-#define _GNU_SOURCE
#include <unistd.h>
#include <fcntl.h>
#include <sys/select.h>
#include <sys/epoll.h>
#include <stdio.h>
+#include <sys/wait.h>
+#include <sys/signalfd.h>
+#include <sys/resource.h>
#include "process.h"
#include "debug.h"
+#include "utils.h"
-static int child_count;
-static int parent_count;
-static int job_request_fd[2];
-static int job_get_permission_fd[2];
+static int epoll_fd;
static unsigned int max_jobs;
-static unsigned int job_count;
-static unsigned int jobs_pending;
-
-int get_child_count(void)
+static unsigned int max_jobs_pending;
+
+struct work_struct {
+ const char *name;
+ int (*work_fn)(void *);
+ void *arg;
+ struct work_struct *next;
+};
+
+struct work_queue {
+ struct work_struct *work;
+ int length;
+ char *name;
+ struct mutex lock;
+};
+
+struct work_queue work_queues[WORK_PRIORITIES_NUM] = {
+ {
+ .name = "high priority",
+ .lock = {
+ .name = "high_prio_queue",
+ .lock = PTHREAD_MUTEX_INITIALIZER,
+ },
+ },
+ {
+ .name = "low priority",
+ .lock = {
+ .name = "low_prio_queue",
+ .lock = PTHREAD_MUTEX_INITIALIZER,
+ },
+ },
+};
+
+struct mutex work_stats_mutex = {
+ .name = "work_stats",
+ .lock = PTHREAD_MUTEX_INITIALIZER,
+};
+static int workers_active;
+static int worker_count;
+
+static int job_notify_fd[2];
+
+static int run_work_on_queue(struct work_queue *queue)
{
- return child_count;
-}
+ struct work_struct *work;
-int get_parent_count(void)
-{
- return parent_count;
+ mutex_lock(&queue->lock);
+
+ if (!queue->work) {
+ mutex_unlock(&queue->lock);
+ return 0;
+ }
+
+ /* Take next work */
+ work = queue->work;
+ queue->work = work->next;
+ queue->length--;
+
+ mutex_unlock(&queue->lock);
+
+ pr_info("Executing work %s from queue %s, %d still pending\n",
+ work->name, queue->name, queue->length);
+
+ work->work_fn(work->arg);
+ pr_info("Work %s done\n", work->name);
+ free(work);
+
+ return 1;
}
-static void sigchild_handler(int signal)
+static void *worker_thread(void *arg)
{
- pr_info("Child has died, go harvest the zombie\n");
- harvest_zombies(0);
+ int ret;
+ char name[16];
+
+ mutex_lock(&work_stats_mutex);
+ snprintf(name, sizeof(name), "worker%d", worker_count);
+ worker_count++;
+ mutex_unlock(&work_stats_mutex);
+
+ pthread_setname_np(pthread_self(), name);
+ pthread_detach(pthread_self());
+
+ pr_info("Worker started\n");
+
+ /* Execute all high priority work from the queue */
+ while (run_work_on_queue(&work_queues[WORK_PRIORITY_HIGH]))
+ ;
+ /*
+ * All high priority work is now done, see if we have enough
+ * workers executing low priority worl. Continue from there if
+ * needed.
+ */
+ mutex_lock(&work_stats_mutex);
+ if (workers_active > max_jobs)
+ goto out_unlock;
+
+ mutex_unlock(&work_stats_mutex);
+
+ /*
+ * Start executing the low priority work. Drop the nice value
+ * as this really is low priority stuff
+ */
+ ret = nice(19);
+ pr_info("Worker priority dropped to %d\n", ret);
+
+ while (run_work_on_queue(&work_queues[WORK_PRIORITY_LOW]))
+ ;
+
+ /* All done, exit */
+ mutex_lock(&work_stats_mutex);
+out_unlock:
+ workers_active--;
+ pr_info("Worker exiting, %d left active\n",
+ workers_active);
+
+ /*
+ * Last exiting worker zeroes the worker_count. This
+ * ensures next time we start spawning worker threads
+ * the first thread will have number zero on its name.
+ */
+ if (!workers_active)
+ worker_count = 0;
+
+ /*
+ * Kick the job poller. If all jobs were active at this point
+ * the scheduler thread will wait indefinitely until someone
+ * tells it to do something. We may now know when next job is
+ * available, so it is better for the scheduler to recalculate
+ * its sleep time.
+ */
+ notify_job_request();
+
+ mutex_unlock(&work_stats_mutex);
+
+ return NULL;
}
-static int setup_sigchild_handler(void)
+int queue_work(unsigned int priority, char *name,
+ int (work_fn)(void *arg), void *arg)
{
- struct sigaction sa;
- int ret;
+ pthread_t *thread;
+ struct work_queue *queue;
+ struct work_struct *work, *last_work;
- sa.sa_handler = sigchild_handler;
- sa.sa_flags = SA_NOCLDSTOP;
+ if (priority >= WORK_PRIORITIES_NUM) {
+ pr_err("Invalid priority: %d\n", priority);
+ return -EINVAL;
+ }
- ret = sigaction(SIGCHLD, &sa, NULL);
- if (ret)
- pr_err("Failed to setup SIGCHLD handler: %m\n");
+ work = calloc(sizeof(*work), 1);
- return ret;
+ work->name = name;
+ work->work_fn = work_fn;
+ work->arg = arg;
+
+ queue = &work_queues[priority];
+
+ /* Insert new work at the end of the work queue */
+ mutex_lock(&queue->lock);
+
+ last_work = queue->work;
+ while (last_work && last_work->next)
+ last_work = last_work->next;
+
+ if (!last_work)
+ queue->work = work;
+ else
+ last_work->next = work;
+
+ pr_info("Inserted work %s in queue %s, with %d pending items\n",
+ work->name, queue->name, queue->length);
+ queue->length++;
+ mutex_unlock(&queue->lock);
+
+ mutex_lock(&work_stats_mutex);
+ pr_info("workers_active: %d, priority: %d\n", workers_active, priority);
+ if (priority != WORK_PRIORITY_HIGH && workers_active >= max_jobs) {
+ mutex_unlock(&work_stats_mutex);
+ return 0;
+ }
+ workers_active++;
+ mutex_unlock(&work_stats_mutex);
+
+ pr_info("Creating new worker thread\n");
+ /* We need a worker thread, create one */
+ thread = calloc(sizeof(*thread), 1);
+ pthread_create(thread, NULL, worker_thread, NULL);
+
+ free(thread);
+
+ return 0;
}
-static int cancel_sigchild_handler(void)
+static int job_notify_handler(struct event_handler *h)
{
- struct sigaction sa;
int ret;
+ char buf[64];
- sa.sa_handler = SIG_DFL;
- sa.sa_flags = SA_NOCLDSTOP;
-
- ret = sigaction(SIGCHLD, &sa, NULL);
- if (ret)
- pr_err("Failed to cancel SIGCHLD handler: %m\n");
+ ret = read(job_notify_fd[0], buf, sizeof(buf));
+ if (ret < 0)
+ pr_err("Failed to read: %m\n");
- return ret;
+ return 0;
}
/*
* Initialize the jobcontrol.
*
* Create the pipes that are used to grant children execution
- * permissions. If max_jobs is zero, count the CPUs from /proc/cpuinfo
- * and use that.
+ * permissions. If max_jobs is zero, count the number of CPUs from
+ * /proc/cpuinfo and use that.
*/
-int init_max_jobs(int max_jobs_requested)
+int init_jobcontrol(int max_jobs_requested)
{
+ static struct event_handler ev;
FILE *file;
int ret;
char buf[256];
char match[8];
- if (pipe2(job_request_fd, O_NONBLOCK | O_CLOEXEC)) {
- pr_err("Failed to create pipe: %m\n");
+ epoll_fd = epoll_create(1);
+ if (epoll_fd == -1) {
+ pr_err("Failed to epoll_create(): %m\n");
return -1;
}
- if (pipe2(job_get_permission_fd, O_CLOEXEC)) {
- pr_err("Failed to create pipe: %m\n");
- return -1;
- }
-
- ret = setup_sigchild_handler();
- if (ret)
- return ret;
-
if (max_jobs_requested > 0) {
max_jobs = max_jobs_requested;
goto no_count_cpus;
max_jobs++;
}
+ ret = pipe(job_notify_fd);
+ if (ret) {
+ pr_err("pipe() failed: %m\n");
+ return ret;
+ }
+
+ ev.fd = job_notify_fd[0];
+ ev.events = EPOLLIN;
+ ev.handle_event = job_notify_handler;
+ ev.name = "job_notify";
+ register_event_handler(&ev, EPOLL_CTL_ADD);
+
open_fail:
read_fail:
fclose(file);
no_count_cpus:
pr_info("Set maximum number of parallel jobs to %d\n", max_jobs);
- return 0;
-}
-
-static int grant_new_job(void)
-{
- int ret;
- char byte = 0;
-
- job_count++;
- pr_info("Granting new job. %d jobs currently and %d pending\n",
- job_count, jobs_pending);
-
- ret = write(job_get_permission_fd[1], &byte, 1);
- if (ret != 1) {
- pr_err("Failed to write 1 byte: %m\n");
- return -1;
- }
+ max_jobs_pending = max_jobs * 10 + 25;
+ pr_info("Set maximum number of pending jobs to %d\n", max_jobs_pending);
return 0;
}
int poll_job_requests(int timeout)
{
struct epoll_event event;
- static int epollfd;
+ struct event_handler *job_handler;
int ret;
- int pid;
-
- if (!epollfd) {
- struct epoll_event ev;
-
- epollfd = epoll_create(1);
- if (epollfd == -1) {
- pr_err("Failed to epoll_create(): %m\n");
- return -1;
- }
-
- ev.events = EPOLLIN;
- ev.data.fd = job_request_fd[0];
- epoll_ctl(epollfd, EPOLL_CTL_ADD, job_request_fd[0], &ev);
- }
/* Convert positive seconds to milliseconds */
timeout = timeout > 0 ? 1000 * timeout : timeout;
- ret = epoll_wait(epollfd, &event, 1, timeout);
+ ret = epoll_wait(epoll_fd, &event, 1, timeout);
if (ret == -1) {
if (errno != EINTR) {
goto out;
}
- ret = read(event.data.fd, &pid, sizeof(pid));
- if (ret < 0) {
- pr_err("Failed to read: %m\n");
- return -1;
- }
+ job_handler = event.data.ptr;
- if (ret == 0) {
- pr_info("Read zero bytes\n");
- return 0;
+ if (!job_handler || !job_handler->handle_event) {
+ pr_err("Corrupted event handler for fd %d\n",
+ event.data.fd);
+ goto out;
}
- if (pid > 0) {
- if (job_count >= max_jobs) {
- jobs_pending++;
- } else {
- ret = grant_new_job();
- goto out;
- }
- } else if (pid < 0) {
- if (job_count > max_jobs)
- pr_err("BUG: Job %u jobs exceeded limit %u\n",
- job_count, max_jobs);
-
- pr_info("Job %d finished\n", -pid);
- job_count--;
- if (jobs_pending) {
- jobs_pending--;
- ret = grant_new_job();
- goto out;
- }
- }
+ pr_debug("Running handler %s to handle events from fd %d\n",
+ job_handler->name, job_handler->fd);
+ job_handler->handle_event(job_handler);
out:
- pr_info("Jobs active: %u, pending: %u\n", job_count, jobs_pending);
+ pr_info("Workers active: %u\n", workers_active);
return ret;
}
-/*
- * Per process flag indicating whether this child has requested fork
- * limiting. If it has, it must also tell the master parent when it
- * has died so that the parent can give next pending job permission to
- * go.
- */
-static int is_limited_fork;
+int notify_job_request(void)
+{
+ int ret;
+ char byte = 0;
+
+ ret = write(job_notify_fd[1], &byte, sizeof(byte));
+ if (ret < 0)
+ pr_err("Failed to write: %m\n");
+
+ return 0;
+}
int do_fork(void)
{
}
if (child) {
- child_count++;
- pr_info("Fork %d, child %d\n", child_count, child);
+ pr_debug("Fork child %d\n", child);
return child;
}
- /*
- * Child processes may want to use waitpid() for synchronizing
- * with their sub-childs. Disable the signal handler by
- * default
- */
- cancel_sigchild_handler();
-
- /*
- * Also do not notify the master parent the death of this
- * child. Only childs that have been created with
- * do_fork_limited() can have this flag set.
- */
- is_limited_fork = 0;
-
- /* reset child's child count */
- child_count = 0;
- parent_count++;
return 0;
}
-static int request_fork(char request)
-{
- int pid = getpid();
-
- pid = request > 0 ? pid : -pid;
-
- return write(job_request_fd[1], &pid, sizeof(pid));
-}
-
-static void limited_fork_exit_handler(void)
-{
- if (is_limited_fork)
- request_fork(-1);
-}
-
-/*
- * Like do_fork(), but allow the child continue only after the global
- * job count is low enough.
- *
- * We allow the parent to continue other more important activities but
- * child respects the limit of global active processes.
- */
-int do_fork_limited(void)
-{
- int child, ret;
- char byte;
-
- child = do_fork();
- if (child)
- return child;
-
- /* Remember to notify the parent when we are done */
- atexit(limited_fork_exit_handler);
- is_limited_fork = 1;
-
- pr_info("Requesting permission to go\n");
-
- /* Signal the parent that we are here, waiting to go */
- request_fork(1);
-
- /*
- * The parent will tell us when we can continue. If there were
- * multiple children waiting for their turn to run parent's
- * write to the pipe will wake up every process reading
- * it. However, only one will be reading the content and
- * getting the permission to run. So we will read as many
- * times as needed until we get our own permission to run.
- */
- do {
- ret = read(job_get_permission_fd[0], &byte, sizeof(byte));
- if (ret == 1)
- break;
-
- } while(1);
-
- pr_info("Continuing\n");
- return child;
-}
-
-int harvest_zombies(int pid)
+int clear_zombie(int pid)
{
int status;
-
- if (child_count == 0)
- return 0;
+ struct rusage rusage;
+ char *status_str = NULL;
+ int code = 0;
if (pid)
- pr_info("Waiting on pid %d, children left: %d\n", pid,
- child_count);
+ pr_debug("Waiting on pid %d\n", pid);
- pid = waitpid(pid, &status, 0);
- if (pid < 0) {
- pr_err("Error on wait(): %m\n");
- child_count--; /* Decrement child count anyway */
- }
- else {
- child_count--;
- pr_info("pid %d: exit code %d. Children left: %d\n", pid,
- status, child_count);
+ do {
+ pid = wait4(pid, &status, 0, &rusage);
+ if (pid < 0) {
+ pr_err("Error on waitid(): %m\n");
+ return 0;
+ }
+ /* Wait until the child has become a zombie */
+ } while (!WIFEXITED(status) && !WIFSIGNALED(status));
+
+ if (WIFEXITED(status)) {
+ status_str = "exited with status";
+ code = WEXITSTATUS(status);
+ } else if (WIFSIGNALED(status)) {
+ status_str = "killed by signal";
+ code = WTERMSIG(status);
}
+ pr_debug("pid %d: %s %d.\n", pid,
+ status_str, code);
+ pr_debug("pid %d: User time: %ld.%03lds, System %ld.%03lds\n", pid,
+ (long)rusage.ru_utime.tv_sec, rusage.ru_utime.tv_usec / 1000,
+ (long)rusage.ru_stime.tv_sec, rusage.ru_stime.tv_usec / 1000);
return 1;
}
if (stdoutfd && pipe(ofd)) {
pr_err("pipe() failed: %m\n");
- return -1;
+ goto close_ifd;
}
if (stderrfd && pipe(efd)) {
pr_err("pipe() failed: %m\n");
- return -1;
+ goto close_ofd;
}
pid = do_fork();
exit(1);
return 0;
+
+close_ofd:
+ if (stdoutfd) {
+ close(ofd[0]);
+ close(ofd[1]);
+ }
+close_ifd:
+ if (stdinfd) {
+ close(ifd[0]);
+ close(ifd[1]);
+ }
+
+ return -1;
}
/*
if (stdinf)
i = &ifd;
- else
+ else
i = 0;
if (stdoutf)
o = &ofd;
- else
+ else
o = 0;
if (stderrf)
e = &efd;
- else
+ else
e = 0;
pid = run_piped(cmd, argv, i, o, e);
if (stdinf) {
- *stdinf = fdopen(ifd, "r");
+ *stdinf = fdopen(ifd, "w");
if (*stdinf == NULL) {
pr_err("Error opening file stream for fd %d: %m\n",
ifd);
* Forks a child and executes a command to run on parallel
*/
-#define max(a,b) (a) < (b) ? (b) : (a)
#define BUF_SIZE (128*1024)
int run(const char *cmd, char *const argv[])
{
fd_set rfds;
int maxfd;
int eof = 0;
- char stdoutstr[32], stderrstr[32], indent[16] = { " " };
-
- indent[get_parent_count() + 1] = 0;
-
- if ((child = do_fork()))
- return child;
child = run_piped(cmd, argv, NULL, &ofd, &efd);
- snprintf(stdoutstr, 32, "%sstdout", green_color);
- snprintf(stderrstr, 32, "%sstderr", red_color);
FD_ZERO(&rfds);
FD_SET(ofd, &rfds);
char *sptr , *eptr;
char rbuf[BUF_SIZE];
int bytes;
- char *typestr;
+ int is_stderr = 0;
maxfd = max(ofd, efd);
error = select(maxfd, &rfds, NULL, NULL, NULL);
}
if (FD_ISSET(ofd, &rfds)) {
- typestr = stdoutstr;
bytes = read(ofd, rbuf, BUF_SIZE);
goto print;
}
if (FD_ISSET(efd, &rfds)) {
- typestr = stderrstr;
+ is_stderr = 1;
bytes = read(efd, rbuf, BUF_SIZE);
goto print;
}
*/
if (bytes == 0) {
bytes = read(efd, rbuf, BUF_SIZE);
- typestr = stderrstr;
+ is_stderr = 1;
eof = 1;
}
sptr = eptr = rbuf;
- while(bytes--) {
+ while (bytes--) {
if (*eptr == '\n') {
*eptr = 0;
- fprintf(stderr, "%s[%5d %s] %s: %s%s\n", indent,
- child, cmd, typestr, sptr, normal_color);
- sptr = eptr;
+ if (is_stderr)
+ pr_err("%s: stderr: %s\n",
+ cmd, sptr);
+ else
+ pr_info("%s: stdout: %s\n",
+ cmd, sptr);
+ sptr = eptr + 1;
}
eptr++;
}
close(ofd);
close(efd);
- harvest_zombies(child);
+ clear_zombie(child);
+
+ return 0;
+}
+
+int register_event_handler(struct event_handler *handler, int op)
+{
+ struct epoll_event ev;
+ int ret;
+ const char *str;
+
+ if (handler->fd <= 0) {
+ pr_err("Invalid file descriptor of %d\n", handler->fd);
+ return -1;
+ }
+
+ if (!handler->handle_event) {
+ pr_err("Handler callback missing\n");
+ return -1;
+ }
+
+ switch (op) {
+ case EPOLL_CTL_ADD:
+ str = "register";
+ break;
+
+ case EPOLL_CTL_MOD:
+ str = "modify";
+ break;
+
+ case EPOLL_CTL_DEL:
+ str = "deregister";
+ break;
+
+ default:
+ pr_err("Invalid op %d\n", op);
+ return -1;
+ }
+
+ pr_info("Doing a epoll %s for handler %s, fd %d\n", str,
+ handler->name, handler->fd);
+
+ ev.data.fd = handler->fd;
+ ev.data.ptr = handler;
+ ev.events = handler->events;
+ ret = epoll_ctl(epoll_fd, op, handler->fd, &ev);
+ if (ret) {
+ pr_err("Failed to do epoll_ctl %s: %m\n", str);
+ return -1;
+ }
+
+ return 0;
+}
+
+void _mutex_lock_acquired(struct mutex *lock, char *file, int line)
+{
+ lock->line = line;
+ lock->file = file;
+ pthread_getname_np(pthread_self(),
+ lock->owner_name, sizeof(lock->owner_name));
+}
+
+int _mutex_lock(struct mutex *lock, char *file, int line)
+{
+ int ret = 0;
+ int contended = 0;
+
+ if (!pthread_mutex_trylock(&lock->lock))
+ goto out_lock;
+
+ contended = 1;
+ pr_debug("Lock contention at %s:%d on lock %s acquired by %s at %s:%d\n",
+ file, line, lock->name,
+ lock->owner_name, lock->file, lock->line);
+
+ ret = pthread_mutex_lock(&lock->lock);
+ if (ret)
+ pr_err("Acquirin lock %s failed: %m, acquired on %s:%d\n",
+ lock->name, lock->file, lock->line);
+
+out_lock:
+ if (contended)
+ pr_debug("Lock %s acquired at %s:%d after contention\n",
+ lock->name, file, line);
+ _mutex_lock_acquired(lock, file, line);
+ return ret;
+}
+
+int _mutex_unlock(struct mutex *lock)
+{
+ lock->line = 0;
+ lock->file = NULL;
+ pthread_mutex_unlock(&lock->lock);
- exit(1);
return 0;
}