Finally got around to reading up on this. At first I was expecting to
see it originally sent one large packet, but from reading up on the
issue it looks like the problem was it sent a packet for each field of
the JDWP header, resulting in too many small packets, and the fix for
this was to coalesce the header into one packet. So this CR doesn't
explain why two packets are sent (when the data is large) instead of one.
Chris
On 2/26/18 8:20 PM, David Holmes wrote:
The two-step send came in with:
https://bugs.openjdk.java.net/browse/JDK-6401245
"Small JDWP packets with the socket transport causes slow debugging on
linux 2.6.15 kernel and newer"
David
-----
On 27/02/2018 9:29 AM, serguei.spit...@oracle.com wrote:
On 2/26/18 15:06, Chris Plummer wrote:
On 2/26/18 3:00 PM, daniil.x.ti...@oracle.com wrote:
On 2/26/18 12:16 PM, Chris Plummer wrote:
On 2/26/18 11:51 AM, daniil.x.ti...@oracle.com wrote:
Hi David and Sergei,
On 2/20/18 10:16 PM, serguei.spit...@oracle.com wrote:
Hi David,
On 2/20/18 20:02, David Holmes wrote:
Hi Daniil,
Good find on this!
What does the actual spec say about the length of things and
how they may be split across multiple packets? Are we
guaranteed that at most two packets will be involved?
The JDWP spec
(https://docs.oracle.com/javase/9/docs/specs/jdwp/jdwp-spec.html)
says nothing about splitting JDWP reply packets at all but the
implementation limits the max number of the sent packets to two
packets max. The implementation is dated back to the initial load
that happened in 2007 and the information about the related Jira
issue is missing.
open/src/jdk.jdwp.agent/share/native/libdt_socket/socketTransport.c
836 data = packet->type.cmd.data;
837 /* Do one send for short packets, two for longer ones */
838 if (data_len <= MAX_DATA_SIZE) {
839 memcpy(header + JDWP_HEADER_SIZE, data, data_len);
840 if (send_fully(socketFD, (char *)&header,
JDWP_HEADER_SIZE + data_len) !=
841 JDWP_HEADER_SIZE + data_len) {
842 RETURN_IO_ERROR("send failed");
843 }
844 } else {
845 memcpy(header + JDWP_HEADER_SIZE, data, MAX_DATA_SIZE);
846 if (send_fully(socketFD, (char *)&header,
JDWP_HEADER_SIZE + MAX_DATA_SIZE) !=
847 JDWP_HEADER_SIZE + MAX_DATA_SIZE) {
848 RETURN_IO_ERROR("send failed");
849 }
850 /* Send the remaining data bytes right out of the data
area. */
851 if (send_fully(socketFD, (char *)data + MAX_DATA_SIZE,
852 data_len - MAX_DATA_SIZE) != data_len -
MAX_DATA_SIZE) {
853 RETURN_IO_ERROR("send failed");
854 }
855 }
Curious. First packet is limited to MAX_DATA_SIZE, 2nd packet has
no size limit. What's the point then of splitting it then? Is
there a desire to get the header transmitted in a smaller packet.
Chris
It looks as the goal was to somehow improve the responsiveness in
case of the large data but I am not sure about this. I could not
locate any traces in Jira related to this implementation.
I was thinking it might be something like that too. Get the header
across the wire quickly. Maybe the user just wants the header (with
size info) initially, and will allocate a large buffer for the rest
if necessary.
It was my guess too.
At least, it is the best explanation for this design that looks
reasonable to me.
Chris
Probably Serguei has some info what is the history behind this design.
I don't know the history here.
This was implemented in very early days, most likely, before JDK 1.5
or even 1.4.2.
Thanks,
Serguei
68 protected byte[] readJdwpString(DataInputStream ds)
throws IOException {
69 byte[] str = null;
70 int len = ds.readInt();
71 if (len > 0) {
72 str = new byte[len];
73 ds.read(str, 0, len);
74 }
might we get a short-read of the string if it is split across
multiple packets?
This and all other reads happen not directly from the socket
input stream but rather from the DataInputStream object that is
constructed in JdwpReply.initFromStream(InputStream) method. With
the proposed fix we do ensure that the created DataInputStream
object contains data from both packets in cases when the reply
was split in two packets.
Nice catch!
Even though this fix is enough to resolve this problem now,
there is a chance,
it can fail in the future when more modules are added to the
platform.
I'm wondering if all these reads should be loops, ensuring we
read the expected amount of data.
Since the implementation of the socket transport limits the max
number of packets the reply might be split in to two packets I
don't think we really need it here.
One further comment - not sure why we need the print out for
when we do read multiple packets?
That would seem to be a debugging aid.
Yes, it helps to understand what happens.
Many tests have a lack of tracing which makes it harder to debug
and understand failures.
That is correct. This additional tracing was added to help to
understand the possible failures in the future.
Thanks,
Serguei
Thanks,
David
Thanks,
Daniil
On 21/02/2018 10:14 AM, Daniil Titov wrote:
Hi Serguei,
A new version of the webrev that has these strings reformatted
is at http://cr.openjdk.java.net/~dtitov/8170541/webrev.02/
Thank you!
Best regards,
Daniil
*From: *"serguei.spit...@oracle.com" <serguei.spit...@oracle.com>
*Date: *Tuesday, February 20, 2018 at 3:00 PM
*To: *Daniil Titov <daniil.x.ti...@oracle.com>,
"serviceability-dev@openjdk.java.net"
<serviceability-dev@openjdk.java.net>
*Subject: *Re: RFR 8170541:
serviceability/jdwp/AllModulesCommandTest.java fails
intermittently on Windows and Solaris
Hi Daniil,
Interesting issue...
Thank you for finding to the root cause so quickly!
The fix looks good.
Could I ask you to reformat these lines to make the L54
shorter ?:
54 System.out.println("[" +
getClass().getName() + "] Only " + bytesRead + " bytes of " +
dataLength +
55 " were read in the first packet.
Reading the rest...");
Thanks,
Serguei
On 2/20/18 09:24, Daniil Titov wrote:
Please review the changes that fix intermittent failure of
serviceability/jdwp/AllModulesCommandTest.java test.
The problem here is that for a large data the JDWP agent
(socketTransport_writePacket() method in
src/jdk.jdwp.agent/share/native/libdt_socket/socketTransport.c )
sends 2 packets and in some cases only the first packet is
received
at the time when the test reads the reply from the JDWP
agent. Since
the test does not check that all data is received in the
first
packet the correlation between commands and replies became
broken
(the unread second packet is read by the next command and
the reply
for the next command is read by the next after next
command and so on).
Bug: https://bugs.openjdk.java.net/browse/JDK-8170541
Webrev: http://cr.openjdk.java.net/~dtitov/8170541/webrev.01
The tests ran successfully with Mach5.
Best regards,
Daniil
