Range-restricted allocations (GPU_BUDDY_RANGE_ALLOCATION) go through
__alloc_range_bias(), a depth-first walk over the buddy tree that visits
the higher-address child first. It skips blocks that are too small,
outside the range or ALLOCATED, but always descends SPLIT blocks, even
when nothing below them is free. Range allocations are placed top-down,
so they pack below the range end and each new allocation walks over all
blocks already allocated there: O(N) per allocation, O(N^2) in total,
under the driver's allocator lock.
This is a property of the range-biased allocation algorithm rather than
of any particular memory configuration. Any driver that repeatedly
restricts placement to a window can encounter this behavior once enough
small allocations accumulate in that window.
One configuration where the problem becomes particularly visible is a
small-BAR system. Drivers restrict every CPU-visible allocation to that
window with a range placement (fpfn = 0, lpfn = BAR size).
xe_pat@pt-caching-random-offsets does about 130k 4K range allocations
and hit the CI 180s inactivity timeout on a debug kernel inside
gpu_buddy_alloc_blocks(). With the full VRAM visible the same test does
not use the range path.
<6> [59.436001] [IGT] xe_pat: starting subtest pt-caching-random-offsets
<6> [239.585293] task:xe_pat state:R running task stack:0
pid:3700 tgid:3700 ppid:1957 task_flags:0x400100 flags:0x00000020
<6> [239.585304] Call Trace:
<6> [239.585307] <TASK>
<6> [239.585309] ? irqentry_exit+0x2a2/0x8d0
<6> [239.585314] ? trace_hardirqs_off_finish+0xc7/0xe0
<6> [239.585320] ? sysvec_apic_timer_interrupt+0x57/0xc0
<6> [239.585326] ? asm_sysvec_apic_timer_interrupt+0x1b/0x20
<6> [239.585334] ? gpu_buddy_alloc_blocks+0x442/0xec0
<6> [239.585341] ? gpu_buddy_alloc_blocks+0xa37/0xec0
<6> [239.585353] ? xe_ttm_vram_mgr_new+0x1c7/0x340 [xe]
<6> [239.585506] ? ttm_resource_alloc+0x50/0xd0 [ttm]
<6> [239.585518] ? ttm_bo_alloc_at_place+0x84/0x180 [ttm]
<6> [239.585527] ? ttm_bo_alloc_resource+0x115/0x5e0 [ttm]
<6> [239.585538] ? ttm_bo_validate+0x94/0x1c0 [ttm]
<6> [239.585546] ? lock_is_held_type+0xa3/0x130
<6> [239.585552] ? ttm_bo_init_reserved+0xa5/0x1f0 [ttm]
<6> [239.585560] ? xe_bo_init_locked+0x20a/0x650 [xe]
<6> [239.585616] ? __pfx_xe_ttm_bo_destroy+0x10/0x10 [xe]
<6> [239.585672] ? stack_depot_save_flags+0x28/0x960
<6> [239.585679] ? __xe_bo_create_locked+0xb1/0xa20 [xe]
<6> [239.585733] ? _raw_spin_unlock_irqrestore+0x52/0x80
<6> [239.585743] ? __create_object+0x67/0xe0
<6> [239.585747] ? trace_hardirqs_on+0x22/0xe0
<6> [239.585753] ? _raw_spin_unlock_irqrestore+0x52/0x80
<6> [239.585760] ? xe_bo_create_pin_map_at_aligned+0x51/0x1f0 [xe]
<6> [239.585813] ? __kmalloc_cache_noprof+0x494/0x670
<6> [239.585819] ? xe_bo_create_pin_map+0x1f/0x40 [xe]
<6> [239.585870] ? xe_pt_create+0xb5/0x430 [xe]
The patch tracks, per block, the largest order of any FREE block in its
subtree and lets __alloc_range_bias() skip subtrees that cannot satisfy
the request. Skipped subtrees contain nothing the old walk could split
or return, and the remaining blocks are visited in the same order, so
the selected block and placement are unchanged. The new s8 field uses
existing tail padding.
Results
n_alloc avg_blocks_visited avg_allocation_time
without fix with fix without fix with fix
4096 4141 46 10.7 us 0.293 us
65536 127041 48 536 us 0.259 us
126976 249911 47 1.46 ms 0.283 us
v2:
- Reword cover letter and commit message to describe this as a general
fix for the algorithmic cost of range-biased allocation rather than a
small-BAR specific issue (Matt A)
- No code changes
v3: (Sashiko)
- Use div_u64() in the KUnit test to fix the 32-bit build
v4: (Matt A)
- Drop the KUnit tests
Sanjay Yadav (1):
gpu/buddy: Prune exhausted subtrees in range-biased allocation
drivers/gpu/buddy.c | 56 +++++++++++++++++++++++++++------------
include/linux/gpu_buddy.h | 7 +++++
2 files changed, 46 insertions(+), 17 deletions(-)
--
2.52.0
