[Spice-devel] [PATCH 3.1/12] Store QXLInstance in CursorItem

Frediano Ziglio fziglio at redhat.com
Mon Nov 2 02:58:28 PST 2015


> 
> On Fri, 2015-10-30 at 03:24 -0400, Frediano Ziglio wrote:
> > > 
> > > From: Marc-André Lureau <marcandre.lureau at gmail.com>
> > > 
> > > Doing so allows us to remove the extra QXLInstance parameter from
> > > cursor_item_unref() and makes the code a bit cleaner.
> > > 
> > > Also add cursor_item_ref().
> > > 
> > > Signed-off-by: Jonathon Jongsma <jjongsma at redhat.com>
> > > ---
> > >  server/cursor-channel.c | 70
> > >  +++++++++++++++++++++++++++----------------------
> > >  server/cursor-channel.h |  9 ++++---
> > >  2 files changed, 44 insertions(+), 35 deletions(-)
> > > 
> > > diff --git a/server/cursor-channel.c b/server/cursor-channel.c
> > > index 6cc2b93..2d7afc6 100644
> > > --- a/server/cursor-channel.c
> > > +++ b/server/cursor-channel.c
> > > @@ -25,55 +25,58 @@
> > >  #include "cache_item.tmpl.c"
> > >  #undef CLIENT_CURSOR_CACHE
> > >  
> > > -static inline CursorItem *alloc_cursor_item(void)
> > > +CursorItem *cursor_item_new(QXLInstance *qxl, RedCursorCmd *cmd,
> > > uint32_t
> > > group_id)
> > >  {
> > >      CursorItem *cursor_item;
> > >  
> > > +    spice_return_val_if_fail(cmd != NULL, NULL);
> > > +
> > >      cursor_item = g_slice_new0(CursorItem);
> > > +    cursor_item->qxl = qxl;
> > >      cursor_item->refs = 1;
> > > +    cursor_item->group_id = group_id;
> > > +    cursor_item->red_cursor = cmd;
> > >  
> > >      return cursor_item;
> > >  }
> > >  
> > > -CursorItem *cursor_item_new(RedCursorCmd *cmd, uint32_t group_id)
> > > +CursorItem *cursor_item_ref(CursorItem *item)
> > >  {
> > > -    CursorItem *cursor_item;
> > > -
> > > -    spice_return_val_if_fail(cmd != NULL, NULL);
> > > -    cursor_item = alloc_cursor_item();
> > > +    spice_return_val_if_fail(item != NULL, NULL);
> > > +    spice_return_val_if_fail(item->refs > 0, NULL);
> > 
> > This catch a memory corruption, should abort the program,
> > I suggest spice_error, spice_critical or spice_assert.
> 
> It can suggest memory corruption, yes. But I'm not sure that it should
> be fatal. This is exactly the same behavior as g_object_ref(), so I
> don't see any good reason to make this warning fatal when the exact
> same problem with a gobject will not be fatal.
> 

Although glib is now "server aware" you have to consider was born as a
client library where some problem are considered in a different way.
Also GObject pointers are a bit more complicated (for instance they support
weak pointers) that could lead to some temporary strange state.
The counters in our code should be handled in a single thread (from worker),
start from 1, incremented for ref, decremented and object freed on zero.
If everything follows these rules counters are always >0.
If they became <= 0 it means to me:
- some unrelated memory corruption (counters written from unexpected code,
  physical memory errors, invalid dma transfers or OS bugs). Not the type
  of errors you can handle in a safe way;
- counter was incremented till overflow. You are spending lot of memory
  with a lot of objects (there are at least 2^31 pointers to this object!)
  something is really wrong in your program; DoS is likely to happen;
- counter was decremented more then expected. You have pointer to now freed
  data. Dandling pointers are bound to heap memory corruptions which is
  a security problem.

In a server context I don't see the point to keep the program going on these
conditions beside opening the doors to hackers.

> > 
> > >  
> > > -    cursor_item->group_id = group_id;
> > > -    cursor_item->red_cursor = cmd;
> > > +    item->refs++;
> > >  
> > > -    return cursor_item;
> > > +    return item;
> > >  }
> > >  
> > > -void cursor_item_unref(QXLInstance *qxl, CursorItem *cursor)
> > > +void cursor_item_unref(CursorItem *item)
> > >  {
> > > -    if (!--cursor->refs) {
> > > -        QXLReleaseInfoExt release_info_ext;
> > > -        RedCursorCmd *cursor_cmd;
> > > -
> > > -        cursor_cmd = cursor->red_cursor;
> > > -        release_info_ext.group_id = cursor->group_id;
> > > -        release_info_ext.info = cursor_cmd->release_info;
> > > -        qxl->st->qif->release_resource(qxl, release_info_ext);
> > > -        red_put_cursor_cmd(cursor_cmd);
> > > -        free(cursor_cmd);
> > > -
> > > -        g_slice_free(CursorItem, cursor);
> > > -    }
> > > +    QXLReleaseInfoExt release_info_ext;
> > > +    RedCursorCmd *cursor_cmd;
> > > +
> > > +    spice_return_if_fail(item != NULL);
> > > +
> > > +    if (--item->refs)
> > > +        return;
> > > +
> > > +    cursor_cmd = item->red_cursor;
> > > +    release_info_ext.group_id = item->group_id;
> > > +    release_info_ext.info = cursor_cmd->release_info;
> > > +    item->qxl->st->qif->release_resource(item->qxl,
> > > release_info_ext);
> > > +    red_put_cursor_cmd(cursor_cmd);
> > > +    free(cursor_cmd);
> > > +
> > > +    g_slice_free(CursorItem, item);
> > > +
> > >  }
> > >  
> > >  static void cursor_set_item(CursorChannel *cursor, CursorItem
> > > *item)
> > >  {
> > >      if (cursor->item)
> > > -        cursor_item_unref(red_worker_get_qxl(cursor
> > > ->common.worker),
> > > cursor->item);
> > > -
> > > -    if (item)
> > > -        item->refs++;
> > > +        cursor_item_unref(cursor->item);
> > >  
> > > -    cursor->item = item;
> > > +    cursor->item = item ? cursor_item_ref(item) : NULL;
> > >  }
> > >  
> > 
> > Are these tests worth doing or can we assume that a NULL pointer is
> > expected (like
> > unref(NULL) and ref(NULL) do nothing) ? This would make
> > cursor_set_item like
> > 
> > static void cursor_set_item(CursorChannel *cursor, CursorItem *item)
> > {
> >     cursor_item_unref(cursor->item);
> > 
> >     cursor->item = cursor_item_ref(item);
> > }
> 
> 
> I'd prefer to stay closer to the semantics of g_object_ref/unref(),
> which warn if the argument is not a valid GObject.
> 
> > 
> > 
> > On naming, sometimes is cursor_item sometimes item. I would prefer
> > cursor_item.
> > 
> > >  static PipeItem *new_cursor_pipe_item(RedChannelClient *rcc, void
> > > *data, int
> > >  num)
> > > @@ -155,7 +158,7 @@ static void
> > > put_cursor_pipe_item(CursorChannelClient
> > > *ccc, CursorPipeItem *pipe_
> > >  
> > >      spice_assert(!pipe_item_is_linked(&pipe_item->base));
> > >  
> > > -    cursor_item_unref(red_worker_get_qxl(ccc->common.worker),
> > > pipe_item->cursor_item);
> > > +    cursor_item_unref(pipe_item->cursor_item);
> > >      free(pipe_item);
> > >  }
> > >  
> > > @@ -411,7 +414,11 @@ void cursor_channel_process_cmd(CursorChannel
> > > *cursor,
> > > RedCursorCmd *cursor_cmd,
> > >      CursorItem *cursor_item;
> > >      int cursor_show = FALSE;
> > >  
> > > -    cursor_item = cursor_item_new(cursor_cmd, group_id);
> > > +    spice_return_if_fail(cursor);
> > > +    spice_return_if_fail(cursor_cmd);
> > > +
> > > +    cursor_item = cursor_item_new(red_worker_get_qxl(cursor
> > > ->common.worker),
> > > +                                  cursor_cmd, group_id);
> > >  
> > >      switch (cursor_cmd->type) {
> > >      case QXL_CURSOR_SET:
> > > @@ -439,7 +446,8 @@ void cursor_channel_process_cmd(CursorChannel
> > > *cursor,
> > > RedCursorCmd *cursor_cmd,
> > >          red_channel_pipes_new_add(&cursor->common.base,
> > >          new_cursor_pipe_item,
> > >                                    (void*)cursor_item);
> > >      }
> > > -    cursor_item_unref(red_worker_get_qxl(cursor->common.worker),
> > > cursor_item);
> > > +
> > > +    cursor_item_unref(cursor_item);
> > >  }
> > >  
> > >  void cursor_channel_reset(CursorChannel *cursor)
> > > diff --git a/server/cursor-channel.h b/server/cursor-channel.h
> > > index 2c19859..8b49df7 100644
> > > --- a/server/cursor-channel.h
> > > +++ b/server/cursor-channel.h
> > > @@ -33,6 +33,7 @@
> > >  #define CURSOR_CACHE_HASH_KEY(id) ((id) & CURSOR_CACHE_HASH_MASK)
> > >  
> > >  typedef struct CursorItem {
> > > +    QXLInstance *qxl;
> > >      uint32_t group_id;
> > >      int refs;
> > >      RedCursorCmd *red_cursor;
> > > @@ -83,14 +84,14 @@ void                 cursor_channel_reset
> > > (CursorChannel *cursor);
> > >  void                 cursor_channel_process_cmd (CursorChannel
> > > *cursor,
> > >  RedCursorCmd *cursor_cmd,
> > >                                                   uint32_t
> > > group_id);
> > >  
> > > -CursorItem*          cursor_item_new            (RedCursorCmd
> > > *cmd, uint32_t
> > > group_id);
> > > -void                 cursor_item_unref          (QXLInstance *qxl,
> > > CursorItem *cursor);
> > > -
> > > -
> > >  CursorChannelClient *cursor_channel_client_new(CommonChannel
> > > *common,
> > >                                                 RedClient *client,
> > > RedsStream
> > >                                                 *stream,
> > >                                                 int mig_target,
> > >                                                 uint32_t
> > > *common_caps, int
> > >                                                 num_common_caps,
> > >                                                 uint32_t *caps, int
> > >                                                 num_caps);
> > >  
> > > +CursorItem*          cursor_item_new            (QXLInstance *qxl,
> > > RedCursorCmd *cmd, uint32_t group_id);
> > > +CursorItem*          cursor_item_ref            (CursorItem
> > > *cursor);
> > > +void                 cursor_item_unref          (CursorItem
> > > *cursor);
> > > +
> > >  #endif /* CURSOR_CHANNEL_H_ */
> > 

Frediano


More information about the Spice-devel mailing list