_wb_all(struct inode *inode)
1995 trace_nfs_writeback_inode_enter(inode);
1997 ret = filemap_write_and_wait(inode->i_mapping);
2000 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2003 pnfs_sync_inode(inode, true);
2007 trace_nfs_writeback_inode_exit(inode, ret);
2010 EXPORT_SYMBOL_GPL(nfs_wb_all);
2012 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2014 struct nfs_page *req;
2017 wait_on_page_writeback(page);
2019 /* blocking call to cancel all requests and join to a single (head)
2021 req = nfs_lock_and_join_requests(page);
2026 /* all requests from this page have been cancelled by
2027 * nfs_lock_and_join_requests, so just remove the head
2028 * request from the inode / page_private pointer and
2030 nfs_inode_remove_request(req);
2031 nfs_unlock_and_release_request(req);
2038 * Write back all requests on one page - we do this before reading it.
2040 int nfs_wb_page(struct inode *inode, struct page *page)
2042 loff_t range_start = page_file_offset(page);
2043 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2044 struct writeback_control wbc = {
2045 .sync_mode = WB_SYNC_ALL,
2047 .range_start = range_start,
2048 .range_end = range_end,
2052 trace_nfs_writeback_page_enter(inode);
2055 wait_on_page_writeback(page);
2056 if (clear_page_dirty_for_io(page)) {
2057 ret = nfs_writepage_locked(page, &wbc);
2063 if (!PagePrivate(page))
2065 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2070 trace_nfs_writeback_page_exit(inode, ret);
2074 #ifdef CONFIG_MIGRATION
2075 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2076 struct page *page, enum migrate_mode mode)
2079 * If PagePrivate is set, then the page is currently associated with
2080 * an in-progress read or write request. Don't try to migrate it.
2082 * FIXME: we could do this in principle, but we'll need a way to ensure
2083 * that we can safely release the inode reference while holding
2086 if (PagePrivate(page))
2089 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2092 return migrate_page(mapping, newpage, page, mode);
2096 int __init nfs_init_writepagecache(void)
2098 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2099 sizeof(struct nfs_pgio_header),
2100 0, SLAB_HWCACHE_ALIGN,
2102 if (nfs_wdata_cachep == NULL)
2105 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2107 if (nfs_wdata_mempool == NULL)
2108 goto out_destroy_write_cache;
2110 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2111 sizeof(struct nfs_commit_data),
2112 0, SLAB_HWCACHE_ALIGN,
2114 if (nfs_cdata_cachep == NULL)
2115 goto out_destroy_write_mempool;
2117 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2119 if (nfs_commit_mempool == NULL)
2120 goto out_destroy_commit_cache;
2123 * NFS congestion size, scale with available memory.
2135 * This allows larger machines to have larger/more transfers.
2136 * Limit the default to 256M
2138 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2139 if (nfs_congestion_kb > 256*1024)
2140 nfs_congestion_kb = 256*1024;
2144 out_destroy_commit_cache:
2145 kmem_cache_destroy(nfs_cdata_cachep);
2146 out_destroy_write_mempool:
2147 mempool_destroy(nfs_wdata_mempool);
2148 out_destroy_write_cache:
2149 kmem_cache_destroy(nfs_wdata_cachep);
2153 void nfs_destroy_writepagecache(void)
2155 mempool_destroy(nfs_commit_mempool);
2156 kmem_cache_destroy(nfs_cdata_cachep);
2157 mempool_destroy(nfs_wdata_mempool);
2158 kmem_cache_destroy(nfs_wdata_cachep);
2161 static const struct nfs_rw_ops nfs_rw_write_ops = {
2162 .rw_alloc_header = nfs_writehdr_alloc,
2163 .rw_free_header = nfs_writehdr_free,
2164 .rw_done = nfs_writeback_done,
2165 .rw_result = nfs_writeback_result,
2166 .rw_initiate = nfs_initiate_write,