1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
use std::collections::BTreeMap;
use volcanic_mem::{Error, VirtualAllocation, VirtualBlock};
struct Rng(u64);
impl Rng {
fn next(&mut self) -> u64 {
self.0 = self.0.wrapping_add(0x9e37_79b9_7f4a_7c15);
let mut z = self.0;
z = (z ^ (z >> 30)).wrapping_mul(0xbf58_476d_1ce4_e5b9);
z = (z ^ (z >> 27)).wrapping_mul(0x94d0_49bb_1331_11eb);
z ^ (z >> 31)
}
fn below(&mut self, n: u64) -> u64 {
self.next() % n
}
/// Log-uniform size in `1..=max`, matching the skew of real GPU workloads.
fn size(&mut self, max: u64) -> u64 {
let bits = 64 - max.leading_zeros() as u64;
(self.next() & ((1u64 << (self.below(bits) + 1)) - 1)).clamp(1, max)
}
}
fn check_no_overlap(live: &BTreeMap<u64, u64>, offset: u64, size: u64) {
if let Some((&start, &len)) = live.range(..=offset).next_back() {
assert!(start + len <= offset, "overlaps [{start}, {})", start + len);
}
if let Some((&start, _)) = live.range(offset..).next() {
assert!(offset + size <= start, "overlaps allocation at {start}");
}
}
/// TLSF must never report out-of-memory while some gap could hold the request.
fn assert_no_fitting_gap(live: &BTreeMap<u64, u64>, block_size: u64, size: u64, alignment: u64) {
let mut gap_start = 0u64;
let ends = live
.iter()
.map(|(&start, &len)| (start, start + len))
.chain([(block_size, block_size)]);
for (start, end) in ends {
let aligned = gap_start.next_multiple_of(alignment);
assert!(
aligned + size > start,
"spurious OOM: [{gap_start}, {start}) fits {size} @ {alignment}"
);
gap_start = end;
}
}
fn stress(seed: u64, block_size: u64, max_size: u64, steps: usize) {
let mut rng = Rng(seed);
let mut block = VirtualBlock::new(block_size).unwrap();
let mut handles: Vec<(VirtualAllocation, u64)> = Vec::new();
let mut live = BTreeMap::new();
let (mut ok, mut full) = (0, 0);
for step in 0..steps {
let grow = handles.is_empty() || rng.below(100) < 55;
if grow {
let size = rng.size(max_size);
let alignment = 1 << rng.below(13);
match block.allocate(size, alignment) {
Ok(allocation) => {
let offset = allocation.offset();
assert_eq!(offset % alignment, 0, "misaligned");
assert!(offset + size <= block_size, "out of bounds");
assert_eq!(block.allocation_size(&allocation), size);
check_no_overlap(&live, offset, size);
live.insert(offset, size);
handles.push((allocation, size));
ok += 1;
}
Err(Error::OutOfMemory) => {
assert_no_fitting_gap(&live, block_size, size, alignment);
full += 1;
}
Err(error) => panic!("unexpected error {error}"),
}
} else {
let index = rng.below(handles.len() as u64) as usize;
let (allocation, size) = handles.swap_remove(index);
assert_eq!(live.remove(&allocation.offset()), Some(size));
block.free(allocation);
}
let used: u64 = live.values().sum();
assert_eq!(block.free_bytes(), block_size - used);
if step % 64 == 0 {
block.validate();
}
}
assert!(
ok > 0 && full > 0,
"workload did not hit both paths: ok={ok} full={full}"
);
for (allocation, _) in handles.drain(..) {
block.free(allocation);
}
block.validate();
assert!(block.is_empty());
assert_eq!(block.free_bytes(), block_size);
let whole = block
.allocate(block_size, 1)
.expect("fully coalesced block");
assert_eq!(whole.offset(), 0);
}
#[test]
fn small_allocations_in_small_block() {
for seed in 0..8 {
stress(seed, 64 << 10, 4 << 10, 50_000);
}
}
#[test]
fn mixed_allocations_in_large_block() {
for seed in 100..104 {
stress(seed, 256 << 20, 16 << 20, 100_000);
}
}
#[test]
fn non_power_of_two_block() {
stress(7, 3 * 1000 * 1000 + 17, 64 << 10, 50_000);
}
#[test]
fn exact_fit_and_alignment_padding() {
let mut block = VirtualBlock::new(1024).unwrap();
let a = block.allocate(100, 1).unwrap();
let b = block.allocate(100, 256).unwrap();
assert_eq!((a.offset(), b.offset()), (0, 256));
let c = block.allocate(156, 1).unwrap();
assert_eq!(c.offset(), 100, "padding before b must be reusable");
let d = block.allocate(668, 1).unwrap();
assert_eq!(d.offset(), 356);
assert!(matches!(block.allocate(1, 1), Err(Error::OutOfMemory)));
block.validate();
for allocation in [b, d, a, c] {
block.free(allocation);
block.validate();
}
assert_eq!(block.allocate(1024, 1024).unwrap().offset(), 0);
}
#[test]
fn rejects_invalid_requests() {
let mut block = VirtualBlock::new(4096).unwrap();
assert!(matches!(block.allocate(0, 1), Err(Error::ZeroSize)));
assert!(matches!(
block.allocate(1, 3),
Err(Error::InvalidAlignment(3))
));
assert!(matches!(block.allocate(4097, 1), Err(Error::OutOfMemory)));
assert!(matches!(VirtualBlock::new(0), Err(Error::ZeroSize)));
}