* Oleg Nesterov <o...@redhat.com> [2012-09-03 17:26:09]:
> Afaics the usage of update_debugctlmsr() and TIF_BLOCKSTEP in
> step.c was always very wrong.
>
> 1. update_debugctlmsr() was simply unneeded. The child sleeps
> TASK_TRACED, __switch_to_xtra(next_p => child) should notice
> TIF_BLOCKSTEP and set/clear DEBUGCTLMSR_BTF after resume if
> needed.
>
> 2. It is wrong. The state of DEBUGCTLMSR_BTF bit in CPU register
> should always match the state of current's TIF_BLOCKSTEP bit.
>
> 3. Even get_debugctlmsr() + update_debugctlmsr() itself does not
> look right. Irq can change other bits in MSR_IA32_DEBUGCTLMSR
> register or the caller can be preempted in between.
>
> 4. It is not safe to play with TIF_BLOCKSTEP if task != current.
> DEBUGCTLMSR_BTF and TIF_BLOCKSTEP should always match each
> other if the task is running. The tracee is stopped but it
> can be SIGKILL'ed right before set/clear_tsk_thread_flag().
>
> However, now that uprobes uses user_enable_single_step(current)
> we can't simply remove update_debugctlmsr(). So this patch adds
> the additional "task == current" check and disables irqs to avoid
> the race with interrupts/preemption.
>
> Unfortunately this patch doesn't solve the last problem, we need
> another fix. Probably we should teach ptrace_stop() to set/clear
> single/block stepping after resume.
>
> And afaics there is yet another problem: perf can play with
> MSR_IA32_DEBUGCTLMSR from nmi, this obviously means that even
> __switch_to_xtra() has problems.
>
> Signed-off-by: Oleg Nesterov <o...@redhat.com>
> ---
> arch/x86/kernel/step.c | 14 +++++++++++++-
> 1 files changed, 13 insertions(+), 1 deletions(-)
>
> diff --git a/arch/x86/kernel/step.c b/arch/x86/kernel/step.c
> index 7a51498..f89cdc6 100644
> --- a/arch/x86/kernel/step.c
> +++ b/arch/x86/kernel/step.c
> @@ -161,6 +161,16 @@ static void set_task_blockstep(struct task_struct *task,
> bool on)
> {
> unsigned long debugctl;
>
> + /*
> + * Ensure irq/preemption can't change debugctl in between.
> + * Note also that both TIF_BLOCKSTEP and debugctl should
> + * be changed atomically wrt preemption.
> + * FIXME: this means that set/clear TIF_BLOCKSTEP is simply
> + * wrong if task != current, SIGKILL can wakeup the stopped
> + * tracee and set/clear can play with the running task, this
> + * can confuse the next __switch_to_xtra().
> + */
> + local_irq_disable();
> debugctl = get_debugctlmsr();
> if (on) {
> debugctl |= DEBUGCTLMSR_BTF;
> @@ -169,7 +179,9 @@ static void set_task_blockstep(struct task_struct *task,
> bool on)
> debugctl &= ~DEBUGCTLMSR_BTF;
> clear_tsk_thread_flag(task, TIF_BLOCKSTEP);
> }
> - update_debugctlmsr(debugctl);
> + if (task == current)
> + update_debugctlmsr(debugctl);
> + local_irq_enable();
> }
>
> /*
>
The changes look simple and neat. But I would prefer somebody with
better x86 knowledgde comment on this.
--
Thanks and Regards
Srikar
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
