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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168
// Copyright © 2019-2020 VMware, Inc. All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0 OR MIT
use alloc::alloc::{alloc, dealloc, Layout};
use core::cell::Cell;
use core::default::Default;
use core::fmt;
use core::mem::{align_of, size_of};
use core::ops::{Drop, FnMut};
use core::slice::from_raw_parts_mut;
#[cfg(not(loom))]
use core::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
#[cfg(loom)]
pub use loom::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
use crossbeam_utils::CachePadded;
use crate::context::MAX_PENDING_OPS;
use crate::replica::MAX_THREADS_PER_REPLICA;
/// The default size of the shared log in bytes. If constructed using the
/// default constructor, the log will be these many bytes in size. Currently
/// set to 32 MiB based on the ASPLOS 2017 paper.
const DEFAULT_LOG_BYTES: usize = 32 * 1024 * 1024;
const_assert!(DEFAULT_LOG_BYTES >= 1 && (DEFAULT_LOG_BYTES & (DEFAULT_LOG_BYTES - 1) == 0));
/// The maximum number of replicas that can be registered with the log.
#[cfg(not(loom))]
pub const MAX_REPLICAS_PER_LOG: usize = 192;
#[cfg(loom)] // Otherwise uses too much stack space wich crashes in loom...
pub const MAX_REPLICAS_PER_LOG: usize = 3;
/// Constant required for garbage collection. When the tail and the head are
/// these many entries apart on the circular buffer, garbage collection will
/// be performed by one of the replicas registered with the log.
///
/// For the GC algorithm to work, we need to ensure that we can support the
/// largest possible append after deciding to perform GC. This largest possible
/// append is when every thread within a replica has a full batch of writes
/// to be appended to the shared log.
const GC_FROM_HEAD: usize = MAX_PENDING_OPS * MAX_THREADS_PER_REPLICA;
const_assert!(GC_FROM_HEAD >= 1 && (GC_FROM_HEAD & (GC_FROM_HEAD - 1) == 0));
/// Threshold after how many iterations we log a warning for busy spinning loops.
///
/// This helps with debugging to figure out where things may end up blocking.
/// Should be a power of two to avoid divisions.
const WARN_THRESHOLD: usize = 1 << 28;
/// An entry that sits on the log. Each entry consists of three fields: The operation to
/// be performed when a thread reaches this entry on the log, the replica that appended
/// this operation, and a flag indicating whether this entry is valid.
///
/// `T` is the type on the operation - typically an enum class containing opcodes as well as
/// arguments. It is required that this type be sized and cloneable.
#[derive(Default)]
#[repr(align(64))]
struct Entry<T>
where
T: Sized + Clone,
{
/// The operation that this entry represents.
operation: Option<T>,
/// Identifies the replica that issued the above operation.
replica: usize,
/// Indicates whether this entry represents a valid operation when on the log.
alivef: AtomicBool,
}
/// A log of operations that is typically accessed by multiple
/// [Replica](struct.Replica.html).
///
/// Operations can be added to the log by calling the `append()` method and
/// providing a list of operations to be performed.
///
/// Operations already on the log can be executed by calling the `exec()` method
/// and providing a replica-id along with a closure. Newly added operations
/// since the replica last called `exec()` will be executed by invoking the
/// supplied closure for each one of them.
///
/// Accepts one generic type parameter; `T` defines the type of operations and
/// their arguments that will go on the log and would typically be an enum
/// class.
///
/// This struct is aligned to 64 bytes optimizing cache access.\
///
/// # Note
/// As a client, typically there is no need to call any methods on the Log aside
/// from `new`. Only in the rare circumstance someone would implement their own
/// Replica would it be necessary to call any of the Log's methods.
#[repr(align(64))]
pub struct Log<'a, T>
where
T: Sized + Clone,
{
/// Raw pointer to the actual underlying log. Required for dealloc.
rawp: *mut u8,
/// Size of the underlying log in bytes. Required for dealloc.
rawb: usize,
/// The maximum number of entries that can be held inside the log.
size: usize,
/// A reference to the actual log. Nothing but a slice of entries.
slog: &'a [Cell<Entry<T>>],
/// Logical index into the above slice at which the log starts.
head: CachePadded<AtomicUsize>,
/// Logical index into the above slice at which the log ends.
/// New appends go here.
tail: CachePadded<AtomicUsize>,
/// Completed tail maintains an index <= tail that points to a
/// log entry after which there are no completed operations across
/// all replicas registered against this log.
ctail: CachePadded<AtomicUsize>,
/// Array consisting of the local tail of each replica registered with the log.
/// Required for garbage collection; since replicas make progress over the log
/// independently, we want to make sure that we don't garbage collect operations
/// that haven't been executed by all replicas.
ltails: [CachePadded<AtomicUsize>; MAX_REPLICAS_PER_LOG],
/// Identifier that will be allocated to the next replica that registers with
/// this Log. Also required to correctly index into ltails above.
next: CachePadded<AtomicUsize>,
/// Array consisting of local alive masks for each registered replica. Required
/// because replicas make independent progress over the log, so we need to
/// track log wrap-arounds for each of them separately.
lmasks: [CachePadded<Cell<bool>>; MAX_REPLICAS_PER_LOG],
}
impl<'a, T> fmt::Debug for Log<'a, T>
where
T: Sized + Clone,
{
fn fmt(&self, fmt: &mut fmt::Formatter) -> fmt::Result {
fmt.debug_struct("Log")
.field("head", &self.tail)
.field("tail", &self.head)
.field("size", &self.size)
.finish()
}
}
/// The Log is Send. The *mut u8 (`rawp`) is never dereferenced.
unsafe impl<'a, T> Send for Log<'a, T> where T: Sized + Clone {}
/// The Log is Sync. We know this because: `head` and `tail` are atomic variables, `append()`
/// reserves entries using a CAS, and exec() does not concurrently mutate entries on the log.
unsafe impl<'a, T> Sync for Log<'a, T> where T: Sized + Clone {}
impl<'a, T> Log<'a, T>
where
T: Sized + Clone,
{
/// Constructs and returns a log of size `bytes` bytes.
/// A size between 1-2 MiB usually works well in most cases.
///
/// # Example
///
/// ```
/// use node_replication::Log;
///
/// // Operation type that will go onto the log.
/// #[derive(Clone)]
/// enum Operation {
/// Read,
/// Write(u64),
/// Invalid,
/// }
///
/// // Creates a 1 Mega Byte sized log.
/// let l = Log::<Operation>::new(1 * 1024 * 1024);
/// ```
///
/// This method also allocates memory for the log upfront. No further allocations
/// will be performed once this method returns.
pub fn new<'b>(bytes: usize) -> Log<'b, T> {
// Calculate the number of entries that will go into the log, and retrieve a
// slice to it from the allocated region of memory.
let mut num = bytes / Log::<T>::entry_size();
// Make sure the log is large enough to allow for periodic garbage collection.
if num < 2 * GC_FROM_HEAD {
num = 2 * GC_FROM_HEAD;
}
// Round off to the next power of two if required. If we overflow, then set
// the number of entries to the minimum required for GC. This is unlikely since
// we'd need a log size > 2^63 entries for this to happen.
if !num.is_power_of_two() {
num = num.checked_next_power_of_two().unwrap_or(2 * GC_FROM_HEAD)
};
// Now that we have the actual number of entries, allocate the log.
let b = num * Log::<T>::entry_size();
let mem = unsafe {
alloc(
Layout::from_size_align(b, align_of::<Cell<Entry<T>>>())
.expect("Alignment error while allocating the shared log!"),
)
};
if mem.is_null() {
panic!("Failed to allocate memory for the shared log!");
}
let raw = unsafe { from_raw_parts_mut(mem as *mut Cell<Entry<T>>, num) };
// Initialize all log entries by calling the default constructor.
for e in raw.iter_mut() {
unsafe {
::core::ptr::write(
e,
Cell::new(Entry {
operation: None,
replica: 0usize,
alivef: AtomicBool::new(false),
}),
);
}
}
#[allow(clippy::declare_interior_mutable_const)]
const LMASK_DEFAULT: CachePadded<Cell<bool>> = CachePadded::new(Cell::new(true));
#[cfg(not(loom))]
{
#[allow(clippy::declare_interior_mutable_const)]
const LTAIL_DEFAULT: CachePadded<AtomicUsize> = CachePadded::new(AtomicUsize::new(0));
Log {
rawp: mem,
rawb: b,
size: num,
slog: raw,
head: CachePadded::new(AtomicUsize::new(0usize)),
tail: CachePadded::new(AtomicUsize::new(0usize)),
ctail: CachePadded::new(AtomicUsize::new(0usize)),
ltails: [LTAIL_DEFAULT; MAX_REPLICAS_PER_LOG],
next: CachePadded::new(AtomicUsize::new(1usize)),
lmasks: [LMASK_DEFAULT; MAX_REPLICAS_PER_LOG],
}
}
// AtomicUsize::new is not const in loom. This code block (including arr
// dependency) becomes redundant once
// https://github.com/tokio-rs/loom/issues/170 is fixed:
#[cfg(loom)]
{
use arr_macro::arr;
Log {
rawp: mem,
rawb: b,
size: num,
slog: raw,
head: CachePadded::new(AtomicUsize::new(0usize)),
tail: CachePadded::new(AtomicUsize::new(0usize)),
ctail: CachePadded::new(AtomicUsize::new(0usize)),
ltails: arr![CachePadded::new(AtomicUsize::new(0)); 3], // MAX_REPLICAS_PER_LOG
next: CachePadded::new(AtomicUsize::new(1usize)),
lmasks: [LMASK_DEFAULT; MAX_REPLICAS_PER_LOG],
}
}
}
/// Returns the size of an entry in bytes.
fn entry_size() -> usize {
size_of::<Cell<Entry<T>>>()
}
/// Registers a replica with the log. Returns an identifier that the replica
/// can use to execute operations on the log.
///
/// # Example
///
/// ```ignore
/// use node_replication::Log;
///
/// // Operation type that will go onto the log.
/// #[derive(Clone)]
/// enum Operation {
/// Read,
/// Write(u64),
/// Invalid,
/// }
///
/// // Creates a 1 Mega Byte sized log.
/// let l = Log::<Operation>::new(1 * 1024 * 1024);
///
/// // Registers against the log. `idx` can now be used to append operations
/// // to the log, and execute these operations.
/// let idx = l.register().expect("Failed to register with the Log.");
/// ```
pub(crate) fn register(&self) -> Option<usize> {
// Loop until we either run out of identifiers or we manage to increment `next`.
loop {
let n = self.next.load(Ordering::Relaxed);
// Check if we've exceeded the maximum number of replicas the log can support.
if n >= MAX_REPLICAS_PER_LOG {
return None;
};
if self
.next
.compare_exchange_weak(n, n + 1, Ordering::SeqCst, Ordering::SeqCst)
!= Ok(n)
{
continue;
};
return Some(n);
}
}
/// Adds a batch of operations to the shared log.
///
/// # Example
///
/// ```ignore
/// use node_replication::Log;
///
/// // Operation type that will go onto the log.
/// #[derive(Clone)]
/// enum Operation {
/// Read,
/// Write(u64),
/// }
///
/// let l = Log::<Operation>::new(1 * 1024 * 1024);
/// let idx = l.register().expect("Failed to register with the Log.");
///
/// // The set of operations we would like to append. The order will
/// // be preserved by the interface.
/// let ops = [Operation::Write(100), Operation::Read];
///
/// // `append()` might have to garbage collect the log. When doing so,
/// // it might encounter operations added in by another replica/thread.
/// // This closure allows us to consume those operations. `id` identifies
/// // the replica that added in those operations.
/// let f = |op: Operation, id: usize| {
/// match(op) {
/// Operation::Read => println!("Read by {}", id),
/// Operation::Write(x) => println!("Write({}) by {}", x, id),
/// }
/// };
///
/// // Append the operations. These operations will be marked with `idx`,
/// // and will be linearized at the tail of the log.
/// l.append(&ops, idx, f);
/// ```
///
/// If there isn't enough space to perform the append, this method busy
/// waits until the head is advanced. Accepts a replica `idx`; all appended
/// operations/entries will be marked with this replica-identifier. Also
/// accepts a closure `s`; when waiting for GC, this closure is passed into
/// exec() to ensure that this replica does'nt cause a deadlock.
///
/// # Note
/// Documentation for this function is hidden since `append` is currently not
/// intended as a public interface. It is marked as public due to being
/// used by the benchmarking code.
#[inline(always)]
#[doc(hidden)]
pub fn append<F: FnMut(T, usize)>(&self, ops: &[T], idx: usize, mut s: F) {
let nops = ops.len();
let mut iteration = 1;
let mut waitgc = 1;
// Keep trying to reserve entries and add operations to the log until
// we succeed in doing so.
loop {
if iteration % WARN_THRESHOLD == 0 {
warn!(
"append(ops.len()={}, {}) takes too many iterations ({}) to complete...",
ops.len(),
idx,
iteration,
);
}
iteration += 1;
let tail = self.tail.load(Ordering::Relaxed);
let head = self.head.load(Ordering::Relaxed);
// If there are fewer than `GC_FROM_HEAD` entries on the log, then just
// try again. The replica that reserved entry (h + self.size - GC_FROM_HEAD)
// is currently trying to advance the head of the log. Keep refreshing the
// replica against the log to make sure that it isn't deadlocking GC.
if tail > head + self.size - GC_FROM_HEAD {
if waitgc % WARN_THRESHOLD == 0 {
warn!(
"append(ops.len()={}, {}) takes too many iterations ({}) waiting for gc...",
ops.len(),
idx,
waitgc,
);
}
waitgc += 1;
self.exec(idx, &mut s);
#[cfg(loom)]
loom::thread::yield_now();
continue;
}
// If on adding in the above entries there would be fewer than `GC_FROM_HEAD`
// entries left on the log, then we need to advance the head of the log.
let mut advance = false;
if tail + nops > head + self.size - GC_FROM_HEAD {
advance = true
};
// Try reserving slots for the operations. If that fails, then restart
// from the beginning of this loop.
if self.tail.compare_exchange_weak(
tail,
tail + nops,
Ordering::Acquire,
Ordering::Acquire,
) != Ok(tail)
{
continue;
};
// Successfully reserved entries on the shared log. Add the operations in.
for (i, op) in ops.iter().enumerate().take(nops) {
let e = self.slog[self.index(tail + i)].as_ptr();
let mut m = self.lmasks[idx - 1].get();
// This entry was just reserved so it should be dead (!= m). However, if
// the log has wrapped around, then the alive mask has flipped. In this
// case, we flip the mask we were originally going to write into the
// allocated entry. We cannot flip lmasks[idx - 1] because this replica
// might still need to execute a few entries before the wrap around.
if unsafe { (*e).alivef.load(Ordering::Relaxed) == m } {
m = !m;
}
unsafe { (*e).operation = Some(op.clone()) };
unsafe { (*e).replica = idx };
unsafe { (*e).alivef.store(m, Ordering::Release) };
}
// If needed, advance the head of the log forward to make room on the log.
if advance {
self.advance_head(idx, &mut s);
}
return;
}
}
/// Executes a passed in closure (`d`) on all operations starting from
/// a replica's local tail on the shared log. The replica is identified through an
/// `idx` passed in as an argument.
///
/// # Example
///
/// ```ignore
/// use node_replication::Log;
///
/// // Operation type that will go onto the log.
/// #[derive(Clone)]
/// enum Operation {
/// Read,
/// Write(u64),
/// }
///
/// let l = Log::<Operation>::new(1 * 1024 * 1024);
/// let idx = l.register().expect("Failed to register with the Log.");
/// let ops = [Operation::Write(100), Operation::Read];
///
/// let f = |op: Operation, id: usize| {
/// match(op) {
/// Operation::Read => println!("Read by {}", id),
/// Operation::Write(x) => println!("Write({}) by {}", x, id),
/// }
/// };
/// l.append(&ops, idx, f);
///
/// // This closure is executed on every operation appended to the
/// // since the last call to `exec()` by this replica/thread.
/// let mut d = 0;
/// let mut g = |op: Operation, id: usize| {
/// match(op) {
/// // The write happened before the read.
/// Operation::Read => assert_eq!(100, d),
/// Operation::Write(x) => d += 100,
/// }
/// };
/// l.exec(idx, &mut g);
/// ```
///
/// The passed in closure is expected to take in two arguments: The operation
/// from the shared log to be executed and the replica that issued it.
#[inline(always)]
pub(crate) fn exec<F: FnMut(T, usize)>(&self, idx: usize, d: &mut F) {
// Load the logical log offset from which we must execute operations.
let ltail = self.ltails[idx - 1].load(Ordering::Relaxed);
// Check if we have any work to do by comparing our local tail with the log's
// global tail. If they're equal, then we're done here and can simply return.
let gtail = self.tail.load(Ordering::Relaxed);
if ltail == gtail {
return;
}
let h = self.head.load(Ordering::Relaxed);
// Make sure we're within the shared log. If we aren't, then panic.
if ltail > gtail || ltail < h {
panic!("Local tail not within the shared log!")
};
// Execute all operations from the passed in offset to the shared log's tail. Check if
// the entry is live first; we could have a replica that has reserved entries, but not
// filled them into the log yet.
for i in ltail..gtail {
let mut iteration = 1;
let e = self.slog[self.index(i)].as_ptr();
while unsafe { (*e).alivef.load(Ordering::Acquire) != self.lmasks[idx - 1].get() } {
if iteration % WARN_THRESHOLD == 0 {
warn!(
"alivef not being set for self.index(i={}) = {} (self.lmasks[{}] is {})...",
i,
self.index(i),
idx - 1,
self.lmasks[idx - 1].get()
);
}
iteration += 1;
#[cfg(loom)]
loom::thread::yield_now();
}
unsafe { d((*e).operation.as_ref().unwrap().clone(), (*e).replica) };
// Looks like we're going to wrap around now; flip this replica's local mask.
if self.index(i) == self.size - 1 {
self.lmasks[idx - 1].set(!self.lmasks[idx - 1].get());
//trace!("idx: {} lmask: {}", idx, self.lmasks[idx - 1].get());
}
}
// Update the completed tail after we've executed these operations.
// Also update this replica's local tail.
self.ctail.fetch_max(gtail, Ordering::Relaxed);
self.ltails[idx - 1].store(gtail, Ordering::Relaxed);
}
/// Returns a physical index given a logical index into the shared log.
#[inline(always)]
fn index(&self, logical: usize) -> usize {
logical & (self.size - 1)
}
/// Advances the head of the log forward. If a replica has stopped making progress,
/// then this method will never return. Accepts a closure that is passed into exec()
/// to ensure that this replica does not deadlock GC.
#[inline(always)]
fn advance_head<F: FnMut(T, usize)>(&self, rid: usize, mut s: &mut F) {
// Keep looping until we can advance the head and create some free space
// on the log. If one of the replicas has stopped making progress, then
// this method might never return.
let mut iteration = 1;
loop {
let r = self.next.load(Ordering::Relaxed);
let global_head = self.head.load(Ordering::Relaxed);
let f = self.tail.load(Ordering::Relaxed);
let mut min_local_tail = self.ltails[0].load(Ordering::Relaxed);
// Find the smallest local tail across all replicas.
for idx in 1..r {
let cur_local_tail = self.ltails[idx - 1].load(Ordering::Relaxed);
if min_local_tail > cur_local_tail {
min_local_tail = cur_local_tail
};
}
// If we cannot advance the head further, then start
// from the beginning of this loop again. Before doing so, try consuming
// any new entries on the log to prevent deadlock.
if min_local_tail == global_head {
if iteration % WARN_THRESHOLD == 0 {
warn!("Spending a long time in `advance_head`, are we starving?");
}
iteration += 1;
self.exec(rid, &mut s);
#[cfg(loom)]
loom::thread::yield_now();
continue;
}
// There are entries that can be freed up; update the head offset.
self.head.store(min_local_tail, Ordering::Relaxed);
// Make sure that we freed up enough space so that threads waiting for
// GC in append can make progress. Otherwise, try to make progress again.
// If we're making progress again, then try consuming entries on the log.
if f < min_local_tail + self.size - GC_FROM_HEAD {
return;
} else {
self.exec(rid, &mut s);
}
}
}
/// Resets the log. Required for microbenchmarking the log; with this method, we
/// can re-use the log across experimental runs without having to re-allocate the
/// log over and over again.
///
/// # Safety
///
/// *To be used for testing/benchmarking only, hence marked unsafe*. Before calling
/// this method, please make sure that there aren't any replicas/threads actively
/// issuing/executing operations to/from this log.
#[doc(hidden)]
#[inline(always)]
pub unsafe fn reset(&self) {
// First, reset global metadata.
self.head.store(0, Ordering::SeqCst);
self.tail.store(0, Ordering::SeqCst);
self.next.store(1, Ordering::SeqCst);
// Next, reset replica-local metadata.
for r in 0..MAX_REPLICAS_PER_LOG {
self.ltails[r].store(0, Ordering::Relaxed);
self.lmasks[r].set(true);
}
// Next, free up all log entries. Use pointers to avoid memcpy and speed up
// the reset of the log here.
for i in 0..self.size {
let e = self.slog[self.index(i)].as_ptr();
(*e).alivef.store(false, Ordering::Release);
}
}
/// This method checks if the replica is in sync to execute a read-only operation
/// right away. It does so by comparing the replica's local tail with the log's
/// completed tail.
///
/// # Example
///
/// ```ignore
/// use node_replication::Log;
///
/// // Operation type that will go onto the log.
/// #[derive(Clone)]
/// enum Operation {
/// Read,
/// Write(u64),
/// }
///
/// // We register two replicas here, `idx1` and `idx2`.
/// let l = Log::<Operation>::new(1 * 1024 * 1024);
/// let idx1 = l.register().expect("Failed to register with the Log.");
/// let idx2 = l.register().expect("Failed to register with the Log.");
/// let ops = [Operation::Write(100), Operation::Read];
///
/// let f = |op: Operation, id: usize| {
/// match(op) {
/// Operation::Read => println!("Read by {}", id),
/// Operation::Write(x) => println!("Write({}) by {}", x, id),
/// }
/// };
/// l.append(&ops, idx2, f);
///
/// let mut d = 0;
/// let mut g = |op: Operation, id: usize| {
/// match(op) {
/// // The write happened before the read.
/// Operation::Read => assert_eq!(100, d),
/// Operation::Write(x) => d += 100,
/// }
/// };
/// l.exec(idx2, &mut g);
///
/// // This assertion fails because `idx1` has not executed operations
/// // that were appended by `idx2`.
/// assert_eq!(false, l.is_replica_synced_for_reads(idx1, l.get_ctail()));
///
/// let mut e = 0;
/// let mut g = |op: Operation, id: usize| {
/// match(op) {
/// // The write happened before the read.
/// Operation::Read => assert_eq!(100, e),
/// Operation::Write(x) => e += 100,
/// }
/// };
/// l.exec(idx1, &mut g);
///
/// // `idx1` is all synced up, so this assertion passes.
/// assert_eq!(true, l.is_replica_synced_for_reads(idx1, l.get_ctail()));
/// ```
#[inline(always)]
pub(crate) fn is_replica_synced_for_reads(&self, idx: usize, ctail: usize) -> bool {
self.ltails[idx - 1].load(Ordering::Relaxed) >= ctail
}
/// This method returns the current ctail value for the log.
#[inline(always)]
pub(crate) fn get_ctail(&self) -> usize {
self.ctail.load(Ordering::Relaxed)
}
}
impl<'a, T> Default for Log<'a, T>
where
T: Sized + Clone,
{
/// Default constructor for the shared log.
fn default() -> Self {
Log::new(DEFAULT_LOG_BYTES)
}
}
impl<'a, T> Drop for Log<'a, T>
where
T: Sized + Clone,
{
/// Destructor for the shared log.
fn drop(&mut self) {
unsafe {
dealloc(
self.rawp,
Layout::from_size_align(self.rawb, align_of::<Cell<Entry<T>>>())
.expect("Alignment error while deallocating the shared log!"),
)
};
}
}
#[cfg(test)]
mod tests {
// Import std so that we have an allocator for our unit tests.
extern crate std;
use super::*;
use std::sync::Arc;
// Define operations along with their arguments that go onto the log.
#[derive(Clone)] // Traits required by the log interface.
#[derive(Debug, PartialEq)] // Traits required for testing.
enum Operation {
Read,
Write(u64),
Invalid,
}
// Required so that we can unit test Entry.
impl Default for Operation {
fn default() -> Operation {
Operation::Invalid
}
}
// Test that we can default construct entries correctly.
#[test]
fn test_entry_create_default() {
let e = Entry::<Operation>::default();
assert_eq!(e.operation, None);
assert_eq!(e.replica, 0);
assert_eq!(e.alivef.load(Ordering::Relaxed), false);
}
// Test that our entry_size() method returns the correct size.
#[test]
fn test_log_entry_size() {
assert_eq!(Log::<Operation>::entry_size(), 64);
}
// Tests if a small log can be correctly constructed.
#[test]
fn test_log_create() {
let l = Log::<Operation>::new(1024 * 1024);
let n = (1024 * 1024) / Log::<Operation>::entry_size();
assert_eq!(l.rawb, 1024 * 1024);
assert_eq!(l.size, n);
assert_eq!(l.slog.len(), n);
assert_eq!(l.head.load(Ordering::Relaxed), 0);
assert_eq!(l.tail.load(Ordering::Relaxed), 0);
assert_eq!(l.next.load(Ordering::Relaxed), 1);
assert_eq!(l.ctail.load(Ordering::Relaxed), 0);
for i in 0..MAX_REPLICAS_PER_LOG {
assert_eq!(l.ltails[i].load(Ordering::Relaxed), 0);
}
for i in 0..MAX_REPLICAS_PER_LOG {
assert_eq!(l.lmasks[i].get(), true);
}
}
// Tests if the constructor allocates enough space for GC.
#[test]
fn test_log_min_size() {
let l = Log::<Operation>::new(1024);
assert_eq!(l.rawb, 2 * GC_FROM_HEAD * Log::<Operation>::entry_size());
assert_eq!(l.size, 2 * GC_FROM_HEAD);
assert_eq!(l.slog.len(), 2 * GC_FROM_HEAD);
}
// Tests that the constructor allocates a log whose number of entries
// are a power of two.
#[test]
fn test_log_power_of_two() {
let l = Log::<Operation>::new(524 * 1024);
let n = ((524 * 1024) / Log::<Operation>::entry_size()).checked_next_power_of_two();
assert_eq!(l.rawb, n.unwrap() * Log::<Operation>::entry_size());
assert_eq!(l.size, n.unwrap());
assert_eq!(l.slog.len(), n.unwrap());
}
// Tests if the log can be successfully default constructed.
#[test]
fn test_log_create_default() {
let l = Log::<Operation>::default();
let n = DEFAULT_LOG_BYTES / Log::<Operation>::entry_size();
assert_eq!(l.rawb, DEFAULT_LOG_BYTES);
assert_eq!(l.size, n);
assert_eq!(l.slog.len(), n);
assert_eq!(l.head.load(Ordering::Relaxed), 0);
assert_eq!(l.tail.load(Ordering::Relaxed), 0);
assert_eq!(l.next.load(Ordering::Relaxed), 1);
assert_eq!(l.ctail.load(Ordering::Relaxed), 0);
for i in 0..MAX_REPLICAS_PER_LOG {
assert_eq!(l.ltails[i].load(Ordering::Relaxed), 0);
}
for i in 0..MAX_REPLICAS_PER_LOG {
assert_eq!(l.lmasks[i].get(), true);
}
}
// Tests if we can correctly index into the shared log.
#[test]
fn test_log_index() {
let l = Log::<Operation>::new(2 * 1024 * 1024);
assert_eq!(l.index(99000), 696);
}
// Tests if we can correctly register with the shared log.
#[test]
fn test_log_register() {
let l = Log::<Operation>::new(1024);
assert_eq!(l.register(), Some(1));
assert_eq!(l.next.load(Ordering::Relaxed), 2);
}
// Tests that we cannot register more than the max replicas with the log.
#[test]
fn test_log_register_none() {
let l = Log::<Operation>::new(1024);
l.next.store(MAX_REPLICAS_PER_LOG, Ordering::Relaxed);
assert!(l.register().is_none());
assert_eq!(l.next.load(Ordering::Relaxed), MAX_REPLICAS_PER_LOG);
}
// Test that we can correctly append an entry into the log.
#[test]
fn test_log_append() {
let l = Log::<Operation>::default();
let o = [Operation::Read];
l.append(&o, 1, |_o: Operation, _i: usize| {});
assert_eq!(l.head.load(Ordering::Relaxed), 0);
assert_eq!(l.tail.load(Ordering::Relaxed), 1);
let slog = l.slog[0].take();
assert_eq!(slog.operation, Some(Operation::Read));
assert_eq!(slog.replica, 1);
}
// Test that multiple entries can be appended to the log.
#[test]
fn test_log_append_multiple() {
let l = Log::<Operation>::default();
let o = [Operation::Read, Operation::Write(119)];
l.append(&o, 1, |_o: Operation, _i: usize| {});
assert_eq!(l.head.load(Ordering::Relaxed), 0);
assert_eq!(l.tail.load(Ordering::Relaxed), 2);
}
// Tests that we can advance the head of the log to the smallest of all replica-local tails.
#[test]
fn test_log_advance_head() {
let l = Log::<Operation>::default();
l.next.store(5, Ordering::Relaxed);
l.ltails[0].store(1023, Ordering::Relaxed);
l.ltails[1].store(224, Ordering::Relaxed);
l.ltails[2].store(4096, Ordering::Relaxed);
l.ltails[3].store(799, Ordering::Relaxed);
l.advance_head(0, &mut |_o: Operation, _i: usize| {});
assert_eq!(l.head.load(Ordering::Relaxed), 224);
}
// Tests that the head of the log is advanced when we're close to filling up the entire log.
#[test]
fn test_log_append_gc() {
let l = Log::<Operation>::default();
let o: [Operation; 4] = unsafe {
let mut a: [Operation; 4] = ::std::mem::MaybeUninit::zeroed().assume_init();
for i in &mut a[..] {
::std::ptr::write(i, Operation::Read);
}
a
};
l.next.store(2, Ordering::Relaxed);
l.tail.store(l.size - GC_FROM_HEAD - 1, Ordering::Relaxed);
l.ltails[0].store(1024, Ordering::Relaxed);
l.append(&o, 1, |_o: Operation, _i: usize| {});
assert_eq!(l.head.load(Ordering::Relaxed), 1024);
assert_eq!(l.tail.load(Ordering::Relaxed), l.size - GC_FROM_HEAD + 3);
}
// Tests that on log wrap around, the local mask stays
// the same because entries have not been executed yet.
#[test]
fn test_log_append_wrap() {
let l = Log::<Operation>::default();
let o: [Operation; 1024] = unsafe {
let mut a: [Operation; 1024] = ::std::mem::MaybeUninit::zeroed().assume_init();
for i in &mut a[..] {
::std::ptr::write(i, Operation::Read);
}
a
};
l.next.store(2, Ordering::Relaxed);
l.head.store(2 * 8192, Ordering::Relaxed);
l.tail.store(l.size - 10, Ordering::Relaxed);
l.append(&o, 1, |_o: Operation, _i: usize| {});
assert_eq!(l.lmasks[0].get(), true);
assert_eq!(l.tail.load(Ordering::Relaxed), l.size + 1014);
}
// Test that we can execute operations appended to the log.
#[test]
fn test_log_exec() {
let l = Log::<Operation>::default();
let o = [Operation::Read];
let mut f = |op: Operation, i: usize| {
assert_eq!(op, Operation::Read);
assert_eq!(i, 1);
};
l.append(&o, 1, |_o: Operation, _i: usize| {});
l.exec(1, &mut f);
assert_eq!(
l.tail.load(Ordering::Relaxed),
l.ctail.load(Ordering::Relaxed)
);
assert_eq!(
l.tail.load(Ordering::Relaxed),
l.ltails[0].load(Ordering::Relaxed)
);
}
// Test that exec() doesn't do anything when the log is empty.
#[test]
fn test_log_exec_empty() {
let l = Log::<Operation>::default();
let mut f = |_o: Operation, _i: usize| {
assert!(false);
};
l.exec(1, &mut f);
}
// Test that exec() doesn't do anything if we're already up-to-date.
#[test]
fn test_log_exec_zero() {
let l = Log::<Operation>::default();
let o = [Operation::Read];
let mut f = |op: Operation, i: usize| {
assert_eq!(op, Operation::Read);
assert_eq!(i, 1);
};
let mut g = |_op: Operation, _i: usize| {
assert!(false);
};
l.append(&o, 1, |_o: Operation, _i: usize| {});
l.exec(1, &mut f);
l.exec(1, &mut g);
}
// Test that multiple entries on the log can be executed correctly.
#[test]
fn test_log_exec_multiple() {
let l = Log::<Operation>::default();
let o = [Operation::Read, Operation::Write(119)];
let mut s = 0;
let mut f = |op: Operation, _i: usize| match op {
Operation::Read => s += 121,
Operation::Write(v) => s += v,
Operation::Invalid => assert!(false),
};
l.append(&o, 1, |_o: Operation, _i: usize| {});
l.exec(1, &mut f);
assert_eq!(s, 240);
assert_eq!(
l.tail.load(Ordering::Relaxed),
l.ctail.load(Ordering::Relaxed)
);
assert_eq!(
l.tail.load(Ordering::Relaxed),
l.ltails[0].load(Ordering::Relaxed)
);
}
// Test that the replica local mask is updated correctly when executing over
// a wrapped around log.
#[test]
fn test_log_exec_wrap() {
let l = Log::<Operation>::default();
let o: [Operation; 1024] = unsafe {
let mut a: [Operation; 1024] = ::std::mem::MaybeUninit::zeroed().assume_init();
for i in &mut a[..] {
::std::ptr::write(i, Operation::Read);
}
a
};
let mut f = |op: Operation, i: usize| {
assert_eq!(op, Operation::Read);
assert_eq!(i, 1);
};
l.append(&o, 1, |_o: Operation, _i: usize| {}); // Required for GC to work correctly.
l.next.store(2, Ordering::SeqCst);
l.head.store(2 * 8192, Ordering::SeqCst);
l.tail.store(l.size - 10, Ordering::SeqCst);
l.append(&o, 1, |_o: Operation, _i: usize| {});
l.ltails[0].store(l.size - 10, Ordering::SeqCst);
l.exec(1, &mut f);
assert_eq!(l.lmasks[0].get(), false);
assert_eq!(l.tail.load(Ordering::Relaxed), l.size + 1014);
}
// Tests that exec() panics if the head of the log advances beyond the tail.
#[test]
#[should_panic]
fn test_exec_panic() {
let l = Log::<Operation>::default();
let o: [Operation; 1024] = unsafe {
let mut a: [Operation; 1024] = ::std::mem::MaybeUninit::zeroed().assume_init();
for i in &mut a[..] {
::std::ptr::write(i, Operation::Read);
}
a
};
let mut f = |_op: Operation, _i: usize| {
assert!(false);
};
l.append(&o, 1, |_o: Operation, _i: usize| {});
l.head.store(8192, Ordering::SeqCst);
l.exec(1, &mut f);
}
// Tests that operations are cloned when added to the log, and that
// they are correctly dropped once overwritten.
#[test]
fn test_log_change_refcount() {
let l = Log::<Arc<Operation>>::default();
let o1 = [Arc::new(Operation::Read)];
let o2 = [Arc::new(Operation::Read)];
assert_eq!(Arc::strong_count(&o1[0]), 1);
assert_eq!(Arc::strong_count(&o2[0]), 1);
l.append(&o1[..], 1, |_o: Arc<Operation>, _i: usize| {});
assert_eq!(Arc::strong_count(&o1[0]), 2);
l.append(&o1[..], 1, |_o: Arc<Operation>, _i: usize| {});
assert_eq!(Arc::strong_count(&o1[0]), 3);
unsafe { l.reset() };
// Over here, we overwrite entries that were written to by the two
// previous appends. This decreases the refcount of o1 and increases
// the refcount of o2.
l.append(&o2[..], 1, |_o: Arc<Operation>, _i: usize| {});
assert_eq!(Arc::strong_count(&o1[0]), 2);
assert_eq!(Arc::strong_count(&o2[0]), 2);
l.append(&o2[..], 1, |_o: Arc<Operation>, _i: usize| {});
assert_eq!(Arc::strong_count(&o1[0]), 1);
assert_eq!(Arc::strong_count(&o2[0]), 3);
}
// Tests that operations are cloned when added to the log, and that
// they are correctly dropped once overwritten after the GC.
#[test]
fn test_log_refcount_change_with_gc() {
let entry_size = 64;
let total_entries = 16384;
assert_eq!(Log::<Arc<Operation>>::entry_size(), entry_size);
let size: usize = total_entries * entry_size;
let l = Log::<Arc<Operation>>::new(size);
let o1 = [Arc::new(Operation::Read)];
let o2 = [Arc::new(Operation::Read)];
assert_eq!(Arc::strong_count(&o1[0]), 1);
assert_eq!(Arc::strong_count(&o2[0]), 1);
for i in 1..(total_entries + 1) {
l.append(&o1[..], 1, |_o: Arc<Operation>, _i: usize| {});
assert_eq!(Arc::strong_count(&o1[0]), i + 1);
}
assert_eq!(Arc::strong_count(&o1[0]), total_entries + 1);
for i in 1..(total_entries + 1) {
l.append(&o2[..], 1, |_o: Arc<Operation>, _i: usize| {});
assert_eq!(Arc::strong_count(&o1[0]), (total_entries + 1) - i);
assert_eq!(Arc::strong_count(&o2[0]), i + 1);
}
assert_eq!(Arc::strong_count(&o1[0]), 1);
assert_eq!(Arc::strong_count(&o2[0]), total_entries + 1);
}
// Tests that is_replica_synced_for_read() works correctly; it returns
// false when a replica is not synced up and true when it is.
#[test]
fn test_replica_synced_for_read() {
let l = Log::<Operation>::default();
let one = l.register().unwrap();
let two = l.register().unwrap();
assert_eq!(one, 1);
assert_eq!(two, 2);
let o = [Operation::Read];
let mut f = |op: Operation, i: usize| {
assert_eq!(op, Operation::Read);
assert_eq!(i, 1);
};
l.append(&o, one, |_o: Operation, _i: usize| {});
l.exec(one, &mut f);
assert_eq!(l.is_replica_synced_for_reads(one, l.get_ctail()), true);
assert_eq!(l.is_replica_synced_for_reads(two, l.get_ctail()), false);
l.exec(two, &mut f);
assert_eq!(l.is_replica_synced_for_reads(two, l.get_ctail()), true);
}
}