Re: [Qemu-devel] [PATCH RFC kernel] balloon: speed up inflating/deflating process

[Cornelia Huck]

On Fri, 20 May 2016 17:59:46 +0800
Liang Li <address@hidden> wrote:

> The implementation of the current virtio-balloon is not very efficient,
> Bellow is test result of time spends on inflating the balloon to 3GB of
> a 4GB idle guest:
> 
> a. allocating pages (6.5%, 103ms)
> b. sending PFNs to host (68.3%, 787ms)
> c. address translation (6.1%, 96ms)
> d. madvise (19%, 300ms)
> 
> It takes about 1577ms for the whole inflating process to complete. The
> test shows that the bottle neck is the stage b and stage d.
> 
> If using a bitmap to send the page info instead of the PFNs, we can
> reduce the overhead spends on stage b quite a lot. Furthermore, it's
> possible to do the address translation and do the madvise with a bulk
> of pages, instead of the current page per page way, so the overhead of
> stage c and stage d can also be reduced a lot.
> 
> This patch is the kernel side implementation which is intended to speed
> up the inflating & deflating process by adding a new feature to the
> virtio-balloon device. And now, inflating the balloon to 3GB of a 4GB
> idle guest only takes 175ms, it's about 9 times as fast as before.
> 
> TODO: optimize stage a by allocating/freeing a chunk of pages instead
> of a single page at a time.

Not commenting on the approach, but...

> 
> Signed-off-by: Liang Li <address@hidden>
> ---
>  drivers/virtio/virtio_balloon.c     | 199 
> ++++++++++++++++++++++++++++++++++--
>  include/uapi/linux/virtio_balloon.h |   1 +
>  mm/page_alloc.c                     |   6 ++
>  3 files changed, 198 insertions(+), 8 deletions(-)
> 

>  static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq)
>  {
> -     struct scatterlist sg;
>       unsigned int len;
> 
> -     sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns);
> +     if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_PAGE_BITMAP)) {
> +             u32 page_shift = PAGE_SHIFT;
> +             unsigned long start_pfn, end_pfn, flags = 0, bmap_len;
> +             struct scatterlist sg[5];
> +
> +             start_pfn = rounddown(vb->start_pfn, BITS_PER_LONG);
> +             end_pfn = roundup(vb->end_pfn, BITS_PER_LONG);
> +             bmap_len = (end_pfn - start_pfn) / BITS_PER_LONG * sizeof(long);
> +
> +             sg_init_table(sg, 5);
> +             sg_set_buf(&sg[0], &flags, sizeof(flags));
> +             sg_set_buf(&sg[1], &start_pfn, sizeof(start_pfn));
> +             sg_set_buf(&sg[2], &page_shift, sizeof(page_shift));
> +             sg_set_buf(&sg[3], &bmap_len, sizeof(bmap_len));
> +             sg_set_buf(&sg[4], vb->page_bitmap +
> +                              (start_pfn / BITS_PER_LONG), bmap_len);
> +             virtqueue_add_outbuf(vq, sg, 5, vb, GFP_KERNEL);
> +

...you need to take care of the endianness of the data you put on the
queue, otherwise virtio-1 on big endian won't work. (There's just been
a patch for that problem.)