miri/alloc_addresses/reuse_pool.rs
1//! Manages a pool of addresses that can be reused.
2
3use rand::Rng;
4use rustc_abi::{Align, Size};
5
6use crate::concurrency::VClock;
7use crate::{MemoryKind, MiriConfig, ThreadId};
8
9const MAX_POOL_SIZE: usize = 64;
10
11/// The pool strikes a balance between exploring more possible executions and making it more likely
12/// to find bugs. The hypothesis is that bugs are more likely to occur when reuse happens for
13/// allocations with the same layout, since that can trigger e.g. ABA issues in a concurrent data
14/// structure. Therefore we only reuse allocations when size and alignment match exactly.
15#[derive(Debug)]
16pub struct ReusePool {
17 address_reuse_rate: f64,
18 address_reuse_cross_thread_rate: f64,
19 /// The i-th element in `pool` stores allocations of alignment `2^i`. We store these reusable
20 /// allocations as address-size pairs, the list must be sorted by the size and then the thread ID.
21 ///
22 /// Each of these maps has at most MAX_POOL_SIZE elements, and since alignment is limited to
23 /// less than 64 different possible values, that bounds the overall size of the pool.
24 ///
25 /// We also store the ID and the data-race clock of the thread that donated this pool element,
26 /// to ensure synchronization with the thread that picks up this address.
27 pool: Vec<Vec<(u64, Size, ThreadId, VClock)>>,
28}
29
30impl ReusePool {
31 pub fn new(config: &MiriConfig) -> Self {
32 ReusePool {
33 address_reuse_rate: config.address_reuse_rate,
34 address_reuse_cross_thread_rate: config.address_reuse_cross_thread_rate,
35 pool: vec![],
36 }
37 }
38
39 /// Call this when we are using up a lot of the address space: if memory reuse is enabled at all,
40 /// this will bump the intra-thread reuse rate to 100% so that we can keep running this program as
41 /// long as possible.
42 pub fn address_space_shortage(&mut self) {
43 if self.address_reuse_rate > 0.0 {
44 self.address_reuse_rate = 1.0;
45 }
46 }
47
48 fn subpool(&mut self, align: Align) -> &mut Vec<(u64, Size, ThreadId, VClock)> {
49 let pool_idx: usize = align.bytes().trailing_zeros().try_into().unwrap();
50 if self.pool.len() <= pool_idx {
51 self.pool.resize(pool_idx + 1, Vec::new());
52 }
53 &mut self.pool[pool_idx]
54 }
55
56 pub fn add_addr(
57 &mut self,
58 rng: &mut impl Rng,
59 addr: u64,
60 size: Size,
61 align: Align,
62 kind: MemoryKind,
63 thread: ThreadId,
64 clock: impl FnOnce() -> VClock,
65 ) {
66 // Let's see if we even want to remember this address.
67 // We don't remember stack addresses since there's so many of them (so the perf impact is big).
68 if kind == MemoryKind::Stack || !rng.random_bool(self.address_reuse_rate) {
69 return;
70 }
71 let clock = clock();
72 // Determine the pool to add this to, and where in the pool to put it.
73 let subpool = self.subpool(align);
74 let pos = subpool.partition_point(|(_addr, other_size, other_thread, _)| {
75 (*other_size, *other_thread) < (size, thread)
76 });
77 // Make sure the pool does not grow too big.
78 if subpool.len() >= MAX_POOL_SIZE {
79 // Pool full. Replace existing element, or last one if this would be even bigger.
80 let clamped_pos = pos.min(subpool.len() - 1);
81 subpool[clamped_pos] = (addr, size, thread, clock);
82 return;
83 }
84 // Add address to pool, at the right position.
85 subpool.insert(pos, (addr, size, thread, clock));
86 }
87
88 /// Returns the address to use and optionally a clock we have to synchronize with.
89 pub fn take_addr(
90 &mut self,
91 rng: &mut impl Rng,
92 size: Size,
93 align: Align,
94 kind: MemoryKind,
95 thread: ThreadId,
96 ) -> Option<(u64, Option<VClock>)> {
97 // Determine whether we'll even attempt a reuse. As above, we don't do reuse for stack addresses.
98 if kind == MemoryKind::Stack || !rng.random_bool(self.address_reuse_rate) {
99 return None;
100 }
101 let cross_thread_reuse = rng.random_bool(self.address_reuse_cross_thread_rate);
102 // Determine the pool to take this from.
103 let subpool = self.subpool(align);
104 // Let's see if we can find something of the right size. We want to find the full range of
105 // such items, beginning with the first, so we can't use `binary_search_by_key`. If we do
106 // *not* want to consider other thread's allocations, we effectively use the lexicographic
107 // order on `(size, thread)`.
108 let begin = subpool.partition_point(|(_addr, other_size, other_thread, _)| {
109 *other_size < size
110 || (*other_size == size && !cross_thread_reuse && *other_thread < thread)
111 });
112 let mut end = begin;
113 while let Some((_addr, other_size, other_thread, _)) = subpool.get(end) {
114 if *other_size != size {
115 break;
116 }
117 if !cross_thread_reuse && *other_thread != thread {
118 // We entered the allocations of another thread.
119 break;
120 }
121 end += 1;
122 }
123 if end == begin {
124 // Could not find any item of the right size.
125 return None;
126 }
127 // Pick a random element with the desired size.
128 let idx = rng.random_range(begin..end);
129 // Remove it from the pool and return.
130 let (chosen_addr, chosen_size, chosen_thread, clock) = subpool.remove(idx);
131 debug_assert!(chosen_size >= size && chosen_addr % align.bytes() == 0);
132 debug_assert!(cross_thread_reuse || chosen_thread == thread);
133 // No synchronization needed if we reused from the current thread.
134 Some((chosen_addr, if chosen_thread == thread { None } else { Some(clock) }))
135 }
136}