Re: [PATCH v2 2/2] mm: speed up mremap by 500x on large regions

[Anton Ivanov <anton.ivanov@kot-begemot.co.uk>]

On 12/10/2018 22:40, Kirill A. Shutemov wrote:
> On Fri, Oct 12, 2018 at 05:42:24PM +0100, Anton Ivanov wrote:
>> On 10/12/18 3:48 PM, Anton Ivanov wrote:
>>> On 12/10/2018 15:37, Kirill A. Shutemov wrote:
>>>> On Fri, Oct 12, 2018 at 03:09:49PM +0100, Anton Ivanov wrote:
>>>>> On 10/12/18 2:37 AM, Joel Fernandes (Google) wrote:
>>>>>> Android needs to mremap large regions of memory during
>>>>>> memory management
>>>>>> related operations. The mremap system call can be really
>>>>>> slow if THP is
>>>>>> not enabled. The bottleneck is move_page_tables, which is copying each
>>>>>> pte at a time, and can be really slow across a large map.
>>>>>> Turning on THP
>>>>>> may not be a viable option, and is not for us. This patch
>>>>>> speeds up the
>>>>>> performance for non-THP system by copying at the PMD level
>>>>>> when possible.
>>>>>>
>>>>>> The speed up is three orders of magnitude. On a 1GB mremap, the mremap
>>>>>> completion times drops from 160-250 millesconds to 380-400
>>>>>> microseconds.
>>>>>>
>>>>>> Before:
>>>>>> Total mremap time for 1GB data: 242321014 nanoseconds.
>>>>>> Total mremap time for 1GB data: 196842467 nanoseconds.
>>>>>> Total mremap time for 1GB data: 167051162 nanoseconds.
>>>>>>
>>>>>> After:
>>>>>> Total mremap time for 1GB data: 385781 nanoseconds.
>>>>>> Total mremap time for 1GB data: 388959 nanoseconds.
>>>>>> Total mremap time for 1GB data: 402813 nanoseconds.
>>>>>>
>>>>>> Incase THP is enabled, the optimization is skipped. I also flush the
>>>>>> tlb every time we do this optimization since I couldn't find a way to
>>>>>> determine if the low-level PTEs are dirty. It is seen that the cost of
>>>>>> doing so is not much compared the improvement, on both
>>>>>> x86-64 and arm64.
>>>>>>
>>>>>> Cc: minchan@kernel.org
>>>>>> Cc: pantin@google.com
>>>>>> Cc: hughd@google.com
>>>>>> Cc: lokeshgidra@google.com
>>>>>> Cc: dancol@google.com
>>>>>> Cc: mhocko@kernel.org
>>>>>> Cc: kirill@shutemov.name
>>>>>> Cc: akpm@linux-foundation.org
>>>>>> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
>>>>>> ---
>>>>>>  A A  mm/mremap.c | 62
>>>>>> +++++++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>>  A A  1 file changed, 62 insertions(+)
>>>>>>
>>>>>> diff --git a/mm/mremap.c b/mm/mremap.c
>>>>>> index 9e68a02a52b1..d82c485822ef 100644
>>>>>> --- a/mm/mremap.c
>>>>>> +++ b/mm/mremap.c
>>>>>> @@ -191,6 +191,54 @@ static void move_ptes(struct
>>>>>> vm_area_struct *vma, pmd_t *old_pmd,
>>>>>>  A A A A A A A A A A  drop_rmap_locks(vma);
>>>>>>  A A  }
>>>>>> +static bool move_normal_pmd(struct vm_area_struct *vma,
>>>>>> unsigned long old_addr,
>>>>>> +A A A A A A A A A  unsigned long new_addr, unsigned long old_end,
>>>>>> +A A A A A A A A A  pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
>>>>>> +{
>>>>>> +A A A  spinlock_t *old_ptl, *new_ptl;
>>>>>> +A A A  struct mm_struct *mm = vma->vm_mm;
>>>>>> +
>>>>>> +A A A  if ((old_addr & ~PMD_MASK) || (new_addr & ~PMD_MASK)
>>>>>> +A A A A A A A  || old_end - old_addr < PMD_SIZE)
>>>>>> +A A A A A A A  return false;
>>>>>> +
>>>>>> +A A A  /*
>>>>>> +A A A A  * The destination pmd shouldn't be established, free_pgtables()
>>>>>> +A A A A  * should have release it.
>>>>>> +A A A A  */
>>>>>> +A A A  if (WARN_ON(!pmd_none(*new_pmd)))
>>>>>> +A A A A A A A  return false;
>>>>>> +
>>>>>> +A A A  /*
>>>>>> +A A A A  * We don't have to worry about the ordering of src and dst
>>>>>> +A A A A  * ptlocks because exclusive mmap_sem prevents deadlock.
>>>>>> +A A A A  */
>>>>>> +A A A  old_ptl = pmd_lock(vma->vm_mm, old_pmd);
>>>>>> +A A A  if (old_ptl) {
>>>>>> +A A A A A A A  pmd_t pmd;
>>>>>> +
>>>>>> +A A A A A A A  new_ptl = pmd_lockptr(mm, new_pmd);
>>>>>> +A A A A A A A  if (new_ptl != old_ptl)
>>>>>> +A A A A A A A A A A A  spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
>>>>>> +
>>>>>> +A A A A A A A  /* Clear the pmd */
>>>>>> +A A A A A A A  pmd = *old_pmd;
>>>>>> +A A A A A A A  pmd_clear(old_pmd);
>>>>>> +
>>>>>> +A A A A A A A  VM_BUG_ON(!pmd_none(*new_pmd));
>>>>>> +
>>>>>> +A A A A A A A  /* Set the new pmd */
>>>>>> +A A A A A A A  set_pmd_at(mm, new_addr, new_pmd, pmd);
>>>>> UML does not have set_pmd_at at all
>>>> Every architecture does. :)
>>> I tried to build it patching vs 4.19-rc before I made this statement and
>>> ran into that.
>>>
>>> Presently it does not.
>>>
>>> https://elixir.bootlin.com/linux/v4.19-rc7/ident/set_pmd_at - UML is not
>>> on the list.
>> Once this problem as well as the omissions in the include changes for UML in
>> patch one have been fixed it appears to be working.
>>
>> What it needs is attached.
> Well, the optization is only suitable for arch that has 3 or more levels
> of page tables. Otherwise it will not have [non-folded] pmd.
>
> And in this case arch/um already should have set_pmd_at(), see
> 3_LEVEL_PGTABLES.
>
> To port on 2-level paging, it has to be handled on pgd level. It
> complicates the code and will not bring much value.
>
UML has 3 level page tables on 64 bit.

A.