The New VFS

From SambaWiki

Summary

The effort to modernize Samba's VFS interface has reached a major milestone with the next release Samba 4.14.

Starting with version 4.14 Samba provides core infrastructure code that allows basing all access to the server's filesystem on file handles and not on paths. An example of this is using =fstat()= instead of =stat()=, or =SMB_VFS_FSTAT()= instead of =SMB_VFS_STAT()= in Samba parlance.

Historically Samba's fileserver code had to deal a lot with processing path based SMB requests. While the SMB protocol itself has been streamlined to be purely handle based starting with SMB2, large parts of infrastructure code remains in place that will "degrade" handle based SMB2 requests to path based filesystem access.

In order to fully leverage the handle based nature of the SMB2 protocol we came up with a straight forward way to convert this infrastructure code.

At the core, we introduced a helper function that opens a file handle that only serves as a path reference and hence can not be used for any sort of access to file data.

Samba's internal file handle structure is of type =struct files_struct= and all variable pointing to objects of such type are typically called =fsp=. Until very recently the only function that would open such a file handle and return an fsp was =SMB_VFS_CREATE_FILE()=.

Internally =SMB_VFS_CREATE_FILE()= consisted of processing through Samba's VFS open function to open the low level file and then going through Samba's Windows NTFS emulation code.

The key point of the new helper function which is called =openat_pathref_fsp()= is that it skips the NTFS emulation logic. Additionally, the handle is restricted internally to be only usable as a path reference but not for any sort of IO. On Linux this is achieved by using the =O_PATH= =open()= flag, on systems without =O_PATH= support other mechanisms are used described in more detail below.

Path processing in Samba typically means processing client supplied paths by Samba's core path processing function =filename_convert()= which returs a pointer to an object of type =struct smb_filename=. Pointers to such objects are then passed around, often passing many layers of code.

By attaching an =fsp= file handle returned from =openat_pathref_fsp()= to all =struct smb_filename= objects returned from =filename_convert()=, the whole infrastructure code has immediate access to a file handle and so the large infrastructure codebase can be converted to use handle based VFS functions whenever VFS access is done in a piecemeal fashion.

Samba and O_PATH

Background

On Linux the =O_PATH= flag to =open()= can be used to open a filehandle on a file or directory with interesting properties: [fn:manpage]

- the file-handle indicates a location in the filesystem tree, - no permission checks are done by the kernel on the filesystem object and - only operations that act purely at the file descriptor level are allowed.

The file itself is not opened, and other file operations (e.g., ~read(2)~, ~write(2)~, ~fchmod(2)~, ~fchown(2)~, ~fgetxattr(2)~, ~ioctl(2)~, ~mmap(2)~) fail with the error ~EBADF~.

The following subset of operations that is relevant to Samba is allowed:

- ~close(2)~, - ~fchdir(2)~, if the file descriptor refers to a directory, - ~fstat(2)~, - ~fstatfs(2)~ and - passing the file descriptor as the dirfd argument of ~openat()~ and the other "*at()" system calls. This includes ~linkat(2)~ with AT_EMPTY_PATH (or via procfs using AT_SYMLINK_FOLLOW) even if the file is not a directory.

Opening a file or directory with the ~O_PATH~ flag requires no permissions on the object itself (but does require execute permission on the directories in the path prefix). By contrast, obtaining a reference to a filesystem object by opening it with the ~O_RDONLY~ flag requires that the caller have read permission on the object, even when the subsequent operation (e.g., ~fchdir(2)~, ~fstat(2)~) does not require read permission on the object.

If for example Samba receives an SMB request to open a file requesting ~SEC_FILE_READ_ATTRIBUTE~ access rights because the client wants to read the file's metadata from the handle, Samba will have to call ~open()~ with at least ~O_RDONLY~ access rights.

Usecases for O_PATH in Samba

The ~O_PATH~ flag is currently not used in Samba. By leveraging this Linux specific flags we can avoid permission mismatches as described above.

Additionally ~O_PATH~ allows basing all filesystem accesses done by the fileserver on handle based syscalls by opening all client pathnames with ~O_PATH~ and consistently using for example ~fstat()~ instead of ~stat()~ throughout the codebase.

Subsequent parts of this document will call such file-handles opened with O_PATH *path referencing file-handles* or *pathref*s for short.

When to open with O_PATH

In Samba the decision whether to call POSIX ~open()~ on a client pathname or whether to leave the low-level handle at -1 (what we call a stat-open) is based on the client requested SMB acccess mask.

The set of access rights that trigger an ~open()~ includes ~READ_CONTROL_ACCESS~. As a result, the open() will be done with at least ~O_RDONLY~. If the filesystem supports NT style ACLs natively (like GPFS or ZFS), the filesystem may grant the user requested right ~READ_CONTROL_ACCESS~, but it may not grant ~READ_DATA~ (~O_RDONLY~).

Currently the full set of access rights that trigger opening a file is:

- FILE_READ_DATA - FILE_WRITE_DATA - FILE_APPEND_DATA - FILE_EXECUTE - WRITE_DAC_ACCESS - WRITE_OWNER_ACCESS - SEC_FLAG_SYSTEM_SECURITY - READ_CONTROL_ACCESS

In the future we can remove the following rights from the list on systems that support O_PATH:

- WRITE_DAC_ACCESS - WRITE_OWNER_ACCESS - SEC_FLAG_SYSTEM_SECURITY - READ_CONTROL_ACCESS

Fallback on systems without O_PATH support

The code of higher level file-handle consumers must be kept simple and
streamlined, avoiding special casing the handling of the file-handles opened
with or without ~O_PATH~. To achieve this, a fallback that allows opening a
file-handle with the same higher level semantics even if the system doesn't
support ~O_PATH~ is needed.
The way this is implemented on such systems is impersonating the root user for
the ~open()~ syscall. In order to avoid privelege escalations security issues,
we must carefully control the use these file-handles.
The low level filehandle is stored in a public struct ~struct file_handle~ that
is part of the widely used ~struct files_struct~. Consumers used to simply
access the fd directly by derefencing pointers to ~struct files_struct~.
In order to guard access to such file-handles we do two things:
- tag the pathref file-handles and
- control access to the file-handle by making the structure ~struct
  file_handle~ private, only allowing access with accessor functions that
  implement a security boundary.
In order to avoid bypassing restrictive permissions on intermediate directories
of a client path, the root user is only impersonated after changing directory
to the parent directory of the client requested pathname.
Two functions can then be used to fetch the low-level system file-handle from a
~struct files_struct~:
- ~fsp_get_io_fd(fsp)~: enforces fsp is NOT a pathref file-handle and
- ~fsp_get_pathref_fd(fsp)~: allows fsp to be either a pathref file-handle or a
  traditional POSIX file-handle opened with O_RDONLY or any other POSIX open
  flag.
Note that the name ~fsp_get_pathref_fd()~ may sound confusing at first given
that the fsp can be either a pathref fsp or a "normal/full" fsp, but as any
full file-handle can be used for IO and as path reference, the name
correctly reflects the intended usage of the caller.

When to use fsp_get_io_fd() or fsp_get_pathref_fd()

The general guideline is:
- if you do something like ~fstat(fd)~, use ~fsp_get_pathref_fd()~,
- if you do something like ~*at(dirfd, ...)~, use ~fsp_get_pathref_fd()~,
- if you want to print the fd for example in =DEBUG= messages, use ~fsp_get_pathref_fd()~,
- if you want to call ~close(fd)~, use ~fsp_get_pathref_fd()~,
- if you're doing a logical comparison of fd values, use ~fsp_get_pathref_fd()~.
In any other case use ~fsp_get_io_fd()~.