Re: [PATCH 0/7] perf, x86: Haswell LBR call stack support

["Yan, Zheng" <zheng.z.yan@intel.com>]

On 06/27/2013 12:48 AM, Stephane Eranian wrote:
> On Wed, Jun 26, 2013 at 1:54 PM, Peter Zijlstra <peterz@infradead.org> wrote:
>> On Tue, Jun 25, 2013 at 04:47:12PM +0800, Yan, Zheng wrote:
>>> From: "Yan, Zheng" <zheng.z.yan@intel.com>
>>>
>>> Haswell has a new feature that utilizes the existing Last Branch Record
>>> facility to record call chains. When the feature is enabled, function
>>> call will be collected as normal, but as return instructions are executed
>>> the last captured branch record is popped from the on-chip LBR registers.
>>> The LBR call stack facility can help perf to get call chains of progam
>>> without frame pointer. When perf tool requests PERF_SAMPLE_CALLCHAIN +
>>> PERF_SAMPLE_BRANCH_USER, this feature is dynamically enabled by default.
>>> This feature can be disabled/enabled through an attribute file in the cpu
>>> pmu sysfs directory.
>>>
>>> The LBR call stack has following known limitations
>>>  1. Zero length calls are not filtered out by hardware
>>>  2. Exception handing such as setjmp/longjmp will have calls/returns not
>>>     match
>>>  3. Pushing different return address onto the stack will have calls/returns
>>>     not match
>>>
>>
>> You fail to mention what happens when the callstack is deeper than the
>> LBR is big -- a rather common issue I'd think.
>>
> LBR is statistical callstack. By nature, it cannot capture the entire chain.
> 
>> From what I gather if you push when full, the TOS rotates and eats the
>> tail allowing you to add another entry to the head.
>>
>> If you pop when empty; nothing happens.
>>
> Not sure they know "empty" from "non empty", they just move the LBR_TOS
> by one entry on returns.

When pop, it decreases LBR_TOS by one and clear the popped LBR_FROM/LBR_TO MSRs.
If pop when empty, you will get an empty callchains.

Regards
Yan, Zheng

> 
>> So on pretty much every program you'd be lucky to get the top of the
>> callstack but can end up with nearly nothing.
>>
> You will get the calls closest to the interrupt.
> 
>> Given that, and the other limitations I don't think its a fair
>> replacement for user callchains.
> 
> Well, the one advantage I see is that it works on stripped/optimized
> binaries without fp or dwarf info. Compared to dwarf and the stack
> snapshot, it does incur less overhead most likely. But yes, it comes
> with limitations.
>