NFS: Fix the 'desynchronized value of nfs_i.ncommit' error
Trond Myklebust [Fri, 20 Apr 2007 20:12:45 +0000 (16:12 -0400)]
Redirtying a request that is already marked for commit will screw up the
accounting for NR_UNSTABLE_NFS as well as nfs_i.ncommit.
Ensure that all requests on the commit queue are labelled with the
PG_NEED_COMMIT flag, and avoid moving them onto the dirty list inside
nfs_page_mark_flush().

Also inline nfs_mark_request_dirty() into nfs_page_mark_flush() for
atomicity reasons. Avoid dropping the spinlock until we're done marking the
request in the radix tree and have added it to the ->dirty list.

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs/nfs/write.c

index 8e94246..ce5b4a9 100644 (file)
@@ -38,7 +38,6 @@
 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
                                            struct page *,
                                            unsigned int, unsigned int);
-static void nfs_mark_request_dirty(struct nfs_page *req);
 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
 static const struct rpc_call_ops nfs_write_partial_ops;
 static const struct rpc_call_ops nfs_write_full_ops;
@@ -255,7 +254,8 @@ static void nfs_end_page_writeback(struct page *page)
 static int nfs_page_mark_flush(struct page *page)
 {
        struct nfs_page *req;
-       spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
+       struct nfs_inode *nfsi = NFS_I(page->mapping->host);
+       spinlock_t *req_lock = &nfsi->req_lock;
        int ret;
 
        spin_lock(req_lock);
@@ -279,11 +279,23 @@ static int nfs_page_mark_flush(struct page *page)
                        return ret;
                spin_lock(req_lock);
        }
-       spin_unlock(req_lock);
+       if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
+               /* This request is marked for commit */
+               spin_unlock(req_lock);
+               nfs_unlock_request(req);
+               return 1;
+       }
        if (nfs_set_page_writeback(page) == 0) {
                nfs_list_remove_request(req);
-               nfs_mark_request_dirty(req);
-       }
+               /* add the request to the inode's dirty list. */
+               radix_tree_tag_set(&nfsi->nfs_page_tree,
+                               req->wb_index, NFS_PAGE_TAG_DIRTY);
+               nfs_list_add_request(req, &nfsi->dirty);
+               nfsi->ndirty++;
+               spin_unlock(req_lock);
+               __mark_inode_dirty(page->mapping->host, I_DIRTY_PAGES);
+       } else
+               spin_unlock(req_lock);
        ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
        nfs_unlock_request(req);
        return ret;
@@ -406,24 +418,6 @@ static void nfs_inode_remove_request(struct nfs_page *req)
        nfs_release_request(req);
 }
 
-/*
- * Add a request to the inode's dirty list.
- */
-static void
-nfs_mark_request_dirty(struct nfs_page *req)
-{
-       struct inode *inode = req->wb_context->dentry->d_inode;
-       struct nfs_inode *nfsi = NFS_I(inode);
-
-       spin_lock(&nfsi->req_lock);
-       radix_tree_tag_set(&nfsi->nfs_page_tree,
-                       req->wb_index, NFS_PAGE_TAG_DIRTY);
-       nfs_list_add_request(req, &nfsi->dirty);
-       nfsi->ndirty++;
-       spin_unlock(&nfsi->req_lock);
-       __mark_inode_dirty(inode, I_DIRTY_PAGES);
-}
-
 static void
 nfs_redirty_request(struct nfs_page *req)
 {
@@ -438,7 +432,7 @@ nfs_dirty_request(struct nfs_page *req)
 {
        struct page *page = req->wb_page;
 
-       if (page == NULL)
+       if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
                return 0;
        return !PageWriteback(req->wb_page);
 }
@@ -456,6 +450,7 @@ nfs_mark_request_commit(struct nfs_page *req)
        spin_lock(&nfsi->req_lock);
        nfs_list_add_request(req, &nfsi->commit);
        nfsi->ncommit++;
+       set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
        spin_unlock(&nfsi->req_lock);
        inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
        __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
@@ -470,7 +465,7 @@ int nfs_write_need_commit(struct nfs_write_data *data)
 static inline
 int nfs_reschedule_unstable_write(struct nfs_page *req)
 {
-       if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
+       if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
                nfs_mark_request_commit(req);
                return 1;
        }
@@ -557,6 +552,7 @@ static void nfs_cancel_commit_list(struct list_head *head)
                req = nfs_list_entry(head->next);
                dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
                nfs_list_remove_request(req);
+               clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
                nfs_inode_remove_request(req);
                nfs_unlock_request(req);
        }
@@ -1295,6 +1291,7 @@ static void nfs_commit_done(struct rpc_task *task, void *calldata)
        while (!list_empty(&data->pages)) {
                req = nfs_list_entry(data->pages.next);
                nfs_list_remove_request(req);
+               clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
                dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 
                dprintk("NFS: commit (%s/%Ld %d@%Ld)",