/* Copyright 2024 Joshua Bakita * SPDX-License-Identifier: MIT * * File outline: * - Runlist, preemption, and channel control (FIFO) * - Basic GPU information (MC) * - Detailed GPU information (PTOP, FUSE, and CE) * - PRAMIN, BAR1/2, and page table status * - Helper functions for nvdebug */ #include // For dev_get_drvdata() #include // For struct seq_file #include // For PDE_DATA() macro #include // For KERNEL_VERSION and LINUX_VERSION_CODE #include // Fully defined in include/nvgpu/gk20a.h. We only pass around pointers to // this, so declare as incomplete type to avoid pulling in the nvgpu headers. struct gk20a; /* Runlist Channel A timeslice group (TSG) is composed of channels. Each channel is a FIFO queue of GPU commands. These commands are typically queued from userspace. Prior to Volta, channels could also exist independent of a TSG. These are called "bare channels" in the Jetson nvgpu driver. `INST_PTR` points to a GPU Instance Block which contains FIFO states, virtual address space configuration for this context, and a pointer to the page tables. All channels in a TSG point to the same GPU Instance Block (?). "RUNQUEUE_SELECTOR determines to which runqueue the channel belongs, and thereby which PBDMA will run the channel. Increasing values select increasingly numbered PBDMA IDs serving the runlist. If the selector value exceeds the number of PBDMAs on the runlist, the hardware will silently reassign the channel to run on the first PBDMA as though RUNQUEUE_SELECTOR had been set to 0. (In current hardware, this is used by SCG on the graphics runlist only to determine which FE pipe should service a given channel. A value of 0 targets the first FE pipe, which can process all FE driven engines: Graphics, Compute, Inline2Memory, and TwoD. A value of 1 targets the second FE pipe, which can only process Compute work. Note that GRCE work is allowed on either runqueue." (NVIDIA) Note that it appears runqueue 1 is the default for CUDA work on the Jetson Xavier. ENTRY_TYPE (T) : type of this entry: ENTRY_TYPE_CHAN CHID (ID) : identifier of the channel to run (overlays ENTRY_ID) RUNQUEUE_SELECTOR (Q) : selects which PBDMA should run this channel if more than one PBDMA is supported by the runlist, additionally, "A value of 0 targets the first FE pipe, which can process all FE driven engines: Graphics, Compute, Inline2Memory, and TwoD. A value of 1 targets the second FE pipe, which can only process Compute work. Note that GRCE work is allowed on either runqueue.)" INST_PTR_LO : lower 20 bits of the 4k-aligned instance block pointer INST_PTR_HI : upper 32 bit of instance block pointer INST_TARGET (TGI) : aperture of the instance block USERD_PTR_LO : upper 24 bits of the low 32 bits, of the 512-byte-aligned USERD pointer USERD_PTR_HI : upper 32 bits of USERD pointer USERD_TARGET (TGU) : aperture of the USERD data structure Channels were around since at least Fermi, but were rearranged with Volta to add a USERD pointer, a longer INST pointer, and a runqueue selector flag. */ enum ENTRY_TYPE {ENTRY_TYPE_CHAN = 0, ENTRY_TYPE_TSG = 1}; enum INST_TARGET {TARGET_VID_MEM = 0, TARGET_SYS_MEM_COHERENT = 2, TARGET_SYS_MEM_NONCOHERENT = 3}; static inline const char *target_to_text(enum INST_TARGET t) { switch (t) { case TARGET_VID_MEM: return "VID_MEM"; case TARGET_SYS_MEM_COHERENT: return "SYS_MEM_COHERENT"; case TARGET_SYS_MEM_NONCOHERENT: return "SYS_MEM_NONCOHERENT"; default: return "INVALID"; } } // Support: Volta, Ampere, Turing struct gv100_runlist_chan { // 0:63 enum ENTRY_TYPE entry_type:1; uint32_t runqueue_selector:1; uint32_t padding:2; enum INST_TARGET inst_target:2; uint32_t padding2:2; uint32_t userd_ptr_lo:24; uint32_t userd_ptr_hi:32; // 64:128 uint32_t chid:12; uint32_t inst_ptr_lo:20; uint32_t inst_ptr_hi:32; } __attribute__((packed)); // Support: Fermi, Kepler*, Maxwell, Pascal // *In Kepler, inst fields may be unpopulated? struct gm107_runlist_chan { uint32_t chid:12; uint32_t padding0:1; enum ENTRY_TYPE entry_type:1; uint32_t padding1:18; uint32_t inst_ptr_lo:20; enum INST_TARGET inst_target:2; // Totally guessing on this uint32_t padding2:10; } __attribute__((packed)); #define gk110_runlist_chan gm107_runlist_chan /* Runlist TSG (TimeSlice Group) The runlist is composed of timeslice groups (TSG). Each TSG corresponds to a single virtual address space on the GPU and contains `TSG_LENGTH` channels. These channels and virtual address space are accessible to the GPU host unit for use until the timeslice expires or a TSG switch is forcibly initiated via a write to `NV_PFIFO_PREEMPT`. timeslice = (TSG_TIMESLICE_TIMEOUT << TSG_TIMESLICE_SCALE) * 1024 nanoseconds ENTRY_TYPE (T) : type of this entry: ENTRY_TYPE_TSG TIMESLICE_SCALE : scale factor for the TSG's timeslice TIMESLICE_TIMEOUT : timeout amount for the TSG's timeslice TSG_LENGTH : number of channels that are part of this timeslice group TSGID : identifier of the Timeslice group (overlays ENTRY_ID) TSGs appear to have been introduced with Kepler and stayed the same until they were rearranged at the time of channel rearrangement to support longer GPU instance addresses with Volta. */ // Support: Volta, Turing*, Ampere* // *These treat bits 4:11 (8 bits) as GFID (unused) struct gv100_runlist_tsg { // 0:63 enum ENTRY_TYPE entry_type:1; uint64_t padding:15; uint32_t timeslice_scale:4; uint64_t padding2:4; uint32_t timeslice_timeout:8; uint32_t tsg_length:8; uint32_t padding3:24; // 64:128 uint32_t tsgid:12; uint64_t padding4:52; } __attribute__((packed)); #define MAX_TSGID (1 << 12) // Support: Kepler (v2?), Maxwell, Pascal // Same fields as Volta except tsg_length is 6 bits rather than 8 // Last 32 bits appear to contain an undocumented inst ptr struct gk110_runlist_tsg { uint32_t tsgid:12; uint32_t padding0:1; enum ENTRY_TYPE entry_type:1; uint32_t timeslice_scale:4; uint32_t timeslice_timeout:8; uint32_t tsg_length:6; uint32_t padding1:32; } __attribute__((packed)); enum PREEMPT_TYPE {PREEMPT_TYPE_CHANNEL = 0, PREEMPT_TYPE_TSG = 1}; /* Preempt a TSG or Channel by ID ID/CHID : Id of TSG or channel to preempt IS_PENDING : Is a context switch pending? (read-only) TYPE : PREEMPT_TYPE_CHANNEL or PREEMPT_TYPE_TSG Support: Kepler, Maxwell, Pascal, Volta, Turing */ #define NV_PFIFO_PREEMPT 0x00002634 typedef union { struct { uint32_t id:12; uint32_t padding:8; bool is_pending:1; uint32_t padding2:3; enum PREEMPT_TYPE type:2; uint32_t padding3:6; } __attribute__((packed)); uint32_t raw; } pfifo_preempt_t; /* "Initiate a preempt of the engine by writing the bit associated with its runlist to NV_PFIFO_RUNLIST_PREEMPT... Do not poll NV_PFIFO_RUNLIST_PREEMPT for the preempt to complete." Useful for preempting multiple runlists at once. Appears to trigger an interrupt or some other side-effect on the Jetson Xavier, as the built-in nvgpu driver seems to be disturbed by writing to this. To select the runlist dynamically, use the BIT(nr) kernel macro. Example: runlist_preempt_t rl_preempt; rl_preempt.raw = nvdebug_readl(g, NV_PFIFO_RUNLIST_PREEMPT); rl_preempt.raw |= BIT(nr); nvdebug_writel(g, NV_PFIFO_RUNLIST_PREEMPT, rl_preempt.raw); Support: Volta, Turing */ #define NV_PFIFO_RUNLIST_PREEMPT 0x00002638 typedef union { struct { bool runlist_0:1; bool runlist_1:1; bool runlist_2:1; bool runlist_3:1; bool runlist_4:1; bool runlist_5:1; bool runlist_6:1; bool runlist_7:1; bool runlist_8:1; bool runlist_9:1; bool runlist_10:1; bool runlist_11:1; bool runlist_12:1; bool runlist_13:1; uint32_t padding:18; } __attribute__((packed)); uint32_t raw; } runlist_preempt_t; /* Additional information on preempting from NVIDIA's driver (commit b1d0d8ece) * "From h/w team * Engine save can be blocked by eng stalling interrupts. * FIFO interrupts shouldn’t block an engine save from * finishing, but could block FIFO from reporting preempt done. * No immediate reason to reset the engine if FIFO interrupt is * pending. * The hub, priv_ring, and ltc interrupts could block context * switch (or memory), but doesn’t necessarily have to. * For Hub interrupts they just report access counters and page * faults. Neither of these necessarily block context switch * or preemption, but they could. * For example a page fault for graphics would prevent graphics * from saving out. An access counter interrupt is a * notification and has no effect. * SW should handle page faults though for preempt to complete. * PRI interrupt (due to a failed PRI transaction) will result * in ctxsw failure reported to HOST. * LTC interrupts are generally ECC related and if so, * certainly don’t block preemption/ctxsw but they could. * Bus interrupts shouldn’t have anything to do with preemption * state as they are part of the Host EXT pipe, though they may * exhibit a symptom that indicates that GPU is in a bad state. * To be completely fair, when an engine is preempting SW * really should just handle other interrupts as they come in. * It’s generally bad to just poll and wait on a preempt * to complete since there are many things in the GPU which may * cause a system to hang/stop responding." */ /* Runlist Metadata (up through Volta) "Software specifies the GPU contexts that hardware should "run" by writing a list of entries (known as a "runlist") to a 4k-aligned area of memory (beginning at NV_PFIFO_RUNLIST_BASE), and by notifying Host that a new list is available (by writing to NV_PFIFO_RUNLIST). Submission of a new runlist causes Host to expire the timeslice of all work scheduled by the previous runlist, allowing it to schedule the channels present in the new runlist once they are fetched. SW can check the status of the runlist by polling NV_PFIFO_ENG_RUNLIST_PENDING. (see dev_fifo.ref NV_PFIFO_RUNLIST for a full description of the runlist submit mechanism). Runlists can be stored in system memory or video memory (as specified by NV_PFIFO_RUNLIST_BASE_TARGET). If a runlist is stored in video memory, software will have to execute flush or read the last entry written before submitting the runlist to Host to guarantee coherency." (volta/dev_ram.ref.txt) We only document the *_PFIFO_ENG_RUNLIST_*(i) read-only registers here (where i is a runlist index). Runlists are configured via the seperate, writable *_PFIFO_RUNLIST_* register; see open-gpu-doc for more on that. LEN : Number of entries in runlist IS_PENDING : Is runlist committed? PTR : Pointer to start of 4k-aligned runlist (upper 28 of 40 bits) TARGET : Aperture of runlist (video or system memory) Support: Fermi*, Kepler, Maxwell, Pascal, Volta *Fermi may expose this information 8 bytes earlier, starting at 0x227C? */ #define NV_PFIFO_ENG_RUNLIST_BASE_GF100(i) (0x00002280+(i)*8) // Read-only typedef union { struct { // NV_PFIFO_ENG_RUNLIST_BASE_* fields uint32_t ptr:28; enum INST_TARGET target:2; uint32_t padding1:2; // NV_PFIFO_ENG_RUNLIST_* fields uint16_t len:16; uint32_t padding2:4; bool is_pending:1; uint32_t padding3:11; } __attribute__((packed)); uint64_t raw; } eng_runlist_gf100_t; /* Starting with Turing, the seperate registers for reading and writing runlist configuration were dropped in favor of read/write indexed registers. As part of this, the layout was modified to allow for larger runlist pointers (upper 52 of 64 bits). Support: Turing, Ampere, Lovelace?, Hopper? */ // Support: Turing #define NV_PFIFO_RUNLIST_BASE_TU102(i) (0x00002B00+(i)*16) // Read/write #define NV_PFIFO_RUNLIST_SUBMIT_TU102(i) (0x00002B08+(i)*16) // Read/write typedef union { struct { enum INST_TARGET target:2; uint32_t padding:10; uint64_t ptr:28; uint32_t padding2:24; } __attribute__((packed)); uint64_t raw; } runlist_base_tu102_t; typedef union { struct { uint16_t len:16; uint16_t offset:16; uint32_t preempted_tsgid:14; bool valid_preempted_tsgid:1; bool is_pending:1; uint32_t preempted_offset:16; } __attribute__((packed)); uint64_t raw; } runlist_submit_tu102_t; enum CHANNEL_STATUS { CHANNEL_STATUS_IDLE = 0, CHANNEL_STATUS_PENDING = 1, CHANNEL_STATUS_PENDING_CTX_RELOAD = 2, CHANNEL_STATUS_PENDING_ACQUIRE = 3, CHANNEL_STATUS_PENDING_ACQ_CTX_RELOAD = 4, CHANNEL_STATUS_ON_PBDMA = 5, CHANNEL_STATUS_ON_PBDMA_AND_ENG = 6, CHANNEL_STATUS_ON_ENG = 7, CHANNEL_STATUS_ON_ENG_PENDING_ACQUIRE = 8, CHANNEL_STATUS_ON_ENG_PENDING = 9, CHANNEL_STATUS_ON_PBDMA_CTX_RELOAD = 10, CHANNEL_STATUS_ON_PBDMA_AND_ENG_CTX_RELOAD = 11, CHANNEL_STATUS_ON_ENG_CTX_RELOAD = 12, CHANNEL_STATUS_ON_ENG_PENDING_CTX_RELOAD = 13, CHANNEL_STATUS_ON_ENG_PENDING_ACQ_CTX_RELOAD = 14, }; /* Programmable Channel Control System RAM (PCCSR) 512-entry array of channel control and status data structures. Support: Fermi, Maxwell, Pascal, Volta, Turing, [more?] */ #define NV_PCCSR_CHANNEL_INST(i) (0x00800000+(i)*8) #define MAX_CHID 512 typedef union { struct { // 0:31 uint32_t inst_ptr:28; enum INST_TARGET inst_target:2; uint32_t padding0:1; bool inst_bind:1; // 32:64 bool enable:1; bool next:1; uint32_t padding:6; bool force_ctx_reload:1; uint32_t padding2:1; bool enable_set:1; bool enable_clear:1; uint32_t padding3:10; bool pbdma_faulted:1; bool eng_faulted:1; enum CHANNEL_STATUS status:4; bool busy:1; uint32_t padding4:3; } __attribute__((packed)); uint64_t raw; } channel_ctrl_t; /* Control word for runlist enable/disable. RUNLIST_N : Is runlist n disabled? (1 == disabled, 0 == enabled) To select the runlist dynamically, use the BIT(nr) kernel macro. Disabling example: runlist_disable_t rl_disable; rl_disable.raw = nvdebug_readl(g, NV_PFIFO_SCHED_DISABLE); rl_disable.raw |= BIT(nr); nvdebug_writel(g, NV_PFIFO_SCHED_DISABLE, rl_disable.raw); Enabling example: runlist_disable_t rl_disable; rl_disable.raw = nvdebug_readl(g, NV_PFIFO_SCHED_DISABLE); rl_disable.raw &= ~BIT(nr); nvdebug_writel(g, NV_PFIFO_SCHED_DISABLE, rl_disable.raw); Support: Fermi, Kepler, Maxwell, Pascal, Volta, Turing */ #define NV_PFIFO_SCHED_DISABLE 0x00002630 typedef union { struct { bool runlist_0:1; bool runlist_1:1; bool runlist_2:1; bool runlist_3:1; bool runlist_4:1; bool runlist_5:1; bool runlist_6:1; bool runlist_7:1; bool runlist_8:1; bool runlist_9:1; bool runlist_10:1; uint32_t padding:21; } __attribute__((packed)); uint32_t raw; } runlist_disable_t; /* Read GPU descriptors from the Master Controller (MC) MINOR_REVISION : Legacy (only used with Celvin in Nouveau) MAJOR_REVISION : Legacy (only used with Celvin in Nouveau) IMPLEMENTATION : Which implementation of the GPU architecture ARCHITECTURE : Which GPU architecture CHIP_ID = IMPLEMENTATION + ARCHITECTURE << 4 CHIP_ID : Unique ID of all chips since Kelvin Support: Kelvin, Rankline, Curie, Tesla, Fermi, Kepler, Maxwell, Pascal, Volta, Turing, Ampere */ #define NV_MC_BOOT_0 0x00000000 #define NV_CHIP_ID_GP106 0x136 // Discrete GeForce GTX 1060 #define NV_CHIP_ID_GV11B 0x15B // Jetson Xavier embedded GPU #define NV_CHIP_ID_KEPLER 0x0E0 #define NV_CHIP_ID_MAXWELL 0x120 #define NV_CHIP_ID_PASCAL 0x130 #define NV_CHIP_ID_VOLTA 0x140 #define NV_CHIP_ID_VOLTA_INTEGRATED 0x150 #define NV_CHIP_ID_TURING 0x160 #define NV_CHIP_ID_AMPERE 0x170 #define NV_CHIP_ID_HOPPER 0x180 #define NV_CHIP_ID_ADA 0x190 inline static const char* ARCH2NAME(uint32_t arch) { switch (arch) { case 0x01: return "Celsius"; case 0x02: return "Kelvin"; case 0x03: return "Rankline"; case 0x04: case 0x06: // 0x06 is (nForce 6XX integrated only) return "Curie"; // 0x07 is unused/skipped case 0x05: // First Tesla card was released before the nForce 6XX case 0x08: case 0x09: case 0x0A: return "Tesla"; // 0x0B is unused/skipped case 0x0C: case 0x0D: return "Fermi"; case 0x0E: case 0x0F: case 0x11: return "Kepler"; case 0x12: return "Maxwell"; case 0x13: return "Pascal"; case 0x14: case 0x15: // Volta integrated return "Volta"; case 0x16: return "Turing"; case 0x17: return "Ampere"; case 0x18: return "Hopper"; case 0x19: return "Ada Lovelace"; case 0x20: return "Blackwell (?)"; default: if (arch < 0x19) return "[unknown historical architecture]"; else return "[future]"; } } typedef union { // Fields as defined in the NVIDIA reference struct { uint32_t minor_revision:4; uint32_t major_revision:4; uint32_t reserved:4; uint32_t padding0:8; uint32_t implementation:4; uint32_t architecture:5; uint32_t padding1:3; } __attribute__((packed)); uint32_t raw; // Arch << 4 + impl is also often used struct { uint32_t padding2:20; uint32_t chip_id:9; uint32_t padding3:3; } __attribute__((packed)); } mc_boot_0_t; /* GPU engine information and control register offsets (GPU TOPology) Each engine is described by one or more entries (terminated by an entry with the `has_next_entry` flag unset) in the fixed-size PTOP_DEVICE_INFO table. A typical device, such as the graphics/compute engine and any copy engines, are described by three entries, one of each type. The PTOP_DEVICE_INFO table is sparsely populated (entries of type INFO_TYPE_NOT_VALID may be intermingled with valid entries), so any traversal code should check all NV_PTOP_DEVICE_INFO__SIZE_1 entries and not terminate upon reaching the first entry of INFO_TYPE_NOT_VALID. The fields for the Ampere version of the GPU are a strict subset of those for the earlier versions. They are in different positions within the struct and have names ending in _ampere to distinguish them. Other than that, each Ampere device info field is functionally identical to the equivalent field in the previous version. INFO_TYPE : Is this a DATA, ENUM, or ENGINE_TYPE table entry? HAS_NEXT_ENTRY : Does the following entry refer to the same engine? == INFO_TYPE_DATA fields == PRI_BASE : BAR0 base = (PRI_BASE << 12) aka 4k aligned. INST_ID : "Note that some instanced [engines] (such as logical copy engines aka LCE) share a PRI_BASE across all [engines] of the same engine type; such [engines] require an additional offset: instanced base = BAR0 base + stride * INST_ID. FAULT_ID_IS_VALID : Does this engine have its own bind point and fault ID with the MMU? FAULT_ID : "The MMU fault id used by this [engine]. These IDs correspond to the NV_PFAULT_MMU_ENG_ID define list." == INFO_TYPE_ENUM fields == ENGINE_IS_VALID : Is this engine a host engine? ENGINE_ENUM : "[T]he host engine ID for the current [engine] if it is a host engine, meaning Host can send methods to the engine. This id is used to index into any register array whose __SIZE_1 is equal to NV_HOST_NUM_ENGINES. A given ENGINE_ENUM can be present for at most one device in the table. Devices corresponding to all ENGINE_ENUM ids 0 through NV_HOST_NUM_ENGINES - 1 must be present in the device info table." RUNLIST_IS_VALID : Is this engine a host engine with a runlist? RUNLIST_ENUM : "[T]he Host runlist ID on which methods for the current [engine] should be submitted... The runlist id is used to index into any register array whose __SIZE_1 is equal to NV_HOST_NUM_RUNLISTS. [Engines] corresponding to all RUNLIST_ENUM ids 0 through NV_HOST_NUM_RUNLISTS - 1 must be present in the device info table." INTR_IS_VALID : Does this device have an interrupt? INTR_ENUM : Interrupt ID for use with "the NV_PMC_INTR_*_DEVICE register bitfields." RESET_IS_VALID : Does this engine have a reset ID? RESET_ENUM : Reset ID for use indexing the "NV_PMC_ENABLE_DEVICE(i) and NV_PMC_ELPG_ENABLE_DEVICE(i) register bitfields." == INFO_TYPE_ENGINE_TYPE fields == ENGINE_TYPE : What type of engine is this? (see ENGINE_TYPES_NAMES) Support: Kepler, Maxwell, Pascal, Volta, Ampere See dev_top.ref.txt of NVIDIA's open-gpu-doc for more info. */ #define NV_PTOP_DEVICE_INFO_GA100(i) (0x00022800+(i)*4) #define NV_PTOP_DEVICE_INFO_GK104(i) (0x00022700+(i)*4) #define NV_PTOP_DEVICE_INFO__SIZE_1_GA100(g) (nvdebug_readl(g, 0x0224fc) >> 20) #define NV_PTOP_DEVICE_INFO__SIZE_1_GK104 64 enum DEVICE_INFO_TYPE {INFO_TYPE_NOT_VALID = 0, INFO_TYPE_DATA = 1, INFO_TYPE_ENUM = 2, INFO_TYPE_ENGINE_TYPE = 3}; enum ENGINE_TYPES { ENGINE_GRAPHICS = 0, // GRAPHICS [/compute] ENGINE_COPY0 = 1, // [raw/physical] COPY #0 ENGINE_COPY1 = 2, // [raw/physical] COPY #1 ENGINE_COPY2 = 3, // [raw/physical] COPY #2 ENGINE_MSPDEC = 8, // Picture DECoder ENGINE_MSPPP = 9, // [Video] Picture Post Processor ENGINE_MSVLD = 10, // [Video] Variable Length Decoder ENGINE_MSENC = 11, // [Video] ENCoding ENGINE_VIC = 12, // Video Image Compositor ENGINE_SEC = 13, // SEquenCer [?] ENGINE_NVENC0 = 14, // Nvidia Video ENCoder #0 ENGINE_NVENC1 = 15, // Nvidia Video ENCoder #1 ENGINE_NVDEC = 16, // Nvidia Video DECoder ENGINE_IOCTRL = 18, // I/O ConTRoLler [of NVLINK at least] ENGINE_LCE = 19, // Logical Copy Engine ENGINE_GSP = 20, // Gpu System Processor (Volta+) ENGINE_NVJPG = 21, // NVidia JPeG [Decoder] (Turing+) ENGINE_OFA = 22, // Optical Flow Accelerator (Turing+) ENGINE_FLA = 23, // [NVLink] Fabric Logical Addressing [?] }; #define ENGINE_TYPES_LEN 24 static const char* const ENGINE_TYPES_NAMES[ENGINE_TYPES_LEN] = { "Graphics/Compute", "COPY0", "COPY1", "COPY2", "Unknown Engine ID#4", "Unknown Engine ID#5", "Unknown Engine ID#6", "Unknown Engine ID#7", "MSPDEC: Picture Decoder", "MSPPP: Post Processing", "MSVLD: Variable Length Decoder", "MSENC: Encoder", "VIC: Video Image Compositor", "SEC: Sequencer", "NVENC0: NVIDIA Video Encoder #0", "NVENC1: NVIDIA Video Encoder #1", "NVDEC: NVIDIA Video Decoder", "Unknown Engine ID#17", "IOCTRL: I/O Controller", "LCE: Logical Copy Engine", "GSP: GPU System Processor", "NVJPG: NVIDIA JPEG Decoder", "OFA: Optical Flow Accelerator", "FLA: Fabric Logical Addressing", }; // These field are from nvgpu/include/nvgpu/hw/ga100/hw_top_ga100.h typedef union { // _info type fields struct { uint32_t fault_id:11; uint32_t padding0:5; uint32_t inst_id:8; enum ENGINE_TYPES engine_type:7; // "type_enum" bool has_next_entry:1; } __attribute__((packed)); // _info2 type fields struct { uint32_t reset_id:8; uint32_t pri_base:18; // "device_pri_base" uint32_t padding1:4; uint32_t is_engine:1; uint32_t padding2:1; } __attribute__((packed)); struct { uint32_t rleng_id:2; uint32_t padding3:8; uint32_t runlist_pri_base:16; uint32_t padding4:6; } __attribute__((packed)); uint32_t raw; } ptop_device_info_ga100_t; // These field are from open-gpu-doc/manuals/volta/gv100/dev_top.ref.txt typedef union { // DATA type fields struct { enum DEVICE_INFO_TYPE info_type:2; bool fault_id_is_valid:1; uint32_t fault_id:7; uint32_t padding0:2; uint32_t pri_base:12; uint32_t padding1:2; uint32_t inst_id:4; uint32_t is_not_enum2:1; bool has_next_entry:1; } __attribute__((packed)); // ENUM type fields struct { uint32_t padding2:2; bool reset_is_valid:1; bool intr_is_valid:1; bool runlist_is_valid:1; bool engine_is_valid:1; uint32_t padding3:3; uint32_t reset_enum:5; uint32_t padding4:1; uint32_t intr_enum:5; uint32_t padding5:1; uint32_t runlist_enum:4; uint32_t padding6:1; uint32_t engine_enum:4; uint32_t padding7:2; } __attribute__((packed)); // ENGINE_TYPE type fields struct { uint32_t padding8:2; enum ENGINE_TYPES engine_type:29; uint32_t padding9:1; } __attribute__((packed)); uint32_t raw; } ptop_device_info_gk104_t; /* Graphics Processing Cluster (GPC) on-chip information The GPU's Compute/Graphics engine is subdivided into Graphics Processing Clusters (also known as GPU Processing Clusters, starting with Ampere). Each GPC is subdivided into Texture Processing Clusters (TPCs) which contain Streaming Multiprocessors (SMs). The number of these units etched onto the chip may vary from the number enabled and software-visible. These registers expose the number of on-chip GPCs, the number of on-chip TPCs inside a GPC. Support: Fermi through (at least) Blackwell */ #define NV_PTOP_SCAL_NUM_GPCS 0x00022430 #define NV_PTOP_SCAL_NUM_TPC_PER_GPC 0x00022434 /* Graphics Processing Cluster (GPC) enablement information (See above for a description of GPCs and TPCs.) The number of on-chip GPCs and TPCs enabled is driven by: 1) Manufacturing errors which make some units nonfunctional. 2) Commercialization decisions about how many units should be enabled for a specific GPU model. Generally, reason (1) drives disablement early in product manufacturing, whereas, as the manufacturing process matures, (2) steps in to ensure consistency between early-manufactured and late-manufactured products. On-chip fuses are used to dictate which units are enabled and disabled. These registers expose the fuse configuration for GPCs, and the TPCs in each GPC. FUSE_GPC : Bitmask of which GPCs are enabled FUSE_TPC_FOR_GPC(i) : Bitmask of which TPCs are enabled for GPC i Support: Maxwell through Blackwell Note the registers were relocated starting with Ampere. */ #define NV_FUSE_GPC_GM107 0x00021c1c #define NV_FUSE_TPC_FOR_GPC_GM107(i) (0x00021c38+(i)*4) #define NV_FUSE_GPC_GA100 0x00820c1c #define NV_FUSE_TPC_FOR_GPC_GA100(i) (0x00820c38+(i)*4) /* Logical Copy Engine (LCE) Information Every GPU has some number of copy engines which can process transfers to, from, or within a GPU. Up until Maxwell, the hardware engines were directly accessible, and this register exposes how many there are. Starting with Pascal, an additional layer of indirection was added---logical copy engines. Only logical copy engines can be directly dispatched to, and there are normally more logical copy engines than there are physical ones. On Pascal+ this register stores the number of logical copy engines. SCAL_NUM_CES : Number of externally accessible copy engines Errata: Incorrectly reports "3" on Jetson TX1 and TX2. Should report "1" to be consistent with PTOP data. Support: Kepler through (at least) Blackwell Also see dev_ce.ref.txt of NVIDIA's open-gpu-doc for info. */ #define NV_PTOP_SCAL_NUM_CES 0x00022444 // Defined max number of GRCEs for a GPU (TX2 has only one) # define NV_GRCE_MAX 2 // Defined GRCE->CE mapping offsets from nvgpu #define NV_GRCE_FOR_CE_GP100(i) (0x00104034+(i)*4) #define NV_GRCE_FOR_CE_GA100(i) (0x001041c0+(i)*4) // Defined LCE->PCE mapping offset from nvgpu (same as ce_pce2lce_config_r(i) in nvgpu) #define NV_LCE_FOR_PCE_GP100 0x0010402c #define NV_LCE_FOR_PCE_GV100(i) (0x00104040+(i)*4) #define NV_LCE_FOR_PCE_GA100(i) (0x00104100+(i)*4) // Struct for use with nvdebug_reg_range_read() union reg_range { struct { uint32_t offset; uint8_t start_bit; uint8_t stop_bit; }; uint64_t raw; }; /* Physical Copy Engine (PCE) information On Pascal GPUs or newer, this register complements the above information by exposing which, and how many, physical copy engines are enabled on the GPU. CE_PCE_MAP : A bitmask, where a set bit indicates that the PCE for that index is enabled (not floorswept) on this GPU. Count the number of set bits to get the number of PCEs. Note that this may be bogus if the GPU has not been used since reset. Support: Pascal through (at least) Blackwell Also see dev_ce.ref.txt of NVIDIA's open-gpu-doc for info. */ #define NV_CE_PCE_MAP 0x00104028 #define NV_CE_PCE_MAP_SIZE 32 /* Location of the 1Kb instance block with page tables for BAR1 and BAR2. Support: Fermi+ (?), Pascal */ #define NV_PBUS_BAR1_BLOCK 0x00001704 #define NV_PBUS_BAR2_BLOCK 0x00001714 typedef union { struct { uint32_t ptr:28; enum INST_TARGET target:2; uint32_t padding0:1; bool is_virtual:1; } __attribute__((packed)); uint32_t raw; struct { uint32_t map:30; uint32_t padding1:2; } __attribute__((packed)); } bar_config_block_t; /* BAR0 PRAMIN (Private RAM Instance) window configuration BASE : Base of window >> 16 in [TARGET] virtual address space TARGET : Which address space BASE points into Note: This seems to be set to 0x0bff00000 - 0x0c0000000 at least sometimes Support: Tesla 2.0, Fermi, Kepler, Maxwell, Pascal, Turing, Ampere */ #define NV_PBUS_BAR0_WINDOW 0x00001700 #define NV_PRAMIN 0x00700000 // Goes until 0x00800000 (1MB window) #define NV_PRAMIN_LEN 0x00100000 typedef union { struct { uint32_t base:24; enum INST_TARGET target:2; uint32_t padding0:6; } __attribute__((packed)); uint32_t raw; } bar0_window_t; // Support: Tesla 2.0, Fermi, Kepler, Maxwell, Pascal, Turing, Ampere #define NV_PRAMIN_PDB_CONFIG_OFF 0x200 typedef union { struct { uint32_t target:2; uint32_t vol:1; uint32_t padding0:1; uint32_t fault_replay_tex:1; uint32_t fault_replay_gcc:1; uint32_t padding1:4; bool is_ver2:1; bool is_64k_big_page:1; // 128Kb otherwise uint32_t page_dir_lo:20; uint32_t page_dir_hi:32; } __attribute__((packed)); uint64_t raw; } page_dir_config_t; /* NVIDIA GMMU (GPU Memory Management Unit) uses page tables that are mostly straight-forward starting with Pascal ("page table version 2"), except for a few quirks (like 16-byte PDE0 entries, but all other entries are 8 bytes). All you really need to know is that any given Page Directory Entry (PDE) contains a pointer to the start of a 4k page densely filled with PDEs or Page Table Entries (PTEs). == Page Table Refresher == Page tables convert virtual addresses to physical addresses, and they do this via a tree structure. Leafs (PTEs) contain a physical address, and the path from root to leaf is defined by the virtual address. Non-leaf nodes are PDEs. When decending, the virtual address is sliced into pieces, and one slice is used at each level (as an index) to select the next-visited node (in level+1). V2 of NVIDIA's page table format uses 4 levels of PDEs and a final level of PTEs. How the virtual address is sliced to yield an index into each level and a page offset is shown by Fig 1. == Figure 1 == Page Offset (12 bits) <---------------------------------------+ Page Table Entry (PTE) (9 bits) <--------------------+ | Page Directory Entry (PDE) 0 (8 bits) <-----+ | | PDE1 (9 bits) <--------------------+ | | | PDE2 (9 bits) <-----------+ | | | | PDE3 (2 bits) <--+ | | | | | ^ ^ ^ ^ ^ ^ Virtual addr: [48, 47] [46, 38] [37, 29] [28, 21] [20, 12] [11, 0] The following arrays merely represent different projections of Fig. 1, and only one is strictly needed to reconstruct all the others. However, due to the complexity of page tables, we include all of these to aid in readability. */ // How many nodes/entries per level in V2 of NVIDIA's page table format static const int NV_MMU_PT_V2_SZ[5] = {4, 512, 512, 256, 512}; // Size in bytes of an entry at a particular level static const int NV_MMU_PT_V2_ENTRY_SZ[5] = {8, 8, 8, 16, 8}; // Which bit index is the least significant in indexing each page level static const int NV_MMU_PT_V2_LSB[5] = {47, 38, 29, 21, 12}; // Important: Aperture keys are different with PDEs enum PD_TARGET { PD_AND_TARGET_INVALID = 0, // b000 PD_AND_TARGET_VID_MEM = 2, // b010 PD_AND_TARGET_SYS_MEM_COHERENT = 4, // b100 PD_AND_TARGET_SYS_MEM_NONCOHERENT = 6, // b110 PTE_AND_TARGET_VID_MEM = 1, // b001 PTE_AND_TARGET_PEER = 3, // b011 PTE_AND_TARGET_SYS_MEM_COHERENT = 5, // b101 PTE_AND_TARGET_SYS_MEM_NONCOHERENT = 7, // b111 }; static inline const char *pd_target_to_text(enum PD_TARGET t) { switch (t) { case PD_AND_TARGET_INVALID: return "INVALID"; case PD_AND_TARGET_VID_MEM: case PTE_AND_TARGET_VID_MEM: return "VID_MEM"; case PTE_AND_TARGET_PEER: return "PEER"; case PD_AND_TARGET_SYS_MEM_COHERENT: case PTE_AND_TARGET_SYS_MEM_COHERENT: return "SYS_MEM_COHERENT"; case PD_AND_TARGET_SYS_MEM_NONCOHERENT: case PTE_AND_TARGET_SYS_MEM_NONCOHERENT: return "SYS_MEM_NONCOHERENT"; default: return "UNKNOWN"; } } // Page Directory Entry/Page Table Entry V2 type // Note: As the meaning of target (bits 2:1) changes depending on if the entry // is a PTE or not, this combines them into a single target field to // simplify comparisons. // Support: Pascal, Volta, Turing, Ampere, Ada // // V3 introduced with Hopper, but Hopper and Blackwell also support V2 // typedef union { // Page Directory Entry (PDE) struct { enum PD_TARGET target:3; bool is_volatile:1; uint32_t padding1:4; uint32_t addr:24; uint32_t __unused1; } __attribute__((packed)); // Page Table Entry (PTE) struct { bool is_pte:1; enum INST_TARGET aperture:2; uint32_t __is_volatile:1; bool is_encrypted:1; bool is_privileged:1; bool is_readonly:1; bool atomics_disabled:1; uint32_t __addr:24; uint32_t __unused2; } __attribute__((packed)); // For wide addresses in PTEs or PDEs; only used if target is SYS_MEM struct { uint32_t __overlap:8; uint64_t addr_w:46; uint32_t __unused3:10; } __attribute__((packed)); uint64_t raw_w; } page_dir_entry_t; // Page Directory Entry/Page Table Entry V1 type // Support: Fermi, Kepler, Maxwell enum V1_PD_TARGET { PD_TARGET_INVALID = 0, PD_TARGET_VID_MEM = 1, PD_TARGET_SYS_MEM_COHERENT = 2, PD_TARGET_SYS_MEM_NONCOHERENT = 3, }; // Page Directory Entry (PDE) typedef union { // Large page fields struct { // 0:32 enum V1_PD_TARGET target:2; uint32_t padding0:2; uint64_t addr:28; // May be wider? // 32:63 uint32_t padding2:3; uint32_t is_volatile:1; // Might have counted wrong? uint32_t padding3:28; } __attribute__((packed)); // Small page fields struct { // 0:32 uint32_t padding00:32; // 32:63 enum V1_PD_TARGET alt_target:2; uint32_t alt_is_volatile:1; // Might have counted wrong? uint32_t padding03:1; uint64_t alt_addr:28; } __attribute__((packed)); uint64_t raw; } page_dir_entry_v1_t; // Page Table Entry (PTE) // Reconstructed from info in Jetson nvgpu driver typedef union { struct { // 0:32 bool is_present:1; bool is_privileged:1; bool is_readonly:1; uint32_t padding0:1; uint64_t addr:28; // 32:63 bool is_volatile:1; enum INST_TARGET:2; uint32_t padding1:1; uint32_t kind:8; uint32_t comptag:17; uint32_t padding2:1; bool is_read_disabled:1; bool is_write_disabled:1; } __attribute__((packed)); uint64_t raw; } page_tbl_entry_v1_t; //enum V0_PDE_TYPE {NOT_PRESENT = 0, PAGE_64K = 1, PAGE_16K = 2, PAGE_4K = 3}; //enum V0_PDE_SIZE {PDE_SZ_128K = 0, PDE_SZ_32K = 1, PDE_SZ_16K = 2, PDE_SZ_8K = 3}; //static const int V0_PDE_SIZE2NUM[4] = {128*1024, 32*1024, 16*1024, 8*1024}; /* PDE V0 (nv50/Tesla) typedef union { struct { enum V1_PDE_TYPE type:2; enum INST_TARGET target:2; uint32_t padding0:1; enum V1_PDE_SIZE sublevel_size:2; uint32_t padding1:5; uint32_t addr:28; uint32_t padding2:24; } __attribute__((packed)); uint64_t raw; } page_dir_entry_v1_t;*/ /* PTE V0 (nv50) typedef union { struct { bool is_present:1; uint32_t padding3:2; bool is_readonly:1; enum INST_TARGET target:2; bool is_privileged:1; uint32_t contig_blk_sz:3; uint32_t padding4:2; uint32_t addr:28; uint32_t storage_type:7; // ??? uint32_t compression_mode:2; // ??? uint32_t compression_tag:12; // ??? bool is_long_partition_cycle:1; // ??? bool is_encrypted:1; uint32_t padding5:1; } __attribute__((packed)); uint64_t raw; } page_tbl_entry_v1_t;*/ // TODO(jbakita): Maybe put the above GPU types in a different file. #define NV_PCI_VENDOR 0x10de struct nvdebug_state { // Pointer to the mapped base address of the GPU control registers (obtained // via ioremap() originally). For embedded GPUs, we extract this from their // struct nvgpu_os_linux. For discrete GPUs, we create our own mapping of // BAR0 with pci_iomap(). Access via nvgpu_readl/writel functions. void __iomem *regs; // Depending on the architecture, BAR2 or BAR3 are used to access PRAMIN union { void __iomem *bar2; void __iomem *bar3; }; int chip_id; // Additional state from the built-in driver. Only set iff // chip_id == NV_CHIP_ID_GV11B struct gk20a *g; // Pointer to PCI device needed for pci_iounmap struct pci_dev *pcid; }; /*const struct runlist_funcs { u8 size; enum ENTRY_TYPE (*entry_type)(struct nvdebug_state *, void *); uint32_t (*chid)(struct nvdebug_state *, void *); uint32_t (*inst_ptr_lo)(struct nvdebug_state *, void *); enum INST_TARGET (*inst_target)(struct nvdebug_state *, void *): uint32_t (*tsgid)(struct nvdebug_state *, void *); uint32_t (*timeslice_scale)(struct nvdebug_state *, void *); uint32_t (*timeslice_timeout)(struct nvdebug_state *, void *); uint32_t (*tsg_length)(struct nvdebug_state *, void *); };*/ // This disgusting macro is a crutch to work around the fact that runlists were // different prior to Volta. #define VERSIONED_RL_ACCESSOR(_ENTRY_TYPE, type, prop) \ __attribute__((unused)) \ static type (prop)(const struct nvdebug_state *g, const void *raw) { \ if (g->chip_id >= NV_CHIP_ID_VOLTA) { \ const struct gv100_runlist_ ## _ENTRY_TYPE *entry = (struct gv100_runlist_ ## _ENTRY_TYPE*)raw; \ return entry->prop; \ } else if (g->chip_id >= NV_CHIP_ID_KEPLER) { \ const struct gk110_runlist_ ## _ENTRY_TYPE *entry = (struct gk110_runlist_ ## _ENTRY_TYPE*)raw; \ return entry->prop; \ } else { \ printk(KERN_WARNING "[nvdebug] " #prop " unavailable on GPU ID %x, which is older than Kepler.\n", g->chip_id); \ return (type)0; \ } \ } VERSIONED_RL_ACCESSOR(chan, uint32_t, chid); VERSIONED_RL_ACCESSOR(chan, uint32_t, inst_ptr_lo); VERSIONED_RL_ACCESSOR(chan, enum INST_TARGET, inst_target); VERSIONED_RL_ACCESSOR(tsg, uint32_t, tsgid); VERSIONED_RL_ACCESSOR(tsg, enum ENTRY_TYPE, entry_type); VERSIONED_RL_ACCESSOR(tsg, uint32_t, timeslice_scale); VERSIONED_RL_ACCESSOR(tsg, uint32_t, timeslice_timeout); VERSIONED_RL_ACCESSOR(tsg, uint32_t, tsg_length); #define NV_RL_ENTRY_SIZE(g) \ ((g)->chip_id >= NV_CHIP_ID_VOLTA ? sizeof(struct gv100_runlist_tsg) : sizeof(struct gk110_runlist_tsg)) #define for_chan_in_tsg(g, chan, tsg) \ for (chan = (typeof(chan))(((u8*)tsg) + NV_RL_ENTRY_SIZE(g)); \ (u8*)chan < ((u8*)tsg) + (1 + tsg_length(g, tsg)) * NV_RL_ENTRY_SIZE(g); \ chan = (typeof(chan))(((u8*)chan) + NV_RL_ENTRY_SIZE(g))) #define next_tsg(g, tsg) \ (typeof(tsg))((u8*)(tsg) + NV_RL_ENTRY_SIZE(g) * (tsg_length(g, tsg) + 1)) struct runlist_iter { // Pointer to either a TSG or channel entry (they're the same size) void *curr_entry; // This should be set to tsg_length + 1 when a TSG is reached, and // decremented each time _next() is called. This allows us to // track which channels are and are not part of the TSG. int entries_left_in_tsg; // Number of entries in runlist int len; }; #define NVDEBUG_MAX_DEVICES 8 extern struct nvdebug_state g_nvdebug_state[NVDEBUG_MAX_DEVICES]; // Defined in runlist.c int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl_iter); int preempt_tsg(struct nvdebug_state *g, uint32_t tsg_id); int preempt_runlist(struct nvdebug_state *g, uint32_t rl_id); // Defined in mmu.c void __iomem *phy2PRAMIN(struct nvdebug_state* g, uint64_t phy); uint64_t search_page_directory( struct nvdebug_state *g, void __iomem *pde_offset, void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t), uint64_t addr_to_find); uint64_t search_v1_page_directory( struct nvdebug_state *g, void __iomem *pde_offset, void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t), uint64_t addr_to_find); static inline struct gk20a *get_gk20a(struct device *dev) { // XXX: Only works because gk20a* is the first member of gk20a_platform return *((struct gk20a**)dev_get_drvdata(dev)); } // We us the data field of the proc_dir_entry ("PDE" in this function) to store // our index into the g_nvdebug_state array static inline int seq2gpuidx(struct seq_file *s) { const struct file *f = s->file; return (uintptr_t)PDE_DATA(file_inode(f)); } static inline int file2gpuidx(const struct file *f) { return (uintptr_t)PDE_DATA(file_inode(f)); } static inline int file2parentgpuidx(const struct file *f) { // Should be safe to call on ProcFS entries, as our parent should (?) // still exist if we're called. If not, there are worse races in this // module. return (uintptr_t)PDE_DATA(file_dentry(f)->d_parent->d_inode); } #if LINUX_VERSION_CODE >= KERNEL_VERSION(5,6,0) // Commit 643eb158a3 in nvgpu moved the mapped registers to the second entry // of the gk20a struct (after a function pointer). This change was made as L4T // was upgraded from Linux 4.9 to 5.10 (r32 -> r34+) // Note that this is wrong if nvgpu was built without CONFIG_NVGPU_NON_FUSA // i.e. if FUSA was enabled, this is wrong. #define gk20a_regs(gk20a) (*(void**)((void*)gk20a + sizeof(void(*)(void)))) #else #include // For struct nvgpu_os_linux, which holds regs #define gk20a_regs(gk20a) (container_of(gk20a, struct nvgpu_os_linux, g)->regs) #endif // Similar to nvgpu_readl() // (except we don't try to resolve situations where regs is NULL) static inline u32 nvdebug_readl(struct nvdebug_state *s, u32 r) { u32 ret; if (unlikely(!s->regs || (s->g && !gk20a_regs(s->g)))) { printk(KERN_ERR "[nvdebug] nvdebug_readl: Unable to read; registers unavailable. Is GPU on?\n"); return -1; } ret = readl(s->regs + r); // It seems like the GPU returns this as a flag value for bad addresses if (ret == 0xbadf5040) { printk(KERN_ERR "[nvdebug] nvdebug_readl: Unable to read from register offset %#x; bad data\n", r); return -1; } return ret; } // quadword version of nvdebug_readl() static inline u64 nvdebug_readq(struct nvdebug_state *s, u32 r) { u64 ret; if (unlikely(!s->regs || (s->g && !gk20a_regs(s->g)))) { printk(KERN_ERR "[nvdebug] nvdebug_readq: Unable to read; registers unavailable. Is GPU on?\n"); return -1; } // readq seems to always return the uppermost 32 bits as 0, so workaround with readl ret = readl(s->regs + r); ret |= ((u64)readl(s->regs + r + 4)) << 32; // It seems like the GPU returns this as a flag value for bad addresses if ((ret & 0xffffffffull) == 0xbadf5040ull) { printk(KERN_ERR "[nvdebug] nvdebug_readq: Unable to read from register offset %#x; bad data\n", r); return -1; } return ret; } // Similar to nvgpu_writel() static inline void nvdebug_writel(struct nvdebug_state *s, u32 r, u32 v) { if (unlikely(!s->regs || (s->g && !gk20a_regs(s->g)))) { printk(KERN_ERR "[nvdebug] nvdebug_writel: Unable to write; registers unavailable. Is GPU on?\n"); return; } writel_relaxed(v, s->regs + r); wmb(); } // quadword version of nvdebug_writel() // XXX: This probably doesn't work XXX: Untested static inline void nvdebug_writeq(struct nvdebug_state *s, u32 r, u64 v) { if (unlikely(!s->regs || (s->g && !gk20a_regs(s->g)))) { printk(KERN_ERR "[nvdebug] nvdebug_writeq: Unable to write; registers unavailable. Is GPU on?\n"); return; } writeq_relaxed(v, s->regs + r); wmb(); } // Defined in bus.c int addr_to_pramin_mut(struct nvdebug_state *g, uint64_t addr, enum INST_TARGET target); int get_bar2_pdb(struct nvdebug_state *g, void **pdb, bool *is_v2_pdb);