\brief Scheduler public header
*/
-#ifndef HH_Scheduler_
-#define HH_Scheduler_ 1
+#ifndef HH_SENF_Scheduler_Scheduler_
+#define HH_SENF_Scheduler_Scheduler_ 1
// Custom includes
+#include <boost/utility.hpp>
#include "../Utils/Logger/SenfLog.hh"
#include "FdEvent.hh"
#include "TimerEvent.hh"
#include "SignalEvent.hh"
+#include "EventHook.hh"
//#include "scheduler.mpp"
///////////////////////////////hh.p////////////////////////////////////////
The %scheduler API is comprised of two parts:
- \li Specific event classes, one for each type of event.
- \li Some generic functions implemented in the \ref senf::scheduler namespace.
+ \li Specific \ref sched_objects, one for each type of event.
+ \li Some <a href="#autotoc-7.">generic functions</a> implemented in the \ref senf::scheduler
+ namespace.
Events are registered via the respective event class. The (global) functions are used to enter
the application main-loop or query for global information.
+ \autotoc
+
+
+ \section sched_objects Event classes
+
+ The Scheduler is based on the RAII principle: Every event is represented by a class
+ instance. The event is registered in the constructor and removed by the destructor of that
+ instance. This implementation automatically links the lifetime of an event with the lifetime of
+ the object resposible for it's creation.
+
+ Every event registration is represented by an instance of an event specific class:
+
+ \li senf::scheduler::FdEvent for file descriptor events
+ \li senf::scheduler::TimerEvent for single-shot deadline timer events
+ \li senf::scheduler::SignalEvent for UNIX signal events
+ \li senf::scheduler::EventHook for a special event hook
+
+ These instance are owned and managed by the user of the scheduler \e not by the scheduler so the
+ RAII concept can be used.
+
+ \code
+ class SomeServer
+ {
+ SomeSocketHandle handle_;
+ senf::scheduler::FdEvent event_;
+
+ public:
+ SomeServer(SomeSocketHandle handle)
+ : handle_ (handle),
+ event_ ("SomeServer handler", senf::membind(&SomeServer::readData, this),
+ handle, senf::scheduler::FdEvent::EV_READ)
+ {}
+
+ void readData(int events)
+ {
+ // read data from handle_, check for eof and so on.
+ }
+ };
+ \endcode
+
+ The event is defined as a class member variable. When the event member is initialized in the
+ constructor, the event is automatically registered (except if the optional \a initiallyEnabled
+ flag argument is set to \c false). The Destructor will automatically remove the event from the
+ scheduler and ensure, that no dead code is called accidentally.
+
+ The process is the same for the other event types or when registering multiple events. For
+ detailed information on the constructor arguments and other features see the event class
+ documentation referenced below.
+
\section sched_handlers Specifying handlers
messages to identify the failing handler.
- \section sched_fd Registering events
+ \section sched_exec Executing the Scheduler
+
+ To enter the scheduler main-loop, call
+
+ \code
+ senf::scheduler::process();
+ \endcode
+
+ This call will only return in two cases:
- Events are registered by allocating an instance of the corresponding event class:
+ \li When a handler calls senf::scheduler::terminate()
+ \li When there is no active file descriptor or timer event.
- \li senf::scheduler::FdEvent for file descriptor events
- \li senf::scheduler::TimerEvent for single-shot deadline timer events
- \li senf::scheduler::SignalEvent for UNIX signal events
+ Additional <a href="#autotoc-7.">generic functions</a> provide information and %scheduler
+ parameters.
+
+ \section sched_container Event objects and container classes
+
+ As the event objects are \e not copyable, they cannot be placed into ordinary
+ containers. However, it is quite simple to use pointer containers to hold event instances:
- The destructor of each of these classes will ensure, that the event will be properly
- unregistered. The registration can be enabled, disabled or changed using appropriate
- members. See the event class for details on a specific type of event.
+ \code
+ #include <boost/ptr_container/ptr_map.hpp>
+ #include <boost/bind.hpp>
+ class Foo
+ {
+ public:
+ void add(int fd)
+ {
+ fdEvents.insert(
+ fd,
+ new senf::scheduler::FdEvent("foo", boost::bind(&callback, this, fd, _1), fd,
+ senf::scheduler::FdEvent::EV_READ) );
+ }
+
+ void callback(int fd, int events)
+ {
+ FdEvent & event (fdEvents_[fd]);
+
+ // ...
+
+ if (complete)
+ fdEvents_.remove(fd)
+ }
+
+ private:
+ boost::ptr_map<int, FdEvent> fdEvents_;
+ };
+ \endcode
+
+ The pointer container API is (almost) completely identical to the corresponding standard library
+ container API. The only difference is, that all elements added to the container \e must be
+ created via \c new and that the pointer containers themselves are \e not copyable (ok, they are,
+ if the elements are cloneable ...). See <a
+ href="http://www.boost.org/doc/libs/1_36_0/libs/ptr_container/doc/ptr_container.html">Boost.PointerContainer</a>
+ for the pointer container library reference.
+
+
+ \section sched_signals Signals and the Watchdog
+
+ To secure against blocking callbacks, the %scheduler implementation includes a watchdog
+ timer. This timer will produce a warning message on the standard error stream when a single
+ callback is executing for more than the watchdog timeout value. Since the scheduler
+ implementation is completely single threaded, we cannot terminate the callback but at least we
+ can produce an informative message and optionally the program can be aborted.
+
+ The watchdog is controlled using the watchdogTimeout(), watchdogEvents() and watchdogAbort().
+ functions.
+
+ The watchdog is implemented using a free running interval timer. The watchdog signal (\c SIGURG)
+ must \e not be blocked. If signals need to be blocked for some reason, those regions will not be
+ checked by the watchdog. If a callback blocks, the watchdog has no chance to interrupt the
+ process.
+
+ \warning Since the watchdog is free running for performance reasons, every callback must expect
+ signals to happen. Signals \e will certainly happen since the watchdog signal is generated
+ periodically (which does not necessarily generate a watchdog event ...)
+
+ Additional signals (\c SIGALRM) may occur when using using hires timers on kernel/glibc
+ combinations which do not support timerfd(). On such systems, hires timers are implemented using
+ POSIX timers which generate a considerable number of additional signals.
+
\todo Fix the file support to use threads (?) fork (?) and a pipe so it works reliably even
over e.g. NFS.
*/
\li a callback calls terminate()
\li the run queue becomes empty.
*/
- void process();
+ void process();
+
+ /** \brief \c true, if scheduler is running, \c false otherwise */
+ bool running();
/** \brief Called by callbacks to terminate the main loop
*/
void terminate();
- /** \brief Return date/time of last event
+ /** \brief Immediately rescheduler
+
+ Calling yield() will cause the scheduler to terminate the current queue run and immediately
+ rescheduler all pending tasks.
+ */
+ void yield();
+
+ /** \brief Return timestamp of last event
This is the timestamp, the last event has been signaled. This is the real time at which the
event is delivered \e not the time it should have been delivered (in the case of timers).
*/
ClockService::clock_type eventTime();
- /** \brief Set task watchdog timeout */
- void taskTimeout(unsigned ms);
+ /** \brief Return (approximate) current time
+
+ This call will return the current time as far as it is already known to the scheduler. If
+ the scheduler is running, this will return eventTime(), otherwise it will return
+ ClockService::now(). While the scheduler is running, this will reduce the number of system
+ calls.
+ */
+ ClockService::clock_type now();
- /** \brief Current task watchdog timeout */
- unsigned taskTimeout();
+ /** \brief Set watchdog timeout to \a ms milliseconds.
+
+ Setting the watchdog timeout to 0 will disable the watchdog.
+ */
+ void watchdogTimeout(unsigned ms);
- /** \brief Number of watchdog events */
- unsigned hangCount();
+ /** \brief Current watchdog timeout in milliseconds */
+ unsigned watchdogTimeout();
+
+ /** \brief Number of watchdog events
+
+ calling watchtogEvents() will reset the counter to 0
+ */
+ unsigned watchdogEvents();
+
+ /** \brief Enable/disable abort on watchdog event.
+
+ Calling watchdogAbort(\c true) will enable aborting the program execution on a watchdog
+ event.
+ */
+ void watchdogAbort(bool flag);
+
+ /** \brief Get current watchdog abort on event status */
+ bool watchdogAbort();
+
+ /** \brief Switch to using hi resolution timers
+
+ By default, timers are implemented directly using epoll. This however restricts the timer
+ resolution to that of the kernel HZ value.
+
+ High resolution timers are implemented either using POSIX timers or, when available, using
+ the Linux special \c timerfd() syscall.
+
+ POSIX timers are delivered using signals. A high timer load this increases the signal load
+ considerably. \c timerfd()'s are delivered on a file descriptor and thus don't have such a
+ scalability issue.
+
+ \warning The timer source must not be switched from a scheduler callback
+ */
+ void hiresTimers();
+
+ /** \brief Switch back to using epoll for timing
+ \see hiresTimers()
+ */
+ void loresTimers();
+
+ /** \brief return \c true, if \c timerfd() timing is available, \c false otherwise
+ \see hiresTimers()
+ */
+ bool haveScalableHiresTimers();
+
+ /** \brief Return \c true, if using hires times, \c false otherwise
+ \see hiresTimers() */
+ bool usingHiresTimers();
/** \brief Restart scheduler
*/
void restart();
+ /** \brief Return \c true, if no event is registered, \c false otherwise. */
+ bool empty();
+
/** \brief %scheduler specific time source for Utils/Logger framework
This time source may be used to provide timing information for log messages within the
Utils/Logger framework. This time source will use Scheduler::eventTime() to provide timing
information.
+ \code
+ senf::log::timeSource<senf::scheduler::LogTimeSource>();
+ \endcode
+
Using this information reduces the number of necessary ClockService::now() calls and thus
the number of system calls.
*/
senf::log::time_type operator()() const;
};
+ /** \brief Temporarily block all signals
+
+ This class is used to temporarily block all signals in a critical section.
+
+ \code
+ // Begin critical section
+ {
+ senf::scheduler::BlockSignals signalBlocker;
+
+ // critical code executed with all signals blocked
+ }
+ // End critical section
+ \endcode
+
+ You need to take care not to block since even the watchdog timer will be disabled while
+ executing within a critical section.
+ */
+ class BlockSignals
+ : boost::noncopyable
+ {
+ public:
+ BlockSignals(bool initiallyBlocked=true);
+ ///< Block signals until end of scope
+ /**< \param[in] initiallyBlocked set to \c false to not
+ automatically block signals initially */
+ ~BlockSignals(); ///< Release all signal blocks
+
+ void block(); ///< Block signals if not blocked
+ void unblock(); ///< Unblock signals if blocked
+ bool blocked() const; ///< \c true, if signals currently blocked, \c false
+ ///< otherwise
+
+ private:
+ bool blocked_;
+ sigset_t allSigs_;
+ sigset_t savedSigs_;
+ };
+
}}
///////////////////////////////hh.e////////////////////////////////////////
projects / senf.git / blobdiff