Hi,
I have a basic shared library written in rust, that uses wasmtime
and only creates the default Engine using let engine = Engine::default()
.
When I call dlopen("libmy_library.so", RTLD_LAZY);
from c++ code, i get the following error from dlerror()
: cannot allocate memory in static TLS block
When I remove the call to Engine::default()
(and thus, have no remaining dependencies on wasmtime
), the library loads successfully.
The same behavior repeats with versions 5.0.0
-> 7.0.0
(haven't tried earlier ones).
Any thoughts on what could be the culprit, or how I should debug/solve this?
I think you may be running into this issue https://github.com/bytecodealliance/wasmtime/issues/3654
One possible fix is to shrink the size of the TLS variables in Cranelift, probably with a box, since you're the second one to run into this.
Although as bjorn3 mentions there may be deeper toolchain fixes to apply here too
When compiling the library are you doing something that sets the tls model to initial exec? If so using dlopen on it is not allowed. It may happen to work if the library uses little tls space as in that case it is allocated in the static tls block, but this block is tiny and thus will result in an error if more tls space is used, like cranelift-codegen does.
Are you using any kind of build tool that may pass extra compile args?
bjorn3 said:
Are you using any kind of build tool that may pass extra compile args?
thanks for replying.
i'm using dotnet build
with no special arguments
which, in turn, invokes cargo build --release
on the library that references wasmtime
Is your .NET code linked into libmy_library.so or is the .NET code loading libmy_library.so? It looks like .NET itself enables -ftls-model=initial-exec
for compiling the runtime: https://github.com/dotnet/runtime/blob/57ae91cf6fc5672e34705b1a272cf268761d505a/src/mono/mono/utils/mono-tls.h#L183
thank you for the tip.
AFAICS, that flag doesn't get propagated to the build script that invokes cargo build --release --target-dir <target_dir> --package <package_name>
if this is a known issue that is expected to be fixed in the underlying libraries at some point (hinted by Alex's comments), that's fine, it's not urgent from my perspective. otherwise, if it's by design / isn't planned to be addressed - i would recommend there be some documentation/recommendation on how to overcome it.
Last updated: Nov 22 2024 at 16:03 UTC