* adding raises defect across the codebase
* use unittest2
* add windows deps caching
* update mingw link
* die on failed peerinfo initialization
* use result.expect instead of get
* use expect more consistently and rework inits
* use expect more consistently
* throw on missing public key
* remove unused closure annotation
* merge master
* gossipsub: unsubscribe fixes
* fix KeyError when updating metric of unsubscribed topic
* fix unsubscribe message not being sent to all peers causing them to
keep thinking we're still subscribed
* release memory earlier in a few places
* floodsub fix
In `async` functions, a closure environment is created for variables
that cross an await boundary - this closure environment is kept in
memory for the lifetime of the associated future - this means that
although _some_ variables are no longer used, they still take up memory
for a long time.
In Nimbus, message validation is processed in batches meaning the future
of an incoming gossip message stays around for quite a while - this
leads to memory consumption peaks of 100-200 mb when there are many
attestations in the pipeline.
To avoid excessive memory usage, it's generally better to move non-async
code into proc's such that the variables therein can be released earlier
- this includes the many hidden variables introduced by macro and
template expansion (ie chronicles that does expensive exception
handling)
* move seen table salt to floodsub, use there as well
* shorten seen table salt to size of hash
* avoid unnecessary memory allocations and copies in a few places
* factor out message scoring
* avoid reencoding outgoing message for every peer
* keep checking validators until reject (in case there's both reject and
ignore)
* `readOnce` avoids `readExactly` overhead for single-byte read
* genericAssign -> assign2
* properly propagate initiator information for gossipsub
* Fix pubsubpeer lifetime management
* restore old behavior
* tests fixing
* clamp backoff time value received
* fix member name collisions
* internal test fixes
* better names and explaining of the importance of transport direction
* fixes
* Refactor gossipsub into multiple modules
* splitup further gossipsub
* move more mesh related stuff to behavior
* fix internal tests
* fix PubSubPeer.outbound flag, make it more reliable
* use discard rather then _
* Remove unused connections in pubsubpeer, also removed wrong usages, add a disconnect bad peers parameter
* handle exceptions in disconnectPeer
* small fix
* use the proper disconnection procedure for gossip peers
* fixes, more metrics add test about disconnection
* hot fix possible null pointers in switch
* silly isnil sugar
* Fix and test gossip directPeer connections
To break a potential read/write deadlock, gossipsub uses an unbounded
queue for writes - when peers are too slow to process this queue, it may
end up growing without bounds causing high memory usage.
Here, we introduce a maximum write queue length after which the peer is
disconnected - the queue is generous enough that any "normal" usage
should be fine - writes that are `await`:ed are not affected, only
writes that are launched in an `asyncSpawn` task or similar.
* avoid unnecessary copy of message when there are no send observers
* release message memory earlier in gossipsub
* simplify pubsubpeer logging
When messages can't be sent to peer, we try to establish a send
connection - this causes messages to stack up as more and more unsent
messages are blocked on the dial lock.
* remove dial lock
* run reconnection loop in background task
This change modifies how the backpressure algorithm in bufferstream
works - in particular, instead of working byte-by-byte, it will now work
seq-by-seq.
When data arrives, it usually does so in packets - in the current
bufferstream, the packet is read then split into bytes which are fed one
by one to the bufferstream. On the reading side, the bytes are popped of
the bufferstream, again byte by byte, to satisfy `readOnce` requests -
this introduces a lot of synchronization traffic because the checks for
full buffer and for async event handling must be done for every byte.
In this PR, a queue of length 1 is used instead - this means there will
at most exist one "packet" in `pushTo`, one in the queue and one in the
slush buffer that is used to store incomplete reads.
* avoid byte-by-byte copy to buffer, with synchronization in-between
* reuse AsyncQueue synchronization logic instead of rolling own
* avoid writeHandler callback - implement `write` method instead
* simplify EOF signalling by only setting EOF flag in queue reader (and
reset)
* remove BufferStream pipes (unused)
* fixes drainBuffer deadlock when drain is called from within read loop
and thus blocks draining
* fix lpchannel init order
* mcache fixes
* remove timed cache - the window shifting already removes old messages
* ref -> object
* avoid unnecessary allocations with `[]` operator
* simplify init
* fix several gossipsub/floodsub issues
* floodsub, gossipsub: don't rebroadcast messages that fail validation
(!)
* floodsub, gossipsub: don't crash when unsubscribing from unknown
topics (!)
* gossipsub: don't send message to peers that are not interested in the
topic, when messages don't share topic list
* floodsub: don't repeat all messages for each message when
rebroadcasting
* floodsub: allow sending empty data
* floodsub: fix inefficient unsubscribe
* sync floodsub/gossipsub logging
* gossipsub: include incoming messages in mcache (!)
* gossipsub: don't rebroadcast already-seen messages (!)
* pubsubpeer: remove incoming/outgoing seen caches - these are already
handled in gossipsub, floodsub and will cause trouble when peers try to
resubscribe / regraft topics (because control messages will have same
digest)
* timedcache: reimplement without timers (fixes timer leaks and extreme
inefficiency due to per-message closures, futures etc)
* timedcache: ref -> obj
* remove send lock
When mplex receives data it will block until a reader has processed the
data. Thus, when a large message is received, such as a gossipsub
subscription table, all of mplex will be blocked until all reading is
finished.
However, if at the same time a `dial` to establish a gossipsub send
connection is ongoing, that `dial` will be blocked because mplex is no
longer reading data - specifically, it might indeed be the connection
that's processing the previous data that is waiting for a send
connection.
There are other problems with the current code:
* If an exception is raised, it is not necessarily raised for the same
connection as `p.sendConn`, so resetting `p.sendConn` in the exception
handling is wrong
* `p.isConnected` is checked before taking the lock - thus, if it
returns false, a new dial will be started. If a new task enters `send`
before dial is finished, it will also determine `p.isConnected` is
false, then get stuck on the lock - when the previous task finishes and
releases the lock, the new task will _also_ dial and thus reset
`p.sendConn` causing a leak.
* prefer existing connection
simplifies flow
* move pubsub of off switch, pass switch into pubsub
* use join on lpstreams
* properly cleanup up failed peers
* fix tests
* fix peertable hasPeerId
* fix tests
* rework sending, remove helpers from pubsubpeer, unify in broadcast
* further split broadcast into send
* use send where appropriate
* use formatIt
* improve trace
Co-authored-by: Giovanni Petrantoni <giovanni@fragcolor.xyz>
Tanguy
diegomrsantos
Miran
tersec
Tanguy
Dmitriy Ryajov
Jacek Sieka
Giovanni Petrantoni
Giovanni Petrantoni
Eugene Kabanov
Zahary Karadjov