Engine I/O: exclusive output files, read-checked probes, resolved clip-group paths (#1328)

* fix(engine): create output files exclusively and check read access before probing

* address review: canonicalize through the checked io gateway, treat any occupied export name as taken

* address review: read group audio through the validated file handle

* address review: document that a write-only root writes a new numbered WAV per save

* address review: clip-group audio is opened without waiting for a FIFO writer

crates/engine/src/clip_group_file.rs

@@ -87,20 +87,27 @@ fn decode_group_audio(
     cmd_id: &str,
 ) -> Result<(Vec<(String, soundcraft_audio_io::AudioBuffer)>, BTreeMap<u64, usize>)> {
     let (mut decoded, mut index, mut file_of) = (Vec::new(), BTreeMap::new(), BTreeMap::new());
+    // The folder as the file system knows it, so a link among the group's files can't lead out.
+    let dir = if dir.as_os_str().is_empty() { Path::new(".") } else { dir };
+    let root = crate::io::canonical_path(dir)?;
     for (old, rel) in audio_entries {
         let full = dir.join(rel);
+        let escapes = |rel: &str| EngineError::BadParams(cmd_id.into(), format!("audio path escapes the group folder: {rel}"));
         if !full.starts_with(dir)
             || Path::new(rel)
                 .components()
                 .any(|c| matches!(c, std::path::Component::ParentDir | std::path::Component::RootDir | std::path::Component::Prefix(_)))
         {
-            return Err(EngineError::BadParams(cmd_id.into(), format!("audio path escapes the group folder: {rel}")));
+            return Err(escapes(rel));
+        }
+        let key = crate::io::canonical_path(&full)?;
+        if !key.starts_with(&root) {
+            return Err(escapes(rel));
         }
-        let key = crate::io::file_key(&full);
         let n = match index.get(&key) {
             Some(n) => *n,
             None => {
-                let bytes = crate::io::read_file(&full)?;
+                let bytes = crate::io::read_file_inside(&full, &root)?;
                 let (_, buf) = soundcraft_audio_io::decode(&bytes, Some("wav")).map_err(|err| EngineError::Io(format!("{rel}: {err}")))?;
                 let buf = crate::io::at_rate(buf, rate, crate::io::MAX_IMPORT_SAMPLES)
                     .map_err(|r| EngineError::Io(format!("{rel}: {}", r.message("too large to import"))))?;
@@ -268,9 +275,59 @@ mod tests {
         std::fs::write(&path, file.to_string()).unwrap();
         let r = e.execute("file.import_clip_groups", &json!({"path": path, "at": 1000, "track": "Bass"}));
         assert!(matches!(r, Err(crate::EngineError::BadParams(..))), "absolute paths are refused: {r:?}");
+        // A link inside the folder that leads out of it is refused once resolved.
+        #[cfg(unix)]
+        {
+            let outside = std::env::temp_dir().join(format!("sc-grp-outside-{}.wav", std::process::id()));
+            std::fs::copy(dir.join("Wait... what_0.wav"), &outside).unwrap();
+            std::os::unix::fs::symlink(&outside, dir.join("link.wav")).unwrap();
+            for rel in file["audio"].as_object_mut().unwrap().values_mut() {
+                *rel = json!("link.wav");
+            }
+            std::fs::write(&path, file.to_string()).unwrap();
+            let r = e.execute("file.import_clip_groups", &json!({"path": path, "at": 1000, "track": "Bass"}));
+            assert!(matches!(r, Err(crate::EngineError::BadParams(..))), "links out of the folder are refused: {r:?}");
+            let _ = std::fs::remove_file(&outside);
+        }
         let _ = std::fs::remove_dir_all(&dir);
     }
 
+    /// Under a read root, a link in the group folder to a folder outside the root is refused the same way
+    /// whether or not the file asked for exists there, so the group cannot be used to probe for names.
+    #[cfg(unix)]
+    #[test]
+    fn links_out_of_the_read_root_are_refused_alike_for_existing_and_missing_files() {
+        let base = std::env::temp_dir().join(format!("sc-grp-root-{}", std::process::id()));
+        let _ = std::fs::remove_dir_all(&base);
+        let (dir, outside) = (base.join("group"), base.join("outside"));
+        std::fs::create_dir_all(&dir).unwrap();
+        std::fs::create_dir_all(&outside).unwrap();
+        let path = dir.join("drums.scgrp").to_string_lossy().to_string();
+        let mut e = crate::demo::demo_engine();
+        let kick = e.session().tracks.iter().find(|t| t.name == "Kick").map(|t| t.id).unwrap();
+        let ids: Vec<_> = e.session().track(kick).unwrap().clips().iter().take(1).map(|c| c.id.0).collect();
+        e.execute("file.export_clip_groups", &json!({"path": path, "clips": ids})).unwrap();
+        let mut file: serde_json::Value = serde_json::from_str(&std::fs::read_to_string(&path).unwrap()).unwrap();
+        let exported = file["audio"].as_object().unwrap().values().next().unwrap().as_str().unwrap().to_string();
+        std::fs::copy(dir.join(&exported), outside.join("real.wav")).unwrap();
+        std::os::unix::fs::symlink(&outside, dir.join("out")).unwrap();
+        let roots = crate::file_access::AutomationRoots::new(&[&dir], &[]).unwrap().unwrap();
+        let mut refusals = Vec::new();
+        for target in ["out/real.wav", "out/missing.wav"] {
+            for rel in file["audio"].as_object_mut().unwrap().values_mut() {
+                *rel = json!(target);
+            }
+            std::fs::write(&path, file.to_string()).unwrap();
+            let r = crate::file_access::confine(Some(&roots), || {
+                e.execute("file.import_clip_groups", &json!({"path": path, "at": 1000, "track": "Bass"}))
+            });
+            refusals.push(r.unwrap_err().to_string());
+        }
+        assert!(refusals[0].contains("outside the read root"), "{}", refusals[0]);
+        assert_eq!(refusals[0].replace("real.wav", "X"), refusals[1].replace("missing.wav", "X"), "existing and missing names are refused alike");
+        let _ = std::fs::remove_dir_all(&base);
+    }
+
     #[test]
     fn imported_clips_carry_no_auto_rename_scheme() {
         let dir = std::env::temp_dir().join(format!("sc-grp-names-{}", std::process::id()));

crates/engine/src/io.rs

@@ -65,6 +65,69 @@ fn read_file_within(path: &Path, session: bool) -> Result<Vec<u8>> {
     Ok(data)
 }
 
+/// `O_NONBLOCK` of the platforms that have a FIFO to wait on (std has no name for it, and this crate has no unsafe
+/// code or libc); 0 elsewhere, where a FIFO open still waits for a writer.
+#[cfg(unix)]
+const O_NONBLOCK: i32 = if cfg!(all(
+    any(target_os = "linux", target_os = "android"),
+    not(any(
+        target_arch = "mips",
+        target_arch = "mips32r6",
+        target_arch = "mips64",
+        target_arch = "mips64r6",
+        target_arch = "sparc",
+        target_arch = "sparc64"
+    ))
+)) {
+    0x800
+} else if cfg!(any(
+    target_os = "macos",
+    target_os = "ios",
+    target_os = "freebsd",
+    target_os = "netbsd",
+    target_os = "openbsd",
+    target_os = "dragonfly"
+)) {
+    0x4
+} else {
+    0
+};
+
+/// [`read_file`] for a file that must lie inside `root` (a canonical folder): the file is opened
+/// once, and the bytes come from that handle only after the handle is shown to be the file the
+/// path resolves to inside `root`. A file swapped in for the path after a check by name is
+/// refused instead of read. Where the handle's identity can't be read (non-Unix targets) the
+/// resolved path is checked against `root` alone.
+pub(crate) fn read_file_inside(path: &Path, root: &Path) -> Result<Vec<u8>> {
+    use std::io::Read;
+    crate::file_access::check_read(path).map_err(EngineError::Io)?;
+    let err = |err: std::io::Error| EngineError::Io(format!("{}: {err}", path.display()));
+    // opened without waiting for a FIFO's writer (the type is checked on the opened handle, below)
+    let mut options = std::fs::OpenOptions::new();
+    options.read(true);
+    #[cfg(unix)]
+    std::os::unix::fs::OpenOptionsExt::custom_flags(&mut options, O_NONBLOCK);
+    let mut file = options.open(path).map_err(err)?;
+    if !file.metadata().map_err(err)?.is_file() {
+        return Err(EngineError::Io(format!("{}: not a regular file", path.display())));
+    }
+    let key = canonical_path(path)?;
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::MetadataExt;
+        let (a, b) = (file.metadata().map_err(err)?, std::fs::metadata(&key).map_err(err)?);
+        if (a.dev(), a.ino()) != (b.dev(), b.ino()) {
+            return Err(EngineError::Io(format!("{}: changed while it was opened", path.display())));
+        }
+    }
+    if !key.starts_with(root) {
+        return Err(EngineError::Io(format!("{}: outside the folder it must stay in", path.display())));
+    }
+    let mut bytes = Vec::new();
+    file.read_to_end(&mut bytes).map_err(err)?;
+    Ok(bytes)
+}
+
 /// [`read_file`] as UTF-8 text.
 pub fn read_text(path: impl AsRef<Path>) -> Result<String> {
     let path = path.as_ref();
@@ -192,6 +255,17 @@ pub(crate) fn make_dir(dir: impl AsRef<Path>) -> Result<()> {
     std::fs::create_dir_all(dir).map_err(|err| EngineError::Io(format!("{}: {err}", dir.display())))
 }
 
+/// `path` with its links followed, inside the automation read root when there is one. The root is
+/// checked on the path as named (links followed, a missing tail resolved by its nearest folder)
+/// before anything is resolved, so a name that leads out of the root is refused the same way
+/// whether or not it exists there.
+pub(crate) fn canonical_path(path: impl AsRef<Path>) -> Result<PathBuf> {
+    let path = path.as_ref();
+    let checked = if path.as_os_str().is_empty() { Path::new(".") } else { path };
+    crate::file_access::check_read(checked).map_err(EngineError::Io)?;
+    std::fs::canonicalize(path).map_err(|err| EngineError::Io(format!("{}: {err}", path.display())))
+}
+
 /// Write `bytes` to the file at `path`, inside the automation write root when there is one.
 pub(crate) fn write_path(path: impl AsRef<Path>, bytes: impl AsRef<[u8]>) -> Result<()> {
     let path = path.as_ref();
@@ -446,7 +520,7 @@ impl Out {
         if cfg!(target_arch = "wasm32") {
             return Ok(Out::Memory(std::io::Cursor::new(Vec::new())));
         }
-        Ok(Out::File(std::io::BufWriter::with_capacity(1 << 20, std::fs::File::create(part)?)))
+        Ok(Out::File(std::io::BufWriter::with_capacity(1 << 20, create_new(part)?)))
     }
 
     /// Puts the finished file at `path` and returns its size: the part file is synced and renamed
@@ -578,8 +652,28 @@ pub fn export_clips(e: &Engine, ids: &[soundcraft_model::ClipId], dir: &str) ->
         let audio = soundcraft_mix::render_clips(s, t, c.range());
         let buf = AudioBuffer { sample_rate: s.sample_rate.hz(), channels: audio };
         let bytes = soundcraft_audio_io::encode(&buf, &EncodeOptions::default()).map_err(|err| EngineError::Io(err.to_string()))?;
-        let path = clip_file(dir, &sanitize_name(&c.name), &mut names, Path::exists)?;
-        write_file(&path.to_string_lossy(), &bytes)?;
+        let stem = sanitize_name(&c.name);
+        if cfg!(target_arch = "wasm32") {
+            let path = clip_file(dir, &stem, &mut names, Path::exists)?;
+            write_file(&path.to_string_lossy(), &bytes)?;
+            continue;
+        }
+        // The name is taken by creating the file, so whatever entry already has it (a link
+        // to elsewhere included) is left alone rather than written through. An occupied name is
+        // taken whatever it is (a link whose target is outside the write root too): it is only
+        // looked at, never resolved, so the root check does not stop the search for a free name.
+        // The folder itself was checked above.
+        let (mut file, path) = loop {
+            let path = clip_file(dir, &stem, &mut names, |p| std::fs::symlink_metadata(p).is_ok())?;
+            crate::file_access::check_write(&path).map_err(EngineError::Io)?;
+            match std::fs::OpenOptions::new().write(true).create_new(true).open(&path) {
+                Ok(f) => break (f, path),
+                // taken since it was looked at: `names` has it now, so the next name is tried
+                Err(err) if err.kind() == std::io::ErrorKind::AlreadyExists => {}
+                Err(err) => return Err(EngineError::Io(format!("{}: {err}", path.display()))),
+            }
+        };
+        std::io::Write::write_all(&mut file, &bytes).map_err(|err| EngineError::Io(format!("{}: {err}", path.display())))?;
     }
     Ok(names.taken.len())
 }
@@ -644,13 +738,13 @@ fn write_session(session: &Session, path: &str, overwrite: bool) -> Result<(Sess
         {
             src.path = f.join(&src.path).to_string_lossy().into_owned();
         }
-        let exists_rel = same_dir && !Path::new(&src.path).is_absolute() && dir.join(&src.path).exists();
+        let exists_rel = same_dir && !Path::new(&src.path).is_absolute() && file_found(dir.join(&src.path));
         let shared = shared_paths.get(i).copied().unwrap_or(false);
         if exists_rel && !src.unsaved && !shared {
             continue;
         }
         let in_folder = Path::new(&src.path).strip_prefix(&audio_dir).ok().and_then(Path::to_str).map(str::to_string);
-        if let Some(f) = in_folder.filter(|_| !src.unsaved && Path::new(&src.path).is_file()) {
+        if let Some(f) = in_folder.filter(|_| !src.unsaved && file_found(&src.path)) {
             src.path = format!("Audio Files/{f}");
             continue;
         }
@@ -672,8 +766,11 @@ fn write_session(session: &Session, path: &str, overwrite: bool) -> Result<(Sess
         let mut stem = sanitize_name(&src.name);
         if overwrite {
             // One file per source id so autosaves never collide or pile up.
+            // (With only an automation write root, `file_found` cannot read an existing file, so each save
+            // writes a new numbered WAV: reuse never reads a file the roots do not allow, which keeps a save
+            // from revealing whether an unreadable file matches the audio.)
             stem = format!("{stem}_{}", src.id.0);
-            if audio_dir.join(format!("{stem}.wav")).exists() {
+            if file_found(audio_dir.join(format!("{stem}.wav"))) {
                 src.path = format!("Audio Files/{stem}.wav");
                 src.unsaved = false;
                 files.insert(Arc::as_ptr(audio), src.path.clone());
@@ -756,11 +853,22 @@ fn sync_dir(dir: &Path) {
 
 fn write_synced(path: &Path, bytes: &[u8]) -> std::io::Result<()> {
     use std::io::Write;
-    let mut f = std::fs::File::create(path)?;
+    let mut f = create_new(path)?;
     f.write_all(bytes)?;
     f.sync_all()
 }
 
+/// Creates `path` as a new file. A leftover of that name (a part file from an interrupted write)
+/// is removed first, but an existing entry is never opened: a link planted at a predictable name
+/// in the destination folder can't redirect the write elsewhere.
+fn create_new(path: &Path) -> std::io::Result<std::fs::File> {
+    match std::fs::remove_file(path) {
+        Err(err) if err.kind() != std::io::ErrorKind::NotFound => return Err(err),
+        _ => {}
+    }
+    std::fs::OpenOptions::new().write(true).create_new(true).open(path)
+}
+
 /// Copies `from` to `to`, which must not exist yet; an existing entry, even a link, is never followed.
 fn copy_new(from: &Path, to: &Path) -> std::io::Result<()> {
     let mut src = std::fs::File::open(from)?;
@@ -840,7 +948,8 @@ fn same_path(a: &Path, b: &Path) -> bool {
 
 /// Store `bytes` as the first of `<stem>.wav`, `<stem>_01.wav`, … in `dir` that is free or already
 /// holds exactly these bytes, so saving never replaces audio; returns the name and whether it was
-/// written.
+/// written. Where the automation roots allow writing in `dir` but not reading (only a write root),
+/// an existing file is never compared, so each save writes a new numbered WAV.
 fn store_wav(dir: &Path, stem: &str, bytes: &[u8]) -> Result<(String, bool)> {
     use std::io::Write;
     for k in 0..10_000u32 {
@@ -855,7 +964,12 @@ fn store_wav(dir: &Path, stem: &str, bytes: &[u8]) -> Result<(String, bool)> {
                 return Ok((name, true));
             }
             Err(err) if err.kind() == std::io::ErrorKind::AlreadyExists => {
-                if std::fs::metadata(&path).is_ok_and(|m| m.len() == bytes.len() as u64) && std::fs::read(&path).is_ok_and(|b| b == bytes) {
+                // Reused only when it may be read: whether it equals the audio must not be
+                // learnt about a file the roots allow writing beside but not reading.
+                if crate::file_access::check_read(&path).is_ok()
+                    && std::fs::metadata(&path).is_ok_and(|m| m.len() == bytes.len() as u64)
+                    && std::fs::read(&path).is_ok_and(|b| b == bytes)
+                {
                     return Ok((name, false));
                 }
             }
@@ -1054,6 +1168,89 @@ pub fn plan_takes(e: &Engine, unsaved_dir: &Path, input_channels: usize) -> Vec<
         .collect()
 }
 
+#[cfg(test)]
+mod exclusive_tests {
+    use super::*;
+
+    fn scratch(name: &str) -> PathBuf {
+        let dir = std::env::temp_dir().join(format!("soundcraft-io-{name}-{}", std::process::id()));
+        let _ = std::fs::remove_dir_all(&dir);
+        std::fs::create_dir_all(&dir).unwrap();
+        dir
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn a_planted_link_at_the_part_name_is_not_written_through() {
+        let dir = scratch("part-link");
+        let (target, part) = (dir.join("elsewhere"), dir.join("session.scraft.tmp"));
+        std::os::unix::fs::symlink(&target, &part).unwrap();
+        write_synced(&part, b"new").unwrap();
+        assert!(!target.exists(), "the link's target was created");
+        assert_eq!(std::fs::read(&part).unwrap(), b"new", "the link itself was replaced by the file");
+        // A leftover regular part file is still replaced.
+        write_synced(&part, b"newer").unwrap();
+        assert_eq!(std::fs::read(&part).unwrap(), b"newer");
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn a_file_read_within_a_root_must_resolve_inside_it() {
+        let dir = scratch("within");
+        let (inside, outside) = (dir.join("group"), dir.join("elsewhere"));
+        std::fs::create_dir_all(&inside).unwrap();
+        std::fs::create_dir_all(&outside).unwrap();
+        std::fs::write(inside.join("a.wav"), b"a").unwrap();
+        std::fs::write(outside.join("b.wav"), b"b").unwrap();
+        std::os::unix::fs::symlink(outside.join("b.wav"), inside.join("link.wav")).unwrap();
+        let root = canonical_path(&inside).unwrap();
+        assert_eq!(read_file_inside(&inside.join("a.wav"), &root).unwrap(), b"a");
+        assert!(read_file_inside(&inside.join("link.wav"), &root).is_err());
+        // a FIFO with no writer is refused, not waited on (Linux: the coreutils tool makes it)
+        #[cfg(target_os = "linux")]
+        if std::process::Command::new("mkfifo").arg(inside.join("pipe.wav")).status().is_ok_and(|s| s.success()) {
+            let err = read_file_inside(&inside.join("pipe.wav"), &root).unwrap_err().to_string();
+            assert!(err.contains("not a regular file"), "{err}");
+        }
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn export_skips_a_name_held_by_a_dangling_link() {
+        let dir = scratch("export-link");
+        let e = crate::demo::demo_engine();
+        let (t, c) = e.session().tracks.iter().find_map(|t| t.clips().iter().find(|c| c.is_audio()).map(|c| (t.id, c.clone()))).unwrap();
+        let _ = t;
+        let target = dir.join("elsewhere.wav");
+        std::os::unix::fs::symlink(&target, dir.join(format!("{}.wav", sanitize_name(&c.name)))).unwrap();
+        assert_eq!(export_clips(&e, &[c.id], &dir.to_string_lossy()).unwrap(), 1);
+        assert!(!target.exists(), "the export followed the link");
+        assert!(dir.join(format!("{}-2.wav", sanitize_name(&c.name))).is_file(), "the next free name was used");
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    /// With a write root, a link out of it that holds the name is still just an occupied name.
+    #[cfg(unix)]
+    #[test]
+    fn export_skips_a_link_out_of_the_write_root_too() {
+        let dir = scratch("export-root-link");
+        let e = crate::demo::demo_engine();
+        let c = e.session().tracks.iter().find_map(|t| t.clips().iter().find(|c| c.is_audio()).cloned()).unwrap();
+        let elsewhere = scratch("export-root-elsewhere");
+        let target = elsewhere.join("target.wav");
+        std::os::unix::fs::symlink(&target, dir.join(format!("{}.wav", sanitize_name(&c.name)))).unwrap();
+        let roots = crate::file_access::AutomationRoots::new(&[], &[&dir]).unwrap().unwrap();
+        let n = crate::file_access::confine(Some(&roots), || export_clips(&e, &[c.id], &dir.to_string_lossy())).unwrap();
+        assert_eq!(n, 1);
+        assert!(!target.exists(), "the export followed the link out of the root");
+        assert!(dir.join(format!("{}-2.wav", sanitize_name(&c.name))).is_file(), "the next free name was used");
+        let _ = std::fs::remove_dir_all(&dir);
+        let _ = std::fs::remove_dir_all(&elsewhere);
+    }
+}
+
 #[cfg(test)]
 mod stem_tests {
     #[test]