Re: [Qemu-devel] [PATCH v9 1/9] memory: add section range info for IOMMU notifier

[David Gibson]

On Mon, Apr 10, 2017 at 03:09:50PM +0800, Peter Xu wrote:
> On Mon, Apr 10, 2017 at 02:39:22PM +1000, David Gibson wrote:
> > On Fri, Apr 07, 2017 at 06:59:07PM +0800, Peter Xu wrote:
> > > In this patch, IOMMUNotifier.{start|end} are introduced to store section
> > > information for a specific notifier. When notification occurs, we not
> > > only check the notification type (MAP|UNMAP), but also check whether the
> > > notified iova range overlaps with the range of specific IOMMU notifier,
> > > and skip those notifiers if not in the listened range.
> > > 
> > > When removing an region, we need to make sure we removed the correct
> > > VFIOGuestIOMMU by checking the IOMMUNotifier.start address as well.
> > > 
> > > This patch is solving the problem that vfio-pci devices receive
> > > duplicated UNMAP notification on x86 platform when vIOMMU is there. The
> > > issue is that x86 IOMMU has a (0, 2^64-1) IOMMU region, which is
> > > splitted by the (0xfee00000, 0xfeefffff) IRQ region. AFAIK
> > > this (splitted IOMMU region) is only happening on x86.
> > > 
> > > This patch also helps vhost to leverage the new interface as well, so
> > > that vhost won't get duplicated cache flushes. In that sense, it's an
> > > slight performance improvement.
> > > 
> > > Suggested-by: David Gibson <address@hidden>
> > > Reviewed-by: Eric Auger <address@hidden>
> > > Reviewed-by: Michael S. Tsirkin <address@hidden>
> > > Acked-by: Alex Williamson <address@hidden>
> > > Signed-off-by: Peter Xu <address@hidden>
> > > ---
> > >  hw/vfio/common.c      | 12 +++++++++---
> > >  hw/virtio/vhost.c     | 10 ++++++++--
> > >  include/exec/memory.h | 19 ++++++++++++++++++-
> > >  memory.c              |  9 +++++++++
> > >  4 files changed, 44 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/hw/vfio/common.c b/hw/vfio/common.c
> > > index f3ba9b9..6b33b9f 100644
> > > --- a/hw/vfio/common.c
> > > +++ b/hw/vfio/common.c
> > > @@ -478,8 +478,13 @@ static void vfio_listener_region_add(MemoryListener 
> > > *listener,
> > >          giommu->iommu_offset = section->offset_within_address_space -
> > >                                 section->offset_within_region;
> > >          giommu->container = container;
> > > -        giommu->n.notify = vfio_iommu_map_notify;
> > > -        giommu->n.notifier_flags = IOMMU_NOTIFIER_ALL;
> > > +        llend = int128_add(int128_make64(section->offset_within_region),
> > > +                           section->size);
> > > +        llend = int128_sub(llend, int128_one());
> > > +        iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
> > > +                            IOMMU_NOTIFIER_ALL,
> > > +                            section->offset_within_region,
> > > +                            int128_get64(llend));
> > 
> > Seems to me it would make sense to put the fiddling around to convert
> > the MemoryRegionSection into the necessary values would make sense to
> > go inside iommu_notifier_init().
> 
> But will we always get one MemoryRegionSection if we are not in any of
> the region_{add|del} callback? E.g., what if we want to init an IOMMU
> notifier that covers just the whole IOMMU region range?

I suppose so.  It's just the only likely users of the interface I can
see will be always doing this from region_{add,del}.

> Considering above, I would still slightly prefer current interface.

Ok, well my opinion on the matter isn't terribly strong.

> 
> > 
> > >          QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
> > >  
> > >          memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
> > > @@ -550,7 +555,8 @@ static void vfio_listener_region_del(MemoryListener 
> > > *listener,
> > >          VFIOGuestIOMMU *giommu;
> > >  
> > >          QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
> > > -            if (giommu->iommu == section->mr) {
> > > +            if (giommu->iommu == section->mr &&
> > > +                giommu->n.start == section->offset_within_region) {
> > 
> > This test should be sufficient, but I'd be a bit more comfortable if
> > there was a test and assert() that the end matches as well.  I also
> > wonder if remove-matching-notifier helper would be useful here.
> > Although vhost doesn't appear to ever remove, so maybe it's premature.
> 
> Oh... vhost does remove it, but I just forgot to touch it up :( ...
> Thanks for pointing out.
> 
> Marcel, if this is the only comment, would you mind squash below
> change into current patch? Thanks,
> 
> ----8<----
> 
> diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
> index 185b95b..0001e60 100644
> --- a/hw/virtio/vhost.c
> +++ b/hw/virtio/vhost.c
> @@ -771,7 +771,8 @@ static void vhost_iommu_region_del(MemoryListener 
> *listener,
>      }
>  
>      QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
> -        if (iommu->mr == section->mr) {
> +        if (iommu->mr == section->mr &&
> +            iommu->n.start == section->offset_within_region) {
>              memory_region_unregister_iommu_notifier(iommu->mr,
>                                                      &iommu->n);
>              QLIST_REMOVE(iommu, iommu_next);
> 
> ---->8----
> 
> > 
> > >                  memory_region_unregister_iommu_notifier(giommu->iommu,
> > >                                                          &giommu->n);
> > >                  QLIST_REMOVE(giommu, giommu_next);
> > > diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
> > > index 613494d..185b95b 100644
> > > --- a/hw/virtio/vhost.c
> > > +++ b/hw/virtio/vhost.c
> > > @@ -736,14 +736,20 @@ static void vhost_iommu_region_add(MemoryListener 
> > > *listener,
> > >      struct vhost_dev *dev = container_of(listener, struct vhost_dev,
> > >                                           iommu_listener);
> > >      struct vhost_iommu *iommu;
> > > +    Int128 end;
> > >  
> > >      if (!memory_region_is_iommu(section->mr)) {
> > >          return;
> > >      }
> > >  
> > >      iommu = g_malloc0(sizeof(*iommu));
> > > -    iommu->n.notify = vhost_iommu_unmap_notify;
> > > -    iommu->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
> > > +    end = int128_add(int128_make64(section->offset_within_region),
> > > +                     section->size);
> > > +    end = int128_sub(end, int128_one());
> > > +    iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
> > > +                        IOMMU_NOTIFIER_UNMAP,
> > > +                        section->offset_within_region,
> > > +                        int128_get64(end));
> > >      iommu->mr = section->mr;
> > >      iommu->iommu_offset = section->offset_within_address_space -
> > >                            section->offset_within_region;
> > > diff --git a/include/exec/memory.h b/include/exec/memory.h
> > > index f20b191..0840c89 100644
> > > --- a/include/exec/memory.h
> > > +++ b/include/exec/memory.h
> > > @@ -77,13 +77,30 @@ typedef enum {
> > >  
> > >  #define IOMMU_NOTIFIER_ALL (IOMMU_NOTIFIER_MAP | IOMMU_NOTIFIER_UNMAP)
> > >  
> > > +struct IOMMUNotifier;
> > > +typedef void (*IOMMUNotify)(struct IOMMUNotifier *notifier,
> > > +                            IOMMUTLBEntry *data);
> > > +
> > >  struct IOMMUNotifier {
> > > -    void (*notify)(struct IOMMUNotifier *notifier, IOMMUTLBEntry *data);
> > > +    IOMMUNotify notify;
> > >      IOMMUNotifierFlag notifier_flags;
> > > +    /* Notify for address space range start <= addr <= end */
> > > +    hwaddr start;
> > > +    hwaddr end;
> > >      QLIST_ENTRY(IOMMUNotifier) node;
> > >  };
> > >  typedef struct IOMMUNotifier IOMMUNotifier;
> > >  
> > > +static inline void iommu_notifier_init(IOMMUNotifier *n, IOMMUNotify fn,
> > > +                                       IOMMUNotifierFlag flags,
> > > +                                       hwaddr start, hwaddr end)
> > > +{
> > > +    n->notify = fn;
> > > +    n->notifier_flags = flags;
> > > +    n->start = start;
> > > +    n->end = end;
> > > +}
> > > +
> > >  /* New-style MMIO accessors can indicate that the transaction failed.
> > >   * A zero (MEMTX_OK) response means success; anything else is a failure
> > >   * of some kind. The memory subsystem will bitwise-OR together results
> > > diff --git a/memory.c b/memory.c
> > > index 4c95aaf..75ac595 100644
> > > --- a/memory.c
> > > +++ b/memory.c
> > > @@ -1606,6 +1606,7 @@ void 
> > > memory_region_register_iommu_notifier(MemoryRegion *mr,
> > >  
> > >      /* We need to register for at least one bitfield */
> > >      assert(n->notifier_flags != IOMMU_NOTIFIER_NONE);
> > > +    assert(n->start <= n->end);
> > >      QLIST_INSERT_HEAD(&mr->iommu_notify, n, node);
> > >      memory_region_update_iommu_notify_flags(mr);
> > >  }
> > > @@ -1667,6 +1668,14 @@ void memory_region_notify_iommu(MemoryRegion *mr,
> > >      }
> > >  
> > >      QLIST_FOREACH(iommu_notifier, &mr->iommu_notify, node) {
> > > +        /*
> > > +         * Skip the notification if the notification does not overlap
> > > +         * with registered range.
> > > +         */
> > > +        if (iommu_notifier->start > entry.iova + entry.addr_mask + 1 ||
> > > +            iommu_notifier->end < entry.iova) {
> > > +            continue;
> > > +        }
> > >          if (iommu_notifier->notifier_flags & request_flags) {
> > >              iommu_notifier->notify(iommu_notifier, &entry);
> > >          }
> > 
> 
> -- peterx
> 

-- 
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

signature.asc
Description: PGP signature