Implement LLMQ based InstantSend #2735

codablock · 2019-03-01T08:02:43Z

This PR includes an initial implementation of LLMQ based InstantSend. It is meant to be present in parallel to the old InstantSend system.

Switching between the old and the new system happens via SPORK_2_INSTANTSEND_ENABLED, which became a new meaning. When set to 0 (as on testnet/mainet right now), it will use the old system. When set to 1, it will use the new system. When set to anything else, it's considered disabled. This is only meant to be temporary until the old system is removed, at which time we can revert back to the old meaning of spork 2.

The new system is very likely not going to be ready for v14, so we won't activate it. Reason is that more testing is required, especially in regard to performance and load.

The new system has a few differences to the old system:

There is no explicit lock request required. All TXs are meant to be ixlocked, no matter how many inputs or how much funds they move. We might need to add limits to the amount of inputs, or add some kind of prioritization logic that prefers to sign small TXs first, so that DDoS is not that easily pulled off.
The minimum number of confirmations required for TX inputs is lifted when the input is part of a ChainLocked chain (it's enough to have a CLSIG for the current tip). In case inputs were too young and there were no ChainLocks, the system will retry later when with confirmations or when a ChainLock comes in.
The system will also lock TXs for which inputs are not confirmed on-chain, but only when the parent TXs are in the local mempool and also locked by InstantSend. This means that InstantSend can be chained now! In case the conditions are not met, the system will retry locking later when it determines that parents got locked.
ixlocks do not timeout anymore. They are only removed from the system when TXs get confirmed through ChainLocked blocks. This gives some risk in regard to RAM filling up, but only when ChainLocks are disabled. I will also add (cached) persistence for ixlocks in a later PR, so that the risk of filled RAM is removed.

I did quite a bit of load testing on private devnets and also plan to make a public devnet in the next days so that people can play around. Results of the load testing so far:

When there is no load on the network, ixlocks get created in 2-3 seconds, even with simulated network latency and low avg bandwidth
When there is high load, for example 1000 TXs sent in a few seconds, it takes around 35 seconds to lock all TXs. When latency (e.g. 50ms) and low bandwidth (e.g. 2mbit) is simulated, it takes about 90 seconds for 1000 TXs to get IX confirmed. I'm not very happy with these numbers, and it looks like we'll need a rewrite of the signing session code (hint: UDP + direkt push instead of custom inv system). This is the main reason I believe that LLMQ based InstantSend might not be ready when v14 is deployed.
Bandwidth usage is pretty high. I reduced it to the absolute minimum with the optimizations that got merged in the last days, but IMHO its still too high. Only solution I see right now is what I mentioned in 2. already (UDP).

As an alternative to the signing session code rewrite for UDP and targeting v15 (or whatever the next version would be), another option is to add another minimum TX fee for large TXs. This would allow to add LLMQ based InstantSend to v14 and then later optimize it so that all TXs can use it, independent from their size.

UdjinM6

A few quick comments

src/Makefile.am

src/chainparams.cpp

src/llmq/quorums_instantx.cpp

codablock · 2019-03-01T15:53:29Z

Forgot to mention in the PR description that this PR also changes the behavior of ChainLocks when the new InstantSend system is activated. It will only try to ChainLock the tip when all included TXs were locked before or are old enough to be sure that there won't be conflicting InstantSend locks popping up.

This disincentives miners to include TXs that are not publicly known/propagated for now, as they risk creating a block that never gets ChainLocked. I'll later implement InstantSend locking of transactions that were published through mined blocks, which will then allow retroactive ChainLocking of the affected blocks.

UdjinM6

a couple more

src/dsnotificationinterface.h

src/llmq/quorums_chainlocks.cpp

src/llmq/quorums_instantsend.cpp

src/llmq/quorums_instantsend.h

src/llmq/quorums_signing.cpp

codablock · 2019-03-07T05:42:25Z

Rebased on develop and force-pushed

We can reconstruct recovered sigs from other P2P messages to avoid re-validation of those. We will do this later in InstantSend code.

This also includes handling of TXs that were previously orphanced

The new system does not require explicit lock requests, so we downgrade TXLOCKREQUEST to TX and start propagating it instead of the original.

…g a lock

…ndler

Later commits will introduce checks for "safe TXs" which might abort the signing on first try, but succeed a few seconds later, so we periodically retry to sign the tip.

Safe means that the TX is either ixlocked or known since at least 10 minutes. Also change miner code to only include safe TXs in block templates.

…/old system

…d fix CheckCanLock to respect mempool limits)

Return value is unused and the method actually never returned something.

…l.py Otherwise we overload Travis and tests start to randomly fail.

Otherwise we'll miss checks for ancestors.

…moveFinalISLock

UdjinM6

👍

ACK

UdjinM6 reviewed Mar 1, 2019

View reviewed changes

src/Makefile.am Outdated Show resolved Hide resolved

src/chainparams.cpp Outdated Show resolved Hide resolved

src/llmq/quorums_instantx.cpp Outdated Show resolved Hide resolved

codablock force-pushed the pr_llmq_instantsend branch from b650a05 to 72d220c Compare March 1, 2019 15:12

codablock added this to the 14.0 milestone Mar 1, 2019

UdjinM6 reviewed Mar 1, 2019

View reviewed changes

src/dsnotificationinterface.h Outdated Show resolved Hide resolved

src/llmq/quorums_chainlocks.cpp Outdated Show resolved Hide resolved

src/llmq/quorums_instantsend.cpp Outdated Show resolved Hide resolved

src/llmq/quorums_instantsend.h Outdated Show resolved Hide resolved

UdjinM6 reviewed Mar 4, 2019

View reviewed changes

src/llmq/quorums_signing.cpp Outdated Show resolved Hide resolved

codablock force-pushed the pr_llmq_instantsend branch from c141700 to d185f90 Compare March 4, 2019 14:24

codablock force-pushed the pr_llmq_instantsend branch from 991c4e6 to 2da8d74 Compare March 7, 2019 05:42

codablock added 21 commits March 7, 2019 21:14

Introduce NotifyChainLock signal and invoke it when CLSIGs get processed

2bbac8f

Implement PushReconstructedRecoveredSig in CSigningManager

5bbc122

We can reconstruct recovered sigs from other P2P messages to avoid re-validation of those. We will do this later in InstantSend code.

Implement CInstantSendManager and related P2P messages

83dbcc4

Handle incoming TXs by calling CInstantXManager::ProcessTx

1959f3e

This also includes handling of TXs that were previously orphanced

Downgrade TXLOCKREQUEST to TX when new IX system is active

5ff4db0

The new system does not require explicit lock requests, so we downgrade TXLOCKREQUEST to TX and start propagating it instead of the original.

Disable old IX code when the new system is active

34a8b29

Show "verified via LLMQ based InstantSend" in GUI TX status

3a6cc2c

Whenever we check for locked TXs, also check for the new system havin…

1d2d370

…g a lock

Also call ProcessTx from sendrawtransaction and RelayWalletTransaction

68cfdc9

Disable explicit lock requests when the new IX system is active

dc97835

Enforce IX locks from the new system

7945192

Track txids of new blocks and first-seen time of TXs in CChainLocksHa…

bd7edc8

…ndler

Move ChainLock signing into TrySignChainTip and call it periodically

0a5e8eb

Later commits will introduce checks for "safe TXs" which might abort the signing on first try, but succeed a few seconds later, so we periodically retry to sign the tip.

Cheaper/Faster bailout from TrySignChainTip when already signed before

96291e7

Only sign ChainLocks when all included TXs are "safe"

2a7a5c6

Safe means that the TX is either ixlocked or known since at least 10 minutes. Also change miner code to only include safe TXs in block templates.

Fix no-wallet build

baf8b81

Fix crash in BlockAssembler::addPackageTxs

f44f09c

Use scheduleFromNow instead of schedule+boost::chrono

5b8344e

Set llmqForInstantSend for mainnet and testnet

e2f99f4

Add override keywork to CDSNotificationInterface::NotifyChainLock

9e90811

Combine loops in CChainLocksHandler::NewPoWValidBlock

2806907

codablock and others added 12 commits March 7, 2019 21:15

Introduce spork SPORK_20_INSTANTSEND_LLMQ_BASED to switch between new…

f5dcb00

…/old system

Rename IXLOCK to ISLOCK and InstantX to InstantSend

2299ee2

Move IS block filtering into ConnectBlock

55152cb

update p2p-instantsend.py to test both "old" and "new" InstantSend

a8da11a

update p2p-autoinstantsend.py to test both "old" and "new" InstantSend

843b6d7

update autoix-mempool.py to test both "old" and "new" InstantSend (an…

41a71fe

…d fix CheckCanLock to respect mempool limits)

Adjust LLMQ based InstantSend tests for new spork20

3e60d2d

Sync blocks after generating in mine_quorum()

f8f867a

Let ProcessPendingReconstructedRecoveredSigs return void instead of bool

fae33e0

Return value is unused and the method actually never returned something.

Completely disable InstantSend while filling mempool in autoix-mempoo…

4d3365d

…l.py Otherwise we overload Travis and tests start to randomly fail.

Move safe TX checks into TestForBlock and TestPackageTransactions

041a1c2

Otherwise we'll miss checks for ancestors.

Actually remove from finalInstantSendLocks in CInstantSendManager::Re…

06fc655

…moveFinalISLock

codablock force-pushed the pr_llmq_instantsend branch from 8d7bcc6 to 06fc655 Compare March 7, 2019 20:15

UdjinM6 approved these changes Mar 7, 2019

View reviewed changes

codablock merged commit 64ae912 into dashpay:develop Mar 8, 2019

codablock deleted the pr_llmq_instantsend branch March 8, 2019 10:32

UdjinM6 mentioned this pull request Mar 9, 2019

Backports 0.15 pr7 #2727

Merged

codablock mentioned this pull request Mar 12, 2019

[0.13.x] Backport: Move IS block filtering into ConnectBlock #2766

Merged

UdjinM6 added the P2P Some notable changes on p2p level label Mar 20, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Implement LLMQ based InstantSend #2735

Implement LLMQ based InstantSend #2735

Uh oh!

codablock commented Mar 1, 2019

Uh oh!

UdjinM6 left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

codablock commented Mar 1, 2019 •

edited

Loading

Uh oh!

UdjinM6 left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

codablock commented Mar 7, 2019

Uh oh!

UdjinM6 left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Implement LLMQ based InstantSend #2735

Implement LLMQ based InstantSend #2735

Uh oh!

Conversation

codablock commented Mar 1, 2019

Uh oh!

UdjinM6 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

codablock commented Mar 1, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

UdjinM6 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

codablock commented Mar 7, 2019

Uh oh!

UdjinM6 left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codablock commented Mar 1, 2019 •

edited

Loading