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(void)
75 struct nfs_commit_data *p;
77 p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
79 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
82 memset(p, 0, sizeof(*p));
84 INIT_LIST_HEAD(&p->pages);
87 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
89 void nfs_commit_free(struct nfs_commit_data *p)
91 mempool_free(p, nfs_commit_mempool);
93 EXPORT_SYMBOL_GPL(nfs_commit_free);
95 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
97 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_KERNEL);
99 memset(p, 0, sizeof(*p));
100 p->rw_mode = FMODE_WRITE;
104 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
106 mempool_free(hdr, nfs_wdata_mempool);
109 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
111 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
114 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
115 void (*complete)(void *), void *data)
117 ioc->complete = complete;
119 kref_init(&ioc->refcount);
122 static void nfs_io_completion_release(struct kref *kref)
124 struct nfs_io_completion *ioc = container_of(kref,
125 struct nfs_io_completion, refcount);
126 ioc->complete(ioc->data);
130 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
133 kref_get(&ioc->refcount);
136 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
139 kref_put(&ioc->refcount, nfs_io_completion_release);
143 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
145 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
146 kref_get(&req->wb_kref);
147 atomic_long_inc(&NFS_I(inode)->nrequests);
152 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
156 if (!test_bit(PG_REMOVE, &req->wb_flags))
158 ret = nfs_page_group_lock(req);
161 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
162 nfs_page_set_inode_ref(req, inode);
163 nfs_page_group_unlock(req);
167 static struct nfs_page *
168 nfs_page_private_request(struct page *page)
170 if (!PagePrivate(page))
172 return (struct nfs_page *)page_private(page);
176 * nfs_page_find_head_request_locked - find head request associated with @page
178 * must be called while holding the inode lock.
180 * returns matching head request with reference held, or NULL if not found.
182 static struct nfs_page *
183 nfs_page_find_private_request(struct page *page)
185 struct address_space *mapping = page_file_mapping(page);
186 struct nfs_page *req;
188 if (!PagePrivate(page))
190 spin_lock(&mapping->private_lock);
191 req = nfs_page_private_request(page);
193 WARN_ON_ONCE(req->wb_head != req);
194 kref_get(&req->wb_kref);
196 spin_unlock(&mapping->private_lock);
200 static struct nfs_page *
201 nfs_page_find_swap_request(struct page *page)
203 struct inode *inode = page_file_mapping(page)->host;
204 struct nfs_inode *nfsi = NFS_I(inode);
205 struct nfs_page *req = NULL;
206 if (!PageSwapCache(page))
208 mutex_lock(&nfsi->commit_mutex);
209 if (PageSwapCache(page)) {
210 req = nfs_page_search_commits_for_head_request_locked(nfsi,
213 WARN_ON_ONCE(req->wb_head != req);
214 kref_get(&req->wb_kref);
217 mutex_unlock(&nfsi->commit_mutex);
222 * nfs_page_find_head_request - find head request associated with @page
224 * returns matching head request with reference held, or NULL if not found.
226 static struct nfs_page *nfs_page_find_head_request(struct page *page)
228 struct nfs_page *req;
230 req = nfs_page_find_private_request(page);
232 req = nfs_page_find_swap_request(page);
236 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
238 struct inode *inode = page_file_mapping(page)->host;
239 struct nfs_page *req, *head;
243 req = nfs_page_find_head_request(page);
246 head = nfs_page_group_lock_head(req);
248 nfs_release_request(req);
251 ret = nfs_cancel_remove_inode(head, inode);
253 nfs_unlock_and_release_request(head);
256 /* Ensure that nobody removed the request before we locked it */
257 if (head == nfs_page_private_request(page))
259 if (PageSwapCache(page))
261 nfs_unlock_and_release_request(head);
266 /* Adjust the file length if we're writing beyond the end */
267 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
269 struct inode *inode = page_file_mapping(page)->host;
273 spin_lock(&inode->i_lock);
274 i_size = i_size_read(inode);
275 end_index = (i_size - 1) >> PAGE_SHIFT;
276 if (i_size > 0 && page_index(page) < end_index)
278 end = page_file_offset(page) + ((loff_t)offset+count);
281 trace_nfs_size_grow(inode, end);
282 i_size_write(inode, end);
283 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
284 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
286 spin_unlock(&inode->i_lock);
287 nfs_fscache_invalidate(inode, 0);
290 /* A writeback failed: mark the page as bad, and invalidate the page cache */
291 static void nfs_set_pageerror(struct address_space *mapping)
293 struct inode *inode = mapping->host;
295 nfs_zap_mapping(mapping->host, mapping);
296 /* Force file size revalidation */
297 spin_lock(&inode->i_lock);
298 nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
299 NFS_INO_INVALID_CHANGE |
300 NFS_INO_INVALID_SIZE);
301 spin_unlock(&inode->i_lock);
304 static void nfs_mapping_set_error(struct page *page, int error)
306 struct address_space *mapping = page_file_mapping(page);
309 filemap_set_wb_err(mapping, error);
311 errseq_set(&mapping->host->i_sb->s_wb_err,
312 error == -ENOSPC ? -ENOSPC : -EIO);
313 nfs_set_pageerror(mapping);
317 * nfs_page_group_search_locked
318 * @head - head request of page group
319 * @page_offset - offset into page
321 * Search page group with head @head to find a request that contains the
322 * page offset @page_offset.
324 * Returns a pointer to the first matching nfs request, or NULL if no
327 * Must be called with the page group lock held
329 static struct nfs_page *
330 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
332 struct nfs_page *req;
336 if (page_offset >= req->wb_pgbase &&
337 page_offset < (req->wb_pgbase + req->wb_bytes))
340 req = req->wb_this_page;
341 } while (req != head);
347 * nfs_page_group_covers_page
348 * @head - head request of page group
350 * Return true if the page group with head @head covers the whole page,
351 * returns false otherwise
353 static bool nfs_page_group_covers_page(struct nfs_page *req)
355 struct nfs_page *tmp;
356 unsigned int pos = 0;
357 unsigned int len = nfs_page_length(req->wb_page);
359 nfs_page_group_lock(req);
362 tmp = nfs_page_group_search_locked(req->wb_head, pos);
365 pos = tmp->wb_pgbase + tmp->wb_bytes;
368 nfs_page_group_unlock(req);
372 /* We can set the PG_uptodate flag if we see that a write request
373 * covers the full page.
375 static void nfs_mark_uptodate(struct nfs_page *req)
377 if (PageUptodate(req->wb_page))
379 if (!nfs_page_group_covers_page(req))
381 SetPageUptodate(req->wb_page);
384 static int wb_priority(struct writeback_control *wbc)
388 if (wbc->sync_mode == WB_SYNC_ALL)
389 ret = FLUSH_COND_STABLE;
394 * NFS congestion control
397 int nfs_congestion_kb;
399 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
400 #define NFS_CONGESTION_OFF_THRESH \
401 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
403 static void nfs_set_page_writeback(struct page *page)
405 struct inode *inode = page_file_mapping(page)->host;
406 struct nfs_server *nfss = NFS_SERVER(inode);
407 int ret = test_set_page_writeback(page);
409 WARN_ON_ONCE(ret != 0);
411 if (atomic_long_inc_return(&nfss->writeback) >
412 NFS_CONGESTION_ON_THRESH)
413 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
416 static void nfs_end_page_writeback(struct nfs_page *req)
418 struct inode *inode = page_file_mapping(req->wb_page)->host;
419 struct nfs_server *nfss = NFS_SERVER(inode);
422 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
423 nfs_unlock_request(req);
427 end_page_writeback(req->wb_page);
428 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
429 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
433 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
435 * @destroy_list - request list (using wb_this_page) terminated by @old_head
436 * @old_head - the old head of the list
438 * All subrequests must be locked and removed from all lists, so at this point
439 * they are only "active" in this function, and possibly in nfs_wait_on_request
440 * with a reference held by some other context.
443 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
444 struct nfs_page *old_head,
447 while (destroy_list) {
448 struct nfs_page *subreq = destroy_list;
450 destroy_list = (subreq->wb_this_page == old_head) ?
451 NULL : subreq->wb_this_page;
453 /* Note: lock subreq in order to change subreq->wb_head */
454 nfs_page_set_headlock(subreq);
455 WARN_ON_ONCE(old_head != subreq->wb_head);
457 /* make sure old group is not used */
458 subreq->wb_this_page = subreq;
459 subreq->wb_head = subreq;
461 clear_bit(PG_REMOVE, &subreq->wb_flags);
463 /* Note: races with nfs_page_group_destroy() */
464 if (!kref_read(&subreq->wb_kref)) {
465 /* Check if we raced with nfs_page_group_destroy() */
466 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
467 nfs_page_clear_headlock(subreq);
468 nfs_free_request(subreq);
470 nfs_page_clear_headlock(subreq);
473 nfs_page_clear_headlock(subreq);
475 nfs_release_request(old_head);
477 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
478 nfs_release_request(subreq);
479 atomic_long_dec(&NFS_I(inode)->nrequests);
482 /* subreq is now totally disconnected from page group or any
483 * write / commit lists. last chance to wake any waiters */
484 nfs_unlock_and_release_request(subreq);
489 * nfs_join_page_group - destroy subrequests of the head req
490 * @head: the page used to lookup the "page group" of nfs_page structures
491 * @inode: Inode to which the request belongs.
493 * This function joins all sub requests to the head request by first
494 * locking all requests in the group, cancelling any pending operations
495 * and finally updating the head request to cover the whole range covered by
496 * the (former) group. All subrequests are removed from any write or commit
497 * lists, unlinked from the group and destroyed.
500 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
502 struct nfs_page *subreq;
503 struct nfs_page *destroy_list = NULL;
504 unsigned int pgbase, off, bytes;
506 pgbase = head->wb_pgbase;
507 bytes = head->wb_bytes;
508 off = head->wb_offset;
509 for (subreq = head->wb_this_page; subreq != head;
510 subreq = subreq->wb_this_page) {
511 /* Subrequests should always form a contiguous range */
512 if (pgbase > subreq->wb_pgbase) {
513 off -= pgbase - subreq->wb_pgbase;
514 bytes += pgbase - subreq->wb_pgbase;
515 pgbase = subreq->wb_pgbase;
517 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
521 /* Set the head request's range to cover the former page group */
522 head->wb_pgbase = pgbase;
523 head->wb_bytes = bytes;
524 head->wb_offset = off;
526 /* Now that all requests are locked, make sure they aren't on any list.
527 * Commit list removal accounting is done after locks are dropped */
530 nfs_clear_request_commit(subreq);
531 subreq = subreq->wb_this_page;
532 } while (subreq != head);
534 /* unlink subrequests from head, destroy them later */
535 if (head->wb_this_page != head) {
536 /* destroy list will be terminated by head */
537 destroy_list = head->wb_this_page;
538 head->wb_this_page = head;
541 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
545 * nfs_lock_and_join_requests - join all subreqs to the head req
546 * @page: the page used to lookup the "page group" of nfs_page structures
548 * This function joins all sub requests to the head request by first
549 * locking all requests in the group, cancelling any pending operations
550 * and finally updating the head request to cover the whole range covered by
551 * the (former) group. All subrequests are removed from any write or commit
552 * lists, unlinked from the group and destroyed.
554 * Returns a locked, referenced pointer to the head request - which after
555 * this call is guaranteed to be the only request associated with the page.
556 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
557 * error was encountered.
559 static struct nfs_page *
560 nfs_lock_and_join_requests(struct page *page)
562 struct inode *inode = page_file_mapping(page)->host;
563 struct nfs_page *head;
567 * A reference is taken only on the head request which acts as a
568 * reference to the whole page group - the group will not be destroyed
569 * until the head reference is released.
571 head = nfs_find_and_lock_page_request(page);
572 if (IS_ERR_OR_NULL(head))
575 /* lock each request in the page group */
576 ret = nfs_page_group_lock_subrequests(head);
578 nfs_unlock_and_release_request(head);
582 nfs_join_page_group(head, inode);
587 static void nfs_write_error(struct nfs_page *req, int error)
589 trace_nfs_write_error(req, error);
590 nfs_mapping_set_error(req->wb_page, error);
591 nfs_inode_remove_request(req);
592 nfs_end_page_writeback(req);
593 nfs_release_request(req);
597 * Find an associated nfs write request, and prepare to flush it out
598 * May return an error if the user signalled nfs_wait_on_request().
600 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
603 struct nfs_page *req;
606 req = nfs_lock_and_join_requests(page);
613 nfs_set_page_writeback(page);
614 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
616 /* If there is a fatal error that covers this write, just exit */
617 ret = pgio->pg_error;
618 if (nfs_error_is_fatal_on_server(ret))
622 if (!nfs_pageio_add_request(pgio, req)) {
623 ret = pgio->pg_error;
625 * Remove the problematic req upon fatal errors on the server
627 if (nfs_error_is_fatal(ret)) {
628 if (nfs_error_is_fatal_on_server(ret))
632 nfs_redirty_request(req);
635 nfs_add_stats(page_file_mapping(page)->host,
636 NFSIOS_WRITEPAGES, 1);
640 nfs_write_error(req, ret);
644 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
645 struct nfs_pageio_descriptor *pgio)
649 nfs_pageio_cond_complete(pgio, page_index(page));
650 ret = nfs_page_async_flush(pgio, page);
651 if (ret == -EAGAIN) {
652 redirty_page_for_writepage(wbc, page);
653 ret = AOP_WRITEPAGE_ACTIVATE;
659 * Write an mmapped page to the server.
661 static int nfs_writepage_locked(struct page *page,
662 struct writeback_control *wbc)
664 struct nfs_pageio_descriptor pgio;
665 struct inode *inode = page_file_mapping(page)->host;
668 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
669 nfs_pageio_init_write(&pgio, inode, 0,
670 false, &nfs_async_write_completion_ops);
671 err = nfs_do_writepage(page, wbc, &pgio);
673 nfs_pageio_complete(&pgio);
676 if (nfs_error_is_fatal(pgio.pg_error))
677 return pgio.pg_error;
681 int nfs_writepage(struct page *page, struct writeback_control *wbc)
685 ret = nfs_writepage_locked(page, wbc);
686 if (ret != AOP_WRITEPAGE_ACTIVATE)
691 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
695 ret = nfs_do_writepage(page, wbc, data);
696 if (ret != AOP_WRITEPAGE_ACTIVATE)
701 static void nfs_io_completion_commit(void *inode)
703 nfs_commit_inode(inode, 0);
706 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
708 struct inode *inode = mapping->host;
709 struct nfs_pageio_descriptor pgio;
710 struct nfs_io_completion *ioc = NULL;
711 unsigned int mntflags = NFS_SERVER(inode)->flags;
715 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
717 if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
718 wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
719 ioc = nfs_io_completion_alloc(GFP_KERNEL);
721 nfs_io_completion_init(ioc, nfs_io_completion_commit,
723 priority = wb_priority(wbc);
726 nfs_pageio_init_write(&pgio, inode, priority, false,
727 &nfs_async_write_completion_ops);
728 pgio.pg_io_completion = ioc;
729 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
731 nfs_pageio_complete(&pgio);
732 nfs_io_completion_put(ioc);
737 if (nfs_error_is_fatal(err))
745 * Insert a write request into an inode
747 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
749 struct address_space *mapping = page_file_mapping(req->wb_page);
750 struct nfs_inode *nfsi = NFS_I(inode);
752 WARN_ON_ONCE(req->wb_this_page != req);
754 /* Lock the request! */
755 nfs_lock_request(req);
758 * Swap-space should not get truncated. Hence no need to plug the race
759 * with invalidate/truncate.
761 spin_lock(&mapping->private_lock);
762 if (likely(!PageSwapCache(req->wb_page))) {
763 set_bit(PG_MAPPED, &req->wb_flags);
764 SetPagePrivate(req->wb_page);
765 set_page_private(req->wb_page, (unsigned long)req);
767 spin_unlock(&mapping->private_lock);
768 atomic_long_inc(&nfsi->nrequests);
769 /* this a head request for a page group - mark it as having an
770 * extra reference so sub groups can follow suit.
771 * This flag also informs pgio layer when to bump nrequests when
772 * adding subrequests. */
773 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
774 kref_get(&req->wb_kref);
778 * Remove a write request from an inode
780 static void nfs_inode_remove_request(struct nfs_page *req)
782 struct address_space *mapping = page_file_mapping(req->wb_page);
783 struct inode *inode = mapping->host;
784 struct nfs_inode *nfsi = NFS_I(inode);
785 struct nfs_page *head;
787 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
790 spin_lock(&mapping->private_lock);
791 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
792 set_page_private(head->wb_page, 0);
793 ClearPagePrivate(head->wb_page);
794 clear_bit(PG_MAPPED, &head->wb_flags);
796 spin_unlock(&mapping->private_lock);
799 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
800 nfs_release_request(req);
801 atomic_long_dec(&nfsi->nrequests);
806 nfs_mark_request_dirty(struct nfs_page *req)
809 __set_page_dirty_nobuffers(req->wb_page);
813 * nfs_page_search_commits_for_head_request_locked
815 * Search through commit lists on @inode for the head request for @page.
816 * Must be called while holding the inode (which is cinfo) lock.
818 * Returns the head request if found, or NULL if not found.
820 static struct nfs_page *
821 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
824 struct nfs_page *freq, *t;
825 struct nfs_commit_info cinfo;
826 struct inode *inode = &nfsi->vfs_inode;
828 nfs_init_cinfo_from_inode(&cinfo, inode);
830 /* search through pnfs commit lists */
831 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
833 return freq->wb_head;
835 /* Linearly search the commit list for the correct request */
836 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
837 if (freq->wb_page == page)
838 return freq->wb_head;
845 * nfs_request_add_commit_list_locked - add request to a commit list
846 * @req: pointer to a struct nfs_page
847 * @dst: commit list head
848 * @cinfo: holds list lock and accounting info
850 * This sets the PG_CLEAN bit, updates the cinfo count of
851 * number of outstanding requests requiring a commit as well as
854 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
858 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
859 struct nfs_commit_info *cinfo)
861 set_bit(PG_CLEAN, &req->wb_flags);
862 nfs_list_add_request(req, dst);
863 atomic_long_inc(&cinfo->mds->ncommit);
865 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
868 * nfs_request_add_commit_list - add request to a commit list
869 * @req: pointer to a struct nfs_page
870 * @cinfo: holds list lock and accounting info
872 * This sets the PG_CLEAN bit, updates the cinfo count of
873 * number of outstanding requests requiring a commit as well as
876 * The caller must _not_ hold the cinfo->lock, but must be
877 * holding the nfs_page lock.
880 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
882 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
883 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
884 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
886 nfs_mark_page_unstable(req->wb_page, cinfo);
888 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
891 * nfs_request_remove_commit_list - Remove request from a commit list
892 * @req: pointer to a nfs_page
893 * @cinfo: holds list lock and accounting info
895 * This clears the PG_CLEAN bit, and updates the cinfo's count of
896 * number of outstanding requests requiring a commit
897 * It does not update the MM page stats.
899 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
902 nfs_request_remove_commit_list(struct nfs_page *req,
903 struct nfs_commit_info *cinfo)
905 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
907 nfs_list_remove_request(req);
908 atomic_long_dec(&cinfo->mds->ncommit);
910 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
912 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
915 cinfo->inode = inode;
916 cinfo->mds = &NFS_I(inode)->commit_info;
917 cinfo->ds = pnfs_get_ds_info(inode);
919 cinfo->completion_ops = &nfs_commit_completion_ops;
922 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
924 struct nfs_direct_req *dreq)
927 nfs_init_cinfo_from_dreq(cinfo, dreq);
929 nfs_init_cinfo_from_inode(cinfo, inode);
931 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
934 * Add a request to the inode's commit list.
937 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
938 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
940 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
942 nfs_request_add_commit_list(req, cinfo);
946 nfs_clear_page_commit(struct page *page)
948 dec_node_page_state(page, NR_WRITEBACK);
949 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
953 /* Called holding the request lock on @req */
955 nfs_clear_request_commit(struct nfs_page *req)
957 if (test_bit(PG_CLEAN, &req->wb_flags)) {
958 struct nfs_open_context *ctx = nfs_req_openctx(req);
959 struct inode *inode = d_inode(ctx->dentry);
960 struct nfs_commit_info cinfo;
962 nfs_init_cinfo_from_inode(&cinfo, inode);
963 mutex_lock(&NFS_I(inode)->commit_mutex);
964 if (!pnfs_clear_request_commit(req, &cinfo)) {
965 nfs_request_remove_commit_list(req, &cinfo);
967 mutex_unlock(&NFS_I(inode)->commit_mutex);
968 nfs_clear_page_commit(req->wb_page);
972 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
974 if (hdr->verf.committed == NFS_DATA_SYNC)
975 return hdr->lseg == NULL;
976 return hdr->verf.committed != NFS_FILE_SYNC;
979 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
981 nfs_io_completion_get(hdr->io_completion);
984 static void nfs_write_completion(struct nfs_pgio_header *hdr)
986 struct nfs_commit_info cinfo;
987 unsigned long bytes = 0;
989 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
991 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
992 while (!list_empty(&hdr->pages)) {
993 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
995 bytes += req->wb_bytes;
996 nfs_list_remove_request(req);
997 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
998 (hdr->good_bytes < bytes)) {
999 trace_nfs_comp_error(req, hdr->error);
1000 nfs_mapping_set_error(req->wb_page, hdr->error);
1003 if (nfs_write_need_commit(hdr)) {
1004 /* Reset wb_nio, since the write was successful. */
1006 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1007 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1008 hdr->pgio_mirror_idx);
1012 nfs_inode_remove_request(req);
1014 nfs_end_page_writeback(req);
1015 nfs_release_request(req);
1018 nfs_io_completion_put(hdr->io_completion);
1023 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1025 return atomic_long_read(&cinfo->mds->ncommit);
1028 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1030 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1031 struct nfs_commit_info *cinfo, int max)
1033 struct nfs_page *req, *tmp;
1036 list_for_each_entry_safe(req, tmp, src, wb_list) {
1037 kref_get(&req->wb_kref);
1038 if (!nfs_lock_request(req)) {
1039 nfs_release_request(req);
1042 nfs_request_remove_commit_list(req, cinfo);
1043 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1044 nfs_list_add_request(req, dst);
1046 if ((ret == max) && !cinfo->dreq)
1052 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1055 * nfs_scan_commit - Scan an inode for commit requests
1056 * @inode: NFS inode to scan
1057 * @dst: mds destination list
1058 * @cinfo: mds and ds lists of reqs ready to commit
1060 * Moves requests from the inode's 'commit' request list.
1061 * The requests are *not* checked to ensure that they form a contiguous set.
1064 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1065 struct nfs_commit_info *cinfo)
1069 if (!atomic_long_read(&cinfo->mds->ncommit))
1071 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1072 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1073 const int max = INT_MAX;
1075 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1077 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1079 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1084 * Search for an existing write request, and attempt to update
1085 * it to reflect a new dirty region on a given page.
1087 * If the attempt fails, then the existing request is flushed out
1090 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1092 unsigned int offset,
1095 struct nfs_page *req;
1100 end = offset + bytes;
1102 req = nfs_lock_and_join_requests(page);
1103 if (IS_ERR_OR_NULL(req))
1106 rqend = req->wb_offset + req->wb_bytes;
1108 * Tell the caller to flush out the request if
1109 * the offsets are non-contiguous.
1110 * Note: nfs_flush_incompatible() will already
1111 * have flushed out requests having wrong owners.
1113 if (offset > rqend || end < req->wb_offset)
1116 /* Okay, the request matches. Update the region */
1117 if (offset < req->wb_offset) {
1118 req->wb_offset = offset;
1119 req->wb_pgbase = offset;
1122 req->wb_bytes = end - req->wb_offset;
1124 req->wb_bytes = rqend - req->wb_offset;
1129 * Note: we mark the request dirty here because
1130 * nfs_lock_and_join_requests() cannot preserve
1131 * commit flags, so we have to replay the write.
1133 nfs_mark_request_dirty(req);
1134 nfs_unlock_and_release_request(req);
1135 error = nfs_wb_page(inode, page);
1136 return (error < 0) ? ERR_PTR(error) : NULL;
1140 * Try to update an existing write request, or create one if there is none.
1142 * Note: Should always be called with the Page Lock held to prevent races
1143 * if we have to add a new request. Also assumes that the caller has
1144 * already called nfs_flush_incompatible() if necessary.
1146 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1147 struct page *page, unsigned int offset, unsigned int bytes)
1149 struct inode *inode = page_file_mapping(page)->host;
1150 struct nfs_page *req;
1152 req = nfs_try_to_update_request(inode, page, offset, bytes);
1155 req = nfs_create_request(ctx, page, offset, bytes);
1158 nfs_inode_add_request(inode, req);
1163 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1164 unsigned int offset, unsigned int count)
1166 struct nfs_page *req;
1168 req = nfs_setup_write_request(ctx, page, offset, count);
1170 return PTR_ERR(req);
1171 /* Update file length */
1172 nfs_grow_file(page, offset, count);
1173 nfs_mark_uptodate(req);
1174 nfs_mark_request_dirty(req);
1175 nfs_unlock_and_release_request(req);
1179 int nfs_flush_incompatible(struct file *file, struct page *page)
1181 struct nfs_open_context *ctx = nfs_file_open_context(file);
1182 struct nfs_lock_context *l_ctx;
1183 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1184 struct nfs_page *req;
1185 int do_flush, status;
1187 * Look for a request corresponding to this page. If there
1188 * is one, and it belongs to another file, we flush it out
1189 * before we try to copy anything into the page. Do this
1190 * due to the lack of an ACCESS-type call in NFSv2.
1191 * Also do the same if we find a request from an existing
1195 req = nfs_page_find_head_request(page);
1198 l_ctx = req->wb_lock_context;
1199 do_flush = req->wb_page != page ||
1200 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1201 if (l_ctx && flctx &&
1202 !(list_empty_careful(&flctx->flc_posix) &&
1203 list_empty_careful(&flctx->flc_flock))) {
1204 do_flush |= l_ctx->lockowner != current->files;
1206 nfs_release_request(req);
1209 status = nfs_wb_page(page_file_mapping(page)->host, page);
1210 } while (status == 0);
1215 * Avoid buffered writes when a open context credential's key would
1218 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1220 * Return 0 and set a credential flag which triggers the inode to flush
1221 * and performs NFS_FILE_SYNC writes if the key will expired within
1222 * RPC_KEY_EXPIRE_TIMEO.
1225 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1227 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1229 if (nfs_ctx_key_to_expire(ctx, inode) &&
1230 !rcu_access_pointer(ctx->ll_cred))
1231 /* Already expired! */
1237 * Test if the open context credential key is marked to expire soon.
1239 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1241 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1242 struct rpc_cred *cred, *new, *old = NULL;
1243 struct auth_cred acred = {
1249 cred = rcu_dereference(ctx->ll_cred);
1250 if (cred && !(cred->cr_ops->crkey_timeout &&
1251 cred->cr_ops->crkey_timeout(cred)))
1255 new = auth->au_ops->lookup_cred(auth, &acred, 0);
1260 if (IS_ERR_OR_NULL(new)) {
1263 } else if (new->cr_ops->crkey_timeout &&
1264 new->cr_ops->crkey_timeout(new))
1268 old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1269 RCU_INITIALIZER(new)), 1);
1277 * If the page cache is marked as unsafe or invalid, then we can't rely on
1278 * the PageUptodate() flag. In this case, we will need to turn off
1279 * write optimisations that depend on the page contents being correct.
1281 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1282 unsigned int pagelen)
1284 struct nfs_inode *nfsi = NFS_I(inode);
1286 if (nfs_have_delegated_attributes(inode))
1288 if (nfsi->cache_validity &
1289 (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1292 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1295 if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1297 return PageUptodate(page) != 0;
1301 is_whole_file_wrlock(struct file_lock *fl)
1303 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1304 fl->fl_type == F_WRLCK;
1307 /* If we know the page is up to date, and we're not using byte range locks (or
1308 * if we have the whole file locked for writing), it may be more efficient to
1309 * extend the write to cover the entire page in order to avoid fragmentation
1312 * If the file is opened for synchronous writes then we can just skip the rest
1315 static int nfs_can_extend_write(struct file *file, struct page *page,
1316 struct inode *inode, unsigned int pagelen)
1319 struct file_lock_context *flctx = inode->i_flctx;
1320 struct file_lock *fl;
1322 if (file->f_flags & O_DSYNC)
1324 if (!nfs_write_pageuptodate(page, inode, pagelen))
1326 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1328 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1329 list_empty_careful(&flctx->flc_posix)))
1332 /* Check to see if there are whole file write locks */
1334 spin_lock(&flctx->flc_lock);
1335 if (!list_empty(&flctx->flc_posix)) {
1336 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1338 if (is_whole_file_wrlock(fl))
1340 } else if (!list_empty(&flctx->flc_flock)) {
1341 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1343 if (fl->fl_type == F_WRLCK)
1346 spin_unlock(&flctx->flc_lock);
1351 * Update and possibly write a cached page of an NFS file.
1353 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1354 * things with a page scheduled for an RPC call (e.g. invalidate it).
1356 int nfs_updatepage(struct file *file, struct page *page,
1357 unsigned int offset, unsigned int count)
1359 struct nfs_open_context *ctx = nfs_file_open_context(file);
1360 struct address_space *mapping = page_file_mapping(page);
1361 struct inode *inode = mapping->host;
1362 unsigned int pagelen = nfs_page_length(page);
1365 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1367 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1368 file, count, (long long)(page_file_offset(page) + offset));
1373 if (nfs_can_extend_write(file, page, inode, pagelen)) {
1374 count = max(count + offset, pagelen);
1378 status = nfs_writepage_setup(ctx, page, offset, count);
1380 nfs_set_pageerror(mapping);
1382 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1383 status, (long long)i_size_read(inode));
1387 static int flush_task_priority(int how)
1389 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1391 return RPC_PRIORITY_HIGH;
1393 return RPC_PRIORITY_LOW;
1395 return RPC_PRIORITY_NORMAL;
1398 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1399 struct rpc_message *msg,
1400 const struct nfs_rpc_ops *rpc_ops,
1401 struct rpc_task_setup *task_setup_data, int how)
1403 int priority = flush_task_priority(how);
1405 if (IS_SWAPFILE(hdr->inode))
1406 task_setup_data->flags |= RPC_TASK_SWAPPER;
1407 task_setup_data->priority = priority;
1408 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1409 trace_nfs_initiate_write(hdr);
1412 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1413 * call this on each, which will prepare them to be retried on next
1414 * writeback using standard nfs.
1416 static void nfs_redirty_request(struct nfs_page *req)
1418 /* Bump the transmission count */
1420 nfs_mark_request_dirty(req);
1421 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1422 nfs_end_page_writeback(req);
1423 nfs_release_request(req);
1426 static void nfs_async_write_error(struct list_head *head, int error)
1428 struct nfs_page *req;
1430 while (!list_empty(head)) {
1431 req = nfs_list_entry(head->next);
1432 nfs_list_remove_request(req);
1433 if (nfs_error_is_fatal(error))
1434 nfs_write_error(req, error);
1436 nfs_redirty_request(req);
1440 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1442 nfs_async_write_error(&hdr->pages, 0);
1443 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1444 hdr->args.offset + hdr->args.count - 1);
1447 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1448 .init_hdr = nfs_async_write_init,
1449 .error_cleanup = nfs_async_write_error,
1450 .completion = nfs_write_completion,
1451 .reschedule_io = nfs_async_write_reschedule_io,
1454 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1455 struct inode *inode, int ioflags, bool force_mds,
1456 const struct nfs_pgio_completion_ops *compl_ops)
1458 struct nfs_server *server = NFS_SERVER(inode);
1459 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1461 #ifdef CONFIG_NFS_V4_1
1462 if (server->pnfs_curr_ld && !force_mds)
1463 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1465 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1466 server->wsize, ioflags);
1468 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1470 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1472 struct nfs_pgio_mirror *mirror;
1474 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1475 pgio->pg_ops->pg_cleanup(pgio);
1477 pgio->pg_ops = &nfs_pgio_rw_ops;
1479 nfs_pageio_stop_mirroring(pgio);
1481 mirror = &pgio->pg_mirrors[0];
1482 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1484 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1487 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1489 struct nfs_commit_data *data = calldata;
1491 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1495 * Special version of should_remove_suid() that ignores capabilities.
1497 static int nfs_should_remove_suid(const struct inode *inode)
1499 umode_t mode = inode->i_mode;
1502 /* suid always must be killed */
1503 if (unlikely(mode & S_ISUID))
1504 kill = ATTR_KILL_SUID;
1507 * sgid without any exec bits is just a mandatory locking mark; leave
1508 * it alone. If some exec bits are set, it's a real sgid; kill it.
1510 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1511 kill |= ATTR_KILL_SGID;
1513 if (unlikely(kill && S_ISREG(mode)))
1519 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1520 struct nfs_fattr *fattr)
1522 struct nfs_pgio_args *argp = &hdr->args;
1523 struct nfs_pgio_res *resp = &hdr->res;
1524 u64 size = argp->offset + resp->count;
1526 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1528 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1529 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1532 if (size != fattr->size)
1534 /* Set attribute barrier */
1535 nfs_fattr_set_barrier(fattr);
1536 /* ...and update size */
1537 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1540 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1542 struct nfs_fattr *fattr = &hdr->fattr;
1543 struct inode *inode = hdr->inode;
1545 spin_lock(&inode->i_lock);
1546 nfs_writeback_check_extend(hdr, fattr);
1547 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1548 spin_unlock(&inode->i_lock);
1550 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1553 * This function is called when the WRITE call is complete.
1555 static int nfs_writeback_done(struct rpc_task *task,
1556 struct nfs_pgio_header *hdr,
1557 struct inode *inode)
1562 * ->write_done will attempt to use post-op attributes to detect
1563 * conflicting writes by other clients. A strict interpretation
1564 * of close-to-open would allow us to continue caching even if
1565 * another writer had changed the file, but some applications
1566 * depend on tighter cache coherency when writing.
1568 status = NFS_PROTO(inode)->write_done(task, hdr);
1572 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1573 trace_nfs_writeback_done(task, hdr);
1575 if (hdr->res.verf->committed < hdr->args.stable &&
1576 task->tk_status >= 0) {
1577 /* We tried a write call, but the server did not
1578 * commit data to stable storage even though we
1580 * Note: There is a known bug in Tru64 < 5.0 in which
1581 * the server reports NFS_DATA_SYNC, but performs
1582 * NFS_FILE_SYNC. We therefore implement this checking
1583 * as a dprintk() in order to avoid filling syslog.
1585 static unsigned long complain;
1587 /* Note this will print the MDS for a DS write */
1588 if (time_before(complain, jiffies)) {
1589 dprintk("NFS: faulty NFS server %s:"
1590 " (committed = %d) != (stable = %d)\n",
1591 NFS_SERVER(inode)->nfs_client->cl_hostname,
1592 hdr->res.verf->committed, hdr->args.stable);
1593 complain = jiffies + 300 * HZ;
1597 /* Deal with the suid/sgid bit corner case */
1598 if (nfs_should_remove_suid(inode)) {
1599 spin_lock(&inode->i_lock);
1600 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1601 spin_unlock(&inode->i_lock);
1607 * This function is called when the WRITE call is complete.
1609 static void nfs_writeback_result(struct rpc_task *task,
1610 struct nfs_pgio_header *hdr)
1612 struct nfs_pgio_args *argp = &hdr->args;
1613 struct nfs_pgio_res *resp = &hdr->res;
1615 if (resp->count < argp->count) {
1616 static unsigned long complain;
1618 /* This a short write! */
1619 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1621 /* Has the server at least made some progress? */
1622 if (resp->count == 0) {
1623 if (time_before(complain, jiffies)) {
1625 "NFS: Server wrote zero bytes, expected %u.\n",
1627 complain = jiffies + 300 * HZ;
1629 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1630 task->tk_status = -EIO;
1634 /* For non rpc-based layout drivers, retry-through-MDS */
1635 if (!task->tk_ops) {
1636 hdr->pnfs_error = -EAGAIN;
1640 /* Was this an NFSv2 write or an NFSv3 stable write? */
1641 if (resp->verf->committed != NFS_UNSTABLE) {
1642 /* Resend from where the server left off */
1643 hdr->mds_offset += resp->count;
1644 argp->offset += resp->count;
1645 argp->pgbase += resp->count;
1646 argp->count -= resp->count;
1648 /* Resend as a stable write in order to avoid
1649 * headaches in the case of a server crash.
1651 argp->stable = NFS_FILE_SYNC;
1654 resp->verf->committed = 0;
1655 rpc_restart_call_prepare(task);
1659 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1661 return wait_var_event_killable(&cinfo->rpcs_out,
1662 !atomic_read(&cinfo->rpcs_out));
1665 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1667 atomic_inc(&cinfo->rpcs_out);
1670 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1672 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1673 wake_up_var(&cinfo->rpcs_out);
1679 void nfs_commitdata_release(struct nfs_commit_data *data)
1681 put_nfs_open_context(data->context);
1682 nfs_commit_free(data);
1684 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1686 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1687 const struct nfs_rpc_ops *nfs_ops,
1688 const struct rpc_call_ops *call_ops,
1691 struct rpc_task *task;
1692 int priority = flush_task_priority(how);
1693 struct rpc_message msg = {
1694 .rpc_argp = &data->args,
1695 .rpc_resp = &data->res,
1696 .rpc_cred = data->cred,
1698 struct rpc_task_setup task_setup_data = {
1699 .task = &data->task,
1701 .rpc_message = &msg,
1702 .callback_ops = call_ops,
1703 .callback_data = data,
1704 .workqueue = nfsiod_workqueue,
1705 .flags = RPC_TASK_ASYNC | flags,
1706 .priority = priority,
1708 /* Set up the initial task struct. */
1709 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1710 trace_nfs_initiate_commit(data);
1712 dprintk("NFS: initiated commit call\n");
1714 task = rpc_run_task(&task_setup_data);
1716 return PTR_ERR(task);
1717 if (how & FLUSH_SYNC)
1718 rpc_wait_for_completion_task(task);
1722 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1724 static loff_t nfs_get_lwb(struct list_head *head)
1727 struct nfs_page *req;
1729 list_for_each_entry(req, head, wb_list)
1730 if (lwb < (req_offset(req) + req->wb_bytes))
1731 lwb = req_offset(req) + req->wb_bytes;
1737 * Set up the argument/result storage required for the RPC call.
1739 void nfs_init_commit(struct nfs_commit_data *data,
1740 struct list_head *head,
1741 struct pnfs_layout_segment *lseg,
1742 struct nfs_commit_info *cinfo)
1744 struct nfs_page *first;
1745 struct nfs_open_context *ctx;
1746 struct inode *inode;
1748 /* Set up the RPC argument and reply structs
1749 * NB: take care not to mess about with data->commit et al. */
1752 list_splice_init(head, &data->pages);
1754 first = nfs_list_entry(data->pages.next);
1755 ctx = nfs_req_openctx(first);
1756 inode = d_inode(ctx->dentry);
1758 data->inode = inode;
1759 data->cred = ctx->cred;
1760 data->lseg = lseg; /* reference transferred */
1761 /* only set lwb for pnfs commit */
1763 data->lwb = nfs_get_lwb(&data->pages);
1764 data->mds_ops = &nfs_commit_ops;
1765 data->completion_ops = cinfo->completion_ops;
1766 data->dreq = cinfo->dreq;
1768 data->args.fh = NFS_FH(data->inode);
1769 /* Note: we always request a commit of the entire inode */
1770 data->args.offset = 0;
1771 data->args.count = 0;
1772 data->context = get_nfs_open_context(ctx);
1773 data->res.fattr = &data->fattr;
1774 data->res.verf = &data->verf;
1775 nfs_fattr_init(&data->fattr);
1776 nfs_commit_begin(cinfo->mds);
1778 EXPORT_SYMBOL_GPL(nfs_init_commit);
1780 void nfs_retry_commit(struct list_head *page_list,
1781 struct pnfs_layout_segment *lseg,
1782 struct nfs_commit_info *cinfo,
1785 struct nfs_page *req;
1787 while (!list_empty(page_list)) {
1788 req = nfs_list_entry(page_list->next);
1789 nfs_list_remove_request(req);
1790 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1792 nfs_clear_page_commit(req->wb_page);
1793 nfs_unlock_and_release_request(req);
1796 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1799 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1800 struct nfs_page *req)
1802 __set_page_dirty_nobuffers(req->wb_page);
1806 * Commit dirty pages
1809 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1810 struct nfs_commit_info *cinfo)
1812 struct nfs_commit_data *data;
1813 unsigned short task_flags = 0;
1815 /* another commit raced with us */
1816 if (list_empty(head))
1819 data = nfs_commitdata_alloc();
1821 nfs_retry_commit(head, NULL, cinfo, -1);
1825 /* Set up the argument struct */
1826 nfs_init_commit(data, head, NULL, cinfo);
1827 if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1828 task_flags = RPC_TASK_MOVEABLE;
1829 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1831 RPC_TASK_CRED_NOREF | task_flags);
1835 * COMMIT call returned
1837 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1839 struct nfs_commit_data *data = calldata;
1841 /* Call the NFS version-specific code */
1842 NFS_PROTO(data->inode)->commit_done(task, data);
1843 trace_nfs_commit_done(task, data);
1846 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1848 const struct nfs_writeverf *verf = data->res.verf;
1849 struct nfs_page *req;
1850 int status = data->task.tk_status;
1851 struct nfs_commit_info cinfo;
1852 struct nfs_server *nfss;
1854 while (!list_empty(&data->pages)) {
1855 req = nfs_list_entry(data->pages.next);
1856 nfs_list_remove_request(req);
1858 nfs_clear_page_commit(req->wb_page);
1860 dprintk("NFS: commit (%s/%llu %d@%lld)",
1861 nfs_req_openctx(req)->dentry->d_sb->s_id,
1862 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1864 (long long)req_offset(req));
1867 trace_nfs_commit_error(req, status);
1868 nfs_mapping_set_error(req->wb_page, status);
1869 nfs_inode_remove_request(req);
1871 dprintk_cont(", error = %d\n", status);
1875 /* Okay, COMMIT succeeded, apparently. Check the verifier
1876 * returned by the server against all stored verfs. */
1877 if (nfs_write_match_verf(verf, req)) {
1878 /* We have a match */
1880 nfs_inode_remove_request(req);
1881 dprintk_cont(" OK\n");
1884 /* We have a mismatch. Write the page again */
1885 dprintk_cont(" mismatch\n");
1886 nfs_mark_request_dirty(req);
1887 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1889 nfs_unlock_and_release_request(req);
1890 /* Latency breaker */
1893 nfss = NFS_SERVER(data->inode);
1894 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1895 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1897 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1898 nfs_commit_end(cinfo.mds);
1901 static void nfs_commit_release(void *calldata)
1903 struct nfs_commit_data *data = calldata;
1905 data->completion_ops->completion(data);
1906 nfs_commitdata_release(calldata);
1909 static const struct rpc_call_ops nfs_commit_ops = {
1910 .rpc_call_prepare = nfs_commit_prepare,
1911 .rpc_call_done = nfs_commit_done,
1912 .rpc_release = nfs_commit_release,
1915 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1916 .completion = nfs_commit_release_pages,
1917 .resched_write = nfs_commit_resched_write,
1920 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1921 int how, struct nfs_commit_info *cinfo)
1925 status = pnfs_commit_list(inode, head, how, cinfo);
1926 if (status == PNFS_NOT_ATTEMPTED)
1927 status = nfs_commit_list(inode, head, how, cinfo);
1931 static int __nfs_commit_inode(struct inode *inode, int how,
1932 struct writeback_control *wbc)
1935 struct nfs_commit_info cinfo;
1936 int may_wait = how & FLUSH_SYNC;
1940 nfs_init_cinfo_from_inode(&cinfo, inode);
1941 nfs_commit_begin(cinfo.mds);
1943 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1946 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1950 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1951 if (nscan < wbc->nr_to_write)
1952 wbc->nr_to_write -= nscan;
1954 wbc->nr_to_write = 0;
1956 if (nscan < INT_MAX)
1960 nfs_commit_end(cinfo.mds);
1961 if (ret || !may_wait)
1963 return wait_on_commit(cinfo.mds);
1966 int nfs_commit_inode(struct inode *inode, int how)
1968 return __nfs_commit_inode(inode, how, NULL);
1970 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1972 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1974 struct nfs_inode *nfsi = NFS_I(inode);
1975 int flags = FLUSH_SYNC;
1978 if (wbc->sync_mode == WB_SYNC_NONE) {
1979 /* no commits means nothing needs to be done */
1980 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1981 goto check_requests_outstanding;
1983 /* Don't commit yet if this is a non-blocking flush and there
1984 * are a lot of outstanding writes for this mapping.
1986 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1987 goto out_mark_dirty;
1989 /* don't wait for the COMMIT response */
1993 ret = __nfs_commit_inode(inode, flags, wbc);
1995 if (flags & FLUSH_SYNC)
1997 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1998 goto out_mark_dirty;
2000 check_requests_outstanding:
2001 if (!atomic_read(&nfsi->commit_info.rpcs_out))
2004 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2007 EXPORT_SYMBOL_GPL(nfs_write_inode);
2010 * Wrapper for filemap_write_and_wait_range()
2012 * Needed for pNFS in order to ensure data becomes visible to the
2015 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2016 loff_t lstart, loff_t lend)
2020 ret = filemap_write_and_wait_range(mapping, lstart, lend);
2022 ret = pnfs_sync_inode(mapping->host, true);
2025 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2028 * flush the inode to disk.
2030 int nfs_wb_all(struct inode *inode)
2034 trace_nfs_writeback_inode_enter(inode);
2036 ret = filemap_write_and_wait(inode->i_mapping);
2039 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2042 pnfs_sync_inode(inode, true);
2046 trace_nfs_writeback_inode_exit(inode, ret);
2049 EXPORT_SYMBOL_GPL(nfs_wb_all);
2051 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2053 struct nfs_page *req;
2056 wait_on_page_writeback(page);
2058 /* blocking call to cancel all requests and join to a single (head)
2060 req = nfs_lock_and_join_requests(page);
2065 /* all requests from this page have been cancelled by
2066 * nfs_lock_and_join_requests, so just remove the head
2067 * request from the inode / page_private pointer and
2069 nfs_inode_remove_request(req);
2070 nfs_unlock_and_release_request(req);
2077 * Write back all requests on one page - we do this before reading it.
2079 int nfs_wb_page(struct inode *inode, struct page *page)
2081 loff_t range_start = page_file_offset(page);
2082 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2083 struct writeback_control wbc = {
2084 .sync_mode = WB_SYNC_ALL,
2086 .range_start = range_start,
2087 .range_end = range_end,
2091 trace_nfs_writeback_page_enter(inode);
2094 wait_on_page_writeback(page);
2095 if (clear_page_dirty_for_io(page)) {
2096 ret = nfs_writepage_locked(page, &wbc);
2102 if (!PagePrivate(page))
2104 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2109 trace_nfs_writeback_page_exit(inode, ret);
2113 #ifdef CONFIG_MIGRATION
2114 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2115 struct page *page, enum migrate_mode mode)
2118 * If PagePrivate is set, then the page is currently associated with
2119 * an in-progress read or write request. Don't try to migrate it.
2121 * FIXME: we could do this in principle, but we'll need a way to ensure
2122 * that we can safely release the inode reference while holding
2125 if (PagePrivate(page))
2128 if (PageFsCache(page)) {
2129 if (mode == MIGRATE_ASYNC)
2131 wait_on_page_fscache(page);
2134 return migrate_page(mapping, newpage, page, mode);
2138 int __init nfs_init_writepagecache(void)
2140 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2141 sizeof(struct nfs_pgio_header),
2142 0, SLAB_HWCACHE_ALIGN,
2144 if (nfs_wdata_cachep == NULL)
2147 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2149 if (nfs_wdata_mempool == NULL)
2150 goto out_destroy_write_cache;
2152 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2153 sizeof(struct nfs_commit_data),
2154 0, SLAB_HWCACHE_ALIGN,
2156 if (nfs_cdata_cachep == NULL)
2157 goto out_destroy_write_mempool;
2159 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2161 if (nfs_commit_mempool == NULL)
2162 goto out_destroy_commit_cache;
2165 * NFS congestion size, scale with available memory.
2177 * This allows larger machines to have larger/more transfers.
2178 * Limit the default to 256M
2180 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2181 if (nfs_congestion_kb > 256*1024)
2182 nfs_congestion_kb = 256*1024;
2186 out_destroy_commit_cache:
2187 kmem_cache_destroy(nfs_cdata_cachep);
2188 out_destroy_write_mempool:
2189 mempool_destroy(nfs_wdata_mempool);
2190 out_destroy_write_cache:
2191 kmem_cache_destroy(nfs_wdata_cachep);
2195 void nfs_destroy_writepagecache(void)
2197 mempool_destroy(nfs_commit_mempool);
2198 kmem_cache_destroy(nfs_cdata_cachep);
2199 mempool_destroy(nfs_wdata_mempool);
2200 kmem_cache_destroy(nfs_wdata_cachep);
2203 static const struct nfs_rw_ops nfs_rw_write_ops = {
2204 .rw_alloc_header = nfs_writehdr_alloc,
2205 .rw_free_header = nfs_writehdr_free,
2206 .rw_done = nfs_writeback_done,
2207 .rw_result = nfs_writeback_result,
2208 .rw_initiate = nfs_initiate_write,