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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
//! Zero-copy allocator for intra-process serialized communication.
use std::rc::Rc;
use std::cell::RefCell;
use std::collections::{VecDeque, HashMap, hash_map::Entry};
use crossbeam_channel::{Sender, Receiver};
use timely_bytes::arc::Bytes;
use crate::networking::MessageHeader;
use crate::{Allocate, Push, Pull};
use crate::allocator::{AllocateBuilder, Exchangeable};
use crate::allocator::canary::Canary;
use super::bytes_exchange::{BytesPull, SendEndpoint, MergeQueue};
use super::push_pull::{Pusher, Puller};
/// Builds an instance of a ProcessAllocator.
///
/// Builders are required because some of the state in a `ProcessAllocator` cannot be sent between
/// threads (specifically, the `Rc<RefCell<_>>` local channels). So, we must package up the state
/// shared between threads here, and then provide a method that will instantiate the non-movable
/// members once in the destination thread.
pub struct ProcessBuilder {
index: usize, // number out of peers
peers: usize, // number of peer allocators.
pushers: Vec<Receiver<MergeQueue>>, // for pushing bytes at other workers.
pullers: Vec<Sender<MergeQueue>>, // for pulling bytes from other workers.
}
impl ProcessBuilder {
/// Creates a vector of builders, sharing appropriate state.
///
/// This method requires access to a byte exchanger, from which it mints channels.
pub fn new_vector(count: usize) -> Vec<ProcessBuilder> {
// Channels for the exchange of `MergeQueue` endpoints.
let (pullers_vec, pushers_vec) = crate::promise_futures(count, count);
pushers_vec
.into_iter()
.zip(pullers_vec)
.enumerate()
.map(|(index, (pushers, pullers))|
ProcessBuilder {
index,
peers: count,
pushers,
pullers,
}
)
.collect()
}
/// Builds a `ProcessAllocator`, instantiating `Rc<RefCell<_>>` elements.
pub fn build(self) -> ProcessAllocator {
// Fulfill puller obligations.
let mut recvs = Vec::with_capacity(self.peers);
for puller in self.pullers.into_iter() {
let buzzer = crate::buzzer::Buzzer::default();
let queue = MergeQueue::new(buzzer);
puller.send(queue.clone()).expect("Failed to send MergeQueue");
recvs.push(queue.clone());
}
// Extract pusher commitments.
let mut sends = Vec::with_capacity(self.peers);
for pusher in self.pushers.into_iter() {
let queue = pusher.recv().expect("Failed to receive MergeQueue");
let sendpoint = SendEndpoint::new(queue);
sends.push(Rc::new(RefCell::new(sendpoint)));
}
ProcessAllocator {
index: self.index,
peers: self.peers,
events: Rc::new(RefCell::new(Default::default())),
canaries: Rc::new(RefCell::new(Vec::new())),
channel_id_bound: None,
staged: Vec::new(),
sends,
recvs,
to_local: HashMap::new(),
}
}
}
impl AllocateBuilder for ProcessBuilder {
type Allocator = ProcessAllocator;
/// Builds allocator, consumes self.
fn build(self) -> Self::Allocator {
self.build()
}
}
/// A serializing allocator for inter-thread intra-process communication.
pub struct ProcessAllocator {
index: usize, // number out of peers
peers: usize, // number of peer allocators (for typed channel allocation).
events: Rc<RefCell<Vec<usize>>>,
canaries: Rc<RefCell<Vec<usize>>>,
channel_id_bound: Option<usize>,
// sending, receiving, and responding to binary buffers.
staged: Vec<Bytes>,
sends: Vec<Rc<RefCell<SendEndpoint<MergeQueue>>>>, // sends[x] -> goes to thread x.
recvs: Vec<MergeQueue>, // recvs[x] <- from thread x.
to_local: HashMap<usize, Rc<RefCell<VecDeque<Bytes>>>>, // to worker-local typed pullers.
}
impl Allocate for ProcessAllocator {
fn index(&self) -> usize { self.index }
fn peers(&self) -> usize { self.peers }
fn allocate<T: Exchangeable>(&mut self, identifier: usize) -> (Vec<Box<dyn Push<T>>>, Box<dyn Pull<T>>) {
// Assume and enforce in-order identifier allocation.
if let Some(bound) = self.channel_id_bound {
assert!(bound < identifier);
}
self.channel_id_bound = Some(identifier);
let mut pushes = Vec::<Box<dyn Push<T>>>::with_capacity(self.peers());
for target_index in 0 .. self.peers() {
// message header template.
let header = MessageHeader {
channel: identifier,
source: self.index,
target: target_index,
length: 0,
seqno: 0,
};
// create, box, and stash new process_binary pusher.
pushes.push(Box::new(Pusher::new(header, self.sends[target_index].clone())));
}
let channel =
self.to_local
.entry(identifier)
.or_insert_with(|| Rc::new(RefCell::new(VecDeque::new())))
.clone();
use crate::allocator::counters::Puller as CountPuller;
let canary = Canary::new(identifier, self.canaries.clone());
let puller = Box::new(CountPuller::new(Puller::new(channel, canary), identifier, self.events().clone()));
(pushes, puller)
}
// Perform preparatory work, most likely reading binary buffers from self.recv.
#[inline(never)]
fn receive(&mut self) {
// Check for channels whose `Puller` has been dropped.
let mut canaries = self.canaries.borrow_mut();
for dropped_channel in canaries.drain(..) {
let _dropped =
self.to_local
.remove(&dropped_channel)
.expect("non-existent channel dropped");
// Borrowed channels may be non-empty, if the dataflow was forcibly
// dropped. The contract is that if a dataflow is dropped, all other
// workers will drop the dataflow too, without blocking indefinitely
// on events from it.
// assert!(dropped.borrow().is_empty());
}
std::mem::drop(canaries);
let mut events = self.events.borrow_mut();
for recv in self.recvs.iter_mut() {
recv.drain_into(&mut self.staged);
}
for mut bytes in self.staged.drain(..) {
// We expect that `bytes` contains an integral number of messages.
// No splitting occurs across allocations.
while bytes.len() > 0 {
if let Some(header) = MessageHeader::try_read(&mut bytes[..]) {
// Get the header and payload, ditch the header.
let mut peel = bytes.extract_to(header.required_bytes());
let _ = peel.extract_to(40);
// Increment message count for channel.
// Safe to do this even if the channel has been dropped.
events.push(header.channel);
// Ensure that a queue exists.
match self.to_local.entry(header.channel) {
Entry::Vacant(entry) => {
// We may receive data before allocating, and shouldn't block.
if self.channel_id_bound.map(|b| b < header.channel).unwrap_or(true) {
entry.insert(Rc::new(RefCell::new(VecDeque::new())))
.borrow_mut()
.push_back(peel);
}
}
Entry::Occupied(mut entry) => {
entry.get_mut().borrow_mut().push_back(peel);
}
}
}
else {
println!("failed to read full header!");
}
}
}
}
// Perform postparatory work, most likely sending un-full binary buffers.
fn release(&mut self) {
// Publish outgoing byte ledgers.
for send in self.sends.iter_mut() {
send.borrow_mut().publish();
}
// OPTIONAL: Tattle on channels sitting on borrowed data.
// OPTIONAL: Perhaps copy borrowed data into owned allocation.
// for (index, list) in self.to_local.iter() {
// let len = list.borrow_mut().len();
// if len > 0 {
// eprintln!("Warning: worker {}, undrained channel[{}].len() = {}", self.index, index, len);
// }
// }
}
fn events(&self) -> &Rc<RefCell<Vec<usize>>> {
&self.events
}
fn await_events(&self, duration: Option<std::time::Duration>) {
if self.events.borrow().is_empty() {
if let Some(duration) = duration {
std::thread::park_timeout(duration);
}
else {
std::thread::park();
}
}
}
}