Re: [PATCH] perf/core: Fix the same task check in perf_event_set_output

[Adrian Hunter]

On 22/03/23 15:42, Peter Zijlstra wrote:
> On Wed, Mar 22, 2023 at 12:59:28PM +0200, Adrian Hunter wrote:
>> On 11/07/22 21:07, kan.liang@xxxxxxxxxxxxxxx wrote:
>>> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>>>
>>> With the --per-thread option, perf record errors out when sampling with
>>> a hardware event and a software event as below.
>>>
>>>  $ perf record -e cycles,dummy --per-thread ls
>>>  failed to mmap with 22 (Invalid argument)
>>>
>>> The same task is sampled with the two events. The IOC_OUTPUT is invoked
>>> to share the mmap memory of the task between the events. In the
>>> perf_event_set_output(), the event->ctx is used to check whether the
>>> two events are attached to the same task. However, a hardware event and
>>> a software event are from different task context. The check always
>>> fails.
>>>
>>> The task struct is stored in the event->hw.target for each per-thread
>>> event. It can be used to determine whether two events are attached to
>>> the same task.
>>>
>>> The patch can also fix another issue reported months ago.
>>> https://lore.kernel.org/all/92645262-D319-4068-9C44-2409EF44888E@xxxxxxxxx/
>>> The event->ctx is not ready when the perf_event_set_output() is invoked
>>> in the perf_event_open(), while the event->hw.target has been assigned
>>> at the moment.
>>>
>>> The problem should be a long time issue since commit c3f00c70276d
>>> ("perf: Separate find_get_context() from event initialization"). The
>>> event->hw.target doesn't exist at that time. Here, the patch which
>>> introduces the event->hw.target is used by the Fixes tag.
>>>
>>> The problem should still exists between the broken patch and the
>>> event->hw.target patch. This patch does not intend to fix that case.
>>>
>>> Fixes: 50f16a8bf9d7 ("perf: Remove type specific target pointers")
>>> Reviewed-by: Zhengjun Xing <zhengjun.xing@xxxxxxxxxxxxxxx>
>>> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>>
>> Did this slip through the cracks, or is there more complexity
>> to this case than just sharing the rb?
> 
> Both; I very much missed it, but looking at it now, I'm not at all sure
> it is correct prior to the whole context rewrite we did recently.
> 
> So after the rewrite every cpu/task only has a single
> perf_event_context, and your change below is actually an equivalence.
> 
> But prior to that a task could have multiple contexts. Now they got
> co-scheduled most of the times and it will probably work, but I'm not
> entirely sure.
> 
> So how about we change the Fixes tag to something like:
> 
> Fixes: c3f00c70276d ("perf: Separate find_get_context() from event initialization") # >= v6.2
> 
> And anybody that wants to back-port this further gets to either do the
> full audit and/or keep the pieces.
> 
> Hmm?

Seems reasonable to me.  Kan?

> 
>>> ---
>>>  kernel/events/core.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>>> index b4d62210c3e5..22df79d3f19d 100644
>>> --- a/kernel/events/core.c
>>> +++ b/kernel/events/core.c
>>> @@ -12080,7 +12080,7 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
>>>  	/*
>>>  	 * If its not a per-cpu rb, it must be the same task.
>>>  	 */
>>> -	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
>>> +	if (output_event->cpu == -1 && output_event->hw.target != event->hw.target)
>>>  		goto out;
>>>  
>>>  	/*
>>