* [PATCH] drm: Alloc high address for drm buddy topdown flag
@ 2023-01-07 15:15 Arunpravin Paneer Selvam
2023-01-09 10:13 ` Christian König
2023-01-10 12:02 ` Matthew Auld
0 siblings, 2 replies; 6+ messages in thread
From: Arunpravin Paneer Selvam @ 2023-01-07 15:15 UTC (permalink / raw)
To: dri-devel, intel-gfx, amd-gfx, matthew.auld, christian.koenig,
alexander.deucher
Cc: Arunpravin Paneer Selvam
As we are observing low numbers in viewperf graphics benchmark, we
are strictly not allowing the top down flag enabled allocations
to steal the memory space from cpu visible region.
The approach is, we are sorting each order list entries in
ascending order and compare the last entry of each order
list in the freelist and return the max block.
This patch improves the viewperf 3D benchmark scores.
Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
---
drivers/gpu/drm/drm_buddy.c | 81 ++++++++++++++++++++++++-------------
1 file changed, 54 insertions(+), 27 deletions(-)
diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
index 11bb59399471..50916b2f2fc5 100644
--- a/drivers/gpu/drm/drm_buddy.c
+++ b/drivers/gpu/drm/drm_buddy.c
@@ -38,6 +38,25 @@ static void drm_block_free(struct drm_buddy *mm,
kmem_cache_free(slab_blocks, block);
}
+static void list_insert_sorted(struct drm_buddy *mm,
+ struct drm_buddy_block *block)
+{
+ struct drm_buddy_block *node;
+ struct list_head *head;
+
+ head = &mm->free_list[drm_buddy_block_order(block)];
+ if (list_empty(head)) {
+ list_add(&block->link, head);
+ return;
+ }
+
+ list_for_each_entry(node, head, link)
+ if (drm_buddy_block_offset(block) < drm_buddy_block_offset(node))
+ break;
+
+ __list_add(&block->link, node->link.prev, &node->link);
+}
+
static void mark_allocated(struct drm_buddy_block *block)
{
block->header &= ~DRM_BUDDY_HEADER_STATE;
@@ -52,8 +71,7 @@ static void mark_free(struct drm_buddy *mm,
block->header &= ~DRM_BUDDY_HEADER_STATE;
block->header |= DRM_BUDDY_FREE;
- list_add(&block->link,
- &mm->free_list[drm_buddy_block_order(block)]);
+ list_insert_sorted(mm, block);
}
static void mark_split(struct drm_buddy_block *block)
@@ -387,20 +405,26 @@ alloc_range_bias(struct drm_buddy *mm,
}
static struct drm_buddy_block *
-get_maxblock(struct list_head *head)
+get_maxblock(struct drm_buddy *mm, unsigned int order)
{
struct drm_buddy_block *max_block = NULL, *node;
+ unsigned int i;
- max_block = list_first_entry_or_null(head,
- struct drm_buddy_block,
- link);
- if (!max_block)
- return NULL;
+ for (i = order; i <= mm->max_order; ++i) {
+ if (!list_empty(&mm->free_list[i])) {
+ node = list_last_entry(&mm->free_list[i],
+ struct drm_buddy_block,
+ link);
+ if (!max_block) {
+ max_block = node;
+ continue;
+ }
- list_for_each_entry(node, head, link) {
- if (drm_buddy_block_offset(node) >
- drm_buddy_block_offset(max_block))
- max_block = node;
+ if (drm_buddy_block_offset(node) >
+ drm_buddy_block_offset(max_block)) {
+ max_block = node;
+ }
+ }
}
return max_block;
@@ -412,20 +436,23 @@ alloc_from_freelist(struct drm_buddy *mm,
unsigned long flags)
{
struct drm_buddy_block *block = NULL;
- unsigned int i;
+ unsigned int tmp;
int err;
- for (i = order; i <= mm->max_order; ++i) {
- if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
- block = get_maxblock(&mm->free_list[i]);
- if (block)
- break;
- } else {
- block = list_first_entry_or_null(&mm->free_list[i],
- struct drm_buddy_block,
- link);
- if (block)
- break;
+ if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
+ block = get_maxblock(mm, order);
+ if (block)
+ /* Store the obtained block order */
+ tmp = drm_buddy_block_order(block);
+ } else {
+ for (tmp = order; tmp <= mm->max_order; ++tmp) {
+ if (!list_empty(&mm->free_list[tmp])) {
+ block = list_last_entry(&mm->free_list[tmp],
+ struct drm_buddy_block,
+ link);
+ if (block)
+ break;
+ }
}
}
@@ -434,18 +461,18 @@ alloc_from_freelist(struct drm_buddy *mm,
BUG_ON(!drm_buddy_block_is_free(block));
- while (i != order) {
+ while (tmp != order) {
err = split_block(mm, block);
if (unlikely(err))
goto err_undo;
block = block->right;
- i--;
+ tmp--;
}
return block;
err_undo:
- if (i != order)
+ if (tmp != order)
__drm_buddy_free(mm, block);
return ERR_PTR(err);
}
--
2.25.1
^ permalink raw reply related [flat|nested] 6+ messages in thread
* Re: [PATCH] drm: Alloc high address for drm buddy topdown flag
2023-01-07 15:15 [PATCH] drm: Alloc high address for drm buddy topdown flag Arunpravin Paneer Selvam
@ 2023-01-09 10:13 ` Christian König
2023-01-09 15:03 ` Alex Deucher
2023-01-10 12:02 ` Matthew Auld
1 sibling, 1 reply; 6+ messages in thread
From: Christian König @ 2023-01-09 10:13 UTC (permalink / raw)
To: Arunpravin Paneer Selvam, dri-devel, intel-gfx, amd-gfx,
matthew.auld, christian.koenig, alexander.deucher
Am 07.01.23 um 16:15 schrieb Arunpravin Paneer Selvam:
> As we are observing low numbers in viewperf graphics benchmark, we
> are strictly not allowing the top down flag enabled allocations
> to steal the memory space from cpu visible region.
>
> The approach is, we are sorting each order list entries in
> ascending order and compare the last entry of each order
> list in the freelist and return the max block.
>
> This patch improves the viewperf 3D benchmark scores.
>
> Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>, but somebody with more insight of the drm buddy allocator should take a closer look at this.
> ---
> drivers/gpu/drm/drm_buddy.c | 81 ++++++++++++++++++++++++-------------
> 1 file changed, 54 insertions(+), 27 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
> index 11bb59399471..50916b2f2fc5 100644
> --- a/drivers/gpu/drm/drm_buddy.c
> +++ b/drivers/gpu/drm/drm_buddy.c
> @@ -38,6 +38,25 @@ static void drm_block_free(struct drm_buddy *mm,
> kmem_cache_free(slab_blocks, block);
> }
>
> +static void list_insert_sorted(struct drm_buddy *mm,
> + struct drm_buddy_block *block)
> +{
> + struct drm_buddy_block *node;
> + struct list_head *head;
> +
> + head = &mm->free_list[drm_buddy_block_order(block)];
> + if (list_empty(head)) {
> + list_add(&block->link, head);
> + return;
> + }
> +
> + list_for_each_entry(node, head, link)
> + if (drm_buddy_block_offset(block) < drm_buddy_block_offset(node))
> + break;
> +
> + __list_add(&block->link, node->link.prev, &node->link);
> +}
> +
> static void mark_allocated(struct drm_buddy_block *block)
> {
> block->header &= ~DRM_BUDDY_HEADER_STATE;
> @@ -52,8 +71,7 @@ static void mark_free(struct drm_buddy *mm,
> block->header &= ~DRM_BUDDY_HEADER_STATE;
> block->header |= DRM_BUDDY_FREE;
>
> - list_add(&block->link,
> - &mm->free_list[drm_buddy_block_order(block)]);
> + list_insert_sorted(mm, block);
> }
>
> static void mark_split(struct drm_buddy_block *block)
> @@ -387,20 +405,26 @@ alloc_range_bias(struct drm_buddy *mm,
> }
>
> static struct drm_buddy_block *
> -get_maxblock(struct list_head *head)
> +get_maxblock(struct drm_buddy *mm, unsigned int order)
> {
> struct drm_buddy_block *max_block = NULL, *node;
> + unsigned int i;
>
> - max_block = list_first_entry_or_null(head,
> - struct drm_buddy_block,
> - link);
> - if (!max_block)
> - return NULL;
> + for (i = order; i <= mm->max_order; ++i) {
> + if (!list_empty(&mm->free_list[i])) {
> + node = list_last_entry(&mm->free_list[i],
> + struct drm_buddy_block,
> + link);
> + if (!max_block) {
> + max_block = node;
> + continue;
> + }
>
> - list_for_each_entry(node, head, link) {
> - if (drm_buddy_block_offset(node) >
> - drm_buddy_block_offset(max_block))
> - max_block = node;
> + if (drm_buddy_block_offset(node) >
> + drm_buddy_block_offset(max_block)) {
> + max_block = node;
> + }
> + }
> }
>
> return max_block;
> @@ -412,20 +436,23 @@ alloc_from_freelist(struct drm_buddy *mm,
> unsigned long flags)
> {
> struct drm_buddy_block *block = NULL;
> - unsigned int i;
> + unsigned int tmp;
> int err;
>
> - for (i = order; i <= mm->max_order; ++i) {
> - if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
> - block = get_maxblock(&mm->free_list[i]);
> - if (block)
> - break;
> - } else {
> - block = list_first_entry_or_null(&mm->free_list[i],
> - struct drm_buddy_block,
> - link);
> - if (block)
> - break;
> + if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
> + block = get_maxblock(mm, order);
> + if (block)
> + /* Store the obtained block order */
> + tmp = drm_buddy_block_order(block);
> + } else {
> + for (tmp = order; tmp <= mm->max_order; ++tmp) {
> + if (!list_empty(&mm->free_list[tmp])) {
> + block = list_last_entry(&mm->free_list[tmp],
> + struct drm_buddy_block,
> + link);
> + if (block)
> + break;
> + }
> }
> }
>
> @@ -434,18 +461,18 @@ alloc_from_freelist(struct drm_buddy *mm,
>
> BUG_ON(!drm_buddy_block_is_free(block));
>
> - while (i != order) {
> + while (tmp != order) {
> err = split_block(mm, block);
> if (unlikely(err))
> goto err_undo;
>
> block = block->right;
> - i--;
> + tmp--;
> }
> return block;
>
> err_undo:
> - if (i != order)
> + if (tmp != order)
> __drm_buddy_free(mm, block);
> return ERR_PTR(err);
> }
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] drm: Alloc high address for drm buddy topdown flag
2023-01-09 10:13 ` Christian König
@ 2023-01-09 15:03 ` Alex Deucher
0 siblings, 0 replies; 6+ messages in thread
From: Alex Deucher @ 2023-01-09 15:03 UTC (permalink / raw)
To: Christian König
Cc: dri-devel, Arunpravin Paneer Selvam, intel-gfx, amd-gfx,
matthew.auld, alexander.deucher, christian.koenig
On Mon, Jan 9, 2023 at 5:13 AM Christian König
<ckoenig.leichtzumerken@gmail.com> wrote:
>
> Am 07.01.23 um 16:15 schrieb Arunpravin Paneer Selvam:
> > As we are observing low numbers in viewperf graphics benchmark, we
> > are strictly not allowing the top down flag enabled allocations
> > to steal the memory space from cpu visible region.
> >
> > The approach is, we are sorting each order list entries in
> > ascending order and compare the last entry of each order
> > list in the freelist and return the max block.
> >
> > This patch improves the viewperf 3D benchmark scores.
> >
> > Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
>
> Acked-by: Christian König <christian.koenig@amd.com>, but somebody with more insight of the drm buddy allocator should take a closer look at this.
I'm not a drm_buddy expert either, but this patch fixes a lot of
issues on both dGPUs and APUs:
Acked-by: Alex Deucher <alexander.deucher@amd.com>
>
>
> > ---
> > drivers/gpu/drm/drm_buddy.c | 81 ++++++++++++++++++++++++-------------
> > 1 file changed, 54 insertions(+), 27 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
> > index 11bb59399471..50916b2f2fc5 100644
> > --- a/drivers/gpu/drm/drm_buddy.c
> > +++ b/drivers/gpu/drm/drm_buddy.c
> > @@ -38,6 +38,25 @@ static void drm_block_free(struct drm_buddy *mm,
> > kmem_cache_free(slab_blocks, block);
> > }
> >
> > +static void list_insert_sorted(struct drm_buddy *mm,
> > + struct drm_buddy_block *block)
> > +{
> > + struct drm_buddy_block *node;
> > + struct list_head *head;
> > +
> > + head = &mm->free_list[drm_buddy_block_order(block)];
> > + if (list_empty(head)) {
> > + list_add(&block->link, head);
> > + return;
> > + }
> > +
> > + list_for_each_entry(node, head, link)
> > + if (drm_buddy_block_offset(block) < drm_buddy_block_offset(node))
> > + break;
> > +
> > + __list_add(&block->link, node->link.prev, &node->link);
> > +}
> > +
> > static void mark_allocated(struct drm_buddy_block *block)
> > {
> > block->header &= ~DRM_BUDDY_HEADER_STATE;
> > @@ -52,8 +71,7 @@ static void mark_free(struct drm_buddy *mm,
> > block->header &= ~DRM_BUDDY_HEADER_STATE;
> > block->header |= DRM_BUDDY_FREE;
> >
> > - list_add(&block->link,
> > - &mm->free_list[drm_buddy_block_order(block)]);
> > + list_insert_sorted(mm, block);
> > }
> >
> > static void mark_split(struct drm_buddy_block *block)
> > @@ -387,20 +405,26 @@ alloc_range_bias(struct drm_buddy *mm,
> > }
> >
> > static struct drm_buddy_block *
> > -get_maxblock(struct list_head *head)
> > +get_maxblock(struct drm_buddy *mm, unsigned int order)
> > {
> > struct drm_buddy_block *max_block = NULL, *node;
> > + unsigned int i;
> >
> > - max_block = list_first_entry_or_null(head,
> > - struct drm_buddy_block,
> > - link);
> > - if (!max_block)
> > - return NULL;
> > + for (i = order; i <= mm->max_order; ++i) {
> > + if (!list_empty(&mm->free_list[i])) {
> > + node = list_last_entry(&mm->free_list[i],
> > + struct drm_buddy_block,
> > + link);
> > + if (!max_block) {
> > + max_block = node;
> > + continue;
> > + }
> >
> > - list_for_each_entry(node, head, link) {
> > - if (drm_buddy_block_offset(node) >
> > - drm_buddy_block_offset(max_block))
> > - max_block = node;
> > + if (drm_buddy_block_offset(node) >
> > + drm_buddy_block_offset(max_block)) {
> > + max_block = node;
> > + }
> > + }
> > }
> >
> > return max_block;
> > @@ -412,20 +436,23 @@ alloc_from_freelist(struct drm_buddy *mm,
> > unsigned long flags)
> > {
> > struct drm_buddy_block *block = NULL;
> > - unsigned int i;
> > + unsigned int tmp;
> > int err;
> >
> > - for (i = order; i <= mm->max_order; ++i) {
> > - if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
> > - block = get_maxblock(&mm->free_list[i]);
> > - if (block)
> > - break;
> > - } else {
> > - block = list_first_entry_or_null(&mm->free_list[i],
> > - struct drm_buddy_block,
> > - link);
> > - if (block)
> > - break;
> > + if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
> > + block = get_maxblock(mm, order);
> > + if (block)
> > + /* Store the obtained block order */
> > + tmp = drm_buddy_block_order(block);
> > + } else {
> > + for (tmp = order; tmp <= mm->max_order; ++tmp) {
> > + if (!list_empty(&mm->free_list[tmp])) {
> > + block = list_last_entry(&mm->free_list[tmp],
> > + struct drm_buddy_block,
> > + link);
> > + if (block)
> > + break;
> > + }
> > }
> > }
> >
> > @@ -434,18 +461,18 @@ alloc_from_freelist(struct drm_buddy *mm,
> >
> > BUG_ON(!drm_buddy_block_is_free(block));
> >
> > - while (i != order) {
> > + while (tmp != order) {
> > err = split_block(mm, block);
> > if (unlikely(err))
> > goto err_undo;
> >
> > block = block->right;
> > - i--;
> > + tmp--;
> > }
> > return block;
> >
> > err_undo:
> > - if (i != order)
> > + if (tmp != order)
> > __drm_buddy_free(mm, block);
> > return ERR_PTR(err);
> > }
>
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] drm: Alloc high address for drm buddy topdown flag
2023-01-07 15:15 [PATCH] drm: Alloc high address for drm buddy topdown flag Arunpravin Paneer Selvam
2023-01-09 10:13 ` Christian König
@ 2023-01-10 12:02 ` Matthew Auld
2023-01-10 15:26 ` Arunpravin Paneer Selvam
2023-01-10 15:50 ` Matthew Auld
1 sibling, 2 replies; 6+ messages in thread
From: Matthew Auld @ 2023-01-10 12:02 UTC (permalink / raw)
To: Arunpravin Paneer Selvam, dri-devel, intel-gfx, amd-gfx,
christian.koenig, alexander.deucher
On 07/01/2023 15:15, Arunpravin Paneer Selvam wrote:
> As we are observing low numbers in viewperf graphics benchmark, we
> are strictly not allowing the top down flag enabled allocations
> to steal the memory space from cpu visible region.
>
> The approach is, we are sorting each order list entries in
> ascending order and compare the last entry of each order
> list in the freelist and return the max block.
Did you also run the selftests? Does everything still pass and complete
in a reasonable amount of time?
>
> This patch improves the viewperf 3D benchmark scores.
>
> Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
> ---
> drivers/gpu/drm/drm_buddy.c | 81 ++++++++++++++++++++++++-------------
> 1 file changed, 54 insertions(+), 27 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
> index 11bb59399471..50916b2f2fc5 100644
> --- a/drivers/gpu/drm/drm_buddy.c
> +++ b/drivers/gpu/drm/drm_buddy.c
> @@ -38,6 +38,25 @@ static void drm_block_free(struct drm_buddy *mm,
> kmem_cache_free(slab_blocks, block);
> }
>
> +static void list_insert_sorted(struct drm_buddy *mm,
> + struct drm_buddy_block *block)
> +{
> + struct drm_buddy_block *node;
> + struct list_head *head;
> +
> + head = &mm->free_list[drm_buddy_block_order(block)];
> + if (list_empty(head)) {
> + list_add(&block->link, head);
> + return;
> + }
> +
> + list_for_each_entry(node, head, link)
> + if (drm_buddy_block_offset(block) < drm_buddy_block_offset(node))
> + break;
> +
> + __list_add(&block->link, node->link.prev, &node->link);
> +}
> +
> static void mark_allocated(struct drm_buddy_block *block)
> {
> block->header &= ~DRM_BUDDY_HEADER_STATE;
> @@ -52,8 +71,7 @@ static void mark_free(struct drm_buddy *mm,
> block->header &= ~DRM_BUDDY_HEADER_STATE;
> block->header |= DRM_BUDDY_FREE;
>
> - list_add(&block->link,
> - &mm->free_list[drm_buddy_block_order(block)]);
> + list_insert_sorted(mm, block);
One advantage of not sorting is when splitting down a large block.
Previously the most-recently-split would be at the start of the list for
the next order down, where potentially the next allocation could use it.
So perhaps less fragmentation if it's all part of one BO. Otherwise I
don't see any other downsides, other than the extra overhead of sorting.
> }
>
> static void mark_split(struct drm_buddy_block *block)
> @@ -387,20 +405,26 @@ alloc_range_bias(struct drm_buddy *mm,
> }
>
> static struct drm_buddy_block *
> -get_maxblock(struct list_head *head)
> +get_maxblock(struct drm_buddy *mm, unsigned int order)
> {
> struct drm_buddy_block *max_block = NULL, *node;
> + unsigned int i;
>
> - max_block = list_first_entry_or_null(head,
> - struct drm_buddy_block,
> - link);
> - if (!max_block)
> - return NULL;
> + for (i = order; i <= mm->max_order; ++i) {
> + if (!list_empty(&mm->free_list[i])) {
> + node = list_last_entry(&mm->free_list[i],
> + struct drm_buddy_block,
> + link);
> + if (!max_block) {
> + max_block = node;
> + continue;
> + }
>
> - list_for_each_entry(node, head, link) {
> - if (drm_buddy_block_offset(node) >
> - drm_buddy_block_offset(max_block))
> - max_block = node;
> + if (drm_buddy_block_offset(node) >
> + drm_buddy_block_offset(max_block)) {
Formatting doesn't look right here.
Going to test this today with some workloads with small-bar and i915
just to see if this improves/impacts anything for us.
> + max_block = node;
> + }
> + }
> }
>
> return max_block;
> @@ -412,20 +436,23 @@ alloc_from_freelist(struct drm_buddy *mm,
> unsigned long flags)
> {
> struct drm_buddy_block *block = NULL;
> - unsigned int i;
> + unsigned int tmp;
> int err;
>
> - for (i = order; i <= mm->max_order; ++i) {
> - if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
> - block = get_maxblock(&mm->free_list[i]);
> - if (block)
> - break;
> - } else {
> - block = list_first_entry_or_null(&mm->free_list[i],
> - struct drm_buddy_block,
> - link);
> - if (block)
> - break;
> + if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
> + block = get_maxblock(mm, order);
> + if (block)
> + /* Store the obtained block order */
> + tmp = drm_buddy_block_order(block);
> + } else {
> + for (tmp = order; tmp <= mm->max_order; ++tmp) {
> + if (!list_empty(&mm->free_list[tmp])) {
> + block = list_last_entry(&mm->free_list[tmp],
> + struct drm_buddy_block,
> + link);
> + if (block)
> + break;
> + }
> }
> }
>
> @@ -434,18 +461,18 @@ alloc_from_freelist(struct drm_buddy *mm,
>
> BUG_ON(!drm_buddy_block_is_free(block));
>
> - while (i != order) {
> + while (tmp != order) {
> err = split_block(mm, block);
> if (unlikely(err))
> goto err_undo;
>
> block = block->right;
> - i--;
> + tmp--;
> }
> return block;
>
> err_undo:
> - if (i != order)
> + if (tmp != order)
> __drm_buddy_free(mm, block);
> return ERR_PTR(err);
> }
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] drm: Alloc high address for drm buddy topdown flag
2023-01-10 12:02 ` Matthew Auld
@ 2023-01-10 15:26 ` Arunpravin Paneer Selvam
2023-01-10 15:50 ` Matthew Auld
1 sibling, 0 replies; 6+ messages in thread
From: Arunpravin Paneer Selvam @ 2023-01-10 15:26 UTC (permalink / raw)
To: Matthew Auld, dri-devel, intel-gfx, amd-gfx, christian.koenig,
alexander.deucher
Hi Matthew,
On 1/10/2023 5:32 PM, Matthew Auld wrote:
> On 07/01/2023 15:15, Arunpravin Paneer Selvam wrote:
>> As we are observing low numbers in viewperf graphics benchmark, we
>> are strictly not allowing the top down flag enabled allocations
>> to steal the memory space from cpu visible region.
>>
>> The approach is, we are sorting each order list entries in
>> ascending order and compare the last entry of each order
>> list in the freelist and return the max block.
>
> Did you also run the selftests? Does everything still pass and
> complete in a reasonable amount of time?
> I will try giving a run
>>
>> This patch improves the viewperf 3D benchmark scores.
>>
>> Signed-off-by: Arunpravin Paneer Selvam
>> <Arunpravin.PaneerSelvam@amd.com>
>> ---
>> drivers/gpu/drm/drm_buddy.c | 81 ++++++++++++++++++++++++-------------
>> 1 file changed, 54 insertions(+), 27 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
>> index 11bb59399471..50916b2f2fc5 100644
>> --- a/drivers/gpu/drm/drm_buddy.c
>> +++ b/drivers/gpu/drm/drm_buddy.c
>> @@ -38,6 +38,25 @@ static void drm_block_free(struct drm_buddy *mm,
>> kmem_cache_free(slab_blocks, block);
>> }
>> +static void list_insert_sorted(struct drm_buddy *mm,
>> + struct drm_buddy_block *block)
>> +{
>> + struct drm_buddy_block *node;
>> + struct list_head *head;
>> +
>> + head = &mm->free_list[drm_buddy_block_order(block)];
>> + if (list_empty(head)) {
>> + list_add(&block->link, head);
>> + return;
>> + }
>> +
>> + list_for_each_entry(node, head, link)
>> + if (drm_buddy_block_offset(block) <
>> drm_buddy_block_offset(node))
>> + break;
>> +
>> + __list_add(&block->link, node->link.prev, &node->link);
>> +}
>> +
>> static void mark_allocated(struct drm_buddy_block *block)
>> {
>> block->header &= ~DRM_BUDDY_HEADER_STATE;
>> @@ -52,8 +71,7 @@ static void mark_free(struct drm_buddy *mm,
>> block->header &= ~DRM_BUDDY_HEADER_STATE;
>> block->header |= DRM_BUDDY_FREE;
>> - list_add(&block->link,
>> - &mm->free_list[drm_buddy_block_order(block)]);
>> + list_insert_sorted(mm, block);
>
> One advantage of not sorting is when splitting down a large block.
> Previously the most-recently-split would be at the start of the list
> for the next order down, where potentially the next allocation could
> use it. So perhaps less fragmentation if it's all part of one BO.
> Otherwise I don't see any other downsides, other than the extra
> overhead of sorting.
>
Allocating from freelist is traversing through right side (i.e top most
address range) and for TOPDOWN flag allocations we just split the top
most large block once and the subsequent TOPDOWN low order allocations
would get block from same already split large block
For the normal allocations, I modified to retrieve the blocks in each
order list from the last entry which has the high probability of getting
top most blocks as we have sorted the blocks in each order list.
Thus the bottom most large blocks are not frequently used, hence we have
more space for the visible region on dGPU.
For APU which has small sized VRAM space, the allocations are now
ordered and we don't allocate randomly from freelist solving
fragmentation issues.
>> }
>> static void mark_split(struct drm_buddy_block *block)
>> @@ -387,20 +405,26 @@ alloc_range_bias(struct drm_buddy *mm,
>> }
>> static struct drm_buddy_block *
>> -get_maxblock(struct list_head *head)
>> +get_maxblock(struct drm_buddy *mm, unsigned int order)
>> {
>> struct drm_buddy_block *max_block = NULL, *node;
>> + unsigned int i;
>> - max_block = list_first_entry_or_null(head,
>> - struct drm_buddy_block,
>> - link);
>> - if (!max_block)
>> - return NULL;
>> + for (i = order; i <= mm->max_order; ++i) {
>> + if (!list_empty(&mm->free_list[i])) {
>> + node = list_last_entry(&mm->free_list[i],
>> + struct drm_buddy_block,
>> + link);
>> + if (!max_block) {
>> + max_block = node;
>> + continue;
>> + }
>> - list_for_each_entry(node, head, link) {
>> - if (drm_buddy_block_offset(node) >
>> - drm_buddy_block_offset(max_block))
>> - max_block = node;
>> + if (drm_buddy_block_offset(node) >
>> + drm_buddy_block_offset(max_block)) {
>
> Formatting doesn't look right here.
I will check.
>
> Going to test this today with some workloads with small-bar and i915
> just to see if this improves/impacts anything for us.
>
>> + max_block = node;
>> + }
>> + }
>> }
>> return max_block;
>> @@ -412,20 +436,23 @@ alloc_from_freelist(struct drm_buddy *mm,
>> unsigned long flags)
>> {
>> struct drm_buddy_block *block = NULL;
>> - unsigned int i;
>> + unsigned int tmp;
>> int err;
>> - for (i = order; i <= mm->max_order; ++i) {
>> - if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
>> - block = get_maxblock(&mm->free_list[i]);
>> - if (block)
>> - break;
>> - } else {
>> - block = list_first_entry_or_null(&mm->free_list[i],
>> - struct drm_buddy_block,
>> - link);
>> - if (block)
>> - break;
>> + if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
>> + block = get_maxblock(mm, order);
>> + if (block)
>> + /* Store the obtained block order */
>> + tmp = drm_buddy_block_order(block);
>> + } else {
>> + for (tmp = order; tmp <= mm->max_order; ++tmp) {
>> + if (!list_empty(&mm->free_list[tmp])) {
>> + block = list_last_entry(&mm->free_list[tmp],
>> + struct drm_buddy_block,
>> + link);
>> + if (block)
>> + break;
>> + }
>> }
>> }
>> @@ -434,18 +461,18 @@ alloc_from_freelist(struct drm_buddy *mm,
>> BUG_ON(!drm_buddy_block_is_free(block));
>> - while (i != order) {
>> + while (tmp != order) {
>> err = split_block(mm, block);
>> if (unlikely(err))
>> goto err_undo;
>> block = block->right;
>> - i--;
>> + tmp--;
>> }
>> return block;
>> err_undo:
>> - if (i != order)
>> + if (tmp != order)
>> __drm_buddy_free(mm, block);
>> return ERR_PTR(err);
>> }
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] drm: Alloc high address for drm buddy topdown flag
2023-01-10 12:02 ` Matthew Auld
2023-01-10 15:26 ` Arunpravin Paneer Selvam
@ 2023-01-10 15:50 ` Matthew Auld
1 sibling, 0 replies; 6+ messages in thread
From: Matthew Auld @ 2023-01-10 15:50 UTC (permalink / raw)
To: Arunpravin Paneer Selvam, dri-devel, intel-gfx, amd-gfx,
christian.koenig, alexander.deucher
On 10/01/2023 12:02, Matthew Auld wrote:
> On 07/01/2023 15:15, Arunpravin Paneer Selvam wrote:
>> As we are observing low numbers in viewperf graphics benchmark, we
>> are strictly not allowing the top down flag enabled allocations
>> to steal the memory space from cpu visible region.
>>
>> The approach is, we are sorting each order list entries in
>> ascending order and compare the last entry of each order
>> list in the freelist and return the max block.
>
> Did you also run the selftests? Does everything still pass and complete
> in a reasonable amount of time?
>
>>
>> This patch improves the viewperf 3D benchmark scores.
>>
>> Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
>> ---
>> drivers/gpu/drm/drm_buddy.c | 81 ++++++++++++++++++++++++-------------
>> 1 file changed, 54 insertions(+), 27 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
>> index 11bb59399471..50916b2f2fc5 100644
>> --- a/drivers/gpu/drm/drm_buddy.c
>> +++ b/drivers/gpu/drm/drm_buddy.c
>> @@ -38,6 +38,25 @@ static void drm_block_free(struct drm_buddy *mm,
>> kmem_cache_free(slab_blocks, block);
>> }
>> +static void list_insert_sorted(struct drm_buddy *mm,
>> + struct drm_buddy_block *block)
>> +{
>> + struct drm_buddy_block *node;
>> + struct list_head *head;
>> +
>> + head = &mm->free_list[drm_buddy_block_order(block)];
>> + if (list_empty(head)) {
>> + list_add(&block->link, head);
>> + return;
>> + }
>> +
>> + list_for_each_entry(node, head, link)
>> + if (drm_buddy_block_offset(block) <
>> drm_buddy_block_offset(node))
>> + break;
>> +
>> + __list_add(&block->link, node->link.prev, &node->link);
>> +}
>> +
>> static void mark_allocated(struct drm_buddy_block *block)
>> {
>> block->header &= ~DRM_BUDDY_HEADER_STATE;
>> @@ -52,8 +71,7 @@ static void mark_free(struct drm_buddy *mm,
>> block->header &= ~DRM_BUDDY_HEADER_STATE;
>> block->header |= DRM_BUDDY_FREE;
>> - list_add(&block->link,
>> - &mm->free_list[drm_buddy_block_order(block)]);
>> + list_insert_sorted(mm, block);
>
> One advantage of not sorting is when splitting down a large block.
> Previously the most-recently-split would be at the start of the list for
> the next order down, where potentially the next allocation could use it.
> So perhaps less fragmentation if it's all part of one BO. Otherwise I
> don't see any other downsides, other than the extra overhead of sorting.
>
>> }
>> static void mark_split(struct drm_buddy_block *block)
>> @@ -387,20 +405,26 @@ alloc_range_bias(struct drm_buddy *mm,
>> }
>> static struct drm_buddy_block *
>> -get_maxblock(struct list_head *head)
>> +get_maxblock(struct drm_buddy *mm, unsigned int order)
>> {
>> struct drm_buddy_block *max_block = NULL, *node;
>> + unsigned int i;
>> - max_block = list_first_entry_or_null(head,
>> - struct drm_buddy_block,
>> - link);
>> - if (!max_block)
>> - return NULL;
>> + for (i = order; i <= mm->max_order; ++i) {
>> + if (!list_empty(&mm->free_list[i])) {
>> + node = list_last_entry(&mm->free_list[i],
>> + struct drm_buddy_block,
>> + link);
>> + if (!max_block) {
>> + max_block = node;
>> + continue;
>> + }
>> - list_for_each_entry(node, head, link) {
>> - if (drm_buddy_block_offset(node) >
>> - drm_buddy_block_offset(max_block))
>> - max_block = node;
>> + if (drm_buddy_block_offset(node) >
>> + drm_buddy_block_offset(max_block)) {
>
> Formatting doesn't look right here.
>
> Going to test this today with some workloads with small-bar and i915
> just to see if this improves/impacts anything for us.
No surprises, and as advertised seems to lead to reduced utilisation of
the mappable part for buffers that don't explicitly need it (TOPDOWN).
Assuming the selftests are still happy,
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
>
>> + max_block = node;
>> + }
>> + }
>> }
>> return max_block;
>> @@ -412,20 +436,23 @@ alloc_from_freelist(struct drm_buddy *mm,
>> unsigned long flags)
>> {
>> struct drm_buddy_block *block = NULL;
>> - unsigned int i;
>> + unsigned int tmp;
>> int err;
>> - for (i = order; i <= mm->max_order; ++i) {
>> - if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
>> - block = get_maxblock(&mm->free_list[i]);
>> - if (block)
>> - break;
>> - } else {
>> - block = list_first_entry_or_null(&mm->free_list[i],
>> - struct drm_buddy_block,
>> - link);
>> - if (block)
>> - break;
>> + if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
>> + block = get_maxblock(mm, order);
>> + if (block)
>> + /* Store the obtained block order */
>> + tmp = drm_buddy_block_order(block);
>> + } else {
>> + for (tmp = order; tmp <= mm->max_order; ++tmp) {
>> + if (!list_empty(&mm->free_list[tmp])) {
>> + block = list_last_entry(&mm->free_list[tmp],
>> + struct drm_buddy_block,
>> + link);
>> + if (block)
>> + break;
>> + }
>> }
>> }
>> @@ -434,18 +461,18 @@ alloc_from_freelist(struct drm_buddy *mm,
>> BUG_ON(!drm_buddy_block_is_free(block));
>> - while (i != order) {
>> + while (tmp != order) {
>> err = split_block(mm, block);
>> if (unlikely(err))
>> goto err_undo;
>> block = block->right;
>> - i--;
>> + tmp--;
>> }
>> return block;
>> err_undo:
>> - if (i != order)
>> + if (tmp != order)
>> __drm_buddy_free(mm, block);
>> return ERR_PTR(err);
>> }
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2023-01-10 15:50 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2023-01-07 15:15 [PATCH] drm: Alloc high address for drm buddy topdown flag Arunpravin Paneer Selvam
2023-01-09 10:13 ` Christian König
2023-01-09 15:03 ` Alex Deucher
2023-01-10 12:02 ` Matthew Auld
2023-01-10 15:26 ` Arunpravin Paneer Selvam
2023-01-10 15:50 ` Matthew Auld
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox