[PATCH 1/2] amdgpu: improve amdgpu_vamgr_init

[PATCH 1/2] amdgpu: improve amdgpu_vamgr_init

[Jammy Zhou]

Make it a generic function independent of the device info.

Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
---
 amdgpu/amdgpu_vamgr.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/amdgpu/amdgpu_vamgr.c b/amdgpu/amdgpu_vamgr.c
index ced4f4f..e044dfa 100644
--- a/amdgpu/amdgpu_vamgr.c
+++ b/amdgpu/amdgpu_vamgr.c
@@ -42,11 +42,12 @@ int amdgpu_va_range_query(amdgpu_device_handle dev,
 	return -EINVAL;
 }
 
-static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, struct amdgpu_device *dev)
+static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
+			      uint64_t max, uint64_t alignment)
 {
-	mgr->va_offset = dev->dev_info.virtual_address_offset;
-	mgr->va_max = dev->dev_info.virtual_address_max;
-	mgr->va_alignment = dev->dev_info.virtual_address_alignment;
+	mgr->va_offset = start;
+	mgr->va_max = max;
+	mgr->va_alignment = alignment;
 
 	list_inithead(&mgr->va_holes);
 	pthread_mutex_init(&mgr->bo_va_mutex, NULL);
@@ -68,7 +69,9 @@ struct amdgpu_bo_va_mgr * amdgpu_vamgr_get_global(struct amdgpu_device *dev)
 	ref = atomic_inc_return(&vamgr.refcount);
 
 	if (ref == 1)
-		amdgpu_vamgr_init(&vamgr, dev);
+		amdgpu_vamgr_init(&vamgr, dev->dev_info.virtual_address_offset,
+				  dev->dev_info.virtual_address_max,
+				  dev->dev_info.virtual_address_alignment);
 	return &vamgr;
 }
 
-- 
1.9.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 2/2] amdgpu: add flag to support 32bit VA address v2

[Jammy Zhou]

The AMDGPU_VA_RANGE_32_BIT flag is added to request VA range in the
32bit address space for amdgpu_va_range_alloc.

The 32bit address space is reserved at initialization time, and managed
with a separate VAMGR as part of the global VAMGR. And if no enough VA
space available in range above 4GB, this reserved range can be used as
fallback.

v2: add comment for AMDGPU_VA_RANGE_32_BIT, and add vamgr to va_range

Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
---
 amdgpu/amdgpu.h          |  5 +++++
 amdgpu/amdgpu_device.c   | 22 ++++++++++++++++++++++
 amdgpu/amdgpu_internal.h |  9 +++++++++
 amdgpu/amdgpu_vamgr.c    | 35 ++++++++++++++++++++++++++++-------
 4 files changed, 64 insertions(+), 7 deletions(-)

diff --git a/amdgpu/amdgpu.h b/amdgpu/amdgpu.h
index a90c1ac..1e633e3 100644
--- a/amdgpu/amdgpu.h
+++ b/amdgpu/amdgpu.h
@@ -1075,6 +1075,11 @@ int amdgpu_read_mm_registers(amdgpu_device_handle dev, unsigned dword_offset,
 			     uint32_t *values);
 
 /**
+ * Flag to request VA address range in the 32bit address space
+*/
+#define AMDGPU_VA_RANGE_32_BIT		0x1
+
+/**
  * Allocate virtual address range
  *
  * \param dev - [in] Device handle. See #amdgpu_device_initialize()
diff --git a/amdgpu/amdgpu_device.c b/amdgpu/amdgpu_device.c
index ba75db0..067ef64 100644
--- a/amdgpu/amdgpu_device.c
+++ b/amdgpu/amdgpu_device.c
@@ -40,6 +40,7 @@
 #include "amdgpu_drm.h"
 #include "amdgpu_internal.h"
 #include "util_hash_table.h"
+#include "util_math.h"
 
 #define PTR_TO_UINT(x) ((unsigned)((intptr_t)(x)))
 #define UINT_TO_PTR(x) ((void *)((intptr_t)(x)))
@@ -126,6 +127,7 @@ int amdgpu_device_initialize(int fd,
 	int flag_auth = 0;
 	int flag_authexist=0;
 	uint32_t accel_working = 0;
+	uint64_t start, max;
 
 	*device_handle = NULL;
 
@@ -204,6 +206,19 @@ int amdgpu_device_initialize(int fd,
 
 	dev->vamgr = amdgpu_vamgr_get_global(dev);
 
+	max = MIN2(dev->dev_info.virtual_address_max, 0xffffffff);
+	start = amdgpu_vamgr_find_va(dev->vamgr,
+				     max - dev->dev_info.virtual_address_offset,
+				     dev->dev_info.virtual_address_alignment, 0);
+	if (start > 0xffffffff)
+		goto free_va; /* shouldn't get here */
+
+	dev->vamgr_32 =  calloc(1, sizeof(struct amdgpu_bo_va_mgr));
+	if (dev->vamgr_32 == NULL)
+		goto free_va;
+	amdgpu_vamgr_init(dev->vamgr_32, start, max,
+			  dev->dev_info.virtual_address_alignment);
+
 	*major_version = dev->major_version;
 	*minor_version = dev->minor_version;
 	*device_handle = dev;
@@ -212,6 +227,11 @@ int amdgpu_device_initialize(int fd,
 
 	return 0;
 
+free_va:
+	r = -ENOMEM;
+	amdgpu_vamgr_free_va(dev->vamgr, start,
+			     max - dev->dev_info.virtual_address_offset);
+
 cleanup:
 	if (dev->fd >= 0)
 		close(dev->fd);
@@ -222,6 +242,8 @@ cleanup:
 
 void amdgpu_device_free_internal(amdgpu_device_handle dev)
 {
+	amdgpu_vamgr_deinit(dev->vamgr_32);
+	free(dev->vamgr_32);
 	amdgpu_vamgr_reference(&dev->vamgr, NULL);
 	util_hash_table_destroy(dev->bo_flink_names);
 	util_hash_table_destroy(dev->bo_handles);
diff --git a/amdgpu/amdgpu_internal.h b/amdgpu/amdgpu_internal.h
index 526a93f..6e65642 100644
--- a/amdgpu/amdgpu_internal.h
+++ b/amdgpu/amdgpu_internal.h
@@ -63,6 +63,7 @@ struct amdgpu_va {
 	uint64_t address;
 	uint64_t size;
 	enum amdgpu_gpu_va_range range;
+	struct amdgpu_bo_va_mgr *vamgr;
 };
 
 struct amdgpu_device {
@@ -80,7 +81,10 @@ struct amdgpu_device {
 	pthread_mutex_t bo_table_mutex;
 	struct drm_amdgpu_info_device dev_info;
 	struct amdgpu_gpu_info info;
+	/** The global VA manager for the whole virtual address space */
 	struct amdgpu_bo_va_mgr *vamgr;
+	/** The VA manager for the 32bit address space */
+	struct amdgpu_bo_va_mgr *vamgr_32;
 };
 
 struct amdgpu_bo {
@@ -124,6 +128,11 @@ struct amdgpu_bo_va_mgr* amdgpu_vamgr_get_global(struct amdgpu_device *dev);
 
 void amdgpu_vamgr_reference(struct amdgpu_bo_va_mgr **dst, struct amdgpu_bo_va_mgr *src);
 
+void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
+		       uint64_t max, uint64_t alignment);
+
+void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr);
+
 uint64_t amdgpu_vamgr_find_va(struct amdgpu_bo_va_mgr *mgr, uint64_t size,
 				uint64_t alignment, uint64_t base_required);
 
diff --git a/amdgpu/amdgpu_vamgr.c b/amdgpu/amdgpu_vamgr.c
index e044dfa..04e09e2 100644
--- a/amdgpu/amdgpu_vamgr.c
+++ b/amdgpu/amdgpu_vamgr.c
@@ -42,7 +42,7 @@ int amdgpu_va_range_query(amdgpu_device_handle dev,
 	return -EINVAL;
 }
 
-static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
+void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
 			      uint64_t max, uint64_t alignment)
 {
 	mgr->va_offset = start;
@@ -53,7 +53,7 @@ static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
 	pthread_mutex_init(&mgr->bo_va_mutex, NULL);
 }
 
-static void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr)
+void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr)
 {
 	struct amdgpu_bo_va_hole *hole;
 	LIST_FOR_EACH_ENTRY(hole, &mgr->va_holes, list) {
@@ -248,23 +248,40 @@ int amdgpu_va_range_alloc(amdgpu_device_handle dev,
 			  amdgpu_va_handle *va_range_handle,
 			  uint64_t flags)
 {
-	va_base_alignment = MAX2(va_base_alignment, dev->vamgr->va_alignment);
-	size = ALIGN(size, vamgr.va_alignment);
+	struct amdgpu_bo_va_mgr *vamgr;
 
-	*va_base_allocated = amdgpu_vamgr_find_va(dev->vamgr, size,
+	if (flags & AMDGPU_VA_RANGE_32_BIT)
+		vamgr = dev->vamgr_32;
+	else
+		vamgr = dev->vamgr;
+
+	va_base_alignment = MAX2(va_base_alignment, vamgr->va_alignment);
+	size = ALIGN(size, vamgr->va_alignment);
+
+	*va_base_allocated = amdgpu_vamgr_find_va(vamgr, size,
 					va_base_alignment, va_base_required);
 
+	if (!(flags & AMDGPU_VA_RANGE_32_BIT) &&
+	    (*va_base_allocated == AMDGPU_INVALID_VA_ADDRESS)) {
+		/* fallback to 32bit address */
+		vamgr = dev->vamgr_32;
+		*va_base_allocated = amdgpu_vamgr_find_va(vamgr, size,
+					va_base_alignment, va_base_required);
+		
+	}
+
 	if (*va_base_allocated != AMDGPU_INVALID_VA_ADDRESS) {
 		struct amdgpu_va* va;
 		va = calloc(1, sizeof(struct amdgpu_va));
 		if(!va){
-			amdgpu_vamgr_free_va(dev->vamgr, *va_base_allocated, size);
+			amdgpu_vamgr_free_va(vamgr, *va_base_allocated, size);
 			return -ENOMEM;
 		}
 		va->dev = dev;
 		va->address = *va_base_allocated;
 		va->size = size;
 		va->range = va_range_type;
+		va->vamgr = vamgr;
 		*va_range_handle = va;
 	} else {
 		return -EINVAL;
@@ -275,9 +292,13 @@ int amdgpu_va_range_alloc(amdgpu_device_handle dev,
 
 int amdgpu_va_range_free(amdgpu_va_handle va_range_handle)
 {
+	struct amdgpu_bo_va_mgr *vamgr;
+
 	if(!va_range_handle || !va_range_handle->address)
 		return 0;
-	amdgpu_vamgr_free_va(va_range_handle->dev->vamgr, va_range_handle->address,
+
+	amdgpu_vamgr_free_va(va_range_handle->vamgr,
+			va_range_handle->address,
 			va_range_handle->size);
 	free(va_range_handle);
 	return 0;
-- 
1.9.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 1/2] amdgpu: improve amdgpu_vamgr_init

[Alex Deucher]

On Mon, Aug 10, 2015 at 8:26 AM, Jammy Zhou <Jammy.Zhou@amd.com> wrote:
> Make it a generic function independent of the device info.
>
> Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
> Reviewed-by: Christian König <christian.koenig@amd.com>

Series is:
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

If there are not additional comments, I'll push this in the next couple of days.

Alex

> ---
>  amdgpu/amdgpu_vamgr.c | 13 ++++++++-----
>  1 file changed, 8 insertions(+), 5 deletions(-)
>
> diff --git a/amdgpu/amdgpu_vamgr.c b/amdgpu/amdgpu_vamgr.c
> index ced4f4f..e044dfa 100644
> --- a/amdgpu/amdgpu_vamgr.c
> +++ b/amdgpu/amdgpu_vamgr.c
> @@ -42,11 +42,12 @@ int amdgpu_va_range_query(amdgpu_device_handle dev,
>         return -EINVAL;
>  }
>
> -static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, struct amdgpu_device *dev)
> +static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
> +                             uint64_t max, uint64_t alignment)
>  {
> -       mgr->va_offset = dev->dev_info.virtual_address_offset;
> -       mgr->va_max = dev->dev_info.virtual_address_max;
> -       mgr->va_alignment = dev->dev_info.virtual_address_alignment;
> +       mgr->va_offset = start;
> +       mgr->va_max = max;
> +       mgr->va_alignment = alignment;
>
>         list_inithead(&mgr->va_holes);
>         pthread_mutex_init(&mgr->bo_va_mutex, NULL);
> @@ -68,7 +69,9 @@ struct amdgpu_bo_va_mgr * amdgpu_vamgr_get_global(struct amdgpu_device *dev)
>         ref = atomic_inc_return(&vamgr.refcount);
>
>         if (ref == 1)
> -               amdgpu_vamgr_init(&vamgr, dev);
> +               amdgpu_vamgr_init(&vamgr, dev->dev_info.virtual_address_offset,
> +                                 dev->dev_info.virtual_address_max,
> +                                 dev->dev_info.virtual_address_alignment);
>         return &vamgr;
>  }
>
> --
> 1.9.1
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 2/2] amdgpu: add flag to support 32bit VA address v2

[Emil Velikov]

Hi Jammy,

This patch will conflict with an earlier series that I'm planning to
merge shortly. Can you squash the following changes and check for errors
with $make check/distcheck.

On 10/08/15 13:26, Jammy Zhou wrote:
> The AMDGPU_VA_RANGE_32_BIT flag is added to request VA range in the
> 32bit address space for amdgpu_va_range_alloc.
> 
> The 32bit address space is reserved at initialization time, and managed
> with a separate VAMGR as part of the global VAMGR. And if no enough VA
> space available in range above 4GB, this reserved range can be used as
> fallback.
> 
> v2: add comment for AMDGPU_VA_RANGE_32_BIT, and add vamgr to va_range
> 
> Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
> Reviewed-by: Christian König <christian.koenig@amd.com>
> ---
>  amdgpu/amdgpu.h          |  5 +++++
>  amdgpu/amdgpu_device.c   | 22 ++++++++++++++++++++++
>  amdgpu/amdgpu_internal.h |  9 +++++++++
>  amdgpu/amdgpu_vamgr.c    | 35 ++++++++++++++++++++++++++++-------
>  4 files changed, 64 insertions(+), 7 deletions(-)
> 

> diff --git a/amdgpu/amdgpu_internal.h b/amdgpu/amdgpu_internal.h
> index 526a93f..6e65642 100644
> --- a/amdgpu/amdgpu_internal.h
> +++ b/amdgpu/amdgpu_internal.h

> @@ -124,6 +128,11 @@ struct amdgpu_bo_va_mgr* amdgpu_vamgr_get_global(struct amdgpu_device *dev);
>  
>  void amdgpu_vamgr_reference(struct amdgpu_bo_va_mgr **dst, struct amdgpu_bo_va_mgr *src);
>  
> +void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
> +		       uint64_t max, uint64_t alignment);
> +
> +void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr);
> +
Please prefix these two with the drm_private macro.

>  uint64_t amdgpu_vamgr_find_va(struct amdgpu_bo_va_mgr *mgr, uint64_t size,
>  				uint64_t alignment, uint64_t base_required);
>  
> diff --git a/amdgpu/amdgpu_vamgr.c b/amdgpu/amdgpu_vamgr.c
> index e044dfa..04e09e2 100644
> --- a/amdgpu/amdgpu_vamgr.c
> +++ b/amdgpu/amdgpu_vamgr.c
> @@ -42,7 +42,7 @@ int amdgpu_va_range_query(amdgpu_device_handle dev,
>  	return -EINVAL;
>  }
>  
> -static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
> +void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
>  			      uint64_t max, uint64_t alignment)
Same here,
>  {
>  	mgr->va_offset = start;
> @@ -53,7 +53,7 @@ static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
>  	pthread_mutex_init(&mgr->bo_va_mutex, NULL);
>  }
>  
> -static void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr)
> +void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr)
and here.

You won't need to hunt the correct include as my earlier series adds it
in these two files.

Thanks
Emil

P.S. I'll take another look at the drmGetDevices patch in a few hours.

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel