Always map the kernel itself using global pages.

This avoids TLB stalls after context switching, that are very visible
from flamegraphs.

src/arch/x86_64/rmm.rs

@@ -35,7 +35,7 @@ unsafe fn page_flags<A: Arch>(virt: VirtualAddress) -> PageFlags<A> {
     use crate::kernel_executable_offsets::*;
     let virt_addr = virt.data();
 
-    if virt_addr >= __text_start() && virt_addr < __text_end() {
+    (if virt_addr >= __text_start() && virt_addr < __text_end() {
         // Remap text read-only, execute
         PageFlags::new().execute(true)
     } else if virt_addr >= __rodata_start() && virt_addr < __rodata_end() {
@@ -44,7 +44,8 @@ unsafe fn page_flags<A: Arch>(virt: VirtualAddress) -> PageFlags<A> {
     } else {
         // Remap everything else read-write, no execute
         PageFlags::new().write(true)
-    }
+    })
+    .global(cfg!(not(feature = "pti")))
 }
 
 unsafe fn inner(