Date:   Fri, 29 Oct 2021 18:09:29 +0000
From:   Song Liu <songliubraving@...com>
To:     Like Xu <like.xu.linux@...il.com>
CC:     Peter Zijlstra <peterz@...radead.org>, bpf <bpf@...r.kernel.org>,
        Networking <netdev@...r.kernel.org>,
        Kernel Team <Kernel-team@...com>,
        "Andrii Nakryiko" <andrii@...nel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        "like.xu@...ux.intel.com" <like.xu@...ux.intel.com>,
        Andi Kleen <andi@...stfloor.org>,
        "Liang, Kan" <kan.liang@...el.com>
Subject: Re: bpf_get_branch_snapshot on qemu-kvm



> On Oct 27, 2021, at 8:09 PM, Like Xu <like.xu.linux@...il.com> wrote:
> 
> On 26/10/2021 3:09 pm, Song Liu wrote:
>>> On Oct 9, 2021, at 2:03 AM, Like Xu <like.xu.linux@...il.com> wrote:
>>> 
>>> On 9/10/2021 1:08 am, Song Liu wrote:
>>>>> On Oct 7, 2021, at 11:36 PM, Like Xu <like.xu.linux@...il.com> wrote:
>>>>> 
>>>>> On 8/10/2021 1:46 pm, Song Liu wrote:
>>>>>>> On Oct 7, 2021, at 8:34 PM, Like Xu <like.xu.linux@...il.com> wrote:
>>>>>>> 
>>>>>>> On 30/9/2021 4:05 am, Song Liu wrote:
>>>>>>>> Hi Kan,
>>>>>>>>> On Sep 29, 2021, at 9:35 AM, Liang, Kan <kan.liang@...el.com> wrote:
>>>>>>>>> 
>>>>>>>>>>>> - get confirmation that clearing GLOBAL_CTRL is suffient to supress
>>>>>>>>>>>>  PEBS, in which case we can simply remove the PEBS_ENABLE clear.
>>>>>>>>>>> 
>>>>>>>>>>> How should we confirm this? Can we run some tests for this? Or do we
>>>>>>>>>>> need hardware experts' input for this?
>>>>>>>>>> 
>>>>>>>>>> I'll put it on the list to ask the hardware people when I talk to them next. But
>>>>>>>>>> maybe Kan or Andi know without asking.
>>>>>>>>> 
>>>>>>>>> If the GLOBAL_CTRL is explicitly disabled, the counters do not count anymore.
>>>>>>>>> It doesn't matter if PEBS is enabled or not.
>>>>>>>>> 
>>>>>>>>> See 6c1c07b33eb0 ("perf/x86/intel: Avoid unnecessary PEBS_ENABLE MSR
>>>>>>>>> access in PMI "). We optimized the PMU handler base on it.
>>>>>>>> Thanks for these information!
>>>>>>>> IIUC, all we need is the following on top of bpf-next/master:
>>>>>>>> diff --git i/arch/x86/events/intel/core.c w/arch/x86/events/intel/core.c
>>>>>>>> index 1248fc1937f82..d0d357e7d6f21 100644
>>>>>>>> --- i/arch/x86/events/intel/core.c
>>>>>>>> +++ w/arch/x86/events/intel/core.c
>>>>>>>> @@ -2209,7 +2209,6 @@ intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int
>>>>>>>>         /* must not have branches... */
>>>>>>>>         local_irq_save(flags);
>>>>>>>>         __intel_pmu_disable_all(false); /* we don't care about BTS */
>>>>>>> 
>>>>>>> If the value passed in is true, does it affect your use case?
>>>>>>> 
>>>>>>>> -       __intel_pmu_pebs_disable_all();
>>>>>>> 
>>>>>>> In that case, we can reuse "static __always_inline void intel_pmu_disable_all(void)"
>>>>>>> regardless of whether PEBS is supported or enabled inside the guest and the host ?
>>>>>>> 
>>>>>>>>         __intel_pmu_lbr_disable();
>>>>>>> 
>>>>>>> How about using intel_pmu_lbr_disable_all() to cover Arch LBR?
>>>>>> We are using LBR without PMI, so there isn't any hardware mechanism to
>>>>>> stop the LBR, we have to stop it in software. There is always a delay
>>>>>> between the event triggers and the LBR is stopped. In this window,
>>>>> 
>>>>> Do you use counters for snapshot branch stack?
>>>>> 
>>>>> Can the assumption of "without PMI" be broken sine Intel does have
>>>>> the hardware mechanism like "freeze LBR on counter overflow
>>>>> (aka, DEBUGCTLMSR_FREEZE_LBRS_ON_PMI)" ?
>>>> We are capturing LBR on software events. For example, when a complex syscall,
>>>> such as sys_bpf() and sys_perf_event_open(), returns -EINVAL, it is not obvious
>>>> what wen wrong. The branch stack at the return (on a kretprobe or fexit) could
>>>> give us additional information.
>>>>> 
>>>>>> the LBR is still running and old entries are being replaced by new entries.
>>>>>> We actually need the old entries before the triggering event, so the key
>>>>>> design goal here is to minimize the number of branch instructions between
>>>>>> the event triggers and the LBR is stopped.
>>>>> 
>>>>> Yes, it makes sense.
>>>>> 
>>>>>> Here, both __intel_pmu_disable_all(false) and __intel_pmu_lbr_disable()
>>>>>> are used to optimize for this goal: the fewer branch instructions the
>>>>>> better.
>>>>> 
>>>>> Is it possible that we have another LBR in-kernel user in addition to the current
>>>>> BPF-LBR snapshot user, such as another BPF-LBR snapshot user or a LBR perf user ?
>>>> I think it is OK to have another user. We just need to capture the LBR entries.
>>>> In fact, we simply enable LBR by opening a perf_event on each CPU. So from the
>>>> kernel's point of view, the LBR is owned used by "another user".
>>>>> 
>>>>> In the intel_pmu_snapshot_[arch]_branch_stack(), what if there is a PMI or NMI handler
>>>>> to be called before __intel_pmu_lbr_disable(), which means more branch instructions
>>>>> (assuming we don't use the FREEZE_LBRS_ON_xxx capability)?
>>>> If we are unlucky and hit an NMI, we may get garbage data. The user will run the
>>>> test again.
>>>>> How about try to disable LBR at the earliest possible time, before __intel_pmu_disable_all(false) ?
>>>> I am not sure which solution is the best here. On bare metal, current version works
>>>> fine (available in bpf-next tree).
>>>>> 
>>>>>> After removing __intel_pmu_pebs_disable_all() from
>>>>>> intel_pmu_snapshot_branch_stack(), we found quite a few LBR entries in
>>>>>> extable related code. With these entries, snapshot branch stack is not
>>>>> 
>>>>> Are you saying that you still need to call
>>>>> __intel_pmu_pebs_disable_all() to maintain precision ?
>>>> I think we don't need pebs_disable_all. In the VM, pebs_disable_all will trigger
>>>> "unchecked MSR access error" warning. After removing it, the warning message is
>>>> gone. However, after we remove pebs_disable_all, we still see too many LBR entries
>>>> are flushed before LBR is stopped. Most of these new entries are in extable code.
>>>> I guess this is because the VM access these MSR differently.
>>> 
>>> Hi Song,
>>> 
>>> Thanks for your detailed input. I saw your workaround "if (is_hypervisor())" on the tree.
>>> 
>>> Even when the guest supports PEBS, this use case fails and the root cause is still
>>> playing hide-and-seek with me. Just check with you to see if you get similar results
>>> when the guest LBR behavior makes the test case fail like this:
>>> 
>>> serial_test_get_branch_snapshot:FAIL:find_looptest_in_lbr unexpected find_looptest_in_lbr: actual 0 <= expected 6
>>> serial_test_get_branch_snapshot:FAIL:check_wasted_entries unexpected check_wasted_entries: actual 32 >= expected 10
>>> #52 get_branch_snapshot:FAIL
>>> 
>>> Also, do you know or rough guess about how extable code relates to the test case ?
>> Sorry for the delayed response. I finally got some time to look into
>> this again. After disabling most debug configs, I managed to get it
>> work in the VM with a simple change as
> 
> Yes, most of the contaminated lbr records come from these guest symbols:
> 
> intel_pmu_snapshot_branch_stack
> native_write_msr
> trace_hardirqs_off
> lockdep_hardirqs_off
> __lock_acquire
> mark_lock
> migrate_disable
> rcu_is_watching
> bpf_get_branch_snapshot
> __bpf_prog_enter
> 
> I think we're fine with the current guest LBR emulation, right?

Yes, current result looks good to me. But it does rely on the .config.  

> 
>> diff --git i/arch/x86/events/intel/core.c w/arch/x86/events/intel/core.c
>> index 1248fc1937f82..3887b579297d7 100644
>> --- i/arch/x86/events/intel/core.c
>> +++ w/arch/x86/events/intel/core.c
>> @@ -2209,7 +2209,6 @@ intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int
>>         /* must not have branches... */
>>         local_irq_save(flags);
>>         __intel_pmu_disable_all(false); /* we don't care about BTS */
>> -       __intel_pmu_pebs_disable_all();
>>         __intel_pmu_lbr_disable();
>>         /*            ... until here */
>>         return __intel_pmu_snapshot_branch_stack(entries, cnt, flags);
> 
> LGTM.
> 
>> (of course we also need to remove the is_hypervisor() check.).
>> But I am not sure whether this is the best fix.
>> I pushed all the change and debug code I used to
>> https://git.kernel.org/pub/scm/linux/kernel/git/song/linux.git/log/?h=get_branch_snapshot_in_vm
>> Could you please take a look at it and share your feedback on this?
> 
> How do we inform the user of bpf_get_branch_snapshot in a reasonable way
> that the lbr data will be inaccurate when using the debug kernel?
> 
> Is it better to check for mutual exclusion in code or to use the user documentation
> to specify this part of the restriction? It affects the user experience.

In uapi/linux/bpf.h, we have 

 * long bpf_get_branch_snapshot(void *entries, u32 size, u64 flags)
 *      Description
 *              Get branch trace from hardware engines like Intel LBR. The
 *              hardware engine is stopped shortly after the helper is
 *              called. Therefore, the user need to filter branch entries
 *              based on the actual use case. To capture branch trace
 *              before the trigger point of the BPF program, the helper
 *              should be called at the beginning of the BPF program.

I guess this is enough. 

> 
>> Specifically, can we fix intel_pmu_snapshot_branch_stack in vm with the
>> change above?
> 
> At least it's a valid fix and we can start from this change.

Thanks! I will send the fix. 

Song

Powered by blists - more mailing lists