cgroup: never call do_group_exit() with task->frozen bit set
authorRoman Gushchin <guro@fb.com>
Wed, 8 May 2019 20:34:20 +0000 (13:34 -0700)
committerTejun Heo <tj@kernel.org>
Thu, 9 May 2019 14:56:47 +0000 (07:56 -0700)
I've got two independent reports that cgroup_task_frozen() check
in cgroup_exit() has been triggered by lkp libhugetlbfs-test and
LTP ptrace01 tests.

For example:
[   44.576072] WARNING: CPU: 1 PID: 3028 at kernel/cgroup/cgroup.c:5932 cgroup_exit+0x148/0x160
[   44.577724] Modules linked in: crct10dif_pclmul crc32_pclmul crc32c_intel ghash_clmulni_intel sr_mod cdrom
bochs_drm sg ttm ata_generic pata_acpi ppdev drm_kms_helper snd_pcm syscopyarea aesni_intel snd_timer
sysfillrect sysimgblt snd crypto_simd cryptd glue_helper soundcore fb_sys_fops joydev drm serio_raw pcspkr
ata_piix libata i2c_piix4 floppy parport_pc parport ip_tables
[   44.583106] CPU: 1 PID: 3028 Comm: ptrace-write-hu Not tainted 5.1.0-rc3-00053-g9262503 #5
[   44.584600] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
[   44.586116] RIP: 0010:cgroup_exit+0x148/0x160
[   44.587135] Code: 0f 84 50 ff ff ff 48 8b 85 c8 0c 00 00 48 8b 78 70 e8 ec 2e 00 00 e9 3b ff ff ff f0 ff 43 60
0f 88 72 21 89 00 e9 48 ff ff ff <0f> 0b e9 1b ff ff ff e8 3c 73 f4 ff 66 90 66 2e 0f 1f 84 00 00 00
[   44.590113] RSP: 0018:ffffb25702dcfd30 EFLAGS: 00010002
[   44.591167] RAX: ffff96a7fee32410 RBX: ffff96a7ff1d6000 RCX: dead000000000200
[   44.592446] RDX: ffff96a7ff1d6080 RSI: ffff96a7fec75290 RDI: ffff96a7fec75290
[   44.593715] RBP: ffff96a7fec745c0 R08: ffff96a7fec74658 R09: 0000000000000000
[   44.594985] R10: 0000000000000000 R11: 0000000000000001 R12: ffff96a7fec75101
[   44.596266] R13: ffff96a7fec745c0 R14: ffff96a7ff3bde30 R15: ffff96a7fec75130
[   44.597550] FS:  0000000000000000(0000) GS:ffff96a7dd700000(0000) knlGS:0000000000000000
[   44.598950] CS:  0010 DS: 002b ES: 002b CR0: 0000000080050033
[   44.600098] CR2: 00000000f7a00000 CR3: 000000000d20e000 CR4: 00000000000406e0
[   44.601417] Call Trace:
[   44.602777]  do_exit+0x337/0xc40
[   44.603677]  do_group_exit+0x3a/0xa0
[   44.604610]  get_signal+0x12e/0x8d0
[   44.605533]  ? __switch_to_asm+0x40/0x70
[   44.606503]  do_signal+0x36/0x650
[   44.607409]  ? __switch_to_asm+0x40/0x70
[   44.608383]  ? __schedule+0x267/0x860
[   44.609329]  exit_to_usermode_loop+0x89/0xf0
[   44.610349]  do_fast_syscall_32+0x251/0x2e3
[   44.611357]  entry_SYSENTER_compat+0x7f/0x91
[   44.612376] ---[ end trace e4ca5cfc4b7f7964 ]---

The problem is caused by the ptrace_signal() call in the for loop
in get_signal(). There is a cgroup_enter_frozen() call inside
ptrace_signal(), so after exit from ptrace_signal() the task->frozen
bit might be set. In this case do_group_exit() can be called with the
task->frozen bit set and trigger the warning. This is only place where
we can leave the loop with the task->frozen bit set and without
setting JOBCTL_TRAP_FREEZE and TIF_SIGPENDING.

To resolve this problem, let's move cgroup_leave_frozen(true) call to
just after the fatal label. If the task is going to die, the frozen
bit must be cleared no matter how we get into this point.

Reported-by: kernel test robot <rong.a.chen@intel.com>
Reported-by: Qian Cai <cai@lca.pw>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/signal.c

index 16b72f4f14dfa11071e87e1662906a8053c44326..8607b11ff93607aec42d567cf65d666628d40351 100644 (file)
@@ -2483,10 +2483,6 @@ relock:
                ksig->info.si_signo = signr = SIGKILL;
                sigdelset(&current->pending.signal, SIGKILL);
                recalc_sigpending();
-               current->jobctl &= ~JOBCTL_TRAP_FREEZE;
-               spin_unlock_irq(&sighand->siglock);
-               if (unlikely(cgroup_task_frozen(current)))
-                       cgroup_leave_frozen(true);
                goto fatal;
        }
 
@@ -2608,8 +2604,10 @@ relock:
                        continue;
                }
 
-               spin_unlock_irq(&sighand->siglock);
        fatal:
+               spin_unlock_irq(&sighand->siglock);
+               if (unlikely(cgroup_task_frozen(current)))
+                       cgroup_leave_frozen(true);
 
                /*
                 * Anything else is fatal, maybe with a core dump.