Expand support for printing LCE<->PCE and GRCE->LCE configurationrtas24-ae

Tested working on Pascal, Volta, Volta Integrated, Turing, Ampere,
and Ada.

Also clean up minor spacing issues, an errantly added file
(nvdebug.mod), and fix some inconsistencies with upstream.

1 files changed, 24 insertions, 19 deletions

diff --git a/device_info_procfs.c b/device_info_procfs.c
index d5350c8..168905f 100644
--- a/device_info_procfs.c
+++ b/device_info_procfs.c
@@ -9,7 +9,7 @@
 // @param off  Requested offset. Updated by number of characters written.
 // @return -errno on error, otherwise number of bytes written to *buf
 // Note: Parent `data` field MUST be the GPU index
-static ssize_t nvdebug_reg32_read(struct file *f, char __user *buf, size_t size, loff_t *off){
+static ssize_t nvdebug_reg32_read(struct file *f, char __user *buf, size_t size, loff_t *off) {
         char out[16];
         int chars_written;
         struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
@@ -22,37 +22,42 @@ static ssize_t nvdebug_reg32_read(struct file *f, char __user *buf, size_t size,
         *off += chars_written;
         return chars_written;
 }
-static ssize_t nvdebug_read4_pascal(struct file *f, char __user *buf, size_t size, loff_t *off){
-        char out[16];
+
+static ssize_t nvdebug_reg_range_read(struct file *f, char __user *buf, size_t size, loff_t *off) {
+        char out[12];
         int chars_written;
+        uint32_t read, mask;
         struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
-        void* data = PDE_DATA(file_inode(f));
-        struct combo local_combo  = *(struct combo*) &data;
+        // See comment in nvdebug_entry.c to understand `union reg_range`
+        union reg_range range;
+        range.raw = (uintptr_t)PDE_DATA(file_inode(f));
 
-        // 32 bit register will always take less than 16 characters to print
-        if (size < 16 || *off != 0)
+        // "0x" + up to 32-bit register as hex + "\n\0" is at most 12 characters
+        if (size < 12 || *off != 0)
                 return 0;
-        if (local_combo.index % 2 == 0)
-                chars_written = scnprintf(out, 16, "%#0x\n", (nvdebug_readl(g, local_combo.offset) & 0x0f));
-        else
-                chars_written = scnprintf(out, 16, "%#0x\n", (nvdebug_readl(g, local_combo.offset) & 0xf0) >> 4);
+
+        // Print bits `start_bit` to `stop_bit` from 32 bits at address `offset`
+        if ((read = nvdebug_readl(g, range.offset)) == -1)
+                return -EOPNOTSUPP;
+        // Setup `mask` used to throw out unused upper bits
+        mask = -1u >> (32 - range.stop_bit + range.start_bit);
+        // Throw out unused lower bits via a shift, apply the mask, and print
+        chars_written = scnprintf(out, 12, "%#0x\n", (read >> range.start_bit) & mask);
         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;
-
-//(nvdebug_readl(g,NV_LCE_FOR_PCE_GP100(*(int*)PDE_DATA(file_inode(f))))
-
-
-
 }
+
 struct file_operations nvdebug_read_reg32_file_ops = {
         .read = nvdebug_reg32_read,
         .llseek = default_llseek,
 };
-// File operation for reading 4 bits in 32 bit register (used for Pascal copy engine offsets) 
-struct file_operations nvdebug_read4_pascal_file_ops = {
-        .read = nvdebug_read4_pascal,
+
+// Generic mechanism used for printing a subset of bits from a register
+// Please store a `union reg_range` rather than a `uintptr_t` in the PDE_DATA
+struct file_operations nvdebug_read_reg_range_file_ops = {
+        .read = nvdebug_reg_range_read,
         .llseek = default_llseek,
 };