You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[21.0.x] Fix backtraces through empty sequences of Wasm frames (#9412)
* Fix backtraces through empty sequences of Wasm frames
This fixes a bug where we would not properly handle contiguous sequences
of Wasm frames that are empty. This was mistakenly believed to be an impossible
scenario, and before the tail-calls proposal it was impossible, however it can
now happen after the following series of events:
* Host calls into Wasm, pushing the entry trampoline frame.
* Entry trampoline calls the actual Wasm function, pushing a Wasm frame.
* Wasm function tail calls to an imported host function, *replacing*
the Wasm frame with the exit trampoline's frame.
Now we have a stack like `[host, entry trampoline, exit trampoline]`, which has
zero Wasm frames between the entry and exit trampolines. If the host function
that the exit trampoline calls out to attempts to capture a backtrace, then --
before this commit -- we would fail an internal assertion and panic. That panic
would then unwind to the first Rust frame that is called by Wasm. With Rust 1.81
and later, Rust automatically inserts a panic handler that prevents the unwind
from continuing into external/foreign code, which is undefined behavior, and
aborts the process. Rust versions before 1.81 would attempt to continue
unwinding, hitting undefined behavior.
This commit fixes the backtrace capturing machinery to handle empty sequences of
Wasm frames, passes the assertion, and avoids unwinding into external/foreign
code.
Co-Authored-By: Alex Crichton <[email protected]>
* Add release notes
* Add more release notes
* Remove usage of wasmtime_test macro
* Enable more proposals
* ignore tail calls tests on s390x, since it didn't implement them in this release
---------
Co-authored-by: Nick Fitzgerald <[email protected]>
0 commit comments