drm/i915: Fix potential overflows in ilk ips calculations
authorVille Syrjälä <ville.syrjala@linux.intel.com>
Wed, 21 Oct 2020 13:14:41 +0000 (16:14 +0300)
committerVille Syrjälä <ville.syrjala@linux.intel.com>
Wed, 21 Oct 2020 20:20:19 +0000 (23:20 +0300)
A bunch of the ips calculations require 64bit math. In particular
'corr' and 'corr2' look like they can overflow on 32bit systems.
Switch to explicit u64 for those.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20201021131443.25616-3-ville.syrjala@linux.intel.com
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
drivers/gpu/drm/i915/gt/intel_rps.c

index e2d7ef5..0db565d 100644 (file)
@@ -1281,8 +1281,9 @@ static unsigned long __ips_gfx_val(struct intel_ips *ips)
 {
        struct intel_rps *rps = container_of(ips, typeof(*rps), ips);
        struct intel_uncore *uncore = rps_to_uncore(rps);
-       unsigned long t, corr, state1, corr2, state2;
+       unsigned int t, state1, state2;
        u32 pxvid, ext_v;
+       u64 corr, corr2;
 
        lockdep_assert_held(&mchdev_lock);
 
@@ -1303,11 +1304,10 @@ static unsigned long __ips_gfx_val(struct intel_ips *ips)
        else /* < 50 */
                corr = t * 301 + 1004;
 
-       corr = corr * 150142 * state1 / 10000 - 78642;
-       corr /= 100000;
-       corr2 = corr * ips->corr;
+       corr = div_u64(corr * 150142 * state1, 10000) - 78642;
+       corr2 = div_u64(corr, 100000) * ips->corr;
 
-       state2 = corr2 * state1 / 10000;
+       state2 = div_u64(corr2 * state1, 10000);
        state2 /= 100; /* convert to mW */
 
        __gen5_ips_update(ips);