[PATCH] dri2: Reference count DRI2 buffers

[Dave Airlie]

2010/8/21 Oldřich Jedlička <oldium.pro at seznam.cz>:
> On Friday 20 August 2010 11:05:40 Christopher James Halse Rogers wrote:
>> On Fri, 2010-08-20 at 07:55 +0200, Oldřich Jedlička wrote:
>> > On Friday 20 August 2010 02:04:35 Christopher James Halse Rogers wrote:
>> > > On Thu, 2010-08-19 at 21:23 +0200, Oldřich Jedlička wrote:
>> > > > On Thursday 19 August 2010 10:57:21 Christopher James Halse Rogers
> wrote:
>> > > > > When a client calls ScheduleSwap we set up a kernel callback when
>> > > > > the relevent vblank event occurs.  However, it's possible for the
>> > > > > client to go away between calling ScheduleSwap and the vblank
>> > > > > event, resulting in the buffers being destroyed before they're
>> > > > > passed to radeon_dri2_frame_event_handler.
>> > > >
>> > > > I was also thinking about the solution and did some xorg-server
>> > > > investigation. Personally I don't like comparing pointer values
>> > > > (ClientPtr), because it could be the same - sequence
>> > > > malloc/free/malloc could return the same pointer value.
>> > >
>> > > Yeah, there is a chance that a stray DRI2_SwapBuffersComplete event
>> > > could be written to an uninterested client.  It certainly won't try to
>> > > write to an invalid client, though, so it shouldn't crash X.  And it is
>> > > reasonably unlikely that a client will go away and a new client will
>> > > both fill the same slot *and* get the same memory address - there's a
>> > > lot of memory allocation going on in the surrounding code.
>> >
>> > Question is if the client can handle this :-)
>> >
>> > > > Here are two other possible solutions:
>> > > >
>> > > > 1. Add a "uniqueId" (increasing number with each client) to
>> > > > ClientRec. Then you can compare something really unique. On the
>> > > > other hand this needs change in xorg-server.
>> > > >
>> > > > 2. Use AddCallback(&ClientStateCallback,
>> > > > our_client_state_changed_method, 0) during driver initialization to
>> > > > detect that any client went away and invalidate its events (add
>> > > > field "valid" in event). That would be even better than solution 1 -
>> > > > no change of xorg-server is needed. Each client could have private
>> > > > data too - double-linked list of pending events - registered with
>> > > > dixRegisterPrivateKey(). The event would be removed from the list
>> > > > only when it is valid (otherwise the prev/next list pointers would
>> > > > be invalid too). Invalid events would be ignored in the handler,
>> > > > they would be only freed.
>> > >
>> > > I thought about that, too.  It seemed a bit excessive for the driver to
>> > > maintain a list of clients as they come and go just for the purpose of
>> > > not sending an event when a client quits.
>> >
>> > There is no need to have a list of clients. Each client would have the
>> > list of events kept in the client's devPrivate area (registered with
>> > dixRegisterPrivateKey, found by dixLookupPrivate) - ony the event list
>> > head is necessary:
>> >
>> > 1. The event would have "prev" and "next" pointers for the purposes of
>> > the list and a new "valid" boolean field.
>> >
>> > 2. Register Client State callback by the call to AddCallback and call
>> > dixRegisterPrivateKey in the driver initialization routine. Add
>> > RemoveCallback in the driver shut-down routine.
>> >
>> > 3. Whenever new client connects, the list head would be set to 0 (done in
>> > the Client State callback). This step is probably unnecessary if the
>> > area is set to 0 via calloc (I'm just not sure).
>> >
>> > 4. Whenever the new event is created, dixLookupPrivate would get the
>> > client's list head and the new event would be added to the list head.
>> >
>> > 5. Whenever the client dies (recognized in the Client State callback),
>> > the list would be walked-through and events invalidated (valid=false).
>> >
>> > 6. For valid events (valid==true) on event callback the event would be
>> > removed from the list (just modify the event's prev's "next" and event's
>> > next's "prev" pointers, eventually modifying the client's list head).
>> >
>> > 7. For invalid events (valid==false) the list would stay unmodified
>> > (because of the list head modification on freed client's memory), only
>> > the event would be freed.
>> >
>> > This looks to me like a few lines of code for each point, nothing big.
>>
>> Ah, right.  That's the reverse of what I was thinking; it's more
>> reasonable.
>>
>> It still seems a bit heavyweight to me for this corner case.
>
> Yeah, looks so. I think it is now the ATI driver developpers turn to say what
> they want - if your patch is good or needs enhancing.
>
>> > > The pointer comparison is quick, cheap, and ensures we won't crash X.
>> >
>> > Yes, that should work most of the time :-) But this might add some
>> > hard-to- reproduce problem with client getting unwanted message.
>> >
>> > > > Personally I like solution 2, because it fully uses xorg-server
>> > > > facilities. But I don't know if this isn't too much or if there
>> > > > exists a simpler solution.
>> > >
>> > > I think that there should actually be solution 3: the DRI2 extension
>> > > handles this for drivers as a part of the swapbuffers/waitmsc common
>> > > code.
>> >
>> > Yes, definitely.
>> >
>> > But it looks like the driver is currently scheduling the event (and
>> > holding wrong data - ClientPtr) and handling it, only notifying DRI2 on
>> > xserver about the result. So that would mean creating some common part
>> > that handles event creation/handling/invalidation. The driver would call
>> > xserver when the event arrives and xserver would call driver back if it
>> > still applies. Or something simillar...
>>
>> Well, the need to reference count buffers could easily be be removed by
>> having DRI2ScheduleSwap take a reference to the buffers and
>> DRI2SwapComplete decrement that, in the same way that it handles
>> pending_swaps and such already.
>>
>> Similarly, if we need to handle the Client gone fun it could be handled
>> there.
>
> That's right. Maybe somebody from xorg-devel list would be interrested.

It would probably be better to move this discussion to xorg-devel, or
at least involve Kristian,.

I'll try and review this thread next week if I can, I suck at dealing
with swapbuffers stuff

Dave.