netbird/gvisor
0
0
Fork 0
mirror of https://github.com/netbirdio/gvisor.git synced 2026-05-22 17:12:49 -07:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #2751 from mrahatm:fuse

Browse Source 
PiperOrigin-RevId: 313300882
This commit is contained in:
gVisor bot
2020-05-26 17:50:13 -07:00
parent b5caef46db 7938d87348
commit 8e2c5d951e
2 changed files with 261 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
pkg/sentry/fs/g3doc/.gitignore
+1
View File
@@ -0,0 +1 @@
*.html
pkg/sentry/fs/g3doc/fuse.md
+260
View File
@@ -0,0 +1,260 @@
# Foreword
This document describes an on-going project to support FUSE filesystems within
the sentry. This is intended to become the final documentation for this
subsystem, and is therefore written in the past tense. However FUSE support is
currently incomplete and the document will be updated as things progress.
# FUSE: Filesystem in Userspace
The sentry supports dispatching filesystem operations to a FUSE server, allowing
FUSE filesystem to be used with a sandbox.
## Overview
FUSE has two main components:
1. A client kernel driver (canonically `fuse.ko` in Linux), which forwards
filesystem operations (usually initiated by syscalls) to the server.
2. A server, which is a userspace daemon that implements the actual filesystem.
The sentry implements the client component, which allows a server daemon running
within the sandbox to implement a filesystem within the sandbox.
A FUSE filesystem is initialized with `mount(2)`, typically with the help of a
utility like `fusermount(1)`. Various mount options exist for establishing
ownership and access permissions on the filesystem, but the most important mount
option is a file descriptor used to establish communication between the client
and server.
The FUSE device FD is obtained by opening `/dev/fuse`. During regular operation,
the client and server use the FUSE protocol described in `fuse(4)` to service
filesystem operations. See the "Protocol" section below for more information
about this protocol. The core of the sentry support for FUSE is the client-side
implementation of this protocol.
## FUSE in the Sentry
The sentry's FUSE client targets VFS2 and has the following components:
- An implementation of `/dev/fuse`.
- A VFS2 filesystem for mapping syscalls to FUSE ops. Since we're targeting
VFS2, one point of contention may be the lack of inodes in VFS2. We can
tentatively implement a kernfs-based filesystem to bridge the gap in APIs.
The kernfs base functionality can serve the role of the Linux inode cache
and, the filesystem can map VFS2 syscalls to kernfs inode operations; see
the `kernfs.Inode` interface.
The FUSE protocol lends itself well to marshaling with `go_marshal`. The various
request and response packets can be defined in the ABI package and converted to
and from the wire format using `go_marshal`.
### Design Goals
- While filesystem performance is always important, the sentry's FUSE support
is primarily concerned with compatibility, with performance as a secondary
concern.
- Avoiding deadlocks from a hung server daemon.
- Consider the potential for denial of service from a malicious server daemon.
Protecting itself from userspace is already a design goal for the sentry,
but needs additional consideration for FUSE. Normally, an operating system
doesn't rely on userspace to make progress with filesystem operations. Since
this changes with FUSE, it opens up the possibility of creating a chain of
dependencies controlled by userspace, which could affect an entire sandbox.
For example: a FUSE op can block a syscall, which could be holding a
subsystem lock, which can then block another task goroutine.
### Milestones
Below are some broad goals to aim for while implementing FUSE in the sentry.
Many FUSE ops can be grouped into broad categories of functionality, and most
ops can be implemented in parallel.
#### Minimal client that can mount a trivial FUSE filesystem.
- Implement `/dev/fuse`.
- Implement basic FUSE ops like `FUSE_INIT`, `FUSE_DESTROY`.
#### Read-only mount with basic file operations
- Implement the majority of file, directory and file descriptor FUSE ops. For
this milestone, we can skip uncommon or complex operations like mmap, mknod,
file locking, poll, and extended attributes. We can stub these out along
with any ops that modify the filesystem. The exact list of required ops are
to be determined, but the goal is to mount a real filesystem as read-only,
and be able to read contents from the filesystem in the sentry.
#### Full read-write support
- Implement the remaining FUSE ops and decide if we can omit rarely used
operations like ioctl.
# Appendix
## FUSE Protocol
The FUSE protocol is a request-response protocol. All requests are initiated by
the client. The wire-format for the protocol is raw c structs serialized to
memory.
All FUSE requests begin with the following request header:
```c
struct fuse_in_header {
uint32_t len; // Length of the request, including this header.
uint32_t opcode; // Requested operation.
uint64_t unique; // A unique identifier for this request.
uint64_t nodeid; // ID of the filesystem object being operated on.
uint32_t uid; // UID of the requesting process.
uint32_t gid; // GID of the requesting process.
uint32_t pid; // PID of the requesting process.
uint32_t padding;
};
```
The request is then followed by a payload specific to the `opcode`.
All responses begin with this response header:
```c
struct fuse_out_header {
uint32_t len; // Length of the response, including this header.
int32_t error; // Status of the request, 0 if success.
uint64_t unique; // The unique identifier from the corresponding request.
};
```
The response payload also depends on the request `opcode`. If `error != 0`, the
response payload must be empty.
### Operations
The following is a list of all FUSE operations used in `fuse_in_header.opcode`
as of Linux v4.4, and a brief description of their purpose. These are defined in
`uapi/linux/fuse.h`. Many of these have a corresponding request and response
payload struct; `fuse(4)` has details for some of these. We also note how these
operations map to the sentry virtual filesystem.
#### FUSE meta-operations
These operations are specific to FUSE and don't have a corresponding action in a
generic filesystem.
- `FUSE_INIT`: This operation initializes a new FUSE filesystem, and is the
first message sent by the client after mount. This is used for version and
feature negotiation. This is related to `mount(2)`.
- `FUSE_DESTROY`: Teardown a FUSE filesystem, related to `unmount(2)`.
- `FUSE_INTERRUPT`: Interrupts an in-flight operation, specified by the
`fuse_in_header.unique` value provided in the corresponding request header.
The client can send at most one of these per request, and will enter an
uninterruptible wait for a reply. The server is expected to reply promptly.
- `FUSE_FORGET`: A hint to the server that server should evict the indicate
node from any caches. This is wired up to `(struct
super_operations).evict_inode` in Linux, which is in turned hooked as the
inode cache shrinker which is typically triggered by system memory pressure.
- `FUSE_BATCH_FORGET`: Batch version of `FUSE_FORGET`.
#### Filesystem Syscalls
These FUSE ops map directly to an equivalent filesystem syscall, or family of
syscalls. The relevant syscalls have a similar name to the operation, unless
otherwise noted.
Node creation:
- `FUSE_MKNOD`
- `FUSE_MKDIR`
- `FUSE_CREATE`: This is equivalent to `open(2)` and `creat(2)`, which
atomically creates and opens a node.
Node attributes and extended attributes:
- `FUSE_GETATTR`
- `FUSE_SETATTR`
- `FUSE_SETXATTR`
- `FUSE_GETXATTR`
- `FUSE_LISTXATTR`
- `FUSE_REMOVEXATTR`
Node link manipulation:
- `FUSE_READLINK`
- `FUSE_LINK`
- `FUSE_SYMLINK`
- `FUSE_UNLINK`
Directory operations:
- `FUSE_RMDIR`
- `FUSE_RENAME`
- `FUSE_RENAME2`
- `FUSE_OPENDIR`: `open(2)` for directories.
- `FUSE_RELEASEDIR`: `close(2)` for directories.
- `FUSE_READDIR`
- `FUSE_READDIRPLUS`
- `FUSE_FSYNCDIR`: `fsync(2)` for directories.
- `FUSE_LOOKUP`: Establishes a unique identifier for a FS node. This is
reminiscent of `VirtualFilesystem.GetDentryAt` in that it resolves a path
component to a node. However the returned identifier is opaque to the
client. The server must remember this mapping, as this is how the client
will reference the node in the future.
File operations:
- `FUSE_OPEN`: `open(2)` for files.
- `FUSE_RELEASE`: `close(2)` for files.
- `FUSE_FSYNC`
- `FUSE_FALLOCATE`
- `FUSE_SETUPMAPPING`: Creates a memory map on a file for `mmap(2)`.
- `FUSE_REMOVEMAPPING`: Removes a memory map for `munmap(2)`.
File locking:
- `FUSE_GETLK`
- `FUSE_SETLK`
- `FUSE_SETLKW`
- `FUSE_COPY_FILE_RANGE`
File descriptor operations:
- `FUSE_IOCTL`
- `FUSE_POLL`
- `FUSE_LSEEK`
Filesystem operations:
- `FUSE_STATFS`
#### Permissions
- `FUSE_ACCESS` is used to check if a node is accessible, as part of many
syscall implementations. Maps to `vfs.FilesystemImpl.AccessAt` in the
sentry.
#### I/O Operations
These ops are used to read and write file pages. They're used to implement both
I/O syscalls like `read(2)`, `write(2)` and `mmap(2)`.
- `FUSE_READ`
- `FUSE_WRITE`
#### Miscellaneous
- `FUSE_FLUSH`: Used by the client to indicate when a file descriptor is
closed. Distinct from `FUSE_FSYNC`, which corresponds to an `fsync(2)`
syscall from the user. Maps to `vfs.FileDescriptorImpl.Release` in the
sentry.
- `FUSE_BMAP`: Old address space API for block defrag. Probably not needed.
- `FUSE_NOTIFY_REPLY`: [TODO: what does this do?]
# References
- `fuse(4)` manpage.
- Linux kernel FUSE documentation:
https://www.kernel.org/doc/html/latest/filesystems/fuse.html