Re: [RFC PATCH 2/2] perf: arm_spe: Enable CONTEXT packets in SPE traces if the profiler runs in CPU mode.

From: German Gomez <german.gomez@arm.com>
To: Will Deacon <will@kernel.org>
Cc: linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org, mark.rutland@arm.com,
	james.clark@arm.com, leo.yan@linaro.org
Subject: Re: [RFC PATCH 2/2] perf: arm_spe: Enable CONTEXT packets in SPE traces if the profiler runs in CPU mode.
Date: Tue, 18 Jan 2022 14:13:42 +0000	[thread overview]
Message-ID: <a5bc60d9-fff0-abf9-c268-dcb75b790bfb@arm.com> (raw)
In-Reply-To: <20220118095258.GA16547@willie-the-truck>

On 18/01/2022 09:52, Will Deacon wrote:
> On Mon, Jan 17, 2022 at 12:44:32PM +0000, German Gomez wrote:
>> Enable CONTEXT packets in SPE traces if the profiler runs in CPU mode.
>> This is no less permissive than the existing behavior for the following
>> reason:
>>
>> If perf_event_paranoid <= 0, then non perfmon_capable() users can open
>> a per-CPU event. With a per-CPU event, unpriviledged users are allowed
>> to profile _all_ processes, even ones owned by root.
>>
>> Without this change, users could see kernel addresses, root processes,
>> etc, but not gather the PIDs of those processes. The PID is probably the
>> least sensitive of all the information.
>>
>> It would be more idiomatic to check the perf_event_paranoid level with
>> perf_allow_cpu(), but this function is not exported so cannot be used
>> from a module. Looking for cpu != -1 is the indirect way of checking
>> the same thing as it could never get to arm_spe_pmu_event_init() without
>> perf_event_paranoid <= 0.
> perf_allow_cpu() is a static inline so there's no need to export it. What's
> missing?

We were still running into build errors:

ERROR: modpost: "security_perf_event_open" [drivers/perf/arm_spe_pmu.ko] undefined!
ERROR: modpost: "sysctl_perf_event_paranoid" [drivers/perf/arm_spe_pmu.ko] undefined

>
> Will