mm/vmalloc: fix possible exhaustion of vmalloc space caused by vm_map_ram allocator

Recently I came across high fragmentation of vm_map_ram allocator:
vmap_block has free space, but still new blocks continue to appear.
Further investigation showed that certain mapping/unmapping sequences
can exhaust vmalloc space.  On small 32bit systems that's not a big
problem, cause purging will be called soon on a first allocation failure
(alloc_vmap_area), but on 64bit machines, e.g.  x86_64 has 45 bits of
vmalloc space, that can be a disaster.

1) I came up with a simple allocation sequence, which exhausts virtual
   space very quickly:

  while (iters) {

                /* Map/unmap big chunk */
                vaddr = vm_map_ram(pages, 16, -1, PAGE_KERNEL);
                vm_unmap_ram(vaddr, 16);

                /* Map/unmap small chunks.
                 *
                 * -1 for hole, which should be left at the end of each block
                 * to keep it partially used, with some free space available */
                for (i = 0; i < (VMAP_BBMAP_BITS - 16) / 8 - 1; i++) {
                        vaddr = vm_map_ram(pages, 8, -1, PAGE_KERNEL);
                        vm_unmap_ram(vaddr, 8);
                }
  }

The idea behind is simple:

 1. We have to map a big chunk, e.g. 16 pages.

 2. Then we have to occupy the remaining space with smaller chunks, i.e.
    8 pages. At the end small hole should remain to keep block in free list,
    but do not let big chunk to occupy remaining space.

 3. Goto 1 - allocation request of 16 pages can't be completed (only 8 slots
    are left free in the block in the #2 step), new block will be allocated,
    all further requests will lay into newly allocated block.

To have some measurement numbers for all further tests I setup ftrace and
enabled 4 basic calls in a function profile:

        echo vm_map_ram              > /sys/kernel/debug/tracing/set_ftrace_filter;
        echo alloc_vmap_area        >> /sys/kernel/debug/tracing/set_ftrace_filter;
        echo vm_unmap_ram           >> /sys/kernel/debug/tracing/set_ftrace_filter;
        echo free_vmap_block        >> /sys/kernel/debug/tracing/set_ftrace_filter;

So for this scenario I got these results:

BEFORE (all new blocks are put to the head of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
  Function                               Hit    Time            Avg             s^2
  --------                               ---    ----            ---             ---
  vm_map_ram                          126000    30683.30 us     0.243 us        30819.36 us
  vm_unmap_ram                        126000    22003.24 us     0.174 us        340.886 us
  alloc_vmap_area                       1000    4132.065 us     4.132 us        0.903 us

AFTER (all new blocks are put to the tail of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
  Function                               Hit    Time            Avg             s^2
  --------                               ---    ----            ---             ---
  vm_map_ram                          126000    28713.13 us     0.227 us        24944.70 us
  vm_unmap_ram                        126000    20403.96 us     0.161 us        1429.872 us
  alloc_vmap_area                        993    3916.795 us     3.944 us        29.370 us
  free_vmap_block                        992    654.157 us      0.659 us        1.273 us

SUMMARY:

The most interesting numbers in those tables are numbers of block
allocations and deallocations: alloc_vmap_area and free_vmap_block
calls, which show that before the change blocks were not freed, and
virtual space and physical memory (vmap_block structure allocations,
etc) were consumed.

Average time which were spent in vm_map_ram/vm_unmap_ram became slightly
better.  That can be explained with a reasonable amount of blocks in a
free list, which we need to iterate to find a suitable free block.

2) Another scenario is a random allocation:

  while (iters) {

                /* Randomly take number from a range [1..32/64] */
                nr = rand(1, VMAP_MAX_ALLOC);
                vaddr = vm_map_ram(pages, nr, -1, PAGE_KERNEL);
                vm_unmap_ram(vaddr, nr);
  }

I chose mersenne twister PRNG to generate persistent random state to
guarantee that both runs have the same random sequence.  For each
vm_map_ram call random number from [1..32/64] was taken to represent
amount of pages which I do map.

I did 10'000 vm_map_ram calls and got these two tables:

BEFORE (all new blocks are put to the head of a free list)

# cat /sys/kernel/debug/tracing/trace_stat/function0
  Function                               Hit    Time            Avg             s^2
  --------                               ---    ----            ---             ---
  vm_map_ram                           10000    10170.01 us     1.017 us        993.609 us
  vm_unmap_ram                         10000    5321.823 us     0.532 us        59.789 us
  alloc_vmap_area                        420    2150.239 us     5.119 us        3.307 us
  free_vmap_block                         37    159.587 us      4.313 us        134.344 us

AFTER (all new blocks are put to the tail of a free list)

# cat /sys/kernel/debug/tracing/trace_stat/function0
  Function                               Hit    Time            Avg             s^2
  --------                               ---    ----            ---             ---
  vm_map_ram                           10000    7745.637 us     0.774 us        395.229 us
  vm_unmap_ram                         10000    5460.573 us     0.546 us        67.187 us
  alloc_vmap_area                        414    2201.650 us     5.317 us        5.591 us
  free_vmap_block                        412    574.421 us      1.394 us        15.138 us

SUMMARY:

'BEFORE' table shows, that 420 blocks were allocated and only 37 were
freed.  Remained 383 blocks are still in a free list, consuming virtual
space and physical memory.

'AFTER' table shows, that 414 blocks were allocated and 412 were really
freed.  2 blocks remained in a free list.

So fragmentation was dramatically reduced.  Why? Because when we put
newly allocated block to the head, all further requests will occupy new
block, regardless remained space in other blocks.  In this scenario all
requests come randomly.  Eventually remained free space will be less
than requested size, free list will be iterated and it is possible that
nothing will be found there - finally new block will be created.  So
exhaustion in random scenario happens for the maximum possible
allocation size: 32 pages for 32-bit system and 64 pages for 64-bit
system.

Also average cost of vm_map_ram was reduced from 1.017 us to 0.774 us.
Again this can be explained by iteration through smaller list of free
blocks.

3) Next simple scenario is a sequential allocation, when the allocation
   order is increased for each block.  This scenario forces allocator to
   reach maximum amount of partially free blocks in a free list:

  while (iters) {

                /* Populate free list with blocks with remaining space */
                for (order = 0; order <= ilog2(VMAP_MAX_ALLOC); order++) {
                        nr = VMAP_BBMAP_BITS / (1 << order);

                        /* Leave a hole */
                        nr -= 1;

                        for (i = 0; i < nr; i++) {
                                vaddr = vm_map_ram(pages, (1 << order), -1, PAGE_KERNEL);
                                vm_unmap_ram(vaddr, (1 << order));
                }

                /* Completely occupy blocks from a free list */
                for (order = 0; order <= ilog2(VMAP_MAX_ALLOC); order++) {
                        vaddr = vm_map_ram(pages, (1 << order), -1, PAGE_KERNEL);
                        vm_unmap_ram(vaddr, (1 << order));
                }
  }

Results which I got:

BEFORE (all new blocks are put to the head of a free list)

# cat /sys/kernel/debug/tracing/trace_stat/function0
  Function                               Hit    Time            Avg             s^2
  --------                               ---    ----            ---             ---
  vm_map_ram                         2032000    399545.2 us     0.196 us        467123.7 us
  vm_unmap_ram                       2032000    363225.7 us     0.178 us        111405.9 us
  alloc_vmap_area                       7001    30627.76 us     4.374 us        495.755 us
  free_vmap_block                       6993    7011.685 us     1.002 us        159.090 us

AFTER (all new blocks are put to the tail of a free list)

# cat /sys/kernel/debug/tracing/trace_stat/function0
  Function                               Hit    Time            Avg             s^2
  --------                               ---    ----            ---             ---
  vm_map_ram                         2032000    394259.7 us     0.194 us        589395.9 us
  vm_unmap_ram                       2032000    292500.7 us     0.143 us        94181.08 us
  alloc_vmap_area                       7000    31103.11 us     4.443 us        703.225 us
  free_vmap_block                       7000    6750.844 us     0.964 us        119.112 us

SUMMARY:

No surprises here, almost all numbers are the same.

Fixing this fragmentation problem I also did some improvements in a
allocation logic of a new vmap block: occupy block immediately and get
rid of extra search in a free list.

Also I replaced dirty bitmap with min/max dirty range values to make the
logic simpler and slightly faster, since two longs comparison costs
less, than loop thru bitmap.

This patchset raises several questions:

 Q: Think the problem you comments is already known so that I wrote comments
    about it as "it could consume lots of address space through fragmentation".
    Could you tell me about your situation and reason why it should be avoided?
                                                                     Gioh Kim

 A: Indeed, there was a commit 364376383 which adds explicit comment about
    fragmentation.  But fragmentation which is described in this comment caused
    by mixing of long-lived and short-lived objects, when a whole block is pinned
    in memory because some page slots are still in use.  But here I am talking
    about blocks which are free, nobody uses them, and allocator keeps them alive
    forever, continuously allocating new blocks.

 Q: I think that if you put newly allocated block to the tail of a free
    list, below example would results in enormous performance degradation.

    new block: 1MB (256 pages)

    while (iters--) {
      vm_map_ram(3 or something else not dividable for 256) * 85
      vm_unmap_ram(3) * 85
    }

    On every iteration, it needs newly allocated block and it is put to the
    tail of a free list so finding it consumes large amount of time.
                                                                    Joonsoo Kim

 A: Second patch in current patchset gets rid of extra search in a free list,
    so new block will be immediately occupied..

    Also, the scenario above is impossible, cause vm_map_ram allocates virtual
    range in orders, i.e. 2^n.  I.e. passing 3 to vm_map_ram you will allocate
    4 slots in a block and 256 slots (capacity of a block) of course dividable
    on 4, so block will be completely occupied.

    But there is a worst case which we can achieve: each free block has a hole
    equal to order size.

    The maximum size of allocation is 64 pages for 64-bit system
    (if you try to map more, original alloc_vmap_area will be called).

    So the maximum order is 6.  That means that worst case, before allocator
    makes a decision to allocate a new block, is to iterate 7 blocks:

    HEAD
    1st block - has 1  page slot  free (order 0)
    2nd block - has 2  page slots free (order 1)
    3rd block - has 4  page slots free (order 2)
    4th block - has 8  page slots free (order 3)
    5th block - has 16 page slots free (order 4)
    6th block - has 32 page slots free (order 5)
    7th block - has 64 page slots free (order 6)
    TAIL

    So the worst scenario on 64-bit system is that each CPU queue can have 7
    blocks in a free list.

    This can happen only and only if you allocate blocks increasing the order.
    (as I did in the function written in the comment of the first patch)
    This is weird and rare case, but still it is possible.  Afterwards you will
    get 7 blocks in a list.

    All further requests should be placed in a newly allocated block or some
    free slots should be found in a free list.
    Seems it does not look dramatically awful.

This patch (of 3):

If suitable block can't be found, new block is allocated and put into a
head of a free list, so on next iteration this new block will be found
first.

That's bad, because old blocks in a free list will not get a chance to be
fully used, thus fragmentation will grow.

Let's consider this simple example:

 #1 We have one block in a free list which is partially used, and where only
    one page is free:

    HEAD |xxxxxxxxx-| TAIL
                   ^
                   free space for 1 page, order 0

 #2 New allocation request of order 1 (2 pages) comes, new block is allocated
    since we do not have free space to complete this request. New block is put
    into a head of a free list:

    HEAD |----------|xxxxxxxxx-| TAIL

 #3 Two pages were occupied in a new found block:

    HEAD |xx--------|xxxxxxxxx-| TAIL
          ^
          two pages mapped here

 #4 New allocation request of order 0 (1 page) comes.  Block, which was created
    on #2 step, is located at the beginning of a free list, so it will be found
    first:

  HEAD |xxX-------|xxxxxxxxx-| TAIL
          ^                 ^
          page mapped here, but better to use this hole

It is obvious, that it is better to complete request of #4 step using the
old block, where free space is left, because in other case fragmentation
will be highly increased.

But fragmentation is not only the case.  The worst thing is that I can
easily create scenario, when the whole vmalloc space is exhausted by
blocks, which are not used, but already dirty and have several free pages.

Let's consider this function which execution should be pinned to one CPU:

static void exhaust_virtual_space(struct page *pages[16], int iters)
{
        /* Firstly we have to map a big chunk, e.g. 16 pages.
         * Then we have to occupy the remaining space with smaller
         * chunks, i.e. 8 pages. At the end small hole should remain.
         * So at the end of our allocation sequence block looks like
         * this:
         *                XX  big chunk
         * |XXxxxxxxx-|    x  small chunk
         *                 -  hole, which is enough for a small chunk,
         *                    but is not enough for a big chunk
         */
        while (iters--) {
                int i;
                void *vaddr;

                /* Map/unmap big chunk */
                vaddr = vm_map_ram(pages, 16, -1, PAGE_KERNEL);
                vm_unmap_ram(vaddr, 16);

                /* Map/unmap small chunks.
                 *
                 * -1 for hole, which should be left at the end of each block
                 * to keep it partially used, with some free space available */
                for (i = 0; i < (VMAP_BBMAP_BITS - 16) / 8 - 1; i++) {
                        vaddr = vm_map_ram(pages, 8, -1, PAGE_KERNEL);
                        vm_unmap_ram(vaddr, 8);
                }
        }
}

On every iteration new block (1MB of vm area in my case) will be
allocated and then will be occupied, without attempt to resolve small
allocation request using previously allocated blocks in a free list.

In case of random allocation (size should be randomly taken from the
range [1..64] in 64-bit case or [1..32] in 32-bit case) situation is the
same: new blocks continue to appear if maximum possible allocation size
(32 or 64) passed to the allocator, because all remaining blocks in a
free list do not have enough free space to complete this allocation
request.

In summary if new blocks are put into the head of a free list eventually
virtual space will be exhausted.

In current patch I simply put newly allocated block to the tail of a
free list, thus reduce fragmentation, giving a chance to resolve
allocation request using older blocks with possible holes left.

Signed-off-by: Roman Pen <r.peniaev@gmail.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>

1 files changed, 1 insertions, 1 deletions

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index a5bbdd3b5d67..84feb5249b12 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -842,7 +842,7 @@ static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
 
 	vbq = &get_cpu_var(vmap_block_queue);
 	spin_lock(&vbq->lock);
-	list_add_rcu(&vb->free_list, &vbq->free);
+	list_add_tail_rcu(&vb->free_list, &vbq->free);
 	spin_unlock(&vbq->lock);
 	put_cpu_var(vmap_block_queue);