On Sat, 2 Aug 2014, Tatsuhiro Tsujikawa wrote:

> This patch adds break statement if is_empty_data is true at the end of while
> loop in readwrite_data. Without this break statement, CPU usage goes to 100%
> if HTTP/2 transfer is interrupted (RST_STREAM) and its recv function keeps
> return 0.

I don't think it is the right change. I'm worried we have cases where there's
still data to send in when the end of reading condition occurs.

We have a funky condition for nghttp2 i the data_pending() function we should
probably tweak instead so that it won't say that there is data pending when we
know there's not!

How did you test this? I'm thinking perhaps another fix is to clear
httpc->closed after it has served its purpose once, like:

--- a/lib/http2.c
+++ b/lib/http2.c
@@ -666,10 +666,11 @@ static ssize_t http2_recv(struct connectdata *conn, int
socki
    struct http_conn *httpc = &conn->proto.httpc;

    (void)sockindex; /* we always do HTTP2 on sockindex 0 */

    if(httpc->closed) {
+ httpc->closed = FALSE;
      return 0;
    }

    /* Nullify here because we call nghttp2_session_send() and they
       might refer to the old buffer. */
@@ -745,10 +746,11 @@ static ssize_t http2_recv(struct connectdata *conn, int
socki
      return len - conn->proto.httpc.len;
    }
    /* If stream is closed, return 0 to signal the http routine to close
       the connection */
    if(httpc->closed) {
+ httpc->closed = FALSE;
      return 0;
    }
    *err = CURLE_AGAIN;
    return -1;
  }

-- 
  / daniel.haxx.se
-------------------------------------------------------------------
List admin: http://cool.haxx.se/list/listinfo/curl-library
Etiquette:  http://curl.haxx.se/mail/etiquette.html