perf: optimize virtio-blk queue size and memory barrier

src/device/blk.rs

@@ -10,7 +10,7 @@ use log::info;
 use zerocopy::{FromBytes, Immutable, IntoBytes, KnownLayout};
 
 const QUEUE: u16 = 0;
-const QUEUE_SIZE: u16 = 16;
+const QUEUE_SIZE: u16 = 256;
 const SUPPORTED_FEATURES: BlkFeature = BlkFeature::RO
     .union(BlkFeature::FLUSH)
     .union(BlkFeature::RING_INDIRECT_DESC)

src/queue.rs

@@ -85,7 +85,7 @@
     pub fn new<T: Transport>(
         transport: &mut T,
         idx: u16,
         indirect: bool,
         event_idx: bool,
         access_platform: bool,
     ) -> Result<Self> {
@@ -201,8 +201,10 @@
         }
 
         // Write barrier so that device sees changes to descriptor table and available ring before
-        // change to available index.
-        fence(Ordering::SeqCst);
+        // change to available index. A Release fence is sufficient: it ensures all prior writes
+        // (descriptor table, avail ring entry) are ordered before the subsequent Release store to
+        // avail.idx. This matches Linux's dma_wmb() / virtio_wmb() semantics.
+        fence(Ordering::Release);
 
         // increase head of avail ring
         self.avail_idx = self.avail_idx.wrapping_add(1);