Reduce overhead in mutex - c

I am learning mutex currently and the following is the test code. It works perfectly fine. It creates another instance after one is done. Yet, it also introduces overhead according to here.
How efficient is locking an unlocked mutex? What is the cost of a mutex?. How can I modify my code to improve the efficiency?
pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
typedef struct _context_t {
uint64_t id;
char *name;
bool initialized;
} context_t;
context_t *ctx = NULL;
context_t *get_instance() {
pthread_mutex_lock(&lock);
if (ctx == NULL) {
ctx = (context_t *)malloc(sizeof(context_t));
assert(ctx != NULL);
ctx->initialized = false;
}
pthread_mutex_unlock(&lock);
return ctx;
}
int id = 0;
void *do_work(void *arg) {
context_t *ctx = get_instance();
if (!ctx->initialized) {
ctx->name = (char *)arg;
ctx->id = ++id;
ctx->initialized = true;
}
printf("name=%s\tid=%ld\n", ctx->name, ctx->id);
return NULL;
}
int main() {
int rc;
pthread_t p1, p2;
rc = pthread_create(&p1, NULL, do_work, "A");
assert(rc == 0);
rc = pthread_create(&p2, NULL, do_work, "B");
assert(rc == 0);
rc = pthread_join(p1, NULL);
assert(rc == 0);
rc = pthread_join(p2, NULL);
assert(rc == 0);
if (ctx) {
free(ctx);
}
return 0;
}

Instead of having two threads racing to create the context_t, you should create it once before the threads start, and pass its address to the threads explicitly.
Note that you can pass multiple arguments via pthread_create() by putting them in a struct and passing its address.
Then you won't need a mutex at all, because the threads will only read from ctx rather than potentially writing to it.

Related

how to create undefined number of threads and use WaitForMultipleObjects() in c on windows

PS: I am very new to threads.
I have a problem where i need to wait for connection requests(completely arbitrary number of times) from clients, accept a connection on a socket, create a worker thread after connection. The created thread then creates a char array, works on it and needs to pass it to the parent process.
I have been able to create the threads in a while loop like
while ((new_socket = accept(srv_sock, (struct sockaddr *)&client, &c)) != INVALID_SOCKET)
{
puts("\nConnection accepted");
_beginthreadex(0, 0, handle_client, &new_socket, 0, 0);
}
I have seen that pthread_join() can be used to pass data from thread to parent process(in unix). My question is, how can I integrate it into a loop in the main process.
I expect the following approach will result in a situation where no more than one connection can be established between client and server at a time,which is not desired.
while ((new_socket = accept(srv_sock, (struct sockaddr *)&client, &c)) != INVALID_SOCKET)
{
puts("\nConnection accepted");
_beginthreadex(0, 0, handle_client, &new_socket, 0, 0);
pthread_join(thread_id,&my_array);
}
EDIT: I would be happy to know if what I want is impossible or if there are alternatives to pthread_join(). or its windows equivalent.
EDIT: I know that pthread_join() is for Unix and have read that WaitForMultipleObjects() is its equivalent for windows. In any case I still haven't been able to figure out a solution.
I have seen that pthread_join() can be used to pass data from thread to parent process.
That is not entirely correct. You can pass a pointer when you exit a thread, and collect that pointer using pthread_join. You have to implement all the logic yourself. The API does not know (or care) what the pointer is. Threads don't have parents and children, they are siblings.
Example for a creator and a reaper:
global
struct VarLengthArray {
size_t count;
MyElem data[1];
};
exiting thread:
// allocate the result
size_t count = ...;
VarLengthArray *retval = malloc(
sizeof(VarLengthArray) +
sizeof(MyElem) * (count > 0 ? count - 1 : 0)
);
// fill the result
retval->count = count;
for (size_t i = 0; i < retval->count; ++i) {
retval->data[i] = ...;
}
pthread_exit(retval);
collecting thread:
// collect the result
void *retval_;
if (pthread_join(thread_one_id, &retval_) != 0) {
// handle error
}
VarLengthArray *retval = retval_;
// use the result
for (size_t i = 0; i < retval->count; ++i) {
printf("retval->[%u] = %s\n", (unsigned) i, retval->data[i].string_value);
}
// deallocate the result
free(retval);
A full example using a condition variable and multiple creators:
#include <limits.h>
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
typedef struct Datum {
struct Datum *next;
char some_data[32];
} Datum;
typedef struct SharedData {
pthread_mutex_t mutex;
pthread_cond_t cond_empty;
unsigned seed;
Datum *head, *tail;
unsigned children_alive;
} SharedData;
static void *thread_logic(void *argv_);
int main(int argc, char **argv) {
unsigned thread_count = 2;
if (argc > 1) {
if (sscanf(argv[1], " %u ", &thread_count) != 1) {
fprintf(stderr, "Usage: %s [thread_count]\n", argv[0]);
return 1;
}
}
// initialize shared data
SharedData shared_data;
pthread_mutex_init(&shared_data.mutex, NULL);
pthread_cond_init(&shared_data.cond_empty, NULL);
shared_data.seed = time(NULL);
shared_data.head = NULL;
shared_data.tail = NULL;
shared_data.children_alive = 0;
// start threads detached, so you don't have to call pthread_join
pthread_t *child_ids = malloc(sizeof(pthread_t) * thread_count);
pthread_attr_t attr;
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
// start the threads
pthread_mutex_lock(&shared_data.mutex);
for (unsigned i = 0; i < thread_count; ++i) {
if (pthread_create(&child_ids[i], &attr, thread_logic, &shared_data) != 0) {
perror("pthread_create");
} else {
++shared_data.children_alive;
}
}
pthread_mutex_unlock(&shared_data.mutex);
pthread_attr_destroy(&attr);
// loop until all threads are dead
while (shared_data.children_alive > 0) {
// a condition variable: wait until there is data you can read
pthread_mutex_lock(&shared_data.mutex);
while (shared_data.head == NULL) {
pthread_cond_wait(&shared_data.cond_empty, &shared_data.mutex);
}
// collect a first datum
Datum *datum = shared_data.head;
if (datum->next != NULL) {
shared_data.head = datum->next;
} else {
shared_data.head = shared_data.tail = NULL;
}
pthread_mutex_unlock(&shared_data.mutex);
// handle the data (outside of the mutex lock)
printf("Got data: %s\n", datum->some_data);
free(datum);
}
return 0;
}
static void *thread_logic(void *shared_data_) {
SharedData *shared_data = shared_data_;
while (1) {
pthread_mutex_lock(&shared_data->mutex);
// create some data
useconds_t timeout = (
(((float) (unsigned) rand_r(&shared_data->seed)) / UINT_MAX) *
1000000
);
Datum *datum = malloc(sizeof(Datum));
datum->next = NULL;
if (timeout < 1000000 / 25) {
--shared_data->children_alive;
snprintf(datum->some_data, sizeof(datum->some_data), "I'm done\n");
} else {
snprintf(
datum->some_data, sizeof(datum->some_data),
"Sleeping for %uus\n", timeout
);
}
// append the datum
if (shared_data->head) {
shared_data->tail->next = datum;
} else {
shared_data->head = datum;
pthread_cond_signal(&shared_data->cond_empty);
}
shared_data->tail = datum;
pthread_mutex_unlock(&shared_data->mutex);
// most likely it takes some time to create the data
// do lengthly tasks outside of the mutex lock
if (timeout < 1000000 / 25) {
return NULL;
} else {
usleep(timeout);
}
}
}

Pthreads: memory getting corrupted

I have a pthread program where thread function accesses the information updated by the main. I am using a mutex to ensure synchronization, but I'm still observing a memory corruption:
Main thread acquires lock and shall update the global data (via print debugs ensured that the data updated properly). Releases lock
thread will acquire lock and access the global data. Here the data is corrupted.
I have the code snippet below. I am not sure if I have missed out something. Please provide your suggestions.
/* Global info*/
struct ThreadParams {
struct ofconn *ofconn;
const struct ofp_header *request;
};
#define NTHREADS 1
pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
struct ThreadParams threadParams;
enum ercode
handle_request(struct conn *conn, const struct header *request)
{
int rc;
printf("\n%s in!!",__func__);
threadParams.conn = conn;
threadParams.request = request;
printf("\n\t%s: Request type %d", __func__,threadParams.request->type);
printf("\n\t%s: Unlocking Mutex!!",__func__);
rc = pthread_mutex_unlock(&mutex);
checkResults("pthread_mutex_unlock()\n", rc);
printf("\n\t%s: Completed\n",__func__);
return 0;
}
void *thread_func()
{
int rc;
enum ercode error;
printf("\n%s In!!",__func__);
printf("\n%s: Locking Mutex!!\n",__func__);
rc = pthread_mutex_lock(&mutex);
checkResults("pthread_mutex_lock()\n", rc);
printf("\n%s:Received Lock with Request type %d",
__func__, threadParams.request->type); // data is corrupted here
struct util_request fsr;
error = _decode__request(&fsr, threadParams.request);
if (error) {
return 1;
}
printf("\n\t%s: UnLocking Mutex!!",__func__);
rc = pthread_mutex_unlock(&mutex);
checkResults("pthread_mutex_lock()\n", rc);
return NULL;
}
int main(void)
{
int rc=0, i;
pthread_t threadid[NTHREADS];
printf("\n%s: Locking Mutex!!\n",__func__);
rc = pthread_mutex_lock(&mutex);
checkResults("pthread_mutex_lock()\n", rc);
for(i=0; i<NTHREADS; ++i) {
rc = pthread_create(&threadid[i], NULL, thread_func, NULL);
checkResults("pthread_create()\n", rc);
}
sleep(5);
printf("\nCalling hanlde__request function\n");
handle_request(ofconn, request);
return 0;
}

why doesn't sem_wait block

static int res1 = 0;
static int res2 = 0;
static int res3 = 0;
static int counter = 0;
static sem_t sem;
void * func_thread1(void *p)
{
sleep(2);
res1 = 1;
printf("func_thread1\n");
sem_post(&sem);
return NULL;
}
void * func_thread2(void *p)
{
sleep(2);
res2 = 2;
printf("func_thread2\n");
sem_post(&sem);
return NULL;
}
void * func_thread3(void *p)
{
sem_wait(&sem);
sem_wait(&sem);
res3 = res1 + res2;
printf("func_thread3\n");
return NULL;
}
void main()
{
sem_init(&sem, 0, counter);
pthread_t pd1, pd2, pd3;
pthread_create(&pd1, NULL, func_thread1, NULL);
pthread_create(&pd2, NULL, func_thread2, NULL);
pthread_create(&pd3, NULL, func_thread3, NULL);
//pthread_join(pd3, NULL);
printf("main_thread\n");
printf("%d", res3);
}
I'm trying to understand how semaphore works.
I'm trying to make the td3 block to wait for the td1 and the td2.
In my opinion, the sem_wait will block twice. If the sem_posts in func_thread1 and in func_thread2 are executed, func_thread3 could continue.
However, it doesn't work unless I add pthread_join(td3, NULL) in main. I think the join is not necessary because sem_wait can block.
So pthread_join is necessary or I use semaphore incorrectly?
pthread_join is mandatory in your implementation.
Otherwise your process finishes (ie. main returns), and all tasks (ie. threads) are killed before thread 3 prints anything.

Pthread runtime errors

I'm having trouble debugging the following program I wrote. The idea is to have two seperate threads; one thread executes a 5 second countdown while the other waits for key input from the user. Whichever thread completes first should cancel the sibling thread and exit the program. However, the following code just hangs.
Any help would be appreciated, but I would be most grateful for an explanation as to the problem.
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h> // For sleep()
#define NUM_THREADS 2
// The stuct to be passed as an argument to the countdown routine
typedef struct countdown_struct {
pthread_t *thread;
signed int num_secs;
} CountdownStruct;
// Struct for passing to the input wait routine
typedef struct wait_struct {
pthread_t *thread;
int *key;
} WaitStruct;
// Countdown routine; simply acts as a timer counting down
void * countdown(void *args)
{
CountdownStruct *cd_str = (CountdownStruct *)args;
signed int secs = cd_str->num_secs;
printf("Will use default setting in %d seconds...", secs);
while (secs >= 0)
{
sleep(1);
secs -= 1;
printf("Will use default setting in %d seconds...", secs);
}
// Cancel the other struct
pthread_cancel(*(cd_str->thread));
return NULL;
}
// Waits for the user to pass input through the tty
void * wait_for_input(void *args)
{
WaitStruct *wait_str = (WaitStruct *) args;
int c = 0;
do {
c = getchar();
} while (!(c == '1' || c == '2'));
*(wait_str->key) = c;
// Cancel the other thread
pthread_cancel(*(wait_str->thread));
return NULL;
}
int main(int argc, char **argv)
{
pthread_t wait_thread;
pthread_t countdown_thread;
pthread_attr_t attr;
int key=0;
long numMillis=5000;
int rc=0;
int status=0;
// Create the structs to be passe as paramaters to both routines
CountdownStruct *cd_str = (CountdownStruct *) malloc(sizeof(CountdownStruct));
if (cd_str == NULL)
{
printf("Couldn't create the countdown struct. Aborting...");
return -1;
}
cd_str->thread = &wait_thread;
cd_str->num_secs = 5;
WaitStruct *wait_str = (WaitStruct *) malloc(sizeof(WaitStruct));
if (wait_str == NULL)
{
printf("Couldn't create the iput wait struct. Aborting...");
return -1;
}
wait_str->thread = &countdown_thread;
wait_str->key = &key;
// Create the joinable attribute
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
// Create both threads
rc = pthread_create(&countdown_thread, &attr, countdown, (void *) cd_str);
if (rc) { printf("Error with the thread creation!"); exit(-1); }
rc = pthread_create(&wait_thread, &attr, wait_for_input, (void *) wait_str);
if (rc) { printf("Error with the thread creation!"); exit(-1); }
// Destroy the pthread_attribute
pthread_attr_destroy(&attr);
// now join on the threads and wait for main
pthread_join(wait_thread, NULL);
pthread_join(countdown_thread, NULL);
// Call pthread_exit
pthread_exit(NULL);
// Free the function structs
free(cd_str);
free(wait_str);
}
Getchar is not required to be a cancellation point. Select and pselect are. Even if you want to continue to use a countdown thread you could still provide a cancellation point in the opposing thread by use of select.
I had reasonable behavior with the following modified wait_for_input()
// Waits for the user to pass input through the tty
void * wait_for_input(void *args)
{
WaitStruct *wait_str = (WaitStruct *) args;
int c = 0;
fd_set readFds;
int numFds=0;
FD_ZERO(&readFds);
do {
struct timeval timeout={.tv_sec=8,.tv_usec=0};
/* select here is primarily to serve as a cancellation
* point. Though there is a possible race condition
* still between getchar() getting called right as the
* the timeout thread calls cancel.().
* Using the timeout option on select would at least
* cover that, but not done here while testing.
*******************************************************/
FD_ZERO(&readFds);
FD_SET(STDOUT_FILENO,&readFds);
numFds=select(STDOUT_FILENO+1,&readFds,NULL,NULL,&timeout);
if(numFds==0 )
{
/* must be timeout if no FD's selected */
break;
}
if(FD_ISSET(STDOUT_FILENO,&readFds))
{
printf("Only get here if key pressed\n");
c = getchar();
}
} while (!(c == '1' || c == '2'));
*(wait_str->key) = c;
// Cancel the other thread
pthread_cancel(*(wait_str->thread));
return NULL;
}

event_new() function fails on hpux itanium

I'm trying to debug a code that is using a libevent library. In that library, there is a function event_new that is suppose to create an event_cb. Somehow after I dispatch the event base, the event_cb cannot be called or accessed. This problem only happens on hpux itanium. This code works on hpux pa-risc, Redhat, AIX, and Solaris. Is there any certain thing that need to be set?
This is part of the code
int ttypread (int fd, Header *h, char **buf)
{
int c,k;
struct user_data user_data;
struct bufferevent *in_buffer;
struct event_config *evconfig;
log_debug("inside ttypread");
in_buffer = NULL;
user_data.fd = fd;
user_data.h = h;
user_data.buf = buf;
log_debug("from user_data, fd = %d",user_data.fd); //the log_debug is a debugging function for me to check the value sent by the system. I use it to compare between each platform
log_debug("from user_data, buf = %s",user_data.buf);
log_debug("from user_data, h.len = %d",user_data.h->len);
log_debug("from user_data, h.type = %d",user_data.h->type);
evconfig = event_config_new();
if (evconfig == NULL) {
log_error("event_config_new failed");
return -1;
}
if (event_config_require_features(evconfig, EV_FEATURE_FDS)!=0) {
log_error("event_config_require_features failed");
return -1;
}
base = event_base_new_with_config(evconfig);
if (!base) {
log_error("ttypread:event_base_new failed");
return -1;
}
const char* method; //these 3 lines are the new line edited
method = event_base_get_method(base);
log_debug("ttyread is using method = %s",method);
ev = event_new(base, fd, EV_READ|EV_PERSIST, ttypread_event_cb, &user_data);
c = event_add(ev, NULL);
log_debug("ttypread passed event_add with c value is %d",c);
in_buffer = bufferevent_socket_new(base, STDIN_FILENO, BEV_OPT_CLOSE_ON_FREE);
log_debug("ttypread passed bufferevent_socket_new");
if(in_buffer == NULL){
log_debug("problem with bufferevent_socket_new");
}
bufferevent_setcb(in_buffer, in_read_cb, NULL, in_event_cb, NULL);
bufferevent_disable(in_buffer, EV_WRITE);
bufferevent_enable(in_buffer, EV_READ);
k =event_base_dispatch(base);
log_debug("event_base have been dispatched"); //when looking at the debugging file, the other plaform will go to ttypread_event_cb function. But for hpux itanium, it stays here.
if (k == 0){
log_debug("event_base_dispatch returned 0");
} else if (k == -1){
log_debug("event_base_dispatch returned -1");
} else {
log_debug("event_base_dispatch returned 1");
}
event_base_free(base);
event_free(ev);
log_debug("finish ttypread");
log_debug("ttypread_ret will return [%d]",ttypread_ret);
return ttypread_ret;
}
void ttypread_event_cb(evutil_socket_t fd, short events, void *arg)
{
int nread;
struct timeval t;
struct user_data *user_data;
user_data = (struct user_data*)arg;
nread = 0;
log_debug("inside ttypread_event_cb");
if (events & EV_READ) {
log_debug("got events & EV_READ");
nread = ttyread(fd, user_data->h, user_data->buf);
if (nread == -1) {
ttypread_ret = -1;
event_del(ev);
event_base_loopexit(base, NULL);
} else if (nread == 0) {
if (access(input_filename, F_OK)!=0) {
log_debug("cannot access [%s]",input_filename);
tcsetattr(0, TCSANOW, &old); /* Return terminal state */
exit(EXIT_SUCCESS);
}
t.tv_sec = 0;
t.tv_usec = 250000;
select(0, 0, 0, 0, &t);
} else {
ttypread_ret = 1;
event_del(ev);
event_base_loopexit(base, NULL);
}
}
else if (events & EV_WRITE) {
log_debug("got events & EV_WRITE");
}
}
Not sure if this help. But just some info on the hpux itanium
uname -a = HP-UX hpux-ita B.11.23 U ia64
If you need any additional info or other declaration on function, just leave a comment and I will edit the question.
EDIT : i've added a function inside ttypread. Somehow for hpux itanium its returning devpoll while other platform are returning poll. Im not sure if this is the problem. But if that is so, is there any way for me to change it?
After checking the result from event_base_get_method, I found out that only on my hpux-itanium used devpoll method. This is how I solve it.
char string[8] = "devpoll";
struct user_data user_data;
struct bufferevent *in_buffer;
struct event_config *evconfig;
const char *method;
const char *devpoll;
devpoll = string;
in_buffer = NULL;
user_data.fd = fd;
user_data.h = h;
user_data.buf = buf;
evconfig = event_config_new();
if (evconfig == NULL) {
log_error("event_config_new failed");
return -1;
}
if (event_config_require_features(evconfig, EV_FEATURE_FDS)!=0) {
log_error("event_config_require_features failed");
return -1;
}
if (event_config_avoid_method(evconfig,devpoll) != 0)
{
log_error("Failed to ignore devpoll method");
}
Force the libevent to ignore using devpoll and use poll instead.

Resources