4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
9 #include <linux/types.h>
10 #include <linux/slab.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
23 #include <linux/export.h>
25 #include <asm/uaccess.h>
27 #include "delegation.h"
36 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
38 #define MIN_POOL_WRITE (32)
39 #define MIN_POOL_COMMIT (4)
42 * Local function declarations
44 static void nfs_redirty_request(struct nfs_page *req);
45 static const struct rpc_call_ops nfs_write_common_ops;
46 static const struct rpc_call_ops nfs_commit_ops;
47 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
48 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
50 static struct kmem_cache *nfs_wdata_cachep;
51 static mempool_t *nfs_wdata_mempool;
52 static struct kmem_cache *nfs_cdata_cachep;
53 static mempool_t *nfs_commit_mempool;
55 struct nfs_commit_data *nfs_commitdata_alloc(void)
57 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
60 memset(p, 0, sizeof(*p));
61 INIT_LIST_HEAD(&p->pages);
65 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
67 void nfs_commit_free(struct nfs_commit_data *p)
69 mempool_free(p, nfs_commit_mempool);
71 EXPORT_SYMBOL_GPL(nfs_commit_free);
73 struct nfs_rw_header *nfs_writehdr_alloc(void)
75 struct nfs_rw_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
78 struct nfs_pgio_header *hdr = &p->header;
80 memset(p, 0, sizeof(*p));
81 INIT_LIST_HEAD(&hdr->pages);
82 INIT_LIST_HEAD(&hdr->rpc_list);
83 spin_lock_init(&hdr->lock);
84 atomic_set(&hdr->refcnt, 0);
88 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
90 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
92 struct nfs_rw_header *whdr = container_of(hdr, struct nfs_rw_header, header);
93 mempool_free(whdr, nfs_wdata_mempool);
95 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
97 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
101 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
104 static struct nfs_page *
105 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
107 struct nfs_page *req = NULL;
109 if (PagePrivate(page))
110 req = (struct nfs_page *)page_private(page);
111 else if (unlikely(PageSwapCache(page))) {
112 struct nfs_page *freq, *t;
114 /* Linearly search the commit list for the correct req */
115 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
116 if (freq->wb_page == page) {
124 kref_get(&req->wb_kref);
129 static struct nfs_page *nfs_page_find_request(struct page *page)
131 struct inode *inode = page_file_mapping(page)->host;
132 struct nfs_page *req = NULL;
134 spin_lock(&inode->i_lock);
135 req = nfs_page_find_request_locked(NFS_I(inode), page);
136 spin_unlock(&inode->i_lock);
140 /* Adjust the file length if we're writing beyond the end */
141 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
143 struct inode *inode = page_file_mapping(page)->host;
147 spin_lock(&inode->i_lock);
148 i_size = i_size_read(inode);
149 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
150 if (i_size > 0 && page_file_index(page) < end_index)
152 end = page_file_offset(page) + ((loff_t)offset+count);
155 i_size_write(inode, end);
156 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
158 spin_unlock(&inode->i_lock);
161 /* A writeback failed: mark the page as bad, and invalidate the page cache */
162 static void nfs_set_pageerror(struct page *page)
164 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
167 /* We can set the PG_uptodate flag if we see that a write request
168 * covers the full page.
170 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
172 if (PageUptodate(page))
176 if (count != nfs_page_length(page))
178 SetPageUptodate(page);
181 static int wb_priority(struct writeback_control *wbc)
183 if (wbc->for_reclaim)
184 return FLUSH_HIGHPRI | FLUSH_STABLE;
185 if (wbc->for_kupdate || wbc->for_background)
186 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
187 return FLUSH_COND_STABLE;
191 * NFS congestion control
194 int nfs_congestion_kb;
196 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
197 #define NFS_CONGESTION_OFF_THRESH \
198 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
200 static void nfs_set_page_writeback(struct page *page)
202 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
203 int ret = test_set_page_writeback(page);
205 WARN_ON_ONCE(ret != 0);
207 if (atomic_long_inc_return(&nfss->writeback) >
208 NFS_CONGESTION_ON_THRESH) {
209 set_bdi_congested(&nfss->backing_dev_info,
214 static void nfs_end_page_writeback(struct page *page)
216 struct inode *inode = page_file_mapping(page)->host;
217 struct nfs_server *nfss = NFS_SERVER(inode);
219 end_page_writeback(page);
220 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
221 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
224 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
226 struct inode *inode = page_file_mapping(page)->host;
227 struct nfs_page *req;
230 spin_lock(&inode->i_lock);
232 req = nfs_page_find_request_locked(NFS_I(inode), page);
235 if (nfs_lock_request(req))
237 /* Note: If we hold the page lock, as is the case in nfs_writepage,
238 * then the call to nfs_lock_request() will always
239 * succeed provided that someone hasn't already marked the
240 * request as dirty (in which case we don't care).
242 spin_unlock(&inode->i_lock);
244 ret = nfs_wait_on_request(req);
247 nfs_release_request(req);
250 spin_lock(&inode->i_lock);
252 spin_unlock(&inode->i_lock);
257 * Find an associated nfs write request, and prepare to flush it out
258 * May return an error if the user signalled nfs_wait_on_request().
260 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
261 struct page *page, bool nonblock)
263 struct nfs_page *req;
266 req = nfs_find_and_lock_request(page, nonblock);
273 nfs_set_page_writeback(page);
274 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
277 if (!nfs_pageio_add_request(pgio, req)) {
278 nfs_redirty_request(req);
279 ret = pgio->pg_error;
285 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
287 struct inode *inode = page_file_mapping(page)->host;
290 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
291 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
293 nfs_pageio_cond_complete(pgio, page_file_index(page));
294 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
295 if (ret == -EAGAIN) {
296 redirty_page_for_writepage(wbc, page);
303 * Write an mmapped page to the server.
305 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
307 struct nfs_pageio_descriptor pgio;
310 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
311 false, &nfs_async_write_completion_ops);
312 err = nfs_do_writepage(page, wbc, &pgio);
313 nfs_pageio_complete(&pgio);
316 if (pgio.pg_error < 0)
317 return pgio.pg_error;
321 int nfs_writepage(struct page *page, struct writeback_control *wbc)
325 ret = nfs_writepage_locked(page, wbc);
330 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
334 ret = nfs_do_writepage(page, wbc, data);
339 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
341 struct inode *inode = mapping->host;
342 unsigned long *bitlock = &NFS_I(inode)->flags;
343 struct nfs_pageio_descriptor pgio;
346 /* Stop dirtying of new pages while we sync */
347 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
348 nfs_wait_bit_killable, TASK_KILLABLE);
352 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
354 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
355 &nfs_async_write_completion_ops);
356 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
357 nfs_pageio_complete(&pgio);
359 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
360 smp_mb__after_clear_bit();
361 wake_up_bit(bitlock, NFS_INO_FLUSHING);
374 * Insert a write request into an inode
376 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
378 struct nfs_inode *nfsi = NFS_I(inode);
380 /* Lock the request! */
381 nfs_lock_request(req);
383 spin_lock(&inode->i_lock);
384 if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
387 * Swap-space should not get truncated. Hence no need to plug the race
388 * with invalidate/truncate.
390 if (likely(!PageSwapCache(req->wb_page))) {
391 set_bit(PG_MAPPED, &req->wb_flags);
392 SetPagePrivate(req->wb_page);
393 set_page_private(req->wb_page, (unsigned long)req);
396 kref_get(&req->wb_kref);
397 spin_unlock(&inode->i_lock);
401 * Remove a write request from an inode
403 static void nfs_inode_remove_request(struct nfs_page *req)
405 struct inode *inode = req->wb_context->dentry->d_inode;
406 struct nfs_inode *nfsi = NFS_I(inode);
408 spin_lock(&inode->i_lock);
409 if (likely(!PageSwapCache(req->wb_page))) {
410 set_page_private(req->wb_page, 0);
411 ClearPagePrivate(req->wb_page);
412 clear_bit(PG_MAPPED, &req->wb_flags);
415 spin_unlock(&inode->i_lock);
416 nfs_release_request(req);
420 nfs_mark_request_dirty(struct nfs_page *req)
422 __set_page_dirty_nobuffers(req->wb_page);
425 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
427 * nfs_request_add_commit_list - add request to a commit list
428 * @req: pointer to a struct nfs_page
429 * @dst: commit list head
430 * @cinfo: holds list lock and accounting info
432 * This sets the PG_CLEAN bit, updates the cinfo count of
433 * number of outstanding requests requiring a commit as well as
436 * The caller must _not_ hold the cinfo->lock, but must be
437 * holding the nfs_page lock.
440 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
441 struct nfs_commit_info *cinfo)
443 set_bit(PG_CLEAN, &(req)->wb_flags);
444 spin_lock(cinfo->lock);
445 nfs_list_add_request(req, dst);
446 cinfo->mds->ncommit++;
447 spin_unlock(cinfo->lock);
449 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
450 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
452 __mark_inode_dirty(req->wb_context->dentry->d_inode,
456 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
459 * nfs_request_remove_commit_list - Remove request from a commit list
460 * @req: pointer to a nfs_page
461 * @cinfo: holds list lock and accounting info
463 * This clears the PG_CLEAN bit, and updates the cinfo's count of
464 * number of outstanding requests requiring a commit
465 * It does not update the MM page stats.
467 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
470 nfs_request_remove_commit_list(struct nfs_page *req,
471 struct nfs_commit_info *cinfo)
473 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
475 nfs_list_remove_request(req);
476 cinfo->mds->ncommit--;
478 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
480 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
483 cinfo->lock = &inode->i_lock;
484 cinfo->mds = &NFS_I(inode)->commit_info;
485 cinfo->ds = pnfs_get_ds_info(inode);
487 cinfo->completion_ops = &nfs_commit_completion_ops;
490 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
492 struct nfs_direct_req *dreq)
495 nfs_init_cinfo_from_dreq(cinfo, dreq);
497 nfs_init_cinfo_from_inode(cinfo, inode);
499 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
502 * Add a request to the inode's commit list.
505 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
506 struct nfs_commit_info *cinfo)
508 if (pnfs_mark_request_commit(req, lseg, cinfo))
510 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
514 nfs_clear_page_commit(struct page *page)
516 dec_zone_page_state(page, NR_UNSTABLE_NFS);
517 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
521 nfs_clear_request_commit(struct nfs_page *req)
523 if (test_bit(PG_CLEAN, &req->wb_flags)) {
524 struct inode *inode = req->wb_context->dentry->d_inode;
525 struct nfs_commit_info cinfo;
527 nfs_init_cinfo_from_inode(&cinfo, inode);
528 if (!pnfs_clear_request_commit(req, &cinfo)) {
529 spin_lock(cinfo.lock);
530 nfs_request_remove_commit_list(req, &cinfo);
531 spin_unlock(cinfo.lock);
533 nfs_clear_page_commit(req->wb_page);
538 int nfs_write_need_commit(struct nfs_pgio_data *data)
540 if (data->verf.committed == NFS_DATA_SYNC)
541 return data->header->lseg == NULL;
542 return data->verf.committed != NFS_FILE_SYNC;
546 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
551 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
553 struct nfs_direct_req *dreq)
558 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
559 struct nfs_commit_info *cinfo)
564 nfs_clear_request_commit(struct nfs_page *req)
569 int nfs_write_need_commit(struct nfs_pgio_data *data)
576 static void nfs_write_completion(struct nfs_pgio_header *hdr)
578 struct nfs_commit_info cinfo;
579 unsigned long bytes = 0;
581 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
583 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
584 while (!list_empty(&hdr->pages)) {
585 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
587 bytes += req->wb_bytes;
588 nfs_list_remove_request(req);
589 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
590 (hdr->good_bytes < bytes)) {
591 nfs_set_pageerror(req->wb_page);
592 nfs_context_set_write_error(req->wb_context, hdr->error);
595 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
596 nfs_mark_request_dirty(req);
599 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
600 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
601 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
605 nfs_inode_remove_request(req);
607 nfs_unlock_request(req);
608 nfs_end_page_writeback(req->wb_page);
609 nfs_release_request(req);
615 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
617 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
619 return cinfo->mds->ncommit;
622 /* cinfo->lock held by caller */
624 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
625 struct nfs_commit_info *cinfo, int max)
627 struct nfs_page *req, *tmp;
630 list_for_each_entry_safe(req, tmp, src, wb_list) {
631 if (!nfs_lock_request(req))
633 kref_get(&req->wb_kref);
634 if (cond_resched_lock(cinfo->lock))
635 list_safe_reset_next(req, tmp, wb_list);
636 nfs_request_remove_commit_list(req, cinfo);
637 nfs_list_add_request(req, dst);
639 if ((ret == max) && !cinfo->dreq)
646 * nfs_scan_commit - Scan an inode for commit requests
647 * @inode: NFS inode to scan
648 * @dst: mds destination list
649 * @cinfo: mds and ds lists of reqs ready to commit
651 * Moves requests from the inode's 'commit' request list.
652 * The requests are *not* checked to ensure that they form a contiguous set.
655 nfs_scan_commit(struct inode *inode, struct list_head *dst,
656 struct nfs_commit_info *cinfo)
660 spin_lock(cinfo->lock);
661 if (cinfo->mds->ncommit > 0) {
662 const int max = INT_MAX;
664 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
666 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
668 spin_unlock(cinfo->lock);
673 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
678 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
679 struct nfs_commit_info *cinfo)
686 * Search for an existing write request, and attempt to update
687 * it to reflect a new dirty region on a given page.
689 * If the attempt fails, then the existing request is flushed out
692 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
697 struct nfs_page *req;
702 if (!PagePrivate(page))
705 end = offset + bytes;
706 spin_lock(&inode->i_lock);
709 req = nfs_page_find_request_locked(NFS_I(inode), page);
713 rqend = req->wb_offset + req->wb_bytes;
715 * Tell the caller to flush out the request if
716 * the offsets are non-contiguous.
717 * Note: nfs_flush_incompatible() will already
718 * have flushed out requests having wrong owners.
721 || end < req->wb_offset)
724 if (nfs_lock_request(req))
727 /* The request is locked, so wait and then retry */
728 spin_unlock(&inode->i_lock);
729 error = nfs_wait_on_request(req);
730 nfs_release_request(req);
733 spin_lock(&inode->i_lock);
736 /* Okay, the request matches. Update the region */
737 if (offset < req->wb_offset) {
738 req->wb_offset = offset;
739 req->wb_pgbase = offset;
742 req->wb_bytes = end - req->wb_offset;
744 req->wb_bytes = rqend - req->wb_offset;
746 spin_unlock(&inode->i_lock);
748 nfs_clear_request_commit(req);
751 spin_unlock(&inode->i_lock);
752 nfs_release_request(req);
753 error = nfs_wb_page(inode, page);
755 return ERR_PTR(error);
759 * Try to update an existing write request, or create one if there is none.
761 * Note: Should always be called with the Page Lock held to prevent races
762 * if we have to add a new request. Also assumes that the caller has
763 * already called nfs_flush_incompatible() if necessary.
765 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
766 struct page *page, unsigned int offset, unsigned int bytes)
768 struct inode *inode = page_file_mapping(page)->host;
769 struct nfs_page *req;
771 req = nfs_try_to_update_request(inode, page, offset, bytes);
774 req = nfs_create_request(ctx, inode, page, offset, bytes);
777 nfs_inode_add_request(inode, req);
782 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
783 unsigned int offset, unsigned int count)
785 struct nfs_page *req;
787 req = nfs_setup_write_request(ctx, page, offset, count);
790 /* Update file length */
791 nfs_grow_file(page, offset, count);
792 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
793 nfs_mark_request_dirty(req);
794 nfs_unlock_and_release_request(req);
798 int nfs_flush_incompatible(struct file *file, struct page *page)
800 struct nfs_open_context *ctx = nfs_file_open_context(file);
801 struct nfs_lock_context *l_ctx;
802 struct nfs_page *req;
803 int do_flush, status;
805 * Look for a request corresponding to this page. If there
806 * is one, and it belongs to another file, we flush it out
807 * before we try to copy anything into the page. Do this
808 * due to the lack of an ACCESS-type call in NFSv2.
809 * Also do the same if we find a request from an existing
813 req = nfs_page_find_request(page);
816 l_ctx = req->wb_lock_context;
817 do_flush = req->wb_page != page || req->wb_context != ctx;
818 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
819 do_flush |= l_ctx->lockowner.l_owner != current->files
820 || l_ctx->lockowner.l_pid != current->tgid;
822 nfs_release_request(req);
825 status = nfs_wb_page(page_file_mapping(page)->host, page);
826 } while (status == 0);
831 * Avoid buffered writes when a open context credential's key would
834 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
836 * Return 0 and set a credential flag which triggers the inode to flush
837 * and performs NFS_FILE_SYNC writes if the key will expired within
838 * RPC_KEY_EXPIRE_TIMEO.
841 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
843 struct nfs_open_context *ctx = nfs_file_open_context(filp);
844 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
846 return rpcauth_key_timeout_notify(auth, ctx->cred);
850 * Test if the open context credential key is marked to expire soon.
852 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
854 return rpcauth_cred_key_to_expire(ctx->cred);
858 * If the page cache is marked as unsafe or invalid, then we can't rely on
859 * the PageUptodate() flag. In this case, we will need to turn off
860 * write optimisations that depend on the page contents being correct.
862 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
864 struct nfs_inode *nfsi = NFS_I(inode);
866 if (nfs_have_delegated_attributes(inode))
868 if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
871 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
874 return PageUptodate(page) != 0;
877 /* If we know the page is up to date, and we're not using byte range locks (or
878 * if we have the whole file locked for writing), it may be more efficient to
879 * extend the write to cover the entire page in order to avoid fragmentation
882 * If the file is opened for synchronous writes then we can just skip the rest
885 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
887 if (file->f_flags & O_DSYNC)
889 if (!nfs_write_pageuptodate(page, inode))
891 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
893 if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
894 inode->i_flock->fl_end == OFFSET_MAX &&
895 inode->i_flock->fl_type != F_RDLCK))
901 * Update and possibly write a cached page of an NFS file.
903 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
904 * things with a page scheduled for an RPC call (e.g. invalidate it).
906 int nfs_updatepage(struct file *file, struct page *page,
907 unsigned int offset, unsigned int count)
909 struct nfs_open_context *ctx = nfs_file_open_context(file);
910 struct inode *inode = page_file_mapping(page)->host;
913 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
915 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
916 file, count, (long long)(page_file_offset(page) + offset));
918 if (nfs_can_extend_write(file, page, inode)) {
919 count = max(count + offset, nfs_page_length(page));
923 status = nfs_writepage_setup(ctx, page, offset, count);
925 nfs_set_pageerror(page);
927 __set_page_dirty_nobuffers(page);
929 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
930 status, (long long)i_size_read(inode));
934 static int flush_task_priority(int how)
936 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
938 return RPC_PRIORITY_HIGH;
940 return RPC_PRIORITY_LOW;
942 return RPC_PRIORITY_NORMAL;
945 int nfs_initiate_write(struct rpc_clnt *clnt,
946 struct nfs_pgio_data *data,
947 const struct rpc_call_ops *call_ops,
950 struct inode *inode = data->header->inode;
951 int priority = flush_task_priority(how);
952 struct rpc_task *task;
953 struct rpc_message msg = {
954 .rpc_argp = &data->args,
955 .rpc_resp = &data->res,
956 .rpc_cred = data->header->cred,
958 struct rpc_task_setup task_setup_data = {
962 .callback_ops = call_ops,
963 .callback_data = data,
964 .workqueue = nfsiod_workqueue,
965 .flags = RPC_TASK_ASYNC | flags,
966 .priority = priority,
970 /* Set up the initial task struct. */
971 NFS_PROTO(inode)->write_setup(data, &msg);
973 dprintk("NFS: %5u initiated write call "
974 "(req %s/%llu, %u bytes @ offset %llu)\n",
977 (unsigned long long)NFS_FILEID(inode),
979 (unsigned long long)data->args.offset);
981 nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
982 &task_setup_data.rpc_client, &msg, data);
984 task = rpc_run_task(&task_setup_data);
989 if (how & FLUSH_SYNC) {
990 ret = rpc_wait_for_completion_task(task);
992 ret = task->tk_status;
998 EXPORT_SYMBOL_GPL(nfs_initiate_write);
1001 * Set up the argument/result storage required for the RPC call.
1003 static void nfs_write_rpcsetup(struct nfs_pgio_data *data,
1004 unsigned int count, unsigned int offset,
1005 int how, struct nfs_commit_info *cinfo)
1007 struct nfs_page *req = data->header->req;
1009 /* Set up the RPC argument and reply structs
1010 * NB: take care not to mess about with data->commit et al. */
1012 data->args.fh = NFS_FH(data->header->inode);
1013 data->args.offset = req_offset(req) + offset;
1014 /* pnfs_set_layoutcommit needs this */
1015 data->mds_offset = data->args.offset;
1016 data->args.pgbase = req->wb_pgbase + offset;
1017 data->args.pages = data->pages.pagevec;
1018 data->args.count = count;
1019 data->args.context = get_nfs_open_context(req->wb_context);
1020 data->args.lock_context = req->wb_lock_context;
1021 data->args.stable = NFS_UNSTABLE;
1022 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1025 case FLUSH_COND_STABLE:
1026 if (nfs_reqs_to_commit(cinfo))
1029 data->args.stable = NFS_FILE_SYNC;
1032 data->res.fattr = &data->fattr;
1033 data->res.count = count;
1034 data->res.verf = &data->verf;
1035 nfs_fattr_init(&data->fattr);
1038 static int nfs_do_write(struct nfs_pgio_data *data,
1039 const struct rpc_call_ops *call_ops,
1042 struct inode *inode = data->header->inode;
1044 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1047 static int nfs_do_multiple_writes(struct list_head *head,
1048 const struct rpc_call_ops *call_ops,
1051 struct nfs_pgio_data *data;
1054 while (!list_empty(head)) {
1057 data = list_first_entry(head, struct nfs_pgio_data, list);
1058 list_del_init(&data->list);
1060 ret2 = nfs_do_write(data, call_ops, how);
1067 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1068 * call this on each, which will prepare them to be retried on next
1069 * writeback using standard nfs.
1071 static void nfs_redirty_request(struct nfs_page *req)
1073 nfs_mark_request_dirty(req);
1074 nfs_unlock_request(req);
1075 nfs_end_page_writeback(req->wb_page);
1076 nfs_release_request(req);
1079 static void nfs_async_write_error(struct list_head *head)
1081 struct nfs_page *req;
1083 while (!list_empty(head)) {
1084 req = nfs_list_entry(head->next);
1085 nfs_list_remove_request(req);
1086 nfs_redirty_request(req);
1090 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1091 .error_cleanup = nfs_async_write_error,
1092 .completion = nfs_write_completion,
1095 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1096 struct nfs_pgio_header *hdr)
1098 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1099 while (!list_empty(&hdr->rpc_list)) {
1100 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
1101 struct nfs_pgio_data, list);
1102 list_del(&data->list);
1103 nfs_pgio_data_release(data);
1105 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1109 * Generate multiple small requests to write out a single
1110 * contiguous dirty area on one page.
1112 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1113 struct nfs_pgio_header *hdr)
1115 struct nfs_page *req = hdr->req;
1116 struct page *page = req->wb_page;
1117 struct nfs_pgio_data *data;
1118 size_t wsize = desc->pg_bsize, nbytes;
1119 unsigned int offset;
1121 struct nfs_commit_info cinfo;
1123 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1125 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1126 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1127 desc->pg_count > wsize))
1128 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1132 nbytes = desc->pg_count;
1134 size_t len = min(nbytes, wsize);
1136 data = nfs_pgio_data_alloc(hdr, 1);
1138 nfs_flush_error(desc, hdr);
1141 data->pages.pagevec[0] = page;
1142 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1143 list_add(&data->list, &hdr->rpc_list);
1147 } while (nbytes != 0);
1148 nfs_list_remove_request(req);
1149 nfs_list_add_request(req, &hdr->pages);
1150 desc->pg_rpc_callops = &nfs_write_common_ops;
1155 * Create an RPC task for the given write request and kick it.
1156 * The page must have been locked by the caller.
1158 * It may happen that the page we're passed is not marked dirty.
1159 * This is the case if nfs_updatepage detects a conflicting request
1160 * that has been written but not committed.
1162 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1163 struct nfs_pgio_header *hdr)
1165 struct nfs_page *req;
1166 struct page **pages;
1167 struct nfs_pgio_data *data;
1168 struct list_head *head = &desc->pg_list;
1169 struct nfs_commit_info cinfo;
1171 data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
1174 nfs_flush_error(desc, hdr);
1178 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1179 pages = data->pages.pagevec;
1180 while (!list_empty(head)) {
1181 req = nfs_list_entry(head->next);
1182 nfs_list_remove_request(req);
1183 nfs_list_add_request(req, &hdr->pages);
1184 *pages++ = req->wb_page;
1187 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1188 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1189 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1191 /* Set up the argument struct */
1192 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1193 list_add(&data->list, &hdr->rpc_list);
1194 desc->pg_rpc_callops = &nfs_write_common_ops;
1198 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1199 struct nfs_pgio_header *hdr)
1201 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1202 return nfs_flush_multi(desc, hdr);
1203 return nfs_flush_one(desc, hdr);
1205 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1207 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1209 struct nfs_rw_header *whdr;
1210 struct nfs_pgio_header *hdr;
1213 whdr = nfs_writehdr_alloc();
1215 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1218 hdr = &whdr->header;
1219 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1220 atomic_inc(&hdr->refcnt);
1221 ret = nfs_generic_flush(desc, hdr);
1223 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1224 desc->pg_rpc_callops,
1226 if (atomic_dec_and_test(&hdr->refcnt))
1227 hdr->completion_ops->completion(hdr);
1231 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1232 .pg_test = nfs_generic_pg_test,
1233 .pg_doio = nfs_generic_pg_writepages,
1236 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1237 struct inode *inode, int ioflags, bool force_mds,
1238 const struct nfs_pgio_completion_ops *compl_ops)
1240 struct nfs_server *server = NFS_SERVER(inode);
1241 const struct nfs_pageio_ops *pg_ops = &nfs_pageio_write_ops;
1243 #ifdef CONFIG_NFS_V4_1
1244 if (server->pnfs_curr_ld && !force_mds)
1245 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1247 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, server->wsize, ioflags);
1249 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1251 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1253 pgio->pg_ops = &nfs_pageio_write_ops;
1254 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1256 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1259 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1261 struct nfs_pgio_data *data = calldata;
1263 err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1265 rpc_exit(task, err);
1268 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1270 struct nfs_commit_data *data = calldata;
1272 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1276 * Handle a write reply that flushes a whole page.
1278 * FIXME: There is an inherent race with invalidate_inode_pages and
1279 * writebacks since the page->count is kept > 1 for as long
1280 * as the page has a write request pending.
1282 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1284 struct nfs_pgio_data *data = calldata;
1286 nfs_writeback_done(task, data);
1289 static void nfs_writeback_release_common(void *calldata)
1291 struct nfs_pgio_data *data = calldata;
1292 struct nfs_pgio_header *hdr = data->header;
1293 int status = data->task.tk_status;
1295 if ((status >= 0) && nfs_write_need_commit(data)) {
1296 spin_lock(&hdr->lock);
1297 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1299 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1300 memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf));
1301 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf)))
1302 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1303 spin_unlock(&hdr->lock);
1305 nfs_pgio_data_release(data);
1308 static const struct rpc_call_ops nfs_write_common_ops = {
1309 .rpc_call_prepare = nfs_write_prepare,
1310 .rpc_call_done = nfs_writeback_done_common,
1311 .rpc_release = nfs_writeback_release_common,
1315 * Special version of should_remove_suid() that ignores capabilities.
1317 static int nfs_should_remove_suid(const struct inode *inode)
1319 umode_t mode = inode->i_mode;
1322 /* suid always must be killed */
1323 if (unlikely(mode & S_ISUID))
1324 kill = ATTR_KILL_SUID;
1327 * sgid without any exec bits is just a mandatory locking mark; leave
1328 * it alone. If some exec bits are set, it's a real sgid; kill it.
1330 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1331 kill |= ATTR_KILL_SGID;
1333 if (unlikely(kill && S_ISREG(mode)))
1340 * This function is called when the WRITE call is complete.
1342 void nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data)
1344 struct nfs_pgio_args *argp = &data->args;
1345 struct nfs_pgio_res *resp = &data->res;
1346 struct inode *inode = data->header->inode;
1349 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1350 task->tk_pid, task->tk_status);
1353 * ->write_done will attempt to use post-op attributes to detect
1354 * conflicting writes by other clients. A strict interpretation
1355 * of close-to-open would allow us to continue caching even if
1356 * another writer had changed the file, but some applications
1357 * depend on tighter cache coherency when writing.
1359 status = NFS_PROTO(inode)->write_done(task, data);
1362 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1364 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1365 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1366 /* We tried a write call, but the server did not
1367 * commit data to stable storage even though we
1369 * Note: There is a known bug in Tru64 < 5.0 in which
1370 * the server reports NFS_DATA_SYNC, but performs
1371 * NFS_FILE_SYNC. We therefore implement this checking
1372 * as a dprintk() in order to avoid filling syslog.
1374 static unsigned long complain;
1376 /* Note this will print the MDS for a DS write */
1377 if (time_before(complain, jiffies)) {
1378 dprintk("NFS: faulty NFS server %s:"
1379 " (committed = %d) != (stable = %d)\n",
1380 NFS_SERVER(inode)->nfs_client->cl_hostname,
1381 resp->verf->committed, argp->stable);
1382 complain = jiffies + 300 * HZ;
1386 if (task->tk_status < 0) {
1387 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1391 /* Deal with the suid/sgid bit corner case */
1392 if (nfs_should_remove_suid(inode))
1393 nfs_mark_for_revalidate(inode);
1395 if (resp->count < argp->count) {
1396 static unsigned long complain;
1398 /* This a short write! */
1399 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1401 /* Has the server at least made some progress? */
1402 if (resp->count == 0) {
1403 if (time_before(complain, jiffies)) {
1405 "NFS: Server wrote zero bytes, expected %u.\n",
1407 complain = jiffies + 300 * HZ;
1409 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1410 task->tk_status = -EIO;
1413 /* Was this an NFSv2 write or an NFSv3 stable write? */
1414 if (resp->verf->committed != NFS_UNSTABLE) {
1415 /* Resend from where the server left off */
1416 data->mds_offset += resp->count;
1417 argp->offset += resp->count;
1418 argp->pgbase += resp->count;
1419 argp->count -= resp->count;
1421 /* Resend as a stable write in order to avoid
1422 * headaches in the case of a server crash.
1424 argp->stable = NFS_FILE_SYNC;
1426 rpc_restart_call_prepare(task);
1431 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1432 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1436 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1440 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1442 nfs_wait_bit_killable,
1444 return (ret < 0) ? ret : 1;
1447 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1449 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1450 smp_mb__after_clear_bit();
1451 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1454 void nfs_commitdata_release(struct nfs_commit_data *data)
1456 put_nfs_open_context(data->context);
1457 nfs_commit_free(data);
1459 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1461 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1462 const struct rpc_call_ops *call_ops,
1465 struct rpc_task *task;
1466 int priority = flush_task_priority(how);
1467 struct rpc_message msg = {
1468 .rpc_argp = &data->args,
1469 .rpc_resp = &data->res,
1470 .rpc_cred = data->cred,
1472 struct rpc_task_setup task_setup_data = {
1473 .task = &data->task,
1475 .rpc_message = &msg,
1476 .callback_ops = call_ops,
1477 .callback_data = data,
1478 .workqueue = nfsiod_workqueue,
1479 .flags = RPC_TASK_ASYNC | flags,
1480 .priority = priority,
1482 /* Set up the initial task struct. */
1483 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1485 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1487 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1488 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1490 task = rpc_run_task(&task_setup_data);
1492 return PTR_ERR(task);
1493 if (how & FLUSH_SYNC)
1494 rpc_wait_for_completion_task(task);
1498 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1501 * Set up the argument/result storage required for the RPC call.
1503 void nfs_init_commit(struct nfs_commit_data *data,
1504 struct list_head *head,
1505 struct pnfs_layout_segment *lseg,
1506 struct nfs_commit_info *cinfo)
1508 struct nfs_page *first = nfs_list_entry(head->next);
1509 struct inode *inode = first->wb_context->dentry->d_inode;
1511 /* Set up the RPC argument and reply structs
1512 * NB: take care not to mess about with data->commit et al. */
1514 list_splice_init(head, &data->pages);
1516 data->inode = inode;
1517 data->cred = first->wb_context->cred;
1518 data->lseg = lseg; /* reference transferred */
1519 data->mds_ops = &nfs_commit_ops;
1520 data->completion_ops = cinfo->completion_ops;
1521 data->dreq = cinfo->dreq;
1523 data->args.fh = NFS_FH(data->inode);
1524 /* Note: we always request a commit of the entire inode */
1525 data->args.offset = 0;
1526 data->args.count = 0;
1527 data->context = get_nfs_open_context(first->wb_context);
1528 data->res.fattr = &data->fattr;
1529 data->res.verf = &data->verf;
1530 nfs_fattr_init(&data->fattr);
1532 EXPORT_SYMBOL_GPL(nfs_init_commit);
1534 void nfs_retry_commit(struct list_head *page_list,
1535 struct pnfs_layout_segment *lseg,
1536 struct nfs_commit_info *cinfo)
1538 struct nfs_page *req;
1540 while (!list_empty(page_list)) {
1541 req = nfs_list_entry(page_list->next);
1542 nfs_list_remove_request(req);
1543 nfs_mark_request_commit(req, lseg, cinfo);
1545 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1546 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1549 nfs_unlock_and_release_request(req);
1552 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1555 * Commit dirty pages
1558 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1559 struct nfs_commit_info *cinfo)
1561 struct nfs_commit_data *data;
1563 data = nfs_commitdata_alloc();
1568 /* Set up the argument struct */
1569 nfs_init_commit(data, head, NULL, cinfo);
1570 atomic_inc(&cinfo->mds->rpcs_out);
1571 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1574 nfs_retry_commit(head, NULL, cinfo);
1575 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1580 * COMMIT call returned
1582 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1584 struct nfs_commit_data *data = calldata;
1586 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1587 task->tk_pid, task->tk_status);
1589 /* Call the NFS version-specific code */
1590 NFS_PROTO(data->inode)->commit_done(task, data);
1593 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1595 struct nfs_page *req;
1596 int status = data->task.tk_status;
1597 struct nfs_commit_info cinfo;
1599 while (!list_empty(&data->pages)) {
1600 req = nfs_list_entry(data->pages.next);
1601 nfs_list_remove_request(req);
1602 nfs_clear_page_commit(req->wb_page);
1604 dprintk("NFS: commit (%s/%llu %d@%lld)",
1605 req->wb_context->dentry->d_sb->s_id,
1606 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1608 (long long)req_offset(req));
1610 nfs_context_set_write_error(req->wb_context, status);
1611 nfs_inode_remove_request(req);
1612 dprintk(", error = %d\n", status);
1616 /* Okay, COMMIT succeeded, apparently. Check the verifier
1617 * returned by the server against all stored verfs. */
1618 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1619 /* We have a match */
1620 nfs_inode_remove_request(req);
1624 /* We have a mismatch. Write the page again */
1625 dprintk(" mismatch\n");
1626 nfs_mark_request_dirty(req);
1627 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1629 nfs_unlock_and_release_request(req);
1631 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1632 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1633 nfs_commit_clear_lock(NFS_I(data->inode));
1636 static void nfs_commit_release(void *calldata)
1638 struct nfs_commit_data *data = calldata;
1640 data->completion_ops->completion(data);
1641 nfs_commitdata_release(calldata);
1644 static const struct rpc_call_ops nfs_commit_ops = {
1645 .rpc_call_prepare = nfs_commit_prepare,
1646 .rpc_call_done = nfs_commit_done,
1647 .rpc_release = nfs_commit_release,
1650 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1651 .completion = nfs_commit_release_pages,
1652 .error_cleanup = nfs_commit_clear_lock,
1655 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1656 int how, struct nfs_commit_info *cinfo)
1660 status = pnfs_commit_list(inode, head, how, cinfo);
1661 if (status == PNFS_NOT_ATTEMPTED)
1662 status = nfs_commit_list(inode, head, how, cinfo);
1666 int nfs_commit_inode(struct inode *inode, int how)
1669 struct nfs_commit_info cinfo;
1670 int may_wait = how & FLUSH_SYNC;
1673 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1675 goto out_mark_dirty;
1676 nfs_init_cinfo_from_inode(&cinfo, inode);
1677 res = nfs_scan_commit(inode, &head, &cinfo);
1681 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1685 goto out_mark_dirty;
1686 error = wait_on_bit(&NFS_I(inode)->flags,
1688 nfs_wait_bit_killable,
1693 nfs_commit_clear_lock(NFS_I(inode));
1695 /* Note: If we exit without ensuring that the commit is complete,
1696 * we must mark the inode as dirty. Otherwise, future calls to
1697 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1698 * that the data is on the disk.
1701 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1705 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1707 struct nfs_inode *nfsi = NFS_I(inode);
1708 int flags = FLUSH_SYNC;
1711 /* no commits means nothing needs to be done */
1712 if (!nfsi->commit_info.ncommit)
1715 if (wbc->sync_mode == WB_SYNC_NONE) {
1716 /* Don't commit yet if this is a non-blocking flush and there
1717 * are a lot of outstanding writes for this mapping.
1719 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1720 goto out_mark_dirty;
1722 /* don't wait for the COMMIT response */
1726 ret = nfs_commit_inode(inode, flags);
1728 if (wbc->sync_mode == WB_SYNC_NONE) {
1729 if (ret < wbc->nr_to_write)
1730 wbc->nr_to_write -= ret;
1732 wbc->nr_to_write = 0;
1737 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1741 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1747 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1749 return nfs_commit_unstable_pages(inode, wbc);
1751 EXPORT_SYMBOL_GPL(nfs_write_inode);
1754 * flush the inode to disk.
1756 int nfs_wb_all(struct inode *inode)
1758 struct writeback_control wbc = {
1759 .sync_mode = WB_SYNC_ALL,
1760 .nr_to_write = LONG_MAX,
1762 .range_end = LLONG_MAX,
1766 trace_nfs_writeback_inode_enter(inode);
1768 ret = sync_inode(inode, &wbc);
1770 trace_nfs_writeback_inode_exit(inode, ret);
1773 EXPORT_SYMBOL_GPL(nfs_wb_all);
1775 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1777 struct nfs_page *req;
1781 wait_on_page_writeback(page);
1782 req = nfs_page_find_request(page);
1785 if (nfs_lock_request(req)) {
1786 nfs_clear_request_commit(req);
1787 nfs_inode_remove_request(req);
1789 * In case nfs_inode_remove_request has marked the
1790 * page as being dirty
1792 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1793 nfs_unlock_and_release_request(req);
1796 ret = nfs_wait_on_request(req);
1797 nfs_release_request(req);
1805 * Write back all requests on one page - we do this before reading it.
1807 int nfs_wb_page(struct inode *inode, struct page *page)
1809 loff_t range_start = page_file_offset(page);
1810 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1811 struct writeback_control wbc = {
1812 .sync_mode = WB_SYNC_ALL,
1814 .range_start = range_start,
1815 .range_end = range_end,
1819 trace_nfs_writeback_page_enter(inode);
1822 wait_on_page_writeback(page);
1823 if (clear_page_dirty_for_io(page)) {
1824 ret = nfs_writepage_locked(page, &wbc);
1830 if (!PagePrivate(page))
1832 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1837 trace_nfs_writeback_page_exit(inode, ret);
1841 #ifdef CONFIG_MIGRATION
1842 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1843 struct page *page, enum migrate_mode mode)
1846 * If PagePrivate is set, then the page is currently associated with
1847 * an in-progress read or write request. Don't try to migrate it.
1849 * FIXME: we could do this in principle, but we'll need a way to ensure
1850 * that we can safely release the inode reference while holding
1853 if (PagePrivate(page))
1856 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1859 return migrate_page(mapping, newpage, page, mode);
1863 int __init nfs_init_writepagecache(void)
1865 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1866 sizeof(struct nfs_rw_header),
1867 0, SLAB_HWCACHE_ALIGN,
1869 if (nfs_wdata_cachep == NULL)
1872 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1874 if (nfs_wdata_mempool == NULL)
1875 goto out_destroy_write_cache;
1877 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1878 sizeof(struct nfs_commit_data),
1879 0, SLAB_HWCACHE_ALIGN,
1881 if (nfs_cdata_cachep == NULL)
1882 goto out_destroy_write_mempool;
1884 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1886 if (nfs_commit_mempool == NULL)
1887 goto out_destroy_commit_cache;
1890 * NFS congestion size, scale with available memory.
1902 * This allows larger machines to have larger/more transfers.
1903 * Limit the default to 256M
1905 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1906 if (nfs_congestion_kb > 256*1024)
1907 nfs_congestion_kb = 256*1024;
1911 out_destroy_commit_cache:
1912 kmem_cache_destroy(nfs_cdata_cachep);
1913 out_destroy_write_mempool:
1914 mempool_destroy(nfs_wdata_mempool);
1915 out_destroy_write_cache:
1916 kmem_cache_destroy(nfs_wdata_cachep);
1920 void nfs_destroy_writepagecache(void)
1922 mempool_destroy(nfs_commit_mempool);
1923 kmem_cache_destroy(nfs_cdata_cachep);
1924 mempool_destroy(nfs_wdata_mempool);
1925 kmem_cache_destroy(nfs_wdata_cachep);