[Spice-devel] [PATCH spice-server v2] Manage lifetime of RedVmcChannel

[Frediano Ziglio]

> 
> On Wed, 2016-11-02 at 08:31 -0400, Frediano Ziglio wrote:
> > > 
> > > 
> > > Hi,
> > > 
> > > On Wed, Nov 02, 2016 at 04:34:47AM -0400, Frediano Ziglio wrote:
> > > > 
> > > > > 
> > > > > 
> > > > > From: Jonathon Jongsma <jjongsma at redhat.com>
> > > > > 
> > > > > Store the channel in the char device object, and unref it in
> > > > > dispose.
> > > > > Previously, we were just creating a channel and allowing it to
> > > > > leak.
> > > > > There may be better long-term solutions, but this works for
> > > > > now.
> > > > > ---
> > > > >  server/spicevmc.c | 27 ++++++++++++++++++++-------
> > > > >  1 file changed, 20 insertions(+), 7 deletions(-)
> > > > > 
> > > > > Changes:
> > > > > - squashed fixup.
> > > > > 
> > > > > diff --git a/server/spicevmc.c b/server/spicevmc.c
> > > > > index 7067bc7..a77e868 100644
> > > > > --- a/server/spicevmc.c
> > > > > +++ b/server/spicevmc.c
> > > > > @@ -47,6 +47,9 @@
> > > > >  #define BUF_SIZE (64 * 1024 + 32)
> > > > >  #define COMPRESS_THRESHOLD 1000
> > > > >  
> > > > > +typedef struct RedVmcChannel RedVmcChannel;
> > > > > +typedef struct RedVmcChannelClass RedVmcChannelClass;
> > > > > +
> > > > >  typedef struct RedVmcPipeItem {
> > > > >      RedPipeItem base;
> > > > >  
> > > > > @@ -75,6 +78,7 @@ typedef struct RedCharDeviceSpiceVmcClass
> > > > > RedCharDeviceSpiceVmcClass;
> > > > >  
> > > > >  struct RedCharDeviceSpiceVmc {
> > > > >      RedCharDevice parent;
> > > > > +    RedVmcChannel *channel;
> > > > >  };
> > > > >  
> > > > >  struct RedCharDeviceSpiceVmcClass
> > > > > @@ -89,9 +93,6 @@ static RedCharDevice
> > > > > *red_char_device_spicevmc_new(SpiceCharDeviceInstance *sin,
> > > > >  
> > > > >  G_DEFINE_TYPE(RedCharDeviceSpiceVmc, red_char_device_spicevmc,
> > > > >  RED_TYPE_CHAR_DEVICE)
> > > > >  
> > > > > -#define RED_CHAR_DEVICE_SPICEVMC_PRIVATE(o) \
> > > > > -    (G_TYPE_INSTANCE_GET_PRIVATE ((o),
> > > > > RED_TYPE_CHAR_DEVICE_SPICEVMC,
> > > > > RedCharDeviceSpiceVmcPrivate))
> > > > > -
> > > > >  #define RED_TYPE_VMC_CHANNEL red_vmc_channel_get_type()
> > > > >  
> > > > >  #define RED_VMC_CHANNEL(obj) \
> > > > > @@ -103,9 +104,6 @@ G_DEFINE_TYPE(RedCharDeviceSpiceVmc,
> > > > > red_char_device_spicevmc, RED_TYPE_CHAR_DEV
> > > > >  #define RED_VMC_CHANNEL_GET_CLASS(obj) \
> > > > >      (G_TYPE_INSTANCE_GET_CLASS((obj), RED_TYPE_VMC_CHANNEL,
> > > > >      RedVmcChannelClass))
> > > > >  
> > > > > -typedef struct RedVmcChannel RedVmcChannel;
> > > > > -typedef struct RedVmcChannelClass RedVmcChannelClass;
> > > > > -
> > > > >  struct RedVmcChannel
> > > > >  {
> > > > >      RedChannel parent;
> > > > > @@ -855,9 +853,13 @@ RedCharDevice
> > > > > *spicevmc_device_connect(RedsState
> > > > > *reds,
> > > > >                                         SpiceCharDeviceInstance
> > > > > *sin,
> > > > >                                         uint8_t channel_type)
> > > > >  {
> > > > > +    RedCharDeviceSpiceVmc *dev_state;
> > > > >      RedVmcChannel *state = red_vmc_channel_new(reds,
> > > > > channel_type, sin);
> > > > >  
> > > > > -    return state->chardev;
> > > > > +    dev_state = RED_CHAR_DEVICE_SPICEVMC(state->chardev);
> > > > > +    dev_state->channel = state;
> > > > > +
> > > > > +    return RED_CHAR_DEVICE(dev_state);
> > > > >  }
> > > > >  
> > > > >  /* Must be called from RedClient handling thread. */
> > > > > @@ -904,10 +906,21 @@ SPICE_GNUC_VISIBLE void
> > > > > spice_server_port_event(SpiceCharDeviceInstance *sin, ui
> > > > >  }
> > > > >  
> > > > >  static void
> > > > > +red_char_device_spicevmc_dispose(GObject *object)
> > > > > +{
> > > > > +    RedCharDeviceSpiceVmc *self =
> > > > > RED_CHAR_DEVICE_SPICEVMC(object);
> > > > > +
> > > > > +    g_clear_object(&self->channel);
> > > > 
> > > > This call produce a double free.
> > > > The channel is already freed in spicevmc_device_disconnect.
> > > 
> > > But g_clear_object() verify if channel is null before the unref and
> > > set
> > > it to NULL afterwards.
> > > 
> > 
> > Doing a NULL check on a dangling pointer is not much helpful.
> 
> Hmm, you're right. If we're going to remove the 'opaque' property (in a
> different patch), I think we do need to keep a pointer to the channel
> inside of the char device. But perhaps it should be a weak reference
> instead. So maybe we could just remove the dispose changes from this
> patch and squash it with that patch?
> 

Yes, I think we should kind of merge part of this patch in the other one.
This could also be fixed setting the pointer to NULL after calling
spicevmc_device_disconnect.

> Also: perhaps the RedsState object should hold a reference to the
> channels added via reds_register_channel() and unref the channel in
> reds_unregister_channel(). It seems that it shouldn't be necessary to
> call red_channel_destroy() explicitly. This should just happen in the
> RedChannel destructor when the last ref is dropped...
> 
> 

There is already a patch in the queue that does this; you wrote it :)
(cfr "RedsState: keep weak reference to channels").
This patch does not seem really in line with your proposal for
spice_server_destroy/reds_exit changes.

I was going to reply to it. There are some issues:
- reds_channel_destroyed is called by RedChannel destructor (dispose/finalize);
  As channels is not protected by a lock this means that all calls to
  reds_channel_destroyed (so in theory all calls to g_object_unref to
  RedChannel hierarchy) has to be done in the "main" thread... or we need to
  add locking! It's easy to check reds_unregister_channel calls, a bit less
  g_object_refs;
- is it the right time to remove the item from the list or is too late?
  Potentially the list is scanned when a new connection arrive to decide which
  RedChannel will handle the connection (and create a new RedChannelClient).
  Not saying that is wrong, surely will avoid dandling pointers but some race
  may occur.

> > 
> > > 
> > > > 
> > > > 
> > > > > 
> > > > > +}
> > > > > +
> > > > > +static void
> > > > >  red_char_device_spicevmc_class_init(RedCharDeviceSpiceVmcClass
> > > > > *klass)
> > > > >  {
> > > > > +    GObjectClass *object_class = G_OBJECT_CLASS(klass);
> > > > >      RedCharDeviceClass *char_dev_class =
> > > > > RED_CHAR_DEVICE_CLASS(klass);
> > > > > 
> > > > > +    object_class->dispose = red_char_device_spicevmc_dispose;
> > > > > +
> > > > >      char_dev_class->read_one_msg_from_device =
> > > > >      spicevmc_chardev_read_msg_from_dev;
> > > > >      char_dev_class->send_msg_to_client =
> > > > >      spicevmc_chardev_send_msg_to_client;
> > > > >      char_dev_class->send_tokens_to_client =
> > > > >      spicevmc_char_dev_send_tokens_to_client;
> > > > 
> > > > Removing the double free from this patch will basically revert
> > > > the
> > > > entire patch except the RED_CHAR_DEVICE_SPICEVMC_PRIVATE cleanup.
> > > > 
> > 

Frediano