2 files changed, 125 insertions, 71 deletions

diff --git a/nvdebug.h b/nvdebug.h
index f0662cd..39b2e6e 100644
--- a/nvdebug.h
+++ b/nvdebug.h
@@ -436,7 +436,9 @@ typedef union {
 #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
@@ -700,29 +702,47 @@ typedef union {
         uint32_t raw;
 } ptop_device_info_gk104_t;
 
-/* Graphics Processing Cluster (GPC) information
+/* 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
 */
-// Support: Fermi through Blackwell
-// Get the number of GPCs **on die**
 #define NV_PTOP_SCAL_NUM_GPCS 0x00022430
-// Get the number of TPCs per GPC **on die**
 #define NV_PTOP_SCAL_NUM_TPC_PER_GPC 0x00022434
-// GPC and TPC masks
-// Support: Maxwell, Pascal, Volta, Turing
-// Bitmask of which GPC **are enabled** of the max on die
-#define NV_FUSE_GPC 0x00021c1c
-// Bitmask of which TPCs **are enabled** on each GPC
-#define NV_FUSE_TPC_FOR_GPC(i) (0x00021c38+(i)*4)
-// Support: Ampere, Ada, Hopper, Blackwell
-//#define NV_FUSE_GPC 0x00820c1c
-//#define NV_FUSE_TPC_FOR_GPC(i) (0x00820c38+(i)*4)
+
+/* 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,
diff --git a/nvdebug_entry.c b/nvdebug_entry.c
index 7593a3a..0cf5344 100644
--- a/nvdebug_entry.c
+++ b/nvdebug_entry.c
@@ -12,7 +12,8 @@
 #include "nvdebug.h"
 #include "stubs.h"
 
-// Enable to intercept and log GPU interrupts
+// Enable to intercept and log GPU interrupts. Historically used to benchmark
+// interrupt latency.
 #define INTERRUPT_DEBUG 0
 
 // MIT is GPL-compatible. We need to be GPL-compatible for symbols like
@@ -31,14 +32,16 @@ extern struct file_operations copy_topology_file_ops;
 extern struct file_operations nvdebug_read_reg32_file_ops;
 extern struct file_operations nvdebug_read_reg_range_file_ops;
 
-// Bus types are global symbols in the kernel
-extern struct bus_type platform_bus_type;
 struct nvdebug_state g_nvdebug_state[NVDEBUG_MAX_DEVICES];
 unsigned int g_nvdebug_devices = 0;
+// Bus types are global symbols in the kernel
+extern struct bus_type platform_bus_type;
 
-// Starting in Kernel 5.6, proc_ops is required instead of file_operations
+// Starting in Kernel 5.6, proc_ops is required instead of file_operations.
+// As file_operations is larger than proc_ops, we can overwrite the memory
+// backing the file_operations struct to follow the proc_ops layout, and then
+// cast on newer kernels.
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(5,6,0)
-// This rewrites the struct to the proc_ops layout on newer kernels
 const struct proc_ops* compat_ops(const struct file_operations* ops) {
         struct proc_ops new_ops = {};
         new_ops.proc_open = ops->open;
@@ -64,7 +67,7 @@ irqreturn_t nvdebug_irq_tap(int irq_num, void * dev) {
 
 // Find any and all NVIDIA GPUs in the system
 // Note: This function fails if any of them are in a bad state
-int probe_and_cache_device(void) {
+int probe_and_cache_devices(void) {
         // platform bus (SoC) iterators
         struct device *dev = NULL;
         struct device *temp_dev;
@@ -143,13 +146,14 @@ int probe_and_cache_device(void) {
 #endif // INTERRUPT_DEBUG
                 i++;
         }
-        // Return the number of devices we found
+        // Return the number of devices found
         if (i > 0)
                 return i;
         return -ENODEV;
 }
 
 // Create files `/proc/gpu#/runlist#`, world readable
+// Support: Fermi, Maxwell, Pascal, Volta, Turing
 int create_runlist_files(int device_id, struct proc_dir_entry *dir) {
         ptop_device_info_gk104_t info;
         struct proc_dir_entry *rl_entry;
@@ -179,16 +183,24 @@ int create_runlist_files(int device_id, struct proc_dir_entry *dir) {
         return 0;
 }
 
-// Create files /proc/gpu#
-// TODO: Don't run this on unsupported GPUs
+// Create files `/proc/gpu#/gpc#_tpc_mask`, world readable
+// Support: Maxwell+
 int create_tpc_mask_files(int device_id, struct proc_dir_entry *dir) {
+        struct nvdebug_state* g = &g_nvdebug_state[device_id];
         char file_name[20];
         int i;
         struct proc_dir_entry *gpc_tpc_mask_entry;
-        // Get a bitmask of which GPCs are disabled
-        uint32_t gpcs_mask = nvdebug_readl(&g_nvdebug_state[device_id], NV_FUSE_GPC);
         // Get maximum number of enabled GPCs for this chip
         uint32_t max_gpcs = nvdebug_readl(&g_nvdebug_state[device_id], NV_PTOP_SCAL_NUM_GPCS);
+        // Get a bitmask of which GPCs are disabled
+        uint32_t gpcs_mask;
+        if (g->chip_id < NV_CHIP_ID_AMPERE)
+                gpcs_mask = nvdebug_readl(g, NV_FUSE_GPC_GM107);
+        else
+                gpcs_mask = nvdebug_readl(g, NV_FUSE_GPC_GA100);
+        // Verify the reads succeeded
+        if (max_gpcs == -1 || gpcs_mask == -1)
+                return -EIO;
         // For each enabled GPC, expose a mask of disabled TPCs
         for (i = 0; i < max_gpcs; i++) {
                 // Do nothing if GPC is disabled
@@ -196,9 +208,14 @@ int create_tpc_mask_files(int device_id, struct proc_dir_entry *dir) {
                         continue;
                 // If GPC is enabled, create an entry to read disabled TPCs mask
                 snprintf(file_name, 20, "gpc%d_tpc_mask", i);
-                gpc_tpc_mask_entry = proc_create_data(
-                        file_name, 0444, dir, compat_ops(&nvdebug_read_reg32_file_ops),
-                        (void*)(uintptr_t)NV_FUSE_TPC_FOR_GPC(i));
+                if (g->chip_id < NV_CHIP_ID_AMPERE)
+                        gpc_tpc_mask_entry = proc_create_data(
+                                file_name, 0444, dir, compat_ops(&nvdebug_read_reg32_file_ops),
+                                (void*)(uintptr_t)NV_FUSE_TPC_FOR_GPC_GM107(i));
+                else
+                        gpc_tpc_mask_entry = proc_create_data(
+                                file_name, 0444, dir, compat_ops(&nvdebug_read_reg32_file_ops),
+                                (void*)(uintptr_t)NV_FUSE_TPC_FOR_GPC_GA100(i));
                 if (!gpc_tpc_mask_entry)
                         return -ENOMEM;
         }
@@ -206,64 +223,84 @@ int create_tpc_mask_files(int device_id, struct proc_dir_entry *dir) {
 }
 
 int __init nvdebug_init(void) {
-        struct proc_dir_entry *dir, *preempt_entry, *disable_channel_entry,
-                              *enable_channel_entry, *switch_to_tsg_entry, *device_info_entry,
-                              *num_gpcs_entry;
-        int rl_create_err, tpc_masks_create_err;
+        struct proc_dir_entry *dir;
+        int err, res;
         // Check that an NVIDIA GPU is present and initialize g_nvdebug_state
-        int res = probe_and_cache_device();
-        if (res < 0)
+        if ((res = probe_and_cache_devices()) < 0)
                 return res;
         g_nvdebug_devices = res;
         // Create seperate ProcFS directories for each gpu
         while (res--) {
                 char device_id_str[7];