tokio/sync/mpsc/list.rs
1//! A concurrent, lock-free, FIFO list.
2
3use crate::loom::sync::atomic::{AtomicPtr, AtomicUsize};
4use crate::loom::thread;
5use crate::sync::mpsc::block::{self, Block};
6
7use std::fmt;
8use std::ptr::NonNull;
9use std::sync::atomic::Ordering::{AcqRel, Acquire, Relaxed, Release};
10
11/// List queue transmit handle.
12pub(crate) struct Tx<T> {
13 /// Tail in the `Block` mpmc list.
14 block_tail: AtomicPtr<Block<T>>,
15
16 /// Position to push the next message. This references a block and offset
17 /// into the block.
18 tail_position: AtomicUsize,
19}
20
21/// List queue receive handle
22pub(crate) struct Rx<T> {
23 /// Pointer to the block being processed.
24 head: NonNull<Block<T>>,
25
26 /// Next slot index to process.
27 index: usize,
28
29 /// Pointer to the next block pending release.
30 free_head: NonNull<Block<T>>,
31}
32
33/// Return value of `Rx::try_pop`.
34pub(crate) enum TryPopResult<T> {
35 /// Successfully popped a value.
36 Ok(T),
37 /// The channel is empty.
38 Empty,
39 /// The channel is empty and closed.
40 Closed,
41 /// The channel is not empty, but the first value is being written.
42 Busy,
43}
44
45pub(crate) fn channel<T>() -> (Tx<T>, Rx<T>) {
46 // Create the initial block shared between the tx and rx halves.
47 let initial_block = Block::new(0);
48 let initial_block_ptr = Box::into_raw(initial_block);
49
50 let tx = Tx {
51 block_tail: AtomicPtr::new(initial_block_ptr),
52 tail_position: AtomicUsize::new(0),
53 };
54
55 let head = NonNull::new(initial_block_ptr).unwrap();
56
57 let rx = Rx {
58 head,
59 index: 0,
60 free_head: head,
61 };
62
63 (tx, rx)
64}
65
66impl<T> Tx<T> {
67 /// Pushes a value into the list.
68 pub(crate) fn push(&self, value: T) {
69 // First, claim a slot for the value. `Acquire` is used here to
70 // synchronize with the `fetch_add` in `reclaim_blocks`.
71 let slot_index = self.tail_position.fetch_add(1, Acquire);
72
73 // Load the current block and write the value
74 let block = self.find_block(slot_index);
75
76 unsafe {
77 // Write the value to the block
78 block.as_ref().write(slot_index, value);
79 }
80 }
81
82 /// Closes the send half of the list.
83 ///
84 /// Similar process as pushing a value, but instead of writing the value &
85 /// setting the ready flag, the `TX_CLOSED` flag is set on the block.
86 pub(crate) fn close(&self) {
87 // First, claim a slot for the value. This is the last slot that will be
88 // claimed.
89 let slot_index = self.tail_position.fetch_add(1, Acquire);
90
91 let block = self.find_block(slot_index);
92
93 unsafe { block.as_ref().tx_close() }
94 }
95
96 fn find_block(&self, slot_index: usize) -> NonNull<Block<T>> {
97 // The start index of the block that contains `index`.
98 let start_index = block::start_index(slot_index);
99
100 // The index offset into the block
101 let offset = block::offset(slot_index);
102
103 // Load the current head of the block
104 let mut block_ptr = self.block_tail.load(Acquire);
105
106 let block = unsafe { &*block_ptr };
107
108 // Calculate the distance between the tail ptr and the target block
109 let distance = block.distance(start_index);
110
111 // Decide if this call to `find_block` should attempt to update the
112 // `block_tail` pointer.
113 //
114 // Updating `block_tail` is not always performed in order to reduce
115 // contention.
116 //
117 // When set, as the routine walks the linked list, it attempts to update
118 // `block_tail`. If the update cannot be performed, `try_updating_tail`
119 // is unset.
120 let mut try_updating_tail = distance > offset;
121
122 // Walk the linked list of blocks until the block with `start_index` is
123 // found.
124 loop {
125 let block = unsafe { &(*block_ptr) };
126
127 if block.is_at_index(start_index) {
128 return unsafe { NonNull::new_unchecked(block_ptr) };
129 }
130
131 let next_block = block
132 .load_next(Acquire)
133 // There is no allocated next block, grow the linked list.
134 .unwrap_or_else(|| block.grow());
135
136 // If the block is **not** final, then the tail pointer cannot be
137 // advanced any more.
138 try_updating_tail &= block.is_final();
139
140 if try_updating_tail {
141 // Advancing `block_tail` must happen when walking the linked
142 // list. `block_tail` may not advance passed any blocks that are
143 // not "final". At the point a block is finalized, it is unknown
144 // if there are any prior blocks that are unfinalized, which
145 // makes it impossible to advance `block_tail`.
146 //
147 // While walking the linked list, `block_tail` can be advanced
148 // as long as finalized blocks are traversed.
149 //
150 // Release ordering is used to ensure that any subsequent reads
151 // are able to see the memory pointed to by `block_tail`.
152 //
153 // Acquire is not needed as any "actual" value is not accessed.
154 // At this point, the linked list is walked to acquire blocks.
155 if self
156 .block_tail
157 .compare_exchange(block_ptr, next_block.as_ptr(), Release, Relaxed)
158 .is_ok()
159 {
160 // Synchronize with any senders
161 let tail_position = self.tail_position.fetch_add(0, Release);
162
163 unsafe {
164 block.tx_release(tail_position);
165 }
166 } else {
167 // A concurrent sender is also working on advancing
168 // `block_tail` and this thread is falling behind.
169 //
170 // Stop trying to advance the tail pointer
171 try_updating_tail = false;
172 }
173 }
174
175 block_ptr = next_block.as_ptr();
176
177 thread::yield_now();
178 }
179 }
180
181 pub(crate) unsafe fn reclaim_block(&self, mut block: NonNull<Block<T>>) {
182 // The block has been removed from the linked list and ownership
183 // is reclaimed.
184 //
185 // Before dropping the block, see if it can be reused by
186 // inserting it back at the end of the linked list.
187 //
188 // First, reset the data
189 block.as_mut().reclaim();
190
191 let mut reused = false;
192
193 // Attempt to insert the block at the end
194 //
195 // Walk at most three times
196 //
197 let curr_ptr = self.block_tail.load(Acquire);
198
199 // The pointer can never be null
200 debug_assert!(!curr_ptr.is_null());
201
202 let mut curr = NonNull::new_unchecked(curr_ptr);
203
204 // TODO: Unify this logic with Block::grow
205 for _ in 0..3 {
206 match curr.as_ref().try_push(&mut block, AcqRel, Acquire) {
207 Ok(()) => {
208 reused = true;
209 break;
210 }
211 Err(next) => {
212 curr = next;
213 }
214 }
215 }
216
217 if !reused {
218 let _ = Box::from_raw(block.as_ptr());
219 }
220 }
221
222 pub(crate) fn is_closed(&self) -> bool {
223 let tail = self.block_tail.load(Acquire);
224
225 unsafe {
226 let tail_block = &*tail;
227 tail_block.is_closed()
228 }
229 }
230}
231
232impl<T> fmt::Debug for Tx<T> {
233 fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
234 fmt.debug_struct("Tx")
235 .field("block_tail", &self.block_tail.load(Relaxed))
236 .field("tail_position", &self.tail_position.load(Relaxed))
237 .finish()
238 }
239}
240
241impl<T> Rx<T> {
242 pub(crate) fn is_empty(&self, tx: &Tx<T>) -> bool {
243 let block = unsafe { self.head.as_ref() };
244 if block.has_value(self.index) {
245 return false;
246 }
247
248 // It is possible that a block has no value "now" but the list is still not empty.
249 // To be sure, it is necessary to check the length of the list.
250 self.len(tx) == 0
251 }
252
253 pub(crate) fn len(&self, tx: &Tx<T>) -> usize {
254 // When all the senders are dropped, there will be a last block in the tail position,
255 // but it will be closed
256 let tail_position = tx.tail_position.load(Acquire);
257 tail_position - self.index - (tx.is_closed() as usize)
258 }
259
260 /// Pops the next value off the queue.
261 pub(crate) fn pop(&mut self, tx: &Tx<T>) -> Option<block::Read<T>> {
262 // Advance `head`, if needed
263 if !self.try_advancing_head() {
264 return None;
265 }
266
267 self.reclaim_blocks(tx);
268
269 unsafe {
270 let block = self.head.as_ref();
271
272 let ret = block.read(self.index);
273
274 if let Some(block::Read::Value(..)) = ret {
275 self.index = self.index.wrapping_add(1);
276 }
277
278 ret
279 }
280 }
281
282 /// Pops the next value off the queue, detecting whether the block
283 /// is busy or empty on failure.
284 ///
285 /// This function exists because `Rx::pop` can return `None` even if the
286 /// channel's queue contains a message that has been completely written.
287 /// This can happen if the fully delivered message is behind another message
288 /// that is in the middle of being written to the block, since the channel
289 /// can't return the messages out of order.
290 pub(crate) fn try_pop(&mut self, tx: &Tx<T>) -> TryPopResult<T> {
291 let tail_position = tx.tail_position.load(Acquire);
292 let result = self.pop(tx);
293
294 match result {
295 Some(block::Read::Value(t)) => TryPopResult::Ok(t),
296 Some(block::Read::Closed) => TryPopResult::Closed,
297 None if tail_position == self.index => TryPopResult::Empty,
298 None => TryPopResult::Busy,
299 }
300 }
301
302 /// Tries advancing the block pointer to the block referenced by `self.index`.
303 ///
304 /// Returns `true` if successful, `false` if there is no next block to load.
305 fn try_advancing_head(&mut self) -> bool {
306 let block_index = block::start_index(self.index);
307
308 loop {
309 let next_block = {
310 let block = unsafe { self.head.as_ref() };
311
312 if block.is_at_index(block_index) {
313 return true;
314 }
315
316 block.load_next(Acquire)
317 };
318
319 let next_block = match next_block {
320 Some(next_block) => next_block,
321 None => {
322 return false;
323 }
324 };
325
326 self.head = next_block;
327
328 thread::yield_now();
329 }
330 }
331
332 fn reclaim_blocks(&mut self, tx: &Tx<T>) {
333 while self.free_head != self.head {
334 unsafe {
335 // Get a handle to the block that will be freed and update
336 // `free_head` to point to the next block.
337 let block = self.free_head;
338
339 let observed_tail_position = block.as_ref().observed_tail_position();
340
341 let required_index = match observed_tail_position {
342 Some(i) => i,
343 None => return,
344 };
345
346 if required_index > self.index {
347 return;
348 }
349
350 // We may read the next pointer with `Relaxed` ordering as it is
351 // guaranteed that the `reclaim_blocks` routine trails the `recv`
352 // routine. Any memory accessed by `reclaim_blocks` has already
353 // been acquired by `recv`.
354 let next_block = block.as_ref().load_next(Relaxed);
355
356 // Update the free list head
357 self.free_head = next_block.unwrap();
358
359 // Push the emptied block onto the back of the queue, making it
360 // available to senders.
361 tx.reclaim_block(block);
362 }
363
364 thread::yield_now();
365 }
366 }
367
368 /// Effectively `Drop` all the blocks. Should only be called once, when
369 /// the list is dropping.
370 pub(super) unsafe fn free_blocks(&mut self) {
371 debug_assert_ne!(self.free_head, NonNull::dangling());
372
373 let mut cur = Some(self.free_head);
374
375 #[cfg(debug_assertions)]
376 {
377 // to trigger the debug assert above so as to catch that we
378 // don't call `free_blocks` more than once.
379 self.free_head = NonNull::dangling();
380 self.head = NonNull::dangling();
381 }
382
383 while let Some(block) = cur {
384 cur = block.as_ref().load_next(Relaxed);
385 drop(Box::from_raw(block.as_ptr()));
386 }
387 }
388}
389
390impl<T> fmt::Debug for Rx<T> {
391 fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
392 fmt.debug_struct("Rx")
393 .field("head", &self.head)
394 .field("index", &self.index)
395 .field("free_head", &self.free_head)
396 .finish()
397 }
398}