test: add tcache_fiber_migration standalone LTO reproducer

Exercises the tsd thread-pointer hoisting bug: ucontext fibers do
free/swapcontext/malloc in one frame on a worker pool, so a fiber routinely
resumes on a different OS thread; under whole-program LTO the inlined fastpath
then frees/allocs against the previous thread's tcache and crashes.

It is a standalone program (no test harness, so it can be static-linked without
symbol clashes), calling jemalloc via JEMALLOC_MANGLE and linked statically with
--whole-archive so the allocator inlines next to the swapcontext.  Built only
when -flto is in the build flags -- without LTO the allocator is not inlined and
the bug cannot reproduce, so the test would be a meaningless always-pass -- and
gated on ELF + static + have_ucontext (a configure link test, since musl
declares but does not implement getcontext/makecontext/swapcontext).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Makefile.in

@@ -59,6 +59,7 @@ enable_autogen := @enable_autogen@
 enable_doc := @enable_doc@
 enable_shared := @enable_shared@
 enable_static := @enable_static@
+have_ucontext := @have_ucontext@
 enable_prof := @enable_prof@
 enable_zone_allocator := @enable_zone_allocator@
 enable_experimental_smallocx := @enable_experimental_smallocx@
@@ -345,6 +346,24 @@ ifeq (@enable_experimental_smallocx@, 1)
 TESTS_INTEGRATION += \
   $(srcroot)test/integration/smallocx.c
 endif
+# tcache_fiber_migration is a standalone LTO reproducer (issue #2890): the
+# dedicated rule below links it against the static archive (--whole-archive) so
+# the allocator fastpath inlines next to the swapcontext.  It needs the static
+# archive, GNU-ld --whole-archive semantics (the ELF gate), and working
+# ucontext fibers -- have_ucontext is a configure link test because musl
+# declares getcontext/makecontext/swapcontext but does not implement them, so
+# the ELF gate alone would break `make check` on Alpine/musl.  Built only when
+# the build uses LTO (-flto in CFLAGS): without inlining across the swapcontext
+# the bug cannot reproduce.
+ifeq (elf, $(ABI))
+ifeq ($(enable_static), 1)
+ifeq (1, $(have_ucontext))
+ifneq (,$(findstring -flto,$(CFLAGS)))
+TESTS_INTEGRATION += $(srcroot)test/integration/tcache_fiber_migration.c
+endif
+endif
+endif
+endif
 ifeq (@enable_cxx@, 1)
 CPP_SRCS := $(srcroot)src/jemalloc_cpp.cpp
 TESTS_INTEGRATION_CPP := $(srcroot)test/integration/cpp/basic.cpp
@@ -578,6 +597,14 @@ $(objroot)test/integration/%$(EXE): $(objroot)test/integration/%.$(O) $(C_TESTLI
 	@mkdir -p $(@D)
 	$(CC) $(TEST_LD_MODE) $(LDTARGET) $(filter %.$(O),$^) $(call RPATH,$(objroot)lib) $(LJEMALLOC) $(LDFLAGS) $(filter-out -lm,$(filter -lrt -pthread -lstdc++,$(LIBS))) $(LM) $(EXTRA_LDFLAGS)
 
+# tcache_fiber_migration (issue #2890) must inline the allocator next to the
+# swapcontext, so link jemalloc statically (--whole-archive), without the test
+# harness/shared lib; whole-program LTO (from the build's flags) does the
+# inlining.  Explicit rule -- overrides the generic integration rule above.
+$(objroot)test/integration/tcache_fiber_migration$(EXE): $(objroot)test/integration/tcache_fiber_migration.$(O) $(objroot)lib/$(LIBJEMALLOC).$(A)
+	@mkdir -p $(@D)
+	$(CC) $(LDTARGET) $(objroot)test/integration/tcache_fiber_migration.$(O) -Wl,--whole-archive $(objroot)lib/$(LIBJEMALLOC).$(A) -Wl,--no-whole-archive $(LDFLAGS) -pthread $(filter-out -lm,$(LIBS)) $(LM) $(EXTRA_LDFLAGS)
+
 $(objroot)test/integration/cpp/%$(EXE): $(objroot)test/integration/cpp/%.$(O) $(C_TESTLIB_INTEGRATION_OBJS) $(C_UTIL_INTEGRATION_OBJS) $(objroot)lib/$(LIBJEMALLOC).$(IMPORTLIB)
 	@mkdir -p $(@D)
 	$(CXX) $(LDTARGET) $(filter %.$(O),$^) $(call RPATH,$(objroot)lib) $(objroot)lib/$(LIBJEMALLOC).$(IMPORTLIB) $(LDFLAGS) $(filter-out -lm,$(LIBS)) -lm $(EXTRA_LDFLAGS)

configure.ac

@@ -1217,6 +1217,25 @@ if test "$enable_shared$enable_static" = "00" ; then
   AC_MSG_ERROR([Please enable one of shared or static builds])
 fi
 
+dnl Whether ucontext fibers (getcontext/makecontext/swapcontext) actually link.
+dnl musl declares them in <ucontext.h> but exports no implementation, so a
+dnl link test is required -- a header check would pass and then fail to link.
+dnl Used to gate the tcache_fiber_migration test.
+JE_COMPILABLE([ucontext], [
+#include <ucontext.h>
+], [
+	ucontext_t uc, ret;
+	getcontext(&uc);
+	makecontext(&uc, (void (*)(void))0, 0);
+	swapcontext(&ret, &uc);
+], [je_cv_ucontext])
+if test "x${je_cv_ucontext}" = "xyes" ; then
+  have_ucontext="1"
+else
+  have_ucontext="0"
+fi
+AC_SUBST([have_ucontext])
+
 dnl Perform no name mangling by default.
 AC_ARG_WITH([mangling],
   [AS_HELP_STRING([--with-mangling=<map>], [Mangle symbols in <map>])],

test/integration/tcache_fiber_migration.c

@@ -0,0 +1,135 @@
+/*
+ * Regression test for issue #2890: under whole-program LTO the inlined allocator
+ * fastpath caches the TLS-derived tcache base in a register across a
+ * swapcontext, so a fiber resumed on another OS thread uses the previous
+ * thread's tcache -> heap corruption.  See JEMALLOC_TLS_ADDR in tsd_internals.h.
+ *
+ * Standalone (no test harness, so it can be static-linked), calling jemalloc via
+ * JEMALLOC_MANGLE -- it rewrites malloc/free to jemalloc's configured-prefix
+ * symbols, so the calls bind to and inline the static allocator, not libc's
+ * wrappers.  Reproduces only with the allocator inlined next to the swapcontext
+ * (a static --whole-archive link plus whole-program LTO), so the Makefile builds
+ * it only when -flto is in the build flags.
+ */
+#include <pthread.h>
+#include <stdbool.h>
+#include <ucontext.h>
+
+#define JEMALLOC_MANGLE
+#include <jemalloc/jemalloc.h>
+
+enum {
+	num_fibers = 128,
+	num_workers = 8,
+	ops_per_fiber = 20 * 1000,
+	fiber_stack_size = 1 << 16,
+	ready_queue_capacity = 512 /* power of two, > num_fibers */
+};
+
+static ucontext_t fiber_context[num_fibers];
+/* Scheduler context to switch back to when a fiber yields; the resuming worker
+ * writes its own context here, so this changes when the fiber migrates. */
+static ucontext_t *return_context[num_fibers];
+static unsigned fiber_remaining_ops[num_fibers];
+static bool fiber_done[num_fibers];
+
+static unsigned ready_queue[ready_queue_capacity];
+static unsigned ready_queue_head;
+static unsigned ready_queue_tail;
+static unsigned live_fibers;
+static pthread_mutex_t queue_mtx = PTHREAD_MUTEX_INITIALIZER;
+
+static void
+push_ready_fiber(unsigned id) {
+	pthread_mutex_lock(&queue_mtx);
+	ready_queue[ready_queue_head++ & (ready_queue_capacity - 1)] = id;
+	pthread_mutex_unlock(&queue_mtx);
+}
+
+/* Returns a ready fiber id, -1 if none ready now, or -2 if all have finished. */
+static int
+pop_ready_fiber(void) {
+	int id;
+	pthread_mutex_lock(&queue_mtx);
+	if (live_fibers == 0) {
+		id = -2;
+	} else if (ready_queue_tail != ready_queue_head) {
+		id = (int)ready_queue[ready_queue_tail++ & (ready_queue_capacity - 1)];
+	} else {
+		id = -1;
+	}
+	pthread_mutex_unlock(&queue_mtx);
+	return id;
+}
+
+static void
+fiber_run(int id) {
+	void *p = malloc(64);
+	while (fiber_remaining_ops[id] > 0) {
+		fiber_remaining_ops[id]--;
+		free(p); /* push to the CURRENT thread's tcache */
+		/* Yield; may be resumed on a different worker thread. */
+		swapcontext(&fiber_context[id], return_context[id]);
+		p = malloc(64); /* pop; must come from the NEW thread's tcache */
+		*(char *)p = (char)id;
+	}
+	free(p);
+	pthread_mutex_lock(&queue_mtx);
+	fiber_done[id] = true;
+	live_fibers--;
+	pthread_mutex_unlock(&queue_mtx);
+	swapcontext(&fiber_context[id], return_context[id]);
+}
+
+static void *
+worker_thread(void *arg) {
+	ucontext_t scheduler_context;
+	(void)arg;
+	for (;;) {
+		int id = pop_ready_fiber();
+		if (id == -2) {
+			break;
+		}
+		if (id < 0) {
+			continue;
+		}
+		return_context[id] = &scheduler_context;
+		swapcontext(&scheduler_context, &fiber_context[id]);
+		if (!fiber_done[id]) {
+			push_ready_fiber((unsigned)id);
+		}
+	}
+	return NULL;
+}
+
+int
+main(void) {
+	void *stacks[num_fibers];
+	pthread_t threads[num_workers];
+	unsigned i;
+
+	live_fibers = num_fibers;
+	for (i = 0; i < num_fibers; i++) {
+		stacks[i] = malloc(fiber_stack_size);
+		fiber_remaining_ops[i] = ops_per_fiber;
+		getcontext(&fiber_context[i]);
+		fiber_context[i].uc_stack.ss_sp = stacks[i];
+		fiber_context[i].uc_stack.ss_size = fiber_stack_size;
+		fiber_context[i].uc_link = NULL;
+		makecontext(&fiber_context[i], (void (*)(void))fiber_run, 1, (int)i);
+		push_ready_fiber(i);
+	}
+
+	for (i = 0; i < num_workers; i++) {
+		pthread_create(&threads[i], NULL, worker_thread, NULL);
+	}
+	for (i = 0; i < num_workers; i++) {
+		pthread_join(threads[i], NULL);
+	}
+	for (i = 0; i < num_fibers; i++) {
+		free(stacks[i]);
+	}
+
+	/* Completing without a tcache-corruption crash is success. */
+	return live_fibers == 0 ? 0 : 1;
+}