Handle jemalloc calls after TSD teardown

This only changes behavior for the tsd_generic path, where
pthread_getspecific() can report no TSD after the pthread key destructor
has finished.  Other TSD backends keep tsd_teardown_done() as a constant
false, so the added checks compile out there.

Avoid recreating TSD for late deallocations.  Preserve existing
reincarnation behavior for late allocations and nonzero reallocations.

Add Linux CI coverage for force_tls=0, with and without --enable-debug,
to exercise the Android-equivalent generic TSD path.

Author: spredolac

.github/workflows/linux-ci.yml

@@ -508,6 +508,16 @@ jobs:
               CXX: g++
               CONFIGURE_FLAGS: "--enable-debug --enable-experimental-smallocx --enable-stats --enable-prof"
               EXTRA_CFLAGS: "-Werror -Wno-array-bounds"
+          - env:
+              CC: gcc
+              CXX: g++
+              CONFIGURE_FLAGS: force_tls=0
+              EXTRA_CFLAGS: "-Werror -Wno-array-bounds"
+          - env:
+              CC: gcc
+              CXX: g++
+              CONFIGURE_FLAGS: "force_tls=0 --enable-debug"
+              EXTRA_CFLAGS: "-Werror -Wno-array-bounds"
           - env:
               CC: gcc
               CXX: g++

include/jemalloc/internal/tsd.h

@@ -4,7 +4,8 @@
 /*
  * We put the platform-specific data declarations and inlines into their own
  * header files to avoid cluttering this file.  They define tsd_boot0,
- * tsd_boot1, tsd_boot, tsd_booted_get, tsd_get_allocates, tsd_get, and tsd_set.
+ * tsd_boot1, tsd_boot, tsd_booted_get, tsd_get_allocates,
+ * tsd_teardown_done, tsd_get, and tsd_set.
  */
 #ifdef JEMALLOC_MALLOC_THREAD_CLEANUP
 #	include "jemalloc/internal/jemalloc_preamble.h"

include/jemalloc/internal/tsd_generic.h

@@ -27,6 +27,7 @@ void *tsd_init_check_recursion(tsd_init_head_t *head, tsd_init_block_t *block);
 void  tsd_init_finish(tsd_init_head_t *head, tsd_init_block_t *block);
 
 extern pthread_key_t   tsd_tsd;
+extern pthread_key_t   tsd_thread_initialized_tsd;
 extern tsd_init_head_t tsd_init_head;
 extern tsd_wrapper_t   tsd_boot_wrapper;
 extern bool            tsd_booted;
@@ -51,6 +52,12 @@ tsd_cleanup_wrapper(void *arg) {
 			return;
 		}
 	}
+	/*
+	 * Leave tsd_thread_initialized_tsd set.  It has no destructor and
+	 * intentionally outlives the TSD wrapper, so later jemalloc calls can
+	 * distinguish this fully-torn-down thread from one that has never
+	 * initialized TSD.
+	 */
 	malloc_tsd_dalloc(wrapper);
 }
 
@@ -63,6 +70,11 @@ tsd_wrapper_set(tsd_wrapper_t *wrapper) {
 		malloc_write("<jemalloc>: Error setting TSD\n");
 		abort();
 	}
+	if (pthread_setspecific(tsd_thread_initialized_tsd,
+	    (void *)&tsd_thread_initialized_tsd) != 0) {
+		malloc_write("<jemalloc>: Error setting TSD\n");
+		abort();
+	}
 }
 
 JEMALLOC_ALWAYS_INLINE tsd_wrapper_t *
@@ -116,6 +128,15 @@ tsd_boot0(void) {
 	if (pthread_key_create(&tsd_tsd, tsd_cleanup_wrapper) != 0) {
 		return true;
 	}
+	/*
+	 * This key has no destructor.  It records that the current thread once
+	 * had jemalloc TSD, so a later NULL tsd_tsd value means teardown has
+	 * finished rather than initialization not having happened yet.
+	 */
+	if (pthread_key_create(&tsd_thread_initialized_tsd, NULL) != 0) {
+		pthread_key_delete(tsd_tsd);
+		return true;
+	}
 	tsd_booted = true;
 	tsd_wrapper_set(&tsd_boot_wrapper);
 	tsd_init_finish(&tsd_init_head, &block);
@@ -160,6 +181,12 @@ tsd_get_allocates(void) {
 	return true;
 }
 
+JEMALLOC_ALWAYS_INLINE bool
+tsd_teardown_done(void) {
+	return tsd_booted && pthread_getspecific(tsd_tsd) == NULL
+	    && pthread_getspecific(tsd_thread_initialized_tsd) != NULL;
+}
+
 /* Get/set. */
 JEMALLOC_ALWAYS_INLINE tsd_t *
 tsd_get(bool init) {

include/jemalloc/internal/tsd_malloc_thread_cleanup.h

@@ -50,6 +50,11 @@ tsd_get_allocates(void) {
 	return false;
 }
 
+JEMALLOC_ALWAYS_INLINE bool
+tsd_teardown_done(void) {
+	return false;
+}
+
 /* Get/set. */
 JEMALLOC_ALWAYS_INLINE tsd_t *
 tsd_get(bool init) {

include/jemalloc/internal/tsd_tls.h

@@ -43,6 +43,11 @@ tsd_get_allocates(void) {
 	return false;
 }
 
+JEMALLOC_ALWAYS_INLINE bool
+tsd_teardown_done(void) {
+	return false;
+}
+
 /* Get/set. */
 JEMALLOC_ALWAYS_INLINE tsd_t *
 tsd_get(bool init) {

include/jemalloc/internal/tsd_win.h

@@ -146,6 +146,11 @@ tsd_get_allocates(void) {
 	return true;
 }
 
+JEMALLOC_ALWAYS_INLINE bool
+tsd_teardown_done(void) {
+	return false;
+}
+
 /* Get/set. */
 JEMALLOC_ALWAYS_INLINE tsd_t *
 tsd_get(bool init) {
@@ -220,6 +225,11 @@ tsd_get_allocates(void) {
 	return false;
 }
 
+JEMALLOC_ALWAYS_INLINE bool
+tsd_teardown_done(void) {
+	return false;
+}
+
 /* Get/set. */
 JEMALLOC_ALWAYS_INLINE tsd_t *
 tsd_get(bool init) {

scripts/README_GH_ACTIONS.md

@@ -27,12 +27,12 @@ The script can generate workflows for different platforms:
 
 ### Linux CI (`linux-ci.yml`)
 - **test-linux** (AMD64): `ubuntu-latest` (x86_64)
-  - ~96 configurations covering GCC, Clang, various flags
+  - ~98 configurations covering GCC, Clang, various flags
 - **test-linux-arm64** (ARM64): `ubuntu-24.04-arm` (aarch64)
   - ~14 configurations including large hugepage tests
   - **Note:** Free ARM64 runners (Public Preview) - may have longer queue times during peak hours
 
-**Total:** 110 configurations
+**Total:** 112 configurations
 
 ### macOS CI (`macos-ci.yml`)
 - **test-macos** (Intel): `macos-15-intel` (x86_64)
@@ -161,7 +161,7 @@ The Windows workflow uses:
 ### Linux Build Process
 - Ubuntu Latest for AMD64, Ubuntu 24.04 for ARM64
 - Installs 32-bit cross-compilation dependencies when needed
-- Most comprehensive test matrix (110 configurations)
+- Most comprehensive test matrix (112 configurations)
 
 ## Relationship to Travis CI
 
@@ -178,4 +178,3 @@ To regenerate all workflows after modifying `gen_gh_actions.py`:
 ```
 
 **Note**: The generated files should not be edited by hand. All changes should be made to `gen_gh_actions.py` and then regenerated.
-

scripts/gen_gh_actions.py

@@ -273,6 +273,18 @@ def generate_linux_job(arch):
             'CXX': 'g++',
             'CONFIGURE_FLAGS': '--enable-debug --enable-experimental-smallocx --enable-stats --enable-prof',
             'EXTRA_CFLAGS': '-Werror -Wno-array-bounds'
+        },
+        {
+            'CC': 'gcc',
+            'CXX': 'g++',
+            'CONFIGURE_FLAGS': 'force_tls=0',
+            'EXTRA_CFLAGS': '-Werror -Wno-array-bounds'
+        },
+        {
+            'CC': 'gcc',
+            'CXX': 'g++',
+            'CONFIGURE_FLAGS': 'force_tls=0 --enable-debug',
+            'EXTRA_CFLAGS': '-Werror -Wno-array-bounds'
         }
     ]
 

src/jemalloc.c

@@ -1025,12 +1025,50 @@ isfree(tsd_t *tsd, void *ptr, size_t usize, tcache_t *tcache, bool slow_path) {
 	thread_dalloc_event(tsd, usize);
 }
 
+JEMALLOC_ALWAYS_INLINE bool
+dealloc_no_tsd(void *ptr) {
+	/*
+	 * On the generic pthread_getspecific() TSD path, a deallocation can run
+	 * after the thread's TSD has been torn down.  Using tsd_fetch_min() here
+	 * would allocate and publish a fresh TSD wrapper mid-teardown,
+	 * reincarnating the state teardown just released (the crashes this change
+	 * fixes).  Instead we hand the object straight back to its arena via
+	 * idalloctm() with a NULL tsdn.
+	 *
+	 * This deliberately bypasses the per-thread bookkeeping that the normal
+	 * ifree()/isfree() slow path performs, none of which is available or
+	 * meaningful without a live TSD.  The trade-offs, all acceptable for
+	 * these rare teardown-time frees (the memory itself is still correctly
+	 * returned to the arena):
+	 *   - Profiling: prof_free() is skipped, so a sampled object is not
+	 *     unregistered from its prof context; its bytes stay counted as live
+	 *     in prof stats and in any final heap/leak dump.
+	 *   - Junk filling: opt_junk_free is not applied to the freed region.
+	 *   - Sized dealloc: for sdallocx() the caller-supplied size is ignored
+	 *     (szind is looked up from the extent map), so the sized-dealloc
+	 *     safety check does not run.
+	 *   - Thread events: thread_dalloc_event() does not fire, so this free
+	 *     does not advance decay / tcache GC or dalloc stats.
+	 */
+	if (!tsd_teardown_done()) {
+		return false;
+	}
+
+	idalloctm(TSDN_NULL, ptr, /* tcache */ NULL, /* alloc_ctx */ NULL,
+	    /* is_internal */ false, /* slow_path */ true);
+	return true;
+}
+
 JEMALLOC_NOINLINE
 void
 free_default(void *ptr) {
 	UTRACE(ptr, 0, 0);
 	if (likely(ptr != NULL)) {
 		int saved_errno = get_errno();
+		if (unlikely(dealloc_no_tsd(ptr))) {
+			set_errno(saved_errno);
+			return;
+		}
 		/*
 		 * We avoid setting up tsd fully (e.g. tcache, arena binding)
 		 * based on only free() calls -- other activities trigger the
@@ -1568,6 +1606,9 @@ do_realloc_nonnull_zero(void *ptr) {
 		return do_rallocx(ptr, 1, MALLOCX_TCACHE_NONE, true);
 	} else if (opt_zero_realloc_action == zero_realloc_action_free) {
 		UTRACE(ptr, 0, 0);
+		if (unlikely(dealloc_no_tsd(ptr))) {
+			return NULL;
+		}
 		tsd_t *tsd = tsd_fetch();
 		check_entry_exit_locking(tsd_tsdn(tsd));
 
@@ -1843,6 +1884,12 @@ je_dallocx(void *ptr, int flags) {
 	assert(ptr != NULL);
 	assert(malloc_initialized() || malloc_is_initializer());
 
+	UTRACE(ptr, 0, 0);
+	if (unlikely(dealloc_no_tsd(ptr))) {
+		LOG("core.dallocx.exit", "");
+		return;
+	}
+
 	tsd_t *tsd = tsd_fetch_min();
 	bool   fast = tsd_fast(tsd);
 	check_entry_exit_locking(tsd_tsdn(tsd));
@@ -1851,7 +1898,6 @@ je_dallocx(void *ptr, int flags) {
 	tcache_t *tcache = tcache_get_from_ind(tsd, tcache_ind, !fast,
 	    /* is_alloc */ false);
 
-	UTRACE(ptr, 0, 0);
 	if (likely(fast)) {
 		tsd_assert_fast(tsd);
 		ifree(tsd, ptr, tcache, false);
@@ -1879,6 +1925,12 @@ sdallocx_default(void *ptr, size_t size, int flags) {
 	assert(ptr != NULL);
 	assert(malloc_initialized() || malloc_is_initializer());
 
+	UTRACE(ptr, 0, 0);
+	if (unlikely(dealloc_no_tsd(ptr))) {
+		set_errno(saved_errno);
+		return;
+	}
+
 	tsd_t *tsd = tsd_fetch_min();
 	bool   fast = tsd_fast(tsd);
 	size_t usize = inallocx(tsd_tsdn(tsd), size, flags);
@@ -1888,7 +1940,6 @@ sdallocx_default(void *ptr, size_t size, int flags) {
 	tcache_t *tcache = tcache_get_from_ind(tsd, tcache_ind, !fast,
 	    /* is_alloc */ false);
 
-	UTRACE(ptr, 0, 0);
 	if (likely(fast)) {
 		tsd_assert_fast(tsd);
 		isfree(tsd, ptr, usize, tcache, false);

src/tsd.c

@@ -54,6 +54,7 @@ struct tsd_init_head_s {
 };
 
 pthread_key_t   tsd_tsd;
+pthread_key_t   tsd_thread_initialized_tsd;
 tsd_init_head_t tsd_init_head = {
     ql_head_initializer(blocks), MALLOC_MUTEX_INITIALIZER};