superrob1500
/
mirror-ghostty
mirror of https://github.com/ghostty-org/ghostty
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

glslang/shim: delete the GUI-thread TPoolAllocator at exit (not popAll) 

Yesterday's popAll attempt (1c2c5760b) was based on TPoolAllocator
documentation, but the implementation only returns pages to
glslang's internal free list — it never gives them back to the
system allocator. heaptrack confirmed: total leaked unchanged at
24.46 MB after popAll, same TPoolAllocator::allocate bucket
sitting at ~12 MB / 565k calls.

The actual release path: glslang::InitializeThreadPoolAllocator
calls `new TPoolAllocator`, so the pool is heap-allocated. Calling
`delete` on it triggers ~TPoolAllocator which `free()`s every
allocated page. Swap the TLS slot to nullptr first via
SetThreadPoolAllocator(nullptr) so a follow-up GetThreadPoolAllocator
(should be none) wouldn't return a dangling pointer.

Order matters: FinalizeProcess runs FIRST so it can walk +
destroy SharedSymbolTables[v][s][p][src][stage] entries (which
hold pointers into pool memory) while the pool is still alive.
Pool teardown runs last.

Safe at atexit because every renderer thread has joined via
Vulkan.threadExit (their pools are independent threadlocals
already cleaned up); the SPV cache was just cleared; and
FinalizeProcess has already finished.

Co-Authored-By: claude-flow <ruv@ruv.net>
pull/12846/head
Nathan 2026-05-26 10:55:07 -05:00
parent 1c2c5760b7
commit 3ea8fc681b
1 changed files with 29 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
pkg/glslang/override/ghastty_vk_shim.cpp
View File

@ -249,25 +249,34 @@ extern "C" void ghastty_glslang_finalize_process(void) {
std::lock_guard<std::mutex> lg(spv_cache_mutex());
spv_cache().clear();
}
// Free this thread's TPoolAllocator pages. heaptrack pointed
// the ~12 MB glslang leak at TPoolAllocator::allocate calls
// rooted in shadertoy.spirvFromGlsl on the GUI thread (since
// ghostty_surface_new runs glslang synchronously from
// MainWindow::newTab) — that pool's pages persist until thread
// exit, but the GUI thread doesn't exit until process
// termination. glslang::FinalizeProcess only frees the
// process-wide SharedSymbolTables, NOT this pool. Call popAll()
// explicitly to release the pages back to the system allocator.
//
// Safe here because (a) we're called from atexit, every render
// thread has joined via Vulkan.threadExit (which also runs its
// own popAll-equivalent via ThreadState.cleanup); (b) the SPV
// cache was cleared above, so no compiled blob references the
// pool; (c) FinalizeProcess below won't reach into this pool
// either.
glslang::GetThreadPoolAllocator().popAll();
// Release glslang's process-wide shared state (the version-
// indexed SharedSymbolTables built at first compile).
// Release glslang's process-wide shared state FIRST. This deletes
// SharedSymbolTables[v][s][p][src][stage] entries that hold
// pointers INTO the thread pool; we want their dtors to run
// while the pool memory is still live.
glslang::FinalizeProcess();
// Now destroy this thread's TPoolAllocator entirely. popAll()
// alone is insufficient — it returns pages to glslang's
// internal free list but never gives them back to the system
// allocator (verified empirically: heaptrack total leaked
// unchanged after popAll). The pool is `new`-allocated in
// glslang::InitializeThreadPoolAllocator, so `delete` calls
// ~TPoolAllocator which `free()`s every page.
//
// heaptrack pointed the ~12 MB glslang leak at
// TPoolAllocator::allocate calls rooted in
// shadertoy.spirvFromGlsl on the GUI thread (since
// ghostty_surface_new runs glslang synchronously from
// MainWindow::newTab) — that pool's pages persist until the
// GUI thread exits, but a Qt app's GUI thread only exits at
// process termination, after atexit. Manual delete here gives
// the pages back before exit.
//
// Safe at atexit because every renderer thread has joined
// via Vulkan.threadExit (their pools are independent
// threadlocals already cleaned up), the SPV cache was just
// cleared, and FinalizeProcess just ran.
glslang::TPoolAllocator* pool = &glslang::GetThreadPoolAllocator();
glslang::SetThreadPoolAllocator(nullptr);
delete pool;
}