Refactor LG_VADDR detection in configure.ac

The LG_VADDR detection block accumulated duplicated logic and the rdtscp
feature probe was interleaved with virtual-address-bit detection.

Consolidate the detection:
- Wrap the whole case in a single "if LG_VADDR = detect" guard instead
  of repeating it in every "case" section.
- Define LG_VADDR exactly once, after the case.
- Check the cache variable directly for the x86_64 error path.
- Move the rdtscp probe into its own x86_64 case so it stays independent
  of LG_VADDR detection (it must run even when --with-lg-vaddr is given).

There is no functional change. Verified on x86_64 that LG_VADDR and
JEMALLOC_HAVE_RDTSCP are identical to before for auto-detection, and that
--with-lg-vaddr still probes rdtscp.

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 \

configure.ac

@@ -533,13 +533,14 @@ esac
 AC_DEFINE_UNQUOTED([HAVE_CPU_SPINWAIT], [$HAVE_CPU_SPINWAIT], [ ])
 AC_DEFINE_UNQUOTED([CPU_SPINWAIT], [$CPU_SPINWAIT], [ ])
 
+dnl LG_VADDR: number of significant virtual address bits. Auto-detected when --with-lg-vaddr is left unset.
 AC_ARG_WITH([lg_vaddr],
   [AS_HELP_STRING([--with-lg-vaddr=<lg-vaddr>], [Number of significant virtual address bits])],
   [LG_VADDR="$with_lg_vaddr"], [LG_VADDR="detect"])
 
-case "${host_cpu}" in
+if test "x$LG_VADDR" = "xdetect"; then
+  case "${host_cpu}" in
     aarch64)
-    if test "x$LG_VADDR" = "xdetect"; then
       AC_MSG_CHECKING([number of significant virtual address bits])
       if test "x${LG_SIZEOF_PTR}" = "x2" ; then
         #aarch64 ILP32
@@ -549,10 +550,8 @@ case "${host_cpu}" in
         LG_VADDR=48
       fi
       AC_MSG_RESULT([$LG_VADDR])
-    fi
       ;;
     x86_64)
-    if test "x$LG_VADDR" = "xdetect"; then
       AC_CACHE_CHECK([number of significant virtual address bits],
                      [je_cv_lg_vaddr],
                      AC_RUN_IFELSE([AC_LANG_PROGRAM(
@@ -592,15 +591,33 @@ typedef unsigned __int32 uint32_t;
                    [je_cv_lg_vaddr=`cat conftest.out`],
                    [je_cv_lg_vaddr=error],
                    [je_cv_lg_vaddr=57]))
+      if test "x${je_cv_lg_vaddr}" = "xerror" ; then
+        AC_MSG_ERROR([cannot determine number of significant virtual address bits])
+      fi
       if test "x${je_cv_lg_vaddr}" != "x" ; then
         LG_VADDR="${je_cv_lg_vaddr}"
       fi
-      if test "x${LG_VADDR}" != "xerror" ; then
-        AC_DEFINE_UNQUOTED([LG_VADDR], [$LG_VADDR], [ ])
+      ;;
+    *)
+      AC_MSG_CHECKING([number of significant virtual address bits])
+      if test "x${LG_SIZEOF_PTR}" = "x3" ; then
+        LG_VADDR=64
+      elif test "x${LG_SIZEOF_PTR}" = "x2" ; then
+        LG_VADDR=32
+      elif test "x${LG_SIZEOF_PTR}" = "xLG_SIZEOF_PTR_WIN" ; then
+        LG_VADDR="(1U << (LG_SIZEOF_PTR_WIN+3))"
       else
-        AC_MSG_ERROR([cannot determine number of significant virtual address bits])
-      fi
+        AC_MSG_ERROR([Unsupported lg(pointer size): ${LG_SIZEOF_PTR}])
       fi
+      AC_MSG_RESULT([$LG_VADDR])
+      ;;
+  esac
+fi
+AC_DEFINE_UNQUOTED([LG_VADDR], [$LG_VADDR], [ ])
+
+dnl rdtscp feature probe for x86_64
+case "${host_cpu}" in
+  x86_64)
     AC_CACHE_CHECK([rdtscp support],
 		   [je_cv_rdtscp],
 		   AC_RUN_IFELSE([AC_LANG_PROGRAM(
@@ -619,23 +636,8 @@ typedef unsigned __int32 uint32_t;
       AC_DEFINE([JEMALLOC_HAVE_RDTSCP], [ ], [ ])
     fi
     ;;
-  *)
-    if test "x$LG_VADDR" = "xdetect"; then
-      AC_MSG_CHECKING([number of significant virtual address bits])
-      if test "x${LG_SIZEOF_PTR}" = "x3" ; then
-        LG_VADDR=64
-      elif test "x${LG_SIZEOF_PTR}" = "x2" ; then
-        LG_VADDR=32
-      elif test "x${LG_SIZEOF_PTR}" = "xLG_SIZEOF_PTR_WIN" ; then
-        LG_VADDR="(1U << (LG_SIZEOF_PTR_WIN+3))"
-      else
-        AC_MSG_ERROR([Unsupported lg(pointer size): ${LG_SIZEOF_PTR}])
-      fi
-      AC_MSG_RESULT([$LG_VADDR])
-    fi
-    ;;
 esac
-AC_DEFINE_UNQUOTED([LG_VADDR], [$LG_VADDR], [ ])
+
 AC_CACHE_CHECK([asm volatile support],
                [je_cv_asm_volatile],
                AC_RUN_IFELSE([AC_LANG_PROGRAM(

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
-			 * not allowed for the auto arena range.
+			 * Setting thread.arena to an arena in the auto 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,
-			    "thread.arena ctl "
-			    "should not be allowed with percpu arena");
+			    0,
+			    "thread.arena within the per-CPU range should "
+			    "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"