[PATCH 00/17] Add OA functionality to Xe

Tue May 21 19:01:21 UTC 2024

On Tue, 2024-05-21 at 11:11 -0700, Dixit, Ashutosh wrote:
> On Tue, 21 May 2024 11:02:15 -0700, Dixit, Ashutosh wrote:
> > 
> > On Tue, 21 May 2024 10:39:17 -0700, Souza, Jose wrote:
> > > 
> > > On Tue, 2024-05-21 at 09:43 -0700, Dixit, Ashutosh wrote:
> > > > On Tue, 21 May 2024 09:29:51 -0700, Souza, Jose wrote:
> > > > > 
> > > > > On Tue, 2024-05-21 at 09:10 -0700, Dixit, Ashutosh wrote:
> > > > > > On Tue, 21 May 2024 07:47:58 -0700, Souza, Jose wrote:
> > > > > > 
> > > > > > Hi Jose,
> > > > > > 
> > > > > > > > Other ask, can you remove this 'Failed to remove unknown OA config'
> > > > > > > > debug message from xe_oa_remove_config_ioctl()?
> > > > > > > 
> > > > > > > Missed 'Insufficient privileges to remove xe OA config', that need to be
> > > > > > > removed too from xe_oa_remove_config_ioctl().
> > > > > > > 
> > > > > > > > Mesa will be using DRM_XE_PERF_OP_REMOVE_CONFIG with config id set to
> > > > > > > > UINT64_MAX to detect if Xe KMD supports OA counters and if application
> > > > > > > > has enough permissions to use it.  So it causes dmesg to be flooded
> > > > > > > > with 'xe 0000:00:02.0: [drm:xe_oa_remove_config_ioctl [xe]] Failed to
> > > > > > > > remove unknown OA config' messages when running tests suites.
> > > > > > > > 
> > > > > > > > Or do you have other suggestion of uAPI that I can use.
> > > > > > 
> > > > > > OK, so you are relying on ENODEV and EACCES errno's from
> > > > > > DRM_XE_PERF_OP_REMOVE_CONFIG to find out (a) if OA is present and (b) if
> > > > > > you need to be root (actually CAP_PERFMON or CAP_SYS_ADMIN).
> > > > > 
> > > > > yep
> > > > > 
> > > > > > 
> > > > > > This logic in Xe should be close to what we have in i915? What does Mesa do
> > > > > > for i915, or what doesn't work in Xe?
> > > > > > 
> > > > > > Here are some pointers:
> > > > > > 
> > > > > > * You can execute DRM_XE_DEVICE_QUERY_OA_UNITS to see if OA is present
> > > > > > 
> > > > > > * Add/remove OA configs and using the global OAG buffer (time based
> > > > > >   sampling or DRM_XE_OA_PROPERTY_SAMPLE_OA set) are priviliged operations
> > > > > >   (need root). Operations which only need OAR/OAC (OA queries, without
> > > > > >   DRM_XE_OA_PROPERTY_SAMPLE_OA) can be executed by non-root.
> > > > > > 
> > > > > > * If "/proc/sys/dev/xe/perf_stream_paranoid" is 0, all operations can be
> > > > > >   executed by non-root users. Otherwise, as I described in the previous
> > > > > >   point.
> > > > > 
> > > > > It is possible that process not started by root has CAP_PERFMON:
> > > > 
> > > > Yes, correct.
> > > > 
> > > > Above I am using "root" loosely as "CAP_PERFMON or CAP_SYS_ADMIN".
> > > > 
> > > > So if root sets 'perf_stream_paranoid' to 0, normal users can do OA
> > > > priviliged operations (as described above).
> > > > 
> > > > If root sets 'perf_stream_paranoid' to 1, only CAP_PERFMON or CAP_SYS_ADMIN
> > > > users can do OA priviliged operations.
> > > 
> > > oh okay so perf_stream_paranoid has a good usage but it do not covers
> > > CAP_PERFMON, see below.
> > > 
> > > > 
> > > > > "Unprivileged processes with enabled CAP_PERFMON capability are treated
> > > > > as privileged processes with respect to perf_events performance
> > > > > monitoring and observability operations,..."
> > > > > 
> > > > > And from what I understood only root can write to perf_stream_paranoid,
> > > > > so I don't see a point in having this file...
> > > > 
> > > > So I don't know how this statement follows?
> > > > 
> > > > root or superuser is the one which gives the permission to non-CAP_PERFMON
> > > > and non-CAP_SYS_ADMIN users to be able to do priviliged OA operations.
> > > 
> > > so if I'm running a process with CAP_PERFMON and read
> > > perf_stream_paranoid and it returns 0
> > 
> > 0 if fine, everyone can use all OA calls. The issue is with 1.
> > 
> > > there is no way for UMD to know
> > > that process is allowed to use Xe KMD OA feature without do a uAPI call
> > > that checks for CAP_PERFMON.
> > > 
> > > That is why I think is better just do a single
> > > DRM_XE_PERF_OP_REMOVE_CONFIG to detect if feature is present and if
> > > process is allowed to use it. But it generates a bunch of debug messages.
> > 
> > A process should be able to find out on its own, without help from Xe
> > driver, what it's capabilities (CAP_PERFMON or CAP_SYS_ADMIN or neither)
> > are:
> > 
> > https://www.google.com/search?q=linux+get+process+capabilities+in+c
> > https://man7.org/linux/man-pages/man3/libcap.3.html
> > 
> > And therefore, along with perf_stream_paranoid, determine which OA calls it
> > can use.
> 
> Also, could you explain why Mesa has to worry about this? As I see it, Mesa
> as library can be linked with processes of different capabilities. And
> depending on perf_stream_paranoid setting on a system, some OA calls might
> or might not be available, the kernel will handle it. So not sure what Mesa
> has to do, except pass the return code from the kernel up to the app.
> 
> So what are we trying to do here in Mesa?

Every Vulkan or OpenGL application will call intel_perf_init_metrics()->oa_metrics_available() during initialization, that will check if application
has OA permissions and tell applications if it support performance counter capabilities of each graphics API.

https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/intel/perf/intel_perf.c?ref_type=heads

https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/intel/vulkan/anv_perf.c

https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/gallium/drivers/iris/iris_monitor.c

> 
> Thanks.
> --
> Ashutosh
> 
> > > > > > So basically I think you just need to check for the perf_stream_paranoid
> > > > > > file above. It will tell you both (a) if OA is present (because we are
> > > > > > going to merge the code which creates this file together with OA) and (b)
> > > > > > if you need to be root for particular operations.
> > > > > > 
> > > > > > Thanks.
> > > > > > --
> > > > > > Ashutosh
> > > > > 
> > > 