dcache: Translating dentry into pathname without taking rename_lock
[platform/adaptation/renesas_rcar/renesas_kernel.git] / fs / dcache.c
index 043c5b4..38b1b09 100644 (file)
@@ -88,6 +88,44 @@ EXPORT_SYMBOL(rename_lock);
 
 static struct kmem_cache *dentry_cache __read_mostly;
 
+/**
+ * read_seqbegin_or_lock - begin a sequence number check or locking block
+ * lock: sequence lock
+ * seq : sequence number to be checked
+ *
+ * First try it once optimistically without taking the lock. If that fails,
+ * take the lock. The sequence number is also used as a marker for deciding
+ * whether to be a reader (even) or writer (odd).
+ * N.B. seq must be initialized to an even number to begin with.
+ */
+static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
+{
+       if (!(*seq & 1)) {      /* Even */
+               *seq = read_seqbegin(lock);
+               rcu_read_lock();
+       } else                  /* Odd */
+               write_seqlock(lock);
+}
+
+/**
+ * read_seqretry_or_unlock - end a seqretry or lock block & return retry status
+ * lock         : sequence lock
+ * seq  : sequence number
+ * Return: 1 to retry operation again, 0 to continue
+ */
+static inline int read_seqretry_or_unlock(seqlock_t *lock, int *seq)
+{
+       if (!(*seq & 1)) {      /* Even */
+               rcu_read_unlock();
+               if (read_seqretry(lock, *seq)) {
+                       (*seq)++;       /* Take writer lock */
+                       return 1;
+               }
+       } else                  /* Odd */
+               write_sequnlock(lock);
+       return 0;
+}
+
 /*
  * This is the single most critical data structure when it comes
  * to the dcache: the hashtable for lookups. Somebody should try
@@ -1184,6 +1222,39 @@ int have_submounts(struct dentry *parent)
 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
@@ -1264,6 +1335,71 @@ void shrink_dcache_parent(struct dentry *parent)
 }
 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
@@ -2546,9 +2682,39 @@ static int prepend(char **buffer, int *buflen, const char *str, int namelen)
        return 0;
 }
 
+/**
+ * prepend_name - prepend a pathname in front of current buffer pointer
+ * buffer: buffer pointer
+ * buflen: allocated length of the buffer
+ * name:   name string and length qstr structure
+ *
+ * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to
+ * make sure that either the old or the new name pointer and length are
+ * fetched. However, there may be mismatch between length and pointer.
+ * The length cannot be trusted, we need to copy it byte-by-byte until
+ * the length is reached or a null byte is found. It also prepends "/" at
+ * the beginning of the name. The sequence number check at the caller will
+ * retry it again when a d_move() does happen. So any garbage in the buffer
+ * due to mismatched pointer and length will be discarded.
+ */
 static int prepend_name(char **buffer, int *buflen, struct qstr *name)
 {
-       return prepend(buffer, buflen, name->name, name->len);
+       const char *dname = ACCESS_ONCE(name->name);
+       u32 dlen = ACCESS_ONCE(name->len);
+       char *p;
+
+       if (*buflen < dlen + 1)
+               return -ENAMETOOLONG;
+       *buflen -= dlen + 1;
+       p = *buffer -= dlen + 1;
+       *p++ = '/';
+       while (dlen--) {
+               char c = *dname++;
+               if (!c)
+                       break;
+               *p++ = c;
+       }
+       return 0;
 }
 
 /**
@@ -2558,7 +2724,14 @@ static int prepend_name(char **buffer, int *buflen, struct qstr *name)
  * @buffer: pointer to the end of the buffer
  * @buflen: pointer to buffer length
  *
- * Caller holds the rename_lock.
+ * The function tries to write out the pathname without taking any lock other
+ * than the RCU read lock to make sure that dentries won't go away. It only
+ * checks the sequence number of the global rename_lock as any change in the
+ * dentry's d_seq will be preceded by changes in the rename_lock sequence
+ * number. If the sequence number had been change, it will restart the whole
+ * pathname back-tracing sequence again. It performs a total of 3 trials of
+ * lockless back-tracing sequences before falling back to take the
+ * rename_lock.
  */
 static int prepend_path(const struct path *path,
                        const struct path *root,
@@ -2567,54 +2740,60 @@ static int prepend_path(const struct path *path,
        struct dentry *dentry = path->dentry;
        struct vfsmount *vfsmnt = path->mnt;
        struct mount *mnt = real_mount(vfsmnt);
-       bool slash = false;
        int error = 0;
+       unsigned seq = 0;
+       char *bptr;
+       int blen;
 
+restart:
+       bptr = *buffer;
+       blen = *buflen;
+       read_seqbegin_or_lock(&rename_lock, &seq);
        while (dentry != root->dentry || vfsmnt != root->mnt) {
                struct dentry * parent;
 
                if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) {
                        /* Global root? */
-                       if (!mnt_has_parent(mnt))
-                               goto global_root;
-                       dentry = mnt->mnt_mountpoint;
-                       mnt = mnt->mnt_parent;
-                       vfsmnt = &mnt->mnt;
-                       continue;
+                       if (mnt_has_parent(mnt)) {
+                               dentry = mnt->mnt_mountpoint;
+                               mnt = mnt->mnt_parent;
+                               vfsmnt = &mnt->mnt;
+                               continue;
+                       }
+                       /*
+                        * Filesystems needing to implement special "root names"
+                        * should do so with ->d_dname()
+                        */
+                       if (IS_ROOT(dentry) &&
+                          (dentry->d_name.len != 1 ||
+                           dentry->d_name.name[0] != '/')) {
+                               WARN(1, "Root dentry has weird name <%.*s>\n",
+                                    (int) dentry->d_name.len,
+                                    dentry->d_name.name);
+                       }
+                       if (!error)
+                               error = is_mounted(vfsmnt) ? 1 : 2;
+                       break;
                }
                parent = dentry->d_parent;
                prefetch(parent);
-               spin_lock(&dentry->d_lock);
-               error = prepend_name(buffer, buflen, &dentry->d_name);
-               spin_unlock(&dentry->d_lock);
-               if (!error)
-                       error = prepend(buffer, buflen, "/", 1);
+               error = prepend_name(&bptr, &blen, &dentry->d_name);
                if (error)
                        break;
 
-               slash = true;
                dentry = parent;
        }
+       if (read_seqretry_or_unlock(&rename_lock, &seq))
+               goto restart;
 
-       if (!error && !slash)
-               error = prepend(buffer, buflen, "/", 1);
-
-       return error;
-
-global_root:
-       /*
-        * Filesystems needing to implement special "root names"
-        * should do so with ->d_dname()
-        */
-       if (IS_ROOT(dentry) &&
-           (dentry->d_name.len != 1 || dentry->d_name.name[0] != '/')) {
-               WARN(1, "Root dentry has weird name <%.*s>\n",
-                    (int) dentry->d_name.len, dentry->d_name.name);
-       }
-       if (!slash)
-               error = prepend(buffer, buflen, "/", 1);
-       if (!error)
-               error = is_mounted(vfsmnt) ? 1 : 2;
+       if (error >= 0 && bptr == *buffer) {
+               if (--blen < 0)
+                       error = -ENAMETOOLONG;
+               else
+                       *--bptr = '/';
+       }
+       *buffer = bptr;
+       *buflen = blen;
        return error;
 }
 
@@ -2643,9 +2822,7 @@ char *__d_path(const struct path *path,
 
        prepend(&res, &buflen, "\0", 1);
        br_read_lock(&vfsmount_lock);
-       write_seqlock(&rename_lock);
        error = prepend_path(path, root, &res, &buflen);
-       write_sequnlock(&rename_lock);
        br_read_unlock(&vfsmount_lock);
 
        if (error < 0)
@@ -2664,9 +2841,7 @@ char *d_absolute_path(const struct path *path,
 
        prepend(&res, &buflen, "\0", 1);
        br_read_lock(&vfsmount_lock);
-       write_seqlock(&rename_lock);
        error = prepend_path(path, &root, &res, &buflen);
-       write_sequnlock(&rename_lock);
        br_read_unlock(&vfsmount_lock);
 
        if (error > 1)
@@ -2732,9 +2907,7 @@ char *d_path(const struct path *path, char *buf, int buflen)
 
        get_fs_root(current->fs, &root);
        br_read_lock(&vfsmount_lock);
-       write_seqlock(&rename_lock);
        error = path_with_deleted(path, &root, &res, &buflen);
-       write_sequnlock(&rename_lock);
        br_read_unlock(&vfsmount_lock);
        if (error < 0)
                res = ERR_PTR(error);
@@ -2769,10 +2942,10 @@ char *simple_dname(struct dentry *dentry, char *buffer, int buflen)
        char *end = buffer + buflen;
        /* these dentries are never renamed, so d_lock is not needed */
        if (prepend(&end, &buflen, " (deleted)", 11) ||
-           prepend_name(&end, &buflen, &dentry->d_name) ||
+           prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) ||
            prepend(&end, &buflen, "/", 1))  
                end = ERR_PTR(-ENAMETOOLONG);
-       return end;  
+       return end;
 }
 
 /*
@@ -2780,30 +2953,36 @@ char *simple_dname(struct dentry *dentry, char *buffer, int buflen)
  */
 static char *__dentry_path(struct dentry *dentry, char *buf, int buflen)
 {
-       char *end = buf + buflen;
-       char *retval;
+       char *end, *retval;
+       int len, seq = 0;
+       int error = 0;
 
-       prepend(&end, &buflen, "\0", 1);
+restart:
+       end = buf + buflen;
+       len = buflen;
+       prepend(&end, &len, "\0", 1);
        if (buflen < 1)
                goto Elong;
        /* Get '/' right */
        retval = end-1;
        *retval = '/';
-
+       read_seqbegin_or_lock(&rename_lock, &seq);
        while (!IS_ROOT(dentry)) {
                struct dentry *parent = dentry->d_parent;
                int error;
 
                prefetch(parent);
-               spin_lock(&dentry->d_lock);
-               error = prepend_name(&end, &buflen, &dentry->d_name);
-               spin_unlock(&dentry->d_lock);
-               if (error != 0 || prepend(&end, &buflen, "/", 1) != 0)
-                       goto Elong;
+               error = prepend_name(&end, &len, &dentry->d_name);
+               if (error)
+                       break;
 
                retval = end;
                dentry = parent;
        }
+       if (read_seqretry_or_unlock(&rename_lock, &seq))
+               goto restart;
+       if (error)
+               goto Elong;
        return retval;
 Elong:
        return ERR_PTR(-ENAMETOOLONG);
@@ -2811,13 +2990,7 @@ Elong:
 
 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
 {
-       char *retval;
-
-       write_seqlock(&rename_lock);
-       retval = __dentry_path(dentry, buf, buflen);
-       write_sequnlock(&rename_lock);
-
-       return retval;
+       return __dentry_path(dentry, buf, buflen);
 }
 EXPORT_SYMBOL(dentry_path_raw);
 
@@ -2826,7 +2999,6 @@ char *dentry_path(struct dentry *dentry, char *buf, int buflen)
        char *p = NULL;
        char *retval;
 
-       write_seqlock(&rename_lock);
        if (d_unlinked(dentry)) {
                p = buf + buflen;
                if (prepend(&p, &buflen, "//deleted", 10) != 0)
@@ -2834,7 +3006,6 @@ char *dentry_path(struct dentry *dentry, char *buf, int buflen)
                buflen++;
        }
        retval = __dentry_path(dentry, buf, buflen);
-       write_sequnlock(&rename_lock);
        if (!IS_ERR(retval) && p)
                *p = '/';       /* restore '/' overriden with '\0' */
        return retval;
@@ -2873,7 +3044,6 @@ SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
 
        error = -ENOENT;
        br_read_lock(&vfsmount_lock);
-       write_seqlock(&rename_lock);
        if (!d_unlinked(pwd.dentry)) {
                unsigned long len;
                char *cwd = page + PAGE_SIZE;
@@ -2881,7 +3051,6 @@ SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
 
                prepend(&cwd, &buflen, "\0", 1);
                error = prepend_path(&pwd, &root, &cwd, &buflen);
-               write_sequnlock(&rename_lock);
                br_read_unlock(&vfsmount_lock);
 
                if (error < 0)
@@ -2902,7 +3071,6 @@ SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
                                error = -EFAULT;
                }
        } else {
-               write_sequnlock(&rename_lock);
                br_read_unlock(&vfsmount_lock);
        }