Improve debugging messages for reverse page translation

2 files changed, 15 insertions, 6 deletions

diff --git a/mmu.c b/mmu.c
index 23adaf2..30e9362 100644
--- a/mmu.c
+++ b/mmu.c
@@ -64,21 +64,22 @@ uint64_t search_page_directory_subtree(struct nvdebug_state *g,
         // Hack to workaround PDE0 being double-size and strangely formatted
         if (NV_MMU_PT_V2_ENTRY_SZ[level] == 16)
                 pde_offset += 8;
-        entry.raw = readl(pde_offset);
+        entry.raw_w = readq(pde_offset);
         // If we reached an invalid (unpopulated) PDE, walk back up the tree
         if (entry.target == PD_AND_TARGET_INVALID)
                 return 0;
         // Succeed when we reach a PTE with the address we want
         if (entry.is_pte) {
-                printk(KERN_INFO "[nvdebug] PTE for phy addr %llx (raw: %x)\n", ((u64)entry.addr) << 12, entry.raw);
+                // TODO: Handle huge pages here
+                printk(KERN_INFO "[nvdebug] PTE for phy addr %#018llx, ap '%s', vol '%d', priv '%d', ro '%d', no_atomics '%d' (raw: %#018llx)\n", ((u64)entry.addr_w) << 12, pd_target_to_text(entry.target), entry.is_volatile, entry.is_privileged, entry.is_readonly, entry.atomics_disabled, entry.raw_w);
                 return (uint64_t)entry.addr << 12 == addr_to_find;
         }
-        printk(KERN_INFO "[nvdebug] Found PDE pointing to %llx in ap '%d' at lvl %d (raw: %x)\n", ((u64)entry.addr) << 12, entry.target, level, entry.raw);
+        printk(KERN_INFO "[nvdebug] Found PDE pointing to %#018llx in ap '%s' vol '%d' at lvl %d (raw: %#018llx)\n", ((u64)entry.addr_w) << 12, pd_target_to_text(entry.target), entry.is_volatile, level, entry.raw_w);
         // Depth-first search of the page table
         for (i = 0; i < NV_MMU_PT_V2_SZ[level + 1]; i++) {
                 next = off2addr(g, ((uint64_t)entry.addr << 12) + NV_MMU_PT_V2_ENTRY_SZ[level + 1] * i);
                 // off2addr can fail
-                if (!next) {
+                if (!next || !entry.addr_w) {
                         printk(KERN_ERR "[nvdebug] %s: Unable to resolve GPU PA to CPU PA\n", __func__);
                         return 0;
                 }
@@ -111,6 +112,7 @@ uint64_t search_page_directory(struct nvdebug_state *g,
                 printk(KERN_WARNING "[nvdebug] Attempting to search for unaligned address %llx in search_page_directory()!\n", addr_to_find);
                 return 0;
         }
+        printk(KERN_INFO "[nvdebug] Searching for addr %#018llx in page table with base %#018llx\n", (u64)addr_to_find, (u64)pde_offset);
         // Search the top-level page directory (PDE3)
         for (i = 0; i < NV_MMU_PT_V2_SZ[0]; i++)
                 if ((res = search_page_directory_subtree(g, pde_offset + NV_MMU_PT_V2_ENTRY_SZ[0] * i, off2addr, addr_to_find, 0)))
diff --git a/nvdebug.h b/nvdebug.h
index fcc6cff..67f2111 100644
--- a/nvdebug.h
+++ b/nvdebug.h
@@ -935,7 +935,6 @@ static inline const char *pd_target_to_text(enum PD_TARGET t) {
 //
 // V3 introduced with Hopper, but Hopper and Blackwell also support V2
 //
-// FIXME: This structure is 32 bits, but PDE/PTEs are actually 64 bits!
 typedef union {
         // Page Directory Entry (PDE)
         struct {
@@ -943,6 +942,7 @@ typedef union {
                 bool is_volatile:1;
                  uint32_t padding1:4;
                 uint32_t addr:24;
+                 uint32_t __unused1;
         } __attribute__((packed));
         // Page Table Entry (PTE)
         struct {
@@ -954,8 +954,15 @@ typedef union {
                 bool is_readonly:1;
                 bool atomics_disabled:1;
                  uint32_t __addr:24;
+                 uint32_t __unused2;
         } __attribute__((packed));
-        uint32_t raw;
+        // 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