On Tue 2024-06-25 17:53:14, John Ogness wrote:
> On 2024-06-26, Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
> > On 2024/06/25 23:17, John Ogness wrote:
> >> On 2024-06-25, Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
> >>> syzbot is reporting circular locking dependency inside __bpf_prog_run(),
> >>> for fault injection calls printk() despite rq lock is already held.
> >>>
> >>> Guard __bpf_prog_run() using printk_deferred_{enter,exit}() (and
> >>> preempt_{disable,enable}() if CONFIG_PREEMPT_RT=n) in order to defer any
> >>> printk() messages.
> >>
> >> Why is the reason for disabling preemption?
> >
> > Because since kernel/printk/printk_safe.c uses a percpu counter for deferring
> > printk(), printk_safe_enter() and printk_safe_exit() have to be called from
> > the same CPU. preempt_disable() before printk_safe_enter() and preempt_enable()
> > after printk_safe_exit() guarantees that printk_safe_enter() and
> > printk_safe_exit() are called from the same CPU.
>
> Yes, but we already have cant_migrate(). Are you suggesting there are
> configurations where cant_migrate() is true but the context can be
> migrated anyway?
IMHO, we want to enter printk_safe only with preemption disabled.
Otherwise, printk() would stay deferred on the given CPU for any
task scheduled in this section.
Best Regards,
Petr
