Merge pull request #455 from tjwoodall/unique-inodes

Unique inodes

Author: giuseppe

direct.c

@@ -165,10 +165,80 @@ direct_readlinkat (struct ovl_layer *l, const char *path, char *buf, size_t bufs
   return TEMP_FAILURE_RETRY (readlinkat (l->fd, path, buf, bufsiz));
 }
 
+static ino_t
+direct_get_nfs_filehandle (const struct ovl_layer *l, const char *path)
+{
+  int mount_id;
+  int ret = name_to_handle_at (l->fd, path, l->fh, &mount_id, 0);
+  if (ret == -1)
+    return 0;
+
+  ino_t h = 0xcbf29ce484222325ULL;
+  for (size_t i = 0; i < l->fh->handle_bytes; i++)
+    {
+      h ^= l->fh->f_handle[i];
+      h *= 0x100000001b3ULL;
+    }
+
+  return h;
+}
+
+/* Returns:
+   -1 - Error
+   0  - NFS filehandles not supported
+   1  - Can use NFS filehandles */
+static int
+has_nfs_filehandles (struct ovl_layer *l, const char *path)
+{
+  struct file_handle *tmp_fh;
+  int mount_id;
+  int ret;
+
+  tmp_fh = malloc (sizeof (*tmp_fh));
+  if (tmp_fh == NULL)
+    return -1;
+
+  tmp_fh->handle_bytes = 0;
+
+  ret = name_to_handle_at (AT_FDCWD, path, tmp_fh, &mount_id, 0);
+  if (ret == -1 && errno == ENOTSUP)
+    {
+      free (tmp_fh);
+      return 0;
+    }
+  /* previous call should fail with EOVERFLOW and handle_bytes replaced with
+   * the size of the handle. EOVERFLOW can also occur if no filehandle is
+   * available in a system that does support file-handle lookup.
+   */
+  if (ret != -1 || errno != EOVERFLOW || tmp_fh->handle_bytes == 0)
+    {
+      free (tmp_fh);
+      return -1;
+    }
+
+  l->fh = realloc (tmp_fh, tmp_fh->handle_bytes + sizeof (*l->fh));
+  if (! l->fh)
+    {
+      free (tmp_fh);
+      return -1;
+    }
+
+  ret = name_to_handle_at (AT_FDCWD, path, l->fh, &mount_id, 0);
+
+  if (ret == -1)
+    {
+      free (l->fh);
+      l->fh = NULL;
+      return 0;
+    }
+  return 1;
+}
+
 static int
 direct_load_data_source (struct ovl_layer *l, const char *opaque, const char *path, int n_layer)
 {
   char tmp[64];
+  struct stat st;
   l->path = realpath (path, NULL);
   if (l->path == NULL)
     {
@@ -184,6 +254,18 @@ direct_load_data_source (struct ovl_layer *l, const char *opaque, const char *pa
       return l->fd;
     }
 
+  if (fstat (l->fd, &st) == -1)
+    {
+      close (l->fd);
+      free (l->path);
+      l->path = NULL;
+      return -1;
+    }
+  else
+    l->st_dev = st.st_dev;
+
+  l->nfs_filehandles = has_nfs_filehandles (l, l->path);
+
   if (fgetxattr (l->fd, XATTR_PRIVILEGED_OVERRIDE_STAT, tmp, sizeof (tmp)) >= 0)
     l->stat_override_mode = STAT_OVERRIDE_PRIVILEGED;
   else if (fgetxattr (l->fd, XATTR_OVERRIDE_CONTAINERS_STAT, tmp, sizeof (tmp)) >= 0)
@@ -230,4 +312,5 @@ struct data_source direct_access_ds = {
   .listxattr = direct_listxattr,
   .readlinkat = direct_readlinkat,
   .support_acls = direct_support_acls,
+  .get_nfs_filehandle = direct_get_nfs_filehandle,
 };

fuse-overlayfs.1

@@ -9,11 +9,11 @@ fuse-overlayfs - overlayfs FUSE implementation
 .SH SYNOPSIS
 .PP
 mounting
-    fuse-overlayfs [-f] [--debug] [-o OPTS] MOUNT_TARGET
+    fuse-overlayfs [\-f] [\-\-debug] [\-o OPTS] MOUNT_TARGET
 
 .PP
 unmounting
-    fusermount -u mountpoint
+    fusermount \-u mountpoint
 
 
 .SH DESCRIPTION
@@ -25,26 +25,26 @@ namespace.
 
 .SH OPTIONS
 .PP
-\fB--debug\fP
+\fB\-\-debug\fP
 Enable debugging mode, can be very noisy.
 
 .PP
-\fB-o lowerdir=low1[:low2...]\fP
+\fB\-o lowerdir=low1[:low2...]\fP
 A list of directories separated by \fB\fC:\fR\&.  Their content is merged.
 
 .PP
-\fB-o upperdir=upperdir\fP
+\fB\-o upperdir=upperdir\fP
 A directory merged on top of all the lowerdirs where all the changes
 done to the file system will be written.
 
 .PP
-\fB-o workdir=workdir\fP
+\fB\-o workdir=workdir\fP
 A directory used internally by fuse-overlays, must be on the same file
 system as the upper dir.
 
 .PP
-\fB-o uidmapping=UID:MAPPED-UID:LEN[,UID2:MAPPED-UID2:LEN2]\fP
-\fB-o gidmapping=GID:MAPPED-GID:LEN[,GID2:MAPPED-GID2:LEN2]\fP
+\fB\-o uidmapping=UID:MAPPED-UID:LEN[,UID2:MAPPED-UID2:LEN2]\fP
+\fB\-o gidmapping=GID:MAPPED-GID:LEN[,GID2:MAPPED-GID2:LEN2]\fP
 Specifies the dynamic UID/GID mapping used by fuse-overlayfs when
 reading/writing files to the system.
 
@@ -62,7 +62,7 @@ without requiring to chown the files.
 For example, given on the host two files like:
 
 .PP
-$ stat -c %u:%g lower/a lower/b
+$ stat \-c %u:%g lower/a lower/b
 0:0
 1:1
 
@@ -76,7 +76,7 @@ $ cat /proc/self/uid_map
 We would see:
 
 .PP
-$ stat -c %u:%g merged/a merged/b
+$ stat \-c %u:%g merged/a merged/b
 65534:65534
 65534:65534
 
@@ -87,32 +87,32 @@ mapped.
 
 .PP
 In the above example, if we mount the fuse-overlayfs file system using:
-\fB\fC-ouidmapping=0:1000:1:1:110000:65536,gidmapping=0:1000:1:1:110000:65536\fR,
+\fB\fC\-ouidmapping=0:1000:1:1:110000:65536,gidmapping=0:1000:1:1:110000:65536\fR,
 which is the namespace configuration specified on a single line, we'd
 see from the same user namespace:
 
 .PP
-$ stat -c %u:%g merged/a merged/b
+$ stat \-c %u:%g merged/a merged/b
 0:0
 1:1
 
 .PP
 Those are the same IDs visible from outside the user namespace.
 
 .PP
-\fB-o squash_to_root\fP
+\fB\-o squash_to_root\fP
 Every file and directory is owned by the root user (0:0).
 
 .PP
-\fB-o squash_to_uid=uid\fP
-\fB-o squash_to_gid=gid\fP
+\fB\-o squash_to_uid=uid\fP
+\fB\-o squash_to_gid=gid\fP
 Every file and directory is owned by the specified uid or gid.
 
 .PP
 It has higher precedence over \fBsquash_to_root\fP\&.
 
 .PP
-\fB-o static_nlink\fP
+\fB\-o static_nlink\fP
 Set st_nlink to the static value 1 for all directories.
 
 .PP
@@ -122,9 +122,51 @@ be a slow operation. With this option enabled, the number of hard
 links reported when running stat for any directory is 1.
 
 .PP
-\fB-o noacl\fP
+\fB\-o noacl\fP
 Disable ACL support in the FUSE file system.
 
+.PP
+\fB\-o xino=off|auto|on\fP
+Controls how \fIst_ino\fP values are generated for files exposed by fuse-overlayfs.
+
+When all lower and upper layers reside on the same underlying device,
+fuse-overlayfs exposes the real inode number from the underlying filesystem.
+When layers span multiple devices, an opaque inode number is generated; by
+default this value is not stable across mounts.
+
+The \fBxino\fP option modifies this behavior:
+
+.PP
+\fB\-o xino=off\fP
+Disables extended inode generation. This matches the default behavior:
+when all layers are on the same device, the underlying inode number is used;
+otherwise an opaque, non‑stable inode number is returned.
+
+.PP
+\fB\-o xino=auto\fP
+Attempts to generate stable inode numbers across mounts by hashing the file
+handle returned by \fIname_to_handle_at\fP(2).
+This mode is used only if all layers support \fIname_to_handle_at\fP(2); if any
+layer does not, behavior falls back to \fBxino=off\fP.
+If all layers are on the same device, the underlying inode number is still
+used, regardless of this setting.
+
+.PP
+\fB\-o xino=on\fP
+Requires that all layers support \fIname_to_handle_at\fP(2). If they do, inode
+numbers are derived from a hash of the file handle and remain stable across
+mounts.
+If any layer does not support \fIname_to_handle_at\fP(2), the mount fails.
+As with other modes, when all layers are on the same device, the underlying
+inode number always takes precedence.
+
+.PP
+\fB\-o ino32_t\fP
+Forces all returned \fIst_ino\fP values to be truncated to 32 bits.
+
+This option exists solely for compatibility with older 32‑bit userspaces that
+cannot correctly handle 64‑bit inode numbers. It has no functional benefit on
+modern systems and should not be used unless required for legacy compatibility.
 
 .SH SEE ALSO
 .PP

fuse-overlayfs.1.md

@@ -100,6 +100,44 @@ links reported when running stat for any directory is 1.
 **-o noacl**
 Disable ACL support in the FUSE file system.
 
+**-o xino=off|auto|on**
+Controls how `st_ino` values are generated for files returned by fuse-overlayfs.
+
+When all lower and upper layers reside on the same underlying device,
+fuse-overlayfs exposes the real inode number from the underlying filesystem.
+When layers span multiple devices, an opaque inode number is generated; by
+default this value is not stable across mounts.
+
+The `xino` option modifies this behavior:
+
+**xino=off**
+Disables extended inode generation. This matches the default behavior:
+when all layers are on the same device, the underlying inode number is used;
+otherwise an opaque, non‑stable inode number is returned.
+
+**xino=auto**
+Attempts to generate stable inode numbers across mounts by hashing the file
+handle returned by `name_to_handle_at(2)`.
+This mode is used only if all layers support `name_to_handle_at(2)`; if any
+layer does not, behavior falls back to `xino=off`.
+If all layers are on the same device, the underlying inode number is still
+used, regardless of this setting.
+
+**xino=on**
+Requires that all layers support `name_to_handle_at(2)`. If they do, inode
+numbers are derived from a hash of the file handle and remain stable across
+mounts.
+If any layer does not support `name_to_handle_at(2)`, the mount fails.
+As with other modes, when all layers are on the same device, the underlying
+inode number always takes precedence.
+
+**-o ino32_t**
+Forces all returned `st_ino` values to be truncated to 32 bits.
+
+This option exists solely for compatibility with older 32‑bit userspaces that
+cannot correctly handle 64‑bit inode numbers. It has no functional benefit on
+modern systems and should not be used unless required for legacy compatibility.
+
 # SEE ALSO
 
 **fuse**(8), **mount**(8), **user_namespaces**(7)

fuse-overlayfs.h

@@ -105,6 +105,9 @@ struct ovl_data
   int squash_to_uid;
   int squash_to_gid;
   int static_nlink;
+  int ino_passthrough;
+  int nfs_filehandles;
+  int ino_t_32;
 
   int volatile_mode;
 
@@ -137,6 +140,10 @@ struct ovl_layer
 
   void *data_source_private_data;
   int stat_override_mode;
+
+  dev_t st_dev;
+  int nfs_filehandles;
+  struct file_handle *fh;
 };
 
 /* a data_source defines the methods for accessing a lower layer.  */
@@ -156,6 +163,7 @@ struct data_source
   int (*getxattr) (struct ovl_layer *l, const char *path, const char *name, char *buf, size_t size);
   ssize_t (*readlinkat) (struct ovl_layer *l, const char *path, char *buf, size_t bufsiz);
   bool (*support_acls) (struct ovl_layer *l);
+  ino_t (*get_nfs_filehandle) (const struct ovl_layer *l, const char *path);
 };
 
 /* passthrough to the file system.  */