From faf43225ff0d1a0091f5e489714de0c26d24d02f Mon Sep 17 00:00:00 2001
From: Peizhao Ou <peizhaoo@uci.edu>
Date: Tue, 10 Feb 2015 18:11:00 -0800
Subject: [PATCH] edits
---
benchmarks.mk | 2 +-
concurrent-hashmap/Makefile | 8 +-
concurrent-hashmap/hashmap.h | 6 +-
concurrent-hashmap/hashmap_wildcard.h | 307 ++++++++++++++++++++++++++
concurrent-hashmap/main_wildcard.cc | 58 +++++
mpmc-queue/note.txt | 20 ++
6 files changed, 398 insertions(+), 3 deletions(-)
create mode 100644 concurrent-hashmap/hashmap_wildcard.h
create mode 100644 concurrent-hashmap/main_wildcard.cc
create mode 100644 mpmc-queue/note.txt
diff --git a/benchmarks.mk b/benchmarks.mk
index 7f82f92..819248e 100644
--- a/benchmarks.mk
+++ b/benchmarks.mk
@@ -9,7 +9,7 @@ LIB_NAME = model
LIB_SO = lib$(LIB_NAME).so
BASE = ../..
-INCLUDE = -I$(BASE)/include -I../include
+INCLUDE = -I$(BASE)/include -I../include -I$(BASE)/scfence -I$(BASE)
# C preprocessor flags
CPPFLAGS += $(INCLUDE) -g
diff --git a/concurrent-hashmap/Makefile b/concurrent-hashmap/Makefile
index 4db5bca..d39e831 100644
--- a/concurrent-hashmap/Makefile
+++ b/concurrent-hashmap/Makefile
@@ -1,11 +1,17 @@
include ../benchmarks.mk
-TESTS := table
+TESTS := table table_wildcard1
all: $(TESTS)
table: main.cc hashmap.h
$(CXX) -o $@ $^ $(SPEC_OBJ) $(CXXFLAGS) -std=c++0x $(LDFLAGS)
+table_wildcard1: main_wildcard1.cc hashmap_wildcard.h
+ $(CXX) -o $@ $^ $(SPEC_OBJ) $(CXXFLAGS) -std=c++0x $(LDFLAGS)
+
+table_wildcard2: main_wildcard1.cc hashmap_wildcard.h
+ $(CXX) -o $@ $^ $(SPEC_OBJ) $(CXXFLAGS) -std=c++0x $(LDFLAGS)
+
clean:
rm -f *.o *.d $(TESTS)
diff --git a/concurrent-hashmap/hashmap.h b/concurrent-hashmap/hashmap.h
index a66005c..936f284 100644
--- a/concurrent-hashmap/hashmap.h
+++ b/concurrent-hashmap/hashmap.h
@@ -14,7 +14,7 @@
#include <stdlib.h>
#include <mutex>
-#include "cdsannotate.h"
+#include "sc_annotation.h"
#define relaxed memory_order_relaxed
#define release memory_order_release
@@ -164,7 +164,11 @@ class HashMap {
// lock, we ignore this operation for the SC analysis, and otherwise we
// take it into consideration
+ SC_BEGIN();
Entry *firstPtr = first->load(acquire);
+ SC_KEEP();
+ SC_END();
+
e = firstPtr;
while (e != NULL) {
if (e->hash == hash && eq(key, e->key)) {
diff --git a/concurrent-hashmap/hashmap_wildcard.h b/concurrent-hashmap/hashmap_wildcard.h
new file mode 100644
index 0000000..557a918
--- /dev/null
+++ b/concurrent-hashmap/hashmap_wildcard.h
@@ -0,0 +1,307 @@
+#ifndef _HASHMAP_H
+#define _HASHMAP_H
+
+#include <iostream>
+#include <atomic>
+#include "stdio.h"
+//#include <common.h>
+#ifdef STANDALONE
+#include <assert.h>
+#define MODEL_ASSERT assert
+#else
+#include <model-assert.h>
+#endif
+#include <stdlib.h>
+#include <mutex>
+
+#include "sc_annotation.h"
+#include "wildcard.h"
+
+#define relaxed memory_order_relaxed
+#define release memory_order_release
+#define acquire memory_order_acquire
+#define acq_rel memory_order_acq_rel
+#define seq_cst memory_order_seq_cst
+
+using namespace std;
+
+/**
+ For the sake of simplicity, we do not use template but some toy structs to
+ represent the Key and Value.
+*/
+struct Key {
+ // Probably represent the coordinate (x, y, z)
+ int x;
+ int y;
+ int z;
+
+ int hashCode() {
+ return x + 31 * y + 31 * 31 * z;
+ }
+
+ bool equals(Key *other) {
+ if (!other)
+ return false;
+ return x == other->x && y == other->y && z == other->z;
+ }
+
+ Key(int x, int y, int z) :
+ x(x),
+ y(y),
+ z(z)
+ {
+
+ }
+};
+
+struct Value {
+ // Probably represent the speed vector (vX, vY, vZ)
+ int vX;
+ int vY;
+ int vZ;
+
+ Value(int vX, int vY, int vZ) :
+ vX(vX),
+ vY(vY),
+ vZ(vZ)
+ {
+
+ }
+
+ bool equals(Value *other) {
+ if (!other)
+ return false;
+ return vX == other->vX && vY == other->vY && vZ == other->vZ;
+ }
+};
+
+class Entry {
+ public:
+ Key *key;
+ atomic<Value*> value;
+ int hash;
+ atomic<Entry*> next;
+
+ Entry(int h, Key *k, Value *v, Entry *n) {
+ this->hash = h;
+ this->key = k;
+ this->value.store(v, wildcard(1)); // relaxed
+ this->next.store(n, wildcard(2)); // relaxed
+ }
+};
+
+class Segment {
+ public:
+ int count;
+ mutex segMutex;
+
+ void lock() {
+ segMutex.lock();
+ }
+
+ void unlock() {
+ segMutex.unlock();
+ }
+
+ Segment() {
+ this->count = 0;
+ }
+};
+
+class HashMap {
+ public:
+ atomic<Entry*> *table;
+
+ int capacity;
+ int size;
+
+ static const int CONCURRENCY_LEVEL = 16;
+
+ static const int SEGMENT_MASK = CONCURRENCY_LEVEL - 1;
+
+ Segment *segments[CONCURRENCY_LEVEL];
+
+ static const int DEFAULT_INITIAL_CAPACITY = 16;
+
+ // Not gonna consider resize now...
+
+ HashMap() {
+ this->size = 0;
+ this->capacity = DEFAULT_INITIAL_CAPACITY;
+ this->table = new atomic<Entry*>[capacity];
+ for (int i = 0; i < capacity; i++) {
+ atomic_init(&table[i], NULL);
+ }
+ for (int i = 0; i < CONCURRENCY_LEVEL; i++) {
+ segments[i] = new Segment;
+ }
+ }
+
+ int hashKey(Key *x) {
+ int h = x->hashCode();
+ // Use logical right shift
+ unsigned int tmp = (unsigned int) h;
+ return ((h << 7) - h + (tmp >> 9) + (tmp >> 17));
+ }
+
+ bool eq(Key *x, Key *y) {
+ return x == y || x->equals(y);
+ }
+
+ Value* get(Key *key) {
+ MODEL_ASSERT (key);
+ int hash = hashKey(key);
+
+ // Try first without locking...
+ atomic<Entry*> *tab = table;
+ int index = hash & (capacity - 1);
+ atomic<Entry*> *first = &tab[index];
+ Entry *e;
+ Value *res = NULL;
+
+ // Should be a load acquire
+ // This load action here makes it problematic for the SC analysis, what
+ // we need to do is as follows: if the get() method ever acquires the
+ // lock, we ignore this operation for the SC analysis, and otherwise we
+ // take it into consideration
+
+ SC_BEGIN();
+ Entry *firstPtr = first->load(wildcard(3)); // acquire
+ SC_KEEP();
+ SC_END();
+
+ e = firstPtr;
+ while (e != NULL) {
+ if (e->hash == hash && eq(key, e->key)) {
+ res = e->value.load(wildcard(4)); // seq_cst
+ if (res != NULL)
+ return res;
+ else
+ break;
+ // Loading the next entry
+ e = e->next.load(wildcard(5)); // acquire
+ }
+ }
+
+ // Recheck under synch if key apparently not there or interference
+ Segment *seg = segments[hash & SEGMENT_MASK];
+ seg->lock(); // Critical region begins
+ // Not considering resize now, so ignore the reload of table...
+
+ // Synchronized by locking, no need to be load acquire
+ Entry *newFirstPtr = first->load(wildcard(6)); // relaxed
+ if (e != NULL || firstPtr != newFirstPtr) {
+ e = newFirstPtr;
+ while (e != NULL) {
+ if (e->hash == hash && eq(key, e->key)) {
+ res = e->value.load(wildcard(7)); // seq_cst
+ seg->unlock(); // Critical region ends
+ return res;
+ }
+ // Synchronized by locking
+ e = e->next.load(wildcard(8)); // relaxed
+ }
+ }
+ seg->unlock(); // Critical region ends
+ return NULL;
+ }
+
+ Value* put(Key *key, Value *value) {
+ // Don't allow NULL key or value
+ MODEL_ASSERT (key && value);
+
+ int hash = hashKey(key);
+ Segment *seg = segments[hash & SEGMENT_MASK];
+ atomic<Entry*> *tab;
+
+ seg->lock(); // Critical region begins
+ tab = table;
+ int index = hash & (capacity - 1);
+
+ atomic<Entry*> *first = &tab[index];
+ Entry *e;
+ Value *oldValue = NULL;
+
+ // The written of the entry is synchronized by locking
+ Entry *firstPtr = first->load(wildcard(9)); // relaxed
+ e = firstPtr;
+ while (e != NULL) {
+ if (e->hash == hash && eq(key, e->key)) {
+ // FIXME: This could be a relaxed (because locking synchronize
+ // with the previous put())??
+ oldValue = e->value.load(wildcard(10)); // acquire
+ e->value.store(value, wildcard(11)); // seq_cst
+ seg->unlock(); // Don't forget to unlock before return
+ return oldValue;
+ }
+ // Synchronized by locking
+ e = e->next.load(wildcard(12)); // relaxed
+ }
+
+ // Add to front of list
+ Entry *newEntry = new Entry(hash, key, value, firstPtr);
+ // Publish the newEntry to others
+ first->store(newEntry, wildcard(13)); // release
+ seg->unlock(); // Critical region ends
+ return NULL;
+ }
+
+ Value* remove(Key *key, Value *value) {
+ MODEL_ASSERT (key);
+ int hash = hashKey(key);
+ Segment *seg = segments[hash & SEGMENT_MASK];
+ atomic<Entry*> *tab;
+
+ seg->lock(); // Critical region begins
+ tab = table;
+ int index = hash & (capacity - 1);
+
+ atomic<Entry*> *first = &tab[index];
+ Entry *e;
+ Value *oldValue = NULL;
+
+ // The written of the entry is synchronized by locking
+ Entry *firstPtr = first->load(relaxed);
+ e = firstPtr;
+
+ while (true) {
+ if (e != NULL) {
+ seg->unlock(); // Don't forget to unlock
+ return NULL;
+ }
+ if (e->hash == hash && eq(key, e->key))
+ break;
+ // Synchronized by locking
+ e = e->next.load(relaxed);
+ }
+
+ // FIXME: This could be a relaxed (because locking synchronize
+ // with the previous put())??
+ oldValue = e->value.load(acquire);
+ // If the value parameter is NULL, we will remove the entry anyway
+ if (value != NULL && value->equals(oldValue)) {
+ seg->unlock();
+ return NULL;
+ }
+
+ // Force the get() to grab the lock and retry
+ e->value.store(NULL, relaxed);
+
+ // The strategy to remove the entry is to keep the entries after the
+ // removed one and copy the ones before it
+ Entry *head = e->next.load(relaxed);
+ Entry *p;
+ p = first->load(relaxed);
+ while (p != e) {
+ head = new Entry(p->hash, p->key, p->value.load(relaxed), head);
+ p = p->next.load(relaxed);
+ }
+
+ // Publish the new head to readers
+ first->store(head, release);
+ seg->unlock(); // Critical region ends
+ return oldValue;
+ }
+};
+
+#endif
diff --git a/concurrent-hashmap/main_wildcard.cc b/concurrent-hashmap/main_wildcard.cc
new file mode 100644
index 0000000..4a6f904
--- /dev/null
+++ b/concurrent-hashmap/main_wildcard.cc
@@ -0,0 +1,58 @@
+#include <iostream>
+#include <threads.h>
+#include "hashmap_wildcard.h"
+
+HashMap *table;
+
+Key *k1, *k2;
+Value *r1, *r2, *r3, *r4, *v1, *v2;
+
+void printKey(Key *key) {
+ if (key)
+ printf("pos = (%d, %d, %d)\n", key->x, key->y, key->z);
+ else
+ printf("pos = NULL\n");
+}
+
+void printValue(Value *value) {
+ if (value)
+ printf("velocity = (%d, %d, %d)\n", value->vX, value->vY, value->vZ);
+ else
+ printf("velocity = NULL\n");
+}
+
+void threadA(void *arg) {
+ k1 = new Key(1, 1, 1);
+ k2 = new Key(3, 4, 5);
+ v1 = new Value(10, 10, 10);
+ r1 = table->put(k1, v1);
+ //printValue(r1);
+ r2 = table->get(k2);
+ printf("Thrd A:\n");
+ printValue(r2);
+}
+
+void threadB(void *arg) {
+ k1 = new Key(1, 1, 1);
+ k2 = new Key(3, 4, 5);
+ v2 = new Value(30, 40, 50);
+ r3 = table->put(k2, v2);
+ //printValue(r3);
+ r4 = table->get(k1);
+ printf("Thrd B:\n");
+ printValue(r4);
+}
+
+int user_main(int argc, char *argv[]) {
+ thrd_t t1, t2;
+ table = new HashMap;
+
+ thrd_create(&t1, threadA, NULL);
+ thrd_create(&t2, threadB, NULL);
+ thrd_join(t1);
+ thrd_join(t2);
+
+ return 0;
+}
+
+
diff --git a/mpmc-queue/note.txt b/mpmc-queue/note.txt
new file mode 100644
index 0000000..e8597d3
--- /dev/null
+++ b/mpmc-queue/note.txt
@@ -0,0 +1,20 @@
+#1 Why this data structure is non-SC (but it's SC at abstraction).
+
+The underlying pattern of the non-SCness are as the following:
+
+T1 T2
+x = 1; y = 1;
+r1 = y; r2 = x;
+
+Suppose there are two identical threads, both having a writer followed by a
+reader, as follows:
+
+T1 T2
+m_written.CAS; m_rdwr.CAS;
+
+m_rdwr.CAS; m_written.CAS; // In write_publish()
+
+m_written.load; // a m_rdwr.load; // The first in the read_fetch() // b
+
+a & b can both read old value and it's non-SC. However, the later loads and CAS
+in the loops will finally fix this problem.
--
2.34.1