[V6 00/11] perf: New conditional branch filter

[Stephane Eranian]

On Mon, Jun 2, 2014 at 6:04 PM, Anshuman Khandual
<khandual at linux.vnet.ibm.com> wrote:
> On 06/02/2014 06:29 PM, Stephane Eranian wrote:
>> On Wed, May 28, 2014 at 10:04 AM, Anshuman Khandual
>> <khandual at linux.vnet.ibm.com> wrote:
>>> On 05/27/2014 05:39 PM, Stephane Eranian wrote:
>>>> I have been looking at those patches and ran some tests.
>>>> And I found a few issues so far.
>>>>
>>>> I am running:
>>>> $ perf record -j any_ret -e cycles:u test_program
>>>> $ perf report -D
>>>>
>>>> Most entries are okay and match the filter, however some do not make sense:
>>>>
>>>> 3642586996762 0x15d0 [0x108]: PERF_RECORD_SAMPLE(IP, 2): 17921/17921:
>>>> 0x10001170 period: 613678 addr: 0
>>>> .... branch stack: nr:9
>>>> .....  0: 00000000100011cc -> 0000000010000e38
>>>> .....  1: 0000000010001150 -> 00000000100011bc
>>>> .....  2: 0000000010001208 -> 0000000010000e38
>>>> .....  3: 0000000010001160 -> 00000000100011f8
>>>> .....  4: 00000000100011cc -> 0000000010000e38
>>>> .....  5: 0000000010001150 -> 00000000100011bc
>>>> .....  6: 0000000010001208 -> 0000000010000e38
>>>> .....  7: 0000000010001160 -> 00000000100011f8
>>>> .....  8: 0000000000000000 -> 0000000010001160
>>>> ^^^^^^
>>>> Entry 8 does not make sense, unless 0x0 is a valid return branch
>>>> instruction address.
>>>> If an address is invalid, the whole entry needs to be eliminated. It
>>>> is okay to have
>>>> less than the max number of entries supported by HW.
>>>
>>> Hey Stephane,
>>>
>>> Okay. The same behaviour is also reflected in the test results what I have
>>> shared in the patchset. Here is that section.
>>>
>>> (3) perf record -j any_ret -e branch-misses:u ./cprog
>>>
>>> # Overhead  Command  Source Shared Object          Source Symbol  Target Shared Object          Target Symbol
>>> # ........  .......  ....................  .....................  ....................  .....................
>>> #
>>>     15.61%    cprog  [unknown]             [.] 00000000           cprog                 [.] sw_3_1
>>>      6.28%    cprog  cprog                 [.] symbol2            cprog                 [.] hw_1_2
>>>      6.28%    cprog  cprog                 [.] ctr_addr           cprog                 [.] sw_4_1
>>>      6.26%    cprog  cprog                 [.] success_3_1_3      cprog                 [.] sw_3_1
>>>      6.24%    cprog  cprog                 [.] symbol1            cprog                 [.] hw_1_1
>>>      6.24%    cprog  cprog                 [.] sw_4_2             cprog                 [.] callme
>>>      6.21%    cprog  [unknown]             [.] 00000000           cprog                 [.] callme
>>>      6.19%    cprog  cprog                 [.] lr_addr            cprog                 [.] sw_4_2
>>>      3.16%    cprog  cprog                 [.] hw_1_2             cprog                 [.] callme
>>>      3.15%    cprog  cprog                 [.] success_3_1_1      cprog                 [.] sw_3_1
>>>      3.15%    cprog  cprog                 [.] sw_4_1             cprog                 [.] callme
>>>      3.14%    cprog  cprog                 [.] callme             cprog                 [.] main
>>>      3.13%    cprog  cprog                 [.] hw_1_1             cprog                 [.] callme
>>>
>>> So a lot of samples above have 0x0 as the "from" address. This originates from the code
>>> section here inside the function "power_pmu_bhrb_read", where we hit two back to back
>>
>> Could you explain the back-to-back case a bit more here?
>> Back-to-back returns to me means something like:
>>
>> int foo()
>> {
>>   ...
>>    return bar();
>> }
>>
>> int bar()
>> {
>>   return 0;
>> }
>>
>> Not counting the leaf optimization here, bar return to foo which
>> immediately returns: 2 back-2-back returns.
>> Is that the case you're talking about here?
>>
>
> No. Filtering of return branches has been implemented in SW only. So PMU as such does not capture
> return only branches. It captures all the branches what it encounters. During the capture process
> PMU might *record* two back to back "target addresses" (without capturing the from address for the
> first one) for which we are unable to figure out the "from address". This leaves us with one branch
> record where we have the target address not from address and so we make it zero. With the current
> logic all branch records with "from address" as zero get filtered through and become the part of the
> final set. I was not too sure how to deal with these cases.
>
So PPC8 captures all branches, no HW filter. But then in SW you filter
out non return branches.
Given you're description, I have to believe that sometimes the HW does
not even capture the
from address. If so, then in that case, I think it is best to drop the
sample. Because the target
address may be the target of an indirect branch for which there is no
way to find the source.
In other words, the record cannot be exploited.

But why does the HW not capture some from addresses?
I am worried this might create some bias in the samples.

>>> target addresses. So we zero out the from address for the first target address and re-read
>>> the second address over again. So thats how we get zero as the from address. This is how the
>>> HW capture the samples. I was reluctant to drop these samples but I agree that these kind of
>>> samples can be dropped if we need to.
>>>
>> I think we need to make it as simple as possible for tools, i.e.,
>> avoid having to decode the
>> disassembly to figure out what happened. Here address 0 is not exploitable.
>
> Thats right. Dropping the branch record where we have only the target address not the from address
> might just solve this problem.
>