path: root/kernel/gsp_rm_tu104.h

/*
 * TU104 Hardware Tables for GSP-RM Emulation
 *
 * Static device/engine/interrupt tables for the emulated TU104 GPU.
 * Derived from device/engine interface definitions, observed GSP-RM
 * behaviour, and hardware reference headers (hwref/turing/tu104).
 *
 * These tables replace what the real GSP-RM firmware would read from
 * PTOP hardware registers during boot.
 */

#ifndef _GSP_RM_TU104_H_
#define _GSP_RM_TU104_H_

#include <linux/types.h>

/* ============================================================
 * Wire structures (from 570.144 reference, ABI-checked)
 * ============================================================ */

/* NV2080_CTRL_FIFO_DEVICE_ENTRY (ctrl2080fifo.h) - 100 bytes */
struct tu104_fifo_device_entry {
	u32 engineData[16];	/* ENGINE_INFO_TYPE_* indexed */
	u32 pbdmaIds[2];
	u32 pbdmaFaultIds[2];
	u32 numPbdmas;
	char engineName[16];
};

/* NV2080_CTRL_FIFO_GET_DEVICE_INFO_TABLE_PARAMS - 3212 bytes max */
#define TU104_FIFO_TABLE_MAX_ENTRIES	32
#define TU104_FIFO_TABLE_MAX_DEVICES	256

struct tu104_fifo_get_device_info_params {
	u32 baseIndex;
	u32 numEntries;
	u8  bMore;
	u8  _pad[3];
	struct tu104_fifo_device_entry entries[TU104_FIFO_TABLE_MAX_ENTRIES];
};

/* NV2080_CTRL_INTERNAL_DEVICE_INFO (ctrl2080internal.h) - 48 bytes */
struct tu104_internal_device_info {
	u32 faultId;
	u32 instanceId;
	u32 typeEnum;
	u32 resetId;
	u32 devicePriBase;
	u32 isEngine;
	u32 rlEngId;
	u32 runlistPriBase;
	u32 groupId;
	u32 ginTargetId;
	u32 deviceBroadcastPriBase;
	u32 groupLocalInstanceId;
};

/* NV2080_CTRL_INTERNAL_GET_DEVICE_INFO_TABLE_PARAMS - up to 512 entries */
#define TU104_INTERNAL_TABLE_MAX_ENTRIES	512

struct tu104_internal_get_device_info_params {
	u32 numEntries;
	struct tu104_internal_device_info deviceInfoTable[TU104_INTERNAL_TABLE_MAX_ENTRIES];
};

/* NV2080_CTRL_INTERNAL_INTR_GET_KERNEL_TABLE_PARAMS */
#define TU104_INTR_TABLE_MAX_SIZE	128

struct tu104_intr_table_entry {
	u16 engineIdx;
	u16 _pad0;
	u32 pmcIntrMask;
	u32 vectorStall;
	u32 vectorNonStall;
};

/* NV2080_INTR_CATEGORY_SUBTREE_MAP (ctrl2080internal.h) - 2 bytes */
struct tu104_intr_subtree_map {
	u8 subtreeStart;
	u8 subtreeEnd;
};

/* NV2080_INTR_CATEGORY enum */
#define TU104_INTR_CATEGORY_DEFAULT			0
#define TU104_INTR_CATEGORY_ESCHED_DRIVEN_ENGINE		1
#define TU104_INTR_CATEGORY_ESCHED_DRIVEN_ENGINE_NOTIFICATION	2
#define TU104_INTR_CATEGORY_RUNLIST			3
#define TU104_INTR_CATEGORY_RUNLIST_NOTIFICATION		4
#define TU104_INTR_CATEGORY_UVM_OWNED			5
#define TU104_INTR_CATEGORY_UVM_SHARED			6
#define TU104_INTR_CATEGORY_ENUM_COUNT			7
#define TU104_INTR_INVALID_SUBTREE			0xFF

struct tu104_intr_get_kernel_table_params {
	u32 tableLen;
	struct tu104_intr_table_entry table[TU104_INTR_TABLE_MAX_SIZE];
	struct tu104_intr_subtree_map subtreeMap[TU104_INTR_CATEGORY_ENUM_COUNT];
};

/* NV2080_CTRL_GPU_CONSTRUCTED_FALCON_INFO (ctrl2080gpu.h) - 20 bytes */
struct tu104_constructed_falcon_info {
	u32 engDesc;
	u32 ctxAttr;
	u32 ctxBufferSize;
	u32 addrSpaceList;
	u32 registerBase;
};

#define TU104_MAX_CONSTRUCTED_FALCONS	0x40

struct tu104_get_constructed_falcon_info_params {
	u32 numConstructedFalcons;
	struct tu104_constructed_falcon_info constructedFalconsTable[TU104_MAX_CONSTRUCTED_FALCONS];
};

/* ============================================================
 * ENGINE_INFO_TYPE indices (engine_info.h) - frozen vGPU ABI
 * ============================================================ */
#define EI_ENG_DESC			0
#define EI_FIFO_TAG			1
#define EI_RM_ENGINE_TYPE		2
#define EI_RUNLIST			3
#define EI_MMU_FAULT_ID			4
#define EI_RC_MASK			5
#define EI_RESET			6
#define EI_INTR				7
#define EI_MC				8
#define EI_DEV_TYPE_ENUM		9
#define EI_INSTANCE_ID			10
#define EI_RUNLIST_PRI_BASE		11
#define EI_IS_HOST_DRIVEN_ENGINE	12
#define EI_RUNLIST_ENGINE_ID		13
#define EI_CHRAM_PRI_BASE		14
#define EI_KERNEL_RM_MAX		15

/* ============================================================
 * ENG_DESC values (MKENGDESC = classId << 8 | instance)
 * ============================================================ */
#define TU104_ENG_GR0		0xd334df00	/* classId(Graphics)=0xd334df, inst 0 */
#define TU104_ENG_CE0		0x793ceb00	/* classId(OBJCE)=0x793ceb, inst 0 */
#define TU104_ENG_CE1		0x793ceb01
#define TU104_ENG_CE2		0x793ceb02
#define TU104_ENG_CE3		0x793ceb03
#define TU104_ENG_CE4		0x793ceb04
#define TU104_ENG_CES		0x793ceb0e	/* classId(OBJCE) inst 14 (async) */
#define TU104_ENG_NVDEC0	0x8f99e100	/* classId(OBJBSP)=0x8f99e1, inst 0 */
#define TU104_ENG_NVDEC1	0x8f99e101
#define TU104_ENG_NVENC0	0xe97b6c00	/* classId(OBJMSENC)=0xe97b6c, inst 0 */
#define TU104_ENG_SEC2		0x28c40800	/* classId(OBJSEC2)=0x28c408, inst 0 */
#define TU104_ENG_NVJPG		0x2b3a5400	/* classId(OBJNVJPG)=0x2b3a54, inst 0 */
#define TU104_ENG_SW		0x95a6f500	/* classId(OBJSWENG)=0x95a6f5, inst 0 */
#define TU104_ENG_FBFLCN	0x8a20bf00	/* classId(OBJFBFLCN)=0x8a20bf, inst 0 */

/* ============================================================
 * RM_ENGINE_TYPE values (gpu_engine_type.h)
 * ============================================================ */
#define TU104_RM_ENGINE_GR0		0x01
#define TU104_RM_ENGINE_CE0		0x09
#define TU104_RM_ENGINE_CE1		0x0a
#define TU104_RM_ENGINE_CE2		0x0b
#define TU104_RM_ENGINE_CE3		0x0c
#define TU104_RM_ENGINE_CE4		0x0d
#define TU104_RM_ENGINE_CES		0x17	/* RM_ENGINE_TYPE_COPY14 */
#define TU104_RM_ENGINE_NVDEC0		0x1d
#define TU104_RM_ENGINE_NVDEC1		0x1e
#define TU104_RM_ENGINE_NVENC0		0x25
#define TU104_RM_ENGINE_SW		0x2d
#define TU104_RM_ENGINE_SEC2		0x31
#define TU104_RM_ENGINE_HOST		0x32
#define TU104_RM_ENGINE_PMU		0x34
#define TU104_RM_ENGINE_FBFLCN		0x35
#define TU104_RM_ENGINE_NVJPG0		0x36

/* ============================================================
 * MC_ENGINE_IDX values (engine_idx.h)
 * ============================================================ */
#define TU104_MC_ENGINE_GR		84
#define TU104_MC_ENGINE_CE0		15
#define TU104_MC_ENGINE_CE1		16
#define TU104_MC_ENGINE_CE2		17
#define TU104_MC_ENGINE_CE3		18
#define TU104_MC_ENGINE_CE4		19
#define TU104_MC_ENGINE_CES		29	/* MC_ENGINE_IDX_CE14 */
#define TU104_MC_ENGINE_NVDEC0		65
#define TU104_MC_ENGINE_NVDEC1		66
#define TU104_MC_ENGINE_NVENC0		38
#define TU104_MC_ENGINE_SEC2		47
#define TU104_MC_ENGINE_NVJPG		51

/* ============================================================
 * NV_PFAULT_MMU_ENG_ID values (tu104/dev_fault.h)
 * ============================================================ */
#define TU104_FAULT_ID_GR		64
#define TU104_FAULT_ID_CE0		15
#define TU104_FAULT_ID_CE1		16
#define TU104_FAULT_ID_CE2		17
#define TU104_FAULT_ID_CE3		18
#define TU104_FAULT_ID_CE4		19
#define TU104_FAULT_ID_CES		29
#define TU104_FAULT_ID_NVDEC0		10
#define TU104_FAULT_ID_NVDEC1		25
#define TU104_FAULT_ID_NVENC0		11
#define TU104_FAULT_ID_SEC2		14
#define TU104_FAULT_ID_NVJPG0		24
#define TU104_FAULT_ID_GSP		2
#define TU104_FAULT_ID_PMU		6

/* PBDMA fault ID base (NV_PFIFO_CFG0_PBDMA_FAULT_ID_INIT = 32) */
#define TU104_PBDMA_FAULT_ID_BASE	32
#define TU104_NUM_PBDMAS		12

/* ============================================================
 * ROBUST_CHANNEL values (rc.h)
 * ============================================================ */
#define TU104_RC_GRAPHICS		40
#define TU104_RC_CE0			22
#define TU104_RC_CE1			23
#define TU104_RC_CE2			24
#define TU104_RC_CE3			25
#define TU104_RC_CE4			26
#define TU104_RC_NVDEC0			14
#define TU104_RC_NVDEC1			15
#define TU104_RC_NVENC0			11
#define TU104_RC_SEC2			2
#define TU104_RC_NVJPG0			32

/* ============================================================
 * NV_PTOP_DEVICE_INFO_TYPE_ENUM (tu104/dev_top.h)
 * ============================================================ */
#define TU104_PTOP_TYPE_GRAPHICS	0
#define TU104_PTOP_TYPE_LCE		19
#define TU104_PTOP_TYPE_NVDEC		16
#define TU104_PTOP_TYPE_NVENC		14
#define TU104_PTOP_TYPE_NVJPG		21
#define TU104_PTOP_TYPE_SEC		13
#define TU104_PTOP_TYPE_GSP		20
#define TU104_PTOP_TYPE_IOCTRL		18

/* ============================================================
 * FIFO_TAG values (hardware engine enum, 4 bits)
 * ============================================================ */
#define TU104_TAG_GR			0
#define TU104_TAG_CE0			1
#define TU104_TAG_CE1			2
#define TU104_TAG_CE2			3
#define TU104_TAG_CE3			4
#define TU104_TAG_CE4			5
#define TU104_TAG_CES			6
#define TU104_TAG_NVDEC0		10
#define TU104_TAG_NVDEC1		11
#define TU104_TAG_NVENC0		13
#define TU104_TAG_NVJPG			14
#define TU104_TAG_SEC2			15

/* ============================================================
 * TU104 FIFO Engine Table (11 hardware engines + 1 SW)
 *
 * Runlist assignment: each engine gets its own runlist.
 * Runlist 0=GR, 1-5=CE0-CE4, 6-7=NVDEC0-1, 8=NVENC0, 9=SEC2, 10=NVJPG
 * PBDMA assignment: 1:1 with runlists (PBDMA N services runlist N).
 * SW engine shares NVJPG's runlist/PBDMA (per fifoConstructSwEngine).
 * ============================================================ */

static const struct tu104_fifo_device_entry tu104_engine_table[] = {
	/* [0] GR0 - Graphics */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_GR0,
			[EI_FIFO_TAG]       = TU104_TAG_GR,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_GR0,
			[EI_RUNLIST]        = 0,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_GR,
			[EI_RC_MASK]        = TU104_RC_GRAPHICS,
			[EI_RESET]          = 12,	/* NV_PMC_ENABLE_PGRAPH bit */
			[EI_INTR]           = 0,
			[EI_MC]             = TU104_MC_ENGINE_GR,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_GRAPHICS,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {0, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 0, 0},
		.numPbdmas     = 1,
		.engineName    = "GRAPHICS",
	},
	/* [1] CE0 - GR-synchronous CE (GRCE 0); shares the GR PBDMA */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_CE0,
			[EI_FIFO_TAG]       = TU104_TAG_CE0,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_CE0,
			[EI_RUNLIST]        = 0,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_CE0,
			[EI_RC_MASK]        = TU104_RC_CE0,
			[EI_RESET]          = 1,
			[EI_INTR]           = 1,
			[EI_MC]             = TU104_MC_ENGINE_CE0,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_LCE,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {0, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 0, 0},
		.numPbdmas     = 1,
		.engineName    = "CE0",
	},
	/* [2] CE1 - GR-synchronous CE (GRCE 1); shares GR's runlist/PBDMA like CE0 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_CE1,
			[EI_FIFO_TAG]       = TU104_TAG_CE1,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_CE1,
			[EI_RUNLIST]        = 0,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_CE1,
			[EI_RC_MASK]        = TU104_RC_CE1,
			[EI_RESET]          = 2,
			[EI_INTR]           = 2,
			[EI_MC]             = TU104_MC_ENGINE_CE1,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_LCE,
			[EI_INSTANCE_ID]    = 1,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {0, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 0, 0},
		.numPbdmas     = 1,
		.engineName    = "CE1",
	},
	/* [3] CE2 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_CE2,
			[EI_FIFO_TAG]       = TU104_TAG_CE2,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_CE2,
			[EI_RUNLIST]        = 2,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_CE2,
			[EI_RC_MASK]        = TU104_RC_CE2,
			[EI_RESET]          = 3,
			[EI_INTR]           = 3,
			[EI_MC]             = TU104_MC_ENGINE_CE2,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_LCE,
			[EI_INSTANCE_ID]    = 2,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {2, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 2, 0},
		.numPbdmas     = 1,
		.engineName    = "CE2",
	},
	/* [4] CE3 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_CE3,
			[EI_FIFO_TAG]       = TU104_TAG_CE3,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_CE3,
			[EI_RUNLIST]        = 3,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_CE3,
			[EI_RC_MASK]        = TU104_RC_CE3,
			[EI_RESET]          = 4,
			[EI_INTR]           = 4,
			[EI_MC]             = TU104_MC_ENGINE_CE3,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_LCE,
			[EI_INSTANCE_ID]    = 3,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {3, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 3, 0},
		.numPbdmas     = 1,
		.engineName    = "CE3",
	},
	/* [5] CE4 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_CE4,
			[EI_FIFO_TAG]       = TU104_TAG_CE4,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_CE4,
			[EI_RUNLIST]        = 4,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_CE4,
			[EI_RC_MASK]        = TU104_RC_CE4,
			[EI_RESET]          = 5,
			[EI_INTR]           = 5,
			[EI_MC]             = TU104_MC_ENGINE_CE4,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_LCE,
			[EI_INSTANCE_ID]    = 4,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {4, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 4, 0},
		.numPbdmas     = 1,
		.engineName    = "CE4",
	},
	/* [6] CES - async copy engine (RM COPY14 / OBJCE instance 14).
	 * Clients (libcuda) refuse to finish context creation without it: the
	 * async/secondary engine record stays empty and the engine-context RM
	 * alloc is issued with class 0. Not on a runlist of its own; shares the
	 * CE runlist. */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_CES,
			[EI_FIFO_TAG]       = TU104_TAG_CES,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_CES,
			[EI_RUNLIST]        = 5,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_CES,
			[EI_RC_MASK]        = TU104_RC_CE4,
			[EI_RESET]          = 5,
			[EI_INTR]           = 5,
			[EI_MC]             = TU104_MC_ENGINE_CES,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_LCE,
			[EI_INSTANCE_ID]    = 14,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {5, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 5, 0},
		.numPbdmas     = 1,
		.engineName    = "CES",
	},
	/* [7] NVDEC0 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_NVDEC0,
			[EI_FIFO_TAG]       = TU104_TAG_NVDEC0,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_NVDEC0,
			[EI_RUNLIST]        = 6,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_NVDEC0,
			[EI_RC_MASK]        = TU104_RC_NVDEC0,
			[EI_RESET]          = 14,
			[EI_INTR]           = 6,
			[EI_MC]             = TU104_MC_ENGINE_NVDEC0,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_NVDEC,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {6, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 6, 0},
		.numPbdmas     = 1,
		.engineName    = "NVDEC0",
	},
	/* [7] NVDEC1 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_NVDEC1,
			[EI_FIFO_TAG]       = TU104_TAG_NVDEC1,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_NVDEC1,
			[EI_RUNLIST]        = 7,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_NVDEC1,
			[EI_RC_MASK]        = TU104_RC_NVDEC1,
			[EI_RESET]          = 15,
			[EI_INTR]           = 7,
			[EI_MC]             = TU104_MC_ENGINE_NVDEC1,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_NVDEC,
			[EI_INSTANCE_ID]    = 1,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {7, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 7, 0},
		.numPbdmas     = 1,
		.engineName    = "NVDEC1",
	},
	/* [8] NVENC0 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_NVENC0,
			[EI_FIFO_TAG]       = TU104_TAG_NVENC0,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_NVENC0,
			[EI_RUNLIST]        = 8,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_NVENC0,
			[EI_RC_MASK]        = TU104_RC_NVENC0,
			[EI_RESET]          = 17,
			[EI_INTR]           = 8,
			[EI_MC]             = TU104_MC_ENGINE_NVENC0,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_NVENC,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {8, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 8, 0},
		.numPbdmas     = 1,
		.engineName    = "NVENC0",
	},
	/* [9] SEC2 */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_SEC2,
			[EI_FIFO_TAG]       = TU104_TAG_SEC2,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_SEC2,
			[EI_RUNLIST]        = 9,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_SEC2,
			[EI_RC_MASK]        = TU104_RC_SEC2,
			[EI_RESET]          = 19,
			[EI_INTR]           = 9,
			[EI_MC]             = TU104_MC_ENGINE_SEC2,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_SEC,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {9, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 9, 0},
		.numPbdmas     = 1,
		.engineName    = "SEC2",
	},
	/* [10] NVJPG */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_NVJPG,
			[EI_FIFO_TAG]       = TU104_TAG_NVJPG,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_NVJPG0,
			[EI_RUNLIST]        = 10,
			[EI_MMU_FAULT_ID]   = TU104_FAULT_ID_NVJPG0,
			[EI_RC_MASK]        = TU104_RC_NVJPG0,
			[EI_RESET]          = 20,
			[EI_INTR]           = 10,
			[EI_MC]             = TU104_MC_ENGINE_NVJPG,
			[EI_DEV_TYPE_ENUM]  = TU104_PTOP_TYPE_NVJPG,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 1,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {10, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 10, 0},
		.numPbdmas     = 1,
		.engineName    = "NVJPG",
	},
	/* [11] SW (software engine - shares NVJPG's runlist per fifoConstructSwEngine) */
	{
		.engineData = {
			[EI_ENG_DESC]       = TU104_ENG_SW,
			[EI_FIFO_TAG]       = 0xffffffff,
			[EI_RM_ENGINE_TYPE] = TU104_RM_ENGINE_SW,
			[EI_RUNLIST]        = 10,	/* NVJPG's runlist */
			[EI_MMU_FAULT_ID]   = 0xffffffff,
			[EI_RC_MASK]        = 0,
			[EI_RESET]          = 0,
			[EI_INTR]           = 0,
			[EI_MC]             = 0,
			[EI_DEV_TYPE_ENUM]  = 0,
			[EI_INSTANCE_ID]    = 0,
			[EI_RUNLIST_PRI_BASE]     = 0,
			[EI_IS_HOST_DRIVEN_ENGINE] = 0,
			[EI_RUNLIST_ENGINE_ID]    = 0,
			[EI_CHRAM_PRI_BASE]       = 0,
			[EI_KERNEL_RM_MAX]        = 0,
		},
		.pbdmaIds      = {10, 0},
		.pbdmaFaultIds = {TU104_PBDMA_FAULT_ID_BASE + 10, 0},
		.numPbdmas     = 1,
		.engineName    = "SOFTWARE",
	},
};

#define TU104_ENGINE_TABLE_SIZE	13

/* ============================================================
 * TU104 Interrupt Table (from intr_phys_tu102.c, kernel table)
 *
 * These are the CPU-RM (guest) visible interrupts.
 * NV_PMC_INTR_INVALID_MASK = 0 (no PMC mask for these).
 * NV_INTR_VECTOR_INVALID = 0xFFFFFFFF.
 * ============================================================ */

#define TU104_NV_INTR_VECTOR_INVALID	0xFFFFFFFF

static const struct tu104_intr_table_entry tu104_intr_kernel_table[] = {
	/* MC_ENGINE_IDX_REPLAYABLE_FAULT = 59 */
	{ .engineIdx = 59, .pmcIntrMask = 0,
	  .vectorStall = 64,    /* NV_PFB_PRI_MMU_INT_VECTOR_FAULT_NOTIFY_REPLAYABLE */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_REPLAYABLE_FAULT_ERROR = 62 */
	{ .engineIdx = 62, .pmcIntrMask = 0,
	  .vectorStall = 131,   /* NV_PFB_PRI_MMU_INT_VECTOR_FAULT_ERROR_REPLAYABLE */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_ACCESS_CNTR = 60 */
	{ .engineIdx = 60, .pmcIntrMask = 0,
	  .vectorStall = 72,    /* HUB_ACCESS_CNTR_INTR_VECTOR */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_CPU_DOORBELL = 73 */
	{ .engineIdx = 73, .pmcIntrMask = 0,
	  .vectorStall = 129,   /* NV_CTRL_CPU_DOORBELL_VECTORID_VALUE_CONSTANT */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_PRIV_DOORBELL = 74 */
	{ .engineIdx = 74, .pmcIntrMask = 0,
	  .vectorStall = 130,   /* NV_CTRL_VIRTUAL_INTR_VECTORID_VALUE_CONSTANT */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_GR_FECS_LOG = 156 */
	{ .engineIdx = 156, .pmcIntrMask = 0,
	  .vectorStall = TU104_NV_INTR_VECTOR_INVALID,
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_GSP = 50 */
	{ .engineIdx = 50, .pmcIntrMask = 0,
	  .vectorStall = 155,   /* GSP_INTR_0_VECTOR */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
	/* MC_ENGINE_IDX_DISP = 2 */
	{ .engineIdx = 2, .pmcIntrMask = 0,
	  .vectorStall = 154,   /* PDISP_VECTOR */
	  .vectorNonStall = TU104_NV_INTR_VECTOR_INVALID },
};

#define TU104_INTR_KERNEL_TABLE_SIZE	8

/* ============================================================
 * TU104 Constructed Falcon Table
 *
 * Only FBFLCN (framebuffer falcon) is a "constructed falcon" on Turing
 * (SEC2, GSP, PMU are skipped as they have their own kernel objects).
 * ============================================================ */

static const struct tu104_constructed_falcon_info tu104_falcon_table[] = {
	{
		.engDesc       = TU104_ENG_FBFLCN,
		.ctxAttr       = 0,
		.ctxBufferSize = 0x400,		/* 1KB context buffer */
		.addrSpaceList = 0,
		.registerBase  = 0x009a4000,	/* NV_PFBFALCON base */
	},
};

#define TU104_FALCON_TABLE_SIZE	1

/* ============================================================
 * TU104 Internal Device Info Table
 *
 * Engines + key non-engine devices. The guest driver queries this
 * by typeEnum to find device register bases and fault IDs.
 * ============================================================ */

static const struct tu104_internal_device_info tu104_internal_device_table[] = {
	/* Engines (isEngine=1) */
	{ .faultId = TU104_FAULT_ID_GR,    .instanceId = 0, .typeEnum = TU104_PTOP_TYPE_GRAPHICS, .resetId = 12, .devicePriBase = 0x00400000, .isEngine = 1, .rlEngId = 0,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 0 },
	{ .faultId = TU104_FAULT_ID_CE0,   .instanceId = 0, .typeEnum = TU104_PTOP_TYPE_LCE,      .resetId = 1,  .devicePriBase = 0x00040000, .isEngine = 1, .rlEngId = 1,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 0 },
	{ .faultId = TU104_FAULT_ID_CE1,   .instanceId = 1, .typeEnum = TU104_PTOP_TYPE_LCE,      .resetId = 2,  .devicePriBase = 0x00041000, .isEngine = 1, .rlEngId = 2,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 1 },
	{ .faultId = TU104_FAULT_ID_CE2,   .instanceId = 2, .typeEnum = TU104_PTOP_TYPE_LCE,      .resetId = 3,  .devicePriBase = 0x00042000, .isEngine = 1, .rlEngId = 3,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 2 },
	{ .faultId = TU104_FAULT_ID_CE3,   .instanceId = 3, .typeEnum = TU104_PTOP_TYPE_LCE,      .resetId = 4,  .devicePriBase = 0x00043000, .isEngine = 1, .rlEngId = 4,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 3 },
	{ .faultId = TU104_FAULT_ID_CE4,   .instanceId = 4, .typeEnum = TU104_PTOP_TYPE_LCE,      .resetId = 5,  .devicePriBase = 0x00044000, .isEngine = 1, .rlEngId = 5,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 4 },
	{ .faultId = TU104_FAULT_ID_NVDEC0, .instanceId = 0, .typeEnum = TU104_PTOP_TYPE_NVDEC,   .resetId = 14, .devicePriBase = 0x00060000, .isEngine = 1, .rlEngId = 6,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 0 },
	{ .faultId = TU104_FAULT_ID_NVDEC1, .instanceId = 1, .typeEnum = TU104_PTOP_TYPE_NVDEC,   .resetId = 15, .devicePriBase = 0x00061000, .isEngine = 1, .rlEngId = 7,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 1 },
	{ .faultId = TU104_FAULT_ID_NVENC0, .instanceId = 0, .typeEnum = TU104_PTOP_TYPE_NVENC,   .resetId = 17, .devicePriBase = 0x00070000, .isEngine = 1, .rlEngId = 8,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 0 },
	{ .faultId = TU104_FAULT_ID_SEC2,  .instanceId = 0, .typeEnum = TU104_PTOP_TYPE_SEC,      .resetId = 19, .devicePriBase = 0x00840000, .isEngine = 1, .rlEngId = 9,  .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 0 },
	{ .faultId = TU104_FAULT_ID_NVJPG0, .instanceId = 0, .typeEnum = TU104_PTOP_TYPE_NVJPG,   .resetId = 20, .devicePriBase = 0x00080000, .isEngine = 1, .rlEngId = 10, .runlistPriBase = 0, .groupId = 0, .ginTargetId = 0, .deviceBroadcastPriBase = 0, .groupLocalInstanceId = 0 },
};

#define TU104_INTERNAL_DEVICE_TABLE_SIZE	11

/* ============================================================
 * GspStaticConfigInfo field offsets (ABI-verified, 1656 bytes total)
 * ============================================================ */
#define TU104_GSCI_OFF_grCapsBits		0	/* 23 bytes */
#define TU104_GSCI_OFF_gidInfo			24	/* 268 bytes */
#define TU104_GSCI_OFF_SKUInfo			292	/* 48 bytes */
#define TU104_GSCI_OFF_fbRegionInfoParams	344	/* 776 bytes */
#define TU104_GSCI_OFF_sriovCaps		1120	/* 80 bytes */
#define TU104_GSCI_OFF_sriovMaxGfid		1200	/* 4 bytes */
#define TU104_GSCI_OFF_engineCaps		1204	/* 12 bytes */
#define TU104_GSCI_OFF_poisonFuseEnabled	1216	/* 1 byte */
#define TU104_GSCI_OFF_fb_length		1224	/* 8 bytes */
#define TU104_GSCI_OFF_fbio_mask		1232	/* 8 bytes */
#define TU104_GSCI_OFF_fb_bus_width		1240	/* 4 bytes */
#define TU104_GSCI_OFF_fb_ram_type		1244	/* 4 bytes */
#define TU104_GSCI_OFF_fbp_mask			1248	/* 8 bytes */
#define TU104_GSCI_OFF_l2_cache_size		1256	/* 4 bytes */
#define TU104_GSCI_OFF_gpuNameString		1260	/* 64 bytes */
#define TU104_GSCI_OFF_gpuShortNameString	1324	/* 64 bytes */
#define TU104_GSCI_OFF_gpuNameString_Unicode	1388	/* 128 bytes */
#define TU104_GSCI_OFF_bIsMobile		1524	/* 1 byte */
#define TU104_GSCI_OFF_bar1PdeBase		1536	/* 8 bytes */
#define TU104_GSCI_OFF_bar2PdeBase		1544	/* 8 bytes */
#define TU104_GSCI_OFF_bVbiosValid		1552	/* 1 byte */
#define TU104_GSCI_OFF_vbiosSubVendor		1556	/* 4 bytes */
#define TU104_GSCI_OFF_vbiosSubDevice		1560	/* 4 bytes */
#define TU104_GSCI_OFF_hInternalClient		1600	/* 4 bytes */
#define TU104_GSCI_OFF_hInternalDevice		1604	/* 4 bytes */
#define TU104_GSCI_OFF_hInternalSubdevice	1608	/* 4 bytes */
#define TU104_GSCI_OFF_ecidInfo			1616	/* 24 bytes */
#define TU104_GSCI_OFF_fwWprLayoutOffset	1640	/* 16 bytes */
#define TU104_GSCI_SIZE				1656

/* Internal handle values (g_gpu_nvoc.h, g_resserv_nvoc.h) */
#define TU104_H_INTERNAL_CLIENT			0xC1E00000
#define TU104_H_INTERNAL_DEVICE			0xABCD0080
#define TU104_H_INTERNAL_SUBDEVICE		0xABCD2080

/* GSP_RM_CONTROL sub-commands */
#define TU104_CTRL_INTERNAL_GET_DEVICE_INFO_TABLE	0x20800a40
#define TU104_CTRL_FIFO_GET_DEVICE_INFO_TABLE	0x20801112
#define TU104_CTRL_INTERNAL_INTR_GET_KERNEL_TABLE	0x20800a5c
#define TU104_CTRL_GPU_GET_CONSTRUCTED_FALCON_INFO	0x208001b0
#define TU104_CTRL_INTERNAL_FIFO_GET_NUM_CHANNELS	0x20800a61
#define TU104_CTRL_BUS_GET_PCI_BAR_INFO		0x20801803
#define TU104_CTRL_INTERNAL_DISPLAY_GET_STATIC_INFO	0x20800a01
#define TU104_CTRL_INTERNAL_DISPLAY_GET_IP_VERSION	0x20800a4b
#define TU104_CTRL_I2C_TRANSACTION			0x402c0105
#define TU104_CTRL_CE_GET_FAULT_METHOD_BUFFER_SIZE	0x20802a08
#define TU104_CTRL_CE_GET_PHYSICAL_CAPS			0x20802a07
#define TU104_CTRL_CE_GET_ALL_PHYSICAL_CAPS		0x20802a0b
#define TU104_CTRL_CE_GET_ALL_CAPS			0x20802a0a
#define TU104_CTRL_CE_GET_CE_PCE_MASK			0x20802a02

/* CE capability bits (NV2080_CTRL_CE_CAPS_*, ctrl2080ce.h: byte:bit) */
#define TU104_CE_CAPS0_GRCE		0x01	/* synchronous with GR (GRCE) */
#define TU104_CE_CAPS0_SYSMEM_READ	0x04
#define TU104_CE_CAPS0_SYSMEM_WRITE	0x08
#define TU104_CE_CAPS0_SYSMEM		0x20
#define TU104_CE_CAPS0_P2P		0x40
#define TU104_CE_CAPS0_BL_GT_64K	0x80
#define TU104_CE_CAPS0		(TU104_CE_CAPS0_SYSMEM_READ | \
				 TU104_CE_CAPS0_SYSMEM_WRITE | \
				 TU104_CE_CAPS0_SYSMEM | \
				 TU104_CE_CAPS0_P2P | \
				 TU104_CE_CAPS0_BL_GT_64K)
#define TU104_CE_CAPS1		0x00

/* ENGINE_TYPE_COPY0 = 0x9, COPY(i) = COPY0 + i (cl2080_notification.h) */
#define TU104_ENGINE_TYPE_COPY0	0x00000009
#define TU104_NUM_CES		5
/* The GR-synchronous CEs (GRCEs): the client picks the last GRCE as its
 * compute companion ("ceForCompute") and every non-GRCE CE becomes an
 * asynchronous CE. Measured on the same silicon: the host TU104 advertises
 * ASYNC_ENGINE_COUNT=3 out of 5 CEs, i.e. exactly two GRCEs (CE0, CE1).
 * With only one GRCE the client's context walk ends with an empty
 * secondary-engine record and the engine RM alloc is issued with class 0. */
#define TU104_GRCE_MASK		((1u << 0) | (1u << 1))
#define TU104_IS_GRCE(idx)	(!!(TU104_GRCE_MASK & (1u << (idx))))
#define TU104_CTRL_INTERNAL_GMMU_GET_STATIC_INFO	0x20800a59
#define TU104_CTRL_STATIC_KGR_GET_GLOBAL_SM_ORDER	0x20800a22
#define TU104_CTRL_STATIC_KGR_GET_INFO		0x20800a2a
#define TU104_CTRL_STATIC_KGR_GET_CONTEXT_BUFFERS_INFO	0x20800a32

/* NV2080_CTRL_INTERNAL_ENGINE_CONTEXT_BUFFER_INFO (engine context buffer info)
 * NV2080_CTRL_INTERNAL_STATIC_GR_CONTEXT_BUFFERS_INFO: engine[ENGINE_ID_COUNT]
 * NV2080_CTRL_INTERNAL_STATIC_GR_GET_CONTEXT_BUFFERS_INFO_PARAMS:
 *     engineContextBuffersInfo[GR_MAX_ENGINES]
 *
 * The driver's kgraphicsInitializeDeferredStaticData() asks the physical RM
 * (GSP-RM) for this table; kgrctxAllocMainCtxBuffer() then sizes the golden
 * image ctx buffer from engine[GRAPHICS].size. A zeroed reply produces a
 * 0-byte (0-page) allocation -> NV_ERR_NO_MEMORY (mem_desc.c:1353).
 *
 * The physical RM probes these sizes from the ctxsw ucode at GR init
 * (grSetupInitialCtxState -> grDiscoverCtxBufferSize) and reports
 * NV_U32_MAX for engine IDs the chip does not support. The values below are
 * the emulated TU104 layout (own buffers, no real GR engine behind them):
 * sizes are page-aligned scratch regions, NV_U32_MAX = "not supported". */
#define TU104_CTX_ENGINE_ID_COUNT	26	/* _ENGINE_ID_COUNT (0x1a) */
#define TU104_CTX_GPU_MAX_ENGINES	8	/* INTERNAL_GR_MAX_ENGINES */

/* NV0080_CTRL_FIFO_GET_ENGINE_CONTEXT_PROPERTIES_ENGINE_ID_* */
#define TU104_CTX_ENG_GRAPHICS			0x00
#define TU104_CTX_ENG_VLD			0x01
#define TU104_CTX_ENG_VIDEO			0x02
#define TU104_CTX_ENG_MPEG			0x03
#define TU104_CTX_ENG_CAPTURE			0x04
#define TU104_CTX_ENG_DISPLAY			0x05
#define TU104_CTX_ENG_ENCRYPTION		0x06
#define TU104_CTX_ENG_POSTPROCESS		0x07
#define TU104_CTX_ENG_GRAPHICS_ZCULL		0x08
#define TU104_CTX_ENG_GRAPHICS_PM		0x09
#define TU104_CTX_ENG_COMPUTE_PREEMPT		0x0a
#define TU104_CTX_ENG_GRAPHICS_PREEMPT		0x0b
#define TU104_CTX_ENG_GRAPHICS_SPILL		0x0c
#define TU104_CTX_ENG_GRAPHICS_PAGEPOOL		0x0d
#define TU104_CTX_ENG_GRAPHICS_BETACB		0x0e
#define TU104_CTX_ENG_GRAPHICS_RTV		0x0f
#define TU104_CTX_ENG_GRAPHICS_PATCH		0x10
#define TU104_CTX_ENG_GRAPHICS_BUNDLE_CB	0x11
#define TU104_CTX_ENG_GRAPHICS_PAGEPOOL_GLOBAL	0x12
#define TU104_CTX_ENG_GRAPHICS_ATTRIBUTE_CB	0x13
#define TU104_CTX_ENG_GRAPHICS_RTV_CB_GLOBAL	0x14
#define TU104_CTX_ENG_GRAPHICS_GFXP_POOL	0x15
#define TU104_CTX_ENG_GRAPHICS_GFXP_CTRL_BLK	0x16
#define TU104_CTX_ENG_GRAPHICS_FECS_EVENT	0x17
#define TU104_CTX_ENG_GRAPHICS_PRIV_ACCESS_MAP	0x18
#define TU104_CTX_ENG_GRAPHICS_SETUP		0x19

/* grabbag: emulated ctx buffer sizes (bytes) */
#define TU104_CTX_SIZE_MAIN			0x200000
#define TU104_CTX_SIZE_ZCULL			0x100000
#define TU104_CTX_SIZE_PM			0x8000
#define TU104_CTX_SIZE_PREEMPT			0x1000
#define TU104_CTX_SIZE_SPILL			0x1000
#define TU104_CTX_SIZE_PAGEPOOL			0x10000
#define TU104_CTX_SIZE_BETACB			0x1000
#define TU104_CTX_SIZE_RTV			0x1000
#define TU104_CTX_SIZE_PATCH			0x2000
#define TU104_CTX_SIZE_BUNDLE_CB		0x1000
#define TU104_CTX_SIZE_ATTRIBUTE_CB		0x1000
#define TU104_CTX_SIZE_RTV_CB_GLOBAL		0x1000
#define TU104_CTX_SIZE_FECS_EVENT		0x1000
#define TU104_CTX_SIZE_PRIV_ACCESS_MAP		0x1000

struct tu104_ctx_buffer_info {
	u32 size;
	u32 alignment;
};

struct tu104_gr_context_buffers_info {
	struct tu104_ctx_buffer_info engine[TU104_CTX_ENGINE_ID_COUNT];
};

struct tu104_gr_get_context_buffers_info_params {
	struct tu104_gr_context_buffers_info
		engineContextBuffersInfo[TU104_CTX_GPU_MAX_ENGINES];
};

/* FB_GET_INFO_V2 (ctrl2080fb.h): the client sends a list of requested FB info
 * indices; the physical RM must echo each index back with its data.
 * _fbGetFbInfos() asserts pFbInfos[i].index == returned index, so a zeroed
 * reply fires kern_mem_sys_ctrl.c:911 asserts and leaves all FB info zero
 * (CUDA/UVMLite read total/usable memory, BAR1 size, ECC/LTC info from here). */
#define TU104_CTRL_FB_GET_INFO_V2	0x20801303
#define TU104_FB_INFO_MAX_LIST_SIZE	0x80

struct tu104_fb_info_entry {
	u32 index;
	u32 data;
};

struct tu104_fb_get_info_v2_params {
	u32 fbInfoListSize;
	struct tu104_fb_info_entry fbInfoList[TU104_FB_INFO_MAX_LIST_SIZE];
};

/* NV2080_CTRL_FB_INFO_INDEX_* (subset we populate) */
#define TU104_FB_INFO_IDX_TILE_REGION_COUNT		0x00
#define TU104_FB_INFO_IDX_COMPRESSION_SIZE		0x01
#define TU104_FB_INFO_IDX_PARTITION_COUNT		0x04
#define TU104_FB_INFO_IDX_BAR1_SIZE			0x05
#define TU104_FB_INFO_IDX_RAM_SIZE			0x07
#define TU104_FB_INFO_IDX_TOTAL_RAM_SIZE		0x08
#define TU104_FB_INFO_IDX_HEAP_SIZE			0x09
#define TU104_FB_INFO_IDX_MAPPABLE_HEAP_SIZE		0x0a
#define TU104_FB_INFO_IDX_BUS_WIDTH			0x0b
#define TU104_FB_INFO_IDX_RAM_TYPE			0x0d
#define TU104_FB_INFO_IDX_BANK_COUNT			0x0e
#define TU104_FB_INFO_IDX_FB_TAX_SIZE_KB		0x10
#define TU104_FB_INFO_IDX_LARGEST_FREE_REGION_SIZE_KB	0x12
#define TU104_FB_INFO_IDX_PARTITION_MASK		0x14
#define TU104_FB_INFO_IDX_HEAP_FREE			0x16
#define TU104_FB_INFO_IDX_RAM_LOCATION			0x17
#define TU104_FB_INFO_IDX_FBP_COUNT			0x19
#define TU104_FB_INFO_IDX_FBP_MASK			0x1a
#define TU104_FB_INFO_IDX_L2CACHE_SIZE			0x1b
#define TU104_FB_INFO_IDX_MEMORYINFO_VENDOR_ID		0x1c
#define TU104_FB_INFO_IDX_BAR1_AVAIL_SIZE		0x1d
#define TU104_FB_INFO_IDX_BAR1_MAX_CONTIG_AVAIL_SIZE	0x1f
#define TU104_FB_INFO_IDX_USABLE_RAM_SIZE		0x20
#define TU104_FB_INFO_IDX_LTC_COUNT			0x22
#define TU104_FB_INFO_IDX_LTS_COUNT			0x23
#define TU104_FB_INFO_IDX_PSEUDO_CHANNEL_MODE		0x25
#define TU104_FB_INFO_IDX_ALLOW_PAGE_RETIREMENT		0x2a
#define TU104_FB_INFO_IDX_LTC_MASK			0x2b

#define TU104_RAM_TYPE_GDDR6				0x11
#define TU104_RAM_LOCATION_GPU_DEDICATED		0x00
#define TU104_PSEUDO_CHANNEL_MODE_DISABLED		0x01
#define TU104_MEMORYINFO_VENDOR_ID_SAMSUNG		0x01
#define TU104_L2CACHE_SIZE				0x400000 /* TU104: 4MB */
#define TU104_LTC_COUNT					8
#define TU104_LTS_COUNT					16

/* BUS_GET_INFO_V2 (ctrl2080bus.h) - same flat index/data list pattern as FB.
 * UVM's gpuGetSysmemLinkInfo() reads SYSMEM_CONNECTION_TYPE (PCIE) and then
 * PCIE_GPU_LINK_CAPS for the link bandwidth; a zeroed reply makes
 * getPCIELinkRateMBps() fail with NV_ERR_INVALID_STATE. */
#define TU104_CTRL_BUS_GET_INFO_V2	0x20801823
#define TU104_BUS_INFO_MAX_LIST_SIZE	0x34

struct tu104_bus_info_entry {
	u32 index;
	u32 data;
};

struct tu104_bus_get_info_v2_params {
	u32 busInfoListSize;
	struct tu104_bus_info_entry busInfoList[TU104_BUS_INFO_MAX_LIST_SIZE];
};

/* NV2080_CTRL_BUS_INFO_INDEX_* (subset we populate) */
#define TU104_BUS_INFO_IDX_TYPE				0x00
#define TU104_BUS_INFO_IDX_INTLINE			0x01
#define TU104_BUS_INFO_IDX_PCIE_GPU_LINK_CAPS		0x03
#define TU104_BUS_INFO_IDX_PCIE_ROOT_LINK_CAPS		0x04
#define TU104_BUS_INFO_IDX_PCIE_UPSTREAM_LINK_CAPS	0x05
#define TU104_BUS_INFO_IDX_PCIE_DOWNSTREAM_LINK_CAPS	0x06
#define TU104_BUS_INFO_IDX_PCIE_GPU_LINK_CTRL_STATUS	0x07
#define TU104_BUS_INFO_IDX_PCIE_ROOT_LINK_CTRL_STATUS	0x08
#define TU104_BUS_INFO_IDX_PCIE_UPSTREAM_LINK_CTRL	0x09
#define TU104_BUS_INFO_IDX_PCIE_DOWNSTREAM_LINK_CTRL	0x0a
#define TU104_BUS_INFO_IDX_PCIE_GEN_INFO		0x2d
#define TU104_BUS_INFO_IDX_SYSMEM_CONNECTION_TYPE	0x33

#define TU104_BUS_INFO_TYPE_PCI_EXPRESS			0x03
#define TU104_BUS_INFO_SYSMEM_CONN_PCIE			0x00
/* PCIE link caps / GEN info layout: MAX_SPEED 3:0 (8000MBPS=3), MAX_WIDTH
 * 9:4 (lanes), ASPM 11:10, GEN 15:12, CURR_LEVEL 19:16, GPU_GEN 23:20
 * (all GENx = 2 for Gen3), SPEED_CHANGES 24. */
#define TU104_PCIE_LINK_CAP_GEN3_X16	(0x3u | (16u << 4) | \
					 (2u << 12) | (2u << 16) | \
					 (2u << 20))
/* LINK_CTRL_STATUS: ASPM 1:0 (0=disabled), LINK_SPEED 19:16,
 * LINK_WIDTH 25:20 (LINK_WIDTH_X16=0x10). */
#define TU104_PCIE_LINK_CTRL_STATUS_GEN3_X16	((0x3u << 16) | (0x10u << 20))


/* NV0080_CTRL_GR_INFO entry + infoList indices (ctrl0080gr.h) */
struct tu104_gr_info_entry {
	u32 index;
	u32 data;
};
#define TU104_GR_INFO_MAX_SIZE		58	/* NV0080_CTRL_GR_INFO_MAX_SIZE (0x3a) */
#define TU104_GR_INFO_MAX_ENGINES	8

struct tu104_gr_static_info {
	struct tu104_gr_info_entry infoList[TU104_GR_INFO_MAX_SIZE];
};

/* infoList indices we populate (ctrl0080gr.h) */
#define TU104_GR_INFO_IDX_LITTER_NUM_GPCS		0x14
#define TU104_GR_INFO_IDX_LITTER_NUM_TPC_PER_GPC	0x17
#define TU104_GR_INFO_IDX_GPU_CORE_COUNT		0x1d
#define TU104_GR_INFO_IDX_LITTER_NUM_SM_PER_TPC		0x20
#define TU104_GR_INFO_IDX_FAMILY_MAX_TPC_PER_GPC	0x2a
#define TU104_GR_INFO_IDX_MAX_SUBCONTEXT_COUNT		0x2c
#define TU104_GR_INFO_IDX_MAX_LEGACY_SUBCONTEXT_COUNT	0x2d
#define TU104_GR_INFO_IDX_MAX_PER_ENGINE_SUBCONTEXT_COUNT	0x2e

/* TU104 max subcontexts (NV_PGRAPH_PRI_FE_CHIP_DEF_INFO_MAX_VEID_COUNT_INIT) */
#define TU104_MAX_SUBCONTEXT_COUNT	64

/* NV2080_CTRL_INTERNAL_STATIC_GR_FLOORSWEEPING_MASKS (ctrl2080internal.h)
 * + STATIC_KGR_GET_FLOORSWEEPING_MASKS cmd = 0x20800a26.
 * Driver computes numGpcs from floorsweepingMasks.gpcMask - zero kills
 * krcWatchdogInit ("numGpcs > 0" assert -> RmInitAdapter bails). */
#define TU104_CTRL_STATIC_KGR_GET_FLOORSWEEPING_MASKS	0x20800a26

/* Missing RM_CONTROL commands observed in guest driver (570.144) */
#define TU104_CTRL_INTERNAL_MEMSYS_L2_INVALIDATE_EVICT	0x20800a6c
#define TU104_CTRL_INTERNAL_GR_GET_FECS_TRACE_HW_ENABLE	0x20800a38

/* NV0073 (SYSTEM) interface */
#define TU104_CTRL_SYSTEM_MAP_SHARED_DATA		0x730151
#define TU104_GR_MAX_GPC		12
#define TU104_GR_MAX_TPC_PER_GPC_COUNT	10

struct tu104_gr_floorsweeping_masks {
	u32 gpcMask;
	u32 tpcMask[TU104_GR_MAX_GPC];
	u32 tpcCount[TU104_GR_MAX_GPC];
	u32 physGpcMask;
	u32 mmuPerGpc[TU104_GR_MAX_GPC];
	u32 tpcToPesMap[TU104_GR_MAX_TPC_PER_GPC_COUNT];
	u32 numPesPerGpc[TU104_GR_MAX_GPC];
	u32 zcullMask[TU104_GR_MAX_GPC];
	u32 physGfxGpcMask;
	u32 numGfxTpc;
};

/* TU104 SM topology: 6 GPCs x 4 TPCs x 2 SMs = 48 SMs, 24 TPCs (RTX 2080 SUPER) */
#define TU104_NUM_GPC		6
#define TU104_TPC_PER_GPC	4
#define TU104_SM_PER_TPC	2
#define TU104_NUM_SM		(TU104_NUM_GPC * TU104_TPC_PER_GPC * TU104_SM_PER_TPC)
#define TU104_NUM_TPC		(TU104_NUM_GPC * TU104_TPC_PER_GPC)

/* NV2080_CTRL_INTERNAL_STATIC_GR_GLOBAL_SM_ORDER (ctrl2080internal.h) */
#define TU104_GR_MAX_SM		240
#define TU104_GR_MAX_ENGINES	8

struct tu104_gr_sm_id {
	u16 gpcId;
	u16 localTpcId;
	u16 localSmId;
	u16 globalTpcId;
	u16 virtualGpcId;
	u16 migratableTpcId;
};

struct tu104_gr_global_sm_order {
	struct tu104_gr_sm_id globalSmId[TU104_GR_MAX_SM];
	u16 numSm;
	u16 numTpc;
};

/* CE fault method buffer size (ceGetFaultMethodBufferSize_GV100 formula,
 * Turing inherits: 135 * (13*totalCes + 2), page-aligned; TU104 has 5 PCEs) */
#define TU104_CE_FAULT_METHOD_BUFFER_SIZE		12288

/* GMMU static info (kgmmuSetAndGetDefaultFaultBufferSize_TU102 semantics):
 * the hardware default is reported in NV_VIRTUAL_FUNCTION_PRIV_MMU_FAULT_
 * BUFFER_SIZE_VAL units of NVC369_BUF_SIZE (32B). UVM's faultbufConstruct
 * allocates replayableFaultBufferSize and fails with NV_ERR_NO_MEMORY when
 * it reads back 0. Turing defaults are 64KB replayable / 4KB legacy; both
 * must stay within NV2080_CTRL_INTERNAL_GMMU_FAULT_BUFFER_MAX_PAGES (256). */
#define TU104_GMMU_REPLAYABLE_FAULT_BUFFER_SIZE		0x10000
#define TU104_GMMU_NONREPLAYABLE_FAULT_BUFFER_SIZE	0x1000

/* Turing display IP version (DISP v04.00 per DispIpHal halspec) */
#define TU104_DISPLAY_IP_VERSION				0x04000000

/* Channels per runlist reported to the guest (consumer Turing convention) */
#define TU104_FIFO_NUM_CHANNELS_PER_RUNLIST	64

/* FB region info params: numFBRegions(4) + 16 × 48B entries = 772 → aligned 776 */
struct tu104_fb_region {
	u64 base;
	u64 limit;
	u64 reserved;
	u32 performance;
	u8  supportCompressed;
	u8  supportISO;
	u8  bProtected;
	u8  blackList[17];
	u8  _pad[6];
};

#endif /* _GSP_RM_TU104_H_ */