Date:   Fri, 24 May 2019 23:32:01 +0200
From:   Daniel Borkmann <daniel@...earbox.net>
To:     Yonghong Song <yhs@...com>,
        "bpf@...r.kernel.org" <bpf@...r.kernel.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>
Cc:     Alexei Starovoitov <ast@...com>, Kernel Team <Kernel-team@...com>,
        Peter Zijlstra <peterz@...radead.org>
Subject: Re: [PATCH bpf-next v2 1/3] bpf: implement bpf_send_signal() helper

On 05/24/2019 01:54 AM, Yonghong Song wrote:
> 
> 
> On 5/23/19 4:08 PM, Daniel Borkmann wrote:
>> On 05/23/2019 11:30 PM, Yonghong Song wrote:
>>> On 5/23/19 2:07 PM, Yonghong Song wrote:
>>>> On 5/23/19 9:28 AM, Daniel Borkmann wrote:
>>>>> On 05/23/2019 05:58 PM, Yonghong Song wrote:
>>>>>> On 5/23/19 8:41 AM, Daniel Borkmann wrote:
>>>>>>> On 05/22/2019 07:39 AM, Yonghong Song wrote:
>>>>>>>> This patch tries to solve the following specific use case.
>>>>>>>>
>>>>>>>> Currently, bpf program can already collect stack traces
>>>>>>>> through kernel function get_perf_callchain()
>>>>>>>> when certain events happens (e.g., cache miss counter or
>>>>>>>> cpu clock counter overflows). But such stack traces are
>>>>>>>> not enough for jitted programs, e.g., hhvm (jited php).
>>>>>>>> To get real stack trace, jit engine internal data structures
>>>>>>>> need to be traversed in order to get the real user functions.
>>>>>>>>
>>>>>>>> bpf program itself may not be the best place to traverse
>>>>>>>> the jit engine as the traversing logic could be complex and
>>>>>>>> it is not a stable interface either.
>>>>>>>>
>>>>>>>> Instead, hhvm implements a signal handler,
>>>>>>>> e.g. for SIGALARM, and a set of program locations which
>>>>>>>> it can dump stack traces. When it receives a signal, it will
>>>>>>>> dump the stack in next such program location.
>>>>>>>>
>>>>>>>> Such a mechanism can be implemented in the following way:
>>>>>>>>       . a perf ring buffer is created between bpf program
>>>>>>>>         and tracing app.
>>>>>>>>       . once a particular event happens, bpf program writes
>>>>>>>>         to the ring buffer and the tracing app gets notified.
>>>>>>>>       . the tracing app sends a signal SIGALARM to the hhvm.
>>>>>>>>
>>>>>>>> But this method could have large delays and causing profiling
>>>>>>>> results skewed.
>>>>>>>>
>>>>>>>> This patch implements bpf_send_signal() helper to send
>>>>>>>> a signal to hhvm in real time, resulting in intended stack traces.
>>>>>>>>
>>>>>>>> Signed-off-by: Yonghong Song <yhs@...com>
>>>>>>>> ---
>>>>>>>>      include/uapi/linux/bpf.h | 17 +++++++++-
>>>>>>>>      kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++
>>>>>>>>      2 files changed, 83 insertions(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>>>>>>>> index 63e0cf66f01a..68d4470523a0 100644
>>>>>>>> --- a/include/uapi/linux/bpf.h
>>>>>>>> +++ b/include/uapi/linux/bpf.h
>>>>>>>> @@ -2672,6 +2672,20 @@ union bpf_attr {
>>>>>>>>       *		0 on success.
>>>>>>>>       *
>>>>>>>>       *		**-ENOENT** if the bpf-local-storage cannot be found.
>>>>>>>> + *
>>>>>>>> + * int bpf_send_signal(u32 sig)
>>>>>>>> + *	Description
>>>>>>>> + *		Send signal *sig* to the current task.
>>>>>>>> + *	Return
>>>>>>>> + *		0 on success or successfully queued.
>>>>>>>> + *
>>>>>>>> + *		**-EBUSY** if work queue under nmi is full.
>>>>>>>> + *
>>>>>>>> + *		**-EINVAL** if *sig* is invalid.
>>>>>>>> + *
>>>>>>>> + *		**-EPERM** if no permission to send the *sig*.
>>>>>>>> + *
>>>>>>>> + *		**-EAGAIN** if bpf program can try again.
>>>>>>>>       */
>>>>>>>>      #define __BPF_FUNC_MAPPER(FN)		\
>>>>>>>>      	FN(unspec),			\
>>>>>>>> @@ -2782,7 +2796,8 @@ union bpf_attr {
>>>>>>>>      	FN(strtol),			\
>>>>>>>>      	FN(strtoul),			\
>>>>>>>>      	FN(sk_storage_get),		\
>>>>>>>> -	FN(sk_storage_delete),
>>>>>>>> +	FN(sk_storage_delete),		\
>>>>>>>> +	FN(send_signal),
>>>>>>>>      
>>>>>>>>      /* integer value in 'imm' field of BPF_CALL instruction selects which helper
>>>>>>>>       * function eBPF program intends to call
>>>>>>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>>>>>>> index f92d6ad5e080..f8cd0db7289f 100644
>>>>>>>> --- a/kernel/trace/bpf_trace.c
>>>>>>>> +++ b/kernel/trace/bpf_trace.c
>>>>>>>> @@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = {
>>>>>>>>      	.arg3_type	= ARG_ANYTHING,
>>>>>>>>      };
>>>>>>>>      
>>>>>>>> +struct send_signal_irq_work {
>>>>>>>> +	struct irq_work irq_work;
>>>>>>>> +	u32 sig;
>>>>>>>> +};
>>>>>>>> +
>>>>>>>> +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work);
>>>>>>>> +
>>>>>>>> +static void do_bpf_send_signal(struct irq_work *entry)
>>>>>>>> +{
>>>>>>>> +	struct send_signal_irq_work *work;
>>>>>>>> +
>>>>>>>> +	work = container_of(entry, struct send_signal_irq_work, irq_work);
>>>>>>>> +	group_send_sig_info(work->sig, SEND_SIG_PRIV, current, PIDTYPE_TGID);
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +BPF_CALL_1(bpf_send_signal, u32, sig)
>>>>>>>> +{
>>>>>>>> +	struct send_signal_irq_work *work = NULL;
>>>>>>>> +
>>>>>>>> +	/* Similar to bpf_probe_write_user, task needs to be
>>>>>>>> +	 * in a sound condition and kernel memory access be
>>>>>>>> +	 * permitted in order to send signal to the current
>>>>>>>> +	 * task.
>>>>>>>> +	 */
>>>>>>>> +	if (unlikely(current->flags & (PF_KTHREAD | PF_EXITING)))
>>>>>>>> +		return -EPERM;
>>>>>>>> +	if (unlikely(uaccess_kernel()))
>>>>>>>> +		return -EPERM;
>>>>>>>> +	if (unlikely(!nmi_uaccess_okay()))
>>>>>>>> +		return -EPERM;
>>>>>>>> +
>>>>>>>> +	if (in_nmi()) {
>>>>>>>
>>>>>>> Hm, bit confused, can't this only be done out of process context in
>>>>>>> general since only there current points to e.g. hhvm? I'm probably
>>>>>>> missing something. Could you elaborate?
>>>>>>
>>>>>> That is true. If in nmi, it is out of process context and in nmi
>>>>>> context, we use an irq_work here since group_send_sig_info() has
>>>>>> spinlock inside. The bpf program (e.g., a perf_event program) needs to
>>>>>> check it is with right current (e.g., by pid) before calling
>>>>>> this helper.
>>>>>>
>>>>>> Does this address your question?
>>>>
>>>> Thanks, Daniel. The below are really good questions which I did not
>>>> really think through with irq_work.
>>>>
>>>>> Hm, but how is it guaranteed that 'current' inside the callback is still
>>>>> the very same you intend to send the signal to?
>>>>
>>>> I went through irq_work infrastructure. It looks that "current" may
>>>> change. irq_work is registered as an interrupt on x86.
>>>> After nmi, some lower priority
>>>> interrupts get chances to execute including irq_work. But there are some
>>>> other interrupts like timer_interrupt and reschedule_interrupt may
>>>> change "current". But since we are still in interrupt context, task
>>>> should not be destroyed, so the task structure pointer is still valid.
>>>>
>>>> I will pass "current" task struct pointer to irq_work as well. This
>>>> is similar to what we did in stackmap.c:
>>>>      work->sem = &current->mm->mmap_sem;
>>>>      irq_work_queue(&work->irq_work);
>>>> At irq_work_run() time, the previous "current" in nmi should still be
>>>> valid.
>>>>
>>>>> What happens if you're in softirq and send SIGKILL to yourself? Is this
>>>>> ignored/handled gracefully in such case?
>>>>
>>>> It is not ignored. It handled gracefully in this case. I tried my
>>>> example to send SIGKILL. The call stack looks below.
>>>>
>>>> [   24.211943]  bpf_send_signal+0x9/0xb0
>>>> [   24.212427]  bpf_prog_fec6e7cc664d5b91_bpf_send_signal_test+0x228/0x1000
>>>> [   24.213249]  ? bpf_overflow_handler+0xb7/0x180
>>>> [   24.213853]  ? __perf_event_overflow+0x51/0xe0
>>>> [   24.214385]  ? perf_swevent_hrtimer+0x10a/0x160
>>>> [   24.214965]  ? __update_load_avg_cfs_rq+0x1a9/0x1c0
>>>> [   24.215609]  ? task_tick_fair+0x50/0x690
>>>> [   24.216104]  ? run_timer_softirq+0x208/0x490
>>>> [   24.216637]  ? timerqueue_del+0x1e/0x40
>>>> [   24.217111]  ? task_clock_event_del+0x10/0x10
>>>> [   24.217658]  ? __hrtimer_run_queues+0x10d/0x2c0
>>>> [   24.218217]  ? hrtimer_interrupt+0x122/0x270
>>>> [   24.218765]  ? rcu_irq_enter+0x31/0x110
>>>> [   24.219223]  ? smp_apic_timer_interrupt+0x67/0x160
>>>> [   24.219842]  ? apic_timer_interrupt+0xf/0x20
>>>> [   24.220383]  </IRQ>
>>>> [   24.220655]  ? event_sched_out.isra.108+0x150/0x150
>>>> [   24.221271]  ? smp_call_function_single+0xdc/0x100
>>>> [   24.221898]  ? perf_event_sysfs_show+0x20/0x20
>>>> [   24.222469]  ? cpu_function_call+0x42/0x60
>>>> [   24.222982]  ? cpu_clock_event_read+0x10/0x10
>>>> [   24.223525]  ? event_function_call+0xe6/0xf0
>>>> [   24.224053]  ? event_sched_out.isra.108+0x150/0x150
>>>> [   24.224657]  ? perf_remove_from_context+0x20/0x70
>>>> [   24.225262]  ? perf_event_release_kernel+0x106/0x2e0
>>>> [   24.225896]  ? perf_release+0xc/0x10
>>>> [   24.226347]  ? __fput+0xc9/0x230
>>>> [   24.226767]  ? task_work_run+0x83/0xb0
>>>> [   24.227243]  ? do_exit+0x300/0xc50
>>>> [   24.227674]  ? syscall_trace_enter+0x1c9/0x2d0
>>>> [   24.228223]  ? do_group_exit+0x39/0xb0
>>>> [   24.228695]  ? __x64_sys_exit_group+0x14/0x20
>>>> [   24.229270]  ? do_syscall_64+0x49/0x130
>>>> [   24.229762]  ? entry_SYSCALL_64_after_hwframe+0x44/0xa9
>>>>
>>>> I see the task is killed and other process is not impacted and
>>>> no kernel crash/warning.
>>
>> Hm, but I rather meant when you have the case that we're in_serving_softirq()
>> e.g. when processing packets on rx and you send a signal to yourself. Shouldn't
>> we bail out from the helper in such situation as well?
> 
> Just want to clarify. Are you concerned with safety or correctness?
> 
> For safety, if we do send signal here, we may wreck the system?
> 
> For correctness, you mean the information we got to send a signal
> to process is not quite right if in_serving_softirq()? F.e,
> the performance counter overflow may be caused by softirq rather
> the process itself? So in this case, we should only send signal
> to process when in process context, and in nmi (not serving softirq)?
> 
> If for correctness, do you think we should add a "flags" parameter
> to the bpf_send_signal() helper such that:
>     . default not checking is_serving_softirq()
>     . bit0: if set, bail out if is_serving_softirq()
>     . other bits: reserved

Scratch my thought, we do bail out in case of PF_KTHREAD, so should be
okay. Was thinking in terms of both, not wrecking the system / messing
with kthreads and with regards to correctness.

Powered by blists - more mailing lists