* [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
@ 2026-07-31 7:07 Arunpravin Paneer Selvam
2026-07-31 7:07 ` [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test Arunpravin Paneer Selvam
` (2 more replies)
0 siblings, 3 replies; 5+ messages in thread
From: Arunpravin Paneer Selvam @ 2026-07-31 7:07 UTC (permalink / raw)
To: matthew.auld, christian.koenig, dri-devel, intel-gfx, intel-xe,
amd-gfx
Cc: alexander.deucher, Arunpravin Paneer Selvam
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[].
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 | 1318 ++++++++++++++++++++--------
drivers/gpu/tests/gpu_buddy_test.c | 32 +-
include/linux/gpu_buddy.h | 97 +-
3 files changed, 1034 insertions(+), 413 deletions(-)
diff --git a/drivers/gpu/buddy.c b/drivers/gpu/buddy.c
index dc81fe0301ce..a5d68cd8b78d 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 +35,441 @@
#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, so the underlying slab alloc almost always
+ * succeeds via reclaim; the reserve only backstops the rare case where
+ * it still returns NULL (e.g. the current task is an OOM victim),
+ * guaranteeing a non-NULL extent without __GFP_NOFAIL. Because each
+ * alloc can independently wait for reclaim, the reserve need not scale
+ * with the number of 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 that has one. 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 selection order therefore depends on the driver's
+ * free policy:
+ *
+ * 1) If the driver never clears freed blocks, no free block ever holds
+ * clear bytes, so a clear request always falls back to a dirty block.
+ * 2) If the driver clears every freed block, cleared ranges accumulate at
+ * the high end of the address space, so picking the right-most block
+ * yields clear -> mixed -> dirty.
+ * 3) If the driver clears freed blocks selectively, fully-clear blocks are
+ * still preferred over mixed ones at the same order, and the right-most
+ * candidate wins, giving a clear -> mixed -> dirty 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 +503,97 @@ 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 +604,10 @@ 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 +628,6 @@ 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 +640,64 @@ 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)
{
- return RB_EMPTY_ROOT(root);
+ 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)
+{
+ 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 +707,12 @@ 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 +721,7 @@ 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);
}
@@ -167,26 +730,11 @@ 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)
{
@@ -199,21 +747,36 @@ static void mark_allocated(struct gpu_buddy *mm,
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 +816,31 @@ __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;
- if (gpu_buddy_block_is_clear(block))
- mark_cleared(parent);
+ buddy_state = gpu_block_cached_state(buddy);
+
+ 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 +852,11 @@ 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 +871,7 @@ 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 +905,14 @@ 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 +938,8 @@ 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 +961,8 @@ 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 +980,7 @@ 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 +988,10 @@ 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 +1000,8 @@ 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 +1013,7 @@ 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 +1029,18 @@ 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;
}
@@ -572,42 +1055,39 @@ 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);
@@ -620,13 +1100,18 @@ EXPORT_SYMBOL(gpu_buddy_reset_clear);
void gpu_buddy_free_block(struct gpu_buddy *mm,
struct gpu_buddy_block *block)
{
+ u64 size = gpu_buddy_block_size(mm, block);
+ u64 offset = gpu_buddy_block_offset(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);
- __gpu_buddy_free(mm, block, false);
+ mm->avail += size;
+ if (!gpu_buddy_block_is_clear(block))
+ gpu_dirty_tracker_mark_dirty(&mm->dirty, offset, size);
+
+ gpu_buddy_sync_clear_avail(mm);
+ __gpu_buddy_free(mm, block);
}
EXPORT_SYMBOL(gpu_buddy_free_block);
@@ -641,9 +1126,9 @@ static void __gpu_buddy_free_list(struct gpu_buddy *mm,
list_for_each_entry_safe(block, on, objects, link) {
if (mark_clear)
- mark_cleared(block);
+ block->header |= GPU_BUDDY_HEADER_CLEAR;
else if (mark_dirty)
- clear_reset(block);
+ block->header &= ~GPU_BUDDY_HEADER_CLEAR;
gpu_buddy_free_block(mm, block);
cond_resched();
}
@@ -679,13 +1164,6 @@ 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)
{
@@ -696,7 +1174,7 @@ 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);
}
}
@@ -704,8 +1182,7 @@ 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;
@@ -715,7 +1192,15 @@ __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 {
@@ -751,9 +1236,6 @@ __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)) {
/*
@@ -771,8 +1253,38 @@ __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);
@@ -787,48 +1299,32 @@ __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) {
- max_block = block;
- continue;
- }
-
- if (gpu_buddy_block_offset(block) >
- gpu_buddy_block_offset(max_block)) {
+ if (!max_block ||
+ gpu_buddy_block_offset(block) > gpu_buddy_block_offset(max_block))
max_block = block;
- }
}
return max_block;
@@ -840,45 +1336,34 @@ 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) {
@@ -886,7 +1371,26 @@ 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;
@@ -913,12 +1417,10 @@ 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) {
@@ -951,12 +1453,10 @@ 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;
@@ -964,19 +1464,15 @@ 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;
}
@@ -1015,6 +1511,7 @@ 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)
{
@@ -1051,16 +1548,33 @@ 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;
}
}
@@ -1105,6 +1619,7 @@ 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)
{
@@ -1114,20 +1629,23 @@ 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;
@@ -1137,45 +1655,40 @@ 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;
@@ -1209,6 +1722,7 @@ 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;
@@ -1251,22 +1765,38 @@ 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);
}
@@ -1275,6 +1805,22 @@ 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,
@@ -1282,18 +1828,36 @@ __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);
}
/**
@@ -1354,7 +1918,7 @@ 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;
@@ -1380,12 +1944,15 @@ 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);
/*
@@ -1395,8 +1962,6 @@ 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,
@@ -1406,48 +1971,48 @@ 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);
@@ -1542,6 +2107,7 @@ EXPORT_SYMBOL(gpu_buddy_print);
static void gpu_buddy_module_exit(void)
{
+ kmem_cache_destroy(slab_extents);
kmem_cache_destroy(slab_blocks);
}
@@ -1551,7 +2117,15 @@ 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 89698563c61b..f8e56da5058e 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 e037714563d8..899b84298cd8 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: 9c950822f0fa923ccd344d7a143872d25efe89a3
--
2.34.1
^ permalink raw reply related [flat|nested] 5+ messages in thread* [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test
2026-07-31 7:07 [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker Arunpravin Paneer Selvam
@ 2026-07-31 7:07 ` Arunpravin Paneer Selvam
2026-07-31 7:19 ` sashiko-bot
2026-07-31 7:21 ` [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker sashiko-bot
2026-07-31 7:28 ` ✗ Fi.CI.BUILD: failure for series starting with [v7,1/2] " Patchwork
2 siblings, 1 reply; 5+ messages in thread
From: Arunpravin Paneer Selvam @ 2026-07-31 7:07 UTC (permalink / raw)
To: matthew.auld, christian.koenig, dri-devel, intel-gfx, intel-xe,
amd-gfx
Cc: alexander.deucher, Arunpravin Paneer Selvam
Add gpu_test_buddy_dirty_tracker_performance to demonstrate the key
advantage of the decoupled dirty-tracker design over the old dual-tree
/ force_merge approach.
The test runs two scenarios on a 4 GiB pool after alternating
clear/dirty fragmentation at 4 KiB granularity:
1. Contiguous 4 GiB alloc: the old design requires __force_merge()
to rebuild max_order from scratch; the new design coalesces during
free() so the alloc is O(log N).
old (force_merge) - 71 ms
dirty tracker design - 17 ms
2. Repeated 256 KiB alloc throughput: the old design pays
__force_merge() on every alloc; the new design does not.
old (force_merge) - 95 ms
dirty tracker design - 24 ms
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/tests/gpu_buddy_test.c | 110 +++++++++++++++++++++++++++++
1 file changed, 110 insertions(+)
diff --git a/drivers/gpu/tests/gpu_buddy_test.c b/drivers/gpu/tests/gpu_buddy_test.c
index f8e56da5058e..3fc2bfa66ee1 100644
--- a/drivers/gpu/tests/gpu_buddy_test.c
+++ b/drivers/gpu/tests/gpu_buddy_test.c
@@ -283,6 +283,115 @@ static void gpu_test_buddy_fragmentation_performance(struct kunit *test)
gpu_buddy_fini(&mm);
}
+static void gpu_test_buddy_dirty_tracker_performance(struct kunit *test)
+{
+ struct gpu_buddy_block *block, *tmp;
+ unsigned long elapsed_ms;
+ LIST_HEAD(clear_blocks);
+ LIST_HEAD(dirty_blocks);
+ LIST_HEAD(allocated);
+ struct gpu_buddy mm;
+ LIST_HEAD(results);
+ ktime_t start, end;
+ int i, count;
+
+ /*
+ * Contiguous alloc latency after alternating clear/dirty fragmentation
+ *
+ * Fill a 4 GiB pool with 4 KiB allocations, partition them into
+ * alternating cleared and dirty sets, then free both. In the old
+ * dual-tree design every adjacent buddy pair has one cleared half and
+ * one dirty half, so the pair sits on opposite sides of the clear/dirty
+ * merge barrier and cannot be coalesced at free() time. The pool
+ * stays fully fragmented and the subsequent contiguous 4 GiB allocation
+ * has to invoke __force_merge() to climb back up to max_order before
+ * it can succeed. With the dirty-tracker design buddy pairs coalesce
+ * unconditionally during free(), so the pool is already at max_order
+ * before the timed alloc begins and __force_merge() is not needed.
+ */
+ KUNIT_ASSERT_FALSE_MSG(test, gpu_buddy_init(&mm, SZ_4G, SZ_4K),
+ "buddy_init failed\n");
+
+ for (i = 0; i < SZ_4G / SZ_4K; i++)
+ KUNIT_ASSERT_FALSE_MSG(test,
+ gpu_buddy_alloc_blocks(&mm, 0, SZ_4G, SZ_4K, SZ_4K,
+ &allocated, 0),
+ "buddy_alloc hit an error size=%u\n", SZ_4K);
+
+ count = 0;
+ list_for_each_entry_safe(block, tmp, &allocated, link) {
+ if (count++ % 2 == 0)
+ list_move_tail(&block->link, &clear_blocks);
+ else
+ list_move_tail(&block->link, &dirty_blocks);
+ }
+
+ gpu_buddy_free_list(&mm, &clear_blocks, GPU_BUDDY_CLEARED);
+ gpu_buddy_free_list(&mm, &dirty_blocks, 0);
+
+ start = ktime_get();
+ KUNIT_ASSERT_FALSE_MSG(test,
+ gpu_buddy_alloc_blocks(&mm, 0, SZ_4G, SZ_4G, SZ_4K,
+ &results,
+ GPU_BUDDY_CONTIGUOUS_ALLOCATION),
+ "contiguous alloc failed\n");
+ end = ktime_get();
+ elapsed_ms = ktime_to_ms(ktime_sub(end, start));
+
+ kunit_info(test, "Contiguous alloc after fragmentation: %lu ms\n",
+ elapsed_ms);
+
+ gpu_buddy_free_list(&mm, &results, 0);
+ gpu_buddy_fini(&mm);
+
+ /*
+ * Repeated alloc throughput from a maximally fragmented pool
+ *
+ * Fill a 4 GiB pool with 4 KiB allocations, free even-indexed blocks
+ * as cleared and odd-indexed blocks as dirty. The alternating pattern
+ * ensures every adjacent buddy pair has one cleared half and one dirty
+ * half, so each pair lands on opposite sides of the old merge barrier.
+ * Each of the 16 384 x 256 KiB allocations in the timed loop has to
+ * pay the __force_merge() cost on the alloc path under the old design.
+ * With the dirty-tracker design the pool collapses to one max_order
+ * block during free(), so each alloc is a simple O(log N) split.
+ */
+ KUNIT_ASSERT_FALSE_MSG(test, gpu_buddy_init(&mm, SZ_4G, SZ_4K),
+ "buddy_init failed\n");
+
+ for (i = 0; i < SZ_4G / SZ_4K; i++)
+ KUNIT_ASSERT_FALSE_MSG(test,
+ gpu_buddy_alloc_blocks(&mm, 0, SZ_4G, SZ_4K, SZ_4K,
+ &allocated, 0),
+ "buddy_alloc hit an error size=%u\n", SZ_4K);
+
+ count = 0;
+ list_for_each_entry_safe(block, tmp, &allocated, link) {
+ if (count++ % 2 == 0)
+ list_move_tail(&block->link, &clear_blocks);
+ else
+ list_move_tail(&block->link, &dirty_blocks);
+ }
+
+ gpu_buddy_free_list(&mm, &clear_blocks, GPU_BUDDY_CLEARED);
+ gpu_buddy_free_list(&mm, &dirty_blocks, 0);
+
+ start = ktime_get();
+ for (i = 0; i < SZ_4G / SZ_256K; i++)
+ KUNIT_ASSERT_FALSE_MSG(test,
+ gpu_buddy_alloc_blocks(&mm, 0, SZ_4G, SZ_256K, SZ_4K,
+ &results, 0),
+ "buddy_alloc hit an error size=%u\n", SZ_256K);
+ end = ktime_get();
+ elapsed_ms = ktime_to_ms(ktime_sub(end, start));
+
+ kunit_info(test, "Repeated 256 KiB allocs from fragmented pool: %lu ms\n",
+ elapsed_ms);
+
+ gpu_buddy_free_list(&mm, &results, 0);
+ gpu_buddy_fini(&mm);
+}
+
static void gpu_test_buddy_alloc_range_bias(struct kunit *test)
{
u32 mm_size, size, ps, bias_size, bias_start, bias_end, bias_rem;
@@ -1437,6 +1546,7 @@ static struct kunit_case gpu_buddy_tests[] = {
KUNIT_CASE(gpu_test_buddy_alloc_range),
KUNIT_CASE(gpu_test_buddy_alloc_range_bias),
KUNIT_CASE_SLOW(gpu_test_buddy_fragmentation_performance),
+ KUNIT_CASE_SLOW(gpu_test_buddy_dirty_tracker_performance),
KUNIT_CASE(gpu_test_buddy_alloc_exceeds_max_order),
KUNIT_CASE(gpu_test_buddy_offset_aligned_allocation),
KUNIT_CASE(gpu_test_buddy_subtree_offset_alignment_stress),
--
2.34.1
^ permalink raw reply related [flat|nested] 5+ messages in thread* Re: [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test
2026-07-31 7:07 ` [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test Arunpravin Paneer Selvam
@ 2026-07-31 7:19 ` sashiko-bot
0 siblings, 0 replies; 5+ messages in thread
From: sashiko-bot @ 2026-07-31 7:19 UTC (permalink / raw)
To: Arunpravin Paneer Selvam; +Cc: dri-devel
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [Medium] Massive memory footprint in KUnit test triggers OOM in constrained environments
- [Medium] Benchmark logic bypasses the measured path in the baseline allocator
--
--- Patch [2]: [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test ---
commit 7af49c21f49c14e81ea1bc47365402c2ec976aec
Author: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
gpu/tests/buddy: add dirty tracker performance KUnit test
This commit adds a KUnit test `gpu_test_buddy_dirty_tracker_performance`
to demonstrate the performance benefits of the decoupled dirty-tracker
design compared to the dual-tree approach. It measures contiguous
allocation latency and repeated allocation throughput after fragmentation.
> diff --git a/drivers/gpu/tests/gpu_buddy_test.c b/drivers/gpu/tests/gpu_buddy_test.c
> index f8e56da5058e6..3fc2bfa66ee15 100644
> --- a/drivers/gpu/tests/gpu_buddy_test.c
> +++ b/drivers/gpu/tests/gpu_buddy_test.c
[ ... ]
> +static void gpu_test_buddy_dirty_tracker_performance(struct kunit *test)
> +{
> + struct gpu_buddy_block *block, *tmp;
> + unsigned long elapsed_ms;
[ ... ]
> + KUNIT_ASSERT_FALSE_MSG(test, gpu_buddy_init(&mm, SZ_4G, SZ_4K),
> + "buddy_init failed\n");
> +
> + for (i = 0; i < SZ_4G / SZ_4K; i++)
> + KUNIT_ASSERT_FALSE_MSG(test,
> + gpu_buddy_alloc_blocks(&mm, 0, SZ_4G, SZ_4K, SZ_4K,
> + &allocated, 0),
> + "buddy_alloc hit an error size=%u\n", SZ_4K);
[Severity: Medium]
Will this massive memory footprint trigger an Out-Of-Memory panic in
constrained environments?
Allocating a 4 GiB pool using 4 KiB block granularity requires creating over
1 million struct gpu_buddy_block nodes dynamically from the slab. This will
consume well over 100 MiB of kernel memory.
Default UML environments for KUnit often have a memory limit of 256 MiB or
less, and this exhaustive allocation is likely to crash the test suite.
[ ... ]
> + /*
> + * Repeated alloc throughput from a maximally fragmented pool
> + *
> + * Fill a 4 GiB pool with 4 KiB allocations, free even-indexed blocks
> + * as cleared and odd-indexed blocks as dirty. The alternating pattern
> + * ensures every adjacent buddy pair has one cleared half and one dirty
> + * half, so each pair lands on opposite sides of the old merge barrier.
> + * Each of the 16 384 x 256 KiB allocations in the timed loop has to
> + * pay the __force_merge() cost on the alloc path under the old design.
> + * With the dirty-tracker design the pool collapses to one max_order
> + * block during free(), so each alloc is a simple O(log N) split.
> + */
[ ... ]
> + start = ktime_get();
> + for (i = 0; i < SZ_4G / SZ_256K; i++)
> + KUNIT_ASSERT_FALSE_MSG(test,
> + gpu_buddy_alloc_blocks(&mm, 0, SZ_4G, SZ_256K, SZ_4K,
> + &results, 0),
> + "buddy_alloc hit an error size=%u\n", SZ_256K);
[Severity: Medium]
Does this benchmark logic bypass the measured path in the baseline allocator?
The commit message states this measures the __force_merge() cost on the
alloc path under the old design. However, the flags parameter is passed as 0
here, which omits GPU_BUDDY_CONTIGUOUS_ALLOCATION.
Without the contiguous flag, the baseline allocator does not enforce that
the 256 KiB allocation must be a single block. Instead, it will fall back to
satisfying the 256 KiB request with 64 independent 4 KiB blocks.
Because 4 KiB blocks are readily available in the fragmented pool, the
allocation succeeds without ever invoking the __force_merge() fallback.
Are we actually comparing the cost of 64 tree lookups against 1 tree lookup,
rather than measuring the __force_merge() overhead?
> + end = ktime_get();
> + elapsed_ms = ktime_to_ms(ktime_sub(end, start));
> +
> + kunit_info(test, "Repeated 256 KiB allocs from fragmented pool: %lu ms\n",
> + elapsed_ms);
> +
> + gpu_buddy_free_list(&mm, &results, 0);
> + gpu_buddy_fini(&mm);
> +}
--
Sashiko AI review · https://sashiko.dev/#/patchset/20260731070741.2654251-1-Arunpravin.PaneerSelvam@amd.com?part=2
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
2026-07-31 7:07 [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker Arunpravin Paneer Selvam
2026-07-31 7:07 ` [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test Arunpravin Paneer Selvam
@ 2026-07-31 7:21 ` sashiko-bot
2026-07-31 7:28 ` ✗ Fi.CI.BUILD: failure for series starting with [v7,1/2] " Patchwork
2 siblings, 0 replies; 5+ messages in thread
From: sashiko-bot @ 2026-07-31 7:21 UTC (permalink / raw)
To: Arunpravin Paneer Selvam; +Cc: dri-devel
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Mempool deadlock in `gpu_buddy_reset_clear` and free paths during extent allocation when multiple extents are required within a single locked section.
--
--- Patch [1]: [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker ---
commit e1f55d34ee0b3bd852cd9853e793c8cbdf7fc6f6
Author: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
This commit replaces the dual-tree buddy allocator design with a single free
tree and an out-of-band dirty tracker using an augmented interval rbtree. It
also moves extent allocations to a mempool to guarantee progress and avoid
deadlocks associated with __GFP_NOFAIL.
> diff --git a/drivers/gpu/buddy.c b/drivers/gpu/buddy.c
> --- a/drivers/gpu/buddy.c
> +++ b/drivers/gpu/buddy.c
> @@ -34,6 +35,441 @@
> #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, so the underlying slab alloc almost always
> + * succeeds via reclaim; the reserve only backstops the rare case where
> + * it still returns NULL (e.g. the current task is an OOM victim),
> + * guaranteeing a non-NULL extent without __GFP_NOFAIL. Because each
> + * alloc can independently wait for reclaim, the reserve need not scale
> + * with the number of extents added in one locked section (e.g. by
> + * gpu_buddy_reset_clear()).
> + */
> +#define GPU_DIRTY_EXTENT_POOL_MIN 1
[ ... ]
> +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 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;
> +}
[ ... ]
> @@ -572,42 +1055,39 @@ 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);
[ ... ]
> + 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));
[Severity: High]
Could this loop trigger a mempool deadlock during memory pressure?
In gpu_dirty_tracker_init(), the extent pool is created with a minimum
size of GPU_DIRTY_EXTENT_POOL_MIN (1). However, this loop in
gpu_buddy_reset_clear() holds the driver lock and iterates over the free
tree, potentially calling gpu_dirty_tracker_mark_dirty() multiple times.
If the underlying slab allocation fails during extent_alloc(),
mempool_alloc() will use the single reserved extent. If a second allocation
fails within the same locked section, mempool_alloc() with GFP_KERNEL will
sleep and retry indefinitely.
Since the driver lock is held, no other thread can free blocks to return
extents to the mempool, resulting in a deadlock. Does the mempool reserve
need to scale with the maximum number of extents that can be allocated
in a single locked section?
--
Sashiko AI review · https://sashiko.dev/#/patchset/20260731070741.2654251-1-Arunpravin.PaneerSelvam@amd.com?part=1
^ permalink raw reply [flat|nested] 5+ messages in thread* ✗ Fi.CI.BUILD: failure for series starting with [v7,1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
2026-07-31 7:07 [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker Arunpravin Paneer Selvam
2026-07-31 7:07 ` [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test Arunpravin Paneer Selvam
2026-07-31 7:21 ` [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker sashiko-bot
@ 2026-07-31 7:28 ` Patchwork
2 siblings, 0 replies; 5+ messages in thread
From: Patchwork @ 2026-07-31 7:28 UTC (permalink / raw)
To: Arunpravin Paneer Selvam; +Cc: intel-gfx
== Series Details ==
Series: series starting with [v7,1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
URL : https://patchwork.freedesktop.org/series/171418/
State : failure
== Summary ==
Error: patch https://patchwork.freedesktop.org/api/1.0/series/171418/revisions/1/mbox/ not applied
Applying: gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
Using index info to reconstruct a base tree...
M drivers/gpu/buddy.c
Falling back to patching base and 3-way merge...
Auto-merging drivers/gpu/buddy.c
CONFLICT (content): Merge conflict in drivers/gpu/buddy.c
error: Failed to merge in the changes.
hint: Use 'git am --show-current-patch=diff' to see the failed patch
Patch failed at 0001 gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker
When you have resolved this problem, run "git am --continue".
If you prefer to skip this patch, run "git am --skip" instead.
To restore the original branch and stop patching, run "git am --abort".
Build failed, no error log produced
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2026-07-31 7:29 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-07-31 7:07 [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker Arunpravin Paneer Selvam
2026-07-31 7:07 ` [PATCH v7 2/2] gpu/tests/buddy: add dirty tracker performance KUnit test Arunpravin Paneer Selvam
2026-07-31 7:19 ` sashiko-bot
2026-07-31 7:21 ` [PATCH v7 1/2] gpu/buddy: replace dual-tree/force_merge with decoupled dirty tracker sashiko-bot
2026-07-31 7:28 ` ✗ Fi.CI.BUILD: failure for series starting with [v7,1/2] " Patchwork
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.