[Nouveau] [PATCH 01/10] drm/nv50: decode PGRAPH status registers on TLB flush fail

Marcin Slusarz marcin.slusarz at gmail.com
Mon Aug 20 10:02:46 PDT 2012


On Mon, Aug 20, 2012 at 04:27:18PM +1000, Ben Skeggs wrote:
> On Sun, Aug 19, 2012 at 10:59:56PM +0200, Marcin Slusarz wrote:
> > Now it outputs:
> > nouveau E[  PGRAPH][0000:02:00.0] PGRAPH TLB flush idle timeout fail
> > nouveau E[  PGRAPH][0000:02:00.0] PGRAPH_STATUS: BUSY DISPATCH VFETCH CCACHE_UNK4 STRMOUT_GSCHED_UNK5 UNK14XX UNK1CXX CLIPID ZCULL ENG2D UNK34XX TPRAST TPROP ROP (0x011fde03)
> > nouveau E[  PGRAPH][0000:02:00.0] PGRAPH_VSTATUS_0: CCACHE (0x00145b4d)
> > nouveau E[  PGRAPH][0000:02:00.0] PGRAPH_VSTATUS_1: (0x0000002d)
> > nouveau E[  PGRAPH][0000:02:00.0] PGRAPH_VSTATUS_2: ENG2D ROP (0x0034db40)
> > 
> > instead of:
> > [drm] nouveau 0000:02:00.0: PGRAPH TLB flush idle timeout fail: 0x011fde03 0x00145b4d 0x0000002d 0x0034db40
> 
> I didn't push these first few patches yet as I have a couple of thoughts on it.
> 
> Mostly I was thinking we should probably have a:
> 
> nv_printf(obj, lvl, fmt, args...)
> 
> to replace the printk's which just does a continued print while still obeying
> the debug level?

Well, this patch does not use nv_printk_enabled, so these concerns should
not apply to it, right? I'll fix up bitfield list and resend it.

Speaking of nv_printf, I have a couple of problems with it:
- it would repeatedly recheck whether to print something or not
- it would always evaluate its arguments
- nouveau_enum_print / nouveau_bitfield_print would need object/level arguments
- you still would need nv_printk_enabled for levels >= DEBUG (like in mxms.c)

It's a lot of wasted CPU time...

> > 
> > Based on envytools docs.
> > 
> > Signed-off-by: Marcin Slusarz <marcin.slusarz at gmail.com>
> > ---
> >  drivers/gpu/drm/nouveau/core/engine/graph/nv50.c |   78 ++++++++++++++++++++-
> >  1 files changed, 74 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/nv50.c b/drivers/gpu/drm/nouveau/core/engine/graph/nv50.c
> > index c93b525..f60aec9 100644
> > --- a/drivers/gpu/drm/nouveau/core/engine/graph/nv50.c
> > +++ b/drivers/gpu/drm/nouveau/core/engine/graph/nv50.c
> > @@ -184,6 +184,62 @@ nv50_graph_tlb_flush(struct nouveau_engine *engine)
> >  	return 0;
> >  }
> >  
> > +static struct nouveau_bitfield nv50_pgraph_status[] = {
> > +	{ 1<<0, "BUSY" }, /* set when any bit is set */
> > +	{ 1<<1, "DISPATCH" },
> > +	{ 1<<2, "UNK2" },
> > +	{ 1<<3, "UNK3" },
> > +	{ 1<<4, "UNK4" },
> > +	{ 1<<5, "UNK5" },
> > +	{ 1<<6, "M2MF" },
> > +	{ 1<<7, "UNK7" },
> > +	{ 1<<8, "CTXPROG" },
> > +	{ 1<<9, "VFETCH" },
> > +	{ 1<<10, "CCACHE_UNK4" },
> > +	{ 1<<11, "STRMOUT_GSCHED_UNK5" },
> > +	{ 1<<12, "UNK14XX" },
> > +	{ 1<<13, "UNK24XX_CSCHED" },
> > +	{ 1<<14, "UNK1CXX" },
> > +	{ 1<<15, "CLIPID" },
> > +	{ 1<<16, "ZCULL" },
> > +	{ 1<<17, "ENG2D" },
> > +	{ 1<<18, "UNK34XX" },
> > +	{ 1<<19, "TPRAST" },
> > +	{ 1<<20, "TPROP" },
> > +	{ 1<<21, "TEX" },
> > +	{ 1<<22, "TPVP" },
> > +	{ 1<<23, "MP" },
> > +	{ 1<<24, "ROP" },
> > +	{}
> > +};
> hex bitfields please :)

Sure. I thought it would be easier to read.

> > +
> > +static const char *const nv50_pgraph_vstatus_0[] = {
> > +	"VFETCH", "CCACHE", "UNK4", "UNK5", "GSCHED", "STRMOUT", "UNK14XX", NULL
> > +};
> > +
> > +static const char *const nv50_pgraph_vstatus_1[] = {
> > +	"TPRAST", "TPROP", "TEXTURE", "TPVP", "MP", NULL
> > +};
> > +
> > +static const char *const nv50_pgraph_vstatus_2[] = {
> > +	"UNK24XX", "CSCHED", "UNK1CXX", "CLIPID", "ZCULL", "ENG2D", "UNK34XX",
> > +	"ROP", NULL
> > +};
> > +
> > +static void nouveau_pgraph_vstatus_print(const char *const units[], u32 status)
> > +{
> > +	int i;
> > +	u32 tmp = status;
> > +	for (i = 0; units[i] && tmp; i++) {
> > +		if ((tmp & 7) == 1)
> > +			pr_cont("%s ", units[i]);
> > +		tmp >>= 3;
> > +	}
> > +	if (tmp)
> > +		pr_cont("invalid: %x ", tmp);
> > +	pr_cont("(0x%08x)\n", status);
> > +}
> > +
> >  static int
> >  nv84_graph_tlb_flush(struct nouveau_engine *engine)
> >  {
> > @@ -219,10 +275,24 @@ nv84_graph_tlb_flush(struct nouveau_engine *engine)
> >  		 !(timeout = ptimer->read(ptimer) - start > 2000000000));
> >  
> >  	if (timeout) {
> > -		nv_error(priv, "PGRAPH TLB flush idle timeout fail: "
> > -			      "0x%08x 0x%08x 0x%08x 0x%08x\n",
> > -			 nv_rd32(priv, 0x400700), nv_rd32(priv, 0x400380),
> > -			 nv_rd32(priv, 0x400384), nv_rd32(priv, 0x400388));
> > +		nv_error(priv, "PGRAPH TLB flush idle timeout fail\n");
> > +
> > +		nv_error(priv, "PGRAPH_STATUS: ");
> > +		tmp = nv_rd32(priv, 0x400700);
> > +		nouveau_bitfield_print(nv50_pgraph_status, tmp);
> > +		pr_cont(" (0x%08x)\n", tmp);
> > +
> > +		nv_error(priv, "PGRAPH_VSTATUS_0: ");
> > +		nouveau_pgraph_vstatus_print(nv50_pgraph_vstatus_0,
> > +				nv_rd32(priv, 0x400380));
> > +
> > +		nv_error(priv, "PGRAPH_VSTATUS_1: ");
> > +		nouveau_pgraph_vstatus_print(nv50_pgraph_vstatus_1,
> > +				nv_rd32(priv, 0x400384));
> > +
> > +		nv_error(priv, "PGRAPH_VSTATUS_2: ");
> > +		nouveau_pgraph_vstatus_print(nv50_pgraph_vstatus_2,
> > +				nv_rd32(priv, 0x400388));
> >  	}
> >  
> >  	nv50_vm_flush_engine(&engine->base, 0x00);
> > -- 


More information about the Nouveau mailing list