summaryrefslogtreecommitdiffhomepage
path: root/pkg/lisafs
diff options
context:
space:
mode:
authorJamie Liu <jamieliu@google.com>2020-09-15 13:17:26 -0700
committergVisor bot <gvisor-bot@google.com>2020-09-15 13:19:36 -0700
commit86b31a80777dd906b275fb4b21888e93836f8a9b (patch)
tree51eeb9bb3e2679de26c3dfd283659cdc1469acf1 /pkg/lisafs
parent84d48c0fddc35e82bf8760acd366509581513274 (diff)
RFC: design for a 9P replacement
Tentatively `lisafs` (LInux SAndbox FileSystem). PiperOrigin-RevId: 331839246
Diffstat (limited to 'pkg/lisafs')
-rw-r--r--pkg/lisafs/README.md363
1 files changed, 363 insertions, 0 deletions
diff --git a/pkg/lisafs/README.md b/pkg/lisafs/README.md
new file mode 100644
index 000000000..51d0d40e5
--- /dev/null
+++ b/pkg/lisafs/README.md
@@ -0,0 +1,363 @@
+# Replacing 9P
+
+## Background
+
+The Linux filesystem model consists of the following key aspects (modulo mounts,
+which are outside the scope of this discussion):
+
+- A `struct inode` represents a "filesystem object", such as a directory or a
+ regular file. "Filesystem object" is most precisely defined by the practical
+ properties of an inode, such as an immutable type (regular file, directory,
+ symbolic link, etc.) and its independence from the path originally used to
+ obtain it.
+
+- A `struct dentry` represents a node in a filesystem tree. Semantically, each
+ dentry is immutably associated with an inode representing the filesystem
+ object at that position. (Linux implements optimizations involving reuse of
+ unreferenced dentries, which allows their associated inodes to change, but
+ this is outside the scope of this discussion.)
+
+- A `struct file` represents an open file description (hereafter FD) and is
+ needed to perform I/O. Each FD is immutably associated with the dentry
+ through which it was opened.
+
+The current gVisor virtual filesystem implementation (hereafter VFS1) closely
+imitates the Linux design:
+
+- `struct inode` => `fs.Inode`
+
+- `struct dentry` => `fs.Dirent`
+
+- `struct file` => `fs.File`
+
+gVisor accesses most external filesystems through a variant of the 9P2000.L
+protocol, including extensions for performance (`walkgetattr`) and for features
+not supported by vanilla 9P2000.L (`flushf`, `lconnect`). The 9P protocol family
+is inode-based; 9P fids represent a file (equivalently "file system object"),
+and the protocol is structured around alternatively obtaining fids to represent
+files (with `walk` and, in gVisor, `walkgetattr`) and performing operations on
+those fids.
+
+In the sections below, a **shared** filesystem is a filesystem that is *mutably*
+accessible by multiple concurrent clients, such that a **non-shared** filesystem
+is a filesystem that is either read-only or accessible by only a single client.
+
+## Problems
+
+### Serialization of Path Component RPCs
+
+Broadly speaking, VFS1 traverses each path component in a pathname, alternating
+between verifying that each traversed dentry represents an inode that represents
+a searchable directory and moving to the next dentry in the path.
+
+In the context of a remote filesystem, the structure of this traversal means
+that - modulo caching - a path involving N components requires at least N-1
+*sequential* RPCs to obtain metadata for intermediate directories, incurring
+significant latency. (In vanilla 9P2000.L, 2(N-1) RPCs are required: N-1 `walk`
+and N-1 `getattr`. We added the `walkgetattr` RPC to reduce this overhead.) On
+non-shared filesystems, this overhead is primarily significant during
+application startup; caching mitigates much of this overhead at steady state. On
+shared filesystems, where correct caching requires revalidation (requiring RPCs
+for each revalidated directory anyway), this overhead is consistently ruinous.
+
+### Inefficient RPCs
+
+9P is not exceptionally economical with RPCs in general. In addition to the
+issue described above:
+
+- Opening an existing file in 9P involves at least 2 RPCs: `walk` to produce
+ an unopened fid representing the file, and `lopen` to open the fid.
+
+- Creating a file also involves at least 2 RPCs: `walk` to produce an unopened
+ fid representing the parent directory, and `lcreate` to create the file and
+ convert the fid to an open fid representing the created file. In practice,
+ both the Linux and gVisor 9P clients expect to have an unopened fid for the
+ created file (necessitating an additional `walk`), as well as attributes for
+ the created file (necessitating an additional `getattr`), for a total of 4
+ RPCs. (In a shared filesystem, where whether a file already exists can
+ change between RPCs, a correct implementation of `open(O_CREAT)` would have
+ to alternate between these two paths (plus `clunk`ing the temporary fid
+ between alternations, since the nature of the `fid` differs between the two
+ paths). Neither Linux nor gVisor implement the required alternation, so
+ `open(O_CREAT)` without `O_EXCL` can spuriously fail with `EEXIST` on both.)
+
+- Closing (`clunk`ing) a fid requires an RPC. VFS1 issues this RPC
+ asynchronously in an attempt to reduce critical path latency, but scheduling
+ overhead makes this not clearly advantageous in practice.
+
+- `read` and `readdir` can return partial reads without a way to indicate EOF,
+ necessitating an additional final read to detect EOF.
+
+- Operations that affect filesystem state do not consistently return updated
+ filesystem state. In gVisor, the client implementation attempts to handle
+ this by tracking what it thinks updated state "should" be; this is complex,
+ and especially brittle for timestamps (which are often not arbitrarily
+ settable). In Linux, the client implemtation invalidates cached metadata
+ whenever it performs such an operation, and reloads it when a dentry
+ corresponding to an inode with no valid cached metadata is revalidated; this
+ is simple, but necessitates an additional `getattr`.
+
+### Dentry/Inode Ambiguity
+
+As noted above, 9P's documentation tends to imply that unopened fids represent
+an inode. In practice, most filesystem APIs present very limited interfaces for
+working with inodes at best, such that the interpretation of unopened fids
+varies:
+
+- Linux's 9P client associates unopened fids with (dentry, uid) pairs. When
+ caching is enabled, it also associates each inode with the first fid opened
+ writably that references that inode, in order to support page cache
+ writeback.
+
+- gVisor's 9P client associates unopened fids with inodes, and also caches
+ opened fids in inodes in a manner similar to Linux.
+
+- The runsc fsgofer associates unopened fids with both "dentries" (host
+ filesystem paths) and "inodes" (host file descriptors); which is used
+ depends on the operation invoked on the fid.
+
+For non-shared filesystems, this confusion has resulted in correctness issues
+that are (in gVisor) currently handled by a number of coarse-grained locks that
+serialize renames with all other filesystem operations. For shared filesystems,
+this means inconsistent behavior in the presence of concurrent mutation.
+
+## Design
+
+Almost all Linux filesystem syscalls describe filesystem resources in one of two
+ways:
+
+- Path-based: A filesystem position is described by a combination of a
+ starting position and a sequence of path components relative to that
+ position, where the starting position is one of:
+
+ - The VFS root (defined by mount namespace and chroot), for absolute paths
+
+ - The VFS position of an existing FD, for relative paths passed to `*at`
+ syscalls (e.g. `statat`)
+
+ - The current working directory, for relative paths passed to non-`*at`
+ syscalls and `*at` syscalls with `AT_FDCWD`
+
+- File-description-based: A filesystem object is described by an existing FD,
+ passed to a `f*` syscall (e.g. `fstat`).
+
+Many of our issues with 9P arise from its (and VFS') interposition of a model
+based on inodes between the filesystem syscall API and filesystem
+implementations. We propose to replace 9P with a protocol that does not feature
+inodes at all, and instead closely follows the filesystem syscall API by
+featuring only path-based and FD-based operations, with minimal deviations as
+necessary to ameliorate deficiencies in the syscall interface (see below). This
+approach addresses the issues described above:
+
+- Even on shared filesystems, most application filesystem syscalls are
+ translated to a single RPC (possibly excepting special cases described
+ below), which is a logical lower bound.
+
+- The behavior of application syscalls on shared filesystems is
+ straightforwardly predictable: path-based syscalls are translated to
+ path-based RPCs, which will re-lookup the file at that path, and FD-based
+ syscalls are translated to FD-based RPCs, which use an existing open file
+ without performing another lookup. (This is at least true on gofers that
+ proxy the host local filesystem; other filesystems that lack support for
+ e.g. certain operations on FDs may have different behavior, but this
+ divergence is at least still predictable and inherent to the underlying
+ filesystem implementation.)
+
+Note that this approach is only feasible in gVisor's next-generation virtual
+filesystem (VFS2), which does not assume the existence of inodes and allows the
+remote filesystem client to translate whole path-based syscalls into RPCs. Thus
+one of the unavoidable tradeoffs associated with such a protocol vs. 9P is the
+inability to construct a Linux client that is performance-competitive with
+gVisor.
+
+### File Permissions
+
+Many filesystem operations are side-effectual, such that file permissions must
+be checked before such operations take effect. The simplest approach to file
+permission checking is for the sentry to obtain permissions from the remote
+filesystem, then apply permission checks in the sentry before performing the
+application-requested operation. However, this requires an additional RPC per
+application syscall (which can't be mitigated by caching on shared filesystems).
+Alternatively, we may delegate file permission checking to gofers. In general,
+file permission checks depend on the following properties of the accessor:
+
+- Filesystem UID/GID
+
+- Supplementary GIDs
+
+- Effective capabilities in the accessor's user namespace (i.e. the accessor's
+ effective capability set)
+
+- All UIDs and GIDs mapped in the accessor's user namespace (which determine
+ if the accessor's capabilities apply to accessed files)
+
+We may choose to delay implementation of file permission checking delegation,
+although this is potentially costly since it doubles the number of required RPCs
+for most operations on shared filesystems. We may also consider compromise
+options, such as only delegating file permission checks for accessors in the
+root user namespace.
+
+### Symbolic Links
+
+gVisor usually interprets symbolic link targets in its VFS rather than on the
+filesystem containing the symbolic link; thus e.g. a symlink to
+"/proc/self/maps" on a remote filesystem resolves to said file in the sentry's
+procfs rather than the host's. This implies that:
+
+- Remote filesystem servers that proxy filesystems supporting symlinks must
+ check if each path component is a symlink during path traversal.
+
+- Absolute symlinks require that the sentry restart the operation at its
+ contextual VFS root (which is task-specific and may not be on a remote
+ filesystem at all), so if a remote filesystem server encounters an absolute
+ symlink during path traversal on behalf of a path-based operation, it must
+ terminate path traversal and return the symlink target.
+
+- Relative symlinks begin target resolution in the parent directory of the
+ symlink, so in theory most relative symlinks can be handled automatically
+ during the path traversal that encounters the symlink, provided that said
+ traversal is supplied with the number of remaining symlinks before `ELOOP`.
+ However, the new path traversed by the symlink target may cross VFS mount
+ boundaries, such that it's only safe for remote filesystem servers to
+ speculatively follow relative symlinks for side-effect-free operations such
+ as `stat` (where the sentry can simply ignore results that are inapplicable
+ due to crossing mount boundaries). We may choose to delay implementation of
+ this feature, at the cost of an additional RPC per relative symlink (note
+ that even if the symlink target crosses a mount boundary, the sentry will
+ need to `stat` the path to the mount boundary to confirm that each traversed
+ component is an accessible directory); until it is implemented, relative
+ symlinks may be handled like absolute symlinks, by terminating path
+ traversal and returning the symlink target.
+
+The possibility of symlinks (and the possibility of a compromised sentry) means
+that the sentry may issue RPCs with paths that, in the absence of symlinks,
+would traverse beyond the root of the remote filesystem. For example, the sentry
+may issue an RPC with a path like "/foo/../..", on the premise that if "/foo" is
+a symlink then the resulting path may be elsewhere on the remote filesystem. To
+handle this, path traversal must also track its current depth below the remote
+filesystem root, and terminate path traversal if it would ascend beyond this
+point.
+
+### Path Traversal
+
+Since path-based VFS operations will translate to path-based RPCs, filesystem
+servers will need to handle path traversal. From the perspective of a given
+filesystem implementation in the server, there are two basic approaches to path
+traversal:
+
+- Inode-walk: For each path component, obtain a handle to the underlying
+ filesystem object (e.g. with `open(O_PATH)`), check if that object is a
+ symlink (as described above) and that that object is accessible by the
+ caller (e.g. with `fstat()`), then continue to the next path component (e.g.
+ with `openat()`). This ensures that the checked filesystem object is the one
+ used to obtain the next object in the traversal, which is intuitively
+ appealing. However, while this approach works for host local filesystems, it
+ requires features that are not widely supported by other filesystems.
+
+- Path-walk: For each path component, use a path-based operation to determine
+ if the filesystem object currently referred to by that path component is a
+ symlink / is accessible. This is highly portable, but suffers from quadratic
+ behavior (at the level of the underlying filesystem implementation, the
+ first path component will be traversed a number of times equal to the number
+ of path components in the path).
+
+The implementation should support either option by delegating path traversal to
+filesystem implementations within the server (like VFS and the remote filesystem
+protocol itself), as inode-walking is still safe, efficient, amenable to FD
+caching, and implementable on non-shared host local filesystems (a sufficiently
+common case as to be worth considering in the design).
+
+Both approaches are susceptible to race conditions that may permit sandboxed
+filesystem escapes:
+
+- Under inode-walk, a malicious application may cause a directory to be moved
+ (with `rename`) during path traversal, such that the filesystem
+ implementation incorrectly determines whether subsequent inodes are located
+ in paths that should be visible to sandboxed applications.
+
+- Under path-walk, a malicious application may cause a non-symlink file to be
+ replaced with a symlink during path traversal, such that following path
+ operations will incorrectly follow the symlink.
+
+Both race conditions can, to some extent, be mitigated in filesystem server
+implementations by synchronizing path traversal with the hazardous operations in
+question. However, shared filesystems are frequently used to share data between
+sandboxed and unsandboxed applications in a controlled way, and in some cases a
+malicious sandboxed application may be able to take advantage of a hazardous
+filesystem operation performed by an unsandboxed application. In some cases,
+filesystem features may be available to ensure safety even in such cases (e.g.
+[the new openat2() syscall](https://man7.org/linux/man-pages/man2/openat2.2.html)),
+but it is not clear how to solve this problem in general. (Note that this issue
+is not specific to our design; rather, it is a fundamental limitation of
+filesystem sandboxing.)
+
+### Filesystem Multiplexing
+
+A given sentry may need to access multiple distinct remote filesystems (e.g.
+different volumes for a given container). In many cases, there is no advantage
+to serving these filesystems from distinct filesystem servers, or accessing them
+through distinct connections (factors such as maximum RPC concurrency should be
+based on available host resources). Therefore, the protocol should support
+multiplexing of distinct filesystem trees within a single session. 9P supports
+this by allowing multiple calls to the `attach` RPC to produce fids representing
+distinct filesystem trees, but this is somewhat clunky; we propose a much
+simpler mechanism wherein each message that conveys a path also conveys a
+numeric filesystem ID that identifies a filesystem tree.
+
+## Alternatives Considered
+
+### Additional Extensions to 9P
+
+There are at least three conceptual aspects to 9P:
+
+- Wire format: messages with a 4-byte little-endian size prefix, strings with
+ a 2-byte little-endian size prefix, etc. Whether the wire format is worth
+ retaining is unclear; in particular, it's unclear that the 9P wire format
+ has a significant advantage over protobufs, which are substantially easier
+ to extend. Note that the official Go protobuf implementation is widely known
+ to suffer from a significant number of performance deficiencies, so if we
+ choose to switch to protobuf, we may need to use an alternative toolchain
+ such as `gogo/protobuf` (which is also widely used in the Go ecosystem, e.g.
+ by Kubernetes).
+
+- Filesystem model: fids, qids, etc. Discarding this is one of the motivations
+ for this proposal.
+
+- RPCs: Twalk, Tlopen, etc. In addition to previously-described
+ inefficiencies, most of these are dependent on the filesystem model and
+ therefore must be discarded.
+
+### FUSE
+
+The FUSE (Filesystem in Userspace) protocol is frequently used to provide
+arbitrary userspace filesystem implementations to a host Linux kernel.
+Unfortunately, FUSE is also inode-based, and therefore doesn't address any of
+the problems we have with 9P.
+
+### virtio-fs
+
+virtio-fs is an ongoing project aimed at improving Linux VM filesystem
+performance when accessing Linux host filesystems (vs. virtio-9p). In brief, it
+is based on:
+
+- Using a FUSE client in the guest that communicates over virtio with a FUSE
+ server in the host.
+
+- Using DAX to map the host page cache into the guest.
+
+- Using a file metadata table in shared memory to avoid VM exits for metadata
+ updates.
+
+None of these improvements seem applicable to gVisor:
+
+- As explained above, FUSE is still inode-based, so it is still susceptible to
+ most of the problems we have with 9P.
+
+- Our use of host file descriptors already allows us to leverage the host page
+ cache for file contents.
+
+- Our need for shared filesystem coherence is usually based on a user
+ requirement that an out-of-sandbox filesystem mutation is guaranteed to be
+ visible by all subsequent observations from within the sandbox, or vice
+ versa; it's not clear that this can be guaranteed without a synchronous
+ signaling mechanism like an RPC.