PVF Host and Workers

The PVF host is responsible for handling requests to prepare and execute PVF code blobs, which it sends to PVF workers running in their own child processes. These workers are spawned from the polkadot-prepare-worker and polkadot-execute-worker binaries.

While the workers are generally long-living, they also spawn one-off secure job processes that perform the jobs. See "Job Processes" section below.

High-Level Flow

Some notes about the graph:

1. Once a job has finished, the response will flow back up the way it came.
2. In the case of execution, the host will send a request for preparation to the Prepare Queue if needed. In that case, only after the preparation succeeds does the Execute Queue continue with validation.
3. Multiple requests for preparing the same artifact are coalesced, so that the work is only done once.

Goals

This system has two high-level goals that we will touch on here: determinism and security.

Determinism

One high-level goal is to make PVF operations as deterministic as possible, to reduce the rate of disputes. Disputes can happen due to e.g. a job timing out on one machine, but not another. While we do not have full determinism, there are some dispute reduction mechanisms in place right now.

Retrying execution requests

If the execution request fails during preparation, we will retry if it is possible that the preparation error was transient (e.g. if the error was a panic or time out). We will only retry preparation if another request comes in after 15 minutes, to ensure any potential transient conditions had time to be resolved. We will retry up to 5 times.

If the actual execution of the artifact fails, we will retry once if it was a possibly transient error, to allow the conditions that led to the error to hopefully resolve. We use a more brief delay here (1 second as opposed to 15 minutes for preparation (see above)), because a successful execution must happen in a short amount of time.

If the execution fails during the backing phase, we won't retry to reduce the chance of supporting nondeterministic candidates. This reduces the chance of nondeterministic blocks getting backed and honest backers getting slashed.

We currently know of the following specific cases that will lead to a retried execution request:

1. OOM: We have memory limits to try to prevent attackers from exhausting host memory. If the memory limit is hit, we kill the job process and retry the job. Alternatively, the host might have been temporarily low on memory due to other processes running on the same machine. NOTE: This case will lead to voting against the candidate (and possibly a dispute) if the retry is still not successful.
2. Syscall violations: If the job attempts a system call that is blocked by the sandbox's security policy, the job process is immediately killed and we retry. NOTE: In the future, if we have a proper way to detect that the job died due to a security violation, it might make sense not to retry in this case.
3. Artifact missing: The prepared artifact might have been deleted due to operator error or some bug in the system.
4. Job errors: For example, the job process panicked for some indeterminate reason, which may or may not be independent of the candidate or PVF.
5. Internal errors: See "Internal Errors" section. In this case, after the retry we abstain from voting.
6. RuntimeConstruction error. The precheck handles a general case of a wrong artifact but doesn't guarantee its consistency between the preparation and the execution. If something happened with the artifact between the preparation of the artifact and its execution (e.g. the artifact was corrupted on disk or a dirty node upgrade happened when the prepare worker has a wasmtime version different from the execute worker's wasmtime version). We treat such an error as possibly transient due to local issues and retry one time.

Preparation timeouts

We use timeouts for both preparation and execution jobs to limit the amount of time they can take. As the time for a job can vary depending on the machine and load on the machine, this can potentially lead to disputes where some validators successfully execute a PVF and others don't.

One dispute mitigation we have in place is a more lenient timeout for preparation during execution than during pre-checking. The rationale is that the PVF has already passed pre-checking, so we know it should be valid, and we allow it to take longer than expected, as this is likely due to an issue with the machine and not the PVF.

CPU clock timeouts

Another timeout-related mitigation we employ is to measure the time taken by jobs using CPU time, rather than wall clock time. This is because the CPU time of a process is less variable under different system conditions. When the overall system is under heavy load, the wall clock time of a job is affected more than the CPU time.

Internal errors

An internal, or local, error is one that we treat as independent of the PVF and/or candidate, i.e. local to the running machine. If this happens, then we will first retry the job and if the errors persists, then we simply do not vote. This prevents slashes, since otherwise our vote may not agree with that of the other validators.

In general, for errors not raising a dispute we have to be very careful. This is only sound, if either:

1. We ruled out that error in pre-checking. If something is not checked in pre-checking, even if independent of the candidate and PVF, we must raise a dispute.
2. We are 100% confident that it is a hardware/local issue: Like corrupted file, etc.

Reasoning: Otherwise it would be possible to register a PVF where candidates can not be checked, but we don't get a dispute - so nobody gets punished. Second, we end up with a finality stall that is not going to resolve!

Note that any error from the job process we cannot treat as internal. The job runs untrusted code and an attacker can therefore return arbitrary errors. If they were to return errors that we treat as internal, they could make us abstain from voting. Since we are unsure if such errors are legitimate, we will first retry the candidate, and if the issue persists we are forced to vote invalid.

Security

With on-demand parachains, it is much easier to submit PVFs to the chain for preparation and execution. This makes it easier for erroneous disputes and slashing to occur, whether intentional (as a result of a malicious attacker) or not (a bug or operator error occurred).

Therefore, another goal of ours is to harden our security around PVFs, in order to protect the economic interests of validators and increase overall confidence in the system.

Possible attacks / threat model

Webassembly is already sandboxed, but there have already been reported multiple CVEs enabling remote code execution. See e.g. these two advisories from Mar 2023 and Jul 2022.

So what are we actually worried about? Things that come to mind:

1. Consensus faults - If an attacker can get some source of randomness they could vote against with 50% chance and cause unresolvable disputes.
2. Targeted slashes - An attacker can target certain validators (e.g. some validators running on vulnerable hardware) and make them vote invalid and get them slashed.
3. Mass slashes - With some source of randomness they can do an untargeted attack. I.e. a baddie can do significant economic damage by voting against with 1/3 chance, without even stealing keys or completely replacing the binary.
4. Stealing keys - That would be pretty bad. Should not be possible with sandboxing. We should at least not allow filesystem-access or network access.
5. Taking control over the validator. E.g. replacing the polkadot binary with a polkadot-evil binary. Should again not be possible with the above sandboxing in place.
6. Intercepting and manipulating packages - Effect very similar to the above, hard to do without also being able to do 4 or 5.

We do not protect against (1), (2), and (3), because there are too many sources of randomness for an attacker to exploit.

We provide very good protection against (4), (5), and (6).

Job Processes

As mentioned above, our architecture includes long-living worker processes and one-off job processes. This separation is important so that the handling of untrusted code can be limited to the job processes. A hijacked job process can therefore not interfere with other jobs running in separate processes.

Furthermore, if an unexpected execution error occurred in the execution worker and not the job itself, we generally can be confident that it has nothing to do with the candidate, so we can abstain from voting. On the other hand, a hijacked job is able to send back erroneous responses for candidates, so we know that we should not abstain from voting on such errors from jobs. Otherwise, an attacker could trigger a finality stall. (See "Internal Errors" section above.)

Restricting file-system access

A basic security mechanism is to make sure that any process directly interfacing with untrusted code does not have unnecessary access to the file-system. This provides some protection against attackers accessing sensitive data or modifying data on the host machine.

Currently this is only supported on Linux.

Restricting networking

We also disable networking on PVF threads by disabling certain syscalls, such as the creation of sockets. This prevents attackers from either downloading payloads or communicating sensitive data from the validator's machine to the outside world.

Currently this is only supported on Linux.

Clearing env vars

We clear environment variables before handling untrusted code, because why give attackers potentially sensitive data unnecessarily? And even if everything else is locked down, env vars can potentially provide a source of randomness (see point 1, "Consensus faults" above).