]> nv-tegra.nvidia Code Review - linux-2.6.git/blobdiff - arch/x86/mm/pageattr.c
x86: PAT infrastructure patch
[linux-2.6.git] / arch / x86 / mm / pageattr.c
index 72880993af89bb9b23656e68e4e61f326d6cb5ac..f7d5ca170c22141db4e3360b15fc57966399177f 100644 (file)
@@ -8,6 +8,9 @@
 #include <linux/sched.h>
 #include <linux/slab.h>
 #include <linux/mm.h>
+#include <linux/interrupt.h>
+#include <linux/seq_file.h>
+#include <linux/debugfs.h>
 
 #include <asm/e820.h>
 #include <asm/processor.h>
 #include <asm/sections.h>
 #include <asm/uaccess.h>
 #include <asm/pgalloc.h>
+#include <asm/proto.h>
 
+/*
+ * The current flushing context - we pass it instead of 5 arguments:
+ */
 struct cpa_data {
        unsigned long   vaddr;
        pgprot_t        mask_set;
        pgprot_t        mask_clr;
        int             numpages;
        int             flushtlb;
+       unsigned long   pfn;
 };
 
-enum {
-       CPA_NO_SPLIT = 0,
-       CPA_SPLIT,
-};
+#ifdef CONFIG_X86_64
+
+static inline unsigned long highmap_start_pfn(void)
+{
+       return __pa(_text) >> PAGE_SHIFT;
+}
+
+static inline unsigned long highmap_end_pfn(void)
+{
+       return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
+}
+
+#endif
+
+#ifdef CONFIG_DEBUG_PAGEALLOC
+# define debug_pagealloc 1
+#else
+# define debug_pagealloc 0
+#endif
 
 static inline int
 within(unsigned long addr, unsigned long start, unsigned long end)
@@ -119,34 +142,19 @@ static void cpa_flush_range(unsigned long start, int numpages, int cache)
                /*
                 * Only flush present addresses:
                 */
-               if (pte && pte_present(*pte))
+               if (pte && (pte_val(*pte) & _PAGE_PRESENT))
                        clflush_cache_range((void *) addr, PAGE_SIZE);
        }
 }
 
-#define HIGH_MAP_START __START_KERNEL_map
-#define HIGH_MAP_END   (__START_KERNEL_map + KERNEL_TEXT_SIZE)
-
-
-/*
- * Converts a virtual address to a X86-64 highmap address
- */
-static unsigned long virt_to_highmap(void *address)
-{
-#ifdef CONFIG_X86_64
-       return __pa((unsigned long)address) + HIGH_MAP_START - phys_base;
-#else
-       return (unsigned long)address;
-#endif
-}
-
 /*
  * Certain areas of memory on x86 require very specific protection flags,
  * for example the BIOS area or kernel text. Callers don't always get this
  * right (again, ioremap() on BIOS memory is not uncommon) so this function
  * checks and fixes these known static required protection bits.
  */
-static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
+static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
+                                  unsigned long pfn)
 {
        pgprot_t forbidden = __pgprot(0);
 
@@ -154,34 +162,24 @@ static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
         * The BIOS area between 640k and 1Mb needs to be executable for
         * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
         */
-       if (within(__pa(address), BIOS_BEGIN, BIOS_END))
+       if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
                pgprot_val(forbidden) |= _PAGE_NX;
 
        /*
         * The kernel text needs to be executable for obvious reasons
-        * Does not cover __inittext since that is gone later on
+        * Does not cover __inittext since that is gone later on. On
+        * 64bit we do not enforce !NX on the low mapping
         */
        if (within(address, (unsigned long)_text, (unsigned long)_etext))
                pgprot_val(forbidden) |= _PAGE_NX;
-       /*
-        * Do the same for the x86-64 high kernel mapping
-        */
-       if (within(address, virt_to_highmap(_text), virt_to_highmap(_etext)))
-               pgprot_val(forbidden) |= _PAGE_NX;
-
 
-#ifdef CONFIG_DEBUG_RODATA
-       /* The .rodata section needs to be read-only */
-       if (within(address, (unsigned long)__start_rodata,
-                               (unsigned long)__end_rodata))
-               pgprot_val(forbidden) |= _PAGE_RW;
        /*
-        * Do the same for the x86-64 high kernel mapping
+        * The .rodata section needs to be read-only. Using the pfn
+        * catches all aliases.
         */
-       if (within(address, virt_to_highmap(__start_rodata),
-                               virt_to_highmap(__end_rodata)))
+       if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
+                  __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
                pgprot_val(forbidden) |= _PAGE_RW;
-#endif
 
        prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 
@@ -196,7 +194,7 @@ static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
  * or when the present bit is not set. Otherwise we would return a
  * pointer to a nonexisting mapping.
  */
-pte_t *lookup_address(unsigned long address, int *level)
+pte_t *lookup_address(unsigned long address, unsigned int *level)
 {
        pgd_t *pgd = pgd_offset_k(address);
        pud_t *pud;
@@ -206,9 +204,15 @@ pte_t *lookup_address(unsigned long address, int *level)
 
        if (pgd_none(*pgd))
                return NULL;
+
        pud = pud_offset(pgd, address);
        if (pud_none(*pud))
                return NULL;
+
+       *level = PG_LEVEL_1G;
+       if (pud_large(*pud) || !pud_present(*pud))
+               return (pte_t *)pud;
+
        pmd = pmd_offset(pud, address);
        if (pmd_none(*pmd))
                return NULL;
@@ -218,9 +222,13 @@ pte_t *lookup_address(unsigned long address, int *level)
                return (pte_t *)pmd;
 
        *level = PG_LEVEL_4K;
+
        return pte_offset_kernel(pmd, address);
 }
 
+/*
+ * Set the new pmd in all the pgds we know about:
+ */
 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 {
        /* change init_mm */
@@ -243,24 +251,15 @@ static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 #endif
 }
 
-static int try_preserve_large_page(pte_t *kpte, unsigned long address,
-                                  struct cpa_data *cpa)
+static int
+try_preserve_large_page(pte_t *kpte, unsigned long address,
+                       struct cpa_data *cpa)
 {
-       unsigned long nextpage_addr, numpages, pmask, psize, flags;
+       unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
        pte_t new_pte, old_pte, *tmp;
        pgprot_t old_prot, new_prot;
-       int level, res = CPA_SPLIT;
-
-       /*
-        * An Athlon 64 X2 showed hard hangs if we tried to preserve
-        * largepages and changed the PSE entry from RW to RO.
-        *
-        * As AMD CPUs have a long series of erratas in this area,
-        * (and none of the known ones seem to explain this hang),
-        * disable this code until the hang can be debugged:
-        */
-       if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
-               return res;
+       int i, do_split = 1;
+       unsigned int level;
 
        spin_lock_irqsave(&pgd_lock, flags);
        /*
@@ -276,9 +275,14 @@ static int try_preserve_large_page(pte_t *kpte, unsigned long address,
                psize = PMD_PAGE_SIZE;
                pmask = PMD_PAGE_MASK;
                break;
+#ifdef CONFIG_X86_64
        case PG_LEVEL_1G:
+               psize = PUD_PAGE_SIZE;
+               pmask = PUD_PAGE_MASK;
+               break;
+#endif
        default:
-               res = -EINVAL;
+               do_split = -EINVAL;
                goto out_unlock;
        }
 
@@ -299,14 +303,36 @@ static int try_preserve_large_page(pte_t *kpte, unsigned long address,
 
        pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
        pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
-       new_prot = static_protections(new_prot, address);
+
+       /*
+        * old_pte points to the large page base address. So we need
+        * to add the offset of the virtual address:
+        */
+       pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
+       cpa->pfn = pfn;
+
+       new_prot = static_protections(new_prot, address, pfn);
+
+       /*
+        * We need to check the full range, whether
+        * static_protection() requires a different pgprot for one of
+        * the pages in the range we try to preserve:
+        */
+       addr = address + PAGE_SIZE;
+       pfn++;
+       for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
+               pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
+
+               if (pgprot_val(chk_prot) != pgprot_val(new_prot))
+                       goto out_unlock;
+       }
 
        /*
         * If there are no changes, return. maxpages has been updated
         * above:
         */
        if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
-               res = CPA_NO_SPLIT;
+               do_split = 0;
                goto out_unlock;
        }
 
@@ -326,54 +352,149 @@ static int try_preserve_large_page(pte_t *kpte, unsigned long address,
                new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
                __set_pmd_pte(kpte, address, new_pte);
                cpa->flushtlb = 1;
-               res = CPA_NO_SPLIT;
+               do_split = 0;
        }
 
 out_unlock:
        spin_unlock_irqrestore(&pgd_lock, flags);
-       return res;
+
+       return do_split;
+}
+
+static LIST_HEAD(page_pool);
+static unsigned long pool_size, pool_pages, pool_low;
+static unsigned long pool_used, pool_failed;
+
+static void cpa_fill_pool(struct page **ret)
+{
+       gfp_t gfp = GFP_KERNEL;
+       unsigned long flags;
+       struct page *p;
+
+       /*
+        * Avoid recursion (on debug-pagealloc) and also signal
+        * our priority to get to these pagetables:
+        */
+       if (current->flags & PF_MEMALLOC)
+               return;
+       current->flags |= PF_MEMALLOC;
+
+       /*
+        * Allocate atomically from atomic contexts:
+        */
+       if (in_atomic() || irqs_disabled() || debug_pagealloc)
+               gfp =  GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
+
+       while (pool_pages < pool_size || (ret && !*ret)) {
+               p = alloc_pages(gfp, 0);
+               if (!p) {
+                       pool_failed++;
+                       break;
+               }
+               /*
+                * If the call site needs a page right now, provide it:
+                */
+               if (ret && !*ret) {
+                       *ret = p;
+                       continue;
+               }
+               spin_lock_irqsave(&pgd_lock, flags);
+               list_add(&p->lru, &page_pool);
+               pool_pages++;
+               spin_unlock_irqrestore(&pgd_lock, flags);
+       }
+
+       current->flags &= ~PF_MEMALLOC;
+}
+
+#define SHIFT_MB               (20 - PAGE_SHIFT)
+#define ROUND_MB_GB            ((1 << 10) - 1)
+#define SHIFT_MB_GB            10
+#define POOL_PAGES_PER_GB      16
+
+void __init cpa_init(void)
+{
+       struct sysinfo si;
+       unsigned long gb;
+
+       si_meminfo(&si);
+       /*
+        * Calculate the number of pool pages:
+        *
+        * Convert totalram (nr of pages) to MiB and round to the next
+        * GiB. Shift MiB to Gib and multiply the result by
+        * POOL_PAGES_PER_GB:
+        */
+       if (debug_pagealloc) {
+               gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
+               pool_size = POOL_PAGES_PER_GB * gb;
+       } else {
+               pool_size = 1;
+       }
+       pool_low = pool_size;
+
+       cpa_fill_pool(NULL);
+       printk(KERN_DEBUG
+              "CPA: page pool initialized %lu of %lu pages preallocated\n",
+              pool_pages, pool_size);
 }
 
 static int split_large_page(pte_t *kpte, unsigned long address)
 {
-       pgprot_t ref_prot;
-       gfp_t gfp_flags = GFP_KERNEL;
-       unsigned long flags, addr, pfn;
+       unsigned long flags, pfn, pfninc = 1;
+       unsigned int i, level;
        pte_t *pbase, *tmp;
+       pgprot_t ref_prot;
        struct page *base;
-       unsigned int i, level;
-
-#ifdef CONFIG_DEBUG_PAGEALLOC
-       gfp_flags = GFP_ATOMIC | __GFP_NOWARN;
-#endif
-       base = alloc_pages(gfp_flags, 0);
-       if (!base)
-               return -ENOMEM;
 
+       /*
+        * Get a page from the pool. The pool list is protected by the
+        * pgd_lock, which we have to take anyway for the split
+        * operation:
+        */
        spin_lock_irqsave(&pgd_lock, flags);
+       if (list_empty(&page_pool)) {
+               spin_unlock_irqrestore(&pgd_lock, flags);
+               base = NULL;
+               cpa_fill_pool(&base);
+               if (!base)
+                       return -ENOMEM;
+               spin_lock_irqsave(&pgd_lock, flags);
+       } else {
+               base = list_first_entry(&page_pool, struct page, lru);
+               list_del(&base->lru);
+               pool_pages--;
+
+               if (pool_pages < pool_low)
+                       pool_low = pool_pages;
+       }
+
        /*
         * Check for races, another CPU might have split this page
         * up for us already:
         */
        tmp = lookup_address(address, &level);
-       if (tmp != kpte) {
-               WARN_ON_ONCE(1);
+       if (tmp != kpte)
                goto out_unlock;
-       }
 
-       address = __pa(address);
-       addr = address & PMD_PAGE_MASK;
        pbase = (pte_t *)page_address(base);
 #ifdef CONFIG_X86_32
        paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 #endif
        ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 
+#ifdef CONFIG_X86_64
+       if (level == PG_LEVEL_1G) {
+               pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
+               pgprot_val(ref_prot) |= _PAGE_PSE;
+       }
+#endif
+
        /*
         * Get the target pfn from the original entry:
         */
        pfn = pte_pfn(*kpte);
-       for (i = 0; i < PTRS_PER_PTE; i++, pfn++)
+       for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
                set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
 
        /*
@@ -392,53 +513,60 @@ static int split_large_page(pte_t *kpte, unsigned long address)
        base = NULL;
 
 out_unlock:
+       /*
+        * If we dropped out via the lookup_address check under
+        * pgd_lock then stick the page back into the pool:
+        */
+       if (base) {
+               list_add(&base->lru, &page_pool);
+               pool_pages++;
+       } else
+               pool_used++;
        spin_unlock_irqrestore(&pgd_lock, flags);
 
-       if (base)
-               __free_pages(base, 0);
-
        return 0;
 }
 
-static int __change_page_attr(unsigned long address, struct cpa_data *cpa)
+static int __change_page_attr(struct cpa_data *cpa, int primary)
 {
-       struct page *kpte_page;
-       int level, res;
-       pte_t *kpte;
+       unsigned long address = cpa->vaddr;
+       int do_split, err;
+       unsigned int level;
+       pte_t *kpte, old_pte;
 
 repeat:
        kpte = lookup_address(address, &level);
        if (!kpte)
-               return -EINVAL;
+               return primary ? -EINVAL : 0;
 
-       kpte_page = virt_to_page(kpte);
-       BUG_ON(PageLRU(kpte_page));
-       BUG_ON(PageCompound(kpte_page));
+       old_pte = *kpte;
+       if (!pte_val(old_pte)) {
+               if (!primary)
+                       return 0;
+               printk(KERN_WARNING "CPA: called for zero pte. "
+                      "vaddr = %lx cpa->vaddr = %lx\n", address,
+                      cpa->vaddr);
+               WARN_ON(1);
+               return -EINVAL;
+       }
 
        if (level == PG_LEVEL_4K) {
-               pte_t new_pte, old_pte = *kpte;
+               pte_t new_pte;
                pgprot_t new_prot = pte_pgprot(old_pte);
-
-               if(!pte_val(old_pte)) {
-                       printk(KERN_WARNING "CPA: called for zero pte. "
-                              "vaddr = %lx cpa->vaddr = %lx\n", address,
-                               cpa->vaddr);
-                       WARN_ON(1);
-                       return -EINVAL;
-               }
+               unsigned long pfn = pte_pfn(old_pte);
 
                pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
                pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 
-               new_prot = static_protections(new_prot, address);
+               new_prot = static_protections(new_prot, address, pfn);
 
                /*
                 * We need to keep the pfn from the existing PTE,
                 * after all we're only going to change it's attributes
                 * not the memory it points to
                 */
-               new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
-
+               new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
+               cpa->pfn = pfn;
                /*
                 * Do we really change anything ?
                 */
@@ -454,90 +582,80 @@ repeat:
         * Check, whether we can keep the large page intact
         * and just change the pte:
         */
-       res = try_preserve_large_page(kpte, address, cpa);
-       if (res < 0)
-               return res;
-
+       do_split = try_preserve_large_page(kpte, address, cpa);
        /*
         * When the range fits into the existing large page,
         * return. cp->numpages and cpa->tlbflush have been updated in
         * try_large_page:
         */
-       if (res == CPA_NO_SPLIT)
-               return 0;
+       if (do_split <= 0)
+               return do_split;
 
        /*
         * We have to split the large page:
         */
-       res = split_large_page(kpte, address);
-       if (res)
-               return res;
-       cpa->flushtlb = 1;
-       goto repeat;
+       err = split_large_page(kpte, address);
+       if (!err) {
+               cpa->flushtlb = 1;
+               goto repeat;
+       }
+
+       return err;
 }
 
-/**
- * change_page_attr_addr - Change page table attributes in linear mapping
- * @address: Virtual address in linear mapping.
- * @prot:    New page table attribute (PAGE_*)
- *
- * Change page attributes of a page in the direct mapping. This is a variant
- * of change_page_attr() that also works on memory holes that do not have
- * mem_map entry (pfn_valid() is false).
- *
- * See change_page_attr() documentation for more details.
- *
- * Modules and drivers should use the set_memory_* APIs instead.
- */
+static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
 
-static int change_page_attr_addr(struct cpa_data *cpa)
+static int cpa_process_alias(struct cpa_data *cpa)
 {
-       int err;
-       unsigned long address = cpa->vaddr;
+       struct cpa_data alias_cpa;
+       int ret = 0;
 
-#ifdef CONFIG_X86_64
-       unsigned long phys_addr = __pa(address);
+       if (cpa->pfn > max_pfn_mapped)
+               return 0;
 
        /*
-        * If we are inside the high mapped kernel range, then we
-        * fixup the low mapping first. __va() returns the virtual
-        * address in the linear mapping:
+        * No need to redo, when the primary call touched the direct
+        * mapping already:
         */
-       if (within(address, HIGH_MAP_START, HIGH_MAP_END))
-               address = (unsigned long) __va(phys_addr);
-#endif
+       if (!within(cpa->vaddr, PAGE_OFFSET,
+                   PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
 
-       err = __change_page_attr(address, cpa);
-       if (err)
-               return err;
+               alias_cpa = *cpa;
+               alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
+
+               ret = __change_page_attr_set_clr(&alias_cpa, 0);
+       }
 
 #ifdef CONFIG_X86_64
+       if (ret)
+               return ret;
+       /*
+        * No need to redo, when the primary call touched the high
+        * mapping already:
+        */
+       if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
+               return 0;
+
        /*
         * If the physical address is inside the kernel map, we need
         * to touch the high mapped kernel as well:
         */
-       if (within(phys_addr, 0, KERNEL_TEXT_SIZE)) {
-               /*
-                * Calc the high mapping address. See __phys_addr()
-                * for the non obvious details.
-                *
-                * Note that NX and other required permissions are
-                * checked in static_protections().
-                */
-               address = phys_addr + HIGH_MAP_START - phys_base;
+       if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
+               return 0;
 
-               /*
-                * Our high aliases are imprecise, because we check
-                * everything between 0 and KERNEL_TEXT_SIZE, so do
-                * not propagate lookup failures back to users:
-                */
-               __change_page_attr(address, cpa);
-       }
+       alias_cpa = *cpa;
+       alias_cpa.vaddr =
+               (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
+
+       /*
+        * The high mapping range is imprecise, so ignore the return value.
+        */
+       __change_page_attr_set_clr(&alias_cpa, 0);
 #endif
-       return err;
+       return ret;
 }
 
-static int __change_page_attr_set_clr(struct cpa_data *cpa)
+static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
 {
        int ret, numpages = cpa->numpages;
 
@@ -547,10 +665,17 @@ static int __change_page_attr_set_clr(struct cpa_data *cpa)
                 * preservation check.
                 */
                cpa->numpages = numpages;
-               ret = change_page_attr_addr(cpa);
+
+               ret = __change_page_attr(cpa, checkalias);
                if (ret)
                        return ret;
 
+               if (checkalias) {
+                       ret = cpa_process_alias(cpa);
+                       if (ret)
+                               return ret;
+               }
+
                /*
                 * Adjust the number of pages with the result of the
                 * CPA operation. Either a large page has been
@@ -573,7 +698,7 @@ static int change_page_attr_set_clr(unsigned long addr, int numpages,
                                    pgprot_t mask_set, pgprot_t mask_clr)
 {
        struct cpa_data cpa;
-       int ret, cache;
+       int ret, cache, checkalias;
 
        /*
         * Check, if we are requested to change a not supported
@@ -584,19 +709,31 @@ static int change_page_attr_set_clr(unsigned long addr, int numpages,
        if (!pgprot_val(mask_set) && !pgprot_val(mask_clr))
                return 0;
 
+       /* Ensure we are PAGE_SIZE aligned */
+       if (addr & ~PAGE_MASK) {
+               addr &= PAGE_MASK;
+               /*
+                * People should not be passing in unaligned addresses:
+                */
+               WARN_ON_ONCE(1);
+       }
+
        cpa.vaddr = addr;
        cpa.numpages = numpages;
        cpa.mask_set = mask_set;
        cpa.mask_clr = mask_clr;
        cpa.flushtlb = 0;
 
-       ret = __change_page_attr_set_clr(&cpa);
+       /* No alias checking for _NX bit modifications */
+       checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
+
+       ret = __change_page_attr_set_clr(&cpa, checkalias);
 
        /*
         * Check whether we really changed something:
         */
        if (!cpa.flushtlb)
-               return ret;
+               goto out;
 
        /*
         * No need to flush, when we did not set any of the caching
@@ -615,6 +752,9 @@ static int change_page_attr_set_clr(unsigned long addr, int numpages,
        else
                cpa_flush_all(cache);
 
+out:
+       cpa_fill_pool(NULL);
+
        return ret;
 }
 
@@ -633,14 +773,14 @@ static inline int change_page_attr_clear(unsigned long addr, int numpages,
 int set_memory_uc(unsigned long addr, int numpages)
 {
        return change_page_attr_set(addr, numpages,
-                                   __pgprot(_PAGE_PCD | _PAGE_PWT));
+                                   __pgprot(_PAGE_CACHE_UC));
 }
 EXPORT_SYMBOL(set_memory_uc);
 
 int set_memory_wb(unsigned long addr, int numpages)
 {
        return change_page_attr_clear(addr, numpages,
-                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
+                                     __pgprot(_PAGE_CACHE_MASK));
 }
 EXPORT_SYMBOL(set_memory_wb);
 
@@ -726,7 +866,7 @@ static int __set_pages_p(struct page *page, int numpages)
                                .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
                                .mask_clr = __pgprot(0)};
 
-       return __change_page_attr_set_clr(&cpa);
+       return __change_page_attr_set_clr(&cpa, 1);
 }
 
 static int __set_pages_np(struct page *page, int numpages)
@@ -736,7 +876,7 @@ static int __set_pages_np(struct page *page, int numpages)
                                .mask_set = __pgprot(0),
                                .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
 
-       return __change_page_attr_set_clr(&cpa);
+       return __change_page_attr_set_clr(&cpa, 1);
 }
 
 void kernel_map_pages(struct page *page, int numpages, int enable)
@@ -755,8 +895,12 @@ void kernel_map_pages(struct page *page, int numpages, int enable)
                return;
 
        /*
-        * The return value is ignored - the calls cannot fail,
-        * large pages are disabled at boot time:
+        * The return value is ignored as the calls cannot fail.
+        * Large pages are kept enabled at boot time, and are
+        * split up quickly with DEBUG_PAGEALLOC. If a splitup
+        * fails here (due to temporary memory shortage) no damage
+        * is done because we just keep the largepage intact up
+        * to the next attempt when it will likely be split up:
         */
        if (enable)
                __set_pages_p(page, numpages);
@@ -768,9 +912,71 @@ void kernel_map_pages(struct page *page, int numpages, int enable)
         * but that can deadlock->flush only current cpu:
         */
        __flush_tlb_all();
+
+       /*
+        * Try to refill the page pool here. We can do this only after
+        * the tlb flush.
+        */
+       cpa_fill_pool(NULL);
+}
+
+#ifdef CONFIG_DEBUG_FS
+static int dpa_show(struct seq_file *m, void *v)
+{
+       seq_puts(m, "DEBUG_PAGEALLOC\n");
+       seq_printf(m, "pool_size     : %lu\n", pool_size);
+       seq_printf(m, "pool_pages    : %lu\n", pool_pages);
+       seq_printf(m, "pool_low      : %lu\n", pool_low);
+       seq_printf(m, "pool_used     : %lu\n", pool_used);
+       seq_printf(m, "pool_failed   : %lu\n", pool_failed);
+
+       return 0;
+}
+
+static int dpa_open(struct inode *inode, struct file *filp)
+{
+       return single_open(filp, dpa_show, NULL);
 }
+
+static const struct file_operations dpa_fops = {
+       .open           = dpa_open,
+       .read           = seq_read,
+       .llseek         = seq_lseek,
+       .release        = single_release,
+};
+
+int __init debug_pagealloc_proc_init(void)
+{
+       struct dentry *de;
+
+       de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
+                                &dpa_fops);
+       if (!de)
+               return -ENOMEM;
+
+       return 0;
+}
+__initcall(debug_pagealloc_proc_init);
 #endif
 
+#ifdef CONFIG_HIBERNATION
+
+bool kernel_page_present(struct page *page)
+{
+       unsigned int level;
+       pte_t *pte;
+
+       if (PageHighMem(page))
+               return false;
+
+       pte = lookup_address((unsigned long)page_address(page), &level);
+       return (pte_val(*pte) & _PAGE_PRESENT);
+}
+
+#endif /* CONFIG_HIBERNATION */
+
+#endif /* CONFIG_DEBUG_PAGEALLOC */
+
 /*
  * The testcases use internal knowledge of the implementation that shouldn't
  * be exposed to the rest of the kernel. Include these directly here.