Lower layer is missing packets. about picotcp HOT 7 CLOSED

First observation is that Nagle is trying to enqueue full segments.
For example if the queue has 800 bytes free and 1000 bytes segment arrives the algorithm will reject that
when it should enqueue the 800 bytes and let the upper layer know this.

from picotcp.

This is happening as an unfortunate case where

• the output queue can't take another segment anymore
• the hold queue also can't take another segment anymore
• we are waiting for the ACK to bless us with free space in the output queue.

As a solution I would propose setting the pico_err to try again and don't let the write function fail that fast.

from picotcp.

Using pico_err solution failed to fix the issue, the queues still get filled.
The wireshark log looks like there are a lot of retransmissions and the hold queue can't get more frames.
This could be because there is a problem with the way we ack packets (the wireshark shows there is a segment that is constantly retransmitted and even if the other side is sending us the ack packets we fail to ack the segment).

from picotcp.

The behavior seems to be caused because we are not receiving packets anymore.
A breakpoint in ENET_IRQ is not reached.
No more Rx packets, means no more ACKs for tcp.

from picotcp.

The Rx interrupt is not generated anymore because the Rx Dma buffer is full.
(RxProduceIndex + 1) % MaxNumberOfElements = RxConsumeIndex
in this case the DMA rejects frames and expects for us to consume before generating a new Rx Interrupt.

This happens because the queue that holds the events coming from IRQ or application is full.
At this point we can see that we are generating events faster than we process them.(main thread seems to have a higher priority than pico).

I think that this event handling should be avoided, so we don't lose events anymore and even if we do lose events
,to be able to recover from this.

from picotcp.

Tested with a bit different design for the event queue.
With this one no more interrupts are getting lost and the memory test finishes with succes

• with Nagle enabled : ~5 Mbit/s
• without Nagle : ~7 Mbit/s

Making the final changes, this bug is finished.

from picotcp.

Bug squashed by pushes
master 22b0995
master 9f18218
mbed 58:5f6eedbbbd5b.

One less headache, closed.

from picotcp.