Segfault with 1 arena through `mi_manage_os_memory_ex` and multiple per-thread heaps through `mi_heap_new_in_arena`

[MIvanchev]

I get the following stacktrace when using a single arena like that

    mi_option_enable(mi_option_verbose);
    mi_option_enable(mi_option_show_errors);

    if(_not(mi_manage_os_memory_ex(
                 mem,
                 MEM_ARENA_SIZE,
                 true, // commited
                 false,
                 true, // zero
                 -1,
                 true,
                 &arena
             ))) {
        print_err("Failed to initialize mimalloc (a memory allocator).");
        exit(1);
    }

   // Per thread

   static __thread mi_heap_t *g_threadHeap;
   g_threadHeap = mi_heap_new_in_arena(arena);

   // Alloc

   ret = mi_heap_realloc(g_threadHeap, ptr, newSize);

0xf7f5058f ???
	???:0
0x820b128 mi_segment_os_alloc
	src/3rdparty/mimalloc-main/src/segment.c:863
0x820b128 mi_segment_alloc
	src/3rdparty/mimalloc-main/src/segment.c:925
0x820d5a5 mi_segment_reclaim_or_alloc
	src/3rdparty/mimalloc-main/src/segment.c:1541
0x820d5a5 mi_segments_page_alloc
	src/3rdparty/mimalloc-main/src/segment.c:1560
0x820da5f _mi_segment_page_alloc
	src/3rdparty/mimalloc-main/src/segment.c:1671
0x820db35 mi_page_fresh_alloc
	src/3rdparty/mimalloc-main/src/page.c:309
0x820e2d1 mi_page_fresh
	src/3rdparty/mimalloc-main/src/page.c:333
0x820e2d1 mi_page_queue_find_free_ex
	src/3rdparty/mimalloc-main/src/page.c:868
0x820e959 _mi_malloc_generic
	src/3rdparty/mimalloc-main/src/page.c:1051
0x820fbc9 mi_heap_malloc_small_zero
	src/3rdparty/mimalloc-main/src/alloc.c:147
0x820fbc9 _mi_heap_malloc_zero_ex
	src/3rdparty/mimalloc-main/src/alloc.c:176
0x821338b mi_heap_umalloc
	src/3rdparty/mimalloc-main/src/alloc.c:241
0x821338b _mi_heap_realloc_zero
	src/3rdparty/mimalloc-main/src/alloc.c:308
0x8213569 mi_heap_realloc
	src/3rdparty/mimalloc-main/src/alloc.c:329

Can someone tell me what I am doing wrong or if this is a bug? The program WORKS with the debug flags (no optimizations etc.). I am using the latest main revision.

[MIvanchev]

The library is compiled statically through static.c

[MIvanchev]

So the error happens if I compile with -O2 at which point the a huge number of warnings is emitted

src/3rdparty/mimalloc-main/include/mimalloc/internal.h:640:28: warning: inlining failed in call to 'mi_tf_delayed': function not inlinable [-Winline]
  640 | static inline mi_delayed_t mi_tf_delayed(mi_thread_free_t tf) {

[MIvanchev]

-O1 is the maximum the lib will work with it seems. This is 32-bit x86.

[daanx]

Very strange -- I wonder if there is some 32-bit or compiler issue since mi_tf_delayed is a single logical and and it should be inlineable otherwise. What is the C compiler + version ? What is the OS? The crash is also in a strange location so not sure what is happening.

[MIvanchev]

@daanx this is VoidLinux, gcc (GCC) 14.2.1 20250405. It appears that at -O2 the symbol mi_segment_calculate_slices is unresolvable on some threads. It works on the main and another one, but the 3rd fails. Again, only with '-O2'. no-inline has no effect.

[daanx]

Not sure how to make progress on this -- maybe it is a compiler/linker issue? Have you tried the v3 version of mimalloc -- that one no longer uses segments and may just work?

[MIvanchev]

I am using v3 all along @daanx, let me refactor some stuff and come back to you, I'll try to figure out why code is missing. Let's not close for now.

[daanx]

I am using v3 all along @daanx, let me refactor some stuff and come back to you, I'll try to figure out why code is missing. Let's not close for now.

Ah, if you are using v3, you are somehow mixing sources as segment.c is not part of v3 (but only on v2). Maybe try a clean clone of dev3 and see how it goes? Maybe there is something cached by the compiler?

[MIvanchev]

I was/am using a fresh copy from here (master) and using compiling static.c along with the rest of the project :) I have a hunch it could be something related to weak symbols. I have a vague memory of debugging something like that in Vulkan some years ago.

[daanx]

Ah, main is still downstream of dev2 (v2) for stability reasons, so you are using v2 at the moment.
Given the strange "non inlineable" error I guess there is indeed something cached?

If you can, give v3 a try (latest dev3 branch) and see what happens?

[MIvanchev]

I absolutely will, thanks for the clarification =) I don't think the non inlineable is the culprit, those errors are normal in the optimization modes. Thank you for your patience, really appreciate it.

[MIvanchev]

dev3 fails almost immediately in both -O1 and -O2 with

3rdparty/mimalloc-dev3/src/theap.c":267, _mi_theap_create
assertion: "memid.mem.os.size >= size"

My code is exactly as above, I just compiled with dev3

[daanx]

Ah, this is with heap exclusive arena's. It seems that assertion is wrong though (my bad) -- can you comment it out and try again?

[MIvanchev]

Yes @daanx, commenting out the assertion worked in all modes (so far, I'll keep you posted), thank you so much for your zeal :) This is a great software and perfect for my use case.