Headline
GHSA-w24w-wp77-qffm: CometBFT may duplicate transactions in the mempool's data structures
Impact
The mempool maintains two data structures to keep track of outstanding transactions: a list and a map. These two data structures are supposed to be in sync all the time in the sense that the map tracks the index (if any) of the transaction in the list.
Unfortunately, it is possible to have them out of sync. When this happens, the list may contain several copies of the same transaction. Because the map tracks a single index, it is then no longer possible to remove all the copies of the transaction from the list. This happens even if the duplicated transaction is later committed in a block. The only way to remove the transaction is by restarting the node.
These are the steps to cause the above duplication problem. Everything should happen within one height, that is no FinalizeBlock or BeginBlock ABCI calls should happen while these steps are reproduced:
- send transaction tx1 to the target full node via RPC
- send N more different transactions to the target full node, where N should be higher than the node’s configured value for
cache_sizeinconfig.toml - send transaction tx1 again to the target full node
One of the copies of tx1 is now stuck in the mempool’s data structures. Effectively causing a memory leak, and having that node gossiping that transaction to its peers forever.
The above problem can be repeated on and on until a sizable number of transactions are stuck in the mempool, in order to try to bring down the target node.
This problem is present in releases: v0.37.0, and v0.37.1, as well as in v0.34.28, and all previous releases of the CometBFT repo. It will be fixed in releases v0.34.29 and v0.37.2.
Patches
The PR containing the fix is here.
Workarounds
- Increasing the value of
cache_sizeinconfig.tomlmakes it very difficult to effectively attack a full node. - Not exposing the transaction submission RPC’s would mitigate the probability of a successful attack, as the attacker would then have to create a modified (byzantine) full node to be able to perform the attack via p2p.
References
Impact
The mempool maintains two data structures to keep track of outstanding transactions: a list and a map.
These two data structures are supposed to be in sync all the time in the sense that the map tracks the index (if any) of the transaction in the list.
Unfortunately, it is possible to have them out of sync. When this happens, the list may contain several copies of the same transaction.
Because the map tracks a single index, it is then no longer possible to remove all the copies of the transaction from the list.
This happens even if the duplicated transaction is later committed in a block.
The only way to remove the transaction is by restarting the node.
These are the steps to cause the above duplication problem. Everything should happen within one height, that is no FinalizeBlock or BeginBlock ABCI calls should happen while these steps are reproduced:
- send transaction tx1 to the target full node via RPC
- send N more different transactions to the target full node, where N should be higher than the node’s configured value for cache_size in config.toml
- send transaction tx1 again to the target full node
One of the copies of tx1 is now stuck in the mempool’s data structures. Effectively causing a memory leak, and having that node gossiping that transaction to its peers forever.
The above problem can be repeated on and on until a sizable number of transactions are stuck in the mempool, in order to try to bring down the target node.
This problem is present in releases: v0.37.0, and v0.37.1, as well as in v0.34.28, and all previous releases of the CometBFT repo. It will be fixed in releases v0.34.29 and v0.37.2.
Patches
The PR containing the fix is here.
Workarounds
- Increasing the value of cache_size in config.toml makes it very difficult to effectively attack a full node.
- Not exposing the transaction submission RPC’s would mitigate the probability of a successful attack, as the attacker would then have to create a modified (byzantine) full node to be able to perform the attack via p2p.
References
- PR that introduced the map to track transactions in the mempool.
- PR containing the fix.
References
- GHSA-w24w-wp77-qffm
- https://nvd.nist.gov/vuln/detail/CVE-2023-34451
- cometbft/cometbft#890
- tendermint/tendermint#2778
Related news
CometBFT is a Byzantine Fault Tolerant (BFT) middleware that takes a state transition machine and replicates it on many machines. The mempool maintains two data structures to keep track of outstanding transactions: a list and a map. These two data structures are supposed to be in sync all the time in the sense that the map tracks the index (if any) of the transaction in the list. In `v0.37.0`, and `v0.37.1`, as well as in `v0.34.28`, and all previous releases of the CometBFT repo2, it is possible to have them out of sync. When this happens, the list may contain several copies of the same transaction. Because the map tracks a single index, it is then no longer possible to remove all the copies of the transaction from the list. This happens even if the duplicated transaction is later committed in a block. The only way to remove the transaction is by restarting the node. The above problem can be repeated on and on until a sizable number of transactions are stuck in the mempool, in order ...