[PATCH v3 1/1] drm/mm: optimize rb_hole_addr rbtree search
Nirmoy
nirmodas at amd.com
Mon May 4 15:33:59 UTC 2020
On 5/4/20 3:35 PM, Christian König wrote:
> Am 04.05.20 um 12:42 schrieb Nirmoy Das:
>> Userspace can severely fragment rb_hole_addr rbtree by manipulating
>> alignment while allocating buffers. Fragmented rb_hole_addr rbtree
>> would result in large delays while allocating buffer object for a
>> userspace application. It takes long time to find suitable hole
>> because if we fail to find a suitable hole in the first attempt
>> then we look for neighbouring nodes using rb_prev()/rb_next().
>> Traversing rbtree using rb_prev()/rb_next() can take really long
>> time if the tree is fragmented.
>>
>> This patch improves searches in fragmented rb_hole_addr rbtree by
>> modifying it to an augmented rbtree which will store an extra field
>> in drm_mm_node, subtree_max_hole. Each drm_mm_node now stores maximum
>> hole size for its subtree in drm_mm_node->subtree_max_hole. Using
>> drm_mm_node->subtree_max_hole, it is possible to eliminate complete
>> subtree
>
> "eliminate a complete subtree", but I'm not a native speaker either so
> take this with a grain of salt.
No, you are right.
>
>> if that subtree is unable to serve a request hence reducing number of
>> rb_prev()/rb_next() used.
>>
>> With this patch applied, 1 million bo allocations on amdgpu took ~8
>> sec and
>
> "1 million bo allocs on amdgpu took ~8 sec, compared to 50k bo allocs
> took 28 sec without".
>
> E.g. keep the numbers in the same order for easy compare.
Thanks, this looks easier to understand.
>
>> without the patch the test code took 28 sec for only 50k bo allocs.
>>
>> partial test code:
>> int test_fragmentation(void)
>> {
>>
>> int i = 0;
>> uint32_t minor_version;
>> uint32_t major_version;
>>
>> struct amdgpu_bo_alloc_request request = {};
>> amdgpu_bo_handle vram_handle[MAX_ALLOC] = {};
>> amdgpu_device_handle device_handle;
>>
>> request.alloc_size = 4096;
>> request.phys_alignment = 8192;
>> request.preferred_heap = AMDGPU_GEM_DOMAIN_VRAM;
>>
>> int fd = open("/dev/dri/card0", O_RDWR | O_CLOEXEC);
>> amdgpu_device_initialize(fd, &major_version, &minor_version,
>> &device_handle);
>>
>> for (i = 0; i < MAX_ALLOC; i++) {
>> amdgpu_bo_alloc(device_handle, &request,
>> &vram_handle[i]);
>> }
>>
>> for (i = 0; i < MAX_ALLOC; i++)
>> amdgpu_bo_free(vram_handle[i]);
>>
>> return 0;
>> }
>>
>> v2:
>> Use RB_DECLARE_CALLBACKS_MAX to maintain subtree_max_hole
>> v3:
>> insert_hole_addr() should be static a function
>> fix return value of next_hole_high_addr()/next_hole_low_addr()
>> Reported-by: kbuild test robot <lkp at intel.com>
>>
>> Signed-off-by: Nirmoy Das <nirmoy.das at amd.com>
>> Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>
>> ---
>> drivers/gpu/drm/drm_mm.c | 133 +++++++++++++++++++++++++++++++++------
>> include/drm/drm_mm.h | 1 +
>> 2 files changed, 115 insertions(+), 19 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/drm_mm.c b/drivers/gpu/drm/drm_mm.c
>> index 8981abe8b7c9..f4ca1ff80af9 100644
>> --- a/drivers/gpu/drm/drm_mm.c
>> +++ b/drivers/gpu/drm/drm_mm.c
>> @@ -212,20 +212,6 @@ static void drm_mm_interval_tree_add_node(struct
>> drm_mm_node *hole_node,
>> &drm_mm_interval_tree_augment);
>> }
>> -#define RB_INSERT(root, member, expr) do { \
>> - struct rb_node **link = &root.rb_node, *rb = NULL; \
>> - u64 x = expr(node); \
>> - while (*link) { \
>> - rb = *link; \
>> - if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \
>> - link = &rb->rb_left; \
>> - else \
>> - link = &rb->rb_right; \
>> - } \
>> - rb_link_node(&node->member, rb, link); \
>> - rb_insert_color(&node->member, &root); \
>> -} while (0)
>> -
>> #define HOLE_SIZE(NODE) ((NODE)->hole_size)
>> #define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
>> @@ -255,16 +241,42 @@ static void insert_hole_size(struct
>> rb_root_cached *root,
>> rb_insert_color_cached(&node->rb_hole_size, root, first);
>> }
>> +RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks,
>> + struct drm_mm_node, rb_hole_addr,
>> + u64, subtree_max_hole, HOLE_SIZE)
>
> Well that is a really nice solution. I've reviewed the implementation
> and couldn't find anything obvious wrong.
>
> But I'm neither an expert on the drm_mm code nor the augmented RB tree
> implementation. So Acked-by: Christian König
> <christian.koenig at amd.com> for this patch.
Thanks I will resend after updating the commit message.
>
> As a follow up I think we should extend this into keeping track of
> subtree_max_hole_align. With HOLE_SIZE_ALIGN(NODE) defined as
> (((NODE)->hole_size) << 6) | ffs(HOLE_ADDR(NODE)).
This should make it more faster .
Regards,
Nirmoy
>
> This way we can also handle the ugly case of 4K allocation which are
> 8K aligned.
>
> Regards,
> Christian.
>
>> +
>> +static void insert_hole_addr(struct rb_root *root, struct
>> drm_mm_node *node)
>> +{
>> + struct rb_node **link = &root->rb_node, *rb_parent = NULL;
>> + u64 start = HOLE_ADDR(node), subtree_max_hole =
>> node->subtree_max_hole;
>> + struct drm_mm_node *parent;
>> +
>> + while (*link) {
>> + rb_parent = *link;
>> + parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr);
>> + if (parent->subtree_max_hole < subtree_max_hole)
>> + parent->subtree_max_hole = subtree_max_hole;
>> + if (start < HOLE_ADDR(parent))
>> + link = &parent->rb_hole_addr.rb_left;
>> + else
>> + link = &parent->rb_hole_addr.rb_right;
>> + }
>> +
>> + rb_link_node(&node->rb_hole_addr, rb_parent, link);
>> + rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks);
>> +}
>> +
>> static void add_hole(struct drm_mm_node *node)
>> {
>> struct drm_mm *mm = node->mm;
>> node->hole_size =
>> __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
>> + node->subtree_max_hole = node->hole_size;
>> DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
>> insert_hole_size(&mm->holes_size, node);
>> - RB_INSERT(mm->holes_addr, rb_hole_addr, HOLE_ADDR);
>> + insert_hole_addr(&mm->holes_addr, node);
>> list_add(&node->hole_stack, &mm->hole_stack);
>> }
>> @@ -275,8 +287,10 @@ static void rm_hole(struct drm_mm_node *node)
>> list_del(&node->hole_stack);
>> rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size);
>> - rb_erase(&node->rb_hole_addr, &node->mm->holes_addr);
>> + rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr,
>> + &augment_callbacks);
>> node->hole_size = 0;
>> + node->subtree_max_hole = 0;
>> DRM_MM_BUG_ON(drm_mm_hole_follows(node));
>> }
>> @@ -361,9 +375,90 @@ first_hole(struct drm_mm *mm,
>> }
>> }
>> +/**
>> + * next_hole_high_addr - returns next hole for a DRM_MM_INSERT_HIGH
>> mode request
>> + * @entry: previously selected drm_mm_node
>> + * @size: size of the a hole needed for the request
>> + *
>> + * This function will verify whether left subtree of @entry has hole
>> big enough
>> + * to fit the requtested size. If so, it will return previous node
>> of @entry or
>> + * else it will return parent node of @entry
>> + *
>> + * It will also skip the complete left subtree if subtree_max_hole
>> of that
>> + * subtree is same as the subtree_max_hole of the @entry.
>> + *
>> + * Returns:
>> + * previous node of @entry if left subtree of @entry can serve the
>> request or
>> + * else return parent of @entry
>> + */
>> +static struct drm_mm_node *
>> +next_hole_high_addr(struct drm_mm_node *entry, u64 size)
>> +{
>> + struct rb_node *rb_node, *left_rb_node, *parent_rb_node;
>> + struct drm_mm_node *left_node;
>> +
>> + if (!entry)
>> + return NULL;
>> +
>> + rb_node = &entry->rb_hole_addr;
>> + if (rb_node->rb_left) {
>> + left_rb_node = rb_node->rb_left;
>> + parent_rb_node = rb_parent(rb_node);
>> + left_node = rb_entry(left_rb_node,
>> + struct drm_mm_node, rb_hole_addr);
>> + if ((left_node->subtree_max_hole < size ||
>> + entry->size == entry->subtree_max_hole) &&
>> + parent_rb_node && parent_rb_node->rb_left != rb_node)
>> + return rb_hole_addr_to_node(parent_rb_node);
>> + }
>> +
>> + return rb_hole_addr_to_node(rb_prev(rb_node));
>> +}
>> +
>> +/**
>> + * next_hole_low_addr - returns next hole for a DRM_MM_INSERT_LOW
>> mode request
>> + * @entry: previously selected drm_mm_node
>> + * @size: size of the a hole needed for the request
>> + *
>> + * This function will verify whether right subtree of @entry has
>> hole big enough
>> + * to fit the requtested size. If so, it will return next node of
>> @entry or
>> + * else it will return parent node of @entry
>> + *
>> + * It will also skip the complete right subtree if subtree_max_hole
>> of that
>> + * subtree is same as the subtree_max_hole of the @entry.
>> + *
>> + * Returns:
>> + * next node of @entry if right subtree of @entry can serve the
>> request or
>> + * else return parent of @entry
>> + */
>> +static struct drm_mm_node *
>> +next_hole_low_addr(struct drm_mm_node *entry, u64 size)
>> +{
>> + struct rb_node *rb_node, *right_rb_node, *parent_rb_node;
>> + struct drm_mm_node *right_node;
>> +
>> + if (!entry)
>> + return NULL;
>> +
>> + rb_node = &entry->rb_hole_addr;
>> + if (rb_node->rb_right) {
>> + right_rb_node = rb_node->rb_right;
>> + parent_rb_node = rb_parent(rb_node);
>> + right_node = rb_entry(right_rb_node,
>> + struct drm_mm_node, rb_hole_addr);
>> + if ((right_node->subtree_max_hole < size ||
>> + entry->size == entry->subtree_max_hole) &&
>> + parent_rb_node && parent_rb_node->rb_right != rb_node)
>> + return rb_hole_addr_to_node(parent_rb_node);
>> + }
>> +
>> + return rb_hole_addr_to_node(rb_next(rb_node));
>> +}
>> +
>> static struct drm_mm_node *
>> next_hole(struct drm_mm *mm,
>> struct drm_mm_node *node,
>> + u64 size,
>> enum drm_mm_insert_mode mode)
>> {
>> switch (mode) {
>> @@ -372,10 +467,10 @@ next_hole(struct drm_mm *mm,
>> return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
>> case DRM_MM_INSERT_LOW:
>> - return rb_hole_addr_to_node(rb_next(&node->rb_hole_addr));
>> + return next_hole_low_addr(node, size);
>> case DRM_MM_INSERT_HIGH:
>> - return rb_hole_addr_to_node(rb_prev(&node->rb_hole_addr));
>> + return next_hole_high_addr(node, size);
>> case DRM_MM_INSERT_EVICT:
>> node = list_next_entry(node, hole_stack);
>> @@ -489,7 +584,7 @@ int drm_mm_insert_node_in_range(struct drm_mm *
>> const mm,
>> remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
>> for (hole = first_hole(mm, range_start, range_end, size, mode);
>> hole;
>> - hole = once ? NULL : next_hole(mm, hole, mode)) {
>> + hole = once ? NULL : next_hole(mm, hole, size, mode)) {
>> u64 hole_start = __drm_mm_hole_node_start(hole);
>> u64 hole_end = hole_start + hole->hole_size;
>> u64 adj_start, adj_end;
>> diff --git a/include/drm/drm_mm.h b/include/drm/drm_mm.h
>> index ee8b0e80ca90..a01bc6fac83c 100644
>> --- a/include/drm/drm_mm.h
>> +++ b/include/drm/drm_mm.h
>> @@ -168,6 +168,7 @@ struct drm_mm_node {
>> struct rb_node rb_hole_addr;
>> u64 __subtree_last;
>> u64 hole_size;
>> + u64 subtree_max_hole;
>> unsigned long flags;
>> #define DRM_MM_NODE_ALLOCATED_BIT 0
>> #define DRM_MM_NODE_SCANNED_BIT 1
>
More information about the dri-devel
mailing list