vfs: check unlinked ancestors before mount
[platform/adaptation/renesas_rcar/renesas_kernel.git] / fs / dcache.c
index 96655f4..761e31b 100644 (file)
@@ -472,7 +472,7 @@ relock:
         * inform the fs via d_prune that this dentry is about to be
         * unhashed and destroyed.
         */
-       if (dentry->d_flags & DCACHE_OP_PRUNE)
+       if ((dentry->d_flags & DCACHE_OP_PRUNE) && !d_unhashed(dentry))
                dentry->d_op->d_prune(dentry);
 
        dentry_lru_del(dentry);
@@ -727,6 +727,14 @@ restart:
        hlist_for_each_entry(dentry, &inode->i_dentry, d_alias) {
                spin_lock(&dentry->d_lock);
                if (!dentry->d_lockref.count) {
+                       /*
+                        * inform the fs via d_prune that this dentry
+                        * is about to be unhashed and destroyed.
+                        */
+                       if ((dentry->d_flags & DCACHE_OP_PRUNE) &&
+                           !d_unhashed(dentry))
+                               dentry->d_op->d_prune(dentry);
+
                        __dget_dlock(dentry);
                        __d_drop(dentry);
                        spin_unlock(&dentry->d_lock);
@@ -911,7 +919,8 @@ static void shrink_dcache_for_umount_subtree(struct dentry *dentry)
                         * inform the fs that this dentry is about to be
                         * unhashed and destroyed.
                         */
-                       if (dentry->d_flags & DCACHE_OP_PRUNE)
+                       if ((dentry->d_flags & DCACHE_OP_PRUNE) &&
+                           !d_unhashed(dentry))
                                dentry->d_op->d_prune(dentry);
 
                        dentry_lru_del(dentry);
@@ -1022,34 +1031,56 @@ static struct dentry *try_to_ascend(struct dentry *old, int locked, unsigned seq
        return new;
 }
 
+/**
+ * enum d_walk_ret - action to talke during tree walk
+ * @D_WALK_CONTINUE:   contrinue walk
+ * @D_WALK_QUIT:       quit walk
+ * @D_WALK_NORETRY:    quit when retry is needed
+ * @D_WALK_SKIP:       skip this dentry and its children
+ */
+enum d_walk_ret {
+       D_WALK_CONTINUE,
+       D_WALK_QUIT,
+       D_WALK_NORETRY,
+       D_WALK_SKIP,
+};
 
-/*
- * Search for at least 1 mount point in the dentry's subdirs.
- * We descend to the next level whenever the d_subdirs
- * list is non-empty and continue searching.
- */
 /**
- * have_submounts - check for mounts over a dentry
- * @parent: dentry to check.
+ * d_walk - walk the dentry tree
+ * @parent:    start of walk
+ * @data:      data passed to @enter() and @finish()
+ * @enter:     callback when first entering the dentry
+ * @finish:    callback when successfully finished the walk
  *
- * Return true if the parent or its subdirectories contain
- * a mount point
+ * The @enter() and @finish() callbacks are called with d_lock held.
  */
-int have_submounts(struct dentry *parent)
+static void d_walk(struct dentry *parent, void *data,
+                  enum d_walk_ret (*enter)(void *, struct dentry *),
+                  void (*finish)(void *))
 {
        struct dentry *this_parent;
        struct list_head *next;
        unsigned seq;
        int locked = 0;
+       enum d_walk_ret ret;
+       bool retry = true;
 
        seq = read_seqbegin(&rename_lock);
 again:
        this_parent = parent;
-
-       if (d_mountpoint(parent))
-               goto positive;
        spin_lock(&this_parent->d_lock);
+
+       ret = enter(data, this_parent);
+       switch (ret) {
+       case D_WALK_CONTINUE:
+               break;
+       case D_WALK_QUIT:
+       case D_WALK_SKIP:
+               goto out_unlock;
+       case D_WALK_NORETRY:
+               retry = false;
+               break;
+       }
 repeat:
        next = this_parent->d_subdirs.next;
 resume:
@@ -1059,12 +1090,22 @@ resume:
                next = tmp->next;
 
                spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
-               /* Have we found a mount point ? */
-               if (d_mountpoint(dentry)) {
+
+               ret = enter(data, dentry);
+               switch (ret) {
+               case D_WALK_CONTINUE:
+                       break;
+               case D_WALK_QUIT:
                        spin_unlock(&dentry->d_lock);
-                       spin_unlock(&this_parent->d_lock);
-                       goto positive;
+                       goto out_unlock;
+               case D_WALK_NORETRY:
+                       retry = false;
+                       break;
+               case D_WALK_SKIP:
+                       spin_unlock(&dentry->d_lock);
+                       continue;
                }
+
                if (!list_empty(&dentry->d_subdirs)) {
                        spin_unlock(&this_parent->d_lock);
                        spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
@@ -1085,29 +1126,97 @@ resume:
                next = child->d_u.d_child.next;
                goto resume;
        }
-       spin_unlock(&this_parent->d_lock);
-       if (!locked && read_seqretry(&rename_lock, seq))
-               goto rename_retry;
-       if (locked)
-               write_sequnlock(&rename_lock);
-       return 0; /* No mount points found in tree */
-positive:
-       if (!locked && read_seqretry(&rename_lock, seq))
+       if (!locked && read_seqretry(&rename_lock, seq)) {
+               spin_unlock(&this_parent->d_lock);
                goto rename_retry;
+       }
+       if (finish)
+               finish(data);
+
+out_unlock:
+       spin_unlock(&this_parent->d_lock);
        if (locked)
                write_sequnlock(&rename_lock);
-       return 1;
+       return;
 
 rename_retry:
+       if (!retry)
+               return;
        if (locked)
                goto again;
        locked = 1;
        write_seqlock(&rename_lock);
        goto again;
 }
+
+/*
+ * Search for at least 1 mount point in the dentry's subdirs.
+ * We descend to the next level whenever the d_subdirs
+ * list is non-empty and continue searching.
+ */
+
+/**
+ * have_submounts - check for mounts over a dentry
+ * @parent: dentry to check.
+ *
+ * Return true if the parent or its subdirectories contain
+ * a mount point
+ */
+
+static enum d_walk_ret check_mount(void *data, struct dentry *dentry)
+{
+       int *ret = data;
+       if (d_mountpoint(dentry)) {
+               *ret = 1;
+               return D_WALK_QUIT;
+       }
+       return D_WALK_CONTINUE;
+}
+
+int have_submounts(struct dentry *parent)
+{
+       int ret = 0;
+
+       d_walk(parent, &ret, check_mount, NULL);
+
+       return ret;
+}
 EXPORT_SYMBOL(have_submounts);
 
 /*
+ * Called by mount code to set a mountpoint and check if the mountpoint is
+ * reachable (e.g. NFS can unhash a directory dentry and then the complete
+ * subtree can become unreachable).
+ *
+ * Only one of check_submounts_and_drop() and d_set_mounted() must succeed.  For
+ * this reason take rename_lock and d_lock on dentry and ancestors.
+ */
+int d_set_mounted(struct dentry *dentry)
+{
+       struct dentry *p;
+       int ret = -ENOENT;
+       write_seqlock(&rename_lock);
+       for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
+               /* Need exclusion wrt. check_submounts_and_drop() */
+               spin_lock(&p->d_lock);
+               if (unlikely(d_unhashed(p))) {
+                       spin_unlock(&p->d_lock);
+                       goto out;
+               }
+               spin_unlock(&p->d_lock);
+       }
+       spin_lock(&dentry->d_lock);
+       if (!d_unlinked(dentry)) {
+               dentry->d_flags |= DCACHE_MOUNTED;
+               ret = 0;
+       }
+       spin_unlock(&dentry->d_lock);
+out:
+       write_sequnlock(&rename_lock);
+       return ret;
+}
+
+/*
  * Search the dentry child list of the specified parent,
  * and move any unused dentries to the end of the unused
  * list for prune_dcache(). We descend to the next level
@@ -1121,93 +1230,46 @@ EXPORT_SYMBOL(have_submounts);
  * drop the lock and return early due to latency
  * constraints.
  */
-static int select_parent(struct dentry *parent, struct list_head *dispose)
-{
-       struct dentry *this_parent;
-       struct list_head *next;
-       unsigned seq;
-       int found = 0;
-       int locked = 0;
-
-       seq = read_seqbegin(&rename_lock);
-again:
-       this_parent = parent;
-       spin_lock(&this_parent->d_lock);
-repeat:
-       next = this_parent->d_subdirs.next;
-resume:
-       while (next != &this_parent->d_subdirs) {
-               struct list_head *tmp = next;
-               struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child);
-               next = tmp->next;
 
-               spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
+struct select_data {
+       struct dentry *start;
+       struct list_head dispose;
+       int found;
+};
 
-               /*
-                * move only zero ref count dentries to the dispose list.
-                *
-                * Those which are presently on the shrink list, being processed
-                * by shrink_dentry_list(), shouldn't be moved.  Otherwise the
-                * loop in shrink_dcache_parent() might not make any progress
-                * and loop forever.
-                */
-               if (dentry->d_lockref.count) {
-                       dentry_lru_del(dentry);
-               } else if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
-                       dentry_lru_move_list(dentry, dispose);
-                       dentry->d_flags |= DCACHE_SHRINK_LIST;
-                       found++;
-               }
-               /*
-                * We can return to the caller if we have found some (this
-                * ensures forward progress). We'll be coming back to find
-                * the rest.
-                */
-               if (found && need_resched()) {
-                       spin_unlock(&dentry->d_lock);
-                       goto out;
-               }
+static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
+{
+       struct select_data *data = _data;
+       enum d_walk_ret ret = D_WALK_CONTINUE;
 
-               /*
-                * Descend a level if the d_subdirs list is non-empty.
-                */
-               if (!list_empty(&dentry->d_subdirs)) {
-                       spin_unlock(&this_parent->d_lock);
-                       spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
-                       this_parent = dentry;
-                       spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
-                       goto repeat;
-               }
+       if (data->start == dentry)
+               goto out;
 
-               spin_unlock(&dentry->d_lock);
-       }
        /*
-        * All done at this level ... ascend and resume the search.
+        * move only zero ref count dentries to the dispose list.
+        *
+        * Those which are presently on the shrink list, being processed
+        * by shrink_dentry_list(), shouldn't be moved.  Otherwise the
+        * loop in shrink_dcache_parent() might not make any progress
+        * and loop forever.
         */
-       if (this_parent != parent) {
-               struct dentry *child = this_parent;
-               this_parent = try_to_ascend(this_parent, locked, seq);
-               if (!this_parent)
-                       goto rename_retry;
-               next = child->d_u.d_child.next;
-               goto resume;
+       if (dentry->d_lockref.count) {
+               dentry_lru_del(dentry);
+       } else if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
+               dentry_lru_move_list(dentry, &data->dispose);
+               dentry->d_flags |= DCACHE_SHRINK_LIST;
+               data->found++;
+               ret = D_WALK_NORETRY;
        }
+       /*
+        * We can return to the caller if we have found some (this
+        * ensures forward progress). We'll be coming back to find
+        * the rest.
+        */
+       if (data->found && need_resched())
+               ret = D_WALK_QUIT;
 out:
-       spin_unlock(&this_parent->d_lock);
-       if (!locked && read_seqretry(&rename_lock, seq))
-               goto rename_retry;
-       if (locked)
-               write_sequnlock(&rename_lock);
-       return found;
-
-rename_retry:
-       if (found)
-               return found;
-       if (locked)
-               goto again;
-       locked = 1;
-       write_seqlock(&rename_lock);
-       goto again;
+       return ret;
 }
 
 /**
@@ -1216,18 +1278,90 @@ rename_retry:
  *
  * Prune the dcache to remove unused children of the parent dentry.
  */
-void shrink_dcache_parent(struct dentry * parent)
+void shrink_dcache_parent(struct dentry *parent)
 {
-       LIST_HEAD(dispose);
-       int found;
+       for (;;) {
+               struct select_data data;
 
-       while ((found = select_parent(parent, &dispose)) != 0) {
-               shrink_dentry_list(&dispose);
+               INIT_LIST_HEAD(&data.dispose);
+               data.start = parent;
+               data.found = 0;
+
+               d_walk(parent, &data, select_collect, NULL);
+               if (!data.found)
+                       break;
+
+               shrink_dentry_list(&data.dispose);
                cond_resched();
        }
 }
 EXPORT_SYMBOL(shrink_dcache_parent);
 
+static enum d_walk_ret check_and_collect(void *_data, struct dentry *dentry)
+{
+       struct select_data *data = _data;
+
+       if (d_mountpoint(dentry)) {
+               data->found = -EBUSY;
+               return D_WALK_QUIT;
+       }
+
+       return select_collect(_data, dentry);
+}
+
+static void check_and_drop(void *_data)
+{
+       struct select_data *data = _data;
+
+       if (d_mountpoint(data->start))
+               data->found = -EBUSY;
+       if (!data->found)
+               __d_drop(data->start);
+}
+
+/**
+ * check_submounts_and_drop - prune dcache, check for submounts and drop
+ *
+ * All done as a single atomic operation relative to has_unlinked_ancestor().
+ * Returns 0 if successfully unhashed @parent.  If there were submounts then
+ * return -EBUSY.
+ *
+ * @dentry: dentry to prune and drop
+ */
+int check_submounts_and_drop(struct dentry *dentry)
+{
+       int ret = 0;
+
+       /* Negative dentries can be dropped without further checks */
+       if (!dentry->d_inode) {
+               d_drop(dentry);
+               goto out;
+       }
+
+       for (;;) {
+               struct select_data data;
+
+               INIT_LIST_HEAD(&data.dispose);
+               data.start = dentry;
+               data.found = 0;
+
+               d_walk(dentry, &data, check_and_collect, check_and_drop);
+               ret = data.found;
+
+               if (!list_empty(&data.dispose))
+                       shrink_dentry_list(&data.dispose);
+
+               if (ret <= 0)
+                       break;
+
+               cond_resched();
+       }
+
+out:
+       return ret;
+}
+EXPORT_SYMBOL(check_submounts_and_drop);
+
 /**
  * __d_alloc   -       allocate a dcache entry
  * @sb: filesystem it will belong to
@@ -2919,68 +3053,24 @@ int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
        return result;
 }
 
-void d_genocide(struct dentry *root)
+static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
 {
-       struct dentry *this_parent;
-       struct list_head *next;
-       unsigned seq;
-       int locked = 0;
+       struct dentry *root = data;
+       if (dentry != root) {
+               if (d_unhashed(dentry) || !dentry->d_inode)
+                       return D_WALK_SKIP;
 
-       seq = read_seqbegin(&rename_lock);
-again:
-       this_parent = root;
-       spin_lock(&this_parent->d_lock);
-repeat:
-       next = this_parent->d_subdirs.next;
-resume:
-       while (next != &this_parent->d_subdirs) {
-               struct list_head *tmp = next;
-               struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child);
-               next = tmp->next;
-
-               spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
-               if (d_unhashed(dentry) || !dentry->d_inode) {
-                       spin_unlock(&dentry->d_lock);
-                       continue;
-               }
-               if (!list_empty(&dentry->d_subdirs)) {
-                       spin_unlock(&this_parent->d_lock);
-                       spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
-                       this_parent = dentry;
-                       spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
-                       goto repeat;
-               }
                if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
                        dentry->d_flags |= DCACHE_GENOCIDE;
                        dentry->d_lockref.count--;
                }
-               spin_unlock(&dentry->d_lock);
        }
-       if (this_parent != root) {
-               struct dentry *child = this_parent;
-               if (!(this_parent->d_flags & DCACHE_GENOCIDE)) {
-                       this_parent->d_flags |= DCACHE_GENOCIDE;
-                       this_parent->d_lockref.count--;
-               }
-               this_parent = try_to_ascend(this_parent, locked, seq);
-               if (!this_parent)
-                       goto rename_retry;
-               next = child->d_u.d_child.next;
-               goto resume;
-       }
-       spin_unlock(&this_parent->d_lock);
-       if (!locked && read_seqretry(&rename_lock, seq))
-               goto rename_retry;
-       if (locked)
-               write_sequnlock(&rename_lock);
-       return;
+       return D_WALK_CONTINUE;
+}
 
-rename_retry:
-       if (locked)
-               goto again;
-       locked = 1;
-       write_seqlock(&rename_lock);
-       goto again;
+void d_genocide(struct dentry *parent)
+{
+       d_walk(parent, parent, d_genocide_kill, NULL);
 }
 
 void d_tmpfile(struct dentry *dentry, struct inode *inode)