drm/amdgpu: move static CSA address to top of address space v2

Move the CSA area to the top of the VA space to avoid clashing with
HMM/ATC in the lower range on GFX9.

v2: wrong sign noticed by Roger, rebase on CSA_VADDR cleanup, handle VA
hole on GFX9 as well.
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Monk Liu <monk.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c

@@ -24,6 +24,18 @@
 #include "amdgpu.h"
 #define MAX_KIQ_REG_WAIT	100000000 /* in usecs */
 
+uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev)
+{
+	uint64_t addr = adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT;
+
+	addr -= AMDGPU_VA_RESERVED_SIZE;
+
+	if (addr >= AMDGPU_VA_HOLE_START)
+		addr |= AMDGPU_VA_HOLE_END;
+
+	return addr;
+}
+
 bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev)
 {
 	/* By now all MMIO pages except mailbox are blocked */
@@ -55,14 +67,14 @@ void amdgpu_free_static_csa(struct amdgpu_device *adev) {
 
 /*
  * amdgpu_map_static_csa should be called during amdgpu_vm_init
- * it maps virtual address "AMDGPU_VA_RESERVED_SIZE - AMDGPU_CSA_SIZE"
- * to this VM, and each command submission of GFX should use this virtual
- * address within META_DATA init package to support SRIOV gfx preemption.
+ * it maps virtual address amdgpu_csa_vaddr() to this VM, and each command
+ * submission of GFX should use this virtual address within META_DATA init
+ * package to support SRIOV gfx preemption.
  */
-
 int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 			  struct amdgpu_bo_va **bo_va)
 {
+	uint64_t csa_addr = amdgpu_csa_vaddr(adev) & AMDGPU_VA_HOLE_MASK;
 	struct ww_acquire_ctx ticket;
 	struct list_head list;
 	struct amdgpu_bo_list_entry pd;
@@ -90,7 +102,7 @@ int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		return -ENOMEM;
 	}
 
-	r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, AMDGPU_CSA_VADDR,
+	r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, csa_addr,
 				AMDGPU_CSA_SIZE);
 	if (r) {
 		DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", r);
@@ -99,7 +111,7 @@ int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		return r;
 	}
 
-	r = amdgpu_vm_bo_map(adev, *bo_va, AMDGPU_CSA_VADDR, 0, AMDGPU_CSA_SIZE,
+	r = amdgpu_vm_bo_map(adev, *bo_va, csa_addr, 0, AMDGPU_CSA_SIZE,
 			     AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
 			     AMDGPU_PTE_EXECUTABLE);
 

drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h

@@ -251,8 +251,7 @@ struct amdgpu_virt {
 	uint32_t gim_feature;
 };
 
-#define AMDGPU_CSA_SIZE    (8 * 1024)
-#define AMDGPU_CSA_VADDR   (AMDGPU_VA_RESERVED_SIZE - AMDGPU_CSA_SIZE)
+#define AMDGPU_CSA_SIZE		(8 * 1024)
 
 #define amdgpu_sriov_enabled(adev) \
 ((adev)->virt.caps & AMDGPU_SRIOV_CAPS_ENABLE_IOV)
@@ -279,6 +278,8 @@ static inline bool is_virtual_machine(void)
 }
 
 struct amdgpu_vm;
+
+uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev);
 bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev);
 int amdgpu_allocate_static_csa(struct amdgpu_device *adev);
 int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

@@ -7132,11 +7132,11 @@ static void gfx_v8_0_ring_emit_ce_meta(struct amdgpu_ring *ring)
 	} ce_payload = {};
 
 	if (ring->adev->virt.chained_ib_support) {
-		ce_payload_addr = AMDGPU_CSA_VADDR +
+		ce_payload_addr = amdgpu_csa_vaddr(ring->adev) +
 			offsetof(struct vi_gfx_meta_data_chained_ib, ce_payload);
 		cnt_ce = (sizeof(ce_payload.chained) >> 2) + 4 - 2;
 	} else {
-		ce_payload_addr = AMDGPU_CSA_VADDR +
+		ce_payload_addr = amdgpu_csa_vaddr(ring->adev) +
 			offsetof(struct vi_gfx_meta_data, ce_payload);
 		cnt_ce = (sizeof(ce_payload.regular) >> 2) + 4 - 2;
 	}
@@ -7160,7 +7160,7 @@ static void gfx_v8_0_ring_emit_de_meta(struct amdgpu_ring *ring)
 		struct vi_de_ib_state_chained_ib chained;
 	} de_payload = {};
 
-	csa_addr = AMDGPU_CSA_VADDR;
+	csa_addr = amdgpu_csa_vaddr(ring->adev);
 	gds_addr = csa_addr + 4096;
 	if (ring->adev->virt.chained_ib_support) {
 		de_payload.chained.gds_backup_addrlo = lower_32_bits(gds_addr);

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c

@@ -3865,7 +3865,7 @@ static void gfx_v9_0_ring_emit_ce_meta(struct amdgpu_ring *ring)
 	int cnt;
 
 	cnt = (sizeof(ce_payload) >> 2) + 4 - 2;
-	csa_addr = AMDGPU_CSA_VADDR;
+	csa_addr = amdgpu_csa_vaddr(ring->adev);
 
 	amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, cnt));
 	amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(2) |
@@ -3883,7 +3883,7 @@ static void gfx_v9_0_ring_emit_de_meta(struct amdgpu_ring *ring)
 	uint64_t csa_addr, gds_addr;
 	int cnt;
 
-	csa_addr = AMDGPU_CSA_VADDR;
+	csa_addr = amdgpu_csa_vaddr(ring->adev);
 	gds_addr = csa_addr + 4096;
 	de_payload.gds_backup_addrlo = lower_32_bits(gds_addr);
 	de_payload.gds_backup_addrhi = upper_32_bits(gds_addr);