On Fri, Nov 06, 2020 at 11:52:05AM +0300, Dan Carpenter wrote:
> The exit_pi_state_list() function calls put_pi_state() with IRQs
> disabled and is not expecting that IRQs will be enabled inside the
> function. Use the _irqsave() so that IRQs are restored to the original
> state instead of enabled unconditionally.
>
> Fixes: 153fbd1226fb ("futex: Fix more put_pi_state() vs. exit_pi_state_list() races")
> Signed-off-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> ---
> This is from static analysis and not tested. I am not very familiar
> with futex code.
It it exceedingly rare if at all possible to trigger this, but yes, your
patch is correct.
Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
>
> kernel/futex.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/futex.c b/kernel/futex.c
> index f8614ef4ff31..ca84745713bc 100644
> --- a/kernel/futex.c
> +++ b/kernel/futex.c
> @@ -788,8 +788,9 @@ static void put_pi_state(struct futex_pi_state *pi_state)
> */
> if (pi_state->owner) {
> struct task_struct *owner;
> + unsigned long flags;
>
> - raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
> + raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
> owner = pi_state->owner;
> if (owner) {
> raw_spin_lock(&owner->pi_lock);
> @@ -797,7 +798,7 @@ static void put_pi_state(struct futex_pi_state *pi_state)
> raw_spin_unlock(&owner->pi_lock);
> }
> rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
> - raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
> + raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
> }
>
> if (current->pi_state_cache) {
> --
> 2.28.0
>
