/* Copyright (C) 2012 J.F.Dockes
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the
* Free Software Foundation, Inc.,
* 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*/
#ifndef _WORKQUEUE_H_INCLUDED_
#define _WORKQUEUE_H_INCLUDED_
#include <pthread.h>
#include <time.h>
#include <string>
#include <queue>
#include "unordered_defs.h"
using std::queue;
using std::string;
#include "debuglog.h"
#include "ptmutex.h"
/// Store per-worker-thread data. Just an initialized timespec, and
/// used at the moment.
class WQTData {
public:
WQTData() {wstart.tv_sec = 0; wstart.tv_nsec = 0;}
struct timespec wstart;
};
/**
* A WorkQueue manages the synchronisation around a queue of work items,
* where a number of client threads queue tasks and a number of worker
* threads take and execute them. The goal is to introduce some level
* of parallelism between the successive steps of a previously single
* threaded pipeline. For example data extraction / data preparation / index
* update, but this could have other uses.
*
* There is no individual task status return. In case of fatal error,
* the client or worker sets an end condition on the queue. A second
* queue could conceivably be used for returning individual task
* status.
*/
template <class T> class WorkQueue {
public:
/** Create a WorkQueue
* @param name for message printing
* @param hi number of tasks on queue before clients blocks. Default 0
* meaning no limit. hi == -1 means that the queue is disabled.
* @param lo minimum count of tasks before worker starts. Default 1.
*/
WorkQueue(const string& name, size_t hi = 0, size_t lo = 1)
: m_name(name), m_high(hi), m_low(lo),
m_workers_exited(0), m_clients_waiting(0), m_workers_waiting(0),
m_tottasks(0), m_nowake(0), m_workersleeps(0), m_clientsleeps(0)
{
m_ok = (pthread_cond_init(&m_ccond, 0) == 0) &&
(pthread_cond_init(&m_wcond, 0) == 0);
}
~WorkQueue()
{
LOGDEB2(("WorkQueue::~WorkQueue:%s\n", m_name.c_str()));
if (!m_worker_threads.empty())
setTerminateAndWait();
}
/** Start the worker threads.
*
* @param nworkers number of threads copies to start.
* @param start_routine thread function. It should loop
* taking (QueueWorker::take()) and executing tasks.
* @param arg initial parameter to thread function.
* @return true if ok.
*/
bool start(int nworkers, void *(*start_routine)(void *), void *arg)
{
PTMutexLocker lock(m_mutex);
for (int i = 0; i < nworkers; i++) {
int err;
pthread_t thr;
if ((err = pthread_create(&thr, 0, start_routine, arg))) {
LOGERR(("WorkQueue:%s: pthread_create failed, err %d\n",
m_name.c_str(), err));
return false;
}
m_worker_threads.insert(pair<pthread_t, WQTData>(thr, WQTData()));
}
return true;
}
/** Add item to work queue, called from client.
*
* Sleeps if there are already too many.
*/
bool put(T t)
{
PTMutexLocker lock(m_mutex);
if (!lock.ok() || !ok()) {
LOGERR(("WorkQueue::put:%s: !ok or mutex_lock failed\n",
m_name.c_str()));
return false;
}
while (ok() && m_high > 0 && m_queue.size() >= m_high) {
m_clientsleeps++;
// Keep the order: we test ok() AFTER the sleep...
m_clients_waiting++;
if (pthread_cond_wait(&m_ccond, lock.getMutex()) || !ok()) {
m_clients_waiting--;
return false;
}
m_clients_waiting--;
}
m_queue.push(t);
if (m_workers_waiting > 0) {
// Just wake one worker, there is only one new task.
pthread_cond_signal(&m_wcond);
} else {
m_nowake++;
}
return true;
}
/** Wait until the queue is inactive. Called from client.
*
* Waits until the task queue is empty and the workers are all
* back sleeping. Used by the client to wait for all current work
* to be completed, when it needs to perform work that couldn't be
* done in parallel with the worker's tasks, or before shutting
* down. Work can be resumed after calling this. Note that the
* only thread which can call it safely is the client just above
* (which can control the task flow), else there could be
* tasks in the intermediate queues.
* To rephrase: there is no warranty on return that the queue is actually
* idle EXCEPT if the caller knows that no jobs are still being created.
* It would be possible to transform this into a safe call if some kind
* of suspend condition was set on the queue by waitIdle(), to be reset by
* some kind of "resume" call. Not currently the case.
*/
bool waitIdle()
{
PTMutexLocker lock(m_mutex);
if (!lock.ok() || !ok()) {
LOGERR(("WorkQueue::waitIdle:%s: not ok or can't lock\n",
m_name.c_str()));
return false;
}
// We're done when the queue is empty AND all workers are back
// waiting for a task.
while (ok() && (m_queue.size() > 0 ||
m_workers_waiting != m_worker_threads.size())) {
m_clients_waiting++;
if (pthread_cond_wait(&m_ccond, lock.getMutex())) {
m_clients_waiting--;
m_ok = false;
LOGERR(("WorkQueue::waitIdle:%s: cond_wait failed\n",
m_name.c_str()));
return false;
}
m_clients_waiting--;
}
return ok();
}
/** Tell the workers to exit, and wait for them.
*
* Does not bother about tasks possibly remaining on the queue, so
* should be called after waitIdle() for an orderly shutdown.
*/
void* setTerminateAndWait()
{
PTMutexLocker lock(m_mutex);
LOGDEB(("setTerminateAndWait:%s\n", m_name.c_str()));
if (m_worker_threads.empty()) {
// Already called ?
return (void*)0;
}
// Wait for all worker threads to have called workerExit()
m_ok = false;
while (m_workers_exited < m_worker_threads.size()) {
pthread_cond_broadcast(&m_wcond);
m_clients_waiting++;
if (pthread_cond_wait(&m_ccond, lock.getMutex())) {
LOGERR(("WorkQueue::setTerminate:%s: cond_wait failed\n",
m_name.c_str()));
m_clients_waiting--;
return (void*)0;
}
m_clients_waiting--;
}
LOGINFO(("%s: tasks %u nowakes %u wsleeps %u csleeps %u\n",
m_name.c_str(), m_tottasks, m_nowake, m_workersleeps,
m_clientsleeps));
// Perform the thread joins and compute overall status
// Workers return (void*)1 if ok
void *statusall = (void*)1;
STD_UNORDERED_MAP<pthread_t, WQTData>::iterator it;
while (!m_worker_threads.empty()) {
void *status;
it = m_worker_threads.begin();
pthread_join(it->first, &status);
if (status == (void *)0)
statusall = status;
m_worker_threads.erase(it);
}
// Reset to start state.
m_workers_exited = m_clients_waiting = m_workers_waiting =
m_tottasks = m_nowake = m_workersleeps = m_clientsleeps = 0;
m_ok = true;
LOGDEB(("setTerminateAndWait:%s done\n", m_name.c_str()));
return statusall;
}
/** Take task from queue. Called from worker.
*
* Sleeps if there are not enough. Signal if we go to sleep on empty
* queue: client may be waiting for our going idle.
*/
bool take(T* tp, size_t *szp = 0)
{
PTMutexLocker lock(m_mutex);
if (!lock.ok() || !ok()) {
LOGDEB(("WorkQueue::take:%s: not ok\n", m_name.c_str()));
return false;
}
while (ok() && m_queue.size() < m_low) {
m_workersleeps++;
m_workers_waiting++;
if (m_queue.empty())
pthread_cond_broadcast(&m_ccond);
if (pthread_cond_wait(&m_wcond, lock.getMutex()) || !ok()) {
// !ok is a normal condition when shutting down
if (ok())
LOGERR(("WorkQueue::take:%s: cond_wait failed or !ok\n",
m_name.c_str()));
m_workers_waiting--;
return false;
}
m_workers_waiting--;
}
m_tottasks++;
*tp = m_queue.front();
if (szp)
*szp = m_queue.size();
m_queue.pop();
if (m_clients_waiting > 0) {
// No reason to wake up more than one client thread
pthread_cond_signal(&m_ccond);
} else {
m_nowake++;
}
return true;
}
/** Advertise exit and abort queue. Called from worker
*
* This would happen after an unrecoverable error, or when
* the queue is terminated by the client. Workers never exit normally,
* except when the queue is shut down (at which point m_ok is set to
* false by the shutdown code anyway). The thread must return/exit
* immediately after calling this.
*/
void workerExit()
{
LOGDEB(("workerExit:%s\n", m_name.c_str()));
PTMutexLocker lock(m_mutex);
m_workers_exited++;
m_ok = false;
pthread_cond_broadcast(&m_ccond);
}
size_t qsize()
{
PTMutexLocker lock(m_mutex);
size_t sz = m_queue.size();
return sz;
}
private:
bool ok()
{
bool isok = m_ok && m_workers_exited == 0 && !m_worker_threads.empty();
if (!isok) {
LOGDEB(("WorkQueue:ok:%s: not ok m_ok %d m_workers_exited %d "
"m_worker_threads size %d\n", m_name.c_str(),
m_ok, m_workers_exited, int(m_worker_threads.size())));
}
return isok;
}
long long nanodiff(const struct timespec& older,
const struct timespec& newer)
{
return (newer.tv_sec - older.tv_sec) * 1000000000LL
+ newer.tv_nsec - older.tv_nsec;
}
// Configuration
string m_name;
size_t m_high;
size_t m_low;
// Status
// Worker threads having called exit
unsigned int m_workers_exited;
bool m_ok;
// Per-thread data. The data is not used currently, this could be
// a set<pthread_t>
STD_UNORDERED_MAP<pthread_t, WQTData> m_worker_threads;
// Synchronization
queue<T> m_queue;
pthread_cond_t m_ccond;
pthread_cond_t m_wcond;
PTMutexInit m_mutex;
// Client/Worker threads currently waiting for a job
unsigned int m_clients_waiting;
unsigned int m_workers_waiting;
// Statistics
unsigned int m_tottasks;
unsigned int m_nowake;
unsigned int m_workersleeps;
unsigned int m_clientsleeps;
};
#endif /* _WORKQUEUE_H_INCLUDED_ */