[buyemune-sys]

Why

Under solo, all of the process's malloc is served by musl's mallocng, which takes a single lock for every call. Multithreaded guests hit it first. Measured on a glibc program with N threads, 4M malloc/free per thread, 2 cores:

threads	glibc native	solo, mallocng	solo, mimalloc
1	0.153 s	0.283 s	0.048 s
2	0.158 s	1.103 s	0.051 s
4	0.323 s	1.799 s	0.097 s

On Tarantool 3.8.1 (a glibc binary in a solo bundle), an ffi malloc/free loop is 2× slower than native with mallocng and 4× faster than native with mimalloc. Other workloads (memtx, net.box, LuaJIT) stay within noise either way.

Why in the stub and not in the guest

An allocator linked into the guest (MI_OVERRIDE inside a glibc program) breaks under solo. musl's internal allocations (strdup, realpath(NULL), asprintf, getaddrinfo, …) bypass it and later reach its free():

strdup in mimalloc heap: 1, realpath: 0, asprintf: 0
mimalloc: error: mi_free: invalid pointer: 0x7AEEE26610B0
Segmentation fault

Natively on glibc this works, because glibc itself calls malloc through the PLT. When mimalloc replaces malloc in the stub itself, musl, the bridge and every guest share one heap.

What's done

• -Dmimalloc (opt-in; the default build is unchanged) links the vendored mimalloc 3.5.3 (ext/mimalloc, tag v3.5.3 = d4881d33) into the stub in place of mallocng.
• mimalloc is compiled as C++ (upstream's standard MI_USE_CXX mode). Its atomics then come from the vendored libc++ <atomic>, so the -nostdinc world does not need the compiler's <stdatomic.h>.
• Transparent huge pages are off (MI_DEFAULT_ALLOW_THP=0): that is the host's decision.
• README: mimalloc is added to the list of vendored versions.

Test

• tst/glibc_guest_test.c gains a check of a property that must hold for any allocator the stub links: the guest frees memory musl allocated on its behalf (realpath(NULL), asprintf) and memory allocated by another thread. Needle: guest heap shared. It passes on main too (one mallocng there); its job is to catch a wrongly wired allocator.
• The Alpine job gains a Test with mimalloc step (./build -Dmimalloc test).
• Locally both ./build test and ./build -Dmimalloc test pass. In the stub, nm shows malloc == mi_malloc and no mallocng symbols.

[buyemune-sys]

This one is meh, I'll understand if you'll close it as 'wontfix'.