eth/downloader: fix the stall checks/drops during sync

[karalabe]

This is a fix for an interesting corner case regression introduced by the EDGE release. Our code before the EDGE release tried to estimate the capacity of a peer and fetch data proportional to it. If the retrieval timed out, it assumed we were a bit over zealous with te request and reduced the peer capacity to zero (i.e. 1 data item per request). If this also timed out afterwards, the peer was deemed useless beyond retention (during sync, which needs performance) and dropped it.

The EDGE release introduced a fancier throughput and latency measurement algo, which among other constantly tries to request just a bit more than the capacity. This way it can correctly separate the latency from the bandwidth. This means however, that it will never request just one data item, rather always a minimum of two. Our stall checker code however only assumed 1 to be stalling. So EDGE effectively disabled the stall drop.

The big problem with this is that if I connect to two very bad peers, which time out always, then instead of dropping them off eventually, we switch between one or the other, trying to request the same 2 data items will infinity. We'll never break out of this loop as long as they are there, blocking up sync.

[robotally]

Vote	Count	Reviewers
👍	0
👎	0

Updated: Mon Jul 25 10:08:07 UTC 2016

[karalabe]

@fjl Please review this sooner rather than later :)

[fjl]

👍