Run Wasm code on a separate stack #2807

Amanieu · 2022-02-25T16:44:43Z

This uses the corosensei crate to
run Wasm code on a separate stack from the main thread stack.

In trap handlers for stack overflows and memory out of bounds accesses,
we can now check whether we are executing on the Wasm stack and reset
execution back to the main thread stack when returning from the trap
handler.

When Wasm code needs to perform an operation which may modify internal
data structures (e.g. growing a memory) then execution must switch back
to the main thread stack using on_host_stack. This is necessary to avoid
leaving internal data structure in an inconsistent state when a stack
overflow happens.

In the future, this can also be used to suspend execution of a Wasm
module (#1127) by modeling it as an async function call.

Fixes #2757
Fixes #2562

syrusakbary · 2022-02-25T21:45:46Z

lib/vm/src/trap/handlers.c

-void wasmer_unwind(void *JmpBuf) {
-  platform_jmp_buf *buf = (platform_jmp_buf*) JmpBuf;
-  platform_longjmp(*buf, 1);
-}


Nice work removing the dependency on this c file!

That C file had some hack to allow correct debugging on mac. It would be nice if that hack culd be put somewhere else, as a comment or something optionnal?

You can use an LLDB command to mask signals:

process handle SIGBUS --notify false --stop false

no, that's not the issue. It's not about signal trapping.
that line of code
task_set_exception_ports(mach_task_self(), EXC_MASK_BAD_ACCESS, MACH_PORT_NULL, EXCEPTION_DEFAULT, 0);
allow EXC_BAD_ACCESS exception to be routed and trigger segfault. Without it, you get stuck at the exception and can debug anything.

Is there a reason why this line is hidden behind an #if 0 by default? It seems that mono always does this.

No specific reason, I just let it disabled unless activated for debug to avoid changing previous behaviour. But I suppose it can just be enabled.

syrusakbary

Should we eliminate vm/probestack.rs along with the PROBESTACK const, wasmer_vm_probestack, and LibCall::Probestack (in vm/libcalls.rs) ?

They seem to not be used any longer (although I'm not really sure)

Amanieu · 2022-02-25T22:53:28Z

Should we eliminate vm/probestack.rs along with the PROBESTACK const, wasmer_vm_probestack, and LibCall::Probestack (in vm/libcalls.rs) ?

They seem to not be used any longer (although I'm not really sure)

Cranelift still uses it for stack probing.

tests/ignores.txt

ptitSeb

we need to remember to add some workaround for the macOS task_set_exception_ports(mach_task_self(), EXC_MASK_BAD_ACCESS, MACH_PORT_NULL, EXCEPTION_DEFAULT, 0); debuggin issue...

syrusakbary · 2022-03-07T20:35:48Z

bors r+

2807: Run Wasm code on a separate stack r=syrusakbary a=Amanieu This uses the [corosensei](https://crates.io/crates/corosensei) crate to run Wasm code on a separate stack from the main thread stack. In trap handlers for stack overflows and memory out of bounds accesses, we can now check whether we are executing on the Wasm stack and reset execution back to the main thread stack when returning from the trap handler. When Wasm code needs to perform an operation which may modify internal data structures (e.g. growing a memory) then execution must switch back to the main thread stack using on_host_stack. This is necessary to avoid leaving internal data structure in an inconsistent state when a stack overflow happens. In the future, this can also be used to suspend execution of a Wasm module (#1127) by modeling it as an async function call. Fixes #2757 Fixes #2562 Co-authored-by: Amanieu d'Antras <[email protected]>

syrusakbary · 2022-03-07T20:37:53Z

bors r-

bors · 2022-03-07T20:37:54Z

Canceled.

syrusakbary · 2022-03-07T20:40:35Z

Some context on why we did cancel:

We still need to add the macOS debugging workaround and publish a new version of corosensei.

Once that's addressed we should be good to merge

Stack probes must be done before the stack pointer is adjusted. This ensures that the stack pointer is still within the bounds of the stack when inspected by the signal handler.

This uses the [corosensei](https://crates.io/crates/corosensei) crate to run Wasm code on a separate stack from the main thread stack. In trap handlers for stack overflows and memory out of bounds accesses, we can now check whether we are executing on the Wasm stack and reset execution back to the main thread stack when returning from the trap handler. When Wasm code needs to perform an operation which may modify internal data structures (e.g. growing a memory) then execution must switch back to the main thread stack using on_host_stack. This is necessary to avoid leaving internal data structure in an inconsistent state when a stack overflow happens. In the future, this can also be used to suspend execution of a Wasm module (#1127) by modeling it as an async function call. Fixes #2757 Fixes #2562

This is now handled by the generic trampoline code in the engine.

Amanieu · 2022-03-16T13:24:35Z

bors r+

bors · 2022-03-16T14:58:28Z

Build succeeded:

Amanieu requested a review from syrusakbary as a code owner February 25, 2022 16:44

Amanieu force-pushed the corosensei branch 2 times, most recently from f1262fb to 3d83d55 Compare February 25, 2022 17:19

syrusakbary reviewed Feb 25, 2022

View reviewed changes

Amanieu force-pushed the corosensei branch 2 times, most recently from ef38f0d to 290c84e Compare March 2, 2022 22:59

syrusakbary reviewed Mar 2, 2022

View reviewed changes

tests/ignores.txt Outdated Show resolved Hide resolved

Amanieu force-pushed the corosensei branch from 290c84e to d63a7b2 Compare March 2, 2022 23:03

ptitSeb approved these changes Mar 3, 2022

View reviewed changes

Amanieu force-pushed the corosensei branch 2 times, most recently from a747bdd to bee530b Compare March 4, 2022 15:49

Amanieu added 12 commits March 14, 2022 22:14

Use LLVM's inline-asm stack probing instead of an external function

7467ec3

Fix stack probing in the singlepass compiler

054e524

Stack probes must be done before the stack pointer is adjusted. This ensures that the stack pointer is still within the bounds of the stack when inspected by the signal handler.

Fix unused field lint

83fd790

Fix stack probing typo

2b86fe8

Fix calling host functions directly from the API

f9d1cfd

Update nightly Rust version used to build headless

58bab0f

Try to enable more tests

af3d948

Remove the custom probestack trampoline in compiler-cranelift

657bbae

This is now handled by the generic trampoline code in the engine.

Fix x86_64 libcall trampoline in engine-universal

a280914

Don't try generating a backtrace on stack overflow

3d52bf2

Update corosensei to 0.1.2

742c20a

Add back workaround for LLDB debugging on macOS

c4e1c5f

Amanieu force-pushed the corosensei branch from e002b7b to c4e1c5f Compare March 14, 2022 22:14

Amanieu mentioned this pull request Mar 16, 2022

Feat/singlepass dylib #2802

Merged

bors bot merged commit fc794ab into master Mar 16, 2022

bors bot deleted the corosensei branch March 16, 2022 15:15

This was referenced Apr 11, 2022

Remove several sources of non-determinism massalabs/massa-sc-runtime#124

Merged

Ensure determinism massalabs/massa-sc-runtime#94

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Run Wasm code on a separate stack #2807

Run Wasm code on a separate stack #2807

Amanieu commented Feb 25, 2022

syrusakbary Feb 25, 2022

ptitSeb Mar 1, 2022

Amanieu Mar 1, 2022

ptitSeb Mar 1, 2022 •

edited

Loading

Amanieu Mar 1, 2022

ptitSeb Mar 1, 2022

syrusakbary left a comment

Amanieu commented Feb 25, 2022

ptitSeb left a comment

syrusakbary commented Mar 7, 2022

syrusakbary commented Mar 7, 2022

bors bot commented Mar 7, 2022

syrusakbary commented Mar 7, 2022

Amanieu commented Mar 16, 2022

bors bot commented Mar 16, 2022

Run Wasm code on a separate stack #2807

Run Wasm code on a separate stack #2807

Conversation

Amanieu commented Feb 25, 2022

syrusakbary Feb 25, 2022

Choose a reason for hiding this comment

ptitSeb Mar 1, 2022

Choose a reason for hiding this comment

Amanieu Mar 1, 2022

Choose a reason for hiding this comment

ptitSeb Mar 1, 2022 • edited Loading

Choose a reason for hiding this comment

Amanieu Mar 1, 2022

Choose a reason for hiding this comment

ptitSeb Mar 1, 2022

Choose a reason for hiding this comment

syrusakbary left a comment

Choose a reason for hiding this comment

Amanieu commented Feb 25, 2022

ptitSeb left a comment

Choose a reason for hiding this comment

syrusakbary commented Mar 7, 2022

syrusakbary commented Mar 7, 2022

bors bot commented Mar 7, 2022

syrusakbary commented Mar 7, 2022

Amanieu commented Mar 16, 2022

bors bot commented Mar 16, 2022

ptitSeb Mar 1, 2022 •

edited

Loading