Re: Runtime trouble with commit dbd952127d (seccomp: introduce writer locking)

[Guenter Roeck <linux@roeck-us.net>]

On 08/10/2014 12:33 PM, Oleg Nesterov wrote:
> On 08/09, Kees Cook wrote:
>> On Sat, Aug 9, 2014 at 6:47 PM, Guenter Roeck <private@roeck-us.net> wrote:
>>> Hi folks,
>>>
>>> I am having some trouble with commit dbd952127d (seccomp: introduce
>>> writer locking) when running my qemu tests on the upstream kernel.
>> Eek, sorry this is causing you trouble!
>>
>>> With powerpc, I get the following crash.
>>>
>>> ftrace: allocating 20093 entries in 59 pages
>>> ------------[ cut here ]------------
>>> kernel BUG at kernel/fork.c:1108!
>> For your tree, does this resolve to copy_seccomp()'s
>>
>>          BUG_ON(!spin_is_locked(&current->sighand->siglock));
>>
>> line?
> This is almost off-topic, and I too do not understand whats going on...
Some progress. SMP must be disabled for the problem to be seen.

The underlying spinlock structure, specifically arch_spinlock_t, is from
include/linux/spinlock_types_up.h (not as one would innocently assume
from arch/powerpc/include/asm/spinlock_types.h).

In include/linux/spinlock_types_up.h, arch_spinlock_t is defined as

typedef struct {
         /* no debug version on UP */
} arch_rwlock_t;

if spinlock debugging is disabled. With this definition, it is obviously
not possible to detect if the spinlock is locked or not. Actually, the
same file defines

#define arch_spin_is_locked(lock)       ((void)(lock), 0)

so the BUG is really not at all surprising.

That means that the broken configuration is (CONFIG_DEBUG_SPINLOCK=n, CONFIG_SMP=n).
It also means that the BUG_ON checks introduced with the seccomp commit
will cause this configuration to fail hard at least for architectures where CONFIG_SMP
can be disabled, and if those architectures use include/linux/spinlock_types_up.h.

Guenter