On Fri, Oct 12, 2018 at 10:32:34AM +0200, David Hildenbrand wrote:
> On 12/10/2018 05:24, David Gibson wrote:
> > The virtio-balloon always works in units of 4kiB (BALLOON_PAGE_SIZE), but
> > on the host side, we can only actually discard memory in units of the host
> > page size.
> > 
> > At present we handle this very badly: we silently ignore balloon requests
> > that aren't host page aligned, and for requests that are host page aligned
> > we discard the entire host page.  The latter potentially corrupts guest
> > memory if its page size is smaller than the host's.
> > 
> > We could just disable the balloon if the host page size is not 4kiB, but
> > that would break a the special case where host and guest have the same page
> > size, but that's larger than 4kiB.  Thius case currently works by accident:
> > when the guest puts its page into the balloon, it will submit balloon
> > requests for each 4kiB subpage.  Most will be ignored, but the one which
> > happens to be host page aligned will discard the whole lot.
> > 
> > This occurs in practice routinely for POWER KVM systems, since both host
> > and guest typically use 64kiB pages.
> > 
> > To make this safe, without breaking that useful case, we need to
> > accumulate 4kiB balloon requests until we have a whole contiguous host page
> > at which point we can discard it.
> > 
> > We could in principle do that across all guest memory, but it would require
> > a large bitmap to track.  This patch represents a compromise: instead we
> > track ballooned subpages for a single contiguous host page at a time.  This
> > means that if the guest discards all 4kiB chunks of a host page in
> > succession, we will discard it.  In particular that means the balloon will
> > continue to work for the (host page size) == (guest page size) > 4kiB case.
> > 
> > If the guest scatters 4kiB requests across different host pages, we don't
> > discard anything, and issue a warning.  Not ideal, but at least we don't
> > corrupt guest memory as the previous version could.
> > 
> > Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> > ---
> >  hw/virtio/virtio-balloon.c         | 67 +++++++++++++++++++++++++-----
> >  include/hw/virtio/virtio-balloon.h |  3 ++
> >  2 files changed, 60 insertions(+), 10 deletions(-)
> > 
> > diff --git a/hw/virtio/virtio-balloon.c b/hw/virtio/virtio-balloon.c
> > index 4435905c87..39573ef2e3 100644
> > --- a/hw/virtio/virtio-balloon.c
> > +++ b/hw/virtio/virtio-balloon.c
> > @@ -33,33 +33,80 @@
> >  
> >  #define BALLOON_PAGE_SIZE  (1 << VIRTIO_BALLOON_PFN_SHIFT)
> >  
> > +typedef struct PartiallyBalloonedPage {
> > +    RAMBlock *rb;
> > +    ram_addr_t base;
> > +    unsigned long bitmap[];
> 
> BTW, it might be easier to only remember the last inflated page and
> incrementing it when you see the successor.

That would be marginally simpler, but I was preferring not to rely on
the guest always doing things in a particular order.

> 
> initialize last_page to -1ull on realize/reset
> 
> 
> if (QEMU_IS_ALIGNED(addr, PAGE_SIZE)) {
> 	/* start of a new potential page block */
> 	last_page == addr;
> } else if (addr == last_page + BALLOON_PAGE_SIZE) {
> 	/* next successor */
> 	last_page == addr;
> 	if (QEMU_IS_ALIGNED(last_page + BALLOON_PAGE_SIZE, PAGE_SIZE)) {
> 		ramblock_discard()....
> 	}
> } else {
> 	last_page = -1ull;
> }
> 
> 
> > +} PartiallyBalloonedPage;
> > +
> >  static void balloon_inflate_page(VirtIOBalloon *balloon,
> >                                   MemoryRegion *mr, hwaddr offset)
> >  {
> >      void *addr = memory_region_get_ram_ptr(mr) + offset;
> >      RAMBlock *rb;
> >      size_t rb_page_size;
> > -    ram_addr_t ram_offset;
> > +    int subpages;
> > +    ram_addr_t ram_offset, host_page_base;
> >  
> >      /* XXX is there a better way to get to the RAMBlock than via a
> >       * host address? */
> >      rb = qemu_ram_block_from_host(addr, false, &ram_offset);
> >      rb_page_size = qemu_ram_pagesize(rb);
> > +    host_page_base = ram_offset & ~(rb_page_size - 1);
> > +
> > +    if (rb_page_size == BALLOON_PAGE_SIZE) {
> > +        /* Easy case */
> >  
> > -    /* Silently ignore hugepage RAM blocks */
> > -    if (rb_page_size != getpagesize()) {
> > +        ram_block_discard_range(rb, ram_offset, rb_page_size);
> > +        /* We ignore errors from ram_block_discard_range(), because it
> > +         * has already reported them, and failing to discard a balloon
> > +         * page is not fatal */
> >          return;
> >      }
> >  
> > -    /* Silently ignore unaligned requests */
> > -    if (ram_offset & (rb_page_size - 1)) {
> > -        return;
> > +    /* Hard case
> > +     *
> > +     * We've put a piece of a larger host page into the balloon - we
> > +     * need to keep track until we have a whole host page to
> > +     * discard
> > +     */
> > +    subpages = rb_page_size / BALLOON_PAGE_SIZE;
> > +
> > +    if (balloon->pbp
> > +        && (rb != balloon->pbp->rb
> > +            || host_page_base != balloon->pbp->base)) {
> > +        /* We've partially ballooned part of a host page, but now
> > +         * we're trying to balloon part of a different one.  Too hard,
> > +         * give up on the old partial page */
> > +        warn_report("Unable to insert a partial page into virtio-balloon");
> > +        free(balloon->pbp);
> > +        balloon->pbp = NULL;
> >      }
> >  
> > -    ram_block_discard_range(rb, ram_offset, rb_page_size);
> > -    /* We ignore errors from ram_block_discard_range(), because it has
> > -     * already reported them, and failing to discard a balloon page is
> > -     * not fatal */
> > +    if (!balloon->pbp) {
> > +        /* Starting on a new host page */
> > +        size_t bitlen = BITS_TO_LONGS(subpages) * sizeof(unsigned long);
> > +        balloon->pbp = g_malloc0(sizeof(PartiallyBalloonedPage) + bitlen);
> > +        balloon->pbp->rb = rb;
> > +        balloon->pbp->base = host_page_base;
> > +    }
> > +
> > +    bitmap_set(balloon->pbp->bitmap,
> > +               (ram_offset - balloon->pbp->base) / BALLOON_PAGE_SIZE,
> > +               subpages);
> > +
> > +    if (bitmap_full(balloon->pbp->bitmap, subpages)) {
> > +        /* We've accumulated a full host page, we can actually discard
> > +         * it now */
> > +
> > +        ram_block_discard_range(rb, balloon->pbp->base, rb_page_size);
> > +        /* We ignore errors from ram_block_discard_range(), because it
> > +         * has already reported them, and failing to discard a balloon
> > +         * page is not fatal */
> > +
> > +        free(balloon->pbp);
> > +        balloon->pbp = NULL;
> > +    }
> >  }
> >  
> >  static const char *balloon_stat_names[] = {
> > diff --git a/include/hw/virtio/virtio-balloon.h b/include/hw/virtio/virtio-balloon.h
> > index e0df3528c8..99dcd6d105 100644
> > --- a/include/hw/virtio/virtio-balloon.h
> > +++ b/include/hw/virtio/virtio-balloon.h
> > @@ -30,6 +30,8 @@ typedef struct virtio_balloon_stat_modern {
> >         uint64_t val;
> >  } VirtIOBalloonStatModern;
> >  
> > +typedef struct PartiallyBalloonedPage PartiallyBalloonedPage;
> > +
> >  typedef struct VirtIOBalloon {
> >      VirtIODevice parent_obj;
> >      VirtQueue *ivq, *dvq, *svq;
> > @@ -42,6 +44,7 @@ typedef struct VirtIOBalloon {
> >      int64_t stats_last_update;
> >      int64_t stats_poll_interval;
> >      uint32_t host_features;
> > +    PartiallyBalloonedPage *pbp;
> >  } VirtIOBalloon;
> >  
> >  #endif
> > 
> 
> 

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson