Nagle appears to be a significant contributor to latency now that the static
sleeps are gone. Most of our messages are relatively large compared to
IP + TCP so I do not expect this to create enormous overhead.
This may also reduce traffic burstyness somewhat.
- to match the peers.dat handling also supply a debug.log entry for how
many entries were loaded from banlist.dat and how long it took
- add a GUI init message for loading the banlist (same as with peers.dat)
- move the same message for peers.dat upwards in the code, to be able to
reuse the timing variable nStart and also just log, if our read from
peers.dat didn't fail
- only start working on/with banlist data, if reading in the banlist from
disk didn't fail
- as CNode::setBannedIsDirty is false (default) when reading fails, we
don't need to explicitly set it to false to prevent writing
banlist.dat in that case either
`nMinPingUsecTime` was left uninitialized in CNode.
The correct initialization for a minimum-until-now is int64_t's max value, so initialize it to that.
Thanks @MarcoFalke for noticing.
When running the rpc tests in Wine, nodes often fail to listen on localhost
due to a stale socket from a previous run. This aligns the behavior with other
platforms.
While CBloomFilter is usually used with an explicitly set nTweak,
CRollingBloomFilter is only used internally. Requiring every caller to
set nTweak is error-prone and redundant; better to have the class handle
that for you with a high-quality randomness source.
Additionally when clearing the filter it makes sense to change nTweak as
well to recover from a bad setting, e.g. due to insufficient randomness
at initialization, so the clear() method is replaced by a reset() method
that sets a new, random, nTweak value.
- rpcwallet: No need to lock twice here
- openssl: Clang doesn't understand selective lock/unlock here. Ignore it.
- CNode: Fix a legitimate (though very unlikely) locking bug.
This sets aside a number of connection slots for whitelisted peers,
useful for ensuring your local users and miners can always get in,
even if your limit on inbound connections has already been reached.
Use a probabilistic bloom filter to keep track of which addresses
we think we have given our peers, instead of a list.
This uses much less memory, at the cost of sometimes failing to
relay an address to a peer-- worst case if the bloom filter happens
to be as full as it gets, 1-in-1,000.
Measured memory usage of a full mruset setAddrKnown: 650Kbytes
Constant memory usage of CRollingBloomFilter addrKnown: 37Kbytes.
This will also help heap fragmentation, because the 37K of storage
is allocated when a CNode is created (when a connection to a peer
is established) and then there is no per-item-remembered memory
allocation.
I plan on testing by restarting a full node with an empty peers.dat,
running a while with -debug=addrman and -debug=net, and making sure
that the 'addr' message traffic out is reasonable.
(suggestions for better tests welcome)
Gregory Maxwell