[runtime/buffer] avoid write-locking the Append buffer during blob writes

runtime/fuzz/fuzz_targets/buffer.rs

@@ -322,7 +322,7 @@ fn fuzz(input: FuzzInput) {
 
                 FuzzOperation::AppendCloneBlob => {
                     if let Some(ref append) = append_buffer {
-                        let _ = append.clone_blob();
+                        let _ = append.clone_blob().await;
                     }
                 }
 

runtime/src/utils/buffer/append.rs

@@ -7,10 +7,20 @@ use std::{num::NonZeroUsize, sync::Arc};
 
 /// A [Blob] wrapper that supports appending new data that is both read and write cached, and
 /// provides buffer-pool managed read caching of older data.
+///
+/// # Concurrent Access
+///
+/// This implementation allows readers to proceed while flush I/O is in progress, as long as they
+/// are reading from the write buffer or the pool cache. Readers that need to access data from the
+/// underlying blob (cache miss) will wait for any in-progress write to complete.
+///
+/// The implementation involves two locks: one for the write buffer (and blob size metadata), and
+/// one for the underlying blob itself. To avoid deadlocks, the buffer lock is always acquired
+/// before the blob lock.
 #[derive(Clone)]
 pub struct Append<B: Blob> {
-    /// The underlying blob being wrapped.
-    blob: B,
+    /// The underlying blob being wrapped, protected by a lock for I/O coordination.
+    blob: Arc<RwLock<B>>,
 
     /// Unique id assigned by the buffer pool.
     id: u64,
@@ -58,7 +68,7 @@ impl<B: Blob> Append<B> {
         }
 
         Ok(Self {
-            blob,
+            blob: Arc::new(RwLock::new(blob)),
             id: pool_ref.next_id().await,
             pool_ref,
             buffer: Arc::new(RwLock::new((buffer, size))),
@@ -71,7 +81,8 @@ impl<B: Blob> Append<B> {
         let buf = buf.into();
 
         // Acquire a write lock on the buffer and blob_size.
-        let (buffer, blob_size) = &mut *self.buffer.write().await;
+        let mut guard = self.buffer.write().await;
+        let (buffer, _) = &mut *guard;
 
         // Ensure the write doesn't overflow.
         buffer
@@ -81,7 +92,7 @@ impl<B: Blob> Append<B> {
 
         if buffer.append(buf.as_ref()) {
             // Buffer is over capacity, flush it to the underlying blob.
-            return self.flush(buffer, blob_size).await;
+            return self.flush_internal(guard).await;
         }
 
         Ok(())
@@ -97,16 +108,54 @@ impl<B: Blob> Append<B> {
 
     /// Flush the append buffer to the underlying blob, caching each page worth of written data in
     /// the buffer pool.
-    async fn flush(&self, buffer: &mut Buffer, blob_size: &mut u64) -> Result<(), Error> {
-        // Take the buffered data, if any.
-        let Some((mut buf, offset)) = buffer.take() else {
+    ///
+    /// This method acquires the blob write lock before releasing the buffer lock, ensuring readers
+    /// that need blob access will wait for the write to complete.
+    async fn flush_internal(
+        &self,
+        mut guard: crate::RwLockWriteGuard<'_, (Buffer, u64)>,
+    ) -> Result<(), Error> {
+        let (buffer, blob_size) = &mut *guard;
+
+        // Prepare the data to be written.
+        let Some((buf, write_offset)) = self.prepare_flush_data(buffer, *blob_size).await else {
             return Ok(());
         };
 
-        // Insert the flushed data into the buffer pool. This step isn't just to ensure recently
-        // written data remains cached for future reads, but is in fact required to purge
-        // potentially stale cache data which might result from the edge the case of rewinding a
-        // blob across a page boundary.
+        // Update blob_size *before* releasing the lock. We do this optimistically; if the write
+        // fails below, the program will return an error and likely abort/panic, so maintaining
+        // exact consistency on error isn't strictly required.
+        *blob_size = write_offset + buf.len() as u64;
+
+        // Acquire blob write lock BEFORE releasing buffer lock. This ensures no reader can access
+        // the blob until the write completes.
+        let blob_guard = self.blob.write().await;
+
+        // Release buffer lock, allowing concurrent buffered reads while the write is in progress.
+        // Any attempts to read from the blob will block until the write completes.
+        drop(guard);
+
+        // Perform the write while holding only blob lock.
+        blob_guard.write_at(buf, write_offset).await?;
+
+        Ok(())
+    }
+
+    /// Prepares data from the buffer to be flushed to the blob.
+    ///
+    /// This method:
+    /// 1. Takes the data from the write buffer.
+    /// 2. Caches it in the buffer pool.
+    /// 3. Returns the data to be written and the offset to write it at (if any).
+    async fn prepare_flush_data(
+        &self,
+        buffer: &mut Buffer,
+        blob_size: u64,
+    ) -> Option<(Vec<u8>, u64)> {
+        // Take the buffered data, if any.
+        let (mut buf, offset) = buffer.take()?;
+
+        // Insert the flushed data into the buffer pool.
         let remaining = self.pool_ref.cache(self.id, &buf, offset).await;
 
         // If there's any data left over that doesn't constitute an entire page, re-buffer it into
@@ -121,22 +170,20 @@ impl<B: Blob> Append<B> {
 
         // Early exit if there's no new data to write.
         if new_data_start >= buf.len() {
-            return Ok(());
+            return None;
         }
 
         if new_data_start > 0 {
             buf.drain(0..new_data_start);
         }
-        let new_data_len = buf.len() as u64;
-        self.blob.write_at(buf, *blob_size).await?;
-        *blob_size += new_data_len;
 
-        Ok(())
+        // Return the data to write, and the offset where to write it within the blob.
+        Some((buf, blob_size))
     }
 
     /// Clones and returns the underlying blob.
-    pub fn clone_blob(&self) -> B {
-        self.blob.clone()
+    pub async fn clone_blob(&self) -> B {
+        self.blob.read().await.clone()
     }
 }
 
@@ -155,7 +202,8 @@ impl<B: Blob> Blob for Append<B> {
             .ok_or(Error::OffsetOverflow)?;
 
         // Acquire a read lock on the buffer.
-        let (buffer, _) = &*self.buffer.read().await;
+        let guard = self.buffer.read().await;
+        let (buffer, _) = &*guard;
 
         // If the data required is beyond the size of the blob, return an error.
         if end_offset > buffer.size() {
@@ -164,13 +212,40 @@ impl<B: Blob> Blob for Append<B> {
 
         // Extract any bytes from the buffer that overlap with the requested range.
         let remaining = buffer.extract(buf.as_mut(), offset);
+
+        // Release buffer lock before potential I/O.
+        drop(guard);
+
         if remaining == 0 {
             return Ok(buf);
         }
 
-        // If there are bytes remaining to be read, use the buffer pool to get them.
+        // Fast path: try to read *only from pool cache without acquiring blob lock. This allows
+        // concurrent reads even while a flush is in progress.
+        let cached = self
+            .pool_ref
+            .read_cached(self.id, &mut buf.as_mut()[..remaining], offset)
+            .await;
+
+        if cached == remaining {
+            // All bytes found in cache.
+            return Ok(buf);
+        }
+
+        // Slow path: cache miss (partial or full), acquire blob read lock to ensure any in-flight
+        // write completes before we read from the blob.
+        let blob_guard = self.blob.read().await;
+
+        // Read remaining bytes that were not already obtained from the earlier cache read.
+        let uncached_offset = offset + cached as u64;
+        let uncached_len = remaining - cached;
         self.pool_ref
-            .read(&self.blob, self.id, &mut buf.as_mut()[..remaining], offset)
+            .read(
+                &*blob_guard,
+                self.id,
+                &mut buf.as_mut()[cached..cached + uncached_len],
+                uncached_offset,
+            )
             .await?;
 
         Ok(buf)
@@ -184,11 +259,15 @@ impl<B: Blob> Blob for Append<B> {
     }
 
     async fn sync(&self) -> Result<(), Error> {
+        // Flush any buffered data. When flush_internal returns, the write_at has completed
+        // and data is in the OS buffer.
         {
-            let (buffer, blob_size) = &mut *self.buffer.write().await;
-            self.flush(buffer, blob_size).await?;
+            let guard = self.buffer.write().await;
+            self.flush_internal(guard).await?;
         }
-        self.blob.sync().await
+        // Sync the OS buffer to disk. We need the blob read lock here since sync() requires
+        // access to the blob, but only a read lock since we're not modifying blob state.
+        self.blob.read().await.sync().await
     }
 
     /// Resize the blob to the provided `size`.
@@ -199,17 +278,30 @@ impl<B: Blob> Blob for Append<B> {
         // always updated should the blob grow back to the point where we have new data for the same
         // page, if any old data hasn't expired naturally by then.
 
-        // Acquire a write lock on the buffer.
-        let (buffer, blob_size) = &mut *self.buffer.write().await;
+        // Acquire buffer lock first.
+        // NOTE: We MUST acquire the buffer lock before the blob lock to avoid deadlocks with
+        // `append`, which acquires buffer then blob (via `flush_internal`).
+        let mut guard = self.buffer.write().await;
+        let (buffer, blob_size) = &mut *guard;
 
-        // Flush any buffered bytes to the underlying blob. (Note that a fancier implementation
-        // might avoid flushing those bytes that are backed up over by the next step, if any.)
-        self.flush(buffer, blob_size).await?;
+        // Acquire blob write lock to prevent concurrent reads throughout the resize.
+        let blob_guard = self.blob.write().await;
+
+        // Flush any buffered bytes first, using the helper.
+        // We hold both locks here, so no concurrent operations can happen.
+        if let Some((buf, write_offset)) = self.prepare_flush_data(buffer, *blob_size).await {
+            // Write the data to the blob.
+            let len = buf.len() as u64;
+            blob_guard.write_at(buf, write_offset).await?;
+
+            // Update blob_size to reflect the flush.
+            *blob_size = write_offset + len;
+        }
 
         // Resize the underlying blob.
-        self.blob.resize(size).await?;
+        blob_guard.resize(size).await?;
 
-        // Update the physical blob size.
+        // Update the blob size.
         *blob_size = size;
 
         // Reset the append buffer to the new size, ensuring its page alignment.
@@ -218,7 +310,7 @@ impl<B: Blob> Blob for Append<B> {
         buffer.data.clear();
         if leftover_size != 0 {
             let page_buf = vec![0; leftover_size as usize];
-            let buf = self.blob.read_at(page_buf, buffer.offset).await?;
+            let buf = blob_guard.read_at(page_buf, buffer.offset).await?;
             assert!(!buffer.append(buf.as_ref()));
         }
 

runtime/src/utils/buffer/pool.rs

@@ -102,6 +102,35 @@ impl PoolRef {
         Pool::offset_to_page(self.page_size, offset)
     }
 
+    /// Try to read the specified bytes from the buffer pool cache only. Returns the number of
+    /// bytes successfully read from cache and copied to `buf` before a page fault, if any.
+    ///
+    /// This method never reads from the underlying blob - it only checks the cache.
+    ///
+    /// # Warning
+    ///
+    /// Attempts to read any of the last (blob_size % page_size) "trailing bytes" of the blob will
+    /// always return 0 since the buffer pool only deals with page sized chunks.
+    pub(super) async fn read_cached(
+        &self,
+        blob_id: u64,
+        mut buf: &mut [u8],
+        mut offset: u64,
+    ) -> usize {
+        let original_len = buf.len();
+        let buffer_pool = self.pool.read().await;
+        while !buf.is_empty() {
+            let count = buffer_pool.read_at(self.page_size, blob_id, buf, offset);
+            if count == 0 {
+                // Cache miss - return how many bytes we successfully read
+                break;
+            }
+            offset += count as u64;
+            buf = &mut buf[count..];
+        }
+        original_len - buf.len()
+    }
+
     /// Read the specified bytes, preferentially from the buffer pool cache. Bytes not found in the
     /// buffer pool will be read from the provided `blob` and cached for future reads.
     ///

storage/src/journal/contiguous/fixed.rs

@@ -498,15 +498,15 @@ impl<E: Storage + Metrics, A: CodecFixed<Cfg = ()>> Journal<E, A> {
         assert!(start_blob <= self.tail_index);
         let blobs = self.blobs.range(start_blob..).collect::<Vec<_>>();
         let full_size = items_per_blob * Self::CHUNK_SIZE_U64;
-        let mut blob_plus = blobs
-            .into_iter()
-            .map(|(blob_index, blob)| (*blob_index, blob.clone_blob(), full_size))
-            .collect::<Vec<_>>();
+        let mut blob_plus = Vec::with_capacity(blobs.len() + 1);
+        for (blob_index, blob) in blobs {
+            blob_plus.push((*blob_index, blob.clone_blob().await, full_size));
+        }
 
         // Include the tail blob.
         self.tail.sync().await?; // make sure no data is buffered
         let tail_size = self.tail.size().await;
-        blob_plus.push((self.tail_index, self.tail.clone_blob(), tail_size));
+        blob_plus.push((self.tail_index, self.tail.clone_blob().await, tail_size));
         let start_offset = (start_pos % items_per_blob) * Self::CHUNK_SIZE_U64;
 
         // Replay all blobs in order and stream items as they are read (to avoid occupying too much