assertion failed: "theap->tld->thread_id == 0 || theap->tld->thread_id == tid"

[hbirler]

I was trying the new first class heaps (mimalloc 3.2.8, installed via vcpkg), but I encountered some issues.

I can reliably get the following error in debug mode:

assertion failed: at "alloc.c":136, mi_theap_malloc_small_zero_nonnull
assertion: "theap->tld->thread_id == 0 || theap->tld->thread_id == tid"

In release mode, a SIGSEGV is triggered.

My reproducer

#include <cassert>
#include <cstdlib>
#include <print>
#include <thread>
#include <vector>
#include <mimalloc.h>
#include <barrier>

using namespace std;

// We repeatedly
// * (Destroy previous heap)
// * Create a new heap
// * Allocate from the heap with threads
// This seems to trigger assertions or segfaults in mimalloc
int main()
{
    int numWorkers = 32;
    std::barrier startBar(numWorkers + 1);
    std::barrier doneBar(numWorkers + 1);
    vector<jthread> threads;
    mi_heap_t* heap = nullptr;
    for (int i = 0; i < numWorkers; i++)
    {
        threads.emplace_back([&startBar, &doneBar, &heap](std::stop_token tok)
        {
            while (true)
            {
                startBar.arrive_and_wait();
                if (tok.stop_requested())
                    return;
                mi_heap_t* localHeap = heap;
                if (true)
                {
                    for (int j = 0; j < 100; j++)
                    {
                        assert(localHeap);
                        auto* ptr = mi_heap_malloc(localHeap, 17);
                        mi_free(ptr);
                    }
                }
                doneBar.arrive_and_wait();
            }
        });
    }

    for (size_t j = 0; j < 100; j++)
    {
        std::print(stderr, "Round {}\n", j);
        if (heap)
            mi_heap_destroy(heap);
        heap = mi_heap_new();
        startBar.arrive_and_wait();
        doneBar.arrive_and_wait();
    }

    for (auto& t : threads)
        t.request_stop();

    startBar.arrive_and_wait();

    return 0;
}

[hbirler]

I initially had a more complicated use case where not all threads actually allocate in each round (only a random subset does) and not all pointers were freed. But I later noticed that even the case when all threads always allocate lead to the issue.

[hbirler]

This seems to be distinct from #1230 as adding the lines

void* dummy = mi_heap_malloc(heap, 1);
mi_free(dummy);

after mi_heap_new() does not seem to fix the issue.