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_write_header *nfs_writehdr_alloc(void)
75 struct nfs_write_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);
89 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
91 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
92 unsigned int pagecount)
94 struct nfs_write_data *data, *prealloc;
96 prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
97 if (prealloc->header == NULL)
100 data = kzalloc(sizeof(*data), GFP_KERNEL);
104 if (nfs_pgarray_set(&data->pages, pagecount)) {
106 atomic_inc(&hdr->refcnt);
108 if (data != prealloc)
116 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
118 struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
119 mempool_free(whdr, nfs_wdata_mempool);
121 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
123 void nfs_writedata_release(struct nfs_write_data *wdata)
125 struct nfs_pgio_header *hdr = wdata->header;
126 struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
128 put_nfs_open_context(wdata->args.context);
129 if (wdata->pages.pagevec != wdata->pages.page_array)
130 kfree(wdata->pages.pagevec);
131 if (wdata == &write_header->rpc_data) {
132 wdata->header = NULL;
135 if (atomic_dec_and_test(&hdr->refcnt))
136 hdr->completion_ops->completion(hdr);
137 /* Note: we only free the rpc_task after callbacks are done.
138 * See the comment in rpc_free_task() for why
142 EXPORT_SYMBOL_GPL(nfs_writedata_release);
144 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
148 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
151 static struct nfs_page *
152 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
154 struct nfs_page *req = NULL;
156 if (PagePrivate(page))
157 req = (struct nfs_page *)page_private(page);
158 else if (unlikely(PageSwapCache(page))) {
159 struct nfs_page *freq, *t;
161 /* Linearly search the commit list for the correct req */
162 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
163 if (freq->wb_page == page) {
171 kref_get(&req->wb_kref);
176 static struct nfs_page *nfs_page_find_request(struct page *page)
178 struct inode *inode = page_file_mapping(page)->host;
179 struct nfs_page *req = NULL;
181 spin_lock(&inode->i_lock);
182 req = nfs_page_find_request_locked(NFS_I(inode), page);
183 spin_unlock(&inode->i_lock);
187 /* Adjust the file length if we're writing beyond the end */
188 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
190 struct inode *inode = page_file_mapping(page)->host;
194 spin_lock(&inode->i_lock);
195 i_size = i_size_read(inode);
196 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
197 if (i_size > 0 && page_file_index(page) < end_index)
199 end = page_file_offset(page) + ((loff_t)offset+count);
202 i_size_write(inode, end);
203 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
205 spin_unlock(&inode->i_lock);
208 /* A writeback failed: mark the page as bad, and invalidate the page cache */
209 static void nfs_set_pageerror(struct page *page)
211 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
214 /* We can set the PG_uptodate flag if we see that a write request
215 * covers the full page.
217 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
219 if (PageUptodate(page))
223 if (count != nfs_page_length(page))
225 SetPageUptodate(page);
228 static int wb_priority(struct writeback_control *wbc)
230 if (wbc->for_reclaim)
231 return FLUSH_HIGHPRI | FLUSH_STABLE;
232 if (wbc->for_kupdate || wbc->for_background)
233 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
234 return FLUSH_COND_STABLE;
238 * NFS congestion control
241 int nfs_congestion_kb;
243 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
244 #define NFS_CONGESTION_OFF_THRESH \
245 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
247 static void nfs_set_page_writeback(struct page *page)
249 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
250 int ret = test_set_page_writeback(page);
252 WARN_ON_ONCE(ret != 0);
254 if (atomic_long_inc_return(&nfss->writeback) >
255 NFS_CONGESTION_ON_THRESH) {
256 set_bdi_congested(&nfss->backing_dev_info,
261 static void nfs_end_page_writeback(struct page *page)
263 struct inode *inode = page_file_mapping(page)->host;
264 struct nfs_server *nfss = NFS_SERVER(inode);
266 end_page_writeback(page);
267 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
268 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
271 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
273 struct inode *inode = page_file_mapping(page)->host;
274 struct nfs_page *req;
277 spin_lock(&inode->i_lock);
279 req = nfs_page_find_request_locked(NFS_I(inode), page);
282 if (nfs_lock_request(req))
284 /* Note: If we hold the page lock, as is the case in nfs_writepage,
285 * then the call to nfs_lock_request() will always
286 * succeed provided that someone hasn't already marked the
287 * request as dirty (in which case we don't care).
289 spin_unlock(&inode->i_lock);
291 ret = nfs_wait_on_request(req);
294 nfs_release_request(req);
297 spin_lock(&inode->i_lock);
299 spin_unlock(&inode->i_lock);
304 * Find an associated nfs write request, and prepare to flush it out
305 * May return an error if the user signalled nfs_wait_on_request().
307 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
308 struct page *page, bool nonblock)
310 struct nfs_page *req;
313 req = nfs_find_and_lock_request(page, nonblock);
320 nfs_set_page_writeback(page);
321 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
324 if (!nfs_pageio_add_request(pgio, req)) {
325 nfs_redirty_request(req);
326 ret = pgio->pg_error;
332 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
334 struct inode *inode = page_file_mapping(page)->host;
337 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
338 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
340 nfs_pageio_cond_complete(pgio, page_file_index(page));
341 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
342 if (ret == -EAGAIN) {
343 redirty_page_for_writepage(wbc, page);
350 * Write an mmapped page to the server.
352 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
354 struct nfs_pageio_descriptor pgio;
357 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
358 false, &nfs_async_write_completion_ops);
359 err = nfs_do_writepage(page, wbc, &pgio);
360 nfs_pageio_complete(&pgio);
363 if (pgio.pg_error < 0)
364 return pgio.pg_error;
368 int nfs_writepage(struct page *page, struct writeback_control *wbc)
372 ret = nfs_writepage_locked(page, wbc);
377 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
381 ret = nfs_do_writepage(page, wbc, data);
386 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
388 struct inode *inode = mapping->host;
389 unsigned long *bitlock = &NFS_I(inode)->flags;
390 struct nfs_pageio_descriptor pgio;
393 /* Stop dirtying of new pages while we sync */
394 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
395 nfs_wait_bit_killable, TASK_KILLABLE);
399 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
401 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
402 &nfs_async_write_completion_ops);
403 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
404 nfs_pageio_complete(&pgio);
406 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
407 smp_mb__after_clear_bit();
408 wake_up_bit(bitlock, NFS_INO_FLUSHING);
421 * Insert a write request into an inode
423 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
425 struct nfs_inode *nfsi = NFS_I(inode);
427 /* Lock the request! */
428 nfs_lock_request(req);
430 spin_lock(&inode->i_lock);
431 if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
434 * Swap-space should not get truncated. Hence no need to plug the race
435 * with invalidate/truncate.
437 if (likely(!PageSwapCache(req->wb_page))) {
438 set_bit(PG_MAPPED, &req->wb_flags);
439 SetPagePrivate(req->wb_page);
440 set_page_private(req->wb_page, (unsigned long)req);
443 kref_get(&req->wb_kref);
444 spin_unlock(&inode->i_lock);
448 * Remove a write request from an inode
450 static void nfs_inode_remove_request(struct nfs_page *req)
452 struct inode *inode = req->wb_context->dentry->d_inode;
453 struct nfs_inode *nfsi = NFS_I(inode);
455 spin_lock(&inode->i_lock);
456 if (likely(!PageSwapCache(req->wb_page))) {
457 set_page_private(req->wb_page, 0);
458 ClearPagePrivate(req->wb_page);
459 clear_bit(PG_MAPPED, &req->wb_flags);
462 spin_unlock(&inode->i_lock);
463 nfs_release_request(req);
467 nfs_mark_request_dirty(struct nfs_page *req)
469 __set_page_dirty_nobuffers(req->wb_page);
472 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
474 * nfs_request_add_commit_list - add request to a commit list
475 * @req: pointer to a struct nfs_page
476 * @dst: commit list head
477 * @cinfo: holds list lock and accounting info
479 * This sets the PG_CLEAN bit, updates the cinfo count of
480 * number of outstanding requests requiring a commit as well as
483 * The caller must _not_ hold the cinfo->lock, but must be
484 * holding the nfs_page lock.
487 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
488 struct nfs_commit_info *cinfo)
490 set_bit(PG_CLEAN, &(req)->wb_flags);
491 spin_lock(cinfo->lock);
492 nfs_list_add_request(req, dst);
493 cinfo->mds->ncommit++;
494 spin_unlock(cinfo->lock);
496 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
497 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
499 __mark_inode_dirty(req->wb_context->dentry->d_inode,
503 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
506 * nfs_request_remove_commit_list - Remove request from a commit list
507 * @req: pointer to a nfs_page
508 * @cinfo: holds list lock and accounting info
510 * This clears the PG_CLEAN bit, and updates the cinfo's count of
511 * number of outstanding requests requiring a commit
512 * It does not update the MM page stats.
514 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
517 nfs_request_remove_commit_list(struct nfs_page *req,
518 struct nfs_commit_info *cinfo)
520 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
522 nfs_list_remove_request(req);
523 cinfo->mds->ncommit--;
525 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
527 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
530 cinfo->lock = &inode->i_lock;
531 cinfo->mds = &NFS_I(inode)->commit_info;
532 cinfo->ds = pnfs_get_ds_info(inode);
534 cinfo->completion_ops = &nfs_commit_completion_ops;
537 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
539 struct nfs_direct_req *dreq)
542 nfs_init_cinfo_from_dreq(cinfo, dreq);
544 nfs_init_cinfo_from_inode(cinfo, inode);
546 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
549 * Add a request to the inode's commit list.
552 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
553 struct nfs_commit_info *cinfo)
555 if (pnfs_mark_request_commit(req, lseg, cinfo))
557 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
561 nfs_clear_page_commit(struct page *page)
563 dec_zone_page_state(page, NR_UNSTABLE_NFS);
564 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
568 nfs_clear_request_commit(struct nfs_page *req)
570 if (test_bit(PG_CLEAN, &req->wb_flags)) {
571 struct inode *inode = req->wb_context->dentry->d_inode;
572 struct nfs_commit_info cinfo;
574 nfs_init_cinfo_from_inode(&cinfo, inode);
575 if (!pnfs_clear_request_commit(req, &cinfo)) {
576 spin_lock(cinfo.lock);
577 nfs_request_remove_commit_list(req, &cinfo);
578 spin_unlock(cinfo.lock);
580 nfs_clear_page_commit(req->wb_page);
585 int nfs_write_need_commit(struct nfs_write_data *data)
587 if (data->verf.committed == NFS_DATA_SYNC)
588 return data->header->lseg == NULL;
589 return data->verf.committed != NFS_FILE_SYNC;
593 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
598 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
600 struct nfs_direct_req *dreq)
605 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
606 struct nfs_commit_info *cinfo)
611 nfs_clear_request_commit(struct nfs_page *req)
616 int nfs_write_need_commit(struct nfs_write_data *data)
623 static void nfs_write_completion(struct nfs_pgio_header *hdr)
625 struct nfs_commit_info cinfo;
626 unsigned long bytes = 0;
628 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
630 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
631 while (!list_empty(&hdr->pages)) {
632 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
634 bytes += req->wb_bytes;
635 nfs_list_remove_request(req);
636 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
637 (hdr->good_bytes < bytes)) {
638 nfs_set_pageerror(req->wb_page);
639 nfs_context_set_write_error(req->wb_context, hdr->error);
642 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
643 nfs_mark_request_dirty(req);
646 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
647 memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
648 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
652 nfs_inode_remove_request(req);
654 nfs_unlock_request(req);
655 nfs_end_page_writeback(req->wb_page);
656 nfs_release_request(req);
662 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
664 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
666 return cinfo->mds->ncommit;
669 /* cinfo->lock held by caller */
671 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
672 struct nfs_commit_info *cinfo, int max)
674 struct nfs_page *req, *tmp;
677 list_for_each_entry_safe(req, tmp, src, wb_list) {
678 if (!nfs_lock_request(req))
680 kref_get(&req->wb_kref);
681 if (cond_resched_lock(cinfo->lock))
682 list_safe_reset_next(req, tmp, wb_list);
683 nfs_request_remove_commit_list(req, cinfo);
684 nfs_list_add_request(req, dst);
686 if ((ret == max) && !cinfo->dreq)
693 * nfs_scan_commit - Scan an inode for commit requests
694 * @inode: NFS inode to scan
695 * @dst: mds destination list
696 * @cinfo: mds and ds lists of reqs ready to commit
698 * Moves requests from the inode's 'commit' request list.
699 * The requests are *not* checked to ensure that they form a contiguous set.
702 nfs_scan_commit(struct inode *inode, struct list_head *dst,
703 struct nfs_commit_info *cinfo)
707 spin_lock(cinfo->lock);
708 if (cinfo->mds->ncommit > 0) {
709 const int max = INT_MAX;
711 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
713 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
715 spin_unlock(cinfo->lock);
720 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
725 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
726 struct nfs_commit_info *cinfo)
733 * Search for an existing write request, and attempt to update
734 * it to reflect a new dirty region on a given page.
736 * If the attempt fails, then the existing request is flushed out
739 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
744 struct nfs_page *req;
749 if (!PagePrivate(page))
752 end = offset + bytes;
753 spin_lock(&inode->i_lock);
756 req = nfs_page_find_request_locked(NFS_I(inode), page);
760 rqend = req->wb_offset + req->wb_bytes;
762 * Tell the caller to flush out the request if
763 * the offsets are non-contiguous.
764 * Note: nfs_flush_incompatible() will already
765 * have flushed out requests having wrong owners.
768 || end < req->wb_offset)
771 if (nfs_lock_request(req))
774 /* The request is locked, so wait and then retry */
775 spin_unlock(&inode->i_lock);
776 error = nfs_wait_on_request(req);
777 nfs_release_request(req);
780 spin_lock(&inode->i_lock);
783 /* Okay, the request matches. Update the region */
784 if (offset < req->wb_offset) {
785 req->wb_offset = offset;
786 req->wb_pgbase = offset;
789 req->wb_bytes = end - req->wb_offset;
791 req->wb_bytes = rqend - req->wb_offset;
793 spin_unlock(&inode->i_lock);
795 nfs_clear_request_commit(req);
798 spin_unlock(&inode->i_lock);
799 nfs_release_request(req);
800 error = nfs_wb_page(inode, page);
802 return ERR_PTR(error);
806 * Try to update an existing write request, or create one if there is none.
808 * Note: Should always be called with the Page Lock held to prevent races
809 * if we have to add a new request. Also assumes that the caller has
810 * already called nfs_flush_incompatible() if necessary.
812 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
813 struct page *page, unsigned int offset, unsigned int bytes)
815 struct inode *inode = page_file_mapping(page)->host;
816 struct nfs_page *req;
818 req = nfs_try_to_update_request(inode, page, offset, bytes);
821 req = nfs_create_request(ctx, inode, page, offset, bytes);
824 nfs_inode_add_request(inode, req);
829 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
830 unsigned int offset, unsigned int count)
832 struct nfs_page *req;
834 req = nfs_setup_write_request(ctx, page, offset, count);
837 /* Update file length */
838 nfs_grow_file(page, offset, count);
839 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
840 nfs_mark_request_dirty(req);
841 nfs_unlock_and_release_request(req);
845 int nfs_flush_incompatible(struct file *file, struct page *page)
847 struct nfs_open_context *ctx = nfs_file_open_context(file);
848 struct nfs_lock_context *l_ctx;
849 struct nfs_page *req;
850 int do_flush, status;
852 * Look for a request corresponding to this page. If there
853 * is one, and it belongs to another file, we flush it out
854 * before we try to copy anything into the page. Do this
855 * due to the lack of an ACCESS-type call in NFSv2.
856 * Also do the same if we find a request from an existing
860 req = nfs_page_find_request(page);
863 l_ctx = req->wb_lock_context;
864 do_flush = req->wb_page != page || req->wb_context != ctx;
865 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
866 do_flush |= l_ctx->lockowner.l_owner != current->files
867 || l_ctx->lockowner.l_pid != current->tgid;
869 nfs_release_request(req);
872 status = nfs_wb_page(page_file_mapping(page)->host, page);
873 } while (status == 0);
878 * Avoid buffered writes when a open context credential's key would
881 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
883 * Return 0 and set a credential flag which triggers the inode to flush
884 * and performs NFS_FILE_SYNC writes if the key will expired within
885 * RPC_KEY_EXPIRE_TIMEO.
888 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
890 struct nfs_open_context *ctx = nfs_file_open_context(filp);
891 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
893 return rpcauth_key_timeout_notify(auth, ctx->cred);
897 * Test if the open context credential key is marked to expire soon.
899 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
901 return rpcauth_cred_key_to_expire(ctx->cred);
905 * If the page cache is marked as unsafe or invalid, then we can't rely on
906 * the PageUptodate() flag. In this case, we will need to turn off
907 * write optimisations that depend on the page contents being correct.
909 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
911 struct nfs_inode *nfsi = NFS_I(inode);
913 if (nfs_have_delegated_attributes(inode))
915 if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
918 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
921 return PageUptodate(page) != 0;
924 /* If we know the page is up to date, and we're not using byte range locks (or
925 * if we have the whole file locked for writing), it may be more efficient to
926 * extend the write to cover the entire page in order to avoid fragmentation
929 * If the file is opened for synchronous writes then we can just skip the rest
932 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
934 if (file->f_flags & O_DSYNC)
936 if (!nfs_write_pageuptodate(page, inode))
938 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
940 if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
941 inode->i_flock->fl_end == OFFSET_MAX &&
942 inode->i_flock->fl_type != F_RDLCK))
948 * Update and possibly write a cached page of an NFS file.
950 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
951 * things with a page scheduled for an RPC call (e.g. invalidate it).
953 int nfs_updatepage(struct file *file, struct page *page,
954 unsigned int offset, unsigned int count)
956 struct nfs_open_context *ctx = nfs_file_open_context(file);
957 struct inode *inode = page_file_mapping(page)->host;
960 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
962 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
963 file, count, (long long)(page_file_offset(page) + offset));
965 if (nfs_can_extend_write(file, page, inode)) {
966 count = max(count + offset, nfs_page_length(page));
970 status = nfs_writepage_setup(ctx, page, offset, count);
972 nfs_set_pageerror(page);
974 __set_page_dirty_nobuffers(page);
976 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
977 status, (long long)i_size_read(inode));
981 static int flush_task_priority(int how)
983 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
985 return RPC_PRIORITY_HIGH;
987 return RPC_PRIORITY_LOW;
989 return RPC_PRIORITY_NORMAL;
992 int nfs_initiate_write(struct rpc_clnt *clnt,
993 struct nfs_write_data *data,
994 const struct rpc_call_ops *call_ops,
997 struct inode *inode = data->header->inode;
998 int priority = flush_task_priority(how);
999 struct rpc_task *task;
1000 struct rpc_message msg = {
1001 .rpc_argp = &data->args,
1002 .rpc_resp = &data->res,
1003 .rpc_cred = data->header->cred,
1005 struct rpc_task_setup task_setup_data = {
1007 .task = &data->task,
1008 .rpc_message = &msg,
1009 .callback_ops = call_ops,
1010 .callback_data = data,
1011 .workqueue = nfsiod_workqueue,
1012 .flags = RPC_TASK_ASYNC | flags,
1013 .priority = priority,
1017 /* Set up the initial task struct. */
1018 NFS_PROTO(inode)->write_setup(data, &msg);
1020 dprintk("NFS: %5u initiated write call "
1021 "(req %s/%llu, %u bytes @ offset %llu)\n",
1024 (unsigned long long)NFS_FILEID(inode),
1026 (unsigned long long)data->args.offset);
1028 nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1029 &task_setup_data.rpc_client, &msg, data);
1031 task = rpc_run_task(&task_setup_data);
1033 ret = PTR_ERR(task);
1036 if (how & FLUSH_SYNC) {
1037 ret = rpc_wait_for_completion_task(task);
1039 ret = task->tk_status;
1045 EXPORT_SYMBOL_GPL(nfs_initiate_write);
1048 * Set up the argument/result storage required for the RPC call.
1050 static void nfs_write_rpcsetup(struct nfs_write_data *data,
1051 unsigned int count, unsigned int offset,
1052 int how, struct nfs_commit_info *cinfo)
1054 struct nfs_page *req = data->header->req;
1056 /* Set up the RPC argument and reply structs
1057 * NB: take care not to mess about with data->commit et al. */
1059 data->args.fh = NFS_FH(data->header->inode);
1060 data->args.offset = req_offset(req) + offset;
1061 /* pnfs_set_layoutcommit needs this */
1062 data->mds_offset = data->args.offset;
1063 data->args.pgbase = req->wb_pgbase + offset;
1064 data->args.pages = data->pages.pagevec;
1065 data->args.count = count;
1066 data->args.context = get_nfs_open_context(req->wb_context);
1067 data->args.lock_context = req->wb_lock_context;
1068 data->args.stable = NFS_UNSTABLE;
1069 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1072 case FLUSH_COND_STABLE:
1073 if (nfs_reqs_to_commit(cinfo))
1076 data->args.stable = NFS_FILE_SYNC;
1079 data->res.fattr = &data->fattr;
1080 data->res.count = count;
1081 data->res.verf = &data->verf;
1082 nfs_fattr_init(&data->fattr);
1085 static int nfs_do_write(struct nfs_write_data *data,
1086 const struct rpc_call_ops *call_ops,
1089 struct inode *inode = data->header->inode;
1091 return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1094 static int nfs_do_multiple_writes(struct list_head *head,
1095 const struct rpc_call_ops *call_ops,
1098 struct nfs_write_data *data;
1101 while (!list_empty(head)) {
1104 data = list_first_entry(head, struct nfs_write_data, list);
1105 list_del_init(&data->list);
1107 ret2 = nfs_do_write(data, call_ops, how);
1114 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1115 * call this on each, which will prepare them to be retried on next
1116 * writeback using standard nfs.
1118 static void nfs_redirty_request(struct nfs_page *req)
1120 nfs_mark_request_dirty(req);
1121 nfs_unlock_request(req);
1122 nfs_end_page_writeback(req->wb_page);
1123 nfs_release_request(req);
1126 static void nfs_async_write_error(struct list_head *head)
1128 struct nfs_page *req;
1130 while (!list_empty(head)) {
1131 req = nfs_list_entry(head->next);
1132 nfs_list_remove_request(req);
1133 nfs_redirty_request(req);
1137 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1138 .error_cleanup = nfs_async_write_error,
1139 .completion = nfs_write_completion,
1142 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1143 struct nfs_pgio_header *hdr)
1145 set_bit(NFS_IOHDR_REDO, &hdr->flags);
1146 while (!list_empty(&hdr->rpc_list)) {
1147 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1148 struct nfs_write_data, list);
1149 list_del(&data->list);
1150 nfs_writedata_release(data);
1152 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1156 * Generate multiple small requests to write out a single
1157 * contiguous dirty area on one page.
1159 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1160 struct nfs_pgio_header *hdr)
1162 struct nfs_page *req = hdr->req;
1163 struct page *page = req->wb_page;
1164 struct nfs_write_data *data;
1165 size_t wsize = desc->pg_bsize, nbytes;
1166 unsigned int offset;
1168 struct nfs_commit_info cinfo;
1170 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1172 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1173 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1174 desc->pg_count > wsize))
1175 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1179 nbytes = desc->pg_count;
1181 size_t len = min(nbytes, wsize);
1183 data = nfs_writedata_alloc(hdr, 1);
1185 nfs_flush_error(desc, hdr);
1188 data->pages.pagevec[0] = page;
1189 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1190 list_add(&data->list, &hdr->rpc_list);
1194 } while (nbytes != 0);
1195 nfs_list_remove_request(req);
1196 nfs_list_add_request(req, &hdr->pages);
1197 desc->pg_rpc_callops = &nfs_write_common_ops;
1202 * Create an RPC task for the given write request and kick it.
1203 * The page must have been locked by the caller.
1205 * It may happen that the page we're passed is not marked dirty.
1206 * This is the case if nfs_updatepage detects a conflicting request
1207 * that has been written but not committed.
1209 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1210 struct nfs_pgio_header *hdr)
1212 struct nfs_page *req;
1213 struct page **pages;
1214 struct nfs_write_data *data;
1215 struct list_head *head = &desc->pg_list;
1216 struct nfs_commit_info cinfo;
1218 data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1221 nfs_flush_error(desc, hdr);
1225 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1226 pages = data->pages.pagevec;
1227 while (!list_empty(head)) {
1228 req = nfs_list_entry(head->next);
1229 nfs_list_remove_request(req);
1230 nfs_list_add_request(req, &hdr->pages);
1231 *pages++ = req->wb_page;
1234 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1235 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1236 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1238 /* Set up the argument struct */
1239 nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1240 list_add(&data->list, &hdr->rpc_list);
1241 desc->pg_rpc_callops = &nfs_write_common_ops;
1245 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1246 struct nfs_pgio_header *hdr)
1248 if (desc->pg_bsize < PAGE_CACHE_SIZE)
1249 return nfs_flush_multi(desc, hdr);
1250 return nfs_flush_one(desc, hdr);
1252 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1254 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1256 struct nfs_write_header *whdr;
1257 struct nfs_pgio_header *hdr;
1260 whdr = nfs_writehdr_alloc();
1262 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1265 hdr = &whdr->header;
1266 nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1267 atomic_inc(&hdr->refcnt);
1268 ret = nfs_generic_flush(desc, hdr);
1270 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1271 desc->pg_rpc_callops,
1273 if (atomic_dec_and_test(&hdr->refcnt))
1274 hdr->completion_ops->completion(hdr);
1278 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1279 .pg_test = nfs_generic_pg_test,
1280 .pg_doio = nfs_generic_pg_writepages,
1283 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1284 struct inode *inode, int ioflags, bool force_mds,
1285 const struct nfs_pgio_completion_ops *compl_ops)
1287 struct nfs_server *server = NFS_SERVER(inode);
1288 const struct nfs_pageio_ops *pg_ops = &nfs_pageio_write_ops;
1290 #ifdef CONFIG_NFS_V4_1
1291 if (server->pnfs_curr_ld && !force_mds)
1292 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1294 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, server->wsize, ioflags);
1296 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1298 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1300 pgio->pg_ops = &nfs_pageio_write_ops;
1301 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1303 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1306 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1308 struct nfs_write_data *data = calldata;
1310 err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1312 rpc_exit(task, err);
1315 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1317 struct nfs_commit_data *data = calldata;
1319 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1323 * Handle a write reply that flushes a whole page.
1325 * FIXME: There is an inherent race with invalidate_inode_pages and
1326 * writebacks since the page->count is kept > 1 for as long
1327 * as the page has a write request pending.
1329 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1331 struct nfs_write_data *data = calldata;
1333 nfs_writeback_done(task, data);
1336 static void nfs_writeback_release_common(void *calldata)
1338 struct nfs_write_data *data = calldata;
1339 struct nfs_pgio_header *hdr = data->header;
1340 int status = data->task.tk_status;
1342 if ((status >= 0) && nfs_write_need_commit(data)) {
1343 spin_lock(&hdr->lock);
1344 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1346 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1347 memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1348 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1349 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1350 spin_unlock(&hdr->lock);
1352 nfs_writedata_release(data);
1355 static const struct rpc_call_ops nfs_write_common_ops = {
1356 .rpc_call_prepare = nfs_write_prepare,
1357 .rpc_call_done = nfs_writeback_done_common,
1358 .rpc_release = nfs_writeback_release_common,
1362 * Special version of should_remove_suid() that ignores capabilities.
1364 static int nfs_should_remove_suid(const struct inode *inode)
1366 umode_t mode = inode->i_mode;
1369 /* suid always must be killed */
1370 if (unlikely(mode & S_ISUID))
1371 kill = ATTR_KILL_SUID;
1374 * sgid without any exec bits is just a mandatory locking mark; leave
1375 * it alone. If some exec bits are set, it's a real sgid; kill it.
1377 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1378 kill |= ATTR_KILL_SGID;
1380 if (unlikely(kill && S_ISREG(mode)))
1387 * This function is called when the WRITE call is complete.
1389 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1391 struct nfs_writeargs *argp = &data->args;
1392 struct nfs_writeres *resp = &data->res;
1393 struct inode *inode = data->header->inode;
1396 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1397 task->tk_pid, task->tk_status);
1400 * ->write_done will attempt to use post-op attributes to detect
1401 * conflicting writes by other clients. A strict interpretation
1402 * of close-to-open would allow us to continue caching even if
1403 * another writer had changed the file, but some applications
1404 * depend on tighter cache coherency when writing.
1406 status = NFS_PROTO(inode)->write_done(task, data);
1409 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1411 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1412 if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1413 /* We tried a write call, but the server did not
1414 * commit data to stable storage even though we
1416 * Note: There is a known bug in Tru64 < 5.0 in which
1417 * the server reports NFS_DATA_SYNC, but performs
1418 * NFS_FILE_SYNC. We therefore implement this checking
1419 * as a dprintk() in order to avoid filling syslog.
1421 static unsigned long complain;
1423 /* Note this will print the MDS for a DS write */
1424 if (time_before(complain, jiffies)) {
1425 dprintk("NFS: faulty NFS server %s:"
1426 " (committed = %d) != (stable = %d)\n",
1427 NFS_SERVER(inode)->nfs_client->cl_hostname,
1428 resp->verf->committed, argp->stable);
1429 complain = jiffies + 300 * HZ;
1433 if (task->tk_status < 0) {
1434 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1438 /* Deal with the suid/sgid bit corner case */
1439 if (nfs_should_remove_suid(inode))
1440 nfs_mark_for_revalidate(inode);
1442 if (resp->count < argp->count) {
1443 static unsigned long complain;
1445 /* This a short write! */
1446 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1448 /* Has the server at least made some progress? */
1449 if (resp->count == 0) {
1450 if (time_before(complain, jiffies)) {
1452 "NFS: Server wrote zero bytes, expected %u.\n",
1454 complain = jiffies + 300 * HZ;
1456 nfs_set_pgio_error(data->header, -EIO, argp->offset);
1457 task->tk_status = -EIO;
1460 /* Was this an NFSv2 write or an NFSv3 stable write? */
1461 if (resp->verf->committed != NFS_UNSTABLE) {
1462 /* Resend from where the server left off */
1463 data->mds_offset += resp->count;
1464 argp->offset += resp->count;
1465 argp->pgbase += resp->count;
1466 argp->count -= resp->count;
1468 /* Resend as a stable write in order to avoid
1469 * headaches in the case of a server crash.
1471 argp->stable = NFS_FILE_SYNC;
1473 rpc_restart_call_prepare(task);
1478 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1479 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1483 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1487 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1489 nfs_wait_bit_killable,
1491 return (ret < 0) ? ret : 1;
1494 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1496 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1497 smp_mb__after_clear_bit();
1498 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1501 void nfs_commitdata_release(struct nfs_commit_data *data)
1503 put_nfs_open_context(data->context);
1504 nfs_commit_free(data);
1506 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1508 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1509 const struct rpc_call_ops *call_ops,
1512 struct rpc_task *task;
1513 int priority = flush_task_priority(how);
1514 struct rpc_message msg = {
1515 .rpc_argp = &data->args,
1516 .rpc_resp = &data->res,
1517 .rpc_cred = data->cred,
1519 struct rpc_task_setup task_setup_data = {
1520 .task = &data->task,
1522 .rpc_message = &msg,
1523 .callback_ops = call_ops,
1524 .callback_data = data,
1525 .workqueue = nfsiod_workqueue,
1526 .flags = RPC_TASK_ASYNC | flags,
1527 .priority = priority,
1529 /* Set up the initial task struct. */
1530 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1532 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1534 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1535 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1537 task = rpc_run_task(&task_setup_data);
1539 return PTR_ERR(task);
1540 if (how & FLUSH_SYNC)
1541 rpc_wait_for_completion_task(task);
1545 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1548 * Set up the argument/result storage required for the RPC call.
1550 void nfs_init_commit(struct nfs_commit_data *data,
1551 struct list_head *head,
1552 struct pnfs_layout_segment *lseg,
1553 struct nfs_commit_info *cinfo)
1555 struct nfs_page *first = nfs_list_entry(head->next);
1556 struct inode *inode = first->wb_context->dentry->d_inode;
1558 /* Set up the RPC argument and reply structs
1559 * NB: take care not to mess about with data->commit et al. */
1561 list_splice_init(head, &data->pages);
1563 data->inode = inode;
1564 data->cred = first->wb_context->cred;
1565 data->lseg = lseg; /* reference transferred */
1566 data->mds_ops = &nfs_commit_ops;
1567 data->completion_ops = cinfo->completion_ops;
1568 data->dreq = cinfo->dreq;
1570 data->args.fh = NFS_FH(data->inode);
1571 /* Note: we always request a commit of the entire inode */
1572 data->args.offset = 0;
1573 data->args.count = 0;
1574 data->context = get_nfs_open_context(first->wb_context);
1575 data->res.fattr = &data->fattr;
1576 data->res.verf = &data->verf;
1577 nfs_fattr_init(&data->fattr);
1579 EXPORT_SYMBOL_GPL(nfs_init_commit);
1581 void nfs_retry_commit(struct list_head *page_list,
1582 struct pnfs_layout_segment *lseg,
1583 struct nfs_commit_info *cinfo)
1585 struct nfs_page *req;
1587 while (!list_empty(page_list)) {
1588 req = nfs_list_entry(page_list->next);
1589 nfs_list_remove_request(req);
1590 nfs_mark_request_commit(req, lseg, cinfo);
1592 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1593 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1596 nfs_unlock_and_release_request(req);
1599 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1602 * Commit dirty pages
1605 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1606 struct nfs_commit_info *cinfo)
1608 struct nfs_commit_data *data;
1610 data = nfs_commitdata_alloc();
1615 /* Set up the argument struct */
1616 nfs_init_commit(data, head, NULL, cinfo);
1617 atomic_inc(&cinfo->mds->rpcs_out);
1618 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1621 nfs_retry_commit(head, NULL, cinfo);
1622 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1627 * COMMIT call returned
1629 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1631 struct nfs_commit_data *data = calldata;
1633 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1634 task->tk_pid, task->tk_status);
1636 /* Call the NFS version-specific code */
1637 NFS_PROTO(data->inode)->commit_done(task, data);
1640 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1642 struct nfs_page *req;
1643 int status = data->task.tk_status;
1644 struct nfs_commit_info cinfo;
1646 while (!list_empty(&data->pages)) {
1647 req = nfs_list_entry(data->pages.next);
1648 nfs_list_remove_request(req);
1649 nfs_clear_page_commit(req->wb_page);
1651 dprintk("NFS: commit (%s/%llu %d@%lld)",
1652 req->wb_context->dentry->d_sb->s_id,
1653 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1655 (long long)req_offset(req));
1657 nfs_context_set_write_error(req->wb_context, status);
1658 nfs_inode_remove_request(req);
1659 dprintk(", error = %d\n", status);
1663 /* Okay, COMMIT succeeded, apparently. Check the verifier
1664 * returned by the server against all stored verfs. */
1665 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1666 /* We have a match */
1667 nfs_inode_remove_request(req);
1671 /* We have a mismatch. Write the page again */
1672 dprintk(" mismatch\n");
1673 nfs_mark_request_dirty(req);
1674 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1676 nfs_unlock_and_release_request(req);
1678 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1679 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1680 nfs_commit_clear_lock(NFS_I(data->inode));
1683 static void nfs_commit_release(void *calldata)
1685 struct nfs_commit_data *data = calldata;
1687 data->completion_ops->completion(data);
1688 nfs_commitdata_release(calldata);
1691 static const struct rpc_call_ops nfs_commit_ops = {
1692 .rpc_call_prepare = nfs_commit_prepare,
1693 .rpc_call_done = nfs_commit_done,
1694 .rpc_release = nfs_commit_release,
1697 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1698 .completion = nfs_commit_release_pages,
1699 .error_cleanup = nfs_commit_clear_lock,
1702 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1703 int how, struct nfs_commit_info *cinfo)
1707 status = pnfs_commit_list(inode, head, how, cinfo);
1708 if (status == PNFS_NOT_ATTEMPTED)
1709 status = nfs_commit_list(inode, head, how, cinfo);
1713 int nfs_commit_inode(struct inode *inode, int how)
1716 struct nfs_commit_info cinfo;
1717 int may_wait = how & FLUSH_SYNC;
1720 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1722 goto out_mark_dirty;
1723 nfs_init_cinfo_from_inode(&cinfo, inode);
1724 res = nfs_scan_commit(inode, &head, &cinfo);
1728 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1732 goto out_mark_dirty;
1733 error = wait_on_bit(&NFS_I(inode)->flags,
1735 nfs_wait_bit_killable,
1740 nfs_commit_clear_lock(NFS_I(inode));
1742 /* Note: If we exit without ensuring that the commit is complete,
1743 * we must mark the inode as dirty. Otherwise, future calls to
1744 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1745 * that the data is on the disk.
1748 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1752 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1754 struct nfs_inode *nfsi = NFS_I(inode);
1755 int flags = FLUSH_SYNC;
1758 /* no commits means nothing needs to be done */
1759 if (!nfsi->commit_info.ncommit)
1762 if (wbc->sync_mode == WB_SYNC_NONE) {
1763 /* Don't commit yet if this is a non-blocking flush and there
1764 * are a lot of outstanding writes for this mapping.
1766 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1767 goto out_mark_dirty;
1769 /* don't wait for the COMMIT response */
1773 ret = nfs_commit_inode(inode, flags);
1775 if (wbc->sync_mode == WB_SYNC_NONE) {
1776 if (ret < wbc->nr_to_write)
1777 wbc->nr_to_write -= ret;
1779 wbc->nr_to_write = 0;
1784 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1788 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1794 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1796 return nfs_commit_unstable_pages(inode, wbc);
1798 EXPORT_SYMBOL_GPL(nfs_write_inode);
1801 * flush the inode to disk.
1803 int nfs_wb_all(struct inode *inode)
1805 struct writeback_control wbc = {
1806 .sync_mode = WB_SYNC_ALL,
1807 .nr_to_write = LONG_MAX,
1809 .range_end = LLONG_MAX,
1813 trace_nfs_writeback_inode_enter(inode);
1815 ret = sync_inode(inode, &wbc);
1817 trace_nfs_writeback_inode_exit(inode, ret);
1820 EXPORT_SYMBOL_GPL(nfs_wb_all);
1822 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1824 struct nfs_page *req;
1828 wait_on_page_writeback(page);
1829 req = nfs_page_find_request(page);
1832 if (nfs_lock_request(req)) {
1833 nfs_clear_request_commit(req);
1834 nfs_inode_remove_request(req);
1836 * In case nfs_inode_remove_request has marked the
1837 * page as being dirty
1839 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1840 nfs_unlock_and_release_request(req);
1843 ret = nfs_wait_on_request(req);
1844 nfs_release_request(req);
1852 * Write back all requests on one page - we do this before reading it.
1854 int nfs_wb_page(struct inode *inode, struct page *page)
1856 loff_t range_start = page_file_offset(page);
1857 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1858 struct writeback_control wbc = {
1859 .sync_mode = WB_SYNC_ALL,
1861 .range_start = range_start,
1862 .range_end = range_end,
1866 trace_nfs_writeback_page_enter(inode);
1869 wait_on_page_writeback(page);
1870 if (clear_page_dirty_for_io(page)) {
1871 ret = nfs_writepage_locked(page, &wbc);
1877 if (!PagePrivate(page))
1879 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1884 trace_nfs_writeback_page_exit(inode, ret);
1888 #ifdef CONFIG_MIGRATION
1889 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1890 struct page *page, enum migrate_mode mode)
1893 * If PagePrivate is set, then the page is currently associated with
1894 * an in-progress read or write request. Don't try to migrate it.
1896 * FIXME: we could do this in principle, but we'll need a way to ensure
1897 * that we can safely release the inode reference while holding
1900 if (PagePrivate(page))
1903 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1906 return migrate_page(mapping, newpage, page, mode);
1910 int __init nfs_init_writepagecache(void)
1912 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1913 sizeof(struct nfs_write_header),
1914 0, SLAB_HWCACHE_ALIGN,
1916 if (nfs_wdata_cachep == NULL)
1919 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1921 if (nfs_wdata_mempool == NULL)
1922 goto out_destroy_write_cache;
1924 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1925 sizeof(struct nfs_commit_data),
1926 0, SLAB_HWCACHE_ALIGN,
1928 if (nfs_cdata_cachep == NULL)
1929 goto out_destroy_write_mempool;
1931 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1933 if (nfs_commit_mempool == NULL)
1934 goto out_destroy_commit_cache;
1937 * NFS congestion size, scale with available memory.
1949 * This allows larger machines to have larger/more transfers.
1950 * Limit the default to 256M
1952 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1953 if (nfs_congestion_kb > 256*1024)
1954 nfs_congestion_kb = 256*1024;
1958 out_destroy_commit_cache:
1959 kmem_cache_destroy(nfs_cdata_cachep);
1960 out_destroy_write_mempool:
1961 mempool_destroy(nfs_wdata_mempool);
1962 out_destroy_write_cache:
1963 kmem_cache_destroy(nfs_wdata_cachep);
1967 void nfs_destroy_writepagecache(void)
1969 mempool_destroy(nfs_commit_mempool);
1970 kmem_cache_destroy(nfs_cdata_cachep);
1971 mempool_destroy(nfs_wdata_mempool);
1972 kmem_cache_destroy(nfs_wdata_cachep);