]> pilppa.com Git - linux-2.6-omap-h63xx.git/commitdiff
mm: rewrite vmap layer
authorNick Piggin <npiggin@suse.de>
Sun, 19 Oct 2008 03:27:03 +0000 (20:27 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 20 Oct 2008 15:52:32 +0000 (08:52 -0700)
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).

The biggest problem with vmap is actually vunmap.  Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache.  This is all done under a global lock.  As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
 This gives terrible quadratic scalability characteristics.

Another problem is that the entire vmap subsystem works under a single
lock.  It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.

This is a rewrite of vmap subsystem to solve those problems.  The existing
vmalloc API is implemented on top of the rewritten subsystem.

The TLB flushing problem is solved by using lazy TLB unmapping.  vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated.  So the addresses aren't allocated again until
a subsequent TLB flush.  A single TLB flush then can flush multiple
vunmaps from each CPU.

XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.

The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.

There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.

To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap.  Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).

As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages.  Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron.  Results are
in nanoseconds per map+touch+unmap.

threads           vanilla         vmap rewrite
1                 14700           2900
2                 33600           3000
4                 49500           2800
8                 70631           2900

So with a 8 cores, the rewritten version is already 25x faster.

In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram...  along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system.  I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now.  vmap is pretty well blown off the
profiles.

Before:
1352059 total                                      0.1401
798784 _write_lock                              8320.6667 <- vmlist_lock
529313 default_idle                             1181.5022
 15242 smp_call_function                         15.8771  <- vmap tlb flushing
  2472 __get_vm_area_node                         1.9312  <- vmap
  1762 remove_vm_area                             4.5885  <- vunmap
   316 map_vm_area                                0.2297  <- vmap
   312 kfree                                      0.1950
   300 _spin_lock                                 3.1250
   252 sn_send_IPI_phys                           0.4375  <- tlb flushing
   238 vmap                                       0.8264  <- vmap
   216 find_lock_page                             0.5192
   196 find_next_bit                              0.3603
   136 sn2_send_IPI                               0.2024
   130 pio_phys_write_mmr                         2.0312
   118 unmap_kernel_range                         0.1229

After:
 78406 total                                      0.0081
 40053 default_idle                              89.4040
 33576 ia64_spinlock_contention                 349.7500
  1650 _spin_lock                                17.1875
   319 __reg_op                                   0.5538
   281 _atomic_dec_and_lock                       1.0977
   153 mutex_unlock                               1.5938
   123 iget_locked                                0.1671
   117 xfs_dir_lookup                             0.1662
   117 dput                                       0.1406
   114 xfs_iget_core                              0.0268
    92 xfs_da_hashname                            0.1917
    75 d_alloc                                    0.0670
    68 vmap_page_range                            0.0462 <- vmap
    58 kmem_cache_alloc                           0.0604
    57 memset                                     0.0540
    52 rb_next                                    0.1625
    50 __copy_user                                0.0208
    49 bitmap_find_free_region                    0.2188 <- vmap
    46 ia64_sn_udelay                             0.1106
    45 find_inode_fast                            0.1406
    42 memcmp                                     0.2188
    42 finish_task_switch                         0.1094
    42 __d_lookup                                 0.0410
    40 radix_tree_lookup_slot                     0.1250
    37 _spin_unlock_irqrestore                    0.3854
    36 xfs_bmapi                                  0.0050
    36 kmem_cache_free                            0.0256
    35 xfs_vn_getattr                             0.0322
    34 radix_tree_lookup                          0.1062
    33 __link_path_walk                           0.0035
    31 xfs_da_do_buf                              0.0091
    30 _xfs_buf_find                              0.0204
    28 find_get_page                              0.0875
    27 xfs_iread                                  0.0241
    27 __strncpy_from_user                        0.2812
    26 _xfs_buf_initialize                        0.0406
    24 _xfs_buf_lookup_pages                      0.0179
    24 vunmap_page_range                          0.0250 <- vunmap
    23 find_lock_page                             0.0799
    22 vm_map_ram                                 0.0087 <- vmap
    20 kfree                                      0.0125
    19 put_page                                   0.0330
    18 __kmalloc                                  0.0176
    17 xfs_da_node_lookup_int                     0.0086
    17 _read_lock                                 0.0885
    17 page_waitqueue                             0.0664

vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.

[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
arch/x86/mm/pageattr.c
arch/x86/xen/enlighten.c
arch/x86/xen/mmu.c
include/linux/vmalloc.h
init/main.c
mm/vmalloc.c

index a9ec89c3fbca32c3fda0e65d8da8846c8f96ed2a..407d8784f669603fd48667c8d2ee22bb8e870f35 100644 (file)
@@ -792,6 +792,8 @@ static int change_page_attr_set_clr(unsigned long *addr, int numpages,
        /* Must avoid aliasing mappings in the highmem code */
        kmap_flush_unused();
 
+       vm_unmap_aliases();
+
        cpa.vaddr = addr;
        cpa.numpages = numpages;
        cpa.mask_set = mask_set;
index 0013a729b41ddc93045510068da9bcbbadf76433..b61534c7a4c4d5894812c5bb97a89f29a0f72338 100644 (file)
@@ -871,6 +871,7 @@ static void xen_alloc_ptpage(struct mm_struct *mm, unsigned long pfn, unsigned l
                        /* make sure there are no stray mappings of
                           this page */
                        kmap_flush_unused();
+                       vm_unmap_aliases();
        }
 }
 
index ae173f6edd8ba044f5cc5372193ea7ce0af2e299..d4d52f5a1cf7bae392801e19367030ebe58ccf79 100644 (file)
@@ -846,6 +846,7 @@ static void __xen_pgd_pin(struct mm_struct *mm, pgd_t *pgd)
                /* re-enable interrupts for kmap_flush_unused */
                xen_mc_issue(0);
                kmap_flush_unused();
+               vm_unmap_aliases();
                xen_mc_batch();
        }
 
index 328eb40227277967c8a6179efb56ad14e03fc233..4c28c4d564e2da1e5e6112e0e5081d4f08ebd942 100644 (file)
@@ -2,6 +2,7 @@
 #define _LINUX_VMALLOC_H
 
 #include <linux/spinlock.h>
+#include <linux/init.h>
 #include <asm/page.h>          /* pgprot_t */
 
 struct vm_area_struct;         /* vma defining user mapping in mm_types.h */
@@ -23,7 +24,6 @@ struct vm_area_struct;                /* vma defining user mapping in mm_types.h */
 #endif
 
 struct vm_struct {
-       /* keep next,addr,size together to speedup lookups */
        struct vm_struct        *next;
        void                    *addr;
        unsigned long           size;
@@ -37,6 +37,19 @@ struct vm_struct {
 /*
  *     Highlevel APIs for driver use
  */
+extern void vm_unmap_ram(const void *mem, unsigned int count);
+extern void *vm_map_ram(struct page **pages, unsigned int count,
+                               int node, pgprot_t prot);
+extern void vm_unmap_aliases(void);
+
+#ifdef CONFIG_MMU
+extern void __init vmalloc_init(void);
+#else
+static inline void vmalloc_init(void)
+{
+}
+#endif
+
 extern void *vmalloc(unsigned long size);
 extern void *vmalloc_user(unsigned long size);
 extern void *vmalloc_node(unsigned long size, int node);
index 27f6bf6108e96908b10df5b20e0011aaebb3266c..4371d11721f618f45ccfc65b799fbbdfebbf8140 100644 (file)
@@ -27,6 +27,7 @@
 #include <linux/gfp.h>
 #include <linux/percpu.h>
 #include <linux/kmod.h>
+#include <linux/vmalloc.h>
 #include <linux/kernel_stat.h>
 #include <linux/start_kernel.h>
 #include <linux/security.h>
@@ -642,6 +643,7 @@ asmlinkage void __init start_kernel(void)
                initrd_start = 0;
        }
 #endif
+       vmalloc_init();
        vfs_caches_init_early();
        cpuset_init_early();
        mem_init();
index bba06c41fc59ed10be118f7af24ecfa71d023c33..712ae47af0bf5f23a4b32ea0497dd2f6e8272085 100644 (file)
@@ -8,6 +8,7 @@
  *  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/debugobjects.h>
 #include <linux/vmalloc.h>
 #include <linux/kallsyms.h>
+#include <linux/list.h>
+#include <linux/rbtree.h>
+#include <linux/radix-tree.h>
+#include <linux/rcupdate.h>
 
+#include <asm/atomic.h>
 #include <asm/uaccess.h>
 #include <asm/tlbflush.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, void *caller);
+/*** Page table manipulation functions ***/
 
 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
 {
@@ -40,8 +42,7 @@ 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;
@@ -55,8 +56,7 @@ 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;
@@ -70,12 +70,10 @@ 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);
@@ -86,35 +84,36 @@ void unmap_kernel_range(unsigned long addr, unsigned long size)
                        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;
@@ -124,14 +123,14 @@ 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;
@@ -141,44 +140,49 @@ 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(unsigned long addr, 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;
+       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;
        } while (pgd++, addr = next, addr != end);
-       flush_cache_vmap((unsigned long) area->addr, end);
-       return err;
+       flush_cache_vmap(addr, end);
+
+       if (unlikely(err))
+               return err;
+       return nr;
 }
-EXPORT_SYMBOL_GPL(map_vm_area);
 
 /*
- * Map a vmalloc()-space virtual address to the physical page.
+ * 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);
-       pud_t *pud;
-       pmd_t *pmd;
-       pte_t *ptep, pte;
 
        /*
         * XXX we might need to change this if we add VIRTUAL_BUG_ON for
@@ -188,10 +192,12 @@ struct page *vmalloc_to_page(const void *vmalloc_addr)
                        !is_module_address(addr));
 
        if (!pgd_none(*pgd)) {
-               pud = pud_offset(pgd, addr);
+               pud_t *pud = pud_offset(pgd, addr);
                if (!pud_none(*pud)) {
-                       pmd = pmd_offset(pud, addr);
+                       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))
@@ -213,13 +219,751 @@ unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 }
 EXPORT_SYMBOL(vmalloc_to_pfn);
 
-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, void *caller)
+
+/*** 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 struct vmap_area *__find_vmap_area(unsigned long addr)
 {
-       struct vm_struct **p, *tmp, *area;
-       unsigned long align = 1;
+       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;
+
+               parent = *p;
+               tmp = rb_entry(parent, struct vmap_area, rb_node);
+               if (va->va_start < tmp->va_end)
+                       p = &(*p)->rb_left;
+               else if (va->va_end > tmp->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 & ~PAGE_MASK);
+
+       addr = ALIGN(vstart, align);
+
+       va = kmalloc_node(sizeof(struct vmap_area),
+                       gfp_mask & GFP_RECLAIM_MASK, node);
+       if (unlikely(!va))
+               return ERR_PTR(-ENOMEM);
+
+retry:
+       spin_lock(&vmap_area_lock);
+       /* 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);
+
+                       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) {
+               spin_unlock(&vmap_area_lock);
+               if (!purged) {
+                       purge_vmap_area_lazy();
+                       purged = 1;
+                       goto retry;
+               }
+               if (printk_ratelimit())
+                       printk(KERN_WARNING "vmap allocation failed: "
+                                "use vmalloc=<size> to increase size.\n");
+               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);
+
+       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);
+}
+
+/*
+ * 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);
+
+/*
+ * 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;
+       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);
+
+       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;
+                       unmap_vmap_area(va);
+                       list_add_tail(&va->purge_list, &valist);
+                       va->flags |= VM_LAZY_FREEING;
+                       va->flags &= ~VM_LAZY_FREE;
+               }
+       }
+       rcu_read_unlock();
+
+       if (nr) {
+               BUG_ON(nr > atomic_read(&vmap_lazy_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(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.
+ */
+static void purge_vmap_area_lazy(void)
+{
+       unsigned long start = ULONG_MAX, end = 0;
+
+       __purge_vmap_area_lazy(&start, &end, 0, 0);
+}
+
+/*
+ * Free and unmap a vmap area
+ */
+static void free_unmap_vmap_area(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()))
+               purge_vmap_area_lazy();
+}
+
+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)
+
+struct vmap_block_queue {
+       spinlock_t lock;
+       struct list_head free;
+       struct list_head dirty;
+       unsigned int nr_dirty;
+};
+
+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);
+       union {
+               struct {
+                       struct list_head free_list;
+                       struct list_head dirty_list;
+               };
+               struct rcu_head rcu_head;
+       };
+};
+
+/* 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 vmap_block *new_vmap_block(gfp_t gfp_mask)
+{
+       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 (unlikely(IS_ERR(va))) {
+               kfree(vb);
+               return ERR_PTR(PTR_ERR(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);
+       INIT_LIST_HEAD(&vb->dirty_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(&vb->free_list, &vbq->free);
+       spin_unlock(&vbq->lock);
+       put_cpu_var(vmap_cpu_blocks);
+
+       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;
+
+       spin_lock(&vb->vbq->lock);
+       if (!list_empty(&vb->free_list))
+               list_del(&vb->free_list);
+       if (!list_empty(&vb->dirty_list))
+               list_del(&vb->dirty_list);
+       spin_unlock(&vb->vbq->lock);
+
+       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_unmap_vmap_area(vb->va);
+       call_rcu(&vb->rcu_head, rcu_free_vb);
+}
+
+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;
+
+       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);
+               i = bitmap_find_free_region(vb->alloc_map,
+                                               VMAP_BBMAP_BITS, order);
+
+               if (i >= 0) {
+                       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_init(&vb->free_list);
+                               spin_unlock(&vbq->lock);
+                       }
+                       spin_unlock(&vb->lock);
+                       break;
+               }
+               spin_unlock(&vb->lock);
+       }
+       put_cpu_var(vmap_cpu_blocks);
+       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);
+       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);
+
+       spin_lock(&vb->lock);
+       bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
+       if (!vb->dirty) {
+               spin_lock(&vb->vbq->lock);
+               list_add(&vb->dirty_list, &vb->vbq->dirty);
+               spin_unlock(&vb->vbq->lock);
+       }
+       vb->dirty += 1UL << order;
+       if (vb->dirty == VMAP_BBMAP_BITS) {
+               BUG_ON(vb->free || !list_empty(&vb->free_list));
+               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;
+
+       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);
+                               vunmap_page_range(s, e);
+                               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);
+
+       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);
+
+void __init vmalloc_init(void)
+{
+       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);
+               INIT_LIST_HEAD(&vbq->dirty);
+               vbq->nr_dirty = 0;
+       }
+}
+
+void unmap_kernel_range(unsigned long addr, unsigned long size)
+{
+       unsigned long end = addr + size;
+       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 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, void *caller)
+{
+       static struct vmap_area *va;
+       struct vm_struct *area;
+       struct vm_struct *tmp, **p;
+       unsigned long align = 1;
 
        BUG_ON(in_interrupt());
        if (flags & VM_IOREMAP) {
@@ -232,13 +976,12 @@ __get_vm_area_node(unsigned long size, unsigned long flags, unsigned long start,
 
                align = 1ul << bit;
        }
-       addr = ALIGN(start, align);
+
        size = PAGE_ALIGN(size);
        if (unlikely(!size))
                return NULL;
 
        area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
-
        if (unlikely(!area))
                return NULL;
 
@@ -247,48 +990,32 @@ __get_vm_area_node(unsigned long size, unsigned long flags, unsigned long start,
         */
        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;
        }
-       if ((size + addr) < addr)
-               goto out;
-       if (addr > end - size)
-               goto out;
-
-found:
-       area->next = *p;
-       *p = area;
 
        area->flags = flags;
-       area->addr = (void *)addr;
+       area->addr = (void *)va->va_start;
        area->size = size;
        area->pages = NULL;
        area->nr_pages = 0;
        area->phys_addr = 0;
        area->caller = caller;
+       va->private = area;
+       va->flags |= VM_VM_AREA;
+
+       write_lock(&vmlist_lock);
+       for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
+               if (tmp->addr >= area->addr)
+                       break;
+       }
+       area->next = *p;
+       *p = area;
        write_unlock(&vmlist_lock);
 
        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,
@@ -328,39 +1055,15 @@ struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
                                  gfp_mask, __builtin_return_address(0));
 }
 
-/* Caller must hold vmlist_lock */
-static struct vm_struct *__find_vm_area(const 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(const 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;
 }
 
 /**
@@ -373,11 +1076,24 @@ found:
  */
 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;
+               free_unmap_vmap_area(va);
+               vm->size -= PAGE_SIZE;
+
+               write_lock(&vmlist_lock);
+               for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
+                       ;
+               *p = tmp->next;
+               write_unlock(&vmlist_lock);
+
+               return vm;
+       }
+       return NULL;
 }
 
 static void __vunmap(const void *addr, int deallocate_pages)
@@ -487,6 +1203,8 @@ void *vmap(struct page **pages, unsigned int count,
 }
 EXPORT_SYMBOL(vmap);
 
+static void *__vmalloc_node(unsigned long size, 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)
 {
@@ -613,10 +1331,8 @@ void *vmalloc_user(unsigned long size)
 
        ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
        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;
 }
@@ -696,10 +1412,8 @@ void *vmalloc_32_user(unsigned long size)
 
        ret = __vmalloc(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL);
        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;
 }
@@ -800,26 +1514,25 @@ 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;
@@ -832,11 +1545,7 @@ 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);