Synchronization

François Trahay

Introduction

Objectives of this lecture:
- How are synchronization primitives implemented?
- How to do without locks?

Atomic operations

Motivation

By default, an instruction modifying a variable is non-atomic
example : x++ gives :
- register = load(x)
- register ++
- x = store (register)

→ Problem if the variable is modified by a other thread simultaneously

Can’t we just use `volatile` ?

Tells the compiler that the variable can change from one access to another:
- modification by another thread
- modification by a signal handler
But volatile does not ensure atomicity

Here is an example of a program that may suffer from overly aggressive optimization by the compiler:

#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>


#if USE_VOLATILE
volatile int a = 0;
#else
int a = 0;
#endif

void* thread1(void*arg) {
  while(a == 0) ;
  printf("Hello\n");
  return NULL;
}

void* thread2(void*arg) {
  a = 1;
  return NULL;
}

int main(int argc, char**argv) {
  pthread_t t1, t2;
  pthread_create(&t1, NULL, thread1, NULL);
  pthread_create(&t2, NULL, thread2, NULL);

  pthread_join(t1, NULL);
  pthread_join(t2, NULL);
  return EXIT_SUCCESS;
}

When compiled with the optimization level -O0 (i.e. without any optimization), thread1 spins waiting, and when thread2 modifies the variable a, it unlocks thread1 which displays Hello:

$ gcc -o volatile volatile.c -Wall -pthread -O0
$ ./volatile 
Hello
$

When compiled with the optimization level -O1, the generated code no longer works:

$ gcc -o volatile volatile.c -Wall -pthread -O1
$ ./volatile 
[waits indefinitely]
^C
$

Analyzing the code generated by the compiler reveals the problem:

$ gcc -o volatile volatile.c -Wall -pthread -O2
$ gdb ./volatile 
[...]
(gdb) disassemble thread1
Dump of assembler code for function thread1:
   0x0000000000000756 <+0>:  auipc  a5,0x2
   0x000000000000075a <+4>:  lw a5,-1778(a5) # 0x2064 <a>
   0x000000000000075e <+8>:  bnez   a5,0x762 <thread1+12>
   0x0000000000000760 <+10>:    j   0x760 <thread1+10>
   0x0000000000000762 <+12>:    add sp,sp,-16
   0x0000000000000764 <+14>:    auipc   a0,0x0
   0x0000000000000768 <+18>:    add a0,a0,36 # 0x788
   0x000000000000076c <+22>:    sd  ra,8(sp)
   0x000000000000076e <+24>:    jal 0x620 <puts@plt>
   0x0000000000000772 <+28>:    ld  ra,8(sp)
   0x0000000000000774 <+30>:    li  a0,0
   0x0000000000000776 <+32>:    add sp,sp,16
   0x0000000000000778 <+34>:    ret
nd of assembler dump.
$

We see here that at the address 0x760, the program jumps to the address 0x760. So it jumps in place indefinitely.

This is explained by the fact that the variable a is not volatile. The compiler therefore thinks it can optimize access to this variable: since the thread1 function only accesses the variable in read-mode, the program loads the variable in a register (here, the a5 register, see the instruction 0x75a), then consults the registry. When thread2 modifies the variable a, the modification is therefore not perceived by thread1!

Declaring the variable as volatile forces the compiler to read the variable each time:

$ gcc -o volatile volatile.c -Wall -pthread -O2 -DUSE_VOLATILE=1
$ gdb volatile
(gdb) disassemble thread1 
Dump of assembler code for function thread1:
   0x0000000000000756 <+0>:  add    sp,sp,-16
   0x0000000000000758 <+2>:  sd ra,8(sp)
   0x000000000000075a <+4>:  auipc  a4,0x2
   0x000000000000075e <+8>:  add    a4,a4,-1782 # 0x2064 <a>
   0x0000000000000762 <+12>:    lw  a5,0(a4)
   0x0000000000000764 <+14>:    beqz    a5,0x762 <thread1+12>
   0x0000000000000766 <+16>:    auipc   a0,0x0
   0x000000000000076a <+20>:    add a0,a0,34 # 0x788
   0x000000000000076e <+24>:    jal 0x620 <puts@plt>
   0x0000000000000772 <+28>:    ld  ra,8(sp)
   0x0000000000000774 <+30>:    li  a0,0
   0x0000000000000776 <+32>:    add sp,sp,16
   0x0000000000000778 <+34>:    ret
End of assembler dump.

Here, the loop while (a == 0) is translated to the lines from 0x762 to 0x764. At each loop iteration, the value of a is loaded, then tested.

Atomic operations

C11 provides a set of atomic operations, including
- atomic_flag_test_and_set
- atomic_compare_exchange_strong
- atomic_fetch_add
- atomic_thread_fence

Test and set

_Bool atomic_flag_test_and_set(volatile atomic_flag* obj);

sets a flag and returns its previous value

Performs atomically:

int atomic_flag_test_and_set(int* flag) {
  int old = *flag;
  *flag = 1;
  return old;
}

Implementing a lock:

void lock(int* lock) {
  while(atomic_flag_test_and_set(lock) == 1) ;
}

Here is an example of a program using a test_and_set based lock:

#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <stdatomic.h>

#define NITER 1000000
#define NTHREADS 4

volatile int lock=0;

int x = 0;
#ifdef NOT_THREAD_SAFE

/* thread-unsafe version */
void do_lock() {
  while(lock) ;
  lock = 1;
}

void do_unlock() {
  lock = 0;
}

#else

/* thread-safe version */
void do_lock() {
  while(atomic_flag_test_and_set(&lock)) ;
}

void do_unlock() {
  lock = 0;
}

#endif	/* NOT_THREAD_SAFE */

void* thread_function(void* arg) {
  for(int i=0; i<NITER; i++) {
    do_lock();
    x++;
    do_unlock();
  }
  return NULL;
}

int main(int argc, char**argv) {
  pthread_t tids[NTHREADS];
  int ret;
  for(int i = 0; i<NTHREADS; i++) {
    ret = pthread_create(&tids[i], NULL, thread_function, NULL);
    assert(ret == 0);
  }
  for(int i = 0; i<NTHREADS; i++) {
    ret = pthread_join(tids[i], NULL);
    assert(ret == 0);
  }

  printf("x = %d\n", x);
  return EXIT_SUCCESS;
}

Compare And Swap (CAS)

_Bool atomic_compare_exchange_strong(volatile A* obj, C* expected, C desired);

compares *obj and *expected
if equal, copy desired into *obj and return true
else, copy the value of *obj into *expected and return false
Performs atomically:

bool CAS(int* obj, int* expected, int desired) {
  if(*obj != *expected) {
    *expected = *obj;
    return false;
  } else {
    *obj = desired;
    return true;
  }
}

Here is an example of a program handling a lock-free list thanks to compare_and_swap:


#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <stdatomic.h>

#define NITER 1000000
#define NTHREADS 4

struct node {
  int value;
  struct node* next;
};

struct node *stack = NULL;

#ifdef NOT_THREAD_SAFE

/* thread-unsafe version */
void push(int value) {
  struct node* n = malloc(sizeof(struct node));
  n->value = value;
  n->next = stack;
  stack = n;
}

int pop() {
  struct node* n = stack;
  int value = 0;
  if(n) {
    value = n->value;
    stack = n->next;
    free(n);
  }
  return value;
}

#else

/* thread-safe version */
void push(int value) {
  struct node* n = malloc(sizeof(struct node));
  n->value = value;
  n->next = stack;

  int done = 0;
  do {
    done = atomic_compare_exchange_strong(&stack, &n->next, n);
  } while(!done);
}

int pop() {
  int value = 0;
  struct node* old_head = NULL;
  struct node* new_head = NULL;
  int done = 0;

  do {
    /* Warning: this function still suffers a race condition (search for
     * "ABA problem" for more information).
     * Fixing this would be too complicated for this simple example.
     */
    old_head = stack;
    if(old_head)
      new_head = old_head->next;
    done = atomic_compare_exchange_strong(&stack, &old_head, new_head);
  } while (!done);

  if(old_head) {
    value = old_head->value;
    free(old_head);
  }
  return value;
}

#endif	/* NOT_THREAD_SAFE */


_Atomic int sum = 0;
void* thread_function(void* arg) {
  for(int i=0; i<NITER; i++) {
    push(1);
  }

  int value;
  while((value=pop()) != 0) {
    sum+=value;
  }

  return NULL;
}

int main(int argc, char**argv) {
  pthread_t tids[NTHREADS];
  for(int i = 0; i<NTHREADS; i++) {
    pthread_create(&tids[i], NULL, thread_function, NULL);
  }
  for(int i = 0; i<NTHREADS; i++) {
    pthread_join(tids[i], NULL);
  }
  printf("sum = %d\n", sum);
  return EXIT_SUCCESS;
}

Fetch and Add

C atomic_fetch_add( volatile A* obj, M arg );
- replace obj with arg+obj
- return the old value of obj
Performs atomically:

int fetch_and_add(int* obj, int value) {
  int old = *obj;
  *obj = old+value;
  return old;
}

Here is an example of a program using fetch_and_add to atomically increment a variable:


#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <stdatomic.h>

#define NITER 1000000
#define NTHREADS 4

volatile int x = 0;

#ifdef NOT_THREAD_SAFE

/* thread-unsafe version */
void inc(volatile int * obj) {
  *obj = (*obj)+1;
}

#else

/* thread-safe version */
void inc(volatile int * obj) {
  atomic_fetch_add(obj, 1);
}

#endif 	/* NOT_THREAD_SAFE */

void* thread_function(void* arg) {
  for(int i=0; i<NITER; i++) {
    inc(&x);
  }
  return NULL;
}

int main(int argc, char**argv) {
  pthread_t tids[NTHREADS];
  for(int i = 0; i<NTHREADS; i++) {
    pthread_create(&tids[i], NULL, thread_function, NULL);
  }
  for(int i = 0; i<NTHREADS; i++) {
    pthread_join(tids[i], NULL);
  }

  printf("x = %d\n", x);
  return EXIT_SUCCESS;
}

Memory Fence (Barrière mémoire)

C atomic_thread_fence( memory_order order );
- performs a memory synchronization
- ensures that all past memory operations are visible by all threads according to the memory model chosen (see C11 memory model)

Synchronization primitives

Properties to consider when choosing a synchronization primitive
- Reactivity: time spent between the release of a lock and the unblocking of a thread waiting for this lock
- Contention: memory traffic generated by threads waiting for a lock
- Equity and risk of famine: if several threads are waiting for a lock, do they all have the same probability of acquire it? Are some threads likely to wait indefinitely?

Busy-waiting synchronization

int pthread_spin_lock(pthread_spinlock_t *lock);

tests the value of the lock until it becomes free, then acquires the lock

int pthread_spin_unlock(pthread_spinlock_t *lock);

Benefits
- Simple to implement (with test_and_set)
- Reactivity
Disadvantages
- Consumes CPU while waiting
- Consumes memory bandwidth while waiting

It is also possible to implement a spinlock using an atomic operation:

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <pthread.h>
#include <stdatomic.h>
#include <assert.h>

#define NITER 1000000
#define NTHREADS 4

struct lock {
  /* if flag=0, the lock is available
   * if flag=1, the lock is taken
   */
  volatile int flag;
};
typedef struct lock lock_t;


void lock(lock_t *l) {
  /* try to set flag to 1.
   * if the flag is already 1, loop and try again
   */
  while(atomic_flag_test_and_set(&l->flag)) ;
}

void unlock(lock_t *l) {
  l->flag = 0;
}

void lock_init(lock_t *l) {
  l->flag = 0;
}


lock_t l;
int x;

void* thread_function(void*arg){
  for(int i=0; i<NITER; i++) {
    lock(&l);
    x++;
    unlock(&l);
  }
  return NULL;
}

int main(int argc, char**argv) {
  lock_init(&l);

  pthread_t tids[NTHREADS];
  int ret;
  for(int i = 0; i<NTHREADS; i++) {
    ret = pthread_create(&tids[i], NULL, thread_function, NULL);
    assert(ret == 0);
  }
  for(int i = 0; i<NTHREADS; i++) {
    ret = pthread_join(tids[i], NULL);
    assert(ret == 0);
  }

  printf("x = %d\n", x);
  printf("expected: %d\n", NTHREADS*NITER);
  return EXIT_SUCCESS;

}

Futex

Fast Userspace Mutex
- System call allowing to build synchronization mechanisms in userland
- Allows waiting without monopolizing the CPU
- A futex is made up of:
  - a value
  - a waiting list
- Available operations (among others)
  - WAIT(int *addr, int value)
    - while(*addr == value) { sleep(); }
    - add the current thread to the waiting list
  - WAKE(int *addr, int value, int num)
    - *addr = value
    - wake up num threads waiting on addr

Implementing a mutex using a futex

mutex: an integer with two possible values: 1 (unlocked), or 0 (locked)
mutex_lock(m):
- Test and unset the mutex
- if mutex is 0, call FUTEX_WAIT
mutex_unlock(m):
- Test and set the mutex
- call FUTEX_WAKE to wake up a thread from the waiting list

Here is an example of a program implementing a mutex using futex:

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <pthread.h>
#include <stdatomic.h>
#include <linux/futex.h>
#include <sys/time.h>
#include <sys/syscall.h>
#include <errno.h>
#include <assert.h>

#define NITER 1000000
#define NTHREADS 4

struct lock {
  int flag;
};
typedef struct lock lock_t;


static int futex(int *uaddr, int futex_op, int val,
		 const struct timespec *timeout, int *uaddr2, int val3) {
  return syscall(SYS_futex, uaddr, futex_op, val,
		 timeout, uaddr2, val3);
}

void lock(lock_t *l) {  
  while (1) {
    /* Is the futex available? */
    int expected = 1;
    if (atomic_compare_exchange_strong(&l->flag, &expected, 0))
      return;      /* Yes */

    /* Futex is not available; wait */
    int s = futex(&l->flag, FUTEX_WAIT, 0, NULL, NULL, 0);
    if (s == -1 && errno != EAGAIN) {
      perror("futex_wait failed");
      abort();
    }
  }
}

void unlock(lock_t *l) {
  int expected = 0;
  atomic_compare_exchange_strong(&l->flag, &expected, 1);
  int s = futex(&l->flag, FUTEX_WAKE, 1, NULL, NULL, 0);
  if (s  == -1) {
    perror("futex_wake failed");
    abort();
  }
}

void lock_init(lock_t *l) {
  l->flag = 1;
}


lock_t l;
int x;

void* thread_function(void*arg){
  for(int i=0; i<NITER; i++) {
    //    printf("%d\n", i);
    lock(&l);
    x++;
    unlock(&l);
  }
  return NULL;
}

int main(int argc, char**argv) {
  lock_init(&l);

  pthread_t tids[NTHREADS];
  int ret;
  for(int i = 0; i<NTHREADS; i++) {
    ret = pthread_create(&tids[i], NULL, thread_function, NULL);
    assert(ret == 0);
  }
  for(int i = 0; i<NTHREADS; i++) {
    ret = pthread_join(tids[i], NULL);
    assert(ret == 0);
  }

  printf("x = %d\n", x);
  printf("expected: %d\n", NTHREADS*NITER);
  return EXIT_SUCCESS;

}

Implementing a monitor using a futex

condition: a counter

struct cond {
  int cpt;
};

void cond_wait(cond_t *c, pthread_mutex_t *m) {
  int value = atomic_load(&c->value);
  pthread_mutex_unlock(m);
  futex(&c->value, FUTEX_WAIT, value);
  pthread_mutex_lock(m);
}

void cond_signal(cond_t *c) {
  atomic_fetch_add(&c->value, 1);
  futex(&c->value, FUTEX_WAKE, 0);
}

Here is an example of a program implementing a condition using futex:

#include <stdlib.h>
#include <unistd.h>
#include <stdio.h>
#include <pthread.h>
#include <sys/syscall.h>
#include <linux/futex.h>
#include <stdatomic.h>
#include <assert.h>

#define N 10

int n_loops = 20;

struct cond {
  int cpt;
};
typedef struct cond cond_t;

static int futex(int *uaddr, int futex_op, int val) {
  return syscall(SYS_futex, uaddr, futex_op, val, NULL, uaddr, 0);
}

void cond_init(cond_t *c) {
  c->cpt = 0;
}

void cond_wait(cond_t *c, pthread_mutex_t *m) {
  int cpt = atomic_load(&c->cpt);
  pthread_mutex_unlock(m);
  futex(&c->cpt, FUTEX_WAIT, cpt);
  pthread_mutex_lock(m);
}

void cond_signal(cond_t *c) {
  atomic_fetch_add(&c->cpt, 1);
  futex(&c->cpt, FUTEX_WAKE, 0);
}



struct monitor{
  int value;
  pthread_mutex_t mutex;
  cond_t cond;
};

int infos[N];
int i_depot, i_extrait;
int nb_produits = 0;
struct monitor places_dispo;
struct monitor info_prete;


void* function_prod(void*arg) {
  static _Atomic int nb_threads=0;
  int my_rank = nb_threads++;
  
  for(int i=0; i<n_loops; i++) {
    int cur_indice;
    int product_id;
    usleep(100);
    pthread_mutex_lock(&places_dispo.mutex);
    while(places_dispo.value == 0) {
      cond_wait(&places_dispo.cond, &places_dispo.mutex);
    }
    places_dispo.value--;
    cur_indice = i_depot++;
    i_depot = i_depot % N;

    product_id = nb_produits++;
    pthread_mutex_unlock(&places_dispo.mutex);

    usleep(500000);
    printf("P%d produit %d dans %d\n", my_rank, product_id, cur_indice);

    pthread_mutex_lock(&info_prete.mutex);
    infos[cur_indice] = product_id;
    info_prete.value ++;
    cond_signal(&info_prete.cond);
    pthread_mutex_unlock(&info_prete.mutex);
  }
  return NULL;
}


void* function_cons(void*arg) {
  static _Atomic int nb_threads=0;
  int my_rank = nb_threads++;
  
  for(int i=0; i<n_loops; i++) {
    int cur_indice;
    int product_id;
    usleep(100);
    pthread_mutex_lock(&info_prete.mutex);
    while(info_prete.value == 0) {
      cond_wait(&info_prete.cond, &info_prete.mutex);
    }
    info_prete.value--;
    product_id = infos[i_extrait];
    cur_indice = i_extrait;
    i_extrait = (i_extrait+1) % N;
    pthread_mutex_unlock(&info_prete.mutex);

    usleep(100000);
    printf("C%d consomme %d depuis %d\n", my_rank, product_id, cur_indice);

    pthread_mutex_lock(&places_dispo.mutex);
    places_dispo.value ++;
    cond_signal(&places_dispo.cond);
    pthread_mutex_unlock(&places_dispo.mutex);
  }
  return NULL;
}

void init_monitor(struct monitor *m, int value) {
  m->value = value;
  pthread_mutex_init(&m->mutex, NULL);
  cond_init(&m->cond);
}

int main(int argc, char**argv) {
  init_monitor(&places_dispo, N);
  init_monitor(&info_prete, 0);
  i_depot = 0;
  i_extrait = 0;

  
  int nthreads_prod=2;
  int nthreads_cons=2;
  pthread_t tid_prod[nthreads_prod];
  pthread_t tid_cons[nthreads_cons];
  int ret;
  
  for(int i=0; i<nthreads_prod; i++) {
    ret = pthread_create(&tid_prod[i], NULL, function_prod, NULL);
    assert(ret == 0);
  }
  for(int i=0; i<nthreads_cons; i++) {
    ret = pthread_create(&tid_cons[i], NULL, function_cons, NULL);
    assert(ret == 0);
  }

  for(int i=0; i<nthreads_prod; i++) {
    ret = pthread_join(tid_prod[i], NULL);
    assert(ret == 0);
  }
  for(int i=0; i<nthreads_cons; i++) {
    ret = pthread_join(tid_cons[i], NULL);
    assert(ret == 0);
  }

  return EXIT_SUCCESS;
}

Using synchronization

Classic problems:
- deadlocks
- lock granularity
- scalability

Deadlock

Situation such that at least two processes are each waiting for a non-shareable resource already allocated to the other
Necessary and sufficient conditions (Coffman, 1971 [coffman1971])
1. Resources accessed under mutual exclusion (non-shareable resources)
2. Waiting processes (processes keep resources that are acquired)
3. Non-preemption of resources
4. Circular chain of blocked processes
Strategies:
- Prevention: acquisition of mutexes in the same order
- Deadlock detection and resolution (eg. with pthread_mutex_timedlock)

Lock granularity

Coarse grain locking
- A lock protects a large portion of the program
- Advantage: easy to implement
- Disadvantage: reduces parallelism
Fine grain locking
- Each lock protects a small portion of the program
- Advantage: possibility of using various resources in parallel
- Disadvantages:
  - Complex to implement without bug (eg. deadlocks, memory corruption)
  - Overhead (locking comes at a cost)

Scalability of a parallel system

Scalability = ability to reduce execution time when adding processing units
Sequential parts of a program reduce the scalability of a program (Amdhal’s law [amdahl1967])
In a parallel program, waiting for a lock introduced sequentiality → Locks can interfere with scalability

Bibliography

[amdahl1967] Amdahl, G. M. (1967, April). Validity of the single processor approach to achieving large scale computing capabilities. In Proceedings of the April 18-20, 1967, spring joint computer conference (pp. 483-485). ISO 690

[coffman1971] Coffman, Edward G., Melanie Elphick, and Arie Shoshani. “System deadlocks.” ACM Computing Surveys (CSUR) 3.2 (1971): 67-78.

Synchronization

Introduction

Atomic operations

Motivation

Can’t we just use volatile ?

Atomic operations

Test and set

Compare And Swap (CAS)

Fetch and Add

Memory Fence (Barrière mémoire)

Synchronization primitives

Busy-waiting synchronization

Futex

Implementing a mutex using a futex

Implementing a monitor using a futex

Using synchronization

Deadlock

Lock granularity

Scalability of a parallel system

Bibliography

Can’t we just use `volatile` ?