/*
- * Copyright (c) 2008, 2009 Nicira Networks.
+ * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
#include <config.h>
#include "poll-loop.h"
-#include <assert.h>
#include <errno.h>
#include <inttypes.h>
#include <poll.h>
#include <stdlib.h>
#include <string.h>
-#include "backtrace.h"
#include "coverage.h"
#include "dynamic-string.h"
#include "fatal-signal.h"
#include "list.h"
+#include "ovs-thread.h"
+#include "seq.h"
+#include "socket-util.h"
#include "timeval.h"
+#include "openvswitch/vlog.h"
+#include "hmap.h"
+#include "hash.h"
-#define THIS_MODULE VLM_poll_loop
-#include "vlog.h"
-
-/* An event that will wake the following call to poll_block(). */
-struct poll_waiter {
- /* Set when the waiter is created. */
- struct list node; /* Element in global waiters list. */
- int fd; /* File descriptor. */
- short int events; /* Events to wait for (POLLIN, POLLOUT). */
- poll_fd_func *function; /* Callback function, if any, or null. */
- void *aux; /* Argument to callback function. */
- struct backtrace *backtrace; /* Optionally, event that created waiter. */
-
- /* Set only when poll_block() is called. */
- struct pollfd *pollfd; /* Pointer to element of the pollfds array
- (null if added from a callback). */
+VLOG_DEFINE_THIS_MODULE(poll_loop);
+
+COVERAGE_DEFINE(poll_create_node);
+COVERAGE_DEFINE(poll_zero_timeout);
+
+struct poll_node {
+ struct hmap_node hmap_node;
+ struct pollfd pollfd; /* Events to pass to time_poll(). */
+ HANDLE wevent; /* Events for WaitForMultipleObjects(). */
+ const char *where; /* Where poll_node was created. */
};
-/* All active poll waiters. */
-static struct list waiters = LIST_INITIALIZER(&waiters);
+struct poll_loop {
+ /* All active poll waiters. */
+ struct hmap poll_nodes;
+
+ /* 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. */
+};
-/* Number of elements in the waiters list. */
-static size_t n_waiters;
+static struct poll_loop *poll_loop(void);
-/* Max time to wait in next call to poll_block(), in milliseconds, or -1 to
- * wait forever. */
-static int timeout = -1;
+/* Look up the node with same fd or wevent. */
+static struct poll_node *
+find_poll_node(struct poll_loop *loop, int fd, HANDLE wevent)
+{
+ struct poll_node *node;
-/* Backtrace of 'timeout''s registration, if debugging is enabled. */
-static struct backtrace timeout_backtrace;
+ /* Both 'fd' and 'wevent' cannot be set. */
+ ovs_assert(!fd != !wevent);
-/* Callback currently running, to allow verifying that poll_cancel() is not
- * being called on a running callback. */
-#ifndef NDEBUG
-static struct poll_waiter *running_cb;
-#endif
+ HMAP_FOR_EACH_WITH_HASH (node, hmap_node,
+ hash_2words(fd, (uint32_t)wevent),
+ &loop->poll_nodes) {
+ if ((fd && node->pollfd.fd == fd)
+ || (wevent && node->wevent == wevent)) {
+ return node;
+ }
+ }
+ return NULL;
+}
+
+/* On Unix based systems:
+ *
+ * Registers 'fd' as waiting for the specified 'events' (which should be
+ * POLLIN or POLLOUT or POLLIN | POLLOUT). The following call to
+ * poll_block() will wake up when 'fd' becomes ready for one or more of the
+ * requested events. The 'fd's are given to poll() function later.
+ *
+ * On Windows system:
+ *
+ * If 'fd' is specified, create a new 'wevent'. Association of 'fd' and
+ * 'wevent' for 'events' happens in poll_block(). If 'wevent' is specified,
+ * it is assumed that it is unrelated to any sockets and poll_block()
+ * will wake up on any event on that 'wevent'. It is an error to pass
+ * both 'wevent' and 'fd'.
+ *
+ * The event registration is one-shot: only the following call to
+ * poll_block() is affected. The event will need to be re-registered after
+ * poll_block() is called if it is to persist.
+ *
+ * ('where' is used in debug logging. Commonly one would use poll_fd_wait() to
+ * automatically provide the caller's source file and line number for
+ * 'where'.) */
+static void
+poll_create_node(int fd, HANDLE wevent, short int events, const char *where)
+{
+ struct poll_loop *loop = poll_loop();
+ struct poll_node *node;
+
+ COVERAGE_INC(poll_create_node);
+
+ /* Both 'fd' and 'wevent' cannot be set. */
+ ovs_assert(!fd != !wevent);
-static struct poll_waiter *new_waiter(int fd, short int events);
+ /* Check for duplicate. If found, "or" the events. */
+ node = find_poll_node(loop, fd, wevent);
+ if (node) {
+ node->pollfd.events |= events;
+ } else {
+ node = xzalloc(sizeof *node);
+ hmap_insert(&loop->poll_nodes, &node->hmap_node,
+ hash_2words(fd, (uint32_t)wevent));
+ node->pollfd.fd = fd;
+ node->pollfd.events = events;
+#ifdef _WIN32
+ if (!wevent) {
+ wevent = CreateEvent(NULL, FALSE, FALSE, NULL);
+ }
+#endif
+ node->wevent = wevent;
+ node->where = where;
+ }
+}
/* Registers 'fd' as waiting for the specified 'events' (which should be POLLIN
* or POLLOUT or POLLIN | POLLOUT). The following call to poll_block() will
* wake up when 'fd' becomes ready for one or more of the requested events.
*
+ * On Windows, 'fd' must be a socket.
+ *
+ * The event registration is one-shot: only the following call to poll_block()
+ * is affected. The event will need to be re-registered after poll_block() is
+ * called if it is to persist.
+ *
+ * ('where' is used in debug logging. Commonly one would use poll_fd_wait() to
+ * automatically provide the caller's source file and line number for
+ * 'where'.) */
+void
+poll_fd_wait_at(int fd, short int events, const char *where)
+{
+ poll_create_node(fd, 0, events, where);
+}
+
+#ifdef _WIN32
+/* Registers for the next call to poll_block() to wake up when 'wevent' is
+ * signaled.
+ *
* The event registration is one-shot: only the following call to poll_block()
* is affected. The event will need to be re-registered after poll_block() is
- * called if it is to persist. */
-struct poll_waiter *
-poll_fd_wait(int fd, short int events)
+ * called if it is to persist.
+ *
+ * ('where' is used in debug logging. Commonly one would use
+ * poll_wevent_wait() to automatically provide the caller's source file and
+ * line number for 'where'.) */
+void
+poll_wevent_wait_at(HANDLE wevent, const char *where)
{
- COVERAGE_INC(poll_fd_wait);
- return new_waiter(fd, events);
+ poll_create_node(0, wevent, 0, where);
}
+#endif /* _WIN32 */
/* Causes the following call to poll_block() to block for no more than 'msec'
* milliseconds. If 'msec' is nonpositive, the following call to poll_block()
*
* The timer registration is one-shot: only the following call to poll_block()
* is affected. The timer will need to be re-registered after poll_block() is
- * called if it is to persist. */
+ * called if it is to persist.
+ *
+ * ('where' is used in debug logging. Commonly one would use poll_timer_wait()
+ * to automatically provide the caller's source file and line number for
+ * 'where'.) */
void
-poll_timer_wait(int msec)
+poll_timer_wait_at(long long int msec, const char *where)
{
- if (timeout < 0 || msec < timeout) {
- timeout = MAX(0, msec);
- if (VLOG_IS_DBG_ENABLED()) {
- backtrace_capture(&timeout_backtrace);
- }
+ long long int now = time_msec();
+ long long int when;
+
+ if (msec <= 0) {
+ /* Wake up immediately. */
+ when = LLONG_MIN;
+ } else if ((unsigned long long int) now + msec <= LLONG_MAX) {
+ /* Normal case. */
+ when = now + msec;
+ } else {
+ /* now + msec would overflow. */
+ when = LLONG_MAX;
+ }
+
+ poll_timer_wait_until_at(when, where);
+}
+
+/* Causes the following call to poll_block() to wake up when the current time,
+ * as returned by time_msec(), reaches 'when' or later. If 'when' is earlier
+ * than the current time, the following call to poll_block() will not block at
+ * all.
+ *
+ * The timer registration is one-shot: only the following call to poll_block()
+ * is affected. The timer will need to be re-registered after poll_block() is
+ * called if it is to persist.
+ *
+ * ('where' is used in debug logging. Commonly one would use
+ * poll_timer_wait_until() to automatically provide the caller's source file
+ * and line number for 'where'.) */
+void
+poll_timer_wait_until_at(long long int when, const char *where)
+{
+ struct poll_loop *loop = poll_loop();
+ if (when < loop->timeout_when) {
+ loop->timeout_when = when;
+ loop->timeout_where = where;
}
}
/* Causes the following call to poll_block() to wake up immediately, without
- * blocking. */
+ * blocking.
+ *
+ * ('where' is used in debug logging. Commonly one would use
+ * poll_immediate_wake() to automatically provide the caller's source file and
+ * line number for 'where'.) */
void
-poll_immediate_wake(void)
+poll_immediate_wake_at(const char *where)
{
- poll_timer_wait(0);
+ poll_timer_wait_at(0, where);
}
-static void PRINTF_FORMAT(2, 3)
-log_wakeup(const struct backtrace *backtrace, const char *format, ...)
+/* Logs, if appropriate, that the poll loop was awakened by an event
+ * registered at 'where' (typically a source file and line number). The other
+ * arguments have two possible interpretations:
+ *
+ * - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
+ * the wakeup. 'timeout' is ignored.
+ *
+ * - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
+ * which the poll loop woke up.
+ */
+static void
+log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
{
- struct ds ds;
- va_list args;
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 10);
+ enum vlog_level level;
+ int cpu_usage;
+ struct ds s;
+
+ cpu_usage = get_cpu_usage();
+ if (VLOG_IS_DBG_ENABLED()) {
+ level = VLL_DBG;
+ } else if (cpu_usage > 50
+ && !thread_is_pmd()
+ && !VLOG_DROP_INFO(&rl)) {
+ level = VLL_INFO;
+ } else {
+ return;
+ }
- ds_init(&ds);
- va_start(args, format);
- ds_put_format_valist(&ds, format, args);
- va_end(args);
+ ds_init(&s);
+ ds_put_cstr(&s, "wakeup due to ");
+ if (pollfd) {
+ char *description = describe_fd(pollfd->fd);
+ if (pollfd->revents & POLLIN) {
+ ds_put_cstr(&s, "[POLLIN]");
+ }
+ if (pollfd->revents & POLLOUT) {
+ ds_put_cstr(&s, "[POLLOUT]");
+ }
+ if (pollfd->revents & POLLERR) {
+ ds_put_cstr(&s, "[POLLERR]");
+ }
+ if (pollfd->revents & POLLHUP) {
+ ds_put_cstr(&s, "[POLLHUP]");
+ }
+ if (pollfd->revents & POLLNVAL) {
+ ds_put_cstr(&s, "[POLLNVAL]");
+ }
+ ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
+ free(description);
+ } else {
+ ds_put_format(&s, "%d-ms timeout", timeout);
+ }
+ if (where) {
+ ds_put_format(&s, " at %s", where);
+ }
+ if (cpu_usage >= 0) {
+ ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
+ }
+ VLOG(level, "%s", ds_cstr(&s));
+ ds_destroy(&s);
+}
- if (backtrace) {
- int i;
+static void
+free_poll_nodes(struct poll_loop *loop)
+{
+ struct poll_node *node, *next;
- ds_put_char(&ds, ':');
- for (i = 0; i < backtrace->n_frames; i++) {
- ds_put_format(&ds, " 0x%"PRIxPTR, backtrace->frames[i]);
+ HMAP_FOR_EACH_SAFE (node, next, hmap_node, &loop->poll_nodes) {
+ hmap_remove(&loop->poll_nodes, &node->hmap_node);
+#ifdef _WIN32
+ if (node->wevent && node->pollfd.fd) {
+ WSAEventSelect(node->pollfd.fd, NULL, 0);
+ CloseHandle(node->wevent);
}
+#endif
+ free(node);
}
- VLOG_DBG("%s", ds_cstr(&ds));
- ds_destroy(&ds);
}
/* Blocks until one or more of the events registered with poll_fd_wait()
* occurs, or until the minimum duration registered with poll_timer_wait()
- * elapses, or not at all if poll_immediate_wake() has been called.
- *
- * Also executes any autonomous subroutines registered with poll_fd_callback(),
- * if their file descriptors have become ready. */
+ * elapses, or not at all if poll_immediate_wake() has been called. */
void
poll_block(void)
{
- static struct pollfd *pollfds;
- static size_t max_pollfds;
-
- struct poll_waiter *pw;
- struct list *node;
- int n_pollfds;
+ struct poll_loop *loop = poll_loop();
+ struct poll_node *node;
+ struct pollfd *pollfds;
+ HANDLE *wevents = NULL;
+ int elapsed;
int retval;
+ int i;
/* Register fatal signal events before actually doing any real work for
* poll_block. */
fatal_signal_wait();
- assert(!running_cb);
- if (max_pollfds < n_waiters) {
- max_pollfds = n_waiters;
- pollfds = xrealloc(pollfds, max_pollfds * sizeof *pollfds);
+ if (loop->timeout_when == LLONG_MIN) {
+ COVERAGE_INC(poll_zero_timeout);
}
- n_pollfds = 0;
- LIST_FOR_EACH (pw, struct poll_waiter, node, &waiters) {
- pw->pollfd = &pollfds[n_pollfds];
- pollfds[n_pollfds].fd = pw->fd;
- pollfds[n_pollfds].events = pw->events;
- pollfds[n_pollfds].revents = 0;
- n_pollfds++;
- }
+ timewarp_run();
+ pollfds = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *pollfds);
- if (!timeout) {
- COVERAGE_INC(poll_zero_timeout);
- }
- retval = time_poll(pollfds, n_pollfds, timeout);
- if (retval < 0) {
- static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
- VLOG_ERR_RL(&rl, "poll: %s", strerror(-retval));
- } else if (!retval && VLOG_IS_DBG_ENABLED()) {
- log_wakeup(&timeout_backtrace, "%d-ms timeout", timeout);
- }
+#ifdef _WIN32
+ wevents = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *wevents);
+#endif
- for (node = waiters.next; node != &waiters; ) {
- pw = CONTAINER_OF(node, struct poll_waiter, node);
- if (!pw->pollfd || !pw->pollfd->revents) {
- if (pw->function) {
- node = node->next;
- continue;
+ /* Populate with all the fds and events. */
+ i = 0;
+ HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
+ pollfds[i] = node->pollfd;
+#ifdef _WIN32
+ wevents[i] = node->wevent;
+ if (node->pollfd.fd && node->wevent) {
+ short int wsa_events = 0;
+ if (node->pollfd.events & POLLIN) {
+ wsa_events |= FD_READ | FD_ACCEPT | FD_CLOSE;
}
- } else {
- if (VLOG_IS_DBG_ENABLED()) {
- log_wakeup(pw->backtrace, "%s%s%s%s%s on fd %d",
- pw->pollfd->revents & POLLIN ? "[POLLIN]" : "",
- pw->pollfd->revents & POLLOUT ? "[POLLOUT]" : "",
- pw->pollfd->revents & POLLERR ? "[POLLERR]" : "",
- pw->pollfd->revents & POLLHUP ? "[POLLHUP]" : "",
- pw->pollfd->revents & POLLNVAL ? "[POLLNVAL]" : "",
- pw->fd);
+ if (node->pollfd.events & POLLOUT) {
+ wsa_events |= FD_WRITE | FD_CONNECT | FD_CLOSE;
}
-
- if (pw->function) {
-#ifndef NDEBUG
- running_cb = pw;
-#endif
- pw->function(pw->fd, pw->pollfd->revents, pw->aux);
-#ifndef NDEBUG
- running_cb = NULL;
+ WSAEventSelect(node->pollfd.fd, node->wevent, wsa_events);
+ }
#endif
+ i++;
+ }
+
+ retval = time_poll(pollfds, hmap_count(&loop->poll_nodes), wevents,
+ 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(loop->timeout_where, NULL, elapsed);
+ } else if (get_cpu_usage() > 50 || VLOG_IS_DBG_ENABLED()) {
+ i = 0;
+ HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
+ if (pollfds[i].revents) {
+ log_wakeup(node->where, &pollfds[i], 0);
}
+ i++;
}
- node = node->next;
- poll_cancel(pw);
}
- timeout = -1;
- timeout_backtrace.n_frames = 0;
+ free_poll_nodes(loop);
+ loop->timeout_when = LLONG_MAX;
+ loop->timeout_where = NULL;
+ free(pollfds);
+ free(wevents);
/* Handle any pending signals before doing anything else. */
fatal_signal_run();
-}
-/* Registers 'function' to be called with argument 'aux' by poll_block() when
- * 'fd' becomes ready for one of the events in 'events', which should be POLLIN
- * or POLLOUT or POLLIN | POLLOUT.
- *
- * The callback registration persists until the event actually occurs. At that
- * point, it is automatically de-registered. The callback function must
- * re-register the event by calling poll_fd_callback() again within the
- * callback, if it wants to be called back again later. */
-struct poll_waiter *
-poll_fd_callback(int fd, short int events, poll_fd_func *function, void *aux)
+ seq_woke();
+}
+\f
+static void
+free_poll_loop(void *loop_)
{
- struct poll_waiter *pw = new_waiter(fd, events);
- pw->function = function;
- pw->aux = aux;
- return pw;
+ struct poll_loop *loop = loop_;
+
+ free_poll_nodes(loop);
+ hmap_destroy(&loop->poll_nodes);
+ free(loop);
}
-/* Cancels the file descriptor event registered with poll_fd_wait() or
- * poll_fd_callback(). 'pw' must be the struct poll_waiter returned by one of
- * those functions.
- *
- * An event registered with poll_fd_wait() may be canceled from its time of
- * registration until the next call to poll_block(). At that point, the event
- * is automatically canceled by the system and its poll_waiter is freed.
- *
- * An event registered with poll_fd_callback() may be canceled from its time of
- * registration until its callback is actually called. At that point, the
- * event is automatically canceled by the system and its poll_waiter is
- * freed. */
-void
-poll_cancel(struct poll_waiter *pw)
+static struct poll_loop *
+poll_loop(void)
{
- if (pw) {
- assert(pw != running_cb);
- list_remove(&pw->node);
- free(pw->backtrace);
- free(pw);
- n_waiters--;
+ 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);
}
-}
-\f
-/* Creates and returns a new poll_waiter for 'fd' and 'events'. */
-static struct poll_waiter *
-new_waiter(int fd, short int events)
-{
- struct poll_waiter *waiter = xzalloc(sizeof *waiter);
- assert(fd >= 0);
- waiter->fd = fd;
- waiter->events = events;
- if (VLOG_IS_DBG_ENABLED()) {
- waiter->backtrace = xmalloc(sizeof *waiter->backtrace);
- backtrace_capture(waiter->backtrace);
+
+ loop = pthread_getspecific(key);
+ if (!loop) {
+ loop = xzalloc(sizeof *loop);
+ hmap_init(&loop->poll_nodes);
+ xpthread_setspecific(key, loop);
}
- list_push_back(&waiters, &waiter->node);
- n_waiters++;
- return waiter;
+ return loop;
}
+
projects / cascardo / ovs.git / blobdiff