Please try drm-tip as the interface intel-gpu-overlay uses has finally been upstreamed, hopefully it's just that.

I suppose drm-tip is this: https://cgit.freedesktop.org/drm-tip

I have tried to use that kernel but intel-gpu-overlay is still showing the super high values. Both older versions of intel-gpu-overlay and turbostat show sane values.

Ok, try changing intel-gpu-tools:

diff --git a/overlay/power.c b/overlay/power.c
index 9ac90fde..e02edec8 100644
--- a/overlay/power.c
+++ b/overlay/power.c
@@ -116,7 +116,8 @@ int power_init(struct power *power)
 
        memset(power, 0, sizeof(*power));
 
-       power->fd = igt_perf_open(rapl_type_id(), rapl_gpu_power());
+       power->fd = -1;
+       //power->fd = igt_perf_open(rapl_type_id(), rapl_gpu_power());
        if (power->fd >= 0) {
                power->rapl_scale = rapl_gpu_power_scale();

With this change it is showing the values as expected, 350 mW instead of 3500000000000 mW.

Add a couple of printfs,

diff --git a/overlay/power.c b/overlay/power.c
index 9ac90fde..e6ac728a 100644
--- a/overlay/power.c
+++ b/overlay/power.c
@@ -117,8 +117,12 @@ int power_init(struct power *power)
        memset(power, 0, sizeof(*power));
 
        power->fd = igt_perf_open(rapl_type_id(), rapl_gpu_power());
+       fprintf(stderr, "rapl_type_id()=%"PRIx64", rapl_gpu_power()=%"PRIx64"\n",
+               rapl_type_id(), rapl_gpu_power());
        if (power->fd >= 0) {
                power->rapl_scale = rapl_gpu_power_scale();
+               fprintf(stderr, "rapl_gpu_power_scale()=%f\n",
+                       rapl_gpu_power_scale());
 
                if (power->rapl_scale != NAN) {
                        power->rapl_scale *= 1e3; /* from nano to micro */

and run with -f (so that it doesn't detach and we can see the output).

The lines

        fprintf(stderr, "rapl_type_id()=%"PRIx64", rapl_gpu_power()=%"PRIx64"\n",
                rapl_type_id(), rapl_gpu_power());

made the overlay fail to build. I have changed that to (using lx is not perfect, but PRIx64 made it fail to build):

	fprintf(stderr, "rapl_type_id()=%lx\n",rapl_type_id());
	fprintf(stderr, "rapl_gpu_power()=%lx\n",rapl_gpu_power());

which should result in the intended output. It was:

rapl_type_id()=d
rapl_gpu_power()=4
rapl_gpu_power_scale()=2,000000

And for completeness:  cat /sys/devices/power/events/energy-gpu.scale

$ sudo cat /sys/devices/power/events/energy-gpu.scale
2.3283064365386962890625e-10

Created attachment 138378 [details] [review]
Use setlocale("C") around strtod

Please try the attached patch.

With this patch, the power consumption is shown correctly.

Thank you for the bug report and the invaluable testing.

commit 3fa0b027304ec28cd24b314349d3731b55dfcc0a (HEAD, upstream/master)
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Mar 27 15:20:51 2018 +0100

    overlay: Call setlocale around strtod
    
    strtod() is locale-dependent. The decimal conversion depends on the radix
    character ('.' for some of us like myself) varies by locale. As the
    kernel reports its values using the "C" locale, we need to switch to
    that when parsing; and switch back before reporting to the user.
    
    Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105712
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
    Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
    Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

Closing this bug, as it was resolved/fixed.