Re: [PATCH v1] mm/vmalloc: fix exact allocations with an alignment > 1

[David Hildenbrand]

On 29.09.21 16:49, Uladzislau Rezki wrote:
> On Wed, Sep 29, 2021 at 4:40 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > On 29.09.21 16:30, Uladzislau Rezki wrote:
> > > > So the idea is that once we run into a dead end because we took a left
> > > > subtree, we rollback to the next possible rigth subtree and try again.
> > > > If we run into another dead end, we repeat ... thus, this can now happen
> > > > more than once.
> > > >
> > > > I assume the only implication is that this can now be slower in some
> > > > corner cases with larger alignment, because it might take longer to find
> > > > something suitable. Fair enough.
> > > Yep, your understanding is correct regarding the tree traversal. If no
> > > suitable block
> > > is found in left sub-tree we roll-back and check right one. So it can
> > > be(the scanning)
> > > more than one time.
> > >
> > > I did some performance analyzing using vmalloc test suite to figure
> > > out a performance
> > > loss for allocations with specific alignment. On that syntactic test i
> > > see approx. 30%
> > > of degradation:
> >
> > How realistic is that test case? I assume most alignment we're dealing
> > with is:
> > * 1/PAGE_SIZE
> > * huge page size (for automatic huge page placing)
> Well that is synthetic test. Most of the alignments are 1 or PAGE_SIZE.
> There are users which use internal API where you can specify an alignment
> you want but those are mainly like KASAN, module alloc, etc.
>
> > > 2.225 microseconds vs 1.496 microseconds. That time includes both
> > > vmalloc() and vfree()
> > > calls. I do not consider it as a big degrade, but from the other hand
> > > we can still adjust the
> > > search length for alignments > one page:
> > >
> > > # add it on top of previous proposal and search length instead of size
> > > length = align > PAGE_SIZE ? size + align:size;
> >
> > That will not allow to place huge pages in the case of kasan. And I
> > consider that more important than optimizing a syntactic test :) My 2 cents.
> Could you please to be more specific? I mean how is it connected with huge
> pages mappings? Huge-pages are which have order > 0. Or you mean that
> a special alignments are needed for mapping huge pages?

Let me try to clarify:


KASAN does an exact allocation when onlining a memory block, 
__vmalloc_node_range() will try placing huge pages first, increasing the 
alignment to e.g., "1 << PMD_SHIFT".

If we increase the search length in find_vmap_lowest_match(), that 
search will fail if the exact allocation is surrounded by other 
allocations. In that case, we won't place a huge page although we could 
-- because find_vmap_lowest_match() would be imprecise for alignments > 
PAGE_SIZE.


Memory blocks we online/offline on x86 are at least 128MB. The KASAN 
"overhead" we have to allocate is 1/8 of that -- 16 MB, so essentially 8 
huge pages.

__vmalloc_node_range() will increase the alignment to 2MB to try placing 
huge pages first. find_vmap_lowest_match() will search within the given 
exact 16MB are a 18MB area (size + align), which won't work. So 
__vmalloc_node_range() will fallback to the original PAGE_SIZE alignment 
and shift=PAGE_SHIFT.

__vmalloc_area_node() will set the set_vm_area_page_order effectively to 
0 --  small pages.

Does that make sense or am I missing something?

--
Thanks,

David / dhildenb