Files
Graphite/document/container/src/backends/opfs.rs
Dennis Kobert 6fe1af3afe Add document-container crate (#4191)
* Add document-container crate: container backends and archive codecs

* Address PR review: path/prefix split, safe size casts, OPFS stream aborts

* Address PR review round 2: mmap read check, UTF8 entry names, prefix normalization

* Make MmappedBytes::new fallible so mmap reads can't silently degrade

* Address PR review round 3: backend contract uniformity (symlinks, remove, list)

* Apply symlink-component check to FolderBackend listing paths

* Address PR review: idempotent OPFS delete logging, tar default-features, shared entry-size cap

* Fix validate_path doc: dotfiles pass, CurDir/ParentDir rejected

* Omit symlink entries from FolderBackend listings for consistency with resolve

* Preserve zip I/O errors and reject non-canonical paths in validate_path

* Extend archive apis to return the archive writer

* Rename document/document-container directory to document/container

* Add archive format sniffing and deserialize_auto

* Drop temporal hedge from checked_entry_size comment

* Tighten verbose doc comments in document-container

* Coalesce consecutive same-path OPFS appends to avoid O(n^2) file copies

* Review

* Update document-container for the deserialize/store_non_blocking renames

---------

Co-authored-by: Timon <me@timon.zip>
2026-06-05 13:28:02 -04:00

476 lines
17 KiB
Rust

//! OPFS (Origin Private File System) backend for browser wasm.
use crate::{AsyncContainer, ByteHolder, ContainerError, Result, validate_path, validate_prefix, with_trailing_slash};
use futures::channel::oneshot;
use js_sys::Uint8Array;
use std::collections::{HashSet, VecDeque};
use std::sync::{Arc, Mutex};
use wasm_bindgen::JsCast;
use wasm_bindgen::prelude::*;
use wasm_bindgen_futures::{JsFuture, spawn_local};
use web_sys::{
Blob, DomException, FileSystemCreateWritableOptions, FileSystemDirectoryHandle, FileSystemFileHandle, FileSystemGetDirectoryOptions, FileSystemGetFileOptions, FileSystemWritableFileStream,
WritableStream,
};
enum Mutation {
Write {
path: String,
bytes: Vec<u8>,
},
Append {
path: String,
bytes: Vec<u8>,
},
Delete {
path: String,
},
/// In-band flush barrier. The worker fulfills the sender once it dequeues this, which (by FIFO
/// order) signals that every mutation enqueued before it has been applied to disk.
Barrier(oneshot::Sender<()>),
}
struct Inner {
directory: FileSystemDirectoryHandle,
/// Paths believed to be on disk, letting `exists_non_blocking` answer without OPFS's missing sync API.
/// Optimistic: a path is inserted when its write is queued and kept even if that write later fails, so it
/// can briefly over-report. Rebuilt from real disk state on the next `open`.
on_disk: HashSet<String>,
queue: VecDeque<Mutation>,
worker_active: bool,
}
pub struct OpfsBackend {
inner: Arc<Mutex<Inner>>,
}
// Safety: only built for browser wasm where JS handles never leave the main thread.
unsafe impl Send for OpfsBackend {}
unsafe impl Sync for OpfsBackend {}
impl OpfsBackend {
/// Open (or create) `directory_name` under the OPFS root.
pub async fn open(directory_name: &str) -> Result<Self> {
let directory = open_directory(directory_name).await.map_err(js_err)?;
let on_disk = enumerate_paths(&directory, "").await.map_err(js_err)?;
Ok(Self {
inner: Arc::new(Mutex::new(Inner {
directory,
on_disk,
queue: VecDeque::new(),
worker_active: false,
})),
})
}
fn directory(&self) -> FileSystemDirectoryHandle {
self.inner.lock().unwrap().directory.clone()
}
/// Wait until all queued non-blocking mutations have hit disk, so awaited reads observe them. Plants a
/// barrier at the back of the queue and awaits the worker reaching it; by FIFO order, everything ahead is
/// then applied. Draining stays the worker's job so it remains the sole mutator.
async fn flush(&self) {
let (sender, receiver) = oneshot::channel();
{
let mut guard = self.inner.lock().unwrap();
guard.queue.push_back(Mutation::Barrier(sender));
kick_worker(&self.inner, &mut guard);
}
// The sender is only dropped without sending if the worker is torn down mid-drain; either way
// there is nothing left to wait for, so a receive error is treated as "already flushed".
let _ = receiver.await;
}
}
impl AsyncContainer for OpfsBackend {
async fn read(&self, path: &str) -> Result<ByteHolder> {
validate_path(path)?;
// Non-blocking writes only land on disk once the queue drains, so flush first and then treat
// disk as authoritative. Draining (rather than folding the queue against an awaited base read)
// avoids racing the background worker, which could otherwise double-apply a queued append.
self.flush().await;
let bytes = read_file(&self.directory(), path).await.map_err(js_err)?;
Ok(ByteHolder::Owned(bytes))
}
async fn write(&self, path: &str, bytes: &[u8]) -> Result<()> {
validate_path(path)?;
// Flush first so this awaited write lands after any non-blocking mutation already queued for the
// same path, then apply directly so the real disk error still propagates to the caller.
self.flush().await;
write_file(&self.directory(), path, bytes).await.map_err(js_err)?;
self.inner.lock().unwrap().on_disk.insert(path.to_string());
Ok(())
}
async fn append(&self, path: &str, bytes: &[u8]) -> Result<()> {
validate_path(path)?;
self.flush().await;
append_file(&self.directory(), path, bytes).await.map_err(js_err)?;
self.inner.lock().unwrap().on_disk.insert(path.to_string());
Ok(())
}
async fn write_sized(&self, path: &str, size: usize, fill: &mut dyn FnMut(&mut [u8]) -> Result<()>) -> Result<()> {
validate_path(path)?;
let mut buffer = vec![0; size];
fill(&mut buffer)?;
self.flush().await;
write_file(&self.directory(), path, &buffer).await.map_err(js_err)?;
self.inner.lock().unwrap().on_disk.insert(path.to_string());
Ok(())
}
async fn list(&self, prefix: &str) -> Result<Vec<String>> {
validate_prefix(prefix)?;
self.flush().await;
list_entries(&self.directory(), prefix, EntryKind::File).await.map_err(|error| list_error(prefix, error))
}
async fn list_dirs(&self, prefix: &str) -> Result<Vec<String>> {
validate_prefix(prefix)?;
self.flush().await;
list_entries(&self.directory(), prefix, EntryKind::Directory).await.map_err(|error| list_error(prefix, error))
}
async fn exists(&self, path: &str) -> bool {
if validate_path(path).is_err() {
return false;
}
self.flush().await;
file_exists(&self.directory(), path).await
}
async fn remove(&self, path: &str) -> Result<()> {
validate_path(path)?;
self.flush().await;
// Idempotent: a missing entry is not an error, matching the other backends.
if let Err(error) = remove_file(&self.directory(), path).await
&& !is_not_found(&error)
{
return Err(js_err(error));
}
self.inner.lock().unwrap().on_disk.remove(path);
Ok(())
}
fn write_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
validate_path(path)?;
let mut guard = self.inner.lock().unwrap();
guard.on_disk.insert(path.to_string());
guard.queue.push_back(Mutation::Write {
path: path.to_string(),
bytes: bytes.to_vec(),
});
kick_worker(&self.inner, &mut guard);
Ok(())
}
fn append_non_blocking(&self, path: &str, bytes: &[u8]) -> Result<()> {
validate_path(path)?;
let mut guard = self.inner.lock().unwrap();
guard.on_disk.insert(path.to_string());
guard.queue.push_back(Mutation::Append {
path: path.to_string(),
bytes: bytes.to_vec(),
});
kick_worker(&self.inner, &mut guard);
Ok(())
}
fn remove_non_blocking(&self, path: &str) -> Result<()> {
validate_path(path)?;
let mut guard = self.inner.lock().unwrap();
guard.on_disk.remove(path);
guard.queue.push_back(Mutation::Delete { path: path.to_string() });
kick_worker(&self.inner, &mut guard);
Ok(())
}
fn exists_non_blocking(&self, path: &str) -> bool {
validate_path(path).is_ok() && self.inner.lock().unwrap().on_disk.contains(path)
}
}
fn kick_worker(inner: &Arc<Mutex<Inner>>, guard: &mut Inner) {
if guard.worker_active {
return;
}
guard.worker_active = true;
let inner = inner.clone();
spawn_local(drain_queue(inner));
}
/// Apply every queued mutation to disk, in FIFO order, until the queue is empty, then mark the worker
/// idle. Spawned once via [`kick_worker`] and runs as the sole mutator; the awaited read paths wait on
/// a [`Mutation::Barrier`] rather than draining themselves, so there is never more than one drainer.
///
/// Consecutive appends to the same path are coalesced into one combined append. Each `createWritable`
/// on OPFS copies the whole existing file, so appending N frames one at a time is O(N^2) in bytes
/// copied (every per-op autosave queues one append per hot op). Concatenating a run into a single
/// append collapses that to one file copy, while staying byte-identical to applying them in order.
async fn drain_queue(inner: Arc<Mutex<Inner>>) {
loop {
// Pop the front mutation, coalescing a run of same-path appends behind it into one batch.
let (directory, mutation) = {
let mut guard = inner.lock().unwrap();
let Some(mut mutation) = guard.queue.pop_front() else {
guard.worker_active = false;
return;
};
if let Mutation::Append { path, bytes } = &mut mutation {
while let Some(Mutation::Append { path: next_path, .. }) = guard.queue.front()
&& next_path == path
{
let Some(Mutation::Append { bytes: next_bytes, .. }) = guard.queue.pop_front() else {
unreachable!()
};
bytes.extend_from_slice(&next_bytes);
}
}
(guard.directory.clone(), mutation)
};
match mutation {
Mutation::Write { path, bytes } => {
if let Err(error) = write_file(&directory, &path, &bytes).await {
log::error!("OPFS background write for {path} failed: {error:?}");
}
}
Mutation::Append { path, bytes } => {
if let Err(error) = append_file(&directory, &path, &bytes).await {
log::error!("OPFS background append for {path} failed: {error:?}");
}
}
Mutation::Delete { path } => {
// Removal is idempotent, so a missing entry is the expected outcome of a redundant delete, not an error.
if let Err(error) = remove_file(&directory, &path).await
&& !is_not_found(&error)
{
log::error!("OPFS background delete for {path} failed: {error:?}");
}
}
// A receive error on the waiter side just means the reader gave up; nothing to apply.
Mutation::Barrier(sender) => {
let _ = sender.send(());
}
}
}
}
fn js_err(error: JsValue) -> ContainerError {
ContainerError::Backend(format!("{error:?}"))
}
/// Resolve `directory_path` (a `/`-separated relative path) under the OPFS root, creating each
/// segment. OPFS rejects directory names containing `/`, so a multi-segment path like
/// `documents/<id>` must be descended one segment at a time rather than passed whole.
async fn open_directory(directory_path: &str) -> std::result::Result<FileSystemDirectoryHandle, JsValue> {
let storage = web_sys::window().ok_or_else(|| JsValue::from_str("no window"))?.navigator().storage();
let mut current: FileSystemDirectoryHandle = JsFuture::from(storage.get_directory()).await?.dyn_into()?;
for segment in directory_path.split('/').filter(|segment| !segment.is_empty()) {
let options = FileSystemGetDirectoryOptions::new();
options.set_create(true);
current = JsFuture::from(current.get_directory_handle_with_options(segment, &options)).await?.dyn_into()?;
}
Ok(current)
}
/// Descend the `/`-separated path against `root` and return the directory handle plus the final segment.
async fn descend<'a>(root: &FileSystemDirectoryHandle, relative: &'a str, create_dirs: bool) -> std::result::Result<(FileSystemDirectoryHandle, &'a str), JsValue> {
let mut current = root.clone();
let mut segments = relative.split('/').filter(|s| !s.is_empty()).collect::<Vec<_>>();
let file = segments.pop().ok_or_else(|| JsValue::from_str("empty path"))?;
for segment in segments {
let options = FileSystemGetDirectoryOptions::new();
options.set_create(create_dirs);
current = JsFuture::from(current.get_directory_handle_with_options(segment, &options)).await?.dyn_into()?;
}
Ok((current, file))
}
async fn read_file(root: &FileSystemDirectoryHandle, path: &str) -> std::result::Result<Vec<u8>, JsValue> {
let (directory, name) = descend(root, path, false).await?;
let handle: FileSystemFileHandle = JsFuture::from(directory.get_file_handle(name)).await?.dyn_into()?;
let file_value = JsFuture::from(handle.get_file()).await?;
let blob: Blob = file_value.dyn_into()?;
let buffer = JsFuture::from(blob.array_buffer()).await?;
Ok(Uint8Array::new(&buffer).to_vec())
}
async fn write_file(root: &FileSystemDirectoryHandle, path: &str, bytes: &[u8]) -> std::result::Result<(), JsValue> {
let (directory, name) = descend(root, path, true).await?;
let options = FileSystemGetFileOptions::new();
options.set_create(true);
let handle: FileSystemFileHandle = JsFuture::from(directory.get_file_handle_with_options(name, &options)).await?.dyn_into()?;
let writable: FileSystemWritableFileStream = JsFuture::from(handle.create_writable()).await?.dyn_into()?;
let stream: WritableStream = writable.clone().unchecked_into();
// Wrap in an async block so any failure, including a synchronous JS throw from `write_with_js_u8_array`,
// aborts the stream before returning instead of leaving a dangling locked writable.
let write = async {
let array = Uint8Array::from(bytes);
JsFuture::from(writable.write_with_js_u8_array(&array)?).await?;
Ok::<(), JsValue>(())
}
.await;
if let Err(error) = write {
let _ = JsFuture::from(stream.abort()).await;
return Err(error);
}
JsFuture::from(stream.close()).await?;
Ok(())
}
async fn append_file(root: &FileSystemDirectoryHandle, path: &str, bytes: &[u8]) -> std::result::Result<(), JsValue> {
let (directory, name) = descend(root, path, true).await?;
let file_options = FileSystemGetFileOptions::new();
file_options.set_create(true);
let handle: FileSystemFileHandle = JsFuture::from(directory.get_file_handle_with_options(name, &file_options)).await?.dyn_into()?;
// Determine the current end-of-file so we can seek there before writing.
let file_value = JsFuture::from(handle.get_file()).await?;
let blob: Blob = file_value.dyn_into()?;
let offset = blob.size();
// `keepExistingData: true` preserves bytes outside the written range; without it OPFS truncates to the written length.
let writable_options = FileSystemCreateWritableOptions::new();
writable_options.set_keep_existing_data(true);
let writable: FileSystemWritableFileStream = JsFuture::from(handle.create_writable_with_options(&writable_options)).await?.dyn_into()?;
let stream: WritableStream = writable.clone().unchecked_into();
// Wrap in an async block so a synchronous JS throw from `seek_with_f64`/`write_with_js_u8_array`
// aborts the stream instead of bypassing the abort via `?`.
let seek_and_write = async {
JsFuture::from(writable.seek_with_f64(offset)?).await?;
let array = Uint8Array::from(bytes);
JsFuture::from(writable.write_with_js_u8_array(&array)?).await?;
Ok::<(), JsValue>(())
}
.await;
if let Err(error) = seek_and_write {
let _ = JsFuture::from(stream.abort()).await;
return Err(error);
}
JsFuture::from(stream.close()).await?;
Ok(())
}
async fn remove_file(root: &FileSystemDirectoryHandle, path: &str) -> std::result::Result<(), JsValue> {
let (directory, name) = descend(root, path, false).await?;
JsFuture::from(directory.remove_entry(name)).await?;
Ok(())
}
async fn file_exists(root: &FileSystemDirectoryHandle, path: &str) -> bool {
let Ok((directory, name)) = descend(root, path, false).await else {
return false;
};
JsFuture::from(directory.get_file_handle(name)).await.is_ok()
}
#[derive(Clone, Copy)]
enum EntryKind {
File,
Directory,
}
async fn list_entries(root: &FileSystemDirectoryHandle, prefix: &str, want: EntryKind) -> std::result::Result<Vec<String>, JsValue> {
// `.` and the empty string both name the container root.
let prefix = if prefix == "." { "" } else { prefix };
let directory = if prefix.is_empty() {
root.clone()
} else {
let mut current = root.clone();
for segment in prefix.split('/').filter(|s| !s.is_empty()) {
let options = FileSystemGetDirectoryOptions::new();
options.set_create(false);
current = match JsFuture::from(current.get_directory_handle_with_options(segment, &options)).await {
Ok(value) => value.dyn_into()?,
Err(error) if is_not_found(&error) => return Ok(Vec::new()),
Err(error) => return Err(error),
};
}
current
};
let entries = directory.entries();
let iterator: js_sys::AsyncIterator = entries.unchecked_into();
let mut results = Vec::new();
let prefix_with_slash = with_trailing_slash(prefix);
let want_kind = match want {
EntryKind::File => web_sys::FileSystemHandleKind::File,
EntryKind::Directory => web_sys::FileSystemHandleKind::Directory,
};
loop {
let next: js_sys::IteratorNext = JsFuture::from(iterator.next()?).await?.unchecked_into();
if next.done() {
break;
}
let pair: js_sys::Array = next.value().unchecked_into();
let Some(name) = pair.get(0).as_string() else { continue };
let handle: web_sys::FileSystemHandle = pair.get(1).unchecked_into();
if handle.kind() != want_kind {
continue;
}
results.push(format!("{prefix_with_slash}{name}"));
}
Ok(results)
}
/// Walk every file under `prefix` (recursively) and collect their full container paths.
/// Used at open time to populate the in-memory tracking set.
async fn enumerate_paths(root: &FileSystemDirectoryHandle, prefix: &str) -> std::result::Result<HashSet<String>, JsValue> {
let mut paths = HashSet::new();
let mut to_visit = vec![prefix.to_string()];
while let Some(current_prefix) = to_visit.pop() {
for file in list_entries(root, &current_prefix, EntryKind::File).await? {
paths.insert(file);
}
for dir in list_entries(root, &current_prefix, EntryKind::Directory).await? {
to_visit.push(dir);
}
}
Ok(paths)
}
fn is_not_found(error: &JsValue) -> bool {
error.dyn_ref::<DomException>().is_some_and(|error| error.name() == "NotFoundError")
}
/// OPFS raises `TypeMismatchError` when a path segment used as a directory is actually a file.
fn is_type_mismatch(error: &JsValue) -> bool {
error.dyn_ref::<DomException>().is_some_and(|error| error.name() == "TypeMismatchError")
}
/// Map a listing error: a prefix that names a file becomes [`ContainerError::NotADirectory`], anything
/// else passes through as a backend error.
fn list_error(prefix: &str, error: JsValue) -> ContainerError {
if is_type_mismatch(&error) {
ContainerError::NotADirectory(prefix.to_string())
} else {
js_err(error)
}
}