Some minor changes to virtual memory handling

+ Should start to think about how virtual memory should be handled
  efficiently, I'm sort of flailing around and not actually doing
  anything beneficial. The mapping system I've so far implemented
  _works_, but it's not good and I'm not happy with it. I might come
  back and fix/improve it, but I think I'll go ahead and do other things for
  a while, might give me a better overhead view of what the virtual
  memory system should do.

+ TODO: should probably document the memory layout I'm going with at the
  moment.

1 files changed, 63 insertions, 5 deletions

diff --git a/arch/riscv/common/vmem.c b/arch/riscv/common/vmem.c
index 7b2d66b..6b022e1 100644
--- a/arch/riscv/common/vmem.c
+++ b/arch/riscv/common/vmem.c
@@ -1,33 +1,91 @@
 #include <apos/string.h>
 #include <apos/pmem.h>
 #include <apos/vmem.h>
+#include <apos/mem.h>
+#include <apos/debug.h>
 #include <pages.h>
 
 #define pte_ppn(pte) (((pm_t)(pte)) >> 10)
 #define pte_flags(pte) (((pm_t)(pte)) & 0xff)
 #define to_pte(p, f) ((pm_to_pnum(p) << 10) + (f))
 #define pte_addr(pte) (pnum_to_paddr(pte_ppn(pte)))
+#define vm_to_index(a, o) (pm_to_index(a, o))
+
+#if defined(KERNEL)
+
+static vm_t *tmp_pte = 0;
+void arch_init_vmem(struct vm_branch_t *branch, vm_t pte)
+{
+	(void)branch;
+	tmp_pte = (vm_t *)pte;
+}
+
+/* this seems like an awful idea, should probably try to come up with a smarter
+ * approach? flushing the cache every time is probably pretty slow and also
+ * this code doesn't take into account the cpu ASID, although that's probably
+ * not relevant here */
+
+/* what I probably should do is something like Linux, where the kernel has a
+ * direct (or linear, I suppose?) so that all direct memory accesses are also
+ * accessible from virtual memory. Not entirely sure how I should do that,
+ * wouldn't there be situations where the RAM is larger than the allocated
+ * memory region?
+ *
+ * Or maybe some kind of walker, where I maintain a virtual page table that
+ * mirrors the physical page table? Not sure yet, but this seems to work for
+ * _now_.
+ */
+#define QUOTE2(x) #x
+#define QUOTE(x) QUOTE2(x)
+static void *set_vptr(uint8_t flags, pm_t a)
+{
+	*tmp_pte = to_pte(a, flags);
+	__asm__ ("sfence.vma %0, %1"
+			:: "rk" (0), "rK" (TMP_PTE)
+			: "memory");
+
+	return (void *)TMP_PTE;
+}
+
+#else
+struct vm_branch_t *arch_get_tmp_pte(struct vm_branch_t *branch)
+{
+	enum mm_order_t top = __mm_max_order;
+	while(top){
+		size_t idx = vm_to_index(TMP_PTE, top);
+		branch = (struct vm_branch_t *)pte_addr(branch->leaf[idx]);
+		top--;
+	}
+
+	return branch;
+}
+
+#define set_vptr(flags, a) (a)
+#endif
 
-/* probably not actually this simple, right? */
 void map_vmem(struct vm_branch_t *branch,
 		pm_t paddr, vm_t vaddr, uint8_t flags, enum mm_order_t order)
 {
 	enum mm_order_t top = __mm_max_order;
 	while (top != order) {
-		size_t idx = pm_to_index(vaddr, top);
+		size_t idx = vm_to_index(vaddr, top);
 
 		if (!branch->leaf[idx]) {
 			pm_t new_leaf = alloc_page(MM_KPAGE, 0);
 			branch->leaf[idx] =
 				(struct vm_branch_t *)to_pte(new_leaf, VM_V);
-			memset((void *)new_leaf, 0, sizeof(struct vm_branch_t));
+
+			void *leaf_ptr = (void *)set_vptr(VM_R | VM_W | VM_V, new_leaf);
+			memset(leaf_ptr, 0, sizeof(struct vm_branch_t));
 		}
 
-		branch = (struct vm_branch_t *)pte_addr(branch->leaf[idx]);
+		pm_t pte = (pm_t)branch->leaf[idx];
+		pm_t branch_pptr = (pm_t)pte_addr(pte);
+		branch = (struct vm_branch_t *)set_vptr(VM_R | VM_W | VM_V, branch_pptr);
 		top--;
 	}
 
-	size_t idx = pm_to_index(vaddr, top);
+	size_t idx = vm_to_index(vaddr, top);
 	branch->leaf[idx] = (struct vm_branch_t *)to_pte(paddr, flags);
 }