Merge pull request #23 from mtak-/starvation

starvation freedom - and use criterion for rbtree benchmarks

Author: mtak-

build.rs

@@ -0,0 +1,9 @@
+macro_rules! stats {
+    ($($(#[$attr:meta])* $names:ident: $kinds:tt @ $env_var:ident),* $(,)*) => {
+        concat!($("cargo:rerun-if-env-changed=", "SWYM_", stringify!($env_var),"\n"),*)
+    };
+}
+
+fn main() {
+    println!(include!("./src/stats_list.rs"));
+}

ci/rbtree.sh

@@ -30,43 +30,11 @@ cargo check --features debug-alloc,nightly,stats,$RTM --benches --bins --example
 # run tests
 ./x.py test
 RUST_TEST_THREADS=1 cargo test --features stats,nightly,$RTM --lib --tests
-RUST_TEST_THREADS=1 RUSTFLAGS="${RUSTFLAGS} ${ASAN_FLAG}" \
-    time cargo test \
-        --features debug-alloc,stats,$RTM
 
-# benchmarks
-./x.py bench insert:: --features nightly,$RTM 
+# TODO: address sanitizer doesn't work with criterion?
+# RUSTFLAGS="${RUSTFLAGS} ${ASAN_FLAG}"
+RUST_TEST_THREADS=1 \
+    time cargo test --features debug-alloc,stats,$RTM
 
-# these benchmarks are run one at a time due to high memory usage
-./x.py bench --features nightly,$RTM rbtree::contains_key_01
-./x.py bench --features nightly,$RTM rbtree::contains_key_02
-./x.py bench --features nightly,$RTM rbtree::contains_key_03
-./x.py bench --features nightly,$RTM rbtree::contains_key_04
-./x.py bench --features nightly,$RTM rbtree::contains_key_05
-./x.py bench --features nightly,$RTM rbtree::contains_key_06
-./x.py bench --features nightly,$RTM rbtree::contains_key_07
-./x.py bench --features nightly,$RTM rbtree::contains_key_08
-./x.py bench --features nightly,$RTM rbtree::entry_01
-./x.py bench --features nightly,$RTM rbtree::entry_02
-./x.py bench --features nightly,$RTM rbtree::entry_03
-./x.py bench --features nightly,$RTM rbtree::entry_04
-./x.py bench --features nightly,$RTM rbtree::entry_05
-./x.py bench --features nightly,$RTM rbtree::entry_06
-./x.py bench --features nightly,$RTM rbtree::entry_07
-./x.py bench --features nightly,$RTM rbtree::entry_08
-./x.py bench --features nightly,$RTM rbtree::get_01
-./x.py bench --features nightly,$RTM rbtree::get_02
-./x.py bench --features nightly,$RTM rbtree::get_03
-./x.py bench --features nightly,$RTM rbtree::get_04
-./x.py bench --features nightly,$RTM rbtree::get_05
-./x.py bench --features nightly,$RTM rbtree::get_06
-./x.py bench --features nightly,$RTM rbtree::get_07
-./x.py bench --features nightly,$RTM rbtree::get_08
-./x.py bench --features nightly,$RTM rbtree::insert_01
-./x.py bench --features nightly,$RTM rbtree::insert_02
-./x.py bench --features nightly,$RTM rbtree::insert_03
-./x.py bench --features nightly,$RTM rbtree::insert_04
-./x.py bench --features nightly,$RTM rbtree::insert_05
-./x.py bench --features nightly,$RTM rbtree::insert_06
-./x.py bench --features nightly,$RTM rbtree::insert_07
-./x.py bench --features nightly,$RTM rbtree::insert_08
+# benchmarks
+./x.py bench --features nightly,$RTM

src/internal.rs

@@ -11,6 +11,7 @@ pub mod bloom;
 mod commit;
 mod gc;
 mod parking;
+mod starvation;
 
 pub mod epoch;
 pub mod read_log;

src/internal/alloc/fvec.rs

@@ -1,6 +1,6 @@
 use core::ops::{Deref, DerefMut};
 
-const START_SIZE: usize = 1024;
+const START_SIZE: usize = 0;
 
 #[derive(Debug)]
 pub struct FVec<T> {

src/internal/commit.rs

@@ -18,7 +18,6 @@ const MAX_HTX_RETRIES: u8 = 3;
 impl<'tcell> Logs<'tcell> {
     #[inline]
     pub unsafe fn remove_writes_from_reads(&mut self) {
-        #[allow(unused_mut)]
         let mut count = 0;
 
         let write_log = &mut self.write_log;
@@ -107,7 +106,8 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
 
     #[inline]
     unsafe fn commit_empty_write_log(self) -> bool {
-        let (_, logs) = self.into_inner();
+        let (_, logs, progress) = self.into_inner();
+        progress.progressed();
         // RwTx validates reads as they occur. As a result, if there are no writes, then we have
         // no work to do in our commit algorithm.
         //
@@ -136,6 +136,7 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
     #[inline]
     fn commit_slow(self) -> bool {
         let mut retry_count = 0;
+        self.progress().wait_for_starvers();
         match self.start_htx(&mut retry_count) {
             Ok(htx) => {
                 let success = self.commit_hard(htx);
@@ -147,7 +148,7 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
                 self.commit_soft()
             }
             Err(BoundedHtxErr::AbortOrConflict) => {
-                stats::htm_conflicts(retry_count as _);
+                stats::htm_abort(retry_count as _);
                 false
             }
         }
@@ -156,7 +157,7 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
     #[inline(never)]
     fn commit_hard(self, htx: HardwareTx) -> bool {
         unsafe {
-            let (synch, logs) = self.into_inner();
+            let (synch, logs, progress) = self.into_inner();
             let current = synch.current_epoch();
             logs.read_log.validate_reads_htm(current, &htx);
             let park_status = logs.write_log.write_and_lock_htm(&htx, current);
@@ -168,6 +169,8 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
 
             logs.read_log.clear();
             logs.write_log.clear_no_drop();
+
+            progress.progressed();
             if unlikely!(park_status == ParkStatus::HasParked) {
                 crate::internal::parking::unpark();
             }
@@ -234,7 +237,7 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
 
     #[inline]
     unsafe fn validation_success(self, park_status: ParkStatus) -> bool {
-        let (synch, logs) = self.into_inner();
+        let (synch, logs, progress) = self.into_inner();
 
         // The writes must be performed before the EPOCH_CLOCK is tick'ed.
         // Reads can get away with performing less work with this ordering.
@@ -251,6 +254,7 @@ impl<'tx, 'tcell> PinRw<'tx, 'tcell> {
 
         logs.read_log.clear();
         logs.write_log.clear_no_drop();
+        progress.progressed();
         if unlikely!(park_status == ParkStatus::HasParked) {
             crate::internal::parking::unpark();
         }

src/internal/epoch.rs

@@ -52,7 +52,7 @@ const FIRST: Storage = TICK_SIZE + UNPARK_BIT;
 /// The smallest difference between points on the EpochClock.
 ///
 /// Two is used because the first bit of EpochLock is reserved as the UNPARK_BIT
-const TICK_SIZE: Storage = 1 << 1;
+pub const TICK_SIZE: Storage = 1 << 1;
 
 /// The least significant bit is set when _no_ threads are parked waiting for modifications to an
 /// EpochLock.
@@ -123,6 +123,11 @@ impl QuiesceEpoch {
         NonZeroStorage::new(epoch).map(QuiesceEpoch)
     }
 
+    #[inline]
+    pub fn get(self) -> NonZeroStorage {
+        self.0
+    }
+
     /// Returns the maximum value that a QuiesceEpoch can hold. This is useful for finding the
     /// minimum of a set of epochs.
     #[inline]
@@ -487,6 +492,7 @@ impl ThreadEpoch {
 
 /// A monotonically increasing clock.
 #[derive(Debug)]
+#[repr(align(64))]
 pub struct EpochClock(HtmStorage);
 
 /// The world clock. The source of truth, and synchronization for swym. Every write transaction

src/internal/optim.rs

@@ -34,6 +34,7 @@ pub fn _likely(b: bool) -> bool {
     b
 }
 
+#[cold]
 pub fn _abort() -> ! {
     std::process::abort();
 }

src/internal/parking.rs

@@ -5,7 +5,6 @@ use crate::{
     },
     stats,
 };
-use crossbeam_utils::Backoff;
 use parking_lot_core::{FilterOp, ParkResult, ParkToken, DEFAULT_UNPARK_TOKEN};
 use swym_htm::{BoundedHtxErr, HardwareTx};
 
@@ -26,7 +25,7 @@ fn parkable<'tx, 'tcell>(pin: PinMutRef<'tx, 'tcell>) -> bool {
 
 #[inline(never)]
 #[cold]
-pub fn park<'tx, 'tcell>(mut pin: PinRw<'tx, 'tcell>, backoff: &Backoff) {
+pub fn park<'tx, 'tcell>(mut pin: PinRw<'tx, 'tcell>) {
     debug_assert!(
         parkable(pin.reborrow()),
         "`AWAIT_RETRY` on a transaction that has an empty read set causes the thread to sleep \
@@ -50,12 +49,10 @@ pub fn park<'tx, 'tcell>(mut pin: PinRw<'tx, 'tcell>, backoff: &Backoff) {
             debug_assert_eq!(token, DEFAULT_UNPARK_TOKEN);
             let parked_size = logs.read_log.len() + logs.write_log.epoch_locks().count();
             stats::parked_size(parked_size);
-            backoff.reset()
         }
         ParkResult::Invalid => {
             let parked_size = logs.read_log.len() + logs.write_log.epoch_locks().count();
             stats::park_failure_size(parked_size);
-            backoff.snooze()
         }
         ParkResult::TimedOut => {
             if cfg!(debug_assertions) {

src/internal/read_log.rs

@@ -13,7 +13,7 @@ use crate::{
 };
 use swym_htm::HardwareTx;
 
-const READ_CAPACITY: usize = 1024;
+const READ_CAPACITY: usize = 0;
 
 #[derive(Debug)]
 pub struct ReadLog<'tcell> {