On Tuesday 15 January 2008 16:19, Robert Hailey wrote:
>
> On Jan 12, 2008, at 2:01 PM, Matthew Toseland wrote:
>
> > On Friday 11 January 2008 22:46, Robert Hailey wrote:
> >>
> >> On Jan 11, 2008, at 4:06 PM, Matthew Toseland wrote:
> >>
> >>> On Wednesday 09 January 2008 22:10, robert at freenetproject.org wrote:
> >>>> Author: robert
> >>>> Date: 2008-01-09 22:10:36 +0000 (Wed, 09 Jan 2008)
> >>>> New Revision: 16987
> >>>>
> >>>> Modified:
> >>>> trunk/freenet/src/freenet/io/xfer/BlockTransmitter.java
> >>>> Log:
> >>>> don't assume everything has been transmitted when the transmitter
> >>>> oversleeps
> >>>> - should get rid of "haven't heard from receiver in 1m55s." message
> >>>> - the transmitter still oversleeps (2 minutes?!), and is not woken
> >>>> up by
> >>> notifyAll() because it is in the throttle function: delay()
> >>>
> >>> Why does the below help?
> >>
> >> I think it was actually because delay() was deadlocking, not
> >> oversleeping. In any case, this helps then because we will not then
> >> see an old timeAllSent and assume that the receiver is dead ("haven't
> >> heard from the receiver").
> >
> > I still don't understand this - timeAllSent isn't set until not only
> > have we
> > sent all the packets but we've also received all the packets.
> > Therefore,
> > packetReceived() won't be called after that, because
> > packetReceived() means
> > we've received a new packet from the previous sender. I can see that
> > we'd
> > want to reset it in some other cases, but not here.
>
> You are 99.9% right. Logically it should be totally redundant (if ever
> we are receiving a packet, we have not sent everything), but this
> actually can have an effect if we receive duplicate packets (which due
> to the requeueing bug & retransmit requests was possible). Arguably it
> would be better to not call the callback on receiving a duplicate
> packet, but both of these fixes are overkill for a problem about to go
> away (the only reason I could see that the BlockReceiver would be
> asking for a missing packet is if it was endlessly requeued).
Well, we really don't want it being sent twice to impact on downstream nodes,
so the correct fix is still to not call the callback on receiving a
duplicate, which iirc is trivial, isn't it, since we have a bitmap of
received blocks in PartiallyReceivedBlock?
>
> On a side note, it may be the case that the block transmitter threads
> are not properly exiting in the trunk, I need to take a closer look at
> this before the next build.
Please do.
>
> --
> Robert Hailey
>
>
> >>>> Modified: trunk/freenet/src/freenet/io/xfer/BlockTransmitter.java
> >>>> ===================================================================
> >>>> --- trunk/freenet/src/freenet/io/xfer/BlockTransmitter.java
> >>>> 2008-01-09
> >>> 18:33:46 UTC (rev 16986)
> >>>> +++ trunk/freenet/src/freenet/io/xfer/BlockTransmitter.java
> >>>> 2008-01-09
> >>> 22:10:36 UTC (rev 16987)
> >>>> @@ -192,6 +192,7 @@
> >>>> public void packetReceived(int
> >>>> packetNo) {
> >>>> synchronized(_senderThread) {
> >>>> _unsent.addLast(new
> >>>> Integer(packetNo));
> >>>> + timeAllSent =
> >>>> -1;
> >>>>
> >>>> _sentPackets.setBit(packetNo, false);
> >>>>
> >>>> _senderThread.notifyAll();
> >>>> }
> >>>
> _______________________________________________
> Devl mailing list
> Devl at freenetproject.org
> http://emu.freenetproject.org/cgi-bin/mailman/listinfo/devl
>
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL:
<https://emu.freenetproject.org/pipermail/devl/attachments/20080115/3870757b/attachment.pgp>
