// $Id$
//
-// Copyright (C) 2006
-// Fraunhofer Institut fuer offene Kommunikationssysteme (FOKUS)
-// Kompetenzzentrum fuer Satelitenkommunikation (SatCom)
-// Stefan Bund <stefan.bund@fokus.fraunhofer.de>
+// Copyright (C) 2006
+// Fraunhofer Institute for Open Communication Systems (FOKUS)
+// Competence Center NETwork research (NET), St. Augustin, GERMANY
+// Stefan Bund <g0dil@berlios.de>
//
// 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
// 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
/** \file
-
- \idea Implement signal handling (See source for more discussion
- about this)
+ \brief Scheduler non-inline non-template implementation
\idea Multithreading support: To support multithreading, the
static member Scheduler::instance() must return a thread-local
value (that is Scheduler::instance() must allocate one Scheduler
- instance per thread)
-
- \fixme Test2
+ instance per thread). Another possibility would be to distribute
+ the async load unto several threads (one scheduler for multiple
+ threads)
*/
// Here a basic concept of how to add signal support to the scheduler:
//
-// Every signal to be reported by the scheduler will be asigned a
-// generic signal handler by the scheduler. This signal handler will
-// use longjmp (juck) to report this signal back to the scheduler
-// main-loop.
-//
-// To make this safe, the main-loop will look something like:
-//
-// int signal = setjmp(jmpBuffer_);
-// if (signal == 0) {
-// // unblock all signals which are registered with the
-// // scheduler
-// // call epoll
-// // block all relevant signals again
-// }
-//
-// // now handle the event
-//
-// The signal handler is then simply defined as
-//
-// static void Scheduler::sigHandler(int signal)
-// {
-// // make sure to restore the signal handler here if
-// // necessary
-// longjmp(Scheduler::instance().jmpBuffer_,signal);
-// }
-//
-// You should use sigaction to register the signal handlers and define
-// a sa_mask so all Scheduler-registered signals are automatically
-// *blocked* whenever one of the signals is called (including the
-// called signal!). This ensures, that no two signals can be delivered
-// on top of each other. And of course any signal registered with the
-// scheduler must be blocked as soon as it is registered with the
-// scheduler.
-
-// Definition of non-inline non-template functions
+// ... no, I had overlooked one race condition. So back to the signal-pipe approach ...
#include "Scheduler.hh"
//#include "Scheduler.ih"
// Custom includes
+#include "../Utils/senfassert.hh"
#include <errno.h>
#include <sys/epoll.h>
-#include "Utils/Exception.hh"
-#include "Utils/MicroTime.hh"
+#include <unistd.h>
+#include <fcntl.h>
+#include "../Utils/Exception.hh"
static const int EPollInitialSize = 16;
#define prefix_
///////////////////////////////cc.p////////////////////////////////////////
-prefix_ senf::Scheduler::Scheduler & senf::Scheduler::instance()
+prefix_ senf::Scheduler::Scheduler()
+ : files_(0), timerIdCounter_(0), epollFd_ (epoll_create(EPollInitialSize)), terminate_(false),
+ eventTime_(0), eventEarly_(ClockService::milliseconds(11)), eventAdjust_(0)
{
- static Scheduler instance;
- return instance;
+ if (epollFd_<0)
+ throwErrno();
+
+ if (::pipe(sigpipe_) < 0)
+ throwErrno();
+
+ int flags (::fcntl(sigpipe_[1],F_GETFL));
+ if (flags < 0)
+ throwErrno();
+ flags |= O_NONBLOCK;
+ if (::fcntl(sigpipe_[1], F_SETFL, flags) < 0)
+ throwErrno();
+
+ ::epoll_event ev;
+ ::memset(&ev, 0, sizeof(ev));
+ ev.events = EV_READ;
+ ev.data.fd = sigpipe_[0];
+ if (::epoll_ctl(epollFd_, EPOLL_CTL_ADD, sigpipe_[0], &ev) < 0)
+ throwErrno();
}
-prefix_ void senf::Scheduler::timeout(unsigned long timeout, TimerCallback const & cb)
+prefix_ void senf::Scheduler::registerSignal(unsigned signal, SimpleCallback const & cb)
{
- timerQueue_.push(TimerSpec(now()+1000*timeout,cb));
+ ::sigset_t sig;
+ ::sigemptyset(&sig);
+ if (::sigaddset(&sig, signal) < 0)
+ throw InvalidSignalNumberException();
+ ::sigprocmask(SIG_BLOCK, &sig, 0);
+ ::sigaddset(&sigset_, signal);
+ if (sigHandlers_.size() <= signal)
+ sigHandlers_.resize(signal+1);
+ sigHandlers_[signal] = cb;
+
+ registerSigHandlers();
}
-prefix_ senf::Scheduler::Scheduler()
- : epollFd_(epoll_create(EPollInitialSize))
+prefix_ void senf::Scheduler::unregisterSignal(unsigned signal)
{
- if (epollFd_<0)
- throw SystemException(errno);
+ if (::sigdelset(&sigset_, signal) < 0)
+ throw InvalidSignalNumberException();
+ sigHandlers_[signal] = 0;
+ ::signal(signal, SIG_DFL);
+ registerSigHandlers();
}
-prefix_ void senf::Scheduler::do_add(int fd, SimpleCallback const & cb, int eventMask)
+prefix_ void senf::Scheduler::do_add(int fd, FdCallback const & cb, int eventMask)
{
+ if (eventMask == 0)
+ return;
+
FdTable::iterator i (fdTable_.find(fd));
int action (EPOLL_CTL_MOD);
if (i == fdTable_.end()) {
action = EPOLL_CTL_ADD;
i = fdTable_.insert(std::make_pair(fd, EventSpec())).first;
}
+ if (i->second.epollMask() == 0) {
+ action = EPOLL_CTL_ADD;
+ fdErase_.erase( std::remove(fdErase_.begin(), fdErase_.end(), unsigned(fd)),
+ fdErase_.end() );
+ }
if (eventMask & EV_READ) i->second.cb_read = cb;
if (eventMask & EV_PRIO) i->second.cb_prio = cb;
if (eventMask & EV_WRITE) i->second.cb_write = cb;
- if (eventMask & EV_HUP) i->second.cb_hup = cb;
- if (eventMask & EV_ERR) i->second.cb_err = cb;
epoll_event ev;
memset(&ev,0,sizeof(ev));
ev.events = i->second.epollMask();
ev.data.fd = fd;
-
- if (epoll_ctl(epollFd_, action, fd, &ev)<0)
- throw SystemException(errno);
+
+ if (! i->second.file && epoll_ctl(epollFd_, action, fd, &ev) < 0) {
+ if (errno == EPERM) {
+ // Argh ... epoll does not support ordinary files :-( :-(
+ i->second.file = true;
+ ++ files_;
+ }
+ else
+ throwErrno("::epoll_ctl()");
+ }
}
prefix_ void senf::Scheduler::do_remove(int fd, int eventMask)
{
+ if (eventMask == 0)
+ return;
+
FdTable::iterator i (fdTable_.find(fd));
- if (i == fdTable_.end())
+ if (i == fdTable_.end())
return;
if (eventMask & EV_READ) i->second.cb_read = 0;
if (eventMask & EV_PRIO) i->second.cb_prio = 0;
if (eventMask & EV_WRITE) i->second.cb_write = 0;
- if (eventMask & EV_HUP) i->second.cb_hup = 0;
- if (eventMask & EV_ERR) i->second.cb_err = 0;
epoll_event ev;
memset(&ev,0,sizeof(ev));
ev.events = i->second.epollMask();
ev.data.fd = fd;
-
+
int action (EPOLL_CTL_MOD);
+ bool file (i->second.file);
if (ev.events==0) {
action = EPOLL_CTL_DEL;
- fdTable_.erase(i);
+ fdErase_.push_back(fd);
}
- if (epoll_ctl(epollFd_, action, fd, &ev)<0)
- throw SystemException(errno);
+ if (! file && epoll_ctl(epollFd_, action, fd, &ev) < 0)
+ throwErrno("::epoll_ctl()");
+ if (file)
+ -- files_;
+}
+
+prefix_ void senf::Scheduler::registerSigHandlers()
+{
+ for (unsigned signal (1); signal < sigHandlers_.size(); ++signal) {
+ if (sigHandlers_[signal]) {
+ struct ::sigaction sa;
+ sa.sa_sigaction = & Scheduler::sigHandler;
+ sa.sa_mask = sigset_;
+ sa.sa_flags = SA_SIGINFO;
+ if (signal == SIGCHLD)
+ sa.sa_flags |= SA_NOCLDSTOP;
+ if (::sigaction(signal, &sa, 0) < 0)
+ throwErrno();
+ }
+ }
}
+prefix_ void senf::Scheduler::sigHandler(int signal, ::siginfo_t * siginfo, void *)
+{
+ // This is a bit unsafe. Better write single bytes and place the siginfo into an explicit
+ // queue. Since signals are only unblocked during epoll_wait, we even wouldn't need to
+ // synchronize access to that queue any further.
+
+ ::write(instance().sigpipe_[1], siginfo, sizeof(*siginfo));
+
+ // We ignore errors. The file handle is set to non-blocking IO. If any failure occurs (pipe
+ // full), the signal will be dropped. That's like kernel signal handling which may also drop
+ // signals.
+}
prefix_ int senf::Scheduler::EventSpec::epollMask()
const
if (cb_read) mask |= EPOLLIN;
if (cb_prio) mask |= EPOLLPRI;
if (cb_write) mask |= EPOLLOUT;
- if (cb_hup) mask |= EPOLLHUP;
- if (cb_err) mask |= EPOLLERR;
return mask;
}
prefix_ void senf::Scheduler::process()
{
terminate_ = false;
+ eventTime_ = ClockService::now();
while (! terminate_) {
- MicroTime timeNow = now();
- while ( ! timerQueue_.empty() && timerQueue_.top().timeout <= timeNow ) {
- timerQueue_.top().cb();
- timerQueue_.pop();
- }
- if (terminate_)
- return;
- int timeout = timerQueue_.empty() ? -1 : int((timerQueue_.top().timeout - timeNow)/1000);
-
- struct epoll_event ev;
- int events = epoll_wait(epollFd_, &ev, 1, timeout);
- if (events<0)
- // Hmm ... man epoll says, it will NOT return with EINTR ??
- throw SystemException(errno);
- if (events==0)
- // Timeout .. it will be run when reachiung the top of the loop
- continue;
-
- FdTable::iterator i = fdTable_.find(ev.data.fd);
- BOOST_ASSERT (i != fdTable_.end() );
- EventSpec const & spec (i->second);
+ // Since a callback may have disabled further timers, we need to check for canceled timeouts
+ // again.
- if (ev.events & EPOLLIN) {
- BOOST_ASSERT(spec.cb_read);
- spec.cb_read(EV_READ);
- }
- else if (ev.events & EPOLLPRI) {
- BOOST_ASSERT(spec.cb_prio);
- spec.cb_prio(EV_PRIO);
+ while (! timerQueue_.empty()) {
+ TimerMap::iterator i (timerQueue_.top());
+ if (! i->second.canceled)
+ break;
+ timerMap_.erase(i);
+ timerQueue_.pop();
}
- else if (ev.events & EPOLLOUT) {
- BOOST_ASSERT(spec.cb_write);
- spec.cb_write(EV_WRITE);
+
+ for (FdEraseList::iterator i (fdErase_.begin()); i != fdErase_.end(); ++i)
+ fdTable_.erase(*i);
+ fdErase_.clear();
+
+ int timeout (-1);
+ if (files_ > 0)
+ timeout = 0;
+ else {
+ if (timerQueue_.empty()) {
+ if (fdTable_.empty())
+ break;
+ }
+ else {
+ ClockService::clock_type delta (
+ (timerQueue_.top()->second.timeout - eventTime_ + eventAdjust_)/1000000UL);
+ timeout = delta < 0 ? 0 : delta;
+ }
}
- else if (ev.events & EPOLLHUP) {
- if (spec.cb_hup)
- spec.cb_hup(EV_HUP);
- else if (ev.events & EPOLLERR) {
- if (spec.cb_write) spec.cb_write(EV_HUP);
- if (spec.cb_read) spec.cb_read(EV_HUP);
- }
+ ///\todo Handle more than one epoll_event per call
+ struct epoll_event ev;
+
+ ::sigprocmask(SIG_UNBLOCK, &sigset_, 0);
+ int events (epoll_wait(epollFd_, &ev, 1, timeout));
+ ::sigprocmask(SIG_BLOCK, &sigset_, 0);
+
+ if (events<0)
+ if (errno != EINTR)
+ throwErrno();
+
+ eventTime_ = ClockService::now();
+
+ // We always run timeout handlers. This is important, even if a file-descriptor is signaled
+ // since some descriptors (e.g. real files) will *always* be ready and we still may want to
+ // handle timers. Time handlers are run before file events to not delay them unnecessarily.
+
+ while (! timerQueue_.empty()) {
+ TimerMap::iterator i (timerQueue_.top());
+ if (i->second.canceled)
+ ;
+ else if (i->second.timeout <= eventTime_ + eventEarly_)
+ i->second.cb();
+ else
+ break;
+ timerQueue_.pop();
+ timerMap_.erase(i);
}
- else if (ev.events & EPOLLERR && ! ev.events & EPOLLHUP) {
- if (spec.cb_err)
- spec.cb_err(EV_ERR);
- else {
- if (spec.cb_write) spec.cb_write(EV_ERR);
- if (spec.cb_read) spec.cb_read(EV_ERR);
- }
+
+ // Check the signal queue
+ if (events > 0 && ev.data.fd == sigpipe_[0]) {
+ ::siginfo_t siginfo;
+ if (::read(sigpipe_[0], &siginfo, sizeof(siginfo)) < int(sizeof(siginfo))) {
+ // We ignore truncated records which may only occur if the signal
+ // queue became filled up
+ SENF_LOG((senf::log::IMPORTANT)("Truncated signal record!"));
+ continue;
+ }
+ if (siginfo.si_signo < int(sigHandlers_.size()) && sigHandlers_[siginfo.si_signo])
+ sigHandlers_[siginfo.si_signo]();
+ continue;
}
+ for (FdTable::iterator i = fdTable_.begin(); i != fdTable_.end(); ++i) {
+ EventSpec & spec (i->second);
+
+ if (! (spec.file || (events > 0 && i->first == ev.data.fd)))
+ continue;
+
+ unsigned extraFlags (0);
+ unsigned mask (spec.file ? spec.epollMask() : ev.events);
+
+ if (mask & EPOLLHUP) extraFlags |= EV_HUP;
+ if (mask & EPOLLERR) extraFlags |= EV_ERR;
+
+ if (mask & EPOLLIN) {
+ SENF_ASSERT(spec.cb_read);
+ spec.cb_read(EventId(EV_READ | extraFlags));
+ }
+ else if (mask & EPOLLPRI) {
+ SENF_ASSERT(spec.cb_prio);
+ spec.cb_prio(EventId(EV_PRIO | extraFlags));
+ }
+ else if (mask & EPOLLOUT) {
+ SENF_ASSERT(spec.cb_write);
+ spec.cb_write(EventId(EV_WRITE | extraFlags));
+ }
+ else {
+ // This branch is only taken, if HUP or ERR is signaled but none of IN/OUT/PRI.
+ // In this case we will signal all registered callbacks. The callbacks must be
+ // prepared to be called multiple times if they are registered to more than
+ // one event.
+ if (spec.cb_write)
+ spec.cb_write(EventId(extraFlags));
+ if (spec.cb_prio)
+ spec.cb_prio(EventId(extraFlags));
+ if (spec.cb_read)
+ spec.cb_read(EventId(extraFlags));
+ }
+ }
}
}
+///////////////////////////////////////////////////////////////////////////
+// senf::SchedulerLogTimeSource
+
+prefix_ boost::posix_time::ptime senf::SchedulerLogTimeSource::operator()()
+ const
+{
+ return ClockService::abstime(Scheduler::instance().eventTime());
+}
+
///////////////////////////////cc.e////////////////////////////////////////
#undef prefix_
\f
// Local Variables:
// mode: c++
+// fill-column: 100
// c-file-style: "senf"
+// indent-tabs-mode: nil
+// ispell-local-dictionary: "american"
+// compile-command: "scons -u test"
+// comment-column: 40
// End: