Add /proc/gpu#/local_memory API for getting VRAM size

3 files changed, 59 insertions, 0 deletions

diff --git a/device_info_procfs.c b/device_info_procfs.c
index 8fe9709..c8903fc 100644
--- a/device_info_procfs.c
+++ b/device_info_procfs.c
@@ -61,6 +61,30 @@ struct file_operations nvdebug_read_reg_range_file_ops = {
         .llseek = default_llseek,
 };
 
+static ssize_t local_memory_read(struct file *f, char __user *buf, size_t size, loff_t *off) {
+        struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
+        char out[30];
+        int chars_written;
+        memory_range_t mem_range;
+        if (size < 30 || *off != 0)
+                return 0;
+        mem_range.raw = nvdebug_readl(g, NV_FB_MMU_LOCAL_MEMORY_RANGE);
+        if (mem_range.raw == -1)
+                return -EIO;
+        // 64-bit size has at most 19 characters + 8 for text and termination
+        chars_written = scnprintf(out, 30, "%lld bytes\n", memory_range_to_bytes(mem_range));
+        if (copy_to_user(buf, out, chars_written))
+                printk(KERN_WARNING "Unable to copy all data for %s\n", file_dentry(f)->d_name.name);
+        *off += chars_written;
+        return chars_written;
+}
+
+// Read out size of on-device VRAM
+struct file_operations local_memory_file_ops = {
+        .read = local_memory_read,
+        .llseek = default_llseek,
+};
+
 typedef struct {
         int idx; // Current index in the device_info table
         int length; // Length of device_info table (including unpopulated entries)
diff --git a/nvdebug.h b/nvdebug.h
index ff35f70..b13d311 100644
--- a/nvdebug.h
+++ b/nvdebug.h
@@ -1145,7 +1145,34 @@ typedef union {
 } page_tbl_entry_v0_t;
 */
 
+/* VRAM Information
 
+  If ECC is disabled:
+    bytes = (magnitude << scale) * 1024 * 1024
+  If ECC is enabled:
+    bytes = ((magnitude << scale) * 1024 * 1024) / 16 * 15
+
+  Support: Pascal, Volta, Turing, [more?]
+ */
+#define NV_FB_MMU_LOCAL_MEMORY_RANGE 0x00100ce0
+typedef union  {
+        struct {
+                uint32_t scale:4;
+                uint32_t mag:6;
+                uint32_t:20;
+                bool is_ecc:1;
+                uint32_t:1;
+        } __attribute__((packed));
+        uint32_t raw;
+} memory_range_t;
+
+static inline uint64_t memory_range_to_bytes(memory_range_t range) {
+        // ECC takes a byte out of available memory for parity data
+        if (range.is_ecc)
+                return ((range.mag << range.scale) * 1024ull * 1024ull) / 16 * 15;
+        else
+                return (range.mag << range.scale) * 1024ull * 1024ull;
+}
 
 /* Begin nvdebug types and functions */
 
diff --git a/nvdebug_entry.c b/nvdebug_entry.c
index 0caa289..f3dcdbb 100644
--- a/nvdebug_entry.c
+++ b/nvdebug_entry.c
@@ -31,6 +31,7 @@ extern struct file_operations device_info_file_ops;
 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;
+extern struct file_operations local_memory_file_ops;
 
 struct nvdebug_state g_nvdebug_state[NVDEBUG_MAX_DEVICES];
 unsigned int g_nvdebug_devices = 0;
@@ -303,6 +304,13 @@ int __init nvdebug_init(void) {
                                         (void*)NV_FUSE_GPC_GM107))
                                 goto out_nomem;
                 }
+                // Create file `/proc/gpu#/local_memory`, world readable (Pascal+)
+                if (g_nvdebug_state[res].chip_id >= NV_CHIP_ID_PASCAL) {
+                        if (!proc_create_data(
+                                        "local_memory", 0444, dir, compat_ops(&local_memory_file_ops),
+                                        (void*)0x00100ce0))
+                                goto out_nomem;
+                        }
                 // Create files exposing LCE and PCE configuration (Pascal+)
                 if (g_nvdebug_state[res].chip_id >= NV_CHIP_ID_PASCAL) {
                         // Create file `/proc/gpu#/copy_topology`, world readable