]> bbs.cooldavid.org Git - net-next-2.6.git/blobdiff - fs/notify/fsnotify.c
fsnotify: fix list walk order
[net-next-2.6.git] / fs / notify / fsnotify.c
index 54d58d5f72c1c4fe22c3fceeb443f74f13b89cc6..2169aa593d5fa24695fe1e342cb731690814f172 100644 (file)
@@ -140,64 +140,74 @@ void __fsnotify_parent(struct path *path, struct dentry *dentry, __u32 mask)
 }
 EXPORT_SYMBOL_GPL(__fsnotify_parent);
 
-void __fsnotify_flush_ignored_mask(struct inode *inode, void *data, int data_is)
+static int send_to_group(struct inode *to_tell, struct vfsmount *mnt,
+                        struct fsnotify_mark *inode_mark,
+                        struct fsnotify_mark *vfsmount_mark,
+                        __u32 mask, void *data,
+                        int data_is, u32 cookie,
+                        const unsigned char *file_name,
+                        struct fsnotify_event **event)
 {
-       struct fsnotify_mark *mark;
-       struct hlist_node *node;
-
-       if (!hlist_empty(&inode->i_fsnotify_marks)) {
-               spin_lock(&inode->i_lock);
-               hlist_for_each_entry(mark, node, &inode->i_fsnotify_marks, i.i_list) {
-                       if (!(mark->flags & FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY))
-                               mark->ignored_mask = 0;
-               }
-               spin_unlock(&inode->i_lock);
+       struct fsnotify_group *group = NULL;
+       __u32 inode_test_mask = 0;
+       __u32 vfsmount_test_mask = 0;
+
+       if (unlikely(!inode_mark && !vfsmount_mark)) {
+               BUG();
+               return 0;
        }
 
-       if (data_is == FSNOTIFY_EVENT_PATH) {
-               struct vfsmount *mnt;
+       /* clear ignored on inode modification */
+       if (mask & FS_MODIFY) {
+               if (inode_mark &&
+                   !(inode_mark->flags & FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY))
+                       inode_mark->ignored_mask = 0;
+               if (vfsmount_mark &&
+                   !(vfsmount_mark->flags & FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY))
+                       vfsmount_mark->ignored_mask = 0;
+       }
 
-               mnt = ((struct path *)data)->mnt;
-               if (mnt && !hlist_empty(&mnt->mnt_fsnotify_marks)) {
-                       spin_lock(&mnt->mnt_root->d_lock);
-                       hlist_for_each_entry(mark, node, &mnt->mnt_fsnotify_marks, m.m_list) {
-                               if (!(mark->flags & FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY))
-                                       mark->ignored_mask = 0;
-                       }
-                       spin_unlock(&mnt->mnt_root->d_lock);
-               }
+       /* does the inode mark tell us to do something? */
+       if (inode_mark) {
+               group = inode_mark->group;
+               inode_test_mask = (mask & ~FS_EVENT_ON_CHILD);
+               inode_test_mask &= inode_mark->mask;
+               inode_test_mask &= ~inode_mark->ignored_mask;
        }
-}
 
-static void send_to_group(struct fsnotify_group *group, struct inode *to_tell,
-                         struct vfsmount *mnt, __u32 mask, void *data,
-                         int data_is, u32 cookie, const char *file_name,
-                         struct fsnotify_event **event)
-{
-       if (!group->ops->should_send_event(group, to_tell, mnt, mask,
-                                          data, data_is))
-               return;
+       /* does the vfsmount_mark tell us to do something? */
+       if (vfsmount_mark) {
+               vfsmount_test_mask = (mask & ~FS_EVENT_ON_CHILD);
+               group = vfsmount_mark->group;
+               vfsmount_test_mask &= vfsmount_mark->mask;
+               vfsmount_test_mask &= ~vfsmount_mark->ignored_mask;
+               if (inode_mark)
+                       vfsmount_test_mask &= ~inode_mark->ignored_mask;
+       }
+
+       pr_debug("%s: group=%p to_tell=%p mnt=%p mask=%x inode_mark=%p"
+                " inode_test_mask=%x vfsmount_mark=%p vfsmount_test_mask=%x"
+                " data=%p data_is=%d cookie=%d event=%p\n",
+                __func__, group, to_tell, mnt, mask, inode_mark,
+                inode_test_mask, vfsmount_mark, vfsmount_test_mask, data,
+                data_is, cookie, *event);
+
+       if (!inode_test_mask && !vfsmount_test_mask)
+               return 0;
+
+       if (group->ops->should_send_event(group, to_tell, inode_mark,
+                                         vfsmount_mark, mask, data,
+                                         data_is) == false)
+               return 0;
+
        if (!*event) {
                *event = fsnotify_create_event(to_tell, mask, data,
                                                data_is, file_name,
                                                cookie, GFP_KERNEL);
-               /*
-                * shit, we OOM'd and now we can't tell, maybe
-                * someday someone else will want to do something
-                * here
-                */
                if (!*event)
-                       return;
+                       return -ENOMEM;
        }
-       group->ops->handle_event(group, *event);
-}
-
-static bool needed_by_vfsmount(__u32 test_mask, struct vfsmount *mnt)
-{
-       if (!mnt)
-               return false;
-
-       return (test_mask & mnt->mnt_fsnotify_mask);
+       return group->ops->handle_event(group, inode_mark, vfsmount_mark, *event);
 }
 
 /*
@@ -206,73 +216,112 @@ static bool needed_by_vfsmount(__u32 test_mask, struct vfsmount *mnt)
  * out to all of the registered fsnotify_group.  Those groups can then use the
  * notification event in whatever means they feel necessary.
  */
-void fsnotify(struct inode *to_tell, __u32 mask, void *data, int data_is, const char *file_name, u32 cookie)
+int fsnotify(struct inode *to_tell, __u32 mask, void *data, int data_is,
+            const unsigned char *file_name, u32 cookie)
 {
-       struct fsnotify_group *group;
+       struct hlist_node *inode_node = NULL, *vfsmount_node = NULL;
+       struct fsnotify_mark *inode_mark = NULL, *vfsmount_mark = NULL;
+       struct fsnotify_group *inode_group, *vfsmount_group;
        struct fsnotify_event *event = NULL;
-       struct vfsmount *mnt = NULL;
-       int idx;
+       struct vfsmount *mnt;
+       int idx, ret = 0;
+       bool used_inode, used_vfsmount;
        /* global tests shouldn't care about events on child only the specific event */
        __u32 test_mask = (mask & ~FS_EVENT_ON_CHILD);
 
-       /* if no fsnotify listeners, nothing to do */
-       if (list_empty(&fsnotify_inode_groups) &&
-           list_empty(&fsnotify_vfsmount_groups))
-                return;
-       if (mask & FS_MODIFY)
-               __fsnotify_flush_ignored_mask(to_tell, data, data_is);
-
-       /* if none of the directed listeners or vfsmount listeners care */
-       if (!(test_mask & fsnotify_inode_mask) &&
-           !(test_mask & fsnotify_vfsmount_mask))
-                return;
        if (data_is == FSNOTIFY_EVENT_PATH)
                mnt = ((struct path *)data)->mnt;
-
-       /* if this inode's directed listeners don't care and nothing on the vfsmount
-        * listeners list cares, nothing to do */
-       if (!(test_mask & to_tell->i_fsnotify_mask) &&
-           !needed_by_vfsmount(test_mask, mnt))
-                return;
+       else
+               mnt = NULL;
 
        /*
-        * SRCU!!  the groups list is very very much read only and the path is
-        * very hot.  The VAST majority of events are not going to need to do
-        * anything other than walk the list so it's crazy to pre-allocate.
+        * if this is a modify event we may need to clear the ignored masks
+        * otherwise return if neither the inode nor the vfsmount care about
+        * this type of event.
         */
-       idx = srcu_read_lock(&fsnotify_grp_srcu);
-
-       if (test_mask & to_tell->i_fsnotify_mask) {
-               list_for_each_entry_rcu(group, &fsnotify_inode_groups, inode_group_list) {
-                       if (test_mask & group->mask) {
-                               send_to_group(group, to_tell, NULL, mask, data, data_is,
-                                             cookie, file_name, &event);
-                       }
-               }
+       if (!(mask & FS_MODIFY) &&
+           !(test_mask & to_tell->i_fsnotify_mask) &&
+           !(mnt && test_mask & mnt->mnt_fsnotify_mask))
+               return 0;
+
+       idx = srcu_read_lock(&fsnotify_mark_srcu);
+
+       if ((mask & FS_MODIFY) ||
+           (test_mask & to_tell->i_fsnotify_mask))
+               inode_node = srcu_dereference(to_tell->i_fsnotify_marks.first,
+                                             &fsnotify_mark_srcu);
+
+       if (mnt && ((mask & FS_MODIFY) ||
+                   (test_mask & mnt->mnt_fsnotify_mask))) {
+               vfsmount_node = srcu_dereference(mnt->mnt_fsnotify_marks.first,
+                                                &fsnotify_mark_srcu);
+               inode_node = srcu_dereference(to_tell->i_fsnotify_marks.first,
+                                             &fsnotify_mark_srcu);
        }
-       if (needed_by_vfsmount(test_mask, mnt)) {
-               list_for_each_entry_rcu(group, &fsnotify_vfsmount_groups, vfsmount_group_list) {
-                       if (test_mask & group->mask) {
-                               send_to_group(group, to_tell, mnt, mask, data, data_is,
-                                             cookie, file_name, &event);
-                       }
+
+       while (inode_node || vfsmount_node) {
+               used_inode = used_vfsmount = false;
+               inode_group = vfsmount_group = NULL;
+
+               if (inode_node) {
+                       inode_mark = hlist_entry(srcu_dereference(inode_node, &fsnotify_mark_srcu),
+                                                struct fsnotify_mark, i.i_list);
+                       inode_group = inode_mark->group;
+               }
+
+               if (vfsmount_node) {
+                       vfsmount_mark = hlist_entry(srcu_dereference(vfsmount_node, &fsnotify_mark_srcu),
+                                                       struct fsnotify_mark, m.m_list);
+                       vfsmount_group = vfsmount_mark->group;
+               }
+
+               if (inode_group > vfsmount_group) {
+                       /* handle inode */
+                       send_to_group(to_tell, NULL, inode_mark, NULL, mask, data,
+                                     data_is, cookie, file_name, &event);
+                       used_inode = true;
+               } else if (vfsmount_group > inode_group) {
+                       send_to_group(to_tell, mnt, NULL, vfsmount_mark, mask, data,
+                                     data_is, cookie, file_name, &event);
+                       used_vfsmount = true;
+               } else {
+                       send_to_group(to_tell, mnt, inode_mark, vfsmount_mark,
+                                     mask, data, data_is, cookie, file_name,
+                                     &event);
+                       used_vfsmount = true;
+                       used_inode = true;
                }
+
+               if (used_inode)
+                       inode_node = srcu_dereference(inode_node->next,
+                                                     &fsnotify_mark_srcu);
+               if (used_vfsmount)
+                       vfsmount_node = srcu_dereference(vfsmount_node->next,
+                                                        &fsnotify_mark_srcu);
        }
 
-       srcu_read_unlock(&fsnotify_grp_srcu, idx);
+       srcu_read_unlock(&fsnotify_mark_srcu, idx);
        /*
         * fsnotify_create_event() took a reference so the event can't be cleaned
         * up while we are still trying to add it to lists, drop that one.
         */
        if (event)
                fsnotify_put_event(event);
+
+       return ret;
 }
 EXPORT_SYMBOL_GPL(fsnotify);
 
 static __init int fsnotify_init(void)
 {
-       return init_srcu_struct(&fsnotify_grp_srcu);
+       int ret;
+
+       BUG_ON(hweight32(ALL_FSNOTIFY_EVENTS) != 23);
+
+       ret = init_srcu_struct(&fsnotify_mark_srcu);
+       if (ret)
+               panic("initializing fsnotify_mark_srcu");
+
+       return 0;
 }
-subsys_initcall(fsnotify_init);
+core_initcall(fsnotify_init);