mm: factor out main logic of access_process_vm
[linux-2.6.git] / mm / vmalloc.c
index ddf8714..f9b1667 100644 (file)
@@ -8,24 +8,30 @@
  *  Numa awareness, Christoph Lameter, SGI, June 2005
  */
 
+#include <linux/vmalloc.h>
 #include <linux/mm.h>
 #include <linux/module.h>
 #include <linux/highmem.h>
+#include <linux/sched.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include <linux/interrupt.h>
-
-#include <linux/vmalloc.h>
-
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+#include <linux/debugobjects.h>
+#include <linux/kallsyms.h>
+#include <linux/list.h>
+#include <linux/rbtree.h>
+#include <linux/radix-tree.h>
+#include <linux/rcupdate.h>
+#include <linux/pfn.h>
+#include <linux/kmemleak.h>
+#include <asm/atomic.h>
 #include <asm/uaccess.h>
 #include <asm/tlbflush.h>
+#include <asm/shmparam.h>
 
-
-DEFINE_RWLOCK(vmlist_lock);
-struct vm_struct *vmlist;
-
-static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
-                           int node);
+/*** Page table manipulation functions ***/
 
 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
 {
@@ -38,8 +44,7 @@ static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
        } while (pte++, addr += PAGE_SIZE, addr != end);
 }
 
-static inline void vunmap_pmd_range(pud_t *pud, unsigned long addr,
-                                               unsigned long end)
+static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
 {
        pmd_t *pmd;
        unsigned long next;
@@ -53,8 +58,7 @@ static inline void vunmap_pmd_range(pud_t *pud, unsigned long addr,
        } while (pmd++, addr = next, addr != end);
 }
 
-static inline void vunmap_pud_range(pgd_t *pgd, unsigned long addr,
-                                               unsigned long end)
+static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
 {
        pud_t *pud;
        unsigned long next;
@@ -68,51 +72,49 @@ static inline void vunmap_pud_range(pgd_t *pgd, unsigned long addr,
        } while (pud++, addr = next, addr != end);
 }
 
-void unmap_kernel_range(unsigned long addr, unsigned long size)
+static void vunmap_page_range(unsigned long addr, unsigned long end)
 {
        pgd_t *pgd;
        unsigned long next;
-       unsigned long start = addr;
-       unsigned long end = addr + size;
 
        BUG_ON(addr >= end);
        pgd = pgd_offset_k(addr);
-       flush_cache_vunmap(addr, end);
        do {
                next = pgd_addr_end(addr, end);
                if (pgd_none_or_clear_bad(pgd))
                        continue;
                vunmap_pud_range(pgd, addr, next);
        } while (pgd++, addr = next, addr != end);
-       flush_tlb_kernel_range(start, end);
-}
-
-static void unmap_vm_area(struct vm_struct *area)
-{
-       unmap_kernel_range((unsigned long)area->addr, area->size);
 }
 
 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
-                       unsigned long end, pgprot_t prot, struct page ***pages)
+               unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 {
        pte_t *pte;
 
+       /*
+        * nr is a running index into the array which helps higher level
+        * callers keep track of where we're up to.
+        */
+
        pte = pte_alloc_kernel(pmd, addr);
        if (!pte)
                return -ENOMEM;
        do {
-               struct page *page = **pages;
-               WARN_ON(!pte_none(*pte));
-               if (!page)
+               struct page *page = pages[*nr];
+
+               if (WARN_ON(!pte_none(*pte)))
+                       return -EBUSY;
+               if (WARN_ON(!page))
                        return -ENOMEM;
                set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
-               (*pages)++;
+               (*nr)++;
        } while (pte++, addr += PAGE_SIZE, addr != end);
        return 0;
 }
 
-static inline int vmap_pmd_range(pud_t *pud, unsigned long addr,
-                       unsigned long end, pgprot_t prot, struct page ***pages)
+static int vmap_pmd_range(pud_t *pud, unsigned long addr,
+               unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 {
        pmd_t *pmd;
        unsigned long next;
@@ -122,14 +124,14 @@ static inline int vmap_pmd_range(pud_t *pud, unsigned long addr,
                return -ENOMEM;
        do {
                next = pmd_addr_end(addr, end);
-               if (vmap_pte_range(pmd, addr, next, prot, pages))
+               if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
                        return -ENOMEM;
        } while (pmd++, addr = next, addr != end);
        return 0;
 }
 
-static inline int vmap_pud_range(pgd_t *pgd, unsigned long addr,
-                       unsigned long end, pgprot_t prot, struct page ***pages)
+static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
+               unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 {
        pud_t *pud;
        unsigned long next;
@@ -139,40 +141,1108 @@ static inline int vmap_pud_range(pgd_t *pgd, unsigned long addr,
                return -ENOMEM;
        do {
                next = pud_addr_end(addr, end);
-               if (vmap_pmd_range(pud, addr, next, prot, pages))
+               if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
                        return -ENOMEM;
        } while (pud++, addr = next, addr != end);
        return 0;
 }
 
-int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
+/*
+ * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
+ * will have pfns corresponding to the "pages" array.
+ *
+ * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
+ */
+static int vmap_page_range_noflush(unsigned long start, unsigned long end,
+                                  pgprot_t prot, struct page **pages)
 {
        pgd_t *pgd;
        unsigned long next;
-       unsigned long addr = (unsigned long) area->addr;
-       unsigned long end = addr + area->size - PAGE_SIZE;
-       int err;
+       unsigned long addr = start;
+       int err = 0;
+       int nr = 0;
 
        BUG_ON(addr >= end);
        pgd = pgd_offset_k(addr);
        do {
                next = pgd_addr_end(addr, end);
-               err = vmap_pud_range(pgd, addr, next, prot, pages);
+               err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
                if (err)
-                       break;
+                       return err;
        } while (pgd++, addr = next, addr != end);
-       flush_cache_vmap((unsigned long) area->addr, end);
-       return err;
+
+       return nr;
+}
+
+static int vmap_page_range(unsigned long start, unsigned long end,
+                          pgprot_t prot, struct page **pages)
+{
+       int ret;
+
+       ret = vmap_page_range_noflush(start, end, prot, pages);
+       flush_cache_vmap(start, end);
+       return ret;
+}
+
+int is_vmalloc_or_module_addr(const void *x)
+{
+       /*
+        * ARM, x86-64 and sparc64 put modules in a special place,
+        * and fall back on vmalloc() if that fails. Others
+        * just put it in the vmalloc space.
+        */
+#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
+       unsigned long addr = (unsigned long)x;
+       if (addr >= MODULES_VADDR && addr < MODULES_END)
+               return 1;
+#endif
+       return is_vmalloc_addr(x);
+}
+
+/*
+ * Walk a vmap address to the struct page it maps.
+ */
+struct page *vmalloc_to_page(const void *vmalloc_addr)
+{
+       unsigned long addr = (unsigned long) vmalloc_addr;
+       struct page *page = NULL;
+       pgd_t *pgd = pgd_offset_k(addr);
+
+       /*
+        * XXX we might need to change this if we add VIRTUAL_BUG_ON for
+        * architectures that do not vmalloc module space
+        */
+       VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
+
+       if (!pgd_none(*pgd)) {
+               pud_t *pud = pud_offset(pgd, addr);
+               if (!pud_none(*pud)) {
+                       pmd_t *pmd = pmd_offset(pud, addr);
+                       if (!pmd_none(*pmd)) {
+                               pte_t *ptep, pte;
+
+                               ptep = pte_offset_map(pmd, addr);
+                               pte = *ptep;
+                               if (pte_present(pte))
+                                       page = pte_page(pte);
+                               pte_unmap(ptep);
+                       }
+               }
+       }
+       return page;
+}
+EXPORT_SYMBOL(vmalloc_to_page);
+
+/*
+ * Map a vmalloc()-space virtual address to the physical page frame number.
+ */
+unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
+{
+       return page_to_pfn(vmalloc_to_page(vmalloc_addr));
+}
+EXPORT_SYMBOL(vmalloc_to_pfn);
+
+
+/*** Global kva allocator ***/
+
+#define VM_LAZY_FREE   0x01
+#define VM_LAZY_FREEING        0x02
+#define VM_VM_AREA     0x04
+
+struct vmap_area {
+       unsigned long va_start;
+       unsigned long va_end;
+       unsigned long flags;
+       struct rb_node rb_node;         /* address sorted rbtree */
+       struct list_head list;          /* address sorted list */
+       struct list_head purge_list;    /* "lazy purge" list */
+       void *private;
+       struct rcu_head rcu_head;
+};
+
+static DEFINE_SPINLOCK(vmap_area_lock);
+static struct rb_root vmap_area_root = RB_ROOT;
+static LIST_HEAD(vmap_area_list);
+static unsigned long vmap_area_pcpu_hole;
+
+static struct vmap_area *__find_vmap_area(unsigned long addr)
+{
+       struct rb_node *n = vmap_area_root.rb_node;
+
+       while (n) {
+               struct vmap_area *va;
+
+               va = rb_entry(n, struct vmap_area, rb_node);
+               if (addr < va->va_start)
+                       n = n->rb_left;
+               else if (addr > va->va_start)
+                       n = n->rb_right;
+               else
+                       return va;
+       }
+
+       return NULL;
+}
+
+static void __insert_vmap_area(struct vmap_area *va)
+{
+       struct rb_node **p = &vmap_area_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct rb_node *tmp;
+
+       while (*p) {
+               struct vmap_area *tmp_va;
+
+               parent = *p;
+               tmp_va = rb_entry(parent, struct vmap_area, rb_node);
+               if (va->va_start < tmp_va->va_end)
+                       p = &(*p)->rb_left;
+               else if (va->va_end > tmp_va->va_start)
+                       p = &(*p)->rb_right;
+               else
+                       BUG();
+       }
+
+       rb_link_node(&va->rb_node, parent, p);
+       rb_insert_color(&va->rb_node, &vmap_area_root);
+
+       /* address-sort this list so it is usable like the vmlist */
+       tmp = rb_prev(&va->rb_node);
+       if (tmp) {
+               struct vmap_area *prev;
+               prev = rb_entry(tmp, struct vmap_area, rb_node);
+               list_add_rcu(&va->list, &prev->list);
+       } else
+               list_add_rcu(&va->list, &vmap_area_list);
+}
+
+static void purge_vmap_area_lazy(void);
+
+/*
+ * Allocate a region of KVA of the specified size and alignment, within the
+ * vstart and vend.
+ */
+static struct vmap_area *alloc_vmap_area(unsigned long size,
+                               unsigned long align,
+                               unsigned long vstart, unsigned long vend,
+                               int node, gfp_t gfp_mask)
+{
+       struct vmap_area *va;
+       struct rb_node *n;
+       unsigned long addr;
+       int purged = 0;
+
+       BUG_ON(!size);
+       BUG_ON(size & ~PAGE_MASK);
+
+       va = kmalloc_node(sizeof(struct vmap_area),
+                       gfp_mask & GFP_RECLAIM_MASK, node);
+       if (unlikely(!va))
+               return ERR_PTR(-ENOMEM);
+
+retry:
+       addr = ALIGN(vstart, align);
+
+       spin_lock(&vmap_area_lock);
+       if (addr + size - 1 < addr)
+               goto overflow;
+
+       /* XXX: could have a last_hole cache */
+       n = vmap_area_root.rb_node;
+       if (n) {
+               struct vmap_area *first = NULL;
+
+               do {
+                       struct vmap_area *tmp;
+                       tmp = rb_entry(n, struct vmap_area, rb_node);
+                       if (tmp->va_end >= addr) {
+                               if (!first && tmp->va_start < addr + size)
+                                       first = tmp;
+                               n = n->rb_left;
+                       } else {
+                               first = tmp;
+                               n = n->rb_right;
+                       }
+               } while (n);
+
+               if (!first)
+                       goto found;
+
+               if (first->va_end < addr) {
+                       n = rb_next(&first->rb_node);
+                       if (n)
+                               first = rb_entry(n, struct vmap_area, rb_node);
+                       else
+                               goto found;
+               }
+
+               while (addr + size > first->va_start && addr + size <= vend) {
+                       addr = ALIGN(first->va_end + PAGE_SIZE, align);
+                       if (addr + size - 1 < addr)
+                               goto overflow;
+
+                       n = rb_next(&first->rb_node);
+                       if (n)
+                               first = rb_entry(n, struct vmap_area, rb_node);
+                       else
+                               goto found;
+               }
+       }
+found:
+       if (addr + size > vend) {
+overflow:
+               spin_unlock(&vmap_area_lock);
+               if (!purged) {
+                       purge_vmap_area_lazy();
+                       purged = 1;
+                       goto retry;
+               }
+               if (printk_ratelimit())
+                       printk(KERN_WARNING
+                               "vmap allocation for size %lu failed: "
+                               "use vmalloc=<size> to increase size.\n", size);
+               kfree(va);
+               return ERR_PTR(-EBUSY);
+       }
+
+       BUG_ON(addr & (align-1));
+
+       va->va_start = addr;
+       va->va_end = addr + size;
+       va->flags = 0;
+       __insert_vmap_area(va);
+       spin_unlock(&vmap_area_lock);
+
+       return va;
+}
+
+static void rcu_free_va(struct rcu_head *head)
+{
+       struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
+
+       kfree(va);
+}
+
+static void __free_vmap_area(struct vmap_area *va)
+{
+       BUG_ON(RB_EMPTY_NODE(&va->rb_node));
+       rb_erase(&va->rb_node, &vmap_area_root);
+       RB_CLEAR_NODE(&va->rb_node);
+       list_del_rcu(&va->list);
+
+       /*
+        * Track the highest possible candidate for pcpu area
+        * allocation.  Areas outside of vmalloc area can be returned
+        * here too, consider only end addresses which fall inside
+        * vmalloc area proper.
+        */
+       if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
+               vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
+
+       call_rcu(&va->rcu_head, rcu_free_va);
+}
+
+/*
+ * Free a region of KVA allocated by alloc_vmap_area
+ */
+static void free_vmap_area(struct vmap_area *va)
+{
+       spin_lock(&vmap_area_lock);
+       __free_vmap_area(va);
+       spin_unlock(&vmap_area_lock);
+}
+
+/*
+ * Clear the pagetable entries of a given vmap_area
+ */
+static void unmap_vmap_area(struct vmap_area *va)
+{
+       vunmap_page_range(va->va_start, va->va_end);
+}
+
+static void vmap_debug_free_range(unsigned long start, unsigned long end)
+{
+       /*
+        * Unmap page tables and force a TLB flush immediately if
+        * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
+        * bugs similarly to those in linear kernel virtual address
+        * space after a page has been freed.
+        *
+        * All the lazy freeing logic is still retained, in order to
+        * minimise intrusiveness of this debugging feature.
+        *
+        * This is going to be *slow* (linear kernel virtual address
+        * debugging doesn't do a broadcast TLB flush so it is a lot
+        * faster).
+        */
+#ifdef CONFIG_DEBUG_PAGEALLOC
+       vunmap_page_range(start, end);
+       flush_tlb_kernel_range(start, end);
+#endif
+}
+
+/*
+ * lazy_max_pages is the maximum amount of virtual address space we gather up
+ * before attempting to purge with a TLB flush.
+ *
+ * There is a tradeoff here: a larger number will cover more kernel page tables
+ * and take slightly longer to purge, but it will linearly reduce the number of
+ * global TLB flushes that must be performed. It would seem natural to scale
+ * this number up linearly with the number of CPUs (because vmapping activity
+ * could also scale linearly with the number of CPUs), however it is likely
+ * that in practice, workloads might be constrained in other ways that mean
+ * vmap activity will not scale linearly with CPUs. Also, I want to be
+ * conservative and not introduce a big latency on huge systems, so go with
+ * a less aggressive log scale. It will still be an improvement over the old
+ * code, and it will be simple to change the scale factor if we find that it
+ * becomes a problem on bigger systems.
+ */
+static unsigned long lazy_max_pages(void)
+{
+       unsigned int log;
+
+       log = fls(num_online_cpus());
+
+       return log * (32UL * 1024 * 1024 / PAGE_SIZE);
+}
+
+static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
+
+/* for per-CPU blocks */
+static void purge_fragmented_blocks_allcpus(void);
+
+/*
+ * called before a call to iounmap() if the caller wants vm_area_struct's
+ * immediately freed.
+ */
+void set_iounmap_nonlazy(void)
+{
+       atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
+}
+
+/*
+ * Purges all lazily-freed vmap areas.
+ *
+ * If sync is 0 then don't purge if there is already a purge in progress.
+ * If force_flush is 1, then flush kernel TLBs between *start and *end even
+ * if we found no lazy vmap areas to unmap (callers can use this to optimise
+ * their own TLB flushing).
+ * Returns with *start = min(*start, lowest purged address)
+ *              *end = max(*end, highest purged address)
+ */
+static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
+                                       int sync, int force_flush)
+{
+       static DEFINE_SPINLOCK(purge_lock);
+       LIST_HEAD(valist);
+       struct vmap_area *va;
+       struct vmap_area *n_va;
+       int nr = 0;
+
+       /*
+        * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
+        * should not expect such behaviour. This just simplifies locking for
+        * the case that isn't actually used at the moment anyway.
+        */
+       if (!sync && !force_flush) {
+               if (!spin_trylock(&purge_lock))
+                       return;
+       } else
+               spin_lock(&purge_lock);
+
+       if (sync)
+               purge_fragmented_blocks_allcpus();
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(va, &vmap_area_list, list) {
+               if (va->flags & VM_LAZY_FREE) {
+                       if (va->va_start < *start)
+                               *start = va->va_start;
+                       if (va->va_end > *end)
+                               *end = va->va_end;
+                       nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
+                       list_add_tail(&va->purge_list, &valist);
+                       va->flags |= VM_LAZY_FREEING;
+                       va->flags &= ~VM_LAZY_FREE;
+               }
+       }
+       rcu_read_unlock();
+
+       if (nr)
+               atomic_sub(nr, &vmap_lazy_nr);
+
+       if (nr || force_flush)
+               flush_tlb_kernel_range(*start, *end);
+
+       if (nr) {
+               spin_lock(&vmap_area_lock);
+               list_for_each_entry_safe(va, n_va, &valist, purge_list)
+                       __free_vmap_area(va);
+               spin_unlock(&vmap_area_lock);
+       }
+       spin_unlock(&purge_lock);
+}
+
+/*
+ * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
+ * is already purging.
+ */
+static void try_purge_vmap_area_lazy(void)
+{
+       unsigned long start = ULONG_MAX, end = 0;
+
+       __purge_vmap_area_lazy(&start, &end, 0, 0);
+}
+
+/*
+ * Kick off a purge of the outstanding lazy areas.
+ */
+static void purge_vmap_area_lazy(void)
+{
+       unsigned long start = ULONG_MAX, end = 0;
+
+       __purge_vmap_area_lazy(&start, &end, 1, 0);
+}
+
+/*
+ * Free a vmap area, caller ensuring that the area has been unmapped
+ * and flush_cache_vunmap had been called for the correct range
+ * previously.
+ */
+static void free_vmap_area_noflush(struct vmap_area *va)
+{
+       va->flags |= VM_LAZY_FREE;
+       atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
+       if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
+               try_purge_vmap_area_lazy();
+}
+
+/*
+ * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
+ * called for the correct range previously.
+ */
+static void free_unmap_vmap_area_noflush(struct vmap_area *va)
+{
+       unmap_vmap_area(va);
+       free_vmap_area_noflush(va);
+}
+
+/*
+ * Free and unmap a vmap area
+ */
+static void free_unmap_vmap_area(struct vmap_area *va)
+{
+       flush_cache_vunmap(va->va_start, va->va_end);
+       free_unmap_vmap_area_noflush(va);
+}
+
+static struct vmap_area *find_vmap_area(unsigned long addr)
+{
+       struct vmap_area *va;
+
+       spin_lock(&vmap_area_lock);
+       va = __find_vmap_area(addr);
+       spin_unlock(&vmap_area_lock);
+
+       return va;
+}
+
+static void free_unmap_vmap_area_addr(unsigned long addr)
+{
+       struct vmap_area *va;
+
+       va = find_vmap_area(addr);
+       BUG_ON(!va);
+       free_unmap_vmap_area(va);
+}
+
+
+/*** Per cpu kva allocator ***/
+
+/*
+ * vmap space is limited especially on 32 bit architectures. Ensure there is
+ * room for at least 16 percpu vmap blocks per CPU.
+ */
+/*
+ * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
+ * to #define VMALLOC_SPACE            (VMALLOC_END-VMALLOC_START). Guess
+ * instead (we just need a rough idea)
+ */
+#if BITS_PER_LONG == 32
+#define VMALLOC_SPACE          (128UL*1024*1024)
+#else
+#define VMALLOC_SPACE          (128UL*1024*1024*1024)
+#endif
+
+#define VMALLOC_PAGES          (VMALLOC_SPACE / PAGE_SIZE)
+#define VMAP_MAX_ALLOC         BITS_PER_LONG   /* 256K with 4K pages */
+#define VMAP_BBMAP_BITS_MAX    1024    /* 4MB with 4K pages */
+#define VMAP_BBMAP_BITS_MIN    (VMAP_MAX_ALLOC*2)
+#define VMAP_MIN(x, y)         ((x) < (y) ? (x) : (y)) /* can't use min() */
+#define VMAP_MAX(x, y)         ((x) > (y) ? (x) : (y)) /* can't use max() */
+#define VMAP_BBMAP_BITS                VMAP_MIN(VMAP_BBMAP_BITS_MAX,           \
+                                       VMAP_MAX(VMAP_BBMAP_BITS_MIN,   \
+                                               VMALLOC_PAGES / NR_CPUS / 16))
+
+#define VMAP_BLOCK_SIZE                (VMAP_BBMAP_BITS * PAGE_SIZE)
+
+static bool vmap_initialized __read_mostly = false;
+
+struct vmap_block_queue {
+       spinlock_t lock;
+       struct list_head free;
+};
+
+struct vmap_block {
+       spinlock_t lock;
+       struct vmap_area *va;
+       struct vmap_block_queue *vbq;
+       unsigned long free, dirty;
+       DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
+       DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
+       struct list_head free_list;
+       struct rcu_head rcu_head;
+       struct list_head purge;
+};
+
+/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
+static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
+
+/*
+ * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
+ * in the free path. Could get rid of this if we change the API to return a
+ * "cookie" from alloc, to be passed to free. But no big deal yet.
+ */
+static DEFINE_SPINLOCK(vmap_block_tree_lock);
+static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
+
+/*
+ * We should probably have a fallback mechanism to allocate virtual memory
+ * out of partially filled vmap blocks. However vmap block sizing should be
+ * fairly reasonable according to the vmalloc size, so it shouldn't be a
+ * big problem.
+ */
+
+static unsigned long addr_to_vb_idx(unsigned long addr)
+{
+       addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
+       addr /= VMAP_BLOCK_SIZE;
+       return addr;
 }
 
-static struct vm_struct *__get_vm_area_node(unsigned long size, unsigned long flags,
-                                           unsigned long start, unsigned long end,
-                                           int node, gfp_t gfp_mask)
+static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
 {
-       struct vm_struct **p, *tmp, *area;
-       unsigned long align = 1;
+       struct vmap_block_queue *vbq;
+       struct vmap_block *vb;
+       struct vmap_area *va;
+       unsigned long vb_idx;
+       int node, err;
+
+       node = numa_node_id();
+
+       vb = kmalloc_node(sizeof(struct vmap_block),
+                       gfp_mask & GFP_RECLAIM_MASK, node);
+       if (unlikely(!vb))
+               return ERR_PTR(-ENOMEM);
+
+       va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
+                                       VMALLOC_START, VMALLOC_END,
+                                       node, gfp_mask);
+       if (IS_ERR(va)) {
+               kfree(vb);
+               return ERR_CAST(va);
+       }
+
+       err = radix_tree_preload(gfp_mask);
+       if (unlikely(err)) {
+               kfree(vb);
+               free_vmap_area(va);
+               return ERR_PTR(err);
+       }
+
+       spin_lock_init(&vb->lock);
+       vb->va = va;
+       vb->free = VMAP_BBMAP_BITS;
+       vb->dirty = 0;
+       bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
+       bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
+       INIT_LIST_HEAD(&vb->free_list);
+
+       vb_idx = addr_to_vb_idx(va->va_start);
+       spin_lock(&vmap_block_tree_lock);
+       err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
+       spin_unlock(&vmap_block_tree_lock);
+       BUG_ON(err);
+       radix_tree_preload_end();
+
+       vbq = &get_cpu_var(vmap_block_queue);
+       vb->vbq = vbq;
+       spin_lock(&vbq->lock);
+       list_add_rcu(&vb->free_list, &vbq->free);
+       spin_unlock(&vbq->lock);
+       put_cpu_var(vmap_block_queue);
+
+       return vb;
+}
+
+static void rcu_free_vb(struct rcu_head *head)
+{
+       struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
+
+       kfree(vb);
+}
+
+static void free_vmap_block(struct vmap_block *vb)
+{
+       struct vmap_block *tmp;
+       unsigned long vb_idx;
+
+       vb_idx = addr_to_vb_idx(vb->va->va_start);
+       spin_lock(&vmap_block_tree_lock);
+       tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
+       spin_unlock(&vmap_block_tree_lock);
+       BUG_ON(tmp != vb);
+
+       free_vmap_area_noflush(vb->va);
+       call_rcu(&vb->rcu_head, rcu_free_vb);
+}
+
+static void purge_fragmented_blocks(int cpu)
+{
+       LIST_HEAD(purge);
+       struct vmap_block *vb;
+       struct vmap_block *n_vb;
+       struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(vb, &vbq->free, free_list) {
+
+               if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
+                       continue;
+
+               spin_lock(&vb->lock);
+               if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
+                       vb->free = 0; /* prevent further allocs after releasing lock */
+                       vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
+                       bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
+                       bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
+                       spin_lock(&vbq->lock);
+                       list_del_rcu(&vb->free_list);
+                       spin_unlock(&vbq->lock);
+                       spin_unlock(&vb->lock);
+                       list_add_tail(&vb->purge, &purge);
+               } else
+                       spin_unlock(&vb->lock);
+       }
+       rcu_read_unlock();
+
+       list_for_each_entry_safe(vb, n_vb, &purge, purge) {
+               list_del(&vb->purge);
+               free_vmap_block(vb);
+       }
+}
+
+static void purge_fragmented_blocks_thiscpu(void)
+{
+       purge_fragmented_blocks(smp_processor_id());
+}
+
+static void purge_fragmented_blocks_allcpus(void)
+{
+       int cpu;
+
+       for_each_possible_cpu(cpu)
+               purge_fragmented_blocks(cpu);
+}
+
+static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
+{
+       struct vmap_block_queue *vbq;
+       struct vmap_block *vb;
+       unsigned long addr = 0;
+       unsigned int order;
+       int purge = 0;
+
+       BUG_ON(size & ~PAGE_MASK);
+       BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
+       order = get_order(size);
+
+again:
+       rcu_read_lock();
+       vbq = &get_cpu_var(vmap_block_queue);
+       list_for_each_entry_rcu(vb, &vbq->free, free_list) {
+               int i;
+
+               spin_lock(&vb->lock);
+               if (vb->free < 1UL << order)
+                       goto next;
+
+               i = bitmap_find_free_region(vb->alloc_map,
+                                               VMAP_BBMAP_BITS, order);
+
+               if (i < 0) {
+                       if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
+                               /* fragmented and no outstanding allocations */
+                               BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
+                               purge = 1;
+                       }
+                       goto next;
+               }
+               addr = vb->va->va_start + (i << PAGE_SHIFT);
+               BUG_ON(addr_to_vb_idx(addr) !=
+                               addr_to_vb_idx(vb->va->va_start));
+               vb->free -= 1UL << order;
+               if (vb->free == 0) {
+                       spin_lock(&vbq->lock);
+                       list_del_rcu(&vb->free_list);
+                       spin_unlock(&vbq->lock);
+               }
+               spin_unlock(&vb->lock);
+               break;
+next:
+               spin_unlock(&vb->lock);
+       }
+
+       if (purge)
+               purge_fragmented_blocks_thiscpu();
+
+       put_cpu_var(vmap_block_queue);
+       rcu_read_unlock();
+
+       if (!addr) {
+               vb = new_vmap_block(gfp_mask);
+               if (IS_ERR(vb))
+                       return vb;
+               goto again;
+       }
+
+       return (void *)addr;
+}
+
+static void vb_free(const void *addr, unsigned long size)
+{
+       unsigned long offset;
+       unsigned long vb_idx;
+       unsigned int order;
+       struct vmap_block *vb;
+
+       BUG_ON(size & ~PAGE_MASK);
+       BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
+
+       flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
+
+       order = get_order(size);
+
+       offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
+
+       vb_idx = addr_to_vb_idx((unsigned long)addr);
+       rcu_read_lock();
+       vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
+       rcu_read_unlock();
+       BUG_ON(!vb);
+
+       vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
+
+       spin_lock(&vb->lock);
+       BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
+
+       vb->dirty += 1UL << order;
+       if (vb->dirty == VMAP_BBMAP_BITS) {
+               BUG_ON(vb->free);
+               spin_unlock(&vb->lock);
+               free_vmap_block(vb);
+       } else
+               spin_unlock(&vb->lock);
+}
+
+/**
+ * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
+ *
+ * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
+ * to amortize TLB flushing overheads. What this means is that any page you
+ * have now, may, in a former life, have been mapped into kernel virtual
+ * address by the vmap layer and so there might be some CPUs with TLB entries
+ * still referencing that page (additional to the regular 1:1 kernel mapping).
+ *
+ * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
+ * be sure that none of the pages we have control over will have any aliases
+ * from the vmap layer.
+ */
+void vm_unmap_aliases(void)
+{
+       unsigned long start = ULONG_MAX, end = 0;
+       int cpu;
+       int flush = 0;
+
+       if (unlikely(!vmap_initialized))
+               return;
+
+       for_each_possible_cpu(cpu) {
+               struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
+               struct vmap_block *vb;
+
+               rcu_read_lock();
+               list_for_each_entry_rcu(vb, &vbq->free, free_list) {
+                       int i;
+
+                       spin_lock(&vb->lock);
+                       i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
+                       while (i < VMAP_BBMAP_BITS) {
+                               unsigned long s, e;
+                               int j;
+                               j = find_next_zero_bit(vb->dirty_map,
+                                       VMAP_BBMAP_BITS, i);
+
+                               s = vb->va->va_start + (i << PAGE_SHIFT);
+                               e = vb->va->va_start + (j << PAGE_SHIFT);
+                               flush = 1;
+
+                               if (s < start)
+                                       start = s;
+                               if (e > end)
+                                       end = e;
+
+                               i = j;
+                               i = find_next_bit(vb->dirty_map,
+                                                       VMAP_BBMAP_BITS, i);
+                       }
+                       spin_unlock(&vb->lock);
+               }
+               rcu_read_unlock();
+       }
+
+       __purge_vmap_area_lazy(&start, &end, 1, flush);
+}
+EXPORT_SYMBOL_GPL(vm_unmap_aliases);
+
+/**
+ * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
+ * @mem: the pointer returned by vm_map_ram
+ * @count: the count passed to that vm_map_ram call (cannot unmap partial)
+ */
+void vm_unmap_ram(const void *mem, unsigned int count)
+{
+       unsigned long size = count << PAGE_SHIFT;
+       unsigned long addr = (unsigned long)mem;
+
+       BUG_ON(!addr);
+       BUG_ON(addr < VMALLOC_START);
+       BUG_ON(addr > VMALLOC_END);
+       BUG_ON(addr & (PAGE_SIZE-1));
+
+       debug_check_no_locks_freed(mem, size);
+       vmap_debug_free_range(addr, addr+size);
+
+       if (likely(count <= VMAP_MAX_ALLOC))
+               vb_free(mem, size);
+       else
+               free_unmap_vmap_area_addr(addr);
+}
+EXPORT_SYMBOL(vm_unmap_ram);
+
+/**
+ * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
+ * @pages: an array of pointers to the pages to be mapped
+ * @count: number of pages
+ * @node: prefer to allocate data structures on this node
+ * @prot: memory protection to use. PAGE_KERNEL for regular RAM
+ *
+ * Returns: a pointer to the address that has been mapped, or %NULL on failure
+ */
+void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
+{
+       unsigned long size = count << PAGE_SHIFT;
+       unsigned long addr;
+       void *mem;
+
+       if (likely(count <= VMAP_MAX_ALLOC)) {
+               mem = vb_alloc(size, GFP_KERNEL);
+               if (IS_ERR(mem))
+                       return NULL;
+               addr = (unsigned long)mem;
+       } else {
+               struct vmap_area *va;
+               va = alloc_vmap_area(size, PAGE_SIZE,
+                               VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
+               if (IS_ERR(va))
+                       return NULL;
+
+               addr = va->va_start;
+               mem = (void *)addr;
+       }
+       if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
+               vm_unmap_ram(mem, count);
+               return NULL;
+       }
+       return mem;
+}
+EXPORT_SYMBOL(vm_map_ram);
+
+/**
+ * vm_area_register_early - register vmap area early during boot
+ * @vm: vm_struct to register
+ * @align: requested alignment
+ *
+ * This function is used to register kernel vm area before
+ * vmalloc_init() is called.  @vm->size and @vm->flags should contain
+ * proper values on entry and other fields should be zero.  On return,
+ * vm->addr contains the allocated address.
+ *
+ * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
+ */
+void __init vm_area_register_early(struct vm_struct *vm, size_t align)
+{
+       static size_t vm_init_off __initdata;
        unsigned long addr;
 
+       addr = ALIGN(VMALLOC_START + vm_init_off, align);
+       vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
+
+       vm->addr = (void *)addr;
+
+       vm->next = vmlist;
+       vmlist = vm;
+}
+
+void __init vmalloc_init(void)
+{
+       struct vmap_area *va;
+       struct vm_struct *tmp;
+       int i;
+
+       for_each_possible_cpu(i) {
+               struct vmap_block_queue *vbq;
+
+               vbq = &per_cpu(vmap_block_queue, i);
+               spin_lock_init(&vbq->lock);
+               INIT_LIST_HEAD(&vbq->free);
+       }
+
+       /* Import existing vmlist entries. */
+       for (tmp = vmlist; tmp; tmp = tmp->next) {
+               va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
+               va->flags = tmp->flags | VM_VM_AREA;
+               va->va_start = (unsigned long)tmp->addr;
+               va->va_end = va->va_start + tmp->size;
+               __insert_vmap_area(va);
+       }
+
+       vmap_area_pcpu_hole = VMALLOC_END;
+
+       vmap_initialized = true;
+}
+
+/**
+ * map_kernel_range_noflush - map kernel VM area with the specified pages
+ * @addr: start of the VM area to map
+ * @size: size of the VM area to map
+ * @prot: page protection flags to use
+ * @pages: pages to map
+ *
+ * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
+ * specify should have been allocated using get_vm_area() and its
+ * friends.
+ *
+ * NOTE:
+ * This function does NOT do any cache flushing.  The caller is
+ * responsible for calling flush_cache_vmap() on to-be-mapped areas
+ * before calling this function.
+ *
+ * RETURNS:
+ * The number of pages mapped on success, -errno on failure.
+ */
+int map_kernel_range_noflush(unsigned long addr, unsigned long size,
+                            pgprot_t prot, struct page **pages)
+{
+       return vmap_page_range_noflush(addr, addr + size, prot, pages);
+}
+
+/**
+ * unmap_kernel_range_noflush - unmap kernel VM area
+ * @addr: start of the VM area to unmap
+ * @size: size of the VM area to unmap
+ *
+ * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
+ * specify should have been allocated using get_vm_area() and its
+ * friends.
+ *
+ * NOTE:
+ * This function does NOT do any cache flushing.  The caller is
+ * responsible for calling flush_cache_vunmap() on to-be-mapped areas
+ * before calling this function and flush_tlb_kernel_range() after.
+ */
+void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
+{
+       vunmap_page_range(addr, addr + size);
+}
+EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
+
+/**
+ * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
+ * @addr: start of the VM area to unmap
+ * @size: size of the VM area to unmap
+ *
+ * Similar to unmap_kernel_range_noflush() but flushes vcache before
+ * the unmapping and tlb after.
+ */
+void unmap_kernel_range(unsigned long addr, unsigned long size)
+{
+       unsigned long end = addr + size;
+
+       flush_cache_vunmap(addr, end);
+       vunmap_page_range(addr, end);
+       flush_tlb_kernel_range(addr, end);
+}
+
+int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
+{
+       unsigned long addr = (unsigned long)area->addr;
+       unsigned long end = addr + area->size - PAGE_SIZE;
+       int err;
+
+       err = vmap_page_range(addr, end, prot, *pages);
+       if (err > 0) {
+               *pages += err;
+               err = 0;
+       }
+
+       return err;
+}
+EXPORT_SYMBOL_GPL(map_vm_area);
+
+/*** Old vmalloc interfaces ***/
+DEFINE_RWLOCK(vmlist_lock);
+struct vm_struct *vmlist;
+
+static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
+                             unsigned long flags, void *caller)
+{
+       struct vm_struct *tmp, **p;
+
+       vm->flags = flags;
+       vm->addr = (void *)va->va_start;
+       vm->size = va->va_end - va->va_start;
+       vm->caller = caller;
+       va->private = vm;
+       va->flags |= VM_VM_AREA;
+
+       write_lock(&vmlist_lock);
+       for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
+               if (tmp->addr >= vm->addr)
+                       break;
+       }
+       vm->next = *p;
+       *p = vm;
+       write_unlock(&vmlist_lock);
+}
+
+static struct vm_struct *__get_vm_area_node(unsigned long size,
+               unsigned long align, unsigned long flags, unsigned long start,
+               unsigned long end, int node, gfp_t gfp_mask, void *caller)
+{
+       static struct vmap_area *va;
+       struct vm_struct *area;
+
        BUG_ON(in_interrupt());
        if (flags & VM_IOREMAP) {
                int bit = fls(size);
@@ -184,12 +1254,12 @@ static struct vm_struct *__get_vm_area_node(unsigned long size, unsigned long fl
 
                align = 1ul << bit;
        }
-       addr = ALIGN(start, align);
+
        size = PAGE_ALIGN(size);
        if (unlikely(!size))
                return NULL;
 
-       area = kmalloc_node(sizeof(*area), gfp_mask & GFP_LEVEL_MASK, node);
+       area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
        if (unlikely(!area))
                return NULL;
 
@@ -198,53 +1268,34 @@ static struct vm_struct *__get_vm_area_node(unsigned long size, unsigned long fl
         */
        size += PAGE_SIZE;
 
-       write_lock(&vmlist_lock);
-       for (p = &vmlist; (tmp = *p) != NULL ;p = &tmp->next) {
-               if ((unsigned long)tmp->addr < addr) {
-                       if((unsigned long)tmp->addr + tmp->size >= addr)
-                               addr = ALIGN(tmp->size + 
-                                            (unsigned long)tmp->addr, align);
-                       continue;
-               }
-               if ((size + addr) < addr)
-                       goto out;
-               if (size + addr <= (unsigned long)tmp->addr)
-                       goto found;
-               addr = ALIGN(tmp->size + (unsigned long)tmp->addr, align);
-               if (addr > end - size)
-                       goto out;
+       va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
+       if (IS_ERR(va)) {
+               kfree(area);
+               return NULL;
        }
 
-found:
-       area->next = *p;
-       *p = area;
-
-       area->flags = flags;
-       area->addr = (void *)addr;
-       area->size = size;
-       area->pages = NULL;
-       area->nr_pages = 0;
-       area->phys_addr = 0;
-       write_unlock(&vmlist_lock);
-
+       insert_vmalloc_vm(area, va, flags, caller);
        return area;
-
-out:
-       write_unlock(&vmlist_lock);
-       kfree(area);
-       if (printk_ratelimit())
-               printk(KERN_WARNING "allocation failed: out of vmalloc space - use vmalloc=<size> to increase size.\n");
-       return NULL;
 }
 
 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
                                unsigned long start, unsigned long end)
 {
-       return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL);
+       return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
+                                               __builtin_return_address(0));
+}
+EXPORT_SYMBOL_GPL(__get_vm_area);
+
+struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
+                                      unsigned long start, unsigned long end,
+                                      void *caller)
+{
+       return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
+                                 caller);
 }
 
 /**
- *     get_vm_area  -  reserve a contingous kernel virtual area
+ *     get_vm_area  -  reserve a contiguous kernel virtual area
  *     @size:          size of the area
  *     @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
  *
@@ -254,69 +1305,65 @@ struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
  */
 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
 {
-       return __get_vm_area(size, flags, VMALLOC_START, VMALLOC_END);
+       return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
+                               -1, GFP_KERNEL, __builtin_return_address(0));
 }
 
-struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
-                                  int node, gfp_t gfp_mask)
+struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
+                               void *caller)
 {
-       return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
-                                 gfp_mask);
+       return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
+                                               -1, GFP_KERNEL, caller);
 }
 
-/* Caller must hold vmlist_lock */
-static struct vm_struct *__find_vm_area(void *addr)
+static struct vm_struct *find_vm_area(const void *addr)
 {
-       struct vm_struct *tmp;
+       struct vmap_area *va;
 
-       for (tmp = vmlist; tmp != NULL; tmp = tmp->next) {
-                if (tmp->addr == addr)
-                       break;
-       }
-
-       return tmp;
-}
-
-/* Caller must hold vmlist_lock */
-static struct vm_struct *__remove_vm_area(void *addr)
-{
-       struct vm_struct **p, *tmp;
+       va = find_vmap_area((unsigned long)addr);
+       if (va && va->flags & VM_VM_AREA)
+               return va->private;
 
-       for (p = &vmlist ; (tmp = *p) != NULL ;p = &tmp->next) {
-                if (tmp->addr == addr)
-                        goto found;
-       }
        return NULL;
-
-found:
-       unmap_vm_area(tmp);
-       *p = tmp->next;
-
-       /*
-        * Remove the guard page.
-        */
-       tmp->size -= PAGE_SIZE;
-       return tmp;
 }
 
 /**
- *     remove_vm_area  -  find and remove a contingous kernel virtual area
+ *     remove_vm_area  -  find and remove a continuous kernel virtual area
  *     @addr:          base address
  *
  *     Search for the kernel VM area starting at @addr, and remove it.
  *     This function returns the found VM area, but using it is NOT safe
  *     on SMP machines, except for its size or flags.
  */
-struct vm_struct *remove_vm_area(void *addr)
+struct vm_struct *remove_vm_area(const void *addr)
 {
-       struct vm_struct *v;
-       write_lock(&vmlist_lock);
-       v = __remove_vm_area(addr);
-       write_unlock(&vmlist_lock);
-       return v;
+       struct vmap_area *va;
+
+       va = find_vmap_area((unsigned long)addr);
+       if (va && va->flags & VM_VM_AREA) {
+               struct vm_struct *vm = va->private;
+               struct vm_struct *tmp, **p;
+               /*
+                * remove from list and disallow access to this vm_struct
+                * before unmap. (address range confliction is maintained by
+                * vmap.)
+                */
+               write_lock(&vmlist_lock);
+               for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
+                       ;
+               *p = tmp->next;
+               write_unlock(&vmlist_lock);
+
+               vmap_debug_free_range(va->va_start, va->va_end);
+               free_unmap_vmap_area(va);
+               vm->size -= PAGE_SIZE;
+
+               return vm;
+       }
+       return NULL;
 }
 
-static void __vunmap(void *addr, int deallocate_pages)
+static void __vunmap(const void *addr, int deallocate_pages)
 {
        struct vm_struct *area;
 
@@ -324,27 +1371,28 @@ static void __vunmap(void *addr, int deallocate_pages)
                return;
 
        if ((PAGE_SIZE-1) & (unsigned long)addr) {
-               printk(KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
-               WARN_ON(1);
+               WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
                return;
        }
 
        area = remove_vm_area(addr);
        if (unlikely(!area)) {
-               printk(KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
+               WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
                                addr);
-               WARN_ON(1);
                return;
        }
 
        debug_check_no_locks_freed(addr, area->size);
+       debug_check_no_obj_freed(addr, area->size);
 
        if (deallocate_pages) {
                int i;
 
                for (i = 0; i < area->nr_pages; i++) {
-                       BUG_ON(!area->pages[i]);
-                       __free_page(area->pages[i]);
+                       struct page *page = area->pages[i];
+
+                       BUG_ON(!page);
+                       __free_page(page);
                }
 
                if (area->flags & VM_VPAGES)
@@ -361,15 +1409,18 @@ static void __vunmap(void *addr, int deallocate_pages)
  *     vfree  -  release memory allocated by vmalloc()
  *     @addr:          memory base address
  *
- *     Free the virtually contiguous memory area starting at @addr, as
+ *     Free the virtually continuous memory area starting at @addr, as
  *     obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
  *     NULL, no operation is performed.
  *
  *     Must not be called in interrupt context.
  */
-void vfree(void *addr)
+void vfree(const void *addr)
 {
        BUG_ON(in_interrupt());
+
+       kmemleak_free(addr);
+
        __vunmap(addr, 1);
 }
 EXPORT_SYMBOL(vfree);
@@ -383,9 +1434,10 @@ EXPORT_SYMBOL(vfree);
  *
  *     Must not be called in interrupt context.
  */
-void vunmap(void *addr)
+void vunmap(const void *addr)
 {
        BUG_ON(in_interrupt());
+       might_sleep();
        __vunmap(addr, 0);
 }
 EXPORT_SYMBOL(vunmap);
@@ -405,12 +1457,16 @@ void *vmap(struct page **pages, unsigned int count,
 {
        struct vm_struct *area;
 
-       if (count > num_physpages)
+       might_sleep();
+
+       if (count > totalram_pages)
                return NULL;
 
-       area = get_vm_area((count << PAGE_SHIFT), flags);
+       area = get_vm_area_caller((count << PAGE_SHIFT), flags,
+                                       __builtin_return_address(0));
        if (!area)
                return NULL;
+
        if (map_vm_area(area, prot, &pages)) {
                vunmap(area->addr);
                return NULL;
@@ -420,11 +1476,15 @@ void *vmap(struct page **pages, unsigned int count,
 }
 EXPORT_SYMBOL(vmap);
 
-void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
-                               pgprot_t prot, int node)
+static void *__vmalloc_node(unsigned long size, unsigned long align,
+                           gfp_t gfp_mask, pgprot_t prot,
+                           int node, void *caller);
+static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
+                                pgprot_t prot, int node, void *caller)
 {
        struct page **pages;
        unsigned int nr_pages, array_size, i;
+       gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
 
        nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
        array_size = (nr_pages * sizeof(struct page *));
@@ -432,31 +1492,34 @@ void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
        area->nr_pages = nr_pages;
        /* Please note that the recursion is strictly bounded. */
        if (array_size > PAGE_SIZE) {
-               pages = __vmalloc_node(array_size, gfp_mask, PAGE_KERNEL, node);
+               pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
+                               PAGE_KERNEL, node, caller);
                area->flags |= VM_VPAGES;
        } else {
-               pages = kmalloc_node(array_size,
-                               (gfp_mask & GFP_LEVEL_MASK),
-                               node);
+               pages = kmalloc_node(array_size, nested_gfp, node);
        }
        area->pages = pages;
+       area->caller = caller;
        if (!area->pages) {
                remove_vm_area(area->addr);
                kfree(area);
                return NULL;
        }
-       memset(area->pages, 0, array_size);
 
        for (i = 0; i < area->nr_pages; i++) {
+               struct page *page;
+
                if (node < 0)
-                       area->pages[i] = alloc_page(gfp_mask);
+                       page = alloc_page(gfp_mask);
                else
-                       area->pages[i] = alloc_pages_node(node, gfp_mask, 0);
-               if (unlikely(!area->pages[i])) {
+                       page = alloc_pages_node(node, gfp_mask, 0);
+
+               if (unlikely(!page)) {
                        /* Successfully allocated i pages, free them in __vunmap() */
                        area->nr_pages = i;
                        goto fail;
                }
+               area->pages[i] = page;
        }
 
        if (map_vm_area(area, prot, &pages))
@@ -468,44 +1531,86 @@ fail:
        return NULL;
 }
 
-void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
-{
-       return __vmalloc_area_node(area, gfp_mask, prot, -1);
-}
-
 /**
- *     __vmalloc_node  -  allocate virtually contiguous memory
+ *     __vmalloc_node_range  -  allocate virtually contiguous memory
  *     @size:          allocation size
+ *     @align:         desired alignment
+ *     @start:         vm area range start
+ *     @end:           vm area range end
  *     @gfp_mask:      flags for the page level allocator
  *     @prot:          protection mask for the allocated pages
  *     @node:          node to use for allocation or -1
+ *     @caller:        caller's return address
  *
  *     Allocate enough pages to cover @size from the page level
  *     allocator with @gfp_mask flags.  Map them into contiguous
  *     kernel virtual space, using a pagetable protection of @prot.
  */
-static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
-                           int node)
+void *__vmalloc_node_range(unsigned long size, unsigned long align,
+                       unsigned long start, unsigned long end, gfp_t gfp_mask,
+                       pgprot_t prot, int node, void *caller)
 {
        struct vm_struct *area;
+       void *addr;
+       unsigned long real_size = size;
 
        size = PAGE_ALIGN(size);
-       if (!size || (size >> PAGE_SHIFT) > num_physpages)
+       if (!size || (size >> PAGE_SHIFT) > totalram_pages)
                return NULL;
 
-       area = get_vm_area_node(size, VM_ALLOC, node, gfp_mask);
+       area = __get_vm_area_node(size, align, VM_ALLOC, start, end, node,
+                                 gfp_mask, caller);
+
        if (!area)
                return NULL;
 
-       return __vmalloc_area_node(area, gfp_mask, prot, node);
+       addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
+
+       /*
+        * A ref_count = 3 is needed because the vm_struct and vmap_area
+        * structures allocated in the __get_vm_area_node() function contain
+        * references to the virtual address of the vmalloc'ed block.
+        */
+       kmemleak_alloc(addr, real_size, 3, gfp_mask);
+
+       return addr;
+}
+
+/**
+ *     __vmalloc_node  -  allocate virtually contiguous memory
+ *     @size:          allocation size
+ *     @align:         desired alignment
+ *     @gfp_mask:      flags for the page level allocator
+ *     @prot:          protection mask for the allocated pages
+ *     @node:          node to use for allocation or -1
+ *     @caller:        caller's return address
+ *
+ *     Allocate enough pages to cover @size from the page level
+ *     allocator with @gfp_mask flags.  Map them into contiguous
+ *     kernel virtual space, using a pagetable protection of @prot.
+ */
+static void *__vmalloc_node(unsigned long size, unsigned long align,
+                           gfp_t gfp_mask, pgprot_t prot,
+                           int node, void *caller)
+{
+       return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
+                               gfp_mask, prot, node, caller);
 }
 
 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
 {
-       return __vmalloc_node(size, gfp_mask, prot, -1);
+       return __vmalloc_node(size, 1, gfp_mask, prot, -1,
+                               __builtin_return_address(0));
 }
 EXPORT_SYMBOL(__vmalloc);
 
+static inline void *__vmalloc_node_flags(unsigned long size,
+                                       int node, gfp_t flags)
+{
+       return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
+                                       node, __builtin_return_address(0));
+}
+
 /**
  *     vmalloc  -  allocate virtually contiguous memory
  *     @size:          allocation size
@@ -517,11 +1622,28 @@ EXPORT_SYMBOL(__vmalloc);
  */
 void *vmalloc(unsigned long size)
 {
-       return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
+       return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
 }
 EXPORT_SYMBOL(vmalloc);
 
 /**
+ *     vzalloc - allocate virtually contiguous memory with zero fill
+ *     @size:  allocation size
+ *     Allocate enough pages to cover @size from the page level
+ *     allocator and map them into contiguous kernel virtual space.
+ *     The memory allocated is set to zero.
+ *
+ *     For tight control over page level allocator and protection flags
+ *     use __vmalloc() instead.
+ */
+void *vzalloc(unsigned long size)
+{
+       return __vmalloc_node_flags(size, -1,
+                               GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
+}
+EXPORT_SYMBOL(vzalloc);
+
+/**
  * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
  * @size: allocation size
  *
@@ -533,12 +1655,12 @@ void *vmalloc_user(unsigned long size)
        struct vm_struct *area;
        void *ret;
 
-       ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
+       ret = __vmalloc_node(size, SHMLBA,
+                            GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
+                            PAGE_KERNEL, -1, __builtin_return_address(0));
        if (ret) {
-               write_lock(&vmlist_lock);
-               area = __find_vm_area(ret);
+               area = find_vm_area(ret);
                area->flags |= VM_USERMAP;
-               write_unlock(&vmlist_lock);
        }
        return ret;
 }
@@ -557,10 +1679,30 @@ EXPORT_SYMBOL(vmalloc_user);
  */
 void *vmalloc_node(unsigned long size, int node)
 {
-       return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL, node);
+       return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
+                                       node, __builtin_return_address(0));
 }
 EXPORT_SYMBOL(vmalloc_node);
 
+/**
+ * vzalloc_node - allocate memory on a specific node with zero fill
+ * @size:      allocation size
+ * @node:      numa node
+ *
+ * Allocate enough pages to cover @size from the page level
+ * allocator and map them into contiguous kernel virtual space.
+ * The memory allocated is set to zero.
+ *
+ * For tight control over page level allocator and protection flags
+ * use __vmalloc_node() instead.
+ */
+void *vzalloc_node(unsigned long size, int node)
+{
+       return __vmalloc_node_flags(size, node,
+                        GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
+}
+EXPORT_SYMBOL(vzalloc_node);
+
 #ifndef PAGE_KERNEL_EXEC
 # define PAGE_KERNEL_EXEC PAGE_KERNEL
 #endif
@@ -579,13 +1721,14 @@ EXPORT_SYMBOL(vmalloc_node);
 
 void *vmalloc_exec(unsigned long size)
 {
-       return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
+       return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
+                             -1, __builtin_return_address(0));
 }
 
 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
-#define GFP_VMALLOC32 GFP_DMA32
+#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
-#define GFP_VMALLOC32 GFP_DMA
+#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
 #else
 #define GFP_VMALLOC32 GFP_KERNEL
 #endif
@@ -599,7 +1742,8 @@ void *vmalloc_exec(unsigned long size)
  */
 void *vmalloc_32(unsigned long size)
 {
-       return __vmalloc(size, GFP_VMALLOC32, PAGE_KERNEL);
+       return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
+                             -1, __builtin_return_address(0));
 }
 EXPORT_SYMBOL(vmalloc_32);
 
@@ -615,21 +1759,130 @@ void *vmalloc_32_user(unsigned long size)
        struct vm_struct *area;
        void *ret;
 
-       ret = __vmalloc(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL);
+       ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
+                            -1, __builtin_return_address(0));
        if (ret) {
-               write_lock(&vmlist_lock);
-               area = __find_vm_area(ret);
+               area = find_vm_area(ret);
                area->flags |= VM_USERMAP;
-               write_unlock(&vmlist_lock);
        }
        return ret;
 }
 EXPORT_SYMBOL(vmalloc_32_user);
 
+/*
+ * small helper routine , copy contents to buf from addr.
+ * If the page is not present, fill zero.
+ */
+
+static int aligned_vread(char *buf, char *addr, unsigned long count)
+{
+       struct page *p;
+       int copied = 0;
+
+       while (count) {
+               unsigned long offset, length;
+
+               offset = (unsigned long)addr & ~PAGE_MASK;
+               length = PAGE_SIZE - offset;
+               if (length > count)
+                       length = count;
+               p = vmalloc_to_page(addr);
+               /*
+                * To do safe access to this _mapped_ area, we need
+                * lock. But adding lock here means that we need to add
+                * overhead of vmalloc()/vfree() calles for this _debug_
+                * interface, rarely used. Instead of that, we'll use
+                * kmap() and get small overhead in this access function.
+                */
+               if (p) {
+                       /*
+                        * we can expect USER0 is not used (see vread/vwrite's
+                        * function description)
+                        */
+                       void *map = kmap_atomic(p, KM_USER0);
+                       memcpy(buf, map + offset, length);
+                       kunmap_atomic(map, KM_USER0);
+               } else
+                       memset(buf, 0, length);
+
+               addr += length;
+               buf += length;
+               copied += length;
+               count -= length;
+       }
+       return copied;
+}
+
+static int aligned_vwrite(char *buf, char *addr, unsigned long count)
+{
+       struct page *p;
+       int copied = 0;
+
+       while (count) {
+               unsigned long offset, length;
+
+               offset = (unsigned long)addr & ~PAGE_MASK;
+               length = PAGE_SIZE - offset;
+               if (length > count)
+                       length = count;
+               p = vmalloc_to_page(addr);
+               /*
+                * To do safe access to this _mapped_ area, we need
+                * lock. But adding lock here means that we need to add
+                * overhead of vmalloc()/vfree() calles for this _debug_
+                * interface, rarely used. Instead of that, we'll use
+                * kmap() and get small overhead in this access function.
+                */
+               if (p) {
+                       /*
+                        * we can expect USER0 is not used (see vread/vwrite's
+                        * function description)
+                        */
+                       void *map = kmap_atomic(p, KM_USER0);
+                       memcpy(map + offset, buf, length);
+                       kunmap_atomic(map, KM_USER0);
+               }
+               addr += length;
+               buf += length;
+               copied += length;
+               count -= length;
+       }
+       return copied;
+}
+
+/**
+ *     vread() -  read vmalloc area in a safe way.
+ *     @buf:           buffer for reading data
+ *     @addr:          vm address.
+ *     @count:         number of bytes to be read.
+ *
+ *     Returns # of bytes which addr and buf should be increased.
+ *     (same number to @count). Returns 0 if [addr...addr+count) doesn't
+ *     includes any intersect with alive vmalloc area.
+ *
+ *     This function checks that addr is a valid vmalloc'ed area, and
+ *     copy data from that area to a given buffer. If the given memory range
+ *     of [addr...addr+count) includes some valid address, data is copied to
+ *     proper area of @buf. If there are memory holes, they'll be zero-filled.
+ *     IOREMAP area is treated as memory hole and no copy is done.
+ *
+ *     If [addr...addr+count) doesn't includes any intersects with alive
+ *     vm_struct area, returns 0.
+ *     @buf should be kernel's buffer. Because this function uses KM_USER0,
+ *     the caller should guarantee KM_USER0 is not used.
+ *
+ *     Note: In usual ops, vread() is never necessary because the caller
+ *     should know vmalloc() area is valid and can use memcpy().
+ *     This is for routines which have to access vmalloc area without
+ *     any informaion, as /dev/kmem.
+ *
+ */
+
 long vread(char *buf, char *addr, unsigned long count)
 {
        struct vm_struct *tmp;
        char *vaddr, *buf_start = buf;
+       unsigned long buflen = count;
        unsigned long n;
 
        /* Don't allow overflow */
@@ -637,7 +1890,7 @@ long vread(char *buf, char *addr, unsigned long count)
                count = -(unsigned long) addr;
 
        read_lock(&vmlist_lock);
-       for (tmp = vmlist; tmp; tmp = tmp->next) {
+       for (tmp = vmlist; count && tmp; tmp = tmp->next) {
                vaddr = (char *) tmp->addr;
                if (addr >= vaddr + tmp->size - PAGE_SIZE)
                        continue;
@@ -650,32 +1903,72 @@ long vread(char *buf, char *addr, unsigned long count)
                        count--;
                }
                n = vaddr + tmp->size - PAGE_SIZE - addr;
-               do {
-                       if (count == 0)
-                               goto finished;
-                       *buf = *addr;
-                       buf++;
-                       addr++;
-                       count--;
-               } while (--n > 0);
+               if (n > count)
+                       n = count;
+               if (!(tmp->flags & VM_IOREMAP))
+                       aligned_vread(buf, addr, n);
+               else /* IOREMAP area is treated as memory hole */
+                       memset(buf, 0, n);
+               buf += n;
+               addr += n;
+               count -= n;
        }
 finished:
        read_unlock(&vmlist_lock);
-       return buf - buf_start;
+
+       if (buf == buf_start)
+               return 0;
+       /* zero-fill memory holes */
+       if (buf != buf_start + buflen)
+               memset(buf, 0, buflen - (buf - buf_start));
+
+       return buflen;
 }
 
+/**
+ *     vwrite() -  write vmalloc area in a safe way.
+ *     @buf:           buffer for source data
+ *     @addr:          vm address.
+ *     @count:         number of bytes to be read.
+ *
+ *     Returns # of bytes which addr and buf should be incresed.
+ *     (same number to @count).
+ *     If [addr...addr+count) doesn't includes any intersect with valid
+ *     vmalloc area, returns 0.
+ *
+ *     This function checks that addr is a valid vmalloc'ed area, and
+ *     copy data from a buffer to the given addr. If specified range of
+ *     [addr...addr+count) includes some valid address, data is copied from
+ *     proper area of @buf. If there are memory holes, no copy to hole.
+ *     IOREMAP area is treated as memory hole and no copy is done.
+ *
+ *     If [addr...addr+count) doesn't includes any intersects with alive
+ *     vm_struct area, returns 0.
+ *     @buf should be kernel's buffer. Because this function uses KM_USER0,
+ *     the caller should guarantee KM_USER0 is not used.
+ *
+ *     Note: In usual ops, vwrite() is never necessary because the caller
+ *     should know vmalloc() area is valid and can use memcpy().
+ *     This is for routines which have to access vmalloc area without
+ *     any informaion, as /dev/kmem.
+ *
+ *     The caller should guarantee KM_USER1 is not used.
+ */
+
 long vwrite(char *buf, char *addr, unsigned long count)
 {
        struct vm_struct *tmp;
-       char *vaddr, *buf_start = buf;
-       unsigned long n;
+       char *vaddr;
+       unsigned long n, buflen;
+       int copied = 0;
 
        /* Don't allow overflow */
        if ((unsigned long) addr + count < count)
                count = -(unsigned long) addr;
+       buflen = count;
 
        read_lock(&vmlist_lock);
-       for (tmp = vmlist; tmp; tmp = tmp->next) {
+       for (tmp = vmlist; count && tmp; tmp = tmp->next) {
                vaddr = (char *) tmp->addr;
                if (addr >= vaddr + tmp->size - PAGE_SIZE)
                        continue;
@@ -687,18 +1980,21 @@ long vwrite(char *buf, char *addr, unsigned long count)
                        count--;
                }
                n = vaddr + tmp->size - PAGE_SIZE - addr;
-               do {
-                       if (count == 0)
-                               goto finished;
-                       *addr = *buf;
-                       buf++;
-                       addr++;
-                       count--;
-               } while (--n > 0);
+               if (n > count)
+                       n = count;
+               if (!(tmp->flags & VM_IOREMAP)) {
+                       aligned_vwrite(buf, addr, n);
+                       copied++;
+               }
+               buf += n;
+               addr += n;
+               count -= n;
        }
 finished:
        read_unlock(&vmlist_lock);
-       return buf - buf_start;
+       if (!copied)
+               return 0;
+       return buflen;
 }
 
 /**
@@ -706,7 +2002,8 @@ finished:
  *     @vma:           vma to cover (map full range of vma)
  *     @addr:          vmalloc memory
  *     @pgoff:         number of pages into addr before first page to map
- *     @returns:       0 for success, -Exxx on failure
+ *
+ *     Returns:        0 for success, -Exxx on failure
  *
  *     This function checks that addr is a valid vmalloc'ed area, and
  *     that it is big enough to cover the vma. Will return failure if
@@ -720,26 +2017,25 @@ int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
        struct vm_struct *area;
        unsigned long uaddr = vma->vm_start;
        unsigned long usize = vma->vm_end - vma->vm_start;
-       int ret;
 
        if ((PAGE_SIZE-1) & (unsigned long)addr)
                return -EINVAL;
 
-       read_lock(&vmlist_lock);
-       area = __find_vm_area(addr);
+       area = find_vm_area(addr);
        if (!area)
-               goto out_einval_locked;
+               return -EINVAL;
 
        if (!(area->flags & VM_USERMAP))
-               goto out_einval_locked;
+               return -EINVAL;
 
        if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
-               goto out_einval_locked;
-       read_unlock(&vmlist_lock);
+               return -EINVAL;
 
        addr += pgoff << PAGE_SHIFT;
        do {
                struct page *page = vmalloc_to_page(addr);
+               int ret;
+
                ret = vm_insert_page(vma, uaddr, page);
                if (ret)
                        return ret;
@@ -752,11 +2048,7 @@ int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
        /* Prevent "things" like memory migration? VM_flags need a cleanup... */
        vma->vm_flags |= VM_RESERVED;
 
-       return ret;
-
-out_einval_locked:
-       read_unlock(&vmlist_lock);
-       return -EINVAL;
+       return 0;
 }
 EXPORT_SYMBOL(remap_vmalloc_range);
 
@@ -767,3 +2059,469 @@ EXPORT_SYMBOL(remap_vmalloc_range);
 void  __attribute__((weak)) vmalloc_sync_all(void)
 {
 }
+
+
+static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
+{
+       /* apply_to_page_range() does all the hard work. */
+       return 0;
+}
+
+/**
+ *     alloc_vm_area - allocate a range of kernel address space
+ *     @size:          size of the area
+ *
+ *     Returns:        NULL on failure, vm_struct on success
+ *
+ *     This function reserves a range of kernel address space, and
+ *     allocates pagetables to map that range.  No actual mappings
+ *     are created.  If the kernel address space is not shared
+ *     between processes, it syncs the pagetable across all
+ *     processes.
+ */
+struct vm_struct *alloc_vm_area(size_t size)
+{
+       struct vm_struct *area;
+
+       area = get_vm_area_caller(size, VM_IOREMAP,
+                               __builtin_return_address(0));
+       if (area == NULL)
+               return NULL;
+
+       /*
+        * This ensures that page tables are constructed for this region
+        * of kernel virtual address space and mapped into init_mm.
+        */
+       if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
+                               area->size, f, NULL)) {
+               free_vm_area(area);
+               return NULL;
+       }
+
+       /* Make sure the pagetables are constructed in process kernel
+          mappings */
+       vmalloc_sync_all();
+
+       return area;
+}
+EXPORT_SYMBOL_GPL(alloc_vm_area);
+
+void free_vm_area(struct vm_struct *area)
+{
+       struct vm_struct *ret;
+       ret = remove_vm_area(area->addr);
+       BUG_ON(ret != area);
+       kfree(area);
+}
+EXPORT_SYMBOL_GPL(free_vm_area);
+
+#ifdef CONFIG_SMP
+static struct vmap_area *node_to_va(struct rb_node *n)
+{
+       return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
+}
+
+/**
+ * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
+ * @end: target address
+ * @pnext: out arg for the next vmap_area
+ * @pprev: out arg for the previous vmap_area
+ *
+ * Returns: %true if either or both of next and prev are found,
+ *         %false if no vmap_area exists
+ *
+ * Find vmap_areas end addresses of which enclose @end.  ie. if not
+ * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
+ */
+static bool pvm_find_next_prev(unsigned long end,
+                              struct vmap_area **pnext,
+                              struct vmap_area **pprev)
+{
+       struct rb_node *n = vmap_area_root.rb_node;
+       struct vmap_area *va = NULL;
+
+       while (n) {
+               va = rb_entry(n, struct vmap_area, rb_node);
+               if (end < va->va_end)
+                       n = n->rb_left;
+               else if (end > va->va_end)
+                       n = n->rb_right;
+               else
+                       break;
+       }
+
+       if (!va)
+               return false;
+
+       if (va->va_end > end) {
+               *pnext = va;
+               *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
+       } else {
+               *pprev = va;
+               *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
+       }
+       return true;
+}
+
+/**
+ * pvm_determine_end - find the highest aligned address between two vmap_areas
+ * @pnext: in/out arg for the next vmap_area
+ * @pprev: in/out arg for the previous vmap_area
+ * @align: alignment
+ *
+ * Returns: determined end address
+ *
+ * Find the highest aligned address between *@pnext and *@pprev below
+ * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
+ * down address is between the end addresses of the two vmap_areas.
+ *
+ * Please note that the address returned by this function may fall
+ * inside *@pnext vmap_area.  The caller is responsible for checking
+ * that.
+ */
+static unsigned long pvm_determine_end(struct vmap_area **pnext,
+                                      struct vmap_area **pprev,
+                                      unsigned long align)
+{
+       const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
+       unsigned long addr;
+
+       if (*pnext)
+               addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
+       else
+               addr = vmalloc_end;
+
+       while (*pprev && (*pprev)->va_end > addr) {
+               *pnext = *pprev;
+               *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
+       }
+
+       return addr;
+}
+
+/**
+ * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
+ * @offsets: array containing offset of each area
+ * @sizes: array containing size of each area
+ * @nr_vms: the number of areas to allocate
+ * @align: alignment, all entries in @offsets and @sizes must be aligned to this
+ *
+ * Returns: kmalloc'd vm_struct pointer array pointing to allocated
+ *         vm_structs on success, %NULL on failure
+ *
+ * Percpu allocator wants to use congruent vm areas so that it can
+ * maintain the offsets among percpu areas.  This function allocates
+ * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
+ * be scattered pretty far, distance between two areas easily going up
+ * to gigabytes.  To avoid interacting with regular vmallocs, these
+ * areas are allocated from top.
+ *
+ * Despite its complicated look, this allocator is rather simple.  It
+ * does everything top-down and scans areas from the end looking for
+ * matching slot.  While scanning, if any of the areas overlaps with
+ * existing vmap_area, the base address is pulled down to fit the
+ * area.  Scanning is repeated till all the areas fit and then all
+ * necessary data structres are inserted and the result is returned.
+ */
+struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
+                                    const size_t *sizes, int nr_vms,
+                                    size_t align)
+{
+       const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
+       const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
+       struct vmap_area **vas, *prev, *next;
+       struct vm_struct **vms;
+       int area, area2, last_area, term_area;
+       unsigned long base, start, end, last_end;
+       bool purged = false;
+
+       /* verify parameters and allocate data structures */
+       BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
+       for (last_area = 0, area = 0; area < nr_vms; area++) {
+               start = offsets[area];
+               end = start + sizes[area];
+
+               /* is everything aligned properly? */
+               BUG_ON(!IS_ALIGNED(offsets[area], align));
+               BUG_ON(!IS_ALIGNED(sizes[area], align));
+
+               /* detect the area with the highest address */
+               if (start > offsets[last_area])
+                       last_area = area;
+
+               for (area2 = 0; area2 < nr_vms; area2++) {
+                       unsigned long start2 = offsets[area2];
+                       unsigned long end2 = start2 + sizes[area2];
+
+                       if (area2 == area)
+                               continue;
+
+                       BUG_ON(start2 >= start && start2 < end);
+                       BUG_ON(end2 <= end && end2 > start);
+               }
+       }
+       last_end = offsets[last_area] + sizes[last_area];
+
+       if (vmalloc_end - vmalloc_start < last_end) {
+               WARN_ON(true);
+               return NULL;
+       }
+
+       vms = kzalloc(sizeof(vms[0]) * nr_vms, GFP_KERNEL);
+       vas = kzalloc(sizeof(vas[0]) * nr_vms, GFP_KERNEL);
+       if (!vas || !vms)
+               goto err_free;
+
+       for (area = 0; area < nr_vms; area++) {
+               vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
+               vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
+               if (!vas[area] || !vms[area])
+                       goto err_free;
+       }
+retry:
+       spin_lock(&vmap_area_lock);
+
+       /* start scanning - we scan from the top, begin with the last area */
+       area = term_area = last_area;
+       start = offsets[area];
+       end = start + sizes[area];
+
+       if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
+               base = vmalloc_end - last_end;
+               goto found;
+       }
+       base = pvm_determine_end(&next, &prev, align) - end;
+
+       while (true) {
+               BUG_ON(next && next->va_end <= base + end);
+               BUG_ON(prev && prev->va_end > base + end);
+
+               /*
+                * base might have underflowed, add last_end before
+                * comparing.
+                */
+               if (base + last_end < vmalloc_start + last_end) {
+                       spin_unlock(&vmap_area_lock);
+                       if (!purged) {
+                               purge_vmap_area_lazy();
+                               purged = true;
+                               goto retry;
+                       }
+                       goto err_free;
+               }
+
+               /*
+                * If next overlaps, move base downwards so that it's
+                * right below next and then recheck.
+                */
+               if (next && next->va_start < base + end) {
+                       base = pvm_determine_end(&next, &prev, align) - end;
+                       term_area = area;
+                       continue;
+               }
+
+               /*
+                * If prev overlaps, shift down next and prev and move
+                * base so that it's right below new next and then
+                * recheck.
+                */
+               if (prev && prev->va_end > base + start)  {
+                       next = prev;
+                       prev = node_to_va(rb_prev(&next->rb_node));
+                       base = pvm_determine_end(&next, &prev, align) - end;
+                       term_area = area;
+                       continue;
+               }
+
+               /*
+                * This area fits, move on to the previous one.  If
+                * the previous one is the terminal one, we're done.
+                */
+               area = (area + nr_vms - 1) % nr_vms;
+               if (area == term_area)
+                       break;
+               start = offsets[area];
+               end = start + sizes[area];
+               pvm_find_next_prev(base + end, &next, &prev);
+       }
+found:
+       /* we've found a fitting base, insert all va's */
+       for (area = 0; area < nr_vms; area++) {
+               struct vmap_area *va = vas[area];
+
+               va->va_start = base + offsets[area];
+               va->va_end = va->va_start + sizes[area];
+               __insert_vmap_area(va);
+       }
+
+       vmap_area_pcpu_hole = base + offsets[last_area];
+
+       spin_unlock(&vmap_area_lock);
+
+       /* insert all vm's */
+       for (area = 0; area < nr_vms; area++)
+               insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
+                                 pcpu_get_vm_areas);
+
+       kfree(vas);
+       return vms;
+
+err_free:
+       for (area = 0; area < nr_vms; area++) {
+               if (vas)
+                       kfree(vas[area]);
+               if (vms)
+                       kfree(vms[area]);
+       }
+       kfree(vas);
+       kfree(vms);
+       return NULL;
+}
+
+/**
+ * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
+ * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
+ * @nr_vms: the number of allocated areas
+ *
+ * Free vm_structs and the array allocated by pcpu_get_vm_areas().
+ */
+void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
+{
+       int i;
+
+       for (i = 0; i < nr_vms; i++)
+               free_vm_area(vms[i]);
+       kfree(vms);
+}
+#endif /* CONFIG_SMP */
+
+#ifdef CONFIG_PROC_FS
+static void *s_start(struct seq_file *m, loff_t *pos)
+       __acquires(&vmlist_lock)
+{
+       loff_t n = *pos;
+       struct vm_struct *v;
+
+       read_lock(&vmlist_lock);
+       v = vmlist;
+       while (n > 0 && v) {
+               n--;
+               v = v->next;
+       }
+       if (!n)
+               return v;
+
+       return NULL;
+
+}
+
+static void *s_next(struct seq_file *m, void *p, loff_t *pos)
+{
+       struct vm_struct *v = p;
+
+       ++*pos;
+       return v->next;
+}
+
+static void s_stop(struct seq_file *m, void *p)
+       __releases(&vmlist_lock)
+{
+       read_unlock(&vmlist_lock);
+}
+
+static void show_numa_info(struct seq_file *m, struct vm_struct *v)
+{
+       if (NUMA_BUILD) {
+               unsigned int nr, *counters = m->private;
+
+               if (!counters)
+                       return;
+
+               memset(counters, 0, nr_node_ids * sizeof(unsigned int));
+
+               for (nr = 0; nr < v->nr_pages; nr++)
+                       counters[page_to_nid(v->pages[nr])]++;
+
+               for_each_node_state(nr, N_HIGH_MEMORY)
+                       if (counters[nr])
+                               seq_printf(m, " N%u=%u", nr, counters[nr]);
+       }
+}
+
+static int s_show(struct seq_file *m, void *p)
+{
+       struct vm_struct *v = p;
+
+       seq_printf(m, "0x%p-0x%p %7ld",
+               v->addr, v->addr + v->size, v->size);
+
+       if (v->caller)
+               seq_printf(m, " %pS", v->caller);
+
+       if (v->nr_pages)
+               seq_printf(m, " pages=%d", v->nr_pages);
+
+       if (v->phys_addr)
+               seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
+
+       if (v->flags & VM_IOREMAP)
+               seq_printf(m, " ioremap");
+
+       if (v->flags & VM_ALLOC)
+               seq_printf(m, " vmalloc");
+
+       if (v->flags & VM_MAP)
+               seq_printf(m, " vmap");
+
+       if (v->flags & VM_USERMAP)
+               seq_printf(m, " user");
+
+       if (v->flags & VM_VPAGES)
+               seq_printf(m, " vpages");
+
+       show_numa_info(m, v);
+       seq_putc(m, '\n');
+       return 0;
+}
+
+static const struct seq_operations vmalloc_op = {
+       .start = s_start,
+       .next = s_next,
+       .stop = s_stop,
+       .show = s_show,
+};
+
+static int vmalloc_open(struct inode *inode, struct file *file)
+{
+       unsigned int *ptr = NULL;
+       int ret;
+
+       if (NUMA_BUILD) {
+               ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
+               if (ptr == NULL)
+                       return -ENOMEM;
+       }
+       ret = seq_open(file, &vmalloc_op);
+       if (!ret) {
+               struct seq_file *m = file->private_data;
+               m->private = ptr;
+       } else
+               kfree(ptr);
+       return ret;
+}
+
+static const struct file_operations proc_vmalloc_operations = {
+       .open           = vmalloc_open,
+       .read           = seq_read,
+       .llseek         = seq_lseek,
+       .release        = seq_release_private,
+};
+
+static int __init proc_vmalloc_init(void)
+{
+       proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
+       return 0;
+}
+module_init(proc_vmalloc_init);
+#endif
+