Re: [PATCH RFC] virtio-pci: disable vring processing when bus-mastering is disabled

[Michael Roth]

Quoting Michael S. Tsirkin (2019-11-13 04:09:02)
> On Tue, Nov 12, 2019 at 11:43:01PM -0600, Michael Roth wrote:
> > Currently the SLOF firmware for pseries guests will disable/re-enable
> > a PCI device multiple times via IO/MEM/MASTER bits of PCI_COMMAND
> > register after the initial probe/feature negotiation, as it tends to
> > work with a single device at a time at various stages like probing
> > and running block/network bootloaders without doing a full reset
> > in-between.
> > 
> > In QEMU, when PCI_COMMAND_MASTER is disabled we disable the
> > corresponding IOMMU memory region, so DMA accesses (including to vring
> > fields like idx/flags) will no longer undergo the necessary
> > translation. Normally we wouldn't expect this to happen since it would
> > be misbehavior on the driver side to continue driving DMA requests.
> > 
> > However, in the case of pseries, with iommu_platform=on, we trigger the
> > following sequence when tearing down the virtio-blk dataplane ioeventfd
> > in response to the guest unsetting PCI_COMMAND_MASTER:
> > 
> >   #2  0x0000555555922651 in virtqueue_map_desc 
> > (vdev=vdev@entry=0x555556dbcfb0, p_num_sg=p_num_sg@entry=0x7fffe657e1a8, 
> > addr=addr@entry=0x7fffe657e240, iov=iov@entry=0x7fffe6580240, 
> > max_num_sg=max_num_sg@entry=1024, is_write=is_write@entry=false, pa=0, sz=0)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio.c:757
> >   #3  0x0000555555922a89 in virtqueue_pop (vq=vq@entry=0x555556dc8660, 
> > sz=sz@entry=184)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio.c:950
> >   #4  0x00005555558d3eca in virtio_blk_get_request (vq=0x555556dc8660, 
> > s=0x555556dbcfb0)
> >       at /home/mdroth/w/qemu.git/hw/block/virtio-blk.c:255
> >   #5  0x00005555558d3eca in virtio_blk_handle_vq (s=0x555556dbcfb0, 
> > vq=0x555556dc8660)
> >       at /home/mdroth/w/qemu.git/hw/block/virtio-blk.c:776
> >   #6  0x000055555591dd66 in virtio_queue_notify_aio_vq 
> > (vq=vq@entry=0x555556dc8660)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio.c:1550
> >   #7  0x000055555591ecef in virtio_queue_notify_aio_vq (vq=0x555556dc8660)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio.c:1546
> >   #8  0x000055555591ecef in virtio_queue_host_notifier_aio_poll 
> > (opaque=0x555556dc86c8)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio.c:2527
> >   #9  0x0000555555d02164 in run_poll_handlers_once 
> > (ctx=ctx@entry=0x55555688bfc0, timeout=timeout@entry=0x7fffe65844a8)
> >       at /home/mdroth/w/qemu.git/util/aio-posix.c:520
> >   #10 0x0000555555d02d1b in try_poll_mode (timeout=0x7fffe65844a8, 
> > ctx=0x55555688bfc0)
> >       at /home/mdroth/w/qemu.git/util/aio-posix.c:607
> >   #11 0x0000555555d02d1b in aio_poll (ctx=ctx@entry=0x55555688bfc0, 
> > blocking=blocking@entry=true)
> >       at /home/mdroth/w/qemu.git/util/aio-posix.c:639
> >   #12 0x0000555555d0004d in aio_wait_bh_oneshot (ctx=0x55555688bfc0, 
> > cb=cb@entry=0x5555558d5130 <virtio_blk_data_plane_stop_bh>, 
> > opaque=opaque@entry=0x555556de86f0)
> >       at /home/mdroth/w/qemu.git/util/aio-wait.c:71
> >   #13 0x00005555558d59bf in virtio_blk_data_plane_stop (vdev=<optimized 
> > out>)
> >       at /home/mdroth/w/qemu.git/hw/block/dataplane/virtio-blk.c:288
> >   #14 0x0000555555b906a1 in virtio_bus_stop_ioeventfd 
> > (bus=bus@entry=0x555556dbcf38)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio-bus.c:245
> >   #15 0x0000555555b90dbb in virtio_bus_stop_ioeventfd 
> > (bus=bus@entry=0x555556dbcf38)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio-bus.c:237
> >   #16 0x0000555555b92a8e in virtio_pci_stop_ioeventfd (proxy=0x555556db4e40)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio-pci.c:292
> >   #17 0x0000555555b92a8e in virtio_write_config (pci_dev=0x555556db4e40, 
> > address=<optimized out>, val=1048832, len=<optimized out>)
> >       at /home/mdroth/w/qemu.git/hw/virtio/virtio-pci.c:613
> > 
> > I.e. the calling code is only scheduling a one-shot BH for
> > virtio_blk_data_plane_stop_bh, but somehow we end up trying to process
> > an additional virtqueue entry before we get there. This is likely due
> > to the following check in virtio_queue_host_notifier_aio_poll:
> > 
> >   static bool virtio_queue_host_notifier_aio_poll(void *opaque)
> >   {
> >       EventNotifier *n = opaque;
> >       VirtQueue *vq = container_of(n, VirtQueue, host_notifier);
> >       bool progress;
> > 
> >       if (!vq->vring.desc || virtio_queue_empty(vq)) {
> >           return false;
> >       }
> > 
> >       progress = virtio_queue_notify_aio_vq(vq);
> > 
> > namely the call to virtio_queue_empty(). In this case, since no new
> > requests have actually been issued, shadow_avail_idx == last_avail_idx,
> > so we actually try to access the vring via vring_avail_idx() to get
> > the latest non-shadowed idx:
> > 
> >   int virtio_queue_empty(VirtQueue *vq)
> >   {
> >       bool empty;
> >       ...
> > 
> >       if (vq->shadow_avail_idx != vq->last_avail_idx) {
> >           return 0;
> >       }
> > 
> >       rcu_read_lock();
> >       empty = vring_avail_idx(vq) == vq->last_avail_idx;
> >       rcu_read_unlock();
> >       return empty;
> > 
> > but since the IOMMU region has been disabled we get a bogus value (0
> > usually), which causes virtio_queue_empty() to falsely report that
> > there are entries to be processed, which causes errors such as:
> > 
> >   "virtio: zero sized buffers are not allowed"
> > 
> > or
> > 
> >   "virtio-blk missing headers"
> > 
> > and puts the device in an error state.
> > 
> > This patch works around the issue by introducing virtio_set_disabled(),
> > which piggy-backs off the vdev->broken flag we already use to bypass
> > checks like virtio_queue_empty(), and sets/unsets it in response to
> > enabling/disabling bus-mastering.
> > 
> > NOTES:
> > 
> >  - It's possible we could also work around this in SLOF by doing a
> >    full reset instead of relying on PCI_COMMAND to enable/disable, but
> >    in any case the QEMU behavior seems wrong.
> >  - This leaves some other oddities in play, like the fact that
> >    DRIVER_OK also gets unset in response to bus-mastering being
> >    disabled, but not restored (however the device seems to continue
> >    working)
> >  - Similarly, we disable the host notifier via
> >    virtio_bus_stop_ioeventfd(), which seems to move the handling out
> >    of virtio-blk dataplane and back into the main IO thread, and it
> >    ends up staying there till a reset (but otherwise continues working
> >    normally)
> > 
> > Cc: David Gibson <address@hidden>,
> > Cc: Alexey Kardashevskiy <address@hidden>
> > Cc: "Michael S. Tsirkin" <address@hidden>
> > Signed-off-by: Michael Roth <address@hidden>
> > ---
> >  hw/virtio/virtio-pci.c     | 12 ++++++++----
> >  hw/virtio/virtio.c         |  7 ++++++-
> >  include/hw/virtio/virtio.h |  1 +
> >  3 files changed, 15 insertions(+), 5 deletions(-)
> > 
> > diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> > index c6b47a9c73..394d409fb9 100644
> > --- a/hw/virtio/virtio-pci.c
> > +++ b/hw/virtio/virtio-pci.c
> > @@ -608,10 +608,14 @@ static void virtio_write_config(PCIDevice *pci_dev, 
> > uint32_t address,
> >          pcie_cap_flr_write_config(pci_dev, address, val, len);
> >      }
> >  
> > -    if (range_covers_byte(address, len, PCI_COMMAND) &&
> > -        !(pci_dev->config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
> > -        virtio_pci_stop_ioeventfd(proxy);
> > -        virtio_set_status(vdev, vdev->status & ~VIRTIO_CONFIG_S_DRIVER_OK);
> > +    if (range_covers_byte(address, len, PCI_COMMAND)) {
> > +        if (!(pci_dev->config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
> > +            virtio_set_disabled(vdev, true);
> > +            virtio_pci_stop_ioeventfd(proxy);
> > +            virtio_set_status(vdev, vdev->status & 
> > ~VIRTIO_CONFIG_S_DRIVER_OK);
> > +        } else {
> > +            virtio_set_disabled(vdev, false);
> > +        }
> >      }
> >  
> >      if (proxy->config_cap &&
> > diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
> > index 527df03bfd..46580c357d 100644
> > --- a/hw/virtio/virtio.c
> > +++ b/hw/virtio/virtio.c
> > @@ -2575,6 +2575,11 @@ void virtio_device_set_child_bus_name(VirtIODevice 
> > *vdev, char *bus_name)
> >      vdev->bus_name = g_strdup(bus_name);
> >  }
> >  
> > +void virtio_set_disabled(VirtIODevice *vdev, bool disable)
> > +{
> > +    vdev->broken = disable;
> > +}
> > +
> >  void GCC_FMT_ATTR(2, 3) virtio_error(VirtIODevice *vdev, const char *fmt, 
> > ...)
> >  {
> >      va_list ap;
> 
> Hmm. I think just clear and immediate set of bus master while device was
> not doing any DMA actually should be fine and should not require a
> reset.  Now it's true that right now guests reset first thing which will
> clear the broken flag, but I'd say it's wrong to require this specific
> order.
> I think the easiest thing is to add a "disabled" flag.

Agreed. Some comments on that below.

> 
> 
> > @@ -2588,7 +2593,7 @@ void GCC_FMT_ATTR(2, 3) virtio_error(VirtIODevice 
> > *vdev, const char *fmt, ...)
> >          virtio_notify_config(vdev);
> >      }
> >  
> > -    vdev->broken = true;
> > +    virtio_set_disabled(vdev, true);
> >  }
> >  
> >  static void virtio_memory_listener_commit(MemoryListener *listener)
> > diff --git a/include/hw/virtio/virtio.h b/include/hw/virtio/virtio.h
> > index 48e8d04ff6..921945b7e8 100644
> > --- a/include/hw/virtio/virtio.h
> > +++ b/include/hw/virtio/virtio.h
> > @@ -168,6 +168,7 @@ void virtio_init(VirtIODevice *vdev, const char *name,
> >                           uint16_t device_id, size_t config_size);
> >  void virtio_cleanup(VirtIODevice *vdev);
> >  
> > +void virtio_set_disabled(VirtIODevice *vdev, bool disable);
> >  void virtio_error(VirtIODevice *vdev, const char *fmt, ...) 
> > GCC_FMT_ATTR(2, 3);
> >  
> >  /* Set the child bus name. */
> 
> 
> One open question here is cross version migration.
> What exactly happens if we migrate to an old qemu?
> from an old qemu?

Currently vdev->broken is already migrated via a subsection if it is set.

So, with the current proposed patch I would expect the following:

 old -> new:
   dev->broken set due to proposed virtio_set_disabled():
     - N/A
   dev->broken set due to other/existing reasons:
     - value is migrated, migration succeeds
     - device continues not working until a reset, or bus-mastering
       enabled/re-enabled. (which is good, but also makes me realize
       that a device that was put into broken state for reasons
       other than virtio_set_disabled() should not get 'unbroken'
       simply because bus-master bit was cycled. a separate flag
       is probably needed)
     - PASS
   dev->broken not set:
     - not migrated
     - PASS
 new -> old:
   dev->broken set due to proposed virtio_set_disabled():
     - migration succeeds to any QEMU that already has handling for
       dev->broken.
     - migration fails for any QEMU that doesn't have handling for
       dev->broken as it does now, but will a higher chance of
       triggering
     - device won't work until it is reset. for most guests we will
       probably get a reset before the device is used again anyway. for
       the SLOF case the device will stay broken after bus-mastering
       re-enabled, but that's the case for existing QEMU now anyway
     - PASS, but with increased chance of migration breakage for
       QEMU's that don't have handling for dev->broken.
   dev->broken set due to other/existing reasons:
     - migration succeeds to any QEMU that already has handling for
       dev->broken. device will require a reset as it does now.
     - migration fails for any QEMU that doesn't have handling for
       dev->broken as it does now
     - PASS
   dev->broken not set:
     - not migrated
     - PASS

With a new dev->disabled flag, which we'd likely send using a subsection
like with dev->broken, I would expect the following:

 old -> new:
   dev->disabled set:
     - N/A
   dev->disabled not set:
     - not migrated
     - if source disables BM just before migration we will likely go
       into an error state that either sets dev->broken or puts the
       device in some other possibly bad state. either case would require
       subsequent reset, just as they would without migration
     - PASS
 new -> old:
   dev->disabled set:
     - migration stream fails
     - FAIL, but as expected, and doesn't seem common currently outside
       of SLOF early-boot else we'd probably have more reports of
       breakage from vring access while BM isn't set.
   dev->broken not set:
     - not migrated
     - PASS

So re-using 'broken' is slightly better from a migration standpoint, but
as noted above it is probably wrong to unset 'broken' just because BM bit
gets cycled, so a new 'disabled' flag is probably needed.

We could probably get by with just adding a check for dev->disabled in 
virtio_queue_empty(), or even earlier in
virtio_queue_host_notifier_aio_poll(), but it seems more proper to also
add it in most of the same places we currently check for dev->broken.

That seems somewhat redundant though, so I think maybe the best approach
is to:

 - replace most dev->broken checks with checks for dev->disabled
 - set dev->disabled whenever dev->broken gets set
 - add a check in virtio_set_disabled() that only allows us to re-enable if
   dev->broken hasn't also been set.

I'll work on a follow-up using that approach if it seems reasonable to
you.

Thanks!

> 
> > -- 
> > 2.17.1
>