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 WARN_ON_ONCE(old_head != subreq->wb_head);
430 /* make sure old group is not used */
431 subreq->wb_this_page = subreq;
433 clear_bit(PG_REMOVE, &subreq->wb_flags);
435 /* Note: races with nfs_page_group_destroy() */
436 if (!kref_read(&subreq->wb_kref)) {
437 /* Check if we raced with nfs_page_group_destroy() */
438 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
439 nfs_free_request(subreq);
443 subreq->wb_head = subreq;
444 nfs_release_request(old_head);
446 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
447 nfs_release_request(subreq);
448 atomic_long_dec(&NFS_I(inode)->nrequests);
451 /* subreq is now totally disconnected from page group or any
452 * write / commit lists. last chance to wake any waiters */
453 nfs_unlock_and_release_request(subreq);
458 * nfs_lock_and_join_requests - join all subreqs to the head req and return
459 * a locked reference, cancelling any pending
460 * operations for this page.
462 * @page - the page used to lookup the "page group" of nfs_page structures
464 * This function joins all sub requests to the head request by first
465 * locking all requests in the group, cancelling any pending operations
466 * and finally updating the head request to cover the whole range covered by
467 * the (former) group. All subrequests are removed from any write or commit
468 * lists, unlinked from the group and destroyed.
470 * Returns a locked, referenced pointer to the head request - which after
471 * this call is guaranteed to be the only request associated with the page.
472 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
473 * error was encountered.
475 static struct nfs_page *
476 nfs_lock_and_join_requests(struct page *page)
478 struct inode *inode = page_file_mapping(page)->host;
479 struct nfs_page *head, *subreq;
480 struct nfs_page *destroy_list = NULL;
481 unsigned int total_bytes;
486 * A reference is taken only on the head request which acts as a
487 * reference to the whole page group - the group will not be destroyed
488 * until the head reference is released.
490 head = nfs_page_find_head_request(page);
494 /* lock the page head first in order to avoid an ABBA inefficiency */
495 if (!nfs_lock_request(head)) {
496 ret = nfs_wait_on_request(head);
497 nfs_release_request(head);
503 /* Ensure that nobody removed the request before we locked it */
504 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
505 nfs_unlock_and_release_request(head);
509 ret = nfs_page_group_lock(head);
511 goto release_request;
513 /* lock each request in the page group */
514 total_bytes = head->wb_bytes;
515 for (subreq = head->wb_this_page; subreq != head;
516 subreq = subreq->wb_this_page) {
518 if (!kref_get_unless_zero(&subreq->wb_kref)) {
519 if (subreq->wb_offset == head->wb_offset + total_bytes)
520 total_bytes += subreq->wb_bytes;
524 while (!nfs_lock_request(subreq)) {
526 * Unlock page to allow nfs_page_group_sync_on_bit()
529 nfs_page_group_unlock(head);
530 ret = nfs_wait_on_request(subreq);
532 ret = nfs_page_group_lock(head);
534 nfs_unroll_locks(inode, head, subreq);
535 nfs_release_request(subreq);
536 goto release_request;
540 * Subrequests are always contiguous, non overlapping
541 * and in order - but may be repeated (mirrored writes).
543 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
544 /* keep track of how many bytes this group covers */
545 total_bytes += subreq->wb_bytes;
546 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
547 ((subreq->wb_offset + subreq->wb_bytes) >
548 (head->wb_offset + total_bytes)))) {
549 nfs_page_group_unlock(head);
550 nfs_unroll_locks(inode, head, subreq);
551 nfs_unlock_and_release_request(subreq);
553 goto release_request;
557 /* Now that all requests are locked, make sure they aren't on any list.
558 * Commit list removal accounting is done after locks are dropped */
561 nfs_clear_request_commit(subreq);
562 subreq = subreq->wb_this_page;
563 } while (subreq != head);
565 /* unlink subrequests from head, destroy them later */
566 if (head->wb_this_page != head) {
567 /* destroy list will be terminated by head */
568 destroy_list = head->wb_this_page;
569 head->wb_this_page = head;
571 /* change head request to cover whole range that
572 * the former page group covered */
573 head->wb_bytes = total_bytes;
576 /* Postpone destruction of this request */
577 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
578 set_bit(PG_INODE_REF, &head->wb_flags);
579 kref_get(&head->wb_kref);
580 atomic_long_inc(&NFS_I(inode)->nrequests);
583 nfs_page_group_unlock(head);
585 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
587 /* Did we lose a race with nfs_inode_remove_request()? */
588 if (!(PagePrivate(page) || PageSwapCache(page))) {
589 nfs_unlock_and_release_request(head);
593 /* still holds ref on head from nfs_page_find_head_request
594 * and still has lock on head from lock loop */
598 nfs_unlock_and_release_request(head);
602 static void nfs_write_error(struct nfs_page *req, int error)
604 nfs_set_pageerror(page_file_mapping(req->wb_page));
605 nfs_mapping_set_error(req->wb_page, error);
606 nfs_inode_remove_request(req);
607 nfs_end_page_writeback(req);
608 nfs_release_request(req);
612 * Find an associated nfs write request, and prepare to flush it out
613 * May return an error if the user signalled nfs_wait_on_request().
615 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
618 struct nfs_page *req;
621 req = nfs_lock_and_join_requests(page);
628 nfs_set_page_writeback(page);
629 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
631 /* If there is a fatal error that covers this write, just exit */
632 ret = pgio->pg_error;
633 if (nfs_error_is_fatal_on_server(ret))
637 if (!nfs_pageio_add_request(pgio, req)) {
638 ret = pgio->pg_error;
640 * Remove the problematic req upon fatal errors on the server
642 if (nfs_error_is_fatal(ret)) {
643 if (nfs_error_is_fatal_on_server(ret))
647 nfs_redirty_request(req);
650 nfs_add_stats(page_file_mapping(page)->host,
651 NFSIOS_WRITEPAGES, 1);
655 nfs_write_error(req, ret);
659 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
660 struct nfs_pageio_descriptor *pgio)
664 nfs_pageio_cond_complete(pgio, page_index(page));
665 ret = nfs_page_async_flush(pgio, page);
666 if (ret == -EAGAIN) {
667 redirty_page_for_writepage(wbc, page);
668 ret = AOP_WRITEPAGE_ACTIVATE;
674 * Write an mmapped page to the server.
676 static int nfs_writepage_locked(struct page *page,
677 struct writeback_control *wbc)
679 struct nfs_pageio_descriptor pgio;
680 struct inode *inode = page_file_mapping(page)->host;
683 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
684 nfs_pageio_init_write(&pgio, inode, 0,
685 false, &nfs_async_write_completion_ops);
686 err = nfs_do_writepage(page, wbc, &pgio);
688 nfs_pageio_complete(&pgio);
691 if (nfs_error_is_fatal(pgio.pg_error))
692 return pgio.pg_error;
696 int nfs_writepage(struct page *page, struct writeback_control *wbc)
700 ret = nfs_writepage_locked(page, wbc);
701 if (ret != AOP_WRITEPAGE_ACTIVATE)
706 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
710 ret = nfs_do_writepage(page, wbc, data);
711 if (ret != AOP_WRITEPAGE_ACTIVATE)
716 static void nfs_io_completion_commit(void *inode)
718 nfs_commit_inode(inode, 0);
721 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
723 struct inode *inode = mapping->host;
724 struct nfs_pageio_descriptor pgio;
725 struct nfs_io_completion *ioc;
728 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
730 ioc = nfs_io_completion_alloc(GFP_KERNEL);
732 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
734 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
735 &nfs_async_write_completion_ops);
736 pgio.pg_io_completion = ioc;
737 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
739 nfs_pageio_complete(&pgio);
740 nfs_io_completion_put(ioc);
745 if (nfs_error_is_fatal(err))
753 * Insert a write request into an inode
755 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
757 struct address_space *mapping = page_file_mapping(req->wb_page);
758 struct nfs_inode *nfsi = NFS_I(inode);
760 WARN_ON_ONCE(req->wb_this_page != req);
762 /* Lock the request! */
763 nfs_lock_request(req);
766 * Swap-space should not get truncated. Hence no need to plug the race
767 * with invalidate/truncate.
769 spin_lock(&mapping->private_lock);
770 if (!nfs_have_writebacks(inode) &&
771 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
772 inode_inc_iversion_raw(inode);
773 if (likely(!PageSwapCache(req->wb_page))) {
774 set_bit(PG_MAPPED, &req->wb_flags);
775 SetPagePrivate(req->wb_page);
776 set_page_private(req->wb_page, (unsigned long)req);
778 spin_unlock(&mapping->private_lock);
779 atomic_long_inc(&nfsi->nrequests);
780 /* this a head request for a page group - mark it as having an
781 * extra reference so sub groups can follow suit.
782 * This flag also informs pgio layer when to bump nrequests when
783 * adding subrequests. */
784 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
785 kref_get(&req->wb_kref);
789 * Remove a write request from an inode
791 static void nfs_inode_remove_request(struct nfs_page *req)
793 struct address_space *mapping = page_file_mapping(req->wb_page);
794 struct inode *inode = mapping->host;
795 struct nfs_inode *nfsi = NFS_I(inode);
796 struct nfs_page *head;
798 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
801 spin_lock(&mapping->private_lock);
802 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
803 set_page_private(head->wb_page, 0);
804 ClearPagePrivate(head->wb_page);
805 clear_bit(PG_MAPPED, &head->wb_flags);
807 spin_unlock(&mapping->private_lock);
810 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
811 nfs_release_request(req);
812 atomic_long_dec(&nfsi->nrequests);
817 nfs_mark_request_dirty(struct nfs_page *req)
820 __set_page_dirty_nobuffers(req->wb_page);
824 * nfs_page_search_commits_for_head_request_locked
826 * Search through commit lists on @inode for the head request for @page.
827 * Must be called while holding the inode (which is cinfo) lock.
829 * Returns the head request if found, or NULL if not found.
831 static struct nfs_page *
832 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
835 struct nfs_page *freq, *t;
836 struct nfs_commit_info cinfo;
837 struct inode *inode = &nfsi->vfs_inode;
839 nfs_init_cinfo_from_inode(&cinfo, inode);
841 /* search through pnfs commit lists */
842 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
844 return freq->wb_head;
846 /* Linearly search the commit list for the correct request */
847 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
848 if (freq->wb_page == page)
849 return freq->wb_head;
856 * nfs_request_add_commit_list_locked - add request to a commit list
857 * @req: pointer to a struct nfs_page
858 * @dst: commit list head
859 * @cinfo: holds list lock and accounting info
861 * This sets the PG_CLEAN bit, updates the cinfo count of
862 * number of outstanding requests requiring a commit as well as
865 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
869 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
870 struct nfs_commit_info *cinfo)
872 set_bit(PG_CLEAN, &req->wb_flags);
873 nfs_list_add_request(req, dst);
874 atomic_long_inc(&cinfo->mds->ncommit);
876 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
879 * nfs_request_add_commit_list - add request to a commit list
880 * @req: pointer to a struct nfs_page
881 * @cinfo: holds list lock and accounting info
883 * This sets the PG_CLEAN bit, updates the cinfo count of
884 * number of outstanding requests requiring a commit as well as
887 * The caller must _not_ hold the cinfo->lock, but must be
888 * holding the nfs_page lock.
891 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
893 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
894 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
895 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
897 nfs_mark_page_unstable(req->wb_page, cinfo);
899 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
902 * nfs_request_remove_commit_list - Remove request from a commit list
903 * @req: pointer to a nfs_page
904 * @cinfo: holds list lock and accounting info
906 * This clears the PG_CLEAN bit, and updates the cinfo's count of
907 * number of outstanding requests requiring a commit
908 * It does not update the MM page stats.
910 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
913 nfs_request_remove_commit_list(struct nfs_page *req,
914 struct nfs_commit_info *cinfo)
916 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
918 nfs_list_remove_request(req);
919 atomic_long_dec(&cinfo->mds->ncommit);
921 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
923 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
926 cinfo->inode = inode;
927 cinfo->mds = &NFS_I(inode)->commit_info;
928 cinfo->ds = pnfs_get_ds_info(inode);
930 cinfo->completion_ops = &nfs_commit_completion_ops;
933 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
935 struct nfs_direct_req *dreq)
938 nfs_init_cinfo_from_dreq(cinfo, dreq);
940 nfs_init_cinfo_from_inode(cinfo, inode);
942 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
945 * Add a request to the inode's commit list.
948 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
949 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
951 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
953 nfs_request_add_commit_list(req, cinfo);
957 nfs_clear_page_commit(struct page *page)
959 dec_node_page_state(page, NR_UNSTABLE_NFS);
960 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
964 /* Called holding the request lock on @req */
966 nfs_clear_request_commit(struct nfs_page *req)
968 if (test_bit(PG_CLEAN, &req->wb_flags)) {
969 struct nfs_open_context *ctx = nfs_req_openctx(req);
970 struct inode *inode = d_inode(ctx->dentry);
971 struct nfs_commit_info cinfo;
973 nfs_init_cinfo_from_inode(&cinfo, inode);
974 mutex_lock(&NFS_I(inode)->commit_mutex);
975 if (!pnfs_clear_request_commit(req, &cinfo)) {
976 nfs_request_remove_commit_list(req, &cinfo);
978 mutex_unlock(&NFS_I(inode)->commit_mutex);
979 nfs_clear_page_commit(req->wb_page);
983 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
985 if (hdr->verf.committed == NFS_DATA_SYNC)
986 return hdr->lseg == NULL;
987 return hdr->verf.committed != NFS_FILE_SYNC;
990 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
992 nfs_io_completion_get(hdr->io_completion);
995 static void nfs_write_completion(struct nfs_pgio_header *hdr)
997 struct nfs_commit_info cinfo;
998 unsigned long bytes = 0;
1000 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1002 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1003 while (!list_empty(&hdr->pages)) {
1004 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1006 bytes += req->wb_bytes;
1007 nfs_list_remove_request(req);
1008 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1009 (hdr->good_bytes < bytes)) {
1010 nfs_set_pageerror(page_file_mapping(req->wb_page));
1011 nfs_mapping_set_error(req->wb_page, hdr->error);
1014 if (nfs_write_need_commit(hdr)) {
1015 /* Reset wb_nio, since the write was successful. */
1017 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1018 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1019 hdr->pgio_mirror_idx);
1023 nfs_inode_remove_request(req);
1025 nfs_end_page_writeback(req);
1026 nfs_release_request(req);
1029 nfs_io_completion_put(hdr->io_completion);
1034 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1036 return atomic_long_read(&cinfo->mds->ncommit);
1039 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1041 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1042 struct nfs_commit_info *cinfo, int max)
1044 struct nfs_page *req, *tmp;
1048 list_for_each_entry_safe(req, tmp, src, wb_list) {
1049 kref_get(&req->wb_kref);
1050 if (!nfs_lock_request(req)) {
1053 /* Prevent deadlock with nfs_lock_and_join_requests */
1054 if (!list_empty(dst)) {
1055 nfs_release_request(req);
1058 /* Ensure we make progress to prevent livelock */
1059 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1060 status = nfs_wait_on_request(req);
1061 nfs_release_request(req);
1062 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1067 nfs_request_remove_commit_list(req, cinfo);
1068 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1069 nfs_list_add_request(req, dst);
1071 if ((ret == max) && !cinfo->dreq)
1077 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1080 * nfs_scan_commit - Scan an inode for commit requests
1081 * @inode: NFS inode to scan
1082 * @dst: mds destination list
1083 * @cinfo: mds and ds lists of reqs ready to commit
1085 * Moves requests from the inode's 'commit' request list.
1086 * The requests are *not* checked to ensure that they form a contiguous set.
1089 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1090 struct nfs_commit_info *cinfo)
1094 if (!atomic_long_read(&cinfo->mds->ncommit))
1096 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1097 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1098 const int max = INT_MAX;
1100 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1102 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1104 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1109 * Search for an existing write request, and attempt to update
1110 * it to reflect a new dirty region on a given page.
1112 * If the attempt fails, then the existing request is flushed out
1115 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1117 unsigned int offset,
1120 struct nfs_page *req;
1125 end = offset + bytes;
1127 req = nfs_lock_and_join_requests(page);
1128 if (IS_ERR_OR_NULL(req))
1131 rqend = req->wb_offset + req->wb_bytes;
1133 * Tell the caller to flush out the request if
1134 * the offsets are non-contiguous.
1135 * Note: nfs_flush_incompatible() will already
1136 * have flushed out requests having wrong owners.
1138 if (offset > rqend || end < req->wb_offset)
1141 /* Okay, the request matches. Update the region */
1142 if (offset < req->wb_offset) {
1143 req->wb_offset = offset;
1144 req->wb_pgbase = offset;
1147 req->wb_bytes = end - req->wb_offset;
1149 req->wb_bytes = rqend - req->wb_offset;
1154 * Note: we mark the request dirty here because
1155 * nfs_lock_and_join_requests() cannot preserve
1156 * commit flags, so we have to replay the write.
1158 nfs_mark_request_dirty(req);
1159 nfs_unlock_and_release_request(req);
1160 error = nfs_wb_page(inode, page);
1161 return (error < 0) ? ERR_PTR(error) : NULL;
1165 * Try to update an existing write request, or create one if there is none.
1167 * Note: Should always be called with the Page Lock held to prevent races
1168 * if we have to add a new request. Also assumes that the caller has
1169 * already called nfs_flush_incompatible() if necessary.
1171 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1172 struct page *page, unsigned int offset, unsigned int bytes)
1174 struct inode *inode = page_file_mapping(page)->host;
1175 struct nfs_page *req;
1177 req = nfs_try_to_update_request(inode, page, offset, bytes);
1180 req = nfs_create_request(ctx, page, offset, bytes);
1183 nfs_inode_add_request(inode, req);
1188 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1189 unsigned int offset, unsigned int count)
1191 struct nfs_page *req;
1193 req = nfs_setup_write_request(ctx, page, offset, count);
1195 return PTR_ERR(req);
1196 /* Update file length */
1197 nfs_grow_file(page, offset, count);
1198 nfs_mark_uptodate(req);
1199 nfs_mark_request_dirty(req);
1200 nfs_unlock_and_release_request(req);
1204 int nfs_flush_incompatible(struct file *file, struct page *page)
1206 struct nfs_open_context *ctx = nfs_file_open_context(file);
1207 struct nfs_lock_context *l_ctx;
1208 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1209 struct nfs_page *req;
1210 int do_flush, status;
1212 * Look for a request corresponding to this page. If there
1213 * is one, and it belongs to another file, we flush it out
1214 * before we try to copy anything into the page. Do this
1215 * due to the lack of an ACCESS-type call in NFSv2.
1216 * Also do the same if we find a request from an existing
1220 req = nfs_page_find_head_request(page);
1223 l_ctx = req->wb_lock_context;
1224 do_flush = req->wb_page != page ||
1225 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1226 if (l_ctx && flctx &&
1227 !(list_empty_careful(&flctx->flc_posix) &&
1228 list_empty_careful(&flctx->flc_flock))) {
1229 do_flush |= l_ctx->lockowner != current->files;
1231 nfs_release_request(req);
1234 status = nfs_wb_page(page_file_mapping(page)->host, page);
1235 } while (status == 0);
1240 * Avoid buffered writes when a open context credential's key would
1243 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1245 * Return 0 and set a credential flag which triggers the inode to flush
1246 * and performs NFS_FILE_SYNC writes if the key will expired within
1247 * RPC_KEY_EXPIRE_TIMEO.
1250 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1252 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1254 if (nfs_ctx_key_to_expire(ctx, inode) &&
1256 /* Already expired! */
1262 * Test if the open context credential key is marked to expire soon.
1264 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1266 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1267 struct rpc_cred *cred = ctx->ll_cred;
1268 struct auth_cred acred = {
1272 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1274 ctx->ll_cred = NULL;
1278 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1279 if (!cred || IS_ERR(cred))
1281 ctx->ll_cred = cred;
1282 return !!(cred->cr_ops->crkey_timeout &&
1283 cred->cr_ops->crkey_timeout(cred));
1287 * If the page cache is marked as unsafe or invalid, then we can't rely on
1288 * the PageUptodate() flag. In this case, we will need to turn off
1289 * write optimisations that depend on the page contents being correct.
1291 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1293 struct nfs_inode *nfsi = NFS_I(inode);
1295 if (nfs_have_delegated_attributes(inode))
1297 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1300 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1303 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1305 return PageUptodate(page) != 0;
1309 is_whole_file_wrlock(struct file_lock *fl)
1311 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1312 fl->fl_type == F_WRLCK;
1315 /* If we know the page is up to date, and we're not using byte range locks (or
1316 * if we have the whole file locked for writing), it may be more efficient to
1317 * extend the write to cover the entire page in order to avoid fragmentation
1320 * If the file is opened for synchronous writes then we can just skip the rest
1323 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1326 struct file_lock_context *flctx = inode->i_flctx;
1327 struct file_lock *fl;
1329 if (file->f_flags & O_DSYNC)
1331 if (!nfs_write_pageuptodate(page, inode))
1333 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1335 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1336 list_empty_careful(&flctx->flc_posix)))
1339 /* Check to see if there are whole file write locks */
1341 spin_lock(&flctx->flc_lock);
1342 if (!list_empty(&flctx->flc_posix)) {
1343 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1345 if (is_whole_file_wrlock(fl))
1347 } else if (!list_empty(&flctx->flc_flock)) {
1348 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1350 if (fl->fl_type == F_WRLCK)
1353 spin_unlock(&flctx->flc_lock);
1358 * Update and possibly write a cached page of an NFS file.
1360 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1361 * things with a page scheduled for an RPC call (e.g. invalidate it).
1363 int nfs_updatepage(struct file *file, struct page *page,
1364 unsigned int offset, unsigned int count)
1366 struct nfs_open_context *ctx = nfs_file_open_context(file);
1367 struct address_space *mapping = page_file_mapping(page);
1368 struct inode *inode = mapping->host;
1371 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1373 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1374 file, count, (long long)(page_file_offset(page) + offset));
1379 if (nfs_can_extend_write(file, page, inode)) {
1380 count = max(count + offset, nfs_page_length(page));
1384 status = nfs_writepage_setup(ctx, page, offset, count);
1386 nfs_set_pageerror(mapping);
1388 __set_page_dirty_nobuffers(page);
1390 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1391 status, (long long)i_size_read(inode));
1395 static int flush_task_priority(int how)
1397 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1399 return RPC_PRIORITY_HIGH;
1401 return RPC_PRIORITY_LOW;
1403 return RPC_PRIORITY_NORMAL;
1406 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1407 struct rpc_message *msg,
1408 const struct nfs_rpc_ops *rpc_ops,
1409 struct rpc_task_setup *task_setup_data, int how)
1411 int priority = flush_task_priority(how);
1413 task_setup_data->priority = priority;
1414 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1415 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1419 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1420 * call this on each, which will prepare them to be retried on next
1421 * writeback using standard nfs.
1423 static void nfs_redirty_request(struct nfs_page *req)
1425 /* Bump the transmission count */
1427 nfs_mark_request_dirty(req);
1428 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1429 nfs_end_page_writeback(req);
1430 nfs_release_request(req);
1433 static void nfs_async_write_error(struct list_head *head, int error)
1435 struct nfs_page *req;
1437 while (!list_empty(head)) {
1438 req = nfs_list_entry(head->next);
1439 nfs_list_remove_request(req);
1440 if (nfs_error_is_fatal(error))
1441 nfs_write_error(req, error);
1443 nfs_redirty_request(req);
1447 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1449 nfs_async_write_error(&hdr->pages, 0);
1450 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1451 hdr->args.offset + hdr->args.count - 1);
1454 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1455 .init_hdr = nfs_async_write_init,
1456 .error_cleanup = nfs_async_write_error,
1457 .completion = nfs_write_completion,
1458 .reschedule_io = nfs_async_write_reschedule_io,
1461 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1462 struct inode *inode, int ioflags, bool force_mds,
1463 const struct nfs_pgio_completion_ops *compl_ops)
1465 struct nfs_server *server = NFS_SERVER(inode);
1466 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1468 #ifdef CONFIG_NFS_V4_1
1469 if (server->pnfs_curr_ld && !force_mds)
1470 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1472 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1473 server->wsize, ioflags);
1475 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1477 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1479 struct nfs_pgio_mirror *mirror;
1481 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1482 pgio->pg_ops->pg_cleanup(pgio);
1484 pgio->pg_ops = &nfs_pgio_rw_ops;
1486 nfs_pageio_stop_mirroring(pgio);
1488 mirror = &pgio->pg_mirrors[0];
1489 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1491 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1494 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1496 struct nfs_commit_data *data = calldata;
1498 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1502 * Special version of should_remove_suid() that ignores capabilities.
1504 static int nfs_should_remove_suid(const struct inode *inode)
1506 umode_t mode = inode->i_mode;
1509 /* suid always must be killed */
1510 if (unlikely(mode & S_ISUID))
1511 kill = ATTR_KILL_SUID;
1514 * sgid without any exec bits is just a mandatory locking mark; leave
1515 * it alone. If some exec bits are set, it's a real sgid; kill it.
1517 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1518 kill |= ATTR_KILL_SGID;
1520 if (unlikely(kill && S_ISREG(mode)))
1526 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1527 struct nfs_fattr *fattr)
1529 struct nfs_pgio_args *argp = &hdr->args;
1530 struct nfs_pgio_res *resp = &hdr->res;
1531 u64 size = argp->offset + resp->count;
1533 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1535 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1536 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1539 if (size != fattr->size)
1541 /* Set attribute barrier */
1542 nfs_fattr_set_barrier(fattr);
1543 /* ...and update size */
1544 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1547 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1549 struct nfs_fattr *fattr = &hdr->fattr;
1550 struct inode *inode = hdr->inode;
1552 spin_lock(&inode->i_lock);
1553 nfs_writeback_check_extend(hdr, fattr);
1554 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1555 spin_unlock(&inode->i_lock);
1557 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1560 * This function is called when the WRITE call is complete.
1562 static int nfs_writeback_done(struct rpc_task *task,
1563 struct nfs_pgio_header *hdr,
1564 struct inode *inode)
1569 * ->write_done will attempt to use post-op attributes to detect
1570 * conflicting writes by other clients. A strict interpretation
1571 * of close-to-open would allow us to continue caching even if
1572 * another writer had changed the file, but some applications
1573 * depend on tighter cache coherency when writing.
1575 status = NFS_PROTO(inode)->write_done(task, hdr);
1579 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1580 trace_nfs_writeback_done(inode, task->tk_status,
1581 hdr->args.offset, hdr->res.verf);
1583 if (hdr->res.verf->committed < hdr->args.stable &&
1584 task->tk_status >= 0) {
1585 /* We tried a write call, but the server did not
1586 * commit data to stable storage even though we
1588 * Note: There is a known bug in Tru64 < 5.0 in which
1589 * the server reports NFS_DATA_SYNC, but performs
1590 * NFS_FILE_SYNC. We therefore implement this checking
1591 * as a dprintk() in order to avoid filling syslog.
1593 static unsigned long complain;
1595 /* Note this will print the MDS for a DS write */
1596 if (time_before(complain, jiffies)) {
1597 dprintk("NFS: faulty NFS server %s:"
1598 " (committed = %d) != (stable = %d)\n",
1599 NFS_SERVER(inode)->nfs_client->cl_hostname,
1600 hdr->res.verf->committed, hdr->args.stable);
1601 complain = jiffies + 300 * HZ;
1605 /* Deal with the suid/sgid bit corner case */
1606 if (nfs_should_remove_suid(inode)) {
1607 spin_lock(&inode->i_lock);
1608 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1609 spin_unlock(&inode->i_lock);
1615 * This function is called when the WRITE call is complete.
1617 static void nfs_writeback_result(struct rpc_task *task,
1618 struct nfs_pgio_header *hdr)
1620 struct nfs_pgio_args *argp = &hdr->args;
1621 struct nfs_pgio_res *resp = &hdr->res;
1623 if (resp->count < argp->count) {
1624 static unsigned long complain;
1626 /* This a short write! */
1627 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1629 /* Has the server at least made some progress? */
1630 if (resp->count == 0) {
1631 if (time_before(complain, jiffies)) {
1633 "NFS: Server wrote zero bytes, expected %u.\n",
1635 complain = jiffies + 300 * HZ;
1637 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1638 task->tk_status = -EIO;
1642 /* For non rpc-based layout drivers, retry-through-MDS */
1643 if (!task->tk_ops) {
1644 hdr->pnfs_error = -EAGAIN;
1648 /* Was this an NFSv2 write or an NFSv3 stable write? */
1649 if (resp->verf->committed != NFS_UNSTABLE) {
1650 /* Resend from where the server left off */
1651 hdr->mds_offset += resp->count;
1652 argp->offset += resp->count;
1653 argp->pgbase += resp->count;
1654 argp->count -= resp->count;
1656 /* Resend as a stable write in order to avoid
1657 * headaches in the case of a server crash.
1659 argp->stable = NFS_FILE_SYNC;
1661 rpc_restart_call_prepare(task);
1665 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1667 return wait_var_event_killable(&cinfo->rpcs_out,
1668 !atomic_read(&cinfo->rpcs_out));
1671 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1673 atomic_inc(&cinfo->rpcs_out);
1676 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1678 if (atomic_dec_and_test(&cinfo->rpcs_out))
1679 wake_up_var(&cinfo->rpcs_out);
1682 void nfs_commitdata_release(struct nfs_commit_data *data)
1684 put_nfs_open_context(data->context);
1685 nfs_commit_free(data);
1687 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1689 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1690 const struct nfs_rpc_ops *nfs_ops,
1691 const struct rpc_call_ops *call_ops,
1694 struct rpc_task *task;
1695 int priority = flush_task_priority(how);
1696 struct rpc_message msg = {
1697 .rpc_argp = &data->args,
1698 .rpc_resp = &data->res,
1699 .rpc_cred = data->cred,
1701 struct rpc_task_setup task_setup_data = {
1702 .task = &data->task,
1704 .rpc_message = &msg,
1705 .callback_ops = call_ops,
1706 .callback_data = data,
1707 .workqueue = nfsiod_workqueue,
1708 .flags = RPC_TASK_ASYNC | flags,
1709 .priority = priority,
1711 /* Set up the initial task struct. */
1712 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1713 trace_nfs_initiate_commit(data);
1715 dprintk("NFS: initiated commit call\n");
1717 task = rpc_run_task(&task_setup_data);
1719 return PTR_ERR(task);
1720 if (how & FLUSH_SYNC)
1721 rpc_wait_for_completion_task(task);
1725 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1727 static loff_t nfs_get_lwb(struct list_head *head)
1730 struct nfs_page *req;
1732 list_for_each_entry(req, head, wb_list)
1733 if (lwb < (req_offset(req) + req->wb_bytes))
1734 lwb = req_offset(req) + req->wb_bytes;
1740 * Set up the argument/result storage required for the RPC call.
1742 void nfs_init_commit(struct nfs_commit_data *data,
1743 struct list_head *head,
1744 struct pnfs_layout_segment *lseg,
1745 struct nfs_commit_info *cinfo)
1747 struct nfs_page *first = nfs_list_entry(head->next);
1748 struct nfs_open_context *ctx = nfs_req_openctx(first);
1749 struct inode *inode = d_inode(ctx->dentry);
1751 /* Set up the RPC argument and reply structs
1752 * NB: take care not to mess about with data->commit et al. */
1754 list_splice_init(head, &data->pages);
1756 data->inode = inode;
1757 data->cred = ctx->cred;
1758 data->lseg = lseg; /* reference transferred */
1759 /* only set lwb for pnfs commit */
1761 data->lwb = nfs_get_lwb(&data->pages);
1762 data->mds_ops = &nfs_commit_ops;
1763 data->completion_ops = cinfo->completion_ops;
1764 data->dreq = cinfo->dreq;
1766 data->args.fh = NFS_FH(data->inode);
1767 /* Note: we always request a commit of the entire inode */
1768 data->args.offset = 0;
1769 data->args.count = 0;
1770 data->context = get_nfs_open_context(ctx);
1771 data->res.fattr = &data->fattr;
1772 data->res.verf = &data->verf;
1773 nfs_fattr_init(&data->fattr);
1775 EXPORT_SYMBOL_GPL(nfs_init_commit);
1777 void nfs_retry_commit(struct list_head *page_list,
1778 struct pnfs_layout_segment *lseg,
1779 struct nfs_commit_info *cinfo,
1782 struct nfs_page *req;
1784 while (!list_empty(page_list)) {
1785 req = nfs_list_entry(page_list->next);
1786 nfs_list_remove_request(req);
1787 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1789 nfs_clear_page_commit(req->wb_page);
1790 nfs_unlock_and_release_request(req);
1793 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1796 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1797 struct nfs_page *req)
1799 __set_page_dirty_nobuffers(req->wb_page);
1803 * Commit dirty pages
1806 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1807 struct nfs_commit_info *cinfo)
1809 struct nfs_commit_data *data;
1811 /* another commit raced with us */
1812 if (list_empty(head))
1815 data = nfs_commitdata_alloc(true);
1817 /* Set up the argument struct */
1818 nfs_init_commit(data, head, NULL, cinfo);
1819 atomic_inc(&cinfo->mds->rpcs_out);
1820 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1821 data->mds_ops, how, 0);
1825 * COMMIT call returned
1827 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1829 struct nfs_commit_data *data = calldata;
1831 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1832 task->tk_pid, task->tk_status);
1834 /* Call the NFS version-specific code */
1835 NFS_PROTO(data->inode)->commit_done(task, data);
1836 trace_nfs_commit_done(data);
1839 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1841 const struct nfs_writeverf *verf = data->res.verf;
1842 struct nfs_page *req;
1843 int status = data->task.tk_status;
1844 struct nfs_commit_info cinfo;
1845 struct nfs_server *nfss;
1847 while (!list_empty(&data->pages)) {
1848 req = nfs_list_entry(data->pages.next);
1849 nfs_list_remove_request(req);
1851 nfs_clear_page_commit(req->wb_page);
1853 dprintk("NFS: commit (%s/%llu %d@%lld)",
1854 nfs_req_openctx(req)->dentry->d_sb->s_id,
1855 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1857 (long long)req_offset(req));
1860 nfs_mapping_set_error(req->wb_page, status);
1861 nfs_inode_remove_request(req);
1863 dprintk_cont(", error = %d\n", status);
1867 /* Okay, COMMIT succeeded, apparently. Check the verifier
1868 * returned by the server against all stored verfs. */
1869 if (verf->committed > NFS_UNSTABLE &&
1870 !nfs_write_verifier_cmp(&req->wb_verf, &verf->verifier)) {
1871 /* We have a match */
1873 nfs_inode_remove_request(req);
1874 dprintk_cont(" OK\n");
1877 /* We have a mismatch. Write the page again */
1878 dprintk_cont(" mismatch\n");
1879 nfs_mark_request_dirty(req);
1880 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1882 nfs_unlock_and_release_request(req);
1883 /* Latency breaker */
1886 nfss = NFS_SERVER(data->inode);
1887 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1888 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1890 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1891 nfs_commit_end(cinfo.mds);
1894 static void nfs_commit_release(void *calldata)
1896 struct nfs_commit_data *data = calldata;
1898 data->completion_ops->completion(data);
1899 nfs_commitdata_release(calldata);
1902 static const struct rpc_call_ops nfs_commit_ops = {
1903 .rpc_call_prepare = nfs_commit_prepare,
1904 .rpc_call_done = nfs_commit_done,
1905 .rpc_release = nfs_commit_release,
1908 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1909 .completion = nfs_commit_release_pages,
1910 .resched_write = nfs_commit_resched_write,
1913 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1914 int how, struct nfs_commit_info *cinfo)
1918 status = pnfs_commit_list(inode, head, how, cinfo);
1919 if (status == PNFS_NOT_ATTEMPTED)
1920 status = nfs_commit_list(inode, head, how, cinfo);
1924 static int __nfs_commit_inode(struct inode *inode, int how,
1925 struct writeback_control *wbc)
1928 struct nfs_commit_info cinfo;
1929 int may_wait = how & FLUSH_SYNC;
1932 nfs_init_cinfo_from_inode(&cinfo, inode);
1933 nfs_commit_begin(cinfo.mds);
1935 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1938 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1942 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1943 if (nscan < wbc->nr_to_write)
1944 wbc->nr_to_write -= nscan;
1946 wbc->nr_to_write = 0;
1948 if (nscan < INT_MAX)
1952 nfs_commit_end(cinfo.mds);
1953 if (ret || !may_wait)
1955 return wait_on_commit(cinfo.mds);
1958 int nfs_commit_inode(struct inode *inode, int how)
1960 return __nfs_commit_inode(inode, how, NULL);
1962 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1964 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1966 struct nfs_inode *nfsi = NFS_I(inode);
1967 int flags = FLUSH_SYNC;
1970 if (wbc->sync_mode == WB_SYNC_NONE) {
1971 /* no commits means nothing needs to be done */
1972 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1973 goto check_requests_outstanding;
1975 /* Don't commit yet if this is a non-blocking flush and there
1976 * are a lot of outstanding writes for this mapping.
1978 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1979 goto out_mark_dirty;
1981 /* don't wait for the COMMIT response */
1985 ret = __nfs_commit_inode(inode, flags, wbc);
1987 if (flags & FLUSH_SYNC)
1989 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1990 goto out_mark_dirty;
1992 check_requests_outstanding:
1993 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1996 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1999 EXPORT_SYMBOL_GPL(nfs_write_inode);
2002 * Wrapper for filemap_write_and_wait_range()
2004 * Needed for pNFS in order to ensure data becomes visible to the
2007 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2008 loff_t lstart, loff_t lend)
2012 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2014 ret = pnfs_sync_inode(mapping->host, true);
2017 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2020 * flush the inode to disk.
2022 int nfs_wb_all(struct inode *inode)
2026 trace_nfs_writeback_inode_enter(inode);
2028 ret = filemap_write_and_wait(inode->i_mapping);
2031 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2034 pnfs_sync_inode(inode, true);
2038 trace_nfs_writeback_inode_exit(inode, ret);
2041 EXPORT_SYMBOL_GPL(nfs_wb_all);
2043 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2045 struct nfs_page *req;
2048 wait_on_page_writeback(page);
2050 /* blocking call to cancel all requests and join to a single (head)
2052 req = nfs_lock_and_join_requests(page);
2057 /* all requests from this page have been cancelled by
2058 * nfs_lock_and_join_requests, so just remove the head
2059 * request from the inode / page_private pointer and
2061 nfs_inode_remove_request(req);
2062 nfs_unlock_and_release_request(req);
2069 * Write back all requests on one page - we do this before reading it.
2071 int nfs_wb_page(struct inode *inode, struct page *page)
2073 loff_t range_start = page_file_offset(page);
2074 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2075 struct writeback_control wbc = {
2076 .sync_mode = WB_SYNC_ALL,
2078 .range_start = range_start,
2079 .range_end = range_end,
2083 trace_nfs_writeback_page_enter(inode);
2086 wait_on_page_writeback(page);
2087 if (clear_page_dirty_for_io(page)) {
2088 ret = nfs_writepage_locked(page, &wbc);
2094 if (!PagePrivate(page))
2096 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2101 trace_nfs_writeback_page_exit(inode, ret);
2105 #ifdef CONFIG_MIGRATION
2106 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2107 struct page *page, enum migrate_mode mode)
2110 * If PagePrivate is set, then the page is currently associated with
2111 * an in-progress read or write request. Don't try to migrate it.
2113 * FIXME: we could do this in principle, but we'll need a way to ensure
2114 * that we can safely release the inode reference while holding
2117 if (PagePrivate(page))
2120 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2123 return migrate_page(mapping, newpage, page, mode);
2127 int __init nfs_init_writepagecache(void)
2129 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2130 sizeof(struct nfs_pgio_header),
2131 0, SLAB_HWCACHE_ALIGN,
2133 if (nfs_wdata_cachep == NULL)
2136 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2138 if (nfs_wdata_mempool == NULL)
2139 goto out_destroy_write_cache;
2141 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2142 sizeof(struct nfs_commit_data),
2143 0, SLAB_HWCACHE_ALIGN,
2145 if (nfs_cdata_cachep == NULL)
2146 goto out_destroy_write_mempool;
2148 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2150 if (nfs_commit_mempool == NULL)
2151 goto out_destroy_commit_cache;
2154 * NFS congestion size, scale with available memory.
2166 * This allows larger machines to have larger/more transfers.
2167 * Limit the default to 256M
2169 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2170 if (nfs_congestion_kb > 256*1024)
2171 nfs_congestion_kb = 256*1024;
2175 out_destroy_commit_cache:
2176 kmem_cache_destroy(nfs_cdata_cachep);
2177 out_destroy_write_mempool:
2178 mempool_destroy(nfs_wdata_mempool);
2179 out_destroy_write_cache:
2180 kmem_cache_destroy(nfs_wdata_cachep);
2184 void nfs_destroy_writepagecache(void)
2186 mempool_destroy(nfs_commit_mempool);
2187 kmem_cache_destroy(nfs_cdata_cachep);
2188 mempool_destroy(nfs_wdata_mempool);
2189 kmem_cache_destroy(nfs_wdata_cachep);
2192 static const struct nfs_rw_ops nfs_rw_write_ops = {
2193 .rw_alloc_header = nfs_writehdr_alloc,
2194 .rw_free_header = nfs_writehdr_free,
2195 .rw_done = nfs_writeback_done,
2196 .rw_result = nfs_writeback_result,
2197 .rw_initiate = nfs_initiate_write,