Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

SIMD-0161: SBPF versioning and feature gating #161

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
Open
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
138 changes: 138 additions & 0 deletions proposals/0161-sbpf-feature-gating.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,138 @@
---
simd: '0161'
title: SBPF versioning and feature gating
authors:
- Alexander Meißner
- Alessandro Decina
category: Standard
type: Core
status: Draft
created: 2024-07-15
---

## Summary

An explicit versioning system for SBPF programs, that allows enabling or
disabling features on a per-program basis.

## Motivation

SBPF has evolved over the years and will continue to evolve. Changes are
necessary among other reasons to improve performance, address security concerns
and provide new, better features to dapp developers.

Today, the only way to introduce changes to the program runtime is via feature
gates [link that explains what feature gates are]. For example we have used
feature gates in the past to restrict new deployments of programs owned by
deprecated loaders. Feature gates alone are not sufficient to evolve SBPF
though - we are missing a mechanism to enable (or disable) specific features on
a per-program basis.

As an example, over two years ago we decided to change the way the stack works
for SBPF programs - from allocating a fixed 4kB per function frame, to letting
functions dynamically specify how much stack they need. This change was
implemented in both the program runtime and the toolchain (LLVM), but as of
today it has not yet been deployed, because it's essentially too hard to do so:
upon executing a program, should the program runtime use the old or the new
stack allocation strategy?

We propose to introduce an explicit versioning scheme for SBPF programs:
programs will encode which SBPF version they are compiled for; based on this
version, the program runtime will alter its behavior. In order to prevent an
"extension hell" it will not be possible to opt into specific changes via
flags, instead an entire set of changes is reduced to a single SBPF version.

## New Terminology

SBPF version: the SBPF version number a program was compiled for.

## Detailed Design

Every program must signal which SBPF version it was compiled for via
the `e_flags` field in the
[ELF header](https://refspecs.linuxfoundation.org/elf/gabi4+/ch4.eheader.html).
Block explorers are recommended to display this SBPF version field for program
accounts.

For each SBPF version two feature gates will control its availability:

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure that we need such a strict system. I think we could probably have several versions active and deployable at the same time, without forcing people to update toolchain and redeploy or pay extra costs.

Take the following examples:

  • dynamic frames: from the vm perspective, it is cheaper to use dynamic frames, because programs potentially page in less memory if an X-deep call stack fits in 1 page vs taking X pages because of fixed framing. We might want to push people to update to increase performance.

  • the sub imm operand swap change: this doesn't materially change anything. It's a nice thing for us to fix for aesthetic reasons, but it doesn't really change anything for developers.

I think I'd prefer an ad-hoc system where depending on the impact of a new SBPF version, we decide whether and how to rug earlier versions (in practice I expect this to happen very rarely).

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Mandating that all deployments are targeting the newest version does not force anybody to redeploy. It only means that dapp devs have to keep their toolchain updated when they re/deploy. You are right in that it is not necessary to be this strict, but it has the nice effect of allowing us to get more testing exposure for the newest toolchain version.


- One to enable deployment and execution of that SBPF version, which includes
disabling deployment (but not execution) of all earlier versions.
- Another to first disincentivize (via CU costs) and later disable execution
of that SBPF version, so that the runtime and virtual machine can reduce their
complexity again. This feature gate can be shared by multiple SBPF versions,
effectively deprecating larger blocks of versions in order to reduce the amount
of redeployment required from dapp developers.

### Example

| SBPF version | becomes deployable and executable | ceases to be deployable |
| ------------ | --------------------------------- | ----------------------- |
| v1 | | Feature gate A |
| v2 | Feature gate A | Feature gate B |
Comment on lines +69 to +72
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this purely an example, or are we considering only one version to be able to deploy at a time? It might be helpful to include the number of simultaneous SBPF versions that the network will allow.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes the proposal right now says that there can only be one deployable version at a time (but many executable versions). See the discussion above.

| v3 | Feature gate B | Feature gate C |
| v4 | Feature gate C | Feature gate D |
| v5 | Feature gate D | Feature gate E |
| v6 | Feature gate E | Feature gate F |
| v7 | Feature gate F | |

| SBPF version | ceases to be executable |
| ------------ | ----------------------- |
| v1 | Feature gate G |
| v2 | Feature gate G |
| v3 | Feature gate G |
| v4 | Feature gate H |
| v5 | Feature gate H |
| v6 | Feature gate H |
| v7 | |

## Alternatives Considered

### Use the runtime global feature gating mechanism

This would mean all programs need to be redeployed simultaneously at feature
activation and that they would each require an alternative version to be
prepared and uploaded to be switched to (redeployed) on feature activation.
This would increase the complexity of program management in loaders and require
dapp developers to double the locked tokens in their deployed programs.

It also means that validator implementations would have to handle a massive
amount of loading, parsing, relocation, verification and possibly compilation
of programs in a single block. It took the Agave team two years to develop,
implement, test and deploy this capability. It is desirable for other validator
implementations to avoid having to do the same.

### Coupling the feature gating to the loader, not the individual program accounts

The version number could be implied by the loader a program was deployed in,
effectively shifting the encoding from the program account content to the owner
field. This would require a new loader address for each SBPF version, as the
program management and execution logic could be shared for the most part.
However, new loaders would necessitate a way to migrate programs between
loaders upon redeployment.

## Impact

Active dapp developers who redeploy their programs regularly will (upon feature
activation) be forced to also keep their toolchain up-to-date, which they most
likely do anyway.

Less active dapp developers who do not redeploy their programs will be nagged
by their users to redeploy their programs once the deprecation CU cost ramps
up.

Finalized programs which simply can not be redeployed will instead have their
user base slowly migrate their funds to alternative programs.

Validator implementations will have to support multiple code paths in the time
between the introduction feature activation and the fixed slot offset after the
deprecation feature activation.

## Security Considerations

None.

## Drawbacks

Programs will break and if they are finalized their addresses and deployment
cost will be burnt.
Loading