fix(update): never apply unverified staging — close the OTA truncated-binary race

Root cause of .198 bricking on the 1.7.103 OTA: two concurrent
update.download RPCs shared one staging file, cancel_download wiped
staging mid-flight, a third download began re-filling it, and
apply_update mv'd the 3-second-old 17MB partial of the 49MB binary
into /usr/local/bin -> SEGV boot loop (236 restarts, no rollback).

- Single-flight UPDATE_OP_LOCK across download/apply; concurrent
  callers get an explicit 'already running' error.
- apply_update now requires the .download-complete marker AND
  re-verifies every staged component (size + SHA-256 + BLAKE3)
  against the manifest before touching the system.
- cancel_download only wipes staging when no operation holds the
  lock; otherwise it just flags the in-flight loop to bail.
- Fixed the 'file already complete' path in
  download_component_resumable, which skipped verification and fell
  through to the bogus 'download failed without a captured error'.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
archipelago 2026-07-18 19:33:52 -04:00
parent 573b469191
commit 837cfdfd1f

View File

@ -24,6 +24,16 @@ pub static DOWNLOAD_CANCEL: AtomicBool = AtomicBool::new(false);
/// confidence than "looks stuck at 0%".
pub static DOWNLOAD_PROGRESS_AT: AtomicU64 = AtomicU64::new(0);
/// Serializes the mutating update operations (download, apply, and the
/// staging wipe in cancel). The .198 v1.7.103 bricking (2026-07-18) was
/// exactly this race: two concurrent `update.download` RPCs shared one
/// staging file, a cancel wiped staging mid-flight, a third download began
/// re-filling it, and `apply_update` mv'd the 3-second-old 17MB partial of
/// a 49MB binary into /usr/local/bin → SEGV boot loop. Writers take this
/// via `try_lock` so a concurrent caller gets an explicit "already running"
/// error instead of silently interleaving.
static UPDATE_OP_LOCK: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());
fn now_ms() -> u64 {
use std::time::{SystemTime, UNIX_EPOCH};
SystemTime::now()
@ -976,6 +986,9 @@ pub async fn dismiss_update(data_dir: &Path) -> Result<()> {
/// verified over the complete file at the end of each component, so a
/// partially-corrupt resume still fails cleanly.
pub async fn download_update(data_dir: &Path) -> Result<DownloadProgress> {
let _op = UPDATE_OP_LOCK.try_lock().map_err(|_| {
anyhow::anyhow!("another update operation (download or apply) is already running")
})?;
let mut state = load_state(data_dir).await?;
if state.available_update.is_none() {
state = check_for_updates(data_dir).await?;
@ -1133,7 +1146,6 @@ async fn download_component_resumable(
dest: &Path,
prior_total: u64,
) -> Result<()> {
use sha2::{Digest, Sha256};
use tokio::io::AsyncWriteExt;
const MAX_ATTEMPTS: u32 = 6;
const BACKOFFS: [u64; 5] = [5, 15, 30, 60, 120];
@ -1145,8 +1157,19 @@ async fn download_component_resumable(
Err(_) => 0,
};
if existing_len >= component.size_bytes {
// File is already complete — break out and go verify.
break;
// File is already complete (a resumed run finished it, or a
// leftover from an earlier attempt) — verify it instead of
// trusting it. The old code `break`d here, which skipped
// verification entirely AND landed on the error return below
// ("download failed without a captured error").
match verify_component_on_disk(component, dest).await {
Ok(()) => return Ok(()),
Err(e) => {
let _ = tokio::fs::remove_file(dest).await;
last_err = Some(e);
continue;
}
}
}
if attempt > 1 {
let delay = BACKOFFS[(attempt as usize - 2).min(BACKOFFS.len() - 1)];
@ -1294,44 +1317,83 @@ async fn download_component_resumable(
continue;
}
// Full file — verify hash.
let bytes = tokio::fs::read(dest)
.await
.context("read staging file for hash check")?;
let hash = hex::encode(Sha256::digest(&bytes));
if hash == component.sha256 {
// DHT Phase 1: if the manifest also pins a BLAKE3 digest, it must
// match too. SHA-256 stays the mandatory gate during migration;
// BLAKE3 is the hash the iroh swarm will fetch/verify by, so a
// present-but-wrong BLAKE3 means the bytes aren't swarm-consistent
// — treat it like a SHA mismatch and re-download.
if let Some(b3) = component.blake3.as_deref() {
let expected = b3.trim().strip_prefix("blake3:").unwrap_or(b3.trim());
let actual = crate::content_hash::blake3_hex(&bytes);
if !actual.eq_ignore_ascii_case(expected) {
let _ = tokio::fs::remove_file(dest).await;
last_err = Some(anyhow::anyhow!(
"BLAKE3 mismatch for {}: expected {}, got {}",
component.name,
expected,
actual
));
continue;
}
// Full file — verify hashes. On mismatch the file on disk is
// garbage: nuke it and start over from scratch on the next attempt.
match verify_component_on_disk(component, dest).await {
Ok(()) => return Ok(()),
Err(e) => {
let _ = tokio::fs::remove_file(dest).await;
last_err = Some(e);
}
return Ok(());
}
// SHA mismatch — the file on disk is garbage. Nuke it and
// start over from scratch on the next attempt.
let _ = tokio::fs::remove_file(dest).await;
last_err = Some(anyhow::anyhow!(
}
Err(last_err.unwrap_or_else(|| anyhow::anyhow!("download failed without a captured error")))
}
/// Verify a fully-downloaded component file on disk: SHA-256 is the
/// mandatory gate; when the manifest also pins a BLAKE3 digest it must
/// match too (BLAKE3 is the hash the iroh swarm fetches/verifies by, so
/// a present-but-wrong BLAKE3 means the bytes aren't swarm-consistent —
/// treated exactly like a SHA mismatch). Err = mismatch; the caller
/// decides whether to remove the file and retry.
async fn verify_component_on_disk(component: &ComponentUpdate, dest: &Path) -> Result<()> {
use sha2::{Digest, Sha256};
let bytes = tokio::fs::read(dest)
.await
.context("read staging file for hash check")?;
let hash = hex::encode(Sha256::digest(&bytes));
if hash != component.sha256 {
anyhow::bail!(
"SHA256 mismatch for {}: expected {}, got {}",
component.name,
component.sha256,
hash
));
);
}
Err(last_err.unwrap_or_else(|| anyhow::anyhow!("download failed without a captured error")))
if let Some(b3) = component.blake3.as_deref() {
let expected = b3.trim().strip_prefix("blake3:").unwrap_or(b3.trim());
let actual = crate::content_hash::blake3_hex(&bytes);
if !actual.eq_ignore_ascii_case(expected) {
anyhow::bail!(
"BLAKE3 mismatch for {}: expected {}, got {}",
component.name,
expected,
actual
);
}
}
Ok(())
}
/// Re-verify every manifest component against the bytes actually sitting
/// in staging, immediately before install. The download path verifies as
/// it goes, but staging can change between download and apply — on .198
/// (v1.7.103, 2026-07-18) a concurrent download was re-filling a wiped
/// staging dir when apply ran, and a 17MB partial of the 49MB binary got
/// installed. This apply-time gate is the one that must never be skipped.
async fn verify_staged_components(staging_dir: &Path, manifest: &UpdateManifest) -> Result<()> {
for component in &manifest.components {
let dest = staging_dir.join(&component.name);
let len = tokio::fs::metadata(&dest).await.map(|m| m.len()).unwrap_or(0);
if len != component.size_bytes {
anyhow::bail!(
"staged component {} is {} bytes but the manifest says {} — \
refusing to apply (incomplete or concurrently-rewritten download)",
component.name,
len,
component.size_bytes
);
}
verify_component_on_disk(component, &dest)
.await
.with_context(|| {
format!(
"staged component {} failed verification — refusing to apply",
component.name
)
})?;
}
Ok(())
}
/// Cancel an in-flight download. Sets the cancellation flag so the
@ -1343,11 +1405,21 @@ pub async fn cancel_download(data_dir: &Path) -> Result<()> {
DOWNLOAD_CANCEL.store(true, Ordering::Relaxed);
DOWNLOAD_BYTES.store(0, Ordering::Relaxed);
DOWNLOAD_TOTAL.store(0, Ordering::Relaxed);
// Only wipe staging when no download/apply holds the op lock. Wiping
// under a live operation is how .198 ended up applying a re-filling
// staging dir; with the lock held elsewhere we just set the cancel
// flag and let the in-flight loop bail at its next chunk boundary
// (partials are size+hash revalidated on the next resume anyway).
let staging = data_dir.join("update-staging");
let wiped = if staging.exists() {
tokio::fs::remove_dir_all(&staging).await.is_ok()
} else {
false
let wiped = match UPDATE_OP_LOCK.try_lock() {
Ok(_op) => {
if staging.exists() {
tokio::fs::remove_dir_all(&staging).await.is_ok()
} else {
false
}
}
Err(_) => false,
};
// Clear the "downloaded, ready to apply" marker too — a canceled
// download is not a staged update.
@ -1398,11 +1470,34 @@ pub(crate) async fn host_sudo(args: &[&str]) -> Result<std::process::ExitStatus>
/// Apply a downloaded update. Backs up current binaries, replaces with staged versions.
pub async fn apply_update(data_dir: &Path) -> Result<()> {
let _op = UPDATE_OP_LOCK.try_lock().map_err(|_| {
anyhow::anyhow!("another update operation (download or apply) is already running")
})?;
let staging_dir = data_dir.join("update-staging");
if !staging_dir.exists() {
anyhow::bail!("No staged update found. Download first.");
}
// Gate 1: the completion marker is written only after EVERY component
// downloaded and hash-verified. A staging dir without it is a partial
// or in-flight download — exactly what got installed on .198.
if !has_staged_update(data_dir).await {
anyhow::bail!(
"Staged update is incomplete (no completion marker) — download the update again before applying"
);
}
// Gate 2: re-verify the actual staged bytes against the manifest.
let manifest = load_state(data_dir)
.await?
.available_update
.ok_or_else(|| {
anyhow::anyhow!(
"no update manifest in state to verify staged files against — re-download the update"
)
})?;
verify_staged_components(&staging_dir, &manifest).await?;
let backup_dir = data_dir.join("update-backup");
fs::create_dir_all(&backup_dir)
.await
@ -2443,6 +2538,65 @@ mod tests {
assert!(!persisted.update_in_progress);
}
#[tokio::test]
async fn test_apply_refuses_unmarked_staging() {
// Regression: .198 v1.7.103 bricking — apply ran against a staging
// dir that a concurrent download was still filling. Without the
// .download-complete marker, apply must refuse before touching
// anything.
let dir = tempfile::tempdir().unwrap();
let staging = dir.path().join("update-staging");
tokio::fs::create_dir_all(&staging).await.unwrap();
tokio::fs::write(staging.join("archipelago"), b"partial")
.await
.unwrap();
let err = apply_update(dir.path()).await.unwrap_err();
assert!(
err.to_string().contains("completion marker"),
"got: {err:#}"
);
}
#[tokio::test]
async fn test_apply_refuses_staged_bytes_that_mismatch_manifest() {
// Marker present (a complete download once existed) but the staged
// bytes no longer match the manifest — apply must re-verify and
// refuse rather than install whatever is on disk.
let dir = tempfile::tempdir().unwrap();
let staging = dir.path().join("update-staging");
tokio::fs::create_dir_all(&staging).await.unwrap();
tokio::fs::write(staging.join(STAGED_COMPLETE_MARKER), b"1")
.await
.unwrap();
tokio::fs::write(staging.join("archipelago"), b"truncated-garbage")
.await
.unwrap();
let state = UpdateState {
available_update: Some(UpdateManifest {
version: "999.0.0".to_string(),
release_date: "2026-07-18".to_string(),
changelog: vec![],
components: vec![ComponentUpdate {
name: "archipelago".to_string(),
current_version: "1.0.0".to_string(),
new_version: "999.0.0".to_string(),
download_url: "http://example.invalid/archipelago".to_string(),
sha256: "0".repeat(64),
size_bytes: 49_949_048,
blake3: None,
}],
}),
update_in_progress: true,
..UpdateState::default()
};
save_state(dir.path(), &state).await.unwrap();
let err = apply_update(dir.path()).await.unwrap_err();
assert!(
err.to_string().contains("refusing to apply"),
"got: {err:#}"
);
}
#[tokio::test]
async fn test_dismiss_update_clears_available() {
let dir = tempfile::tempdir().unwrap();