[PATCH v4 09/66] mm/mmap: Use the maple tree in find_vma() instead of the rbtree.

[Liam Howlett]

* Vlastimil Babka <vbabka at suse.cz> [211215 08:05]:
> On 12/1/21 15:29, Liam Howlett wrote:
> > From: "Liam R. Howlett" <Liam.Howlett at Oracle.com>
> > 
> > Using the maple tree interface mt_find() will handle the RCU locking and
> > will start searching at the address up to the limit, ULONG_MAX in this
> > case.
> > 
> > Add kernel documentation to this API.
> > 
> > Signed-off-by: Liam R. Howlett <Liam.Howlett at Oracle.com>
> 
> Acked-by: Vlastimil Babka <vbabka at suse.cz>
> 
> Note below:
> 
> > ---
> >  mm/mmap.c | 27 +++++++++------------------
> >  1 file changed, 9 insertions(+), 18 deletions(-)
> > 
> > diff --git a/mm/mmap.c b/mm/mmap.c
> > index de78fc0ce809..6a7502f74190 100644
> > --- a/mm/mmap.c
> > +++ b/mm/mmap.c
> > @@ -2429,10 +2429,16 @@ get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
> >  
> >  EXPORT_SYMBOL(get_unmapped_area);
> >  
> > -/* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
> > +/**
> > + * find_vma() - Find the VMA for a given address, or the next vma.
> > + * @mm: The mm_struct to check
> > + * @addr: The address
> > + *
> > + * Returns: The VMA associated with addr, or the next vma.
> > + * May return %NULL in the case of no vma at addr or above.
> > + */
> >  struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
> >  {
> > -	struct rb_node *rb_node;
> >  	struct vm_area_struct *vma;
> >  
> >  	mmap_assert_locked(mm);
> > @@ -2441,22 +2447,7 @@ struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
> >  	if (likely(vma))
> >  		return vma;
> >  
> > -	rb_node = mm->mm_rb.rb_node;
> > -
> > -	while (rb_node) {
> > -		struct vm_area_struct *tmp;
> > -
> > -		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
> > -
> > -		if (tmp->vm_end > addr) {
> > -			vma = tmp;
> > -			if (tmp->vm_start <= addr)
> > -				break;
> > -			rb_node = rb_node->rb_left;
> > -		} else
> > -			rb_node = rb_node->rb_right;
> > -	}
> > -
> > +	vma = mt_find(&mm->mm_mt, &addr, ULONG_MAX);
> 
> This updates addr to the end of vma->vm_end.
> 
> >  	if (vma)
> >  		vmacache_update(addr, vma);
> 
> And here vmacache is updated with the updated addr, which wasn't done
> before. This AFAIU has two consequences:
> 
> - the original addr will not be cached, so this whole vma lookup will not be
> cached and will always resort to maple tree search. Possibly affecting any
> measurements you made. Especially the vmacache removal later in the series
> might now look like it makes not much of a performance difference - but
> vmcache is effectively dysfunctional.
> 
> - the future lookup of address vma->vm_end will return this vma, although
> the address doesn't belong to it. Wouldn't be surprised if this caused
> infrequent mysterious bugs?
> 
> Both will resolve with vmacache removal later, but better still fix this.

This is certainly worth fixing.  I am surprised that I did not hit a
random bug as you said above, I would expect it to occur as well.
Although, I didn't run each patch individually for long unless I was
tracking down an issue on a rebase.

As for the performance, I will retest the performance for the vmacache
included and excluded but since the overall performance of the patch set
is still valid, I don't expect a change here.

This is a really good catch, thanks.
Liam