* [PATCH v3 1/2] fuse: allow FUSE_SYNCFS for privileged userspace servers
2026-08-20 13:01 [PATCH v3 0/2] fuse: allow FUSE_SYNCFS for privileged userspace servers Jimmy Zuber
@ 2026-08-20 13:01 ` Jimmy Zuber
2026-08-20 13:01 ` [PATCH v3 2/2] selftests/fuse: add test for FUSE_HAS_SYNCFS privilege gating Jimmy Zuber
1 sibling, 0 replies; 3+ messages in thread
From: Jimmy Zuber @ 2026-08-20 13:01 UTC (permalink / raw)
To: Miklos Szeredi
Cc: fuse-devel, linux-fsdevel, linux-api, linux-kernel, Shuah Khan,
linux-kselftest
Propagating syncfs()/sync() to a FUSE server via FUSE_SYNCFS lets the
server flush its own cached or intermediate state when userspace asks the
filesystem to sync. This is currently enabled only for virtiofs and
fuseblk, because an untrusted server can use it to stall sync()
indefinitely (see commit 2d82ab251ef0 ("virtiofs: propagate sync() to file
server"), and commit d3906d8f3cee ("fuse: enable FUSE_SYNCFS for all
fuseblk servers")). Both of those mount types require host privilege to
set up, so the server is trusted not to abuse it.
There is nothing virtiofs- or block-specific about wanting to handle
syncfs(), though. A plain /dev/fuse server is just as entitled to
participate in the sync() path -- so that data it has buffered reaches
stable storage when the user asks for it -- provided it is equally
trusted.
The trust property that virtiofs and fuseblk satisfy is that neither can
be mounted without CAP_SYS_ADMIN in the initial user namespace (neither
sets FS_USERNS_MOUNT). A plain fuse mount does set FS_USERNS_MOUNT, so its
existence guarantees no such privilege; the privilege has to be checked
rather than assumed.
Add an opt-in INIT flag, FUSE_HAS_SYNCFS, and honor it only when the
server opened /dev/fuse with CAP_SYS_ADMIN in the initial user namespace,
recorded at mount time via file_ns_capable(). This is the same privilege
that mounting virtiofs or fuseblk requires, applied to the process that
actually services (and could stall) the connection. Checking the device
opener's capability -- rather than the mount's user namespace -- avoids
treating an unprivileged server that merely happens to run in the initial
user namespace (e.g. a normal sshfs mount) as trusted.
The flag is only advertised to servers that pass this check, so an
unprivileged server is never invited to opt in (and is ignored by
fuse_syncfs_enable() if it sets the flag anyway).
Signed-off-by: Jimmy Zuber <jamz@amazon.com>
Assisted-by: Claude:claude-opus-4-8 [Claude-Code]
---
fs/fuse/fuse_i.h | 9 +++++++++
fs/fuse/inode.c | 28 ++++++++++++++++++++++++++++
include/uapi/linux/fuse.h | 12 +++++++++++-
3 files changed, 48 insertions(+), 1 deletion(-)
diff --git a/fs/fuse/fuse_i.h b/fs/fuse/fuse_i.h
index c8d4c5f3af7e..97d356588d00 100644
--- a/fs/fuse/fuse_i.h
+++ b/fs/fuse/fuse_i.h
@@ -391,6 +391,7 @@ struct fuse_fs_context {
bool no_control:1;
bool no_force_umount:1;
bool legacy_opts_show:1;
+ bool syncfs_capable:1;
enum fuse_dax_mode dax_mode;
unsigned int max_read;
unsigned int blksize;
@@ -675,6 +676,14 @@ struct fuse_conn {
/** @sync_fs: Propagate syncfs() to server */
unsigned int sync_fs:1;
+ /**
+ * @syncfs_capable: the privilege required to honor FUSE_HAS_SYNCFS was
+ * present when /dev/fuse was opened (CAP_SYS_ADMIN in the initial user
+ * namespace), i.e. the same privilege that mounting virtiofs/fuseblk
+ * requires.
+ */
+ unsigned int syncfs_capable:1;
+
/** @init_security: Initialize security xattrs when creating a new inode */
unsigned int init_security:1;
diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c
index 1c6ee01c6796..970d4d2e9833 100644
--- a/fs/fuse/inode.c
+++ b/fs/fuse/inode.c
@@ -803,6 +803,15 @@ static int fuse_opt_fd(struct fs_context *fsc, struct file *file)
if (file->f_cred->user_ns != fsc->user_ns)
return invalfc(fsc, "wrong user namespace for fuse device");
+ /*
+ * Record whether the server opened /dev/fuse with CAP_SYS_ADMIN in the
+ * initial user namespace -- the same privilege that mounting virtiofs
+ * or fuseblk requires. Only such servers are trusted to receive
+ * FUSE_SYNCFS (see fuse_syncfs_enable()).
+ */
+ ctx->syncfs_capable = file_ns_capable(file, &init_user_ns,
+ CAP_SYS_ADMIN);
+
ctx->fud = fuse_dev_grab(file);
return 0;
@@ -1269,6 +1278,16 @@ struct fuse_init_args {
struct fuse_mount *fm;
};
+/*
+ * A server can stall syncfs()/sync(), so only honor FUSE_HAS_SYNCFS for
+ * servers that opened /dev/fuse with CAP_SYS_ADMIN in the initial user
+ * namespace -- the same privilege required to mount virtiofs or fuseblk.
+ */
+static bool fuse_syncfs_enable(struct fuse_conn *fc, u64 flags)
+{
+ return (flags & FUSE_HAS_SYNCFS) && fc->syncfs_capable;
+}
+
static void process_init_reply(struct fuse_args *args, int error)
{
struct fuse_init_args *ia = container_of(args, typeof(*ia), args);
@@ -1410,6 +1429,9 @@ static void process_init_reply(struct fuse_args *args, int error)
if (flags & FUSE_REQUEST_TIMEOUT)
timeout = arg->request_timeout;
+
+ if (fuse_syncfs_enable(fc, flags))
+ fc->sync_fs = 1;
} else {
ra_pages = fc->max_read / PAGE_SIZE;
fc->no_lock = 1;
@@ -1479,6 +1501,11 @@ static struct fuse_init_args *fuse_new_init(struct fuse_mount *fm)
flags |= FUSE_SUBMOUNTS;
if (IS_ENABLED(CONFIG_FUSE_PASSTHROUGH))
flags |= FUSE_PASSTHROUGH;
+ /* Only offered to sufficiently privileged servers; see
+ * fuse_syncfs_enable().
+ */
+ if (fm->fc->syncfs_capable)
+ flags |= FUSE_HAS_SYNCFS;
/*
* This is just an information flag for fuse server. No need to check
@@ -1774,6 +1801,7 @@ int fuse_fill_super_common(struct super_block *sb, struct fuse_fs_context *ctx)
fc->default_permissions = ctx->default_permissions;
fc->allow_other = ctx->allow_other;
+ fc->syncfs_capable = ctx->syncfs_capable;
fc->user_id = ctx->user_id;
fc->group_id = ctx->group_id;
fc->legacy_opts_show = ctx->legacy_opts_show;
diff --git a/include/uapi/linux/fuse.h b/include/uapi/linux/fuse.h
index 7435e09c87fe..10a7f31c4bdf 100644
--- a/include/uapi/linux/fuse.h
+++ b/include/uapi/linux/fuse.h
@@ -248,6 +248,9 @@
* - add bufpool offset field to fuse_uring_ent_in_out struct
* - add FUSE_URING_ZERO_COPY, FUSE_URING_ENT_ZERO_COPY, and
* FOPEN_IO_URING_ZERO_COPY flag
+ *
+ * 7.47
+ * - add FUSE_HAS_SYNCFS opt-in flag for privileged userspace servers
*/
#ifndef _LINUX_FUSE_H
@@ -283,7 +286,7 @@
#define FUSE_KERNEL_VERSION 7
/** Minor version number of this interface */
-#define FUSE_KERNEL_MINOR_VERSION 46
+#define FUSE_KERNEL_MINOR_VERSION 47
/** The node ID of the root inode */
#define FUSE_ROOT_ID 1
@@ -464,6 +467,12 @@ struct fuse_file_lock {
* FUSE_REQUEST_TIMEOUT: kernel supports timing out requests.
* init_out.request_timeout contains the timeout (in secs)
* FUSE_HAS_IO_URING_BUFPOOL: kernel supports io-uring buffer pools
+ * FUSE_HAS_SYNCFS: server requests that syncfs()/sync() be propagated as
+ * FUSE_SYNCFS requests. Since an untrusted server can use this
+ * to stall sync(), it is only honored when /dev/fuse was opened
+ * with CAP_SYS_ADMIN in the initial user namespace (the same
+ * privilege that mounting virtiofs or fuseblk requires).
+ * Insufficiently privileged servers ignore it.
*/
#define FUSE_ASYNC_READ (1 << 0)
#define FUSE_POSIX_LOCKS (1 << 1)
@@ -512,6 +521,7 @@ struct fuse_file_lock {
#define FUSE_OVER_IO_URING (1ULL << 41)
#define FUSE_REQUEST_TIMEOUT (1ULL << 42)
#define FUSE_HAS_IO_URING_BUFPOOL (1ULL << 43)
+#define FUSE_HAS_SYNCFS (1ULL << 44)
/**
* CUSE INIT request/reply flags
--
2.50.1
^ permalink raw reply related [flat|nested] 3+ messages in thread* [PATCH v3 2/2] selftests/fuse: add test for FUSE_HAS_SYNCFS privilege gating
2026-08-20 13:01 [PATCH v3 0/2] fuse: allow FUSE_SYNCFS for privileged userspace servers Jimmy Zuber
2026-08-20 13:01 ` [PATCH v3 1/2] " Jimmy Zuber
@ 2026-08-20 13:01 ` Jimmy Zuber
1 sibling, 0 replies; 3+ messages in thread
From: Jimmy Zuber @ 2026-08-20 13:01 UTC (permalink / raw)
To: Miklos Szeredi
Cc: fuse-devel, linux-fsdevel, linux-api, linux-kernel, Shuah Khan,
linux-kselftest
Add a selftest that talks the raw FUSE protocol over /dev/fuse (rather
than via libfuse, which negotiates INIT internally) so it can both choose
whether to advertise FUSE_HAS_SYNCFS and directly observe whether a
FUSE_SYNCFS opcode is forwarded by the kernel.
Three cases are covered:
T1: host-root mount, server sets FUSE_HAS_SYNCFS
-> FUSE_SYNCFS must reach the server.
T2: host-root mount, server does not opt in
-> FUSE_SYNCFS must not be sent (back-compat).
T3: server opts in but opened /dev/fuse without CAP_SYS_ADMIN while still
in the initial user namespace
-> FUSE_SYNCFS must be withheld. This is the case that distinguishes
gating on the server's privilege from gating on the mount's user
namespace.
Signed-off-by: Jimmy Zuber <jamz@amazon.com>
Assisted-by: Claude:claude-opus-4-8 [Claude-Code]
---
.../selftests/filesystems/fuse/.gitignore | 1 +
.../selftests/filesystems/fuse/Makefile | 2 +-
.../selftests/filesystems/fuse/test_syncfs.c | 309 ++++++++++++++++++
3 files changed, 311 insertions(+), 1 deletion(-)
create mode 100644 tools/testing/selftests/filesystems/fuse/test_syncfs.c
diff --git a/tools/testing/selftests/filesystems/fuse/.gitignore b/tools/testing/selftests/filesystems/fuse/.gitignore
index 3e72e742d08e..4e92f363c74a 100644
--- a/tools/testing/selftests/filesystems/fuse/.gitignore
+++ b/tools/testing/selftests/filesystems/fuse/.gitignore
@@ -1,3 +1,4 @@
# SPDX-License-Identifier: GPL-2.0-only
fuse_mnt
fusectl_test
+test_syncfs
diff --git a/tools/testing/selftests/filesystems/fuse/Makefile b/tools/testing/selftests/filesystems/fuse/Makefile
index 612aad69a93a..cbba01635226 100644
--- a/tools/testing/selftests/filesystems/fuse/Makefile
+++ b/tools/testing/selftests/filesystems/fuse/Makefile
@@ -2,7 +2,7 @@
CFLAGS += -Wall -O2 -g $(KHDR_INCLUDES)
-TEST_GEN_PROGS := fusectl_test
+TEST_GEN_PROGS := fusectl_test test_syncfs
TEST_GEN_FILES := fuse_mnt
include ../../lib.mk
diff --git a/tools/testing/selftests/filesystems/fuse/test_syncfs.c b/tools/testing/selftests/filesystems/fuse/test_syncfs.c
new file mode 100644
index 000000000000..d3ca2613b45c
--- /dev/null
+++ b/tools/testing/selftests/filesystems/fuse/test_syncfs.c
@@ -0,0 +1,309 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Test that FUSE_SYNCFS is propagated to a userspace server only when the
+ * server opts in with FUSE_HAS_SYNCFS *and* opened /dev/fuse with
+ * CAP_SYS_ADMIN in the initial user namespace.
+ *
+ * Unlike the libfuse-based selftests, this talks the raw FUSE wire protocol
+ * over /dev/fuse so it can (a) choose whether to advertise FUSE_HAS_SYNCFS in
+ * the INIT reply and (b) observe directly whether a FUSE_SYNCFS opcode arrives.
+ */
+#define _GNU_SOURCE
+#include <errno.h>
+#include <fcntl.h>
+#include <poll.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/eventfd.h>
+#include <sys/mount.h>
+#include <sys/stat.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <sys/uio.h>
+#include <sys/wait.h>
+#include <linux/capability.h>
+#include <linux/fuse.h>
+
+#include "../../kselftest.h"
+
+#define FUSE_ROOT_ID 1
+
+/*
+ * Add or drop CAP_SYS_ADMIN in the effective set via raw capget/capset
+ * (avoids a libcap dependency). Used to construct a server that opens
+ * /dev/fuse without the privilege FUSE_HAS_SYNCFS requires.
+ */
+static int set_sysadmin(int on)
+{
+ struct __user_cap_header_struct hdr = {
+ .version = _LINUX_CAPABILITY_VERSION_3,
+ .pid = 0,
+ };
+ struct __user_cap_data_struct data[_LINUX_CAPABILITY_U32S_3] = {};
+ unsigned int idx = CAP_SYS_ADMIN >> 5;
+ __u32 bit = 1U << (CAP_SYS_ADMIN & 31);
+
+ if (syscall(SYS_capget, &hdr, data))
+ return -1;
+ if (on)
+ data[idx].effective |= bit;
+ else
+ data[idx].effective &= ~bit;
+ return syscall(SYS_capset, &hdr, data);
+}
+
+/*
+ * eventfd the server child writes once when it receives FUSE_SYNCFS; the
+ * parent poll()s it to observe (or rule out) propagation.
+ */
+static int syncfs_evfd;
+
+static void reply(int fd, uint64_t unique, int error, void *data, size_t len)
+{
+ struct fuse_out_header oh = {
+ .len = sizeof(oh) + len,
+ .error = error,
+ .unique = unique,
+ };
+ struct iovec iov[2] = {
+ { &oh, sizeof(oh) },
+ { data, len },
+ };
+
+ if (writev(fd, iov, data ? 2 : 1) < 0)
+ ksft_print_msg("server writev failed: %s\n", strerror(errno));
+}
+
+static void fill_attr(struct fuse_attr *a, uint64_t ino, uint32_t mode,
+ uint32_t nlink)
+{
+ memset(a, 0, sizeof(*a));
+ a->ino = ino;
+ a->mode = mode;
+ a->nlink = nlink;
+ a->blksize = 4096;
+}
+
+/*
+ * Minimal FUSE server. Advertises FUSE_HAS_SYNCFS in its INIT reply iff
+ * @advertise is set. Signals syncfs_evfd when a FUSE_SYNCFS opcode arrives.
+ */
+#define SERVER_MAX_WRITE 65536
+static void run_server(int fd, int advertise)
+{
+ /*
+ * The kernel rejects reads (EINVAL) whose buffer is smaller than
+ * max_write + header, so size generously for the max_write we
+ * advertise in the INIT reply below.
+ */
+ static char buf[SERVER_MAX_WRITE + 4096];
+
+ for (;;) {
+ ssize_t n = read(fd, buf, sizeof(buf));
+ struct fuse_in_header *ih = (void *)buf;
+
+ if (n < 0) {
+ if (errno == EINTR || errno == EAGAIN)
+ continue;
+ return; /* device closed on unmount/abort */
+ }
+ if (n < (ssize_t)sizeof(*ih))
+ continue;
+
+ switch (ih->opcode) {
+ case FUSE_INIT: {
+ struct fuse_init_in *in = (void *)(ih + 1);
+ struct fuse_init_out out = {0};
+ uint64_t flags = FUSE_INIT_EXT;
+
+ out.major = FUSE_KERNEL_VERSION;
+ out.minor = FUSE_KERNEL_MINOR_VERSION;
+ out.max_readahead = in->max_readahead;
+ out.max_write = SERVER_MAX_WRITE;
+ out.max_background = 16;
+ out.congestion_threshold = 12;
+ if (advertise)
+ flags |= FUSE_HAS_SYNCFS;
+ out.flags = flags;
+ out.flags2 = flags >> 32;
+ reply(fd, ih->unique, 0, &out, sizeof(out));
+ break;
+ }
+ case FUSE_GETATTR: {
+ struct fuse_attr_out out = {0};
+
+ fill_attr(&out.attr, FUSE_ROOT_ID, S_IFDIR | 0755, 2);
+ reply(fd, ih->unique, 0, &out, sizeof(out));
+ break;
+ }
+ case FUSE_SYNCFS: {
+ uint64_t one = 1;
+
+ if (write(syncfs_evfd, &one, sizeof(one)) < 0)
+ ksft_print_msg("server eventfd write failed: %s\n",
+ strerror(errno));
+ reply(fd, ih->unique, 0, NULL, 0);
+ break;
+ }
+ default:
+ /*
+ * Anything else (e.g. OPENDIR from opening the mount
+ * root) is not needed to drive this test; -ENOSYS lets
+ * the kernel proceed.
+ */
+ reply(fd, ih->unique, -ENOSYS, NULL, 0);
+ break;
+ }
+ }
+}
+
+/*
+ * Mount a fuse fs backed by a forked server, issue syncfs(), and report
+ * whether the server observed FUSE_SYNCFS. Returns 0 on success, -1 if the
+ * environment could not support the test (caller should skip).
+ *
+ * If @unpriv_open is set, /dev/fuse is opened with CAP_SYS_ADMIN dropped
+ * (regained only for the mount() syscall), so the device opener -- the
+ * server principal FUSE_HAS_SYNCFS is gated on -- lacks the privilege even
+ * though it remains in the initial user namespace.
+ */
+static int do_mount_and_syncfs(const char *mnt, int advertise, int unpriv_open,
+ int *seen)
+{
+ struct pollfd pfd = { .events = POLLIN };
+ char opts[256];
+ int fd, mfd = -1, i;
+ pid_t pid;
+
+ syncfs_evfd = eventfd(0, EFD_CLOEXEC);
+ if (syncfs_evfd < 0)
+ return -1;
+
+ if (unpriv_open && set_sysadmin(0))
+ goto out_evfd;
+ fd = open("/dev/fuse", O_RDWR);
+ if (unpriv_open && set_sysadmin(1)) {
+ if (fd >= 0)
+ close(fd);
+ goto out_evfd;
+ }
+ if (fd < 0)
+ goto out_evfd;
+
+ mkdir(mnt, 0755);
+ snprintf(opts, sizeof(opts),
+ "fd=%d,rootmode=40000,user_id=%d,group_id=%d",
+ fd, getuid(), getgid());
+
+ if (mount("fuse", mnt, "fuse", 0, opts) < 0)
+ goto out_fd;
+
+ pid = fork();
+ if (pid < 0)
+ goto out_umount;
+ if (pid == 0) {
+ run_server(fd, advertise);
+ _exit(0);
+ }
+
+ /*
+ * The parent does not service the fuse fd; the child does. Close our
+ * copy so the kernel sees a single server, and so that if the child
+ * dies the connection aborts instead of hanging us forever.
+ */
+ close(fd);
+
+ /*
+ * mount() returns before the server has answered FUSE_INIT, so the
+ * first open() can race and fail with ENOTCONN; retry until the
+ * handshake settles.
+ */
+ for (i = 0; i < 1000; i++) {
+ mfd = open(mnt, O_RDONLY | O_DIRECTORY);
+ if (mfd >= 0)
+ break;
+ usleep(1000);
+ }
+ if (mfd >= 0) {
+ syncfs(mfd);
+ close(mfd);
+ }
+
+ /*
+ * No waiting is needed: the server writes syncfs_evfd before it replies
+ * to FUSE_SYNCFS, and that reply is what unblocks the synchronous
+ * syncfs() above. So once syncfs() has returned, the eventfd is already
+ * signalled if the opcode was propagated, and will never be otherwise.
+ * poll() with a zero timeout therefore decides both cases immediately.
+ */
+ pfd.fd = syncfs_evfd;
+ *seen = poll(&pfd, 1, 0) > 0 && (pfd.revents & POLLIN);
+
+ kill(pid, SIGKILL);
+ waitpid(pid, NULL, 0);
+ umount2(mnt, MNT_DETACH);
+ close(syncfs_evfd);
+ return 0;
+
+out_umount:
+ umount2(mnt, MNT_DETACH);
+out_fd:
+ close(fd);
+out_evfd:
+ close(syncfs_evfd);
+ return -1;
+}
+
+int main(void)
+{
+ char mnt[] = "/tmp/fuse_syncfs_XXXXXX";
+ int seen, ret;
+
+ ksft_print_header();
+ ksft_set_plan(3);
+
+ /* Hard watchdog: never let a stuck syncfs hang the test runner. */
+ signal(SIGALRM, SIG_DFL);
+ alarm(60);
+
+ if (geteuid() != 0)
+ ksft_exit_skip("test requires root to mount fuse\n");
+
+ if (!mkdtemp(mnt))
+ ksft_exit_fail_msg("mkdtemp failed\n");
+
+ /* T1: host-root mount, server opts in -> syncfs must reach server. */
+ ret = do_mount_and_syncfs(mnt, 1, 0, &seen);
+ if (ret < 0)
+ ksft_test_result_skip("T1: could not mount fuse\n");
+ else
+ ksft_test_result(seen,
+ "T1 host-root + FUSE_HAS_SYNCFS: server receives FUSE_SYNCFS\n");
+
+ /* T2: host-root mount, server does NOT opt in -> no FUSE_SYNCFS. */
+ ret = do_mount_and_syncfs(mnt, 0, 0, &seen);
+ if (ret < 0)
+ ksft_test_result_skip("T2: could not mount fuse\n");
+ else
+ ksft_test_result(!seen,
+ "T2 host-root, no opt-in: server does NOT receive FUSE_SYNCFS\n");
+
+ /*
+ * T3: server opts in but opened /dev/fuse without CAP_SYS_ADMIN while
+ * still in the initial user namespace -> kernel must withhold
+ * FUSE_SYNCFS. This is the case that distinguishes gating on the
+ * server's privilege from gating on the mount's user namespace.
+ */
+ ret = do_mount_and_syncfs(mnt, 1, 1, &seen);
+ if (ret < 0)
+ ksft_test_result_skip("T3: could not mount fuse unprivileged\n");
+ else
+ ksft_test_result(!seen,
+ "T3 init_userns, opener lacks CAP_SYS_ADMIN: FUSE_SYNCFS withheld\n");
+
+ rmdir(mnt);
+ ksft_finished();
+}
--
2.50.1
^ permalink raw reply related [flat|nested] 3+ messages in thread