]> nv-tegra.nvidia Code Review - linux-3.10.git/blobdiff - mm/vmalloc.c
mm: swap: implement generic handler for swap_activate
[linux-3.10.git] / mm / vmalloc.c
index 78ec9d8bc57ce4fe516a52be62c10ca625369d52..2bb90b1d241cc872da1e13dd80b449b2d323e812 100644 (file)
@@ -26,7 +26,7 @@
 #include <linux/rcupdate.h>
 #include <linux/pfn.h>
 #include <linux/kmemleak.h>
-#include <asm/atomic.h>
+#include <linux/atomic.h>
 #include <asm/uaccess.h>
 #include <asm/tlbflush.h>
 #include <asm/shmparam.h>
@@ -256,13 +256,20 @@ struct vmap_area {
        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 vm_struct *vm;
        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 struct rb_root vmap_area_root = RB_ROOT;
+
+/* The vmap cache globals are protected by vmap_area_lock */
+static struct rb_node *free_vmap_cache;
+static unsigned long cached_hole_size;
+static unsigned long cached_vstart;
+static unsigned long cached_align;
+
 static unsigned long vmap_area_pcpu_hole;
 
 static struct vmap_area *__find_vmap_area(unsigned long addr)
@@ -331,9 +338,11 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
        struct rb_node *n;
        unsigned long addr;
        int purged = 0;
+       struct vmap_area *first;
 
        BUG_ON(!size);
        BUG_ON(size & ~PAGE_MASK);
+       BUG_ON(!is_power_of_2(align));
 
        va = kmalloc_node(sizeof(struct vmap_area),
                        gfp_mask & GFP_RECLAIM_MASK, node);
@@ -341,91 +350,127 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
                return ERR_PTR(-ENOMEM);
 
 retry:
-       addr = ALIGN(vstart, align);
-
        spin_lock(&vmap_area_lock);
-       if (addr + size - 1 < addr)
-               goto overflow;
+       /*
+        * Invalidate cache if we have more permissive parameters.
+        * cached_hole_size notes the largest hole noticed _below_
+        * the vmap_area cached in free_vmap_cache: if size fits
+        * into that hole, we want to scan from vstart to reuse
+        * the hole instead of allocating above free_vmap_cache.
+        * Note that __free_vmap_area may update free_vmap_cache
+        * without updating cached_hole_size or cached_align.
+        */
+       if (!free_vmap_cache ||
+                       size < cached_hole_size ||
+                       vstart < cached_vstart ||
+                       align < cached_align) {
+nocache:
+               cached_hole_size = 0;
+               free_vmap_cache = NULL;
+       }
+       /* record if we encounter less permissive parameters */
+       cached_vstart = vstart;
+       cached_align = align;
+
+       /* find starting point for our search */
+       if (free_vmap_cache) {
+               first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
+               addr = ALIGN(first->va_end, align);
+               if (addr < vstart)
+                       goto nocache;
+               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;
+       } else {
+               addr = ALIGN(vstart, align);
+               if (addr + size - 1 < addr)
+                       goto overflow;
+
+               n = vmap_area_root.rb_node;
+               first = NULL;
 
-               do {
+               while (n) {
                        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;
+                               if (tmp->va_start <= addr)
+                                       break;
+                               n = n->rb_left;
+                       } else
                                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;
-               }
+       /* from the starting point, walk areas until a suitable hole is found */
+       while (addr + size > first->va_start && addr + size <= vend) {
+               if (addr + cached_hole_size < first->va_start)
+                       cached_hole_size = first->va_start - addr;
+               addr = ALIGN(first->va_end, align);
+               if (addr + size - 1 < addr)
+                       goto overflow;
 
-               while (addr + size > first->va_start && addr + size <= vend) {
-                       addr = ALIGN(first->va_end + PAGE_SIZE, align);
-                       if (addr + size - 1 < addr)
-                               goto overflow;
+               if (list_is_last(&first->list, &vmap_area_list))
+                       goto found;
 
-                       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);
+               first = list_entry(first->list.next,
+                               struct vmap_area, list);
        }
 
-       BUG_ON(addr & (align-1));
+found:
+       if (addr + size > vend)
+               goto overflow;
 
        va->va_start = addr;
        va->va_end = addr + size;
        va->flags = 0;
        __insert_vmap_area(va);
+       free_vmap_cache = &va->rb_node;
        spin_unlock(&vmap_area_lock);
 
-       return va;
-}
+       BUG_ON(va->va_start & (align-1));
+       BUG_ON(va->va_start < vstart);
+       BUG_ON(va->va_end > vend);
 
-static void rcu_free_va(struct rcu_head *head)
-{
-       struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
+       return va;
 
+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);
 }
 
 static void __free_vmap_area(struct vmap_area *va)
 {
        BUG_ON(RB_EMPTY_NODE(&va->rb_node));
+
+       if (free_vmap_cache) {
+               if (va->va_end < cached_vstart) {
+                       free_vmap_cache = NULL;
+               } else {
+                       struct vmap_area *cache;
+                       cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
+                       if (va->va_start <= cache->va_start) {
+                               free_vmap_cache = rb_prev(&va->rb_node);
+                               /*
+                                * We don't try to update cached_hole_size or
+                                * cached_align, but it won't go very wrong.
+                                */
+                       }
+               }
+       }
        rb_erase(&va->rb_node, &vmap_area_root);
        RB_CLEAR_NODE(&va->rb_node);
        list_del_rcu(&va->list);
@@ -439,7 +484,7 @@ static void __free_vmap_area(struct vmap_area *va)
        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);
+       kfree_rcu(va, rcu_head);
 }
 
 /*
@@ -680,9 +725,10 @@ static void free_unmap_vmap_area_addr(unsigned long addr)
 #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_BBMAP_BITS                \
+               VMAP_MIN(VMAP_BBMAP_BITS_MAX,   \
+               VMAP_MAX(VMAP_BBMAP_BITS_MIN,   \
+                       VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
 
 #define VMAP_BLOCK_SIZE                (VMAP_BBMAP_BITS * PAGE_SIZE)
 
@@ -748,7 +794,7 @@ static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
        va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
                                        VMALLOC_START, VMALLOC_END,
                                        node, gfp_mask);
-       if (unlikely(IS_ERR(va))) {
+       if (IS_ERR(va)) {
                kfree(vb);
                return ERR_CAST(va);
        }
@@ -785,13 +831,6 @@ static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
        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;
@@ -804,7 +843,7 @@ static void free_vmap_block(struct vmap_block *vb)
        BUG_ON(tmp != vb);
 
        free_vmap_area_noflush(vb->va);
-       call_rcu(&vb->rcu_head, rcu_free_vb);
+       kfree_rcu(vb, rcu_head);
 }
 
 static void purge_fragmented_blocks(int cpu)
@@ -865,6 +904,14 @@ static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
 
        BUG_ON(size & ~PAGE_MASK);
        BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
+       if (WARN_ON(size == 0)) {
+               /*
+                * Allocating 0 bytes isn't what caller wants since
+                * get_order(0) returns funny result. Just warn and terminate
+                * early.
+                */
+               return NULL;
+       }
        order = get_order(size);
 
 again:
@@ -1078,6 +1125,32 @@ void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t pro
 }
 EXPORT_SYMBOL(vm_map_ram);
 
+/**
+ * vm_area_add_early - add vmap area early during boot
+ * @vm: vm_struct to add
+ *
+ * This function is used to add fixed kernel vm area to vmlist before
+ * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
+ * should contain proper values and the other fields should be zero.
+ *
+ * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
+ */
+void __init vm_area_add_early(struct vm_struct *vm)
+{
+       struct vm_struct *tmp, **p;
+
+       BUG_ON(vmap_initialized);
+       for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
+               if (tmp->addr >= vm->addr) {
+                       BUG_ON(tmp->addr < vm->addr + vm->size);
+                       break;
+               } else
+                       BUG_ON(tmp->addr + tmp->size > vm->addr);
+       }
+       vm->next = *p;
+       *p = vm;
+}
+
 /**
  * vm_area_register_early - register vmap area early during boot
  * @vm: vm_struct to register
@@ -1100,8 +1173,7 @@ void __init vm_area_register_early(struct vm_struct *vm, size_t align)
 
        vm->addr = (void *)addr;
 
-       vm->next = vmlist;
-       vmlist = vm;
+       vm_area_add_early(vm);
 }
 
 void __init vmalloc_init(void)
@@ -1121,9 +1193,10 @@ void __init vmalloc_init(void)
        /* 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->flags = VM_VM_AREA;
                va->va_start = (unsigned long)tmp->addr;
                va->va_end = va->va_start + tmp->size;
+               va->vm = tmp;
                __insert_vmap_area(va);
        }
 
@@ -1175,6 +1248,7 @@ 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
@@ -1213,18 +1287,22 @@ EXPORT_SYMBOL_GPL(map_vm_area);
 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)
+static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
+                             unsigned long flags, const 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->vm = vm;
        va->flags |= VM_VM_AREA;
+}
 
+static void insert_vmalloc_vmlist(struct vm_struct *vm)
+{
+       struct vm_struct *tmp, **p;
+
+       vm->flags &= ~VM_UNLIST;
        write_lock(&vmlist_lock);
        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
                if (tmp->addr >= vm->addr)
@@ -1235,11 +1313,18 @@ static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
        write_unlock(&vmlist_lock);
 }
 
+static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
+                             unsigned long flags, const void *caller)
+{
+       setup_vmalloc_vm(vm, va, flags, caller);
+       insert_vmalloc_vmlist(vm);
+}
+
 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)
+               unsigned long end, int node, gfp_t gfp_mask, const void *caller)
 {
-       static struct vmap_area *va;
+       struct vmap_area *va;
        struct vm_struct *area;
 
        BUG_ON(in_interrupt());
@@ -1273,7 +1358,18 @@ static struct vm_struct *__get_vm_area_node(unsigned long size,
                return NULL;
        }
 
-       insert_vmalloc_vm(area, va, flags, caller);
+       /*
+        * When this function is called from __vmalloc_node_range,
+        * we do not add vm_struct to vmlist here to avoid
+        * accessing uninitialized members of vm_struct such as
+        * pages and nr_pages fields. They will be set later.
+        * To distinguish it from others, we use a VM_UNLIST flag.
+        */
+       if (flags & VM_UNLIST)
+               setup_vmalloc_vm(area, va, flags, caller);
+       else
+               insert_vmalloc_vm(area, va, flags, caller);
+
        return area;
 }
 
@@ -1287,7 +1383,7 @@ 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)
+                                      const void *caller)
 {
        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
                                  caller);
@@ -1309,19 +1405,27 @@ struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
 }
 
 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
-                               void *caller)
+                               const void *caller)
 {
        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
                                                -1, GFP_KERNEL, caller);
 }
 
-static struct vm_struct *find_vm_area(const void *addr)
+/**
+ *     find_vm_area  -  find a continuous kernel virtual area
+ *     @addr:          base address
+ *
+ *     Search for the kernel VM area starting at @addr, and return it.
+ *     It is up to the caller to do all required locking to keep the returned
+ *     pointer valid.
+ */
+struct vm_struct *find_vm_area(const void *addr)
 {
        struct vmap_area *va;
 
        va = find_vmap_area((unsigned long)addr);
        if (va && va->flags & VM_VM_AREA)
-               return va->private;
+               return va->vm;
 
        return NULL;
 }
@@ -1340,18 +1444,21 @@ struct vm_struct *remove_vm_area(const void *addr)
 
        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);
+               struct vm_struct *vm = va->vm;
+
+               if (!(vm->flags & VM_UNLIST)) {
+                       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);
@@ -1477,10 +1584,11 @@ EXPORT_SYMBOL(vmap);
 
 static void *__vmalloc_node(unsigned long size, unsigned long align,
                            gfp_t gfp_mask, pgprot_t prot,
-                           int node, void *caller);
+                           int node, const void *caller);
 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
-                                pgprot_t prot, int node, void *caller)
+                                pgprot_t prot, int node, const void *caller)
 {
+       const int order = 0;
        struct page **pages;
        unsigned int nr_pages, array_size, i;
        gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
@@ -1507,11 +1615,12 @@ static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
 
        for (i = 0; i < area->nr_pages; i++) {
                struct page *page;
+               gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
 
                if (node < 0)
-                       page = alloc_page(gfp_mask);
+                       page = alloc_page(tmp_mask);
                else
-                       page = alloc_pages_node(node, gfp_mask, 0);
+                       page = alloc_pages_node(node, tmp_mask, order);
 
                if (unlikely(!page)) {
                        /* Successfully allocated i pages, free them in __vunmap() */
@@ -1526,29 +1635,19 @@ static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
        return area->addr;
 
 fail:
+       warn_alloc_failed(gfp_mask, order,
+                         "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
+                         (area->nr_pages*PAGE_SIZE), area->size);
        vfree(area->addr);
        return NULL;
 }
 
-void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
-{
-       void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
-                                        __builtin_return_address(0));
-
-       /*
-        * 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, area->size - PAGE_SIZE, 3, gfp_mask);
-
-       return addr;
-}
-
 /**
- *     __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
@@ -1558,9 +1657,9 @@ void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
  *     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)
+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, const void *caller)
 {
        struct vm_struct *area;
        void *addr;
@@ -1568,15 +1667,22 @@ static void *__vmalloc_node(unsigned long size, unsigned long align,
 
        size = PAGE_ALIGN(size);
        if (!size || (size >> PAGE_SHIFT) > totalram_pages)
-               return NULL;
-
-       area = __get_vm_area_node(size, align, VM_ALLOC, VMALLOC_START,
-                                 VMALLOC_END, node, gfp_mask, caller);
+               goto fail;
 
+       area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
+                                 start, end, node, gfp_mask, caller);
        if (!area)
-               return NULL;
+               goto fail;
 
        addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
+       if (!addr)
+               return NULL;
+
+       /*
+        * In this function, newly allocated vm_struct is not added
+        * to vmlist at __get_vm_area_node(). so, it is added here.
+        */
+       insert_vmalloc_vmlist(area);
 
        /*
         * A ref_count = 3 is needed because the vm_struct and vmap_area
@@ -1586,6 +1692,33 @@ static void *__vmalloc_node(unsigned long size, unsigned long align,
        kmemleak_alloc(addr, real_size, 3, gfp_mask);
 
        return addr;
+
+fail:
+       warn_alloc_failed(gfp_mask, 0,
+                         "vmalloc: allocation failure: %lu bytes\n",
+                         real_size);
+       return NULL;
+}
+
+/**
+ *     __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, const 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)
@@ -1790,9 +1923,9 @@ static int aligned_vread(char *buf, char *addr, unsigned long count)
                         * we can expect USER0 is not used (see vread/vwrite's
                         * function description)
                         */
-                       void *map = kmap_atomic(p, KM_USER0);
+                       void *map = kmap_atomic(p);
                        memcpy(buf, map + offset, length);
-                       kunmap_atomic(map, KM_USER0);
+                       kunmap_atomic(map);
                } else
                        memset(buf, 0, length);
 
@@ -1829,9 +1962,9 @@ static int aligned_vwrite(char *buf, char *addr, unsigned long count)
                         * we can expect USER0 is not used (see vread/vwrite's
                         * function description)
                         */
-                       void *map = kmap_atomic(p, KM_USER0);
+                       void *map = kmap_atomic(p);
                        memcpy(map + offset, buf, length);
-                       kunmap_atomic(map, KM_USER0);
+                       kunmap_atomic(map);
                }
                addr += length;
                buf += length;
@@ -1858,9 +1991,7 @@ static int aligned_vwrite(char *buf, char *addr, unsigned long count)
  *     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.
+ *     vm_struct area, returns 0. @buf should be kernel's buffer.
  *
  *     Note: In usual ops, vread() is never necessary because the caller
  *     should know vmalloc() area is valid and can use memcpy().
@@ -1934,16 +2065,12 @@ finished:
  *     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.
+ *     vm_struct area, returns 0. @buf should be kernel's buffer.
  *
  *     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)
@@ -2054,23 +2181,30 @@ 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. */
+       pte_t ***p = data;
+
+       if (p) {
+               *(*p) = pte;
+               (*p)++;
+       }
        return 0;
 }
 
 /**
  *     alloc_vm_area - allocate a range of kernel address space
  *     @size:          size of the area
+ *     @ptes:          returns the PTEs for the address space
  *
  *     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.
+ *     are created.
+ *
+ *     If @ptes is non-NULL, pointers to the PTEs (in init_mm)
+ *     allocated for the VM area are returned.
  */
-struct vm_struct *alloc_vm_area(size_t size)
+struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
 {
        struct vm_struct *area;
 
@@ -2084,15 +2218,11 @@ struct vm_struct *alloc_vm_area(size_t size)
         * 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)) {
+                               size, f, ptes ? &ptes : 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);
@@ -2196,17 +2326,16 @@ static unsigned long pvm_determine_end(struct vmap_area **pnext,
  * @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
- * @gfp_mask: allocation mask
  *
  * 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.  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.
+ * 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
@@ -2217,7 +2346,7 @@ static unsigned long pvm_determine_end(struct vmap_area **pnext,
  */
 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
                                     const size_t *sizes, int nr_vms,
-                                    size_t align, gfp_t gfp_mask)
+                                    size_t align)
 {
        const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
@@ -2227,8 +2356,6 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
        unsigned long base, start, end, last_end;
        bool purged = false;
 
-       gfp_mask &= GFP_RECLAIM_MASK;
-
        /* 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++) {
@@ -2261,14 +2388,14 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
                return NULL;
        }
 
-       vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
-       vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
+       vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
+       vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
        if (!vas || !vms)
-               goto err_free;
+               goto err_free2;
 
        for (area = 0; area < nr_vms; area++) {
-               vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
-               vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
+               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;
        }
@@ -2362,11 +2489,10 @@ found:
 
 err_free:
        for (area = 0; area < nr_vms; area++) {
-               if (vas)
-                       kfree(vas[area]);
-               if (vms)
-                       kfree(vms[area]);
+               kfree(vas[area]);
+               kfree(vms[area]);
        }
+err_free2:
        kfree(vas);
        kfree(vms);
        return NULL;