BufferedInputStream#fill() hanging in SocketInputStream#read0()
Alan.Bateman at oracle.com
Tue Sep 7 02:34:44 PDT 2010
Andreas Kohn wrote:
> in an application that makes heavy use of JAX-RS, and for that reason
> small HTTP connections, I frequently observe situations where
> BufferedInputStream would hang like this:
> "ae03a305-557e-4db6-b9b8-2bf50f056aaf" prio=10 tid=0x00002aab0175a000 nid=0x30dc runnable [0x0000000048cb3000]
> java.lang.Thread.State: RUNNABLE
> at java.net.SocketInputStream.socketRead0(Native Method)
> at java.net.SocketInputStream.read(SocketInputStream.java:129)
> at java.io.BufferedInputStream.fill(BufferedInputStream.java:218)
> at java.io.BufferedInputStream.read1(BufferedInputStream.java:258)
> at java.io.BufferedInputStream.read(BufferedInputStream.java:317)
> - locked <0x00002aaab605b670> (a java.io.BufferedInputStream)
> at sun.net.www.http.HttpClient.parseHTTPHeader(HttpClient.java:687)
> at sun.net.www.http.HttpClient.parseHTTP(HttpClient.java:632)
> at sun.net.www.protocol.http.HttpURLConnection.getInputStream(HttpURLConnection.java:1072)
> Typically when looking into those hangs with a debugger it seems that the HTTP response
> was received fully, but the BufferedIInputStream tries to fill its buffer with more bytes than
> were originally requested by the HttpClient. The HTTP server is now waiting for the client to
> continue sending requests, while the client is stuck, hoping that the server will send more bytes
> to fill its buffer with.
> Currently the only option I have in this case is restart the application (even killing the server
> does not lead to a connection reset for some reason!), which in many cases is quite unacceptable.
> Disabling Keep-Alive connections helps if the hanging occurs when reading response bodies, but not
> when the hanging occurs while the header is being read.
> I believe this essentially the issue 6192696 (BufferedInputStream.read(byte, int, int)
> can block if the entire buffer can't be filled). I understand that this issue had had multiple
> fix-attempts, which could not be done due reliance on a proper and performing #available() implementation.
> But, wouldn't it be enough to just prevent fill() from filling more than the user wanted to
> read in the first place? For the HTTP example the Content-Length gives that information, and the
> server will not send more than that, but on the other hand it is safe for the connection to block
> until those bytes are received.
> Attached is a patch that implements that idea, am I missing something here? Note that I left #read()
> without arguments alone.
> I've been running with this patch for a few weeks now in test environments, and it seems to have no
> negative impact on classloading speed as detailed in the original bug comment trail.
It might be good to bring up the http client part of this discussion on
the net-dev mailing list. In the above stack trace, it looks like the
http client is attempting to look at the first 8 bytes of the response
to verify that it is a valid HTTP response. With your patch then I
assume it will only read up to 8 bytes. Once it gets past this then it
proceeds one byte at a time through the message headers looking for the
end of the header fields. If I understand your changes correctly, then
read() is unchanged, and so reading through the message headers will
result in at least one fill, maybe more. The length of the message
headers isn't known in advance and with a persistent connection then
there isn't an immediate EOF. So my point is that maybe the http client
code should be re-examined and not use a BIS, at least not when
processing the message headers.
More information about the core-libs-dev