Merge pull request #1888 from ctiller/flow-like-lava-to-a-barnyard

Transport/call flow control interface

src/core/surface/byte_buffer_queue.c

@@ -62,6 +62,7 @@ int grpc_bbq_empty(grpc_byte_buffer_queue *q) {
 }
 
 void grpc_bbq_push(grpc_byte_buffer_queue *q, grpc_byte_buffer *buffer) {
+  q->bytes += grpc_byte_buffer_length(buffer);
   bba_push(&q->filling, buffer);
 }
 
@@ -72,8 +73,11 @@ void grpc_bbq_flush(grpc_byte_buffer_queue *q) {
   }
 }
 
+size_t grpc_bbq_bytes(grpc_byte_buffer_queue *q) { return q->bytes; }
+
 grpc_byte_buffer *grpc_bbq_pop(grpc_byte_buffer_queue *q) {
   grpc_bbq_array temp_array;
+  grpc_byte_buffer *out;
 
   if (q->drain_pos == q->draining.count) {
     if (q->filling.count == 0) {
@@ -87,5 +91,7 @@ grpc_byte_buffer *grpc_bbq_pop(grpc_byte_buffer_queue *q) {
     q->draining = temp_array;
   }
 
-  return q->draining.data[q->drain_pos++];
+  out = q->draining.data[q->drain_pos++];
+  q->bytes -= grpc_byte_buffer_length(out);
+  return out;
 }

src/core/surface/byte_buffer_queue.h

@@ -49,6 +49,7 @@ typedef struct {
   size_t drain_pos;
   grpc_bbq_array filling;
   grpc_bbq_array draining;
+  size_t bytes;
 } grpc_byte_buffer_queue;
 
 void grpc_bbq_destroy(grpc_byte_buffer_queue *q);
@@ -56,5 +57,6 @@ grpc_byte_buffer *grpc_bbq_pop(grpc_byte_buffer_queue *q);
 void grpc_bbq_flush(grpc_byte_buffer_queue *q);
 int grpc_bbq_empty(grpc_byte_buffer_queue *q);
 void grpc_bbq_push(grpc_byte_buffer_queue *q, grpc_byte_buffer *bb);
+size_t grpc_bbq_bytes(grpc_byte_buffer_queue *q);
 
 #endif  /* GRPC_INTERNAL_CORE_SURFACE_BYTE_BUFFER_QUEUE_H */

src/core/surface/call.c

@@ -513,6 +513,8 @@ static void unlock(grpc_call *call) {
   int completing_requests = 0;
   int start_op = 0;
   int i;
+  const gpr_uint32 MAX_RECV_PEEK_AHEAD = 65536;
+  size_t buffered_bytes;
   int cancel_alarm = 0;
 
   memset(&op, 0, sizeof(op));
@@ -528,6 +530,17 @@ static void unlock(grpc_call *call) {
     op.recv_ops = &call->recv_ops;
     op.recv_state = &call->recv_state;
     op.on_done_recv = &call->on_done_recv;
+    if (grpc_bbq_empty(&call->incoming_queue) && call->reading_message) {
+      op.max_recv_bytes = call->incoming_message_length -
+                          call->incoming_message.length + MAX_RECV_PEEK_AHEAD;
+    } else {
+      buffered_bytes = grpc_bbq_bytes(&call->incoming_queue);
+      if (buffered_bytes > MAX_RECV_PEEK_AHEAD) {
+        op.max_recv_bytes = 0;
+      } else {
+        op.max_recv_bytes = MAX_RECV_PEEK_AHEAD - buffered_bytes;
+      }
+    }
     call->receiving = 1;
     GRPC_CALL_INTERNAL_REF(call, "receiving");
     start_op = 1;

src/core/transport/chttp2/frame_window_update.c

@@ -94,8 +94,8 @@ grpc_chttp2_parse_error grpc_chttp2_window_update_parser_parse(
     }
     GPR_ASSERT(is_last);
 
-    if (transport_parsing->incoming_stream_id) {
-      if (stream_parsing) {
+    if (transport_parsing->incoming_stream_id != 0) {
+      if (stream_parsing != NULL) {
         GRPC_CHTTP2_FLOWCTL_TRACE_STREAM("update", transport_parsing,
                                          stream_parsing, outgoing_window_update,
                                          p->amount);

src/core/transport/chttp2/internal.h

@@ -353,7 +353,19 @@ typedef struct {
 
   /** window available for us to send to peer */
   gpr_int64 outgoing_window;
-  /** window available for peer to send to us - updated after parse */
+  /** The number of bytes the upper layers have offered to receive.
+      As the upper layer offers more bytes, this value increases.
+      As bytes are read, this value decreases. */
+  gpr_uint32 max_recv_bytes;
+  /** The number of bytes the upper layer has offered to read but we have
+      not yet announced to HTTP2 flow control.
+      As the upper layers offer to read more bytes, this value increases.
+      As we advertise incoming flow control window, this value decreases. */
+  gpr_uint32 unannounced_incoming_window;
+  /** The number of bytes of HTTP2 flow control we have advertised.
+      As we advertise incoming flow control window, this value increases.
+      As bytes are read, this value decreases.
+      Updated after parse. */
   gpr_uint32 incoming_window;
   /** stream ops the transport user would like to send */
   grpc_stream_op_buffer *outgoing_sopb;
@@ -391,6 +403,8 @@ typedef struct {
   grpc_stream_op_buffer sopb;
   /** how strongly should we indicate closure with the next write */
   grpc_chttp2_send_closed send_closed;
+  /** how much window should we announce? */
+  gpr_uint32 announce_window;
 } grpc_chttp2_stream_writing;
 
 struct grpc_chttp2_stream_parsing {
@@ -501,7 +515,9 @@ void grpc_chttp2_list_add_writable_window_update_stream(
     grpc_chttp2_stream_global *stream_global);
 int grpc_chttp2_list_pop_writable_window_update_stream(
     grpc_chttp2_transport_global *transport_global,
-    grpc_chttp2_stream_global **stream_global);
+    grpc_chttp2_transport_writing *transport_writing,
+    grpc_chttp2_stream_global **stream_global,
+    grpc_chttp2_stream_writing **stream_writing);
 void grpc_chttp2_list_remove_writable_window_update_stream(
     grpc_chttp2_transport_global *transport_global,
     grpc_chttp2_stream_global *stream_global);

src/core/transport/chttp2/parsing.c

@@ -173,7 +173,14 @@ void grpc_chttp2_publish_reads(
       GRPC_CHTTP2_FLOWCTL_TRACE_STREAM(
           "parsed", transport_parsing, stream_parsing, incoming_window_delta,
           -(gpr_int64)stream_parsing->incoming_window_delta);
+      GRPC_CHTTP2_FLOWCTL_TRACE_STREAM(
+          "parsed", transport_parsing, stream_global, max_recv_bytes,
+          -(gpr_int64)stream_parsing->incoming_window_delta);
       stream_global->incoming_window -= stream_parsing->incoming_window_delta;
+      GPR_ASSERT(stream_global->max_recv_bytes >= 
+          stream_parsing->incoming_window_delta);
+      stream_global->max_recv_bytes -= 
+          stream_parsing->incoming_window_delta;
       stream_parsing->incoming_window_delta = 0;
       grpc_chttp2_list_add_writable_window_update_stream(transport_global,
                                                          stream_global);

src/core/transport/chttp2/stream_lists.c

@@ -139,6 +139,7 @@ static void stream_list_add(grpc_chttp2_transport *t, grpc_chttp2_stream *s,
 void grpc_chttp2_list_add_writable_stream(
     grpc_chttp2_transport_global *transport_global,
     grpc_chttp2_stream_global *stream_global) {
+  GPR_ASSERT(stream_global->id != 0);
   stream_list_add(TRANSPORT_FROM_GLOBAL(transport_global),
                   STREAM_FROM_GLOBAL(stream_global), GRPC_CHTTP2_LIST_WRITABLE);
 }
@@ -204,6 +205,7 @@ int grpc_chttp2_list_pop_written_stream(
 void grpc_chttp2_list_add_writable_window_update_stream(
     grpc_chttp2_transport_global *transport_global,
     grpc_chttp2_stream_global *stream_global) {
+  GPR_ASSERT(stream_global->id != 0);
   stream_list_add(TRANSPORT_FROM_GLOBAL(transport_global),
                   STREAM_FROM_GLOBAL(stream_global),
                   GRPC_CHTTP2_LIST_WRITABLE_WINDOW_UPDATE);
@@ -211,11 +213,14 @@ void grpc_chttp2_list_add_writable_window_update_stream(
 
 int grpc_chttp2_list_pop_writable_window_update_stream(
     grpc_chttp2_transport_global *transport_global,
-    grpc_chttp2_stream_global **stream_global) {
+    grpc_chttp2_transport_writing *transport_writing,
+    grpc_chttp2_stream_global **stream_global,
+    grpc_chttp2_stream_writing **stream_writing) {
   grpc_chttp2_stream *stream;
   int r = stream_list_pop(TRANSPORT_FROM_GLOBAL(transport_global), &stream,
                           GRPC_CHTTP2_LIST_WRITABLE_WINDOW_UPDATE);
   *stream_global = &stream->global;
+  *stream_writing = &stream->writing;
   return r;
 }
 

src/core/transport/chttp2/writing.c

@@ -66,11 +66,9 @@ int grpc_chttp2_unlocking_check_writes(
   /* for each grpc_chttp2_stream that's become writable, frame it's data
      (according to
      available window sizes) and add to the output buffer */
-  while (transport_global->outgoing_window &&
-         grpc_chttp2_list_pop_writable_stream(transport_global,
+  while (grpc_chttp2_list_pop_writable_stream(transport_global,
                                               transport_writing, &stream_global,
-                                              &stream_writing) &&
-         stream_global->outgoing_window > 0) {
+                                              &stream_writing)) {
     stream_writing->id = stream_global->id;
     window_delta = grpc_chttp2_preencode(
         stream_global->outgoing_sopb->ops, &stream_global->outgoing_sopb->nops,
@@ -106,20 +104,21 @@ int grpc_chttp2_unlocking_check_writes(
   /* for each grpc_chttp2_stream that wants to update its window, add that
    * window here */
   while (grpc_chttp2_list_pop_writable_window_update_stream(transport_global,
-                                                            &stream_global)) {
-    window_delta =
-        transport_global->settings[GRPC_LOCAL_SETTINGS]
-                                  [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE] -
-        stream_global->incoming_window;
-    if (!stream_global->read_closed && window_delta > 0) {
-      gpr_slice_buffer_add(
-          &transport_writing->outbuf,
-          grpc_chttp2_window_update_create(stream_global->id, window_delta));
+                                                            transport_writing,
+                                                            &stream_global,
+                                                            &stream_writing)) {
+    stream_writing->id = stream_global->id;
+    if (!stream_global->read_closed && stream_global->unannounced_incoming_window > 0) {
+      stream_writing->announce_window = stream_global->unannounced_incoming_window;
       GRPC_CHTTP2_FLOWCTL_TRACE_STREAM("write", transport_global, stream_global,
-                                       incoming_window, window_delta);
-      stream_global->incoming_window += window_delta;
+                                       incoming_window, stream_global->unannounced_incoming_window);
+      GRPC_CHTTP2_FLOWCTL_TRACE_STREAM("write", transport_global, stream_global,
+                                       unannounced_incoming_window, -(gpr_int64)stream_global->unannounced_incoming_window);
+      stream_global->incoming_window += stream_global->unannounced_incoming_window;
+      stream_global->unannounced_incoming_window = 0;
       grpc_chttp2_list_add_incoming_window_updated(transport_global,
                                                    stream_global);
+      grpc_chttp2_list_add_writing_stream(transport_writing, stream_writing);
     }
   }
 
@@ -169,10 +168,19 @@ static void finalize_outbuf(grpc_chttp2_transport_writing *transport_writing) {
 
   while (
       grpc_chttp2_list_pop_writing_stream(transport_writing, &stream_writing)) {
-    grpc_chttp2_encode(stream_writing->sopb.ops, stream_writing->sopb.nops,
-                       stream_writing->send_closed != GRPC_DONT_SEND_CLOSED,
-                       stream_writing->id, &transport_writing->hpack_compressor,
-                       &transport_writing->outbuf);
+    if (stream_writing->sopb.nops > 0 || stream_writing->send_closed != GRPC_DONT_SEND_CLOSED) {
+      grpc_chttp2_encode(stream_writing->sopb.ops, stream_writing->sopb.nops,
+                         stream_writing->send_closed != GRPC_DONT_SEND_CLOSED,
+                         stream_writing->id, &transport_writing->hpack_compressor,
+                         &transport_writing->outbuf);
+    }
+    if (stream_writing->announce_window > 0) {
+      gpr_slice_buffer_add(
+          &transport_writing->outbuf,
+          grpc_chttp2_window_update_create(
+              stream_writing->id, stream_writing->announce_window));
+      stream_writing->announce_window = 0;
+    }
     stream_writing->sopb.nops = 0;
     if (stream_writing->send_closed == GRPC_SEND_CLOSED_WITH_RST_STREAM) {
       gpr_slice_buffer_add(&transport_writing->outbuf,
@@ -197,7 +205,8 @@ void grpc_chttp2_cleanup_writing(
 
   while (grpc_chttp2_list_pop_written_stream(
       transport_global, transport_writing, &stream_global, &stream_writing)) {
-    if (stream_global->outgoing_sopb->nops == 0) {
+    if (stream_global->outgoing_sopb != NULL &&
+        stream_global->outgoing_sopb->nops == 0) {
       stream_global->outgoing_sopb = NULL;
       grpc_chttp2_schedule_closure(transport_global,
                                    stream_global->send_done_closure, 1);

src/core/transport/chttp2_transport.c

@@ -358,7 +358,9 @@ static int init_stream(grpc_transport *gt, grpc_stream *gs,
     s->global.outgoing_window =
         t->global.settings[GRPC_PEER_SETTINGS]
                           [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE];
-    s->parsing.incoming_window = s->global.incoming_window =
+    s->global.max_recv_bytes = 
+        s->parsing.incoming_window = 
+        s->global.incoming_window =
         t->global.settings[GRPC_SENT_SETTINGS]
                           [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE];
     *t->accepting_stream = s;
@@ -562,6 +564,8 @@ static void maybe_start_some_streams(
     stream_global->incoming_window =
         transport_global->settings[GRPC_SENT_SETTINGS]
                                   [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE];
+    stream_global->max_recv_bytes = 
+        GPR_MAX(stream_global->incoming_window, stream_global->max_recv_bytes);
     grpc_chttp2_stream_map_add(
         &TRANSPORT_FROM_GLOBAL(transport_global)->new_stream_map,
         stream_global->id, STREAM_FROM_GLOBAL(stream_global));
@@ -570,6 +574,9 @@ static void maybe_start_some_streams(
     grpc_chttp2_list_add_incoming_window_updated(transport_global,
                                                  stream_global);
     grpc_chttp2_list_add_writable_stream(transport_global, stream_global);
+    grpc_chttp2_list_add_writable_window_update_stream(transport_global,
+                                                       stream_global);
+
   }
   /* cancel out streams that will never be started */
   while (transport_global->next_stream_id >= MAX_CLIENT_STREAM_ID &&
@@ -620,12 +627,23 @@ static void perform_stream_op_locked(
     stream_global->publish_sopb = op->recv_ops;
     stream_global->publish_sopb->nops = 0;
     stream_global->publish_state = op->recv_state;
+    if (stream_global->max_recv_bytes < op->max_recv_bytes) {
+      GRPC_CHTTP2_FLOWCTL_TRACE_STREAM("op", transport_global, stream_global,
+          max_recv_bytes, op->max_recv_bytes - stream_global->max_recv_bytes);
+      GRPC_CHTTP2_FLOWCTL_TRACE_STREAM(
+          "op", transport_global, stream_global, unannounced_incoming_window,
+          op->max_recv_bytes - stream_global->max_recv_bytes);
+      stream_global->unannounced_incoming_window += op->max_recv_bytes - stream_global->max_recv_bytes;
+      stream_global->max_recv_bytes = op->max_recv_bytes;
+    }
     grpc_chttp2_incoming_metadata_live_op_buffer_end(
         &stream_global->outstanding_metadata);
-    grpc_chttp2_list_add_read_write_state_changed(transport_global,
-                                                  stream_global);
-    grpc_chttp2_list_add_writable_window_update_stream(transport_global,
-                                                       stream_global);
+    if (stream_global->id != 0) {
+      grpc_chttp2_list_add_read_write_state_changed(transport_global,
+                                                    stream_global);
+      grpc_chttp2_list_add_writable_window_update_stream(transport_global,
+                                                         stream_global);
+    }
   }
 
   if (op->bind_pollset) {
@@ -1038,7 +1056,7 @@ void grpc_chttp2_flowctl_trace(const char *file, int line, const char *reason,
     identifier = gpr_strdup(context_scope);
   }
   gpr_log(GPR_INFO,
-          "FLOWCTL: %s %-10s %8s %-23s %8lld %c %8lld = %8lld %-10s [%s:%d]",
+          "FLOWCTL: %s %-10s %8s %-27s %8lld %c %8lld = %8lld %-10s [%s:%d]",
           is_client ? "client" : "server", identifier, context_thread, var,
           current_value, delta < 0 ? '-' : '+', delta < 0 ? -delta : delta,
           current_value + delta, reason, file, line);

src/core/transport/transport.h

@@ -72,6 +72,10 @@ typedef struct grpc_transport_stream_op {
 
   grpc_stream_op_buffer *recv_ops;
   grpc_stream_state *recv_state;
+  /** The number of bytes this peer is currently prepared to receive.
+      These bytes will be eventually used to replenish per-stream flow control
+      windows. */
+  gpr_uint32 max_recv_bytes;
   grpc_iomgr_closure *on_done_recv;
 
   grpc_pollset *bind_pollset;

src/core/transport/transport_op_string.c

@@ -128,7 +128,8 @@ char *grpc_transport_stream_op_string(grpc_transport_stream_op *op) {
   if (op->recv_ops) {
     if (!first) gpr_strvec_add(&b, gpr_strdup(" "));
     first = 0;
-    gpr_strvec_add(&b, gpr_strdup("RECV"));
+    gpr_asprintf(&tmp, "RECV:max_recv_bytes=%d", op->max_recv_bytes);
+    gpr_strvec_add(&b, tmp);
   }
 
   if (op->bind_pollset) {