Merge pull request #101 from sysprog21/nested-clone

Sync child TID and retain CoW across nested fork

Author: jserv

src/core/guest.c

@@ -441,12 +441,13 @@ int guest_init(guest_t *g, uint64_t size, uint32_t ipa_bits)
 int guest_init_from_shm(guest_t *g,
                         int shm_fd,
                         uint64_t size,
-                        uint32_t ipa_bits)
+                        uint32_t ipa_bits,
+                        bool retain_shared)
 {
     uint64_t t0;
 
     memset(g, 0, sizeof(*g));
-    g->shm_fd = -1; /* Child does not own the shm */
+    g->shm_fd = -1; /* Child does not own the shm unless retain_shared */
     g->ipa_base = GUEST_IPA_BASE;
     g->elf_load_min = ELF_DEFAULT_BASE;
     g->brk_base = BRK_BASE_DEFAULT;
@@ -471,13 +472,21 @@ int guest_init_from_shm(guest_t *g,
     }
     g->pt_pool_next = g->pt_pool_base;
 
-    /* Map the shm fd MAP_PRIVATE: copy-on-write semantics. Reads see the
-     * parent's frozen snapshot; writes are private to this process. macOS CoW
-     * is page-granular: only modified pages are duplicated.
+    /* Two mapping modes:
+     *   retain_shared: shm_fd is an independent APFS clone of the parent's
+     *     memory (already isolated from the parent). Map MAP_SHARED so the
+     *     child's writes land in the clone file, then keep the fd so the child
+     *     can fclonefileat it for its own nested CoW fork. guest_destroy closes
+     *     it.
+     *   otherwise: shm_fd may be the parent's live fd (clonefile fallback). Map
+     *     MAP_PRIVATE so writes stay private to this process, then close the
+     * fd. macOS CoW is page-granular either way: only modified pages are
+     * duplicated.
      */
+    int map_flags = retain_shared ? MAP_SHARED : MAP_PRIVATE;
     t0 = startup_trace_now_ns();
     g->host_base =
-        mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, shm_fd, 0);
+        mmap(NULL, size, PROT_READ | PROT_WRITE, map_flags, shm_fd, 0);
     startup_trace_step("shm_mmap", t0);
     if (g->host_base == MAP_FAILED) {
         perror("guest: mmap shm");
@@ -486,8 +495,10 @@ int guest_init_from_shm(guest_t *g,
         return -1;
     }
 
-    /* Close the shm fd; the mapping keeps the pages alive */
-    close(shm_fd);
+    if (retain_shared)
+        g->shm_fd = shm_fd; /* Child owns the clone; guest_destroy closes it */
+    else
+        close(shm_fd); /* MAP_PRIVATE mapping keeps the pages alive */
 
     /* Create HVF VM with the same IPA width as the parent */
     hv_return_t ret = HV_ERROR;
@@ -506,6 +517,10 @@ int guest_init_from_shm(guest_t *g,
         log_error("guest: hv_vm_create (shm) failed: %d", (int) ret);
         munmap(g->host_base, size);
         g->host_base = NULL;
+        if (g->shm_fd >= 0) {
+            close(g->shm_fd);
+            g->shm_fd = -1;
+        }
         return -1;
     }
 
@@ -518,6 +533,10 @@ int guest_init_from_shm(guest_t *g,
         hv_vm_destroy();
         munmap(g->host_base, size);
         g->host_base = NULL;
+        if (g->shm_fd >= 0) {
+            close(g->shm_fd);
+            g->shm_fd = -1;
+        }
         return -1;
     }
 

src/core/guest.h

@@ -774,16 +774,22 @@ static inline bool guest_addr_in_infra(const guest_t *g, uint64_t addr)
  */
 int guest_init(guest_t *g, uint64_t size, uint32_t ipa_bits);
 
-/* Initialize guest from a POSIX shared memory fd (CoW fork path). Maps shm_fd
- * MAP_PRIVATE (copy-on-write), creates HVF VM, maps to hypervisor. The child
- * gets an instant CoW snapshot of parent's guest memory without copying. shm_fd
- * is closed after mapping.
+/* Initialize guest from a shared memory fd (CoW fork path). Creates the HVF VM
+ * and maps the fd to the hypervisor. The child gets an instant CoW snapshot of
+ * the parent's guest memory without copying.
+ *
+ * retain_shared selects the mapping: when true, shm_fd is an independent APFS
+ * clone, so it is mapped MAP_SHARED and retained in g->shm_fd (guest_destroy
+ * closes it) so the child can fclonefileat it for nested CoW fork. When false,
+ * shm_fd may be the parent's live fd, so it is mapped MAP_PRIVATE and closed
+ * after mapping. This function takes ownership of shm_fd on every path.
  * Returns 0 on success, -1 on failure.
  */
 int guest_init_from_shm(guest_t *g,
                         int shm_fd,
                         uint64_t size,
-                        uint32_t ipa_bits);
+                        uint32_t ipa_bits,
+                        bool retain_shared);
 
 /* Tear down VM and free guest memory. */
 void guest_destroy(guest_t *g);

src/runtime/fork-state.h

@@ -60,6 +60,22 @@ typedef struct {
     uint64_t rosetta_entry;
     uint64_t kbuf_gpa;
     uint64_t ttbr1;
+    /* Clone TID-sync state for the fork path. glibc's fork wrapper passes
+     * CLONE_CHILD_SETTID | CLONE_CHILD_CLEARTID so the child writes its new TID
+     * into the TCB and clears it on exit. The posix_spawn child has no access
+     * to the original clone() arguments, so the parent forwards them here:
+     * clone_flags carries the CHILD_SETTID / CHILD_CLEARTID bits and ctid_gva
+     * the guest address. Zero for callers (e.g. raw fork(2)) that pass neither.
+     */
+    uint64_t clone_flags;
+    uint64_t ctid_gva;
+    /* Nonzero when the shm fd sent below is an independent fclonefileat clone
+     * (not the parent's live fd). Only then may the child map it MAP_SHARED and
+     * retain it for its own nested CoW fork; the live-fd fallback must stay
+     * MAP_PRIVATE so the child does not share writes with the parent.
+     */
+    uint32_t shm_is_clone;
+    uint32_t _pad2;
 } ipc_header_t;
 
 typedef struct {

src/runtime/forkipc.c

@@ -48,6 +48,18 @@
 #include "debug/log.h"
 #include "debug/syscall-hist.h"
 
+/* Linux clone flags. Shared by the fork-child TID-sync emulation below and
+ * sys_clone further down.
+ */
+#define LINUX_CLONE_VM 0x00000100
+#define LINUX_CLONE_VFORK 0x00004000
+#define LINUX_CLONE_THREAD 0x00010000
+#define LINUX_CLONE_SETTLS 0x00080000
+#define LINUX_CLONE_PARENT_SETTID 0x00100000
+#define LINUX_CLONE_CHILD_CLEARTID 0x00200000
+#define LINUX_CLONE_CHILD_SETTID 0x01000000
+/* LINUX_SIGCHLD defined in syscall_signal.h (included above) */
+
 /* fork_child_main. */
 
 static int fork_child_vfork_notify_fd = -1;
@@ -166,7 +178,8 @@ int fork_child_main(int ipc_fd,
             close(ipc_fd);
             return 1;
         }
-        if (guest_init_from_shm(&g, shm_fd, hdr.guest_size, hdr.ipa_bits) < 0) {
+        if (guest_init_from_shm(&g, shm_fd, hdr.guest_size, hdr.ipa_bits,
+                                hdr.shm_is_clone != 0) < 0) {
             log_error("fork-child: guest_init_from_shm failed");
             close(ipc_fd);
             return 1;
@@ -363,6 +376,30 @@ int fork_child_main(int ipc_fd,
      */
     thread_register_main(vcpu, vexit, hdr.child_pid, regs.sp_el1);
 
+    /* Emulate CLONE_CHILD_SETTID for the fork child. glibc's fork wrapper
+     * passes CLONE_CHILD_SETTID | CLONE_CHILD_CLEARTID so the child's TCB
+     * caches its own TID; without the SETTID write the child keeps the parent's
+     * cached TID and modern glibc trips stack-canary / TLS checks ("stack
+     * smashing detected"). The write goes through guest memory, valid for both
+     * the CoW and region-copy paths. A faulting ctid_gva is the guest's own bad
+     * pointer: warn and continue, matching how the kernel ignores a
+     * child_tidptr fault.
+     *
+     * CLONE_CHILD_CLEARTID is deliberately not honored here. The clear-and-wake
+     * on exit only matters to an in-process joiner waiting on the futex (that
+     * is how the worker-thread exit path serves pthread_join). A fork child is
+     * a separate process with its own address space, so its ctid lives in
+     * memory no other process can observe -- the parent reaps it via
+     * wait4/SIGCHLD, not a cross-process futex. Registering clear_child_tid
+     * would be inert.
+     */
+    if (hdr.clone_flags & LINUX_CLONE_CHILD_SETTID) {
+        int32_t tid32 = (int32_t) hdr.child_pid;
+        if (guest_write_small(&g, hdr.ctid_gva, &tid32, sizeof(tid32)) < 0)
+            log_warn("fork-child: CHILD_SETTID write to 0x%llx failed",
+                     (unsigned long long) hdr.ctid_gva);
+    }
+
     /* Re-publish identity into the child's shim-globals cache: the CoW / region
      * copy inherits the parent's pid/uid values, and the shim's identity fast
      * path would otherwise return the parent's pid to the child. Identity is
@@ -420,16 +457,6 @@ int fork_child_main(int ipc_fd,
 
 /* sys_clone. */
 
-/* Linux clone flags */
-#define LINUX_CLONE_VM 0x00000100
-#define LINUX_CLONE_VFORK 0x00004000
-#define LINUX_CLONE_THREAD 0x00010000
-#define LINUX_CLONE_SETTLS 0x00080000
-#define LINUX_CLONE_PARENT_SETTID 0x00100000
-#define LINUX_CLONE_CHILD_CLEARTID 0x00200000
-#define LINUX_CLONE_CHILD_SETTID 0x01000000
-/* LINUX_SIGCHLD defined in syscall_signal.h (included above) */
-
 /* Namespace flags. elfuse implements no namespace isolation. Both sys_clone and
  * sys_clone3 reject them.
  */
@@ -1528,6 +1555,10 @@ int64_t sys_clone(hv_vcpu_t vcpu,
         .rosetta_entry = g->rosetta_entry,
         .kbuf_gpa = g->kbuf_gpa,
         .ttbr1 = g->ttbr1,
+        .clone_flags =
+            flags & (LINUX_CLONE_CHILD_SETTID | LINUX_CLONE_CHILD_CLEARTID),
+        .ctid_gva = ctid_gva,
+        .shm_is_clone = (snapshot_shm_fd >= 0) ? 1 : 0,
     };
     if (fork_ipc_write_all(ipc_sock, &hdr, sizeof(hdr)) < 0) {
         log_error("clone: failed to send header");

tests/manifest.txt

@@ -95,6 +95,7 @@ test-signal-thread
 
 [section] Fork edge cases
 test-clone3                    # diff=skip
+test-clone-childtid
 test-fork-exec $TESTDIR/echo-test
 test-fork-lowbase
 

tests/test-clone-childtid.c

@@ -0,0 +1,107 @@
+/* Test CLONE_CHILD_SETTID / CLONE_CHILD_CLEARTID on the fork (posix_spawn) path
+ *
+ * Copyright 2026 elfuse contributors
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Issue #99: glibc's fork wrapper clones with CLONE_CHILD_SETTID |
+ * CLONE_CHILD_CLEARTID | SIGCHLD. The child's TID must be written into the
+ * ctid address so glibc's TCB caches the right value. This calls clone()
+ * directly with those exact flags (no CLONE_VM/THREAD/VFORK, so elfuse takes
+ * the fork helper-process path) and checks the child observes its own TID at
+ * the ctid slot -- glibc-version-independent, unlike the canary symptom.
+ *
+ * Raw syscall throughout: glibc's own clone wrapper does not expose the ctid
+ * arg, and we want to exercise elfuse's handling rather than libc's.
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+#include <unistd.h>
+#include <sched.h>
+#include <sys/syscall.h>
+#include <sys/wait.h>
+#include <linux/sched.h>
+
+#ifndef CLONE_CHILD_CLEARTID
+#define CLONE_CHILD_CLEARTID 0x00200000
+#endif
+#ifndef CLONE_CHILD_SETTID
+#define CLONE_CHILD_SETTID 0x01000000
+#endif
+
+static volatile int child_tid_slot;
+
+int main(void)
+{
+    /* aarch64 clone(2): clone(flags, stack, parent_tid, tls, child_tid). */
+    unsigned long flags = CLONE_CHILD_SETTID | CLONE_CHILD_CLEARTID | SIGCHLD;
+    long rc = syscall(SYS_clone, flags, (void *) 0, (void *) 0, (void *) 0,
+                      (void *) &child_tid_slot);
+    if (rc < 0) {
+        printf("test-clone-childtid: clone failed -- FAIL\n");
+        return 1;
+    }
+
+    if (rc == 0) {
+        /* Child: the kernel (here, elfuse) must have written our TID into the
+         * ctid slot before we resumed.
+         */
+        pid_t tid = (pid_t) syscall(SYS_gettid);
+        if (child_tid_slot != tid) {
+            /* Cannot printf reliably from a possibly-confused child; encode the
+             * result in the exit status instead.
+             */
+            _exit(child_tid_slot == 0 ? 2 : 3);
+        }
+
+        /* Nested clone: the child forks a grandchild with the same flags. This
+         * exercises the child-side CoW shm retention (issue #99 part 2): the
+         * child must be able to clone its own memory again, and the grandchild
+         * must likewise see a fresh TID at its ctid slot.
+         */
+        static volatile int grand_tid_slot;
+        long grc = syscall(SYS_clone, flags, (void *) 0, (void *) 0, (void *) 0,
+                           (void *) &grand_tid_slot);
+        if (grc < 0)
+            _exit(4);
+        if (grc == 0) {
+            pid_t gtid = (pid_t) syscall(SYS_gettid);
+            _exit(grand_tid_slot == gtid ? 0 : 5);
+        }
+        int gstatus;
+        if (waitpid((pid_t) grc, &gstatus, 0) < 0)
+            _exit(6);
+        if (!WIFEXITED(gstatus) || WEXITSTATUS(gstatus) != 0)
+            _exit(7);
+        _exit(0);
+    }
+
+    int status;
+    if (waitpid((pid_t) rc, &status, 0) < 0) {
+        printf("test-clone-childtid: waitpid failed -- FAIL\n");
+        return 1;
+    }
+    if (!WIFEXITED(status)) {
+        printf(
+            "test-clone-childtid: child did not exit cleanly (0x%x) -- FAIL\n",
+            status);
+        return 1;
+    }
+    switch (WEXITSTATUS(status)) {
+    case 0:
+        printf("test-clone-childtid: child saw its TID at ctid -- PASS\n");
+        return 0;
+    case 2:
+        printf(
+            "test-clone-childtid: ctid slot still 0 (SETTID ignored) -- "
+            "FAIL\n");
+        return 1;
+    case 3:
+        printf("test-clone-childtid: ctid slot holds wrong TID -- FAIL\n");
+        return 1;
+    default:
+        printf("test-clone-childtid: unexpected child exit %d -- FAIL\n",
+               WEXITSTATUS(status));
+        return 1;
+    }
+}

tests/test-matrix.sh

@@ -453,6 +453,7 @@ run_unit_tests()
 
     printf "\nProcess tests\n"
     test_check "$runner" "test-fork" "PASS" "$bindir/test-fork"
+    test_check "$runner" "test-clone-childtid" "PASS" "$bindir/test-clone-childtid"
     test_check "$runner" "test-exec" "exec-works" "$bindir/test-exec" "$bindir/echo-test" exec-works
     test_check "$runner" "test-fork-exec" "PASS" "$bindir/test-fork-exec" "$bindir/echo-test"
     test_check "$runner" "test-cloexec" "PASS" "$bindir/test-cloexec"