[Mesa-dev] [PATCH 3/4] i965: Prevent infinite finish_batch recursion

[Ben Widawsky]

On Sat, Feb 28, 2015 at 12:31:59AM -0800, Kenneth Graunke wrote:
> On Friday, February 27, 2015 10:22:10 AM Ben Widawsky wrote:
> > From the comments in the code:
> > 
> >    Called from intel_batchbuffer_flush before emitting MI_BATCHBUFFER_END and
> >    sending it off.
> > 
> > This fixes a possible, unlikely infinite recursion in our batch flush path. More
> > importantly it allows me to add some code here.
> > 
> > The relevant part of the call chain for flush
> > intel_batchbuffer_flush()->brw_finish_batch(). The problem arises if anything in
> > the time from intel_batchbuffer_flush, until brw_finish_batch ends up calling
> > itself. This can happen as a result of a call to
> > intel_batchbuffer_begin()->intel_batchbuffer_require_space()->intel_batchbuffer_flush().
> > 
> > There are two possible cases today which can spawn this recursion.
> > 1. There is a ring switch occurring (impossible, see below)
> > 2. The ring is out of space (fairly unlikely)
> 
> NAK.  brw_finish_batch is for putting any last-minute commands in the
> batchbuffer before sending it off.
> 
> Such commands must fit in the BATCH_RESERVED space (which is probably
> already large enough since SNB has a zillion workarounds).  This
> eliminates reason #2.
> 
> #1 means that you're trying to emit commands for the wrong ring, i.e.
> trying to do RCS workarounds in a BLT batch.  This isn't what you want.

Sorry you wasted your time looking at this. I had fixed both of what you
mentioned, but hadn't sent it out yet. It's fixed with a better commit
message here:
http://cgit.freedesktop.org/~bwidawsk/mesa/commit/?h=workarounds&id=f87c17e34eeac9d5e3f54c9b182090f8aede0284