Re: [PATCH v17 4/6] virtio-balloon: VIRTIO_BALLOON_F_SG

[Wei Wang]

On 11/03/2017 07:25 PM, Tetsuo Handa wrote:
> Wei Wang wrote:
> > @@ -164,6 +284,8 @@ static unsigned fill_balloon(struct virtio_balloon *vb, size_t num)
> >   			break;
> >   		}
> >   
> > +		if (use_sg && xb_set_page(vb, page, &pfn_min, &pfn_max) < 0)
> Isn't this leaking "page" ?


Right, thanks, will add __free_page(page) here.

> > @@ -184,8 +307,12 @@ static unsigned fill_balloon(struct virtio_balloon *vb, size_t num)
> >   
> >   	num_allocated_pages = vb->num_pfns;
> >   	/* Did we get any? */
> > -	if (vb->num_pfns != 0)
> > -		tell_host(vb, vb->inflate_vq);
> > +	if (vb->num_pfns) {
> > +		if (use_sg)
> > +			tell_host_sgs(vb, vb->inflate_vq, pfn_min, pfn_max);
> Please describe why tell_host_sgs() can work without __GFP_DIRECT_RECLAIM allocation,
> for tell_host_sgs() is called with vb->balloon_lock mutex held.

Essentially, 
tell_host_sgs()-->send_balloon_page_sg()-->add_one_sg()-->virtqueue_add_inbuf( 
, , num=1 ,,GFP_KERNEL)
won't need any memory allocation, because we always add one sg (i.e. 
num=1) each time. That memory
allocation option is only used when multiple sgs are added (i.e. num > 
1) and the implementation inside virtqueue_add_inbuf
need allocation of indirect descriptor table.

We could also add some comments above the function to explain a little 
about this if necessary.

> > @@ -223,7 +353,13 @@ static unsigned leak_balloon(struct virtio_balloon *vb, size_t num)
> >   		page = balloon_page_dequeue(vb_dev_info);
> >   		if (!page)
> >   			break;
> > -		set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
> > +		if (use_sg) {
> > +			if (xb_set_page(vb, page, &pfn_min, &pfn_max) < 0)
> Isn't this leaking "page" ?

Yes, will make it:

    if (xb_set_page(vb, page, &pfn_min, &pfn_max) < 0) {
        balloon_page_enqueue(..., page);
        break;
    }

> If this is inside vb->balloon_lock mutex (isn't this?), xb_set_page() must not
> use __GFP_DIRECT_RECLAIM allocation, for leak_balloon_sg_oom() will be blocked
> on vb->balloon_lock mutex.

OK. Since the preload() doesn't need too much memory (< 4K in total), 
how about GFP_NOWAIT here?


Best,
Wei