Linux 5.17+ support and allow including nvdebug.h independently

- Move Linux-specific functions to nvdebug_linux.h and .c
- Workaround PDE_DATA() being pde_data() on Linux 5.17+

8 files changed, 140 insertions, 108 deletions

diff --git a/Makefile b/Makefile
index c2d2e96..972bc0b 100644
--- a/Makefile
+++ b/Makefile
@@ -1,6 +1,6 @@
 obj-m += nvdebug.o
 nvdebug-objs = runlist_procfs.o device_info_procfs.o runlist.o mmu.o \
-               nvdebug_entry.o bus.o copy_topology_procfs.o
+               nvdebug_entry.o bus.o nvdebug_linux.o copy_topology_procfs.o
 KBUILD_CFLAGS += -DGIT_HASH=\"$(shell git --git-dir=$(PWD)/.git rev-parse --short HEAD)\"
 # -mfentry above if not building due to mcount missing
 
diff --git a/copy_topology_procfs.c b/copy_topology_procfs.c
index cfedce7..a1aa0ad 100644
--- a/copy_topology_procfs.c
+++ b/copy_topology_procfs.c
@@ -1,6 +1,6 @@
 // Copyright 2024 Joshua Bakita
 
-#include "nvdebug.h"
+#include "nvdebug_linux.h"
 
 // Maximum number of LCEs that we will print
 #define MAX_LCES 32
diff --git a/device_info_procfs.c b/device_info_procfs.c
index 168905f..8fe9709 100644
--- a/device_info_procfs.c
+++ b/device_info_procfs.c
@@ -1,4 +1,4 @@
-#include "nvdebug.h"
+#include "nvdebug_linux.h"
 #include <linux/seq_file.h> // For seq_* functions and types
 #include <linux/uaccess.h> // For copy_to_user()
 
@@ -16,7 +16,7 @@ static ssize_t nvdebug_reg32_read(struct file *f, char __user *buf, size_t size,
         if (size < 16 || *off != 0)
                 return 0;
         // 32 bit register will always take less than 16 characters to print
-        chars_written = scnprintf(out, 16, "%#0x\n", nvdebug_readl(g, (uintptr_t)PDE_DATA(file_inode(f))));
+        chars_written = scnprintf(out, 16, "%#0x\n", nvdebug_readl(g, (uintptr_t)pde_data(file_inode(f))));
         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;
@@ -30,7 +30,7 @@ static ssize_t nvdebug_reg_range_read(struct file *f, char __user *buf, size_t s
         struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
         // See comment in nvdebug_entry.c to understand `union reg_range`
         union reg_range range;
-        range.raw = (uintptr_t)PDE_DATA(file_inode(f));
+        range.raw = (uintptr_t)pde_data(file_inode(f));
 
         // "0x" + up to 32-bit register as hex + "\n\0" is at most 12 characters
         if (size < 12 || *off != 0)
@@ -55,7 +55,7 @@ struct file_operations nvdebug_read_reg32_file_ops = {
 };
 
 // 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
+// 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,
diff --git a/nvdebug.h b/nvdebug.h
index eff1470..ff35f70 100644
--- a/nvdebug.h
+++ b/nvdebug.h
@@ -7,13 +7,11 @@
  * - Detailed GPU information (PTOP, FUSE, and CE)
  * - PRAMIN, BAR1/2, and page table status
  * - Helper functions for nvdebug
+ *
+ * This function should not depend on any Linux-internal headers, and may be
+ * included outside of nvdebug.
  */
-
-#include <linux/device.h>   // For dev_get_drvdata()
-#include <linux/seq_file.h> // For struct seq_file
-#include <linux/proc_fs.h>  // For PDE_DATA() macro
-#include <linux/version.h>  // For KERNEL_VERSION and LINUX_VERSION_CODE
-#include <asm/io.h>
+#include <linux/types.h>
 
 // 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.
@@ -1147,8 +1145,11 @@ typedef union {
 } page_tbl_entry_v0_t;
 */
 
-// TODO(jbakita): Maybe put the above GPU types in a different file.
 
+
+/* Begin nvdebug types and functions */
+
+// Vendor ID for PCI devices manufactured by NVIDIA
 #define NV_PCI_VENDOR 0x10de
 struct nvdebug_state {
         // Pointer to the mapped base address of the GPU control registers (obtained
@@ -1195,7 +1196,6 @@ struct nvdebug_state {
                         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; \
                 } \
         }
@@ -1250,96 +1250,18 @@ uint64_t search_v1_page_directory(
         struct nvdebug_state *g,
         page_dir_config_t pd_config,
         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 <os/linux/os_linux.h>  // 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