Disable automatic load CCS load balancing

The hardware should not dynamically balance the load between CCS engines. Wa_14019159160 recommends disabling it across all platforms.

Fixes d2eae8e98d59 ("drm/i915/dg2: Drop force_probe requirement")

Signed-off-by Andi Shyti <[email protected]>

Cc Chris Wilson <[email protected]>

Cc Joonas Lahtinen <[email protected]>

Cc Matt Roper <[email protected]>

Cc <[email protected]> # v6.2+

---
 drivers/gpu/drm/i915/gt/intel_gt_regs.h     | 1 +
 drivers/gpu/drm/i915/gt/intel_workarounds.c | 5 +++++
 2 files changed, 6 insertions(+)

drivers/gpu/drm/i915/gt/intel_gt_regs.h +1 -0

index 50962cfd1353..cf709f6c05ae 100644
@@ -1478,6 +1478,7 @@
 
 #define GEN12_RCU_MODE				_MMIO(0x14800)
 #define   GEN12_RCU_MODE_CCS_ENABLE		REG_BIT(0)
+#define   XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE	REG_BIT(1)
 
 #define CHV_FUSE_GT				_MMIO(VLV_GUNIT_BASE + 0x2168)
 #define   CHV_FGT_DISABLE_SS0			(1 << 10)

drivers/gpu/drm/i915/gt/intel_workarounds.c +0 -0

index d67d44611c28..a2e78cf0b5f5 100644

drivers/gpu/drm/i915/gt/intel_workarounds.c +2 -0

+		 */
+		wa_masked_en(wal, GEN12_RCU_MODE, XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE);
 	}
 
 	if (IS_DG2_G11(i915)) {

Signature · 1 line

2.43.0

Quoted 32 lines · Andi Shyti

The hardware should not dynamically balance the load between CCS engines. Wa_14019159160 recommends disabling it across all platforms.

Fixes d2eae8e98d59 ("drm/i915/dg2: Drop force_probe requirement")

Signed-off-by Andi Shyti <[email protected]>

Cc Chris Wilson <[email protected]>

Cc Joonas Lahtinen <[email protected]>

Cc Matt Roper <[email protected]>

Cc <[email protected]> # v6.2+

---
 drivers/gpu/drm/i915/gt/intel_gt_regs.h     | 1 +
 drivers/gpu/drm/i915/gt/intel_workarounds.c | 5 +++++
 2 files changed, 6 insertions(+)

drivers/gpu/drm/i915/gt/intel_gt_regs.h +1 -0

index 50962cfd1353..cf709f6c05ae 100644
@@ -1478,6 +1478,7 @@
 
 #define GEN12_RCU_MODE				_MMIO(0x14800)
 #define   GEN12_RCU_MODE_CCS_ENABLE		REG_BIT(0)
+#define   XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE	REG_BIT(1)
 
 #define CHV_FUSE_GT				_MMIO(VLV_GUNIT_BASE + 0x2168)
 #define   CHV_FGT_DISABLE_SS0			(1 << 10)

drivers/gpu/drm/i915/gt/intel_workarounds.c +0 -0

index d67d44611c28..a2e78cf0b5f5 100644

drivers/gpu/drm/i915/gt/intel_workarounds.c general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_li · line 2950

Quoted 6 lines

drivers/gpu/drm/i915/gt/intel_workarounds.c +3 -0

@@ -2945,6 +2945,11 @@ general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_li
 
 		/* Wa_18028616096 */
 		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW, UGM_FRAGMENT_THRESHOLD_TO_3);
+
+		/*
+		 * Wa_14019159160: disable the automatic CCS load balancing

I'm still a bit concerned that this doesn't really match what this specific workaround is asking us to do. There seems to be an agreement on various internal email threads that we need to disable load balancing, but there's no single specific workaround that officially documents that decision.

This specific workaround asks us to do a bunch of different things, and the third item it asks for is to disable load balancing in very specific cases (i.e., while the RCS is active at the same time as one or more CCS engines). Taking this workaround in isolation, it would be valid to keep load balancing active if you were just using the CCS engines and leaving the RCS idle, or if balancing was turned on/off by the GuC scheduler according to engine use at the moment, as the documented workaround seems to assume will be the case.

So in general I think we do need to disable load balancing based on other offline discussion, but blaming that entire change on Wa_14019159160 seems a bit questionable since it's not really what this specific workaround is asking us to do and someone may come back and try to "correct" the implementation of this workaround in the future without realizing there are other factors too. It would be great if we could get hardware teams to properly document this expectation somewhere (either in a separate dedicated workaround, or in the MMIO tuning guide) so that we'll have a more direct and authoritative source for such a large behavioral change.

Hi Matt,

Quoted 54 lines · Matt Roper

Quoted 32 lines · Andi Shyti

The hardware should not dynamically balance the load between CCS engines. Wa_14019159160 recommends disabling it across all platforms.

Fixes d2eae8e98d59 ("drm/i915/dg2: Drop force_probe requirement")

Signed-off-by Andi Shyti <[email protected]>

Cc Chris Wilson <[email protected]>

Cc Joonas Lahtinen <[email protected]>

Cc Matt Roper <[email protected]>

Cc <[email protected]> # v6.2+

---
  drivers/gpu/drm/i915/gt/intel_gt_regs.h     | 1 +
  drivers/gpu/drm/i915/gt/intel_workarounds.c | 5 +++++
  2 files changed, 6 insertions(+)

drivers/gpu/drm/i915/gt/intel_gt_regs.h +1 -0

index 50962cfd1353..cf709f6c05ae 100644
@@ -1478,6 +1478,7 @@
 
  #define GEN12_RCU_MODE				_MMIO(0x14800)
  #define   GEN12_RCU_MODE_CCS_ENABLE		REG_BIT(0)
+#define   XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE	REG_BIT(1)
 
  #define CHV_FUSE_GT				_MMIO(VLV_GUNIT_BASE + 0x2168)
  #define   CHV_FGT_DISABLE_SS0			(1 << 10)

drivers/gpu/drm/i915/gt/intel_workarounds.c +0 -0

index d67d44611c28..a2e78cf0b5f5 100644

drivers/gpu/drm/i915/gt/intel_workarounds.c general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_li · line 2950

Quoted 6 lines

drivers/gpu/drm/i915/gt/intel_workarounds.c +3 -0

@@ -2945,6 +2945,11 @@ general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_li
 
  		/* Wa_18028616096 */
  		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW, UGM_FRAGMENT_THRESHOLD_TO_3);
+
+		/*
+		 * Wa_14019159160: disable the automatic CCS load balancing

I'm still a bit concerned that this doesn't really match what this specific workaround is asking us to do. There seems to be an agreement on various internal email threads that we need to disable load balancing, but there's no single specific workaround that officially documents that decision.

This specific workaround asks us to do a bunch of different things, and the third item it asks for is to disable load balancing in very specific cases (i.e., while the RCS is active at the same time as one or more CCS engines). Taking this workaround in isolation, it would be valid to keep load balancing active if you were just using the CCS engines and leaving the RCS idle, or if balancing was turned on/off by the GuC scheduler according to engine use at the moment, as the documented workaround seems to assume will be the case.

Quoted 10 lines

So in general I think we do need to disable load balancing based on other offline discussion, but blaming that entire change on Wa_14019159160 seems a bit questionable since it's not really what this specific workaround is asking us to do and someone may come back and try to "correct" the implementation of this workaround in the future without realizing there are other factors too. It would be great if we could get hardware teams to properly document this expectation somewhere (either in a separate dedicated workaround, or in the MMIO tuning guide) so that we'll have a more direct and authoritative source for such a large behavioral change.

On one had I think you are right, on the other hand I think this workaround has not properly developed in what we have been describing later.

Perhaps, one solution would be to create a new generic workaround for all platforms with more than one CCS and put everyone at peace. But I don't know the process.

Are you able to help here? Or Joonas?

Thanks, Matt!
Andi

Quoted 28 lines

Quoted 18 lines

drivers/gpu/drm/i915/gt/intel_workarounds.c general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_li · line 2950

Quoted 3 lines

drivers/gpu/drm/i915/gt/intel_workarounds.c +3 -0

+
+		/*
+		 * Wa_14019159160: disable the automatic CCS load balancing

I'm still a bit concerned that this doesn't really match what this specific workaround is asking us to do. There seems to be an agreement on various internal email threads that we need to disable load balancing, but there's no single specific workaround that officially documents that decision.

This specific workaround asks us to do a bunch of different things, and the third item it asks for is to disable load balancing in very specific cases (i.e., while the RCS is active at the same time as one or more CCS engines). Taking this workaround in isolation, it would be valid to keep load balancing active if you were just using the CCS engines and leaving the RCS idle, or if balancing was turned on/off by the GuC scheduler according to engine use at the moment, as the documented workaround seems to assume will be the case.

Quoted 10 lines

So in general I think we do need to disable load balancing based on other offline discussion, but blaming that entire change on Wa_14019159160 seems a bit questionable since it's not really what this specific workaround is asking us to do and someone may come back and try to "correct" the implementation of this workaround in the future without realizing there are other factors too. It would be great if we could get hardware teams to properly document this expectation somewhere (either in a separate dedicated workaround, or in the MMIO tuning guide) so that we'll have a more direct and authoritative source for such a large behavioral change.

Quoted 3 lines

On one had I think you are right, on the other hand I think this workaround has not properly developed in what we have been describing later.

I think it is not so much that the w/a is 'not properly developed'. It's more that this w/a plus others when taken in combination plus knowledge of future directions has led to an architectural decision that is beyond the scope of the w/a.

As such, I think Matt is definitely correct. Tagging a code change with a w/a number when that change does something very different to what is described in the w/a is wrong and a maintenance issue waiting to happen.

At the very least, you should just put in a comment explaining the situation. E.g.:

/*
* Wa_14019159160: This w/a plus others cause significant issues with the use of
* load balancing. Hence an architectural level decision was taking to simply
* disable automatic CCS load balancing completely.
*/

Enable only one CCS engine by default with all the compute sices allocated to it.

While generating the list of UABI engines to be exposed to the user, exclude any additional CCS engines beyond the first instance.

This change can be tested with igt i915_query.

Fixes d2eae8e98d59 ("drm/i915/dg2: Drop force_probe requirement")

Requires: 97aba5e46038 ("drm/i915/gt: Refactor uabi engine class/instance list creation")

Signed-off-by Andi Shyti <[email protected]>

Cc Chris Wilson <[email protected]>

Cc Joonas Lahtinen <[email protected]>

Cc Matt Roper <[email protected]>

Cc <[email protected]> # v6.2+

---
 drivers/gpu/drm/i915/gt/intel_engine_user.c | 11 ++++++++++
 drivers/gpu/drm/i915/gt/intel_gt.c          | 23 +++++++++++++++++++++
 drivers/gpu/drm/i915/gt/intel_gt_regs.h     |  5 +++++
 3 files changed, 39 insertions(+)

drivers/gpu/drm/i915/gt/intel_engine_user.c +11 -0

index 11cc06c0c785..9ef1c4ce252d 100644
@@ -208,6 +208,7 @@ void intel_engines_driver_register(struct drm_i915_private *i915)
 	struct list_head *it, *next;
 	struct rb_node **p, *prev;
 	LIST_HEAD(engines);
+	u16 uabi_ccs = 0;
 
 	sort_engines(i915, &engines);
 

@@ -244,6 +245,16 @@ void intel_engines_driver_register(struct drm_i915_private *i915)
 		if (uabi_class > I915_LAST_UABI_ENGINE_CLASS)
 			continue;
 
+		/*
+		 * The load is balanced among all the available compute
+		 * slices. Expose only the first instance of the compute
+		 * engine.
+		 */
+		if (IS_DG2(i915) &&
+		    uabi_class == I915_ENGINE_CLASS_COMPUTE &&
+		    uabi_ccs++)
+			continue;
+
 		GEM_BUG_ON(uabi_class >=
 			   ARRAY_SIZE(i915->engine_uabi_class_count));
 		i915->engine_uabi_class_count[uabi_class]++;

drivers/gpu/drm/i915/gt/intel_gt.c +8 -0

index a425db5ed3a2..0aac97439552 100644
@@ -168,6 +168,26 @@ static void init_unused_rings(struct intel_gt *gt)
 	}
 }
 
+static void intel_gt_apply_ccs_mode(struct intel_gt *gt)
+{
+	u32 mode;
+	int cslice;
+
+	if (!IS_DG2(gt->i915))
+		return;
+

drivers/gpu/drm/i915/gt/intel_gt.c +1 -0

+
 	/*
 	 * At least 830 can leave some of the unused rings
 	 * "active" (ie. head != tail) after resume which

drivers/gpu/drm/i915/gt/intel_gt_regs.h +5 -0

index cf709f6c05ae..8224dd99c7d7 100644
@@ -1480,6 +1480,11 @@
 #define   GEN12_RCU_MODE_CCS_ENABLE		REG_BIT(0)
 #define   XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE	REG_BIT(1)
 
+#define XEHP_CCS_MODE				_MMIO(0x14804)
+#define   XEHP_CCS_MODE_CSLICE_MASK		REG_GENMASK(2, 0) /* CCS0-3 + rsvd */
+#define   XEHP_CCS_MODE_CSLICE_WIDTH		ilog2(XEHP_CCS_MODE_CSLICE_MASK + 1)
+#define   XEHP_CCS_MODE_CSLICE(cslice, ccs)	(ccs << (cslice * XEHP_CCS_MODE_CSLICE_WIDTH))
+
 #define CHV_FUSE_GT				_MMIO(VLV_GUNIT_BASE + 0x2168)
 #define   CHV_FGT_DISABLE_SS0			(1 << 10)
 #define   CHV_FGT_DISABLE_SS1			(1 << 11)

Signature · 1 line

2.43.0

Quoted 16 lines

drivers/gpu/drm/i915/gt/intel_gt.c +16 -0

+static void intel_gt_apply_ccs_mode(struct intel_gt *gt)
+{
+	u32 mode;
+	int cslice;
+
+	if (!IS_DG2(gt->i915))
+		return;
+
+	/* Set '0' as a default CCS id to all the cslices */
+	mode = 0;
+
+	for (cslice = 0; cslice < hweight32(CCS_MASK(gt)); cslice++)
+		/* Write 0x7 if no CCS context dispatches to this cslice */
+		if (!(CCS_MASK(gt) & BIT(cslice)))
+			mode |= XEHP_CCS_MODE_CSLICE(cslice,
+						     XEHP_CCS_MODE_CSLICE_MASK);

Review

Quoted 2 lines

(quoted diff) +2 -0

+
+	intel_uncore_write(gt->uncore, XEHP_CCS_MODE, mode);

This is still going to hook all available cslices up to hardware engine ccs0. But what you actually want is to hook them all up to what userspace sees as CCS0 (i.e., the first CCS engine that wasn't fused off). Hardware's engine numbering and userspace's numbering aren't the same.

Also, if you have a part that only has hardware ccs1/cslice1 for example, you're not going to set cslices 2 & 3 to 0x7 properly.

So probably what you want is something like this (untested):

static void intel_gt_apply_ccs_mode(struct intel_gt *gt)

{
        u32 mode = 0;
        int first_ccs = __ffs(CCS_MASK(gt));

        /*
         * Re-assign every present cslice to the first available CCS
         * engine; mark unavailable cslices as unused.
         */
        for (int cslice = 0; cslice < 4; cslice++) {
                if (CCS_MASK(gt) & BIT(cslice))
                        mode |= XEHP_CCS_MODE_CSLICE(cslice, first_ccs);
                else
                        mode |= XEHP_CCS_MODE_CSLICE(cslice,
                                                     XEHP_CCS_MODE_CSLICE_MASK);
        }

        intel_uncore_write(gt->uncore, XEHP_CCS_MODE, mode);
}

Quoted 9 lines

drivers/gpu/drm/i915/gt/intel_gt.c +2 -0

+}
+
 int intel_gt_init_hw(struct intel_gt *gt)
 {
 	struct drm_i915_private *i915 = gt->i915;

@@ -195,6 +215,9 @@ int intel_gt_init_hw(struct intel_gt *gt)
 
 	intel_gt_init_swizzling(gt);
 

drivers/gpu/drm/i915/gt/intel_gt.c int intel_gt_init_hw(struct intel_gt *gt) · line 219

Quoted 2 lines

drivers/gpu/drm/i915/gt/intel_gt.c +2 -0

+	/* Configure CCS mode */
+	intel_gt_apply_ccs_mode(gt);

This is only setting this once during init. The value gets lost on every RCS/CCS reset, so we need to make sure it gets reapplied when necessary. That means you either need to add this to the GuC regset, or you need to implement the programming as a "fake workaround" so that the workaround framework will take care of the re-application for you.