upstream commit 944001201ca0196bcdb088129e5866a9f379d08c (plus some defines) [2.6.32 backport] diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 0d05c6f..b87f65d 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -4967,6 +4967,16 @@ i915_gem_load(struct drm_device *dev) list_add(&dev_priv->mm.shrink_list, &shrink_list); spin_unlock(&shrink_list_lock); + /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ + if (IS_GEN3(dev)) { + u32 tmp = I915_READ(MI_ARB_STATE); + if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) { + /* arb state is a masked write, so set bit + bit in mask */ + tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT); + I915_WRITE(MI_ARB_STATE, tmp); + } + } + /* Old X drivers will take 0-2 for front, back, depth buffers */ dev_priv->fence_reg_start = 3; diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h index 4cbc521..4543975 100644 --- a/drivers/gpu/drm/i915/i915_reg.h +++ b/drivers/gpu/drm/i915/i915_reg.h @@ -357,6 +357,70 @@ #define LM_BURST_LENGTH 0x00000700 #define LM_FIFO_WATERMARK 0x0000001F #define MI_ARB_STATE 0x020e4 /* 915+ only */ +#define MI_ARB_MASK_SHIFT 16 /* shift for enable bits */ + +/* Make render/texture TLB fetches lower priorty than associated data + * fetches. This is not turned on by default + */ +#define MI_ARB_RENDER_TLB_LOW_PRIORITY (1 << 15) + +/* Isoch request wait on GTT enable (Display A/B/C streams). + * Make isoch requests stall on the TLB update. May cause + * display underruns (test mode only) + */ +#define MI_ARB_ISOCH_WAIT_GTT (1 << 14) + +/* Block grant count for isoch requests when block count is + * set to a finite value. + */ +#define MI_ARB_BLOCK_GRANT_MASK (3 << 12) +#define MI_ARB_BLOCK_GRANT_8 (0 << 12) /* for 3 display planes */ +#define MI_ARB_BLOCK_GRANT_4 (1 << 12) /* for 2 display planes */ +#define MI_ARB_BLOCK_GRANT_2 (2 << 12) /* for 1 display plane */ +#define MI_ARB_BLOCK_GRANT_0 (3 << 12) /* don't use */ + +/* Enable render writes to complete in C2/C3/C4 power states. + * If this isn't enabled, render writes are prevented in low + * power states. That seems bad to me. + */ +#define MI_ARB_C3_LP_WRITE_ENABLE (1 << 11) + +/* This acknowledges an async flip immediately instead + * of waiting for 2TLB fetches. + */ +#define MI_ARB_ASYNC_FLIP_ACK_IMMEDIATE (1 << 10) + +/* Enables non-sequential data reads through arbiter + */ +#define MI_ARB_DUAL_DATA_PHASE_DISABLE (1 << 9) + +/* Disable FSB snooping of cacheable write cycles from binner/render + * command stream + */ +#define MI_ARB_CACHE_SNOOP_DISABLE (1 << 8) + +/* Arbiter time slice for non-isoch streams */ +#define MI_ARB_TIME_SLICE_MASK (7 << 5) +#define MI_ARB_TIME_SLICE_1 (0 << 5) +#define MI_ARB_TIME_SLICE_2 (1 << 5) +#define MI_ARB_TIME_SLICE_4 (2 << 5) +#define MI_ARB_TIME_SLICE_6 (3 << 5) +#define MI_ARB_TIME_SLICE_8 (4 << 5) +#define MI_ARB_TIME_SLICE_10 (5 << 5) +#define MI_ARB_TIME_SLICE_14 (6 << 5) +#define MI_ARB_TIME_SLICE_16 (7 << 5) + +/* Low priority grace period page size */ +#define MI_ARB_LOW_PRIORITY_GRACE_4KB (0 << 4) /* default */ +#define MI_ARB_LOW_PRIORITY_GRACE_8KB (1 << 4) + +/* Disable display A/B trickle feed */ +#define MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE (1 << 2) + +/* Set display plane priority */ +#define MI_ARB_DISPLAY_PRIORITY_A_B (0 << 0) /* display A > display B */ +#define MI_ARB_DISPLAY_PRIORITY_B_A (1 << 0) /* display B > display A */ + #define CACHE_MODE_0 0x02120 /* 915+ only */ #define CM0_MASK_SHIFT 16 #define CM0_IZ_OPT_DISABLE (1<<6) --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1045,6 +1045,13 @@ extern int i915_wait_ring(struct drm_dev #define IS_I9XX(dev) (INTEL_INFO(dev)->is_i9xx) #define IS_MOBILE(dev) (INTEL_INFO(dev)->is_mobile) +#define IS_GEN3(dev) (IS_I915G(dev) || \ + IS_I915GM(dev) || \ + IS_I945G(dev) || \ + IS_I945GM(dev) || \ + IS_G33(dev) || \ + IS_PINEVIEW(dev)) + #define I915_NEED_GFX_HWS(dev) (INTEL_INFO(dev)->need_gfx_hws) /* With the 945 and later, Y tiling got adjusted so that it was 32 128-byte --