Transitioning from a `nvexec::stream_context` to another seg faults

I'm using c9d272a37754dc585cd40dbbf149e4aab4d3fe15.

I'm trying to transition from one nvexec::stream_context to another, but it seg faults.

Here is the reproducer:

TEST(try, reproducer)
{
    ::nvexec::stream_context stream_ctx_0{};
    ::nvexec::stream_context stream_ctx_1{};

    auto snd = ::stdexec::schedule(stream_ctx_0.get_scheduler())
        | ::stdexec::then([=] () -> int {
            if (::nvexec::is_on_gpu())
                return 1;
            else
                return 0;
        })
        | ::stdexec::continues_on(stream_ctx_1.get_scheduler())
        | ::stdexec::then([=](const int val) -> int {
            if (::nvexec::is_on_gpu() && val == 1)
                return 2;
            else
                return 0;
        });

    const auto [result] = ::stdexec::sync_wait(std::move(snd)).value();

    ASSERT_EQ(result, 2);
}

Note that if I change to ::stdexec::continues_on(stream_ctx_0.get_scheduler()) (i.e. transitioning from stream_ctx_0 to itself), it works fine.

Here is the backtrace I get from cuda-gdb:

[----------] 1 test from try
[ RUN      ] try.reproducer
[New Thread 0x7fff995fe000 (LWP 33576)]
[New Thread 0x7fff98dfd000 (LWP 33577)]
[New Thread 0x7fff79fff000 (LWP 33578)]
[Thread 0x7fff98dfd000 (LWP 33577) exited]

Thread 21 "tests_nvexec_ad" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7fff995fe000 (LWP 33576)]
0x00007ffff274276b in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
(cuda-gdb) bt
#0  0x00007ffff274276b in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#1  0x00007ffff284214e in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#2  0x00007ffff27f97c9 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#3  0x00007ffff27df6f0 in cuMemFreeAsync () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#4  0x00007ffff2016540 in ?? () from /usr/local/cuda/lib64/libcudart.so.12
#5  0x00007ffff207e05f in cudaFreeAsync () from /usr/local/cuda/lib64/libcudart.so.12
#6  0x000055555556ecb1 in nvexec::_strm::continuation_task_t<nvexec::_strm::_continues_on::operation_state_t<nvexec::_strm::then_sender_t<nvexec::_strm::stream_scheduler::sender_t, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda()#1}>, nvexec::_strm::_sched_from::receiver_t<nvexec::_strm::continues_on_sender_t<nvexec::_strm::stream_scheduler, nvexec::_strm::then_sender_t<nvexec::_strm::stream_scheduler::sender_t, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda()#1}> >, nvexec::_strm::_then::receiver_t<4ul, nvexec::_strm::propagate_receiver_t<nvexec::_strm::_sync_wait::receiver_t<nvexec::_strm::then_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, nvexec::_strm::continues_on_sender_t<nvexec::_strm::stream_scheduler, nvexec::_strm::then_sender_t<nvexec::_strm::stream_scheduler::sender_t, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda()#1}> > >, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda(int)#1}> > >, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda(int)#1}> > >::__t::receiver_t, nvexec::variant_t<cuda::std::__4::tuple<nvexec::_strm::set_noop>, cuda::std::__4::tuple<stdexec::__rcvrs::set_value_t, int>, cuda::std::__4::tuple<stdexec::__rcvrs::set_error_t, cudaError>, cuda::std::__4::tuple<stdexec::__rcvrs::set_error_t, std::__exception_ptr::exception_ptr> > >::continuation_task_t(nvexec::_strm::_continues_on::operation_state_t<nvexec::_strm::then_sender_t<nvexec::_strm::stream_scheduler::sender_t, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda()#1}>, nvexec::_strm::_sched_from::receiver_t<nvexec::_strm::continues_on_sender_t<nvexec::_strm::stream_scheduler, nvexec::_strm::then_sender_t<nvexec::_strm::stream_scheduler::sender_t, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda()#1}> >, nvexec::_strm::_then::receiver_t<4ul, nvexec::_strm::propagate_receiver_t<nvexec::_strm::_sync_wait::receiver_t<nvexec::_strm::then_sender_t<nvexec::_strm::schedule_from_sender_t<nvexec::_strm::stream_scheduler, nvexec::_strm::continues_on_sender_t<nvexec::_strm::stream_scheduler, nvexec::_strm::then_sender_t<nvexec::_strm::stream_scheduler::sender_t, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda()#1}> > >, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda(int)#1}> > >, tests::nvexec::adaptors::try_reproducer_Test::TestBody()::{lambda(int)#1}> > >::__t::receiver_t, nvexec::variant_t<cuda::std::__4::tuple<nvexec::_strm::set_noop>, cuda::std::__4::tuple<stdexec::__rcvrs::set_value_t, int>, cuda::std::__4::tuple<stdexec::__rcvrs::set_error_t, cudaError>, cuda::std::__4::tuple<stdexec::__rcvrs::set_error_t, std::__exception_ptr::exception_ptr> >*, CUstream_st*, std::pmr::memory_resource*)::{lambda(nvexec::_strm::queue::task_base_t*)#2}::__invoke(nvexec::_strm::queue::task_base_t*) ()

I am on a AMPERE86 GPU with Cuda 12.8, compiling with:

Ubuntu clang version 20.1.2 (++20250331083337+bc65196c0919-1exp120250331203353.96)

Any idea ? I am assuming that my code is legit, but I might be wrong as well. Thanks!

Start with the supplied reproducer and inspect the nvexec::_strm::_continues_on and continuation_task_t paths named in the cuda-gdb backtrace. Run the test under CUDA 12.8 with cuda-gdb, then verify that transitioning between distinct stream_context objects completes without a segfault and produces result 2.