Skip to main content

ic_backup/ops/persistence/download_journal/ic_snapshot_artifact/
mod.rs

1//! Bounded local snapshot streaming joined to the existing durable download owner.
2
3mod upload;
4mod verification;
5pub use upload::IcSnapshotUploadArtifactError;
6
7use super::{DownloadJournalError, DownloadJournalGuard};
8use crate::{
9    hash::hex_bytes,
10    model::{
11        artifacts::{ArtifactChecksumRecord, ChecksumError},
12        download_journal::ArtifactStateRecord,
13        ic_snapshot_coverage::{IcSnapshotDataCoverage, IcSnapshotDataCoverageError},
14        ic_snapshot_data::IcSnapshotDataReply,
15        ic_snapshot_metadata::{IcSnapshotMetadataReply, MAX_IC_SNAPSHOT_METADATA_BYTES},
16    },
17    ops::{
18        artifacts::{ArtifactError, checksum_directory, checksum_relative_files},
19        persistence::write_json_durable,
20    },
21};
22use ic_management_canister_types::SnapshotDataKind;
23use rustix::fs::{self as unix_fs, AtFlags, Dir, FileType, Mode, OFlags};
24use sha2::{Digest, Sha256};
25use std::{
26    collections::BTreeSet,
27    fmt,
28    fs::{self, File},
29    io::{self, Write},
30    os::unix::fs::MetadataExt,
31    path::{Path, PathBuf},
32};
33use thiserror::Error;
34
35const FORMAT: &[u8] = b"ic-backup/ic-snapshot-artifact/v1\n";
36const REGIONS: [&str; 3] = ["wasm-module.bin", "wasm-memory.bin", "stable-memory.bin"];
37
38/// Exclusive ephemeral writer for one original journal artifact and exact IC metadata.
39///
40/// Each successful append retains decoded bytes in private descriptor-opened files.
41/// It reuses the model's coverage owner and keeps only three hash states plus at most
42/// 1,024 chunk checksums. It borrows the original layout/journal for its whole lifetime.
43/// Appends consume the writer: any rejection or partial IO failure closes it and retains
44/// the partial tree, without a completion transition or permission to repeat reads.
45/// There is no partial-transfer reconstruction or automatic cleanup.
46///
47/// The integration must qualify the generic token's association with the raw IC ID,
48/// authentic capture/metadata/data, original per-call spending, fresh read permission,
49/// complete backend transfer and stable noncooperating byte/command custody. This writer
50/// performs no provider calls and grants no upload/load/start, settlement or release.
51pub struct IcSnapshotArtifactWriter<'journal, 'layout, 'metadata> {
52    journal: &'journal mut DownloadJournalGuard<'layout>,
53    coverage: IcSnapshotDataCoverage<'metadata>,
54    snapshot: String,
55    path: PathBuf,
56    parent: File,
57    directory: File,
58    regions: [File; 3],
59    hashes: [Sha256; 3],
60    checksums: Vec<(PathBuf, ArtifactChecksumRecord)>,
61}
62
63impl<'layout> DownloadJournalGuard<'layout> {
64    /// Create private staging for an existing Created artifact and exact decoded metadata.
65    ///
66    /// The exact generic backend token remains unchanged and is not parsed as a raw IC ID.
67    /// The integration supplies their authoritative association. Target and timestamp must
68    /// match the retained entry; its independently observed total snapshot size is preserved,
69    /// never inferred from a sum of metadata regions. Exact original metadata wire bytes and
70    /// request arguments are retained alongside fixed region/chunk files in a distinct v1
71    /// artifact layout. Existing staging/canonical destinations reject without adoption.
72    /// # Errors
73    /// Rejects wrong identity/state/raw metadata, occupied or unsafe paths and IO failure.
74    /// Failure retains partial bytes and leaves the original journal unchanged.
75    pub fn stage_ic_snapshot_artifact<'journal, 'metadata>(
76        &'journal mut self,
77        snapshot: &str,
78        metadata: &'metadata IcSnapshotMetadataReply<'metadata>,
79        raw_metadata: &[u8],
80    ) -> Result<IcSnapshotArtifactWriter<'journal, 'layout, 'metadata>, IcSnapshotArtifactError>
81    {
82        self.check_usable()?;
83        let entry = self
84            .record
85            .artifact(metadata.request().target(), snapshot)
86            .map_err(DownloadJournalError::from)?;
87        if entry.state() != ArtifactStateRecord::Created
88            || entry.snapshot_taken_at_timestamp() != metadata.metadata().taken_at_timestamp
89            || raw_metadata.len() > MAX_IC_SNAPSHOT_METADATA_BYTES
90            || ArtifactChecksumRecord::from_bytes(raw_metadata) != *metadata.payload_checksum()
91        {
92            return Err(IcSnapshotArtifactError::OriginalMismatch);
93        }
94        self.check_artifact_parent()?;
95        let path = self.layout.root().join(entry.staging_path());
96        let canonical = self.layout.root().join(entry.artifact_path());
97        match fs::symlink_metadata(canonical) {
98            Err(error) if error.kind() == io::ErrorKind::NotFound => {}
99            Err(error) => return Err(error.into()),
100            Ok(_) => return Err(io::Error::from(io::ErrorKind::AlreadyExists).into()),
101        }
102        let parent = open_directory(&self.layout.root().join("artifacts"))?;
103        let name = path
104            .file_name()
105            .ok_or(IcSnapshotArtifactError::CustodyChanged)?;
106        unix_fs::mkdirat(&parent, name, Mode::from_bits_truncate(0o700))
107            .map_err(io::Error::from)?;
108        let directory = File::from(
109            unix_fs::openat(
110                &parent,
111                name,
112                OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
113                Mode::empty(),
114            )
115            .map_err(io::Error::from)?,
116        );
117        let mut checksums = Vec::new();
118        for (name, bytes) in [
119            ("format", FORMAT),
120            ("metadata.candid", raw_metadata),
121            ("metadata-arguments.candid", metadata.request().arguments()),
122        ] {
123            create_file(&directory, name)?.write_all(bytes)?;
124            checksums.push((name.into(), ArtifactChecksumRecord::from_bytes(bytes)));
125        }
126        let regions = [
127            create_file(&directory, REGIONS[0])?,
128            create_file(&directory, REGIONS[1])?,
129            create_file(&directory, REGIONS[2])?,
130        ];
131        let writer = IcSnapshotArtifactWriter {
132            journal: self,
133            coverage: IcSnapshotDataCoverage::new(metadata),
134            snapshot: snapshot.to_owned(),
135            path,
136            parent,
137            directory,
138            regions,
139            hashes: std::array::from_fn(|_| Sha256::new()),
140            checksums,
141        };
142        writer.check_custody()?;
143        Ok(writer)
144    }
145}
146
147impl<'layout> IcSnapshotArtifactWriter<'_, 'layout, '_> {
148    /// Read the original declared coverage without changing it.
149    #[must_use]
150    pub const fn coverage(&self) -> &IcSnapshotDataCoverage<'_> {
151        &self.coverage
152    }
153
154    /// Read the fixed private staging path; stable external custody remains required.
155    #[must_use]
156    pub fn path(&self) -> &Path {
157        &self.path
158    }
159
160    /// Append one exact admitted reply without retaining its memory buffer.
161    ///
162    /// Regions may interleave; each region must remain contiguous from zero. Chunk files
163    /// use exact metadata hash names and exclusive creation. No fsync or journal transition
164    /// occurs until explicit finish. Any error consumes this owner and retains all bytes.
165    /// Before writing, all region names must still select the held regular files at
166    /// their exact already-covered lengths. This sequential check is not a byte fence.
167    /// # Errors
168    /// Rejects changed custody, coverage conflicts and filesystem failures.
169    pub fn append(
170        mut self,
171        reply: &IcSnapshotDataReply<'_, '_>,
172    ) -> Result<Self, IcSnapshotArtifactError> {
173        self.check_custody()?;
174        self.check_region_custody()?;
175        self.coverage.admit(reply)?;
176        let index = match reply.request().kind() {
177            SnapshotDataKind::WasmModule { .. } => 0,
178            SnapshotDataKind::WasmMemory { .. } => 1,
179            SnapshotDataKind::StableMemory { .. } => 2,
180            SnapshotDataKind::WasmChunk { hash } => {
181                let name = format!("chunk-{}.bin", hex_bytes(hash));
182                create_file(&self.directory, &name)?.write_all(reply.chunk())?;
183                self.checksums
184                    .push((name.into(), reply.chunk_checksum().clone()));
185                self.check_custody()?;
186                return Ok(self);
187            }
188        };
189        self.regions[index].write_all(reply.chunk())?;
190        self.hashes[index].update(reply.chunk());
191        self.check_custody()?;
192        Ok(self)
193    }
194
195    /// Verify complete local bytes, retain their exact checksum and durably publish them.
196    ///
197    /// Complete declared coverage is required. The fixed closed tree must match hashes
198    /// computed from the actual admitted bytes, including original metadata and request.
199    /// Existing model transitions derive `Downloaded` and `ChecksumVerified` together; their
200    /// exact checksum is persisted atomically before the existing durable publisher runs.
201    /// On interruption, reopen the journal and use its ordinary `ChecksumVerified` publication
202    /// recovery without another read or transfer. Created partial trees cannot be resumed
203    /// through this writer; preserve them for operator-owned disposition. Ordinary Durable
204    /// replay reads only retained progress, without fresh byte verification.
205    ///
206    /// The caller still owns authentic backend completeness, token/raw-ID association and
207    /// stable custody. A returned checksum is local evidence, not a terminal/effect permit.
208    /// # Errors
209    /// Rejects incomplete coverage, extra/changed/unsafe bytes, replaced custody and failed
210    /// persistence/publication. Failure can leave a verified or published tree; retain it.
211    /// Closing directory-custody rejection can follow a retained `Durable` transition.
212    pub fn finish(self) -> Result<ArtifactChecksumRecord, IcSnapshotArtifactError> {
213        self.finish_with(DownloadJournalGuard::finalize_artifact)
214    }
215
216    fn finish_with(
217        mut self,
218        publish: impl FnOnce(
219            &mut DownloadJournalGuard<'layout>,
220            &str,
221            &str,
222        ) -> Result<(), DownloadJournalError>,
223    ) -> Result<ArtifactChecksumRecord, IcSnapshotArtifactError> {
224        if self.coverage.complete().is_none() {
225            return Err(IcSnapshotArtifactError::IncompleteCoverage);
226        }
227        self.check_custody()?;
228        for (name, hash) in REGIONS.into_iter().zip(self.hashes.iter()) {
229            self.checksums.push((
230                name.into(),
231                ArtifactChecksumRecord::from_digest(hash.clone().finalize().into()),
232            ));
233        }
234        self.check_closed_tree()?;
235        let expected = checksum_relative_files(std::mem::take(&mut self.checksums))
236            .map_err(ArtifactError::from)?;
237        checksum_directory(&self.path)?.verify(expected.hash())?;
238        self.check_custody()?;
239        let target = self.coverage.metadata().request().target();
240        let mut next = self.journal.next(
241            target,
242            &self.snapshot,
243            ArtifactStateRecord::Downloaded,
244            None,
245        )?;
246        next.advance(
247            target,
248            &self.snapshot,
249            ArtifactStateRecord::ChecksumVerified,
250            Some(expected.clone()),
251        )
252        .map_err(DownloadJournalError::from)?;
253        self.journal.store(next, write_json_durable)?;
254        publish(self.journal, target, &self.snapshot)?;
255        self.journal.check_usable()?;
256        check_directory_identity(
257            self.path
258                .parent()
259                .ok_or(IcSnapshotArtifactError::CustodyChanged)?,
260            &self.parent,
261        )?;
262        let entry = self
263            .journal
264            .record
265            .artifact(target, &self.snapshot)
266            .map_err(DownloadJournalError::from)?;
267        check_directory_identity(
268            &self.journal.layout.root().join(entry.artifact_path()),
269            &self.directory,
270        )?;
271        Ok(expected)
272    }
273
274    fn check_custody(&self) -> Result<(), IcSnapshotArtifactError> {
275        self.journal.check_usable()?;
276        for (path, held) in [
277            (
278                self.path
279                    .parent()
280                    .ok_or(IcSnapshotArtifactError::CustodyChanged)?,
281                &self.parent,
282            ),
283            (self.path.as_path(), &self.directory),
284        ] {
285            check_directory_identity(path, held)?;
286        }
287        Ok(())
288    }
289
290    fn check_closed_tree(&self) -> Result<(), IcSnapshotArtifactError> {
291        check_closed_tree(&self.directory, &self.checksums)
292    }
293
294    fn check_region_custody(&self) -> Result<(), IcSnapshotArtifactError> {
295        for ((name, file), expected) in REGIONS
296            .iter()
297            .zip(&self.regions)
298            .zip(self.coverage.covered_region_bytes())
299        {
300            let held = unix_fs::fstat(file).map_err(io::Error::from)?;
301            let current = unix_fs::statat(&self.directory, *name, AtFlags::SYMLINK_NOFOLLOW)
302                .map_err(io::Error::from)?;
303            let same_file = (current.st_dev, current.st_ino) == (held.st_dev, held.st_ino);
304            let exact_extent = current.st_size == held.st_size
305                && u64::try_from(held.st_size).ok() == Some(expected);
306            if !FileType::from_raw_mode(current.st_mode).is_file() || !same_file || !exact_extent {
307                return Err(IcSnapshotArtifactError::FileShape);
308            }
309        }
310        Ok(())
311    }
312}
313
314fn check_closed_tree(
315    directory: &File,
316    checksums: &[(PathBuf, ArtifactChecksumRecord)],
317) -> Result<(), IcSnapshotArtifactError> {
318    let mut expected = checksums
319        .iter()
320        .map(|(path, _)| path.as_os_str().as_encoded_bytes().to_vec())
321        .collect::<BTreeSet<_>>();
322    let mut directory = Dir::read_from(directory).map_err(io::Error::from)?;
323    while let Some(entry) = directory.read() {
324        let entry = entry.map_err(io::Error::from)?;
325        let name = entry.file_name().to_bytes();
326        if matches!(name, b"." | b"..") {
327            continue;
328        }
329        if !expected.remove(name) {
330            return Err(IcSnapshotArtifactError::UnexpectedEntry);
331        }
332    }
333    if !expected.is_empty() {
334        return Err(IcSnapshotArtifactError::UnexpectedEntry);
335    }
336    Ok(())
337}
338
339impl fmt::Debug for IcSnapshotArtifactWriter<'_, '_, '_> {
340    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
341        f.debug_struct("IcSnapshotArtifactWriter")
342            .field("coverage", &self.coverage)
343            .finish_non_exhaustive()
344    }
345}
346
347fn open_directory(path: &Path) -> io::Result<File> {
348    unix_fs::open(
349        path,
350        OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
351        Mode::empty(),
352    )
353    .map(File::from)
354    .map_err(io::Error::from)
355}
356
357fn check_directory_identity(path: &Path, held: &File) -> Result<(), IcSnapshotArtifactError> {
358    let current = fs::symlink_metadata(path)?;
359    let original = held.metadata()?;
360    if !current.is_dir() || current.dev() != original.dev() || current.ino() != original.ino() {
361        return Err(IcSnapshotArtifactError::CustodyChanged);
362    }
363    Ok(())
364}
365
366fn create_file(directory: &File, name: &str) -> io::Result<File> {
367    unix_fs::openat(
368        directory,
369        name,
370        OFlags::WRONLY | OFlags::CREATE | OFlags::EXCL | OFlags::NOFOLLOW | OFlags::CLOEXEC,
371        Mode::from_bits_truncate(0o600),
372    )
373    .map(File::from)
374    .map_err(io::Error::from)
375}
376
377/// Typed local artifact failures; no wire payload or raw snapshot identity is retained.
378#[derive(Debug, Error)]
379pub enum IcSnapshotArtifactError {
380    /// Original journal state, timestamp or raw metadata differs.
381    #[error("IC snapshot artifact original evidence mismatch")]
382    OriginalMismatch,
383    /// Complete declared data coverage has not been retained.
384    #[error("IC snapshot artifact coverage is incomplete")]
385    IncompleteCoverage,
386    /// The original artifact parent or staging directory has been replaced.
387    #[error("IC snapshot artifact directory custody changed")]
388    CustodyChanged,
389    /// The fixed closed artifact tree contains missing or additional entries.
390    #[error("IC snapshot artifact tree entries differ")]
391    UnexpectedEntry,
392    /// A retained direct child differs in regular-file identity or bounded length.
393    #[error("IC snapshot artifact file type or length differs")]
394    FileShape,
395    /// Retained original plan, complete durable selection or journal admission failed.
396    #[error(transparent)]
397    Integrity(#[from] super::DownloadIntegrityError),
398    /// Exact metadata/range/chunk coverage admission failed.
399    #[error(transparent)]
400    Coverage(#[from] IcSnapshotDataCoverageError),
401    /// Local expected admitted bytes differ from the current tree.
402    #[error(transparent)]
403    Checksum(#[from] ChecksumError),
404    /// Original download journal or durable publication failed.
405    #[error(transparent)]
406    Journal(#[from] DownloadJournalError),
407    /// Descriptor-based artifact traversal failed.
408    #[error(transparent)]
409    Artifact(#[from] ArtifactError),
410    /// Private descriptor creation or byte IO failed.
411    #[error(transparent)]
412    Io(#[from] io::Error),
413}
414
415#[cfg(test)]
416mod tests;