[PATCH] pid_namespace: make zap_pid_ns_processes() freezable

[Aviv Vaknin]

When the init of a pid namespace exits, zap_pid_ns_processes() waits
until every other pid in the namespace has been freed. That wait has no
upper bound: a tracer outside the namespace that doesn't reap its traced
zombies keeps their pids allocated (see the comment above the final
loop). While a task waits there, every suspend attempt fails:

  Freezing user space processes failed after 20.006 seconds (2 tasks refusing to freeze, wq_busy=0):
  task:pidns-freeze-re state:S stack:14200 pid:104   tgid:104   ppid:99     task_flags:0x40014c flags:0x00080000
  task:pidns-freeze-re state:S stack:14160 pid:109   tgid:105   ppid:104    task_flags:0x40044c flags:0x00080802

and keeps failing until the tracer reaps or exits, because neither wait
can be frozen:

- The first loop sleeps in kernel_wait4(). The freezer's fake signal
  ends that wait with -ERESTARTSYS, then the loop clears TIF_SIGPENDING
  and waits again without ever calling try_to_freeze().

- The final loop sleeps in TASK_INTERRUPTIBLE without TASK_FREEZABLE.
  Worse, nothing clears the TIF_SIGPENDING that the fake signal set, so
  from the first suspend attempt on schedule() returns at once and the
  task spins at 100% CPU for as long as the wait lasts.

Neither loop has ever been freezable. The spin came with
commit b9a985db9896 ("pid_ns: Sleep in TASK_INTERRUPTIBLE in zap_pid_ns_processes"),
which replaced TASK_UNINTERRUPTIBLE to keep such long waits from
triggering the hung task detector.

This happens in practice with Chromium-based browsers, whose renderers
are pid namespace inits. If a renderer crashes while the browser shuts
down, crashpad (outside the namespace) ptrace-attaches its threads to
dump it. When the renderer is SIGKILLed mid-dump, crashpad keeps the
dead threads as unreaped traced zombies, and it can then block in
waitpid() on the renderer's last thread, which is itself waiting in
zap_pid_ns_processes() for those zombies. One such renderer blocks
suspend system-wide until crashpad is killed or the machine reboots.
The userspace side is being reported to Chromium separately.

Call try_to_freeze() after kernel_wait4() in the first loop, and sleep
in TASK_IDLE | TASK_FREEZABLE in the final loop, as coredump_task_exit()
already does on the same exit path. Like TASK_INTERRUPTIBLE, TASK_IDLE
keeps the hung task detector quiet and adds no load, but it is not
woken by signals, so a pending signal can no longer turn the wait into
a busy loop. TASK_FREEZABLE lets the freezer freeze the task in place;
a free_pid() wakeup that arrives while it is frozen is kept in
->saved_state and acted on at thaw.

Tested on v7.3-rc5 in QEMU with PROVE_LOCKING, DEBUG_ATOMIC_SLEEP and
DETECT_HUNG_TASK (10 s timeout), using a reproducer in which a tracer
PTRACE_SEIZEs two threads of a nested pid namespace init and doesn't
reap them (it works unprivileged too, from a user namespace), and using
the real browser and crashpad trigger:

- before: freezing fails after 20 s, and afterwards the final loop
  spins (500 ticks of system time per 5 s).
- after: freezing completes in 0.001 s, the waiting task sits in state
  I without using CPU, no hung task or lockdep report appears over
  30-45 s, and the task exits as soon as the zombies are reaped.

Fixes: 3eb07c8c8adb ("pid namespaces: destroy pid namespace on init's death")
Fixes: b9a985db9896 ("pid_ns: Sleep in TASK_INTERRUPTIBLE in zap_pid_ns_processes")
Cc: stable@xxxxxxxxxxxxxxx # 6.1+
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Aviv Vaknin <vaknins33@xxxxxxxxx>
---
Based on vfs kernel-7.4.signal (also in next-20260930), on top of
5c0985dd3c32 ("pid_namespace: prevent TIF_NOTIFY_SIGNAL from
interrupting the reaper"). That commit keeps TIF_NOTIFY_SIGNAL from
waking the reaper; this one is about the freezer, whose fake signal
sets TIF_SIGPENDING, which the guard doesn't cover. Retested on this
base with the reproducer below (same debug config): freezing completes
in 0.000 s, no spin, no warnings. A version for v7.3-rc5 is available
if wanted.

Reproducer (run as root in a VM, then follow its instructions):

  // SPDX-License-Identifier: GPL-2.0
  /*
   * pidns-freeze-repro: a task waiting in zap_pid_ns_processes() blocks suspend.
   *
   *   main (tracer, root pid ns)
   *    `- A: pid 1 of pid ns A, single-threaded
   *       `- B: pid 1 of nested pid ns B, multithreaded
   *
   * main PTRACE_SEIZEs non-leader threads of B, then A exits.  zap(A) kills B;
   * the seized threads stay traced zombies until main reaps them, keeping ns B's
   * pids allocated.  B's last thread waits in zap's final loop, A in zap's first
   * loop (kernel_wait4() for B).  While "stuck" is shown, run as root:
   *   echo freezer > /sys/power/pm_test; echo mem > /sys/power/state
   * Unpatched, freezing fails after 20 s ("2 tasks refusing to freeze") and B's
   * thread then busy-loops in zap.  At the end main reaps, ending the hang.
   *
   * Build: gcc -O2 -Wall -pthread -o pidns-freeze-repro pidns-freeze-repro.c
   * Usage: ./pidns-freeze-repro [hold_seconds]  (root, in a VM; default: Enter)
   * With -DUSERNS, A also gets a new user ns, so no privilege is needed.
   */
  #define _GNU_SOURCE
  #include <dirent.h>
  #include <pthread.h>
  #include <sched.h>
  #include <signal.h>
  #include <stdarg.h>
  #include <stdio.h>
  #include <stdlib.h>
  #include <string.h>
  #include <sys/ptrace.h>
  #include <sys/wait.h>
  #include <unistd.h>

  #ifdef USERNS
  #define NS_FLAGS (CLONE_NEWUSER | CLONE_NEWPID)
  #else
  #define NS_FLAGS CLONE_NEWPID
  #endif
  #define NTHREADS 4	/* B's threads besides its leader */
  #define NSEIZE   2	/* how many of them main seizes */

  static int ready[2], go[2];	/* B -> main: threads are up; main -> A: exit */
  static char stack_a[256 << 10], stack_b[256 << 10];

  static void *idle_thread(void *arg) { for (;;) pause(); return arg; }

  static int init_b(void *arg)		/* pid 1 of ns B */
  {
  	pthread_t t;

  	for (int i = 0; i < NTHREADS; i++)
  		pthread_create(&t, NULL, idle_thread, arg);
  	if (write(ready[1], "r", 1) != 1)
  		_exit(1);
  	for (;;) pause();
  }

  static int init_a(void *arg)		/* pid 1 of ns A */
  {
  	close(go[1]);			/* only main may hold the write end */
  	if (clone(init_b, stack_b + sizeof(stack_b), CLONE_NEWPID | SIGCHLD, arg) < 0)
  		_exit(1);
  	while (read(go[0], &arg, 1) > 0)	/* EOF: main has seized */
  		;
  	_exit(0);			/* -> zap_pid_ns_processes(A) */
  }

  /* Read a small /proc file into buf ("" if it is gone). */
  static char *slurp(char *buf, size_t len, const char *fmt, ...)
  {
  	char path[300];
  	va_list ap;
  	FILE *f;

  	va_start(ap, fmt);
  	vsnprintf(path, sizeof(path), fmt, ap);
  	va_end(ap);
  	buf[0] = 0;
  	if ((f = fopen(path, "r"))) {
  		buf[fread(buf, 1, len - 1, f)] = 0;
  		fclose(f);
  	}
  	return buf;
  }

  /* Global pid of A's child B: the process whose PPid is A. */
  static pid_t child_of(pid_t parent)
  {
  	DIR *d = opendir("/proc");
  	struct dirent *e;
  	char buf[512], *p;
  	int ppid;

  	while (d && (e = readdir(d))) {
  		p = strrchr(slurp(buf, sizeof(buf), "/proc/%s/stat", e->d_name), ')');
  		if (p && sscanf(p + 2, "%*c %d", &ppid) == 1 && ppid == parent)
  			return atoi(e->d_name);	/* (leaks d; fine here) */
  	}
  	return -1;
  }

  /* Print a task's state and wchan; return whether it is exiting (PF_EXITING). */
  static int show(const char *what, pid_t pid, pid_t tid)
  {
  	char st[512], wchan[64], *p, c = '?';
  	unsigned int flags = 0;

  	p = strrchr(slurp(st, sizeof(st), "/proc/%d/task/%d/stat", pid, tid), ')');
  	if (p)
  		sscanf(p + 2, "%c %*d %*d %*d %*d %*d %u", &c, &flags);
  	slurp(wchan, sizeof(wchan), "/proc/%d/task/%d/wchan", pid, tid);
  	printf("  %s tid %d: state %c, flags %#x, wchan %s\n", what, tid, c, flags, wchan);
  	return flags & 0x4;
  }

  int main(int argc, char **argv)
  {
  	pid_t a, b, tid, seized[NSEIZE];
  	int n = 0, status;
  	struct dirent *e;
  	char buf[64];
  	DIR *d;

  	if (pipe(ready) || pipe(go))
  		return perror("pipe"), 1;
  	if ((a = clone(init_a, stack_a + sizeof(stack_a), NS_FLAGS | SIGCHLD, NULL)) < 0)
  		return perror("clone (run as root)"), 1;
  	if (read(ready[0], buf, 1) != 1 || (b = child_of(a)) < 0)
  		return fprintf(stderr, "setup failed\n"), 1;

  	/* From outside both namespaces, seize non-leader threads of B. */
  	snprintf(buf, sizeof(buf), "/proc/%d/task", b);
  	for (d = opendir(buf); d && (e = readdir(d)) && n < NSEIZE; )
  		if ((tid = atoi(e->d_name)) > 0 && tid != b && !ptrace(PTRACE_SEIZE, tid, 0, 0))
  			seized[n++] = tid;
  	printf("A=%d B=%d: seized %d threads of B\n", a, b, n);

  	close(go[1]);				/* A exits now */
  	sleep(2);
  	if (!n || waitpid(a, &status, WNOHANG) != 0 || !show("A", a, a))
  		return printf("not reproduced\n"), 1;
  	for (rewinddir(d); (e = readdir(d)); )
  		if ((tid = atoi(e->d_name)) > 0)
  			show("B", b, tid);
  	printf("stuck: A and B can't finish exiting until main reaps its %d traced threads\n", n);
  	printf("now run: echo freezer > /sys/power/pm_test; echo mem > /sys/power/state\n");
  	fflush(stdout);
  	if (argc > 1)
  		sleep(atoi(argv[1]));
  	else
  		getchar();

  	/* Reap the traced zombies: ns B empties, then B and A finish exiting. */
  	for (int i = 0, left = n; left && i < 500; i++, usleep(10000))
  		for (int k = 0; k < n; k++)
  			if (seized[k] && waitpid(seized[k], &status, __WALL | WNOHANG) == seized[k])
  				seized[k] = 0, left--;
  	for (int i = 0; i < 500; i++, usleep(10000))
  		if (waitpid(a, &status, WNOHANG) == a)
  			return printf("reaped the zombies: hang resolved\n"), 0;
  	printf("still stuck after reaping\n");
  	return 2;
  }

 kernel/pid_namespace.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/kernel/pid_namespace.c b/kernel/pid_namespace.c
index 8bc9edb40..abdc05462 100644
--- a/kernel/pid_namespace.c
+++ b/kernel/pid_namespace.c
@@ -23,6 +23,7 @@
 #include <linux/sched/task.h>
 #include <linux/sched/signal.h>
 #include <linux/idr.h>
+#include <linux/freezer.h>
 #include <linux/nstree.h>
 #include <uapi/linux/wait.h>
 #include "pid_sysctl.h"
@@ -243,6 +244,11 @@ void zap_pid_ns_processes(struct pid_namespace *pid_ns)
 	do {
 		clear_thread_flag(TIF_SIGPENDING);
 		rc = kernel_wait4(-1, NULL, __WALL, NULL);
+		/*
+		 * The freezer's fake signal ends the wait with -ERESTARTSYS;
+		 * freeze here, or one stuck pid namespace blocks suspend.
+		 */
+		try_to_freeze();
 	} while (rc != -ECHILD);
 
 	/*
@@ -269,7 +275,13 @@ void zap_pid_ns_processes(struct pid_namespace *pid_ns)
 	 * free_pid() will awaken this task.
 	 */
 	for (;;) {
-		set_current_state(TASK_INTERRUPTIBLE);
+		/*
+		 * TASK_IDLE: no hung task warning or load for a wait that can
+		 * last as long as a tracer keeps a zombie, and a pending signal
+		 * (e.g. the freezer's fake one) can't turn it into a busy loop.
+		 * TASK_FREEZABLE: let the freezer freeze us while we wait.
+		 */
+		set_current_state(TASK_IDLE | TASK_FREEZABLE);
 		if (pid_ns->pid_allocated == init_pids)
 			break;
 		schedule();

base-commit: ed14a591175bb5f56c2936b082cffb9e4b935e6d
-- 
2.55.0