On Mon, Mar 22, 2021 at 11:36:19PM +0100, Uladzislau Rezki wrote:
> On Mon, Mar 22, 2021 at 07:38:20PM +0000, Matthew Wilcox (Oracle) wrote:
> > If we're trying to allocate 4MB of memory, the table will be 8KiB in size
> > (1024 pointers * 8 bytes per pointer), which can usually be satisfied
> > by a kmalloc (which is significantly faster). Instead of changing this
> > open-coded implementation, just use kvmalloc().
> >
> > Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > ---
> > mm/vmalloc.c | 7 +------
> > 1 file changed, 1 insertion(+), 6 deletions(-)
> >
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 96444d64129a..32b640a84250 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -2802,13 +2802,8 @@ static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
> > gfp_mask |= __GFP_HIGHMEM;
> >
> > /* Please note that the recursion is strictly bounded. */
> > - if (array_size > PAGE_SIZE) {
> > - pages = __vmalloc_node(array_size, 1, nested_gfp, node,
> > + pages = kvmalloc_node_caller(array_size, nested_gfp, node,
> > area->caller);
> > - } else {
> > - pages = kmalloc_node(array_size, nested_gfp, node);
> > - }
> > -
> > if (!pages) {
> > free_vm_area(area);
> > return NULL;
> > --
> > 2.30.2
> Makes sense to me. Though i expected a bigger difference:
>
> # patch
> single CPU, 4MB allocation, loops: 1000000 avg: 85293854 usec
>
> # default
> single CPU, 4MB allocation, loops: 1000000 avg: 89275857 usec
Well, 4.5% isn't something to leave on the table ... but yeah, I was
expecting more in the 10-20% range. It may be more significant if
there's contention on the spinlocks (like if this crazy ksmbd is calling
vmalloc(4MB) on multiple nodes simultaneously).
I suspect the vast majority of the time is spent calling alloc_pages_node()
1024 times. Have you looked at Mel's patch to do ... well, exactly what
vmalloc() wants?
https://lore.kernel.org/linux-mm/20210322091845.16437-1-mgorman@xxxxxxxxxxxxxxxxxxx/
> One question. Should we care much about fragmentation? I mean
> with the patch, allocations > 2MB will do request to SLAB bigger
> then PAGE_SIZE.
We're pretty good about allocating memory in larger chunks these days.
Looking at my laptop's slabinfo,
kmalloc-8k 219 232 8192 4 8 : tunables 0 0 0 : sla
bdata 58 58 0
That's using 8 pages per slab, so that's order-3 allocations. There's a
few more of those:
$ sudo grep '8 :' /proc/slabinfo |wc
42 672 4508
so I have confidence that kvmalloc() will manage to use kmalloc up to 16MB
vmalloc allocations, and after that it'll tend to fall back to vmalloc.
