i386: prepare shared mm/pgtable.c
authorThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:13:53 +0000 (11:13 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:13:53 +0000 (11:13 +0200)
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/i386/mm/Makefile
arch/i386/mm/pgtable.c [deleted file]
arch/i386/mm/pgtable_32.c [new file with mode: 0644]

index c5ae6432203daaabb1a7d1bf893c20b63a5448fa..78c242ccaed0aeddc25ac5d538d6acaccae8b094 100644 (file)
@@ -2,7 +2,7 @@
 # Makefile for the linux i386-specific parts of the memory manager.
 #
 
-obj-y  := init_32.o pgtable.o fault.o ioremap_32.o extable_32.o pageattr.o mmap_32.o
+obj-y  := init_32.o pgtable_32.o fault.o ioremap_32.o extable_32.o pageattr.o mmap_32.o
 
 obj-$(CONFIG_NUMA) += discontig_32.o
 obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o
diff --git a/arch/i386/mm/pgtable.c b/arch/i386/mm/pgtable.c
deleted file mode 100644 (file)
index 01437c4..0000000
+++ /dev/null
@@ -1,373 +0,0 @@
-/*
- *  linux/arch/i386/mm/pgtable.c
- */
-
-#include <linux/sched.h>
-#include <linux/kernel.h>
-#include <linux/errno.h>
-#include <linux/mm.h>
-#include <linux/swap.h>
-#include <linux/smp.h>
-#include <linux/highmem.h>
-#include <linux/slab.h>
-#include <linux/pagemap.h>
-#include <linux/spinlock.h>
-#include <linux/module.h>
-#include <linux/quicklist.h>
-
-#include <asm/system.h>
-#include <asm/pgtable.h>
-#include <asm/pgalloc.h>
-#include <asm/fixmap.h>
-#include <asm/e820.h>
-#include <asm/tlb.h>
-#include <asm/tlbflush.h>
-
-void show_mem(void)
-{
-       int total = 0, reserved = 0;
-       int shared = 0, cached = 0;
-       int highmem = 0;
-       struct page *page;
-       pg_data_t *pgdat;
-       unsigned long i;
-       unsigned long flags;
-
-       printk(KERN_INFO "Mem-info:\n");
-       show_free_areas();
-       printk(KERN_INFO "Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
-       for_each_online_pgdat(pgdat) {
-               pgdat_resize_lock(pgdat, &flags);
-               for (i = 0; i < pgdat->node_spanned_pages; ++i) {
-                       page = pgdat_page_nr(pgdat, i);
-                       total++;
-                       if (PageHighMem(page))
-                               highmem++;
-                       if (PageReserved(page))
-                               reserved++;
-                       else if (PageSwapCache(page))
-                               cached++;
-                       else if (page_count(page))
-                               shared += page_count(page) - 1;
-               }
-               pgdat_resize_unlock(pgdat, &flags);
-       }
-       printk(KERN_INFO "%d pages of RAM\n", total);
-       printk(KERN_INFO "%d pages of HIGHMEM\n", highmem);
-       printk(KERN_INFO "%d reserved pages\n", reserved);
-       printk(KERN_INFO "%d pages shared\n", shared);
-       printk(KERN_INFO "%d pages swap cached\n", cached);
-
-       printk(KERN_INFO "%lu pages dirty\n", global_page_state(NR_FILE_DIRTY));
-       printk(KERN_INFO "%lu pages writeback\n",
-                                       global_page_state(NR_WRITEBACK));
-       printk(KERN_INFO "%lu pages mapped\n", global_page_state(NR_FILE_MAPPED));
-       printk(KERN_INFO "%lu pages slab\n",
-               global_page_state(NR_SLAB_RECLAIMABLE) +
-               global_page_state(NR_SLAB_UNRECLAIMABLE));
-       printk(KERN_INFO "%lu pages pagetables\n",
-                                       global_page_state(NR_PAGETABLE));
-}
-
-/*
- * Associate a virtual page frame with a given physical page frame 
- * and protection flags for that frame.
- */ 
-static void set_pte_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
-{
-       pgd_t *pgd;
-       pud_t *pud;
-       pmd_t *pmd;
-       pte_t *pte;
-
-       pgd = swapper_pg_dir + pgd_index(vaddr);
-       if (pgd_none(*pgd)) {
-               BUG();
-               return;
-       }
-       pud = pud_offset(pgd, vaddr);
-       if (pud_none(*pud)) {
-               BUG();
-               return;
-       }
-       pmd = pmd_offset(pud, vaddr);
-       if (pmd_none(*pmd)) {
-               BUG();
-               return;
-       }
-       pte = pte_offset_kernel(pmd, vaddr);
-       if (pgprot_val(flags))
-               /* <pfn,flags> stored as-is, to permit clearing entries */
-               set_pte(pte, pfn_pte(pfn, flags));
-       else
-               pte_clear(&init_mm, vaddr, pte);
-
-       /*
-        * It's enough to flush this one mapping.
-        * (PGE mappings get flushed as well)
-        */
-       __flush_tlb_one(vaddr);
-}
-
-/*
- * Associate a large virtual page frame with a given physical page frame 
- * and protection flags for that frame. pfn is for the base of the page,
- * vaddr is what the page gets mapped to - both must be properly aligned. 
- * The pmd must already be instantiated. Assumes PAE mode.
- */ 
-void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
-{
-       pgd_t *pgd;
-       pud_t *pud;
-       pmd_t *pmd;
-
-       if (vaddr & (PMD_SIZE-1)) {             /* vaddr is misaligned */
-               printk(KERN_WARNING "set_pmd_pfn: vaddr misaligned\n");
-               return; /* BUG(); */
-       }
-       if (pfn & (PTRS_PER_PTE-1)) {           /* pfn is misaligned */
-               printk(KERN_WARNING "set_pmd_pfn: pfn misaligned\n");
-               return; /* BUG(); */
-       }
-       pgd = swapper_pg_dir + pgd_index(vaddr);
-       if (pgd_none(*pgd)) {
-               printk(KERN_WARNING "set_pmd_pfn: pgd_none\n");
-               return; /* BUG(); */
-       }
-       pud = pud_offset(pgd, vaddr);
-       pmd = pmd_offset(pud, vaddr);
-       set_pmd(pmd, pfn_pmd(pfn, flags));
-       /*
-        * It's enough to flush this one mapping.
-        * (PGE mappings get flushed as well)
-        */
-       __flush_tlb_one(vaddr);
-}
-
-static int fixmaps;
-unsigned long __FIXADDR_TOP = 0xfffff000;
-EXPORT_SYMBOL(__FIXADDR_TOP);
-
-void __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t flags)
-{
-       unsigned long address = __fix_to_virt(idx);
-
-       if (idx >= __end_of_fixed_addresses) {
-               BUG();
-               return;
-       }
-       set_pte_pfn(address, phys >> PAGE_SHIFT, flags);
-       fixmaps++;
-}
-
-/**
- * reserve_top_address - reserves a hole in the top of kernel address space
- * @reserve - size of hole to reserve
- *
- * Can be used to relocate the fixmap area and poke a hole in the top
- * of kernel address space to make room for a hypervisor.
- */
-void reserve_top_address(unsigned long reserve)
-{
-       BUG_ON(fixmaps > 0);
-       printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
-              (int)-reserve);
-       __FIXADDR_TOP = -reserve - PAGE_SIZE;
-       __VMALLOC_RESERVE += reserve;
-}
-
-pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
-{
-       return (pte_t *)__get_free_page(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO);
-}
-
-struct page *pte_alloc_one(struct mm_struct *mm, unsigned long address)
-{
-       struct page *pte;
-
-#ifdef CONFIG_HIGHPTE
-       pte = alloc_pages(GFP_KERNEL|__GFP_HIGHMEM|__GFP_REPEAT|__GFP_ZERO, 0);
-#else
-       pte = alloc_pages(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO, 0);
-#endif
-       return pte;
-}
-
-void pmd_ctor(void *pmd, struct kmem_cache *cache, unsigned long flags)
-{
-       memset(pmd, 0, PTRS_PER_PMD*sizeof(pmd_t));
-}
-
-/*
- * List of all pgd's needed for non-PAE so it can invalidate entries
- * in both cached and uncached pgd's; not needed for PAE since the
- * kernel pmd is shared. If PAE were not to share the pmd a similar
- * tactic would be needed. This is essentially codepath-based locking
- * against pageattr.c; it is the unique case in which a valid change
- * of kernel pagetables can't be lazily synchronized by vmalloc faults.
- * vmalloc faults work because attached pagetables are never freed.
- * -- wli
- */
-DEFINE_SPINLOCK(pgd_lock);
-struct page *pgd_list;
-
-static inline void pgd_list_add(pgd_t *pgd)
-{
-       struct page *page = virt_to_page(pgd);
-       page->index = (unsigned long)pgd_list;
-       if (pgd_list)
-               set_page_private(pgd_list, (unsigned long)&page->index);
-       pgd_list = page;
-       set_page_private(page, (unsigned long)&pgd_list);
-}
-
-static inline void pgd_list_del(pgd_t *pgd)
-{
-       struct page *next, **pprev, *page = virt_to_page(pgd);
-       next = (struct page *)page->index;
-       pprev = (struct page **)page_private(page);
-       *pprev = next;
-       if (next)
-               set_page_private(next, (unsigned long)pprev);
-}
-
-
-
-#if (PTRS_PER_PMD == 1)
-/* Non-PAE pgd constructor */
-static void pgd_ctor(void *pgd)
-{
-       unsigned long flags;
-
-       /* !PAE, no pagetable sharing */
-       memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
-
-       spin_lock_irqsave(&pgd_lock, flags);
-
-       /* must happen under lock */
-       clone_pgd_range((pgd_t *)pgd + USER_PTRS_PER_PGD,
-                       swapper_pg_dir + USER_PTRS_PER_PGD,
-                       KERNEL_PGD_PTRS);
-       paravirt_alloc_pd_clone(__pa(pgd) >> PAGE_SHIFT,
-                               __pa(swapper_pg_dir) >> PAGE_SHIFT,
-                               USER_PTRS_PER_PGD,
-                               KERNEL_PGD_PTRS);
-       pgd_list_add(pgd);
-       spin_unlock_irqrestore(&pgd_lock, flags);
-}
-#else  /* PTRS_PER_PMD > 1 */
-/* PAE pgd constructor */
-static void pgd_ctor(void *pgd)
-{
-       /* PAE, kernel PMD may be shared */
-
-       if (SHARED_KERNEL_PMD) {
-               clone_pgd_range((pgd_t *)pgd + USER_PTRS_PER_PGD,
-                               swapper_pg_dir + USER_PTRS_PER_PGD,
-                               KERNEL_PGD_PTRS);
-       } else {
-               unsigned long flags;
-
-               memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
-               spin_lock_irqsave(&pgd_lock, flags);
-               pgd_list_add(pgd);
-               spin_unlock_irqrestore(&pgd_lock, flags);
-       }
-}
-#endif /* PTRS_PER_PMD */
-
-static void pgd_dtor(void *pgd)
-{
-       unsigned long flags; /* can be called from interrupt context */
-
-       if (SHARED_KERNEL_PMD)
-               return;
-
-       paravirt_release_pd(__pa(pgd) >> PAGE_SHIFT);
-       spin_lock_irqsave(&pgd_lock, flags);
-       pgd_list_del(pgd);
-       spin_unlock_irqrestore(&pgd_lock, flags);
-}
-
-#define UNSHARED_PTRS_PER_PGD                          \
-       (SHARED_KERNEL_PMD ? USER_PTRS_PER_PGD : PTRS_PER_PGD)
-
-/* If we allocate a pmd for part of the kernel address space, then
-   make sure its initialized with the appropriate kernel mappings.
-   Otherwise use a cached zeroed pmd.  */
-static pmd_t *pmd_cache_alloc(int idx)
-{
-       pmd_t *pmd;
-
-       if (idx >= USER_PTRS_PER_PGD) {
-               pmd = (pmd_t *)__get_free_page(GFP_KERNEL);
-
-               if (pmd)
-                       memcpy(pmd,
-                              (void *)pgd_page_vaddr(swapper_pg_dir[idx]),
-                              sizeof(pmd_t) * PTRS_PER_PMD);
-       } else
-               pmd = kmem_cache_alloc(pmd_cache, GFP_KERNEL);
-
-       return pmd;
-}
-
-static void pmd_cache_free(pmd_t *pmd, int idx)
-{
-       if (idx >= USER_PTRS_PER_PGD)
-               free_page((unsigned long)pmd);
-       else
-               kmem_cache_free(pmd_cache, pmd);
-}
-
-pgd_t *pgd_alloc(struct mm_struct *mm)
-{
-       int i;
-       pgd_t *pgd = quicklist_alloc(0, GFP_KERNEL, pgd_ctor);
-
-       if (PTRS_PER_PMD == 1 || !pgd)
-               return pgd;
-
-       for (i = 0; i < UNSHARED_PTRS_PER_PGD; ++i) {
-               pmd_t *pmd = pmd_cache_alloc(i);
-
-               if (!pmd)
-                       goto out_oom;
-
-               paravirt_alloc_pd(__pa(pmd) >> PAGE_SHIFT);
-               set_pgd(&pgd[i], __pgd(1 + __pa(pmd)));
-       }
-       return pgd;
-
-out_oom:
-       for (i--; i >= 0; i--) {
-               pgd_t pgdent = pgd[i];
-               void* pmd = (void *)__va(pgd_val(pgdent)-1);
-               paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
-               pmd_cache_free(pmd, i);
-       }
-       quicklist_free(0, pgd_dtor, pgd);
-       return NULL;
-}
-
-void pgd_free(pgd_t *pgd)
-{
-       int i;
-
-       /* in the PAE case user pgd entries are overwritten before usage */
-       if (PTRS_PER_PMD > 1)
-               for (i = 0; i < UNSHARED_PTRS_PER_PGD; ++i) {
-                       pgd_t pgdent = pgd[i];
-                       void* pmd = (void *)__va(pgd_val(pgdent)-1);
-                       paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
-                       pmd_cache_free(pmd, i);
-               }
-       /* in the non-PAE case, free_pgtables() clears user pgd entries */
-       quicklist_free(0, pgd_dtor, pgd);
-}
-
-void check_pgt_cache(void)
-{
-       quicklist_trim(0, pgd_dtor, 25, 16);
-}
-
diff --git a/arch/i386/mm/pgtable_32.c b/arch/i386/mm/pgtable_32.c
new file mode 100644 (file)
index 0000000..01437c4
--- /dev/null
@@ -0,0 +1,373 @@
+/*
+ *  linux/arch/i386/mm/pgtable.c
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/mm.h>
+#include <linux/swap.h>
+#include <linux/smp.h>
+#include <linux/highmem.h>
+#include <linux/slab.h>
+#include <linux/pagemap.h>
+#include <linux/spinlock.h>
+#include <linux/module.h>
+#include <linux/quicklist.h>
+
+#include <asm/system.h>
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/fixmap.h>
+#include <asm/e820.h>
+#include <asm/tlb.h>
+#include <asm/tlbflush.h>
+
+void show_mem(void)
+{
+       int total = 0, reserved = 0;
+       int shared = 0, cached = 0;
+       int highmem = 0;
+       struct page *page;
+       pg_data_t *pgdat;
+       unsigned long i;
+       unsigned long flags;
+
+       printk(KERN_INFO "Mem-info:\n");
+       show_free_areas();
+       printk(KERN_INFO "Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
+       for_each_online_pgdat(pgdat) {
+               pgdat_resize_lock(pgdat, &flags);
+               for (i = 0; i < pgdat->node_spanned_pages; ++i) {
+                       page = pgdat_page_nr(pgdat, i);
+                       total++;
+                       if (PageHighMem(page))
+                               highmem++;
+                       if (PageReserved(page))
+                               reserved++;
+                       else if (PageSwapCache(page))
+                               cached++;
+                       else if (page_count(page))
+                               shared += page_count(page) - 1;
+               }
+               pgdat_resize_unlock(pgdat, &flags);
+       }
+       printk(KERN_INFO "%d pages of RAM\n", total);
+       printk(KERN_INFO "%d pages of HIGHMEM\n", highmem);
+       printk(KERN_INFO "%d reserved pages\n", reserved);
+       printk(KERN_INFO "%d pages shared\n", shared);
+       printk(KERN_INFO "%d pages swap cached\n", cached);
+
+       printk(KERN_INFO "%lu pages dirty\n", global_page_state(NR_FILE_DIRTY));
+       printk(KERN_INFO "%lu pages writeback\n",
+                                       global_page_state(NR_WRITEBACK));
+       printk(KERN_INFO "%lu pages mapped\n", global_page_state(NR_FILE_MAPPED));
+       printk(KERN_INFO "%lu pages slab\n",
+               global_page_state(NR_SLAB_RECLAIMABLE) +
+               global_page_state(NR_SLAB_UNRECLAIMABLE));
+       printk(KERN_INFO "%lu pages pagetables\n",
+                                       global_page_state(NR_PAGETABLE));
+}
+
+/*
+ * Associate a virtual page frame with a given physical page frame 
+ * and protection flags for that frame.
+ */ 
+static void set_pte_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+       pte_t *pte;
+
+       pgd = swapper_pg_dir + pgd_index(vaddr);
+       if (pgd_none(*pgd)) {
+               BUG();
+               return;
+       }
+       pud = pud_offset(pgd, vaddr);
+       if (pud_none(*pud)) {
+               BUG();
+               return;
+       }
+       pmd = pmd_offset(pud, vaddr);
+       if (pmd_none(*pmd)) {
+               BUG();
+               return;
+       }
+       pte = pte_offset_kernel(pmd, vaddr);
+       if (pgprot_val(flags))
+               /* <pfn,flags> stored as-is, to permit clearing entries */
+               set_pte(pte, pfn_pte(pfn, flags));
+       else
+               pte_clear(&init_mm, vaddr, pte);
+
+       /*
+        * It's enough to flush this one mapping.
+        * (PGE mappings get flushed as well)
+        */
+       __flush_tlb_one(vaddr);
+}
+
+/*
+ * Associate a large virtual page frame with a given physical page frame 
+ * and protection flags for that frame. pfn is for the base of the page,
+ * vaddr is what the page gets mapped to - both must be properly aligned. 
+ * The pmd must already be instantiated. Assumes PAE mode.
+ */ 
+void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+
+       if (vaddr & (PMD_SIZE-1)) {             /* vaddr is misaligned */
+               printk(KERN_WARNING "set_pmd_pfn: vaddr misaligned\n");
+               return; /* BUG(); */
+       }
+       if (pfn & (PTRS_PER_PTE-1)) {           /* pfn is misaligned */
+               printk(KERN_WARNING "set_pmd_pfn: pfn misaligned\n");
+               return; /* BUG(); */
+       }
+       pgd = swapper_pg_dir + pgd_index(vaddr);
+       if (pgd_none(*pgd)) {
+               printk(KERN_WARNING "set_pmd_pfn: pgd_none\n");
+               return; /* BUG(); */
+       }
+       pud = pud_offset(pgd, vaddr);
+       pmd = pmd_offset(pud, vaddr);
+       set_pmd(pmd, pfn_pmd(pfn, flags));
+       /*
+        * It's enough to flush this one mapping.
+        * (PGE mappings get flushed as well)
+        */
+       __flush_tlb_one(vaddr);
+}
+
+static int fixmaps;
+unsigned long __FIXADDR_TOP = 0xfffff000;
+EXPORT_SYMBOL(__FIXADDR_TOP);
+
+void __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t flags)
+{
+       unsigned long address = __fix_to_virt(idx);
+
+       if (idx >= __end_of_fixed_addresses) {
+               BUG();
+               return;
+       }
+       set_pte_pfn(address, phys >> PAGE_SHIFT, flags);
+       fixmaps++;
+}
+
+/**
+ * reserve_top_address - reserves a hole in the top of kernel address space
+ * @reserve - size of hole to reserve
+ *
+ * Can be used to relocate the fixmap area and poke a hole in the top
+ * of kernel address space to make room for a hypervisor.
+ */
+void reserve_top_address(unsigned long reserve)
+{
+       BUG_ON(fixmaps > 0);
+       printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
+              (int)-reserve);
+       __FIXADDR_TOP = -reserve - PAGE_SIZE;
+       __VMALLOC_RESERVE += reserve;
+}
+
+pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
+{
+       return (pte_t *)__get_free_page(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO);
+}
+
+struct page *pte_alloc_one(struct mm_struct *mm, unsigned long address)
+{
+       struct page *pte;
+
+#ifdef CONFIG_HIGHPTE
+       pte = alloc_pages(GFP_KERNEL|__GFP_HIGHMEM|__GFP_REPEAT|__GFP_ZERO, 0);
+#else
+       pte = alloc_pages(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO, 0);
+#endif
+       return pte;
+}
+
+void pmd_ctor(void *pmd, struct kmem_cache *cache, unsigned long flags)
+{
+       memset(pmd, 0, PTRS_PER_PMD*sizeof(pmd_t));
+}
+
+/*
+ * List of all pgd's needed for non-PAE so it can invalidate entries
+ * in both cached and uncached pgd's; not needed for PAE since the
+ * kernel pmd is shared. If PAE were not to share the pmd a similar
+ * tactic would be needed. This is essentially codepath-based locking
+ * against pageattr.c; it is the unique case in which a valid change
+ * of kernel pagetables can't be lazily synchronized by vmalloc faults.
+ * vmalloc faults work because attached pagetables are never freed.
+ * -- wli
+ */
+DEFINE_SPINLOCK(pgd_lock);
+struct page *pgd_list;
+
+static inline void pgd_list_add(pgd_t *pgd)
+{
+       struct page *page = virt_to_page(pgd);
+       page->index = (unsigned long)pgd_list;
+       if (pgd_list)
+               set_page_private(pgd_list, (unsigned long)&page->index);
+       pgd_list = page;
+       set_page_private(page, (unsigned long)&pgd_list);
+}
+
+static inline void pgd_list_del(pgd_t *pgd)
+{
+       struct page *next, **pprev, *page = virt_to_page(pgd);
+       next = (struct page *)page->index;
+       pprev = (struct page **)page_private(page);
+       *pprev = next;
+       if (next)
+               set_page_private(next, (unsigned long)pprev);
+}
+
+
+
+#if (PTRS_PER_PMD == 1)
+/* Non-PAE pgd constructor */
+static void pgd_ctor(void *pgd)
+{
+       unsigned long flags;
+
+       /* !PAE, no pagetable sharing */
+       memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
+
+       spin_lock_irqsave(&pgd_lock, flags);
+
+       /* must happen under lock */
+       clone_pgd_range((pgd_t *)pgd + USER_PTRS_PER_PGD,
+                       swapper_pg_dir + USER_PTRS_PER_PGD,
+                       KERNEL_PGD_PTRS);
+       paravirt_alloc_pd_clone(__pa(pgd) >> PAGE_SHIFT,
+                               __pa(swapper_pg_dir) >> PAGE_SHIFT,
+                               USER_PTRS_PER_PGD,
+                               KERNEL_PGD_PTRS);
+       pgd_list_add(pgd);
+       spin_unlock_irqrestore(&pgd_lock, flags);
+}
+#else  /* PTRS_PER_PMD > 1 */
+/* PAE pgd constructor */
+static void pgd_ctor(void *pgd)
+{
+       /* PAE, kernel PMD may be shared */
+
+       if (SHARED_KERNEL_PMD) {
+               clone_pgd_range((pgd_t *)pgd + USER_PTRS_PER_PGD,
+                               swapper_pg_dir + USER_PTRS_PER_PGD,
+                               KERNEL_PGD_PTRS);
+       } else {
+               unsigned long flags;
+
+               memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
+               spin_lock_irqsave(&pgd_lock, flags);
+               pgd_list_add(pgd);
+               spin_unlock_irqrestore(&pgd_lock, flags);
+       }
+}
+#endif /* PTRS_PER_PMD */
+
+static void pgd_dtor(void *pgd)
+{
+       unsigned long flags; /* can be called from interrupt context */
+
+       if (SHARED_KERNEL_PMD)
+               return;
+
+       paravirt_release_pd(__pa(pgd) >> PAGE_SHIFT);
+       spin_lock_irqsave(&pgd_lock, flags);
+       pgd_list_del(pgd);
+       spin_unlock_irqrestore(&pgd_lock, flags);
+}
+
+#define UNSHARED_PTRS_PER_PGD                          \
+       (SHARED_KERNEL_PMD ? USER_PTRS_PER_PGD : PTRS_PER_PGD)
+
+/* If we allocate a pmd for part of the kernel address space, then
+   make sure its initialized with the appropriate kernel mappings.
+   Otherwise use a cached zeroed pmd.  */
+static pmd_t *pmd_cache_alloc(int idx)
+{
+       pmd_t *pmd;
+
+       if (idx >= USER_PTRS_PER_PGD) {
+               pmd = (pmd_t *)__get_free_page(GFP_KERNEL);
+
+               if (pmd)
+                       memcpy(pmd,
+                              (void *)pgd_page_vaddr(swapper_pg_dir[idx]),
+                              sizeof(pmd_t) * PTRS_PER_PMD);
+       } else
+               pmd = kmem_cache_alloc(pmd_cache, GFP_KERNEL);
+
+       return pmd;
+}
+
+static void pmd_cache_free(pmd_t *pmd, int idx)
+{
+       if (idx >= USER_PTRS_PER_PGD)
+               free_page((unsigned long)pmd);
+       else
+               kmem_cache_free(pmd_cache, pmd);
+}
+
+pgd_t *pgd_alloc(struct mm_struct *mm)
+{
+       int i;
+       pgd_t *pgd = quicklist_alloc(0, GFP_KERNEL, pgd_ctor);
+
+       if (PTRS_PER_PMD == 1 || !pgd)
+               return pgd;
+
+       for (i = 0; i < UNSHARED_PTRS_PER_PGD; ++i) {
+               pmd_t *pmd = pmd_cache_alloc(i);
+
+               if (!pmd)
+                       goto out_oom;
+
+               paravirt_alloc_pd(__pa(pmd) >> PAGE_SHIFT);
+               set_pgd(&pgd[i], __pgd(1 + __pa(pmd)));
+       }
+       return pgd;
+
+out_oom:
+       for (i--; i >= 0; i--) {
+               pgd_t pgdent = pgd[i];
+               void* pmd = (void *)__va(pgd_val(pgdent)-1);
+               paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
+               pmd_cache_free(pmd, i);
+       }
+       quicklist_free(0, pgd_dtor, pgd);
+       return NULL;
+}
+
+void pgd_free(pgd_t *pgd)
+{
+       int i;
+
+       /* in the PAE case user pgd entries are overwritten before usage */
+       if (PTRS_PER_PMD > 1)
+               for (i = 0; i < UNSHARED_PTRS_PER_PGD; ++i) {
+                       pgd_t pgdent = pgd[i];
+                       void* pmd = (void *)__va(pgd_val(pgdent)-1);
+                       paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
+                       pmd_cache_free(pmd, i);
+               }
+       /* in the non-PAE case, free_pgtables() clears user pgd entries */
+       quicklist_free(0, pgd_dtor, pgd);
+}
+
+void check_pgt_cache(void)
+{
+       quicklist_trim(0, pgd_dtor, 25, 16);
+}
+