1 // SPDX-License-Identifier: GPL-2.0-only
5 * Write file data over NFS.
7 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
10 #include <linux/types.h>
11 #include <linux/slab.h>
13 #include <linux/pagemap.h>
14 #include <linux/file.h>
15 #include <linux/writeback.h>
16 #include <linux/swap.h>
17 #include <linux/migrate.h>
19 #include <linux/sunrpc/clnt.h>
20 #include <linux/nfs_fs.h>
21 #include <linux/nfs_mount.h>
22 #include <linux/nfs_page.h>
23 #include <linux/backing-dev.h>
24 #include <linux/export.h>
25 #include <linux/freezer.h>
26 #include <linux/wait.h>
27 #include <linux/iversion.h>
29 #include <linux/uaccess.h>
30 #include <linux/sched/mm.h>
32 #include "delegation.h"
41 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
43 #define MIN_POOL_WRITE (32)
44 #define MIN_POOL_COMMIT (4)
46 struct nfs_io_completion {
47 void (*complete)(void *data);
53 * Local function declarations
55 static void nfs_redirty_request(struct nfs_page *req);
56 static const struct rpc_call_ops nfs_commit_ops;
57 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
58 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
59 static const struct nfs_rw_ops nfs_rw_write_ops;
60 static void nfs_inode_remove_request(struct nfs_page *req);
61 static void nfs_clear_request_commit(struct nfs_page *req);
62 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
64 static struct nfs_page *
65 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
68 static struct kmem_cache *nfs_wdata_cachep;
69 static mempool_t *nfs_wdata_mempool;
70 static struct kmem_cache *nfs_cdata_cachep;
71 static mempool_t *nfs_commit_mempool;
73 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
75 struct nfs_commit_data *p;
78 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
80 /* It is OK to do some reclaim, not no safe to wait
81 * for anything to be returned to the pool.
82 * mempool_alloc() cannot handle that particular combination,
83 * so we need two separate attempts.
85 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
87 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
88 __GFP_NOWARN | __GFP_NORETRY);
93 memset(p, 0, sizeof(*p));
94 INIT_LIST_HEAD(&p->pages);
97 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
99 void nfs_commit_free(struct nfs_commit_data *p)
101 mempool_free(p, nfs_commit_mempool);
103 EXPORT_SYMBOL_GPL(nfs_commit_free);
105 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
107 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_KERNEL);
109 memset(p, 0, sizeof(*p));
110 p->rw_mode = FMODE_WRITE;
114 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
116 mempool_free(hdr, nfs_wdata_mempool);
119 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
121 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
124 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
125 void (*complete)(void *), void *data)
127 ioc->complete = complete;
129 kref_init(&ioc->refcount);
132 static void nfs_io_completion_release(struct kref *kref)
134 struct nfs_io_completion *ioc = container_of(kref,
135 struct nfs_io_completion, refcount);
136 ioc->complete(ioc->data);
140 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
143 kref_get(&ioc->refcount);
146 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
149 kref_put(&ioc->refcount, nfs_io_completion_release);
152 static struct nfs_page *
153 nfs_page_private_request(struct page *page)
155 if (!PagePrivate(page))
157 return (struct nfs_page *)page_private(page);
161 * nfs_page_find_head_request_locked - find head request associated with @page
163 * must be called while holding the inode lock.
165 * returns matching head request with reference held, or NULL if not found.
167 static struct nfs_page *
168 nfs_page_find_private_request(struct page *page)
170 struct address_space *mapping = page_file_mapping(page);
171 struct nfs_page *req;
173 if (!PagePrivate(page))
175 spin_lock(&mapping->private_lock);
176 req = nfs_page_private_request(page);
178 WARN_ON_ONCE(req->wb_head != req);
179 kref_get(&req->wb_kref);
181 spin_unlock(&mapping->private_lock);
185 static struct nfs_page *
186 nfs_page_find_swap_request(struct page *page)
188 struct inode *inode = page_file_mapping(page)->host;
189 struct nfs_inode *nfsi = NFS_I(inode);
190 struct nfs_page *req = NULL;
191 if (!PageSwapCache(page))
193 mutex_lock(&nfsi->commit_mutex);
194 if (PageSwapCache(page)) {
195 req = nfs_page_search_commits_for_head_request_locked(nfsi,
198 WARN_ON_ONCE(req->wb_head != req);
199 kref_get(&req->wb_kref);
202 mutex_unlock(&nfsi->commit_mutex);
207 * nfs_page_find_head_request - find head request associated with @page
209 * returns matching head request with reference held, or NULL if not found.
211 static struct nfs_page *nfs_page_find_head_request(struct page *page)
213 struct nfs_page *req;
215 req = nfs_page_find_private_request(page);
217 req = nfs_page_find_swap_request(page);
221 /* Adjust the file length if we're writing beyond the end */
222 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
224 struct inode *inode = page_file_mapping(page)->host;
228 spin_lock(&inode->i_lock);
229 i_size = i_size_read(inode);
230 end_index = (i_size - 1) >> PAGE_SHIFT;
231 if (i_size > 0 && page_index(page) < end_index)
233 end = page_file_offset(page) + ((loff_t)offset+count);
236 i_size_write(inode, end);
237 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
238 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
240 spin_unlock(&inode->i_lock);
243 /* A writeback failed: mark the page as bad, and invalidate the page cache */
244 static void nfs_set_pageerror(struct address_space *mapping)
246 struct inode *inode = mapping->host;
248 nfs_zap_mapping(mapping->host, mapping);
249 /* Force file size revalidation */
250 spin_lock(&inode->i_lock);
251 NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
252 NFS_INO_REVAL_PAGECACHE |
253 NFS_INO_INVALID_SIZE;
254 spin_unlock(&inode->i_lock);
257 static void nfs_mapping_set_error(struct page *page, int error)
260 mapping_set_error(page_file_mapping(page), error);
264 * nfs_page_group_search_locked
265 * @head - head request of page group
266 * @page_offset - offset into page
268 * Search page group with head @head to find a request that contains the
269 * page offset @page_offset.
271 * Returns a pointer to the first matching nfs request, or NULL if no
274 * Must be called with the page group lock held
276 static struct nfs_page *
277 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
279 struct nfs_page *req;
283 if (page_offset >= req->wb_pgbase &&
284 page_offset < (req->wb_pgbase + req->wb_bytes))
287 req = req->wb_this_page;
288 } while (req != head);
294 * nfs_page_group_covers_page
295 * @head - head request of page group
297 * Return true if the page group with head @head covers the whole page,
298 * returns false otherwise
300 static bool nfs_page_group_covers_page(struct nfs_page *req)
302 struct nfs_page *tmp;
303 unsigned int pos = 0;
304 unsigned int len = nfs_page_length(req->wb_page);
306 nfs_page_group_lock(req);
309 tmp = nfs_page_group_search_locked(req->wb_head, pos);
312 pos = tmp->wb_pgbase + tmp->wb_bytes;
315 nfs_page_group_unlock(req);
319 /* We can set the PG_uptodate flag if we see that a write request
320 * covers the full page.
322 static void nfs_mark_uptodate(struct nfs_page *req)
324 if (PageUptodate(req->wb_page))
326 if (!nfs_page_group_covers_page(req))
328 SetPageUptodate(req->wb_page);
331 static int wb_priority(struct writeback_control *wbc)
335 if (wbc->sync_mode == WB_SYNC_ALL)
336 ret = FLUSH_COND_STABLE;
341 * NFS congestion control
344 int nfs_congestion_kb;
346 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
347 #define NFS_CONGESTION_OFF_THRESH \
348 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
350 static void nfs_set_page_writeback(struct page *page)
352 struct inode *inode = page_file_mapping(page)->host;
353 struct nfs_server *nfss = NFS_SERVER(inode);
354 int ret = test_set_page_writeback(page);
356 WARN_ON_ONCE(ret != 0);
358 if (atomic_long_inc_return(&nfss->writeback) >
359 NFS_CONGESTION_ON_THRESH)
360 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
363 static void nfs_end_page_writeback(struct nfs_page *req)
365 struct inode *inode = page_file_mapping(req->wb_page)->host;
366 struct nfs_server *nfss = NFS_SERVER(inode);
369 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
370 nfs_unlock_request(req);
374 end_page_writeback(req->wb_page);
375 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
376 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
380 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
382 * this is a helper function for nfs_lock_and_join_requests
384 * @inode - inode associated with request page group, must be holding inode lock
385 * @head - head request of page group, must be holding head lock
386 * @req - request that couldn't lock and needs to wait on the req bit lock
388 * NOTE: this must be called holding page_group bit lock
389 * which will be released before returning.
391 * returns 0 on success, < 0 on error.
394 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
395 struct nfs_page *req)
397 struct nfs_page *tmp;
399 /* relinquish all the locks successfully grabbed this run */
400 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
401 if (!kref_read(&tmp->wb_kref))
403 nfs_unlock_and_release_request(tmp);
408 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
410 * @destroy_list - request list (using wb_this_page) terminated by @old_head
411 * @old_head - the old head of the list
413 * All subrequests must be locked and removed from all lists, so at this point
414 * they are only "active" in this function, and possibly in nfs_wait_on_request
415 * with a reference held by some other context.
418 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
419 struct nfs_page *old_head,
422 while (destroy_list) {
423 struct nfs_page *subreq = destroy_list;
425 destroy_list = (subreq->wb_this_page == old_head) ?
426 NULL : subreq->wb_this_page;
428 /* Note: lock subreq in order to change subreq->wb_head */
429 nfs_page_set_headlock(subreq);
430 WARN_ON_ONCE(old_head != subreq->wb_head);
432 /* make sure old group is not used */
433 subreq->wb_this_page = subreq;
434 subreq->wb_head = subreq;
436 clear_bit(PG_REMOVE, &subreq->wb_flags);
438 /* Note: races with nfs_page_group_destroy() */
439 if (!kref_read(&subreq->wb_kref)) {
440 /* Check if we raced with nfs_page_group_destroy() */
441 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
442 nfs_page_clear_headlock(subreq);
443 nfs_free_request(subreq);
445 nfs_page_clear_headlock(subreq);
448 nfs_page_clear_headlock(subreq);
450 nfs_release_request(old_head);
452 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
453 nfs_release_request(subreq);
454 atomic_long_dec(&NFS_I(inode)->nrequests);
457 /* subreq is now totally disconnected from page group or any
458 * write / commit lists. last chance to wake any waiters */
459 nfs_unlock_and_release_request(subreq);
464 * nfs_lock_and_join_requests - join all subreqs to the head req and return
465 * a locked reference, cancelling any pending
466 * operations for this page.
468 * @page - the page used to lookup the "page group" of nfs_page structures
470 * This function joins all sub requests to the head request by first
471 * locking all requests in the group, cancelling any pending operations
472 * and finally updating the head request to cover the whole range covered by
473 * the (former) group. All subrequests are removed from any write or commit
474 * lists, unlinked from the group and destroyed.
476 * Returns a locked, referenced pointer to the head request - which after
477 * this call is guaranteed to be the only request associated with the page.
478 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
479 * error was encountered.
481 static struct nfs_page *
482 nfs_lock_and_join_requests(struct page *page)
484 struct inode *inode = page_file_mapping(page)->host;
485 struct nfs_page *head, *subreq;
486 struct nfs_page *destroy_list = NULL;
487 unsigned int total_bytes;
492 * A reference is taken only on the head request which acts as a
493 * reference to the whole page group - the group will not be destroyed
494 * until the head reference is released.
496 head = nfs_page_find_head_request(page);
500 /* lock the page head first in order to avoid an ABBA inefficiency */
501 if (!nfs_lock_request(head)) {
502 ret = nfs_wait_on_request(head);
503 nfs_release_request(head);
509 /* Ensure that nobody removed the request before we locked it */
510 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
511 nfs_unlock_and_release_request(head);
515 ret = nfs_page_group_lock(head);
517 goto release_request;
519 /* lock each request in the page group */
520 total_bytes = head->wb_bytes;
521 for (subreq = head->wb_this_page; subreq != head;
522 subreq = subreq->wb_this_page) {
524 if (!kref_get_unless_zero(&subreq->wb_kref)) {
525 if (subreq->wb_offset == head->wb_offset + total_bytes)
526 total_bytes += subreq->wb_bytes;
530 while (!nfs_lock_request(subreq)) {
532 * Unlock page to allow nfs_page_group_sync_on_bit()
535 nfs_page_group_unlock(head);
536 ret = nfs_wait_on_request(subreq);
538 ret = nfs_page_group_lock(head);
540 nfs_unroll_locks(inode, head, subreq);
541 nfs_release_request(subreq);
542 goto release_request;
546 * Subrequests are always contiguous, non overlapping
547 * and in order - but may be repeated (mirrored writes).
549 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
550 /* keep track of how many bytes this group covers */
551 total_bytes += subreq->wb_bytes;
552 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
553 ((subreq->wb_offset + subreq->wb_bytes) >
554 (head->wb_offset + total_bytes)))) {
555 nfs_page_group_unlock(head);
556 nfs_unroll_locks(inode, head, subreq);
557 nfs_unlock_and_release_request(subreq);
559 goto release_request;
563 /* Now that all requests are locked, make sure they aren't on any list.
564 * Commit list removal accounting is done after locks are dropped */
567 nfs_clear_request_commit(subreq);
568 subreq = subreq->wb_this_page;
569 } while (subreq != head);
571 /* unlink subrequests from head, destroy them later */
572 if (head->wb_this_page != head) {
573 /* destroy list will be terminated by head */
574 destroy_list = head->wb_this_page;
575 head->wb_this_page = head;
577 /* change head request to cover whole range that
578 * the former page group covered */
579 head->wb_bytes = total_bytes;
582 /* Postpone destruction of this request */
583 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
584 set_bit(PG_INODE_REF, &head->wb_flags);
585 kref_get(&head->wb_kref);
586 atomic_long_inc(&NFS_I(inode)->nrequests);
589 nfs_page_group_unlock(head);
591 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
593 /* Did we lose a race with nfs_inode_remove_request()? */
594 if (!(PagePrivate(page) || PageSwapCache(page))) {
595 nfs_unlock_and_release_request(head);
599 /* still holds ref on head from nfs_page_find_head_request
600 * and still has lock on head from lock loop */
604 nfs_unlock_and_release_request(head);
608 static void nfs_write_error(struct nfs_page *req, int error)
610 nfs_set_pageerror(page_file_mapping(req->wb_page));
611 nfs_mapping_set_error(req->wb_page, error);
612 nfs_inode_remove_request(req);
613 nfs_end_page_writeback(req);
614 nfs_release_request(req);
618 * Find an associated nfs write request, and prepare to flush it out
619 * May return an error if the user signalled nfs_wait_on_request().
621 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
624 struct nfs_page *req;
627 req = nfs_lock_and_join_requests(page);
634 nfs_set_page_writeback(page);
635 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
637 /* If there is a fatal error that covers this write, just exit */
638 ret = pgio->pg_error;
639 if (nfs_error_is_fatal_on_server(ret))
643 if (!nfs_pageio_add_request(pgio, req)) {
644 ret = pgio->pg_error;
646 * Remove the problematic req upon fatal errors on the server
648 if (nfs_error_is_fatal(ret)) {
649 if (nfs_error_is_fatal_on_server(ret))
653 nfs_redirty_request(req);
656 nfs_add_stats(page_file_mapping(page)->host,
657 NFSIOS_WRITEPAGES, 1);
661 nfs_write_error(req, ret);
665 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
666 struct nfs_pageio_descriptor *pgio)
670 nfs_pageio_cond_complete(pgio, page_index(page));
671 ret = nfs_page_async_flush(pgio, page);
672 if (ret == -EAGAIN) {
673 redirty_page_for_writepage(wbc, page);
674 ret = AOP_WRITEPAGE_ACTIVATE;
680 * Write an mmapped page to the server.
682 static int nfs_writepage_locked(struct page *page,
683 struct writeback_control *wbc)
685 struct nfs_pageio_descriptor pgio;
686 struct inode *inode = page_file_mapping(page)->host;
689 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
690 nfs_pageio_init_write(&pgio, inode, 0,
691 false, &nfs_async_write_completion_ops);
692 err = nfs_do_writepage(page, wbc, &pgio);
694 nfs_pageio_complete(&pgio);
698 int nfs_writepage(struct page *page, struct writeback_control *wbc)
702 ret = nfs_writepage_locked(page, wbc);
703 if (ret != AOP_WRITEPAGE_ACTIVATE)
708 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
712 ret = nfs_do_writepage(page, wbc, data);
713 if (ret != AOP_WRITEPAGE_ACTIVATE)
718 static void nfs_io_completion_commit(void *inode)
720 nfs_commit_inode(inode, 0);
723 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
725 struct inode *inode = mapping->host;
726 struct nfs_pageio_descriptor pgio;
727 struct nfs_io_completion *ioc;
730 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
732 ioc = nfs_io_completion_alloc(GFP_KERNEL);
734 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
736 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
737 &nfs_async_write_completion_ops);
738 pgio.pg_io_completion = ioc;
739 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
741 nfs_pageio_complete(&pgio);
742 nfs_io_completion_put(ioc);
752 * Insert a write request into an inode
754 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
756 struct address_space *mapping = page_file_mapping(req->wb_page);
757 struct nfs_inode *nfsi = NFS_I(inode);
759 WARN_ON_ONCE(req->wb_this_page != req);
761 /* Lock the request! */
762 nfs_lock_request(req);
765 * Swap-space should not get truncated. Hence no need to plug the race
766 * with invalidate/truncate.
768 spin_lock(&mapping->private_lock);
769 if (!nfs_have_writebacks(inode) &&
770 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
771 inode_inc_iversion_raw(inode);
772 if (likely(!PageSwapCache(req->wb_page))) {
773 set_bit(PG_MAPPED, &req->wb_flags);
774 SetPagePrivate(req->wb_page);
775 set_page_private(req->wb_page, (unsigned long)req);
777 spin_unlock(&mapping->private_lock);
778 atomic_long_inc(&nfsi->nrequests);
779 /* this a head request for a page group - mark it as having an
780 * extra reference so sub groups can follow suit.
781 * This flag also informs pgio layer when to bump nrequests when
782 * adding subrequests. */
783 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
784 kref_get(&req->wb_kref);
788 * Remove a write request from an inode
790 static void nfs_inode_remove_request(struct nfs_page *req)
792 struct address_space *mapping = page_file_mapping(req->wb_page);
793 struct inode *inode = mapping->host;
794 struct nfs_inode *nfsi = NFS_I(inode);
795 struct nfs_page *head;
797 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
800 spin_lock(&mapping->private_lock);
801 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
802 set_page_private(head->wb_page, 0);
803 ClearPagePrivate(head->wb_page);
804 clear_bit(PG_MAPPED, &head->wb_flags);
806 spin_unlock(&mapping->private_lock);
809 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
810 nfs_release_request(req);
811 atomic_long_dec(&nfsi->nrequests);
816 nfs_mark_request_dirty(struct nfs_page *req)
819 __set_page_dirty_nobuffers(req->wb_page);
823 * nfs_page_search_commits_for_head_request_locked
825 * Search through commit lists on @inode for the head request for @page.
826 * Must be called while holding the inode (which is cinfo) lock.
828 * Returns the head request if found, or NULL if not found.
830 static struct nfs_page *
831 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
834 struct nfs_page *freq, *t;
835 struct nfs_commit_info cinfo;
836 struct inode *inode = &nfsi->vfs_inode;
838 nfs_init_cinfo_from_inode(&cinfo, inode);
840 /* search through pnfs commit lists */
841 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
843 return freq->wb_head;
845 /* Linearly search the commit list for the correct request */
846 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
847 if (freq->wb_page == page)
848 return freq->wb_head;
855 * nfs_request_add_commit_list_locked - add request to a commit list
856 * @req: pointer to a struct nfs_page
857 * @dst: commit list head
858 * @cinfo: holds list lock and accounting info
860 * This sets the PG_CLEAN bit, updates the cinfo count of
861 * number of outstanding requests requiring a commit as well as
864 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
868 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
869 struct nfs_commit_info *cinfo)
871 set_bit(PG_CLEAN, &req->wb_flags);
872 nfs_list_add_request(req, dst);
873 atomic_long_inc(&cinfo->mds->ncommit);
875 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
878 * nfs_request_add_commit_list - add request to a commit list
879 * @req: pointer to a struct nfs_page
880 * @cinfo: holds list lock and accounting info
882 * This sets the PG_CLEAN bit, updates the cinfo count of
883 * number of outstanding requests requiring a commit as well as
886 * The caller must _not_ hold the cinfo->lock, but must be
887 * holding the nfs_page lock.
890 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
892 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
893 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
894 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
896 nfs_mark_page_unstable(req->wb_page, cinfo);
898 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
901 * nfs_request_remove_commit_list - Remove request from a commit list
902 * @req: pointer to a nfs_page
903 * @cinfo: holds list lock and accounting info
905 * This clears the PG_CLEAN bit, and updates the cinfo's count of
906 * number of outstanding requests requiring a commit
907 * It does not update the MM page stats.
909 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
912 nfs_request_remove_commit_list(struct nfs_page *req,
913 struct nfs_commit_info *cinfo)
915 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
917 nfs_list_remove_request(req);
918 atomic_long_dec(&cinfo->mds->ncommit);
920 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
922 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
925 cinfo->inode = inode;
926 cinfo->mds = &NFS_I(inode)->commit_info;
927 cinfo->ds = pnfs_get_ds_info(inode);
929 cinfo->completion_ops = &nfs_commit_completion_ops;
932 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
934 struct nfs_direct_req *dreq)
937 nfs_init_cinfo_from_dreq(cinfo, dreq);
939 nfs_init_cinfo_from_inode(cinfo, inode);
941 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
944 * Add a request to the inode's commit list.
947 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
948 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
950 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
952 nfs_request_add_commit_list(req, cinfo);
956 nfs_clear_page_commit(struct page *page)
958 dec_node_page_state(page, NR_UNSTABLE_NFS);
959 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
963 /* Called holding the request lock on @req */
965 nfs_clear_request_commit(struct nfs_page *req)
967 if (test_bit(PG_CLEAN, &req->wb_flags)) {
968 struct nfs_open_context *ctx = nfs_req_openctx(req);
969 struct inode *inode = d_inode(ctx->dentry);
970 struct nfs_commit_info cinfo;
972 nfs_init_cinfo_from_inode(&cinfo, inode);
973 mutex_lock(&NFS_I(inode)->commit_mutex);
974 if (!pnfs_clear_request_commit(req, &cinfo)) {
975 nfs_request_remove_commit_list(req, &cinfo);
977 mutex_unlock(&NFS_I(inode)->commit_mutex);
978 nfs_clear_page_commit(req->wb_page);
982 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
984 if (hdr->verf.committed == NFS_DATA_SYNC)
985 return hdr->lseg == NULL;
986 return hdr->verf.committed != NFS_FILE_SYNC;
989 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
991 nfs_io_completion_get(hdr->io_completion);
994 static void nfs_write_completion(struct nfs_pgio_header *hdr)
996 struct nfs_commit_info cinfo;
997 unsigned long bytes = 0;
999 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1001 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1002 while (!list_empty(&hdr->pages)) {
1003 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1005 bytes += req->wb_bytes;
1006 nfs_list_remove_request(req);
1007 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1008 (hdr->good_bytes < bytes)) {
1009 nfs_set_pageerror(page_file_mapping(req->wb_page));
1010 nfs_mapping_set_error(req->wb_page, hdr->error);
1013 if (nfs_write_need_commit(hdr)) {
1014 /* Reset wb_nio, since the write was successful. */
1016 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1017 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1018 hdr->pgio_mirror_idx);
1022 nfs_inode_remove_request(req);
1024 nfs_end_page_writeback(req);
1025 nfs_release_request(req);
1028 nfs_io_completion_put(hdr->io_completion);
1033 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1035 return atomic_long_read(&cinfo->mds->ncommit);
1038 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1040 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1041 struct nfs_commit_info *cinfo, int max)
1043 struct nfs_page *req, *tmp;
1046 list_for_each_entry_safe(req, tmp, src, wb_list) {
1047 kref_get(&req->wb_kref);
1048 if (!nfs_lock_request(req)) {
1049 nfs_release_request(req);
1052 nfs_request_remove_commit_list(req, cinfo);
1053 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1054 nfs_list_add_request(req, dst);
1056 if ((ret == max) && !cinfo->dreq)
1062 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1065 * nfs_scan_commit - Scan an inode for commit requests
1066 * @inode: NFS inode to scan
1067 * @dst: mds destination list
1068 * @cinfo: mds and ds lists of reqs ready to commit
1070 * Moves requests from the inode's 'commit' request list.
1071 * The requests are *not* checked to ensure that they form a contiguous set.
1074 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1075 struct nfs_commit_info *cinfo)
1079 if (!atomic_long_read(&cinfo->mds->ncommit))
1081 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1082 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1083 const int max = INT_MAX;
1085 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1087 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1089 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1094 * Search for an existing write request, and attempt to update
1095 * it to reflect a new dirty region on a given page.
1097 * If the attempt fails, then the existing request is flushed out
1100 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1102 unsigned int offset,
1105 struct nfs_page *req;
1110 end = offset + bytes;
1112 req = nfs_lock_and_join_requests(page);
1113 if (IS_ERR_OR_NULL(req))
1116 rqend = req->wb_offset + req->wb_bytes;
1118 * Tell the caller to flush out the request if
1119 * the offsets are non-contiguous.
1120 * Note: nfs_flush_incompatible() will already
1121 * have flushed out requests having wrong owners.
1123 if (offset > rqend || end < req->wb_offset)
1126 /* Okay, the request matches. Update the region */
1127 if (offset < req->wb_offset) {
1128 req->wb_offset = offset;
1129 req->wb_pgbase = offset;
1132 req->wb_bytes = end - req->wb_offset;
1134 req->wb_bytes = rqend - req->wb_offset;
1139 * Note: we mark the request dirty here because
1140 * nfs_lock_and_join_requests() cannot preserve
1141 * commit flags, so we have to replay the write.
1143 nfs_mark_request_dirty(req);
1144 nfs_unlock_and_release_request(req);
1145 error = nfs_wb_page(inode, page);
1146 return (error < 0) ? ERR_PTR(error) : NULL;
1150 * Try to update an existing write request, or create one if there is none.
1152 * Note: Should always be called with the Page Lock held to prevent races
1153 * if we have to add a new request. Also assumes that the caller has
1154 * already called nfs_flush_incompatible() if necessary.
1156 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1157 struct page *page, unsigned int offset, unsigned int bytes)
1159 struct inode *inode = page_file_mapping(page)->host;
1160 struct nfs_page *req;
1162 req = nfs_try_to_update_request(inode, page, offset, bytes);
1165 req = nfs_create_request(ctx, page, offset, bytes);
1168 nfs_inode_add_request(inode, req);
1173 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1174 unsigned int offset, unsigned int count)
1176 struct nfs_page *req;
1178 req = nfs_setup_write_request(ctx, page, offset, count);
1180 return PTR_ERR(req);
1181 /* Update file length */
1182 nfs_grow_file(page, offset, count);
1183 nfs_mark_uptodate(req);
1184 nfs_mark_request_dirty(req);
1185 nfs_unlock_and_release_request(req);
1189 int nfs_flush_incompatible(struct file *file, struct page *page)
1191 struct nfs_open_context *ctx = nfs_file_open_context(file);
1192 struct nfs_lock_context *l_ctx;
1193 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1194 struct nfs_page *req;
1195 int do_flush, status;
1197 * Look for a request corresponding to this page. If there
1198 * is one, and it belongs to another file, we flush it out
1199 * before we try to copy anything into the page. Do this
1200 * due to the lack of an ACCESS-type call in NFSv2.
1201 * Also do the same if we find a request from an existing
1205 req = nfs_page_find_head_request(page);
1208 l_ctx = req->wb_lock_context;
1209 do_flush = req->wb_page != page ||
1210 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1211 if (l_ctx && flctx &&
1212 !(list_empty_careful(&flctx->flc_posix) &&
1213 list_empty_careful(&flctx->flc_flock))) {
1214 do_flush |= l_ctx->lockowner != current->files;
1216 nfs_release_request(req);
1219 status = nfs_wb_page(page_file_mapping(page)->host, page);
1220 } while (status == 0);
1225 * Avoid buffered writes when a open context credential's key would
1228 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1230 * Return 0 and set a credential flag which triggers the inode to flush
1231 * and performs NFS_FILE_SYNC writes if the key will expired within
1232 * RPC_KEY_EXPIRE_TIMEO.
1235 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1237 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1239 if (nfs_ctx_key_to_expire(ctx, inode) &&
1241 /* Already expired! */
1247 * Test if the open context credential key is marked to expire soon.
1249 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1251 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1252 struct rpc_cred *cred = ctx->ll_cred;
1253 struct auth_cred acred = {
1257 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1259 ctx->ll_cred = NULL;
1263 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1264 if (!cred || IS_ERR(cred))
1266 ctx->ll_cred = cred;
1267 return !!(cred->cr_ops->crkey_timeout &&
1268 cred->cr_ops->crkey_timeout(cred));
1272 * If the page cache is marked as unsafe or invalid, then we can't rely on
1273 * the PageUptodate() flag. In this case, we will need to turn off
1274 * write optimisations that depend on the page contents being correct.
1276 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1278 struct nfs_inode *nfsi = NFS_I(inode);
1280 if (nfs_have_delegated_attributes(inode))
1282 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1285 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1288 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1290 return PageUptodate(page) != 0;
1294 is_whole_file_wrlock(struct file_lock *fl)
1296 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1297 fl->fl_type == F_WRLCK;
1300 /* If we know the page is up to date, and we're not using byte range locks (or
1301 * if we have the whole file locked for writing), it may be more efficient to
1302 * extend the write to cover the entire page in order to avoid fragmentation
1305 * If the file is opened for synchronous writes then we can just skip the rest
1308 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1311 struct file_lock_context *flctx = inode->i_flctx;
1312 struct file_lock *fl;
1314 if (file->f_flags & O_DSYNC)
1316 if (!nfs_write_pageuptodate(page, inode))
1318 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1320 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1321 list_empty_careful(&flctx->flc_posix)))
1324 /* Check to see if there are whole file write locks */
1326 spin_lock(&flctx->flc_lock);
1327 if (!list_empty(&flctx->flc_posix)) {
1328 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1330 if (is_whole_file_wrlock(fl))
1332 } else if (!list_empty(&flctx->flc_flock)) {
1333 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1335 if (fl->fl_type == F_WRLCK)
1338 spin_unlock(&flctx->flc_lock);
1343 * Update and possibly write a cached page of an NFS file.
1345 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1346 * things with a page scheduled for an RPC call (e.g. invalidate it).
1348 int nfs_updatepage(struct file *file, struct page *page,
1349 unsigned int offset, unsigned int count)
1351 struct nfs_open_context *ctx = nfs_file_open_context(file);
1352 struct address_space *mapping = page_file_mapping(page);
1353 struct inode *inode = mapping->host;
1356 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1358 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1359 file, count, (long long)(page_file_offset(page) + offset));
1364 if (nfs_can_extend_write(file, page, inode)) {
1365 count = max(count + offset, nfs_page_length(page));
1369 status = nfs_writepage_setup(ctx, page, offset, count);
1371 nfs_set_pageerror(mapping);
1373 __set_page_dirty_nobuffers(page);
1375 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1376 status, (long long)i_size_read(inode));
1380 static int flush_task_priority(int how)
1382 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1384 return RPC_PRIORITY_HIGH;
1386 return RPC_PRIORITY_LOW;
1388 return RPC_PRIORITY_NORMAL;
1391 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1392 struct rpc_message *msg,
1393 const struct nfs_rpc_ops *rpc_ops,
1394 struct rpc_task_setup *task_setup_data, int how)
1396 int priority = flush_task_priority(how);
1398 task_setup_data->priority = priority;
1399 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1400 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1404 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1405 * call this on each, which will prepare them to be retried on next
1406 * writeback using standard nfs.
1408 static void nfs_redirty_request(struct nfs_page *req)
1410 /* Bump the transmission count */
1412 nfs_mark_request_dirty(req);
1413 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1414 nfs_end_page_writeback(req);
1415 nfs_release_request(req);
1418 static void nfs_async_write_error(struct list_head *head, int error)
1420 struct nfs_page *req;
1422 while (!list_empty(head)) {
1423 req = nfs_list_entry(head->next);
1424 nfs_list_remove_request(req);
1425 if (nfs_error_is_fatal_on_server(error))
1426 nfs_write_error(req, error);
1428 nfs_redirty_request(req);
1432 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1434 nfs_async_write_error(&hdr->pages, 0);
1435 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1436 hdr->args.offset + hdr->args.count - 1);
1439 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1440 .init_hdr = nfs_async_write_init,
1441 .error_cleanup = nfs_async_write_error,
1442 .completion = nfs_write_completion,
1443 .reschedule_io = nfs_async_write_reschedule_io,
1446 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1447 struct inode *inode, int ioflags, bool force_mds,
1448 const struct nfs_pgio_completion_ops *compl_ops)
1450 struct nfs_server *server = NFS_SERVER(inode);
1451 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1453 #ifdef CONFIG_NFS_V4_1
1454 if (server->pnfs_curr_ld && !force_mds)
1455 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1457 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1458 server->wsize, ioflags);
1460 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1462 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1464 struct nfs_pgio_mirror *mirror;
1466 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1467 pgio->pg_ops->pg_cleanup(pgio);
1469 pgio->pg_ops = &nfs_pgio_rw_ops;
1471 nfs_pageio_stop_mirroring(pgio);
1473 mirror = &pgio->pg_mirrors[0];
1474 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1476 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1479 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1481 struct nfs_commit_data *data = calldata;
1483 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1487 * Special version of should_remove_suid() that ignores capabilities.
1489 static int nfs_should_remove_suid(const struct inode *inode)
1491 umode_t mode = inode->i_mode;
1494 /* suid always must be killed */
1495 if (unlikely(mode & S_ISUID))
1496 kill = ATTR_KILL_SUID;
1499 * sgid without any exec bits is just a mandatory locking mark; leave
1500 * it alone. If some exec bits are set, it's a real sgid; kill it.
1502 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1503 kill |= ATTR_KILL_SGID;
1505 if (unlikely(kill && S_ISREG(mode)))
1511 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1512 struct nfs_fattr *fattr)
1514 struct nfs_pgio_args *argp = &hdr->args;
1515 struct nfs_pgio_res *resp = &hdr->res;
1516 u64 size = argp->offset + resp->count;
1518 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1520 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1521 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1524 if (size != fattr->size)
1526 /* Set attribute barrier */
1527 nfs_fattr_set_barrier(fattr);
1528 /* ...and update size */
1529 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1532 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1534 struct nfs_fattr *fattr = &hdr->fattr;
1535 struct inode *inode = hdr->inode;
1537 spin_lock(&inode->i_lock);
1538 nfs_writeback_check_extend(hdr, fattr);
1539 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1540 spin_unlock(&inode->i_lock);
1542 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1545 * This function is called when the WRITE call is complete.
1547 static int nfs_writeback_done(struct rpc_task *task,
1548 struct nfs_pgio_header *hdr,
1549 struct inode *inode)
1554 * ->write_done will attempt to use post-op attributes to detect
1555 * conflicting writes by other clients. A strict interpretation
1556 * of close-to-open would allow us to continue caching even if
1557 * another writer had changed the file, but some applications
1558 * depend on tighter cache coherency when writing.
1560 status = NFS_PROTO(inode)->write_done(task, hdr);
1564 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1565 trace_nfs_writeback_done(inode, task->tk_status,
1566 hdr->args.offset, hdr->res.verf);
1568 if (hdr->res.verf->committed < hdr->args.stable &&
1569 task->tk_status >= 0) {
1570 /* We tried a write call, but the server did not
1571 * commit data to stable storage even though we
1573 * Note: There is a known bug in Tru64 < 5.0 in which
1574 * the server reports NFS_DATA_SYNC, but performs
1575 * NFS_FILE_SYNC. We therefore implement this checking
1576 * as a dprintk() in order to avoid filling syslog.
1578 static unsigned long complain;
1580 /* Note this will print the MDS for a DS write */
1581 if (time_before(complain, jiffies)) {
1582 dprintk("NFS: faulty NFS server %s:"
1583 " (committed = %d) != (stable = %d)\n",
1584 NFS_SERVER(inode)->nfs_client->cl_hostname,
1585 hdr->res.verf->committed, hdr->args.stable);
1586 complain = jiffies + 300 * HZ;
1590 /* Deal with the suid/sgid bit corner case */
1591 if (nfs_should_remove_suid(inode)) {
1592 spin_lock(&inode->i_lock);
1593 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1594 spin_unlock(&inode->i_lock);
1600 * This function is called when the WRITE call is complete.
1602 static void nfs_writeback_result(struct rpc_task *task,
1603 struct nfs_pgio_header *hdr)
1605 struct nfs_pgio_args *argp = &hdr->args;
1606 struct nfs_pgio_res *resp = &hdr->res;
1608 if (resp->count < argp->count) {
1609 static unsigned long complain;
1611 /* This a short write! */
1612 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1614 /* Has the server at least made some progress? */
1615 if (resp->count == 0) {
1616 if (time_before(complain, jiffies)) {
1618 "NFS: Server wrote zero bytes, expected %u.\n",
1620 complain = jiffies + 300 * HZ;
1622 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1623 task->tk_status = -EIO;
1627 /* For non rpc-based layout drivers, retry-through-MDS */
1628 if (!task->tk_ops) {
1629 hdr->pnfs_error = -EAGAIN;
1633 /* Was this an NFSv2 write or an NFSv3 stable write? */
1634 if (resp->verf->committed != NFS_UNSTABLE) {
1635 /* Resend from where the server left off */
1636 hdr->mds_offset += resp->count;
1637 argp->offset += resp->count;
1638 argp->pgbase += resp->count;
1639 argp->count -= resp->count;
1641 /* Resend as a stable write in order to avoid
1642 * headaches in the case of a server crash.
1644 argp->stable = NFS_FILE_SYNC;
1646 rpc_restart_call_prepare(task);
1650 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1652 return wait_var_event_killable(&cinfo->rpcs_out,
1653 !atomic_read(&cinfo->rpcs_out));
1656 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1658 atomic_inc(&cinfo->rpcs_out);
1661 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1663 if (atomic_dec_and_test(&cinfo->rpcs_out))
1664 wake_up_var(&cinfo->rpcs_out);
1667 void nfs_commitdata_release(struct nfs_commit_data *data)
1669 put_nfs_open_context(data->context);
1670 nfs_commit_free(data);
1672 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1674 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1675 const struct nfs_rpc_ops *nfs_ops,
1676 const struct rpc_call_ops *call_ops,
1679 struct rpc_task *task;
1680 int priority = flush_task_priority(how);
1681 struct rpc_message msg = {
1682 .rpc_argp = &data->args,
1683 .rpc_resp = &data->res,
1684 .rpc_cred = data->cred,
1686 struct rpc_task_setup task_setup_data = {
1687 .task = &data->task,
1689 .rpc_message = &msg,
1690 .callback_ops = call_ops,
1691 .callback_data = data,
1692 .workqueue = nfsiod_workqueue,
1693 .flags = RPC_TASK_ASYNC | flags,
1694 .priority = priority,
1696 /* Set up the initial task struct. */
1697 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1698 trace_nfs_initiate_commit(data);
1700 dprintk("NFS: initiated commit call\n");
1702 task = rpc_run_task(&task_setup_data);
1704 return PTR_ERR(task);
1705 if (how & FLUSH_SYNC)
1706 rpc_wait_for_completion_task(task);
1710 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1712 static loff_t nfs_get_lwb(struct list_head *head)
1715 struct nfs_page *req;
1717 list_for_each_entry(req, head, wb_list)
1718 if (lwb < (req_offset(req) + req->wb_bytes))
1719 lwb = req_offset(req) + req->wb_bytes;
1725 * Set up the argument/result storage required for the RPC call.
1727 void nfs_init_commit(struct nfs_commit_data *data,
1728 struct list_head *head,
1729 struct pnfs_layout_segment *lseg,
1730 struct nfs_commit_info *cinfo)
1732 struct nfs_page *first = nfs_list_entry(head->next);
1733 struct nfs_open_context *ctx = nfs_req_openctx(first);
1734 struct inode *inode = d_inode(ctx->dentry);
1736 /* Set up the RPC argument and reply structs
1737 * NB: take care not to mess about with data->commit et al. */
1739 list_splice_init(head, &data->pages);
1741 data->inode = inode;
1742 data->cred = ctx->cred;
1743 data->lseg = lseg; /* reference transferred */
1744 /* only set lwb for pnfs commit */
1746 data->lwb = nfs_get_lwb(&data->pages);
1747 data->mds_ops = &nfs_commit_ops;
1748 data->completion_ops = cinfo->completion_ops;
1749 data->dreq = cinfo->dreq;
1751 data->args.fh = NFS_FH(data->inode);
1752 /* Note: we always request a commit of the entire inode */
1753 data->args.offset = 0;
1754 data->args.count = 0;
1755 data->context = get_nfs_open_context(ctx);
1756 data->res.fattr = &data->fattr;
1757 data->res.verf = &data->verf;
1758 nfs_fattr_init(&data->fattr);
1760 EXPORT_SYMBOL_GPL(nfs_init_commit);
1762 void nfs_retry_commit(struct list_head *page_list,
1763 struct pnfs_layout_segment *lseg,
1764 struct nfs_commit_info *cinfo,
1767 struct nfs_page *req;
1769 while (!list_empty(page_list)) {
1770 req = nfs_list_entry(page_list->next);
1771 nfs_list_remove_request(req);
1772 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1774 nfs_clear_page_commit(req->wb_page);
1775 nfs_unlock_and_release_request(req);
1778 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1781 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1782 struct nfs_page *req)
1784 __set_page_dirty_nobuffers(req->wb_page);
1788 * Commit dirty pages
1791 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1792 struct nfs_commit_info *cinfo)
1794 struct nfs_commit_data *data;
1796 /* another commit raced with us */
1797 if (list_empty(head))
1800 data = nfs_commitdata_alloc(true);
1802 /* Set up the argument struct */
1803 nfs_init_commit(data, head, NULL, cinfo);
1804 atomic_inc(&cinfo->mds->rpcs_out);
1805 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1806 data->mds_ops, how, 0);
1810 * COMMIT call returned
1812 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1814 struct nfs_commit_data *data = calldata;
1816 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1817 task->tk_pid, task->tk_status);
1819 /* Call the NFS version-specific code */
1820 NFS_PROTO(data->inode)->commit_done(task, data);
1821 trace_nfs_commit_done(data);
1824 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1826 const struct nfs_writeverf *verf = data->res.verf;
1827 struct nfs_page *req;
1828 int status = data->task.tk_status;
1829 struct nfs_commit_info cinfo;
1830 struct nfs_server *nfss;
1832 while (!list_empty(&data->pages)) {
1833 req = nfs_list_entry(data->pages.next);
1834 nfs_list_remove_request(req);
1836 nfs_clear_page_commit(req->wb_page);
1838 dprintk("NFS: commit (%s/%llu %d@%lld)",
1839 nfs_req_openctx(req)->dentry->d_sb->s_id,
1840 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1842 (long long)req_offset(req));
1845 nfs_mapping_set_error(req->wb_page, status);
1846 nfs_inode_remove_request(req);
1848 dprintk_cont(", error = %d\n", status);
1852 /* Okay, COMMIT succeeded, apparently. Check the verifier
1853 * returned by the server against all stored verfs. */
1854 if (verf->committed > NFS_UNSTABLE &&
1855 !nfs_write_verifier_cmp(&req->wb_verf, &verf->verifier)) {
1856 /* We have a match */
1858 nfs_inode_remove_request(req);
1859 dprintk_cont(" OK\n");
1862 /* We have a mismatch. Write the page again */
1863 dprintk_cont(" mismatch\n");
1864 nfs_mark_request_dirty(req);
1865 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1867 nfs_unlock_and_release_request(req);
1868 /* Latency breaker */
1871 nfss = NFS_SERVER(data->inode);
1872 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1873 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1875 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1876 nfs_commit_end(cinfo.mds);
1879 static void nfs_commit_release(void *calldata)
1881 struct nfs_commit_data *data = calldata;
1883 data->completion_ops->completion(data);
1884 nfs_commitdata_release(calldata);
1887 static const struct rpc_call_ops nfs_commit_ops = {
1888 .rpc_call_prepare = nfs_commit_prepare,
1889 .rpc_call_done = nfs_commit_done,
1890 .rpc_release = nfs_commit_release,
1893 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1894 .completion = nfs_commit_release_pages,
1895 .resched_write = nfs_commit_resched_write,
1898 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1899 int how, struct nfs_commit_info *cinfo)
1903 status = pnfs_commit_list(inode, head, how, cinfo);
1904 if (status == PNFS_NOT_ATTEMPTED)
1905 status = nfs_commit_list(inode, head, how, cinfo);
1909 static int __nfs_commit_inode(struct inode *inode, int how,
1910 struct writeback_control *wbc)
1913 struct nfs_commit_info cinfo;
1914 int may_wait = how & FLUSH_SYNC;
1918 nfs_init_cinfo_from_inode(&cinfo, inode);
1919 nfs_commit_begin(cinfo.mds);
1921 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1924 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1928 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1929 if (nscan < wbc->nr_to_write)
1930 wbc->nr_to_write -= nscan;
1932 wbc->nr_to_write = 0;
1934 if (nscan < INT_MAX)
1938 nfs_commit_end(cinfo.mds);
1939 if (ret || !may_wait)
1941 return wait_on_commit(cinfo.mds);
1944 int nfs_commit_inode(struct inode *inode, int how)
1946 return __nfs_commit_inode(inode, how, NULL);
1948 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1950 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1952 struct nfs_inode *nfsi = NFS_I(inode);
1953 int flags = FLUSH_SYNC;
1956 if (wbc->sync_mode == WB_SYNC_NONE) {
1957 /* no commits means nothing needs to be done */
1958 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1959 goto check_requests_outstanding;
1961 /* Don't commit yet if this is a non-blocking flush and there
1962 * are a lot of outstanding writes for this mapping.
1964 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1965 goto out_mark_dirty;
1967 /* don't wait for the COMMIT response */
1971 ret = __nfs_commit_inode(inode, flags, wbc);
1973 if (flags & FLUSH_SYNC)
1975 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1976 goto out_mark_dirty;
1978 check_requests_outstanding:
1979 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1982 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1985 EXPORT_SYMBOL_GPL(nfs_write_inode);
1988 * Wrapper for filemap_write_and_wait_range()
1990 * Needed for pNFS in order to ensure data becomes visible to the
1993 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1994 loff_t lstart, loff_t lend)
1998 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2000 ret = pnfs_sync_inode(mapping->host, true);
2003 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2006 * flush the inode to disk.
2008 int nfs_wb_all(struct inode *inode)
2012 trace_nfs_writeback_inode_enter(inode);
2014 ret = filemap_write_and_wait(inode->i_mapping);
2017 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2020 pnfs_sync_inode(inode, true);
2024 trace_nfs_writeback_inode_exit(inode, ret);
2027 EXPORT_SYMBOL_GPL(nfs_wb_all);
2029 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2031 struct nfs_page *req;
2034 wait_on_page_writeback(page);
2036 /* blocking call to cancel all requests and join to a single (head)
2038 req = nfs_lock_and_join_requests(page);
2043 /* all requests from this page have been cancelled by
2044 * nfs_lock_and_join_requests, so just remove the head
2045 * request from the inode / page_private pointer and
2047 nfs_inode_remove_request(req);
2048 nfs_unlock_and_release_request(req);
2055 * Write back all requests on one page - we do this before reading it.
2057 int nfs_wb_page(struct inode *inode, struct page *page)
2059 loff_t range_start = page_file_offset(page);
2060 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2061 struct writeback_control wbc = {
2062 .sync_mode = WB_SYNC_ALL,
2064 .range_start = range_start,
2065 .range_end = range_end,
2069 trace_nfs_writeback_page_enter(inode);
2072 wait_on_page_writeback(page);
2073 if (clear_page_dirty_for_io(page)) {
2074 ret = nfs_writepage_locked(page, &wbc);
2080 if (!PagePrivate(page))
2082 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2087 trace_nfs_writeback_page_exit(inode, ret);
2091 #ifdef CONFIG_MIGRATION
2092 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2093 struct page *page, enum migrate_mode mode)
2096 * If PagePrivate is set, then the page is currently associated with
2097 * an in-progress read or write request. Don't try to migrate it.
2099 * FIXME: we could do this in principle, but we'll need a way to ensure
2100 * that we can safely release the inode reference while holding
2103 if (PagePrivate(page))
2106 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2109 return migrate_page(mapping, newpage, page, mode);
2113 int __init nfs_init_writepagecache(void)
2115 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2116 sizeof(struct nfs_pgio_header),
2117 0, SLAB_HWCACHE_ALIGN,
2119 if (nfs_wdata_cachep == NULL)
2122 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2124 if (nfs_wdata_mempool == NULL)
2125 goto out_destroy_write_cache;
2127 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2128 sizeof(struct nfs_commit_data),
2129 0, SLAB_HWCACHE_ALIGN,
2131 if (nfs_cdata_cachep == NULL)
2132 goto out_destroy_write_mempool;
2134 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2136 if (nfs_commit_mempool == NULL)
2137 goto out_destroy_commit_cache;
2140 * NFS congestion size, scale with available memory.
2152 * This allows larger machines to have larger/more transfers.
2153 * Limit the default to 256M
2155 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2156 if (nfs_congestion_kb > 256*1024)
2157 nfs_congestion_kb = 256*1024;
2161 out_destroy_commit_cache:
2162 kmem_cache_destroy(nfs_cdata_cachep);
2163 out_destroy_write_mempool:
2164 mempool_destroy(nfs_wdata_mempool);
2165 out_destroy_write_cache:
2166 kmem_cache_destroy(nfs_wdata_cachep);
2170 void nfs_destroy_writepagecache(void)
2172 mempool_destroy(nfs_commit_mempool);
2173 kmem_cache_destroy(nfs_cdata_cachep);
2174 mempool_destroy(nfs_wdata_mempool);
2175 kmem_cache_destroy(nfs_wdata_cachep);
2178 static const struct nfs_rw_ops nfs_rw_write_ops = {
2179 .rw_alloc_header = nfs_writehdr_alloc,
2180 .rw_free_header = nfs_writehdr_free,
2181 .rw_done = nfs_writeback_done,
2182 .rw_result = nfs_writeback_result,
2183 .rw_initiate = nfs_initiate_write,