Bug 180631

Created attachment 30233 [details]
The text export from Wireshark

I used a Wireshark to capture of HTTP communication between Konqueror and abclinuxu.cz site - http://www.abclinuxu.cz/blog/techblog/2009/1/entropa. The filter value was 'http and ip.addr==195.70.150.7'. You can see there is a big delay between every server response and another Konqueror's GET.

Can confirm that. Very interesting.

I can confirm it too in 4.1.96. Refresh works - page is loaded instantly! 

Created attachment 30275 [details]
Konqueror debug log - waiting

Also I have log after refresh instead waiting so if it helps I can upload too.

Potential fix below. The problem seems to occur because we get the 304 response + headers in the single read. Then, we parse the response, unget the headers, and then when trying to get the headers next do a blocking read. This make it return unget buffer first:

Index: http.cpp
===================================================================
--- http.cpp	(revision 919318)
+++ http.cpp	(working copy)
@@ -1861,6 +1861,8 @@
     m_unreadBuf.clear();
 }
 
+// Note: the implementation of unread/readBuffered assumes that unread will only 
+// be used when there is extra data we don't want to handle, and not to wait for more data.
 void HTTPProtocol::unread(char *buf, size_t size)
 {
     // implement LIFO (stack) semantics
@@ -1886,6 +1888,10 @@
             buf[i] = m_unreadBuf.constData()[bufSize - i - 1];
         }
         m_unreadBuf.truncate(bufSize - bytesRead);
+
+        // if we have an unread buffer, return here, since we may already have enough data to 
+        // complete the response, so we don't want to wait for more.
+        return bytesRead;
     }
     if (bytesRead < size) {
         int rawRead = TCPSlaveBase::read(buf + bytesRead, size - bytesRead);

It works fine now, thanks man!

(Btw, it's the `kioslave/http/http.cpp' file in kdelibs)

@Maksim: thanks for the fix, now it's OK. It's possible to commit it to svn?

SVN commit 926999 by orlovich:

Make sure we don't do any extra blocking reads if we still have buffered data to process; 
fixes freezes on websites with extremely small reply headers, like abclinuxu.cz

BUG:180631


 M  +6 -0      http.cpp  


WebSVN link: http://websvn.kde.org/?view=rev&revision=926999