net: don't preserve 1xx responses in parser buffer.

net/http/http_stream_parser.h

Index: net/http/http_stream_parser.h
diff --git a/net/http/http_stream_parser.h b/net/http/http_stream_parser.h
index a41e393b485f3fd6f6e758777c80988e82e7bd76..43e1514c2fecef2bee58ab4772c47b0b73b3bc99 100644
--- a/net/http/http_stream_parser.h
+++ b/net/http/http_stream_parser.h
@@ -178,7 +178,8 @@ class NET_EXPORT_PRIVATE HttpStreamParser {
   scoped_refptr<GrowableIOBuffer> read_buf_;
 
   // Offset of the first unused byte in |read_buf_|.  May be nonzero due to
-  // a 1xx header, or body data in the same packet as header data.
+  // body data in the same packet as header data but is zero when reading
+  // headers.
   int read_buf_unused_offset_;

[SkyLined, 2013/09/30 18:27:47]

This may be a stretch, but since this is only used to refer to the start of the
body in the buffer, could we:
- rename this "read_buf_body_offset_",
- set it to some invalid constant (eg. -1) while reading headers,
- set it to a non-zero positive offset only once headers have been read and the
response is expected to contain a body?
I think that would make the code easier to understand and the DCHECKS more
obvious.

[agl, 2013/09/30 20:59:28]

I haven't done this for the following reasons:

1) It's not the offset of the body, it's the current offset: it's moved as we
read the body in DoReadBody. I would expect the proposed name to be used for a
variable that's constant.

2) I haven't switched the magic value to -1 because DoReadHeadersComplete can
move to STATE_READ_BODY_COMPLETE in the event of truncated headers and
DoReadBodyComplete does arithmetic with read_buf_unused_offset_ which is easier
if it's zero in the unused case.

 
   // The amount beyond |read_buf_unused_offset_| where the status line starts;