drm/i915/icl: Fix context RPCS programming
authorTvrtko Ursulin <tvrtko.ursulin@intel.com>
Mon, 3 Sep 2018 11:30:07 +0000 (12:30 +0100)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Tue, 4 Sep 2018 13:49:46 +0000 (14:49 +0100)
There are two issues with the current RPCS programming for Icelake:

Expansion of the slice count bitfield has been missed, as well as the
required programming workaround for the subslice count bitfield size
limitation.

1)

Bitfield width for configuring the active slice count has grown so we need
to program the GEN8_R_PWR_CLK_STATE accordingly.

Current code was always requesting eight times the number of slices (due
writing to a bitfield starting three bits higher than it should). These
requests were luckily a) capped by the hardware to the available number of
slices, and b) we haven't yet exported the code to ask for reduced slice
configurations.

Due both of the above there was no impact from this incorrect programming
but we should still fix it.

2)

Due subslice count bitfield being only three bits wide and furthermore
capped to a maximum documented value of four, special programming
workaround is needed to enable more than four subslices.

With this programming driver has to consider the GT configuration as
2x4x8, while the hardware internally translates this to 1x8x8.

A limitation stemming from this is that either a subslice count between
one and four can be selected, or a subslice count equaling the total
number of subslices in all selected slices. In other words, odd subslice
counts greater than four are impossible, as are odd subslice counts
greater than a single slice subslice count.

This also had no impact in the current code base due breakage from 1)
always reqesting more than one slice.

While fixing this we also add some asserts to flag up any future bitfield
overflows.

v2:
 * Use a local in all branches for clarity. (Lionel)

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Bspec: 12247
Reported-by: tony.ye@intel.com
Suggested-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: tony.ye@intel.com
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180903113007.2643-1-tvrtko.ursulin@linux.intel.com
drivers/gpu/drm/i915/i915_reg.h
drivers/gpu/drm/i915/intel_lrc.c

index f2321785cbd61bc2b1a36e0bac7efbca46244a74..09bc8e730ee1260553ecdf30a0ed170f6e3fc968 100644 (file)
@@ -344,6 +344,8 @@ static inline bool i915_mmio_reg_valid(i915_reg_t reg)
 #define   GEN8_RPCS_S_CNT_ENABLE       (1 << 18)
 #define   GEN8_RPCS_S_CNT_SHIFT                15
 #define   GEN8_RPCS_S_CNT_MASK         (0x7 << GEN8_RPCS_S_CNT_SHIFT)
+#define   GEN11_RPCS_S_CNT_SHIFT       12
+#define   GEN11_RPCS_S_CNT_MASK                (0x3f << GEN11_RPCS_S_CNT_SHIFT)
 #define   GEN8_RPCS_SS_CNT_ENABLE      (1 << 11)
 #define   GEN8_RPCS_SS_CNT_SHIFT       8
 #define   GEN8_RPCS_SS_CNT_MASK                (0x7 << GEN8_RPCS_SS_CNT_SHIFT)
index f8ceb9c99dd62f4a35123bbfbdcdce8122c45519..def467c2451b36c85915b7457ccf2bc7cb73e734 100644 (file)
@@ -2480,6 +2480,9 @@ int logical_xcs_ring_init(struct intel_engine_cs *engine)
 static u32
 make_rpcs(struct drm_i915_private *dev_priv)
 {
+       bool subslice_pg = INTEL_INFO(dev_priv)->sseu.has_subslice_pg;
+       u8 slices = hweight8(INTEL_INFO(dev_priv)->sseu.slice_mask);
+       u8 subslices = hweight8(INTEL_INFO(dev_priv)->sseu.subslice_mask[0]);
        u32 rpcs = 0;
 
        /*
@@ -2489,6 +2492,38 @@ make_rpcs(struct drm_i915_private *dev_priv)
        if (INTEL_GEN(dev_priv) < 9)
                return 0;
 
+       /*
+        * Since the SScount bitfield in GEN8_R_PWR_CLK_STATE is only three bits
+        * wide and Icelake has up to eight subslices, specfial programming is
+        * needed in order to correctly enable all subslices.
+        *
+        * According to documentation software must consider the configuration
+        * as 2x4x8 and hardware will translate this to 1x8x8.
+        *
+        * Furthemore, even though SScount is three bits, maximum documented
+        * value for it is four. From this some rules/restrictions follow:
+        *
+        * 1.
+        * If enabled subslice count is greater than four, two whole slices must
+        * be enabled instead.
+        *
+        * 2.
+        * When more than one slice is enabled, hardware ignores the subslice
+        * count altogether.
+        *
+        * From these restrictions it follows that it is not possible to enable
+        * a count of subslices between the SScount maximum of four restriction,
+        * and the maximum available number on a particular SKU. Either all
+        * subslices are enabled, or a count between one and four on the first
+        * slice.
+        */
+       if (IS_GEN11(dev_priv) && slices == 1 && subslices >= 4) {
+               GEM_BUG_ON(subslices & 1);
+
+               subslice_pg = false;
+               slices *= 2;
+       }
+
        /*
         * Starting in Gen9, render power gating can leave
         * slice/subslice/EU in a partially enabled state. We
@@ -2496,24 +2531,50 @@ make_rpcs(struct drm_i915_private *dev_priv)
         * enablement.
        */
        if (INTEL_INFO(dev_priv)->sseu.has_slice_pg) {
-               rpcs |= GEN8_RPCS_S_CNT_ENABLE;
-               rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.slice_mask) <<
-                       GEN8_RPCS_S_CNT_SHIFT;
-               rpcs |= GEN8_RPCS_ENABLE;
+               u32 mask, val = slices;
+
+               if (INTEL_GEN(dev_priv) >= 11) {
+                       mask = GEN11_RPCS_S_CNT_MASK;
+                       val <<= GEN11_RPCS_S_CNT_SHIFT;
+               } else {
+                       mask = GEN8_RPCS_S_CNT_MASK;
+                       val <<= GEN8_RPCS_S_CNT_SHIFT;
+               }
+
+               GEM_BUG_ON(val & ~mask);
+               val &= mask;
+
+               rpcs |= GEN8_RPCS_ENABLE | GEN8_RPCS_S_CNT_ENABLE | val;
        }
 
-       if (INTEL_INFO(dev_priv)->sseu.has_subslice_pg) {
-               rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
-               rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.subslice_mask[0]) <<
-                       GEN8_RPCS_SS_CNT_SHIFT;
-               rpcs |= GEN8_RPCS_ENABLE;
+       if (subslice_pg) {
+               u32 val = subslices;
+
+               val <<= GEN8_RPCS_SS_CNT_SHIFT;
+
+               GEM_BUG_ON(val & ~GEN8_RPCS_SS_CNT_MASK);
+               val &= GEN8_RPCS_SS_CNT_MASK;
+
+               rpcs |= GEN8_RPCS_ENABLE | GEN8_RPCS_SS_CNT_ENABLE | val;
        }
 
        if (INTEL_INFO(dev_priv)->sseu.has_eu_pg) {
-               rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
-                       GEN8_RPCS_EU_MIN_SHIFT;
-               rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
-                       GEN8_RPCS_EU_MAX_SHIFT;
+               u32 val;
+
+               val = INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
+                     GEN8_RPCS_EU_MIN_SHIFT;
+               GEM_BUG_ON(val & ~GEN8_RPCS_EU_MIN_MASK);
+               val &= GEN8_RPCS_EU_MIN_MASK;
+
+               rpcs |= val;
+
+               val = INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
+                     GEN8_RPCS_EU_MAX_SHIFT;
+               GEM_BUG_ON(val & ~GEN8_RPCS_EU_MAX_MASK);
+               val &= GEN8_RPCS_EU_MAX_MASK;
+
+               rpcs |= val;
+
                rpcs |= GEN8_RPCS_ENABLE;
        }