Repository navigation

Heap delete/destroy: one teardown protocol (supersedes #22–#26) - #27

Merged

Jarred-Sumner merged 34 commits into

bun-dev3-v2from

claude/heap-lifecycle

Aug 23, 2026

Commits on Aug 22, 2026

threadlocal: return NULL from a thread-local get after the thread-locals were freed
Jarred-Sumner
committed
f3b763f
fix assertion check in realloc, issue #1371

daanx
authored and
Jarred-Sumner
committed
4cb2e6e
fix numa node count on linux: return the count, not the highest node index

devnexen
authored and
Jarred-Sumner
committed
16cd368
arena: re-abandon a page to the abandoned map even when the freeing thread has no theap for its heap (upstream issue #1364)
Jarred-Sumner
committed
721ffff
page-map: acquire-load the submap entry in _mi_ptr_page; osx: drop the thread-initialized assertion in zone_free
Jarred-Sumner
committed
89b54c8
unix: only opt mappings out of THP when the system setting is always

robobun
authored and
Jarred-Sumner
committed
195ee24
Check the free lists on the cold walks instead of faulting on a corrupted link

robobun
authored and
Jarred-Sumner
committed
a8dfced
Do not dereference page->heap on the cross-thread free path while a concurrent mi_heap_delete can free the heap

robobun
authored and
Jarred-Sumner
committed
1630e86
Make the arena_pages bit clear the last access to the heap when a page is freed

robobun
authored and
Jarred-Sumner
committed
0ac2519
mi_heap_delete: detach the theaps before the pages are moved, free them after

robobun
authored and
Jarred-Sumner
committed
ec23898
test-fork-user-heap: arm the stall before the stall thread is created

robobun
authored and
Jarred-Sumner
committed
7cc5564
test-heap-aba: regression test for the theap cache ABA on a reused heap address

robobun
authored and
Jarred-Sumner
committed
0cde8c7
heap delete/destroy: one teardown protocol instead of per-reader fixes
Jarred-Sumner
committed
8286bfb

Navigation Menu

Repository navigation

Heap delete/destroy: one teardown protocol (supersedes #22–#26) - #27

Heap delete/destroy: one teardown protocol (supersedes #22–#26)#27
Jarred-Sumner merged 34 commits into
bun-dev3-v2from
claude/heap-lifecycle

Commits on Aug 22, 2026

threadlocal: return NULL from a thread-local get after the thread-locals were freed

fix assertion check in realloc, issue #1371

fix numa node count on linux: return the count, not the highest node index

arena: re-abandon a page to the abandoned map even when the freeing thread has no theap for its heap (upstream issue #1364)

page-map: acquire-load the submap entry in _mi_ptr_page; osx: drop the thread-initialized assertion in zone_free

unix: only opt mappings out of THP when the system setting is always

Check the free lists on the cold walks instead of faulting on a corrupted link

Do not dereference page->heap on the cross-thread free path while a concurrent mi_heap_delete can free the heap

Make the arena_pages bit clear the last access to the heap when a page is freed

mi_heap_delete: detach the theaps before the pages are moved, free them after

test-fork-user-heap: arm the stall before the stall thread is created

test-heap-aba: regression test for the theap cache ABA on a reused heap address

heap delete/destroy: one teardown protocol instead of per-reader fixes

Commits on Aug 23, 2026

test-heap-teardown; free an abandoned page without touching its heap after it is unpublished

init: spell out the tail of the mi_tld_detached initializer (-Wmissing-field-initializers)

arena: account a re-abandon on the heap instead of the freeing thread's theap for it; tests: atomics for the heap-aba / heap-delete-race handshakes

Pin GitHub Actions to full-length commit SHAs, see PR #1368

stats/snapshot: read live counters and bitmaps atomically when printing or snapshotting; debug test hooks are atomics

mi_on_thread_idle_start: do not touch the park fields when already parked

ci: drop the Alpine jobs in this fork (setup-alpine's nested action is not SHA-pinned, which the org policy rejects for the whole workflow)

Fix the three tests that failed in the default (malloc override) configuration

test-snapshot: include <process.h> for getpid on Windows; ci: fail-fast off

Revert the scavenger exit handshake; test-stress-subprocs: no mi_subproc_destroy when malloc is overridden

Make the test suite pass on every CI configuration, not just Linux/C/static

tests: geometry-independent purge-holes expectations, heap-mt 1000 iterations, C++ builds of the heap tests on MSVC; stats: mi_stats_copy reads live counters atomically; cmake: -save-temps=obj

scavenger: start when a second thread initializes (not from inside a free), initialize its mutex/condvar eagerly; tests: guarded/32-bit/16K-page expectations

test-purge-holes: sweep-skips picks its page and victim from the actual layout (16 KiB pages, 32-bit); large-pages accepts singleton pages that the sweep cannot reach

page-map: keep the relaxed submap load; lighter TSAN stress; thp-optout not under sanitizers

arena: a page whose page-map registration failed may be freed (debug assertion on the allocation failure path, reached with MI_GUARDED when mprotect fails)

test-park-handoff: 2000 live blocks under MI_GUARDED (one guard mapping per block would exceed vm.max_map_count)

Pass the message to the error handler (mi_register_error_message); no static initializer for the scavenger's mutex/condvar

Do not check-and-repair free lists in the cold walks; only keep the sweep from acting on a list that does not check out

Tests for the three fixes that had none: the sweep on a double-freed list, the scavenger's lazy start, and its signal mask

Take the double-free guard out of the sweep walk again