poll-loop: Make poll loop data structures per-thread.

Signed-off-by: Ben Pfaff <blp@nicira.com>

lib/poll-loop.c

@@ -25,6 +25,7 @@
 #include "dynamic-string.h"
 #include "fatal-signal.h"
 #include "list.h"
+#include "ovs-thread.h"
 #include "socket-util.h"
 #include "timeval.h"
 #include "vlog.h"
@@ -39,25 +40,20 @@ VLOG_DEFINE_THIS_MODULE(poll_loop);
 COVERAGE_DEFINE(poll_fd_wait);
 COVERAGE_DEFINE(poll_zero_timeout);
 
-/* An event that will wake the following call to poll_block(). */
-struct poll_waiter {
-    const char *where;          /* Where the waiter was created. */
+struct poll_loop {
+    /* All active poll waiters. */
+    struct pollfd *pollfds;     /* Events to pass to poll(). */
+    const char **where;         /* Where each pollfd was created. */
+    size_t n_waiters;           /* Number of elems in 'where' and 'pollfds'. */
+    size_t allocated_waiters;   /* Allocated elems in 'where' and 'pollfds'. */
+
+    /* Time at which to wake up the next call to poll_block(), LLONG_MIN to
+     * wake up immediately, or LLONG_MAX to wait forever. */
+    long long int timeout_when; /* In msecs as returned by time_msec(). */
+    const char *timeout_where;  /* Where 'timeout_when' was set. */
 };
 
-/* All active poll waiters. */
-static struct poll_waiter *waiters;
-static struct pollfd *pollfds;
-static size_t n_waiters, allocated_waiters;
-
-/* Time at which to wake up the next call to poll_block(), in milliseconds as
- * returned by time_msec(), LLONG_MIN to wake up immediately, or LLONG_MAX to
- * wait forever. */
-static long long int timeout_when = LLONG_MAX;
-
-/* Location where waiter created. */
-static const char *timeout_where;
-
-static void new_waiter(int fd, short int events, const char *where);
+static struct poll_loop *poll_loop(void);
 
 /* Registers 'fd' as waiting for the specified 'events' (which should be POLLIN
  * or POLLOUT or POLLIN | POLLOUT).  The following call to poll_block() will
@@ -72,8 +68,21 @@ static void new_waiter(int fd, short int events, const char *where);
 void
 poll_fd_wait(int fd, short int events, const char *where)
 {
+    struct poll_loop *loop = poll_loop();
+
     COVERAGE_INC(poll_fd_wait);
-    new_waiter(fd, events, where);
+    if (loop->n_waiters >= loop->allocated_waiters) {
+        loop->where = x2nrealloc(loop->where, &loop->allocated_waiters,
+                                 sizeof *loop->where);
+        loop->pollfds = xrealloc(loop->pollfds,
+                                 (loop->allocated_waiters
+                                  * sizeof *loop->pollfds));
+    }
+
+    loop->where[loop->n_waiters] = where;
+    loop->pollfds[loop->n_waiters].fd = fd;
+    loop->pollfds[loop->n_waiters].events = events;
+    loop->n_waiters++;
 }
 
 /* Causes the following call to poll_block() to block for no more than 'msec'
@@ -120,9 +129,10 @@ poll_timer_wait(long long int msec, const char *where)
 void
 poll_timer_wait_until(long long int when, const char *where)
 {
-    if (when < timeout_when) {
-        timeout_when = when;
-        timeout_where = where;
+    struct poll_loop *loop = poll_loop();
+    if (when < loop->timeout_when) {
+        loop->timeout_when = when;
+        loop->timeout_where = where;
     }
 }
 
@@ -204,6 +214,7 @@ log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
 void
 poll_block(void)
 {
+    struct poll_loop *loop = poll_loop();
     int elapsed;
     int retval;
 
@@ -211,45 +222,62 @@ poll_block(void)
      * poll_block. */
     fatal_signal_wait();
 
-    if (timeout_when == LLONG_MIN) {
+    if (loop->timeout_when == LLONG_MIN) {
         COVERAGE_INC(poll_zero_timeout);
     }
-    retval = time_poll(pollfds, n_waiters, timeout_when, &elapsed);
+
+    retval = time_poll(loop->pollfds, loop->n_waiters,
+                       loop->timeout_when, &elapsed);
     if (retval < 0) {
         static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
         VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(-retval));
     } else if (!retval) {
-        log_wakeup(timeout_where, NULL, elapsed);
+        log_wakeup(loop->timeout_where, NULL, elapsed);
     } else if (get_cpu_usage() > 50 || VLOG_IS_DBG_ENABLED()) {
         size_t i;
 
-        for (i = 0; i < n_waiters; i++) {
-            if (pollfds[i].revents) {
-                log_wakeup(waiters[i].where, &pollfds[i], 0);
+        for (i = 0; i < loop->n_waiters; i++) {
+            if (loop->pollfds[i].revents) {
+                log_wakeup(loop->where[i], &loop->pollfds[i], 0);
             }
         }
     }
 
-    timeout_when = LLONG_MAX;
-    timeout_where = NULL;
-    n_waiters = 0;
+    loop->timeout_when = LLONG_MAX;
+    loop->timeout_where = NULL;
+    loop->n_waiters = 0;
 
     /* Handle any pending signals before doing anything else. */
     fatal_signal_run();
 }
 
-/* Creates a new poll_waiter for 'fd' and 'events', recording 'where' as the
- * location where the event was registered. */
 static void
-new_waiter(int fd, short int events, const char *where)
+free_poll_loop(void *loop_)
 {
-    if (n_waiters >= allocated_waiters) {
-        waiters = x2nrealloc(waiters, &allocated_waiters, sizeof *waiters);
-        pollfds = xrealloc(pollfds, allocated_waiters * sizeof *pollfds);
+    struct poll_loop *loop = loop_;
+
+    free(loop->pollfds);
+    free(loop->where);
+    free(loop);
 }
 
-    waiters[n_waiters].where = where;
-    pollfds[n_waiters].fd = fd;
-    pollfds[n_waiters].events = events;
-    n_waiters++;
+static struct poll_loop *
+poll_loop(void)
+{
+    static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
+    static pthread_key_t key;
+    struct poll_loop *loop;
+
+    if (ovsthread_once_start(&once)) {
+        xpthread_key_create(&key, free_poll_loop);
+        ovsthread_once_done(&once);
     }
+
+    loop = pthread_getspecific(key);
+    if (!loop) {
+        loop = xzalloc(sizeof *loop);
+        pthread_setspecific(key, loop);
+    }
+    return loop;
+}
+

lib/poll-loop.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2009, 2010, 2011 Nicira, Inc.
+ * Copyright (c) 2008, 2009, 2010, 2011, 2013 Nicira, Inc.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,14 +16,20 @@
 
 /* High-level wrapper around the "poll" system call.
  *
- * Intended usage is for the program's main loop to go about its business
+ * The intended usage is for each thread's main loop to go about its business
  * servicing whatever events it needs to.  Then, when it runs out of immediate
  * tasks, it calls each subordinate module's "wait" function, which in turn
  * calls one (or more) of the functions poll_fd_wait(), poll_immediate_wake(),
  * and poll_timer_wait() to register to be awakened when the appropriate event
  * occurs.  Then the main loop calls poll_block(), which blocks until one of
- * the registered events happens. */
-
+ * the registered events happens.
+ *
+ *
+ * Thread-safety
+ * =============
+ *
+ * The poll set is per-thread, so all functions in this module are thread-safe.
+ */
 #ifndef POLL_LOOP_H
 #define POLL_LOOP_H 1