Re: [Qemu-devel] [PATCH v4] tests/vhost-user-bridge: implement logging of dirty pages

[Michael S. Tsirkin]

On Tue, Nov 17, 2015 at 01:40:12PM +0200, Victor Kaplansky wrote:
> On Tue, Nov 17, 2015 at 01:04:33PM +0200, Michael S. Tsirkin wrote:
> > On Tue, Nov 17, 2015 at 12:04:06PM +0200, Victor Kaplansky wrote:
> > > During migration devices continue writing to the guest's memory.
> > > The writes has to be reported to QEMU. This change implements
> > > minimal support in vhost-user-bridge required for successful
> > > migration of a guest with virtio-net device.
> > > 
> > > Signed-off-by: Victor Kaplansky <address@hidden>
> > > ---
> > > v4:
> > >     - implement set_vring_enable()
> > >     - stop the device on vubr_get_vring_base_exec() and
> > >       start on setting kick_fd as a work-around to QEMU bug
> > >       enabling vrings too early.
> > 
> > You do want to address that FIXME but at least it's
> > a documented bug.
> > 
> > > v3:
> > >     - Get rid of vhost_log_chunk_t. Just use uint8_t.
> > >     - Implement vubr_set_log_fd_exec().
> > >     - Kick the log if log_call_fd has been set up.
> > >     - Mark bits in log table atomically to enable more then one
> > >       simultaneous vhost-user backend.
> > >     - Fix the calculations of required log table size in an
> > >       assert.
> > >     - Fix the coding style: only single space before assignment
> > >       operator.
> > >     - Add a comment on the hack to determine that queues are
> > >       ready for processing.
> > >     - Other minor cosmetic fixes.
> > > v2:
> > >     - use log_guest_addr for used ring reported by qemu instead of
> > >       translating.
> > >     - use mmap_size and mmap_offset defined in new
> > >       VHOST_USER_SET_LOG_BASE interface. See the patch
> > >       "vhost-user: modify SET_LOG_BASE to pass mmap size and
> > >       offset".
> > >     - start logging dirty pages only after the appropriate feature
> > >       is set by a VHOST_USER_GET_PROTOCOL_FEATURES request.
> > >     - updated TODO list.
> > > 
> > >  tests/vhost-user-bridge.c | 220 
> > > ++++++++++++++++++++++++++++++++++++++++------
> > >  1 file changed, 195 insertions(+), 25 deletions(-)
> > > 
> > > diff --git a/tests/vhost-user-bridge.c b/tests/vhost-user-bridge.c
> > > index 864f69e..7bdfc98 100644
> > > --- a/tests/vhost-user-bridge.c
> > > +++ b/tests/vhost-user-bridge.c
> > > @@ -466,11 +507,39 @@ vubr_virtqueue_kick(VubrVirtq *vq)
> > >  }
> > >  
> > >  static void
> > > +vubr_log_page(uint8_t *log_table, uint64_t page)
> > > +{
> > > +    DPRINT("Logged dirty guest page: %"PRId64"\n", page);
> > > +    atomic_or(&log_table[page / 8], 1 << (page % 8));
> > 
> > One thing you definitely want here is a check
> > against log size.
> > Otherwise it's a security hole:
> > remote can corrupt your memory by making you
> > set random bits outside the mapped buffer.
> > 
> > What to do on error? Probably abort.
> > 
> 
> vubr_log_write() has an assert checking the address.
> In the next version I'll replace if by an if statement.

assert is ok I think.

> > > +}
> > > +
> > > +static void
> > > +vubr_log_write(VubrDev *dev, uint64_t address, uint64_t length)
> > > +{
> > > +    uint64_t page;
> > > +
> > > +    if (!(dev->features & (1ULL << VHOST_F_LOG_ALL)) ||
> > > +        !dev->log_table || !length) {
> > > +        return;
> > > +    }
> > > +
> > > +    assert(dev->log_size > ((address + length - 1) / VHOST_LOG_PAGE / 
> > > 8));
> > > +
> > > +    page = address / VHOST_LOG_PAGE;
> > > +    while (page * VHOST_LOG_PAGE < address + length) {
> > > +        vubr_log_page(dev->log_table, page);
> > > +        page += VHOST_LOG_PAGE;
> > > +    }
> > > +    vubr_log_kick(dev);
> > > +}
> > > +
> 
> -- Victor