[PATCH] DRI2: fixup handling of last_swap_target
mario.kleiner at tuebingen.mpg.de
Mon Mar 8 11:13:21 PST 2010
On Mar 8, 2010, at 6:07 PM, Jesse Barnes wrote:
> On Sun, 7 Mar 2010 21:16:21 +0100
> Florian Mickler <florian at mickler.org> wrote:
>> On Sun, 7 Mar 2010 09:10:51 -0800
>> Jesse Barnes <jbarnes at virtuousgeek.org> wrote:
>>> On Sun, 7 Mar 2010 08:44:51 +0100
>>> Mario Kleiner <mario.kleiner at tuebingen.mpg.de> wrote:
>>> Yeah, this could also happen without OML I think, if a few swaps
>>> queued (resulting in a block) and then a SGI_video_sync call
>>> I'll fix it up.
>> could this also happen with a single glxgears instance and no other
>> 3d clients running?
> I don't *think* glxgears does this; it's been running fine for me at
> But you could definitely be hitting one of the MSC races; you can add
> some debug output to I830DRI2ScheduleSwap in src/i830_dri.c in the DDX
> driver to see what's going on. Presumably one of the paths there is
> putting the client to sleep and never waking it.
The glxgears.c of mesa only does draw -> glXSwapBuffers -> draw ->
glXSwapBuffers -> ...
It doesn't use any of the new api's, so the only way it can block is
probably via the DRI2ThrottleClient call when requesting a "new"
I see at least one problematic thing. In the xserver's dri2.c file in
DRI2SwapBuffers(), the pPriv->swapsPending++ statement is *after* the
call into the ddx's I830DRI2ScheduleSwap() function, instead of
*before* it. If a swap completes inside the ddx, it will call
DRI2SwapComplete() which will do a pPriv->swapsPending-- before it
was incremented to mark a swap as pending. That is, if swapsPending
was zero before, it will now wrap around to 0xffffffff. This will
happen if the fallback_blit path inside the ddx is used, e.g., if the
drawable is not visible (yet).
Not sure if this by itself is sufficient for the hang, because the
call to pPriv->swapsPending++ after return from I830DRI2ScheduleSwap
() should fix this, but if something else goes wrong and an error
path is taken or an event not delivered, then swapsPending could get
stuck at 0xffffffff or a similar high number, which would trigger
DRI2ThrottleClient and block the client connection without ever
waking it up again, as clients are only woken up if a swap completes,
which can't happen if no swap is pending. -> hang in _XReply on the
So we should probable move pPriv->swapsPending++ before the call to...
ret = (*ds->ScheduleSwap)(client, pDraw, pDestBuffer, pSrcBuffer,
swap_target, divisor, remainder, func,
... in DRI2SwapBuffers() and probably add a pPriv->swapsPending--
into the error handling path directly after the call to ds-
More information about the xorg-devel