Re: [PATCH] random: Don't freeze in add_hwgenerator_randomness() if stopping kthread

[Maciej S. Szmigiero]

On 15.11.2019 18:44, Stephen Boyd wrote:
> Quoting Maciej S. Szmigiero (2019-11-10 05:55:42)
>> Since commit 59b569480dc8
>> ("random: Use wait_event_freezable() in add_hwgenerator_randomness()")
>> there is a race in add_hwgenerator_randomness() between freezing and
>> stopping the calling kthread.
>>
>> This commit changed wait_event_interruptible() call with
>> kthread_freezable_should_stop() as a condition into wait_event_freezable()
>> with just kthread_should_stop() as a condition to fix a warning that
>> kthread_freezable_should_stop() might sleep inside the wait.
>>
>> wait_event_freezable() ultimately calls __refrigerator() with its
>> check_kthr_stop argument set to false, which causes it to keep the kthread
>> frozen even if somebody calls kthread_stop() on it.
>>
>> Calling wait_event_freezable() with kthread_should_stop() as a condition
>> is racy because it doesn't take into account the situation where this
>> condition becomes true on a kthread marked for freezing only after this
>> condition has already been checked.
>>
>> Calling freezing() should avoid the issue that the commit 59b569480dc8 has
>> fixed, as it is only a checking function, it doesn't actually do the
>> freezing.
>>
>> add_hwgenerator_randomness() has two post-boot users: in khwrng the
>> kthread will be frozen anyway by call to kthread_freezable_should_stop()
>> in its main loop, while its second user (ath9k-hwrng) is not freezable at
>> all.
>>
>> This change allows a VM with virtio-rng loaded to write s2disk image
>> successfully.
>>
>> Fixes: 59b569480dc8 ("random: Use wait_event_freezable() in add_hwgenerator_randomness()")
>> Signed-off-by: Maciej S. Szmigiero <mail@xxxxxxxxxxxxxxxxxxxxx>
>> ---
>>  drivers/char/random.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/char/random.c b/drivers/char/random.c
>> index de434feb873a..2f87910dd498 100644
>> --- a/drivers/char/random.c
>> +++ b/drivers/char/random.c
>> @@ -2500,8 +2500,8 @@ void add_hwgenerator_randomness(const char *buffer, size_t count,
>>          * We'll be woken up again once below random_write_wakeup_thresh,
>>          * or when the calling thread is about to terminate.
>>          */
>> -       wait_event_freezable(random_write_wait,
>> -                       kthread_should_stop() ||
>> +       wait_event_interruptible(random_write_wait,
>> +                       kthread_should_stop() || freezing(current) ||
>>                         ENTROPY_BITS(&input_pool) <= random_write_wakeup_bits);
> 
> Is it a problem that this wakes up, sees that it should freeze but then
> calls mix_pool_bytes() and credit_entropy_bits()? It looks like
> credit_entropy_bits() will try to wakeup a reader task that is already
> frozen (see the wake_up_interruptible(&random_read_wait) call).

If a reader (user space) task is frozen then it is no longer waiting
on this waitqueue - at least if I understand correctly how the freezer
works for user space tasks, that is by interrupting waits via a fake
signal.

> At one> point I was checking to see if the task was freezing and avoided calling
> those functions so we could get back to kthread_freezable_should_stop()
> in the kthread and actually freeze.

Yes, but I think mixing some extra valid data into random buffer in this
very rare situation shouldn't hurt.
 
>>         mix_pool_bytes(poolp, buffer, count);
>>         credit_entropy_bits(poolp, entropy);
> 
> It's almost like we need a wait_event_freezable_stoppable() API that
> will freeze if freezing() and break out if the kthread is stopped and
> otherwise wait for a wakeup to test the condition. Basically the same
> sort of API that we have for wait_event_freezable() but we pass true for
> the check_kthr_stop flag. Then we can have something like this:
> 
> 	wait_event_freezable_stoppable(
> 		ENTROPY_BITS(&input_pool) <= random_write_wakeup_bits);
> 	if (!kthread_should_stop()) {
> 		mix_pool_bytes(...);
> 		credit_entropy_bits(...);
> 	}
> 

This API could be added but will there be other users for it?
I mean I think it might not be worth adding a new core kernel API for
such a specific, and workaroundable, issue.

But, on the other hand, maybe this would be a cleaner solution, even
though it would be more complicated.

By the way, the same goes for something like set_freezable_only()
function for the second khwrng freezing issue that I have mentioned
in my Nov 10 UTC message.

Maciej