Queue blk requests when the ring is empty

Instead of cancelling block requests due to no space on the ring
that lead to corruption of the upper layer, block until there is
space.

drivers/virtio-blk.cc

@@ -171,6 +171,13 @@ void virtio_blk::response_worker() {
             delete req;
             queue->get_buf_finalize();
         }
+
+        // wake up the requesting thread in case the ring was full before
+        _request_thread_lock.lock();
+        if (_waiting_request_thread) {
+            _waiting_request_thread->wake();
+        }
+        _request_thread_lock.unlock();
     }
 }
 
@@ -195,12 +202,6 @@ int virtio_blk::make_virtio_request(struct bio* bio)
         }
 
         vring* queue = get_virt_queue(0);
-        if (!queue->avail_ring_not_empty()) {
-            virtio_w("Warn: No space on ring");
-            biodone(bio, false);
-            return ENOMEM;
-        }
-
         virtio_blk_request_type type;
 
         switch (bio->bio_cmd) {
@@ -251,10 +252,13 @@ int virtio_blk::make_virtio_request(struct bio* bio)
         req->res.status = 0;
         queue->_sg_vec.push_back(vring::sg_node(mmu::virt_to_phys(&req->res), sizeof (struct virtio_blk_res), vring_desc::VRING_DESC_F_WRITE));
 
-        if (!queue->add_buf(req)) {
-            biodone(bio, false);
-            delete req;
-            return EBUSY;
+        while (!queue->add_buf(req)) {
+            _waiting_request_thread = sched::thread::current();
+            std::atomic_thread_fence(std::memory_order_seq_cst);
+            sched::thread::wait_until([queue] {queue->get_buf_gc(); return queue->avail_ring_has_room(queue->_sg_vec.size());});
+            _request_thread_lock.lock();
+            _waiting_request_thread = nullptr;
+            _request_thread_lock.unlock();
         }
 
         queue->kick(); // should be out of the loop but I like plenty of irqs for the test

drivers/virtio-blk.hh

@@ -178,7 +178,10 @@ namespace virtio {
         bool _ro;
         // This mutex protects parallel make_request invocations
         mutex _lock;
-
+        sched::thread* _waiting_request_thread = nullptr;
+        // The mutex protects the above thread ptr from going away of
+        // the waker
+        mutex _request_thread_lock;
     };