Reduce the page table pool to 4%. Batch tlb invalidations.

2 files changed, 44 insertions, 2 deletions

diff --git a/viengoos/sysdeps/x86_64/thread-arch.c b/viengoos/sysdeps/x86_64/thread-arch.c
index 8c7f9d9..7086288 100644
--- a/viengoos/sysdeps/x86_64/thread-arch.c
+++ b/viengoos/sysdeps/x86_64/thread-arch.c
@@ -167,6 +167,7 @@ thread_resume (void)
   }
 #endif
 
+  vm_as_install (current_thread);
 
 #ifndef NDEBUG
   if (current_thread->pml4)
diff --git a/viengoos/sysdeps/x86_64/vm.c b/viengoos/sysdeps/x86_64/vm.c
index 754f02f..0a4e261 100644
--- a/viengoos/sysdeps/x86_64/vm.c
+++ b/viengoos/sysdeps/x86_64/vm.c
@@ -86,7 +86,10 @@ pool_free (void *page)
 static void
 pool_init (void)
 {
-  int frames = memory_total / 8;
+  /* Use about 4% of memory for page tables.  On a machine with 1GB of
+     RAM, this translates to being able to map about 10GB of memory,
+     assuming that leaf page tables are half full.  */
+  int frames = memory_total / 25;
 
   debug (0, "Reserving %d (of %"PRId32") frames for page tables",
 	 frames, memory_total);
@@ -204,11 +207,49 @@ vm_as_maybe_init (struct thread *thread)
 /* The current address space.  */
 static struct thread *current_as;
 
+/* The (direct) cost of a cr3 in terms of invlpg's is 5-10 (at least
+   on AMD64 processors).  There are indirect costs associated with
+   reloading cr3 as opposed to invalidating individual pages, namely,
+   that potentially useful translations will be flushed.  As such, we
+   use the upper bound.  */
+#define CR3_RELOAD_IN_INVLPGS 10
+static uintptr_t shootdown_queue[CR3_RELOAD_IN_INVLPGS];
+static int shootdown_queue_length;
+
+static void
+shootdown_enqueue (uintptr_t addr)
+{
+  assert ((addr & (PAGESIZE - 1)) == 0);
+  assert (addr < KERNEL_OFFSET);
+
+  if (shootdown_queue_length >= CR3_RELOAD_IN_INVLPGS)
+    shootdown_queue_length = CR3_RELOAD_IN_INVLPGS + 1;
+  else
+    shootdown_queue[shootdown_queue_length ++] = addr;
+}
+
 void
 vm_as_install (struct thread *thread)
 {
   if (current_as == thread)
-    return;
+    {
+      if (unlikely (shootdown_queue_length))
+	{
+	  if (shootdown_queue_length > CR3_RELOAD_IN_INVLPGS)
+	    asm ("mov %0, %%cr3" :: "r" (current_thread->pml4));
+	  else
+	    {
+	      int i;
+	      for (i = 0; i < CR3_RELOAD_IN_INVLPGS; i ++)
+		pte_shootdown (shootdown_queue[i]);
+	    }
+
+	  shootdown_queue_length = 0;
+	}
+
+      return;
+    }
+  shootdown_queue_length = 0;
 
   vm_as_maybe_init (thread);