[Intel-gfx] [PATCH] drm/i915: Implement workaround for broken CS tlb on i830/845

[Ville Syrjälä]

On Mon, Dec 17, 2012 at 04:12:18PM +0100, Daniel Vetter wrote:
> On Mon, Dec 17, 2012 at 04:46:17PM +0200, Ville Syrjälä wrote:
> > On Mon, Dec 17, 2012 at 02:43:03PM +0100, Daniel Vetter wrote:
> > > On Mon, Dec 17, 2012 at 01:33:01PM +0000, Chris Wilson wrote:
> > > > On Sun, 16 Dec 2012 22:46:00 +0000, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > > > > @@ -1087,6 +1087,9 @@ i915_error_first_batchbuffer(struct drm_i915_private *dev_priv,
> > > > >  	if (!ring->get_seqno)
> > > > >  		return NULL;
> > > > >  
> > > > > +	if (HAS_BROKEN_CS_TLB(dev_priv->dev))
> > > > > +		return i915_error_object_create(dev_priv, ring->private);
> > > > 
> > > > Hmm, this is complicated by userspace opting out of the CS w/a, and
> > > > imposes quite a burden upon our simple seq interface.
> > > 
> > > Right, I've written this without userspace being able to opt out in mind
> > > ... for the seq -ENOMEM, I guess it's just time to bite the bullet. Loads
> > > of the error_states for the ilk fallout couldn't be dumped (but could be
> > > captured) by bug reporters already :(
> > 
> > Assuming you're talking about debugfs error_state returning -ENOMEM due
> > to seq_file's massive kmalloc(), I had a couple of ideas for fixing it
> > in seq_file itself.
> > 
> > 1) just use vmalloc()
> > 
> > 2) use multiple pages instead of one big allocation
> > 
> >  seq_printf() {
> >    try to print the line
> >    if not enough space {
> >      mark the end of valid data in current page
> >      allocate a new page
> >      print again
> >    }
> >  }
> > 
> > And adjust seq_read()/seq_lseek accordingly.
> > 
> > Of course then you can't print anything > PAGE_SIZE,
> > but that seems unlikely anyway, and if really needed
> > it could try to allocate something larger than a page
> > when needed.
> 
> Afaik that's pretty much what the real seq_file interface does,

AFAICS it just starts w/ a buffer size of one page, and it just
retries in a loop calling ->show() and doubling the buffer size
every time seq_printf() indicated an overflow. So it wants to
print the whole thing into a single contiguous buffer.

-- 
Ville Syrjälä
Intel OTC