drm/i915/bdw: Handle forcewake for writes on gen8
authorBen Widawsky <benjamin.widawsky@intel.com>
Sun, 3 Nov 2013 04:07:00 +0000 (21:07 -0700)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 7 Nov 2013 21:16:04 +0000 (22:16 +0100)
GEN8 removes the GT FIFO which we've all come to know and love. Instead
it offers a wider range of optimized registers which always keep a
shadowed copy, and are fed to the GPU when it wakes.

How this is implemented in hardware is still somewhat of a mystery. As
far as I can tell, the basic design is as follows:

If the register is not optimized, you must use the old forcewake
mechanism to bring the GT out of sleep. [1]

If register is in the optimized list the write will signal that the
GT should begin to come out of whatever sleep state it is in.

While the GT is coming out of sleep, the requested write will be stored
in an intermediate shadow register.

Do to the fact that the implementation details are not clear, I see
several risks:
1. Order is not preserved as it is with GT FIFO. If we issue multiple
writes to optimized registers, where order matters, we may need to
serialize it with forcewake.
2. The optimized registers have only 1 shadowed slot, meaning if we
issue multiple writes to the same register, and those values need to
reach the GPU in order, forcewake will be required.

[1] We could implement a SW queue the way the GT FIFO used to work if
desired.

NOTE: Compile tested only until we get real silicon.

v2:
- Use a default case to make future platforms also work.
- Get rid of IS_BROADWELL since that's not yet defined, but we want to
  MMIO as soon as possible.

v3: Apply suggestions from Mika's review:
- s/optimized/shadowed/
- invert the logic of the helper so that it does what it says (the
  code itself was correct, just confusing to read).

v4:
- Squash in lost break.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v1)
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/intel_uncore.c

index f6fae35c568e53036faf150f4791e950c5bfaf3f..ccbbd6d5f10fd44901a3e674d7b12c296c7d4aa4 100644 (file)
@@ -93,7 +93,7 @@ static void __gen6_gt_force_wake_mt_get(struct drm_i915_private *dev_priv)
 {
        u32 forcewake_ack;
 
-       if (IS_HASWELL(dev_priv->dev))
+       if (IS_HASWELL(dev_priv->dev) || IS_GEN8(dev_priv->dev))
                forcewake_ack = FORCEWAKE_ACK_HSW;
        else
                forcewake_ack = FORCEWAKE_MT_ACK;
@@ -459,6 +459,46 @@ hsw_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace)
        spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags); \
 }
 
+static const u32 gen8_shadowed_regs[] = {
+       FORCEWAKE_MT,
+       GEN6_RPNSWREQ,
+       GEN6_RC_VIDEO_FREQ,
+       RING_TAIL(RENDER_RING_BASE),
+       RING_TAIL(GEN6_BSD_RING_BASE),
+       RING_TAIL(VEBOX_RING_BASE),
+       RING_TAIL(BLT_RING_BASE),
+       /* TODO: Other registers are not yet used */
+};
+
+static bool is_gen8_shadowed(struct drm_i915_private *dev_priv, u32 reg)
+{
+       int i;
+       for (i = 0; i < ARRAY_SIZE(gen8_shadowed_regs); i++)
+               if (reg == gen8_shadowed_regs[i])
+                       return true;
+
+       return false;
+}
+
+#define __gen8_write(x) \
+static void \
+gen8_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
+       bool __needs_put = !is_gen8_shadowed(dev_priv, reg); \
+       REG_WRITE_HEADER; \
+       if (__needs_put) { \
+               dev_priv->uncore.funcs.force_wake_get(dev_priv); \
+       } \
+       __raw_i915_write##x(dev_priv, reg, val); \
+       if (__needs_put) { \
+               dev_priv->uncore.funcs.force_wake_put(dev_priv); \
+       } \
+       spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags); \
+}
+
+__gen8_write(8)
+__gen8_write(16)
+__gen8_write(32)
+__gen8_write(64)
 __hsw_write(8)
 __hsw_write(16)
 __hsw_write(32)
@@ -476,6 +516,7 @@ __gen4_write(16)
 __gen4_write(32)
 __gen4_write(64)
 
+#undef __gen8_write
 #undef __hsw_write
 #undef __gen6_write
 #undef __gen5_write
@@ -534,6 +575,16 @@ void intel_uncore_init(struct drm_device *dev)
        }
 
        switch (INTEL_INFO(dev)->gen) {
+       default:
+               dev_priv->uncore.funcs.mmio_writeb  = gen8_write8;
+               dev_priv->uncore.funcs.mmio_writew  = gen8_write16;
+               dev_priv->uncore.funcs.mmio_writel  = gen8_write32;
+               dev_priv->uncore.funcs.mmio_writeq  = gen8_write64;
+               dev_priv->uncore.funcs.mmio_readb  = gen6_read8;
+               dev_priv->uncore.funcs.mmio_readw  = gen6_read16;
+               dev_priv->uncore.funcs.mmio_readl  = gen6_read32;
+               dev_priv->uncore.funcs.mmio_readq  = gen6_read64;
+               break;
        case 7:
        case 6:
                if (IS_HASWELL(dev)) {