nodeDB.querySeeds was not safe for concurrent use but could be called
concurrenty on multiple goroutines in the following case:
- the table was empty
- a timed refresh started
- a lookup was started and initiated refresh
These conditions are unlikely to coincide during normal use, but are
much more likely to occur all at once when the user's machine just woke
from sleep. The root cause of the issue is that querySeeds reused the
same leveldb iterator until it was exhausted.
This commit moves the refresh scheduling logic into its own goroutine
(so only one refresh is ever active) and changes querySeeds to not use
a persistent iterator. The seed node selection is now more random and
ignores nodes that have not been contacted in the last 5 days.
PR #1621 changed Table locking so the mutex is not held while a
contested node is being pinged. If multiple nodes ping the local node
during this time window, multiple ping packets will be sent to the
contested node. The changes in this commit prevent multiple packets by
tracking whether the node is being replaced.
If the timeout fired (even just nanoseconds) before the deadline of the
next pending reply, the timer was not rescheduled. The timer would've
been rescheduled anyway once the next packet was sent, but there were
cases where no next packet could ever be sent due to the locking issue
fixed in the previous commit.
As timing-related bugs go, this issue had been present for a long time
and I could never reproduce it. The test added in this commit did
reproduce the issue on about one out of 15 runs.
Table.mutex was being held while waiting for a reply packet, which
effectively made many parts of the whole stack block on that packet,
including the net_peerCount RPC call.
Lookup calls would spin out of control when network connectivity was
lost. The throttling that was in place only took effect when the table
returned zero results, which doesn't happen very often.
The new throttling should not have a negative impact when the host is
online. Lookups against the network take some time and dials for all
results must complete or hit the cache before a new one is started. This
usually takes longer than four seconds, leaving online lookups
unaffected.
Fixes#1296
As of this commit, we no longer rely on the protocol handler to report
write errors in a timely fashion. When a write fails, shutdown is
initiated immediately and no new writes can start. This will also
prevent new writes from starting after Server.Stop has been called.
The most visible change is event-based dialing, which should be an
improvement over the timer-based system that we have at the moment.
The dialer gets a chance to compute new tasks whenever peers change or
dials complete. This is better than checking peers on a timer because
dials happen faster. The dialer can now make more precise decisions
about whom to dial based on the peer set and we can test those
decisions without actually opening any sockets.
Peer management is easier to test because the tests can inject
connections at checkpoints (after enc handshake, after protocol
handshake).
Most of the handshake stuff is now part of the RLPx code. It could be
exported or move to its own package because it is no longer entangled
with Server logic.
The previous limit was 10MB which is unacceptable for all kinds
of reasons, the most important one being that we don't want to
allow the remote side to make us allocate 10MB at handshake time.
The returned reason is currently not used except for the log
message. This change makes the log messages a bit more useful.
The handshake code also returns the remote reason.
People stil get confused about the messages. This commit changes
the levels so that the only thing printed at the default level (info)
is a successful mapping.
The test listens for multicast UDP packets on the default interface
because I couldn't get it to work reliably on loopback without massive
changes to goupnp. This means that the test might fail when there is a
UPnP-enabled router attached on that interface. I checked that locally
by looping the test and it passes reliably because the local SSDP server
always responds faster.
Concurrent calls to Interface methods on autodisc could return a "not
discovered" error if the discovery did not finish before the call.
autodisc.wait expected the done channel to carry the found Interface
but it was closed instead.
The fix is to use sync.Once for now, which is easier to get right.
And there is a test. Finally.
This will have to change again when we introduce re-discovery.
We don't have a UDP which specifies any messages that will be 4KB. Aside from being implemented for months and a necessity for encryption and piggy-backing packets, 1280bytes is ideal, and, means this TODO can be completed!
Why 1280 bytes?
* It's less than the default MTU for most WAN/LAN networks. That means fewer fragmented datagrams (esp on well-connected networks).
* Fragmented datagrams and dropped packets suck and add latency while OS waits for a dropped fragment to never arrive (blocking readLoop())
* Most of our packets are < 1280 bytes.
* 1280 bytes is minimum datagram size and MTU for IPv6 -- on IPv6, a datagram < 1280bytes will *never* be fragmented.
UDP datagrams are dropped. A lot! And fragmented datagrams are worse. If a datagram has a 30% chance of being dropped, then a fragmented datagram has a 60% chance of being dropped. More importantly, we have signed packets and can't do anything with a packet unless we receive the entire datagram because the signature can't be verified. The same is true when we have encrypted packets.
So the solution here to picking an ideal buffer size for receiving datagrams is a number under 1400bytes. And the lower-bound value for IPv6 of 1280 bytes make's it a non-decision. On IPv4 most ISPs and 3g/4g/let networks have an MTU just over 1400 -- and *never* over 1500. Never -- that means packets over 1500 (in reality: ~1450) bytes are fragmented. And probably dropped a lot.
Just to prove the point, here are pings sending non-fragmented packets over wifi/ISP, and a second set of pings via cell-phone tethering. It's important to note that, if *any* router between my system and the EC2 node has a lower MTU, the message would not go through:
On wifi w/normal ISP:
localhost:Debug $ ping -D -s 1450 52.6.250.242
PING 52.6.250.242 (52.6.250.242): 1450 data bytes
1458 bytes from 52.6.250.242: icmp_seq=0 ttl=42 time=104.831 ms
1458 bytes from 52.6.250.242: icmp_seq=1 ttl=42 time=119.004 ms
^C
--- 52.6.250.242 ping statistics ---
2 packets transmitted, 2 packets received, 0.0% packet loss
round-trip min/avg/max/stddev = 104.831/111.918/119.004/7.087 ms
localhost:Debug $ ping -D -s 1480 52.6.250.242
PING 52.6.250.242 (52.6.250.242): 1480 data bytes
ping: sendto: Message too long
ping: sendto: Message too long
Request timeout for icmp_seq 0
ping: sendto: Message too long
Request timeout for icmp_seq 1
Tethering to O2:
localhost:Debug $ ping -D -s 1480 52.6.250.242
PING 52.6.250.242 (52.6.250.242): 1480 data bytes
ping: sendto: Message too long
ping: sendto: Message too long
Request timeout for icmp_seq 0
^C
--- 52.6.250.242 ping statistics ---
2 packets transmitted, 0 packets received, 100.0% packet loss
localhost:Debug $ ping -D -s 1450 52.6.250.242
PING 52.6.250.242 (52.6.250.242): 1450 data bytes
1458 bytes from 52.6.250.242: icmp_seq=0 ttl=42 time=107.844 ms
1458 bytes from 52.6.250.242: icmp_seq=1 ttl=42 time=105.127 ms
1458 bytes from 52.6.250.242: icmp_seq=2 ttl=42 time=120.483 ms
1458 bytes from 52.6.250.242: icmp_seq=3 ttl=42 time=102.136 ms
With the introduction of static/trusted nodes, the peer count
can go above MaxPeers. Update the capacity check to handle this.
While here, decouple the trusted nodes check from the handshake
by passing a function instead.
The previous metric was pubkey1^pubkey2, as specified in the Kademlia
paper. We missed that EC public keys are not uniformly distributed.
Using the hash of the public keys addresses that. It also makes it
a bit harder to generate node IDs that are close to a particular node.
