IGT bits for small-bar

Add helpers for object creation and querying the cpu_visible related bits.

v2: Make it backwards compat with older kernels

Signed-off-by Matthew Auld <[email protected]>

Cc Gwan-gyeong Mun <[email protected]>

Reviewed-by Gwan-gyeong Mun <[email protected]>

---
 lib/xe/xe_query.c | 86 ++++++++++++++++++++++++++++++++++++++++++++++-
 lib/xe/xe_query.h |  6 ++++
 2 files changed, 91 insertions(+), 1 deletion(-)

lib/xe/xe_query.c +32 -1

index f535ad853..5412fbeb5 100644
@@ -140,6 +140,17 @@ static uint64_t gt_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
 	return 0;
 }
 
+static uint64_t gt_visible_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
+				     const struct drm_xe_query_gts *gts, int gt)
+{
+	int region_idx = ffs(native_region_for_gt(gts, gt)) - 1;
+
+	if (XE_IS_CLASS_VRAM(&mem_usage->regions[region_idx]))
+		return mem_usage->regions[region_idx].cpu_visible_size;
+
+	return 0;
+}
+
 static bool __mem_has_vram(struct drm_xe_query_mem_usage *mem_usage)
 {
 	for (int i = 0; i < mem_usage->num_regions; i++)

@@ -246,9 +257,14 @@ struct xe_device *xe_device_get(int fd)
 	xe_dev->hw_engines = xe_query_engines_new(fd, &xe_dev->number_hw_engines);
 	xe_dev->mem_usage = xe_query_mem_usage_new(fd);
 	xe_dev->vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->vram_size));
-	for (int gt = 0; gt < xe_dev->number_gt; gt++)
+	xe_dev->visible_vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->visible_vram_size));
+	for (int gt = 0; gt < xe_dev->number_gt; gt++) {
 		xe_dev->vram_size[gt] = gt_vram_size(xe_dev->mem_usage,
 						     xe_dev->gts, gt);
+		xe_dev->visible_vram_size[gt] =
+			gt_visible_vram_size(xe_dev->mem_usage,
+					     xe_dev->gts, gt);
+	}
 	xe_dev->default_alignment = __mem_default_alignment(xe_dev->mem_usage);
 	xe_dev->has_vram = __mem_has_vram(xe_dev->mem_usage);
 

@@ -383,6 +399,33 @@ uint64_t vram_memory(int fd, int gt)
 	return xe_has_vram(fd) ? native_region_for_gt(xe_dev->gts, gt) : 0;
 }
 
+static uint64_t __xe_visible_vram_size(int fd, int gt)
+{
+	struct xe_device *xe_dev;
+
+	xe_dev = find_in_cache(fd);
+	igt_assert(xe_dev);
+
+	return xe_dev->visible_vram_size[gt];
+}
+
+/**
+ * visible_vram_memory:
+ * @fd: xe device fd
+ * @gt: gt id
+ *

lib/xe/xe_query.c +13 -0

+}
+
 /**
  * vram_if_possible:
  * @fd: xe device fd

@@ -396,6 +439,28 @@ uint64_t vram_if_possible(int fd, int gt)
 	return vram_memory(fd, gt) ?: system_memory(fd);
 }
 
+/**
+ * visible_vram_if_possible:
+ * @fd: xe device fd
+ * @gt: gt id
+ *
+ * Returns vram memory bitmask for xe device @fd and @gt id or system memory if
+ * there's no vram memory available for @gt. Also attaches the
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM to ensure that CPU access is possible
+ * when using vram.
+ */
+uint64_t visible_vram_if_possible(int fd, int gt)

lib/xe/xe_query.c +20 -0

+
 /**
  * xe_hw_engines:
  * @fd: xe device fd

@@ -536,6 +601,24 @@ uint64_t xe_vram_size(int fd, int gt)
 	return xe_dev->vram_size[gt];
 }
 
+/**
+ * xe_visible_vram_size:
+ * @fd: xe device fd
+ * @gt: gt
+ *
+ * Returns size of visible vram of xe device @fd.
+ */
+uint64_t xe_visible_vram_size(int fd, int gt)
+{
+	uint64_t visible_size;
+
+	visible_size = __xe_visible_vram_size(fd, gt);
+	if (!visible_size) /* older kernel */
+		visible_size = xe_vram_size(fd, gt);
+
+	return visible_size;
+}
+
 /**
  * xe_get_default_alignment:
  * @fd: xe device fd

@@ -552,6 +635,7 @@ xe_dev_FN(xe_get_default_alignment, default_alignment, uint32_t);
  */
 xe_dev_FN(xe_va_bits, va_bits, uint32_t);
 
+
 /**
  * xe_dev_id:
  * @fd: xe device fd

lib/xe/xe_query.h +6 -0

index 68ca5a680..1b74c58ab 100644
@@ -47,6 +47,9 @@ struct xe_device {
 	/** @vram_size: array of vram sizes for all gts */
 	uint64_t *vram_size;
 
+	/** @visible_vram_size: array of visible vram sizes for all gts */
+	uint64_t *visible_vram_size;
+
 	/** @default_alignment: safe alignment regardless region location */
 	uint32_t default_alignment;
 

@@ -80,7 +83,9 @@ unsigned int xe_number_gt(int fd);
 uint64_t all_memory_regions(int fd);
 uint64_t system_memory(int fd);
 uint64_t vram_memory(int fd, int gt);
+uint64_t visible_vram_memory(int fd, int gt);
 uint64_t vram_if_possible(int fd, int gt);
+uint64_t visible_vram_if_possible(int fd, int gt);
 struct drm_xe_engine_class_instance *xe_hw_engines(int fd);
 struct drm_xe_engine_class_instance *xe_hw_engine(int fd, int idx);
 struct drm_xe_query_mem_region *xe_mem_region(int fd, uint64_t region);

@@ -91,6 +96,7 @@ struct drm_xe_query_config *xe_config(int fd);
 unsigned int xe_number_hw_engines(int fd);
 bool xe_has_vram(int fd);
 uint64_t xe_vram_size(int fd, int gt);
+uint64_t xe_visible_vram_size(int fd, int gt);
 uint32_t xe_get_default_alignment(int fd);
 uint32_t xe_va_bits(int fd);
 uint16_t xe_dev_id(int fd);

Quoted 69 lines · Matthew Auld

Add helpers for object creation and querying the cpu_visible related bits.

v2: Make it backwards compat with older kernels

Signed-off-by Matthew Auld <[email protected]>

Cc Gwan-gyeong Mun <[email protected]>

Reviewed-by Gwan-gyeong Mun <[email protected]>

---
 lib/xe/xe_query.c | 86 ++++++++++++++++++++++++++++++++++++++++++++++-
 lib/xe/xe_query.h |  6 ++++
 2 files changed, 91 insertions(+), 1 deletion(-)

lib/xe/xe_query.c +32 -1

index f535ad853..5412fbeb5 100644
@@ -140,6 +140,17 @@ static uint64_t gt_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
 	return 0;
 }
 
+static uint64_t gt_visible_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
+				     const struct drm_xe_query_gts *gts, int gt)
+{
+	int region_idx = ffs(native_region_for_gt(gts, gt)) - 1;
+
+	if (XE_IS_CLASS_VRAM(&mem_usage->regions[region_idx]))
+		return mem_usage->regions[region_idx].cpu_visible_size;
+
+	return 0;
+}
+
 static bool __mem_has_vram(struct drm_xe_query_mem_usage *mem_usage)
 {
 	for (int i = 0; i < mem_usage->num_regions; i++)

@@ -246,9 +257,14 @@ struct xe_device *xe_device_get(int fd)
 	xe_dev->hw_engines = xe_query_engines_new(fd, &xe_dev->number_hw_engines);
 	xe_dev->mem_usage = xe_query_mem_usage_new(fd);
 	xe_dev->vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->vram_size));
-	for (int gt = 0; gt < xe_dev->number_gt; gt++)
+	xe_dev->visible_vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->visible_vram_size));
+	for (int gt = 0; gt < xe_dev->number_gt; gt++) {
 		xe_dev->vram_size[gt] = gt_vram_size(xe_dev->mem_usage,
 						     xe_dev->gts, gt);
+		xe_dev->visible_vram_size[gt] =
+			gt_visible_vram_size(xe_dev->mem_usage,
+					     xe_dev->gts, gt);
+	}
 	xe_dev->default_alignment = __mem_default_alignment(xe_dev->mem_usage);
 	xe_dev->has_vram = __mem_has_vram(xe_dev->mem_usage);
 

@@ -383,6 +399,33 @@ uint64_t vram_memory(int fd, int gt)
 	return xe_has_vram(fd) ? native_region_for_gt(xe_dev->gts, gt) : 0;
 }
 
+static uint64_t __xe_visible_vram_size(int fd, int gt)
+{
+	struct xe_device *xe_dev;
+
+	xe_dev = find_in_cache(fd);
+	igt_assert(xe_dev);
+
+	return xe_dev->visible_vram_size[gt];
+}
+
+/**
+ * visible_vram_memory:
+ * @fd: xe device fd
+ * @gt: gt id
+ *

lib/xe/xe_query.c uint64_t vram_memory(int fd, int gt) · line 426

Quoted 10 lines

lib/xe/xe_query.c +10 -0

+ * Returns vram memory bitmask for xe device @fd and @gt id, with
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM also set, to ensure that CPU access is
+ * possible.
+ */
+uint64_t visible_vram_memory(int fd, int gt)
+{
+	if (__xe_visible_vram_size(fd, gt))
+		return vram_memory(fd, gt) | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
+	else
+		return vram_memory(fd, gt); /* older kernel */

We don't need to maintain any compability with older Xe kernels for now...

Quoted 24 lines

lib/xe/xe_query.c +17 -0

+}
+
 /**
  * vram_if_possible:
  * @fd: xe device fd

@@ -396,6 +439,28 @@ uint64_t vram_if_possible(int fd, int gt)
 	return vram_memory(fd, gt) ?: system_memory(fd);
 }
 
+/**
+ * visible_vram_if_possible:
+ * @fd: xe device fd
+ * @gt: gt id
+ *
+ * Returns vram memory bitmask for xe device @fd and @gt id or system memory if
+ * there's no vram memory available for @gt. Also attaches the
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM to ensure that CPU access is possible
+ * when using vram.
+ */
+uint64_t visible_vram_if_possible(int fd, int gt)
+{
+	uint64_t regions = all_memory_regions(fd);
+	uint64_t system_memory = regions & 0x1;
+	uint64_t vram = regions & (0x2 << gt);

lib/xe/xe_query.c uint64_t vram_if_possible(int fd, int gt) · line 462

Quoted 6 lines

lib/xe/xe_query.c +6 -0

+
+	if (__xe_visible_vram_size(fd, gt))
+		return vram ? vram | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM : system_memory;
+	else
+		return vram ? vram : system_memory; /* older kernel */
+}

so this can be dropped and current callers should use visible_vram_memory().

Quoted 70 lines

lib/xe/xe_query.c +20 -0

+
 /**
  * xe_hw_engines:
  * @fd: xe device fd

@@ -536,6 +601,24 @@ uint64_t xe_vram_size(int fd, int gt)
 	return xe_dev->vram_size[gt];
 }
 
+/**
+ * xe_visible_vram_size:
+ * @fd: xe device fd
+ * @gt: gt
+ *
+ * Returns size of visible vram of xe device @fd.
+ */
+uint64_t xe_visible_vram_size(int fd, int gt)
+{
+	uint64_t visible_size;
+
+	visible_size = __xe_visible_vram_size(fd, gt);
+	if (!visible_size) /* older kernel */
+		visible_size = xe_vram_size(fd, gt);
+
+	return visible_size;
+}
+
 /**
  * xe_get_default_alignment:
  * @fd: xe device fd

@@ -552,6 +635,7 @@ xe_dev_FN(xe_get_default_alignment, default_alignment, uint32_t);
  */
 xe_dev_FN(xe_va_bits, va_bits, uint32_t);
 
+
 /**
  * xe_dev_id:
  * @fd: xe device fd

lib/xe/xe_query.h +6 -0

index 68ca5a680..1b74c58ab 100644
@@ -47,6 +47,9 @@ struct xe_device {
 	/** @vram_size: array of vram sizes for all gts */
 	uint64_t *vram_size;
 
+	/** @visible_vram_size: array of visible vram sizes for all gts */
+	uint64_t *visible_vram_size;
+
 	/** @default_alignment: safe alignment regardless region location */
 	uint32_t default_alignment;
 

@@ -80,7 +83,9 @@ unsigned int xe_number_gt(int fd);
 uint64_t all_memory_regions(int fd);
 uint64_t system_memory(int fd);
 uint64_t vram_memory(int fd, int gt);
+uint64_t visible_vram_memory(int fd, int gt);
 uint64_t vram_if_possible(int fd, int gt);
+uint64_t visible_vram_if_possible(int fd, int gt);
 struct drm_xe_engine_class_instance *xe_hw_engines(int fd);
 struct drm_xe_engine_class_instance *xe_hw_engine(int fd, int idx);
 struct drm_xe_query_mem_region *xe_mem_region(int fd, uint64_t region);

@@ -91,6 +96,7 @@ struct drm_xe_query_config *xe_config(int fd);
 unsigned int xe_number_hw_engines(int fd);
 bool xe_has_vram(int fd);
 uint64_t xe_vram_size(int fd, int gt);
+uint64_t xe_visible_vram_size(int fd, int gt);
 uint32_t xe_get_default_alignment(int fd);
 uint32_t xe_va_bits(int fd);
 uint16_t xe_dev_id(int fd);

Quoted 69 lines · Matthew Auld

Add helpers for object creation and querying the cpu_visible related bits.

v2: Make it backwards compat with older kernels

Signed-off-by Matthew Auld <[email protected]>

Cc Gwan-gyeong Mun <[email protected]>

Reviewed-by Gwan-gyeong Mun <[email protected]>

---
  lib/xe/xe_query.c | 86 ++++++++++++++++++++++++++++++++++++++++++++++-
  lib/xe/xe_query.h |  6 ++++
  2 files changed, 91 insertions(+), 1 deletion(-)

lib/xe/xe_query.c +32 -1

index f535ad853..5412fbeb5 100644
@@ -140,6 +140,17 @@ static uint64_t gt_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
  	return 0;
  }
 
+static uint64_t gt_visible_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
+				     const struct drm_xe_query_gts *gts, int gt)
+{
+	int region_idx = ffs(native_region_for_gt(gts, gt)) - 1;
+
+	if (XE_IS_CLASS_VRAM(&mem_usage->regions[region_idx]))
+		return mem_usage->regions[region_idx].cpu_visible_size;
+
+	return 0;
+}
+
  static bool __mem_has_vram(struct drm_xe_query_mem_usage *mem_usage)
  {
  	for (int i = 0; i < mem_usage->num_regions; i++)

@@ -246,9 +257,14 @@ struct xe_device *xe_device_get(int fd)
  	xe_dev->hw_engines = xe_query_engines_new(fd, &xe_dev->number_hw_engines);
  	xe_dev->mem_usage = xe_query_mem_usage_new(fd);
  	xe_dev->vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->vram_size));
-	for (int gt = 0; gt < xe_dev->number_gt; gt++)
+	xe_dev->visible_vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->visible_vram_size));
+	for (int gt = 0; gt < xe_dev->number_gt; gt++) {
  		xe_dev->vram_size[gt] = gt_vram_size(xe_dev->mem_usage,
  						     xe_dev->gts, gt);
+		xe_dev->visible_vram_size[gt] =
+			gt_visible_vram_size(xe_dev->mem_usage,
+					     xe_dev->gts, gt);
+	}
  	xe_dev->default_alignment = __mem_default_alignment(xe_dev->mem_usage);
  	xe_dev->has_vram = __mem_has_vram(xe_dev->mem_usage);
 

@@ -383,6 +399,33 @@ uint64_t vram_memory(int fd, int gt)
  	return xe_has_vram(fd) ? native_region_for_gt(xe_dev->gts, gt) : 0;
  }
 
+static uint64_t __xe_visible_vram_size(int fd, int gt)
+{
+	struct xe_device *xe_dev;
+
+	xe_dev = find_in_cache(fd);
+	igt_assert(xe_dev);
+
+	return xe_dev->visible_vram_size[gt];
+}
+
+/**
+ * visible_vram_memory:
+ * @fd: xe device fd
+ * @gt: gt id
+ *

lib/xe/xe_query.c uint64_t vram_memory(int fd, int gt) · line 426

Quoted 10 lines

lib/xe/xe_query.c +10 -0

+ * Returns vram memory bitmask for xe device @fd and @gt id, with
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM also set, to ensure that CPU access is
+ * possible.
+ */
+uint64_t visible_vram_memory(int fd, int gt)
+{
+	if (__xe_visible_vram_size(fd, gt))
+		return vram_memory(fd, gt) | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
+	else
+		return vram_memory(fd, gt); /* older kernel */

We don't need to maintain any compability with older Xe kernels for now...

Quoted 24 lines

lib/xe/xe_query.c +17 -0

+}
+
  /**
   * vram_if_possible:
   * @fd: xe device fd

@@ -396,6 +439,28 @@ uint64_t vram_if_possible(int fd, int gt)
  	return vram_memory(fd, gt) ?: system_memory(fd);
  }
 
+/**
+ * visible_vram_if_possible:
+ * @fd: xe device fd
+ * @gt: gt id
+ *
+ * Returns vram memory bitmask for xe device @fd and @gt id or system memory if
+ * there's no vram memory available for @gt. Also attaches the
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM to ensure that CPU access is possible
+ * when using vram.
+ */
+uint64_t visible_vram_if_possible(int fd, int gt)
+{
+	uint64_t regions = all_memory_regions(fd);
+	uint64_t system_memory = regions & 0x1;
+	uint64_t vram = regions & (0x2 << gt);

lib/xe/xe_query.c uint64_t vram_if_possible(int fd, int gt) · line 462

Quoted 6 lines

lib/xe/xe_query.c +6 -0

+
+	if (__xe_visible_vram_size(fd, gt))
+		return vram ? vram | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM : system_memory;
+	else
+		return vram ? vram : system_memory; /* older kernel */
+}

so this can be dropped and current callers should use visible_vram_memory().

Quoted 70 lines

lib/xe/xe_query.c +20 -0

+
  /**
   * xe_hw_engines:
   * @fd: xe device fd

@@ -536,6 +601,24 @@ uint64_t xe_vram_size(int fd, int gt)
  	return xe_dev->vram_size[gt];
  }
 
+/**
+ * xe_visible_vram_size:
+ * @fd: xe device fd
+ * @gt: gt
+ *
+ * Returns size of visible vram of xe device @fd.
+ */
+uint64_t xe_visible_vram_size(int fd, int gt)
+{
+	uint64_t visible_size;
+
+	visible_size = __xe_visible_vram_size(fd, gt);
+	if (!visible_size) /* older kernel */
+		visible_size = xe_vram_size(fd, gt);
+
+	return visible_size;
+}
+
  /**
   * xe_get_default_alignment:
   * @fd: xe device fd

@@ -552,6 +635,7 @@ xe_dev_FN(xe_get_default_alignment, default_alignment, uint32_t);
   */
  xe_dev_FN(xe_va_bits, va_bits, uint32_t);
 
+
  /**
   * xe_dev_id:
   * @fd: xe device fd

lib/xe/xe_query.h +6 -0

index 68ca5a680..1b74c58ab 100644
@@ -47,6 +47,9 @@ struct xe_device {
  	/** @vram_size: array of vram sizes for all gts */
  	uint64_t *vram_size;
 
+	/** @visible_vram_size: array of visible vram sizes for all gts */
+	uint64_t *visible_vram_size;
+
  	/** @default_alignment: safe alignment regardless region location */
  	uint32_t default_alignment;
 

@@ -80,7 +83,9 @@ unsigned int xe_number_gt(int fd);
  uint64_t all_memory_regions(int fd);
  uint64_t system_memory(int fd);
  uint64_t vram_memory(int fd, int gt);
+uint64_t visible_vram_memory(int fd, int gt);
  uint64_t vram_if_possible(int fd, int gt);
+uint64_t visible_vram_if_possible(int fd, int gt);
  struct drm_xe_engine_class_instance *xe_hw_engines(int fd);
  struct drm_xe_engine_class_instance *xe_hw_engine(int fd, int idx);
  struct drm_xe_query_mem_region *xe_mem_region(int fd, uint64_t region);

@@ -91,6 +96,7 @@ struct drm_xe_query_config *xe_config(int fd);
  unsigned int xe_number_hw_engines(int fd);
  bool xe_has_vram(int fd);
  uint64_t xe_vram_size(int fd, int gt);
+uint64_t xe_visible_vram_size(int fd, int gt);
  uint32_t xe_get_default_alignment(int fd);
  uint32_t xe_va_bits(int fd);
  uint16_t xe_dev_id(int fd);

Quoted 107 lines · Jose

Quoted 69 lines · Matthew Auld

Add helpers for object creation and querying the cpu_visible related bits.

v2: Make it backwards compat with older kernels

Signed-off-by Matthew Auld <[email protected]>

Cc Gwan-gyeong Mun <[email protected]>

Reviewed-by Gwan-gyeong Mun <[email protected]>

---
  lib/xe/xe_query.c | 86 ++++++++++++++++++++++++++++++++++++++++++++++-
  lib/xe/xe_query.h |  6 ++++
  2 files changed, 91 insertions(+), 1 deletion(-)

lib/xe/xe_query.c +32 -1

index f535ad853..5412fbeb5 100644
@@ -140,6 +140,17 @@ static uint64_t gt_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
  	return 0;
  }
 
+static uint64_t gt_visible_vram_size(const struct drm_xe_query_mem_usage *mem_usage,
+				     const struct drm_xe_query_gts *gts, int gt)
+{
+	int region_idx = ffs(native_region_for_gt(gts, gt)) - 1;
+
+	if (XE_IS_CLASS_VRAM(&mem_usage->regions[region_idx]))
+		return mem_usage->regions[region_idx].cpu_visible_size;
+
+	return 0;
+}
+
  static bool __mem_has_vram(struct drm_xe_query_mem_usage *mem_usage)
  {
  	for (int i = 0; i < mem_usage->num_regions; i++)

@@ -246,9 +257,14 @@ struct xe_device *xe_device_get(int fd)
  	xe_dev->hw_engines = xe_query_engines_new(fd, &xe_dev->number_hw_engines);
  	xe_dev->mem_usage = xe_query_mem_usage_new(fd);
  	xe_dev->vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->vram_size));
-	for (int gt = 0; gt < xe_dev->number_gt; gt++)
+	xe_dev->visible_vram_size = calloc(xe_dev->number_gt, sizeof(*xe_dev->visible_vram_size));
+	for (int gt = 0; gt < xe_dev->number_gt; gt++) {
  		xe_dev->vram_size[gt] = gt_vram_size(xe_dev->mem_usage,
  						     xe_dev->gts, gt);
+		xe_dev->visible_vram_size[gt] =
+			gt_visible_vram_size(xe_dev->mem_usage,
+					     xe_dev->gts, gt);
+	}
  	xe_dev->default_alignment = __mem_default_alignment(xe_dev->mem_usage);
  	xe_dev->has_vram = __mem_has_vram(xe_dev->mem_usage);
 

@@ -383,6 +399,33 @@ uint64_t vram_memory(int fd, int gt)
  	return xe_has_vram(fd) ? native_region_for_gt(xe_dev->gts, gt) : 0;
  }
 
+static uint64_t __xe_visible_vram_size(int fd, int gt)
+{
+	struct xe_device *xe_dev;
+
+	xe_dev = find_in_cache(fd);
+	igt_assert(xe_dev);
+
+	return xe_dev->visible_vram_size[gt];
+}
+
+/**
+ * visible_vram_memory:
+ * @fd: xe device fd
+ * @gt: gt id
+ *

lib/xe/xe_query.c uint64_t vram_memory(int fd, int gt) · line 426

Quoted 10 lines

lib/xe/xe_query.c +10 -0

+ * Returns vram memory bitmask for xe device @fd and @gt id, with
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM also set, to ensure that CPU access is
+ * possible.
+ */
+uint64_t visible_vram_memory(int fd, int gt)
+{
+	if (__xe_visible_vram_size(fd, gt))
+		return vram_memory(fd, gt) | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
+	else
+		return vram_memory(fd, gt); /* older kernel */

We don't need to maintain any compability with older Xe kernels for now...

Quoted 24 lines

lib/xe/xe_query.c +17 -0

+}
+
  /**
   * vram_if_possible:
   * @fd: xe device fd

@@ -396,6 +439,28 @@ uint64_t vram_if_possible(int fd, int gt)
  	return vram_memory(fd, gt) ?: system_memory(fd);
  }
 
+/**
+ * visible_vram_if_possible:
+ * @fd: xe device fd
+ * @gt: gt id
+ *
+ * Returns vram memory bitmask for xe device @fd and @gt id or system memory if
+ * there's no vram memory available for @gt. Also attaches the
+ * XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM to ensure that CPU access is possible
+ * when using vram.
+ */
+uint64_t visible_vram_if_possible(int fd, int gt)
+{
+	uint64_t regions = all_memory_regions(fd);
+	uint64_t system_memory = regions & 0x1;
+	uint64_t vram = regions & (0x2 << gt);

Quoted 8 lines

lib/xe/xe_query.c uint64_t vram_if_possible(int fd, int gt) · line 462

Quoted 6 lines

lib/xe/xe_query.c +6 -0

+
+	if (__xe_visible_vram_size(fd, gt))
+		return vram ? vram | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM : system_memory;
+	else
+		return vram ? vram : system_memory; /* older kernel */
+}

so this can be dropped and current callers should use visible_vram_memory().

Here I just wanted to keep it compatible so I could merge the IGT in any order. Ideally I want to merge the IGT first and let that soak for a couple of hours, and assuming there is no new breakage I can re-send the kernel patches with a hack patch to force small-bar on some CI machine to get pre-merge testing. Once that all passes I can merge the kernel side. The above compatibility stuff can then be removed after.

Quoted 72 lines

Quoted 71 lines

lib/xe/xe_query.c +20 -0

+
  /**
   * xe_hw_engines:
   * @fd: xe device fd

@@ -536,6 +601,24 @@ uint64_t xe_vram_size(int fd, int gt)
  	return xe_dev->vram_size[gt];
  }
 
+/**
+ * xe_visible_vram_size:
+ * @fd: xe device fd
+ * @gt: gt
+ *
+ * Returns size of visible vram of xe device @fd.
+ */
+uint64_t xe_visible_vram_size(int fd, int gt)
+{
+	uint64_t visible_size;
+
+	visible_size = __xe_visible_vram_size(fd, gt);
+	if (!visible_size) /* older kernel */
+		visible_size = xe_vram_size(fd, gt);
+
+	return visible_size;
+}
+
  /**
   * xe_get_default_alignment:
   * @fd: xe device fd

@@ -552,6 +635,7 @@ xe_dev_FN(xe_get_default_alignment, default_alignment, uint32_t);
   */
  xe_dev_FN(xe_va_bits, va_bits, uint32_t);
 
+
  /**
   * xe_dev_id:
   * @fd: xe device fd

lib/xe/xe_query.h +6 -0

index 68ca5a680..1b74c58ab 100644
@@ -47,6 +47,9 @@ struct xe_device {
  	/** @vram_size: array of vram sizes for all gts */
  	uint64_t *vram_size;
 
+	/** @visible_vram_size: array of visible vram sizes for all gts */
+	uint64_t *visible_vram_size;
+
  	/** @default_alignment: safe alignment regardless region location */
  	uint32_t default_alignment;
 

@@ -80,7 +83,9 @@ unsigned int xe_number_gt(int fd);
  uint64_t all_memory_regions(int fd);
  uint64_t system_memory(int fd);
  uint64_t vram_memory(int fd, int gt);
+uint64_t visible_vram_memory(int fd, int gt);
  uint64_t vram_if_possible(int fd, int gt);
+uint64_t visible_vram_if_possible(int fd, int gt);
  struct drm_xe_engine_class_instance *xe_hw_engines(int fd);
  struct drm_xe_engine_class_instance *xe_hw_engine(int fd, int idx);
  struct drm_xe_query_mem_region *xe_mem_region(int fd, uint64_t region);

@@ -91,6 +96,7 @@ struct drm_xe_query_config *xe_config(int fd);
  unsigned int xe_number_hw_engines(int fd);
  bool xe_has_vram(int fd);
  uint64_t xe_vram_size(int fd, int gt);
+uint64_t xe_visible_vram_size(int fd, int gt);
  uint32_t xe_get_default_alignment(int fd);
  uint32_t xe_va_bits(int fd);
  uint16_t xe_dev_id(int fd);

Convert all the existing tests that require CPU access.

v2:
- Split out the lib changes
- Prefer SZ_256M and SZ_1G in xe_evict
- Simplify and fix the bo_flags handling in test_exec
v3:
- Small fix in xe_evict conversion (missing system_memory(fd))

Signed-off-by Matthew Auld <[email protected]>

Cc José Roberto de Souza <[email protected]>

Cc Kamil Konieczny <[email protected]>

Cc Gwan-gyeong Mun <[email protected]>

---
 tests/xe/xe_dma_buf_sync.c      |  3 ++-
 tests/xe/xe_evict.c             | 34 +++++++++++++++++++++------------
 tests/xe/xe_exec_balancer.c     |  6 +++---
 tests/xe/xe_exec_basic.c        | 17 ++++++++---------
 tests/xe/xe_exec_compute_mode.c |  4 ++--
 tests/xe/xe_exec_fault_mode.c   | 12 ++++++++----
 tests/xe/xe_exec_reset.c        | 13 ++++++++-----
 tests/xe/xe_exec_store.c        |  6 ++++--
 tests/xe/xe_exec_threads.c      |  9 ++++++---
 tests/xe/xe_guc_pc.c            |  3 ++-
 tests/xe/xe_mmap.c              |  4 ++--
 tests/xe/xe_pm.c                |  3 ++-
 tests/xe/xe_prime_self_import.c |  8 ++++----
 tests/xe/xe_vm.c                | 23 +++++++++++++++-------
 14 files changed, 89 insertions(+), 56 deletions(-)

tests/xe/xe_dma_buf_sync.c +2 -1

index c08f8ac18..4e76d85ab 100644
@@ -120,7 +120,8 @@ test_export_dma_buf(struct drm_xe_engine_class_instance *hwe0,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd[0]),
 			xe_get_default_alignment(fd[0]));
 	for (i = 0; i < n_bo; ++i) {
-		bo[i] = xe_bo_create(fd[0], hwe0->gt_id, 0, bo_size);
+		bo[i] = xe_bo_create_flags(fd[0], 0, bo_size,
+					   visible_vram_if_possible(fd[0], hwe0->gt_id));
 		dma_buf_fd[i] = prime_handle_to_fd(fd[0], bo[i]);
 		import_bo[i] = prime_fd_to_handle(fd[1], dma_buf_fd[i]);
 

tests/xe/xe_evict.c +22 -12

index 1a70f1b45..c44cb80dc 100644
@@ -97,15 +97,17 @@ test_evict(int fd, struct drm_xe_engine_class_instance *eci,
                                 i < n_execs / 8 ? 0 : vm;
 
 			if (flags & MULTI_VM) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id, 0,
-							    bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, 0,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else if (flags & THREADED) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id, vm,
-							    bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, vm,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else {
 				__bo = bo[i] = xe_bo_create_flags(fd, _vm,
 								  bo_size,
-								  vram_memory(fd, eci->gt_id) |
+								  visible_vram_memory(fd, eci->gt_id) |
 								  system_memory(fd));
 			}
 		} else {

@@ -278,15 +280,17 @@ test_evict_cm(int fd, struct drm_xe_engine_class_instance *eci,
                                 i < n_execs / 8 ? 0 : vm;
 
 			if (flags & MULTI_VM) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id,
-							    0, bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, 0,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else if (flags & THREADED) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id,
-							    vm, bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, vm,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else {
 				__bo = bo[i] = xe_bo_create_flags(fd, _vm,
 								  bo_size,
-								  vram_memory(fd, eci->gt_id) |
+								  visible_vram_memory(fd, eci->gt_id) |
 								  system_memory(fd));
 			}
 		} else {

@@ -449,9 +453,15 @@ threads(int fd, struct drm_xe_engine_class_instance *eci,
 		pthread_join(threads_data[i].thread, NULL);
 }
 
+#define SZ_256M 0x10000000
+#define SZ_1G   0x40000000
+
 static uint64_t calc_bo_size(uint64_t vram_size, int mul, int div)
 {
-	return (ALIGN(vram_size, 0x40000000)  * mul) / div;
+	if (vram_size >= SZ_1G)
+		return (ALIGN(vram_size, SZ_1G)  * mul) / div;
+	else
+		return (ALIGN(vram_size, SZ_256M)  * mul) / div; /* small-bar */
 }
 
 /**

@@ -664,7 +674,7 @@ igt_main
 	igt_fixture {
 		fd = drm_open_driver(DRIVER_XE);
 		igt_require(xe_has_vram(fd));
-		vram_size = xe_vram_size(fd, 0);
+		vram_size = xe_visible_vram_size(fd, 0);
 		igt_assert(vram_size);
 
 		xe_for_each_hw_engine(fd, hwe)

tests/xe/xe_exec_balancer.c +3 -3

index 8df6ceba8..0b00d93de 100644
@@ -69,7 +69,7 @@ static void test_all_active(int fd, int gt, int class)
 	bo_size = sizeof(*data) * num_placements;
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd), xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, gt, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < num_placements; i++) {

@@ -225,7 +225,7 @@ test_exec(int fd, int gt, int class, int n_engines, int n_execs,
 		}
 		memset(data, 0, bo_size);
 	} else {
-		bo = xe_bo_create(fd, gt, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 

@@ -447,7 +447,7 @@ test_cm(int fd, int gt, int class, int n_engines, int n_execs,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, gt, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

tests/xe/xe_exec_basic.c +8 -9

index af581c327..a4bae93f0 100644
@@ -126,15 +126,14 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
 		}
 		memset(data, 0, bo_size);
 	} else {
-		if (flags & DEFER_ALLOC) {
-			bo = xe_bo_create_flags(fd, n_vm == 1 ? vm[0] : 0,
-						bo_size,
-						vram_if_possible(fd, eci->gt_id) |
-						XE_GEM_CREATE_FLAG_DEFER_BACKING);
-		} else {
-			bo = xe_bo_create(fd, eci->gt_id, n_vm == 1 ? vm[0] : 0,
-					  bo_size);
-		}
+		uint32_t bo_flags;
+
+		bo_flags = visible_vram_if_possible(fd, eci->gt_id);
+		if (flags & DEFER_ALLOC)
+			bo_flags |= XE_GEM_CREATE_FLAG_DEFER_BACKING;
+
+		bo = xe_bo_create_flags(fd, n_vm == 1 ? vm[0] : 0,
+					bo_size, bo_flags);
 		if (!(flags & DEFER_BIND))
 			data = xe_bo_map(fd, bo, bo_size);
 	}

tests/xe/xe_exec_compute_mode.c +2 -2

index 27b538414..ee9756c21 100644
@@ -150,8 +150,8 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, flags & VM_FOR_BO ? vm : 0,
-				  bo_size);
+		bo = xe_bo_create_flags(fd, flags & VM_FOR_BO ? vm : 0,
+					bo_size, visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

tests/xe/xe_exec_fault_mode.c +8 -4

index bf7230c5a..7dcbb3c45 100644
@@ -153,9 +153,11 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
 	} else {
 		if (flags & PREFETCH)
 			bo = xe_bo_create_flags(fd, 0, bo_size,
-						all_memory_regions(fd));
+						all_memory_regions(fd) |
+						visible_vram_if_possible(fd, 0));
 		else
-			bo = xe_bo_create(fd, eci->gt_id, 0, bo_size);
+			bo = xe_bo_create_flags(fd, 0, bo_size,
+						visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

@@ -382,8 +384,10 @@ test_atomic(int fd, struct drm_xe_engine_class_instance *eci,
 	addr_wait = addr + bo_size;
 
 	bo = xe_bo_create_flags(fd, vm, bo_size,
-				all_memory_regions(fd));
-	bo_wait = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+				all_memory_regions(fd) |
+				visible_vram_if_possible(fd, 0));
+	bo_wait = xe_bo_create_flags(fd, vm, bo_size,
+				     visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 	wait = xe_bo_map(fd, bo_wait, bo_size);
 	ptr = &data[0].data;

tests/xe/xe_exec_reset.c +8 -5

index 6ca1cd769..dfbaa6035 100644
@@ -50,7 +50,8 @@ static void test_spin(int fd, struct drm_xe_engine_class_instance *eci)
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	spin = xe_bo_map(fd, bo, bo_size);
 
 	engine = xe_engine_create(fd, vm, eci, 0);

@@ -187,7 +188,7 @@ test_balancer(int fd, int gt, int class, int n_engines, int n_execs,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, gt, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

@@ -379,7 +380,8 @@ test_legacy_mode(int fd, struct drm_xe_engine_class_instance *eci,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

@@ -550,7 +552,8 @@ test_compute_mode(int fd, struct drm_xe_engine_class_instance *eci,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 	memset(data, 0, bo_size);
 

@@ -682,7 +685,7 @@ static void submit_jobs(struct gt_thread_data *t)
 	uint32_t bo;
 	uint32_t *data;
 
-	bo = xe_bo_create(fd, 0, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, 0));
 	data = xe_bo_map(fd, bo, bo_size);
 	data[0] = MI_BATCH_BUFFER_END;
 

tests/xe/xe_exec_store.c +4 -2

index 9640b1567..ab1bde36e 100644
@@ -82,7 +82,8 @@ static void store(int fd)
 			xe_get_default_alignment(fd));
 
 	hw_engine = xe_hw_engine(fd, 1);
-	bo = xe_bo_create(fd, hw_engine->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, hw_engine->gt_id));
 
 	xe_vm_bind_async(fd, vm, hw_engine->gt_id, bo, 0, addr, bo_size, &sync, 1);
 	data = xe_bo_map(fd, bo, bo_size);

@@ -138,7 +139,8 @@ static void store_all(int fd, int gt, int class)
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, 0, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, 0));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	xe_for_each_hw_engine(fd, hwe) {

tests/xe/xe_exec_threads.c +6 -3

index 414d8ee9a..396398984 100644
@@ -106,7 +106,8 @@ test_balancer(int fd, int gt, uint32_t vm, uint64_t addr, uint64_t userptr,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, gt, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, gt));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

@@ -306,7 +307,8 @@ test_compute_mode(int fd, uint32_t vm, uint64_t addr, uint64_t userptr,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, 0, bo_size);
+		bo = xe_bo_create_flags(fd, 0, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

@@ -516,7 +518,8 @@ test_legacy_mode(int fd, uint32_t vm, uint64_t addr, uint64_t userptr,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

tests/xe/xe_guc_pc.c +2 -1

index c34df8d60..6339b3893 100644
@@ -64,7 +64,8 @@ static void exec_basic(int fd, struct drm_xe_engine_class_instance *eci,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

tests/xe/xe_mmap.c +2 -2

index 798facca9..583f39d7a 100644
@@ -118,10 +118,10 @@ igt_main
 		test_mmap(fd, system_memory(fd));
 
 	igt_subtest("vram")
-		test_mmap(fd, vram_memory(fd, 0));
+		test_mmap(fd, visible_vram_memory(fd, 0));
 
 	igt_subtest("vram-system")
-		test_mmap(fd, vram_memory(fd, 0) | system_memory(fd));
+		test_mmap(fd, visible_vram_memory(fd, 0) | system_memory(fd));
 
 	igt_subtest("bad-flags")
 		test_bad_flags(fd);

tests/xe/xe_pm.c +2 -1

index a7f73c4e6..559eccdeb 100644
@@ -254,7 +254,8 @@ test_exec(device_t device, struct drm_xe_engine_class_instance *eci,
 	if (check_rpm && runtime_usage_available(device.pci_xe))
 		rpm_usage = igt_pm_get_runtime_usage(device.pci_xe);
 
-	bo = xe_bo_create(device.fd_xe, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(device.fd_xe, vm, bo_size,
+				visible_vram_if_possible(device.fd_xe, eci->gt_id));
 	data = xe_bo_map(device.fd_xe, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

tests/xe/xe_prime_self_import.c +4 -4

index 0fd79f704..e712e2a9c 100644
@@ -107,7 +107,7 @@ static void test_with_fd_dup(void)
 	fd1 = drm_open_driver(DRIVER_XE);
 	fd2 = drm_open_driver(DRIVER_XE);
 
-	handle = xe_bo_create(fd1, 0, 0, BO_SIZE);
+	handle = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
 
 	dma_buf_fd1 = prime_handle_to_fd(fd1, handle);
 	gem_close(fd1, handle);

@@ -141,8 +141,8 @@ static void test_with_two_bos(void)
 	fd1 = drm_open_driver(DRIVER_XE);
 	fd2 = drm_open_driver(DRIVER_XE);
 
-	handle1 = xe_bo_create(fd1, 0, 0, BO_SIZE);
-	handle2 = xe_bo_create(fd1, 0, 0, BO_SIZE);
+	handle1 = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
+	handle2 = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
 
 	dma_buf_fd = prime_handle_to_fd(fd1, handle1);
 	handle_import = prime_fd_to_handle(fd2, dma_buf_fd);

@@ -211,7 +211,7 @@ static void test_with_one_bo(void)
 	fd1 = drm_open_driver(DRIVER_XE);
 	fd2 = drm_open_driver(DRIVER_XE);
 
-	handle = xe_bo_create(fd1, 0, 0, BO_SIZE);
+	handle = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
 
 	dma_buf_fd = prime_handle_to_fd(fd1, handle);
 	handle_import1 = prime_fd_to_handle(fd2, dma_buf_fd);

tests/xe/xe_vm.c +9 -5

index 04d6c3956..982c50f6d 100644
@@ -52,7 +52,8 @@ write_dwords(int fd, uint32_t vm, int n_dwords, uint64_t *addrs)
 	batch_size = (n_dwords * 4 + 1) * sizeof(uint32_t);
 	batch_size = ALIGN(batch_size + xe_cs_prefetch_size(fd),
 			   xe_get_default_alignment(fd));
-	batch_bo = xe_bo_create(fd, 0, vm, batch_size);
+	batch_bo = xe_bo_create_flags(fd, vm, batch_size,
+				      visible_vram_if_possible(fd, 0));
 	batch_map = xe_bo_map(fd, batch_bo, batch_size);
 
 	for (i = 0; i < n_dwords; i++) {

@@ -116,7 +117,7 @@ __test_bind_one_bo(int fd, uint32_t vm, int n_addrs, uint64_t *addrs)
 		vms = malloc(sizeof(*vms) * n_addrs);
 		igt_assert(vms);
 	}
-	bo = xe_bo_create(fd, 0, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, 0));
 	map = xe_bo_map(fd, bo, bo_size);
 	memset(map, 0, bo_size);
 

@@ -554,7 +555,8 @@ shared_pte_page(int fd, struct drm_xe_engine_class_instance *eci, int n_bo,
 			xe_get_default_alignment(fd));
 
 	for (i = 0; i < n_bo; ++i) {
-		bo[i] = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		bo[i] = xe_bo_create_flags(fd, vm, bo_size,
+					   visible_vram_if_possible(fd, eci->gt_id));
 		data[i] = xe_bo_map(fd, bo[i], bo_size);
 	}
 

@@ -723,7 +725,8 @@ test_bind_engines_independent(int fd, struct drm_xe_engine_class_instance *eci)
 	bo_size = sizeof(*data) * N_ENGINES;
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < N_ENGINES; i++) {

@@ -880,7 +883,8 @@ test_bind_array(int fd, struct drm_xe_engine_class_instance *eci, int n_execs,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	if (flags & BIND_ARRAY_BIND_ENGINE_FLAG)

tests/xe/xe_vm.c +5 -1

+
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		map = xe_bo_map(fd, bo, bo_size);
 	}
 

@@ -1350,7 +1358,8 @@ test_munmap_style_unbind(int fd, struct drm_xe_engine_class_instance *eci,
 			    MAP_ANONYMOUS, -1, 0);
 		igt_assert(map != MAP_FAILED);
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		map = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(map, 0, bo_size);

Quoted 2 lines · Matthew Auld

Convert all the existing tests that require CPU access.

Missing conversion in test_bad_flags(), test_bad_extensions(), test_bad_object(), test_with_one_bo_two_files(), test_reimport_close_race(), thread_fn_export_vs_close(), test_llseek_size(), test_llseek_bad().

The mmaps tests are expected to fail but not because region is not cpu mmaped.
For the prime_self_import test it is also good to require visible vram even if the test don't do any operation with CPU, in real world applications you need to give applications the flexibility to access it in CPU as well.

Quoted 463 lines

v2:
- Split out the lib changes
- Prefer SZ_256M and SZ_1G in xe_evict
- Simplify and fix the bo_flags handling in test_exec
v3:
- Small fix in xe_evict conversion (missing system_memory(fd))

Signed-off-by Matthew Auld <[email protected]>

Cc José Roberto de Souza <[email protected]>

Cc Kamil Konieczny <[email protected]>

Cc Gwan-gyeong Mun <[email protected]>

---
 tests/xe/xe_dma_buf_sync.c      |  3 ++-
 tests/xe/xe_evict.c             | 34 +++++++++++++++++++++------------
 tests/xe/xe_exec_balancer.c     |  6 +++---
 tests/xe/xe_exec_basic.c        | 17 ++++++++---------
 tests/xe/xe_exec_compute_mode.c |  4 ++--
 tests/xe/xe_exec_fault_mode.c   | 12 ++++++++----
 tests/xe/xe_exec_reset.c        | 13 ++++++++-----
 tests/xe/xe_exec_store.c        |  6 ++++--
 tests/xe/xe_exec_threads.c      |  9 ++++++---
 tests/xe/xe_guc_pc.c            |  3 ++-
 tests/xe/xe_mmap.c              |  4 ++--
 tests/xe/xe_pm.c                |  3 ++-
 tests/xe/xe_prime_self_import.c |  8 ++++----
 tests/xe/xe_vm.c                | 23 +++++++++++++++-------
 14 files changed, 89 insertions(+), 56 deletions(-)

tests/xe/xe_dma_buf_sync.c +2 -1

index c08f8ac18..4e76d85ab 100644
@@ -120,7 +120,8 @@ test_export_dma_buf(struct drm_xe_engine_class_instance *hwe0,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd[0]),
 			xe_get_default_alignment(fd[0]));
 	for (i = 0; i < n_bo; ++i) {
-		bo[i] = xe_bo_create(fd[0], hwe0->gt_id, 0, bo_size);
+		bo[i] = xe_bo_create_flags(fd[0], 0, bo_size,
+					   visible_vram_if_possible(fd[0], hwe0->gt_id));
 		dma_buf_fd[i] = prime_handle_to_fd(fd[0], bo[i]);
 		import_bo[i] = prime_fd_to_handle(fd[1], dma_buf_fd[i]);
 

tests/xe/xe_evict.c +22 -12

index 1a70f1b45..c44cb80dc 100644
@@ -97,15 +97,17 @@ test_evict(int fd, struct drm_xe_engine_class_instance *eci,
                                 i < n_execs / 8 ? 0 : vm;
 
 			if (flags & MULTI_VM) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id, 0,
-							    bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, 0,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else if (flags & THREADED) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id, vm,
-							    bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, vm,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else {
 				__bo = bo[i] = xe_bo_create_flags(fd, _vm,
 								  bo_size,
-								  vram_memory(fd, eci->gt_id) |
+								  visible_vram_memory(fd, eci->gt_id) |
 								  system_memory(fd));
 			}
 		} else {

@@ -278,15 +280,17 @@ test_evict_cm(int fd, struct drm_xe_engine_class_instance *eci,
                                 i < n_execs / 8 ? 0 : vm;
 
 			if (flags & MULTI_VM) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id,
-							    0, bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, 0,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else if (flags & THREADED) {
-				__bo = bo[i] = xe_bo_create(fd, eci->gt_id,
-							    vm, bo_size);
+				__bo = bo[i] = xe_bo_create_flags(fd, vm,
+								  bo_size,
+								  visible_vram_memory(fd, eci->gt_id));
 			} else {
 				__bo = bo[i] = xe_bo_create_flags(fd, _vm,
 								  bo_size,
-								  vram_memory(fd, eci->gt_id) |
+								  visible_vram_memory(fd, eci->gt_id) |
 								  system_memory(fd));
 			}
 		} else {

@@ -449,9 +453,15 @@ threads(int fd, struct drm_xe_engine_class_instance *eci,
 		pthread_join(threads_data[i].thread, NULL);
 }
 
+#define SZ_256M 0x10000000
+#define SZ_1G   0x40000000
+
 static uint64_t calc_bo_size(uint64_t vram_size, int mul, int div)
 {
-	return (ALIGN(vram_size, 0x40000000)  * mul) / div;
+	if (vram_size >= SZ_1G)
+		return (ALIGN(vram_size, SZ_1G)  * mul) / div;
+	else
+		return (ALIGN(vram_size, SZ_256M)  * mul) / div; /* small-bar */
 }
 
 /**

@@ -664,7 +674,7 @@ igt_main
 	igt_fixture {
 		fd = drm_open_driver(DRIVER_XE);
 		igt_require(xe_has_vram(fd));
-		vram_size = xe_vram_size(fd, 0);
+		vram_size = xe_visible_vram_size(fd, 0);
 		igt_assert(vram_size);
 
 		xe_for_each_hw_engine(fd, hwe)

tests/xe/xe_exec_balancer.c +3 -3

index 8df6ceba8..0b00d93de 100644
@@ -69,7 +69,7 @@ static void test_all_active(int fd, int gt, int class)
 	bo_size = sizeof(*data) * num_placements;
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd), xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, gt, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < num_placements; i++) {

@@ -225,7 +225,7 @@ test_exec(int fd, int gt, int class, int n_engines, int n_execs,
 		}
 		memset(data, 0, bo_size);
 	} else {
-		bo = xe_bo_create(fd, gt, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 

@@ -447,7 +447,7 @@ test_cm(int fd, int gt, int class, int n_engines, int n_execs,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, gt, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

tests/xe/xe_exec_basic.c +8 -9

index af581c327..a4bae93f0 100644
@@ -126,15 +126,14 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
 		}
 		memset(data, 0, bo_size);
 	} else {
-		if (flags & DEFER_ALLOC) {
-			bo = xe_bo_create_flags(fd, n_vm == 1 ? vm[0] : 0,
-						bo_size,
-						vram_if_possible(fd, eci->gt_id) |
-						XE_GEM_CREATE_FLAG_DEFER_BACKING);
-		} else {
-			bo = xe_bo_create(fd, eci->gt_id, n_vm == 1 ? vm[0] : 0,
-					  bo_size);
-		}
+		uint32_t bo_flags;
+
+		bo_flags = visible_vram_if_possible(fd, eci->gt_id);
+		if (flags & DEFER_ALLOC)
+			bo_flags |= XE_GEM_CREATE_FLAG_DEFER_BACKING;
+
+		bo = xe_bo_create_flags(fd, n_vm == 1 ? vm[0] : 0,
+					bo_size, bo_flags);
 		if (!(flags & DEFER_BIND))
 			data = xe_bo_map(fd, bo, bo_size);
 	}

tests/xe/xe_exec_compute_mode.c +2 -2

index 27b538414..ee9756c21 100644
@@ -150,8 +150,8 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, flags & VM_FOR_BO ? vm : 0,
-				  bo_size);
+		bo = xe_bo_create_flags(fd, flags & VM_FOR_BO ? vm : 0,
+					bo_size, visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

tests/xe/xe_exec_fault_mode.c +8 -4

index bf7230c5a..7dcbb3c45 100644
@@ -153,9 +153,11 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
 	} else {
 		if (flags & PREFETCH)
 			bo = xe_bo_create_flags(fd, 0, bo_size,
-						all_memory_regions(fd));
+						all_memory_regions(fd) |
+						visible_vram_if_possible(fd, 0));
 		else
-			bo = xe_bo_create(fd, eci->gt_id, 0, bo_size);
+			bo = xe_bo_create_flags(fd, 0, bo_size,
+						visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

@@ -382,8 +384,10 @@ test_atomic(int fd, struct drm_xe_engine_class_instance *eci,
 	addr_wait = addr + bo_size;
 
 	bo = xe_bo_create_flags(fd, vm, bo_size,
-				all_memory_regions(fd));
-	bo_wait = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+				all_memory_regions(fd) |
+				visible_vram_if_possible(fd, 0));
+	bo_wait = xe_bo_create_flags(fd, vm, bo_size,
+				     visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 	wait = xe_bo_map(fd, bo_wait, bo_size);
 	ptr = &data[0].data;

tests/xe/xe_exec_reset.c +8 -5

index 6ca1cd769..dfbaa6035 100644
@@ -50,7 +50,8 @@ static void test_spin(int fd, struct drm_xe_engine_class_instance *eci)
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	spin = xe_bo_map(fd, bo, bo_size);
 
 	engine = xe_engine_create(fd, vm, eci, 0);

@@ -187,7 +188,7 @@ test_balancer(int fd, int gt, int class, int n_engines, int n_execs,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, gt, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, gt));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

@@ -379,7 +380,8 @@ test_legacy_mode(int fd, struct drm_xe_engine_class_instance *eci,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

@@ -550,7 +552,8 @@ test_compute_mode(int fd, struct drm_xe_engine_class_instance *eci,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 	memset(data, 0, bo_size);
 

@@ -682,7 +685,7 @@ static void submit_jobs(struct gt_thread_data *t)
 	uint32_t bo;
 	uint32_t *data;
 
-	bo = xe_bo_create(fd, 0, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, 0));
 	data = xe_bo_map(fd, bo, bo_size);
 	data[0] = MI_BATCH_BUFFER_END;
 

tests/xe/xe_exec_store.c +4 -2

index 9640b1567..ab1bde36e 100644
@@ -82,7 +82,8 @@ static void store(int fd)
 			xe_get_default_alignment(fd));
 
 	hw_engine = xe_hw_engine(fd, 1);
-	bo = xe_bo_create(fd, hw_engine->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, hw_engine->gt_id));
 
 	xe_vm_bind_async(fd, vm, hw_engine->gt_id, bo, 0, addr, bo_size, &sync, 1);
 	data = xe_bo_map(fd, bo, bo_size);

@@ -138,7 +139,8 @@ static void store_all(int fd, int gt, int class)
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, 0, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, 0));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	xe_for_each_hw_engine(fd, hwe) {

tests/xe/xe_exec_threads.c +6 -3

index 414d8ee9a..396398984 100644
@@ -106,7 +106,8 @@ test_balancer(int fd, int gt, uint32_t vm, uint64_t addr, uint64_t userptr,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, gt, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, gt));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

@@ -306,7 +307,8 @@ test_compute_mode(int fd, uint32_t vm, uint64_t addr, uint64_t userptr,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, 0, bo_size);
+		bo = xe_bo_create_flags(fd, 0, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

@@ -516,7 +518,8 @@ test_legacy_mode(int fd, uint32_t vm, uint64_t addr, uint64_t userptr,
 			igt_assert(data);
 		}
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		data = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(data, 0, bo_size);

tests/xe/xe_guc_pc.c +2 -1

index c34df8d60..6339b3893 100644
@@ -64,7 +64,8 @@ static void exec_basic(int fd, struct drm_xe_engine_class_instance *eci,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

tests/xe/xe_mmap.c +2 -2

index 798facca9..583f39d7a 100644
@@ -118,10 +118,10 @@ igt_main
 		test_mmap(fd, system_memory(fd));
 
 	igt_subtest("vram")
-		test_mmap(fd, vram_memory(fd, 0));
+		test_mmap(fd, visible_vram_memory(fd, 0));
 
 	igt_subtest("vram-system")
-		test_mmap(fd, vram_memory(fd, 0) | system_memory(fd));
+		test_mmap(fd, visible_vram_memory(fd, 0) | system_memory(fd));
 
 	igt_subtest("bad-flags")
 		test_bad_flags(fd);

tests/xe/xe_pm.c +2 -1

index a7f73c4e6..559eccdeb 100644
@@ -254,7 +254,8 @@ test_exec(device_t device, struct drm_xe_engine_class_instance *eci,
 	if (check_rpm && runtime_usage_available(device.pci_xe))
 		rpm_usage = igt_pm_get_runtime_usage(device.pci_xe);
 
-	bo = xe_bo_create(device.fd_xe, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(device.fd_xe, vm, bo_size,
+				visible_vram_if_possible(device.fd_xe, eci->gt_id));
 	data = xe_bo_map(device.fd_xe, bo, bo_size);
 
 	for (i = 0; i < n_engines; i++) {

tests/xe/xe_prime_self_import.c +4 -4

index 0fd79f704..e712e2a9c 100644
@@ -107,7 +107,7 @@ static void test_with_fd_dup(void)
 	fd1 = drm_open_driver(DRIVER_XE);
 	fd2 = drm_open_driver(DRIVER_XE);
 
-	handle = xe_bo_create(fd1, 0, 0, BO_SIZE);
+	handle = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
 
 	dma_buf_fd1 = prime_handle_to_fd(fd1, handle);
 	gem_close(fd1, handle);

@@ -141,8 +141,8 @@ static void test_with_two_bos(void)
 	fd1 = drm_open_driver(DRIVER_XE);
 	fd2 = drm_open_driver(DRIVER_XE);
 
-	handle1 = xe_bo_create(fd1, 0, 0, BO_SIZE);
-	handle2 = xe_bo_create(fd1, 0, 0, BO_SIZE);
+	handle1 = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
+	handle2 = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
 
 	dma_buf_fd = prime_handle_to_fd(fd1, handle1);
 	handle_import = prime_fd_to_handle(fd2, dma_buf_fd);

@@ -211,7 +211,7 @@ static void test_with_one_bo(void)
 	fd1 = drm_open_driver(DRIVER_XE);
 	fd2 = drm_open_driver(DRIVER_XE);
 
-	handle = xe_bo_create(fd1, 0, 0, BO_SIZE);
+	handle = xe_bo_create_flags(fd1, 0, BO_SIZE, visible_vram_if_possible(fd1, 0));
 
 	dma_buf_fd = prime_handle_to_fd(fd1, handle);
 	handle_import1 = prime_fd_to_handle(fd2, dma_buf_fd);

tests/xe/xe_vm.c +9 -5

index 04d6c3956..982c50f6d 100644
@@ -52,7 +52,8 @@ write_dwords(int fd, uint32_t vm, int n_dwords, uint64_t *addrs)
 	batch_size = (n_dwords * 4 + 1) * sizeof(uint32_t);
 	batch_size = ALIGN(batch_size + xe_cs_prefetch_size(fd),
 			   xe_get_default_alignment(fd));
-	batch_bo = xe_bo_create(fd, 0, vm, batch_size);
+	batch_bo = xe_bo_create_flags(fd, vm, batch_size,
+				      visible_vram_if_possible(fd, 0));
 	batch_map = xe_bo_map(fd, batch_bo, batch_size);
 
 	for (i = 0; i < n_dwords; i++) {

@@ -116,7 +117,7 @@ __test_bind_one_bo(int fd, uint32_t vm, int n_addrs, uint64_t *addrs)
 		vms = malloc(sizeof(*vms) * n_addrs);
 		igt_assert(vms);
 	}
-	bo = xe_bo_create(fd, 0, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size, visible_vram_if_possible(fd, 0));
 	map = xe_bo_map(fd, bo, bo_size);
 	memset(map, 0, bo_size);
 

@@ -554,7 +555,8 @@ shared_pte_page(int fd, struct drm_xe_engine_class_instance *eci, int n_bo,
 			xe_get_default_alignment(fd));
 
 	for (i = 0; i < n_bo; ++i) {
-		bo[i] = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		bo[i] = xe_bo_create_flags(fd, vm, bo_size,
+					   visible_vram_if_possible(fd, eci->gt_id));
 		data[i] = xe_bo_map(fd, bo[i], bo_size);
 	}
 

@@ -723,7 +725,8 @@ test_bind_engines_independent(int fd, struct drm_xe_engine_class_instance *eci)
 	bo_size = sizeof(*data) * N_ENGINES;
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 
 	for (i = 0; i < N_ENGINES; i++) {

@@ -880,7 +883,8 @@ test_bind_array(int fd, struct drm_xe_engine_class_instance *eci, int n_execs,
 	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd),
 			xe_get_default_alignment(fd));
 
-	bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+	bo = xe_bo_create_flags(fd, vm, bo_size,
+				visible_vram_if_possible(fd, eci->gt_id));
 	data = xe_bo_map(fd, bo, bo_size);
 

tests/xe/xe_vm.c test_large_binds(int fd, struct drm_xe_engine_class_instance *eci, · line 1080

Quoted 8 lines

tests/xe/xe_vm.c +2 -1

 	if (flags & BIND_ARRAY_BIND_ENGINE_FLAG)

@@ -1072,7 +1076,11 @@ test_large_binds(int fd, struct drm_xe_engine_class_instance *eci,
 		map = aligned_alloc(xe_get_default_alignment(fd), bo_size);
 		igt_assert(map);
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		igt_skip_on(xe_visible_vram_size(fd, 0) && bo_size >
+			    xe_visible_vram_size(fd, 0));

redundant 'xe_visible_vram_size(fd, 0)' if it is 0 it will be smaller than bo_size.

Other than this and the missing cases LGTM.

Quoted 16 lines

tests/xe/xe_vm.c +5 -1

+
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		map = xe_bo_map(fd, bo, bo_size);
 	}
 

@@ -1350,7 +1358,8 @@ test_munmap_style_unbind(int fd, struct drm_xe_engine_class_instance *eci,
 			    MAP_ANONYMOUS, -1, 0);
 		igt_assert(map != MAP_FAILED);
 	} else {
-		bo = xe_bo_create(fd, eci->gt_id, vm, bo_size);
+		bo = xe_bo_create_flags(fd, vm, bo_size,
+					visible_vram_if_possible(fd, eci->gt_id));
 		map = xe_bo_map(fd, bo, bo_size);
 	}
 	memset(map, 0, bo_size);