VirtioFS: statx() returns ENOENT for a just-created path when STATX_BTIME is requested

[lparker-rti]

• I have tried with the latest version of Docker Desktop
• I have tried disabling enabled experimental features
• I have uploaded Diagnostics
• Diagnostics ID: none — I'm not able to share a bundle from this machine. The repro below is self-contained and reproduces on a stock python:3-slim image against any bind-mounted host directory, so it shouldn't need one. Happy to run specific commands and report the output instead.

Expected behavior

After mkdir(path) returns, statx(path, STATX_BTIME) should describe that path, as statx(path, STATX_BASIC_STATS) and stat(path) both already do.

Actual behavior

On a VirtioFS bind mount, statx() returns ENOENT for a just-created path whenever STATX_BTIME (0x800) is in the requested mask. Removing that one bit makes the identical call succeed on the identical path. The entry is really on disk throughout — readdir lists it, and open/access succeed.

It is create-scoped and time-bounded, not a permanent negative cache: the same masked call succeeds once the path is a few milliseconds old.

arch=aarch64  __NR_statx=291
mkdir(), then statx() the same path:
  mask=STATX_BASIC_STATS (0x7ff)                0/20 ENOENT
  mask=STATX_BTIME       (0x800)               20/20 ENOENT
  mask=basic|btime       (0xfff)               20/20 ENOENT
  mask=STATX_BTIME, after 5ms sleep             0/20 ENOENT
os.stat() (newfstatat, no btime) for comparison:
  os.stat()                                     0/20 ENOENT

Reproduce

Stock python:3-slim, no Node, no project code. Any bind-mounted host directory:

mkdir -p /tmp/statxbug && cd /tmp/statxbug
cat > probe.py <<'EOF'
import ctypes, ctypes.util, os, platform, time

NR = {"aarch64": 291, "x86_64": 332}[platform.machine()]
AT_FDCWD, ENOENT = -100, 2
BASIC, BTIME = 0x7FF, 0x800

libc = ctypes.CDLL(ctypes.util.find_library("c"), use_errno=True)
buf = ctypes.create_string_buffer(256)

def statx(path, mask):
    ctypes.set_errno(0)
    rc = libc.syscall(
        ctypes.c_long(NR), ctypes.c_int(AT_FDCWD), ctypes.c_char_p(path.encode()),
        ctypes.c_int(0), ctypes.c_uint(mask), ctypes.byref(buf),
    )
    if rc != 0:
        raise OSError(ctypes.get_errno(), "statx")

def trial(label, mask, delay=0.0):
    fail = 0
    for i in range(20):
        p = "/vol/probe-%x-%d" % (mask, i)
        os.mkdir(p)
        if delay:
            time.sleep(delay)
        try:
            statx(p, mask)
        except OSError as e:
            fail += e.errno == ENOENT
        os.rmdir(p)
    print("  %-44s %2d/20 ENOENT" % (label, fail))

print("arch=%s  __NR_statx=%d" % (platform.machine(), NR))
print("mkdir(), then statx() the same path:")
trial("mask=STATX_BASIC_STATS (0x7ff)", BASIC)
trial("mask=STATX_BTIME       (0x800)", BTIME)
trial("mask=basic|btime       (0xfff)", BASIC | BTIME)
trial("mask=STATX_BTIME, after 5ms sleep", BTIME, 0.005)
print("os.stat() (newfstatat, no btime) for comparison:")
fail = 0
for i in range(20):
    p = "/vol/plain-%d" % i
    os.mkdir(p)
    try:
        os.stat(p)
    except FileNotFoundError:
        fail += 1
    os.rmdir(p)
print("  %-44s %2d/20 ENOENT" % ("os.stat()", fail))
EOF
docker run --rm -v "$PWD:/vol" python:3-slim python3 /vol/probe.py

The mount under test:

3515 3506 0:44 /private/tmp/statxbug /vol rw,nosuid,nodev,relatime - virtiofs host rw,ignore_atime,no_xattr

Why this reaches ordinary users

The kernel picks the FUSE opcode from the requested mask. In fuse_update_get_attr:

if (!fc->no_statx && (request_mask & ~STATX_BASIC_STATS))
        err = fuse_do_statx(idmap, inode, file, stat);   /* FUSE_STATX */
else
        err = fuse_do_getattr(idmap, inode, stat, file); /* FUSE_GETATTR */

So asking for btime is exactly what routes a request to FUSE_STATX, which is the opcode that answers ENOENT here. Everything that stays within STATX_BASIC_STATS takes FUSE_GETATTR and is fine.

libuv requests btime, so every fs.stat() in Node on a VirtioFS bind hits this path. In one Node process on one mount, fresh unique path per iteration:

Node call	result
statSync / lstatSync	10/10 ENOENT
accessSync, existsSync	0/10
readdirSync, opendirSync	0/10
openSync, readFileSync	0/10

Downstream this reads as tooling being broken rather than as a filesystem fault. Two we hit: Storybook's mkdirp stats a directory it just created and dies with ENOTDIR: not a directory, mkdir …/node_modules/.cache/storybook/…/public, and Vitest writes .vite-temp/vitest.config.ts.timestamp-*.mjs and then has the ESM resolver stat it, dying with ERR_MODULE_NOT_FOUND before a single test runs. Both look like broken configs and are not.

A shell mkdir d && stat d does not reproduce it, because process spawn costs more than the window — which makes this awkward to catch by hand.

Information

• Reproduces on every run, 20/20, on a fresh unique path each iteration.
• Recovery time varies with system load — a repeat run showed 4/20 rather than 0/20 for the 5 ms case. The mask result itself is stable across every run: 0/20 for 0x7ff, 20/20 for 0x800.
• Not observed on an ext4 named volume or on the container overlay — 0/20 on both, same probe.
• No workaround from userland that we can see: Node exposes no way to ask libuv to drop btime.
• Possibly the undiagnosed cause behind pnpm fails with spurious ENOENT in shared volume using gRPC FUSE or VirtioFS #6787, though the open/copyfile failures reported there do not reproduce for us, and that issue predates FUSE statx support (Linux 6.6), so it may well be unrelated.

docker version

Client:
 Version:           29.7.2
 API version:       1.55
 Go version:        go1.26.5
 Git commit:        a7dcaa6
 Built:             Wed Aug  5 18:27:50 2026
 OS/Arch:           darwin/arm64
 Context:           desktop-linux

Server: Docker Desktop 4.87.0 (236836)
 Engine:
  Version:          29.7.2
  API version:      1.55 (minimum version 1.40)
  Go version:       go1.26.5
  Git commit:       6a43e3d
  Built:            Wed Aug  5 18:28:35 2026
  OS/Arch:          linux/arm64
  Experimental:     false
 containerd:
  Version:          v2.2.5
  GitCommit:        e53c7c1516c3b2bff98eb76f1f4117477e6f4e66
 runc:
  Version:          1.3.6
  GitCommit:        v1.3.6-0-g491b69ba
 docker-init:
  Version:          0.19.0
  GitCommit:        de40ad0

Host

• macOS 26.5.2 (25F84), arm64
• Guest kernel 7.0.12-linuxkit #1 SMP PREEMPT Wed Aug 12 20:18:49 UTC 2026 aarch64
• File sharing implementation: VirtioFS