Add poison state to sandbox #931
Conversation
ludfjig
added
the
kind/enhancement
ludfjig
requested review from
danbugs,
dblnz,
devigned,
jprendes,
marosset,
simongdavies and
syntactically
as code owners
There was a problem hiding this comment.
Pull Request Overview
This PR introduces sandbox poisoning functionality to prevent further operations when a sandbox is in an inconsistent state that could compromise memory safety. The sandbox becomes poisoned when guest functions abort/panic or when host-initiated execution cancellation occurs.
Key changes:
- Added poisoned state tracking with safety checks across all sandbox operations
- Implemented automatic poison detection for specific error types (GuestAborted, ExecutionCanceledByHost)
- Added recovery mechanisms through snapshot restoration or manual poison clearing
Reviewed Changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| src/hyperlight_host/src/sandbox/initialized_multi_use.rs | Core implementation of poison state tracking, safety checks, and recovery mechanisms |
| src/hyperlight_host/src/error.rs | Added PoisonedSandbox error variant with detailed documentation |
| src/hyperlight_host/tests/integration_test.rs | Updated interrupt tests to clear poison state for continued execution |
| src/hyperlight_host/src/sandbox/snapshot.rs | Added Debug trait to Snapshot struct |
| src/hyperlight_host/src/mem/shared_mem_snapshot.rs | Added Debug trait to SharedMemorySnapshot struct |
ludfjig
force-pushed
the
poison
branch
3 times, most recently
from
61713f3 to
74a9260
Compare
| | HyperlightError::NoMemorySnapshot | ||
| | HyperlightError::ParameterValueConversionFailure(_, _) | ||
| | HyperlightError::PEFileProcessingFailure(_) | ||
| | HyperlightError::PoisonedSandbox |
There was a problem hiding this comment.
It seems a bit strange to me that IsPoisoned doesn't return true when this is the error.
There was a problem hiding this comment.
you're right it's confusing, but it is correct. PoisonedSandbox is what we return in case some other error already have caused poison state.
There was a problem hiding this comment.
Which reminds we should probably better organize our errors into categories.
There was a problem hiding this comment.
isn't by definition poisoned then?
There was a problem hiding this comment.
No because this error doesn't cause poison. PoisionedSandbox is what we returned if it is poisoned.
There was a problem hiding this comment.
if the sandbox returns this error then poisoned() == true? something feels off about not verifying that if this is returned and the internal state should also be verified to be correct.
| /// | ||
| /// **Data Structure Corruption**: | ||
| /// - Hash tables with inconsistent bucket counts vs. actual entries | ||
| /// - Linked lists with broken next/prev pointers |
There was a problem hiding this comment.
Are these things we have that get corrupted? These seems like general information on corruption but not specific to what we are protecting against.
There was a problem hiding this comment.
yes these are just general information warnings
There was a problem hiding this comment.
Would all of these fall under memory corruption?
There was a problem hiding this comment.
IMO, this list feels overly generic and distracts from the rest of the comment but can see what others think
Signed-off-by: Ludvig Liljenberg <[email protected]>
3 participants
This PR adds a poison state to sandbox in order to prevent further operations when the sandbox is left in an inconsistent state that could compromise memory safety, data integrity, or security. The sandbox becomes poisoned when guest functions abort/panic or when host-initiated execution cancellation occurs, leaving behind leaked heap allocations, corrupted data structures, or unreleased resources. For example, interrupting execution while guest is allocating can leave the global allocation lock in an inconsistent state, making future allocations impossible in subsequent runs due to infinite locking/spinning.
Poisoned sandboxes will reject all further operations (guest calls, snapshots, memory mapping) until the inconsistent state is resolved through either restoring to a snapshot or manually (unsafely) clearing the poison state.
Closes #848