diff --git a/src/system/kernel/fs/fd.cpp b/src/system/kernel/fs/fd.cpp
index d19ab6f951..d85450c99e 100644
--- a/src/system/kernel/fs/fd.cpp
+++ b/src/system/kernel/fs/fd.cpp
@@ -29,6 +29,7 @@
 #include <vfs.h>
 #include <wait_for_objects.h>
 
+#include "Vnode.h"
 #include "vfs_tracing.h"
 
 
@@ -722,7 +723,30 @@ common_vector_io(int fd, off_t pos, const iovec* vecs, size_t count, bool write,
 		return B_BAD_VALUE;
 	}
 
+	// See if we can bypass the loop and perform I/O directly. We can only do this
+	// for vnodes that have no cache, as the I/O hook bypasses the cache entirely.
+	struct vnode* vnode = descriptor->u.vnode;
 	status_t status = B_OK;
+	if (!movePosition && pos != -1 && count > 1 && descriptor->type == FDTYPE_FILE
+			&& vnode != NULL && vnode->cache == NULL && vnode->ops->io != NULL) {
+		BStackOrHeapArray<generic_io_vec, 8> iovecs(count);
+		if (!iovecs.IsValid())
+			return B_NO_MEMORY;
+
+		generic_size_t length = 0;
+		for (size_t i = 0; i < count; i++) {
+			iovecs[i].base = (generic_addr_t)vecs[i].iov_base;
+			iovecs[i].length = vecs[i].iov_len;
+			length += vecs[i].iov_len;
+		}
+
+		status = (write ? vfs_write_pages : vfs_read_pages)(vnode,
+			descriptor->cookie, pos, iovecs, count, 0, &length);
+		if (length > 0)
+			return length;
+		return status;
+	}
+
 	ssize_t bytesTransferred = 0;
 	for (size_t i = 0; i < count; i++) {
 		if (vecs[i].iov_base == NULL)