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;
99 p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
101 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
104 memset(p, 0, sizeof(*p));
106 p->rw_mode = FMODE_WRITE;
110 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
112 mempool_free(hdr, nfs_wdata_mempool);
115 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
117 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
120 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
121 void (*complete)(void *), void *data)
123 ioc->complete = complete;
125 kref_init(&ioc->refcount);
128 static void nfs_io_completion_release(struct kref *kref)
130 struct nfs_io_completion *ioc = container_of(kref,
131 struct nfs_io_completion, refcount);
132 ioc->complete(ioc->data);
136 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
139 kref_get(&ioc->refcount);
142 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
145 kref_put(&ioc->refcount, nfs_io_completion_release);
149 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
151 if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
152 kref_get(&req->wb_kref);
153 atomic_long_inc(&NFS_I(inode)->nrequests);
158 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
162 if (!test_bit(PG_REMOVE, &req->wb_flags))
164 ret = nfs_page_group_lock(req);
167 if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
168 nfs_page_set_inode_ref(req, inode);
169 nfs_page_group_unlock(req);
173 static struct nfs_page *
174 nfs_page_private_request(struct page *page)
176 if (!PagePrivate(page))
178 return (struct nfs_page *)page_private(page);
182 * nfs_page_find_head_request_locked - find head request associated with @page
184 * must be called while holding the inode lock.
186 * returns matching head request with reference held, or NULL if not found.
188 static struct nfs_page *
189 nfs_page_find_private_request(struct page *page)
191 struct address_space *mapping = page_file_mapping(page);
192 struct nfs_page *req;
194 if (!PagePrivate(page))
196 spin_lock(&mapping->private_lock);
197 req = nfs_page_private_request(page);
199 WARN_ON_ONCE(req->wb_head != req);
200 kref_get(&req->wb_kref);
202 spin_unlock(&mapping->private_lock);
206 static struct nfs_page *
207 nfs_page_find_swap_request(struct page *page)
209 struct inode *inode = page_file_mapping(page)->host;
210 struct nfs_inode *nfsi = NFS_I(inode);
211 struct nfs_page *req = NULL;
212 if (!PageSwapCache(page))
214 mutex_lock(&nfsi->commit_mutex);
215 if (PageSwapCache(page)) {
216 req = nfs_page_search_commits_for_head_request_locked(nfsi,
219 WARN_ON_ONCE(req->wb_head != req);
220 kref_get(&req->wb_kref);
223 mutex_unlock(&nfsi->commit_mutex);
228 * nfs_page_find_head_request - find head request associated with @page
230 * returns matching head request with reference held, or NULL if not found.
232 static struct nfs_page *nfs_page_find_head_request(struct page *page)
234 struct nfs_page *req;
236 req = nfs_page_find_private_request(page);
238 req = nfs_page_find_swap_request(page);
242 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
244 struct inode *inode = page_file_mapping(page)->host;
245 struct nfs_page *req, *head;
249 req = nfs_page_find_head_request(page);
252 head = nfs_page_group_lock_head(req);
254 nfs_release_request(req);
257 ret = nfs_cancel_remove_inode(head, inode);
259 nfs_unlock_and_release_request(head);
262 /* Ensure that nobody removed the request before we locked it */
263 if (head == nfs_page_private_request(page))
265 if (PageSwapCache(page))
267 nfs_unlock_and_release_request(head);
272 /* Adjust the file length if we're writing beyond the end */
273 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
275 struct inode *inode = page_file_mapping(page)->host;
279 spin_lock(&inode->i_lock);
280 i_size = i_size_read(inode);
281 end_index = (i_size - 1) >> PAGE_SHIFT;
282 if (i_size > 0 && page_index(page) < end_index)
284 end = page_file_offset(page) + ((loff_t)offset+count);
287 i_size_write(inode, end);
288 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
289 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
291 spin_unlock(&inode->i_lock);
294 /* A writeback failed: mark the page as bad, and invalidate the page cache */
295 static void nfs_set_pageerror(struct address_space *mapping)
297 struct inode *inode = mapping->host;
299 nfs_zap_mapping(mapping->host, mapping);
300 /* Force file size revalidation */
301 spin_lock(&inode->i_lock);
302 NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
303 NFS_INO_REVAL_PAGECACHE |
304 NFS_INO_INVALID_SIZE;
305 spin_unlock(&inode->i_lock);
308 static void nfs_mapping_set_error(struct page *page, int error)
310 struct address_space *mapping = page_file_mapping(page);
313 filemap_set_wb_err(mapping, error);
315 errseq_set(&mapping->host->i_sb->s_wb_err,
316 error == -ENOSPC ? -ENOSPC : -EIO);
317 nfs_set_pageerror(mapping);
321 * nfs_page_group_search_locked
322 * @head - head request of page group
323 * @page_offset - offset into page
325 * Search page group with head @head to find a request that contains the
326 * page offset @page_offset.
328 * Returns a pointer to the first matching nfs request, or NULL if no
331 * Must be called with the page group lock held
333 static struct nfs_page *
334 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
336 struct nfs_page *req;
340 if (page_offset >= req->wb_pgbase &&
341 page_offset < (req->wb_pgbase + req->wb_bytes))
344 req = req->wb_this_page;
345 } while (req != head);
351 * nfs_page_group_covers_page
352 * @head - head request of page group
354 * Return true if the page group with head @head covers the whole page,
355 * returns false otherwise
357 static bool nfs_page_group_covers_page(struct nfs_page *req)
359 struct nfs_page *tmp;
360 unsigned int pos = 0;
361 unsigned int len = nfs_page_length(req->wb_page);
363 nfs_page_group_lock(req);
366 tmp = nfs_page_group_search_locked(req->wb_head, pos);
369 pos = tmp->wb_pgbase + tmp->wb_bytes;
372 nfs_page_group_unlock(req);
376 /* We can set the PG_uptodate flag if we see that a write request
377 * covers the full page.
379 static void nfs_mark_uptodate(struct nfs_page *req)
381 if (PageUptodate(req->wb_page))
383 if (!nfs_page_group_covers_page(req))
385 SetPageUptodate(req->wb_page);
388 static int wb_priority(struct writeback_control *wbc)
392 if (wbc->sync_mode == WB_SYNC_ALL)
393 ret = FLUSH_COND_STABLE;
398 * NFS congestion control
401 int nfs_congestion_kb;
403 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
404 #define NFS_CONGESTION_OFF_THRESH \
405 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
407 static void nfs_set_page_writeback(struct page *page)
409 struct inode *inode = page_file_mapping(page)->host;
410 struct nfs_server *nfss = NFS_SERVER(inode);
411 int ret = test_set_page_writeback(page);
413 WARN_ON_ONCE(ret != 0);
415 if (atomic_long_inc_return(&nfss->writeback) >
416 NFS_CONGESTION_ON_THRESH)
417 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
420 static void nfs_end_page_writeback(struct nfs_page *req)
422 struct inode *inode = page_file_mapping(req->wb_page)->host;
423 struct nfs_server *nfss = NFS_SERVER(inode);
426 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
427 nfs_unlock_request(req);
431 end_page_writeback(req->wb_page);
432 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
433 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
437 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
439 * @destroy_list - request list (using wb_this_page) terminated by @old_head
440 * @old_head - the old head of the list
442 * All subrequests must be locked and removed from all lists, so at this point
443 * they are only "active" in this function, and possibly in nfs_wait_on_request
444 * with a reference held by some other context.
447 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
448 struct nfs_page *old_head,
451 while (destroy_list) {
452 struct nfs_page *subreq = destroy_list;
454 destroy_list = (subreq->wb_this_page == old_head) ?
455 NULL : subreq->wb_this_page;
457 /* Note: lock subreq in order to change subreq->wb_head */
458 nfs_page_set_headlock(subreq);
459 WARN_ON_ONCE(old_head != subreq->wb_head);
461 /* make sure old group is not used */
462 subreq->wb_this_page = subreq;
463 subreq->wb_head = subreq;
465 clear_bit(PG_REMOVE, &subreq->wb_flags);
467 /* Note: races with nfs_page_group_destroy() */
468 if (!kref_read(&subreq->wb_kref)) {
469 /* Check if we raced with nfs_page_group_destroy() */
470 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
471 nfs_page_clear_headlock(subreq);
472 nfs_free_request(subreq);
474 nfs_page_clear_headlock(subreq);
477 nfs_page_clear_headlock(subreq);
479 nfs_release_request(old_head);
481 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
482 nfs_release_request(subreq);
483 atomic_long_dec(&NFS_I(inode)->nrequests);
486 /* subreq is now totally disconnected from page group or any
487 * write / commit lists. last chance to wake any waiters */
488 nfs_unlock_and_release_request(subreq);
493 * nfs_join_page_group - destroy subrequests of the head req
494 * @head: the page used to lookup the "page group" of nfs_page structures
495 * @inode: Inode to which the request belongs.
497 * This function joins all sub requests to the head request by first
498 * locking all requests in the group, cancelling any pending operations
499 * and finally updating the head request to cover the whole range covered by
500 * the (former) group. All subrequests are removed from any write or commit
501 * lists, unlinked from the group and destroyed.
504 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
506 struct nfs_page *subreq;
507 struct nfs_page *destroy_list = NULL;
508 unsigned int pgbase, off, bytes;
510 pgbase = head->wb_pgbase;
511 bytes = head->wb_bytes;
512 off = head->wb_offset;
513 for (subreq = head->wb_this_page; subreq != head;
514 subreq = subreq->wb_this_page) {
515 /* Subrequests should always form a contiguous range */
516 if (pgbase > subreq->wb_pgbase) {
517 off -= pgbase - subreq->wb_pgbase;
518 bytes += pgbase - subreq->wb_pgbase;
519 pgbase = subreq->wb_pgbase;
521 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
525 /* Set the head request's range to cover the former page group */
526 head->wb_pgbase = pgbase;
527 head->wb_bytes = bytes;
528 head->wb_offset = off;
530 /* Now that all requests are locked, make sure they aren't on any list.
531 * Commit list removal accounting is done after locks are dropped */
534 nfs_clear_request_commit(subreq);
535 subreq = subreq->wb_this_page;
536 } while (subreq != head);
538 /* unlink subrequests from head, destroy them later */
539 if (head->wb_this_page != head) {
540 /* destroy list will be terminated by head */
541 destroy_list = head->wb_this_page;
542 head->wb_this_page = head;
545 nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
549 * nfs_lock_and_join_requests - join all subreqs to the head req
550 * @page: the page used to lookup the "page group" of nfs_page structures
552 * This function joins all sub requests to the head request by first
553 * locking all requests in the group, cancelling any pending operations
554 * and finally updating the head request to cover the whole range covered by
555 * the (former) group. All subrequests are removed from any write or commit
556 * lists, unlinked from the group and destroyed.
558 * Returns a locked, referenced pointer to the head request - which after
559 * this call is guaranteed to be the only request associated with the page.
560 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
561 * error was encountered.
563 static struct nfs_page *
564 nfs_lock_and_join_requests(struct page *page)
566 struct inode *inode = page_file_mapping(page)->host;
567 struct nfs_page *head;
571 * A reference is taken only on the head request which acts as a
572 * reference to the whole page group - the group will not be destroyed
573 * until the head reference is released.
575 head = nfs_find_and_lock_page_request(page);
576 if (IS_ERR_OR_NULL(head))
579 /* lock each request in the page group */
580 ret = nfs_page_group_lock_subrequests(head);
582 nfs_unlock_and_release_request(head);
586 nfs_join_page_group(head, inode);
591 static void nfs_write_error(struct nfs_page *req, int error)
593 trace_nfs_write_error(req, error);
594 nfs_mapping_set_error(req->wb_page, error);
595 nfs_inode_remove_request(req);
596 nfs_end_page_writeback(req);
597 nfs_release_request(req);
601 * Find an associated nfs write request, and prepare to flush it out
602 * May return an error if the user signalled nfs_wait_on_request().
604 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
607 struct nfs_page *req;
610 req = nfs_lock_and_join_requests(page);
617 nfs_set_page_writeback(page);
618 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
620 /* If there is a fatal error that covers this write, just exit */
621 ret = pgio->pg_error;
622 if (nfs_error_is_fatal_on_server(ret))
626 if (!nfs_pageio_add_request(pgio, req)) {
627 ret = pgio->pg_error;
629 * Remove the problematic req upon fatal errors on the server
631 if (nfs_error_is_fatal(ret)) {
632 if (nfs_error_is_fatal_on_server(ret))
636 nfs_redirty_request(req);
639 nfs_add_stats(page_file_mapping(page)->host,
640 NFSIOS_WRITEPAGES, 1);
644 nfs_write_error(req, ret);
648 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
649 struct nfs_pageio_descriptor *pgio)
653 nfs_pageio_cond_complete(pgio, page_index(page));
654 ret = nfs_page_async_flush(pgio, page);
655 if (ret == -EAGAIN) {
656 redirty_page_for_writepage(wbc, page);
657 ret = AOP_WRITEPAGE_ACTIVATE;
663 * Write an mmapped page to the server.
665 static int nfs_writepage_locked(struct page *page,
666 struct writeback_control *wbc)
668 struct nfs_pageio_descriptor pgio;
669 struct inode *inode = page_file_mapping(page)->host;
672 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
673 nfs_pageio_init_write(&pgio, inode, 0,
674 false, &nfs_async_write_completion_ops);
675 err = nfs_do_writepage(page, wbc, &pgio);
677 nfs_pageio_complete(&pgio);
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;
713 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
715 ioc = nfs_io_completion_alloc(GFP_KERNEL);
717 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
719 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
720 &nfs_async_write_completion_ops);
721 pgio.pg_io_completion = ioc;
722 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
724 nfs_pageio_complete(&pgio);
725 nfs_io_completion_put(ioc);
735 * Insert a write request into an inode
737 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
739 struct address_space *mapping = page_file_mapping(req->wb_page);
740 struct nfs_inode *nfsi = NFS_I(inode);
742 WARN_ON_ONCE(req->wb_this_page != req);
744 /* Lock the request! */
745 nfs_lock_request(req);
748 * Swap-space should not get truncated. Hence no need to plug the race
749 * with invalidate/truncate.
751 spin_lock(&mapping->private_lock);
752 if (!nfs_have_writebacks(inode) &&
753 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
754 inode_inc_iversion_raw(inode);
755 if (likely(!PageSwapCache(req->wb_page))) {
756 set_bit(PG_MAPPED, &req->wb_flags);
757 SetPagePrivate(req->wb_page);
758 set_page_private(req->wb_page, (unsigned long)req);
760 spin_unlock(&mapping->private_lock);
761 atomic_long_inc(&nfsi->nrequests);
762 /* this a head request for a page group - mark it as having an
763 * extra reference so sub groups can follow suit.
764 * This flag also informs pgio layer when to bump nrequests when
765 * adding subrequests. */
766 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
767 kref_get(&req->wb_kref);
771 * Remove a write request from an inode
773 static void nfs_inode_remove_request(struct nfs_page *req)
775 struct address_space *mapping = page_file_mapping(req->wb_page);
776 struct inode *inode = mapping->host;
777 struct nfs_inode *nfsi = NFS_I(inode);
778 struct nfs_page *head;
780 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
783 spin_lock(&mapping->private_lock);
784 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
785 set_page_private(head->wb_page, 0);
786 ClearPagePrivate(head->wb_page);
787 clear_bit(PG_MAPPED, &head->wb_flags);
789 spin_unlock(&mapping->private_lock);
792 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
793 nfs_release_request(req);
794 atomic_long_dec(&nfsi->nrequests);
799 nfs_mark_request_dirty(struct nfs_page *req)
802 __set_page_dirty_nobuffers(req->wb_page);
806 * nfs_page_search_commits_for_head_request_locked
808 * Search through commit lists on @inode for the head request for @page.
809 * Must be called while holding the inode (which is cinfo) lock.
811 * Returns the head request if found, or NULL if not found.
813 static struct nfs_page *
814 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
817 struct nfs_page *freq, *t;
818 struct nfs_commit_info cinfo;
819 struct inode *inode = &nfsi->vfs_inode;
821 nfs_init_cinfo_from_inode(&cinfo, inode);
823 /* search through pnfs commit lists */
824 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
826 return freq->wb_head;
828 /* Linearly search the commit list for the correct request */
829 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
830 if (freq->wb_page == page)
831 return freq->wb_head;
838 * nfs_request_add_commit_list_locked - add request to a commit list
839 * @req: pointer to a struct nfs_page
840 * @dst: commit list head
841 * @cinfo: holds list lock and accounting info
843 * This sets the PG_CLEAN bit, updates the cinfo count of
844 * number of outstanding requests requiring a commit as well as
847 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
851 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
852 struct nfs_commit_info *cinfo)
854 set_bit(PG_CLEAN, &req->wb_flags);
855 nfs_list_add_request(req, dst);
856 atomic_long_inc(&cinfo->mds->ncommit);
858 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
861 * nfs_request_add_commit_list - add request to a commit list
862 * @req: pointer to a struct nfs_page
863 * @cinfo: holds list lock and accounting info
865 * This sets the PG_CLEAN bit, updates the cinfo count of
866 * number of outstanding requests requiring a commit as well as
869 * The caller must _not_ hold the cinfo->lock, but must be
870 * holding the nfs_page lock.
873 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
875 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
876 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
877 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
879 nfs_mark_page_unstable(req->wb_page, cinfo);
881 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
884 * nfs_request_remove_commit_list - Remove request from a commit list
885 * @req: pointer to a nfs_page
886 * @cinfo: holds list lock and accounting info
888 * This clears the PG_CLEAN bit, and updates the cinfo's count of
889 * number of outstanding requests requiring a commit
890 * It does not update the MM page stats.
892 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
895 nfs_request_remove_commit_list(struct nfs_page *req,
896 struct nfs_commit_info *cinfo)
898 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
900 nfs_list_remove_request(req);
901 atomic_long_dec(&cinfo->mds->ncommit);
903 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
905 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
908 cinfo->inode = inode;
909 cinfo->mds = &NFS_I(inode)->commit_info;
910 cinfo->ds = pnfs_get_ds_info(inode);
912 cinfo->completion_ops = &nfs_commit_completion_ops;
915 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
917 struct nfs_direct_req *dreq)
920 nfs_init_cinfo_from_dreq(cinfo, dreq);
922 nfs_init_cinfo_from_inode(cinfo, inode);
924 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
927 * Add a request to the inode's commit list.
930 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
931 struct nfs_commit_info *cinfo, u32 ds_commit_idx)
933 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
935 nfs_request_add_commit_list(req, cinfo);
939 nfs_clear_page_commit(struct page *page)
941 dec_node_page_state(page, NR_WRITEBACK);
942 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
946 /* Called holding the request lock on @req */
948 nfs_clear_request_commit(struct nfs_page *req)
950 if (test_bit(PG_CLEAN, &req->wb_flags)) {
951 struct nfs_open_context *ctx = nfs_req_openctx(req);
952 struct inode *inode = d_inode(ctx->dentry);
953 struct nfs_commit_info cinfo;
955 nfs_init_cinfo_from_inode(&cinfo, inode);
956 mutex_lock(&NFS_I(inode)->commit_mutex);
957 if (!pnfs_clear_request_commit(req, &cinfo)) {
958 nfs_request_remove_commit_list(req, &cinfo);
960 mutex_unlock(&NFS_I(inode)->commit_mutex);
961 nfs_clear_page_commit(req->wb_page);
965 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
967 if (hdr->verf.committed == NFS_DATA_SYNC)
968 return hdr->lseg == NULL;
969 return hdr->verf.committed != NFS_FILE_SYNC;
972 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
974 nfs_io_completion_get(hdr->io_completion);
977 static void nfs_write_completion(struct nfs_pgio_header *hdr)
979 struct nfs_commit_info cinfo;
980 unsigned long bytes = 0;
982 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
984 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
985 while (!list_empty(&hdr->pages)) {
986 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
988 bytes += req->wb_bytes;
989 nfs_list_remove_request(req);
990 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
991 (hdr->good_bytes < bytes)) {
992 trace_nfs_comp_error(req, hdr->error);
993 nfs_mapping_set_error(req->wb_page, hdr->error);
996 if (nfs_write_need_commit(hdr)) {
997 /* Reset wb_nio, since the write was successful. */
999 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1000 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1001 hdr->pgio_mirror_idx);
1005 nfs_inode_remove_request(req);
1007 nfs_end_page_writeback(req);
1008 nfs_release_request(req);
1011 nfs_io_completion_put(hdr->io_completion);
1016 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1018 return atomic_long_read(&cinfo->mds->ncommit);
1021 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1023 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1024 struct nfs_commit_info *cinfo, int max)
1026 struct nfs_page *req, *tmp;
1029 list_for_each_entry_safe(req, tmp, src, wb_list) {
1030 kref_get(&req->wb_kref);
1031 if (!nfs_lock_request(req)) {
1032 nfs_release_request(req);
1035 nfs_request_remove_commit_list(req, cinfo);
1036 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1037 nfs_list_add_request(req, dst);
1039 if ((ret == max) && !cinfo->dreq)
1045 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1048 * nfs_scan_commit - Scan an inode for commit requests
1049 * @inode: NFS inode to scan
1050 * @dst: mds destination list
1051 * @cinfo: mds and ds lists of reqs ready to commit
1053 * Moves requests from the inode's 'commit' request list.
1054 * The requests are *not* checked to ensure that they form a contiguous set.
1057 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1058 struct nfs_commit_info *cinfo)
1062 if (!atomic_long_read(&cinfo->mds->ncommit))
1064 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1065 if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1066 const int max = INT_MAX;
1068 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1070 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1072 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1077 * Search for an existing write request, and attempt to update
1078 * it to reflect a new dirty region on a given page.
1080 * If the attempt fails, then the existing request is flushed out
1083 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1085 unsigned int offset,
1088 struct nfs_page *req;
1093 end = offset + bytes;
1095 req = nfs_lock_and_join_requests(page);
1096 if (IS_ERR_OR_NULL(req))
1099 rqend = req->wb_offset + req->wb_bytes;
1101 * Tell the caller to flush out the request if
1102 * the offsets are non-contiguous.
1103 * Note: nfs_flush_incompatible() will already
1104 * have flushed out requests having wrong owners.
1106 if (offset > rqend || end < req->wb_offset)
1109 /* Okay, the request matches. Update the region */
1110 if (offset < req->wb_offset) {
1111 req->wb_offset = offset;
1112 req->wb_pgbase = offset;
1115 req->wb_bytes = end - req->wb_offset;
1117 req->wb_bytes = rqend - req->wb_offset;
1122 * Note: we mark the request dirty here because
1123 * nfs_lock_and_join_requests() cannot preserve
1124 * commit flags, so we have to replay the write.
1126 nfs_mark_request_dirty(req);
1127 nfs_unlock_and_release_request(req);
1128 error = nfs_wb_page(inode, page);
1129 return (error < 0) ? ERR_PTR(error) : NULL;
1133 * Try to update an existing write request, or create one if there is none.
1135 * Note: Should always be called with the Page Lock held to prevent races
1136 * if we have to add a new request. Also assumes that the caller has
1137 * already called nfs_flush_incompatible() if necessary.
1139 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1140 struct page *page, unsigned int offset, unsigned int bytes)
1142 struct inode *inode = page_file_mapping(page)->host;
1143 struct nfs_page *req;
1145 req = nfs_try_to_update_request(inode, page, offset, bytes);
1148 req = nfs_create_request(ctx, page, offset, bytes);
1151 nfs_inode_add_request(inode, req);
1156 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1157 unsigned int offset, unsigned int count)
1159 struct nfs_page *req;
1161 req = nfs_setup_write_request(ctx, page, offset, count);
1163 return PTR_ERR(req);
1164 /* Update file length */
1165 nfs_grow_file(page, offset, count);
1166 nfs_mark_uptodate(req);
1167 nfs_mark_request_dirty(req);
1168 nfs_unlock_and_release_request(req);
1172 int nfs_flush_incompatible(struct file *file, struct page *page)
1174 struct nfs_open_context *ctx = nfs_file_open_context(file);
1175 struct nfs_lock_context *l_ctx;
1176 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1177 struct nfs_page *req;
1178 int do_flush, status;
1180 * Look for a request corresponding to this page. If there
1181 * is one, and it belongs to another file, we flush it out
1182 * before we try to copy anything into the page. Do this
1183 * due to the lack of an ACCESS-type call in NFSv2.
1184 * Also do the same if we find a request from an existing
1188 req = nfs_page_find_head_request(page);
1191 l_ctx = req->wb_lock_context;
1192 do_flush = req->wb_page != page ||
1193 !nfs_match_open_context(nfs_req_openctx(req), ctx);
1194 if (l_ctx && flctx &&
1195 !(list_empty_careful(&flctx->flc_posix) &&
1196 list_empty_careful(&flctx->flc_flock))) {
1197 do_flush |= l_ctx->lockowner != current->files;
1199 nfs_release_request(req);
1202 status = nfs_wb_page(page_file_mapping(page)->host, page);
1203 } while (status == 0);
1208 * Avoid buffered writes when a open context credential's key would
1211 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1213 * Return 0 and set a credential flag which triggers the inode to flush
1214 * and performs NFS_FILE_SYNC writes if the key will expired within
1215 * RPC_KEY_EXPIRE_TIMEO.
1218 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1220 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1222 if (nfs_ctx_key_to_expire(ctx, inode) &&
1224 /* Already expired! */
1230 * Test if the open context credential key is marked to expire soon.
1232 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1234 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1235 struct rpc_cred *cred = ctx->ll_cred;
1236 struct auth_cred acred = {
1240 if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1242 ctx->ll_cred = NULL;
1246 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1247 if (!cred || IS_ERR(cred))
1249 ctx->ll_cred = cred;
1250 return !!(cred->cr_ops->crkey_timeout &&
1251 cred->cr_ops->crkey_timeout(cred));
1255 * If the page cache is marked as unsafe or invalid, then we can't rely on
1256 * the PageUptodate() flag. In this case, we will need to turn off
1257 * write optimisations that depend on the page contents being correct.
1259 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1261 struct nfs_inode *nfsi = NFS_I(inode);
1263 if (nfs_have_delegated_attributes(inode))
1265 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1268 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1271 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1273 return PageUptodate(page) != 0;
1277 is_whole_file_wrlock(struct file_lock *fl)
1279 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1280 fl->fl_type == F_WRLCK;
1283 /* If we know the page is up to date, and we're not using byte range locks (or
1284 * if we have the whole file locked for writing), it may be more efficient to
1285 * extend the write to cover the entire page in order to avoid fragmentation
1288 * If the file is opened for synchronous writes then we can just skip the rest
1291 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1294 struct file_lock_context *flctx = inode->i_flctx;
1295 struct file_lock *fl;
1297 if (file->f_flags & O_DSYNC)
1299 if (!nfs_write_pageuptodate(page, inode))
1301 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1303 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1304 list_empty_careful(&flctx->flc_posix)))
1307 /* Check to see if there are whole file write locks */
1309 spin_lock(&flctx->flc_lock);
1310 if (!list_empty(&flctx->flc_posix)) {
1311 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1313 if (is_whole_file_wrlock(fl))
1315 } else if (!list_empty(&flctx->flc_flock)) {
1316 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1318 if (fl->fl_type == F_WRLCK)
1321 spin_unlock(&flctx->flc_lock);
1326 * Update and possibly write a cached page of an NFS file.
1328 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1329 * things with a page scheduled for an RPC call (e.g. invalidate it).
1331 int nfs_updatepage(struct file *file, struct page *page,
1332 unsigned int offset, unsigned int count)
1334 struct nfs_open_context *ctx = nfs_file_open_context(file);
1335 struct address_space *mapping = page_file_mapping(page);
1336 struct inode *inode = mapping->host;
1339 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1341 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1342 file, count, (long long)(page_file_offset(page) + offset));
1347 if (nfs_can_extend_write(file, page, inode)) {
1348 count = max(count + offset, nfs_page_length(page));
1352 status = nfs_writepage_setup(ctx, page, offset, count);
1354 nfs_set_pageerror(mapping);
1356 __set_page_dirty_nobuffers(page);
1358 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1359 status, (long long)i_size_read(inode));
1363 static int flush_task_priority(int how)
1365 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1367 return RPC_PRIORITY_HIGH;
1369 return RPC_PRIORITY_LOW;
1371 return RPC_PRIORITY_NORMAL;
1374 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1375 struct rpc_message *msg,
1376 const struct nfs_rpc_ops *rpc_ops,
1377 struct rpc_task_setup *task_setup_data, int how)
1379 int priority = flush_task_priority(how);
1381 task_setup_data->priority = priority;
1382 rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1383 trace_nfs_initiate_write(hdr);
1386 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1387 * call this on each, which will prepare them to be retried on next
1388 * writeback using standard nfs.
1390 static void nfs_redirty_request(struct nfs_page *req)
1392 /* Bump the transmission count */
1394 nfs_mark_request_dirty(req);
1395 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1396 nfs_end_page_writeback(req);
1397 nfs_release_request(req);
1400 static void nfs_async_write_error(struct list_head *head, int error)
1402 struct nfs_page *req;
1404 while (!list_empty(head)) {
1405 req = nfs_list_entry(head->next);
1406 nfs_list_remove_request(req);
1407 if (nfs_error_is_fatal_on_server(error))
1408 nfs_write_error(req, error);
1410 nfs_redirty_request(req);
1414 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1416 nfs_async_write_error(&hdr->pages, 0);
1417 filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1418 hdr->args.offset + hdr->args.count - 1);
1421 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1422 .init_hdr = nfs_async_write_init,
1423 .error_cleanup = nfs_async_write_error,
1424 .completion = nfs_write_completion,
1425 .reschedule_io = nfs_async_write_reschedule_io,
1428 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1429 struct inode *inode, int ioflags, bool force_mds,
1430 const struct nfs_pgio_completion_ops *compl_ops)
1432 struct nfs_server *server = NFS_SERVER(inode);
1433 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1435 #ifdef CONFIG_NFS_V4_1
1436 if (server->pnfs_curr_ld && !force_mds)
1437 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1439 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1440 server->wsize, ioflags);
1442 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1444 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1446 struct nfs_pgio_mirror *mirror;
1448 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1449 pgio->pg_ops->pg_cleanup(pgio);
1451 pgio->pg_ops = &nfs_pgio_rw_ops;
1453 nfs_pageio_stop_mirroring(pgio);
1455 mirror = &pgio->pg_mirrors[0];
1456 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1458 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1461 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1463 struct nfs_commit_data *data = calldata;
1465 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1469 * Special version of should_remove_suid() that ignores capabilities.
1471 static int nfs_should_remove_suid(const struct inode *inode)
1473 umode_t mode = inode->i_mode;
1476 /* suid always must be killed */
1477 if (unlikely(mode & S_ISUID))
1478 kill = ATTR_KILL_SUID;
1481 * sgid without any exec bits is just a mandatory locking mark; leave
1482 * it alone. If some exec bits are set, it's a real sgid; kill it.
1484 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1485 kill |= ATTR_KILL_SGID;
1487 if (unlikely(kill && S_ISREG(mode)))
1493 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1494 struct nfs_fattr *fattr)
1496 struct nfs_pgio_args *argp = &hdr->args;
1497 struct nfs_pgio_res *resp = &hdr->res;
1498 u64 size = argp->offset + resp->count;
1500 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1502 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1503 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1506 if (size != fattr->size)
1508 /* Set attribute barrier */
1509 nfs_fattr_set_barrier(fattr);
1510 /* ...and update size */
1511 fattr->valid |= NFS_ATTR_FATTR_SIZE;
1514 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1516 struct nfs_fattr *fattr = &hdr->fattr;
1517 struct inode *inode = hdr->inode;
1519 spin_lock(&inode->i_lock);
1520 nfs_writeback_check_extend(hdr, fattr);
1521 nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1522 spin_unlock(&inode->i_lock);
1524 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1527 * This function is called when the WRITE call is complete.
1529 static int nfs_writeback_done(struct rpc_task *task,
1530 struct nfs_pgio_header *hdr,
1531 struct inode *inode)
1536 * ->write_done will attempt to use post-op attributes to detect
1537 * conflicting writes by other clients. A strict interpretation
1538 * of close-to-open would allow us to continue caching even if
1539 * another writer had changed the file, but some applications
1540 * depend on tighter cache coherency when writing.
1542 status = NFS_PROTO(inode)->write_done(task, hdr);
1546 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1547 trace_nfs_writeback_done(task, hdr);
1549 if (hdr->res.verf->committed < hdr->args.stable &&
1550 task->tk_status >= 0) {
1551 /* We tried a write call, but the server did not
1552 * commit data to stable storage even though we
1554 * Note: There is a known bug in Tru64 < 5.0 in which
1555 * the server reports NFS_DATA_SYNC, but performs
1556 * NFS_FILE_SYNC. We therefore implement this checking
1557 * as a dprintk() in order to avoid filling syslog.
1559 static unsigned long complain;
1561 /* Note this will print the MDS for a DS write */
1562 if (time_before(complain, jiffies)) {
1563 dprintk("NFS: faulty NFS server %s:"
1564 " (committed = %d) != (stable = %d)\n",
1565 NFS_SERVER(inode)->nfs_client->cl_hostname,
1566 hdr->res.verf->committed, hdr->args.stable);
1567 complain = jiffies + 300 * HZ;
1571 /* Deal with the suid/sgid bit corner case */
1572 if (nfs_should_remove_suid(inode)) {
1573 spin_lock(&inode->i_lock);
1574 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1575 spin_unlock(&inode->i_lock);
1581 * This function is called when the WRITE call is complete.
1583 static void nfs_writeback_result(struct rpc_task *task,
1584 struct nfs_pgio_header *hdr)
1586 struct nfs_pgio_args *argp = &hdr->args;
1587 struct nfs_pgio_res *resp = &hdr->res;
1589 if (resp->count < argp->count) {
1590 static unsigned long complain;
1592 /* This a short write! */
1593 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1595 /* Has the server at least made some progress? */
1596 if (resp->count == 0) {
1597 if (time_before(complain, jiffies)) {
1599 "NFS: Server wrote zero bytes, expected %u.\n",
1601 complain = jiffies + 300 * HZ;
1603 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1604 task->tk_status = -EIO;
1608 /* For non rpc-based layout drivers, retry-through-MDS */
1609 if (!task->tk_ops) {
1610 hdr->pnfs_error = -EAGAIN;
1614 /* Was this an NFSv2 write or an NFSv3 stable write? */
1615 if (resp->verf->committed != NFS_UNSTABLE) {
1616 /* Resend from where the server left off */
1617 hdr->mds_offset += resp->count;
1618 argp->offset += resp->count;
1619 argp->pgbase += resp->count;
1620 argp->count -= resp->count;
1622 /* Resend as a stable write in order to avoid
1623 * headaches in the case of a server crash.
1625 argp->stable = NFS_FILE_SYNC;
1628 resp->verf->committed = 0;
1629 rpc_restart_call_prepare(task);
1633 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1635 return wait_var_event_killable(&cinfo->rpcs_out,
1636 !atomic_read(&cinfo->rpcs_out));
1639 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1641 atomic_inc(&cinfo->rpcs_out);
1644 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1646 if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1647 wake_up_var(&cinfo->rpcs_out);
1653 void nfs_commitdata_release(struct nfs_commit_data *data)
1655 put_nfs_open_context(data->context);
1656 nfs_commit_free(data);
1658 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1660 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1661 const struct nfs_rpc_ops *nfs_ops,
1662 const struct rpc_call_ops *call_ops,
1665 struct rpc_task *task;
1666 int priority = flush_task_priority(how);
1667 struct rpc_message msg = {
1668 .rpc_argp = &data->args,
1669 .rpc_resp = &data->res,
1670 .rpc_cred = data->cred,
1672 struct rpc_task_setup task_setup_data = {
1673 .task = &data->task,
1675 .rpc_message = &msg,
1676 .callback_ops = call_ops,
1677 .callback_data = data,
1678 .workqueue = nfsiod_workqueue,
1679 .flags = RPC_TASK_ASYNC | flags,
1680 .priority = priority,
1682 /* Set up the initial task struct. */
1683 nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1684 trace_nfs_initiate_commit(data);
1686 dprintk("NFS: initiated commit call\n");
1688 task = rpc_run_task(&task_setup_data);
1690 return PTR_ERR(task);
1691 if (how & FLUSH_SYNC)
1692 rpc_wait_for_completion_task(task);
1696 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1698 static loff_t nfs_get_lwb(struct list_head *head)
1701 struct nfs_page *req;
1703 list_for_each_entry(req, head, wb_list)
1704 if (lwb < (req_offset(req) + req->wb_bytes))
1705 lwb = req_offset(req) + req->wb_bytes;
1711 * Set up the argument/result storage required for the RPC call.
1713 void nfs_init_commit(struct nfs_commit_data *data,
1714 struct list_head *head,
1715 struct pnfs_layout_segment *lseg,
1716 struct nfs_commit_info *cinfo)
1718 struct nfs_page *first;
1719 struct nfs_open_context *ctx;
1720 struct inode *inode;
1722 /* Set up the RPC argument and reply structs
1723 * NB: take care not to mess about with data->commit et al. */
1726 list_splice_init(head, &data->pages);
1728 first = nfs_list_entry(data->pages.next);
1729 ctx = nfs_req_openctx(first);
1730 inode = d_inode(ctx->dentry);
1732 data->inode = inode;
1733 data->cred = ctx->cred;
1734 data->lseg = lseg; /* reference transferred */
1735 /* only set lwb for pnfs commit */
1737 data->lwb = nfs_get_lwb(&data->pages);
1738 data->mds_ops = &nfs_commit_ops;
1739 data->completion_ops = cinfo->completion_ops;
1740 data->dreq = cinfo->dreq;
1742 data->args.fh = NFS_FH(data->inode);
1743 /* Note: we always request a commit of the entire inode */
1744 data->args.offset = 0;
1745 data->args.count = 0;
1746 data->context = get_nfs_open_context(ctx);
1747 data->res.fattr = &data->fattr;
1748 data->res.verf = &data->verf;
1749 nfs_fattr_init(&data->fattr);
1750 nfs_commit_begin(cinfo->mds);
1752 EXPORT_SYMBOL_GPL(nfs_init_commit);
1754 void nfs_retry_commit(struct list_head *page_list,
1755 struct pnfs_layout_segment *lseg,
1756 struct nfs_commit_info *cinfo,
1759 struct nfs_page *req;
1761 while (!list_empty(page_list)) {
1762 req = nfs_list_entry(page_list->next);
1763 nfs_list_remove_request(req);
1764 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1766 nfs_clear_page_commit(req->wb_page);
1767 nfs_unlock_and_release_request(req);
1770 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1773 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1774 struct nfs_page *req)
1776 __set_page_dirty_nobuffers(req->wb_page);
1780 * Commit dirty pages
1783 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1784 struct nfs_commit_info *cinfo)
1786 struct nfs_commit_data *data;
1788 /* another commit raced with us */
1789 if (list_empty(head))
1792 data = nfs_commitdata_alloc();
1794 nfs_retry_commit(head, NULL, cinfo, -1);
1798 /* Set up the argument struct */
1799 nfs_init_commit(data, head, NULL, cinfo);
1800 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1801 data->mds_ops, how, RPC_TASK_CRED_NOREF);
1805 * COMMIT call returned
1807 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1809 struct nfs_commit_data *data = calldata;
1811 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1812 task->tk_pid, task->tk_status);
1814 /* Call the NFS version-specific code */
1815 NFS_PROTO(data->inode)->commit_done(task, data);
1816 trace_nfs_commit_done(task, data);
1819 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1821 const struct nfs_writeverf *verf = data->res.verf;
1822 struct nfs_page *req;
1823 int status = data->task.tk_status;
1824 struct nfs_commit_info cinfo;
1825 struct nfs_server *nfss;
1827 while (!list_empty(&data->pages)) {
1828 req = nfs_list_entry(data->pages.next);
1829 nfs_list_remove_request(req);
1831 nfs_clear_page_commit(req->wb_page);
1833 dprintk("NFS: commit (%s/%llu %d@%lld)",
1834 nfs_req_openctx(req)->dentry->d_sb->s_id,
1835 (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1837 (long long)req_offset(req));
1840 trace_nfs_commit_error(req, status);
1841 nfs_mapping_set_error(req->wb_page, status);
1842 nfs_inode_remove_request(req);
1844 dprintk_cont(", error = %d\n", status);
1848 /* Okay, COMMIT succeeded, apparently. Check the verifier
1849 * returned by the server against all stored verfs. */
1850 if (nfs_write_match_verf(verf, req)) {
1851 /* We have a match */
1853 nfs_inode_remove_request(req);
1854 dprintk_cont(" OK\n");
1857 /* We have a mismatch. Write the page again */
1858 dprintk_cont(" mismatch\n");
1859 nfs_mark_request_dirty(req);
1860 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1862 nfs_unlock_and_release_request(req);
1863 /* Latency breaker */
1866 nfss = NFS_SERVER(data->inode);
1867 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1868 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1870 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1871 nfs_commit_end(cinfo.mds);
1874 static void nfs_commit_release(void *calldata)
1876 struct nfs_commit_data *data = calldata;
1878 data->completion_ops->completion(data);
1879 nfs_commitdata_release(calldata);
1882 static const struct rpc_call_ops nfs_commit_ops = {
1883 .rpc_call_prepare = nfs_commit_prepare,
1884 .rpc_call_done = nfs_commit_done,
1885 .rpc_release = nfs_commit_release,
1888 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1889 .completion = nfs_commit_release_pages,
1890 .resched_write = nfs_commit_resched_write,
1893 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1894 int how, struct nfs_commit_info *cinfo)
1898 status = pnfs_commit_list(inode, head, how, cinfo);
1899 if (status == PNFS_NOT_ATTEMPTED)
1900 status = nfs_commit_list(inode, head, how, cinfo);
1904 static int __nfs_commit_inode(struct inode *inode, int how,
1905 struct writeback_control *wbc)
1908 struct nfs_commit_info cinfo;
1909 int may_wait = how & FLUSH_SYNC;
1913 nfs_init_cinfo_from_inode(&cinfo, inode);
1914 nfs_commit_begin(cinfo.mds);
1916 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1919 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1923 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1924 if (nscan < wbc->nr_to_write)
1925 wbc->nr_to_write -= nscan;
1927 wbc->nr_to_write = 0;
1929 if (nscan < INT_MAX)
1933 nfs_commit_end(cinfo.mds);
1934 if (ret || !may_wait)
1936 return wait_on_commit(cinfo.mds);
1939 int nfs_commit_inode(struct inode *inode, int how)
1941 return __nfs_commit_inode(inode, how, NULL);
1943 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1945 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1947 struct nfs_inode *nfsi = NFS_I(inode);
1948 int flags = FLUSH_SYNC;
1951 if (wbc->sync_mode == WB_SYNC_NONE) {
1952 /* no commits means nothing needs to be done */
1953 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1954 goto check_requests_outstanding;
1956 /* Don't commit yet if this is a non-blocking flush and there
1957 * are a lot of outstanding writes for this mapping.
1959 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1960 goto out_mark_dirty;
1962 /* don't wait for the COMMIT response */
1966 ret = __nfs_commit_inode(inode, flags, wbc);
1968 if (flags & FLUSH_SYNC)
1970 } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1971 goto out_mark_dirty;
1973 check_requests_outstanding:
1974 if (!atomic_read(&nfsi->commit_info.rpcs_out))
1977 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1980 EXPORT_SYMBOL_GPL(nfs_write_inode);
1983 * Wrapper for filemap_write_and_wait_range()
1985 * Needed for pNFS in order to ensure data becomes visible to the
1988 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1989 loff_t lstart, loff_t lend)
1993 ret = filemap_write_and_wait_range(mapping, lstart, lend);
1995 ret = pnfs_sync_inode(mapping->host, true);
1998 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2001 * flush the inode to disk.
2003 int nfs_wb_all(struct inode *inode)
2007 trace_nfs_writeback_inode_enter(inode);
2009 ret = filemap_write_and_wait(inode->i_mapping);
2012 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2015 pnfs_sync_inode(inode, true);
2019 trace_nfs_writeback_inode_exit(inode, ret);
2022 EXPORT_SYMBOL_GPL(nfs_wb_all);
2024 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2026 struct nfs_page *req;
2029 wait_on_page_writeback(page);
2031 /* blocking call to cancel all requests and join to a single (head)
2033 req = nfs_lock_and_join_requests(page);
2038 /* all requests from this page have been cancelled by
2039 * nfs_lock_and_join_requests, so just remove the head
2040 * request from the inode / page_private pointer and
2042 nfs_inode_remove_request(req);
2043 nfs_unlock_and_release_request(req);
2050 * Write back all requests on one page - we do this before reading it.
2052 int nfs_wb_page(struct inode *inode, struct page *page)
2054 loff_t range_start = page_file_offset(page);
2055 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2056 struct writeback_control wbc = {
2057 .sync_mode = WB_SYNC_ALL,
2059 .range_start = range_start,
2060 .range_end = range_end,
2064 trace_nfs_writeback_page_enter(inode);
2067 wait_on_page_writeback(page);
2068 if (clear_page_dirty_for_io(page)) {
2069 ret = nfs_writepage_locked(page, &wbc);
2075 if (!PagePrivate(page))
2077 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2082 trace_nfs_writeback_page_exit(inode, ret);
2086 #ifdef CONFIG_MIGRATION
2087 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2088 struct page *page, enum migrate_mode mode)
2091 * If PagePrivate is set, then the page is currently associated with
2092 * an in-progress read or write request. Don't try to migrate it.
2094 * FIXME: we could do this in principle, but we'll need a way to ensure
2095 * that we can safely release the inode reference while holding
2098 if (PagePrivate(page))
2101 if (!nfs_fscache_release_page(page, GFP_KERNEL))
2104 return migrate_page(mapping, newpage, page, mode);
2108 int __init nfs_init_writepagecache(void)
2110 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2111 sizeof(struct nfs_pgio_header),
2112 0, SLAB_HWCACHE_ALIGN,
2114 if (nfs_wdata_cachep == NULL)
2117 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2119 if (nfs_wdata_mempool == NULL)
2120 goto out_destroy_write_cache;
2122 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2123 sizeof(struct nfs_commit_data),
2124 0, SLAB_HWCACHE_ALIGN,
2126 if (nfs_cdata_cachep == NULL)
2127 goto out_destroy_write_mempool;
2129 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2131 if (nfs_commit_mempool == NULL)
2132 goto out_destroy_commit_cache;
2135 * NFS congestion size, scale with available memory.
2147 * This allows larger machines to have larger/more transfers.
2148 * Limit the default to 256M
2150 nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2151 if (nfs_congestion_kb > 256*1024)
2152 nfs_congestion_kb = 256*1024;
2156 out_destroy_commit_cache:
2157 kmem_cache_destroy(nfs_cdata_cachep);
2158 out_destroy_write_mempool:
2159 mempool_destroy(nfs_wdata_mempool);
2160 out_destroy_write_cache:
2161 kmem_cache_destroy(nfs_wdata_cachep);
2165 void nfs_destroy_writepagecache(void)
2167 mempool_destroy(nfs_commit_mempool);
2168 kmem_cache_destroy(nfs_cdata_cachep);
2169 mempool_destroy(nfs_wdata_mempool);
2170 kmem_cache_destroy(nfs_wdata_cachep);
2173 static const struct nfs_rw_ops nfs_rw_write_ops = {
2174 .rw_alloc_header = nfs_writehdr_alloc,
2175 .rw_free_header = nfs_writehdr_free,
2176 .rw_done = nfs_writeback_done,
2177 .rw_result = nfs_writeback_result,
2178 .rw_initiate = nfs_initiate_write,