diff mbox series

[v3,05/10] drm/i915: Prepare for multiple gts

Message ID 20211029032817.3747750-6-matthew.d.roper@intel.com (mailing list archive)
State New, archived
Headers show
Series i915: Initial multi-tile support | expand

Commit Message

Matt Roper Oct. 29, 2021, 3:28 a.m. UTC 
From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

Add some basic plumbing to support more than one dynamically allocated
struct intel_gt.  Up to four gts are supported in i915->gts[], with slot
zero shadowing the existing i915->gt to enable source compatibility with
legacy driver paths.  A for_each_gt macro is added to iterate over the
GTs and will be used by upcoming patches that convert various parts of
the driver to be multi-gt aware.

v2:
 - Rename init function to i915_init_tile_memory() and move it to
   i915_drv.c. (Lucas)
 - Squash in patch from Sandeep to release the per-gt resources during
   driver teardown.

Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Venkata Sandeep Dhanalakota <venkata.s.dhanalakota@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
---
 drivers/gpu/drm/i915/gt/intel_gt.c         | 57 +++++++++++++++++++---
 drivers/gpu/drm/i915/gt/intel_gt.h         |  6 +++
 drivers/gpu/drm/i915/gt/intel_gt_types.h   |  2 +
 drivers/gpu/drm/i915/i915_drv.c            | 30 +++++++++++-
 drivers/gpu/drm/i915/i915_drv.h            |  6 +++
 drivers/gpu/drm/i915/intel_memory_region.h |  3 ++
 6 files changed, 96 insertions(+), 8 deletions(-)

Comments

Andi Shyti Nov. 1, 2021, 11:11 p.m. UTC | #1 
Hi Matt and Tvrtko,

[...]

>  static int
>  intel_gt_tile_setup(struct intel_gt *gt, unsigned int id, phys_addr_t phys_addr)

we don't actually need 'id', it's gt->info.id. It's introduced in
patch 3 with the value '0' but it's not needed.

>  {
> +	struct drm_i915_private *i915 = gt->i915;
> +	struct intel_uncore *uncore;
> +	struct intel_uncore_mmio_debug *mmio_debug;
>  	int ret;
>  
> -	intel_uncore_init_early(gt->uncore, gt);
> +	if (id) {

if (gt->info.id) ?

Andi

> +		uncore = kzalloc(sizeof(*uncore), GFP_KERNEL);
> +		if (!uncore)
> +			return -ENOMEM;
> +
> +		mmio_debug = kzalloc(sizeof(*mmio_debug), GFP_KERNEL);
> +		if (!mmio_debug) {
> +			kfree(uncore);
> +			return -ENOMEM;
> +		}
> +
> +		__intel_gt_init_early(gt, uncore, i915);
> +	} else {
> +		uncore = &i915->uncore;
> +		mmio_debug = &i915->mmio_debug;
> +	}
> +
> +	intel_uncore_init_early(uncore, gt);
>  
>  	ret = intel_uncore_setup_mmio(gt->uncore, phys_addr);
Tvrtko Ursulin Nov. 2, 2021, 9:36 a.m. UTC | #2 
On 01/11/2021 23:11, Andi Shyti wrote:
> Hi Matt and Tvrtko,
> 
> [...]
> 
>>   static int
>>   intel_gt_tile_setup(struct intel_gt *gt, unsigned int id, phys_addr_t phys_addr)
> 
> we don't actually need 'id', it's gt->info.id. It's introduced in
> patch 3 with the value '0' but it's not needed.

I have a suspicion code got munged up over endless rebases and refactors.

This patch is the one which introduces the id member to gt->info. But it is not setting it, even though I suspect the intent was for intel_gt_tile_setup to do that.

Instead gt->info.id is only set to a valid value in last patch of this series inside intel_gt_probe_all:

+		gt->i915 = i915;
+		gt->name = gtdef->name;
+		gt->type = gtdef->type;
+		gt->info.engine_mask = gtdef->engine_mask;
+		gt->info.id = i;
+
+		drm_dbg(&i915->drm, "Setting up %s %u\n", gt->name, gt->info.id);
+		ret = intel_gt_tile_setup(gt, i, phys_addr + gtdef->mapping_base);

And intel_gt_tile_setup then calls __intel_gt_init_early which assigns gt->i915 yet again.

So I'd say there is probably space to bring this all into a more streamlined flow, even more than what you suggest below.

Regards,

Tvrtko
  
>>   {
>> +	struct drm_i915_private *i915 = gt->i915;
>> +	struct intel_uncore *uncore;
>> +	struct intel_uncore_mmio_debug *mmio_debug;
>>   	int ret;
>>   
>> -	intel_uncore_init_early(gt->uncore, gt);
>> +	if (id) {
> 
> if (gt->info.id) ?
> 
> Andi
> 
>> +		uncore = kzalloc(sizeof(*uncore), GFP_KERNEL);
>> +		if (!uncore)
>> +			return -ENOMEM;
>> +
>> +		mmio_debug = kzalloc(sizeof(*mmio_debug), GFP_KERNEL);
>> +		if (!mmio_debug) {
>> +			kfree(uncore);
>> +			return -ENOMEM;
>> +		}
>> +
>> +		__intel_gt_init_early(gt, uncore, i915);
>> +	} else {
>> +		uncore = &i915->uncore;
>> +		mmio_debug = &i915->mmio_debug;
>> +	}
>> +
>> +	intel_uncore_init_early(uncore, gt);
>>   
>>   	ret = intel_uncore_setup_mmio(gt->uncore, phys_addr);
Andi Shyti Nov. 2, 2021, 11:26 a.m. UTC | #3 
Hi Tvrtko,

> > [...]
> > 
> > >   static int
> > >   intel_gt_tile_setup(struct intel_gt *gt, unsigned int id, phys_addr_t phys_addr)
> > 
> > we don't actually need 'id', it's gt->info.id. It's introduced in
> > patch 3 with the value '0' but it's not needed.
> 
> I have a suspicion code got munged up over endless rebases and refactors.
> 
> This patch is the one which introduces the id member to gt->info. But it is not setting it, even though I suspect the intent was for intel_gt_tile_setup to do that.
> 
> Instead gt->info.id is only set to a valid value in last patch of this series inside intel_gt_probe_all:
> 
> +		gt->i915 = i915;
> +		gt->name = gtdef->name;
> +		gt->type = gtdef->type;
> +		gt->info.engine_mask = gtdef->engine_mask;
> +		gt->info.id = i;
> +
> +		drm_dbg(&i915->drm, "Setting up %s %u\n", gt->name, gt->info.id);
> +		ret = intel_gt_tile_setup(gt, i, phys_addr + gtdef->mapping_base);
> 
> And intel_gt_tile_setup then calls __intel_gt_init_early which assigns gt->i915 yet again.
> 
> So I'd say there is probably space to bring this all into a more streamlined flow, even more than what you suggest below.

yes, I noticed them!

Patch 3, 5 and 10 are very much connected with each other: 3
prepares for one tile, 5 prepares for multitile and 10 does the
multitile. While in between other patches are doing other things.

On top of some cleanups we could also rearrange the patches with
some squashing and reordering to have them a bit more linear and
also easier to review.

Andi
Tvrtko Ursulin Nov. 2, 2021, 1:58 p.m. UTC | #4 
On 02/11/2021 11:26, Andi Shyti wrote:
> Hi Tvrtko,
> 
>>> [...]
>>>
>>>>    static int
>>>>    intel_gt_tile_setup(struct intel_gt *gt, unsigned int id, phys_addr_t phys_addr)
>>>
>>> we don't actually need 'id', it's gt->info.id. It's introduced in
>>> patch 3 with the value '0' but it's not needed.
>>
>> I have a suspicion code got munged up over endless rebases and refactors.
>>
>> This patch is the one which introduces the id member to gt->info. But it is not setting it, even though I suspect the intent was for intel_gt_tile_setup to do that.
>>
>> Instead gt->info.id is only set to a valid value in last patch of this series inside intel_gt_probe_all:
>>
>> +		gt->i915 = i915;
>> +		gt->name = gtdef->name;
>> +		gt->type = gtdef->type;
>> +		gt->info.engine_mask = gtdef->engine_mask;
>> +		gt->info.id = i;
>> +
>> +		drm_dbg(&i915->drm, "Setting up %s %u\n", gt->name, gt->info.id);
>> +		ret = intel_gt_tile_setup(gt, i, phys_addr + gtdef->mapping_base);
>>
>> And intel_gt_tile_setup then calls __intel_gt_init_early which assigns gt->i915 yet again.
>>
>> So I'd say there is probably space to bring this all into a more streamlined flow, even more than what you suggest below.
> 
> yes, I noticed them!
> 
> Patch 3, 5 and 10 are very much connected with each other: 3
> prepares for one tile, 5 prepares for multitile and 10 does the
> multitile. While in between other patches are doing other things.
> 
> On top of some cleanups we could also rearrange the patches with
> some squashing and reordering to have them a bit more linear and
> also easier to review.

Yes. Maybe make intel_gt_tile_setup accept more arguments so it can be 
truly used to setup a gt?

   intel_gt_tile_setup(gt, id, name, type, engine_mask)

The usual thing where patch which adds something extends the prototype 
to include more stuff. If that applies here.

I know it is originally my patch but I don't have the time to rework it, 
much less the whole series, so usual dispensation to take over 
authorship if changes are large applies.

Regards,

Tvrtko
Andi Shyti Nov. 2, 2021, 9:08 p.m. UTC | #5 
Hi Tvrtko,

> > > > [...]
> > > > 
> > > > >    static int
> > > > >    intel_gt_tile_setup(struct intel_gt *gt, unsigned int id, phys_addr_t phys_addr)
> > > > 
> > > > we don't actually need 'id', it's gt->info.id. It's introduced in
> > > > patch 3 with the value '0' but it's not needed.
> > > 
> > > I have a suspicion code got munged up over endless rebases and refactors.
> > > 
> > > This patch is the one which introduces the id member to gt->info. But it is not setting it, even though I suspect the intent was for intel_gt_tile_setup to do that.
> > > 
> > > Instead gt->info.id is only set to a valid value in last patch of this series inside intel_gt_probe_all:
> > > 
> > > +		gt->i915 = i915;
> > > +		gt->name = gtdef->name;
> > > +		gt->type = gtdef->type;
> > > +		gt->info.engine_mask = gtdef->engine_mask;
> > > +		gt->info.id = i;
> > > +
> > > +		drm_dbg(&i915->drm, "Setting up %s %u\n", gt->name, gt->info.id);
> > > +		ret = intel_gt_tile_setup(gt, i, phys_addr + gtdef->mapping_base);
> > > 
> > > And intel_gt_tile_setup then calls __intel_gt_init_early which assigns gt->i915 yet again.
> > > 
> > > So I'd say there is probably space to bring this all into a more streamlined flow, even more than what you suggest below.
> > 
> > yes, I noticed them!
> > 
> > Patch 3, 5 and 10 are very much connected with each other: 3
> > prepares for one tile, 5 prepares for multitile and 10 does the
> > multitile. While in between other patches are doing other things.
> > 
> > On top of some cleanups we could also rearrange the patches with
> > some squashing and reordering to have them a bit more linear and
> > also easier to review.
> 
> Yes. Maybe make intel_gt_tile_setup accept more arguments so it can be truly
> used to setup a gt?
> 
>   intel_gt_tile_setup(gt, id, name, type, engine_mask)
> 
> The usual thing where patch which adds something extends the prototype to
> include more stuff. If that applies here.
> 
> I know it is originally my patch but I don't have the time to rework it,
> much less the whole series, so usual dispensation to take over authorship if
> changes are large applies.

as no one is stepping forward, if you and Matt are OK, I can try
to venture in some refactoring of these three patches (3, 5 and
10).

Andi
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c
index 098cd8843c38..d02a09653033 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt.c
+++ b/drivers/gpu/drm/i915/gt/intel_gt.c
@@ -23,10 +23,13 @@ 
 #include "shmem_utils.h"
 #include "pxp/intel_pxp.h"
 
-void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915)
+static void
+__intel_gt_init_early(struct intel_gt *gt,
+		      struct intel_uncore *uncore,
+		      struct drm_i915_private *i915)
 {
 	gt->i915 = i915;
-	gt->uncore = &i915->uncore;
+	gt->uncore = uncore;
 
 	spin_lock_init(&gt->irq_lock);
 
@@ -49,10 +52,15 @@  void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915)
 int intel_gt_probe_lmem(struct intel_gt *gt)
 {
 	struct drm_i915_private *i915 = gt->i915;
+	unsigned int instance = gt->info.id;
 	struct intel_memory_region *mem;
 	int id;
 	int err;
 
+	id = INTEL_REGION_LMEM + instance;
+	if (drm_WARN_ON(&i915->drm, id >= INTEL_REGION_STOLEN_SMEM))
+		return -ENODEV;
+
 	mem = intel_gt_setup_lmem(gt);
 	if (mem == ERR_PTR(-ENODEV))
 		mem = intel_gt_setup_fake_lmem(gt);
@@ -67,9 +75,8 @@  int intel_gt_probe_lmem(struct intel_gt *gt)
 		return err;
 	}
 
-	id = INTEL_REGION_LMEM;
-
 	mem->id = id;
+	mem->instance = instance;
 
 	intel_memory_region_set_name(mem, "local%u", mem->instance);
 
@@ -80,6 +87,11 @@  int intel_gt_probe_lmem(struct intel_gt *gt)
 	return 0;
 }
 
+void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915)
+{
+	__intel_gt_init_early(gt, &i915->uncore, i915);
+}
+
 void intel_gt_init_hw_early(struct intel_gt *gt, struct i915_ggtt *ggtt)
 {
 	gt->ggtt = ggtt;
@@ -903,9 +915,29 @@  u32 intel_gt_read_register_fw(struct intel_gt *gt, i915_reg_t reg)
 static int
 intel_gt_tile_setup(struct intel_gt *gt, unsigned int id, phys_addr_t phys_addr)
 {
+	struct drm_i915_private *i915 = gt->i915;
+	struct intel_uncore *uncore;
+	struct intel_uncore_mmio_debug *mmio_debug;
 	int ret;
 
-	intel_uncore_init_early(gt->uncore, gt);
+	if (id) {
+		uncore = kzalloc(sizeof(*uncore), GFP_KERNEL);
+		if (!uncore)
+			return -ENOMEM;
+
+		mmio_debug = kzalloc(sizeof(*mmio_debug), GFP_KERNEL);
+		if (!mmio_debug) {
+			kfree(uncore);
+			return -ENOMEM;
+		}
+
+		__intel_gt_init_early(gt, uncore, i915);
+	} else {
+		uncore = &i915->uncore;
+		mmio_debug = &i915->mmio_debug;
+	}
+
+	intel_uncore_init_early(uncore, gt);
 
 	ret = intel_uncore_setup_mmio(gt->uncore, phys_addr);
 	if (ret)
@@ -920,6 +952,11 @@  static void
 intel_gt_tile_cleanup(struct intel_gt *gt)
 {
 	intel_uncore_cleanup_mmio(gt->uncore);
+
+	if (gt->info.id) {
+		kfree(gt->uncore);
+		kfree(gt);
+	}
 }
 
 int intel_gt_probe_all(struct drm_i915_private *i915)
@@ -937,13 +974,21 @@  int intel_gt_probe_all(struct drm_i915_private *i915)
 	if (ret)
 		return ret;
 
+	i915->gts[0] = &i915->gt;
+
 	/* TODO: add more tiles */
 	return 0;
 }
 
 void intel_gt_release_all(struct drm_i915_private *i915)
 {
-	intel_gt_tile_cleanup(&i915->gt);
+	struct intel_gt *gt;
+	unsigned int id;
+
+	for_each_gt(i915, id, gt) {
+		intel_gt_tile_cleanup(gt);
+		i915->gts[id] = NULL;
+	}
 }
 
 void intel_gt_info_print(const struct intel_gt_info *info,
diff --git a/drivers/gpu/drm/i915/gt/intel_gt.h b/drivers/gpu/drm/i915/gt/intel_gt.h
index 68cdf042ad88..a59521a8c96c 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt.h
+++ b/drivers/gpu/drm/i915/gt/intel_gt.h
@@ -88,6 +88,12 @@  u32 intel_gt_read_register_fw(struct intel_gt *gt, i915_reg_t reg);
 int intel_gt_probe_all(struct drm_i915_private *i915);
 void intel_gt_release_all(struct drm_i915_private *i915);
 
+#define for_each_gt(i915__, id__, gt__) \
+	for ((id__) = 0; \
+	     (id__) < I915_MAX_TILES; \
+	     (id__)++) \
+		for_each_if(((gt__) = (i915__)->gts[(id__)]))
+
 void intel_gt_info_print(const struct intel_gt_info *info,
 			 struct drm_printer *p);
 
diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h
index 66143316d92e..7311e485faae 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h
@@ -186,6 +186,8 @@  struct intel_gt {
 	phys_addr_t phys_addr;
 
 	struct intel_gt_info {
+		unsigned int id;
+
 		intel_engine_mask_t engine_mask;
 
 		u32 l3bank_mask;
diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index 000b1b070155..fde148d6777e 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -382,10 +382,14 @@  static int i915_driver_early_probe(struct drm_i915_private *dev_priv)
  */
 static void i915_driver_late_release(struct drm_i915_private *dev_priv)
 {
+	struct intel_gt *gt;
+	unsigned int id;
+
 	intel_irq_fini(dev_priv);
 	intel_power_domains_cleanup(dev_priv);
 	i915_gem_cleanup_early(dev_priv);
-	intel_gt_driver_late_release(&dev_priv->gt);
+	for_each_gt(dev_priv, id, gt)
+		intel_gt_driver_late_release(gt);
 	intel_region_ttm_device_fini(dev_priv);
 	vlv_suspend_cleanup(dev_priv);
 	i915_workqueues_cleanup(dev_priv);
@@ -512,6 +516,28 @@  static int i915_set_dma_info(struct drm_i915_private *i915)
 	return ret;
 }
 
+/**
+ * i915_init_tile_memory - initialize per-tile memory
+ * @i915: valid i915 instance
+ *
+ * Current multi-tile platforms include a GT and a memory region within each
+ * tile.  We need to initialize each.
+ */
+static int i915_init_tile_memory(struct drm_i915_private *i915)
+{
+	struct intel_gt *gt;
+	unsigned int id;
+	int ret;
+
+	for_each_gt(i915, id, gt) {
+		ret = intel_gt_probe_lmem(gt);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
 /**
  * i915_driver_hw_probe - setup state requiring device access
  * @dev_priv: device private
@@ -579,7 +605,7 @@  static int i915_driver_hw_probe(struct drm_i915_private *dev_priv)
 
 	intel_gt_init_hw_early(&dev_priv->gt, &dev_priv->ggtt);
 
-	ret = intel_gt_probe_lmem(&dev_priv->gt);
+	ret = i915_init_tile_memory(dev_priv);
 	if (ret)
 		goto err_mem_regions;
 
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 19e6700a4315..10a4817e397d 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1193,6 +1193,12 @@  struct drm_i915_private {
 	/* Abstract the submission mechanism (legacy ringbuffer or execlists) away */
 	struct intel_gt gt;
 
+	/*
+	 * i915->gts[0] == &i915->gt
+	 */
+#define I915_MAX_TILES 4
+	struct intel_gt *gts[I915_MAX_TILES];
+
 	struct {
 		struct i915_gem_contexts {
 			spinlock_t lock; /* locks list */
diff --git a/drivers/gpu/drm/i915/intel_memory_region.h b/drivers/gpu/drm/i915/intel_memory_region.h
index 3feae3353d33..d255e4bffb23 100644
--- a/drivers/gpu/drm/i915/intel_memory_region.h
+++ b/drivers/gpu/drm/i915/intel_memory_region.h
@@ -31,6 +31,9 @@  enum intel_memory_type {
 enum intel_region_id {
 	INTEL_REGION_SMEM = 0,
 	INTEL_REGION_LMEM,
+	INTEL_REGION_LMEM1,
+	INTEL_REGION_LMEM2,
+	INTEL_REGION_LMEM3,
 	INTEL_REGION_STOLEN_SMEM,
 	INTEL_REGION_STOLEN_LMEM,
 	INTEL_REGION_UNKNOWN, /* Should be last */