drm/radeon: add support for MC/VM setup on CIK (v6)

The vm callbacks are the same as the SI ones right now
(same regs and bits). We could share the SI variants, and
I may yet do that, but I figured I would add CIK specific
ones for now in case we need to change anything.

V2: add documentation, minor fixes.
V3: integrate vram offset fixes for APUs
V4: enable 2 level VM PTs
V5: index SH_MEM_* regs properly
V6: add ib_parse()
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/radeon/cik.c

@@ -33,6 +33,7 @@
 
 extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
 extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
+extern void si_vram_gtt_location(struct radeon_device *rdev, struct radeon_mc *mc);
 
 /*
  * Core functions
@@ -1387,3 +1388,363 @@ int cik_asic_reset(struct radeon_device *rdev)
 
 	return cik_gfx_gpu_soft_reset(rdev);
 }
+
+/* MC */
+/**
+ * cik_mc_program - program the GPU memory controller
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Set the location of vram, gart, and AGP in the GPU's
+ * physical address space (CIK).
+ */
+static void cik_mc_program(struct radeon_device *rdev)
+{
+	struct evergreen_mc_save save;
+	u32 tmp;
+	int i, j;
+
+	/* Initialize HDP */
+	for (i = 0, j = 0; i < 32; i++, j += 0x18) {
+		WREG32((0x2c14 + j), 0x00000000);
+		WREG32((0x2c18 + j), 0x00000000);
+		WREG32((0x2c1c + j), 0x00000000);
+		WREG32((0x2c20 + j), 0x00000000);
+		WREG32((0x2c24 + j), 0x00000000);
+	}
+	WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
+
+	evergreen_mc_stop(rdev, &save);
+	if (radeon_mc_wait_for_idle(rdev)) {
+		dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
+	}
+	/* Lockout access through VGA aperture*/
+	WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
+	/* Update configuration */
+	WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
+	       rdev->mc.vram_start >> 12);
+	WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
+	       rdev->mc.vram_end >> 12);
+	WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
+	       rdev->vram_scratch.gpu_addr >> 12);
+	tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
+	tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
+	WREG32(MC_VM_FB_LOCATION, tmp);
+	/* XXX double check these! */
+	WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
+	WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
+	WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
+	WREG32(MC_VM_AGP_BASE, 0);
+	WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
+	WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
+	if (radeon_mc_wait_for_idle(rdev)) {
+		dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
+	}
+	evergreen_mc_resume(rdev, &save);
+	/* we need to own VRAM, so turn off the VGA renderer here
+	 * to stop it overwriting our objects */
+	rv515_vga_render_disable(rdev);
+}
+
+/**
+ * cik_mc_init - initialize the memory controller driver params
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Look up the amount of vram, vram width, and decide how to place
+ * vram and gart within the GPU's physical address space (CIK).
+ * Returns 0 for success.
+ */
+static int cik_mc_init(struct radeon_device *rdev)
+{
+	u32 tmp;
+	int chansize, numchan;
+
+	/* Get VRAM informations */
+	rdev->mc.vram_is_ddr = true;
+	tmp = RREG32(MC_ARB_RAMCFG);
+	if (tmp & CHANSIZE_MASK) {
+		chansize = 64;
+	} else {
+		chansize = 32;
+	}
+	tmp = RREG32(MC_SHARED_CHMAP);
+	switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
+	case 0:
+	default:
+		numchan = 1;
+		break;
+	case 1:
+		numchan = 2;
+		break;
+	case 2:
+		numchan = 4;
+		break;
+	case 3:
+		numchan = 8;
+		break;
+	case 4:
+		numchan = 3;
+		break;
+	case 5:
+		numchan = 6;
+		break;
+	case 6:
+		numchan = 10;
+		break;
+	case 7:
+		numchan = 12;
+		break;
+	case 8:
+		numchan = 16;
+		break;
+	}
+	rdev->mc.vram_width = numchan * chansize;
+	/* Could aper size report 0 ? */
+	rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
+	rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
+	/* size in MB on si */
+	rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE) * 1024 * 1024;
+	rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE) * 1024 * 1024;
+	rdev->mc.visible_vram_size = rdev->mc.aper_size;
+	si_vram_gtt_location(rdev, &rdev->mc);
+	radeon_update_bandwidth_info(rdev);
+
+	return 0;
+}
+
+/*
+ * GART
+ * VMID 0 is the physical GPU addresses as used by the kernel.
+ * VMIDs 1-15 are used for userspace clients and are handled
+ * by the radeon vm/hsa code.
+ */
+/**
+ * cik_pcie_gart_tlb_flush - gart tlb flush callback
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Flush the TLB for the VMID 0 page table (CIK).
+ */
+void cik_pcie_gart_tlb_flush(struct radeon_device *rdev)
+{
+	/* flush hdp cache */
+	WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0);
+
+	/* bits 0-15 are the VM contexts0-15 */
+	WREG32(VM_INVALIDATE_REQUEST, 0x1);
+}
+
+/**
+ * cik_pcie_gart_enable - gart enable
+ *
+ * @rdev: radeon_device pointer
+ *
+ * This sets up the TLBs, programs the page tables for VMID0,
+ * sets up the hw for VMIDs 1-15 which are allocated on
+ * demand, and sets up the global locations for the LDS, GDS,
+ * and GPUVM for FSA64 clients (CIK).
+ * Returns 0 for success, errors for failure.
+ */
+static int cik_pcie_gart_enable(struct radeon_device *rdev)
+{
+	int r, i;
+
+	if (rdev->gart.robj == NULL) {
+		dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
+		return -EINVAL;
+	}
+	r = radeon_gart_table_vram_pin(rdev);
+	if (r)
+		return r;
+	radeon_gart_restore(rdev);
+	/* Setup TLB control */
+	WREG32(MC_VM_MX_L1_TLB_CNTL,
+	       (0xA << 7) |
+	       ENABLE_L1_TLB |
+	       SYSTEM_ACCESS_MODE_NOT_IN_SYS |
+	       ENABLE_ADVANCED_DRIVER_MODEL |
+	       SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
+	/* Setup L2 cache */
+	WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
+	       ENABLE_L2_FRAGMENT_PROCESSING |
+	       ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
+	       ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
+	       EFFECTIVE_L2_QUEUE_SIZE(7) |
+	       CONTEXT1_IDENTITY_ACCESS_MODE(1));
+	WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
+	WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
+	       L2_CACHE_BIGK_FRAGMENT_SIZE(6));
+	/* setup context0 */
+	WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
+	WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
+	WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
+	WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
+			(u32)(rdev->dummy_page.addr >> 12));
+	WREG32(VM_CONTEXT0_CNTL2, 0);
+	WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
+				  RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
+
+	WREG32(0x15D4, 0);
+	WREG32(0x15D8, 0);
+	WREG32(0x15DC, 0);
+
+	/* empty context1-15 */
+	/* FIXME start with 4G, once using 2 level pt switch to full
+	 * vm size space
+	 */
+	/* set vm size, must be a multiple of 4 */
+	WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
+	WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
+	for (i = 1; i < 16; i++) {
+		if (i < 8)
+			WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
+			       rdev->gart.table_addr >> 12);
+		else
+			WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
+			       rdev->gart.table_addr >> 12);
+	}
+
+	/* enable context1-15 */
+	WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
+	       (u32)(rdev->dummy_page.addr >> 12));
+	WREG32(VM_CONTEXT1_CNTL2, 0);
+	WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
+				RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
+
+	/* TC cache setup ??? */
+	WREG32(TC_CFG_L1_LOAD_POLICY0, 0);
+	WREG32(TC_CFG_L1_LOAD_POLICY1, 0);
+	WREG32(TC_CFG_L1_STORE_POLICY, 0);
+
+	WREG32(TC_CFG_L2_LOAD_POLICY0, 0);
+	WREG32(TC_CFG_L2_LOAD_POLICY1, 0);
+	WREG32(TC_CFG_L2_STORE_POLICY0, 0);
+	WREG32(TC_CFG_L2_STORE_POLICY1, 0);
+	WREG32(TC_CFG_L2_ATOMIC_POLICY, 0);
+
+	WREG32(TC_CFG_L1_VOLATILE, 0);
+	WREG32(TC_CFG_L2_VOLATILE, 0);
+
+	if (rdev->family == CHIP_KAVERI) {
+		u32 tmp = RREG32(CHUB_CONTROL);
+		tmp &= ~BYPASS_VM;
+		WREG32(CHUB_CONTROL, tmp);
+	}
+
+	/* XXX SH_MEM regs */
+	/* where to put LDS, scratch, GPUVM in FSA64 space */
+	for (i = 0; i < 16; i++) {
+		WREG32(SRBM_GFX_CNTL, VMID(i));
+		WREG32(SH_MEM_CONFIG, 0);
+		WREG32(SH_MEM_APE1_BASE, 1);
+		WREG32(SH_MEM_APE1_LIMIT, 0);
+		WREG32(SH_MEM_BASES, 0);
+	}
+	WREG32(SRBM_GFX_CNTL, 0);
+
+	cik_pcie_gart_tlb_flush(rdev);
+	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
+		 (unsigned)(rdev->mc.gtt_size >> 20),
+		 (unsigned long long)rdev->gart.table_addr);
+	rdev->gart.ready = true;
+	return 0;
+}
+
+/**
+ * cik_pcie_gart_disable - gart disable
+ *
+ * @rdev: radeon_device pointer
+ *
+ * This disables all VM page table (CIK).
+ */
+static void cik_pcie_gart_disable(struct radeon_device *rdev)
+{
+	/* Disable all tables */
+	WREG32(VM_CONTEXT0_CNTL, 0);
+	WREG32(VM_CONTEXT1_CNTL, 0);
+	/* Setup TLB control */
+	WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
+	       SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
+	/* Setup L2 cache */
+	WREG32(VM_L2_CNTL,
+	       ENABLE_L2_FRAGMENT_PROCESSING |
+	       ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
+	       ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
+	       EFFECTIVE_L2_QUEUE_SIZE(7) |
+	       CONTEXT1_IDENTITY_ACCESS_MODE(1));
+	WREG32(VM_L2_CNTL2, 0);
+	WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
+	       L2_CACHE_BIGK_FRAGMENT_SIZE(6));
+	radeon_gart_table_vram_unpin(rdev);
+}
+
+/**
+ * cik_pcie_gart_fini - vm fini callback
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Tears down the driver GART/VM setup (CIK).
+ */
+static void cik_pcie_gart_fini(struct radeon_device *rdev)
+{
+	cik_pcie_gart_disable(rdev);
+	radeon_gart_table_vram_free(rdev);
+	radeon_gart_fini(rdev);
+}
+
+/* vm parser */
+/**
+ * cik_ib_parse - vm ib_parse callback
+ *
+ * @rdev: radeon_device pointer
+ * @ib: indirect buffer pointer
+ *
+ * CIK uses hw IB checking so this is a nop (CIK).
+ */
+int cik_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
+{
+	return 0;
+}
+
+/*
+ * vm
+ * VMID 0 is the physical GPU addresses as used by the kernel.
+ * VMIDs 1-15 are used for userspace clients and are handled
+ * by the radeon vm/hsa code.
+ */
+/**
+ * cik_vm_init - cik vm init callback
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Inits cik specific vm parameters (number of VMs, base of vram for
+ * VMIDs 1-15) (CIK).
+ * Returns 0 for success.
+ */
+int cik_vm_init(struct radeon_device *rdev)
+{
+	/* number of VMs */
+	rdev->vm_manager.nvm = 16;
+	/* base offset of vram pages */
+	if (rdev->flags & RADEON_IS_IGP) {
+		u64 tmp = RREG32(MC_VM_FB_OFFSET);
+		tmp <<= 22;
+		rdev->vm_manager.vram_base_offset = tmp;
+	} else
+		rdev->vm_manager.vram_base_offset = 0;
+
+	return 0;
+}
+
+/**
+ * cik_vm_fini - cik vm fini callback
+ *
+ * @rdev: radeon_device pointer
+ *
+ * Tear down any asic specific VM setup (CIK).
+ */
+void cik_vm_fini(struct radeon_device *rdev)
+{
+}
+

drivers/gpu/drm/radeon/cikd.h

@@ -28,16 +28,106 @@
 
 #define CIK_RB_BITMAP_WIDTH_PER_SH  2
 
+#define VGA_HDP_CONTROL  				0x328
+#define		VGA_MEMORY_DISABLE				(1 << 4)
+
 #define DMIF_ADDR_CALC  				0xC00
 
+#define	SRBM_GFX_CNTL				        0xE44
+#define		PIPEID(x)					((x) << 0)
+#define		MEID(x)						((x) << 2)
+#define		VMID(x)						((x) << 4)
+#define		QUEUEID(x)					((x) << 8)
+
 #define	SRBM_STATUS2				        0xE4C
 #define	SRBM_STATUS				        0xE50
 
+#define VM_L2_CNTL					0x1400
+#define		ENABLE_L2_CACHE					(1 << 0)
+#define		ENABLE_L2_FRAGMENT_PROCESSING			(1 << 1)
+#define		L2_CACHE_PTE_ENDIAN_SWAP_MODE(x)		((x) << 2)
+#define		L2_CACHE_PDE_ENDIAN_SWAP_MODE(x)		((x) << 4)
+#define		ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE		(1 << 9)
+#define		ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE	(1 << 10)
+#define		EFFECTIVE_L2_QUEUE_SIZE(x)			(((x) & 7) << 15)
+#define		CONTEXT1_IDENTITY_ACCESS_MODE(x)		(((x) & 3) << 19)
+#define VM_L2_CNTL2					0x1404
+#define		INVALIDATE_ALL_L1_TLBS				(1 << 0)
+#define		INVALIDATE_L2_CACHE				(1 << 1)
+#define		INVALIDATE_CACHE_MODE(x)			((x) << 26)
+#define			INVALIDATE_PTE_AND_PDE_CACHES		0
+#define			INVALIDATE_ONLY_PTE_CACHES		1
+#define			INVALIDATE_ONLY_PDE_CACHES		2
+#define VM_L2_CNTL3					0x1408
+#define		BANK_SELECT(x)					((x) << 0)
+#define		L2_CACHE_UPDATE_MODE(x)				((x) << 6)
+#define		L2_CACHE_BIGK_FRAGMENT_SIZE(x)			((x) << 15)
+#define		L2_CACHE_BIGK_ASSOCIATIVITY			(1 << 20)
+#define	VM_L2_STATUS					0x140C
+#define		L2_BUSY						(1 << 0)
+#define VM_CONTEXT0_CNTL				0x1410
+#define		ENABLE_CONTEXT					(1 << 0)
+#define		PAGE_TABLE_DEPTH(x)				(((x) & 3) << 1)
+#define		RANGE_PROTECTION_FAULT_ENABLE_DEFAULT		(1 << 4)
+#define VM_CONTEXT1_CNTL				0x1414
+#define VM_CONTEXT0_CNTL2				0x1430
+#define VM_CONTEXT1_CNTL2				0x1434
+#define	VM_CONTEXT8_PAGE_TABLE_BASE_ADDR		0x1438
+#define	VM_CONTEXT9_PAGE_TABLE_BASE_ADDR		0x143c
+#define	VM_CONTEXT10_PAGE_TABLE_BASE_ADDR		0x1440
+#define	VM_CONTEXT11_PAGE_TABLE_BASE_ADDR		0x1444
+#define	VM_CONTEXT12_PAGE_TABLE_BASE_ADDR		0x1448
+#define	VM_CONTEXT13_PAGE_TABLE_BASE_ADDR		0x144c
+#define	VM_CONTEXT14_PAGE_TABLE_BASE_ADDR		0x1450
+#define	VM_CONTEXT15_PAGE_TABLE_BASE_ADDR		0x1454
+
+#define VM_INVALIDATE_REQUEST				0x1478
+#define VM_INVALIDATE_RESPONSE				0x147c
+
+#define VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR	0x1518
+#define VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR	0x151c
+
+#define	VM_CONTEXT0_PAGE_TABLE_BASE_ADDR		0x153c
+#define	VM_CONTEXT1_PAGE_TABLE_BASE_ADDR		0x1540
+#define	VM_CONTEXT2_PAGE_TABLE_BASE_ADDR		0x1544
+#define	VM_CONTEXT3_PAGE_TABLE_BASE_ADDR		0x1548
+#define	VM_CONTEXT4_PAGE_TABLE_BASE_ADDR		0x154c
+#define	VM_CONTEXT5_PAGE_TABLE_BASE_ADDR		0x1550
+#define	VM_CONTEXT6_PAGE_TABLE_BASE_ADDR		0x1554
+#define	VM_CONTEXT7_PAGE_TABLE_BASE_ADDR		0x1558
+#define	VM_CONTEXT0_PAGE_TABLE_START_ADDR		0x155c
+#define	VM_CONTEXT1_PAGE_TABLE_START_ADDR		0x1560
+
+#define	VM_CONTEXT0_PAGE_TABLE_END_ADDR			0x157C
+#define	VM_CONTEXT1_PAGE_TABLE_END_ADDR			0x1580
+
 #define MC_SHARED_CHMAP						0x2004
 #define		NOOFCHAN_SHIFT					12
 #define		NOOFCHAN_MASK					0x0000f000
 #define MC_SHARED_CHREMAP					0x2008
 
+#define CHUB_CONTROL					0x1864
+#define		BYPASS_VM					(1 << 0)
+
+#define	MC_VM_FB_LOCATION				0x2024
+#define	MC_VM_AGP_TOP					0x2028
+#define	MC_VM_AGP_BOT					0x202C
+#define	MC_VM_AGP_BASE					0x2030
+#define	MC_VM_SYSTEM_APERTURE_LOW_ADDR			0x2034
+#define	MC_VM_SYSTEM_APERTURE_HIGH_ADDR			0x2038
+#define	MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR		0x203C
+
+#define	MC_VM_MX_L1_TLB_CNTL				0x2064
+#define		ENABLE_L1_TLB					(1 << 0)
+#define		ENABLE_L1_FRAGMENT_PROCESSING			(1 << 1)
+#define		SYSTEM_ACCESS_MODE_PA_ONLY			(0 << 3)
+#define		SYSTEM_ACCESS_MODE_USE_SYS_MAP			(1 << 3)
+#define		SYSTEM_ACCESS_MODE_IN_SYS			(2 << 3)
+#define		SYSTEM_ACCESS_MODE_NOT_IN_SYS			(3 << 3)
+#define		SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU	(0 << 5)
+#define		ENABLE_ADVANCED_DRIVER_MODEL			(1 << 6)
+#define	MC_VM_FB_OFFSET					0x2068
+
 #define	MC_ARB_RAMCFG					0x2760
 #define		NOOFBANK_SHIFT					0
 #define		NOOFBANK_MASK					0x00000003
@@ -61,10 +151,16 @@
 #define HDP_MISC_CNTL					0x2F4C
 #define 	HDP_FLUSH_INVALIDATE_CACHE			(1 << 0)
 
+#define	CONFIG_MEMSIZE					0x5428
+
+#define HDP_MEM_COHERENCY_FLUSH_CNTL			0x5480
+
 #define	BIF_FB_EN						0x5490
 #define		FB_READ_EN					(1 << 0)
 #define		FB_WRITE_EN					(1 << 1)
 
+#define HDP_REG_COHERENCY_FLUSH_CNTL			0x54A0
+
 #define	GRBM_CNTL					0x8000
 #define		GRBM_READ_TIMEOUT(x)				((x) << 0)
 
@@ -189,6 +285,24 @@
 
 #define	SQ_CONFIG					0x8C00
 
+#define	SH_MEM_BASES					0x8C28
+/* if PTR32, these are the bases for scratch and lds */
+#define		PRIVATE_BASE(x)					((x) << 0) /* scratch */
+#define		SHARED_BASE(x)					((x) << 16) /* LDS */
+#define	SH_MEM_APE1_BASE				0x8C2C
+/* if PTR32, this is the base location of GPUVM */
+#define	SH_MEM_APE1_LIMIT				0x8C30
+/* if PTR32, this is the upper limit of GPUVM */
+#define	SH_MEM_CONFIG					0x8C34
+#define		PTR32						(1 << 0)
+#define		ALIGNMENT_MODE(x)				((x) << 2)
+#define			SH_MEM_ALIGNMENT_MODE_DWORD			0
+#define			SH_MEM_ALIGNMENT_MODE_DWORD_STRICT		1
+#define			SH_MEM_ALIGNMENT_MODE_STRICT			2
+#define			SH_MEM_ALIGNMENT_MODE_UNALIGNED			3
+#define		DEFAULT_MTYPE(x)				((x) << 4)
+#define		APE1_MTYPE(x)					((x) << 7)
+
 #define	SX_DEBUG_1					0x9060
 
 #define	SPI_CONFIG_CNTL					0x9100
@@ -293,6 +407,17 @@
 #define	TCP_CHAN_STEER_LO				0xac0c
 #define	TCP_CHAN_STEER_HI				0xac10
 
+#define	TC_CFG_L1_LOAD_POLICY0				0xAC68
+#define	TC_CFG_L1_LOAD_POLICY1				0xAC6C
+#define	TC_CFG_L1_STORE_POLICY				0xAC70
+#define	TC_CFG_L2_LOAD_POLICY0				0xAC74
+#define	TC_CFG_L2_LOAD_POLICY1				0xAC78
+#define	TC_CFG_L2_STORE_POLICY0				0xAC7C
+#define	TC_CFG_L2_STORE_POLICY1				0xAC80
+#define	TC_CFG_L2_ATOMIC_POLICY				0xAC84
+#define	TC_CFG_L1_VOLATILE				0xAC88
+#define	TC_CFG_L2_VOLATILE				0xAC8C
+
 #define PA_SC_RASTER_CONFIG                             0x28350
 #       define RASTER_CONFIG_RB_MAP_0                   0
 #       define RASTER_CONFIG_RB_MAP_1                   1

drivers/gpu/drm/radeon/si.c

@@ -3535,8 +3535,8 @@ static void si_mc_program(struct radeon_device *rdev)
 	}
 }
 
-static void si_vram_gtt_location(struct radeon_device *rdev,
-				 struct radeon_mc *mc)
+void si_vram_gtt_location(struct radeon_device *rdev,
+			  struct radeon_mc *mc)
 {
 	if (mc->mc_vram_size > 0xFFC0000000ULL) {
 		/* leave room for at least 1024M GTT */