[v4.9-rt PATCH] ARM: mm: remove tasklist locking from update_sections_early()

[Kees Cook]

On Tue, Apr 18, 2017 at 1:48 PM, Grygorii Strashko
<grygorii.strashko at ti.com> wrote:
> The below backtrace can be observed on -rt kernel with CONFIG_DEBUG_RODATA
> option enabled:
>
>  BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:993
>  in_atomic(): 1, irqs_disabled(): 128, pid: 14, name: migration/0
>  1 lock held by migration/0/14:
>   #0:  (tasklist_lock){+.+...}, at: [<c01183e8>] update_sections_early+0x24/0xdc
>  irq event stamp: 38
>  hardirqs last  enabled at (37): [<c08f6f7c>] _raw_spin_unlock_irq+0x24/0x68
>  hardirqs last disabled at (38): [<c01fdfe8>] multi_cpu_stop+0xd8/0x138
>  softirqs last  enabled at (0): [<c01303ec>] copy_process.part.5+0x238/0x1b64
>  softirqs last disabled at (0): [<  (null)>]   (null)
>  Preemption disabled at: [<c01fe244>] cpu_stopper_thread+0x80/0x10c
>  CPU: 0 PID: 14 Comm: migration/0 Not tainted 4.9.21-rt16-02220-g49e319c #15
>  Hardware name: Generic DRA74X (Flattened Device Tree)
>  [<c0112014>] (unwind_backtrace) from [<c010d370>] (show_stack+0x10/0x14)
>  [<c010d370>] (show_stack) from [<c049beb8>] (dump_stack+0xa8/0xd4)
>  [<c049beb8>] (dump_stack) from [<c01631a0>] (___might_sleep+0x1bc/0x2ac)
>  [<c01631a0>] (___might_sleep) from [<c08f7244>] (__rt_spin_lock+0x1c/0x30)
>  [<c08f7244>] (__rt_spin_lock) from [<c08f77a4>] (rt_read_lock+0x54/0x68)
>  [<c08f77a4>] (rt_read_lock) from [<c01183e8>] (update_sections_early+0x24/0xdc)
>  [<c01183e8>] (update_sections_early) from [<c01184b0>] (__fix_kernmem_perms+0x10/0x1c)
>  [<c01184b0>] (__fix_kernmem_perms) from [<c01fe010>] (multi_cpu_stop+0x100/0x138)
>  [<c01fe010>] (multi_cpu_stop) from [<c01fe24c>] (cpu_stopper_thread+0x88/0x10c)
>  [<c01fe24c>] (cpu_stopper_thread) from [<c015edc4>] (smpboot_thread_fn+0x174/0x31c)
>  [<c015edc4>] (smpboot_thread_fn) from [<c015a988>] (kthread+0xf0/0x108)
>  [<c015a988>] (kthread) from [<c0108818>] (ret_from_fork+0x14/0x3c)
>  Freeing unused kernel memory: 1024K (c0d00000 - c0e00000)
>
> The stop_machine() is called with cpus = NULL from fix_kernmem_perms() and
> mark_rodata_ro() which means only one CPU will execute
> update_sections_early() while all other CPUs will spin and wait. Hence,
> it's safe to remove tasklist locking from update_sections_early(). As part of
> this change also mark functions which are local to this module as static
> to simplify code analize in the future.

Hm, yes, good point. It's only every called while other CPUs are stopped.

>
> Cc: Kees Cook <keescook at chromium.org>
> Cc: Laura Abbott <labbott at redhat.com>
> Signed-off-by: Grygorii Strashko <grygorii.strashko at ti.com>
> ---
>  arch/arm/mm/init.c | 8 +++-----
>  1 file changed, 3 insertions(+), 5 deletions(-)
>
> diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> index 370581a..a77953a 100644
> --- a/arch/arm/mm/init.c
> +++ b/arch/arm/mm/init.c
> @@ -693,30 +693,28 @@ static void update_sections_early(struct section_perm perms[], int n)

Maybe this should be renamed update_sections_stopped()? Or at least
comments added to help see why it's safe.

>  {
>         struct task_struct *t, *s;
>
> -       read_lock(&tasklist_lock);
>         for_each_process(t) {
>                 if (t->flags & PF_KTHREAD)
>                         continue;
>                 for_each_thread(t, s)
>                         set_section_perms(perms, n, true, s->mm);
>         }
> -       read_unlock(&tasklist_lock);
>         set_section_perms(perms, n, true, current->active_mm);
>         set_section_perms(perms, n, true, &init_mm);
>  }
>
> -int __fix_kernmem_perms(void *unused)
> +static int __fix_kernmem_perms(void *unused)
>  {
>         update_sections_early(nx_perms, ARRAY_SIZE(nx_perms));
>         return 0;
>  }
>
> -void fix_kernmem_perms(void)
> +static void fix_kernmem_perms(void)
>  {
>         stop_machine(__fix_kernmem_perms, NULL, NULL);
>  }
>
> -int __mark_rodata_ro(void *unused)
> +static int __mark_rodata_ro(void *unused)
>  {
>         update_sections_early(ro_perms, ARRAY_SIZE(ro_perms));
>         return 0;

Yeah, the static marks are all correct, thanks for fixing these!

-Kees

-- 
Kees Cook
Pixel Security