Merge branch 'work.dcache' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
authorLinus Torvalds <torvalds@linux-foundation.org>
Wed, 4 Apr 2018 19:05:25 +0000 (12:05 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 4 Apr 2018 19:05:25 +0000 (12:05 -0700)
Pull vfs dcache updates from Al Viro:
 "Part of this is what the trylock loop elimination series has turned
  into, part making d_move() preserve the parent (and thus the path) of
  victim, plus some general cleanups"

* 'work.dcache' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (22 commits)
  d_genocide: move export to definition
  fold dentry_lock_for_move() into its sole caller and clean it up
  make non-exchanging __d_move() copy ->d_parent rather than swap them
  oprofilefs: don't oops on allocation failure
  lustre: get rid of pointless casts to struct dentry *
  debugfs_lookup(): switch to lookup_one_len_unlocked()
  fold lookup_real() into __lookup_hash()
  take out orphan externs (empty_string/slash_string)
  split d_path() and friends into a separate file
  dcache.c: trim includes
  fs/dcache: Avoid a try_lock loop in shrink_dentry_list()
  get rid of trylock loop around dentry_kill()
  handle move to LRU in retain_dentry()
  dput(): consolidate the "do we need to retain it?" into an inlined helper
  split the slow part of lock_parent() off
  now lock_parent() can't run into killed dentry
  get rid of trylock loop in locking dentries on shrink list
  d_delete(): get rid of trylock loop
  fs/dcache: Move dentry_kill() below lock_parent()
  fs/dcache: Remove stale comment from dentry_kill()
  ...

drivers/oprofile/oprofilefs.c
drivers/staging/lustre/lustre/llite/dcache.c
fs/Makefile
fs/d_path.c [new file with mode: 0644]
fs/dcache.c
fs/debugfs/inode.c
fs/namei.c
include/linux/dcache.h

index d77ebbfc67c98f8b63cdde527b8a59d7135486bd..4ea08979312ca3e86804d5c7fa3e49be5969606a 100644 (file)
@@ -138,6 +138,9 @@ static int __oprofilefs_create_file(struct dentry *root, char const *name,
        struct dentry *dentry;
        struct inode *inode;
 
+       if (!root)
+               return -ENOMEM;
+
        inode_lock(d_inode(root));
        dentry = d_alloc_name(root, name);
        if (!dentry) {
index 549369739d805736e95c8115ca7cbf48dd098e05..6cd0318062e8dfb5d359136ec1ea1c855c245b1c 100644 (file)
@@ -90,7 +90,7 @@ static int ll_dcompare(const struct dentry *dentry,
               d_count(dentry));
 
        /* mountpoint is always valid */
-       if (d_mountpoint((struct dentry *)dentry))
+       if (d_mountpoint(dentry))
                return 0;
 
        if (d_lustre_invalid(dentry))
@@ -111,7 +111,7 @@ static int ll_ddelete(const struct dentry *de)
        LASSERT(de);
 
        CDEBUG(D_DENTRY, "%s dentry %pd (%p, parent %p, inode %p) %s%s\n",
-              d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
+              d_lustre_invalid(de) ? "deleting" : "keeping",
               de, de, de->d_parent, d_inode(de),
               d_unhashed(de) ? "" : "hashed,",
               list_empty(&de->d_subdirs) ? "" : "subdirs");
@@ -119,7 +119,7 @@ static int ll_ddelete(const struct dentry *de)
        /* kernel >= 2.6.38 last refcount is decreased after this function. */
        LASSERT(d_count(de) == 1);
 
-       if (d_lustre_invalid((struct dentry *)de))
+       if (d_lustre_invalid(de))
                return 1;
        return 0;
 }
index add789ea270ad703e9502efad6e3f4070bbae47b..c9375fd2c8c4d337162c77899f9b183bf3ff106e 100644 (file)
@@ -11,7 +11,7 @@ obj-y :=      open.o read_write.o file_table.o super.o \
                ioctl.o readdir.o select.o dcache.o inode.o \
                attr.o bad_inode.o file.o filesystems.o namespace.o \
                seq_file.o xattr.o libfs.o fs-writeback.o \
-               pnode.o splice.o sync.o utimes.o \
+               pnode.o splice.o sync.o utimes.o d_path.o \
                stack.o fs_struct.o statfs.o fs_pin.o nsfs.o
 
 ifeq ($(CONFIG_BLOCK),y)
diff --git a/fs/d_path.c b/fs/d_path.c
new file mode 100644 (file)
index 0000000..e8fce6b
--- /dev/null
@@ -0,0 +1,470 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#include <linux/syscalls.h>
+#include <linux/export.h>
+#include <linux/uaccess.h>
+#include <linux/fs_struct.h>
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include <linux/prefetch.h>
+#include "mount.h"
+
+static int prepend(char **buffer, int *buflen, const char *str, int namelen)
+{
+       *buflen -= namelen;
+       if (*buflen < 0)
+               return -ENAMETOOLONG;
+       *buffer -= namelen;
+       memcpy(*buffer, str, 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 READ_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.
+ *
+ * Load acquire is needed to make sure that we see that terminating NUL.
+ */
+static int prepend_name(char **buffer, int *buflen, const struct qstr *name)
+{
+       const char *dname = smp_load_acquire(&name->name); /* ^^^ */
+       u32 dlen = READ_ONCE(name->len);
+       char *p;
+
+       *buflen -= dlen + 1;
+       if (*buflen < 0)
+               return -ENAMETOOLONG;
+       p = *buffer -= dlen + 1;
+       *p++ = '/';
+       while (dlen--) {
+               char c = *dname++;
+               if (!c)
+                       break;
+               *p++ = c;
+       }
+       return 0;
+}
+
+/**
+ * prepend_path - Prepend path string to a buffer
+ * @path: the dentry/vfsmount to report
+ * @root: root vfsmnt/dentry
+ * @buffer: pointer to the end of the buffer
+ * @buflen: pointer to buffer length
+ *
+ * The function will first try 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 changed, it will restart
+ * the whole pathname back-tracing sequence again by taking the rename_lock.
+ * In this case, there is no need to take the RCU read lock as the recursive
+ * parent pointer references will keep the dentry chain alive as long as no
+ * rename operation is performed.
+ */
+static int prepend_path(const struct path *path,
+                       const struct path *root,
+                       char **buffer, int *buflen)
+{
+       struct dentry *dentry;
+       struct vfsmount *vfsmnt;
+       struct mount *mnt;
+       int error = 0;
+       unsigned seq, m_seq = 0;
+       char *bptr;
+       int blen;
+
+       rcu_read_lock();
+restart_mnt:
+       read_seqbegin_or_lock(&mount_lock, &m_seq);
+       seq = 0;
+       rcu_read_lock();
+restart:
+       bptr = *buffer;
+       blen = *buflen;
+       error = 0;
+       dentry = path->dentry;
+       vfsmnt = path->mnt;
+       mnt = real_mount(vfsmnt);
+       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)) {
+                       struct mount *parent = READ_ONCE(mnt->mnt_parent);
+                       /* Escaped? */
+                       if (dentry != vfsmnt->mnt_root) {
+                               bptr = *buffer;
+                               blen = *buflen;
+                               error = 3;
+                               break;
+                       }
+                       /* Global root? */
+                       if (mnt != parent) {
+                               dentry = READ_ONCE(mnt->mnt_mountpoint);
+                               mnt = parent;
+                               vfsmnt = &mnt->mnt;
+                               continue;
+                       }
+                       if (!error)
+                               error = is_mounted(vfsmnt) ? 1 : 2;
+                       break;
+               }
+               parent = dentry->d_parent;
+               prefetch(parent);
+               error = prepend_name(&bptr, &blen, &dentry->d_name);
+               if (error)
+                       break;
+
+               dentry = parent;
+       }
+       if (!(seq & 1))
+               rcu_read_unlock();
+       if (need_seqretry(&rename_lock, seq)) {
+               seq = 1;
+               goto restart;
+       }
+       done_seqretry(&rename_lock, seq);
+
+       if (!(m_seq & 1))
+               rcu_read_unlock();
+       if (need_seqretry(&mount_lock, m_seq)) {
+               m_seq = 1;
+               goto restart_mnt;
+       }
+       done_seqretry(&mount_lock, m_seq);
+
+       if (error >= 0 && bptr == *buffer) {
+               if (--blen < 0)
+                       error = -ENAMETOOLONG;
+               else
+                       *--bptr = '/';
+       }
+       *buffer = bptr;
+       *buflen = blen;
+       return error;
+}
+
+/**
+ * __d_path - return the path of a dentry
+ * @path: the dentry/vfsmount to report
+ * @root: root vfsmnt/dentry
+ * @buf: buffer to return value in
+ * @buflen: buffer length
+ *
+ * Convert a dentry into an ASCII path name.
+ *
+ * Returns a pointer into the buffer or an error code if the
+ * path was too long.
+ *
+ * "buflen" should be positive.
+ *
+ * If the path is not reachable from the supplied root, return %NULL.
+ */
+char *__d_path(const struct path *path,
+              const struct path *root,
+              char *buf, int buflen)
+{
+       char *res = buf + buflen;
+       int error;
+
+       prepend(&res, &buflen, "\0", 1);
+       error = prepend_path(path, root, &res, &buflen);
+
+       if (error < 0)
+               return ERR_PTR(error);
+       if (error > 0)
+               return NULL;
+       return res;
+}
+
+char *d_absolute_path(const struct path *path,
+              char *buf, int buflen)
+{
+       struct path root = {};
+       char *res = buf + buflen;
+       int error;
+
+       prepend(&res, &buflen, "\0", 1);
+       error = prepend_path(path, &root, &res, &buflen);
+
+       if (error > 1)
+               error = -EINVAL;
+       if (error < 0)
+               return ERR_PTR(error);
+       return res;
+}
+
+/*
+ * same as __d_path but appends "(deleted)" for unlinked files.
+ */
+static int path_with_deleted(const struct path *path,
+                            const struct path *root,
+                            char **buf, int *buflen)
+{
+       prepend(buf, buflen, "\0", 1);
+       if (d_unlinked(path->dentry)) {
+               int error = prepend(buf, buflen, " (deleted)", 10);
+               if (error)
+                       return error;
+       }
+
+       return prepend_path(path, root, buf, buflen);
+}
+
+static int prepend_unreachable(char **buffer, int *buflen)
+{
+       return prepend(buffer, buflen, "(unreachable)", 13);
+}
+
+static void get_fs_root_rcu(struct fs_struct *fs, struct path *root)
+{
+       unsigned seq;
+
+       do {
+               seq = read_seqcount_begin(&fs->seq);
+               *root = fs->root;
+       } while (read_seqcount_retry(&fs->seq, seq));
+}
+
+/**
+ * d_path - return the path of a dentry
+ * @path: path to report
+ * @buf: buffer to return value in
+ * @buflen: buffer length
+ *
+ * Convert a dentry into an ASCII path name. If the entry has been deleted
+ * the string " (deleted)" is appended. Note that this is ambiguous.
+ *
+ * Returns a pointer into the buffer or an error code if the path was
+ * too long. Note: Callers should use the returned pointer, not the passed
+ * in buffer, to use the name! The implementation often starts at an offset
+ * into the buffer, and may leave 0 bytes at the start.
+ *
+ * "buflen" should be positive.
+ */
+char *d_path(const struct path *path, char *buf, int buflen)
+{
+       char *res = buf + buflen;
+       struct path root;
+       int error;
+
+       /*
+        * We have various synthetic filesystems that never get mounted.  On
+        * these filesystems dentries are never used for lookup purposes, and
+        * thus don't need to be hashed.  They also don't need a name until a
+        * user wants to identify the object in /proc/pid/fd/.  The little hack
+        * below allows us to generate a name for these objects on demand:
+        *
+        * Some pseudo inodes are mountable.  When they are mounted
+        * path->dentry == path->mnt->mnt_root.  In that case don't call d_dname
+        * and instead have d_path return the mounted path.
+        */
+       if (path->dentry->d_op && path->dentry->d_op->d_dname &&
+           (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root))
+               return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
+
+       rcu_read_lock();
+       get_fs_root_rcu(current->fs, &root);
+       error = path_with_deleted(path, &root, &res, &buflen);
+       rcu_read_unlock();
+
+       if (error < 0)
+               res = ERR_PTR(error);
+       return res;
+}
+EXPORT_SYMBOL(d_path);
+
+/*
+ * Helper function for dentry_operations.d_dname() members
+ */
+char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
+                       const char *fmt, ...)
+{
+       va_list args;
+       char temp[64];
+       int sz;
+
+       va_start(args, fmt);
+       sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
+       va_end(args);
+
+       if (sz > sizeof(temp) || sz > buflen)
+               return ERR_PTR(-ENAMETOOLONG);
+
+       buffer += buflen - sz;
+       return memcpy(buffer, temp, sz);
+}
+
+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(&end, &buflen, dentry->d_name.name, dentry->d_name.len) ||
+           prepend(&end, &buflen, "/", 1))  
+               end = ERR_PTR(-ENAMETOOLONG);
+       return end;
+}
+EXPORT_SYMBOL(simple_dname);
+
+/*
+ * Write full pathname from the root of the filesystem into the buffer.
+ */
+static char *__dentry_path(struct dentry *d, char *buf, int buflen)
+{
+       struct dentry *dentry;
+       char *end, *retval;
+       int len, seq = 0;
+       int error = 0;
+
+       if (buflen < 2)
+               goto Elong;
+
+       rcu_read_lock();
+restart:
+       dentry = d;
+       end = buf + buflen;
+       len = buflen;
+       prepend(&end, &len, "\0", 1);
+       /* Get '/' right */
+       retval = end-1;
+       *retval = '/';
+       read_seqbegin_or_lock(&rename_lock, &seq);
+       while (!IS_ROOT(dentry)) {
+               struct dentry *parent = dentry->d_parent;
+
+               prefetch(parent);
+               error = prepend_name(&end, &len, &dentry->d_name);
+               if (error)
+                       break;
+
+               retval = end;
+               dentry = parent;
+       }
+       if (!(seq & 1))
+               rcu_read_unlock();
+       if (need_seqretry(&rename_lock, seq)) {
+               seq = 1;
+               goto restart;
+       }
+       done_seqretry(&rename_lock, seq);
+       if (error)
+               goto Elong;
+       return retval;
+Elong:
+       return ERR_PTR(-ENAMETOOLONG);
+}
+
+char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
+{
+       return __dentry_path(dentry, buf, buflen);
+}
+EXPORT_SYMBOL(dentry_path_raw);
+
+char *dentry_path(struct dentry *dentry, char *buf, int buflen)
+{
+       char *p = NULL;
+       char *retval;
+
+       if (d_unlinked(dentry)) {
+               p = buf + buflen;
+               if (prepend(&p, &buflen, "//deleted", 10) != 0)
+                       goto Elong;
+               buflen++;
+       }
+       retval = __dentry_path(dentry, buf, buflen);
+       if (!IS_ERR(retval) && p)
+               *p = '/';       /* restore '/' overriden with '\0' */
+       return retval;
+Elong:
+       return ERR_PTR(-ENAMETOOLONG);
+}
+
+static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root,
+                                   struct path *pwd)
+{
+       unsigned seq;
+
+       do {
+               seq = read_seqcount_begin(&fs->seq);
+               *root = fs->root;
+               *pwd = fs->pwd;
+       } while (read_seqcount_retry(&fs->seq, seq));
+}
+
+/*
+ * NOTE! The user-level library version returns a
+ * character pointer. The kernel system call just
+ * returns the length of the buffer filled (which
+ * includes the ending '\0' character), or a negative
+ * error value. So libc would do something like
+ *
+ *     char *getcwd(char * buf, size_t size)
+ *     {
+ *             int retval;
+ *
+ *             retval = sys_getcwd(buf, size);
+ *             if (retval >= 0)
+ *                     return buf;
+ *             errno = -retval;
+ *             return NULL;
+ *     }
+ */
+SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
+{
+       int error;
+       struct path pwd, root;
+       char *page = __getname();
+
+       if (!page)
+               return -ENOMEM;
+
+       rcu_read_lock();
+       get_fs_root_and_pwd_rcu(current->fs, &root, &pwd);
+
+       error = -ENOENT;
+       if (!d_unlinked(pwd.dentry)) {
+               unsigned long len;
+               char *cwd = page + PATH_MAX;
+               int buflen = PATH_MAX;
+
+               prepend(&cwd, &buflen, "\0", 1);
+               error = prepend_path(&pwd, &root, &cwd, &buflen);
+               rcu_read_unlock();
+
+               if (error < 0)
+                       goto out;
+
+               /* Unreachable from current root */
+               if (error > 0) {
+                       error = prepend_unreachable(&cwd, &buflen);
+                       if (error)
+                               goto out;
+               }
+
+               error = -ERANGE;
+               len = PATH_MAX + page - cwd;
+               if (len <= size) {
+                       error = len;
+                       if (copy_to_user(buf, cwd, len))
+                               error = -EFAULT;
+               }
+       } else {
+               rcu_read_unlock();
+       }
+
+out:
+       __putname(page);
+       return error;
+}
index 8945e6cabd93f7ce42709f93b90db94dceb74273..593079176123823e0ddcf19ce72257e19fdc923e 100644 (file)
@@ -14,7 +14,7 @@
  * the dcache entry is deleted or garbage collected.
  */
 
-#include <linux/syscalls.h>
+#include <linux/ratelimit.h>
 #include <linux/string.h>
 #include <linux/mm.h>
 #include <linux/fs.h>
 #include <linux/hash.h>
 #include <linux/cache.h>
 #include <linux/export.h>
-#include <linux/mount.h>
-#include <linux/file.h>
-#include <linux/uaccess.h>
 #include <linux/security.h>
 #include <linux/seqlock.h>
-#include <linux/swap.h>
 #include <linux/bootmem.h>
-#include <linux/fs_struct.h>
 #include <linux/bit_spinlock.h>
 #include <linux/rculist_bl.h>
-#include <linux/prefetch.h>
-#include <linux/ratelimit.h>
 #include <linux/list_lru.h>
 #include "internal.h"
 #include "mount.h"
@@ -74,9 +67,7 @@
  *       dentry->d_lock
  *
  * If no ancestor relationship:
- * if (dentry1 < dentry2)
- *   dentry1->d_lock
- *     dentry2->d_lock
+ * arbitrary, since it's serialized on rename_lock
  */
 int sysctl_vfs_cache_pressure __read_mostly = 100;
 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
@@ -440,17 +431,6 @@ static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
        list_lru_isolate_move(lru, &dentry->d_lru, list);
 }
 
-/*
- * dentry_lru_(add|del)_list) must be called with d_lock held.
- */
-static void dentry_lru_add(struct dentry *dentry)
-{
-       if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
-               d_lru_add(dentry);
-       else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
-               dentry->d_flags |= DCACHE_REFERENCED;
-}
-
 /**
  * d_drop - drop a dentry
  * @dentry: dentry to drop
@@ -470,30 +450,29 @@ static void dentry_lru_add(struct dentry *dentry)
  */
 static void ___d_drop(struct dentry *dentry)
 {
-       if (!d_unhashed(dentry)) {
-               struct hlist_bl_head *b;
-               /*
-                * Hashed dentries are normally on the dentry hashtable,
-                * with the exception of those newly allocated by
-                * d_obtain_root, which are always IS_ROOT:
-                */
-               if (unlikely(IS_ROOT(dentry)))
-                       b = &dentry->d_sb->s_roots;
-               else
-                       b = d_hash(dentry->d_name.hash);
+       struct hlist_bl_head *b;
+       /*
+        * Hashed dentries are normally on the dentry hashtable,
+        * with the exception of those newly allocated by
+        * d_obtain_root, which are always IS_ROOT:
+        */
+       if (unlikely(IS_ROOT(dentry)))
+               b = &dentry->d_sb->s_roots;
+       else
+               b = d_hash(dentry->d_name.hash);
 
-               hlist_bl_lock(b);
-               __hlist_bl_del(&dentry->d_hash);
-               hlist_bl_unlock(b);
-               /* After this call, in-progress rcu-walk path lookup will fail. */
-               write_seqcount_invalidate(&dentry->d_seq);
-       }
+       hlist_bl_lock(b);
+       __hlist_bl_del(&dentry->d_hash);
+       hlist_bl_unlock(b);
 }
 
 void __d_drop(struct dentry *dentry)
 {
-       ___d_drop(dentry);
-       dentry->d_hash.pprev = NULL;
+       if (!d_unhashed(dentry)) {
+               ___d_drop(dentry);
+               dentry->d_hash.pprev = NULL;
+               write_seqcount_invalidate(&dentry->d_seq);
+       }
 }
 EXPORT_SYMBOL(__d_drop);
 
@@ -589,10 +568,71 @@ static void __dentry_kill(struct dentry *dentry)
                dentry_free(dentry);
 }
 
+static struct dentry *__lock_parent(struct dentry *dentry)
+{
+       struct dentry *parent;
+       rcu_read_lock();
+       spin_unlock(&dentry->d_lock);
+again:
+       parent = READ_ONCE(dentry->d_parent);
+       spin_lock(&parent->d_lock);
+       /*
+        * We can't blindly lock dentry until we are sure
+        * that we won't violate the locking order.
+        * Any changes of dentry->d_parent must have
+        * been done with parent->d_lock held, so
+        * spin_lock() above is enough of a barrier
+        * for checking if it's still our child.
+        */
+       if (unlikely(parent != dentry->d_parent)) {
+               spin_unlock(&parent->d_lock);
+               goto again;
+       }
+       rcu_read_unlock();
+       if (parent != dentry)
+               spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
+       else
+               parent = NULL;
+       return parent;
+}
+
+static inline struct dentry *lock_parent(struct dentry *dentry)
+{
+       struct dentry *parent = dentry->d_parent;
+       if (IS_ROOT(dentry))
+               return NULL;
+       if (likely(spin_trylock(&parent->d_lock)))
+               return parent;
+       return __lock_parent(dentry);
+}
+
+static inline bool retain_dentry(struct dentry *dentry)
+{
+       WARN_ON(d_in_lookup(dentry));
+
+       /* Unreachable? Get rid of it */
+       if (unlikely(d_unhashed(dentry)))
+               return false;
+
+       if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
+               return false;
+
+       if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
+               if (dentry->d_op->d_delete(dentry))
+                       return false;
+       }
+       /* retain; LRU fodder */
+       dentry->d_lockref.count--;
+       if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
+               d_lru_add(dentry);
+       else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
+               dentry->d_flags |= DCACHE_REFERENCED;
+       return true;
+}
+
 /*
  * Finish off a dentry we've decided to kill.
  * dentry->d_lock must be held, returns with it unlocked.
- * If ref is non-zero, then decrement the refcount too.
  * Returns dentry requiring refcount drop, or NULL if we're done.
  */
 static struct dentry *dentry_kill(struct dentry *dentry)
@@ -602,62 +642,43 @@ static struct dentry *dentry_kill(struct dentry *dentry)
        struct dentry *parent = NULL;
 
        if (inode && unlikely(!spin_trylock(&inode->i_lock)))
-               goto failed;
+               goto slow_positive;
 
        if (!IS_ROOT(dentry)) {
                parent = dentry->d_parent;
                if (unlikely(!spin_trylock(&parent->d_lock))) {
-                       if (inode)
-                               spin_unlock(&inode->i_lock);
-                       goto failed;
+                       parent = __lock_parent(dentry);
+                       if (likely(inode || !dentry->d_inode))
+                               goto got_locks;
+                       /* negative that became positive */
+                       if (parent)
+                               spin_unlock(&parent->d_lock);
+                       inode = dentry->d_inode;
+                       goto slow_positive;
                }
        }
-
        __dentry_kill(dentry);
        return parent;
 
-failed:
+slow_positive:
        spin_unlock(&dentry->d_lock);
-       return dentry; /* try again with same dentry */
-}
-
-static inline struct dentry *lock_parent(struct dentry *dentry)
-{
-       struct dentry *parent = dentry->d_parent;
-       if (IS_ROOT(dentry))
-               return NULL;
-       if (unlikely(dentry->d_lockref.count < 0))
-               return NULL;
-       if (likely(spin_trylock(&parent->d_lock)))
+       spin_lock(&inode->i_lock);
+       spin_lock(&dentry->d_lock);
+       parent = lock_parent(dentry);
+got_locks:
+       if (unlikely(dentry->d_lockref.count != 1)) {
+               dentry->d_lockref.count--;
+       } else if (likely(!retain_dentry(dentry))) {
+               __dentry_kill(dentry);
                return parent;
-       rcu_read_lock();
-       spin_unlock(&dentry->d_lock);
-again:
-       parent = READ_ONCE(dentry->d_parent);
-       spin_lock(&parent->d_lock);
-       /*
-        * We can't blindly lock dentry until we are sure
-        * that we won't violate the locking order.
-        * Any changes of dentry->d_parent must have
-        * been done with parent->d_lock held, so
-        * spin_lock() above is enough of a barrier
-        * for checking if it's still our child.
-        */
-       if (unlikely(parent != dentry->d_parent)) {
-               spin_unlock(&parent->d_lock);
-               goto again;
        }
-       if (parent != dentry) {
-               spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
-               if (unlikely(dentry->d_lockref.count < 0)) {
-                       spin_unlock(&parent->d_lock);
-                       parent = NULL;
-               }
-       } else {
-               parent = NULL;
-       }
-       rcu_read_unlock();
-       return parent;
+       /* we are keeping it, after all */
+       if (inode)
+               spin_unlock(&inode->i_lock);
+       if (parent)
+               spin_unlock(&parent->d_lock);
+       spin_unlock(&dentry->d_lock);
+       return NULL;
 }
 
 /*
@@ -807,27 +828,11 @@ repeat:
        /* Slow case: now with the dentry lock held */
        rcu_read_unlock();
 
-       WARN_ON(d_in_lookup(dentry));
-
-       /* Unreachable? Get rid of it */
-       if (unlikely(d_unhashed(dentry)))
-               goto kill_it;
-
-       if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
-               goto kill_it;
-
-       if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
-               if (dentry->d_op->d_delete(dentry))
-                       goto kill_it;
+       if (likely(retain_dentry(dentry))) {
+               spin_unlock(&dentry->d_lock);
+               return;
        }
 
-       dentry_lru_add(dentry);
-
-       dentry->d_lockref.count--;
-       spin_unlock(&dentry->d_lock);
-       return;
-
-kill_it:
        dentry = dentry_kill(dentry);
        if (dentry) {
                cond_resched();
@@ -976,56 +981,83 @@ restart:
 }
 EXPORT_SYMBOL(d_prune_aliases);
 
-static void shrink_dentry_list(struct list_head *list)
+/*
+ * Lock a dentry from shrink list.
+ * Called under rcu_read_lock() and dentry->d_lock; the former
+ * guarantees that nothing we access will be freed under us.
+ * Note that dentry is *not* protected from concurrent dentry_kill(),
+ * d_delete(), etc.
+ *
+ * Return false if dentry has been disrupted or grabbed, leaving
+ * the caller to kick it off-list.  Otherwise, return true and have
+ * that dentry's inode and parent both locked.
+ */
+static bool shrink_lock_dentry(struct dentry *dentry)
 {
-       struct dentry *dentry, *parent;
+       struct inode *inode;
+       struct dentry *parent;
 
-       while (!list_empty(list)) {
-               struct inode *inode;
-               dentry = list_entry(list->prev, struct dentry, d_lru);
+       if (dentry->d_lockref.count)
+               return false;
+
+       inode = dentry->d_inode;
+       if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
+               spin_unlock(&dentry->d_lock);
+               spin_lock(&inode->i_lock);
                spin_lock(&dentry->d_lock);
-               parent = lock_parent(dentry);
+               if (unlikely(dentry->d_lockref.count))
+                       goto out;
+               /* changed inode means that somebody had grabbed it */
+               if (unlikely(inode != dentry->d_inode))
+                       goto out;
+       }
 
-               /*
-                * The dispose list is isolated and dentries are not accounted
-                * to the LRU here, so we can simply remove it from the list
-                * here regardless of whether it is referenced or not.
-                */
-               d_shrink_del(dentry);
+       parent = dentry->d_parent;
+       if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
+               return true;
 
-               /*
-                * We found an inuse dentry which was not removed from
-                * the LRU because of laziness during lookup. Do not free it.
-                */
-               if (dentry->d_lockref.count > 0) {
-                       spin_unlock(&dentry->d_lock);
-                       if (parent)
-                               spin_unlock(&parent->d_lock);
-                       continue;
-               }
+       spin_unlock(&dentry->d_lock);
+       spin_lock(&parent->d_lock);
+       if (unlikely(parent != dentry->d_parent)) {
+               spin_unlock(&parent->d_lock);
+               spin_lock(&dentry->d_lock);
+               goto out;
+       }
+       spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
+       if (likely(!dentry->d_lockref.count))
+               return true;
+       spin_unlock(&parent->d_lock);
+out:
+       if (inode)
+               spin_unlock(&inode->i_lock);
+       return false;
+}
 
+static void shrink_dentry_list(struct list_head *list)
+{
+       while (!list_empty(list)) {
+               struct dentry *dentry, *parent;
 
-               if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) {
-                       bool can_free = dentry->d_flags & DCACHE_MAY_FREE;
+               dentry = list_entry(list->prev, struct dentry, d_lru);
+               spin_lock(&dentry->d_lock);
+               rcu_read_lock();
+               if (!shrink_lock_dentry(dentry)) {
+                       bool can_free = false;
+                       rcu_read_unlock();
+                       d_shrink_del(dentry);
+                       if (dentry->d_lockref.count < 0)
+                               can_free = dentry->d_flags & DCACHE_MAY_FREE;
                        spin_unlock(&dentry->d_lock);
-                       if (parent)
-                               spin_unlock(&parent->d_lock);
                        if (can_free)
                                dentry_free(dentry);
                        continue;
                }
-
-               inode = dentry->d_inode;
-               if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
-                       d_shrink_add(dentry, list);
-                       spin_unlock(&dentry->d_lock);
-                       if (parent)
-                               spin_unlock(&parent->d_lock);
-                       continue;
-               }
-
+               rcu_read_unlock();
+               d_shrink_del(dentry);
+               parent = dentry->d_parent;
                __dentry_kill(dentry);
-
+               if (parent == dentry)
+                       continue;
                /*
                 * We need to prune ancestors too. This is necessary to prevent
                 * quadratic behavior of shrink_dcache_parent(), but is also
@@ -1033,26 +1065,8 @@ static void shrink_dentry_list(struct list_head *list)
                 * fragmentation.
                 */
                dentry = parent;
-               while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) {
-                       parent = lock_parent(dentry);
-                       if (dentry->d_lockref.count != 1) {
-                               dentry->d_lockref.count--;
-                               spin_unlock(&dentry->d_lock);
-                               if (parent)
-                                       spin_unlock(&parent->d_lock);
-                               break;
-                       }
-                       inode = dentry->d_inode;        /* can't be NULL */
-                       if (unlikely(!spin_trylock(&inode->i_lock))) {
-                               spin_unlock(&dentry->d_lock);
-                               if (parent)
-                                       spin_unlock(&parent->d_lock);
-                               cpu_relax();
-                               continue;
-                       }
-                       __dentry_kill(dentry);
-                       dentry = parent;
-               }
+               while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
+                       dentry = dentry_kill(dentry);
        }
 }
 
@@ -2379,32 +2393,22 @@ EXPORT_SYMBOL(d_hash_and_lookup);
  
 void d_delete(struct dentry * dentry)
 {
-       struct inode *inode;
-       int isdir = 0;
+       struct inode *inode = dentry->d_inode;
+       int isdir = d_is_dir(dentry);
+
+       spin_lock(&inode->i_lock);
+       spin_lock(&dentry->d_lock);
        /*
         * Are we the only user?
         */
-again:
-       spin_lock(&dentry->d_lock);
-       inode = dentry->d_inode;
-       isdir = S_ISDIR(inode->i_mode);
        if (dentry->d_lockref.count == 1) {
-               if (!spin_trylock(&inode->i_lock)) {
-                       spin_unlock(&dentry->d_lock);
-                       cpu_relax();
-                       goto again;
-               }
                dentry->d_flags &= ~DCACHE_CANT_MOUNT;
                dentry_unlink_inode(dentry);
-               fsnotify_nameremove(dentry, isdir);
-               return;
-       }
-
-       if (!d_unhashed(dentry))
+       } else {
                __d_drop(dentry);
-
-       spin_unlock(&dentry->d_lock);
-
+               spin_unlock(&dentry->d_lock);
+               spin_unlock(&inode->i_lock);
+       }
        fsnotify_nameremove(dentry, isdir);
 }
 EXPORT_SYMBOL(d_delete);
@@ -2769,57 +2773,6 @@ static void copy_name(struct dentry *dentry, struct dentry *target)
                kfree_rcu(old_name, u.head);
 }
 
-static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target)
-{
-       /*
-        * XXXX: do we really need to take target->d_lock?
-        */
-       if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent)
-               spin_lock(&target->d_parent->d_lock);
-       else {
-               if (d_ancestor(dentry->d_parent, target->d_parent)) {
-                       spin_lock(&dentry->d_parent->d_lock);
-                       spin_lock_nested(&target->d_parent->d_lock,
-                                               DENTRY_D_LOCK_NESTED);
-               } else {
-                       spin_lock(&target->d_parent->d_lock);
-                       spin_lock_nested(&dentry->d_parent->d_lock,
-                                               DENTRY_D_LOCK_NESTED);
-               }
-       }
-       if (target < dentry) {
-               spin_lock_nested(&target->d_lock, 2);
-               spin_lock_nested(&dentry->d_lock, 3);
-       } else {
-               spin_lock_nested(&dentry->d_lock, 2);
-               spin_lock_nested(&target->d_lock, 3);
-       }
-}
-
-static void dentry_unlock_for_move(struct dentry *dentry, struct dentry *target)
-{
-       if (target->d_parent != dentry->d_parent)
-               spin_unlock(&dentry->d_parent->d_lock);
-       if (target->d_parent != target)
-               spin_unlock(&target->d_parent->d_lock);
-       spin_unlock(&target->d_lock);
-       spin_unlock(&dentry->d_lock);
-}
-
-/*
- * When switching names, the actual string doesn't strictly have to
- * be preserved in the target - because we're dropping the target
- * anyway. As such, we can just do a simple memcpy() to copy over
- * the new name before we switch, unless we are going to rehash
- * it.  Note that if we *do* unhash the target, we are not allowed
- * to rehash it without giving it a new name/hash key - whether
- * we swap or overwrite the names here, resulting name won't match
- * the reality in filesystem; it's only there for d_path() purposes.
- * Note that all of this is happening under rename_lock, so the
- * any hash lookup seeing it in the middle of manipulations will
- * be discarded anyway.  So we do not care what happens to the hash
- * key in that case.
- */
 /*
  * __d_move - move a dentry
  * @dentry: entry to move
@@ -2834,15 +2787,34 @@ static void dentry_unlock_for_move(struct dentry *dentry, struct dentry *target)
 static void __d_move(struct dentry *dentry, struct dentry *target,
                     bool exchange)
 {
+       struct dentry *old_parent, *p;
        struct inode *dir = NULL;
        unsigned n;
-       if (!dentry->d_inode)
-               printk(KERN_WARNING "VFS: moving negative dcache entry\n");
 
-       BUG_ON(d_ancestor(dentry, target));
+       WARN_ON(!dentry->d_inode);
+       if (WARN_ON(dentry == target))
+               return;
+
        BUG_ON(d_ancestor(target, dentry));
+       old_parent = dentry->d_parent;
+       p = d_ancestor(old_parent, target);
+       if (IS_ROOT(dentry)) {
+               BUG_ON(p);
+               spin_lock(&target->d_parent->d_lock);
+       } else if (!p) {
+               /* target is not a descendent of dentry->d_parent */
+               spin_lock(&target->d_parent->d_lock);
+               spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
+       } else {
+               BUG_ON(p == dentry);
+               spin_lock(&old_parent->d_lock);
+               if (p != target)
+                       spin_lock_nested(&target->d_parent->d_lock,
+                                       DENTRY_D_LOCK_NESTED);
+       }
+       spin_lock_nested(&dentry->d_lock, 2);
+       spin_lock_nested(&target->d_lock, 3);
 
-       dentry_lock_for_move(dentry, target);
        if (unlikely(d_in_lookup(target))) {
                dir = target->d_parent->d_inode;
                n = start_dir_add(dir);
@@ -2853,47 +2825,44 @@ static void __d_move(struct dentry *dentry, struct dentry *target,
        write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
 
        /* unhash both */
-       /* ___d_drop does write_seqcount_barrier, but they're OK to nest. */
-       ___d_drop(dentry);
-       ___d_drop(target);
+       if (!d_unhashed(dentry))
+               ___d_drop(dentry);
+       if (!d_unhashed(target))
+               ___d_drop(target);
 
-       /* Switch the names.. */
-       if (exchange)
-               swap_names(dentry, target);
-       else
+       /* ... and switch them in the tree */
+       dentry->d_parent = target->d_parent;
+       if (!exchange) {
                copy_name(dentry, target);
-
-       /* rehash in new place(s) */
-       __d_rehash(dentry);
-       if (exchange)
-               __d_rehash(target);
-       else
                target->d_hash.pprev = NULL;
-
-       /* ... and switch them in the tree */
-       if (IS_ROOT(dentry)) {
-               /* splicing a tree */
-               dentry->d_flags |= DCACHE_RCUACCESS;
-               dentry->d_parent = target->d_parent;
-               target->d_parent = target;
-               list_del_init(&target->d_child);
-               list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
+               dentry->d_parent->d_lockref.count++;
+               if (dentry == old_parent)
+                       dentry->d_flags |= DCACHE_RCUACCESS;
+               else
+                       WARN_ON(!--old_parent->d_lockref.count);
        } else {
-               /* swapping two dentries */
-               swap(dentry->d_parent, target->d_parent);
+               target->d_parent = old_parent;
+               swap_names(dentry, target);
                list_move(&target->d_child, &target->d_parent->d_subdirs);
-               list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
-               if (exchange)
-                       fsnotify_update_flags(target);
-               fsnotify_update_flags(dentry);
+               __d_rehash(target);
+               fsnotify_update_flags(target);
        }
+       list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
+       __d_rehash(dentry);
+       fsnotify_update_flags(dentry);
 
        write_seqcount_end(&target->d_seq);
        write_seqcount_end(&dentry->d_seq);
 
        if (dir)
                end_dir_add(dir, n);
-       dentry_unlock_for_move(dentry, target);
+
+       if (dentry->d_parent != old_parent)
+               spin_unlock(&dentry->d_parent->d_lock);
+       if (dentry != old_parent)
+               spin_unlock(&old_parent->d_lock);
+       spin_unlock(&target->d_lock);
+       spin_unlock(&dentry->d_lock);
 }
 
 /*
@@ -3041,12 +3010,14 @@ struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
                                        inode->i_sb->s_type->name,
                                        inode->i_sb->s_id);
                        } else if (!IS_ROOT(new)) {
+                               struct dentry *old_parent = dget(new->d_parent);
                                int err = __d_unalias(inode, dentry, new);
                                write_sequnlock(&rename_lock);
                                if (err) {
                                        dput(new);
                                        new = ERR_PTR(err);
                                }
+                               dput(old_parent);
                        } else {
                                __d_move(new, dentry, false);
                                write_sequnlock(&rename_lock);
@@ -3061,467 +3032,6 @@ out:
 }
 EXPORT_SYMBOL(d_splice_alias);
 
-static int prepend(char **buffer, int *buflen, const char *str, int namelen)
-{
-       *buflen -= namelen;
-       if (*buflen < 0)
-               return -ENAMETOOLONG;
-       *buffer -= namelen;
-       memcpy(*buffer, str, 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 READ_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.
- *
- * Load acquire is needed to make sure that we see that terminating NUL.
- */
-static int prepend_name(char **buffer, int *buflen, const struct qstr *name)
-{
-       const char *dname = smp_load_acquire(&name->name); /* ^^^ */
-       u32 dlen = READ_ONCE(name->len);
-       char *p;
-
-       *buflen -= dlen + 1;
-       if (*buflen < 0)
-               return -ENAMETOOLONG;
-       p = *buffer -= dlen + 1;
-       *p++ = '/';
-       while (dlen--) {
-               char c = *dname++;
-               if (!c)
-                       break;
-               *p++ = c;
-       }
-       return 0;
-}
-
-/**
- * prepend_path - Prepend path string to a buffer
- * @path: the dentry/vfsmount to report
- * @root: root vfsmnt/dentry
- * @buffer: pointer to the end of the buffer
- * @buflen: pointer to buffer length
- *
- * The function will first try 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 changed, it will restart
- * the whole pathname back-tracing sequence again by taking the rename_lock.
- * In this case, there is no need to take the RCU read lock as the recursive
- * parent pointer references will keep the dentry chain alive as long as no
- * rename operation is performed.
- */
-static int prepend_path(const struct path *path,
-                       const struct path *root,
-                       char **buffer, int *buflen)
-{
-       struct dentry *dentry;
-       struct vfsmount *vfsmnt;
-       struct mount *mnt;
-       int error = 0;
-       unsigned seq, m_seq = 0;
-       char *bptr;
-       int blen;
-
-       rcu_read_lock();
-restart_mnt:
-       read_seqbegin_or_lock(&mount_lock, &m_seq);
-       seq = 0;
-       rcu_read_lock();
-restart:
-       bptr = *buffer;
-       blen = *buflen;
-       error = 0;
-       dentry = path->dentry;
-       vfsmnt = path->mnt;
-       mnt = real_mount(vfsmnt);
-       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)) {
-                       struct mount *parent = READ_ONCE(mnt->mnt_parent);
-                       /* Escaped? */
-                       if (dentry != vfsmnt->mnt_root) {
-                               bptr = *buffer;
-                               blen = *buflen;
-                               error = 3;
-                               break;
-                       }
-                       /* Global root? */
-                       if (mnt != parent) {
-                               dentry = READ_ONCE(mnt->mnt_mountpoint);
-                               mnt = parent;
-                               vfsmnt = &mnt->mnt;
-                               continue;
-                       }
-                       if (!error)
-                               error = is_mounted(vfsmnt) ? 1 : 2;
-                       break;
-               }
-               parent = dentry->d_parent;
-               prefetch(parent);
-               error = prepend_name(&bptr, &blen, &dentry->d_name);
-               if (error)
-                       break;
-
-               dentry = parent;
-       }
-       if (!(seq & 1))
-               rcu_read_unlock();
-       if (need_seqretry(&rename_lock, seq)) {
-               seq = 1;
-               goto restart;
-       }
-       done_seqretry(&rename_lock, seq);
-
-       if (!(m_seq & 1))
-               rcu_read_unlock();
-       if (need_seqretry(&mount_lock, m_seq)) {
-               m_seq = 1;
-               goto restart_mnt;
-       }
-       done_seqretry(&mount_lock, m_seq);
-
-       if (error >= 0 && bptr == *buffer) {
-               if (--blen < 0)
-                       error = -ENAMETOOLONG;
-               else
-                       *--bptr = '/';
-       }
-       *buffer = bptr;
-       *buflen = blen;
-       return error;
-}
-
-/**
- * __d_path - return the path of a dentry
- * @path: the dentry/vfsmount to report
- * @root: root vfsmnt/dentry
- * @buf: buffer to return value in
- * @buflen: buffer length
- *
- * Convert a dentry into an ASCII path name.
- *
- * Returns a pointer into the buffer or an error code if the
- * path was too long.
- *
- * "buflen" should be positive.
- *
- * If the path is not reachable from the supplied root, return %NULL.
- */
-char *__d_path(const struct path *path,
-              const struct path *root,
-              char *buf, int buflen)
-{
-       char *res = buf + buflen;
-       int error;
-
-       prepend(&res, &buflen, "\0", 1);
-       error = prepend_path(path, root, &res, &buflen);
-
-       if (error < 0)
-               return ERR_PTR(error);
-       if (error > 0)
-               return NULL;
-       return res;
-}
-
-char *d_absolute_path(const struct path *path,
-              char *buf, int buflen)
-{
-       struct path root = {};
-       char *res = buf + buflen;
-       int error;
-
-       prepend(&res, &buflen, "\0", 1);
-       error = prepend_path(path, &root, &res, &buflen);
-
-       if (error > 1)
-               error = -EINVAL;
-       if (error < 0)
-               return ERR_PTR(error);
-       return res;
-}
-
-/*
- * same as __d_path but appends "(deleted)" for unlinked files.
- */
-static int path_with_deleted(const struct path *path,
-                            const struct path *root,
-                            char **buf, int *buflen)
-{
-       prepend(buf, buflen, "\0", 1);
-       if (d_unlinked(path->dentry)) {
-               int error = prepend(buf, buflen, " (deleted)", 10);
-               if (error)
-                       return error;
-       }
-
-       return prepend_path(path, root, buf, buflen);
-}
-
-static int prepend_unreachable(char **buffer, int *buflen)
-{
-       return prepend(buffer, buflen, "(unreachable)", 13);
-}
-
-static void get_fs_root_rcu(struct fs_struct *fs, struct path *root)
-{
-       unsigned seq;
-
-       do {
-               seq = read_seqcount_begin(&fs->seq);
-               *root = fs->root;
-       } while (read_seqcount_retry(&fs->seq, seq));
-}
-
-/**
- * d_path - return the path of a dentry
- * @path: path to report
- * @buf: buffer to return value in
- * @buflen: buffer length
- *
- * Convert a dentry into an ASCII path name. If the entry has been deleted
- * the string " (deleted)" is appended. Note that this is ambiguous.
- *
- * Returns a pointer into the buffer or an error code if the path was
- * too long. Note: Callers should use the returned pointer, not the passed
- * in buffer, to use the name! The implementation often starts at an offset
- * into the buffer, and may leave 0 bytes at the start.
- *
- * "buflen" should be positive.
- */
-char *d_path(const struct path *path, char *buf, int buflen)
-{
-       char *res = buf + buflen;
-       struct path root;
-       int error;
-
-       /*
-        * We have various synthetic filesystems that never get mounted.  On
-        * these filesystems dentries are never used for lookup purposes, and
-        * thus don't need to be hashed.  They also don't need a name until a
-        * user wants to identify the object in /proc/pid/fd/.  The little hack
-        * below allows us to generate a name for these objects on demand:
-        *
-        * Some pseudo inodes are mountable.  When they are mounted
-        * path->dentry == path->mnt->mnt_root.  In that case don't call d_dname
-        * and instead have d_path return the mounted path.
-        */
-       if (path->dentry->d_op && path->dentry->d_op->d_dname &&
-           (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root))
-               return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
-
-       rcu_read_lock();
-       get_fs_root_rcu(current->fs, &root);
-       error = path_with_deleted(path, &root, &res, &buflen);
-       rcu_read_unlock();
-
-       if (error < 0)
-               res = ERR_PTR(error);
-       return res;
-}
-EXPORT_SYMBOL(d_path);
-
-/*
- * Helper function for dentry_operations.d_dname() members
- */
-char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
-                       const char *fmt, ...)
-{
-       va_list args;
-       char temp[64];
-       int sz;
-
-       va_start(args, fmt);
-       sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
-       va_end(args);
-
-       if (sz > sizeof(temp) || sz > buflen)
-               return ERR_PTR(-ENAMETOOLONG);
-
-       buffer += buflen - sz;
-       return memcpy(buffer, temp, sz);
-}
-
-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(&end, &buflen, dentry->d_name.name, dentry->d_name.len) ||
-           prepend(&end, &buflen, "/", 1))  
-               end = ERR_PTR(-ENAMETOOLONG);
-       return end;
-}
-EXPORT_SYMBOL(simple_dname);
-
-/*
- * Write full pathname from the root of the filesystem into the buffer.
- */
-static char *__dentry_path(struct dentry *d, char *buf, int buflen)
-{
-       struct dentry *dentry;
-       char *end, *retval;
-       int len, seq = 0;
-       int error = 0;
-
-       if (buflen < 2)
-               goto Elong;
-
-       rcu_read_lock();
-restart:
-       dentry = d;
-       end = buf + buflen;
-       len = buflen;
-       prepend(&end, &len, "\0", 1);
-       /* Get '/' right */
-       retval = end-1;
-       *retval = '/';
-       read_seqbegin_or_lock(&rename_lock, &seq);
-       while (!IS_ROOT(dentry)) {
-               struct dentry *parent = dentry->d_parent;
-
-               prefetch(parent);
-               error = prepend_name(&end, &len, &dentry->d_name);
-               if (error)
-                       break;
-
-               retval = end;
-               dentry = parent;
-       }
-       if (!(seq & 1))
-               rcu_read_unlock();
-       if (need_seqretry(&rename_lock, seq)) {
-               seq = 1;
-               goto restart;
-       }
-       done_seqretry(&rename_lock, seq);
-       if (error)
-               goto Elong;
-       return retval;
-Elong:
-       return ERR_PTR(-ENAMETOOLONG);
-}
-
-char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
-{
-       return __dentry_path(dentry, buf, buflen);
-}
-EXPORT_SYMBOL(dentry_path_raw);
-
-char *dentry_path(struct dentry *dentry, char *buf, int buflen)
-{
-       char *p = NULL;
-       char *retval;
-
-       if (d_unlinked(dentry)) {
-               p = buf + buflen;
-               if (prepend(&p, &buflen, "//deleted", 10) != 0)
-                       goto Elong;
-               buflen++;
-       }
-       retval = __dentry_path(dentry, buf, buflen);
-       if (!IS_ERR(retval) && p)
-               *p = '/';       /* restore '/' overriden with '\0' */
-       return retval;
-Elong:
-       return ERR_PTR(-ENAMETOOLONG);
-}
-
-static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root,
-                                   struct path *pwd)
-{
-       unsigned seq;
-
-       do {
-               seq = read_seqcount_begin(&fs->seq);
-               *root = fs->root;
-               *pwd = fs->pwd;
-       } while (read_seqcount_retry(&fs->seq, seq));
-}
-
-/*
- * NOTE! The user-level library version returns a
- * character pointer. The kernel system call just
- * returns the length of the buffer filled (which
- * includes the ending '\0' character), or a negative
- * error value. So libc would do something like
- *
- *     char *getcwd(char * buf, size_t size)
- *     {
- *             int retval;
- *
- *             retval = sys_getcwd(buf, size);
- *             if (retval >= 0)
- *                     return buf;
- *             errno = -retval;
- *             return NULL;
- *     }
- */
-SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
-{
-       int error;
-       struct path pwd, root;
-       char *page = __getname();
-
-       if (!page)
-               return -ENOMEM;
-
-       rcu_read_lock();
-       get_fs_root_and_pwd_rcu(current->fs, &root, &pwd);
-
-       error = -ENOENT;
-       if (!d_unlinked(pwd.dentry)) {
-               unsigned long len;
-               char *cwd = page + PATH_MAX;
-               int buflen = PATH_MAX;
-
-               prepend(&cwd, &buflen, "\0", 1);
-               error = prepend_path(&pwd, &root, &cwd, &buflen);
-               rcu_read_unlock();
-
-               if (error < 0)
-                       goto out;
-
-               /* Unreachable from current root */
-               if (error > 0) {
-                       error = prepend_unreachable(&cwd, &buflen);
-                       if (error)
-                               goto out;
-               }
-
-               error = -ERANGE;
-               len = PATH_MAX + page - cwd;
-               if (len <= size) {
-                       error = len;
-                       if (copy_to_user(buf, cwd, len))
-                               error = -EFAULT;
-               }
-       } else {
-               rcu_read_unlock();
-       }
-
-out:
-       __putname(page);
-       return error;
-}
-
 /*
  * Test whether new_dentry is a subdirectory of old_dentry.
  *
@@ -3585,6 +3095,8 @@ void d_genocide(struct dentry *parent)
        d_walk(parent, parent, d_genocide_kill, NULL);
 }
 
+EXPORT_SYMBOL(d_genocide);
+
 void d_tmpfile(struct dentry *dentry, struct inode *inode)
 {
        inode_dec_link_count(inode);
@@ -3664,8 +3176,6 @@ static void __init dcache_init(void)
 struct kmem_cache *names_cachep __read_mostly;
 EXPORT_SYMBOL(names_cachep);
 
-EXPORT_SYMBOL(d_genocide);
-
 void __init vfs_caches_init_early(void)
 {
        int i;
index 63a998c3f2522d2d727a52bf1a8a9aae4a7c70cd..13b01351dd1cb3f8381dbddd2ee70717792b9ac7 100644 (file)
@@ -270,10 +270,7 @@ struct dentry *debugfs_lookup(const char *name, struct dentry *parent)
        if (!parent)
                parent = debugfs_mount->mnt_root;
 
-       inode_lock(d_inode(parent));
-       dentry = lookup_one_len(name, parent, strlen(name));
-       inode_unlock(d_inode(parent));
-
+       dentry = lookup_one_len_unlocked(name, parent, strlen(name));
        if (IS_ERR(dentry))
                return NULL;
        if (!d_really_is_positive(dentry)) {
index 5c2b953c352db259e648707d5e67dea4457885d0..a09419379f5db033d2c0315ae379180f73d6cff5 100644 (file)
@@ -1474,43 +1474,36 @@ static struct dentry *lookup_dcache(const struct qstr *name,
 }
 
 /*
- * Call i_op->lookup on the dentry.  The dentry must be negative and
- * unhashed.
- *
- * dir->d_inode->i_mutex must be held
+ * Parent directory has inode locked exclusive.  This is one
+ * and only case when ->lookup() gets called on non in-lookup
+ * dentries - as the matter of fact, this only gets called
+ * when directory is guaranteed to have no in-lookup children
+ * at all.
  */
-static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
-                                 unsigned int flags)
-{
-       struct dentry *old;
-
-       /* Don't create child dentry for a dead directory. */
-       if (unlikely(IS_DEADDIR(dir))) {
-               dput(dentry);
-               return ERR_PTR(-ENOENT);
-       }
-
-       old = dir->i_op->lookup(dir, dentry, flags);
-       if (unlikely(old)) {
-               dput(dentry);
-               dentry = old;
-       }
-       return dentry;
-}
-
 static struct dentry *__lookup_hash(const struct qstr *name,
                struct dentry *base, unsigned int flags)
 {
        struct dentry *dentry = lookup_dcache(name, base, flags);
+       struct dentry *old;
+       struct inode *dir = base->d_inode;
 
        if (dentry)
                return dentry;
 
+       /* Don't create child dentry for a dead directory. */
+       if (unlikely(IS_DEADDIR(dir)))
+               return ERR_PTR(-ENOENT);
+
        dentry = d_alloc(base, name);
        if (unlikely(!dentry))
                return ERR_PTR(-ENOMEM);
 
-       return lookup_real(base->d_inode, dentry, flags);
+       old = dir->i_op->lookup(dir, dentry, flags);
+       if (unlikely(old)) {
+               dput(dentry);
+               dentry = old;
+       }
+       return dentry;
 }
 
 static int lookup_fast(struct nameidata *nd,
index 82a99d366aece7acb77a1ecf5ab9809ada60fbe4..94acbde17bb190e3299c35e92ad96525f575812b 100644 (file)
@@ -56,9 +56,7 @@ struct qstr {
 
 #define QSTR_INIT(n,l) { { { .len = l } }, .name = n }
 
-extern const char empty_string[];
 extern const struct qstr empty_name;
-extern const char slash_string[];
 extern const struct qstr slash_name;
 
 struct dentry_stat_t {
@@ -361,7 +359,7 @@ static inline void dont_mount(struct dentry *dentry)
 
 extern void __d_lookup_done(struct dentry *);
 
-static inline int d_in_lookup(struct dentry *dentry)
+static inline int d_in_lookup(const struct dentry *dentry)
 {
        return dentry->d_flags & DCACHE_PAR_LOOKUP;
 }
@@ -489,7 +487,7 @@ static inline bool d_really_is_positive(const struct dentry *dentry)
        return dentry->d_inode != NULL;
 }
 
-static inline int simple_positive(struct dentry *dentry)
+static inline int simple_positive(const struct dentry *dentry)
 {
        return d_really_is_positive(dentry) && !d_unhashed(dentry);
 }