hcq/MaxScale
hcq/MaxScale
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
MaxScale/server/core/internal/worker.hh
Johan Wikman 6e9e83ccaf MXS-1674 Change load granularity to 1 second 
With a granularity of 1 second, the load will from a human
perspective reflect the current situation. That also means
that the maxadmin output shows "natural" steps; 1s, 1m and 1h.
2018-02-21 13:05:58 +02:00

909 lines
26 KiB
C++
Raw Blame History

#pragma once
/*
* Copyright (c) 2016 MariaDB Corporation Ab
*
* Use of this software is governed by the Business Source License included
* in the LICENSE.TXT file and at www.mariadb.com/bsl11.
*
* Change Date: 2020-01-01
*
* On the date above, in accordance with the Business Source License, use
* of this software will be governed by version 2 or later of the General
* Public License.
*/
#include <maxscale/cppdefs.hh>
#include <memory>
#include <vector>
#include <maxscale/platform.h>
#include <maxscale/session.h>
#include <maxscale/utils.hh>
#include "messagequeue.hh"
#include "poll.h"
#include "worker.h"
#include "workertask.hh"
#include "session.hh"
namespace maxscale
{
class Semaphore;
struct WORKER_STATISTICS
{
WORKER_STATISTICS()
{
memset(this, 0, sizeof(WORKER_STATISTICS));
}
enum
{
MAXNFDS = 10,
N_QUEUE_TIMES = 30
};
int64_t n_read; /*< Number of read events */
int64_t n_write; /*< Number of write events */
int64_t n_error; /*< Number of error events */
int64_t n_hup; /*< Number of hangup events */
int64_t n_accept; /*< Number of accept events */
int64_t n_polls; /*< Number of poll cycles */
int64_t n_pollev; /*< Number of polls returning events */
int64_t n_nbpollev; /*< Number of polls returning events */
int64_t n_fds[MAXNFDS]; /*< Number of wakeups with particular n_fds value */
int64_t evq_length; /*< Event queue length */
int64_t evq_max; /*< Maximum event queue length */
int64_t blockingpolls; /*< Number of epoll_waits with a timeout specified */
uint32_t qtimes[N_QUEUE_TIMES + 1];
uint32_t exectimes[N_QUEUE_TIMES + 1];
int64_t maxqtime;
int64_t maxexectime;
};
/**
* WorkerLoad is a class that calculates the load percentage of a worker
* thread, based upon the relative amount of time the worker spends in
* epoll_wait().
*
* If during a time period of length T milliseconds, the worker thread
* spends t milliseconds in epoll_wait(), then the load of the worker is
* calculated as 100 * ((T - t) / T). That is, if the worker spends all
* the time in epoll_wait(), then the load is 0 and if the worker spends
* no time waiting in epoll_wait(), then the load is 100.
*/
class WorkerLoad
{
WorkerLoad(const WorkerLoad&);
WorkerLoad& operator = (const WorkerLoad&);
public:
enum counter_t
{
ONE_SECOND = 1000,
ONE_MINUTE = 60 * ONE_SECOND,
ONE_HOUR = 60 * ONE_MINUTE,
};
enum
{
GRANULARITY = ONE_SECOND
};
/**
* Constructor
*/
WorkerLoad();
/**
* Reset the load calculation. Should be called immediately before the
* worker enters its eternal epoll_wait()-loop.
*/
void reset()
{
uint64_t now = get_time();
m_start_time = now;
m_wait_start = 0;
m_wait_time = 0;
}
/**
* To be used for signaling that the worker is about to call epoll_wait().
*
* @param now The current time.
*/
void about_to_wait(uint64_t now)
{
m_wait_start = now;
}
void about_to_wait()
{
about_to_wait(get_time());
}
/**
* To be used for signaling that the worker has returned from epoll_wait().
*
* @param now The current time.
*/
void about_to_work(uint64_t now);
void about_to_work()
{
about_to_work(get_time());
}
/**
* Returns the last calculated load,
*
* @return A value between 0 and 100.
*/
uint8_t percentage(counter_t counter) const
{
switch (counter)
{
case ONE_SECOND:
return m_load_1_second.value();
case ONE_MINUTE:
return m_load_1_minute.value();
case ONE_HOUR:
return m_load_1_hour.value();
default:
ss_dassert(!true);
return 0;
};
}
/**
* When was the last 1 second period started.
*
* @return The start time.
*/
uint64_t start_time() const
{
return m_start_time;
}
/**
* Returns the current time using CLOCK_MONOTONIC.
*
* @return Current time in milliseconds.
*/
static uint64_t get_time();
private:
/**
* Average is a base class for classes intended to be used for calculating
* averages. An Average may have a dependant Average whose value depends
* upon the value of the first. At certain moments, an Average may trigger
* its dependant Average to update itself.
*/
class Average
{
Average(const Average&);
Average& operator = (const Average&);
public:
/**
* Constructor
*
* @param pDependant An optional dependant average.
*/
Average(Average* pDependant = NULL)
: m_pDependant(pDependant)
, m_value(0)
{}
virtual ~Average();
/**
* Add a value to the Average. The exact meaning depends upon the
* concrete Average class.
*
* If the addition of the value in some sense represents a full cycle
* in the average calculation, then the instance will call add_value()
* on its dependant, otherwise it will call update_value(). In both cases
* with its own value as argument.
*
* @param value The value to be added.
*
* @return True if the addition of the value caused a full cycle
* in the average calculation, false otherwise.
*/
virtual bool add_value(uint8_t value) = 0;
/**
* Update the value of the Average. The exact meaning depends upon the
* concrete Average class. Will also call update_value() of its dependant
* with its own value as argument.
*
* @param value The value to be updated.
*/
virtual void update_value(uint8_t value) = 0;
/**
* Return the average value.
*
* @return The value represented by the Average.
*/
uint8_t value() const
{
return atomic_load_uint32(&m_value);
}
protected:
Average* m_pDependant; /*< The optional dependant Average. */
uint32_t m_value; /*< The current average value. */
protected:
void set_value(uint32_t value)
{
atomic_store_uint32(&m_value, value);
}
};
/**
* An Average consisting of a single value.
*/
class Average1 : public Average
{
public:
Average1(Average* pDependant = NULL)
: Average(pDependant)
{
}
bool add_value(uint8_t value)
{
set_value(value);
// Every addition of a value represents a full cycle.
if (m_pDependant)
{
m_pDependant->add_value(value);
}
return true;
}
void update_value(uint8_t value)
{
set_value(value);
if (m_pDependant)
{
m_pDependant->update_value(value);
}
}
};
/**
* An Average calculated from N values.
*/
template<size_t N>
class AverageN : public Average
{
public:
AverageN(Average* pDependant = NULL)
: Average(pDependant)
, m_end(m_begin + N)
, m_i(m_begin)
, m_sum(0)
, m_nValues(0)
{
}
bool add_value(uint8_t value)
{
if (m_nValues == N)
{
// If as many values that fit has been added, then remove the
// least recent value from the sum.
m_sum -= *m_i;
}
else
{
// Otherwise make a note that a new value is added.
++m_nValues;
}
*m_i = value;
m_sum += *m_i; // Update the sum of all values.
m_i = next(m_i);
uint32_t average = m_sum / m_nValues;
set_value(average);
if (m_pDependant)
{
if (m_i == m_begin)
{
// If we have looped around we have performed a full cycle and will
// add a new value to the dependant average.
m_pDependant->add_value(average);
}
else
{
// Otherwise we just update the most recent value.
m_pDependant->update_value(average);
}
}
return m_i == m_begin;
}
void update_value(uint8_t value)
{
if (m_nValues == 0)
{
// If no values have been added yet, there's nothing to update but we
// need to add the value.
add_value(value);
}
else
{
// Otherwise we update the most recent value.
uint8_t* p = prev(m_i);
m_sum -= *p;
*p = value;
m_sum += *p;
uint32_t average = m_sum / m_nValues;
set_value(average);
if (m_pDependant)
{
m_pDependant->update_value(average);
}
}
}
private:
uint8_t* prev(uint8_t* p)
{
ss_dassert(p >= m_begin);
ss_dassert(p < m_end);
if (p > m_begin)
{
--p;
}
else
{
ss_dassert(p == m_begin);
p = m_end - 1;
}
ss_dassert(p >= m_begin);
ss_dassert(p < m_end);
return p;
}
uint8_t* next(uint8_t* p)
{
ss_dassert(p >= m_begin);
ss_dassert(p < m_end);
++p;
if (p == m_end)
{
p = m_begin;
}
ss_dassert(p >= m_begin);
ss_dassert(p < m_end);
return p;
}
private:
uint8_t m_begin[N]; /*< Buffer containing values from which the average is calculated. */
uint8_t* m_end; /*< Points to one past the end of the buffer. */
uint8_t* m_i; /*< Current position in the buffer. */
uint32_t m_sum; /*< Sum of all values in the buffer. */
uint32_t m_nValues; /*< How many values the buffer contains. */
};
uint64_t m_start_time; /*< When was the current 1-second period started. */
uint64_t m_wait_start; /*< The time when the worker entered epoll_wait(). */
uint64_t m_wait_time; /*< How much time the worker has spent in epoll_wait(). */
AverageN<60> m_load_1_hour; /*< The average load during the last hour. */
AverageN<60> m_load_1_minute; /*< The average load during the last minute. */
Average1 m_load_1_second; /*< The load during the last 1-second period. */
};
class Worker : public MXS_WORKER
, private MessageQueue::Handler
, private MXS_POLL_DATA
{
Worker(const Worker&);
Worker& operator = (const Worker&);
public:
typedef WORKER_STATISTICS STATISTICS;
typedef WorkerTask Task;
typedef WorkerDisposableTask DisposableTask;
typedef Registry<MXS_SESSION> SessionsById;
typedef std::vector<DCB*> Zombies;
typedef WorkerLoad Load;
enum state_t
{
STOPPED,
IDLE,
POLLING,
PROCESSING,
ZPROCESSING
};
enum execute_mode_t
{
EXECUTE_AUTO, /**< Execute tasks immediately */
EXECUTE_QUEUED /**< Only queue tasks for execution */
};
/**
* Initialize the worker mechanism.
*
* To be called once at process startup. This will cause as many workers
* to be created as the number of threads defined.
*
* @return True if the initialization succeeded, false otherwise.
*/
static bool init();
/**
* Finalize the worker mechanism.
*
* To be called once at process shutdown. This will cause all workers
* to be destroyed. When the function is called, no worker should be
* running anymore.
*/
static void finish();
/**
* Returns the id of the worker
*
* @return The id of the worker.
*/
int id() const
{
return m_id;
}
int load(Load::counter_t counter)
{
return m_load.percentage(counter);
}
/**
* Returns the state of the worker.
*
* @return The current state.
*
* @attentions The state might have changed the moment after the function returns.
*/
state_t state() const
{
return m_state;
}
/**
* Returns statistics for this worker.
*
* @return The worker specific statistics.
*
* @attentions The statistics may change at any time.
*/
const STATISTICS& statistics() const
{
return m_statistics;
}
/**
* Returns statistics for all workers.
*
* @return Combined statistics.
*
* @attentions The statistics may no longer be accurate by the time it has
* been returned. The returned values may also not represent a
* 100% consistent set.
*/
static STATISTICS get_statistics();
/**
* Return a specific combined statistic value.
*
* @param what What to return.
*
* @return The corresponding value.
*/
static int64_t get_one_statistic(POLL_STAT what);
/**
* Return this worker's statistics.
*
* @return Local statistics for this worker.
*/
const STATISTICS& get_local_statistics() const
{
return m_statistics;
}
/**
* Return the count of descriptors.
*
* @param pnCurrent On output the current number of descriptors.
* @param pnTotal On output the total number of descriptors.
*/
void get_descriptor_counts(uint32_t* pnCurrent, uint64_t* pnTotal);
/**
* Add a file descriptor to the epoll instance of the worker.
*
* @param fd The file descriptor to be added.
* @param events Mask of epoll event types.
* @param pData The poll data associated with the descriptor:
*
* data->handler : Handler that knows how to deal with events
* for this particular type of 'struct mxs_poll_data'.
* data->thread.id: Will be updated by the worker.
*
* @attention The provided file descriptor must be non-blocking.
* @attention @c pData must remain valid until the file descriptor is
* removed from the worker.
*
* @return True, if the descriptor could be added, false otherwise.
*/
bool add_fd(int fd, uint32_t events, MXS_POLL_DATA* pData);
/**
* Add a file descriptor to the epoll instance shared between all workers.
* Events occuring on the provided file descriptor will be handled by all
* workers. This is primarily intended for listening sockets where the
* only event is EPOLLIN, signaling that accept() can be used on the listening
* socket for creating a connected socket to a client.
*
* @param fd The file descriptor to be added.
* @param events Mask of epoll event types.
* @param pData The poll data associated with the descriptor:
*
* data->handler : Handler that knows how to deal with events
* for this particular type of 'struct mxs_poll_data'.
* data->thread.id: 0
*
* @return True, if the descriptor could be added, false otherwise.
*/
static bool add_shared_fd(int fd, uint32_t events, MXS_POLL_DATA* pData);
/**
* Remove a file descriptor from the worker's epoll instance.
*
* @param fd The file descriptor to be removed.
*
* @return True on success, false on failure.
*/
bool remove_fd(int fd);
/**
* Remove a file descriptor from the epoll instance shared between all workers.
*
* @param fd The file descriptor to be removed.
*
* @return True on success, false on failure.
*/
static bool remove_shared_fd(int fd);
/**
* Register zombie for later deletion.
*
* @param pZombie DCB that will be deleted at end of event loop.
*
* @note The DCB must be owned by this worker.
*/
void register_zombie(DCB* pZombie);
/**
* Main function of worker.
*
* The worker will run the poll loop, until it is told to shut down.
*
* @attention This function will run in the calling thread.
*/
void run();
/**
* Run worker in separate thread.
*
* This function will start a new thread, in which the `run`
* function will be executed.
*
* @param stack_size The stack size of the new thread. A value of 0 means
* that the pthread default should be used.
*
* @return True if the thread could be started, false otherwise.
*/
bool start(size_t stack_size = 0);
/**
* Waits for the worker to finish.
*/
void join();
/**
* Initate shutdown of worker.
*
* @attention A call to this function will only initiate the shutdowm,
* the worker will not have shut down when the function returns.
*
* @attention This function is signal safe.
*/
void shutdown();
/**
* Query whether worker should shutdown.
*
* @return True, if the worker should shut down, false otherwise.
*/
bool should_shutdown() const
{
return m_should_shutdown;
}
/**
* Posts a task to a worker for execution.
*
* @param pTask The task to be executed.
* @param pSem If non-NULL, will be posted once the task's `execute` return.
* @param mode Execution mode
*
* @return True if the task could be posted (i.e. not executed), false otherwise.
*
* @attention The instance must remain valid for as long as it takes for the
* task to be transferred to the worker and its `execute` function
* to be called.
*
* The semaphore can be used for waiting for the task to be finished.
*
* @code
* Semaphore sem;
* MyTask task;
*
* pWorker->execute(&task, &sem);
* sem.wait();
*
* MyResult& result = task.result();
* @endcode
*/
bool post(Task* pTask, Semaphore* pSem = NULL, enum execute_mode_t mode = EXECUTE_AUTO);
/**
* Posts a task to a worker for execution.
*
* @param pTask The task to be executed.
* @param mode Execution mode
*
* @return True if the task could be posted (i.e. not executed), false otherwise.
*
* @attention Once the task has been executed, it will be deleted.
*/
bool post(std::auto_ptr<DisposableTask> sTask, enum execute_mode_t mode = EXECUTE_AUTO);
template<class T>
bool post(std::auto_ptr<T> sTask, enum execute_mode_t mode = EXECUTE_AUTO)
{
return post(std::auto_ptr<DisposableTask>(sTask.release()), mode);
}
/**
* Posts a task to all workers for execution.
*
* @param pTask The task to be executed.
* @param pSem If non-NULL, will be posted once per worker when the task's
* `execute` return.
*
* @return How many workers the task was posted to.
*
* @attention The very same task will be posted to all workers. The task
* should either not have any sharable data or then it should
* have data specific to each worker that can be accessed
* without locks.
*/
static size_t broadcast(Task* pTask, Semaphore* pSem = NULL);
/**
* Posts a task to all workers for execution.
*
* @param pTask The task to be executed.
*
* @return How many workers the task was posted to.
*
* @attention The very same task will be posted to all workers. The task
* should either not have any sharable data or then it should
* have data specific to each worker that can be accessed
* without locks.
*
* @attention Once the task has been executed by all workers, it will
* be deleted.
*/
static size_t broadcast(std::auto_ptr<DisposableTask> sTask);
template<class T>
static size_t broadcast(std::auto_ptr<T> sTask)
{
return broadcast(std::auto_ptr<DisposableTask>(sTask.release()));
}
/**
* Executes a task on all workers in serial mode (the task is executed
* on at most one worker thread at a time). When the function returns
* the task has been executed on all workers.
*
* @param task The task to be executed.
*
* @return How many workers the task was posted to.
*
* @warning This function is extremely inefficient and will be slow compared
* to the other functions. Only use this function when printing thread-specific
* data to stdout.
*/
static size_t execute_serially(Task& task);
/**
* Executes a task on all workers concurrently and waits until all workers
* are done. That is, when the function returns the task has been executed
* by all workers.
*
* @param task The task to be executed.
*
* @return How many workers the task was posted to.
*/
static size_t execute_concurrently(Task& task);
/**
* Post a message to a worker.
*
* @param msg_id The message id.
* @param arg1 Message specific first argument.
* @param arg2 Message specific second argument.
*
* @return True if the message could be sent, false otherwise. If the message
* posting fails, errno is set appropriately.
*
* @attention The return value tells *only* whether the message could be sent,
* *not* that it has reached the worker.
*
* @attention This function is signal safe.
*/
bool post_message(uint32_t msg_id, intptr_t arg1, intptr_t arg2);
/**
* Return a reference to the session registry of this worker.
*
* @return Session registry.
*/
SessionsById& session_registry();
/**
* Broadcast a message to all worker.
*
* @param msg_id The message id.
* @param arg1 Message specific first argument.
* @param arg2 Message specific second argument.
*
* @return The number of messages posted; if less that ne number of workers
* then some postings failed.
*
* @attention The return value tells *only* whether message could be posted,
* *not* that it has reached the worker.
*
* @attentsion Exactly the same arguments are passed to all workers. Take that
* into account if the passed data must be freed.
*
* @attention This function is signal safe.
*/
static size_t broadcast_message(uint32_t msg_id, intptr_t arg1, intptr_t arg2);
/**
* Initate shutdown of all workers.
*
* @attention A call to this function will only initiate the shutdowm,
* the workers will not have shut down when the function returns.
*
* @attention This function is signal safe.
*/
static void shutdown_all();
/**
* Return the worker associated with the provided worker id.
*
* @param worker_id A worker id.
*
* @return The corresponding worker instance, or NULL if the id does
* not correspond to a worker.
*/
static Worker* get(int worker_id);
/**
* Return the worker associated with the current thread.
*
* @return The worker instance, or NULL if the current thread does not have a worker.
*/
static Worker* get_current();
/**
* Return the worker id associated with the current thread.
*
* @return A worker instance, or -1 if the current thread does not have a worker.
*/
static int get_current_id();
/**
* Set the number of non-blocking poll cycles that will be done before
* a blocking poll will take place.
*
* @param nbpolls Number of non-blocking polls to perform before blocking.
*/
static void set_nonblocking_polls(unsigned int nbpolls);
/**
* Maximum time to block in epoll_wait.
*
* @param maxwait Maximum wait time in millliseconds.
*/
static void set_maxwait(unsigned int maxwait);
private:
Worker(int id,
int epoll_fd);
virtual ~Worker();
static Worker* create(int id, int epoll_listener_fd);
void delete_zombies();
bool post_disposable(DisposableTask* pTask, enum execute_mode_t mode = EXECUTE_AUTO);
void handle_message(MessageQueue& queue, const MessageQueue::Message& msg); // override
static void thread_main(void* arg);
void poll_waitevents();
static uint32_t epoll_instance_handler(struct mxs_poll_data* data, int wid, uint32_t events);
uint32_t handle_epoll_events(uint32_t events);
private:
int m_id; /*< The id of the worker. */
state_t m_state; /*< The state of the worker */
int m_epoll_fd; /*< The epoll file descriptor. */
STATISTICS m_statistics; /*< Worker statistics. */
MessageQueue* m_pQueue; /*< The message queue of the worker. */
THREAD m_thread; /*< The thread handle of the worker. */
bool m_started; /*< Whether the thread has been started or not. */
bool m_should_shutdown; /*< Whether shutdown should be performed. */
bool m_shutdown_initiated; /*< Whether shutdown has been initated. */
SessionsById m_sessions; /*< A mapping of session_id->MXS_SESSION. The map
* should contain sessions exclusive to this
* worker and not e.g. listener sessions. For now,
* it's up to the protocol to decide whether a new
* session is added to the map. */
Zombies m_zombies; /*< DCBs to be deleted. */
uint32_t m_nCurrent_descriptors; /*< Current number of descriptors. */
uint64_t m_nTotal_descriptors; /*< Total number of descriptors. */
Load m_load;
};
}
Reference in New Issue View Git Blame Copy Permalink