From 433afcb4fe81e775c15af9d39a6f4db8a53d693a Mon Sep 17 00:00:00 2001
From: Jens Axboe <jens.axboe@oracle.com>
Date: Thu, 22 Feb 2007 10:39:01 +0100
Subject: [PATCH] io_u timeout handling

Further measures to prevent fio getting stuck, even in case of
engine timeout errors.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
---
 fio.c       | 21 +++++++++++++++++----
 fio.h       | 15 +++++++++++++++
 io_u.c      | 51 +++++++++++++++++++++++++++++++++++++++++++++++++++
 ioengines.c | 19 +++++++++++++++++--
 4 files changed, 100 insertions(+), 6 deletions(-)

diff --git a/fio.c b/fio.c
index 0ccd5d85..fad4cca7 100644
--- a/fio.c
+++ b/fio.c
@@ -567,6 +567,8 @@ static int init_io_u(struct thread_data *td)
 		list_add(&io_u->list, &td->io_u_freelist);
 	}
 
+	io_u_init_timeout();
+
 	return 0;
 }
 
@@ -715,6 +717,7 @@ static void *thread_main(void *data)
 	}
 
 	fio_gettime(&td->epoch, NULL);
+	memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 	getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
 	runtime[0] = runtime[1] = 0;
@@ -833,9 +836,14 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 			 * check if someone quit or got killed in an unusual way
 			 */
 			ret = waitpid(td->pid, &status, WNOHANG);
-			if (ret < 0)
+			if (ret < 0) {
+				if (errno == ECHILD) {
+					log_err("fio: pid=%d disappeared\n", td->pid);
+					td_set_runstate(td, TD_REAPED);
+					goto reaped;
+				}
 				perror("waitpid");
-			else if ((ret == td->pid) && WIFSIGNALED(status)) {
+			} else if ((ret == td->pid) && WIFSIGNALED(status)) {
 				int sig = WTERMSIG(status);
 
 				log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
@@ -865,9 +873,14 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 			int status;
 
 			ret = waitpid(td->pid, &status, 0);
-			if (ret < 0)
+			if (ret < 0) {
+				if (errno == ECHILD) {
+					log_err("fio: pid=%d disappeared\n", td->pid);
+					td_set_runstate(td, TD_REAPED);
+					goto reaped;
+				}
 				perror("waitpid");
-			else if (WIFEXITED(status) && WEXITSTATUS(status)) {
+			} else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 				if (!exit_value)
 					exit_value++;
 			}
diff --git a/fio.h b/fio.h
index c168fe70..c1c71465 100644
--- a/fio.h
+++ b/fio.h
@@ -428,8 +428,21 @@ struct thread_data {
 	 */
 	struct list_head io_hist_list;
 	struct list_head io_log_list;
+
+	/*
+	 * timeout handling
+	 */
+	struct timeval timeout_end;
+	struct itimerval timer;
 };
 
+/*
+ * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
+ * the timer on each queue operation.
+ */
+#define IO_U_TIMEOUT_INC	5
+#define IO_U_TIMEOUT		30
+
 #define __td_verror(td, err, msg)					\
 	do {								\
 		if ((td)->error)					\
@@ -628,6 +641,8 @@ extern void requeue_io_u(struct thread_data *, struct io_u **);
 extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *, endio_handler *);
 extern long __must_check io_u_queued_complete(struct thread_data *, int, endio_handler *);
 extern void io_u_queued(struct thread_data *, struct io_u *);
+extern void io_u_init_timeout(void);
+extern void io_u_set_timeout(struct thread_data *);
 
 /*
  * io engine entry points
diff --git a/io_u.c b/io_u.c
index d30ff4b3..4f17decd 100644
--- a/io_u.c
+++ b/io_u.c
@@ -582,3 +582,54 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
 	add_slat_sample(td, io_u->ddir, slat_time);
 }
+
+void io_u_set_timeout(struct thread_data *td)
+{
+	assert(td->cur_depth);
+
+	td->timer.it_interval.tv_sec = 0;
+	td->timer.it_interval.tv_usec = 0;
+	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
+	td->timer.it_value.tv_usec = 0;
+	setitimer(ITIMER_REAL, &td->timer, NULL);
+	fio_gettime(&td->timeout_end, NULL);
+}
+
+static void io_u_timeout_handler(int fio_unused sig)
+{
+	struct thread_data *td, *__td;
+	pid_t pid = getpid();
+	int i;
+
+	log_err("fio: io_u timeout\n");
+
+	/*
+	 * TLS would be nice...
+	 */
+	td = NULL;
+	for_each_td(__td, i) {
+		if (__td->pid == pid) {
+			td = __td;
+			break;
+		}
+	}
+
+	if (!td) {
+		log_err("fio: io_u timeout, can't find job\n");
+		exit(1);
+	}
+
+	if (!td->cur_depth) {
+		log_err("fio: timeout without pending work?\n");
+		return;
+	}
+
+	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
+	td->error = ETIMEDOUT;
+	exit(1);
+}
+
+void io_u_init_timeout(void)
+{
+	signal(SIGALRM, io_u_timeout_handler);
+}
diff --git a/ioengines.c b/ioengines.c
index 17044187..39433207 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -192,17 +192,32 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
 	io_u->flags |= IO_U_F_FLIGHT;
 
-	if (td->io_ops->flags & FIO_SYNCIO)
+	if (td->io_ops->flags & FIO_SYNCIO) {
 		fio_gettime(&io_u->issue_time, NULL);
 
+		/*
+		 * for a sync engine, set the timeout upfront
+		 */
+		if (mtime_since(&td->timeout_end, &io_u->issue_time) < IO_U_TIMEOUT)
+			io_u_set_timeout(td);
+	}
+
 	if (io_u->ddir != DDIR_SYNC)
 		td->io_issues[io_u->ddir]++;
 
 	ret = td->io_ops->queue(td, io_u);
 
-	if ((td->io_ops->flags & FIO_SYNCIO) == 0)
+	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
 		fio_gettime(&io_u->issue_time, NULL);
 
+		/*
+		 * async engine, set the timeout here
+		 */
+		if (ret == FIO_Q_QUEUED &&
+		    mtime_since(&td->timeout_end, &io_u->issue_time) < IO_U_TIMEOUT)
+			io_u_set_timeout(td);
+	}
+
 	return ret;
 }
 
-- 
2.25.1