# Debugging in Wasmtime Nick Fitzgerald 2024-02-01 <!-- Put the link to this slide here so people can follow --> slides: https://hackmd.io/p/-ARJiaL_SAmsWz9DzZUh4Q --- ## Status * We have discussed all these things before in SIG-Debugging * I need to wrap this all up in an RFC * We need to resume and finish in-flight work * Engineering help very welcome! --- ## The Plan * Build a Debug Adapter Protocol (DAP) server into Wasmtime * Leverage existing DAP frontends (VS Code, Emacs, etc) * "Debugging components" to translate between source and Wasm state * Use Winch instead of Cranelift for live debugging ---  --- ## Incremental Steps 1. **DONE:** Generate Wasm core dumps 2. **TODO:** DAP server w/out stepping, backed by core dumps * capture core dump in prod, send it to dev for local debugging 4. **TODO:** DAP server w/ stepping, backed by Winch 5. **TODO:** DAP server w/ time travel * record when running on Cranelift in prod, replay when running on Winch locally? --- ## Single Stepping and Breakpoints * Reminder: Winch only * Insert a `nop` between the machine code for each Wasm instruction * When setting a breakpoint, replace the `nop` with an `int3` (or equiv) * Catch signal in Wasmtime and notify the debugger thread that the debuggee is paused --- ## Recovering Wasm Debuggee State * Winch will need to emit debug info tables * (when in debug mode) * Indexed by PC, yielding: * operand stack depth * which register/SP offset each operand is at * Locals always at frame base (IIUC) --- ## Watch Points * Tables, locals, and globals: * Can implement with breakpoints on every `table.set`, `local.set`, `global.set`, etc instructions * Memory is a bit trickier: * Virtual memory tricks to catch writes and then check if they are inside our target memory range --- ## Step Into * Direct calls: * set a temp breakpoint at the start of the static callee, continue to it * Indirect calls: * break just before the `call_indirect`/`call_ref` transfers control, observe the callee, set a temp breakpoint in the callee --- ## Debugging Components: Why? * Supports debugging guests running in an interpreter compiled to wasm * for example, JS * No need to add new capabilities to guest * compare to: giving guest a socket to speak its own debugging protocol --- ## Debugging Components: Why? * No need to spec DWARF interpretation * Can still wrap existing tooling into debugging components * but don't *have* to, so door is open to future innovation --- ## Debugging Components * Each Wasm has a custom section: `debugging_component` * Contains a URL or file path of this Wasm's debugging component --- ## Debugging Components * The debugging component imports a `WasmDebugging` WIT interface * And exports a `SourceDebugging` WIT interface * Its responsibility is to translate between the two interfaces --- ## Debugging Components * MVP debugging could skip these, and leave them as an incremental milestone * Wasmtime would have a built in / fallback DWARF component * Could implement ourselves with `gimli` or we could wrap `lldb`/`gdb` * `spidermonkey.wasm` would implement this in terms of calls into SpiderMonkey's `Debugger` API --- ## Record and Replay Vision: record a trace from prod (Cranelift) and deterministically debug locally (Winch) --- ## Record and Replay * Recording traces pretty easy * Capture values passed into Wasm by imports * Virtual memory tricks to capture memory writes from host imports * Results of `memory.grow` and `table.grow` * `NaN` bitpatterns * Relaxed SIMD stuff --- ## Record and Replay * Recording even easier at the component layer: * just interface values crossing the root component boundary, that's it. --- ## Record and Replay * Replaying also easy! * hook those same operations to return previously recorded values --- ## Record and Replay * Time travel debugging harder * Requires state snapshotting to create keyframes * Easy for memory, tables, etc... * Harder for stack and locals * But again: this can be pushed off as another, later milestone! * deterministic debugging valuable even without time travel --- ## Next Steps * I need to polish these ideas into an RFC * We need to staff this work --- # Thanks!