Intel-GFX Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: Arunpravin Paneer Selvam <arunpravin.paneerselvam@amd.com>
To: Matthew Auld <matthew.auld@intel.com>,
	christian.koenig@amd.com, dri-devel@lists.freedesktop.org,
	intel-gfx@lists.freedesktop.org, intel-xe@lists.freedesktop.org,
	amd-gfx@lists.freedesktop.org
Cc: alexander.deucher@amd.com, Anand.Raghavendra@amd.com
Subject: Re: [PATCH v9 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
Date: Mon, 17 Aug 2026 18:23:16 +0530	[thread overview]
Message-ID: <1cb5dc7c-c342-4687-a7ce-2f8e2e39109e@amd.com> (raw)
In-Reply-To: <0b92672e-dee7-4ef1-83fd-ea9f31729eca@intel.com>



On 8/17/2026 3:27 PM, Matthew Auld wrote:
> Hey,
>
> Sorry, was OoO last week. I think you missed some feedback/questions 
> here: 
> https://lore.kernel.org/intel-xe/c63399c5-8507-4f4e-9ef1-241efac4d024@intel.com/ 
>
I missed that email and don't see it in my inbox. Thanks for sharing the 
link.
I will go through the comments and reply on the thread.

Thanks,
Arun.
>
> Main question was around GPU_DIRTY_EXTENT_POOL_MIN.
>
> On 13/08/2026 14:37, Arunpravin Paneer Selvam wrote:
>> Hi Matthew,
>>
>> Do you have any objections or concerns with the patches? If not, 
>> should I proceed with merging them,
>> or are you still in the process of reviewing/validating them?
>>
>> Regards,
>> Arun.
>>
>> On 8/11/2026 7:04 PM, Arunpravin Paneer Selvam wrote:
>>> The current buddy allocator maintains separate clear_tree[] and
>>> dirty_tree[] rbtrees per order, preventing coalescing between cleared
>>> and dirty buddies. Under mixed workloads, this creates a merge barrier:
>>> adjacent buddies frequently end up split across trees, forcing reliance
>>> on __force_merge() during allocation.
>>>
>>> __force_merge() performs an O(N x max_order) scan under the VRAM 
>>> manager
>>> lock, leading to allocation stalls and failures for large contiguous
>>> requests even when sufficient total free memory is available.
>>>
>>> Solution
>>>
>>> Replace the dual-tree design with:
>>> - A single free_tree[order] rbtree for dirty and mixed free blocks
>>>    (fully cleared free blocks float outside this tree)
>>> - A lightweight out-of-band dirty tracker (gpu_dirty_tracker)
>>>
>>> Fully cleared free blocks are tracked outside the buddy trees using an
>>> augmented interval rbtree, enabling O(log E) lookup of the largest
>>> cleared extents.
>>>
>>> Buddy coalescing is now unconditional in __gpu_buddy_free(), regardless
>>> of clear/dirty state. This removes the merge barrier and eliminates the
>>> need for __force_merge().
>>>
>>> Benefits
>>>
>>> - Correct high-order allocations after mixed clear/dirty workloads
>>> - Elimination of O(N x max_order) merge cost from the allocation path
>>> - O(log E) cleared-extent lookup replacing O(N) scans
>>> - Predictable allocation latency under fragmentation
>>> - Reduced complexity with a single tree per order
>>>
>>> Test:
>>> dEQP-VK.memory.allocation.basic.size_8KiB.reverse.count_4000
>>>
>>> Below data is from /sys/kernel/debug/dri/1/amdgpu_vram_mm:
>>>
>>> Base (dual-tree), before VKCTS test:
>>>    order- 6 free:   6 MiB,  blocks: 26
>>>    order- 5 free:   1 MiB,  blocks: 15
>>>    order- 4 free: 960 KiB,  blocks: 15
>>>    order- 3 free:   5 MiB,  blocks: 171
>>>    order- 2 free:   2 MiB,  blocks: 176
>>>    order- 1 free:   1 MiB,  blocks: 165
>>>    order- 0 free:  16 KiB,  blocks: 4
>>>
>>> Base (dual-tree), after VKCTS test:
>>>    order- 6 free: 768 KiB,  blocks: 3
>>>    order- 5 free: 499 MiB,  blocks: 3999
>>>    order- 4 free: 250 MiB,  blocks: 4001
>>>    order- 3 free: 129 MiB,  blocks: 4157
>>>    order- 2 free:  65 MiB,  blocks: 4161
>>>    order- 1 free:  63 MiB,  blocks: 8138
>>>    order- 0 free:  20 KiB,  blocks: 5
>>>
>>> Dirty tracker, before VKCTS test:
>>>    order- 6 free:   4 MiB,  blocks: 19
>>>    order- 5 free:   2 MiB,  blocks: 18
>>>    order- 4 free: 704 KiB,  blocks: 11
>>>    order- 3 free:   5 MiB,  blocks: 168
>>>    order- 2 free:   2 MiB,  blocks: 174
>>>    order- 1 free:   1 MiB,  blocks: 167
>>>    order- 0 free:  32 KiB,  blocks: 8
>>>
>>> Dirty tracker, after VKCTS test:
>>>    order- 6 free:   4 MiB,  blocks: 19
>>>    order- 5 free:   2 MiB,  blocks: 18
>>>    order- 4 free: 704 KiB,  blocks: 11
>>>    order- 3 free:   5 MiB,  blocks: 168
>>>    order- 2 free:   2 MiB,  blocks: 174
>>>    order- 1 free:   1 MiB,  blocks: 167
>>>    order- 0 free:  28 KiB,  blocks: 7
>>>
>>> v2:
>>>   - Code-style cleanup and minor refactoring
>>>   - Renamed locals for clarity
>>>
>>> v3:
>>>   - Keep cleared blocks inside free_tree[] instead of floating them.
>>>   - Add subtree_has_dirty rbtree augment for O(log N) dirty-first walk.
>>>
>>> v4:
>>>   - Fixed checkpatch warnings.
>>>   - Optimized gpu_buddy_reset_clear() to a single post-order walk that
>>>     flips block headers and recomputes the rbtree augment in one pass.
>>>   - Propagate subtree_max_size top-down in insert_extent() so ancestors
>>>     are not left with stale values on no-rotation inserts. (sashiko)
>>>   - Drop the whole extent in gpu_dirty_tracker_mark_dirty() when the
>>>     inside-split allocation fails, avoiding a stale clear claim. 
>>> (sashiko)
>>>   - Make gpu_dirty_tracker_find() alignment-aware and fall back to the
>>>     dirty tree on steered failure to avoid spurious -ENOSPC. (sashiko)
>>>
>>> v5:
>>>   - Track dirty extents instead of cleared ones: steer dirty allocs 
>>> onto
>>>     tracked dirty windows and pick clear allocs via a free-tree 
>>> augment,
>>>     avoiding clear-memory wastage by keeping cleared free blocks 
>>> untouched
>>>     during dirty allocation.
>>>
>>> v6:
>>>   - Make __alloc_range_bias() return the highest/right-most address by
>>>     default, establishing top-down as the intended placement for
>>>     range-biased allocations.
>>>   - Honour GPU_BUDDY_CLEAR_ALLOCATION in __alloc_range_bias() by 
>>> steering
>>>     the descent towards clear subtrees for non-top-down clear
>>>     requests. (sashiko)
>>>   - Skip dirty-tracker steering for offset-aligned requests so they 
>>> keep
>>>     their min_block_size alignment. (sashiko)
>>>   - sashiko reported that the __GFP_NOFAIL dirty-extent allocations on
>>>     the free path could deadlock during memory reclaim, since that is a
>>>     GFP_KERNEL allocation on the free path; move to a per-tracker
>>>     mempool so extent nodes are guaranteed without __GFP_NOFAIL.
>>>     (sashiko)
>>>   - Derive each free block's clear/dirty class from the blocks already
>>>     in hand on split, free, alloc, trim and init instead of querying 
>>> the
>>>     dirty tracker, removing the tracker lookups from the hot paths.
>>>
>>> v7:
>>>   - Preserve mixed-block clear state in __gpu_buddy_free() when a mixed
>>>     split child is re-merged after an undone split. (sashiko)
>>>   - Prefer a fully-clear block over a mixed one of the same order via a
>>>     single ordered clear-state max augment on free_tree[].
>>>
>>> v8:
>>>   - Coalesce contiguous dirty blocks in __gpu_buddy_free_list() into 
>>> one
>>>     dirty extent update instead of one mark_dirty() per block. 
>>> (Matthew)
>>>
>>> v9:
>>>   - Reset has_clear on allocation so a mixed block taken whole and 
>>> later
>>>     freed fully dirty is not re-tracked as mixed. (sashiko)
>>>
>>> Assisted-by: Claude:claude-opus-4-8
>>> Cc: Matthew Auld <matthew.auld@intel.com>
>>> Cc: Christian König <christian.koenig@amd.com>
>>> Signed-off-by: Arunpravin Paneer Selvam 
>>> <Arunpravin.PaneerSelvam@amd.com>
>>> ---
>>>   drivers/gpu/buddy.c                | 1370 
>>> ++++++++++++++++++++--------
>>>   drivers/gpu/tests/gpu_buddy_test.c |   32 +-
>>>   include/linux/gpu_buddy.h          |   97 +-
>>>   3 files changed, 1083 insertions(+), 416 deletions(-)
>>>
>>> diff --git a/drivers/gpu/buddy.c b/drivers/gpu/buddy.c
>>> index 4d5ac375a538..7d50156a5bc7 100644
>>> --- a/drivers/gpu/buddy.c
>>> +++ b/drivers/gpu/buddy.c
>>> @@ -8,6 +8,7 @@
>>>   #include <linux/kmemleak.h>
>>>   #include <linux/module.h>
>>>   #include <linux/sizes.h>
>>> +#include <linux/slab.h>
>>>   #include <linux/gpu_buddy.h>
>>> @@ -34,6 +35,447 @@
>>>   #endif
>>>   static struct kmem_cache *slab_blocks;
>>> +static struct kmem_cache *slab_extents;
>>> +
>>> +/*
>>> + * A single reserved extent suffices. Every allocation uses GFP_KERNEL
>>> + * from sleepable context and waits for reclaim, so it almost always
>>> + * succeeds; the reserve only backstops the rare NULL return without
>>> + * __GFP_NOFAIL and need not scale with extents added in one locked
>>> + * section (e.g. by gpu_buddy_reset_clear()).
>>> + */
>>> +#define GPU_DIRTY_EXTENT_POOL_MIN 1
>>> +
>>> +/*
>>> + * Dirty tracker
>>> + * -------------
>>> + *
>>> + * The dirty tracker maintains an augmented interval rbtree of 
>>> contiguous
>>> + * dirty address ranges, decoupled from the buddy free trees.
>>> + * Each node covers a maximal coalesced run; adjacent extents are 
>>> merged
>>> + * on insertion so the tree always holds the smallest possible 
>>> number of
>>> + * extents.  The augmentation field @subtree_max_size lets the 
>>> allocator
>>> + * locate the largest dirty extent in O(log E).
>>> + *
>>> + * Free trees (mm->free_tree[])
>>> + * ----------------------------
>>> + *
>>> + * Per-order augmented rbtrees of FREE buddy blocks, keyed by offset.
>>> + * Every node carries:
>>> + *   - subtree_max_alignment: largest natural alignment in the 
>>> subtree,
>>> + *     used by aligned/range allocations to skip unsuitable 
>>> subtrees in
>>> + *     O(log N).
>>> + *   - subtree_block_state: the highest clear class (DIRTY < MIXED 
>>> < CLEAR)
>>> + *     of any block in the subtree, maintained as a max augment. A 
>>> value of
>>> + *     >= MIXED means a clear-or-mixed block exists; == CLEAR means a
>>> + *     fully-clear block exists.
>>> + *
>>> + * Block classes
>>> + * -------------
>>> + *
>>> + * Each FREE block falls into one of three classes, determined in
>>> + * mark_free() by querying the dirty tracker for the block's range:
>>> + *
>>> + *   clear   -- HEADER_CLEAR set; no dirty extent overlaps the range.
>>> + *   mixed   -- HEADER_CLEAR unset; range has both dirty and clear 
>>> bytes.
>>> + *   dirty   -- HEADER_CLEAR unset; range is fully dirty.
>>> + *
>>> + * Clear allocation
>>> + * ----------------
>>> + *
>>> + * A clear (CLEAR_ALLOCATION) request prefers clear -> mixed -> dirty.
>>> + * Climbing from the requested order up to max_order, 
>>> rbtree_last_clear_free_block()
>>> + * returns, in one O(log N) descent per order, the right-most 
>>> clear- or-mixed block
>>> + * (fully-clear preferred over mixed) at the lowest order whose 
>>> free tree contains
>>> + * such a block. Only if no clear-or-mixed block exists at any 
>>> order >= the
>>> + * requested one does it fall back to a dirty block.
>>> + *
>>> + * Clear state is reported to the driver per whole block via 
>>> HEADER_CLEAR, so a
>>> + * fully-clear block of the requested order lets the driver skip 
>>> the clear pass.
>>> + *
>>> + * The effective allocation preference depends on how the driver 
>>> handles
>>> + * freed blocks:
>>> + *
>>> + *   1) Never clear on free:
>>> + *      No free block contains clear bytes, so clear allocations 
>>> always
>>> + *      fall back to dirty blocks.
>>> + *
>>> + *   2) Always clear on free:
>>> + *      Freed blocks become clear while untouched blocks remain dirty.
>>> + *      Merging clear and dirty buddies produces mixed blocks, 
>>> which are
>>> + *      reclassified when split. Over time, clear blocks become 
>>> dominant,
>>> + *      so clear allocations are typically satisfied from clear 
>>> blocks,
>>> + *      following a clear -> mixed -> dirty preference.
>>> + *
>>> + *   3) Selective clear on free:
>>> + *      For each order examined, fully-clear blocks are preferred over
>>> + *      mixed blocks, and mixed blocks are preferred over dirty 
>>> blocks.
>>> + *      If a clear or mixed block is found at an order, it is selected
>>> + *      without searching higher orders. Dirty blocks are used only 
>>> when
>>> + *      no clear or mixed block exists at any eligible order.
>>> + */
>>> +
>>> +static u64 extent_size(struct gpu_dirty_extent *dirty_extent)
>>> +{
>>> +    return dirty_extent->end - dirty_extent->start;
>>> +}
>>> +
>>> +RB_DECLARE_CALLBACKS_MAX(static, gpu_dirty_augment_cb,
>>> +             struct gpu_dirty_extent, rb,
>>> +             u64, subtree_max_size,
>>> +             extent_size)
>>> +
>>> +static struct gpu_dirty_extent *extent_alloc(struct 
>>> gpu_dirty_tracker *dirty_tracker)
>>> +{
>>> +    /*
>>> +     * The void free/reset paths must record an extent and cannot 
>>> handle
>>> +     * failure, so the mempool reserve guarantees a non-NULL return
>>> +     * without __GFP_NOFAIL. GFP_KERNEL is safe under the buddy 
>>> lock: no
>>> +     * driver frees buddy blocks from a shrinker, so reclaim cannot
>>> +     * recurse into the lock we hold.
>>> +     */
>>> +    return mempool_alloc(dirty_tracker->extent_pool, GFP_KERNEL);
>>> +}
>>> +
>>> +static void extent_free(struct gpu_dirty_tracker *dirty_tracker,
>>> +            struct gpu_dirty_extent *dirty_extent)
>>> +{
>>> +    mempool_free(dirty_extent, dirty_tracker->extent_pool);
>>> +}
>>> +
>>> +/* Return the rightmost extent whose start is strictly below 
>>> @offset. */
>>> +static struct gpu_dirty_extent *
>>> +prev_extent(struct gpu_dirty_tracker *dirty_tracker, u64 offset)
>>> +{
>>> +    struct rb_node *rb = dirty_tracker->root.rb_node;
>>> +    struct gpu_dirty_extent *dirty_extent = NULL;
>>> +
>>> +    while (rb) {
>>> +        struct gpu_dirty_extent *tmp_extent =
>>> +            rb_entry(rb, struct gpu_dirty_extent, rb);
>>> +
>>> +        if (tmp_extent->start < offset) {
>>> +            dirty_extent = tmp_extent;
>>> +            rb = rb->rb_right;
>>> +        } else {
>>> +            rb = rb->rb_left;
>>> +        }
>>> +    }
>>> +
>>> +    return dirty_extent;
>>> +}
>>> +
>>> +/* Return the leftmost extent whose start is at or above @offset. */
>>> +static struct gpu_dirty_extent *
>>> +next_extent(struct gpu_dirty_tracker *dirty_tracker, u64 offset)
>>> +{
>>> +    struct rb_node *rb = dirty_tracker->root.rb_node;
>>> +    struct gpu_dirty_extent *dirty_extent = NULL;
>>> +
>>> +    while (rb) {
>>> +        struct gpu_dirty_extent *tmp_extent =
>>> +            rb_entry(rb, struct gpu_dirty_extent, rb);
>>> +
>>> +        if (tmp_extent->start >= offset) {
>>> +            dirty_extent = tmp_extent;
>>> +            rb = rb->rb_left;
>>> +        } else {
>>> +            rb = rb->rb_right;
>>> +        }
>>> +    }
>>> +
>>> +    return dirty_extent;
>>> +}
>>> +
>>> +static void insert_extent(struct gpu_dirty_tracker *dirty_tracker,
>>> +              struct gpu_dirty_extent *dirty_extent)
>>> +{
>>> +    struct rb_node **link = &dirty_tracker->root.rb_node;
>>> +    struct rb_node *parent = NULL;
>>> +    u64 size = extent_size(dirty_extent);
>>> +
>>> +    while (*link) {
>>> +        struct gpu_dirty_extent *tmp_extent;
>>> +
>>> +        parent = *link;
>>> +        tmp_extent = rb_entry(parent, struct gpu_dirty_extent, rb);
>>> +
>>> +        if (tmp_extent->subtree_max_size < size)
>>> +            tmp_extent->subtree_max_size = size;
>>> +
>>> +        if (dirty_extent->start < tmp_extent->start)
>>> +            link = &parent->rb_left;
>>> +        else
>>> +            link = &parent->rb_right;
>>> +    }
>>> +
>>> +    dirty_extent->subtree_max_size = size;
>>> +    rb_link_node(&dirty_extent->rb, parent, link);
>>> +    rb_insert_augmented(&dirty_extent->rb, &dirty_tracker->root, 
>>> &gpu_dirty_augment_cb);
>>> +}
>>> +
>>> +static void remove_extent(struct gpu_dirty_tracker *dirty_tracker,
>>> +              struct gpu_dirty_extent *dirty_extent)
>>> +{
>>> +    rb_erase_augmented(&dirty_extent->rb, &dirty_tracker->root, 
>>> &gpu_dirty_augment_cb);
>>> +    RB_CLEAR_NODE(&dirty_extent->rb);
>>> +}
>>> +
>>> +static int gpu_dirty_tracker_init(struct gpu_dirty_tracker 
>>> *dirty_tracker)
>>> +{
>>> +    dirty_tracker->root = RB_ROOT;
>>> +    dirty_tracker->total_dirty = 0;
>>> +
>>> +    dirty_tracker->extent_pool =
>>> +        mempool_create_slab_pool(GPU_DIRTY_EXTENT_POOL_MIN, 
>>> slab_extents);
>>> +    if (!dirty_tracker->extent_pool)
>>> +        return -ENOMEM;
>>> +
>>> +    return 0;
>>> +}
>>> +
>>> +static void gpu_dirty_tracker_empty(struct gpu_dirty_tracker 
>>> *dirty_tracker)
>>> +{
>>> +    struct rb_node *rb;
>>> +
>>> +    while ((rb = rb_first(&dirty_tracker->root))) {
>>> +        struct gpu_dirty_extent *dirty_extent =
>>> +            rb_entry(rb, struct gpu_dirty_extent, rb);
>>> +
>>> +        remove_extent(dirty_tracker, dirty_extent);
>>> +        extent_free(dirty_tracker, dirty_extent);
>>> +    }
>>> +
>>> +    dirty_tracker->total_dirty = 0;
>>> +}
>>> +
>>> +static void gpu_dirty_tracker_fini(struct gpu_dirty_tracker 
>>> *dirty_tracker)
>>> +{
>>> +    gpu_dirty_tracker_empty(dirty_tracker);
>>> +    mempool_destroy(dirty_tracker->extent_pool);
>>> +    dirty_tracker->extent_pool = NULL;
>>> +}
>>> +
>>> +/*
>>> + * Mark the range [start, start + size] as dirty. Merge with the 
>>> neighbour on
>>> + * each side if they are contiguous, so the tree never holds two 
>>> adjacent ranges.
>>> + */
>>> +static void gpu_dirty_tracker_mark_dirty(struct gpu_dirty_tracker 
>>> *dirty_tracker,
>>> +                     u64 start, u64 size)
>>> +{
>>> +    struct gpu_dirty_extent *left, *right, *dirty_extent;
>>> +    u64 end = start + size;
>>> +
>>> +    if (!size)
>>> +        return;
>>> +
>>> +    /* Find contiguous neighbours, if any. */
>>> +    left = prev_extent(dirty_tracker, start);
>>> +    if (left && left->end != start)
>>> +        left = NULL;
>>> +
>>> +    right = next_extent(dirty_tracker, end);
>>> +    if (right && right->start != end)
>>> +        right = NULL;
>>> +
>>> +    if (left && right) {
>>> +        /* Merge left + new + right into a single extent. */
>>> +        remove_extent(dirty_tracker, left);
>>> +        remove_extent(dirty_tracker, right);
>>> +        left->end = right->end;
>>> +        extent_free(dirty_tracker, right);
>>> +        insert_extent(dirty_tracker, left);
>>> +    } else if (left) {
>>> +        /* Extend left neighbour rightwards. */
>>> +        remove_extent(dirty_tracker, left);
>>> +        left->end = end;
>>> +        insert_extent(dirty_tracker, left);
>>> +    } else if (right) {
>>> +        /* Extend right neighbour leftwards. */
>>> +        remove_extent(dirty_tracker, right);
>>> +        right->start = start;
>>> +        insert_extent(dirty_tracker, right);
>>> +    } else {
>>> +        /* Standalone extent. */
>>> +        dirty_extent = extent_alloc(dirty_tracker);
>>> +        dirty_extent->start = start;
>>> +        dirty_extent->end   = end;
>>> +        insert_extent(dirty_tracker, dirty_extent);
>>> +    }
>>> +
>>> +    dirty_tracker->total_dirty += size;
>>> +}
>>> +
>>> +/*
>>> + * Remove the range [start, start + size] from the dirty tracker. 
>>> Punch the
>>> + * range out of every overlapping dirty extent, splitting one 
>>> extent in two if
>>> + * the removed range falls strictly inside it.
>>> + */
>>> +static void gpu_dirty_tracker_remove_range(struct gpu_dirty_tracker 
>>> *dirty_tracker,
>>> +                       u64 start, u64 size)
>>> +{
>>> +    struct gpu_dirty_extent *dirty_extent, *next;
>>> +    u64 end = start + size;
>>> +
>>> +    if (!size)
>>> +        return;
>>> +
>>> +    dirty_extent = prev_extent(dirty_tracker, start + 1);
>>> +    if (!dirty_extent)
>>> +        dirty_extent = next_extent(dirty_tracker, start);
>>> +
>>> +    while (dirty_extent && dirty_extent->start < end) {
>>> +        struct rb_node *next_node = rb_next(&dirty_extent->rb);
>>> +        u64 extent_start = dirty_extent->start;
>>> +        u64 extent_end = dirty_extent->end;
>>> +
>>> +        if (next_node)
>>> +            next = rb_entry(next_node, struct gpu_dirty_extent, rb);
>>> +        else
>>> +            next = NULL;
>>> +
>>> +        /* Skip a non-overlapping neighbour returned by 
>>> prev_extent(). */
>>> +        if (extent_end <= start) {
>>> +            dirty_extent = next;
>>> +            continue;
>>> +        }
>>> +
>>> +        if (extent_start < start && extent_end > end) {
>>> +            /*
>>> +             * Removed range lies strictly inside this dirty extent:
>>> +             * split it into the dirty left and right halves.
>>> +             */
>>> +            struct gpu_dirty_extent *right = 
>>> extent_alloc(dirty_tracker);
>>> +
>>> +            remove_extent(dirty_tracker, dirty_extent);
>>> +
>>> +            dirty_extent->end = start;
>>> +            right->start = end;
>>> +            right->end   = extent_end;
>>> +
>>> +            insert_extent(dirty_tracker, dirty_extent);
>>> +            insert_extent(dirty_tracker, right);
>>> +
>>> +            dirty_tracker->total_dirty -= size;
>>> +        } else if (extent_start >= start && extent_end <= end) {
>>> +            /* Extent fully covered: drop it. */
>>> +            remove_extent(dirty_tracker, dirty_extent);
>>> +            extent_free(dirty_tracker, dirty_extent);
>>> +
>>> +            dirty_tracker->total_dirty -= (extent_end - extent_start);
>>> +        } else if (extent_start < start) {
>>> +            /* Extent overlaps from the left: trim its right end. */
>>> +            remove_extent(dirty_tracker, dirty_extent);
>>> +            dirty_extent->end = start;
>>> +            insert_extent(dirty_tracker, dirty_extent);
>>> +
>>> +            dirty_tracker->total_dirty -= (extent_end - start);
>>> +        } else {
>>> +            /* Extent overlaps from the right: trim its left end. */
>>> +            remove_extent(dirty_tracker, dirty_extent);
>>> +            dirty_extent->start = end;
>>> +            insert_extent(dirty_tracker, dirty_extent);
>>> +
>>> +            dirty_tracker->total_dirty -= (end - extent_start);
>>> +        }
>>> +
>>> +        dirty_extent = next;
>>> +    }
>>> +}
>>> +
>>> +static enum gpu_block_state
>>> +gpu_dirty_range_state(struct gpu_dirty_tracker *dirty_tracker,
>>> +              u64 start, u64 size)
>>> +{
>>> +    struct gpu_dirty_extent *dirty_extent;
>>> +    u64 end = start + size;
>>> +
>>> +    dirty_extent = prev_extent(dirty_tracker, start + 1);
>>> +    if (dirty_extent) {
>>> +        if (dirty_extent->start <= start && dirty_extent->end >= end)
>>> +            return GPU_BLOCK_DIRTY;
>>> +        if (dirty_extent->start < end && dirty_extent->end > start)
>>> +            return GPU_BLOCK_MIXED;
>>> +    }
>>> +
>>> +    dirty_extent = next_extent(dirty_tracker, start);
>>> +    if (dirty_extent && dirty_extent->start < end)
>>> +        return GPU_BLOCK_MIXED;
>>> +
>>> +    return GPU_BLOCK_CLEAR;
>>> +}
>>> +
>>> +static struct rb_node *
>>> +dirty_tracker_descend_right(struct rb_node *node, u64 min_size)
>>> +{
>>> +    while (node->rb_right) {
>>> +        struct gpu_dirty_extent *tmp_extent;
>>> +
>>> +        tmp_extent = rb_entry(node->rb_right, struct 
>>> gpu_dirty_extent, rb);
>>> +
>>> +        if (tmp_extent->subtree_max_size < min_size)
>>> +            break;
>>> +        node = node->rb_right;
>>> +    }
>>> +
>>> +    return node;
>>> +}
>>> +
>>> +static struct gpu_dirty_extent *
>>> +gpu_dirty_tracker_find(struct gpu_dirty_tracker *dirty_tracker,
>>> +               u64 min_size, u64 *aligned_start_out)
>>> +{
>>> +    struct rb_node *rb = dirty_tracker->root.rb_node;
>>> +    struct gpu_dirty_extent *root_extent;
>>> +    struct rb_node *parent;
>>> +
>>> +    if (!min_size || !is_power_of_2(min_size))
>>> +        return NULL;
>>> +
>>> +    if (!rb)
>>> +        return NULL;
>>> +
>>> +    root_extent = rb_entry(rb, struct gpu_dirty_extent, rb);
>>> +    if (root_extent->subtree_max_size < min_size)
>>> +        return NULL;
>>> +
>>> +    rb = dirty_tracker_descend_right(rb, min_size);
>>> +
>>> +    while (rb) {
>>> +        struct gpu_dirty_extent *dirty_extent;
>>> +        u64 aligned_start;
>>> +
>>> +        dirty_extent = rb_entry(rb, struct gpu_dirty_extent, rb);
>>> +        aligned_start = ALIGN(dirty_extent->start, min_size);
>>> +
>>> +        /* Check if a min_size block fits after the alignment skip. */
>>> +        if (aligned_start <= dirty_extent->end &&
>>> +            dirty_extent->end - aligned_start >= min_size) {
>>> +            *aligned_start_out = aligned_start;
>>> +            return dirty_extent;
>>> +        }
>>> +
>>> +        if (rb->rb_left) {
>>> +            struct gpu_dirty_extent *tmp_extent;
>>> +
>>> +            tmp_extent = rb_entry(rb->rb_left, struct 
>>> gpu_dirty_extent, rb);
>>> +            if (tmp_extent->subtree_max_size >= min_size) {
>>> +                rb = dirty_tracker_descend_right(rb->rb_left, 
>>> min_size);
>>> +                continue;
>>> +            }
>>> +        }
>>> +
>>> +        /* Walk up until we exit a node via its right child. */
>>> +        parent = rb_parent(rb);
>>> +        while (parent && parent->rb_right != rb) {
>>> +            rb = parent;
>>> +            parent = rb_parent(rb);
>>> +        }
>>> +        rb = parent;
>>> +    }
>>> +
>>> +    return NULL;
>>> +}
>>>   static unsigned int
>>>   gpu_buddy_block_state(struct gpu_buddy_block *block)
>>> @@ -67,10 +509,97 @@ static unsigned int 
>>> gpu_buddy_block_offset_alignment(struct gpu_buddy_block *blo
>>>       return __ffs64(offset);
>>>   }
>>> -RB_DECLARE_CALLBACKS_MAX(static, gpu_buddy_augment_cb,
>>> -             struct gpu_buddy_block, rb,
>>> -             unsigned int, subtree_max_alignment,
>>> -             gpu_buddy_block_offset_alignment);
>>> +static inline enum gpu_block_state
>>> +gpu_block_cached_state(struct gpu_buddy_block *block)
>>> +{
>>> +    if (gpu_buddy_block_is_clear(block))
>>> +        return GPU_BLOCK_CLEAR;
>>> +    if (block->has_clear)
>>> +        return GPU_BLOCK_MIXED;
>>> +    return GPU_BLOCK_DIRTY;
>>> +}
>>> +
>>> +static inline void gpu_buddy_augment_compute(struct gpu_buddy_block 
>>> *block)
>>> +{
>>> +    enum gpu_block_state block_state;
>>> +    struct gpu_buddy_block *right;
>>> +    struct gpu_buddy_block *left;
>>> +    unsigned int max_align;
>>> +
>>> +    max_align = gpu_buddy_block_offset_alignment(block);
>>> +    block_state = gpu_block_cached_state(block);
>>> +
>>> +    left = rb_entry_safe(block->rb.rb_left, struct gpu_buddy_block, 
>>> rb);
>>> +    if (left) {
>>> +        if (left->subtree_max_alignment > max_align)
>>> +            max_align = left->subtree_max_alignment;
>>> +
>>> +        block_state = max(block_state, left->subtree_block_state);
>>> +    }
>>> +
>>> +    right = rb_entry_safe(block->rb.rb_right, struct 
>>> gpu_buddy_block, rb);
>>> +    if (right) {
>>> +        if (right->subtree_max_alignment > max_align)
>>> +            max_align = right->subtree_max_alignment;
>>> +
>>> +        block_state = max(block_state, right->subtree_block_state);
>>> +    }
>>> +
>>> +    block->subtree_max_alignment = max_align;
>>> +    block->subtree_block_state = block_state;
>>> +}
>>> +
>>> +static void gpu_buddy_augment_propagate(struct rb_node *rb, struct 
>>> rb_node *stop)
>>> +{
>>> +    while (rb != stop) {
>>> +        struct gpu_buddy_block *block;
>>> +        unsigned int old_align;
>>> +        enum gpu_block_state old_block_state;
>>> +
>>> +        block = rb_entry(rb, struct gpu_buddy_block, rb);
>>> +        old_align = block->subtree_max_alignment;
>>> +        old_block_state = block->subtree_block_state;
>>> +
>>> +        gpu_buddy_augment_compute(block);
>>> +        if (block->subtree_max_alignment == old_align &&
>>> +            block->subtree_block_state == old_block_state)
>>> +            break;
>>> +
>>> +        rb = rb_parent(&block->rb);
>>> +    }
>>> +}
>>> +
>>> +static void gpu_buddy_augment_copy(struct rb_node *rb_old, struct 
>>> rb_node *rb_new)
>>> +{
>>> +    struct gpu_buddy_block *old;
>>> +    struct gpu_buddy_block *new;
>>> +
>>> +    old = rb_entry(rb_old, struct gpu_buddy_block, rb);
>>> +    new = rb_entry(rb_new, struct gpu_buddy_block, rb);
>>> +
>>> +    new->subtree_max_alignment = old->subtree_max_alignment;
>>> +    new->subtree_block_state = old->subtree_block_state;
>>> +}
>>> +
>>> +static void gpu_buddy_augment_rotate(struct rb_node *rb_old, struct 
>>> rb_node *rb_new)
>>> +{
>>> +    struct gpu_buddy_block *old;
>>> +    struct gpu_buddy_block *new;
>>> +
>>> +    old = rb_entry(rb_old, struct gpu_buddy_block, rb);
>>> +    new = rb_entry(rb_new, struct gpu_buddy_block, rb);
>>> +
>>> +    new->subtree_max_alignment = old->subtree_max_alignment;
>>> +    new->subtree_block_state = old->subtree_block_state;
>>> +
>>> +    gpu_buddy_augment_compute(old);
>>> +}
>>> +
>>> +static const struct rb_augment_callbacks gpu_buddy_augment_cb = {
>>> +    .propagate = gpu_buddy_augment_propagate,
>>> +    .copy      = gpu_buddy_augment_copy,
>>> +    .rotate    = gpu_buddy_augment_rotate,
>>> +};
>>>   static struct gpu_buddy_block *gpu_block_alloc(struct gpu_buddy *mm,
>>>                              struct gpu_buddy_block *parent,
>>> @@ -81,6 +610,10 @@ static struct gpu_buddy_block 
>>> *gpu_block_alloc(struct gpu_buddy *mm,
>>>       BUG_ON(order > GPU_BUDDY_MAX_ORDER);
>>> +    /*
>>> +     * GFP_KERNEL is safe under the buddy lock: no consumer runs a
>>> +     * shrinker that re-enters it during direct reclaim.
>>> +     */
>>>       block = kmem_cache_zalloc(slab_blocks, GFP_KERNEL);
>>>       if (!block)
>>>           return NULL;
>>> @@ -101,13 +634,6 @@ static void gpu_block_free(struct gpu_buddy *mm,
>>>       kmem_cache_free(slab_blocks, block);
>>>   }
>>> -static enum gpu_buddy_free_tree
>>> -get_block_tree(struct gpu_buddy_block *block)
>>> -{
>>> -    return gpu_buddy_block_is_clear(block) ?
>>> -           GPU_BUDDY_CLEAR_TREE : GPU_BUDDY_DIRTY_TREE;
>>> -}
>>> -
>>>   static struct gpu_buddy_block *
>>>   rbtree_get_free_block(const struct rb_node *node)
>>>   {
>>> @@ -120,24 +646,64 @@ rbtree_last_free_block(struct rb_root *root)
>>>       return rbtree_get_free_block(rb_last(root));
>>>   }
>>> -static bool rbtree_is_empty(struct rb_root *root)
>>> +static struct gpu_buddy_block *
>>> +rbtree_last_clear_free_block(struct rb_root *root,
>>> +                 enum gpu_block_state min_block_state)
>>> +{
>>> +    struct rb_node *node = root->rb_node;
>>> +    struct gpu_buddy_block *block = NULL;
>>> +    struct gpu_buddy_block *root_block;
>>> +    enum gpu_block_state target_state;
>>> +
>>> +    root_block = rbtree_get_free_block(node);
>>> +    if (!root_block || root_block->subtree_block_state < 
>>> min_block_state)
>>> +        return NULL;
>>> +
>>> +    target_state = root_block->subtree_block_state;
>>> +
>>> +    while (node) {
>>> +        struct gpu_buddy_block *right_block;
>>> +        struct gpu_buddy_block *node_block;
>>> +
>>> +        node_block = rbtree_get_free_block(node);
>>> +        right_block = rbtree_get_free_block(node->rb_right);
>>> +
>>> +        if (right_block && right_block->subtree_block_state >= 
>>> target_state) {
>>> +            node = node->rb_right;
>>> +            continue;
>>> +        }
>>> +
>>> +        if (gpu_block_cached_state(node_block) == target_state) {
>>> +            block = node_block;
>>> +            break;
>>> +        }
>>> +
>>> +        node = node->rb_left;
>>> +    }
>>> +
>>> +    return block;
>>> +}
>>> +
>>> +static inline void gpu_buddy_sync_clear_avail(struct gpu_buddy *mm)
>>>   {
>>> -    return RB_EMPTY_ROOT(root);
>>> +    mm->clear_avail = mm->avail - mm->dirty.total_dirty;
>>>   }
>>>   static void rbtree_insert(struct gpu_buddy *mm,
>>> -              struct gpu_buddy_block *block,
>>> -              enum gpu_buddy_free_tree tree)
>>> +              struct gpu_buddy_block *block)
>>>   {
>>>       struct rb_node **link, *parent = NULL;
>>> -    unsigned int block_alignment, order;
>>> +    enum gpu_block_state block_state;
>>>       struct gpu_buddy_block *node;
>>> +    unsigned int block_alignment;
>>>       struct rb_root *root;
>>> +    unsigned int order;
>>>       order = gpu_buddy_block_order(block);
>>>       block_alignment = gpu_buddy_block_offset_alignment(block);
>>> +    block_state = gpu_block_cached_state(block);
>>> -    root = &mm->free_trees[tree][order];
>>> +    root = &mm->free_tree[order];
>>>       link = &root->rb_node;
>>>       while (*link) {
>>> @@ -147,10 +713,12 @@ static void rbtree_insert(struct gpu_buddy *mm,
>>>            * Manual augmentation update during insertion traversal. 
>>> Required
>>>            * because rb_insert_augmented() only calls rotate 
>>> callback during
>>>            * rotations. This ensures all ancestors on the insertion 
>>> path have
>>> -         * correct subtree_max_alignment values.
>>> +         * correct subtree_max_alignment / subtree_block_state values.
>>>            */
>>>           if (node->subtree_max_alignment < block_alignment)
>>>               node->subtree_max_alignment = block_alignment;
>>> +        if (node->subtree_block_state < block_state)
>>> +            node->subtree_block_state = block_state;
>>>           if (gpu_buddy_block_offset(block) < 
>>> gpu_buddy_block_offset(node))
>>>               link = &parent->rb_left;
>>> @@ -159,6 +727,7 @@ static void rbtree_insert(struct gpu_buddy *mm,
>>>       }
>>>       block->subtree_max_alignment = block_alignment;
>>> +    block->subtree_block_state = block_state;
>>>       rb_link_node(&block->rb, parent, link);
>>>       rb_insert_augmented(&block->rb, root, &gpu_buddy_augment_cb);
>>>   }
>>> @@ -167,53 +736,55 @@ static void rbtree_remove(struct gpu_buddy *mm,
>>>                 struct gpu_buddy_block *block)
>>>   {
>>>       unsigned int order = gpu_buddy_block_order(block);
>>> -    enum gpu_buddy_free_tree tree;
>>> -    struct rb_root *root;
>>> -
>>> -    tree = get_block_tree(block);
>>> -    root = &mm->free_trees[tree][order];
>>> -    rb_erase_augmented(&block->rb, root, &gpu_buddy_augment_cb);
>>> +    rb_erase_augmented(&block->rb, &mm->free_tree[order], 
>>> &gpu_buddy_augment_cb);
>>>       RB_CLEAR_NODE(&block->rb);
>>>   }
>>> -static void clear_reset(struct gpu_buddy_block *block)
>>> -{
>>> -    block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> -}
>>> -
>>> -static void mark_cleared(struct gpu_buddy_block *block)
>>> -{
>>> -    block->header |= GPU_BUDDY_HEADER_CLEAR;
>>> -}
>>> -
>>>   static void mark_allocated(struct gpu_buddy *mm,
>>>                  struct gpu_buddy_block *block)
>>>   {
>>>       block->header &= ~GPU_BUDDY_HEADER_STATE;
>>>       block->header |= GPU_BUDDY_ALLOCATED;
>>> +    block->has_clear = false;
>>> +
>>>       mm->free_scoreboard[gpu_buddy_block_order(block)]--;
>>>       mm->used_scoreboard[gpu_buddy_block_order(block)]++;
>>>       rbtree_remove(mm, block);
>>>   }
>>> -static void mark_free(struct gpu_buddy *mm,
>>> -              struct gpu_buddy_block *block)
>>> +static void __mark_free(struct gpu_buddy *mm,
>>> +            struct gpu_buddy_block *block,
>>> +            enum gpu_block_state block_state)
>>>   {
>>> -    enum gpu_buddy_free_tree tree;
>>> -
>>>       if (gpu_buddy_block_is_allocated(block))
>>> mm->used_scoreboard[gpu_buddy_block_order(block)]--;
>>>       block->header &= ~GPU_BUDDY_HEADER_STATE;
>>>       block->header |= GPU_BUDDY_FREE;
>>> +    block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> +
>>> +    block->has_clear = (block_state != GPU_BLOCK_DIRTY);
>>> +    if (block_state == GPU_BLOCK_CLEAR)
>>> +        block->header |= GPU_BUDDY_HEADER_CLEAR;
>>> +
>>>       mm->free_scoreboard[gpu_buddy_block_order(block)]++;
>>> -    tree = get_block_tree(block);
>>> -    rbtree_insert(mm, block, tree);
>>> +    rbtree_insert(mm, block);
>>> +}
>>> +
>>> +static void mark_free(struct gpu_buddy *mm,
>>> +              struct gpu_buddy_block *block)
>>> +{
>>> +    enum gpu_block_state block_state;
>>> +
>>> +    block_state = gpu_dirty_range_state(&mm->dirty,
>>> +                        gpu_buddy_block_offset(block),
>>> +                        gpu_buddy_block_size(mm, block));
>>> +    __mark_free(mm, block, block_state);
>>>   }
>>>   static void mark_split(struct gpu_buddy *mm,
>>> @@ -253,37 +824,31 @@ __get_buddy(struct gpu_buddy_block *block)
>>>   }
>>>   static unsigned int __gpu_buddy_free(struct gpu_buddy *mm,
>>> -                     struct gpu_buddy_block *block,
>>> -                     bool force_merge)
>>> +                     struct gpu_buddy_block *block)
>>>   {
>>> +    enum gpu_block_state block_state;
>>>       struct gpu_buddy_block *parent;
>>>       unsigned int order;
>>> -    while ((parent = block->parent)) {
>>> -        struct gpu_buddy_block *buddy;
>>> +    block_state = gpu_block_cached_state(block);
>>> -        buddy = __get_buddy(block);
>>> +    while ((parent = block->parent)) {
>>> +        struct gpu_buddy_block *buddy = __get_buddy(block);
>>>           if (!gpu_buddy_block_is_free(buddy))
>>>               break;
>>> -        if (!force_merge) {
>>> -            /*
>>> -             * Check the block and its buddy clear state and exit
>>> -             * the loop if they both have the dissimilar state.
>>> -             */
>>> -            if (gpu_buddy_block_is_clear(block) !=
>>> -                gpu_buddy_block_is_clear(buddy))
>>> -                break;
>>> +        if (block_state != GPU_BLOCK_MIXED) {
>>> +            enum gpu_block_state buddy_state;
>>> +
>>> +            buddy_state = gpu_block_cached_state(buddy);
>>> -            if (gpu_buddy_block_is_clear(block))
>>> -                mark_cleared(parent);
>>> +            if (buddy_state != block_state)
>>> +                block_state = GPU_BLOCK_MIXED;
>>>           }
>>>           rbtree_remove(mm, buddy);
>>> mm->free_scoreboard[gpu_buddy_block_order(buddy)]--;
>>> -        if (force_merge && gpu_buddy_block_is_clear(buddy))
>>> -            mm->clear_avail -= gpu_buddy_block_size(mm, buddy);
>>>           if (gpu_buddy_block_is_allocated(block))
>>> mm->used_scoreboard[gpu_buddy_block_order(block)]--;
>>> @@ -295,74 +860,11 @@ static unsigned int __gpu_buddy_free(struct 
>>> gpu_buddy *mm,
>>>       }
>>>       order = gpu_buddy_block_order(block);
>>> -    mark_free(mm, block);
>>> +    __mark_free(mm, block, block_state);
>>>       return order;
>>>   }
>>> -static int __force_merge(struct gpu_buddy *mm,
>>> -             u64 start,
>>> -             u64 end,
>>> -             unsigned int min_order)
>>> -{
>>> -    unsigned int tree, order;
>>> -    int i;
>>> -
>>> -    if (!min_order)
>>> -        return -ENOMEM;
>>> -
>>> -    if (min_order > mm->max_order)
>>> -        return -EINVAL;
>>> -
>>> -    for_each_free_tree(tree) {
>>> -        for (i = min_order - 1; i >= 0; i--) {
>>> -            struct rb_node *iter = rb_last(&mm->free_trees[tree][i]);
>>> -
>>> -            while (iter) {
>>> -                struct gpu_buddy_block *block, *buddy;
>>> -                u64 block_start, block_end;
>>> -
>>> -                block = rbtree_get_free_block(iter);
>>> -                iter = rb_prev(iter);
>>> -
>>> -                if (!block || !block->parent)
>>> -                    continue;
>>> -
>>> -                block_start = gpu_buddy_block_offset(block);
>>> -                block_end = block_start + gpu_buddy_block_size(mm, 
>>> block) - 1;
>>> -
>>> -                if (!contains(start, end, block_start, block_end))
>>> -                    continue;
>>> -
>>> -                buddy = __get_buddy(block);
>>> -                if (!gpu_buddy_block_is_free(buddy))
>>> -                    continue;
>>> -
>>> - gpu_buddy_assert(gpu_buddy_block_is_clear(block) !=
>>> -                         gpu_buddy_block_is_clear(buddy));
>>> -
>>> -                /*
>>> -                 * Advance to the next node when the current node 
>>> is the buddy,
>>> -                 * as freeing the block will also remove its buddy 
>>> from the tree.
>>> -                 */
>>> -                if (iter == &buddy->rb)
>>> -                    iter = rb_prev(iter);
>>> -
>>> -                rbtree_remove(mm, block);
>>> - mm->free_scoreboard[gpu_buddy_block_order(block)]--;
>>> -                if (gpu_buddy_block_is_clear(block))
>>> -                    mm->clear_avail -= gpu_buddy_block_size(mm, 
>>> block);
>>> -
>>> -                order = __gpu_buddy_free(mm, block, true);
>>> -                if (order >= min_order)
>>> -                    return 0;
>>> -            }
>>> -        }
>>> -    }
>>> -
>>> -    return -ENOMEM;
>>> -}
>>> -
>>>   /**
>>>    * gpu_buddy_init - init memory manager
>>>    *
>>> @@ -377,7 +879,7 @@ static int __force_merge(struct gpu_buddy *mm,
>>>    */
>>>   int gpu_buddy_init(struct gpu_buddy *mm, u64 size, u64 chunk_size)
>>>   {
>>> -    unsigned int i, j, root_count = 0;
>>> +    unsigned int root_count = 0;
>>>       u64 offset = 0;
>>>       if (size < chunk_size)
>>> @@ -411,22 +913,14 @@ int gpu_buddy_init(struct gpu_buddy *mm, u64 
>>> size, u64 chunk_size)
>>>       if (!mm->used_scoreboard)
>>>           goto out_free_free_scoreboard;
>>> -    mm->free_trees = kmalloc_array(GPU_BUDDY_MAX_FREE_TREES,
>>> -                       sizeof(*mm->free_trees),
>>> -                       GFP_KERNEL);
>>> -    if (!mm->free_trees)
>>> +    mm->free_tree = kcalloc(mm->max_order + 1,
>>> +                sizeof(struct rb_root),
>>> +                GFP_KERNEL);
>>> +    if (!mm->free_tree)
>>>           goto out_free_used_scoreboard;
>>> -    for_each_free_tree(i) {
>>> -        mm->free_trees[i] = kmalloc_array(mm->max_order + 1,
>>> -                          sizeof(struct rb_root),
>>> -                          GFP_KERNEL);
>>> -        if (!mm->free_trees[i])
>>> -            goto out_free_tree;
>>> -
>>> -        for (j = 0; j <= mm->max_order; ++j)
>>> -            mm->free_trees[i][j] = RB_ROOT;
>>> -    }
>>> +    if (gpu_dirty_tracker_init(&mm->dirty))
>>> +        goto out_free_tree;
>>>       mm->n_roots = hweight64(size);
>>> @@ -452,7 +946,8 @@ int gpu_buddy_init(struct gpu_buddy *mm, u64 
>>> size, u64 chunk_size)
>>>           if (!root)
>>>               goto out_free_roots;
>>> -        mark_free(mm, root);
>>> +        gpu_dirty_tracker_mark_dirty(&mm->dirty, offset, root_size);
>>> +        __mark_free(mm, root, GPU_BLOCK_DIRTY);
>>>           BUG_ON(root_count > mm->max_order);
>>>           BUG_ON(gpu_buddy_block_size(mm, root) < chunk_size);
>>> @@ -474,9 +969,8 @@ int gpu_buddy_init(struct gpu_buddy *mm, u64 
>>> size, u64 chunk_size)
>>>           gpu_block_free(mm, mm->roots[root_count]);
>>>       kfree(mm->roots);
>>>   out_free_tree:
>>> -    while (i--)
>>> -        kfree(mm->free_trees[i]);
>>> -    kfree(mm->free_trees);
>>> +    gpu_dirty_tracker_fini(&mm->dirty);
>>> +    kfree(mm->free_tree);
>>>   out_free_used_scoreboard:
>>>       kfree(mm->used_scoreboard);
>>>   out_free_free_scoreboard:
>>> @@ -494,7 +988,7 @@ EXPORT_SYMBOL(gpu_buddy_init);
>>>    */
>>>   void gpu_buddy_fini(struct gpu_buddy *mm)
>>>   {
>>> -    u64 root_size, size, start;
>>> +    u64 root_size, size;
>>>       unsigned int order;
>>>       int i;
>>> @@ -502,14 +996,10 @@ void gpu_buddy_fini(struct gpu_buddy *mm)
>>>       for (i = 0; i < mm->n_roots; ++i) {
>>>           order = ilog2(size) - ilog2(mm->chunk_size);
>>> -        start = gpu_buddy_block_offset(mm->roots[i]);
>>> -        __force_merge(mm, start, start + size, order);
>>> +        root_size = mm->chunk_size << order;
>>> gpu_buddy_assert(gpu_buddy_block_is_free(mm->roots[i]));
>>> -
>>>           gpu_block_free(mm, mm->roots[i]);
>>> -
>>> -        root_size = mm->chunk_size << order;
>>>           size -= root_size;
>>>       }
>>> @@ -518,9 +1008,8 @@ void gpu_buddy_fini(struct gpu_buddy *mm)
>>>       for (i = 0; i <= mm->max_order; ++i)
>>>           gpu_buddy_assert(!mm->used_scoreboard[i]);
>>> -    for_each_free_tree(i)
>>> -        kfree(mm->free_trees[i]);
>>> -    kfree(mm->free_trees);
>>> +    gpu_dirty_tracker_fini(&mm->dirty);
>>> +    kfree(mm->free_tree);
>>>       kfree(mm->roots);
>>>       kfree(mm->free_scoreboard);
>>>       kfree(mm->used_scoreboard);
>>> @@ -532,6 +1021,7 @@ static int split_block(struct gpu_buddy *mm,
>>>   {
>>>       unsigned int block_order = gpu_buddy_block_order(block) - 1;
>>>       u64 offset = gpu_buddy_block_offset(block);
>>> +    enum gpu_block_state parent_state;
>>>       BUG_ON(!gpu_buddy_block_is_free(block));
>>>       BUG_ON(!gpu_buddy_block_order(block));
>>> @@ -547,17 +1037,18 @@ static int split_block(struct gpu_buddy *mm,
>>>           return -ENOMEM;
>>>       }
>>> +    parent_state = gpu_block_cached_state(block);
>>> +
>>>       mark_split(mm, block);
>>> -    if (gpu_buddy_block_is_clear(block)) {
>>> -        mark_cleared(block->left);
>>> -        mark_cleared(block->right);
>>> -        clear_reset(block);
>>> +    if (parent_state == GPU_BLOCK_MIXED) {
>>> +        mark_free(mm, block->left);
>>> +        mark_free(mm, block->right);
>>> +    } else {
>>> +        __mark_free(mm, block->left, parent_state);
>>> +        __mark_free(mm, block->right, parent_state);
>>>       }
>>> -    mark_free(mm, block->left);
>>> -    mark_free(mm, block->right);
>>> -
>>>       return 0;
>>>   }
>>> @@ -572,45 +1063,55 @@ static int split_block(struct gpu_buddy *mm,
>>>    */
>>>   void gpu_buddy_reset_clear(struct gpu_buddy *mm, bool is_clear)
>>>   {
>>> -    enum gpu_buddy_free_tree src_tree, dst_tree;
>>> -    u64 root_size, size, start;
>>> -    unsigned int order;
>>> -    int i;
>>> +    unsigned int i;
>>>       gpu_buddy_driver_lock_held(mm);
>>> -    size = mm->size;
>>> -    for (i = 0; i < mm->n_roots; ++i) {
>>> -        order = ilog2(size) - ilog2(mm->chunk_size);
>>> -        start = gpu_buddy_block_offset(mm->roots[i]);
>>> -        __force_merge(mm, start, start + size, order);
>>> -        root_size = mm->chunk_size << order;
>>> -        size -= root_size;
>>> -    }
>>> -
>>> -    src_tree = is_clear ? GPU_BUDDY_DIRTY_TREE : GPU_BUDDY_CLEAR_TREE;
>>> -    dst_tree = is_clear ? GPU_BUDDY_CLEAR_TREE : GPU_BUDDY_DIRTY_TREE;
>>> +    gpu_dirty_tracker_empty(&mm->dirty);
>>>       for (i = 0; i <= mm->max_order; ++i) {
>>> -        struct rb_root *root = &mm->free_trees[src_tree][i];
>>>           struct gpu_buddy_block *block, *tmp;
>>> -        rbtree_postorder_for_each_entry_safe(block, tmp, root, rb) {
>>> -            rbtree_remove(mm, block);
>>> +        rbtree_postorder_for_each_entry_safe(block, tmp,
>>> +                             &mm->free_tree[i], rb) {
>>>               if (is_clear) {
>>> -                mark_cleared(block);
>>> -                mm->clear_avail += gpu_buddy_block_size(mm, block);
>>> +                if (!gpu_buddy_block_is_clear(block))
>>> +                    block->header |= GPU_BUDDY_HEADER_CLEAR;
>>> +                block->has_clear = true;
>>> +            } else if (gpu_buddy_block_is_clear(block)) {
>>> +                block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> +                block->has_clear = false;
>>> + gpu_dirty_tracker_mark_dirty(&mm->dirty,
>>> + gpu_buddy_block_offset(block),
>>> +                                 gpu_buddy_block_size(mm, block));
>>>               } else {
>>> -                clear_reset(block);
>>> -                mm->clear_avail -= gpu_buddy_block_size(mm, block);
>>> +                block->has_clear = false;
>>> + gpu_dirty_tracker_mark_dirty(&mm->dirty,
>>> + gpu_buddy_block_offset(block),
>>> +                                 gpu_buddy_block_size(mm, block));
>>>               }
>>> -            rbtree_insert(mm, block, dst_tree);
>>> +            gpu_buddy_augment_compute(block);
>>>           }
>>>       }
>>> +
>>> +    gpu_buddy_sync_clear_avail(mm);
>>>   }
>>>   EXPORT_SYMBOL(gpu_buddy_reset_clear);
>>> +static void __gpu_buddy_free_block_internal(struct gpu_buddy *mm,
>>> +                        struct gpu_buddy_block *block)
>>> +{
>>> +    u64 size = gpu_buddy_block_size(mm, block);
>>> +
>>> +    gpu_buddy_driver_lock_held(mm);
>>> +    BUG_ON(!gpu_buddy_block_is_allocated(block));
>>> +
>>> +    mm->avail += size;
>>> +    gpu_buddy_sync_clear_avail(mm);
>>> +    __gpu_buddy_free(mm, block);
>>> +}
>>> +
>>>   /**
>>>    * gpu_buddy_free_block - free a block
>>>    *
>>> @@ -620,13 +1121,12 @@ EXPORT_SYMBOL(gpu_buddy_reset_clear);
>>>   void gpu_buddy_free_block(struct gpu_buddy *mm,
>>>                 struct gpu_buddy_block *block)
>>>   {
>>> -    gpu_buddy_driver_lock_held(mm);
>>> -    BUG_ON(!gpu_buddy_block_is_allocated(block));
>>> -    mm->avail += gpu_buddy_block_size(mm, block);
>>> -    if (gpu_buddy_block_is_clear(block))
>>> -        mm->clear_avail += gpu_buddy_block_size(mm, block);
>>> +    if (!gpu_buddy_block_is_clear(block))
>>> +        gpu_dirty_tracker_mark_dirty(&mm->dirty,
>>> +                         gpu_buddy_block_offset(block),
>>> +                         gpu_buddy_block_size(mm, block));
>>> -    __gpu_buddy_free(mm, block, false);
>>> +    __gpu_buddy_free_block_internal(mm, block);
>>>   }
>>>   EXPORT_SYMBOL(gpu_buddy_free_block);
>>> @@ -689,17 +1189,48 @@ static void __gpu_buddy_free_list(struct 
>>> gpu_buddy *mm,
>>>                     bool mark_dirty)
>>>   {
>>>       struct gpu_buddy_block *block, *on;
>>> +    u64 dirty_start = 0, dirty_size = 0;
>>>       gpu_buddy_assert(!(mark_dirty && mark_clear));
>>>       list_for_each_entry_safe(block, on, objects, link) {
>>> +        u64 offset = gpu_buddy_block_offset(block);
>>> +        u64 size = gpu_buddy_block_size(mm, block);
>>> +
>>>           if (mark_clear)
>>> -            mark_cleared(block);
>>> +            block->header |= GPU_BUDDY_HEADER_CLEAR;
>>>           else if (mark_dirty)
>>> -            clear_reset(block);
>>> -        gpu_buddy_free_block(mm, block);
>>> +            block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> +
>>> +        /*
>>> +         * Coalesce contiguous dirty blocks into one extent update so
>>> +         * a multi-block contiguous free costs a single mark_dirty().
>>> +         * Flush the pending extent and start over on a gap.
>>> +         */
>>> +        if (!gpu_buddy_block_is_clear(block)) {
>>> +            if (dirty_size &&
>>> +                (dirty_start + dirty_size == offset ||
>>> +                 offset + size == dirty_start)) {
>>> +                dirty_start = min(dirty_start, offset);
>>> +                dirty_size += size;
>>> +            } else {
>>> +                if (dirty_size)
>>> + gpu_dirty_tracker_mark_dirty(&mm->dirty,
>>> +                                     dirty_start,
>>> +                                     dirty_size);
>>> +                dirty_start = offset;
>>> +                dirty_size = size;
>>> +            }
>>> +        }
>>> +
>>> +        __gpu_buddy_free_block_internal(mm, block);
>>>           cond_resched();
>>>       }
>>> +
>>> +    if (dirty_size)
>>> +        gpu_dirty_tracker_mark_dirty(&mm->dirty, dirty_start, 
>>> dirty_size);
>>> +
>>> +    gpu_buddy_sync_clear_avail(mm);
>>>       INIT_LIST_HEAD(objects);
>>>   }
>>> @@ -732,13 +1263,6 @@ void gpu_buddy_free_list(struct gpu_buddy *mm,
>>>   }
>>>   EXPORT_SYMBOL(gpu_buddy_free_list);
>>> -static bool block_incompatible(struct gpu_buddy_block *block, 
>>> unsigned int flags)
>>> -{
>>> -    bool needs_clear = flags & GPU_BUDDY_CLEAR_ALLOCATION;
>>> -
>>> -    return needs_clear != gpu_buddy_block_is_clear(block);
>>> -}
>>> -
>>>   static void __gpu_buddy_undo_splits(struct gpu_buddy *mm,
>>>                       struct gpu_buddy_block *block)
>>>   {
>>> @@ -749,7 +1273,7 @@ static void __gpu_buddy_undo_splits(struct 
>>> gpu_buddy *mm,
>>>            gpu_buddy_block_is_free(buddy))) {
>>>           rbtree_remove(mm, block);
>>> mm->free_scoreboard[gpu_buddy_block_order(block)]--;
>>> -        __gpu_buddy_free(mm, block, false);
>>> +        __gpu_buddy_free(mm, block);
>>>       }
>>>   }
>>> @@ -757,8 +1281,7 @@ static struct gpu_buddy_block *
>>>   __alloc_range_bias(struct gpu_buddy *mm,
>>>              u64 start, u64 end,
>>>              unsigned int order,
>>> -           unsigned long flags,
>>> -           bool fallback)
>>> +           unsigned long flags)
>>>   {
>>>       u64 req_size = mm->chunk_size << order;
>>>       struct gpu_buddy_block *block;
>>> @@ -768,7 +1291,15 @@ __alloc_range_bias(struct gpu_buddy *mm,
>>>       end = end - 1;
>>> -    for (i = 0; i < mm->n_roots; ++i)
>>> +    /*
>>> +     * This range-constrained search hands back the highest/right-most
>>> +     * address that satisfies the request: the roots are seeded 
>>> high- to-low
>>> +     * and the right (higher-address) child is descended first, making
>>> +     * top-down the default placement here. A non-top-down clear 
>>> request is
>>> +     * the only exception, where the descent is biased towards 
>>> clear or
>>> +     * clear-containing subtrees to satisfy the clear preference.
>>> +     */
>>> +    for (i = mm->n_roots - 1; i >= 0; --i)
>>>           list_add_tail(&mm->roots[i]->tmp_link, &dfs);
>>>       do {
>>> @@ -804,9 +1335,6 @@ __alloc_range_bias(struct gpu_buddy *mm,
>>>                   continue;
>>>           }
>>> -        if (!fallback && block_incompatible(block, flags))
>>> -            continue;
>>> -
>>>           if (contains(start, end, block_start, block_end) &&
>>>               order == gpu_buddy_block_order(block)) {
>>>               /*
>>> @@ -824,8 +1352,38 @@ __alloc_range_bias(struct gpu_buddy *mm,
>>>                   goto err_undo;
>>>           }
>>> -        list_add(&block->right->tmp_link, &dfs);
>>> -        list_add(&block->left->tmp_link, &dfs);
>>> +        /*
>>> +         * Top-down is a strict address-placement policy, so when 
>>> it is
>>> +         * requested we ignore clear steering and simply descend the
>>> +         * right (higher-address) child first. Only a non-top-down 
>>> clear
>>> +         * request biases the descent towards clear/has_clear 
>>> subtrees.
>>> +         */
>>> +        if ((flags & GPU_BUDDY_CLEAR_ALLOCATION) &&
>>> +            !(flags & GPU_BUDDY_TOPDOWN_ALLOCATION)) {
>>> +            struct gpu_buddy_block *prefer;
>>> +
>>> +            if (gpu_buddy_block_is_clear(block->right))
>>> +                prefer = block->right;
>>> +            else if (gpu_buddy_block_is_clear(block->left))
>>> +                prefer = block->left;
>>> +            else if (block->right->has_clear)
>>> +                prefer = block->right;
>>> +            else if (block->left->has_clear)
>>> +                prefer = block->left;
>>> +            else
>>> +                prefer = block->right;
>>> +
>>> +            if (prefer == block->right) {
>>> +                list_add(&block->left->tmp_link, &dfs);
>>> +                list_add(&block->right->tmp_link, &dfs);
>>> +            } else {
>>> +                list_add(&block->right->tmp_link, &dfs);
>>> +                list_add(&block->left->tmp_link, &dfs);
>>> +            }
>>> +        } else {
>>> +            list_add(&block->left->tmp_link, &dfs);
>>> +            list_add(&block->right->tmp_link, &dfs);
>>> +        }
>>>       } while (1);
>>>       return ERR_PTR(-ENOSPC);
>>> @@ -840,48 +1398,32 @@ __alloc_range_bias(struct gpu_buddy *mm,
>>>       return ERR_PTR(err);
>>>   }
>>> -static struct gpu_buddy_block *
>>> -__gpu_buddy_alloc_range_bias(struct gpu_buddy *mm,
>>> -                 u64 start, u64 end,
>>> -                 unsigned int order,
>>> -                 unsigned long flags)
>>> -{
>>> -    struct gpu_buddy_block *block;
>>> -    bool fallback = false;
>>> -
>>> -    block = __alloc_range_bias(mm, start, end, order,
>>> -                   flags, fallback);
>>> -    if (IS_ERR(block))
>>> -        return __alloc_range_bias(mm, start, end, order,
>>> -                      flags, !fallback);
>>> -
>>> -    return block;
>>> -}
>>> -
>>> +/* Return the highest-address free block of at least @order. */
>>>   static struct gpu_buddy_block *
>>>   get_maxblock(struct gpu_buddy *mm,
>>> -         unsigned int order,
>>> -         enum gpu_buddy_free_tree tree)
>>> +         unsigned int order)
>>>   {
>>> -    struct gpu_buddy_block *max_block = NULL, *block = NULL;
>>> -    struct rb_root *root;
>>> +    struct gpu_buddy_block *max_block;
>>> +    struct gpu_buddy_block *block;
>>>       unsigned int i;
>>> +    /*
>>> +     * Top-down allocation is a strict address-placement policy: 
>>> the block
>>> +     * is chosen purely by offset, regardless of its clear/dirty 
>>> state.
>>> +     * Clear state is re-derived from the dirty tracker once the 
>>> allocation
>>> +     * completes, and the driver is responsible for issuing the 
>>> clear pass
>>> +     * if a clear region is required.
>>> +     */
>>> +    max_block = NULL;
>>> +
>>>       for (i = order; i <= mm->max_order; ++i) {
>>> -        root = &mm->free_trees[tree][i];
>>> -        block = rbtree_last_free_block(root);
>>> +        block = rbtree_last_free_block(&mm->free_tree[i]);
>>>           if (!block)
>>>               continue;
>>> -        if (!max_block) {
>>> +        if (!max_block ||
>>> +            gpu_buddy_block_offset(block) > 
>>> gpu_buddy_block_offset(max_block))
>>>               max_block = block;
>>> -            continue;
>>> -        }
>>> -
>>> -        if (gpu_buddy_block_offset(block) >
>>> -            gpu_buddy_block_offset(max_block)) {
>>> -            max_block = block;
>>> -        }
>>>       }
>>>       return max_block;
>>> @@ -893,45 +1435,34 @@ alloc_from_freetree(struct gpu_buddy *mm,
>>>               unsigned long flags)
>>>   {
>>>       struct gpu_buddy_block *block = NULL;
>>> -    struct rb_root *root;
>>> -    enum gpu_buddy_free_tree tree;
>>>       unsigned int tmp;
>>>       int err;
>>> -    tree = (flags & GPU_BUDDY_CLEAR_ALLOCATION) ?
>>> -        GPU_BUDDY_CLEAR_TREE : GPU_BUDDY_DIRTY_TREE;
>>> -
>>>       if (flags & GPU_BUDDY_TOPDOWN_ALLOCATION) {
>>> -        block = get_maxblock(mm, order, tree);
>>> +        block = get_maxblock(mm, order);
>>>           if (block)
>>> -            /* Store the obtained block order */
>>>               tmp = gpu_buddy_block_order(block);
>>>       } else {
>>> -        for (tmp = order; tmp <= mm->max_order; ++tmp) {
>>> -            /* Get RB tree root for this order and tree */
>>> -            root = &mm->free_trees[tree][tmp];
>>> -            block = rbtree_last_free_block(root);
>>> -            if (block)
>>> -                break;
>>> +        if (flags & GPU_BUDDY_CLEAR_ALLOCATION) {
>>> +            for (tmp = order; tmp <= mm->max_order; ++tmp) {
>>> +                block = rbtree_last_clear_free_block(&mm- 
>>> >free_tree[tmp],
>>> +                                     GPU_BLOCK_MIXED);
>>> +                if (block)
>>> +                    break;
>>> +            }
>>>           }
>>> -    }
>>> -
>>> -    if (!block) {
>>> -        /* Try allocating from the other tree */
>>> -        tree = (tree == GPU_BUDDY_CLEAR_TREE) ?
>>> -            GPU_BUDDY_DIRTY_TREE : GPU_BUDDY_CLEAR_TREE;
>>> -
>>> -        for (tmp = order; tmp <= mm->max_order; ++tmp) {
>>> -            root = &mm->free_trees[tree][tmp];
>>> -            block = rbtree_last_free_block(root);
>>> -            if (block)
>>> -                break;
>>> +        if (!block) {
>>> +            for (tmp = order; tmp <= mm->max_order; ++tmp) {
>>> +                block = rbtree_last_free_block(&mm->free_tree[tmp]);
>>> +                if (block)
>>> +                    break;
>>> +            }
>>>           }
>>> -
>>> -        if (!block)
>>> -            return ERR_PTR(-ENOSPC);
>>>       }
>>> +    if (!block)
>>> +        return ERR_PTR(-ENOSPC);
>>> +
>>>       BUG_ON(!gpu_buddy_block_is_free(block));
>>>       while (tmp != order) {
>>> @@ -939,7 +1470,26 @@ alloc_from_freetree(struct gpu_buddy *mm,
>>>           if (unlikely(err))
>>>               goto err_undo;
>>> -        block = block->right;
>>> +        if ((flags & GPU_BUDDY_CLEAR_ALLOCATION) &&
>>> +            !(flags & GPU_BUDDY_TOPDOWN_ALLOCATION)) {
>>> +            bool right_clear, left_clear;
>>> +
>>> +            right_clear = gpu_buddy_block_is_clear(block->right);
>>> +            left_clear = gpu_buddy_block_is_clear(block->left);
>>> +
>>> +            if (right_clear)
>>> +                block = block->right;
>>> +            else if (left_clear)
>>> +                block = block->left;
>>> +            else if (block->right->has_clear)
>>> +                block = block->right;
>>> +            else if (block->left->has_clear)
>>> +                block = block->left;
>>> +            else
>>> +                block = block->right;
>>> +        } else {
>>> +            block = block->right;
>>> +        }
>>>           tmp--;
>>>       }
>>>       return block;
>>> @@ -966,12 +1516,10 @@ static bool 
>>> gpu_buddy_subtree_can_satisfy(struct rb_node *node,
>>>   static struct gpu_buddy_block *
>>>   gpu_buddy_find_block_aligned(struct gpu_buddy *mm,
>>> -                 enum gpu_buddy_free_tree tree,
>>>                    unsigned int order,
>>> -                 unsigned int alignment,
>>> -                 unsigned long flags)
>>> +                 unsigned int alignment)
>>>   {
>>> -    struct rb_root *root = &mm->free_trees[tree][order];
>>> +    struct rb_root *root = &mm->free_tree[order];
>>>       struct rb_node *rb = root->rb_node;
>>>       while (rb) {
>>> @@ -1004,12 +1552,10 @@ gpu_buddy_find_block_aligned(struct 
>>> gpu_buddy *mm,
>>>   static struct gpu_buddy_block *
>>>   gpu_buddy_offset_aligned_allocation(struct gpu_buddy *mm,
>>>                       u64 size,
>>> -                    u64 min_block_size,
>>> -                    unsigned long flags)
>>> +                    u64 min_block_size)
>>>   {
>>>       struct gpu_buddy_block *block = NULL;
>>>       unsigned int order, tmp, alignment;
>>> -    enum gpu_buddy_free_tree tree;
>>>       unsigned long pages;
>>>       int err;
>>> @@ -1017,19 +1563,15 @@ gpu_buddy_offset_aligned_allocation(struct 
>>> gpu_buddy *mm,
>>>       pages = size >> ilog2(mm->chunk_size);
>>>       order = fls(pages) - 1;
>>> -    tree = (flags & GPU_BUDDY_CLEAR_ALLOCATION) ?
>>> -        GPU_BUDDY_CLEAR_TREE : GPU_BUDDY_DIRTY_TREE;
>>> -
>>> +    /*
>>> +     * Offset-aligned allocation is a strict address-placement 
>>> policy: the
>>> +     * block is chosen purely by its offset alignment, regardless 
>>> of its
>>> +     * clear/dirty state. Clear state is re-derived from the dirty 
>>> tracker
>>> +     * once the allocation completes, and the driver is responsible 
>>> for
>>> +     * issuing the clear pass if a clear region is required.
>>> +     */
>>>       for (tmp = order; tmp <= mm->max_order; ++tmp) {
>>> -        block = gpu_buddy_find_block_aligned(mm, tree, tmp,
>>> -                             alignment, flags);
>>> -        if (!block) {
>>> -            tree = (tree == GPU_BUDDY_CLEAR_TREE) ?
>>> -                GPU_BUDDY_DIRTY_TREE : GPU_BUDDY_CLEAR_TREE;
>>> -            block = gpu_buddy_find_block_aligned(mm, tree, tmp,
>>> -                                 alignment, flags);
>>> -        }
>>> -
>>> +        block = gpu_buddy_find_block_aligned(mm, tmp, alignment);
>>>           if (block)
>>>               break;
>>>       }
>>> @@ -1068,6 +1610,7 @@ gpu_buddy_offset_aligned_allocation(struct 
>>> gpu_buddy *mm,
>>>   static int __alloc_range(struct gpu_buddy *mm,
>>>                struct list_head *dfs,
>>>                u64 start, u64 size,
>>> +             unsigned long flags,
>>>                struct list_head *blocks,
>>>                u64 *total_allocated_on_err)
>>>   {
>>> @@ -1104,16 +1647,33 @@ static int __alloc_range(struct gpu_buddy *mm,
>>>           if (contains(start, end, block_start, block_end)) {
>>>               if (gpu_buddy_block_is_free(block)) {
>>> +                bool block_clear = false;
>>> +                u64 block_offset;
>>> +                u64 block_size;
>>> +
>>> +                block_size = gpu_buddy_block_size(mm, block);
>>> +                block_offset = gpu_buddy_block_offset(block);
>>> +
>>> +                if (flags & GPU_BUDDY_CLEAR_ALLOCATION)
>>> +                    block_clear = gpu_buddy_block_is_clear(block);
>>> +
>>> +                if (!gpu_buddy_block_is_clear(block))
>>> + gpu_dirty_tracker_remove_range(&mm->dirty,
>>> +                                       block_offset,
>>> +                                       block_size);
>>> +
>>>                   mark_allocated(mm, block);
>>> -                total_allocated += gpu_buddy_block_size(mm, block);
>>> -                mm->avail -= gpu_buddy_block_size(mm, block);
>>> -                if (gpu_buddy_block_is_clear(block))
>>> -                    mm->clear_avail -= gpu_buddy_block_size(mm, 
>>> block);
>>> +                total_allocated += block_size;
>>> +                mm->avail -= block_size;
>>> +
>>> +                block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> +                if (block_clear)
>>> +                    block->header |= GPU_BUDDY_HEADER_CLEAR;
>>> +
>>> +                gpu_buddy_sync_clear_avail(mm);
>>> +
>>>                   list_add_tail(&block->link, &allocated);
>>>                   continue;
>>> -            } else if (!mm->clear_avail) {
>>> -                err = -ENOSPC;
>>> -                goto err_free;
>>>               }
>>>           }
>>> @@ -1158,6 +1718,7 @@ static int __alloc_range(struct gpu_buddy *mm,
>>>   static int __gpu_buddy_alloc_range(struct gpu_buddy *mm,
>>>                      u64 start,
>>>                      u64 size,
>>> +                   unsigned long flags,
>>>                      u64 *total_allocated_on_err,
>>>                      struct list_head *blocks)
>>>   {
>>> @@ -1167,20 +1728,23 @@ static int __gpu_buddy_alloc_range(struct 
>>> gpu_buddy *mm,
>>>       for (i = 0; i < mm->n_roots; ++i)
>>>           list_add_tail(&mm->roots[i]->tmp_link, &dfs);
>>> -    return __alloc_range(mm, &dfs, start, size,
>>> +    return __alloc_range(mm, &dfs, start, size, flags,
>>>                    blocks, total_allocated_on_err);
>>>   }
>>>   static int __alloc_contig_try_harder(struct gpu_buddy *mm,
>>>                        u64 size,
>>>                        u64 min_block_size,
>>> +                     unsigned long flags,
>>>                        struct list_head *blocks)
>>>   {
>>>       u64 rhs_offset, lhs_offset, lhs_size, filled;
>>>       struct gpu_buddy_block *block;
>>> -    unsigned int tree, order;
>>>       LIST_HEAD(blocks_lhs);
>>> +    struct rb_root *root;
>>> +    struct rb_node *iter;
>>>       unsigned long pages;
>>> +    unsigned int order;
>>>       u64 modify_size;
>>>       int err;
>>> @@ -1190,45 +1754,40 @@ static int __alloc_contig_try_harder(struct 
>>> gpu_buddy *mm,
>>>       if (order == 0)
>>>           return -ENOSPC;
>>> -    for_each_free_tree(tree) {
>>> -        struct rb_root *root;
>>> -        struct rb_node *iter;
>>> -
>>> -        root = &mm->free_trees[tree][order];
>>> -        if (rbtree_is_empty(root))
>>> -            continue;
>>> +    root = &mm->free_tree[order];
>>> +    if (RB_EMPTY_ROOT(root))
>>> +        return -ENOSPC;
>>> -        iter = rb_last(root);
>>> -        while (iter) {
>>> -            block = rbtree_get_free_block(iter);
>>> -
>>> -            /* Allocate blocks traversing RHS */
>>> -            rhs_offset = gpu_buddy_block_offset(block);
>>> -            err =  __gpu_buddy_alloc_range(mm, rhs_offset, size,
>>> -                               &filled, blocks);
>>> -            if (!err || err != -ENOSPC)
>>> -                return err;
>>> -
>>> -            lhs_size = max((size - filled), min_block_size);
>>> -            if (!IS_ALIGNED(lhs_size, min_block_size))
>>> -                lhs_size = round_up(lhs_size, min_block_size);
>>> -
>>> -            /* Allocate blocks traversing LHS */
>>> -            lhs_offset = gpu_buddy_block_offset(block) - lhs_size;
>>> -            err =  __gpu_buddy_alloc_range(mm, lhs_offset, lhs_size,
>>> -                               NULL, &blocks_lhs);
>>> -            if (!err) {
>>> -                list_splice(&blocks_lhs, blocks);
>>> -                return 0;
>>> -            } else if (err != -ENOSPC) {
>>> -                gpu_buddy_free_list_internal(mm, blocks);
>>> -                return err;
>>> -            }
>>> -            /* Free blocks for the next iteration */
>>> +    iter = rb_last(root);
>>> +    while (iter) {
>>> +        block = rbtree_get_free_block(iter);
>>> +
>>> +        /* Allocate blocks traversing RHS */
>>> +        rhs_offset = gpu_buddy_block_offset(block);
>>> +        err =  __gpu_buddy_alloc_range(mm, rhs_offset, size,
>>> +                           flags, &filled, blocks);
>>> +        if (!err || err != -ENOSPC)
>>> +            return err;
>>> +
>>> +        lhs_size = max((size - filled), min_block_size);
>>> +        if (!IS_ALIGNED(lhs_size, min_block_size))
>>> +            lhs_size = round_up(lhs_size, min_block_size);
>>> +
>>> +        /* Allocate blocks traversing LHS */
>>> +        lhs_offset = gpu_buddy_block_offset(block) - lhs_size;
>>> +        err =  __gpu_buddy_alloc_range(mm, lhs_offset, lhs_size,
>>> +                           flags, NULL, &blocks_lhs);
>>> +        if (!err) {
>>> +            list_splice(&blocks_lhs, blocks);
>>> +            return 0;
>>> +        } else if (err != -ENOSPC) {
>>>               gpu_buddy_free_list_internal(mm, blocks);
>>> -
>>> -            iter = rb_prev(iter);
>>> +            return err;
>>>           }
>>> +        /* Free blocks for the next iteration */
>>> +        gpu_buddy_free_list_internal(mm, blocks);
>>> +
>>> +        iter = rb_prev(iter);
>>>       }
>>>       return -ENOSPC;
>>> @@ -1262,6 +1821,7 @@ int gpu_buddy_block_trim(struct gpu_buddy *mm,
>>>       struct gpu_buddy_block *block;
>>>       u64 block_start, block_end;
>>>       LIST_HEAD(dfs);
>>> +    bool was_clear;
>>>       u64 new_start;
>>>       int err;
>>> @@ -1304,22 +1864,38 @@ int gpu_buddy_block_trim(struct gpu_buddy *mm,
>>>       }
>>>       list_del(&block->link);
>>> -    mark_free(mm, block);
>>> +
>>> +    was_clear = gpu_buddy_block_is_clear(block);
>>> +    block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> +
>>> +    if (!was_clear)
>>> +        gpu_dirty_tracker_mark_dirty(&mm->dirty,
>>> +                         gpu_buddy_block_offset(block),
>>> +                         gpu_buddy_block_size(mm, block));
>>> +
>>> +    __mark_free(mm, block, was_clear ? GPU_BLOCK_CLEAR : 
>>> GPU_BLOCK_DIRTY);
>>>       mm->avail += gpu_buddy_block_size(mm, block);
>>> -    if (gpu_buddy_block_is_clear(block))
>>> -        mm->clear_avail += gpu_buddy_block_size(mm, block);
>>> +    gpu_buddy_sync_clear_avail(mm);
>>>       /* Prevent recursively freeing this node */
>>>       parent = block->parent;
>>>       block->parent = NULL;
>>>       list_add(&block->tmp_link, &dfs);
>>> -    err =  __alloc_range(mm, &dfs, new_start, new_size, blocks, NULL);
>>> +    err =  __alloc_range(mm, &dfs, new_start, new_size,
>>> +                 was_clear ? GPU_BUDDY_CLEAR_ALLOCATION : 0,
>>> +                 blocks, NULL);
>>>       if (err) {
>>>           mark_allocated(mm, block);
>>>           mm->avail -= gpu_buddy_block_size(mm, block);
>>> -        if (gpu_buddy_block_is_clear(block))
>>> -            mm->clear_avail -= gpu_buddy_block_size(mm, block);
>>> +        if (!was_clear) {
>>> +            gpu_dirty_tracker_remove_range(&mm->dirty,
>>> +                               gpu_buddy_block_offset(block),
>>> +                               gpu_buddy_block_size(mm, block));
>>> +        }
>>> +        if (was_clear)
>>> +            block->header |= GPU_BUDDY_HEADER_CLEAR;
>>> +        gpu_buddy_sync_clear_avail(mm);
>>>           list_add(&block->link, blocks);
>>>       }
>>> @@ -1328,6 +1904,22 @@ int gpu_buddy_block_trim(struct gpu_buddy *mm,
>>>   }
>>>   EXPORT_SYMBOL(gpu_buddy_block_trim);
>>> +static bool dirty_steer_window(struct gpu_buddy *mm, u64 req_size,
>>> +                   u64 *start, u64 *end, unsigned long *flags)
>>> +{
>>> +    u64 aligned_start;
>>> +    struct gpu_dirty_extent *ext =
>>> +        gpu_dirty_tracker_find(&mm->dirty, req_size, &aligned_start);
>>> +
>>> +    if (!ext)
>>> +        return false;
>>> +
>>> +    *start  = aligned_start;
>>> +    *end    = ext->end;
>>> +    *flags |= GPU_BUDDY_RANGE_ALLOCATION;
>>> +    return true;
>>> +}
>>> +
>>>   static struct gpu_buddy_block *
>>>   __gpu_buddy_alloc_blocks(struct gpu_buddy *mm,
>>>                u64 start, u64 end,
>>> @@ -1335,18 +1927,36 @@ __gpu_buddy_alloc_blocks(struct gpu_buddy *mm,
>>>                unsigned int order,
>>>                unsigned long flags)
>>>   {
>>> -    if (flags & GPU_BUDDY_RANGE_ALLOCATION)
>>> +    struct gpu_buddy_block *block;
>>> +    bool steered = false;
>>> +
>>> +    /* Allocate from dirty tracker */
>>> +    if (!(flags & GPU_BUDDY_RANGE_ALLOCATION) &&
>>> +        !(flags & GPU_BUDDY_CLEAR_ALLOCATION) &&
>>> +        size >= min_block_size &&
>>> +        mm->clear_avail && mm->dirty.total_dirty) {
>>> +        u64 block_size = mm->chunk_size << order;
>>> +
>>> +        steered = dirty_steer_window(mm, block_size,
>>> +                         &start, &end, &flags);
>>> +    }
>>> +
>>> +    if (flags & GPU_BUDDY_RANGE_ALLOCATION) {
>>>           /* Allocate traversing within the range */
>>> -        return  __gpu_buddy_alloc_range_bias(mm, start, end,
>>> -                             order, flags);
>>> -    else if (size < min_block_size)
>>> +        block = __alloc_range_bias(mm, start, end, order, flags);
>>> +        if (!IS_ERR(block) || !steered)
>>> +            return block;
>>> +
>>> +        flags &= ~GPU_BUDDY_RANGE_ALLOCATION;
>>> +    }
>>> +
>>> +    if (size < min_block_size)
>>>           /* Allocate from an offset-aligned region without size 
>>> rounding */
>>>           return gpu_buddy_offset_aligned_allocation(mm, size,
>>> -                               min_block_size,
>>> -                               flags);
>>> -    else
>>> -        /* Allocate from freetree */
>>> -        return alloc_from_freetree(mm, order, flags);
>>> +                               min_block_size);
>>> +
>>> +    /* Allocate from freetree */
>>> +    return alloc_from_freetree(mm, order, flags);
>>>   }
>>>   /**
>>> @@ -1407,7 +2017,7 @@ int gpu_buddy_alloc_blocks(struct gpu_buddy *mm,
>>>           if (!IS_ALIGNED(start | end, min_block_size))
>>>               return -EINVAL;
>>> -        return __gpu_buddy_alloc_range(mm, start, size, NULL, blocks);
>>> +        return __gpu_buddy_alloc_range(mm, start, size, flags, 
>>> NULL, blocks);
>>>       }
>>>       original_size = size;
>>> @@ -1433,12 +2043,15 @@ int gpu_buddy_alloc_blocks(struct gpu_buddy 
>>> *mm,
>>>           if ((flags & GPU_BUDDY_CONTIGUOUS_ALLOCATION) &&
>>>               !(flags & GPU_BUDDY_RANGE_ALLOCATION))
>>>               return __alloc_contig_try_harder(mm, original_size,
>>> -                             original_min_size, blocks);
>>> +                             original_min_size,
>>> +                             flags, blocks);
>>>           return -EINVAL;
>>>       }
>>>       do {
>>> +        bool block_clear = false;
>>> +
>>>           order = min(order, (unsigned int)fls(pages) - 1);
>>>           BUG_ON(order > mm->max_order);
>>>           /*
>>> @@ -1448,8 +2061,6 @@ int gpu_buddy_alloc_blocks(struct gpu_buddy *mm,
>>>           BUG_ON(size >= min_block_size && order < min_order);
>>>           do {
>>> -            unsigned int fallback_order;
>>> -
>>>               block = __gpu_buddy_alloc_blocks(mm, start,
>>>                                end,
>>>                                size,
>>> @@ -1459,48 +2070,48 @@ int gpu_buddy_alloc_blocks(struct gpu_buddy 
>>> *mm,
>>>               if (!IS_ERR(block))
>>>                   break;
>>> -            if (size < min_block_size) {
>>> -                fallback_order = order;
>>> -            } else if (order == min_order) {
>>> -                fallback_order = min_order;
>>> -            } else {
>>> +            if (size >= min_block_size && order > min_order) {
>>>                   order--;
>>>                   continue;
>>>               }
>>> -            /* Try allocation through force merge method */
>>> -            if (mm->clear_avail &&
>>> -                !__force_merge(mm, start, end, fallback_order)) {
>>> -                block = __gpu_buddy_alloc_blocks(mm, start,
>>> -                                 end,
>>> -                                 size,
>>> -                                 min_block_size,
>>> -                                 fallback_order,
>>> -                                 flags);
>>> -                if (!IS_ERR(block)) {
>>> -                    order = fallback_order;
>>> -                    break;
>>> -                }
>>> -            }
>>> -
>>>               /*
>>>                * Try contiguous block allocation through
>>>                * try harder method.
>>>                */
>>>               if (flags & GPU_BUDDY_CONTIGUOUS_ALLOCATION &&
>>> -                !(flags & GPU_BUDDY_RANGE_ALLOCATION))
>>> -                return __alloc_contig_try_harder(mm,
>>> -                                 original_size,
>>> -                                 original_min_size,
>>> -                                 blocks);
>>> +                !(flags & GPU_BUDDY_RANGE_ALLOCATION)) {
>>> +                err = __alloc_contig_try_harder(mm,
>>> +                                original_size,
>>> +                                original_min_size,
>>> +                                flags,
>>> +                                blocks);
>>> +                if (!err)
>>> +                    return 0;
>>> +                if (err != -ENOSPC)
>>> +                    return err;
>>> +                goto err_free;
>>> +            }
>>>               err = -ENOSPC;
>>>               goto err_free;
>>>           } while (1);
>>> +        if (flags & GPU_BUDDY_CLEAR_ALLOCATION)
>>> +            block_clear = gpu_buddy_block_is_clear(block);
>>> +
>>> +        if (!gpu_buddy_block_is_clear(block))
>>> +            gpu_dirty_tracker_remove_range(&mm->dirty,
>>> +                               gpu_buddy_block_offset(block),
>>> +                               gpu_buddy_block_size(mm, block));
>>> +
>>>           mark_allocated(mm, block);
>>>           mm->avail -= gpu_buddy_block_size(mm, block);
>>> -        if (gpu_buddy_block_is_clear(block))
>>> -            mm->clear_avail -= gpu_buddy_block_size(mm, block);
>>> +
>>> +        block->header &= ~GPU_BUDDY_HEADER_CLEAR;
>>> +        if (block_clear)
>>> +            block->header |= GPU_BUDDY_HEADER_CLEAR;
>>> +
>>> +        gpu_buddy_sync_clear_avail(mm);
>>>           kmemleak_update_trace(block);
>>>           list_add_tail(&block->link, &allocated);
>>> @@ -1595,6 +2206,7 @@ EXPORT_SYMBOL(gpu_buddy_print);
>>>   static void gpu_buddy_module_exit(void)
>>>   {
>>> +    kmem_cache_destroy(slab_extents);
>>>       kmem_cache_destroy(slab_blocks);
>>>   }
>>> @@ -1604,7 +2216,15 @@ static int __init gpu_buddy_module_init(void)
>>>       if (!slab_blocks)
>>>           return -ENOMEM;
>>> +    slab_extents = KMEM_CACHE(gpu_dirty_extent, 0);
>>> +    if (!slab_extents)
>>> +        goto err_destroy_blocks;
>>> +
>>>       return 0;
>>> +
>>> +err_destroy_blocks:
>>> +    kmem_cache_destroy(slab_blocks);
>>> +    return -ENOMEM;
>>>   }
>>>   module_init(gpu_buddy_module_init);
>>> diff --git a/drivers/gpu/tests/gpu_buddy_test.c b/drivers/gpu/tests/ 
>>> gpu_buddy_test.c
>>> index ed4c1c3acb3c..198d8dc4e3f0 100644
>>> --- a/drivers/gpu/tests/gpu_buddy_test.c
>>> +++ b/drivers/gpu/tests/gpu_buddy_test.c
>>> @@ -38,7 +38,7 @@ static void 
>>> gpu_test_buddy_subtree_offset_alignment_stress(struct kunit *test)
>>>       };
>>>       struct list_head allocated[ARRAY_SIZE(alignments)];
>>>       unsigned int i, max_subtree_align = 0;
>>> -    int ret, tree, order;
>>> +    int ret, order;
>>>       struct gpu_buddy mm;
>>>       KUNIT_ASSERT_FALSE_MSG(test, gpu_buddy_init(&mm, mm_size, SZ_4K),
>>> @@ -78,15 +78,11 @@ static void 
>>> gpu_test_buddy_subtree_offset_alignment_stress(struct kunit *test)
>>>           }
>>>           for (order = mm.max_order; order >= 0 && !root; order--) {
>>> -            for (tree = 0; tree < 2; tree++) {
>>> -                node = mm.free_trees[tree][order].rb_node;
>>> -                if (node) {
>>> -                    root = container_of(node,
>>> -                                struct gpu_buddy_block,
>>> -                                rb);
>>> -                    break;
>>> -                }
>>> -            }
>>> +            node = mm.free_tree[order].rb_node;
>>> +            if (node)
>>> +                root = container_of(node,
>>> +                            struct gpu_buddy_block,
>>> +                            rb);
>>>           }
>>>           KUNIT_ASSERT_NOT_NULL(test, root);
>>> @@ -97,15 +93,13 @@ static void 
>>> gpu_test_buddy_subtree_offset_alignment_stress(struct kunit *test)
>>>           gpu_buddy_free_list(&mm, &allocated[i], 0);
>>>           for (order = 0; order <= mm.max_order; order++) {
>>> -            for (tree = 0; tree < 2; tree++) {
>>> -                node = mm.free_trees[tree][order].rb_node;
>>> -                if (!node)
>>> -                    continue;
>>> -
>>> -                block = container_of(node, struct gpu_buddy_block, 
>>> rb);
>>> -                max_subtree_align = max(max_subtree_align,
>>> -                            block->subtree_max_alignment);
>>> -            }
>>> +            node = mm.free_tree[order].rb_node;
>>> +            if (!node)
>>> +                continue;
>>> +
>>> +            block = container_of(node, struct gpu_buddy_block, rb);
>>> +            max_subtree_align = max(max_subtree_align,
>>> +                        block->subtree_max_alignment);
>>>           }
>>>           KUNIT_EXPECT_GE(test, max_subtree_align, 
>>> ilog2(alignments[i]));
>>> diff --git a/include/linux/gpu_buddy.h b/include/linux/gpu_buddy.h
>>> index 2c36124bb696..1fe7b61db484 100644
>>> --- a/include/linux/gpu_buddy.h
>>> +++ b/include/linux/gpu_buddy.h
>>> @@ -8,6 +8,7 @@
>>>   #include <linux/bitops.h>
>>>   #include <linux/list.h>
>>> +#include <linux/mempool.h>
>>>   #include <linux/slab.h>
>>>   #include <linux/sched.h>
>>>   #include <linux/rbtree.h>
>>> @@ -43,8 +44,8 @@
>>>   /**
>>>    * GPU_BUDDY_CLEAR_ALLOCATION - Prefer pre-cleared (zeroed) memory
>>>    *
>>> - * Attempt to allocate from the clear tree first. If insufficient 
>>> clear
>>> - * memory is available, falls back to dirty memory. Useful when the
>>> + * Attempt to allocate outside dirty-tracked ranges first. If 
>>> insufficient
>>> + * clear memory is available, falls back to dirty memory. Useful 
>>> when the
>>>    * caller needs zeroed memory and wants to avoid GPU clear 
>>> operations.
>>>    */
>>>   #define GPU_BUDDY_CLEAR_ALLOCATION        BIT(3)
>>> @@ -53,8 +54,8 @@
>>>    * GPU_BUDDY_CLEARED - Mark returned blocks as cleared
>>>    *
>>>    * Used with gpu_buddy_free_list() to indicate that the memory being
>>> - * freed has been cleared (zeroed). The blocks will be placed in the
>>> - * clear tree for future GPU_BUDDY_CLEAR_ALLOCATION requests.
>>> + * freed has been cleared (zeroed). The blocks will be removed from 
>>> the
>>> + * dirty tracker for future GPU_BUDDY_CLEAR_ALLOCATION requests.
>>>    */
>>>   #define GPU_BUDDY_CLEARED            BIT(4)
>>> @@ -67,15 +68,6 @@
>>>    */
>>>   #define GPU_BUDDY_TRIM_DISABLE            BIT(5)
>>> -enum gpu_buddy_free_tree {
>>> -    GPU_BUDDY_CLEAR_TREE = 0,
>>> -    GPU_BUDDY_DIRTY_TREE,
>>> -    GPU_BUDDY_MAX_FREE_TREES,
>>> -};
>>> -
>>> -#define for_each_free_tree(tree) \
>>> -    for ((tree) = 0; (tree) < GPU_BUDDY_MAX_FREE_TREES; (tree)++)
>>> -
>>>   /**
>>>    * struct gpu_buddy_block - Block within a buddy allocator
>>>    *
>>> @@ -88,6 +80,17 @@ enum gpu_buddy_free_tree {
>>>    * @private: Private data owned by the allocator user (e.g., 
>>> driver- specific data)
>>>    * @link: List node for user ownership while block is allocated
>>>    */
>>> +/*
>>> + * Clear/dirty state of a free block. Ordered so a numerically 
>>> larger value
>>> + * is "more clear" (DIRTY < MIXED < CLEAR) which lets 
>>> subtree_block_state be
>>> + * maintained as a simple max-augment over the per-order free tree.
>>> + */
>>> +enum gpu_block_state {
>>> +    GPU_BLOCK_DIRTY = 0,
>>> +    GPU_BLOCK_MIXED = 1,
>>> +    GPU_BLOCK_CLEAR = 2,
>>> +};
>>> +
>>>   struct gpu_buddy_block {
>>>   /* private: */
>>>       /*
>>> @@ -103,6 +106,13 @@ struct gpu_buddy_block {
>>>   #define   GPU_BUDDY_ALLOCATED       (1 << 10)
>>>   #define   GPU_BUDDY_FREE       (2 << 10)
>>>   #define   GPU_BUDDY_SPLIT       (3 << 10)
>>> +/*
>>> + * GPU_BUDDY_HEADER_CLEAR has two roles:
>>> + *  - FREE state:      set when the block's full range is cleared 
>>> (dirty
>>> + *                     tracker confirmed no overlap).
>>> + *  - ALLOCATED state: set when the block was served from cleared 
>>> memory,
>>> + *                     informing the caller that no GPU clear pass 
>>> is needed.
>>> + */
>>>   #define GPU_BUDDY_HEADER_CLEAR  GENMASK_ULL(9, 9)
>>>   /* Free to be used, if needed in the future */
>>>   #define GPU_BUDDY_HEADER_UNUSED GENMASK_ULL(8, 6)
>>> @@ -128,13 +138,51 @@ struct gpu_buddy_block {
>>>           struct list_head link;
>>>       };
>>>   /* private: */
>>> -    struct list_head tmp_link;
>>> +    enum gpu_block_state subtree_block_state;
>>>       unsigned int subtree_max_alignment;
>>> +    struct list_head tmp_link;
>>> +    bool has_clear;
>>>   };
>>>   /* Order-zero must be at least SZ_4K */
>>>   #define GPU_BUDDY_MAX_ORDER (63 - 12)
>>> +/**
>>> + * struct gpu_dirty_extent - a contiguous dirty address range
>>> + *
>>> + * Tracks a single contiguous address range whose memory content is 
>>> known
>>> + * to be dirty.  Extents are non-overlapping and stored in an 
>>> augmented
>>> + * red-black tree sorted by @start.  The augmented value 
>>> @subtree_max_size
>>> + * allows O(log N) search for an extent of at least a given size.
>>> + */
>>> +struct gpu_dirty_extent {
>>> +/* private: */
>>> +    struct rb_node    rb;
>>> +    u64        start;
>>> +    u64        end;
>>> +    u64        subtree_max_size;
>>> +};
>>> +
>>> +/**
>>> + * struct gpu_dirty_tracker - tracks dirty address intervals
>>> + *
>>> + * Maintains a set of non-overlapping dirty extents as an augmented
>>> + * red-black tree.
>>> + *
>>> + * @total_dirty: Total bytes of dirty memory currently tracked.
>>> + * @extent_pool: Mempool backing extent node allocations. sashiko 
>>> reported
>>> + *         that a __GFP_NOFAIL allocation on the free path could
>>> + *         deadlock during memory reclaim, so a per-tracker mempool is
>>> + *         used to guarantee extent nodes without __GFP_NOFAIL.
>>> + */
>>> +struct gpu_dirty_tracker {
>>> +/* private: */
>>> +    struct rb_root    root;
>>> +    mempool_t    *extent_pool;
>>> +/* public: */
>>> +    u64        total_dirty;
>>> +};
>>> +
>>>   /**
>>>    * struct gpu_buddy - GPU binary buddy allocator
>>>    *
>>> @@ -152,20 +200,25 @@ struct gpu_buddy_block {
>>>    * @chunk_size: Minimum allocation granularity in bytes. Must be 
>>> at least SZ_4K.
>>>    * @size: Total size of the address space managed by this 
>>> allocator in bytes.
>>>    * @avail: Total free space currently available for allocation in 
>>> bytes.
>>> - * @clear_avail: Free space available in the clear tree (zeroed 
>>> memory) in bytes.
>>> - *               This is a subset of @avail.
>>> + * @clear_avail: Free space that is clear (zeroed) in bytes. A 
>>> subset of @avail.
>>> + *               Maintained as @avail - dirty.total_dirty, since 
>>> the tracker
>>> + *               records the dirty extents. Zero at init, as a 
>>> fresh pool is
>>> + *               fully dirty.
>>>    * @lock_dep_map: Annotates gpu_buddy API with a driver provided 
>>> lock.
>>>    */
>>>   struct gpu_buddy {
>>>   /* private: */
>>> +    /* Tracker of dirty address ranges (decoupled from free_tree). */
>>> +    struct gpu_dirty_tracker dirty;
>>>       /*
>>> -     * Array of red-black trees for free block management.
>>> -     * Indexed as free_trees[clear/dirty][order] where:
>>> -     * - Index 0 (GPU_BUDDY_CLEAR_TREE): blocks with zeroed content
>>> -     * - Index 1 (GPU_BUDDY_DIRTY_TREE): blocks with unknown content
>>> -     * Each tree holds free blocks of the corresponding order.
>>> +     * One RB-tree per order containing all free blocks (clear and
>>> +     * dirty alike).  The augment field subtree_block_state (a max 
>>> over
>>> +     * the subtree of each block's state) lets clear allocations
>>> +     * find the right-most fully-clear or mixed block in O(log N).
>>> +     * Dirty free blocks coexist here but are also indexed by the
>>> +     * @dirty tracker for fast dirty allocation lookups.
>>>        */
>>> -    struct rb_root **free_trees;
>>> +    struct rb_root *free_tree;
>>>       /*
>>>        * Array of root blocks representing the top-level blocks of the
>>>        * binary tree(s). Multiple roots exist when the total size is 
>>> not
>>>
>>> base-commit: b961eb36d7b04147104cff2fd8bc0e94f4713324
>>
>


      reply	other threads:[~2026-08-17 12:53 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-08-11 13:34 [PATCH v9 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker Arunpravin Paneer Selvam
2026-08-11 13:34 ` [PATCH v9 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test Arunpravin Paneer Selvam
2026-08-11 14:51 ` ✗ Fi.CI.BUILD: failure for series starting with [v9,1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker Patchwork
2026-08-13 13:37 ` [PATCH v9 1/2] " Arunpravin Paneer Selvam
2026-08-17  9:57   ` Matthew Auld
2026-08-17 12:53     ` Arunpravin Paneer Selvam [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1cb5dc7c-c342-4687-a7ce-2f8e2e39109e@amd.com \
    --to=arunpravin.paneerselvam@amd.com \
    --cc=Anand.Raghavendra@amd.com \
    --cc=alexander.deucher@amd.com \
    --cc=amd-gfx@lists.freedesktop.org \
    --cc=christian.koenig@amd.com \
    --cc=dri-devel@lists.freedesktop.org \
    --cc=intel-gfx@lists.freedesktop.org \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=matthew.auld@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox