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)
205 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
208 /* We can set the PG_uptodate flag if we see that a write request
209 * covers the full page.
211 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
213 if (PageUptodate(page))
217 if (count != nfs_page_length(page))
219 SetPageUptodate(page);
222 static int wb_priority(struct writeback_control *wbc)
224 if (wbc->for_reclaim)
225 return FLUSH_HIGHPRI | FLUSH_STABLE;
226 if (wbc->for_kupdate || wbc->for_background)
227 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
228 return FLUSH_COND_STABLE;
232 * NFS congestion control
235 int nfs_congestion_kb;
237 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
238 #define NFS_CONGESTION_OFF_THRESH \
239 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
241 static void nfs_set_page_writeback(struct page *page)
243 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
244 int ret = test_set_page_writeback(page);
246 WARN_ON_ONCE(ret != 0);
248 if (atomic_long_inc_return(&nfss->writeback) >
249 NFS_CONGESTION_ON_THRESH) {
250 set_bdi_congested(&nfss->backing_dev_info,
255 static void nfs_end_page_writeback(struct page *page)
257 struct inode *inode = page_file_mapping(page)->host;
258 struct nfs_server *nfss = NFS_SERVER(inode);
260 end_page_writeback(page);
261 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
262 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
265 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
267 struct inode *inode = page_file_mapping(page)->host;
268 struct nfs_page *req;
271 spin_lock(&inode->i_lock);
273 req = nfs_page_find_request_locked(NFS_I(inode), page);
276 if (nfs_lock_request(req))
278 /* Note: If we hold the page lock, as is the case in nfs_writepage,
279 * then the call to nfs_lock_request() will always
280 * succeed provided that someone hasn't already marked the
281 * request as dirty (in which case we don't care).
283 spin_unlock(&inode->i_lock);
285 ret = nfs_wait_on_request(req);
288 nfs_release_request(req);
291 spin_lock(&inode->i_lock);
293 spin_unlock(&inode->i_lock);
298 * Find an associated nfs write request, and prepare to flush it out
299 * May return an error if the user signalled nfs_wait_on_request().
301 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
302 struct page *page, bool nonblock)
304 struct nfs_page *req;
307 req = nfs_find_and_lock_request(page, nonblock);
314 nfs_set_page_writeback(page);
315 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
318 if (!nfs_pageio_add_request(pgio, req)) {
319 nfs_redirty_request(req);
320 ret = pgio->pg_error;
326 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
328 struct inode *inode = page_file_mapping(page)->host;
331 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
332 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
334 nfs_pageio_cond_complete(pgio, page_file_index(page));
335 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
336 if (ret == -EAGAIN) {
337 redirty_page_for_writepage(wbc, page);
344 * Write an mmapped page to the server.
346 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
348 struct nfs_pageio_descriptor pgio;
351 NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio,
354 &nfs_async_write_completion_ops);
355 err = nfs_do_writepage(page, wbc, &pgio);
356 nfs_pageio_complete(&pgio);
359 if (pgio.pg_error < 0)
360 return pgio.pg_error;
364 int nfs_writepage(struct page *page, struct writeback_control *wbc)
368 ret = nfs_writepage_locked(page, wbc);
373 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
377 ret = nfs_do_writepage(page, wbc, data);
382 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
384 struct inode *inode = mapping->host;
385 unsigned long *bitlock = &NFS_I(inode)->flags;
386 struct nfs_pageio_descriptor pgio;
389 /* Stop dirtying of new pages while we sync */
390 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
391 nfs_wait_bit_killable, TASK_KILLABLE);
395 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
397 NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
398 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
399 nfs_pageio_complete(&pgio);
401 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
402 smp_mb__after_clear_bit();
403 wake_up_bit(bitlock, NFS_INO_FLUSHING);
416 * Insert a write request into an inode
418 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
420 struct nfs_inode *nfsi = NFS_I(inode);
422 /* Lock the request! */
423 nfs_lock_request(req);
425 spin_lock(&inode->i_lock);
426 if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
429 * Swap-space should not get truncated. Hence no need to plug the race
430 * with invalidate/truncate.
432 if (likely(!PageSwapCache(req->wb_page))) {
433 set_bit(PG_MAPPED, &req->wb_flags);
434 SetPagePrivate(req->wb_page);
435 set_page_private(req->wb_page, (unsigned long)req);
438 kref_get(&req->wb_kref);
439 spin_unlock(&inode->i_lock);
443 * Remove a write request from an inode
445 static void nfs_inode_remove_request(struct nfs_page *req)
447 struct inode *inode = req->wb_context->dentry->d_inode;
448 struct nfs_inode *nfsi = NFS_I(inode);
450 spin_lock(&inode->i_lock);
451 if (likely(!PageSwapCache(req->wb_page))) {
452 set_page_private(req->wb_page, 0);
453 ClearPagePrivate(req->wb_page);
454 clear_bit(PG_MAPPED, &req->wb_flags);
457 spin_unlock(&inode->i_lock);
458 nfs_release_request(req);
462 nfs_mark_request_dirty(struct nfs_page *req)
464 __set_page_dirty_nobuffers(req->wb_page);
467 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
469 * nfs_request_add_commit_list - add request to a commit list
470 * @req: pointer to a struct nfs_page
471 * @dst: commit list head
472 * @cinfo: holds list lock and accounting info
474 * This sets the PG_CLEAN bit, updates the cinfo count of
475 * number of outstanding requests requiring a commit as well as
478 * The caller must _not_ hold the cinfo->lock, but must be
479 * holding the nfs_page lock.
482 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
483 struct nfs_commit_info *cinfo)
485 set_bit(PG_CLEAN, &(req)->wb_flags);
486 spin_lock(cinfo->lock);
487 nfs_list_add_request(req, dst);
488 cinfo->mds->ncommit++;
489 spin_unlock(cinfo->lock);
491 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
492 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
494 __mark_inode_dirty(req->wb_context->dentry->d_inode,
498 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
501 * nfs_request_remove_commit_list - Remove request from a commit list
502 * @req: pointer to a nfs_page
503 * @cinfo: holds list lock and accounting info
505 * This clears the PG_CLEAN bit, and updates the cinfo's count of
506 * number of outstanding requests requiring a commit
507 * It does not update the MM page stats.
509 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
512 nfs_request_remove_commit_list(struct nfs_page *req,
513 struct nfs_commit_info *cinfo)
515 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
517 nfs_list_remove_request(req);
518 cinfo->mds->ncommit--;
520 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
522 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
525 cinfo->lock = &inode->i_lock;
526 cinfo->mds = &NFS_I(inode)->commit_info;
527 cinfo->ds = pnfs_get_ds_info(inode);
529 cinfo->completion_ops = &nfs_commit_completion_ops;
532 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
534 struct nfs_direct_req *dreq)
537 nfs_init_cinfo_from_dreq(cinfo, dreq);
539 nfs_init_cinfo_from_inode(cinfo, inode);
541 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
544 * Add a request to the inode's commit list.
547 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
548 struct nfs_commit_info *cinfo)
550 if (pnfs_mark_request_commit(req, lseg, cinfo))
552 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
556 nfs_clear_page_commit(struct page *page)
558 dec_zone_page_state(page, NR_UNSTABLE_NFS);
559 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
563 nfs_clear_request_commit(struct nfs_page *req)
565 if (test_bit(PG_CLEAN, &req->wb_flags)) {
566 struct inode *inode = req->wb_context->dentry->d_inode;
567 struct nfs_commit_info cinfo;
569 nfs_init_cinfo_from_inode(&cinfo, inode);
570 if (!pnfs_clear_request_commit(req, &cinfo)) {
571 spin_lock(cinfo.lock);
572 nfs_request_remove_commit_list(req, &cinfo);
573 spin_unlock(cinfo.lock);
575 nfs_clear_page_commit(req->wb_page);
580 int nfs_write_need_commit(struct nfs_write_data *data)
582 if (data->verf.committed == NFS_DATA_SYNC)
583 return data->header->lseg == NULL;
584 return data->verf.committed != NFS_FILE_SYNC;
588 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
593 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
595 struct nfs_direct_req *dreq)
600 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
601 struct nfs_commit_info *cinfo)
606 nfs_clear_request_commit(struct nfs_page *req)
611 int nfs_write_need_commit(struct nfs_write_data *data)
618 static void nfs_write_completion(struct nfs_pgio_header *hdr)
620 struct nfs_commit_info cinfo;
621 unsigned long bytes = 0;
623 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
625 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
626 while (!list_empty(&hdr->pages)) {
627 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
629 bytes += req->wb_bytes;
630 nfs_list_remove_request(req);
631 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
632 (hdr->good_bytes < bytes)) {
633 nfs_set_pageerror(req->wb_page);
634 nfs_context_set_write_error(req->wb_context, hdr->error);
637 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
638 nfs_mark_request_dirty(req);
641 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
642 memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
643 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
647 nfs_inode_remove_request(req);
649 nfs_unlock_request(req);
650 nfs_end_page_writeback(req->wb_page);
651 nfs_release_request(req);
657 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
659 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
661 return cinfo->mds->ncommit;
664 /* cinfo->lock held by caller */
666 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
667 struct nfs_commit_info *cinfo, int max)
669 struct nfs_page *req, *tmp;
672 list_for_each_entry_safe(req, tmp, src, wb_list) {
673 if (!nfs_lock_request(req))
675 kref_get(&req->wb_kref);
676 if (cond_resched_lock(cinfo->lock))
677 list_safe_reset_next(req, tmp, wb_list);
678 nfs_request_remove_commit_list(req, cinfo);
679 nfs_list_add_request(req, dst);
681 if ((ret == max) && !cinfo->dreq)
688 * nfs_scan_commit - Scan an inode for commit requests
689 * @inode: NFS inode to scan
690 * @dst: mds destination list
691 * @cinfo: mds and ds lists of reqs ready to commit
693 * Moves requests from the inode's 'commit' request list.
694 * The requests are *not* checked to ensure that they form a contiguous set.
697 nfs_scan_commit(struct inode *inode, struct list_head *dst,
698 struct nfs_commit_info *cinfo)
702 spin_lock(cinfo->lock);
703 if (cinfo->mds->ncommit > 0) {
704 const int max = INT_MAX;
706 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
708 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
710 spin_unlock(cinfo->lock);
715 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
720 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
721 struct nfs_commit_info *cinfo)
728 * Search for an existing write request, and attempt to update
729 * it to reflect a new dirty region on a given page.
731 * If the attempt fails, then the existing request is flushed out
734 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
739 struct nfs_page *req;
744 if (!PagePrivate(page))
747 end = offset + bytes;
748 spin_lock(&inode->i_lock);
751 req = nfs_page_find_request_locked(NFS_I(inode), page);
755 rqend = req->wb_offset + req->wb_bytes;
757 * Tell the caller to flush out the request if
758 * the offsets are non-contiguous.
759 * Note: nfs_flush_incompatible() will already
760 * have flushed out requests having wrong owners.
763 || end < req->wb_offset)
766 if (nfs_lock_request(req))
769 /* The request is locked, so wait and then retry */
770 spin_unlock(&inode->i_lock);
771 error = nfs_wait_on_request(req);
772 nfs_release_request(req);
775 spin_lock(&inode->i_lock);
778 /* Okay, the request matches. Update the region */
779 if (offset < req->wb_offset) {
780 req->wb_offset = offset;
781 req->wb_pgbase = offset;
784 req->wb_bytes = end - req->wb_offset;
786 req->wb_bytes = rqend - req->wb_offset;
788 spin_unlock(&inode->i_lock);
790 nfs_clear_request_commit(req);
793 spin_unlock(&inode->i_lock);
794 nfs_release_request(req);
795 error = nfs_wb_page(inode, page);
797 return ERR_PTR(error);
801 * Try to update an existing write request, or create one if there is none.
803 * Note: Should always be called with the Page Lock held to prevent races
804 * if we have to add a new request. Also assumes that the caller has
805 * already called nfs_flush_incompatible() if necessary.
807 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
808 struct page *page, unsigned int offset, unsigned int bytes)
810 struct inode *inode = page_file_mapping(page)->host;
811 struct nfs_page *req;
813 req = nfs_try_to_update_request(inode, page, offset, bytes);
816 req = nfs_create_request(ctx, inode, page, offset, bytes);
819 nfs_inode_add_request(inode, req);
824 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
825 unsigned int offset, unsigned int count)
827 struct nfs_page *req;
829 req = nfs_setup_write_request(ctx, page, offset, count);
832 /* Update file length */
833 nfs_grow_file(page, offset, count);
834 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
835 nfs_mark_request_dirty(req);
836 nfs_unlock_and_release_request(req);
840 int nfs_flush_incompatible(struct file *file, struct page *page)
842 struct nfs_open_context *ctx = nfs_file_open_context(file);
843 struct nfs_lock_context *l_ctx;
844 struct nfs_page *req;
845 int do_flush, status;
847 * Look for a request corresponding to this page. If there
848 * is one, and it belongs to another file, we flush it out
849 * before we try to copy anything into the page. Do this
850 * due to the lack of an ACCESS-type call in NFSv2.
851 * Also do the same if we find a request from an existing
855 req = nfs_page_find_request(page);
858 l_ctx = req->wb_lock_context;
859 do_flush = req->wb_page != page || req->wb_context != ctx;
861 do_flush |= l_ctx->lockowner.l_owner != current->files
862 || l_ctx->lockowner.l_pid != current->tgid;
864 nfs_release_request(req);
867 status = nfs_wb_page(page_file_mapping(page)->host, page);
868 } while (status == 0);
873 * If the page cache is marked as unsafe or invalid, then we can't rely on
874 * the PageUptodate() flag. In this case, we will need to turn off
875 * write optimisations that depend on the page contents being correct.
877 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
879 if (nfs_have_delegated_attributes(inode))
881 if (NFS_I(inode)->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
884 return PageUptodate(page) != 0;
888 * Update and possibly write a cached page of an NFS file.
890 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
891 * things with a page scheduled for an RPC call (e.g. invalidate it).
893 int nfs_updatepage(struct file *file, struct page *page,
894 unsigned int offset, unsigned int count)
896 struct nfs_open_context *ctx = nfs_file_open_context(file);
897 struct inode *inode = page_file_mapping(page)->host;
900 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
902 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
903 file->f_path.dentry->d_parent->d_name.name,
904 file->f_path.dentry->d_name.name, count,
905 (long long)(page_file_offset(page) + offset));
907 /* If we're not using byte range locks, and we know the page
908 * is up to date, it may be more efficient to extend the write
909 * to cover the entire page in order to avoid fragmentation
912 if (nfs_write_pageuptodate(page, inode) &&
913 inode->i_flock == NULL &&
914 !(file->f_flags & O_DSYNC)) {
915 count = max(count + offset, nfs_page_length(page));
919 status = nfs_writepage_setup(ctx, page, offset, count);
921 nfs_set_pageerror(page);
923 __set_page_dirty_nobuffers(page);
925 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
926 status, (long long)i_size_read(inode));
930 static int flush_task_priority(int how)
932 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
934 return RPC_PRIORITY_HIGH;
936 return RPC_PRIORITY_LOW;
938 return RPC_PRIORITY_NORMAL;
941 int nfs_initiate_write(struct rpc_clnt *clnt,
942 struct nfs_write_data *data,
943 const struct rpc_call_ops *call_ops,
946 struct inode *inode = data->header->inode;
947 int priority = flush_task_priority(how);
948 struct rpc_task *task;
949 struct rpc_message msg = {
950 .rpc_argp = &data->args,
951 .rpc_resp = &data->res,
952 .rpc_cred = data->header->cred,
954 struct rpc_task_setup task_setup_data = {
958 .callback_ops = call_ops,
959 .callback_data = data,
960 .workqueue = nfsiod_workqueue,
961 .flags = RPC_TASK_ASYNC | flags,
962 .priority = priority,
966 /* Set up the initial task struct. */
967 NFS_PROTO(inode)->write_setup(data, &msg);
969 dprintk("NFS: %5u initiated write call "
970 "(req %s/%lld, %u bytes @ offset %llu)\n",
973 (long long)NFS_FILEID(inode),
975 (unsigned long long)data->args.offset);
977 task = rpc_run_task(&task_setup_data);
982 if (how & FLUSH_SYNC) {
983 ret = rpc_wait_for_completion_task(task);
985 ret = task->tk_status;
991 EXPORT_SYMBOL_GPL(nfs_initiate_write);
994 * Set up the argument/result storage required for the RPC call.
996 static void nfs_write_rpcsetup(struct nfs_write_data *data,
997 unsigned int count, unsigned int offset,
998 int how, struct nfs_commit_info *cinfo)
1000 struct nfs_page *req = data->header->req;
1002 /* Set up the RPC argument and reply structs
1003 * NB: take care not to mess about with data->commit et al. */
1005 data->args.fh = NFS_FH(data->header->inode);
1006 data->args.offset = req_offset(req) + offset;
1007 /* pnfs_set_layoutcommit needs this */
1008 data->mds_offset = data->args.offset;
1009 data->args.pgbase = req->wb_pgbase + offset;
1010 data->args.pages = data->pages.pagevec;
1011 data->args.count = count;
1012 data->args.context = get_nfs_open_context(req->wb_context);
1013 data->args.lock_context = req->wb_lock_context;
1014 data->args.stable = NFS_UNSTABLE;
1015 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1018 case FLUSH_COND_STABLE:
1019 if (nfs_reqs_to_commit(cinfo))
1022 data->args.stable = NFS_FILE_SYNC;
1025 data->res.fattr = &data->fattr;
1026 data->res.count = count;
1027 data->res.verf = &data->verf;
1028 nfs_fattr_init(&data->fattr);
1031 static int nfs_do_write(struct nfs_write_data *data,
1032 const struct rpc_call_ops *call_ops,
1035 struct inode *inode = data->header->inode;
1037 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1040 static int nfs_do_multiple_writes(struct list_head *head,
1041 const struct rpc_call_ops *call_ops,
1044 struct nfs_write_data *data;
1047 while (!list_empty(head)) {
1050 data = list_first_entry(head, struct nfs_write_data, list);
1051 list_del_init(&data->list);
1053 ret2 = nfs_do_write(data, call_ops, how);
1060 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1061 * call this on each, which will prepare them to be retried on next
1062 * writeback using standard nfs.
1064 static void nfs_redirty_request(struct nfs_page *req)
1066 nfs_mark_request_dirty(req);
1067 nfs_unlock_request(req);
1068 nfs_end_page_writeback(req->wb_page);
1069 nfs_release_request(req);
1072 static void nfs_async_write_error(struct list_head *head)
1074 struct nfs_page *req;
1076 while (!list_empty(head)) {
1077 req = nfs_list_entry(head->next);
1078 nfs_list_remove_request(req);
1079 nfs_redirty_request(req);
1083 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1084 .error_cleanup = nfs_async_write_error,
1085 .completion = nfs_write_completion,
1088 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1089 struct nfs_pgio_header *hdr)
1091 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1092 while (!list_empty(&hdr->rpc_list)) {
1093 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1094 struct nfs_write_data, list);
1095 list_del(&data->list);
1096 nfs_writedata_release(data);
1098 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1102 * Generate multiple small requests to write out a single
1103 * contiguous dirty area on one page.
1105 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1106 struct nfs_pgio_header *hdr)
1108 struct nfs_page *req = hdr->req;
1109 struct page *page = req->wb_page;
1110 struct nfs_write_data *data;
1111 size_t wsize = desc->pg_bsize, nbytes;
1112 unsigned int offset;
1114 struct nfs_commit_info cinfo;
1116 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1118 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1119 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1120 desc->pg_count > wsize))
1121 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1125 nbytes = desc->pg_count;
1127 size_t len = min(nbytes, wsize);
1129 data = nfs_writedata_alloc(hdr, 1);
1131 nfs_flush_error(desc, hdr);
1134 data->pages.pagevec[0] = page;
1135 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1136 list_add(&data->list, &hdr->rpc_list);
1140 } while (nbytes != 0);
1141 nfs_list_remove_request(req);
1142 nfs_list_add_request(req, &hdr->pages);
1143 desc->pg_rpc_callops = &nfs_write_common_ops;
1148 * Create an RPC task for the given write request and kick it.
1149 * The page must have been locked by the caller.
1151 * It may happen that the page we're passed is not marked dirty.
1152 * This is the case if nfs_updatepage detects a conflicting request
1153 * that has been written but not committed.
1155 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1156 struct nfs_pgio_header *hdr)
1158 struct nfs_page *req;
1159 struct page **pages;
1160 struct nfs_write_data *data;
1161 struct list_head *head = &desc->pg_list;
1162 struct nfs_commit_info cinfo;
1164 data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1167 nfs_flush_error(desc, hdr);
1171 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1172 pages = data->pages.pagevec;
1173 while (!list_empty(head)) {
1174 req = nfs_list_entry(head->next);
1175 nfs_list_remove_request(req);
1176 nfs_list_add_request(req, &hdr->pages);
1177 *pages++ = req->wb_page;
1180 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1181 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1182 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1184 /* Set up the argument struct */
1185 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1186 list_add(&data->list, &hdr->rpc_list);
1187 desc->pg_rpc_callops = &nfs_write_common_ops;
1191 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1192 struct nfs_pgio_header *hdr)
1194 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1195 return nfs_flush_multi(desc, hdr);
1196 return nfs_flush_one(desc, hdr);
1198 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1200 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1202 struct nfs_write_header *whdr;
1203 struct nfs_pgio_header *hdr;
1206 whdr = nfs_writehdr_alloc();
1208 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1211 hdr = &whdr->header;
1212 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1213 atomic_inc(&hdr->refcnt);
1214 ret = nfs_generic_flush(desc, hdr);
1216 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1217 desc->pg_rpc_callops,
1219 if (atomic_dec_and_test(&hdr->refcnt))
1220 hdr->completion_ops->completion(hdr);
1224 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1225 .pg_test = nfs_generic_pg_test,
1226 .pg_doio = nfs_generic_pg_writepages,
1229 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1230 struct inode *inode, int ioflags,
1231 const struct nfs_pgio_completion_ops *compl_ops)
1233 nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1234 NFS_SERVER(inode)->wsize, ioflags);
1236 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1238 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1240 pgio->pg_ops = &nfs_pageio_write_ops;
1241 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1243 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1246 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1248 struct nfs_write_data *data = calldata;
1249 NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1252 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1254 struct nfs_commit_data *data = calldata;
1256 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1260 * Handle a write reply that flushes a whole page.
1262 * FIXME: There is an inherent race with invalidate_inode_pages and
1263 * writebacks since the page->count is kept > 1 for as long
1264 * as the page has a write request pending.
1266 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1268 struct nfs_write_data *data = calldata;
1270 nfs_writeback_done(task, data);
1273 static void nfs_writeback_release_common(void *calldata)
1275 struct nfs_write_data *data = calldata;
1276 struct nfs_pgio_header *hdr = data->header;
1277 int status = data->task.tk_status;
1279 if ((status >= 0) && nfs_write_need_commit(data)) {
1280 spin_lock(&hdr->lock);
1281 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1283 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1284 memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1285 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1286 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1287 spin_unlock(&hdr->lock);
1289 nfs_writedata_release(data);
1292 static const struct rpc_call_ops nfs_write_common_ops = {
1293 .rpc_call_prepare = nfs_write_prepare,
1294 .rpc_call_done = nfs_writeback_done_common,
1295 .rpc_release = nfs_writeback_release_common,
1300 * This function is called when the WRITE call is complete.
1302 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1304 struct nfs_writeargs *argp = &data->args;
1305 struct nfs_writeres *resp = &data->res;
1306 struct inode *inode = data->header->inode;
1309 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1310 task->tk_pid, task->tk_status);
1313 * ->write_done will attempt to use post-op attributes to detect
1314 * conflicting writes by other clients. A strict interpretation
1315 * of close-to-open would allow us to continue caching even if
1316 * another writer had changed the file, but some applications
1317 * depend on tighter cache coherency when writing.
1319 status = NFS_PROTO(inode)->write_done(task, data);
1322 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1324 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1325 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1326 /* We tried a write call, but the server did not
1327 * commit data to stable storage even though we
1329 * Note: There is a known bug in Tru64 < 5.0 in which
1330 * the server reports NFS_DATA_SYNC, but performs
1331 * NFS_FILE_SYNC. We therefore implement this checking
1332 * as a dprintk() in order to avoid filling syslog.
1334 static unsigned long complain;
1336 /* Note this will print the MDS for a DS write */
1337 if (time_before(complain, jiffies)) {
1338 dprintk("NFS: faulty NFS server %s:"
1339 " (committed = %d) != (stable = %d)\n",
1340 NFS_SERVER(inode)->nfs_client->cl_hostname,
1341 resp->verf->committed, argp->stable);
1342 complain = jiffies + 300 * HZ;
1346 if (task->tk_status < 0)
1347 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1348 else if (resp->count < argp->count) {
1349 static unsigned long complain;
1351 /* This a short write! */
1352 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1354 /* Has the server at least made some progress? */
1355 if (resp->count == 0) {
1356 if (time_before(complain, jiffies)) {
1358 "NFS: Server wrote zero bytes, expected %u.\n",
1360 complain = jiffies + 300 * HZ;
1362 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1363 task->tk_status = -EIO;
1366 /* Was this an NFSv2 write or an NFSv3 stable write? */
1367 if (resp->verf->committed != NFS_UNSTABLE) {
1368 /* Resend from where the server left off */
1369 data->mds_offset += resp->count;
1370 argp->offset += resp->count;
1371 argp->pgbase += resp->count;
1372 argp->count -= resp->count;
1374 /* Resend as a stable write in order to avoid
1375 * headaches in the case of a server crash.
1377 argp->stable = NFS_FILE_SYNC;
1379 rpc_restart_call_prepare(task);
1384 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1385 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1389 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1393 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1395 nfs_wait_bit_killable,
1397 return (ret < 0) ? ret : 1;
1400 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1402 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1403 smp_mb__after_clear_bit();
1404 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1407 void nfs_commitdata_release(struct nfs_commit_data *data)
1409 put_nfs_open_context(data->context);
1410 nfs_commit_free(data);
1412 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1414 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1415 const struct rpc_call_ops *call_ops,
1418 struct rpc_task *task;
1419 int priority = flush_task_priority(how);
1420 struct rpc_message msg = {
1421 .rpc_argp = &data->args,
1422 .rpc_resp = &data->res,
1423 .rpc_cred = data->cred,
1425 struct rpc_task_setup task_setup_data = {
1426 .task = &data->task,
1428 .rpc_message = &msg,
1429 .callback_ops = call_ops,
1430 .callback_data = data,
1431 .workqueue = nfsiod_workqueue,
1432 .flags = RPC_TASK_ASYNC | flags,
1433 .priority = priority,
1435 /* Set up the initial task struct. */
1436 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1438 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1440 task = rpc_run_task(&task_setup_data);
1442 return PTR_ERR(task);
1443 if (how & FLUSH_SYNC)
1444 rpc_wait_for_completion_task(task);
1448 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1451 * Set up the argument/result storage required for the RPC call.
1453 void nfs_init_commit(struct nfs_commit_data *data,
1454 struct list_head *head,
1455 struct pnfs_layout_segment *lseg,
1456 struct nfs_commit_info *cinfo)
1458 struct nfs_page *first = nfs_list_entry(head->next);
1459 struct inode *inode = first->wb_context->dentry->d_inode;
1461 /* Set up the RPC argument and reply structs
1462 * NB: take care not to mess about with data->commit et al. */
1464 list_splice_init(head, &data->pages);
1466 data->inode = inode;
1467 data->cred = first->wb_context->cred;
1468 data->lseg = lseg; /* reference transferred */
1469 data->mds_ops = &nfs_commit_ops;
1470 data->completion_ops = cinfo->completion_ops;
1471 data->dreq = cinfo->dreq;
1473 data->args.fh = NFS_FH(data->inode);
1474 /* Note: we always request a commit of the entire inode */
1475 data->args.offset = 0;
1476 data->args.count = 0;
1477 data->context = get_nfs_open_context(first->wb_context);
1478 data->res.fattr = &data->fattr;
1479 data->res.verf = &data->verf;
1480 nfs_fattr_init(&data->fattr);
1482 EXPORT_SYMBOL_GPL(nfs_init_commit);
1484 void nfs_retry_commit(struct list_head *page_list,
1485 struct pnfs_layout_segment *lseg,
1486 struct nfs_commit_info *cinfo)
1488 struct nfs_page *req;
1490 while (!list_empty(page_list)) {
1491 req = nfs_list_entry(page_list->next);
1492 nfs_list_remove_request(req);
1493 nfs_mark_request_commit(req, lseg, cinfo);
1495 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1496 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1499 nfs_unlock_and_release_request(req);
1502 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1505 * Commit dirty pages
1508 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1509 struct nfs_commit_info *cinfo)
1511 struct nfs_commit_data *data;
1513 data = nfs_commitdata_alloc();
1518 /* Set up the argument struct */
1519 nfs_init_commit(data, head, NULL, cinfo);
1520 atomic_inc(&cinfo->mds->rpcs_out);
1521 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1524 nfs_retry_commit(head, NULL, cinfo);
1525 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1530 * COMMIT call returned
1532 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1534 struct nfs_commit_data *data = calldata;
1536 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1537 task->tk_pid, task->tk_status);
1539 /* Call the NFS version-specific code */
1540 NFS_PROTO(data->inode)->commit_done(task, data);
1543 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1545 struct nfs_page *req;
1546 int status = data->task.tk_status;
1547 struct nfs_commit_info cinfo;
1549 while (!list_empty(&data->pages)) {
1550 req = nfs_list_entry(data->pages.next);
1551 nfs_list_remove_request(req);
1552 nfs_clear_page_commit(req->wb_page);
1554 dprintk("NFS: commit (%s/%lld %d@%lld)",
1555 req->wb_context->dentry->d_sb->s_id,
1556 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1558 (long long)req_offset(req));
1560 nfs_context_set_write_error(req->wb_context, status);
1561 nfs_inode_remove_request(req);
1562 dprintk(", error = %d\n", status);
1566 /* Okay, COMMIT succeeded, apparently. Check the verifier
1567 * returned by the server against all stored verfs. */
1568 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1569 /* We have a match */
1570 nfs_inode_remove_request(req);
1574 /* We have a mismatch. Write the page again */
1575 dprintk(" mismatch\n");
1576 nfs_mark_request_dirty(req);
1577 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1579 nfs_unlock_and_release_request(req);
1581 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1582 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1583 nfs_commit_clear_lock(NFS_I(data->inode));
1586 static void nfs_commit_release(void *calldata)
1588 struct nfs_commit_data *data = calldata;
1590 data->completion_ops->completion(data);
1591 nfs_commitdata_release(calldata);
1594 static const struct rpc_call_ops nfs_commit_ops = {
1595 .rpc_call_prepare = nfs_commit_prepare,
1596 .rpc_call_done = nfs_commit_done,
1597 .rpc_release = nfs_commit_release,
1600 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1601 .completion = nfs_commit_release_pages,
1602 .error_cleanup = nfs_commit_clear_lock,
1605 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1606 int how, struct nfs_commit_info *cinfo)
1610 status = pnfs_commit_list(inode, head, how, cinfo);
1611 if (status == PNFS_NOT_ATTEMPTED)
1612 status = nfs_commit_list(inode, head, how, cinfo);
1616 int nfs_commit_inode(struct inode *inode, int how)
1619 struct nfs_commit_info cinfo;
1620 int may_wait = how & FLUSH_SYNC;
1623 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1625 goto out_mark_dirty;
1626 nfs_init_cinfo_from_inode(&cinfo, inode);
1627 res = nfs_scan_commit(inode, &head, &cinfo);
1631 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1635 goto out_mark_dirty;
1636 error = wait_on_bit(&NFS_I(inode)->flags,
1638 nfs_wait_bit_killable,
1643 nfs_commit_clear_lock(NFS_I(inode));
1645 /* Note: If we exit without ensuring that the commit is complete,
1646 * we must mark the inode as dirty. Otherwise, future calls to
1647 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1648 * that the data is on the disk.
1651 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1655 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1657 struct nfs_inode *nfsi = NFS_I(inode);
1658 int flags = FLUSH_SYNC;
1661 /* no commits means nothing needs to be done */
1662 if (!nfsi->commit_info.ncommit)
1665 if (wbc->sync_mode == WB_SYNC_NONE) {
1666 /* Don't commit yet if this is a non-blocking flush and there
1667 * are a lot of outstanding writes for this mapping.
1669 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1670 goto out_mark_dirty;
1672 /* don't wait for the COMMIT response */
1676 ret = nfs_commit_inode(inode, flags);
1678 if (wbc->sync_mode == WB_SYNC_NONE) {
1679 if (ret < wbc->nr_to_write)
1680 wbc->nr_to_write -= ret;
1682 wbc->nr_to_write = 0;
1687 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1691 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1697 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1699 return nfs_commit_unstable_pages(inode, wbc);
1701 EXPORT_SYMBOL_GPL(nfs_write_inode);
1704 * flush the inode to disk.
1706 int nfs_wb_all(struct inode *inode)
1708 struct writeback_control wbc = {
1709 .sync_mode = WB_SYNC_ALL,
1710 .nr_to_write = LONG_MAX,
1712 .range_end = LLONG_MAX,
1715 return sync_inode(inode, &wbc);
1717 EXPORT_SYMBOL_GPL(nfs_wb_all);
1719 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1721 struct nfs_page *req;
1725 wait_on_page_writeback(page);
1726 req = nfs_page_find_request(page);
1729 if (nfs_lock_request(req)) {
1730 nfs_clear_request_commit(req);
1731 nfs_inode_remove_request(req);
1733 * In case nfs_inode_remove_request has marked the
1734 * page as being dirty
1736 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1737 nfs_unlock_and_release_request(req);
1740 ret = nfs_wait_on_request(req);
1741 nfs_release_request(req);
1749 * Write back all requests on one page - we do this before reading it.
1751 int nfs_wb_page(struct inode *inode, struct page *page)
1753 loff_t range_start = page_file_offset(page);
1754 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1755 struct writeback_control wbc = {
1756 .sync_mode = WB_SYNC_ALL,
1758 .range_start = range_start,
1759 .range_end = range_end,
1764 wait_on_page_writeback(page);
1765 if (clear_page_dirty_for_io(page)) {
1766 ret = nfs_writepage_locked(page, &wbc);
1771 if (!PagePrivate(page))
1773 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1782 #ifdef CONFIG_MIGRATION
1783 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1784 struct page *page, enum migrate_mode mode)
1787 * If PagePrivate is set, then the page is currently associated with
1788 * an in-progress read or write request. Don't try to migrate it.
1790 * FIXME: we could do this in principle, but we'll need a way to ensure
1791 * that we can safely release the inode reference while holding
1794 if (PagePrivate(page))
1797 nfs_fscache_release_page(page, GFP_KERNEL);
1799 return migrate_page(mapping, newpage, page, mode);
1803 int __init nfs_init_writepagecache(void)
1805 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1806 sizeof(struct nfs_write_header),
1807 0, SLAB_HWCACHE_ALIGN,
1809 if (nfs_wdata_cachep == NULL)
1812 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1814 if (nfs_wdata_mempool == NULL)
1815 goto out_destroy_write_cache;
1817 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1818 sizeof(struct nfs_commit_data),
1819 0, SLAB_HWCACHE_ALIGN,
1821 if (nfs_cdata_cachep == NULL)
1822 goto out_destroy_write_mempool;
1824 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1826 if (nfs_commit_mempool == NULL)
1827 goto out_destroy_commit_cache;
1830 * NFS congestion size, scale with available memory.
1842 * This allows larger machines to have larger/more transfers.
1843 * Limit the default to 256M
1845 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1846 if (nfs_congestion_kb > 256*1024)
1847 nfs_congestion_kb = 256*1024;
1851 out_destroy_commit_cache:
1852 kmem_cache_destroy(nfs_cdata_cachep);
1853 out_destroy_write_mempool:
1854 mempool_destroy(nfs_wdata_mempool);
1855 out_destroy_write_cache:
1856 kmem_cache_destroy(nfs_wdata_cachep);
1860 void nfs_destroy_writepagecache(void)
1862 mempool_destroy(nfs_commit_mempool);
1863 kmem_cache_destroy(nfs_cdata_cachep);
1864 mempool_destroy(nfs_wdata_mempool);
1865 kmem_cache_destroy(nfs_wdata_cachep);