Conversation
|
the remaining unhandled piece of work here is turning on macos CI, which is, I think, straightforward? I just haven't looked at it yet. |
data-pup
left a comment
data-pup
left a comment
There was a problem hiding this comment.
I can't test this myself, but the changes here look good to me!
iximeow
force-pushed
the
multiplatform-signals
branch
from
da48da4 to
be53c48
Compare
acfoltzer
left a comment
acfoltzer
left a comment
There was a problem hiding this comment.
Really nice work! Just a few cleanup requests
iximeow
force-pushed
the
multiplatform-signals
branch
3 times, most recently
from
8963987 to
781fd9d
Compare
|
🎉 This branch has picked up a few non-signal-handling-related changes now, so given a +1 I'll fix up the commit history here into the ~three commits I'd like to land ("change signal handling around", "fix C API signal stack sizes to the largest minimum we know we have to work with -- 32kb on macos", and "enable mac CI") |
pchickey
left a comment
pchickey
left a comment
There was a problem hiding this comment.
Thanks for getting Mac CI working again!
iximeow
force-pushed
the
multiplatform-signals
branch
4 times, most recently
from
d2088fb to
acfe0a2
Compare
| // pointing the instruction pointer at `lucet_context_set`, then by preparing the argument | ||
| // that `lucet_context_set` should read from `rdi` - the context to switch to. | ||
| // | ||
| // NOTE: it is absolutely critical that `lucet_context_set` does not use the guest stack! |
There was a problem hiding this comment.
full disclosure: I left most of the branch unchanged while adjusting the commits, but this NOTE: is new. I'd realized that there wasn't anything discussing the specific constraints on the function we continue into after handling a signal, and this addresses that.
This is in contrast to calling `Context::set_from_signal` directly, which would skip over typical signal handling teardown some platforms might require. pthread_sigmask are removed here as they only served to reset signal masks that Linux would apply when entering the signal handler. Now that we return normally from the signal handler, on Linux signals are unmasked normally.
* conditionally define signals for invalid memory accesses per platform * add old macos module for version testing binary blob for the module is probably similar to the linux `old_module.so`, for which the source is not known. the macos `old_module.dylib` is built from `fibonacci.wat` as existed in Lucet commit `8639967` with the following command: ``` $ cargo run --bin lucetc ./lucetc/tests/wasm/fibonacci.wat ``` * make signal stacks sufficiently large in C API tests that mac CI passes
also involves fixing some wasi-sdk installation bitrot
iximeow
force-pushed
the
multiplatform-signals
branch
from
acfe0a2 to
78aa0ed
Compare
|
Promise I'm done fiddling with this branch now, it's ready for a final +1! 🤞 |
iximeow
dismissed
acfoltzer’s stale review
changes have been made, and @pchickey has since approved
6 participants
This mostly gets things in place for the Lucet signal handler to work properly on non-linux posix-y platforms, most interestingly: macos and flavors of bsd.
The gist of this change is to not context swap in the signal handler itself, but replace the address the guest continues with by the context swap function. So the signal handler can return on all platforms, passing through
sigreturn-like mechanisms if present, and we have fewer cross-platform compatibility issues.