Memcached source code analysis (threading model)--reference

发布时间:2019-02-25 整理:脚本之家 作者:未知
脚本之家收集整理的这篇文章主要介绍了Memcached source code analysis (threading model)--reference脚本之家小编觉得挺不错的,现在分享给大家,也给大家做个参考。一起跟随脚本之家小编过来看看吧!

Look under the start memcahced threading process memcached multi-threaded mainly by instantiating multiple libevent,are a main thread and n workers thread is the main thread or workers thread all through the the libevent management network event,in fact,each thread is a separate libevent instance The main thread is responsible for monitoring the client to establish a connection request and accept connections workers thread to handle events such as read and write has established good connections Look at general icon: First look main data structures (thread.c): 

C code
  1. / * An item in the connection queue. * /   
  2. typedef   struct conn_queue_item CQ_ITEM;
  3. struct conn_queue_item {
  4. int sfd out of the;
  5. int init_state;
  6. The int event_flags;
  7. int read_buffer_size;
  8. int is_udp;
  9. CQ_ITEM * next;
  10. };

CQ_ITEM is actually the main thread accept returned package to establish a connection fd 

C code
  1. / * A connection queue. * /   
  2. typedef   struct conn_queue CQ;
  3. struct conn_queue {
  4. CQ_ITEM * head;
  5. CQ_ITEM * tail;
  6. pthread_mutex_t lock;
  7. pthread_cond_t cond;
  8. };

CQ is a CQ_ITEM the singly linked list 

C code
  1. typedef   struct {
  2. pthread_t thread_id; / * unique ID of this thread * /   
  3. struct event_base * base; / * libevent handle this thread uses * /   
  4. struct event notify_event; / * listen event for notify pipe * /   
  5. int notify_receive_fd; / * receiving end of notify pipe * /   
  6. int notify_send_fd; / * sending end of notify pipe * /   
  7. CQ new_conn_queue; / * queue of new connections to handle * /   
  8. } LIBEVENT_THREAD;

Memcached thread structure package,you can see each thread contains a CQ queue,a notification pipe pipe The instance event_base and a libevent Another important structure is the most important for each network connection Package conn 

C code
  1. typedef   struct {
  2. int sfd out of the;
  3. int State;
  4. struct event event;
  5. short which;
  6. The char * RBUF;
  7. ... / / Eliminating the status flag and read-write buf information   
  8. } Conn;

memcached mainly through Settings / conversion connected to different states to handle the event (core function drive_machine) See next thread initialization process: The main function of the memcached.c,first initialized on the main thread libevent 

C code
  1. / * Initialize main thread libevent instance * /   
  2. main_base = event_init ();

Then initialize all workers thread and start the startup process will be described in detail later 

C code
  1. / * @H_120_404@Start up worker threads if MT mode * /   
  2. thread_init (settings.num_threads,main_base);

Then the main thread calls (only analyze the the tcp situation,memcached support udp) 

C code
  1. server_socket (settings.port,0)

This method encapsulates create listening socket bound address,set the non-blocking mode and register the listening socket the libevent read event,a series of operations Then the main thread calls 

C code
  1. / * Enter the event loop * /   
  2. event_base_loop (main_base,0);

At this time the main thread start libevent to accept the the external connection request,the entire start-up process is completed Let's look at how to start thread_init all workers thread,look at the core code thread_init 

C code
  1. void thread_init (int nthreads, struct event_base * main_base) {
  2. / /. . . Omission   
  3. threads = malloc (sizeof (LIBEVENT_THREAD) * nthreads);
  4. if (threads) {
  5. perror ("Can't allocate thread descriptors");
  6. Exit (1);
  7. }
  8.   
  9. threads [0]. base = main_base;
  10. threads [0]. thread_id = pthread_self ();
  11.   
  12. for (i = 0; i <nthreads; i + +) {
  13. int fds [2];
  14. if (pipe (fds)) {
  15. perror ("Can't create notify pipe");
  16. Exit (1);
  17. }
  18.   
  19. threads [i notify_receive_fd = fds [0];
  20. threads [i]. notify_send_fd = the FDS [1];
  21.   
  22. setup_thread (& threads [i]);
  23. }
  24.   
  25. / * Create threads after we've done all the libevent setup. * /   
  26. for (i = 1; i <nthreads; i + +) {
  27. create_worker (worker_libevent,& threads [i]);
  28. }
  29. }

threads statement static LIBEVENT_THREAD * threads; The thread_init first malloc thread space,and then the first threads as the main thread,the rest are workers thread is then created for each thread a pipe,this pipe is used as the main thread to inform the workers thread a new connection arrives Following setup_thread for 

C code
  1. static   void setup_thread (LIBEVENT_THREAD * me) {
  2. if (! me-> base) {
  3. me-> base = event_init ();
  4. if (! me-> base) {
  5. fprintf (stderr, "Can't allocate event base \ n");
  6. Exit (1);
  7. }
  8. }
  9.   
  10. / * Listen for notifications from other threads * /   
  11. event_set (& me-> notify_event,me-> notify_receive_fd,
  12. EV_READ | EV_PERSIST,thread_libevent_process,me);
  13. event_base_set (me-> base,& me-> notify_event);
  14.   
  15. if (event_add (& me-> notify_event, 0) == -1) {
  16. fprintf (stderr, "Can't monitor libevent notify pipe \ n");
  17. Exit (1);
  18. }
  19.   
  20. cq_init (& me-> new_conn_queue);
  21. }

the create setup_thread libevent instance of all workers thread (the libevent main thread instance has been established in the main function) Since the threads before [0] base = main_base; first thread (the main thread) will not be here execution event_init () In this method,then is to register all workers thread pipe read end the libevent of read events,wait for the main thread last all workers CQ initialization the create_worker actually is the real start of the thread pthread_create call worker_libevent method,this method is executedevent_base_loop start the thread libevent Here we need to remember that each workers thread only data from the read end of the pipe in its own thread readable trigger,and call thread_libevent_process methods Look at this function 

C code
  1. static   void thread_libevent_process (int fd, short which, void * arg) {
  2. LIBEVENT_THREAD * me = arg;
  3. CQ_ITEM * item;
  4. char buf [1];
  5.   
  6. if (read (fd,buf,1)! = 1)
  7. if (settings.verbose,> 0)
  8. fprintf (stderr, "Can't read from libevent pipe \ n");
  9.   
  10. item = cq_peek (& me-> new_conn_queue);
  11.   
  12. if (NULL! = item) {
  13. conn * c = conn_new (item-> sfd,item-> init_state,item-> event_flags,
  14. item-> read_buffer_size,item-> is_udp,me-> base);
  15. . . . / / Omitted   
  16. }
  17. }

The fd function parameters pipe read end of the thread descriptor first 1 byte of the pipeline notification signal readout (this is necessary in the level trigger mode,if does not handle the event,it will be loop notification know the event to be treated) the cq_peek from the thread CQ queue take the head of the queue a CQ_ITEM,this CQ_ITEM is thrown into the main thread in the queue,item-> SFD is already established connection descriptor,by conn_new function of the descriptor registration the libevent read event,me-> Base on behalf of a thread structure,that is the descriptor event processing the to this workersThreading,the most important elements of conn_new method is: 

C code
  1. the Conn * conn_new (const   int SFD, const   int init_state, const   int event_flags
  2. const   int read_buffer_size, const   bool is_udp, struct event_base * base) {
  3. . . .
  4. event_set (& c-> event,sfd,event_flags,event_handler, (void *) c);
  5. event_base_set (base,& c-> event);
  6. c-> ev_flags = event_flags;
  7. if (event_add (& c-> event,0) == -1) {
  8. the if (conn_add_to_freelist (c)) {
  9. conn_free (c);
  10. }
  11. perror ("event_add");
  12. return NULL;
  13. }
  14. . . .
  15. }

You can see the new connection is registered to an event (actually EV_READ | EV_PERSIST),processed by the current thread (because event_base here the workers thread) When the connection readable data callback event_handler function,actually event_handler in the main call memcached core method drive_machine of Finally,look at the main thread is how to notify workers thread to handle the new connection,the main thread libevent registered readable event listening socket descriptor word,that is,when to establish a connection request,the main thread will handle the callback function is also the event_handler readable event (in fact,the main thread is initialized by conn_new listening socket libevent) Last look at the most central part of the memcached network event processing - drive_machine Need to keep in mind is drive_machine perform multi-threaded environment,the main thread and the workers will executive drive_machine, 

C code
  1. static   void drive_machine (conn * c) {
  2. bool stop = false;
  3. int SFD,flags = 1;
  4. socklen_t addrlen;
  5. struct sockaddr_storage addr;
  6. int res;
  7.   
  8. assert (c! = NULL);
  9.   
  10. while (! stop) {
  11.   
  12. switch (c-> state) {
  13. case conn_listening:
  14. addrlen = sizeof (addr);
  15. if ((sfd = accept (c-> sfd, (struct sockaddr *) & addr,& addrlen)) == -1) {
  16. / / Save n error handling   
  17. break;
  18. }
  19. if ((flags = fcntl (sfd, F_GETFL, 0)) <0 | |
  20. fcntl (sfd,F_SETFL,flags | O_NONBLOCK) <0) {
  21. perror ("setting O_NONBLOCK");
  22. close (SFD);
  23. break;
  24. }
  25. Dispatch_conn_new (sfd,conn_read,EV_READ | EV_PERSIST,
  26. DATA_BUFFER_SIZE, false);
  27. break;
  28.   
  29. case conn_read:
  30. the if (try_read_command,(c) = 0) {
  31. continue;
  32. }
  33. .... / / Omitted   
  34. }
  35. }

First of all,less than in fact be the while loop misleading (most do java students will immediately think of a cycle of the loop) while usually to meet a will break in the case,while taking into account the vertical trigger mode,you must read the error of EWOULDBLOCK Closer to home,drive_machine mainly by the current connection state to determine the what,by to libevent registered callback after the read and write time are the core function,so we registered libevent event,while the event state is written to the conn structure libevent callback will the conn structure as an argument over the method parameter the memcached connected through an enum declaration 

C code
  1. enum conn_states to {
  2. conn_listening, / ** the socket which listens for connections * /   
  3. conn_read, / ** reading in a command line * /   
  4. conn_write, / ** writing out a simple response * /   
  5. conn_nread, / ** reading in a fixed number of bytes * /   
  6. conn_swallow, / ** swallowing unnecessary bytes w / o storing * /   
  7. conn_closing, / ** closing this connection * /   
  8. conn_mwrite, / ** writing out many items sequentially * /   
  9. };

Actual for case conn_listening: This is the main thread to deal with their own Workers threads never do this branch we see the main thread to accept calls Dispatch_conn_new (sfd,DATA_BUFFER_SIZE,false); This function is to inform workers thread to see 

C code
  1. void Dispatch_conn_new (int sfd, int init_state, int event_flags,
  2. int read_buffer_size, int is_udp) {
  3. CQ_ITEM * item = cqi_new ();
  4. int   thread = (last_thread + 1)% settings.num_threads;
  5.   
  6. last_thread = thread;
  7.   
  8. item-> sfd = sfd;
  9. item-> init_state = init_state;
  10. item-> event_flags = event_flags;
  11. item-> read_buffer_size = read_buffer_size;
  12. item-> is_udp = is_udp;
  13.   
  14. cq_push (& threads [thread]. new_conn_queue,item);
  15.   
  16. MEMCACHED_CONN_Dispatch (sfd,threads [thread]. Thread_id);
  17. if (write (threads [thread]. notify_send_fd, "", 1)! = 1) {
  18. perror ("Writing to thread notify pipe");
  19. }
  20. }

You can clearly see,the main thread first create a new CQ_ITEM,then select a thread through the round robin strategy And cq_push this CQ_ITEM into the thread CQ queue,the corresponding workers thread is how do you know it Is through this write (threads [thread]. notify_send_fd,"",1) Write 1-byte data to this thread pipe,then the thread's the libevent immediate callback the thread_libevent_process method (already described above) Then the thread remove the item,Register read time,when the data connection of that section will eventually callback drive_machine method, method case conn_read drive_machine: all workers deal with the main thread only processing conn_listening establish a connection to this This part of the code is indeed more,can not all posted,please refer to the source code,the latest version 1.2.6,the province went to a lot of optimization such as,each CQ_ITEM is malloc will malloc a lot to reduce debris generation and so the details.

reference from:http://www.cprogramdevelop.com/408519/

总结

以上是脚本之家为你收集整理的Memcached source code analysis (threading model)--reference全部内容,希望文章能够帮你解决Memcached source code analysis (threading model)--reference所遇到的程序开发问题。

如果觉得脚本之家网站内容还不错,欢迎将脚本之家网站推荐给程序员好友。

标签: