[RFC] global page support

[Samuel Thibault]

Hi,

Global pages permit to keep the kernel pages in the TLB whatever the
current process, here is a patch for supporting them. On my Pentium M, I
get a 5% speedup on dd < /dev/zero bs=1 > /dev/null .

Samuel

2006-12-19  Samuel Thibault  <samuel.thibault@ens-lyon.org>

        Add support for global pages.
        * i386/i386at/model_dep.c: Include <i386/locore.h>.
        (i386at_init): Set CR4_PGE if available.
        (c_boot_entry): Call discover_x86_cpu_type() before i386at_init().
        * i386/intel/pmap.c: Include <i386/locore.h>.
        (pmap_map_bd): Use INTEL_PTE_GLOBAL if available.
        (pmap_bootstrap): Likewise.

Index: i386/i386at/model_dep.c
===================================================================
RCS file: /cvsroot/hurd/gnumach/i386/i386at/Attic/model_dep.c,v
retrieving revision 1.9.2.10
diff -u -p -r1.9.2.10 model_dep.c
--- i386/i386at/model_dep.c     26 Nov 2006 20:01:47 -0000      1.9.2.10
+++ i386/i386at/model_dep.c     20 Dec 2006 23:41:50 -0000
@@ -50,6 +50,7 @@
 #include <i386/machspl.h>
 #include <i386/pmap.h>
 #include <i386/proc_reg.h>
+#include <i386/locore.h>
 
 /* Location of the kernel's symbol table.
    Both of these are 0 if none is available.  */
@@ -264,6 +265,8 @@ i386at_init()
        kernel_page_dir[lin2pdenum(0)] =
                kernel_page_dir[lin2pdenum(LINEAR_MIN_KERNEL_ADDRESS)];
        set_cr3((unsigned)kernel_page_dir);
+       if (CPU_HAS_FEATURE(CPU_FEATURE_PGE))
+               set_cr4(get_cr4() | CR4_PGE);
        set_cr0(get_cr0() | CR0_PG | CR0_WP);
        flush_instr_queue();
 
@@ -301,6 +304,7 @@ void c_boot_entry(vm_offset_t bi)
 {
        /* Stash the boot_image_info pointer.  */
        boot_info = *(struct multiboot_info*)phystokv(bi);
+       int cpu_type;
 
        /* XXX we currently assume phys_mem_va is always 0 here -
           if it isn't, we must tweak the pointers in the boot_info.  */
@@ -337,6 +341,8 @@ void c_boot_entry(vm_offset_t bi)
        }
 #endif /* MACH_KDB */
 
+       cpu_type = discover_x86_cpu_type ();
+
        /*
         * Do basic VM initialization
         */
@@ -367,7 +373,7 @@ void c_boot_entry(vm_offset_t bi)
        machine_slot[0].running = TRUE;
        machine_slot[0].cpu_subtype = CPU_SUBTYPE_AT386;
 
-       switch (discover_x86_cpu_type ())
+       switch (cpu_type)
          {
          case 3:
          default:
Index: i386/intel/pmap.c
===================================================================
RCS file: /cvsroot/hurd/gnumach/i386/intel/pmap.c,v
retrieving revision 1.4.2.12
diff -u -p -r1.4.2.12 pmap.c
--- i386/intel/pmap.c   20 Nov 2006 22:40:33 -0000      1.4.2.12
+++ i386/intel/pmap.c   20 Dec 2006 23:41:51 -0000
@@ -80,6 +80,7 @@
 #include <machine/thread.h>
 #include <i386/cpu_number.h>
 #include <i386/proc_reg.h>
+#include <i386/locore.h>
 
 #ifdef ORC
 #define        OLIVETTICACHE   1
@@ -547,6 +548,8 @@ vm_offset_t pmap_map_bd(virt, start, end
        register pt_entry_t     *pte;
 
        template = pa_to_pte(start) | INTEL_PTE_VALID;
+       if (CPU_HAS_FEATURE(CPU_FEATURE_PGE))
+               template |= INTEL_PTE_GLOBAL;
        if (prot & VM_PROT_WRITE)
            template |= INTEL_PTE_WRITE;
 
@@ -627,6 +630,7 @@ void pmap_bootstrap()
         */
        {
                vm_offset_t va;
+               pt_entry_t global = CPU_HAS_FEATURE(CPU_FEATURE_PGE) ? 
INTEL_PTE_GLOBAL : 0;
 
                /*
                 * Map virtual memory for all known physical memory, 1-1,
@@ -647,7 +651,7 @@ void pmap_bootstrap()
 
                        /* Initialize the page directory entry.  */
                        *pde = pa_to_pte((vm_offset_t)ptable)
-                               | INTEL_PTE_VALID | INTEL_PTE_WRITE;
+                               | INTEL_PTE_VALID | INTEL_PTE_WRITE | global;
 
                        /* Initialize the page table.  */
                        for (pte = ptable; (va < phys_last_addr) && (pte < 
ptable+NPTES); pte++)
@@ -664,12 +668,12 @@ void pmap_bootstrap()
                                            && (va + INTEL_PGBYTES <= 
(vm_offset_t)etext))
                                        {
                                                WRITE_PTE_FAST(pte, 
pa_to_pte(va)
-                                                       | INTEL_PTE_VALID);
+                                                       | INTEL_PTE_VALID | 
global);
                                        }
                                        else
                                        {
                                                WRITE_PTE_FAST(pte, 
pa_to_pte(va)
-                                                       | INTEL_PTE_VALID | 
INTEL_PTE_WRITE);
+                                                       | INTEL_PTE_VALID | 
INTEL_PTE_WRITE | global);
                                        }
                                        va += INTEL_PGBYTES;
                                }

-- 
Samuel Thibault <samuel.thibault@ens-lyon.org>
Progress (n.): The process through which the Internet has evolved from
smart people in front of dumb terminals to dumb people in front of smart
terminals.