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"
34 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
36 #define MIN_POOL_WRITE (32)
37 #define MIN_POOL_COMMIT (4)
40 * Local function declarations
42 static void nfs_redirty_request(struct nfs_page *req);
43 static const struct rpc_call_ops nfs_write_common_ops;
44 static const struct rpc_call_ops nfs_commit_ops;
45 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
46 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
48 static struct kmem_cache *nfs_wdata_cachep;
49 static mempool_t *nfs_wdata_mempool;
50 static struct kmem_cache *nfs_cdata_cachep;
51 static mempool_t *nfs_commit_mempool;
53 struct nfs_commit_data *nfs_commitdata_alloc(void)
55 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
58 memset(p, 0, sizeof(*p));
59 INIT_LIST_HEAD(&p->pages);
63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
65 void nfs_commit_free(struct nfs_commit_data *p)
67 mempool_free(p, nfs_commit_mempool);
69 EXPORT_SYMBOL_GPL(nfs_commit_free);
71 struct nfs_write_header *nfs_writehdr_alloc(void)
73 struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
76 struct nfs_pgio_header *hdr = &p->header;
78 memset(p, 0, sizeof(*p));
79 INIT_LIST_HEAD(&hdr->pages);
80 INIT_LIST_HEAD(&hdr->rpc_list);
81 spin_lock_init(&hdr->lock);
82 atomic_set(&hdr->refcnt, 0);
87 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
89 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
90 unsigned int pagecount)
92 struct nfs_write_data *data, *prealloc;
94 prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
95 if (prealloc->header == NULL)
98 data = kzalloc(sizeof(*data), GFP_KERNEL);
102 if (nfs_pgarray_set(&data->pages, pagecount)) {
104 atomic_inc(&hdr->refcnt);
106 if (data != prealloc)
114 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
116 struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
117 mempool_free(whdr, nfs_wdata_mempool);
119 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
121 void nfs_writedata_release(struct nfs_write_data *wdata)
123 struct nfs_pgio_header *hdr = wdata->header;
124 struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
126 put_nfs_open_context(wdata->args.context);
127 if (wdata->pages.pagevec != wdata->pages.page_array)
128 kfree(wdata->pages.pagevec);
129 if (wdata != &write_header->rpc_data)
132 wdata->header = NULL;
133 if (atomic_dec_and_test(&hdr->refcnt))
134 hdr->completion_ops->completion(hdr);
136 EXPORT_SYMBOL_GPL(nfs_writedata_release);
138 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
142 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
145 static struct nfs_page *
146 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
148 struct nfs_page *req = NULL;
150 if (PagePrivate(page))
151 req = (struct nfs_page *)page_private(page);
152 else if (unlikely(PageSwapCache(page))) {
153 struct nfs_page *freq, *t;
155 /* Linearly search the commit list for the correct req */
156 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
157 if (freq->wb_page == page) {
165 kref_get(&req->wb_kref);
170 static struct nfs_page *nfs_page_find_request(struct page *page)
172 struct inode *inode = page_file_mapping(page)->host;
173 struct nfs_page *req = NULL;
175 spin_lock(&inode->i_lock);
176 req = nfs_page_find_request_locked(NFS_I(inode), page);
177 spin_unlock(&inode->i_lock);
181 /* Adjust the file length if we're writing beyond the end */
182 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
184 struct inode *inode = page_file_mapping(page)->host;
188 spin_lock(&inode->i_lock);
189 i_size = i_size_read(inode);
190 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
191 if (i_size > 0 && page_file_index(page) < end_index)
193 end = page_file_offset(page) + ((loff_t)offset+count);
196 i_size_write(inode, end);
197 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
199 spin_unlock(&inode->i_lock);
202 /* A writeback failed: mark the page as bad, and invalidate the page cache */
203 static void nfs_set_pageerror(struct page *page)
206 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
209 /* We can set the PG_uptodate flag if we see that a write request
210 * covers the full page.
212 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
214 if (PageUptodate(page))
218 if (count != nfs_page_length(page))
220 SetPageUptodate(page);
223 static int wb_priority(struct writeback_control *wbc)
225 if (wbc->for_reclaim)
226 return FLUSH_HIGHPRI | FLUSH_STABLE;
227 if (wbc->for_kupdate || wbc->for_background)
228 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
229 return FLUSH_COND_STABLE;
233 * NFS congestion control
236 int nfs_congestion_kb;
238 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
239 #define NFS_CONGESTION_OFF_THRESH \
240 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
242 static void nfs_set_page_writeback(struct page *page)
244 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
245 int ret = test_set_page_writeback(page);
247 WARN_ON_ONCE(ret != 0);
249 if (atomic_long_inc_return(&nfss->writeback) >
250 NFS_CONGESTION_ON_THRESH) {
251 set_bdi_congested(&nfss->backing_dev_info,
256 static void nfs_end_page_writeback(struct page *page)
258 struct inode *inode = page_file_mapping(page)->host;
259 struct nfs_server *nfss = NFS_SERVER(inode);
261 end_page_writeback(page);
262 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
263 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
266 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
268 struct inode *inode = page_file_mapping(page)->host;
269 struct nfs_page *req;
272 spin_lock(&inode->i_lock);
274 req = nfs_page_find_request_locked(NFS_I(inode), page);
277 if (nfs_lock_request(req))
279 /* Note: If we hold the page lock, as is the case in nfs_writepage,
280 * then the call to nfs_lock_request() will always
281 * succeed provided that someone hasn't already marked the
282 * request as dirty (in which case we don't care).
284 spin_unlock(&inode->i_lock);
286 ret = nfs_wait_on_request(req);
289 nfs_release_request(req);
292 spin_lock(&inode->i_lock);
294 spin_unlock(&inode->i_lock);
299 * Find an associated nfs write request, and prepare to flush it out
300 * May return an error if the user signalled nfs_wait_on_request().
302 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
303 struct page *page, bool nonblock)
305 struct nfs_page *req;
308 req = nfs_find_and_lock_request(page, nonblock);
315 nfs_set_page_writeback(page);
316 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
319 if (!nfs_pageio_add_request(pgio, req)) {
320 nfs_redirty_request(req);
321 ret = pgio->pg_error;
327 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
329 struct inode *inode = page_file_mapping(page)->host;
332 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
333 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
335 nfs_pageio_cond_complete(pgio, page_file_index(page));
336 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
337 if (ret == -EAGAIN) {
338 redirty_page_for_writepage(wbc, page);
345 * Write an mmapped page to the server.
347 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
349 struct nfs_pageio_descriptor pgio;
352 NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio,
355 &nfs_async_write_completion_ops);
356 err = nfs_do_writepage(page, wbc, &pgio);
357 nfs_pageio_complete(&pgio);
360 if (pgio.pg_error < 0)
361 return pgio.pg_error;
365 int nfs_writepage(struct page *page, struct writeback_control *wbc)
369 ret = nfs_writepage_locked(page, wbc);
374 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
378 ret = nfs_do_writepage(page, wbc, data);
383 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
385 struct inode *inode = mapping->host;
386 unsigned long *bitlock = &NFS_I(inode)->flags;
387 struct nfs_pageio_descriptor pgio;
390 /* Stop dirtying of new pages while we sync */
391 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
392 nfs_wait_bit_killable, TASK_KILLABLE);
396 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
398 NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
399 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
400 nfs_pageio_complete(&pgio);
402 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
403 smp_mb__after_clear_bit();
404 wake_up_bit(bitlock, NFS_INO_FLUSHING);
417 * Insert a write request into an inode
419 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
421 struct nfs_inode *nfsi = NFS_I(inode);
423 /* Lock the request! */
424 nfs_lock_request(req);
426 spin_lock(&inode->i_lock);
427 if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
430 * Swap-space should not get truncated. Hence no need to plug the race
431 * with invalidate/truncate.
433 if (likely(!PageSwapCache(req->wb_page))) {
434 set_bit(PG_MAPPED, &req->wb_flags);
435 SetPagePrivate(req->wb_page);
436 set_page_private(req->wb_page, (unsigned long)req);
439 kref_get(&req->wb_kref);
440 spin_unlock(&inode->i_lock);
444 * Remove a write request from an inode
446 static void nfs_inode_remove_request(struct nfs_page *req)
448 struct inode *inode = req->wb_context->dentry->d_inode;
449 struct nfs_inode *nfsi = NFS_I(inode);
451 spin_lock(&inode->i_lock);
452 if (likely(!PageSwapCache(req->wb_page))) {
453 set_page_private(req->wb_page, 0);
454 ClearPagePrivate(req->wb_page);
455 clear_bit(PG_MAPPED, &req->wb_flags);
458 spin_unlock(&inode->i_lock);
459 nfs_release_request(req);
463 nfs_mark_request_dirty(struct nfs_page *req)
465 __set_page_dirty_nobuffers(req->wb_page);
468 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
470 * nfs_request_add_commit_list - add request to a commit list
471 * @req: pointer to a struct nfs_page
472 * @dst: commit list head
473 * @cinfo: holds list lock and accounting info
475 * This sets the PG_CLEAN bit, updates the cinfo count of
476 * number of outstanding requests requiring a commit as well as
479 * The caller must _not_ hold the cinfo->lock, but must be
480 * holding the nfs_page lock.
483 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
484 struct nfs_commit_info *cinfo)
486 set_bit(PG_CLEAN, &(req)->wb_flags);
487 spin_lock(cinfo->lock);
488 nfs_list_add_request(req, dst);
489 cinfo->mds->ncommit++;
490 spin_unlock(cinfo->lock);
492 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
493 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
495 __mark_inode_dirty(req->wb_context->dentry->d_inode,
499 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
502 * nfs_request_remove_commit_list - Remove request from a commit list
503 * @req: pointer to a nfs_page
504 * @cinfo: holds list lock and accounting info
506 * This clears the PG_CLEAN bit, and updates the cinfo's count of
507 * number of outstanding requests requiring a commit
508 * It does not update the MM page stats.
510 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
513 nfs_request_remove_commit_list(struct nfs_page *req,
514 struct nfs_commit_info *cinfo)
516 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
518 nfs_list_remove_request(req);
519 cinfo->mds->ncommit--;
521 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
523 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
526 cinfo->lock = &inode->i_lock;
527 cinfo->mds = &NFS_I(inode)->commit_info;
528 cinfo->ds = pnfs_get_ds_info(inode);
530 cinfo->completion_ops = &nfs_commit_completion_ops;
533 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
535 struct nfs_direct_req *dreq)
538 nfs_init_cinfo_from_dreq(cinfo, dreq);
540 nfs_init_cinfo_from_inode(cinfo, inode);
542 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
545 * Add a request to the inode's commit list.
548 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
549 struct nfs_commit_info *cinfo)
551 if (pnfs_mark_request_commit(req, lseg, cinfo))
553 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
557 nfs_clear_page_commit(struct page *page)
559 dec_zone_page_state(page, NR_UNSTABLE_NFS);
560 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
564 nfs_clear_request_commit(struct nfs_page *req)
566 if (test_bit(PG_CLEAN, &req->wb_flags)) {
567 struct inode *inode = req->wb_context->dentry->d_inode;
568 struct nfs_commit_info cinfo;
570 nfs_init_cinfo_from_inode(&cinfo, inode);
571 if (!pnfs_clear_request_commit(req, &cinfo)) {
572 spin_lock(cinfo.lock);
573 nfs_request_remove_commit_list(req, &cinfo);
574 spin_unlock(cinfo.lock);
576 nfs_clear_page_commit(req->wb_page);
581 int nfs_write_need_commit(struct nfs_write_data *data)
583 if (data->verf.committed == NFS_DATA_SYNC)
584 return data->header->lseg == NULL;
585 return data->verf.committed != NFS_FILE_SYNC;
589 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
594 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
596 struct nfs_direct_req *dreq)
601 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
602 struct nfs_commit_info *cinfo)
607 nfs_clear_request_commit(struct nfs_page *req)
612 int nfs_write_need_commit(struct nfs_write_data *data)
619 static void nfs_write_completion(struct nfs_pgio_header *hdr)
621 struct nfs_commit_info cinfo;
622 unsigned long bytes = 0;
624 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
626 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
627 while (!list_empty(&hdr->pages)) {
628 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
630 bytes += req->wb_bytes;
631 nfs_list_remove_request(req);
632 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
633 (hdr->good_bytes < bytes)) {
634 nfs_set_pageerror(req->wb_page);
635 nfs_context_set_write_error(req->wb_context, hdr->error);
638 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
639 nfs_mark_request_dirty(req);
642 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
643 memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
644 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
648 nfs_inode_remove_request(req);
650 nfs_unlock_request(req);
651 nfs_end_page_writeback(req->wb_page);
652 nfs_release_request(req);
658 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
660 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
662 return cinfo->mds->ncommit;
665 /* cinfo->lock held by caller */
667 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
668 struct nfs_commit_info *cinfo, int max)
670 struct nfs_page *req, *tmp;
673 list_for_each_entry_safe(req, tmp, src, wb_list) {
674 if (!nfs_lock_request(req))
676 kref_get(&req->wb_kref);
677 if (cond_resched_lock(cinfo->lock))
678 list_safe_reset_next(req, tmp, wb_list);
679 nfs_request_remove_commit_list(req, cinfo);
680 nfs_list_add_request(req, dst);
682 if ((ret == max) && !cinfo->dreq)
689 * nfs_scan_commit - Scan an inode for commit requests
690 * @inode: NFS inode to scan
691 * @dst: mds destination list
692 * @cinfo: mds and ds lists of reqs ready to commit
694 * Moves requests from the inode's 'commit' request list.
695 * The requests are *not* checked to ensure that they form a contiguous set.
698 nfs_scan_commit(struct inode *inode, struct list_head *dst,
699 struct nfs_commit_info *cinfo)
703 spin_lock(cinfo->lock);
704 if (cinfo->mds->ncommit > 0) {
705 const int max = INT_MAX;
707 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
709 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
711 spin_unlock(cinfo->lock);
716 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
721 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
722 struct nfs_commit_info *cinfo)
729 * Search for an existing write request, and attempt to update
730 * it to reflect a new dirty region on a given page.
732 * If the attempt fails, then the existing request is flushed out
735 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
740 struct nfs_page *req;
745 if (!PagePrivate(page))
748 end = offset + bytes;
749 spin_lock(&inode->i_lock);
752 req = nfs_page_find_request_locked(NFS_I(inode), page);
756 rqend = req->wb_offset + req->wb_bytes;
758 * Tell the caller to flush out the request if
759 * the offsets are non-contiguous.
760 * Note: nfs_flush_incompatible() will already
761 * have flushed out requests having wrong owners.
764 || end < req->wb_offset)
767 if (nfs_lock_request(req))
770 /* The request is locked, so wait and then retry */
771 spin_unlock(&inode->i_lock);
772 error = nfs_wait_on_request(req);
773 nfs_release_request(req);
776 spin_lock(&inode->i_lock);
779 /* Okay, the request matches. Update the region */
780 if (offset < req->wb_offset) {
781 req->wb_offset = offset;
782 req->wb_pgbase = offset;
785 req->wb_bytes = end - req->wb_offset;
787 req->wb_bytes = rqend - req->wb_offset;
789 spin_unlock(&inode->i_lock);
791 nfs_clear_request_commit(req);
794 spin_unlock(&inode->i_lock);
795 nfs_release_request(req);
796 error = nfs_wb_page(inode, page);
798 return ERR_PTR(error);
802 * Try to update an existing write request, or create one if there is none.
804 * Note: Should always be called with the Page Lock held to prevent races
805 * if we have to add a new request. Also assumes that the caller has
806 * already called nfs_flush_incompatible() if necessary.
808 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
809 struct page *page, unsigned int offset, unsigned int bytes)
811 struct inode *inode = page_file_mapping(page)->host;
812 struct nfs_page *req;
814 req = nfs_try_to_update_request(inode, page, offset, bytes);
817 req = nfs_create_request(ctx, inode, page, offset, bytes);
820 nfs_inode_add_request(inode, req);
825 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
826 unsigned int offset, unsigned int count)
828 struct nfs_page *req;
830 req = nfs_setup_write_request(ctx, page, offset, count);
833 /* Update file length */
834 nfs_grow_file(page, offset, count);
835 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
836 nfs_mark_request_dirty(req);
837 nfs_unlock_and_release_request(req);
841 int nfs_flush_incompatible(struct file *file, struct page *page)
843 struct nfs_open_context *ctx = nfs_file_open_context(file);
844 struct nfs_lock_context *l_ctx;
845 struct nfs_page *req;
846 int do_flush, status;
848 * Look for a request corresponding to this page. If there
849 * is one, and it belongs to another file, we flush it out
850 * before we try to copy anything into the page. Do this
851 * due to the lack of an ACCESS-type call in NFSv2.
852 * Also do the same if we find a request from an existing
856 req = nfs_page_find_request(page);
859 l_ctx = req->wb_lock_context;
860 do_flush = req->wb_page != page || req->wb_context != ctx;
862 do_flush |= l_ctx->lockowner.l_owner != current->files
863 || l_ctx->lockowner.l_pid != current->tgid;
865 nfs_release_request(req);
868 status = nfs_wb_page(page_file_mapping(page)->host, page);
869 } while (status == 0);
874 * If the page cache is marked as unsafe or invalid, then we can't rely on
875 * the PageUptodate() flag. In this case, we will need to turn off
876 * write optimisations that depend on the page contents being correct.
878 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
880 if (nfs_have_delegated_attributes(inode))
882 if (NFS_I(inode)->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
885 return PageUptodate(page) != 0;
889 * Update and possibly write a cached page of an NFS file.
891 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
892 * things with a page scheduled for an RPC call (e.g. invalidate it).
894 int nfs_updatepage(struct file *file, struct page *page,
895 unsigned int offset, unsigned int count)
897 struct nfs_open_context *ctx = nfs_file_open_context(file);
898 struct inode *inode = page_file_mapping(page)->host;
901 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
903 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
904 file->f_path.dentry->d_parent->d_name.name,
905 file->f_path.dentry->d_name.name, count,
906 (long long)(page_file_offset(page) + offset));
908 /* If we're not using byte range locks, and we know the page
909 * is up to date, it may be more efficient to extend the write
910 * to cover the entire page in order to avoid fragmentation
913 if (nfs_write_pageuptodate(page, inode) &&
914 inode->i_flock == NULL &&
915 !(file->f_flags & O_DSYNC)) {
916 count = max(count + offset, nfs_page_length(page));
920 status = nfs_writepage_setup(ctx, page, offset, count);
922 nfs_set_pageerror(page);
924 __set_page_dirty_nobuffers(page);
926 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
927 status, (long long)i_size_read(inode));
931 static int flush_task_priority(int how)
933 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
935 return RPC_PRIORITY_HIGH;
937 return RPC_PRIORITY_LOW;
939 return RPC_PRIORITY_NORMAL;
942 int nfs_initiate_write(struct rpc_clnt *clnt,
943 struct nfs_write_data *data,
944 const struct rpc_call_ops *call_ops,
947 struct inode *inode = data->header->inode;
948 int priority = flush_task_priority(how);
949 struct rpc_task *task;
950 struct rpc_message msg = {
951 .rpc_argp = &data->args,
952 .rpc_resp = &data->res,
953 .rpc_cred = data->header->cred,
955 struct rpc_task_setup task_setup_data = {
959 .callback_ops = call_ops,
960 .callback_data = data,
961 .workqueue = nfsiod_workqueue,
962 .flags = RPC_TASK_ASYNC | flags,
963 .priority = priority,
967 /* Set up the initial task struct. */
968 NFS_PROTO(inode)->write_setup(data, &msg);
970 dprintk("NFS: %5u initiated write call "
971 "(req %s/%lld, %u bytes @ offset %llu)\n",
974 (long long)NFS_FILEID(inode),
976 (unsigned long long)data->args.offset);
978 task = rpc_run_task(&task_setup_data);
983 if (how & FLUSH_SYNC) {
984 ret = rpc_wait_for_completion_task(task);
986 ret = task->tk_status;
992 EXPORT_SYMBOL_GPL(nfs_initiate_write);
995 * Set up the argument/result storage required for the RPC call.
997 static void nfs_write_rpcsetup(struct nfs_write_data *data,
998 unsigned int count, unsigned int offset,
999 int how, struct nfs_commit_info *cinfo)
1001 struct nfs_page *req = data->header->req;
1003 /* Set up the RPC argument and reply structs
1004 * NB: take care not to mess about with data->commit et al. */
1006 data->args.fh = NFS_FH(data->header->inode);
1007 data->args.offset = req_offset(req) + offset;
1008 /* pnfs_set_layoutcommit needs this */
1009 data->mds_offset = data->args.offset;
1010 data->args.pgbase = req->wb_pgbase + offset;
1011 data->args.pages = data->pages.pagevec;
1012 data->args.count = count;
1013 data->args.context = get_nfs_open_context(req->wb_context);
1014 data->args.lock_context = req->wb_lock_context;
1015 data->args.stable = NFS_UNSTABLE;
1016 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1019 case FLUSH_COND_STABLE:
1020 if (nfs_reqs_to_commit(cinfo))
1023 data->args.stable = NFS_FILE_SYNC;
1026 data->res.fattr = &data->fattr;
1027 data->res.count = count;
1028 data->res.verf = &data->verf;
1029 nfs_fattr_init(&data->fattr);
1032 static int nfs_do_write(struct nfs_write_data *data,
1033 const struct rpc_call_ops *call_ops,
1036 struct inode *inode = data->header->inode;
1038 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1041 static int nfs_do_multiple_writes(struct list_head *head,
1042 const struct rpc_call_ops *call_ops,
1045 struct nfs_write_data *data;
1048 while (!list_empty(head)) {
1051 data = list_first_entry(head, struct nfs_write_data, list);
1052 list_del_init(&data->list);
1054 ret2 = nfs_do_write(data, call_ops, how);
1061 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1062 * call this on each, which will prepare them to be retried on next
1063 * writeback using standard nfs.
1065 static void nfs_redirty_request(struct nfs_page *req)
1067 nfs_mark_request_dirty(req);
1068 nfs_unlock_request(req);
1069 nfs_end_page_writeback(req->wb_page);
1070 nfs_release_request(req);
1073 static void nfs_async_write_error(struct list_head *head)
1075 struct nfs_page *req;
1077 while (!list_empty(head)) {
1078 req = nfs_list_entry(head->next);
1079 nfs_list_remove_request(req);
1080 nfs_redirty_request(req);
1084 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1085 .error_cleanup = nfs_async_write_error,
1086 .completion = nfs_write_completion,
1089 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1090 struct nfs_pgio_header *hdr)
1092 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1093 while (!list_empty(&hdr->rpc_list)) {
1094 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1095 struct nfs_write_data, list);
1096 list_del(&data->list);
1097 nfs_writedata_release(data);
1099 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1103 * Generate multiple small requests to write out a single
1104 * contiguous dirty area on one page.
1106 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1107 struct nfs_pgio_header *hdr)
1109 struct nfs_page *req = hdr->req;
1110 struct page *page = req->wb_page;
1111 struct nfs_write_data *data;
1112 size_t wsize = desc->pg_bsize, nbytes;
1113 unsigned int offset;
1115 struct nfs_commit_info cinfo;
1117 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1119 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1120 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1121 desc->pg_count > wsize))
1122 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1126 nbytes = desc->pg_count;
1128 size_t len = min(nbytes, wsize);
1130 data = nfs_writedata_alloc(hdr, 1);
1132 nfs_flush_error(desc, hdr);
1135 data->pages.pagevec[0] = page;
1136 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1137 list_add(&data->list, &hdr->rpc_list);
1141 } while (nbytes != 0);
1142 nfs_list_remove_request(req);
1143 nfs_list_add_request(req, &hdr->pages);
1144 desc->pg_rpc_callops = &nfs_write_common_ops;
1149 * Create an RPC task for the given write request and kick it.
1150 * The page must have been locked by the caller.
1152 * It may happen that the page we're passed is not marked dirty.
1153 * This is the case if nfs_updatepage detects a conflicting request
1154 * that has been written but not committed.
1156 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1157 struct nfs_pgio_header *hdr)
1159 struct nfs_page *req;
1160 struct page **pages;
1161 struct nfs_write_data *data;
1162 struct list_head *head = &desc->pg_list;
1163 struct nfs_commit_info cinfo;
1165 data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1168 nfs_flush_error(desc, hdr);
1172 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1173 pages = data->pages.pagevec;
1174 while (!list_empty(head)) {
1175 req = nfs_list_entry(head->next);
1176 nfs_list_remove_request(req);
1177 nfs_list_add_request(req, &hdr->pages);
1178 *pages++ = req->wb_page;
1181 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1182 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1183 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1185 /* Set up the argument struct */
1186 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1187 list_add(&data->list, &hdr->rpc_list);
1188 desc->pg_rpc_callops = &nfs_write_common_ops;
1192 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1193 struct nfs_pgio_header *hdr)
1195 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1196 return nfs_flush_multi(desc, hdr);
1197 return nfs_flush_one(desc, hdr);
1199 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1201 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1203 struct nfs_write_header *whdr;
1204 struct nfs_pgio_header *hdr;
1207 whdr = nfs_writehdr_alloc();
1209 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1212 hdr = &whdr->header;
1213 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1214 atomic_inc(&hdr->refcnt);
1215 ret = nfs_generic_flush(desc, hdr);
1217 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1218 desc->pg_rpc_callops,
1220 if (atomic_dec_and_test(&hdr->refcnt))
1221 hdr->completion_ops->completion(hdr);
1225 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1226 .pg_test = nfs_generic_pg_test,
1227 .pg_doio = nfs_generic_pg_writepages,
1230 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1231 struct inode *inode, int ioflags,
1232 const struct nfs_pgio_completion_ops *compl_ops)
1234 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1235 NFS_SERVER(inode)->wsize, ioflags);
1237 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1239 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1241 pgio->pg_ops = &nfs_pageio_write_ops;
1242 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1244 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1247 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1249 struct nfs_write_data *data = calldata;
1250 NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1253 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1255 struct nfs_commit_data *data = calldata;
1257 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1261 * Handle a write reply that flushes a whole page.
1263 * FIXME: There is an inherent race with invalidate_inode_pages and
1264 * writebacks since the page->count is kept > 1 for as long
1265 * as the page has a write request pending.
1267 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1269 struct nfs_write_data *data = calldata;
1271 nfs_writeback_done(task, data);
1274 static void nfs_writeback_release_common(void *calldata)
1276 struct nfs_write_data *data = calldata;
1277 struct nfs_pgio_header *hdr = data->header;
1278 int status = data->task.tk_status;
1280 if ((status >= 0) && nfs_write_need_commit(data)) {
1281 spin_lock(&hdr->lock);
1282 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1284 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1285 memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1286 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1287 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1288 spin_unlock(&hdr->lock);
1290 nfs_writedata_release(data);
1293 static const struct rpc_call_ops nfs_write_common_ops = {
1294 .rpc_call_prepare = nfs_write_prepare,
1295 .rpc_call_done = nfs_writeback_done_common,
1296 .rpc_release = nfs_writeback_release_common,
1301 * This function is called when the WRITE call is complete.
1303 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1305 struct nfs_writeargs *argp = &data->args;
1306 struct nfs_writeres *resp = &data->res;
1307 struct inode *inode = data->header->inode;
1310 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1311 task->tk_pid, task->tk_status);
1314 * ->write_done will attempt to use post-op attributes to detect
1315 * conflicting writes by other clients. A strict interpretation
1316 * of close-to-open would allow us to continue caching even if
1317 * another writer had changed the file, but some applications
1318 * depend on tighter cache coherency when writing.
1320 status = NFS_PROTO(inode)->write_done(task, data);
1323 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1325 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1326 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1327 /* We tried a write call, but the server did not
1328 * commit data to stable storage even though we
1330 * Note: There is a known bug in Tru64 < 5.0 in which
1331 * the server reports NFS_DATA_SYNC, but performs
1332 * NFS_FILE_SYNC. We therefore implement this checking
1333 * as a dprintk() in order to avoid filling syslog.
1335 static unsigned long complain;
1337 /* Note this will print the MDS for a DS write */
1338 if (time_before(complain, jiffies)) {
1339 dprintk("NFS: faulty NFS server %s:"
1340 " (committed = %d) != (stable = %d)\n",
1341 NFS_SERVER(inode)->nfs_client->cl_hostname,
1342 resp->verf->committed, argp->stable);
1343 complain = jiffies + 300 * HZ;
1347 if (task->tk_status < 0)
1348 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1349 else if (resp->count < argp->count) {
1350 static unsigned long complain;
1352 /* This a short write! */
1353 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1355 /* Has the server at least made some progress? */
1356 if (resp->count == 0) {
1357 if (time_before(complain, jiffies)) {
1359 "NFS: Server wrote zero bytes, expected %u.\n",
1361 complain = jiffies + 300 * HZ;
1363 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1364 task->tk_status = -EIO;
1367 /* Was this an NFSv2 write or an NFSv3 stable write? */
1368 if (resp->verf->committed != NFS_UNSTABLE) {
1369 /* Resend from where the server left off */
1370 data->mds_offset += resp->count;
1371 argp->offset += resp->count;
1372 argp->pgbase += resp->count;
1373 argp->count -= resp->count;
1375 /* Resend as a stable write in order to avoid
1376 * headaches in the case of a server crash.
1378 argp->stable = NFS_FILE_SYNC;
1380 rpc_restart_call_prepare(task);
1385 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1386 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1390 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1394 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1396 nfs_wait_bit_killable,
1398 return (ret < 0) ? ret : 1;
1401 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1403 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1404 smp_mb__after_clear_bit();
1405 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1408 void nfs_commitdata_release(struct nfs_commit_data *data)
1410 put_nfs_open_context(data->context);
1411 nfs_commit_free(data);
1413 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1415 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1416 const struct rpc_call_ops *call_ops,
1419 struct rpc_task *task;
1420 int priority = flush_task_priority(how);
1421 struct rpc_message msg = {
1422 .rpc_argp = &data->args,
1423 .rpc_resp = &data->res,
1424 .rpc_cred = data->cred,
1426 struct rpc_task_setup task_setup_data = {
1427 .task = &data->task,
1429 .rpc_message = &msg,
1430 .callback_ops = call_ops,
1431 .callback_data = data,
1432 .workqueue = nfsiod_workqueue,
1433 .flags = RPC_TASK_ASYNC | flags,
1434 .priority = priority,
1436 /* Set up the initial task struct. */
1437 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1439 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1441 task = rpc_run_task(&task_setup_data);
1443 return PTR_ERR(task);
1444 if (how & FLUSH_SYNC)
1445 rpc_wait_for_completion_task(task);
1449 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1452 * Set up the argument/result storage required for the RPC call.
1454 void nfs_init_commit(struct nfs_commit_data *data,
1455 struct list_head *head,
1456 struct pnfs_layout_segment *lseg,
1457 struct nfs_commit_info *cinfo)
1459 struct nfs_page *first = nfs_list_entry(head->next);
1460 struct inode *inode = first->wb_context->dentry->d_inode;
1462 /* Set up the RPC argument and reply structs
1463 * NB: take care not to mess about with data->commit et al. */
1465 list_splice_init(head, &data->pages);
1467 data->inode = inode;
1468 data->cred = first->wb_context->cred;
1469 data->lseg = lseg; /* reference transferred */
1470 data->mds_ops = &nfs_commit_ops;
1471 data->completion_ops = cinfo->completion_ops;
1472 data->dreq = cinfo->dreq;
1474 data->args.fh = NFS_FH(data->inode);
1475 /* Note: we always request a commit of the entire inode */
1476 data->args.offset = 0;
1477 data->args.count = 0;
1478 data->context = get_nfs_open_context(first->wb_context);
1479 data->res.fattr = &data->fattr;
1480 data->res.verf = &data->verf;
1481 nfs_fattr_init(&data->fattr);
1483 EXPORT_SYMBOL_GPL(nfs_init_commit);
1485 void nfs_retry_commit(struct list_head *page_list,
1486 struct pnfs_layout_segment *lseg,
1487 struct nfs_commit_info *cinfo)
1489 struct nfs_page *req;
1491 while (!list_empty(page_list)) {
1492 req = nfs_list_entry(page_list->next);
1493 nfs_list_remove_request(req);
1494 nfs_mark_request_commit(req, lseg, cinfo);
1496 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1497 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1500 nfs_unlock_and_release_request(req);
1503 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1506 * Commit dirty pages
1509 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1510 struct nfs_commit_info *cinfo)
1512 struct nfs_commit_data *data;
1514 data = nfs_commitdata_alloc();
1519 /* Set up the argument struct */
1520 nfs_init_commit(data, head, NULL, cinfo);
1521 atomic_inc(&cinfo->mds->rpcs_out);
1522 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1525 nfs_retry_commit(head, NULL, cinfo);
1526 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1531 * COMMIT call returned
1533 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1535 struct nfs_commit_data *data = calldata;
1537 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1538 task->tk_pid, task->tk_status);
1540 /* Call the NFS version-specific code */
1541 NFS_PROTO(data->inode)->commit_done(task, data);
1544 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1546 struct nfs_page *req;
1547 int status = data->task.tk_status;
1548 struct nfs_commit_info cinfo;
1550 while (!list_empty(&data->pages)) {
1551 req = nfs_list_entry(data->pages.next);
1552 nfs_list_remove_request(req);
1553 nfs_clear_page_commit(req->wb_page);
1555 dprintk("NFS: commit (%s/%lld %d@%lld)",
1556 req->wb_context->dentry->d_sb->s_id,
1557 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1559 (long long)req_offset(req));
1561 nfs_context_set_write_error(req->wb_context, status);
1562 nfs_inode_remove_request(req);
1563 dprintk(", error = %d\n", status);
1567 /* Okay, COMMIT succeeded, apparently. Check the verifier
1568 * returned by the server against all stored verfs. */
1569 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1570 /* We have a match */
1571 nfs_inode_remove_request(req);
1575 /* We have a mismatch. Write the page again */
1576 dprintk(" mismatch\n");
1577 nfs_mark_request_dirty(req);
1578 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1580 nfs_unlock_and_release_request(req);
1582 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1583 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1584 nfs_commit_clear_lock(NFS_I(data->inode));
1587 static void nfs_commit_release(void *calldata)
1589 struct nfs_commit_data *data = calldata;
1591 data->completion_ops->completion(data);
1592 nfs_commitdata_release(calldata);
1595 static const struct rpc_call_ops nfs_commit_ops = {
1596 .rpc_call_prepare = nfs_commit_prepare,
1597 .rpc_call_done = nfs_commit_done,
1598 .rpc_release = nfs_commit_release,
1601 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1602 .completion = nfs_commit_release_pages,
1603 .error_cleanup = nfs_commit_clear_lock,
1606 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1607 int how, struct nfs_commit_info *cinfo)
1611 status = pnfs_commit_list(inode, head, how, cinfo);
1612 if (status == PNFS_NOT_ATTEMPTED)
1613 status = nfs_commit_list(inode, head, how, cinfo);
1617 int nfs_commit_inode(struct inode *inode, int how)
1620 struct nfs_commit_info cinfo;
1621 int may_wait = how & FLUSH_SYNC;
1624 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1626 goto out_mark_dirty;
1627 nfs_init_cinfo_from_inode(&cinfo, inode);
1628 res = nfs_scan_commit(inode, &head, &cinfo);
1632 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1636 goto out_mark_dirty;
1637 error = wait_on_bit(&NFS_I(inode)->flags,
1639 nfs_wait_bit_killable,
1644 nfs_commit_clear_lock(NFS_I(inode));
1646 /* Note: If we exit without ensuring that the commit is complete,
1647 * we must mark the inode as dirty. Otherwise, future calls to
1648 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1649 * that the data is on the disk.
1652 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1656 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1658 struct nfs_inode *nfsi = NFS_I(inode);
1659 int flags = FLUSH_SYNC;
1662 /* no commits means nothing needs to be done */
1663 if (!nfsi->commit_info.ncommit)
1666 if (wbc->sync_mode == WB_SYNC_NONE) {
1667 /* Don't commit yet if this is a non-blocking flush and there
1668 * are a lot of outstanding writes for this mapping.
1670 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1671 goto out_mark_dirty;
1673 /* don't wait for the COMMIT response */
1677 ret = nfs_commit_inode(inode, flags);
1679 if (wbc->sync_mode == WB_SYNC_NONE) {
1680 if (ret < wbc->nr_to_write)
1681 wbc->nr_to_write -= ret;
1683 wbc->nr_to_write = 0;
1688 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1692 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1698 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1700 return nfs_commit_unstable_pages(inode, wbc);
1702 EXPORT_SYMBOL_GPL(nfs_write_inode);
1705 * flush the inode to disk.
1707 int nfs_wb_all(struct inode *inode)
1709 struct writeback_control wbc = {
1710 .sync_mode = WB_SYNC_ALL,
1711 .nr_to_write = LONG_MAX,
1713 .range_end = LLONG_MAX,
1716 return sync_inode(inode, &wbc);
1718 EXPORT_SYMBOL_GPL(nfs_wb_all);
1720 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1722 struct nfs_page *req;
1726 wait_on_page_writeback(page);
1727 req = nfs_page_find_request(page);
1730 if (nfs_lock_request(req)) {
1731 nfs_clear_request_commit(req);
1732 nfs_inode_remove_request(req);
1734 * In case nfs_inode_remove_request has marked the
1735 * page as being dirty
1737 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1738 nfs_unlock_and_release_request(req);
1741 ret = nfs_wait_on_request(req);
1742 nfs_release_request(req);
1750 * Write back all requests on one page - we do this before reading it.
1752 int nfs_wb_page(struct inode *inode, struct page *page)
1754 loff_t range_start = page_file_offset(page);
1755 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1756 struct writeback_control wbc = {
1757 .sync_mode = WB_SYNC_ALL,
1759 .range_start = range_start,
1760 .range_end = range_end,
1765 wait_on_page_writeback(page);
1766 if (clear_page_dirty_for_io(page)) {
1767 ret = nfs_writepage_locked(page, &wbc);
1772 if (!PagePrivate(page))
1774 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1783 #ifdef CONFIG_MIGRATION
1784 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1785 struct page *page, enum migrate_mode mode)
1788 * If PagePrivate is set, then the page is currently associated with
1789 * an in-progress read or write request. Don't try to migrate it.
1791 * FIXME: we could do this in principle, but we'll need a way to ensure
1792 * that we can safely release the inode reference while holding
1795 if (PagePrivate(page))
1798 nfs_fscache_release_page(page, GFP_KERNEL);
1800 return migrate_page(mapping, newpage, page, mode);
1804 int __init nfs_init_writepagecache(void)
1806 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1807 sizeof(struct nfs_write_header),
1808 0, SLAB_HWCACHE_ALIGN,
1810 if (nfs_wdata_cachep == NULL)
1813 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1815 if (nfs_wdata_mempool == NULL)
1816 goto out_destroy_write_cache;
1818 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1819 sizeof(struct nfs_commit_data),
1820 0, SLAB_HWCACHE_ALIGN,
1822 if (nfs_cdata_cachep == NULL)
1823 goto out_destroy_write_mempool;
1825 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1827 if (nfs_commit_mempool == NULL)
1828 goto out_destroy_commit_cache;
1831 * NFS congestion size, scale with available memory.
1843 * This allows larger machines to have larger/more transfers.
1844 * Limit the default to 256M
1846 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1847 if (nfs_congestion_kb > 256*1024)
1848 nfs_congestion_kb = 256*1024;
1852 out_destroy_commit_cache:
1853 kmem_cache_destroy(nfs_cdata_cachep);
1854 out_destroy_write_mempool:
1855 mempool_destroy(nfs_wdata_mempool);
1856 out_destroy_write_cache:
1857 kmem_cache_destroy(nfs_wdata_cachep);
1861 void nfs_destroy_writepagecache(void)
1863 mempool_destroy(nfs_commit_mempool);
1864 kmem_cache_destroy(nfs_cdata_cachep);
1865 mempool_destroy(nfs_wdata_mempool);
1866 kmem_cache_destroy(nfs_wdata_cachep);