[PATCH 00/14] Reduce preallocations for maple tree

[Yin Fengwei]

Hi Peng,

On 6/5/23 11:28, Peng Zhang wrote:
> 
> 
> 在 2023/6/2 16:10, Yin, Fengwei 写道:
>> Hi Liam,
>>
>> On 6/1/2023 10:15 AM, Liam R. Howlett wrote:
>>> Initial work on preallocations showed no regression in performance
>>> during testing, but recently some users (both on [1] and off [android]
>>> list) have reported that preallocating the worst-case number of nodes
>>> has caused some slow down.  This patch set addresses the number of
>>> allocations in a few ways.
>>>
>>> During munmap() most munmap() operations will remove a single VMA, so
>>> leverage the fact that the maple tree can place a single pointer at
>>> range 0 - 0 without allocating.  This is done by changing the index in
>>> the 'sidetree'.
>>>
>>> Re-introduce the entry argument to mas_preallocate() so that a more
>>> intelligent guess of the node count can be made.
>>>
>>> Patches are in the following order:
>>> 0001-0002: Testing framework for benchmarking some operations
>>> 0003-0004: Reduction of maple node allocation in sidetree
>>> 0005:      Small cleanup of do_vmi_align_munmap()
>>> 0006-0013: mas_preallocate() calculation change
>>> 0014:      Change the vma iterator order
>> I did run The AIM:page_test on an IceLake 48C/96T + 192G RAM platform with
>> this patchset.
>>
>> The result has a little bit improvement:
>> Base (next-20230602):
>>    503880
>> Base with this patchset:
>>    519501
>>
>> But they are far from the none-regression result (commit 7be1c1a3c7b1):
>>    718080
>>
>>
>> Some other information I collected:
>> With Base, the mas_alloc_nodes are always hit with request: 7.
>> With this patchset, the request are 1 or 5.
>>
>> I suppose this is the reason for improvement from 503880 to 519501.
>>
>> With commit 7be1c1a3c7b1, mas_store_gfp() in do_brk_flags never triggered
>> mas_alloc_nodes() call. Thanks.
> Hi Fengwei,
> 
> I think it may be related to the inaccurate number of nodes allocated
> in the pre-allocation. I slightly modified the pre-allocation in this
> patchset, but I don't know if it works. It would be great if you could
> help test it, and help pinpoint the cause. Below is the diff, which can
> be applied based on this pachset.
I tried the patch, it could eliminate the call of mas_alloc_nodes() during
the test. But the result of benchmark got a little bit improvement:
  529040

But it's still much less than none-regression result. I will also double
confirm the none-regression result.


Regards
Yin, Fengwei

> 
> Thanks,
> Peng
> 
> diff --git a/lib/maple_tree.c b/lib/maple_tree.c
> index 5ea211c3f186..e67bf2744384 100644
> --- a/lib/maple_tree.c
> +++ b/lib/maple_tree.c
> @@ -5575,9 +5575,11 @@ int mas_preallocate(struct ma_state *mas, void *entry, gfp_t gfp)
>          goto ask_now;
>      }
> 
> -    /* New root needs a singe node */
> -    if (unlikely(mte_is_root(mas->node)))
> -        goto ask_now;
> +    if ((node_size == wr_mas.node_end + 1 &&
> +         mas->offset == wr_mas.node_end) ||
> +        (node_size == wr_mas.node_end &&
> +         wr_mas.offset_end - mas->offset == 1))
> +        return 0;
> 
>      /* Potential spanning rebalance collapsing a node, use worst-case */
>      if (node_size  - 1 <= mt_min_slots[wr_mas.type])
> @@ -5590,7 +5592,6 @@ int mas_preallocate(struct ma_state *mas, void *entry, gfp_t gfp)
>      if (likely(!mas_is_err(mas)))
>          return 0;
> 
> -    mas_set_alloc_req(mas, 0);
>      ret = xa_err(mas->node);
>      mas_reset(mas);
>      mas_destroy(mas);
> 
> 
>>
>>
>> Regards
>> Yin, Fengwei
>>
>>>
>>> [1] https://lore.kernel.org/linux-mm/202305061457.ac15990c-yujie.liu@intel.com/
>>>
>>> Liam R. Howlett (14):
>>>    maple_tree: Add benchmarking for mas_for_each
>>>    maple_tree: Add benchmarking for mas_prev()
>>>    mm: Move unmap_vmas() declaration to internal header
>>>    mm: Change do_vmi_align_munmap() side tree index
>>>    mm: Remove prev check from do_vmi_align_munmap()
>>>    maple_tree: Introduce __mas_set_range()
>>>    mm: Remove re-walk from mmap_region()
>>>    maple_tree: Re-introduce entry to mas_preallocate() arguments
>>>    mm: Use vma_iter_clear_gfp() in nommu
>>>    mm: Set up vma iterator for vma_iter_prealloc() calls
>>>    maple_tree: Move mas_wr_end_piv() below mas_wr_extend_null()
>>>    maple_tree: Update mas_preallocate() testing
>>>    maple_tree: Refine mas_preallocate() node calculations
>>>    mm/mmap: Change vma iteration order in do_vmi_align_munmap()
>>>
>>>   fs/exec.c                        |   1 +
>>>   include/linux/maple_tree.h       |  23 ++++-
>>>   include/linux/mm.h               |   4 -
>>>   lib/maple_tree.c                 |  78 ++++++++++----
>>>   lib/test_maple_tree.c            |  74 +++++++++++++
>>>   mm/internal.h                    |  40 ++++++--
>>>   mm/memory.c                      |  16 ++-
>>>   mm/mmap.c                        | 171 ++++++++++++++++---------------
>>>   mm/nommu.c                       |  45 ++++----
>>>   tools/testing/radix-tree/maple.c |  59 ++++++-----
>>>   10 files changed, 331 insertions(+), 180 deletions(-)
>>>