On Sat Jul 11, 2026 at 7:36 AM BST, Yury Norov wrote:
> Compare the cost of allocating and freeing variable-sized regions using
> a bitmap, IDA and a Maple Tree. All implementations process the same
> randomly generated sequence of region sizes, ranging from 1 to 32 entries,
> until the configured capacity is exhausted.
>
> Run the benchmark at several capacities to show how the approaches
> scale. Report allocation and free times separately because bitmap,
> IDA and Maple Tree removal have substantially different costs.
>
> On x86/kvm, the output example is:
>
> type alloc (ns) free (ns) capacity memory (B)
> bitmap 179573071 342105 1000000 125000
> IDA 46555636 33931498 1000000 134864
> maple 18629665 19741396 1000000 1548304
> bitmap 1630912 30933 100000 12504
> IDA 6144785 3354590 100000 14288
> maple 1745026 1825032 100000 155408
> bitmap 28448 3374 10000 1256
> IDA 418978 333641 10000 1872
> maple 185398 211138 10000 15632
> bitmap 2253 610 1000 128
> IDA 42755 36432 1000 144
> maple 19728 23474 1000 1552
>
> Reported IDA and Maple Tree memory figures exclude slab overhead
> and transient allocations. The Maple Tree figure is additionally
> a lower-bound estimate that assumes fully occupied leaf nodes and
> excludes internal nodes.
>
> IDA has no region-allocation API, so each region is implemented as
> a sequence of single-ID allocations. The IDs remain contiguous
> because this benchmark fills an initially empty IDA monotonically.
>
> The benchmark is motivated by the discussion at the link below about
> choosing the best data structure for the channel ID pool with the
> capacity of 2048 IDs for the nova GPU driver.
>
> Specifically for 2048 IDs the result is:
>
> bitmap 3907 831 2048 256
> IDA 83545 71007 2048 848
> maple 34632 39710 2048 3600
>
> Link:
> https://lore.kernel.org/all/[email protected]/
> Signed-off-by: Yury Norov <[email protected]>
> ---
> v2: add IDA benchmark (Matthew)
> v1: https://lore.kernel.org/all/[email protected]/
>
> MAINTAINERS | 3 +
> lib/Kconfig.debug | 9 ++
> lib/Makefile | 1 +
> lib/region_alloc_benchmark.c | 174 +++++++++++++++++++++++++++++++++++
> 4 files changed, 187 insertions(+)
> create mode 100644 lib/region_alloc_benchmark.c
>
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 7cc4bca5a2c5..9e487a94aba4 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -4615,6 +4615,7 @@ F: lib/bitmap.c
> F: lib/cpumask.c
> F: lib/find_bit.c
> F: lib/find_bit_benchmark.c
> +F: lib/region_alloc_benchmark.c
> F: lib/test_bitmap.c
> F: lib/tests/cpumask_kunit.c
> F: tools/include/linux/bitfield.h
> @@ -15581,6 +15582,7 @@ F: Documentation/core-api/maple_tree.rst
> F: include/linux/maple_tree.h
> F: include/trace/events/maple_tree.h
> F: lib/maple_tree.c
> +F: lib/region_alloc_benchmark.c
> F: lib/test_maple_tree.c
> F: rust/helpers/maple_tree.c
> F: rust/kernel/maple_tree.rs
> @@ -29323,6 +29325,7 @@ F: Documentation/core-api/xarray.rst
> F: include/linux/idr.h
> F: include/linux/xarray.h
> F: lib/idr.c
> +F: lib/region_alloc_benchmark.c
> F: lib/test_xarray.c
> F: lib/xarray.c
> F: tools/testing/radix-tree
> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> index 1244dcac2294..40872263cff1 100644
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -2683,6 +2683,15 @@ config FIND_BIT_BENCHMARK
>
> If unsure, say N.
>
> +config REGION_ALLOC_BENCHMARK
> + tristate "Benchmark bitmap, IDA and Maple Tree region allocation"
> + help
> + This builds a microbenchmark comparing variable-sized region
> + allocation using bitmaps, IDA and Maple Tree. The benchmark
> + runs at module load time.
> +
> + If unsure, say N.
> +
> config FIND_BIT_BENCHMARK_RUST
> tristate "Test find_bit functions in Rust"
> depends on RUST
> diff --git a/lib/Makefile b/lib/Makefile
> index 7f75cc6edf94..adb18810e3f7 100644
> --- a/lib/Makefile
> +++ b/lib/Makefile
> @@ -64,6 +64,7 @@ obj-y += hexdump.o
> obj-$(CONFIG_TEST_HEXDUMP) += test_hexdump.o
> obj-y += kstrtox.o
> obj-$(CONFIG_FIND_BIT_BENCHMARK) += find_bit_benchmark.o
> +obj-$(CONFIG_REGION_ALLOC_BENCHMARK) += region_alloc_benchmark.o
> obj-$(CONFIG_FIND_BIT_BENCHMARK_RUST) += find_bit_benchmark_rust.o
> obj-$(CONFIG_TEST_BPF) += test_bpf.o
> test_dhry-objs := dhry_1.o dhry_2.o dhry_run.o
> diff --git a/lib/region_alloc_benchmark.c b/lib/region_alloc_benchmark.c
> new file mode 100644
> index 000000000000..0e000c4f9926
> --- /dev/null
> +++ b/lib/region_alloc_benchmark.c
> @@ -0,0 +1,174 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/* Benchmark bitmap, IDA and Maple Tree allocation of variable-sized
> regions. */
> +
> +#include <linux/bitmap.h>
> +#include <linux/idr.h>
> +#include <linux/kernel.h>
> +#include <linux/maple_tree.h>
> +#include <linux/module.h>
> +#include <linux/printk.h>
> +#include <linux/random.h>
> +#include <linux/xarray.h>
> +
> +#define MAP_SIZE (1000000UL)
> +#define REGION_MAX_SIZE 32
> +
> +static DECLARE_BITMAP(alloc_bitmap, MAP_SIZE) __initdata;
> +/* One more request guarantees that even an all-ones trace reaches ENOSPC. */
> +static u8 region_sizes[MAP_SIZE + 1] __initdata;
> +static unsigned long region_indexes[MAP_SIZE] __initdata;
> +
> +static unsigned long __init benchmark_bitmap(unsigned long capacity)
> +{
> + unsigned long count, index;
> + ktime_t alloc_time, free_time;
> + size_t memory;
> +
> + bitmap_zero(alloc_bitmap, MAP_SIZE);
> + alloc_time = ktime_get();
To eliminate possible effects of optimization messing up benchmark, please add a
barrier() after start timing and a barrier() before end timing. Probably you'd
want one between each loop iteration too.
> + for (count = 0; count < ARRAY_SIZE(region_sizes); count++) {
> + index = bitmap_find_next_zero_area(alloc_bitmap, capacity, 0,
> + region_sizes[count], 0);
> + if (index >= capacity)
> + break;
> +
> + region_indexes[count] = index;
> + bitmap_set(alloc_bitmap, index, region_sizes[count]);
> + }
> + alloc_time = ktime_get() - alloc_time;
> +
> + index = count;
> + free_time = ktime_get();
> + while (index--)
> + bitmap_clear(alloc_bitmap, region_indexes[index],
> + region_sizes[index]);
> + free_time = ktime_get() - free_time;
> + memory = BITS_TO_LONGS(capacity) * sizeof(unsigned long);
> + pr_err("%-6s %12llu %12llu %8lu %10zu\n", "bitmap", alloc_time,
> + free_time, capacity, memory);
I wonder if the numbers could be more useful if it is somehow normalized (e.g.
by how many iterations are performed?)? This number can be computed by adding
up the prefix of region_sizes until it cover all capacity.
Also, the capacity should probably be made part of the report header for all 3
ID allocators. Given this is randomized benchmark the single line, when taken
out of context, is not very meaningful anyway.
Something like this might be better (fake numbers below, you might also need
some fix point printing):
capacity = 1000000, regions = 100000
type alloc (ns/region) free (ns/region) memory (B/capacity)
bitmap 1795.730 3.421 0.125
IDA 465.556 339.314 0.134
maple 186.296 197.413 1.548
capacity = 100000, regions = 10000
bitmap 163.091 3.093 0.125
IDA 614.478 335.459 0.142
maple 174.502 182.503 1.554
capacity = 10000, regions = 1000
bitmap 28.448 3.374 0.125
IDA 418.978 333.641 0.187
maple 185.398 211.138 1.563
capacity = 1000, regions = 100
bitmap 22.530 6.100 0.128
IDA 427.550 364.320 0.144
maple 197.280 234.740 1.552
Best,
Gary
> +
> + return count;
> +}
> +
> +static size_t __init ida_memory(unsigned long nr_ids)
> +{
> + unsigned long entries = DIV_ROUND_UP(nr_ids, IDA_BITMAP_BITS);
> + unsigned long bitmaps = nr_ids / IDA_BITMAP_BITS;
> + unsigned long nodes = 0;
> +
> + if (nr_ids % IDA_BITMAP_BITS > BITS_PER_XA_VALUE)
> + bitmaps++;
> +
> + while (entries > 1) {
> + entries = DIV_ROUND_UP(entries, XA_CHUNK_SIZE);
> + nodes += entries;
> + }
> +
> + return sizeof(struct ida) + bitmaps * sizeof(struct ida_bitmap) +
> + nodes * sizeof(struct xa_node);
> +}
> +
> +static unsigned long __init benchmark_ida(unsigned long capacity)
> +{
> + struct ida ida = IDA_INIT(ida);
> + unsigned long count, index, offset, nr_ids = 0;
> + ktime_t alloc_time, free_time;
> + int id = -ENOSPC;
> +
> + alloc_time = ktime_get();
> + for (count = 0; count < ARRAY_SIZE(region_sizes); count++) {
> + for (offset = 0; offset < region_sizes[count]; offset++) {
> + id = ida_alloc_max(&ida, capacity - 1, GFP_KERNEL);
> + if (id < 0)
> + break;
> + if (!offset)
> + region_indexes[count] = id;
> + }
> + if (id < 0) {
> + while (offset--)
> + ida_free(&ida, region_indexes[count] + offset);
> + break;
> + }
> + WARN_ON(id != region_indexes[count] + region_sizes[count] - 1);
> + nr_ids += region_sizes[count];
> + }
> + alloc_time = ktime_get() - alloc_time;
> + WARN_ON(id != -ENOSPC);
> +
> + index = count;
> + free_time = ktime_get();
> + while (index--)
> + for (offset = 0; offset < region_sizes[index]; offset++)
> + ida_free(&ida, region_indexes[index] + offset);
> + free_time = ktime_get() - free_time;
> + pr_err("%-6s %12llu %12llu %8lu %10zu\n", "IDA", alloc_time,
> + free_time, capacity, ida_memory(nr_ids));
> + ida_destroy(&ida);
> +
> + return count;
> +}
> +
