futex
[lttv.git] / usertrace-fast / ltt-usertrace-fast.c
index cb1316951f101692dd6a71e6353a0c7859662e0e..5f8c5e08657a2b9e747e7b00af436939e36a8855 100644 (file)
  *   dies, the data is not lost.
  * * Fast thread spawn : a pthread_create() is done by the application for each
  *   new thread.
+ *
+ * We use a timer to check periodically if the parent died. I think it is less
+ * intrusive than a ptrace() on the parent, which would get every signal. The
+ * side effect of this is that we won't be notified if the parent does an
+ * exec(). In this case, we will just sit there until the parent exits.
+ * 
+ *   
  * Copyright 2006 Mathieu Desnoyers
  *
  */
 #include <string.h>
 #include <sys/mman.h>
 #include <signal.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <stdlib.h>
+#include <sys/param.h>
+#include <linux/futex.h>
+#include <sys/time.h>
+
+#include <asm/timex.h> //for get_cycles()
 
 #include "ltt-usertrace-fast.h"
 
-/* TLS for the trace buffer
- * http://www.dis.com/gnu/gcc/C--98-Thread-Local-Edits.html
- *
- * Add after paragraph 4
- *
- *     The storage for an object of thread storage duration shall be statically
- *     initialized before the first statement of the thread startup function. An
- *     object of thread storage duration shall not require dynamic
- *     initialization.
- */
-#if 0
-__thread struct ltt_trace_info ltt_trace_info =
-{
-       .init = 0,
-       .filter = 0,
-       .nesting = ATOMIC_INIT(0),
-       .channel.facilities = 
-               {       ATOMIC_INIT(0),
-                       ATOMIC_INIT(0),
-                       ATOMIC_INIT(0),
-                       ATOMIC_INIT(0)
-               },
-       .channel.cpu = 
-               { ATOMIC_INIT(0),
-                       ATOMIC_INIT(0),
-                       ATOMIC_INIT(0),
-                       ATOMIC_INIT(0)
-               },
-};
-#endif //0
+
+/* Writer (the traced application) */
 
 __thread struct ltt_trace_info *thread_trace_info = NULL;
 
+void ltt_usertrace_fast_buffer_switch(void)
+{
+       struct ltt_trace_info *tmp = thread_trace_info;
+       if(tmp)
+               kill(tmp->daemon_id, SIGUSR1);
+}
+
+/* The cleanup should never be called from a signal handler */
+static void ltt_usertrace_fast_cleanup(void *arg)
+{
+       struct ltt_trace_info *tmp = thread_trace_info;
+       if(tmp) {
+               thread_trace_info = NULL;
+               kill(tmp->daemon_id, SIGUSR2);
+               munmap(tmp, sizeof(*tmp));
+       }
+}
+
+/* Reader (the disk dumper daemon) */
+
+static pid_t traced_pid = 0;
+static pthread_t traced_thread = 0;
+static int parent_exited = 0;
 
 /* signal handling */
+static void handler_sigusr1(int signo)
+{
+       printf("LTT Signal %d received : parent buffer switch.\n", signo);
+}
 
-static void handler(int signo)
+static void handler_sigusr2(int signo)
 {
-       printf("Signal %d received\n", signo);
+       printf("LTT Signal %d received : parent exited.\n", signo);
+       parent_exited = 1;
 }
 
+static void handler_sigalarm(int signo)
+{
+       printf("LTT Signal %d received\n", signo);
 
+       if(getppid() != traced_pid) {
+               /* Parent died */
+               printf("LTT Parent %lu died, cleaning up\n", traced_pid);
+               traced_pid = 0;
+       }
+       alarm(3);
+}
 
-void ltt_usertrace_fast_buffer_switch(void)
+/* Do a buffer switch. Don't switch if buffer is completely empty */
+static void flush_buffer(struct ltt_buf *ltt_buf)
 {
-       kill(thread_trace_info->daemon_id, SIGUSR1);
+
+
+}
+
+static inline int ltt_buffer_get(struct ltt_buf *ltt_buf,
+               unsigned int *offset)
+{
+       unsigned int consumed_old, consumed_idx;
+       consumed_old = atomic_read(&ltt_buf->consumed);
+       consumed_idx = SUBBUF_INDEX(consumed_old, ltt_buf);
+       
+       if(atomic_read(&ltt_buf->commit_count[consumed_idx])
+               != atomic_read(&ltt_buf->reserve_count[consumed_idx])) {
+               return -EAGAIN;
+       }
+       if((SUBBUF_TRUNC(atomic_read(&ltt_buf->offset), ltt_buf)
+                               -SUBBUF_TRUNC(consumed_old, ltt_buf)) == 0) {
+               return -EAGAIN;
+       }
+       
+       *offset = consumed_old;
+
+       return 0;
+}
+
+static inline int ltt_buffer_put(struct ltt_buf *ltt_buf,
+               unsigned int offset)
+{
+       unsigned int consumed_old, consumed_new;
+       int ret;
+
+       consumed_old = offset;
+       consumed_new = SUBBUF_ALIGN(consumed_old, ltt_buf);
+       if(atomic_cmpxchg(&ltt_buf->consumed, consumed_old, consumed_new)
+                       != consumed_old) {
+               /* We have been pushed by the writer : the last buffer read _is_
+                * corrupted!
+                * It can also happen if this is a buffer we never got. */
+               return -EIO;
+       } else {
+               if(atomic_read(&ltt_buf->full) == 1) {
+                       /* tell the client that buffer is now unfull */
+                       ret = futex(&ltt_buf->full, FUTEX_WAKE, 1, NULL, NULL, 0);
+                       if(ret != 1) {
+                               printf("LTT warning : race condition : writer not waiting or too many writers\n");
+                       }
+                       atomic_set(&ltt_buf->full, 0);
+               }
+       }
 }
 
+/* In the writer :
+ *
+ * if(buffer full condition) {
+ *   put myself in the wait queue
+ *   ltt_buf->full = 1;
+ *   schedule
+ * }
+ *{
+       if(buffer_is_full) {
+               atomic_set(&ltt_buf->full, 1);
+               ret = futex(&ltt_buf->full, 1, NULL, NULL, 0);
+       }
+}
+
+ */
+
+static int read_subbuffer(struct ltt_buf *ltt_buf, int fd)
+{
+       int err;
+       printf("LTT read buffer\n");
+
+
+       err = ltt_buffer_get(&shared_trace_info->channel.cpu, &consumed_old);
+       if(err != -EAGAIN && err != 0) {
+               printf("LTT Reserving sub buffer failed\n");
+               goto get_error;
+       }
+
+       err = TEMP_FAILURE_RETRY(write(fd,
+                               ltt_buf->start 
+                                       + (consumed_old & ((ltt_buf->alloc_size)-1)),
+                               ltt_buf->subbuf_size));
+
+       if(err < 0) {
+               perror("Error in writing to file");
+               goto write_error;
+       }
+#if 0
+       err = fsync(pair->trace);
+       if(err < 0) {
+               ret = errno;
+               perror("Error in writing to file");
+               goto write_error;
+       }
+#endif //0
+write_error:
+       err = ltt_buffer_put(&shared_trace_info->channel.cpu, consumed_old);
+
+       if(err != 0) {
+               if(err == -EIO) {
+                       perror("Reader has been pushed by the writer, last subbuffer corrupted.");
+                       /* FIXME : we may delete the last written buffer if we wish. */
+               }
+               goto get_error;
+       }
+
+get_error:
+       return err;
+}
+
+/* This function is called by ltt_rw_init which has signals blocked */
 static void ltt_usertrace_fast_daemon(struct ltt_trace_info *shared_trace_info,
-               sigset_t oldset)
+               sigset_t oldset, pid_t l_traced_pid, pthread_t l_traced_thread)
 {
        struct sigaction act;
        int ret;
+       int fd_fac;
+       int fd_cpu;
+       char outfile_name[PATH_MAX];
+       char identifier_name[PATH_MAX];
 
-       printf("ltt_usertrace_fast_daemon : init is %d, pid is %lu\n",
-                       shared_trace_info->init, getpid());
 
-       act.sa_handler = handler;
+       traced_pid = l_traced_pid;
+       traced_thread = l_traced_thread;
+
+       printf("LTT ltt_usertrace_fast_daemon : init is %d, pid is %lu, traced_pid is %lu\n",
+                       shared_trace_info->init, getpid(), traced_pid);
+
+       act.sa_handler = handler_sigusr1;
        act.sa_flags = 0;
        sigemptyset(&(act.sa_mask));
        sigaddset(&(act.sa_mask), SIGUSR1);
        sigaction(SIGUSR1, &act, NULL);
+
+       act.sa_handler = handler_sigusr2;
+       act.sa_flags = 0;
+       sigemptyset(&(act.sa_mask));
+       sigaddset(&(act.sa_mask), SIGUSR2);
+       sigaction(SIGUSR2, &act, NULL);
+
+       act.sa_handler = handler_sigalarm;
+       act.sa_flags = 0;
+       sigemptyset(&(act.sa_mask));
+       sigaddset(&(act.sa_mask), SIGALRM);
+       sigaction(SIGALRM, &act, NULL);
+
        /* Enable signals */
        ret = pthread_sigmask(SIG_SETMASK, &oldset, NULL);
        if(ret) {
-               printf("Error in pthread_sigmask\n");
+               printf("LTT Error in pthread_sigmask\n");
        }
 
+       alarm(3);
+
+       /* Open output files */
+       umask(00000);
+       ret = mkdir(LTT_USERTRACE_ROOT, 0777);
+       if(ret < 0 && errno != EEXIST) {
+               perror("LTT Error in creating output (mkdir)");
+               exit(-1);
+       }
+       ret = chdir(LTT_USERTRACE_ROOT);
+       if(ret < 0) {
+               perror("LTT Error in creating output (chdir)");
+               exit(-1);
+       }
+       snprintf(identifier_name, PATH_MAX-1,   "%lu.%lu.%llu",
+                       traced_pid, traced_thread, get_cycles());
+       snprintf(outfile_name, PATH_MAX-1,      "facilities-%s", identifier_name);
+       fd_fac = creat(outfile_name, 0644);
+
+       snprintf(outfile_name, PATH_MAX-1,      "cpu-%s", identifier_name);
+       fd_cpu = creat(outfile_name, 0644);
+       
+       
        while(1) {
                pause();
-               printf("Doing a buffer switch read. pid is : %lu\n", getpid());
+               if(traced_pid == 0) break; /* parent died */
+               if(parent_exited) break;
+               printf("LTT Doing a buffer switch read. pid is : %lu\n", getpid());
+       
+               do {
+                       ret = read_buffer(&shared_trace_info->channel.cpu, fd_cpu);
+               } while(ret == 0);
+
+               do {
+                       ret = read_buffer(&shared_trace_info->channel.facilities, fd_fac);
+               } while(ret == 0);
        }
 
+       /* Buffer force switch (flush) */
+       flush_buffer(&shared_trace_info->channel.cpu);
+       do {
+               ret = read_buffer(&shared_trace_info->channel.cpu, fd_cpu);
+       } while(ret == 0);
+
+
+       flush_buffer(&shared_trace_info->channel.facilities);
+       do {
+               ret = read_buffer(&shared_trace_info->channel.facilities, fd_fac);
+       } while(ret == 0);
+
+       close(fd_fac);
+       close(fd_cpu);
+       
+       /* The parent thread is dead and we have finished with the buffer */
+       munmap(shared_trace_info, sizeof(*shared_trace_info));
+       
+       exit(0);
 }
 
-void ltt_thread_init(void)
+
+/* Reader-writer initialization */
+
+static enum ltt_process_role { LTT_ROLE_WRITER, LTT_ROLE_READER }
+       role = LTT_ROLE_WRITER;
+
+
+void ltt_rw_init(void)
 {
        pid_t pid;
        struct ltt_trace_info *shared_trace_info;
        int ret;
        sigset_t set, oldset;
+       pid_t l_traced_pid = getpid();
+       pthread_t l_traced_thread = pthread_self();
 
        /* parent : create the shared memory map */
-       shared_trace_info = thread_trace_info = mmap(0, sizeof(*thread_trace_info),
+       shared_trace_info = mmap(0, sizeof(*thread_trace_info),
                        PROT_READ|PROT_WRITE, MAP_SHARED|MAP_ANONYMOUS, 0, 0);
-       memset(shared_trace_info, 0, sizeof(*thread_trace_info));
-       thread_trace_info->init = 1;
+       memset(shared_trace_info, 0, sizeof(*shared_trace_info));
+       /* Tricky semaphore : is in a shared memory space, so it's ok for a fast
+        * mutex (futex). */
+       atomic_set(&shared_trace_info->channel.facilities.full, 0);
+       shared_trace_info->channel.facilities.alloc_size = LTT_BUF_SIZE_FACILITIES;
+       shared_trace_info->channel.facilities.subbuf_size = LTT_SUBBUF_SIZE_FACILITIES;
+       atomic_set(&shared_trace_info->channel.cpu.full, 0);
+       shared_trace_info->channel.cpu.alloc_size = LTT_BUF_SIZE_CPU;
+       shared_trace_info->channel.cpu.subbuf_size = LTT_SUBBUF_SIZE_CPU;
+       shared_trace_info->init = 1;
 
        /* Disable signals */
   ret = sigfillset(&set);
   if(ret) {
-    printf("Error in sigfillset\n");
+    printf("LTT Error in sigfillset\n");
   } 
        
        
   ret = pthread_sigmask(SIG_BLOCK, &set, &oldset);
   if(ret) {
-    printf("Error in pthread_sigmask\n");
+    printf("LTT Error in pthread_sigmask\n");
   }
-       
+
        pid = fork();
        if(pid > 0) {
                /* Parent */
-               thread_trace_info->daemon_id = pid;
+               shared_trace_info->daemon_id = pid;
+               thread_trace_info = shared_trace_info;
 
                /* Enable signals */
                ret = pthread_sigmask(SIG_SETMASK, &oldset, NULL);
                if(ret) {
-                       printf("Error in pthread_sigmask\n");
+                       printf("LTT Error in pthread_sigmask\n");
                }
        } else if(pid == 0) {
                /* Child */
-               ltt_usertrace_fast_daemon(shared_trace_info, oldset);
+               role = LTT_ROLE_READER;
+               ltt_usertrace_fast_daemon(shared_trace_info, oldset, l_traced_pid,
+                                       l_traced_thread);
                /* Should never return */
                exit(-1);
        } else if(pid < 0) {
                /* fork error */
-               perror("Error in forking ltt-usertrace-fast");
+               perror("LTT Error in forking ltt-usertrace-fast");
        }
 }
 
+static __thread struct _pthread_cleanup_buffer cleanup_buffer;
+
+void ltt_thread_init(void)
+{
+       _pthread_cleanup_push(&cleanup_buffer, ltt_usertrace_fast_cleanup, NULL);
+       ltt_rw_init();
+}
+       
 void __attribute__((constructor)) __ltt_usertrace_fast_init(void)
 {
   printf("LTT usertrace-fast init\n");
 
-       ltt_thread_init();
+       ltt_rw_init();
 }
 
 void __attribute__((destructor)) __ltt_usertrace_fast_fini(void)
 {
-  printf("LTT usertrace-fast fini\n");
-
+       if(role == LTT_ROLE_WRITER) {
+         printf("LTT usertrace-fast fini\n");
+               ltt_usertrace_fast_cleanup(NULL);
+       }
 }
 
This page took 0.032217 seconds and 4 git commands to generate.