Re: [PATCH v7 4/4] drm/i915: cache number of MOCS entries

[Lucas De Marchi <lucas.demarchi@intel.com>]

On Fri, Dec 21, 2018 at 11:56:43AM +0000, Tvrtko Ursulin wrote:
> 
> On 14/12/2018 18:20, Lucas De Marchi wrote:
> > Instead of checking the gen number every time we need to know the max
> > number of entries, just save it into the table struct so we don't need
> > extra branches throughout the code.
> > 
> > Suggested-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> > Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
> > ---
> >   drivers/gpu/drm/i915/intel_mocs.c | 31 ++++++++++++++-----------------
> >   1 file changed, 14 insertions(+), 17 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/intel_mocs.c b/drivers/gpu/drm/i915/intel_mocs.c
> > index dfc4edea020f..22c5f576a3c2 100644
> > --- a/drivers/gpu/drm/i915/intel_mocs.c
> > +++ b/drivers/gpu/drm/i915/intel_mocs.c
> > @@ -32,6 +32,7 @@ struct drm_i915_mocs_entry {
> >   struct drm_i915_mocs_table {
> >   	u32 size;
> > +	u32 n_entries;
> 
> While at it I'd convert both counts to normal unsigned int.
> 
> Another nitpick: I'd also suggest some more descriptive names since I read
> n_entries and size completely opposite than what they are in the code. Maybe
> just s/n_entries/max_entries/ to keep the diff small, or even consider
> changing s/size/used_entries/ or something?

size = ARRAY_SIZE() -- we use that to track accesses to the table so we
don't access beyond the array size.

n_entries = GEN[X]_NUM_ENTRIES -- we use that to track the number of
entries we will program.

So IMO the names look reasonable to me.

> 
> >   	const struct drm_i915_mocs_entry *table;
> >   };
> > @@ -56,9 +57,6 @@ struct drm_i915_mocs_table {
> >   #define GEN9_NUM_MOCS_ENTRIES	62  /* 62 out of 64 - 63 & 64 are reserved. */
> >   #define GEN11_NUM_MOCS_ENTRIES	64  /* 63-64 are reserved, but configured. */
> > -#define NUM_MOCS_ENTRIES(i915) \
> > -	(INTEL_GEN(i915) < 11 ? GEN9_NUM_MOCS_ENTRIES : GEN11_NUM_MOCS_ENTRIES)
> > -
> 
> Do you want to go through patch 3 adds this, patch 4 removes it, or why not
> just squash it into one?

I considered doing that, but then thought the split approach would be
more logical since this patch is about caching the number of entries.
Squashing on the previous patch creates more noise on that patch and
additional headache since I'm not the original author. Not having this
macro in the previous patch basically means squashing this entire patch
there.

So in the end I decided to do it on top. Is that ok to you?


> 
> >   /* (e)LLC caching options */
> >   #define LE_0_PAGETABLE		_LE_CACHEABILITY(0)
> >   #define LE_1_UC			_LE_CACHEABILITY(1)
> > @@ -283,14 +281,17 @@ static bool get_mocs_settings(struct drm_i915_private *dev_priv,
> >   	if (IS_ICELAKE(dev_priv)) {
> >   		table->size  = ARRAY_SIZE(icelake_mocs_table);
> > +		table->n_entries = GEN11_NUM_MOCS_ENTRIES;
> >   		table->table = icelake_mocs_table;
> >   		result = true;
> >   	} else if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
> >   		table->size  = ARRAY_SIZE(skylake_mocs_table);
> > +		table->n_entries = GEN9_NUM_MOCS_ENTRIES;
> >   		table->table = skylake_mocs_table;
> >   		result = true;
> >   	} else if (IS_GEN9_LP(dev_priv)) {
> >   		table->size  = ARRAY_SIZE(broxton_mocs_table);
> > +		table->n_entries = GEN9_NUM_MOCS_ENTRIES;
> >   		table->table = broxton_mocs_table;
> >   		result = true;
> >   	} else {
> > @@ -348,8 +349,6 @@ void intel_mocs_init_engine(struct intel_engine_cs *engine)
> >   	if (!get_mocs_settings(dev_priv, &table))
> >   		return;
> > -	GEM_BUG_ON(table.size > NUM_MOCS_ENTRIES(dev_priv));
> > -
> >   	for (index = 0; index < table.size; index++)
> >   		I915_WRITE(mocs_register(engine->id, index),
> >   			   table.table[index].control_value);
> > @@ -362,7 +361,7 @@ void intel_mocs_init_engine(struct intel_engine_cs *engine)
> >   	 * Entry 0 in the table is uncached - so we are just writing
> >   	 * that value to all the used entries.
> >   	 */
> > -	for (; index < NUM_MOCS_ENTRIES(dev_priv); index++)
> > +	for (; index < table.n_entries; index++)
> >   		I915_WRITE(mocs_register(engine->id, index),
> >   			   table.table[0].control_value);
> >   }
> > @@ -380,19 +379,18 @@ void intel_mocs_init_engine(struct intel_engine_cs *engine)
> >   static int emit_mocs_control_table(struct i915_request *rq,
> >   				   const struct drm_i915_mocs_table *table)
> >   {
> > -	struct drm_i915_private *i915 = rq->i915;
> >   	enum intel_engine_id engine = rq->engine->id;
> >   	unsigned int index;
> >   	u32 *cs;
> > -	if (WARN_ON(table->size > NUM_MOCS_ENTRIES(i915)))
> > +	if (GEM_WARN_ON(table->size > table->n_entries))
> >   		return -ENODEV;
> 
> This (and another below) could go to get_mocs_settings which is now the
> authority to set things up correctly. So fire a warn from there and say no
> mocs for you.

yep, this seems reasonable. I will change that in the next version after
this one settles.

Lucas De Marchi

> 
> > -	cs = intel_ring_begin(rq, 2 + 2 * NUM_MOCS_ENTRIES(i915));
> > +	cs = intel_ring_begin(rq, 2 + 2 * table->n_entries);
> >   	if (IS_ERR(cs))
> >   		return PTR_ERR(cs);
> > -	*cs++ = MI_LOAD_REGISTER_IMM(NUM_MOCS_ENTRIES(i915));
> > +	*cs++ = MI_LOAD_REGISTER_IMM(table->n_entries);
> >   	for (index = 0; index < table->size; index++) {
> >   		*cs++ = i915_mmio_reg_offset(mocs_register(engine, index));
> > @@ -407,7 +405,7 @@ static int emit_mocs_control_table(struct i915_request *rq,
> >   	 * Entry 0 in the table is uncached - so we are just writing
> >   	 * that value to all the used entries.
> >   	 */
> > -	for (; index < NUM_MOCS_ENTRIES(i915); index++) {
> > +	for (; index < table->n_entries; index++) {
> >   		*cs++ = i915_mmio_reg_offset(mocs_register(engine, index));
> >   		*cs++ = table->table[0].control_value;
> >   	}
> > @@ -440,18 +438,17 @@ static inline u32 l3cc_combine(const struct drm_i915_mocs_table *table,
> >   static int emit_mocs_l3cc_table(struct i915_request *rq,
> >   				const struct drm_i915_mocs_table *table)
> >   {
> > -	struct drm_i915_private *i915 = rq->i915;
> >   	unsigned int i;
> >   	u32 *cs;
> > -	if (WARN_ON(table->size > NUM_MOCS_ENTRIES(i915)))
> > +	if (GEM_WARN_ON(table->size > table->n_entries))
> >   		return -ENODEV;
> > -	cs = intel_ring_begin(rq, 2 + NUM_MOCS_ENTRIES(i915));
> > +	cs = intel_ring_begin(rq, 2 + table->n_entries);
> >   	if (IS_ERR(cs))
> >   		return PTR_ERR(cs);
> > -	*cs++ = MI_LOAD_REGISTER_IMM(NUM_MOCS_ENTRIES(i915) / 2);
> > +	*cs++ = MI_LOAD_REGISTER_IMM(table->n_entries / 2);
> >   	for (i = 0; i < table->size/2; i++) {
> >   		*cs++ = i915_mmio_reg_offset(GEN9_LNCFCMOCS(i));
> > @@ -470,7 +467,7 @@ static int emit_mocs_l3cc_table(struct i915_request *rq,
> >   	 * this will be uncached. Leave the last pair uninitialised as
> >   	 * they are reserved by the hardware.
> >   	 */
> > -	for (; i < NUM_MOCS_ENTRIES(i915) / 2; i++) {
> > +	for (; i < table->n_entries / 2; i++) {
> >   		*cs++ = i915_mmio_reg_offset(GEN9_LNCFCMOCS(i));
> >   		*cs++ = l3cc_combine(table, 0, 0);
> >   	}
> > @@ -517,7 +514,7 @@ void intel_mocs_init_l3cc_table(struct drm_i915_private *dev_priv)
> >   	 * this will be uncached. Leave the last pair as initialised as
> >   	 * they are reserved by the hardware.
> >   	 */
> > -	for (; i < (NUM_MOCS_ENTRIES(dev_priv) / 2); i++)
> > +	for (; i < table.n_entries / 2; i++)
> >   		I915_WRITE(GEN9_LNCFCMOCS(i), l3cc_combine(&table, 0, 0));
> >   }
> > 
> 
> Otherwise looks good - thanks for agreeing the suggestion makes sense!
> 
> Regards,
> 
> Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx