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>
27 #include <linux/uaccess.h>
29 #include "delegation.h"
38 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
40 #define MIN_POOL_WRITE (32)
41 #define MIN_POOL_COMMIT (4)
43 struct nfs_io_completion {
44 void (*complete)(void *data);
50 * Local function declarations
52 static void nfs_redirty_request(struct nfs_page *req);
53 static const struct rpc_call_ops nfs_commit_ops;
54 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
55 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
56 static const struct nfs_rw_ops nfs_rw_write_ops;
57 static void nfs_clear_request_commit(struct nfs_page *req);
58 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
60 static struct nfs_page *
61 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
64 static struct kmem_cache *nfs_wdata_cachep;
65 static mempool_t *nfs_wdata_mempool;
66 static struct kmem_cache *nfs_cdata_cachep;
67 static mempool_t *nfs_commit_mempool;
69 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
71 struct nfs_commit_data *p;
74 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
76 /* It is OK to do some reclaim, not no safe to wait
77 * for anything to be returned to the pool.
78 * mempool_alloc() cannot handle that particular combination,
79 * so we need two separate attempts.
81 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
83 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
84 __GFP_NOWARN | __GFP_NORETRY);
89 memset(p, 0, sizeof(*p));
90 INIT_LIST_HEAD(&p->pages);
93 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
95 void nfs_commit_free(struct nfs_commit_data *p)
97 mempool_free(p, nfs_commit_mempool);
99 EXPORT_SYMBOL_GPL(nfs_commit_free);
101 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
103 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
105 memset(p, 0, sizeof(*p));
106 p->rw_mode = FMODE_WRITE;
110 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
112 mempool_free(hdr, nfs_wdata_mempool);
115 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
117 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
120 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
121 void (*complete)(void *), void *data)
123 ioc->complete = complete;
125 kref_init(&ioc->refcount);
128 static void nfs_io_completion_release(struct kref *kref)
130 struct nfs_io_completion *ioc = container_of(kref,
131 struct nfs_io_completion, refcount);
132 ioc->complete(ioc->data);
136 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
139 kref_get(&ioc->refcount);
142 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
145 kref_put(&ioc->refcount, nfs_io_completion_release);
148 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
152 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
155 static struct nfs_page *
156 nfs_page_private_request(struct page *page)
158 if (!PagePrivate(page))
160 return (struct nfs_page *)page_private(page);
164 * nfs_page_find_head_request_locked - find head request associated with @page
166 * must be called while holding the inode lock.
168 * returns matching head request with reference held, or NULL if not found.
170 static struct nfs_page *
171 nfs_page_find_private_request(struct page *page)
173 struct address_space *mapping = page_file_mapping(page);
174 struct nfs_page *req;
176 if (!PagePrivate(page))
178 spin_lock(&mapping->private_lock);
179 req = nfs_page_private_request(page);
181 WARN_ON_ONCE(req->wb_head != req);
182 kref_get(&req->wb_kref);
184 spin_unlock(&mapping->private_lock);
188 static struct nfs_page *
189 nfs_page_find_swap_request(struct page *page)
191 struct inode *inode = page_file_mapping(page)->host;
192 struct nfs_inode *nfsi = NFS_I(inode);
193 struct nfs_page *req = NULL;
194 if (!PageSwapCache(page))
196 mutex_lock(&nfsi->commit_mutex);
197 if (PageSwapCache(page)) {
198 req = nfs_page_search_commits_for_head_request_locked(nfsi,
201 WARN_ON_ONCE(req->wb_head != req);
202 kref_get(&req->wb_kref);
205 mutex_unlock(&nfsi->commit_mutex);
210 * nfs_page_find_head_request - find head request associated with @page
212 * returns matching head request with reference held, or NULL if not found.
214 static struct nfs_page *nfs_page_find_head_request(struct page *page)
216 struct nfs_page *req;
218 req = nfs_page_find_private_request(page);
220 req = nfs_page_find_swap_request(page);
224 /* Adjust the file length if we're writing beyond the end */
225 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
227 struct inode *inode = page_file_mapping(page)->host;
231 spin_lock(&inode->i_lock);
232 i_size = i_size_read(inode);
233 end_index = (i_size - 1) >> PAGE_SHIFT;
234 if (i_size > 0 && page_index(page) < end_index)
236 end = page_file_offset(page) + ((loff_t)offset+count);
239 i_size_write(inode, end);
240 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
242 spin_unlock(&inode->i_lock);
245 /* A writeback failed: mark the page as bad, and invalidate the page cache */
246 static void nfs_set_pageerror(struct page *page)
248 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
252 * nfs_page_group_search_locked
253 * @head - head request of page group
254 * @page_offset - offset into page
256 * Search page group with head @head to find a request that contains the
257 * page offset @page_offset.
259 * Returns a pointer to the first matching nfs request, or NULL if no
262 * Must be called with the page group lock held
264 static struct nfs_page *
265 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
267 struct nfs_page *req;
271 if (page_offset >= req->wb_pgbase &&
272 page_offset < (req->wb_pgbase + req->wb_bytes))
275 req = req->wb_this_page;
276 } while (req != head);
282 * nfs_page_group_covers_page
283 * @head - head request of page group
285 * Return true if the page group with head @head covers the whole page,
286 * returns false otherwise
288 static bool nfs_page_group_covers_page(struct nfs_page *req)
290 struct nfs_page *tmp;
291 unsigned int pos = 0;
292 unsigned int len = nfs_page_length(req->wb_page);
294 nfs_page_group_lock(req);
297 tmp = nfs_page_group_search_locked(req->wb_head, pos);
300 pos = tmp->wb_pgbase + tmp->wb_bytes;
303 nfs_page_group_unlock(req);
307 /* We can set the PG_uptodate flag if we see that a write request
308 * covers the full page.
310 static void nfs_mark_uptodate(struct nfs_page *req)
312 if (PageUptodate(req->wb_page))
314 if (!nfs_page_group_covers_page(req))
316 SetPageUptodate(req->wb_page);
319 static int wb_priority(struct writeback_control *wbc)
323 if (wbc->sync_mode == WB_SYNC_ALL)
324 ret = FLUSH_COND_STABLE;
329 * NFS congestion control
332 int nfs_congestion_kb;
334 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
335 #define NFS_CONGESTION_OFF_THRESH \
336 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
338 static void nfs_set_page_writeback(struct page *page)
340 struct inode *inode = page_file_mapping(page)->host;
341 struct nfs_server *nfss = NFS_SERVER(inode);
342 int ret = test_set_page_writeback(page);
344 WARN_ON_ONCE(ret != 0);
346 if (atomic_long_inc_return(&nfss->writeback) >
347 NFS_CONGESTION_ON_THRESH)
348 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
351 static void nfs_end_page_writeback(struct nfs_page *req)
353 struct inode *inode = page_file_mapping(req->wb_page)->host;
354 struct nfs_server *nfss = NFS_SERVER(inode);
357 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
358 nfs_unlock_request(req);
362 end_page_writeback(req->wb_page);
363 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
364 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
368 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
370 * this is a helper function for nfs_lock_and_join_requests
372 * @inode - inode associated with request page group, must be holding inode lock
373 * @head - head request of page group, must be holding head lock
374 * @req - request that couldn't lock and needs to wait on the req bit lock
376 * NOTE: this must be called holding page_group bit lock
377 * which will be released before returning.
379 * returns 0 on success, < 0 on error.
382 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
383 struct nfs_page *req)
385 struct nfs_page *tmp;
387 /* relinquish all the locks successfully grabbed this run */
388 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
389 if (!kref_read(&tmp->wb_kref))
391 nfs_unlock_and_release_request(tmp);
396 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
398 * @destroy_list - request list (using wb_this_page) terminated by @old_head
399 * @old_head - the old head of the list
401 * All subrequests must be locked and removed from all lists, so at this point
402 * they are only "active" in this function, and possibly in nfs_wait_on_request
403 * with a reference held by some other context.
406 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
407 struct nfs_page *old_head,
410 while (destroy_list) {
411 struct nfs_page *subreq = destroy_list;
413 destroy_list = (subreq->wb_this_page == old_head) ?
414 NULL : subreq->wb_this_page;
416 WARN_ON_ONCE(old_head != subreq->wb_head);
418 /* make sure old group is not used */
419 subreq->wb_this_page = subreq;
421 clear_bit(PG_REMOVE, &subreq->wb_flags);
423 /* Note: races with nfs_page_group_destroy() */
424 if (!kref_read(&subreq->wb_kref)) {
425 /* Check if we raced with nfs_page_group_destroy() */
426 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
427 nfs_free_request(subreq);
431 subreq->wb_head = subreq;
433 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
434 nfs_release_request(subreq);
435 atomic_long_dec(&NFS_I(inode)->nrequests);
438 /* subreq is now totally disconnected from page group or any
439 * write / commit lists. last chance to wake any waiters */
440 nfs_unlock_and_release_request(subreq);
445 * nfs_lock_and_join_requests - join all subreqs to the head req and return
446 * a locked reference, cancelling any pending
447 * operations for this page.
449 * @page - the page used to lookup the "page group" of nfs_page structures
451 * This function joins all sub requests to the head request by first
452 * locking all requests in the group, cancelling any pending operations
453 * and finally updating the head request to cover the whole range covered by
454 * the (former) group. All subrequests are removed from any write or commit
455 * lists, unlinked from the group and destroyed.
457 * Returns a locked, referenced pointer to the head request - which after
458 * this call is guaranteed to be the only request associated with the page.
459 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
460 * error was encountered.
462 static struct nfs_page *
463 nfs_lock_and_join_requests(struct page *page)
465 struct inode *inode = page_file_mapping(page)->host;
466 struct nfs_page *head, *subreq;
467 struct nfs_page *destroy_list = NULL;
468 unsigned int total_bytes;
473 * A reference is taken only on the head request which acts as a
474 * reference to the whole page group - the group will not be destroyed
475 * until the head reference is released.
477 head = nfs_page_find_head_request(page);
481 /* lock the page head first in order to avoid an ABBA inefficiency */
482 if (!nfs_lock_request(head)) {
483 ret = nfs_wait_on_request(head);
484 nfs_release_request(head);
490 /* Ensure that nobody removed the request before we locked it */
491 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
492 nfs_unlock_and_release_request(head);
496 ret = nfs_page_group_lock(head);
498 nfs_unlock_and_release_request(head);
502 /* lock each request in the page group */
503 total_bytes = head->wb_bytes;
504 for (subreq = head->wb_this_page; subreq != head;
505 subreq = subreq->wb_this_page) {
507 if (!kref_get_unless_zero(&subreq->wb_kref))
509 while (!nfs_lock_request(subreq)) {
511 * Unlock page to allow nfs_page_group_sync_on_bit()
514 nfs_page_group_unlock(head);
515 ret = nfs_wait_on_request(subreq);
517 ret = nfs_page_group_lock(head);
519 nfs_unroll_locks(inode, head, subreq);
520 nfs_release_request(subreq);
521 nfs_unlock_and_release_request(head);
526 * Subrequests are always contiguous, non overlapping
527 * and in order - but may be repeated (mirrored writes).
529 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
530 /* keep track of how many bytes this group covers */
531 total_bytes += subreq->wb_bytes;
532 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
533 ((subreq->wb_offset + subreq->wb_bytes) >
534 (head->wb_offset + total_bytes)))) {
535 nfs_page_group_unlock(head);
536 nfs_unroll_locks(inode, head, subreq);
537 nfs_unlock_and_release_request(subreq);
538 nfs_unlock_and_release_request(head);
539 return ERR_PTR(-EIO);
543 /* Now that all requests are locked, make sure they aren't on any list.
544 * Commit list removal accounting is done after locks are dropped */
547 nfs_clear_request_commit(subreq);
548 subreq = subreq->wb_this_page;
549 } while (subreq != head);
551 /* unlink subrequests from head, destroy them later */
552 if (head->wb_this_page != head) {
553 /* destroy list will be terminated by head */
554 destroy_list = head->wb_this_page;
555 head->wb_this_page = head;
557 /* change head request to cover whole range that
558 * the former page group covered */
559 head->wb_bytes = total_bytes;
562 /* Postpone destruction of this request */
563 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
564 set_bit(PG_INODE_REF, &head->wb_flags);
565 kref_get(&head->wb_kref);
566 atomic_long_inc(&NFS_I(inode)->nrequests);
569 nfs_page_group_unlock(head);
571 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
573 /* Did we lose a race with nfs_inode_remove_request()? */
574 if (!(PagePrivate(page) || PageSwapCache(page))) {
575 nfs_unlock_and_release_request(head);
579 /* still holds ref on head from nfs_page_find_head_request
580 * and still has lock on head from lock loop */
584 static void nfs_write_error_remove_page(struct nfs_page *req)
586 nfs_end_page_writeback(req);
587 generic_error_remove_page(page_file_mapping(req->wb_page),
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));
625 /* If there is a fatal error that covers this write, just exit */
626 if (nfs_error_is_fatal_on_server(req->wb_context->error))
629 if (!nfs_pageio_add_request(pgio, req)) {
630 ret = pgio->pg_error;
632 * Remove the problematic req upon fatal errors on the server
634 if (nfs_error_is_fatal(ret)) {
635 nfs_context_set_write_error(req->wb_context, ret);
636 if (nfs_error_is_fatal_on_server(ret))
639 nfs_redirty_request(req);
642 nfs_add_stats(page_file_mapping(page)->host,
643 NFSIOS_WRITEPAGES, 1);
647 nfs_write_error_remove_page(req);
651 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
652 struct nfs_pageio_descriptor *pgio)
656 nfs_pageio_cond_complete(pgio, page_index(page));
657 ret = nfs_page_async_flush(pgio, page);
658 if (ret == -EAGAIN) {
659 redirty_page_for_writepage(wbc, page);
666 * Write an mmapped page to the server.
668 static int nfs_writepage_locked(struct page *page,
669 struct writeback_control *wbc)
671 struct nfs_pageio_descriptor pgio;
672 struct inode *inode = page_file_mapping(page)->host;
675 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
676 nfs_pageio_init_write(&pgio, inode, 0,
677 false, &nfs_async_write_completion_ops);
678 err = nfs_do_writepage(page, wbc, &pgio);
679 nfs_pageio_complete(&pgio);
682 if (pgio.pg_error < 0)
683 return pgio.pg_error;
687 int nfs_writepage(struct page *page, struct writeback_control *wbc)
691 ret = nfs_writepage_locked(page, wbc);
696 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
700 ret = nfs_do_writepage(page, wbc, data);
705 static void nfs_io_completion_commit(void *inode)
707 nfs_commit_inode(inode, 0);
710 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
712 struct inode *inode = mapping->host;
713 struct nfs_pageio_descriptor pgio;
714 struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
717 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
720 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
722 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
723 &nfs_async_write_completion_ops);
724 pgio.pg_io_completion = ioc;
725 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
726 nfs_pageio_complete(&pgio);
727 nfs_io_completion_put(ioc);
740 * Insert a write request into an inode
742 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
744 struct address_space *mapping = page_file_mapping(req->wb_page);
745 struct nfs_inode *nfsi = NFS_I(inode);
747 WARN_ON_ONCE(req->wb_this_page != req);
749 /* Lock the request! */
750 nfs_lock_request(req);
753 * Swap-space should not get truncated. Hence no need to plug the race
754 * with invalidate/truncate.
756 spin_lock(&mapping->private_lock);
757 if (!nfs_have_writebacks(inode) &&
758 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) {
759 spin_lock(&inode->i_lock);
761 spin_unlock(&inode->i_lock);
763 if (likely(!PageSwapCache(req->wb_page))) {
764 set_bit(PG_MAPPED, &req->wb_flags);
765 SetPagePrivate(req->wb_page);
766 set_page_private(req->wb_page, (unsigned long)req);
768 spin_unlock(&mapping->private_lock);
769 atomic_long_inc(&nfsi->nrequests);
770 /* this a head request for a page group - mark it as having an
771 * extra reference so sub groups can follow suit.
772 * This flag also informs pgio layer when to bump nrequests when
773 * adding subrequests. */
774 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
775 kref_get(&req->wb_kref);
779 * Remove a write request from an inode
781 static void nfs_inode_remove_request(struct nfs_page *req)
783 struct address_space *mapping = page_file_mapping(req->wb_page);
784 struct inode *inode = mapping->host;
785 struct nfs_inode *nfsi = NFS_I(inode);
786 struct nfs_page *head;
788 atomic_long_dec(&nfsi->nrequests);
789 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
792 spin_lock(&mapping->private_lock);
793 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
794 set_page_private(head->wb_page, 0);
795 ClearPagePrivate(head->wb_page);
796 clear_bit(PG_MAPPED, &head->wb_flags);
798 spin_unlock(&mapping->private_lock);
801 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
802 nfs_release_request(req);
806 nfs_mark_request_dirty(struct nfs_page *req)
809 __set_page_dirty_nobuffers(req->wb_page);
813 * nfs_page_search_commits_for_head_request_locked
815 * Search through commit lists on @inode for the head request for @page.
816 * Must be called while holding the inode (which is cinfo) lock.
818 * Returns the head request if found, or NULL if not found.
820 static struct nfs_page *
821 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
824 struct nfs_page *freq, *t;
825 struct nfs_commit_info cinfo;
826 struct inode *inode = &nfsi->vfs_inode;
828 nfs_init_cinfo_from_inode(&cinfo, inode);
830 /* search through pnfs commit lists */
831 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
833 return freq->wb_head;
835 /* Linearly search the commit list for the correct request */
836 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
837 if (freq->wb_page == page)
838 return freq->wb_head;
845 * nfs_request_add_commit_list_locked - add request to a commit list
846 * @req: pointer to a struct nfs_page
847 * @dst: commit list head
848 * @cinfo: holds list lock and accounting info
850 * This sets the PG_CLEAN bit, updates the cinfo count of
851 * number of outstanding requests requiring a commit as well as
854 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
858 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
859 struct nfs_commit_info *cinfo)
861 set_bit(PG_CLEAN, &req->wb_flags);
862 nfs_list_add_request(req, dst);
863 atomic_long_inc(&cinfo->mds->ncommit);
865 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
868 * nfs_request_add_commit_list - add request to a commit list
869 * @req: pointer to a struct nfs_page
870 * @dst: commit list head
871 * @cinfo: holds list lock and accounting info
873 * This sets the PG_CLEAN bit, updates the cinfo count of
874 * number of outstanding requests requiring a commit as well as
877 * The caller must _not_ hold the cinfo->lock, but must be
878 * holding the nfs_page lock.
881 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
883 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
884 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
885 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
887 nfs_mark_page_unstable(req->wb_page, cinfo);
889 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
892 * nfs_request_remove_commit_list - Remove request from a commit list
893 * @req: pointer to a nfs_page
894 * @cinfo: holds list lock and accounting info
896 * This clears the PG_CLEAN bit, and updates the cinfo's count of
897 * number of outstanding requests requiring a commit
898 * It does not update the MM page stats.
900 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
903 nfs_request_remove_commit_list(struct nfs_page *req,
904 struct nfs_commit_info *cinfo)
906 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
908 nfs_list_remove_request(req);
909 atomic_long_dec(&cinfo->mds->ncommit);
911 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
913 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
916 cinfo->inode = inode;
917 cinfo->mds = &NFS_I(inode)->commit_info;
918 cinfo->ds = pnfs_get_ds_info(inode);
920 cinfo->completion_ops = &nfs_commit_completion_ops;
923 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
925 struct nfs_direct_req *dreq)
928 nfs_init_cinfo_from_dreq(cinfo, dreq);
930 nfs_init_cinfo_from_inode(cinfo, inode);
932 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
935 * Add a request to the inode's commit list.
938 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
939 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
941 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
943 nfs_request_add_commit_list(req, cinfo);
947 nfs_clear_page_commit(struct page *page)
949 dec_node_page_state(page, NR_UNSTABLE_NFS);
950 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
954 /* Called holding the request lock on @req */
956 nfs_clear_request_commit(struct nfs_page *req)
958 if (test_bit(PG_CLEAN, &req->wb_flags)) {
959 struct inode *inode = d_inode(req->wb_context->dentry);
960 struct nfs_commit_info cinfo;
962 nfs_init_cinfo_from_inode(&cinfo, inode);
963 mutex_lock(&NFS_I(inode)->commit_mutex);
964 if (!pnfs_clear_request_commit(req, &cinfo)) {
965 nfs_request_remove_commit_list(req, &cinfo);
967 mutex_unlock(&NFS_I(inode)->commit_mutex);
968 nfs_clear_page_commit(req->wb_page);
972 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
974 if (hdr->verf.committed == NFS_DATA_SYNC)
975 return hdr->lseg == NULL;
976 return hdr->verf.committed != NFS_FILE_SYNC;
979 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
981 nfs_io_completion_get(hdr->io_completion);
984 static void nfs_write_completion(struct nfs_pgio_header *hdr)
986 struct nfs_commit_info cinfo;
987 unsigned long bytes = 0;
989 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
991 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
992 while (!list_empty(&hdr->pages)) {
993 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
995 bytes += req->wb_bytes;
996 nfs_list_remove_request(req);
997 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
998 (hdr->good_bytes < bytes)) {
999 nfs_set_pageerror(req->wb_page);
1000 nfs_context_set_write_error(req->wb_context, hdr->error);
1003 if (nfs_write_need_commit(hdr)) {
1004 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1005 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1006 hdr->pgio_mirror_idx);
1010 nfs_inode_remove_request(req);
1012 nfs_end_page_writeback(req);
1013 nfs_release_request(req);
1016 nfs_io_completion_put(hdr->io_completion);
1021 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1023 return atomic_long_read(&cinfo->mds->ncommit);
1026 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1028 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1029 struct nfs_commit_info *cinfo, int max)
1031 struct nfs_page *req, *tmp;
1035 list_for_each_entry_safe(req, tmp, src, wb_list) {
1036 kref_get(&req->wb_kref);
1037 if (!nfs_lock_request(req)) {
1040 /* Prevent deadlock with nfs_lock_and_join_requests */
1041 if (!list_empty(dst)) {
1042 nfs_release_request(req);
1045 /* Ensure we make progress to prevent livelock */
1046 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1047 status = nfs_wait_on_request(req);
1048 nfs_release_request(req);
1049 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
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 inode *inode = page_file_mapping(page)->host;
1338 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1340 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1341 file, count, (long long)(page_file_offset(page) + offset));
1346 if (nfs_can_extend_write(file, page, inode)) {
1347 count = max(count + offset, nfs_page_length(page));
1351 status = nfs_writepage_setup(ctx, page, offset, count);
1353 nfs_set_pageerror(page);
1355 __set_page_dirty_nobuffers(page);
1357 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1358 status, (long long)i_size_read(inode));
1362 static int flush_task_priority(int how)
1364 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1366 return RPC_PRIORITY_HIGH;
1368 return RPC_PRIORITY_LOW;
1370 return RPC_PRIORITY_NORMAL;
1373 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1374 struct rpc_message *msg,
1375 const struct nfs_rpc_ops *rpc_ops,
1376 struct rpc_task_setup *task_setup_data, int how)
1378 int priority = flush_task_priority(how);
1380 task_setup_data->priority = priority;
1381 rpc_ops->write_setup(hdr, msg);
1383 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1384 &task_setup_data->rpc_client, msg, hdr);
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)
1401 struct nfs_page *req;
1403 while (!list_empty(head)) {
1404 req = nfs_list_entry(head->next);
1405 nfs_list_remove_request(req);
1406 nfs_redirty_request(req);
1410 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1412 nfs_async_write_error(&hdr->pages);
1415 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1416 .init_hdr = nfs_async_write_init,
1417 .error_cleanup = nfs_async_write_error,
1418 .completion = nfs_write_completion,
1419 .reschedule_io = nfs_async_write_reschedule_io,
1422 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1423 struct inode *inode, int ioflags, bool force_mds,
1424 const struct nfs_pgio_completion_ops *compl_ops)
1426 struct nfs_server *server = NFS_SERVER(inode);
1427 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1429 #ifdef CONFIG_NFS_V4_1
1430 if (server->pnfs_curr_ld && !force_mds)
1431 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1433 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1434 server->wsize, ioflags);
1436 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1438 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1440 struct nfs_pgio_mirror *mirror;
1442 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1443 pgio->pg_ops->pg_cleanup(pgio);
1445 pgio->pg_ops = &nfs_pgio_rw_ops;
1447 nfs_pageio_stop_mirroring(pgio);
1449 mirror = &pgio->pg_mirrors[0];
1450 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1452 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1455 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1457 struct nfs_commit_data *data = calldata;
1459 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1463 * Special version of should_remove_suid() that ignores capabilities.
1465 static int nfs_should_remove_suid(const struct inode *inode)
1467 umode_t mode = inode->i_mode;
1470 /* suid always must be killed */
1471 if (unlikely(mode & S_ISUID))
1472 kill = ATTR_KILL_SUID;
1475 * sgid without any exec bits is just a mandatory locking mark; leave
1476 * it alone. If some exec bits are set, it's a real sgid; kill it.
1478 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1479 kill |= ATTR_KILL_SGID;
1481 if (unlikely(kill && S_ISREG(mode)))
1487 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1488 struct nfs_fattr *fattr)
1490 struct nfs_pgio_args *argp = &hdr->args;
1491 struct nfs_pgio_res *resp = &hdr->res;
1492 u64 size = argp->offset + resp->count;
1494 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1496 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1497 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1500 if (size != fattr->size)
1502 /* Set attribute barrier */
1503 nfs_fattr_set_barrier(fattr);
1504 /* ...and update size */
1505 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1508 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1510 struct nfs_fattr *fattr = &hdr->fattr;
1511 struct inode *inode = hdr->inode;
1513 spin_lock(&inode->i_lock);
1514 nfs_writeback_check_extend(hdr, fattr);
1515 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1516 spin_unlock(&inode->i_lock);
1518 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1521 * This function is called when the WRITE call is complete.
1523 static int nfs_writeback_done(struct rpc_task *task,
1524 struct nfs_pgio_header *hdr,
1525 struct inode *inode)
1530 * ->write_done will attempt to use post-op attributes to detect
1531 * conflicting writes by other clients. A strict interpretation
1532 * of close-to-open would allow us to continue caching even if
1533 * another writer had changed the file, but some applications
1534 * depend on tighter cache coherency when writing.
1536 status = NFS_PROTO(inode)->write_done(task, hdr);
1539 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1541 if (hdr->res.verf->committed < hdr->args.stable &&
1542 task->tk_status >= 0) {
1543 /* We tried a write call, but the server did not
1544 * commit data to stable storage even though we
1546 * Note: There is a known bug in Tru64 < 5.0 in which
1547 * the server reports NFS_DATA_SYNC, but performs
1548 * NFS_FILE_SYNC. We therefore implement this checking
1549 * as a dprintk() in order to avoid filling syslog.
1551 static unsigned long complain;
1553 /* Note this will print the MDS for a DS write */
1554 if (time_before(complain, jiffies)) {
1555 dprintk("NFS: faulty NFS server %s:"
1556 " (committed = %d) != (stable = %d)\n",
1557 NFS_SERVER(inode)->nfs_client->cl_hostname,
1558 hdr->res.verf->committed, hdr->args.stable);
1559 complain = jiffies + 300 * HZ;
1563 /* Deal with the suid/sgid bit corner case */
1564 if (nfs_should_remove_suid(inode))
1565 nfs_mark_for_revalidate(inode);
1570 * This function is called when the WRITE call is complete.
1572 static void nfs_writeback_result(struct rpc_task *task,
1573 struct nfs_pgio_header *hdr)
1575 struct nfs_pgio_args *argp = &hdr->args;
1576 struct nfs_pgio_res *resp = &hdr->res;
1578 if (resp->count < argp->count) {
1579 static unsigned long complain;
1581 /* This a short write! */
1582 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1584 /* Has the server at least made some progress? */
1585 if (resp->count == 0) {
1586 if (time_before(complain, jiffies)) {
1588 "NFS: Server wrote zero bytes, expected %u.\n",
1590 complain = jiffies + 300 * HZ;
1592 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1593 task->tk_status = -EIO;
1597 /* For non rpc-based layout drivers, retry-through-MDS */
1598 if (!task->tk_ops) {
1599 hdr->pnfs_error = -EAGAIN;
1603 /* Was this an NFSv2 write or an NFSv3 stable write? */
1604 if (resp->verf->committed != NFS_UNSTABLE) {
1605 /* Resend from where the server left off */
1606 hdr->mds_offset += resp->count;
1607 argp->offset += resp->count;
1608 argp->pgbase += resp->count;
1609 argp->count -= resp->count;
1611 /* Resend as a stable write in order to avoid
1612 * headaches in the case of a server crash.
1614 argp->stable = NFS_FILE_SYNC;
1616 rpc_restart_call_prepare(task);
1620 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1622 return wait_on_atomic_t(&cinfo->rpcs_out,
1623 nfs_wait_atomic_killable, TASK_KILLABLE);
1626 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1628 atomic_inc(&cinfo->rpcs_out);
1631 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1633 if (atomic_dec_and_test(&cinfo->rpcs_out))
1634 wake_up_atomic_t(&cinfo->rpcs_out);
1637 void nfs_commitdata_release(struct nfs_commit_data *data)
1639 put_nfs_open_context(data->context);
1640 nfs_commit_free(data);
1642 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1644 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1645 const struct nfs_rpc_ops *nfs_ops,
1646 const struct rpc_call_ops *call_ops,
1649 struct rpc_task *task;
1650 int priority = flush_task_priority(how);
1651 struct rpc_message msg = {
1652 .rpc_argp = &data->args,
1653 .rpc_resp = &data->res,
1654 .rpc_cred = data->cred,
1656 struct rpc_task_setup task_setup_data = {
1657 .task = &data->task,
1659 .rpc_message = &msg,
1660 .callback_ops = call_ops,
1661 .callback_data = data,
1662 .workqueue = nfsiod_workqueue,
1663 .flags = RPC_TASK_ASYNC | flags,
1664 .priority = priority,
1666 /* Set up the initial task struct. */
1667 nfs_ops->commit_setup(data, &msg);
1669 dprintk("NFS: initiated commit call\n");
1671 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1672 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1674 task = rpc_run_task(&task_setup_data);
1676 return PTR_ERR(task);
1677 if (how & FLUSH_SYNC)
1678 rpc_wait_for_completion_task(task);
1682 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1684 static loff_t nfs_get_lwb(struct list_head *head)
1687 struct nfs_page *req;
1689 list_for_each_entry(req, head, wb_list)
1690 if (lwb < (req_offset(req) + req->wb_bytes))
1691 lwb = req_offset(req) + req->wb_bytes;
1697 * Set up the argument/result storage required for the RPC call.
1699 void nfs_init_commit(struct nfs_commit_data *data,
1700 struct list_head *head,
1701 struct pnfs_layout_segment *lseg,
1702 struct nfs_commit_info *cinfo)
1704 struct nfs_page *first = nfs_list_entry(head->next);
1705 struct inode *inode = d_inode(first->wb_context->dentry);
1707 /* Set up the RPC argument and reply structs
1708 * NB: take care not to mess about with data->commit et al. */
1710 list_splice_init(head, &data->pages);
1712 data->inode = inode;
1713 data->cred = first->wb_context->cred;
1714 data->lseg = lseg; /* reference transferred */
1715 /* only set lwb for pnfs commit */
1717 data->lwb = nfs_get_lwb(&data->pages);
1718 data->mds_ops = &nfs_commit_ops;
1719 data->completion_ops = cinfo->completion_ops;
1720 data->dreq = cinfo->dreq;
1722 data->args.fh = NFS_FH(data->inode);
1723 /* Note: we always request a commit of the entire inode */
1724 data->args.offset = 0;
1725 data->args.count = 0;
1726 data->context = get_nfs_open_context(first->wb_context);
1727 data->res.fattr = &data->fattr;
1728 data->res.verf = &data->verf;
1729 nfs_fattr_init(&data->fattr);
1731 EXPORT_SYMBOL_GPL(nfs_init_commit);
1733 void nfs_retry_commit(struct list_head *page_list,
1734 struct pnfs_layout_segment *lseg,
1735 struct nfs_commit_info *cinfo,
1738 struct nfs_page *req;
1740 while (!list_empty(page_list)) {
1741 req = nfs_list_entry(page_list->next);
1742 nfs_list_remove_request(req);
1743 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1745 nfs_clear_page_commit(req->wb_page);
1746 nfs_unlock_and_release_request(req);
1749 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1752 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1753 struct nfs_page *req)
1755 __set_page_dirty_nobuffers(req->wb_page);
1759 * Commit dirty pages
1762 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1763 struct nfs_commit_info *cinfo)
1765 struct nfs_commit_data *data;
1767 /* another commit raced with us */
1768 if (list_empty(head))
1771 data = nfs_commitdata_alloc(true);
1773 /* Set up the argument struct */
1774 nfs_init_commit(data, head, NULL, cinfo);
1775 atomic_inc(&cinfo->mds->rpcs_out);
1776 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1777 data->mds_ops, how, 0);
1781 * COMMIT call returned
1783 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1785 struct nfs_commit_data *data = calldata;
1787 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1788 task->tk_pid, task->tk_status);
1790 /* Call the NFS version-specific code */
1791 NFS_PROTO(data->inode)->commit_done(task, data);
1794 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1796 struct nfs_page *req;
1797 int status = data->task.tk_status;
1798 struct nfs_commit_info cinfo;
1799 struct nfs_server *nfss;
1801 while (!list_empty(&data->pages)) {
1802 req = nfs_list_entry(data->pages.next);
1803 nfs_list_remove_request(req);
1805 nfs_clear_page_commit(req->wb_page);
1807 dprintk("NFS: commit (%s/%llu %d@%lld)",
1808 req->wb_context->dentry->d_sb->s_id,
1809 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1811 (long long)req_offset(req));
1813 nfs_context_set_write_error(req->wb_context, status);
1815 nfs_inode_remove_request(req);
1816 dprintk_cont(", error = %d\n", status);
1820 /* Okay, COMMIT succeeded, apparently. Check the verifier
1821 * returned by the server against all stored verfs. */
1822 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1823 /* We have a match */
1825 nfs_inode_remove_request(req);
1826 dprintk_cont(" OK\n");
1829 /* We have a mismatch. Write the page again */
1830 dprintk_cont(" mismatch\n");
1831 nfs_mark_request_dirty(req);
1832 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1834 nfs_unlock_and_release_request(req);
1836 nfss = NFS_SERVER(data->inode);
1837 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1838 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1840 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1841 nfs_commit_end(cinfo.mds);
1844 static void nfs_commit_release(void *calldata)
1846 struct nfs_commit_data *data = calldata;
1848 data->completion_ops->completion(data);
1849 nfs_commitdata_release(calldata);
1852 static const struct rpc_call_ops nfs_commit_ops = {
1853 .rpc_call_prepare = nfs_commit_prepare,
1854 .rpc_call_done = nfs_commit_done,
1855 .rpc_release = nfs_commit_release,
1858 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1859 .completion = nfs_commit_release_pages,
1860 .resched_write = nfs_commit_resched_write,
1863 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1864 int how, struct nfs_commit_info *cinfo)
1868 status = pnfs_commit_list(inode, head, how, cinfo);
1869 if (status == PNFS_NOT_ATTEMPTED)
1870 status = nfs_commit_list(inode, head, how, cinfo);
1874 int nfs_commit_inode(struct inode *inode, int how)
1877 struct nfs_commit_info cinfo;
1878 int may_wait = how & FLUSH_SYNC;
1882 nfs_init_cinfo_from_inode(&cinfo, inode);
1883 nfs_commit_begin(cinfo.mds);
1884 res = nfs_scan_commit(inode, &head, &cinfo);
1886 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1887 nfs_commit_end(cinfo.mds);
1891 goto out_mark_dirty;
1892 error = wait_on_commit(cinfo.mds);
1898 /* Note: If we exit without ensuring that the commit is complete,
1899 * we must mark the inode as dirty. Otherwise, future calls to
1900 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1901 * that the data is on the disk.
1904 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1907 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1909 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1911 struct nfs_inode *nfsi = NFS_I(inode);
1912 int flags = FLUSH_SYNC;
1915 /* no commits means nothing needs to be done */
1916 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1919 if (wbc->sync_mode == WB_SYNC_NONE) {
1920 /* Don't commit yet if this is a non-blocking flush and there
1921 * are a lot of outstanding writes for this mapping.
1923 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1924 goto out_mark_dirty;
1926 /* don't wait for the COMMIT response */
1930 ret = nfs_commit_inode(inode, flags);
1932 if (wbc->sync_mode == WB_SYNC_NONE) {
1933 if (ret < wbc->nr_to_write)
1934 wbc->nr_to_write -= ret;
1936 wbc->nr_to_write = 0;
1941 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1944 EXPORT_SYMBOL_GPL(nfs_write_inode);
1947 * Wrapper for filemap_write_and_wait_range()
1949 * Needed for pNFS in order to ensure data becomes visible to the
1952 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1953 loff_t lstart, loff_t lend)
1957 ret = filemap_write_and_wait_range(mapping, lstart, lend);
1959 ret = pnfs_sync_inode(mapping->host, true);
1962 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1965 * flush the inode to disk.
1967 int nfs_wb_all(struct inode *inode)
1971 trace_nfs_writeback_inode_enter(inode);
1973 ret = filemap_write_and_wait(inode->i_mapping);
1976 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1979 pnfs_sync_inode(inode, true);
1983 trace_nfs_writeback_inode_exit(inode, ret);
1986 EXPORT_SYMBOL_GPL(nfs_wb_all);
1988 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1990 struct nfs_page *req;
1993 wait_on_page_writeback(page);
1995 /* blocking call to cancel all requests and join to a single (head)
1997 req = nfs_lock_and_join_requests(page);
2002 /* all requests from this page have been cancelled by
2003 * nfs_lock_and_join_requests, so just remove the head
2004 * request from the inode / page_private pointer and
2006 nfs_inode_remove_request(req);
2007 nfs_unlock_and_release_request(req);
2014 * Write back all requests on one page - we do this before reading it.
2016 int nfs_wb_page(struct inode *inode, struct page *page)
2018 loff_t range_start = page_file_offset(page);
2019 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2020 struct writeback_control wbc = {
2021 .sync_mode = WB_SYNC_ALL,
2023 .range_start = range_start,
2024 .range_end = range_end,
2028 trace_nfs_writeback_page_enter(inode);
2031 wait_on_page_writeback(page);
2032 if (clear_page_dirty_for_io(page)) {
2033 ret = nfs_writepage_locked(page, &wbc);
2039 if (!PagePrivate(page))
2041 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2046 trace_nfs_writeback_page_exit(inode, ret);
2050 #ifdef CONFIG_MIGRATION
2051 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2052 struct page *page, enum migrate_mode mode)
2055 * If PagePrivate is set, then the page is currently associated with
2056 * an in-progress read or write request. Don't try to migrate it.
2058 * FIXME: we could do this in principle, but we'll need a way to ensure
2059 * that we can safely release the inode reference while holding
2062 if (PagePrivate(page))
2065 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2068 return migrate_page(mapping, newpage, page, mode);
2072 int __init nfs_init_writepagecache(void)
2074 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2075 sizeof(struct nfs_pgio_header),
2076 0, SLAB_HWCACHE_ALIGN,
2078 if (nfs_wdata_cachep == NULL)
2081 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2083 if (nfs_wdata_mempool == NULL)
2084 goto out_destroy_write_cache;
2086 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2087 sizeof(struct nfs_commit_data),
2088 0, SLAB_HWCACHE_ALIGN,
2090 if (nfs_cdata_cachep == NULL)
2091 goto out_destroy_write_mempool;
2093 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2095 if (nfs_commit_mempool == NULL)
2096 goto out_destroy_commit_cache;
2099 * NFS congestion size, scale with available memory.
2111 * This allows larger machines to have larger/more transfers.
2112 * Limit the default to 256M
2114 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2115 if (nfs_congestion_kb > 256*1024)
2116 nfs_congestion_kb = 256*1024;
2120 out_destroy_commit_cache:
2121 kmem_cache_destroy(nfs_cdata_cachep);
2122 out_destroy_write_mempool:
2123 mempool_destroy(nfs_wdata_mempool);
2124 out_destroy_write_cache:
2125 kmem_cache_destroy(nfs_wdata_cachep);
2129 void nfs_destroy_writepagecache(void)
2131 mempool_destroy(nfs_commit_mempool);
2132 kmem_cache_destroy(nfs_cdata_cachep);
2133 mempool_destroy(nfs_wdata_mempool);
2134 kmem_cache_destroy(nfs_wdata_cachep);
2137 static const struct nfs_rw_ops nfs_rw_write_ops = {
2138 .rw_alloc_header = nfs_writehdr_alloc,
2139 .rw_free_header = nfs_writehdr_free,
2140 .rw_done = nfs_writeback_done,
2141 .rw_result = nfs_writeback_result,
2142 .rw_initiate = nfs_initiate_write,