mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
authorRoman Pen <r.peniaev@gmail.com>
Wed, 15 Apr 2015 23:13:55 +0000 (16:13 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 15 Apr 2015 23:35:18 +0000 (16:35 -0700)
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map.  None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.

Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.

But long time ago Zhang Yanfei removed alloc_map by these two commits:

  mm/vmalloc.c: remove dead code in vb_alloc
     3fcd76e8028e0be37b02a2002b4f56755daeda06
  mm/vmalloc.c: remove alloc_map from vmap_block
     b8e748b6c32999f221ea4786557b8e7e6c4e4e7a

In this patch I replaced dirty_map with two range variables: dirty min and
max.  These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.

Why it was made?  Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true.  To avoid complexity seems it is
better to use something simple, like min or max range values.  Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros.  Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block).  Here I replaced the whole bitmap with two
longs.

Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.

Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmalloc.c

index 21ec16b7e6e1592bb346033750f2a0161417ea93..2faaa2976447a104ac5017ce7f4ad1e52b808b0b 100644 (file)
@@ -765,7 +765,7 @@ struct vmap_block {
        spinlock_t lock;
        struct vmap_area *va;
        unsigned long free, dirty;
-       DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
+       unsigned long dirty_min, dirty_max; /*< dirty range */
        struct list_head free_list;
        struct rcu_head rcu_head;
        struct list_head purge;
@@ -851,7 +851,8 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
        BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
        vb->free = VMAP_BBMAP_BITS - (1UL << order);
        vb->dirty = 0;
-       bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
+       vb->dirty_min = VMAP_BBMAP_BITS;
+       vb->dirty_max = 0;
        INIT_LIST_HEAD(&vb->free_list);
 
        vb_idx = addr_to_vb_idx(va->va_start);
@@ -902,7 +903,8 @@ static void purge_fragmented_blocks(int cpu)
                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->dirty_map, VMAP_BBMAP_BITS);
+                       vb->dirty_min = 0;
+                       vb->dirty_max = VMAP_BBMAP_BITS;
                        spin_lock(&vbq->lock);
                        list_del_rcu(&vb->free_list);
                        spin_unlock(&vbq->lock);
@@ -995,6 +997,7 @@ static void vb_free(const void *addr, unsigned long size)
        order = get_order(size);
 
        offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
+       offset >>= PAGE_SHIFT;
 
        vb_idx = addr_to_vb_idx((unsigned long)addr);
        rcu_read_lock();
@@ -1005,7 +1008,10 @@ static void vb_free(const void *addr, unsigned long size)
        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));
+
+       /* Expand dirty range */
+       vb->dirty_min = min(vb->dirty_min, offset);
+       vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
 
        vb->dirty += 1UL << order;
        if (vb->dirty == VMAP_BBMAP_BITS) {
@@ -1044,25 +1050,18 @@ void vm_unmap_aliases(void)
 
                rcu_read_lock();
                list_for_each_entry_rcu(vb, &vbq->free, free_list) {
-                       int i, j;
-
                        spin_lock(&vb->lock);
-                       i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
-                       if (i < VMAP_BBMAP_BITS) {
+                       if (vb->dirty) {
+                               unsigned long va_start = vb->va->va_start;
                                unsigned long s, e;
 
-                               j = find_last_bit(vb->dirty_map,
-                                                       VMAP_BBMAP_BITS);
-                               j = j + 1; /* need exclusive index */
+                               s = va_start + (vb->dirty_min << PAGE_SHIFT);
+                               e = va_start + (vb->dirty_max << PAGE_SHIFT);
 
-                               s = vb->va->va_start + (i << PAGE_SHIFT);
-                               e = vb->va->va_start + (j << PAGE_SHIFT);
-                               flush = 1;
+                               start = min(s, start);
+                               end   = max(e, end);
 
-                               if (s < start)
-                                       start = s;
-                               if (e > end)
-                                       end = e;
+                               flush = 1;
                        }
                        spin_unlock(&vb->lock);
                }