Allow resuming per-CPU arena selection via thread.arena (#2957)

* Allow resuming per-CPU arena selection via thread.arena

With percpu_arena enabled, thread.arena control is one-directional. A
thread can be bound to a manually created arena (an index at or above the
per-CPU auto range) to route a bounded region of work to a dedicated,
long-lived arena, but there is no way back: thread_arena_ctl returns EPERM
for any index within the auto range, and arena_choose_impl only re-selects
a per-CPU arena for threads whose current arena is already in that range.
So once a thread is bound to a manual arena it stays pinned there forever,
and its later allocations land there instead of following the CPU.

Treat setting thread.arena to an index within the per-CPU range as a
request to resume automatic per-CPU selection: hand the thread back to
percpu management (rebinding it to the current CPU's arena via
percpu_arena_update) and return 0 instead of EPERM. Binding to a manual
arena is unchanged. The requested index is advisory; under percpu the
thread is governed by its current CPU, so it resumes on the current CPU's
arena regardless of the value passed.

Add test/unit/percpu_arena_resume covering the manual-arena to resume
round trip with allocs & dallocs happening meanwhile, and update 
test_thread_arena, which asserted the old EPERM.

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

Makefile.in

@@ -276,6 +276,7 @@ TESTS_UNIT := \
 	$(srcroot)test/unit/pack.c \
 	$(srcroot)test/unit/pages.c \
 	$(srcroot)test/unit/peak.c \
+	$(srcroot)test/unit/percpu_arena_resume.c \
 	$(srcroot)test/unit/ph.c \
 	$(srcroot)test/unit/prng.c \
 	$(srcroot)test/unit/prof_accum.c \

src/ctl.c

@@ -2374,10 +2374,18 @@ thread_arena_ctl(tsd_t *tsd, const size_t *mib, size_t miblen, void *oldp,
 	if (have_percpu_arena && PERCPU_ARENA_ENABLED(opt_percpu_arena)) {
 		if (newind < percpu_arena_ind_limit(opt_percpu_arena)) {
 			/*
-			 * If perCPU arena is enabled, thread_arena control is
+			 * Setting thread.arena to an arena in the auto range
-			 * not allowed for the auto arena range.
+			 * means "resume automatic per-CPU selection" rather than
+			 * pinning to a specific per-CPU arena. This lets a caller
+			 * that temporarily switched to a manually managed arena
+			 * (e.g. a scoped guard) hand the thread back to per-CPU
+			 * management. It is otherwise impossible: a thread bound
+			 * to a manual arena is never reclaimed by percpu (see
+			 * arena_choose_impl), so without this it would stay
+			 * pinned forever.
 			 */
-			return EPERM;
+			percpu_arena_update(tsd, percpu_arena_choose());
+			return 0;
 		}
 	}
 

test/unit/mallctl.c

@@ -764,12 +764,17 @@ TEST_BEGIN(test_thread_arena) {
 		    0, "Unexpected mallctl() failure");
 		new_arena_ind = percpu_arena_ind_limit(opt_percpu_arena) - 1;
 		if (old_arena_ind != new_arena_ind) {
+			/*
+			 * Setting thread.arena to an index within the per-CPU
+			 * range resumes automatic per-CPU selection rather than
+			 * failing (see test/unit/percpu_arena_resume.c).
+			 */
 			expect_d_eq(
 			    mallctl("thread.arena", (void *)&old_arena_ind, &sz,
 			        (void *)&new_arena_ind, sizeof(unsigned)),
-			    EPERM,
+			    0,
-			    "thread.arena ctl "
+			    "thread.arena within the per-CPU range should "
-			    "should not be allowed with percpu arena");
+			    "resume per-CPU selection");
 		}
 	}
 }

test/unit/percpu_arena_resume.c

@@ -0,0 +1,81 @@
+#include "test/jemalloc_test.h"
+
+/*
+ * Under percpu_arena, binding a thread to a manual arena (an index at or above
+ * the per-CPU auto range) is one-way: percpu never reclaims it (see
+ * arena_choose_impl). Setting thread.arena back to an index in the auto range
+ * resumes per-CPU selection instead of failing with EPERM.
+ */
+TEST_BEGIN(test_thread_arena_resume_percpu) {
+	test_skip_if(!have_percpu_arena
+	    || !PERCPU_ARENA_ENABLED(opt_percpu_arena));
+
+	unsigned limit = percpu_arena_ind_limit(opt_percpu_arena);
+	/* Bypass the tcache so every allocation and free hits the arena. */
+	const int flags = MALLOCX_TCACHE_NONE;
+
+	void *warm = mallocx(1, 0);
+	expect_ptr_not_null(warm, "Unexpected mallocx() failure");
+	dallocx(warm, 0);
+
+	unsigned cur;
+	size_t sz = sizeof(cur);
+	expect_d_eq(mallctl("thread.arena", (void *)&cur, &sz, NULL, 0), 0,
+	    "Unexpected mallctl() failure");
+	expect_u_lt(cur, limit, "Thread should start on a per-CPU arena");
+
+	unsigned manual;
+	sz = sizeof(manual);
+	expect_d_eq(mallctl("arenas.create", (void *)&manual, &sz, NULL, 0), 0,
+	    "Unexpected arenas.create() failure");
+	expect_u_ge(manual, limit, "A manual arena is outside the per-CPU range");
+
+	unsigned old;
+	sz = sizeof(old);
+	expect_d_eq(mallctl("thread.arena", (void *)&old, &sz, (void *)&manual,
+	    sizeof(manual)), 0, "Binding to a manual arena should be allowed");
+	sz = sizeof(cur);
+	expect_d_eq(mallctl("thread.arena", (void *)&cur, &sz, NULL, 0), 0,
+	    "Unexpected mallctl() failure");
+	expect_u_eq(cur, manual, "Thread should be bound to the manual arena");
+
+	void *p_manual = mallocx(1024, flags);
+	expect_ptr_not_null(p_manual, "Unexpected mallocx() failure");
+	unsigned found;
+	sz = sizeof(found);
+	expect_d_eq(mallctl("arenas.lookup", (void *)&found, &sz,
+	    (void *)&p_manual, sizeof(p_manual)), 0,
+	    "Unexpected arenas.lookup() failure");
+	expect_u_eq(found, manual, "Allocation should come from the manual arena");
+
+	void *scratch = mallocx(1024, flags);
+	expect_ptr_not_null(scratch, "Unexpected mallocx() failure");
+	dallocx(scratch, flags);
+
+	unsigned resume = 0;
+	expect_d_eq(mallctl("thread.arena", NULL, NULL, (void *)&resume,
+	    sizeof(resume)), 0, "Should resume per-CPU selection, not fail");
+	sz = sizeof(cur);
+	expect_d_eq(mallctl("thread.arena", (void *)&cur, &sz, NULL, 0), 0,
+	    "Unexpected mallctl() failure");
+	expect_u_lt(cur, limit, "Thread should be back on a per-CPU arena");
+
+	void *p_percpu = mallocx(1024, flags);
+	expect_ptr_not_null(p_percpu, "Unexpected mallocx() failure");
+	sz = sizeof(found);
+	expect_d_eq(mallctl("arenas.lookup", (void *)&found, &sz,
+	    (void *)&p_percpu, sizeof(p_percpu)), 0,
+	    "Unexpected arenas.lookup() failure");
+	expect_u_lt(found, limit, "Allocation should come from a per-CPU arena");
+	dallocx(p_percpu, flags);
+
+	/* Free the manual-arena region while bound to a different arena. */
+	dallocx(p_manual, flags);
+}
+TEST_END
+
+int
+main(void) {
+	return test(
+	    test_thread_arena_resume_percpu);
+}

test/unit/percpu_arena_resume.sh

@@ -0,0 +1,3 @@
+#!/bin/sh
+
+export MALLOC_CONF="percpu_arena:percpu"