Harden SQLite blob activation safety

Keep VACUUM snapshots private until permissions and durability are established. Refuse to recreate a missing rollback backup after import has begun, and preserve secondary cleanup failures in startup logs.
This commit is contained in:
Scott Werner 2026-08-24 14:01:34 -04:00
parent f9f19213e6
commit 776e719383
6 changed files with 225 additions and 19 deletions

View file

@ -62,8 +62,10 @@ prefix, checks disk headroom sized to the rows not yet imported (a warm
restart with nothing left to import only needs a small fixed headroom; on
filesystems whose free space cannot be determined the check is skipped with
a warning), imports in bounded transactions, compares every legacy blob
byte-for-byte with SQLite, runs a live SQLite integrity check, and completes a
final WAL checkpoint. Boots that import new rows additionally re-verify every
byte-for-byte with SQLite, runs a live SQLite integrity check, and attempts a
final WAL truncate checkpoint. A busy final truncate logs a warning and startup
continues so a later checkpoint can finish after the blocking reader exits.
Boots that import new rows additionally re-verify every
legacy blob against SQLite and validate every SQLite blob row independently.
Any failure stops startup. Warm boots that import no rows skip that full target
scan: the import pass has already byte-compared every retained legacy row, and
@ -73,10 +75,13 @@ never modified and there is no fallback or dual read/write path.
For a non-empty legacy inventory, the first activation also creates the
private sibling backup
`fabro.sqlite3.pre-blob-activation.bak`. Fabro writes and validates a staging
database before publishing the backup without overwriting an existing file.
`fabro.sqlite3.pre-blob-activation.bak`. Fabro writes the staging database
inside a private same-directory area, applies owner-only permissions, flushes
and validates it, then publishes the backup without overwriting an existing file.
A valid retained backup is revalidated on every warm restart and is preserved
as the original pre-activation safety artifact. It is not a promise that an
as the original pre-activation safety artifact. If any legacy row is already
present in SQLite, a missing retained backup stops startup rather than silently
moving that rollback boundary forward. It is not a promise that an
older binary can safely resume after the activated server has accepted new
work; recovery after that boundary is forward-only. Empty legacy inventories
do not need this backup.

View file

@ -50,6 +50,14 @@ pub(crate) enum BlobActivationError {
BackupNotRegular { path: PathBuf },
#[error("activation backup permissions are not private at {path}")]
BackupNotPrivate { path: PathBuf },
#[error(
"activation backup is missing at {path} while {existing_rows} of {legacy_rows} legacy blob rows are already present in SQLite"
)]
MissingBackupAfterImport {
path: PathBuf,
legacy_rows: u64,
existing_rows: u64,
},
#[error("opening or checking activation backup integrity at {path}")]
BackupIntegrity {
path: PathBuf,
@ -133,6 +141,13 @@ pub(crate) async fn activate_blob_storage(
if backup_exists {
validate_backup(&backup_path).await?;
}
if !backup_exists && inventory.pending_rows < inventory.rows {
return Err(BlobActivationError::MissingBackupAfterImport {
path: backup_path,
legacy_rows: inventory.rows,
existing_rows: inventory.rows - inventory.pending_rows,
});
}
let backup_required = inventory.rows > 0 && !backup_exists;
// The resource sampler treats a path with no matching mount as an
// unsupported-but-benign condition (tmpfs or squashfs roots, network
@ -624,6 +639,51 @@ mod tests {
Ok(())
}
#[tokio::test]
async fn missing_backup_after_prior_import_fails_closed() -> TestResult<()> {
let directory = tempfile::tempdir()?;
let sqlite_path = directory.path().join("fabro.sqlite3");
let database = fabro_db::Database::connect(&sqlite_path).await?;
database.migrate().await?;
let object_store: Arc<dyn ObjectStore> = Arc::new(InMemory::new());
let source = fabro_store::test_support::test_database(
Arc::clone(&object_store),
"missing-backup-test",
Duration::from_millis(1),
None,
);
let bytes = b"already-imported";
let hash = fabro_store::test_support::put_legacy_blob(&source, bytes).await?;
sqlx::query("INSERT INTO blobs (hash, data) VALUES (?, ?)")
.bind(hash.to_string())
.bind(bytes.as_slice())
.execute(database.pool())
.await?;
drop(source);
let error = activate_blob_storage(
&database,
&sqlite_path,
object_store,
"missing-backup-test".to_string(),
Duration::from_millis(1),
None,
)
.await
.expect_err("startup must not move the pre-activation rollback boundary");
assert!(matches!(
error,
BlobActivationError::MissingBackupAfterImport {
legacy_rows: 1,
existing_rows: 1,
..
}
));
assert!(!append_to_path(&sqlite_path, BACKUP_SUFFIX).exists());
Ok(())
}
#[tokio::test]
async fn empty_inventory_skips_backup_and_serves_existing_sqlite_rows() -> TestResult<()> {
let directory = tempfile::tempdir()?;

View file

@ -8,12 +8,13 @@ use std::fmt;
use bytes::Bytes;
use fabro_types::BlobHash;
use fabro_util::error;
use futures::TryStreamExt as _;
use sqlx::pool::PoolConnection;
use sqlx::{Acquire as _, Sqlite, SqlitePool};
#[cfg(test)]
use tokio::sync::Barrier;
use tracing::debug;
use tracing::{debug, error};
use crate::Database;
use crate::keys::SlateKey;
@ -709,7 +710,15 @@ impl Database {
}
Err(failure) => {
debug_import_outcome("failed", &report, Some(failure.kind()));
Err(LegacyBlobImportError { report, failure })
let import_error = LegacyBlobImportError { report, failure };
for cleanup_error in import_error.cleanup_errors() {
let rendered = error::collect_chain(cleanup_error).join(": ");
error!(
error = %rendered,
"Legacy blob import cleanup failed"
);
}
Err(import_error)
}
}
}
@ -1766,10 +1775,12 @@ mod tests {
restore_automatic_checkpoint: true,
..ImportControls::default()
};
let capture = CapturedEvents::default();
let error = context
.source
.import_legacy_blobs_with_controls(&context.sqlite, &controls)
.with_subscriber(capture.clone())
.await
.expect_err("both injected failures should fail import");
@ -1786,6 +1797,15 @@ mod tests {
.any(|source| source.downcast_ref::<sqlx::Error>().is_some()),
"restoration source was absent from cleanup errors"
);
let events = capture.events().join("\n");
assert!(
events.contains("Legacy blob import cleanup failed"),
"captured: {events}"
);
assert!(
events.contains("injected automatic checkpoint restoration failure"),
"captured: {events}"
);
Ok(())
}

View file

@ -6,9 +6,9 @@ use fabro_types::{BlobHash, RunId};
use object_store::ObjectStore;
use sqlx::sqlite::{SqliteConnectOptions, SqlitePoolOptions};
use crate::keys::SlateKey;
#[cfg(test)]
use crate::{AuthSessionStore, RunSummaryStore};
use crate::keys::SlateKey;
use crate::{BlobStore, Database, Result};
/// Returns an isolated SQLite blob authority backed by its own in-memory

View file

@ -16,10 +16,10 @@ workspace = true
anyhow.workspace = true
chrono.workspace = true
sqlx.workspace = true
tempfile = "3"
thiserror.workspace = true
tokio.workspace = true
tracing.workspace = true
[dev-dependencies]
tempfile = "3"
tokio = { workspace = true, features = ["macros"] }

View file

@ -227,6 +227,12 @@ pub enum SnapshotStagingError {
#[source]
source: std::io::Error,
},
#[error("creating private snapshot staging area for {path}")]
CreatePrivateStagingArea {
path: PathBuf,
#[source]
source: std::io::Error,
},
#[error("snapshot staging path is not valid UTF-8 at {path}")]
NonUtf8Path { path: PathBuf },
#[error("writing SQLite snapshot {path}")]
@ -247,26 +253,57 @@ pub enum SnapshotStagingError {
#[source]
source: std::io::Error,
},
#[error("publishing private snapshot staging file {path}")]
Publish {
path: PathBuf,
#[source]
source: std::io::Error,
},
}
/// Writes a consistent single-file copy of the live pool to `staging_path`.
///
/// `VACUUM INTO` produces a snapshot that needs no `-wal`/`-shm` siblings to
/// restore. Any stale staging file is removed first and the copy is
/// restricted to private permissions. The caller publishes the staging file
/// into its final path and owns the durability of that rename.
/// restore. The copy is first written inside a private same-directory staging
/// area, then restricted to owner-only permissions and flushed before it is
/// exposed at `staging_path`. The caller publishes the staging file into its
/// final path and owns the durability of that rename.
pub async fn write_snapshot_to_staging(
pool: &DbPool,
staging_path: &Path,
) -> Result<(), SnapshotStagingError> {
write_snapshot_to_staging_inner(pool, staging_path, |_| {}).await
}
async fn write_snapshot_to_staging_inner<F>(
pool: &DbPool,
staging_path: &Path,
after_write: F,
) -> Result<(), SnapshotStagingError>
where
F: FnOnce(&Path),
{
remove_file_if_exists(staging_path)
.await
.map_err(|source| SnapshotStagingError::RemoveStale {
path: staging_path.to_path_buf(),
source,
})?;
let staging_parent = nonempty_parent(staging_path);
// SQLite's VACUUM INTO creates its destination with umask-derived
// permissions. Keep that file behind an owner-only directory until its
// own mode is restricted, so a traversable database directory never
// exposes a partially written snapshot.
let private_staging_area = create_private_staging_area(staging_parent).map_err(|source| {
SnapshotStagingError::CreatePrivateStagingArea {
path: staging_path.to_path_buf(),
source,
}
})?;
let private_staging_path = private_staging_area.path().join("snapshot.sqlite3");
let staging_target =
staging_path
private_staging_path
.to_str()
.ok_or_else(|| SnapshotStagingError::NonUtf8Path {
path: staging_path.to_path_buf(),
@ -279,7 +316,8 @@ pub async fn write_snapshot_to_staging(
path: staging_path.to_path_buf(),
source,
})?;
set_private_permissions(staging_path)
after_write(&private_staging_path);
set_private_permissions(&private_staging_path)
.await
.map_err(|source| SnapshotStagingError::SetPermissions {
path: staging_path.to_path_buf(),
@ -287,7 +325,7 @@ pub async fn write_snapshot_to_staging(
})?;
// The staging file must be durable before the caller renames it into a
// path that later recovery logic treats as a complete snapshot.
let sync_result = match fs::File::open(staging_path).await {
let sync_result = match fs::File::open(&private_staging_path).await {
Ok(file) => file.sync_all().await,
Err(source) => Err(source),
};
@ -295,9 +333,33 @@ pub async fn write_snapshot_to_staging(
path: staging_path.to_path_buf(),
source,
})?;
fs::rename(&private_staging_path, staging_path)
.await
.map_err(|source| SnapshotStagingError::Publish {
path: staging_path.to_path_buf(),
source,
})?;
Ok(())
}
fn create_private_staging_area(parent: &Path) -> std::io::Result<tempfile::TempDir> {
let mut builder = tempfile::Builder::new();
builder.prefix(".fabro-snapshot-");
#[cfg(unix)]
{
use std::os::unix::fs::PermissionsExt as _;
builder.permissions(std::fs::Permissions::from_mode(0o700));
}
builder.tempdir_in(parent)
}
fn nonempty_parent(path: &Path) -> &Path {
path.parent()
.filter(|parent| !parent.as_os_str().is_empty())
.unwrap_or_else(|| Path::new("."))
}
/// Flushes the directory entry metadata for `path`'s parent so a rename into
/// that directory survives power loss. No-op off Unix, where a directory
/// cannot be opened for syncing.
@ -307,10 +369,7 @@ pub async fn write_snapshot_to_staging(
reason = "directory fds have no async open; callers run this on a blocking thread"
)]
pub fn sync_parent_directory(path: &Path) -> std::io::Result<()> {
let Some(parent) = path.parent() else {
return Ok(());
};
std::fs::File::open(parent)?.sync_all()
std::fs::File::open(nonempty_parent(path))?.sync_all()
}
/// Flushes the directory entry metadata for `path`'s parent so a rename into
@ -378,3 +437,65 @@ async fn prepare_private_database_file(path: &Path) -> anyhow::Result<()> {
.with_context(|| format!("creating SQLite database {}", path.display()))?;
Ok(())
}
#[cfg(all(test, unix))]
mod tests {
use std::os::unix::fs::PermissionsExt as _;
use std::sync::atomic::{AtomicBool, Ordering};
use super::*;
#[tokio::test]
async fn snapshot_stays_hidden_until_it_has_private_permissions() -> anyhow::Result<()> {
let root = tempfile::tempdir()?;
let database_directory = root.path().join("traversable-db");
fs::create_dir(&database_directory).await?;
fs::set_permissions(&database_directory, std::fs::Permissions::from_mode(0o755)).await?;
let database = Database::connect(database_directory.join("fabro.sqlite3")).await?;
sqlx::query("CREATE TABLE snapshot_secret (value TEXT NOT NULL)")
.execute(database.pool())
.await?;
sqlx::query("INSERT INTO snapshot_secret (value) VALUES ('kept')")
.execute(database.pool())
.await?;
let staging_path = database_directory.join("snapshot.tmp");
let observed_private_stage = AtomicBool::new(false);
write_snapshot_to_staging_inner(database.pool(), &staging_path, |private_path| {
assert!(
!staging_path.exists(),
"the traversable parent must not expose the snapshot before chmod"
);
let private_directory = private_path
.parent()
.expect("private staging file should have a parent");
let mode = std::fs::metadata(private_directory)
.expect("private staging directory should exist")
.permissions()
.mode()
& 0o777;
assert_eq!(mode, 0o700, "temporary staging directory must be private");
observed_private_stage.store(true, Ordering::Relaxed);
})
.await?;
assert!(observed_private_stage.load(Ordering::Relaxed));
let mode = fs::metadata(&staging_path).await?.permissions().mode() & 0o777;
assert_eq!(mode, 0o600, "published staging file must be private");
let snapshot = SqlitePoolOptions::new()
.max_connections(1)
.connect_with(
SqliteConnectOptions::new()
.filename(&staging_path)
.read_only(true),
)
.await?;
let value: String = sqlx::query_scalar("SELECT value FROM snapshot_secret")
.fetch_one(&snapshot)
.await?;
assert_eq!(value, "kept");
snapshot.close().await;
Ok(())
}
}