[4/6] um: implement time-travel=ext
diff mbox series

Message ID 20200213132651.20506-5-johannes@sipsolutions.net
State New
Headers show
Series
  • um: external time travel
Related show

Commit Message

Johannes Berg Feb. 13, 2020, 1:26 p.m. UTC
From: Johannes Berg <johannes.berg@intel.com>

This implements synchronized time-travel mode which - using a special
application on a unix socket - lets multiple machines take part in a
time-travelling simulation together.

The protocol for the unix domain socket is defined in the new file
include/uapi/linux/um_timetravel.h.

Change-Id: I0c06fae89d77d5b272dfb4c7ea64641b597b5c7b
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
---
 arch/um/drivers/virtio_uml.c          |  68 +++++-
 arch/um/include/linux/time-internal.h |  27 +++
 arch/um/include/shared/os.h           |   1 +
 arch/um/kernel/skas/syscall.c         |   3 +-
 arch/um/kernel/time.c                 | 323 +++++++++++++++++++++++++-
 arch/um/os-Linux/file.c               |  30 +++
 include/uapi/linux/um_timetravel.h    | 128 ++++++++++
 7 files changed, 563 insertions(+), 17 deletions(-)
 create mode 100644 include/uapi/linux/um_timetravel.h

Comments

Hajime Tazaki Feb. 13, 2020, 2:10 p.m. UTC | #1
Hello,

On Thu, 13 Feb 2020 22:26:47 +0900,
Johannes Berg wrote:

> This implements synchronized time-travel mode which - using a special
> application on a unix socket - lets multiple machines take part in a
> time-travelling simulation together.
> 
> The protocol for the unix domain socket is defined in the new file
> include/uapi/linux/um_timetravel.h.

I was wondering if the source code of the special application (central
scheduler?) is public (or not) so that we can test on a local machine.

thanks,
-- Hajime
Johannes Berg Feb. 13, 2020, 3:52 p.m. UTC | #2
Hi,

> > The protocol for the unix domain socket is defined in the new file
> > include/uapi/linux/um_timetravel.h.
> 
> I was wondering if the source code of the special application (central
> scheduler?) is public (or not) so that we can test on a local machine.

So ... sadly, it's a bit of a sad story right now.

I have one, but it's part of a bigger testing framework that we wrote
(and uses the infrastructure from it), that I haven't been able to
release so far.

On this tool, conceptually, the code is very simple, just keep a sorted
"next event" list, and tell the next one that it's their turn.

For a vhost-user device implementation, it's a bit more complex, as it
needs to be able to handle scheduler messages while waiting for a vhost-
user ACK message, similar to what you can see in the UML vhost-user
code.

johannes
Hajime Tazaki Feb. 14, 2020, 5:34 a.m. UTC | #3
Hello,

On Fri, 14 Feb 2020 00:52:39 +0900,
Johannes Berg wrote:

> > > The protocol for the unix domain socket is defined in the new file
> > > include/uapi/linux/um_timetravel.h.
> > 
> > I was wondering if the source code of the special application (central
> > scheduler?) is public (or not) so that we can test on a local machine.
> 
> So ... sadly, it's a bit of a sad story right now.
> 
> I have one, but it's part of a bigger testing framework that we wrote
> (and uses the infrastructure from it), that I haven't been able to
> release so far.

I understand.  Thanks for telling this.

> On this tool, conceptually, the code is very simple, just keep a sorted
> "next event" list, and tell the next one that it's their turn.

I can imagine how it looks like, as I was doing a similar integration
with kernel code and ns-3 network simulator (which isn't LKL).

This is just my 2 cents (and may be out of topic for the patch
review); it would be nice to have an example code of this external
program.  Since the program is independent from arch/um, this is a
great opportunity to extend in future.

One might wish to implement his/her own scheduler for particular
purpose; for instance, a randomized error injection into time progress
to exercise or fuzzing the Linux implementation, or alternate external
program that can handle multiple UML instances across different nodes
(it this already possible?).

In such case, having a reference implementation would be helpful.

> For a vhost-user device implementation, it's a bit more complex, as it
> needs to be able to handle scheduler messages while waiting for a vhost-
> user ACK message, similar to what you can see in the UML vhost-user
> code.

Let me try to look at the code.

-- Hajime
Johannes Berg Feb. 14, 2020, 8:14 a.m. UTC | #4
> > On this tool, conceptually, the code is very simple, just keep a sorted
> > "next event" list, and tell the next one that it's their turn.
> 
> I can imagine how it looks like, as I was doing a similar integration
> with kernel code and ns-3 network simulator (which isn't LKL).
> 
> This is just my 2 cents (and may be out of topic for the patch
> review); it would be nice to have an example code of this external
> program.  Since the program is independent from arch/um, this is a
> great opportunity to extend in future.

Right.

> One might wish to implement his/her own scheduler for particular
> purpose; for instance, a randomized error injection into time progress
> to exercise or fuzzing the Linux implementation, or alternate external
> program that can handle multiple UML instances across different nodes
> (it this already possible?).

It's all unix domain sockets now, so no network transparency. Wouldn't
really be hard to make it speak the same protocol over a TCP socket
though, but for UML it seemed a bit beside the point.

> In such case, having a reference implementation would be helpful.

Agree, I'm trying to see how I can get something released.

Perhaps it could be acceptable in the kernel itself, under the tools
folder?

johannes

Patch
diff mbox series

diff --git a/arch/um/drivers/virtio_uml.c b/arch/um/drivers/virtio_uml.c
index 9b4c5b7c0f0d..be54d368e73d 100644
--- a/arch/um/drivers/virtio_uml.c
+++ b/arch/um/drivers/virtio_uml.c
@@ -26,6 +26,7 @@ 
 #include <linux/virtio.h>
 #include <linux/virtio_config.h>
 #include <linux/virtio_ring.h>
+#include <linux/time-internal.h>
 #include <shared/as-layout.h>
 #include <irq_kern.h>
 #include <init.h>
@@ -64,6 +65,11 @@  struct virtio_uml_device {
 struct virtio_uml_vq_info {
 	int kick_fd, call_fd;
 	char name[32];
+#ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
+	struct virtqueue *vq;
+	vq_callback_t *callback;
+	struct time_travel_event defer;
+#endif
 };
 
 extern unsigned long long physmem_size, highmem;
@@ -118,10 +124,27 @@  static int vhost_user_recv_header(int fd, struct vhost_user_msg *msg)
 
 static int vhost_user_recv(struct virtio_uml_device *vu_dev,
 			   int fd, struct vhost_user_msg *msg,
-			   size_t max_payload_size)
+			   size_t max_payload_size, bool wait)
 {
 	size_t size;
-	int rc = vhost_user_recv_header(fd, msg);
+	int rc;
+
+	/*
+	 * In virtio time-travel mode, we're handling all the vhost-user
+	 * FDs by polling them whenever appropriate. However, we may get
+	 * into a situation where we're sending out an interrupt message
+	 * to a device (e.g. a net device) and need to handle a simulation
+	 * time message while doing so, e.g. one that tells us to update
+	 * our idea of how long we can run without scheduling.
+	 *
+	 * Thus, we need to not just read() from the given fd, but need
+	 * to also handle messages for the simulation time - this function
+	 * does that for us while waiting for the given fd to be readable.
+	 */
+	if (wait)
+		time_travel_wait_readable(fd);
+
+	rc = vhost_user_recv_header(fd, msg);
 
 	if (rc == -ECONNRESET && vu_dev->registered) {
 		struct virtio_uml_platform_data *pdata;
@@ -143,7 +166,8 @@  static int vhost_user_recv_resp(struct virtio_uml_device *vu_dev,
 				struct vhost_user_msg *msg,
 				size_t max_payload_size)
 {
-	int rc = vhost_user_recv(vu_dev, vu_dev->sock, msg, max_payload_size);
+	int rc = vhost_user_recv(vu_dev, vu_dev->sock, msg,
+				 max_payload_size, true);
 
 	if (rc)
 		return rc;
@@ -173,7 +197,8 @@  static int vhost_user_recv_req(struct virtio_uml_device *vu_dev,
 			       struct vhost_user_msg *msg,
 			       size_t max_payload_size)
 {
-	int rc = vhost_user_recv(vu_dev, vu_dev->req_fd, msg, max_payload_size);
+	int rc = vhost_user_recv(vu_dev, vu_dev->req_fd, msg,
+				 max_payload_size, false);
 
 	if (rc)
 		return rc;
@@ -700,6 +725,8 @@  static bool vu_notify(struct virtqueue *vq)
 	const uint64_t n = 1;
 	int rc;
 
+	time_travel_propagate_time();
+
 	if (info->kick_fd < 0) {
 		struct virtio_uml_device *vu_dev;
 
@@ -847,6 +874,23 @@  static int vu_setup_vq_call_fd(struct virtio_uml_device *vu_dev,
 	return rc;
 }
 
+#ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
+static void vu_defer_irq_handle(struct time_travel_event *d)
+{
+	struct virtio_uml_vq_info *info;
+
+	info = container_of(d, struct virtio_uml_vq_info, defer);
+	info->callback(info->vq);
+}
+
+static void vu_defer_irq_callback(struct virtqueue *vq)
+{
+	struct virtio_uml_vq_info *info = vq->priv;
+
+	time_travel_add_irq_event(&info->defer);
+}
+#endif
+
 static struct virtqueue *vu_setup_vq(struct virtio_device *vdev,
 				     unsigned index, vq_callback_t *callback,
 				     const char *name, bool ctx)
@@ -866,6 +910,19 @@  static struct virtqueue *vu_setup_vq(struct virtio_device *vdev,
 	snprintf(info->name, sizeof(info->name), "%s.%d-%s", pdev->name,
 		 pdev->id, name);
 
+#ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
+	/*
+	 * When we get an interrupt, we must bounce it through the simulation
+	 * calendar (the simtime device), except for the simtime device itself
+	 * since that's part of the simulation control.
+	 */
+	if (time_travel_mode == TT_MODE_EXTERNAL && callback) {
+		info->callback = callback;
+		callback = vu_defer_irq_callback;
+		time_travel_set_event_fn(&info->defer, vu_defer_irq_handle);
+	}
+#endif
+
 	vq = vring_create_virtqueue(index, num, PAGE_SIZE, vdev, true, true,
 				    ctx, vu_notify, callback, info->name);
 	if (!vq) {
@@ -874,6 +931,9 @@  static struct virtqueue *vu_setup_vq(struct virtio_device *vdev,
 	}
 	vq->priv = info;
 	num = virtqueue_get_vring_size(vq);
+#ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
+	info->vq = vq;
+#endif
 
 	if (vu_dev->protocol_features &
 			BIT_ULL(VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS)) {
diff --git a/arch/um/include/linux/time-internal.h b/arch/um/include/linux/time-internal.h
index eb1f84616edb..e21655926f08 100644
--- a/arch/um/include/linux/time-internal.h
+++ b/arch/um/include/linux/time-internal.h
@@ -15,6 +15,7 @@  enum time_travel_mode {
 	TT_MODE_OFF,
 	TT_MODE_BASIC,
 	TT_MODE_INFCPU,
+	TT_MODE_EXTERNAL,
 };
 
 #ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
@@ -35,6 +36,24 @@  time_travel_set_event_fn(struct time_travel_event *e,
 {
 	e->fn = fn;
 }
+
+void __time_travel_propagate_time(void);
+
+static inline void time_travel_propagate_time(void)
+{
+	if (time_travel_mode == TT_MODE_EXTERNAL)
+		__time_travel_propagate_time();
+}
+
+void __time_travel_wait_readable(int fd);
+
+static inline void time_travel_wait_readable(int fd)
+{
+	if (time_travel_mode == TT_MODE_EXTERNAL)
+		__time_travel_wait_readable(fd);
+}
+
+void time_travel_add_irq_event(struct time_travel_event *e);
 #else
 struct time_travel_event {
 };
@@ -47,5 +66,13 @@  static inline void time_travel_sleep(unsigned long long duration)
 
 /* this is a macro so the event/function need not exist */
 #define time_travel_set_event_fn(e, fn) do {} while (0)
+
+static inline void time_travel_propagate_time(void)
+{
+}
+
+static inline void time_travel_wait_readable(int fd)
+{
+}
 #endif /* CONFIG_UML_TIME_TRAVEL_SUPPORT */
 #endif /* __TIMER_INTERNAL_H__ */
diff --git a/arch/um/include/shared/os.h b/arch/um/include/shared/os.h
index 506bcd1bca68..65af1f27b968 100644
--- a/arch/um/include/shared/os.h
+++ b/arch/um/include/shared/os.h
@@ -181,6 +181,7 @@  extern int os_falloc_punch(int fd, unsigned long long offset, int count);
 extern int os_eventfd(unsigned int initval, int flags);
 extern int os_sendmsg_fds(int fd, const void *buf, unsigned int len,
 			  const int *fds, unsigned int fds_num);
+int os_poll(unsigned int n, const int *fds);
 
 /* start_up.c */
 extern void os_early_checks(void);
diff --git a/arch/um/kernel/skas/syscall.c b/arch/um/kernel/skas/syscall.c
index 2e82820f7d29..0a12d5a09217 100644
--- a/arch/um/kernel/skas/syscall.c
+++ b/arch/um/kernel/skas/syscall.c
@@ -24,7 +24,8 @@  void handle_syscall(struct uml_pt_regs *r)
 	 * went to sleep, even if said userspace interacts with the kernel in
 	 * various ways.
 	 */
-	if (time_travel_mode == TT_MODE_INFCPU)
+	if (time_travel_mode == TT_MODE_INFCPU ||
+	    time_travel_mode == TT_MODE_EXTERNAL)
 		schedule();
 
 	/* Initialize the syscall number and default return value. */
diff --git a/arch/um/kernel/time.c b/arch/um/kernel/time.c
index cdebe96308d7..15c4825b857e 100644
--- a/arch/um/kernel/time.c
+++ b/arch/um/kernel/time.c
@@ -20,10 +20,12 @@ 
 #include <kern_util.h>
 #include <os.h>
 #include <linux/time-internal.h>
+#include <linux/um_timetravel.h>
 #include <shared/init.h>
 
 #ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
 enum time_travel_mode time_travel_mode;
+EXPORT_SYMBOL_GPL(time_travel_mode);
 
 static bool time_travel_start_set;
 static unsigned long long time_travel_start;
@@ -32,6 +34,12 @@  static LIST_HEAD(time_travel_events);
 static unsigned long long time_travel_timer_interval;
 static unsigned long long time_travel_next_event;
 static struct time_travel_event time_travel_timer_event;
+static int time_travel_ext_fd = -1;
+static unsigned int time_travel_ext_waiting;
+static bool time_travel_ext_prev_request_valid;
+static unsigned long long time_travel_ext_prev_request;
+static bool time_travel_ext_free_until_valid;
+static unsigned long long time_travel_ext_free_until;
 
 static void time_travel_set_time(unsigned long long ns)
 {
@@ -41,6 +49,199 @@  static void time_travel_set_time(unsigned long long ns)
 	time_travel_time = ns;
 }
 
+enum time_travel_message_handling {
+	TTMH_IDLE,
+	TTMH_POLL,
+	TTMH_READ,
+};
+
+static void time_travel_handle_message(struct um_timetravel_msg *msg,
+				       enum time_travel_message_handling mode)
+{
+	struct um_timetravel_msg resp = {
+		.op = UM_TIMETRAVEL_ACK,
+	};
+	int ret;
+
+	/*
+	 * Poll outside the locked section (if we're not called to only read
+	 * the response) so we can get interrupts for e.g. virtio while we're
+	 * here, but then we need to lock to not get interrupted between the
+	 * read of the message and write of the ACK.
+	 */
+	if (mode != TTMH_READ) {
+		while (os_poll(1, &time_travel_ext_fd) != 0) {
+			if (mode == TTMH_IDLE) {
+				BUG_ON(!irqs_disabled());
+				local_irq_enable();
+				local_irq_disable();
+			}
+		}
+	}
+
+	ret = os_read_file(time_travel_ext_fd, msg, sizeof(*msg));
+
+	if (ret == 0)
+		panic("time-travel external link is broken\n");
+	if (ret != sizeof(*msg))
+		panic("invalid time-travel message - %d bytes\n", ret);
+
+	switch (msg->op) {
+	default:
+		WARN_ONCE(1, "time-travel: unexpected message %lld\n",
+			  (unsigned long long)msg->op);
+		break;
+	case UM_TIMETRAVEL_ACK:
+		return;
+	case UM_TIMETRAVEL_RUN:
+		time_travel_set_time(msg->time);
+		break;
+	case UM_TIMETRAVEL_FREE_UNTIL:
+		time_travel_ext_free_until_valid = true;
+		time_travel_ext_free_until = msg->time;
+		break;
+	}
+
+	os_write_file(time_travel_ext_fd, &resp, sizeof(resp));
+}
+
+static u64 time_travel_ext_req(u32 op, u64 time)
+{
+	static int seq;
+	int mseq = ++seq;
+	struct um_timetravel_msg msg = {
+		.op = op,
+		.time = time,
+		.seq = mseq,
+	};
+	unsigned long flags;
+
+	/*
+	 * We need to save interrupts here and only restore when we
+	 * got the ACK - otherwise we can get interrupted and send
+	 * another request while we're still waiting for an ACK, but
+	 * the peer doesn't know we got interrupted and will send
+	 * the ACKs in the same order as the message, but we'd need
+	 * to see them in the opposite order ...
+	 *
+	 * This wouldn't matter *too* much, but some ACKs carry the
+	 * current time (for UM_TIMETRAVEL_GET) and getting another
+	 * ACK without a time would confuse us a lot!
+	 *
+	 * The sequence number assignment that happens here lets us
+	 * debug such message handling issues more easily.
+	 */
+	local_irq_save(flags);
+	os_write_file(time_travel_ext_fd, &msg, sizeof(msg));
+
+	while (msg.op != UM_TIMETRAVEL_ACK)
+		time_travel_handle_message(&msg, TTMH_READ);
+
+	if (msg.seq != mseq)
+		panic("time-travel: ACK message has different seqno! op=%d, seq=%d != %d time=%lld\n",
+		      msg.op, msg.seq, mseq, msg.time);
+
+	if (op == UM_TIMETRAVEL_GET)
+		time_travel_set_time(msg.time);
+	local_irq_restore(flags);
+
+	return msg.time;
+}
+
+void __time_travel_wait_readable(int fd)
+{
+	int fds[2] = { fd, time_travel_ext_fd };
+	int ret;
+
+	if (time_travel_mode != TT_MODE_EXTERNAL)
+		return;
+
+	while ((ret = os_poll(2, fds))) {
+		struct um_timetravel_msg msg;
+
+		if (ret == 1)
+			time_travel_handle_message(&msg, TTMH_READ);
+	}
+}
+EXPORT_SYMBOL_GPL(__time_travel_wait_readable);
+
+static void time_travel_ext_update_request(unsigned long long time)
+{
+	if (time_travel_mode != TT_MODE_EXTERNAL)
+		return;
+
+	/* asked for exactly this time previously */
+	if (time_travel_ext_prev_request_valid &&
+	    time == time_travel_ext_prev_request)
+		return;
+
+	time_travel_ext_prev_request = time;
+	time_travel_ext_prev_request_valid = true;
+	time_travel_ext_req(UM_TIMETRAVEL_REQUEST, time);
+}
+
+void __time_travel_propagate_time(void)
+{
+	time_travel_ext_req(UM_TIMETRAVEL_UPDATE, time_travel_time);
+}
+EXPORT_SYMBOL_GPL(__time_travel_propagate_time);
+
+/* returns true if we must do a wait to the simtime device */
+static bool time_travel_ext_request(unsigned long long time)
+{
+	/*
+	 * If we received an external sync point ("free until") then we
+	 * don't have to request/wait for anything until then, unless
+	 * we're already waiting.
+	 */
+	if (!time_travel_ext_waiting && time_travel_ext_free_until_valid &&
+	    time < time_travel_ext_free_until)
+		return false;
+
+	time_travel_ext_update_request(time);
+	return true;
+}
+
+static void time_travel_ext_wait(bool idle)
+{
+	struct um_timetravel_msg msg = {
+		.op = UM_TIMETRAVEL_ACK,
+	};
+
+	time_travel_ext_prev_request_valid = false;
+	time_travel_ext_waiting++;
+
+	time_travel_ext_req(UM_TIMETRAVEL_WAIT, -1);
+
+	/*
+	 * Here we are deep in the idle loop, so we have to break out of the
+	 * kernel abstraction in a sense and implement this in terms of the
+	 * UML system waiting on the VQ interrupt while sleeping, when we get
+	 * the signal it'll call time_travel_ext_vq_notify_done() completing the
+	 * call.
+	 */
+	while (msg.op != UM_TIMETRAVEL_RUN)
+		time_travel_handle_message(&msg, idle ? TTMH_IDLE : TTMH_POLL);
+
+	time_travel_ext_waiting--;
+
+	/* we might request more stuff while polling - reset when we run */
+	time_travel_ext_prev_request_valid = false;
+}
+
+static void time_travel_ext_get_time(void)
+{
+	time_travel_ext_req(UM_TIMETRAVEL_GET, -1);
+}
+
+static void __time_travel_update_time(unsigned long long ns, bool idle)
+{
+	if (time_travel_mode == TT_MODE_EXTERNAL && time_travel_ext_request(ns))
+		time_travel_ext_wait(idle);
+	else
+		time_travel_set_time(ns);
+}
+
 static struct time_travel_event *time_travel_first_event(void)
 {
 	return list_first_entry_or_null(&time_travel_events,
@@ -85,6 +286,7 @@  static void __time_travel_add_event(struct time_travel_event *e,
 		list_add_tail(&e->list, &time_travel_events);
 
 	tmp = time_travel_first_event();
+	time_travel_ext_update_request(tmp->time);
 	time_travel_next_event = tmp->time;
 }
 
@@ -106,8 +308,21 @@  void time_travel_periodic_timer(struct time_travel_event *e)
 
 static void time_travel_deliver_event(struct time_travel_event *e)
 {
-	/* this is basically just deliver_alarm(), handles IRQs itself */
-	e->fn(e);
+	if (e == &time_travel_timer_event) {
+		/*
+		 * deliver_alarm() does the irq_enter/irq_exit
+		 * by itself, so must handle it specially here
+		 */
+		e->fn(e);
+	} else {
+		unsigned long flags;
+
+		local_irq_save(flags);
+		irq_enter();
+		e->fn(e);
+		irq_exit();
+		local_irq_restore(flags);
+	}
 }
 
 static bool time_travel_del_event(struct time_travel_event *e)
@@ -119,13 +334,13 @@  static bool time_travel_del_event(struct time_travel_event *e)
 	return true;
 }
 
-static void time_travel_update_time(unsigned long long next, bool retearly)
+static void time_travel_update_time(unsigned long long next, bool idle)
 {
 	struct time_travel_event ne = {
 		.onstack = true,
 	};
 	struct time_travel_event *e;
-	bool finished = retearly;
+	bool finished = idle;
 
 	/* add it without a handler - we deal with that specifically below */
 	__time_travel_add_event(&ne, next);
@@ -134,7 +349,7 @@  static void time_travel_update_time(unsigned long long next, bool retearly)
 		e = time_travel_first_event();
 
 		BUG_ON(!e);
-		time_travel_set_time(e->time);
+		__time_travel_update_time(e->time, idle);
 
 		/* new events may have been inserted while we were waiting */
 		if (e == time_travel_first_event()) {
@@ -150,11 +365,29 @@  static void time_travel_update_time(unsigned long long next, bool retearly)
 				time_travel_deliver_event(e);
 			}
 		}
-	} while (!finished);
+
+		e = time_travel_first_event();
+		if (e)
+			time_travel_ext_update_request(e->time);
+	} while (ne.pending && !finished);
 
 	time_travel_del_event(&ne);
 }
 
+void time_travel_add_irq_event(struct time_travel_event *e)
+{
+	BUG_ON(time_travel_mode != TT_MODE_EXTERNAL);
+
+	time_travel_ext_get_time();
+	/*
+	 * We could model interrupt latency here, for now just
+	 * don't have any latency at all and request the exact
+	 * same time (again) to run the interrupt...
+	 */
+	time_travel_add_event(e, time_travel_time);
+}
+EXPORT_SYMBOL_GPL(time_travel_add_irq_event);
+
 static void time_travel_oneshot_timer(struct time_travel_event *e)
 {
 	deliver_alarm();
@@ -199,6 +432,42 @@  static void time_travel_set_interval(unsigned long long interval)
 {
 	time_travel_timer_interval = interval;
 }
+
+static int time_travel_connect_external(const char *socket)
+{
+	const char *sep;
+	unsigned long long id = (unsigned long long)-1;
+	int rc;
+
+	if ((sep = strchr(socket, ':'))) {
+		char buf[25] = {};
+		if (sep - socket > sizeof(buf) - 1)
+			goto invalid_number;
+
+		memcpy(buf, socket, sep - socket);
+		if (kstrtoull(buf, 0, &id)) {
+invalid_number:
+			panic("time-travel: invalid external ID in string '%s'\n",
+			      socket);
+			return -EINVAL;
+		}
+
+		socket = sep + 1;
+	}
+
+	rc = os_connect_socket(socket);
+	if (rc < 0) {
+		panic("time-travel: failed to connect to external socket %s\n",
+		      socket);
+		return rc;
+	}
+
+	time_travel_ext_fd = rc;
+
+	time_travel_ext_req(UM_TIMETRAVEL_START, id);
+
+	return 1;
+}
 #else /* CONFIG_UML_TIME_TRAVEL_SUPPORT */
 #define time_travel_start_set 0
 #define time_travel_start 0
@@ -216,6 +485,9 @@  static void time_travel_set_interval(unsigned long long interval)
 {
 }
 
+/* fail link if this actually gets used */
+extern u64 time_travel_ext_req(u32 op, u64 time);
+
 /* these are empty macros so the struct/fn need not exist */
 #define time_travel_add_event(e, time) do { } while (0)
 #define time_travel_del_event(e) do { } while (0)
@@ -246,7 +518,8 @@  static int itimer_shutdown(struct clock_event_device *evt)
 	if (time_travel_mode != TT_MODE_OFF)
 		time_travel_del_event(&time_travel_timer_event);
 
-	if (time_travel_mode != TT_MODE_INFCPU)
+	if (time_travel_mode != TT_MODE_INFCPU &&
+	    time_travel_mode != TT_MODE_EXTERNAL)
 		os_timer_disable();
 
 	return 0;
@@ -265,7 +538,8 @@  static int itimer_set_periodic(struct clock_event_device *evt)
 				      time_travel_time + interval);
 	}
 
-	if (time_travel_mode != TT_MODE_INFCPU)
+	if (time_travel_mode != TT_MODE_INFCPU &&
+	    time_travel_mode != TT_MODE_EXTERNAL)
 		os_timer_set_interval(interval);
 
 	return 0;
@@ -284,7 +558,8 @@  static int itimer_next_event(unsigned long delta,
 				      time_travel_time + delta);
 	}
 
-	if (time_travel_mode != TT_MODE_INFCPU)
+	if (time_travel_mode != TT_MODE_INFCPU &&
+	    time_travel_mode != TT_MODE_EXTERNAL)
 		return os_timer_one_shot(delta);
 
 	return 0;
@@ -335,8 +610,14 @@  static u64 timer_read(struct clocksource *cs)
 		 * stuck in loops that expect time to move more than the
 		 * exact requested sleep amount, e.g. python's socket server,
 		 * see https://bugs.python.org/issue37026.
+		 *
+		 * However, don't do that when we're in interrupt or such as
+		 * then we might recurse into our own processing, and get to
+		 * even more waiting, and that's not good - it messes up the
+		 * "what do I do next" and onstack event we use to know when
+		 * to return from time_travel_update_time().
 		 */
-		if (!irqs_disabled())
+		if (!irqs_disabled() && !in_interrupt() && !in_softirq())
 			time_travel_update_time(time_travel_time +
 						TIMER_MULTIPLIER,
 						false);
@@ -383,6 +664,8 @@  void read_persistent_clock64(struct timespec64 *ts)
 
 	if (time_travel_start_set)
 		nsecs = time_travel_start + time_travel_time;
+	else if (time_travel_mode == TT_MODE_EXTERNAL)
+		nsecs = time_travel_ext_req(UM_TIMETRAVEL_GET_TOD, -1);
 	else
 		nsecs = os_persistent_clock_emulation();
 
@@ -399,7 +682,8 @@  void __init time_init(void)
 #ifdef CONFIG_UML_TIME_TRAVEL_SUPPORT
 unsigned long calibrate_delay_is_known(void)
 {
-	if (time_travel_mode == TT_MODE_INFCPU)
+	if (time_travel_mode == TT_MODE_INFCPU ||
+	    time_travel_mode == TT_MODE_EXTERNAL)
 		return 1;
 	return 0;
 }
@@ -413,6 +697,13 @@  int setup_time_travel(char *str)
 		return 1;
 	}
 
+	if (strncmp(str, "=ext:", 5) == 0) {
+		time_travel_mode = TT_MODE_EXTERNAL;
+		timer_clockevent.name = "time-travel-timer-external";
+		timer_clocksource.name = "time-travel-clock-external";
+		return time_travel_connect_external(str + 5);
+	}
+
 	if (!*str) {
 		time_travel_mode = TT_MODE_BASIC;
 		timer_clockevent.name = "time-travel-timer";
@@ -437,7 +728,15 @@  __uml_help(setup_time_travel,
 "are no wall clock timers, and any CPU processing happens - as seen from the\n"
 "guest - instantly. This can be useful for accurate simulation regardless of\n"
 "debug overhead, physical CPU speed, etc. but is somewhat dangerous as it can\n"
-"easily lead to getting stuck (e.g. if anything in the system busy loops).\n");
+"easily lead to getting stuck (e.g. if anything in the system busy loops).\n"
+"\n"
+"time-travel=ext:[ID:]/path/to/socket\n"
+"This enables time travel mode similar to =inf-cpu, except the system will\n"
+"use the given socket to coordinate with a central scheduler, in order to\n"
+"have more than one system simultaneously be on simulated time. The virtio\n"
+"driver code in UML knows about this so you can also simulate networks and\n"
+"devices using it, assuming the device has the right capabilities.\n"
+"The optional ID is a 64-bit integer that's sent to the central scheduler.\n");
 
 int setup_time_travel_start(char *str)
 {
diff --git a/arch/um/os-Linux/file.c b/arch/um/os-Linux/file.c
index 5133e3afb96f..408db12ea4ed 100644
--- a/arch/um/os-Linux/file.c
+++ b/arch/um/os-Linux/file.c
@@ -5,6 +5,7 @@ 
 
 #include <stdio.h>
 #include <unistd.h>
+#include <stdlib.h>
 #include <errno.h>
 #include <fcntl.h>
 #include <signal.h>
@@ -16,6 +17,7 @@ 
 #include <sys/un.h>
 #include <sys/types.h>
 #include <sys/eventfd.h>
+#include <poll.h>
 #include <os.h>
 
 static void copy_stat(struct uml_stat *dst, const struct stat64 *src)
@@ -664,3 +666,31 @@  int os_sendmsg_fds(int fd, const void *buf, unsigned int len, const int *fds,
 		return -errno;
 	return err;
 }
+
+int os_poll(unsigned int n, const int *fds)
+{
+	/* currently need 2 FDs at most so avoid dynamic allocation */
+	struct pollfd pollfds[2] = {};
+	unsigned int i;
+	int ret;
+
+	if (n > ARRAY_SIZE(pollfds))
+		return -EINVAL;
+
+	for (i = 0; i < n; i++) {
+		pollfds[i].fd = fds[i];
+		pollfds[i].events = POLLIN;
+	}
+
+	ret = poll(pollfds, n, -1);
+	if (ret < 0)
+		return -errno;
+
+	/* Return the index of the available FD */
+	for (i = 0; i < n; i++) {
+		if (pollfds[i].revents)
+			return i;
+	}
+
+	return -EIO;
+}
diff --git a/include/uapi/linux/um_timetravel.h b/include/uapi/linux/um_timetravel.h
new file mode 100644
index 000000000000..ca3238222b6d
--- /dev/null
+++ b/include/uapi/linux/um_timetravel.h
@@ -0,0 +1,128 @@ 
+/*
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ *
+ * Copyright (C) 2019 Intel Corporation
+ */
+#ifndef _UAPI_LINUX_UM_TIMETRAVEL_H
+#define _UAPI_LINUX_UM_TIMETRAVEL_H
+#include <linux/types.h>
+
+/**
+ * struct um_timetravel_msg - UM time travel message
+ *
+ * This is the basic message type, going in both directions.
+ *
+ * This is the message passed between the host (user-mode Linux instance)
+ * and the calendar (the application on the other side of the socket) in
+ * order to implement common scheduling.
+ *
+ * Whenever UML has an event it will request runtime for it from the
+ * calendar, and then wait for its turn until it can run, etc. Note
+ * that it will only ever request the single next runtime, i.e. multiple
+ * REQUEST messages override each other.
+ */
+struct um_timetravel_msg {
+	/**
+	 * @op: operation value from &enum um_timetravel_ops
+	 */
+	__u32 op;
+
+	/**
+	 * @seq: sequence number for the message - shall be reflected in
+	 *	the ACK response, and should be checked while processing
+	 *	the response to see if it matches
+	 */
+	__u32 seq;
+
+	/**
+	 * @time: time in nanoseconds
+	 */
+	__u64 time;
+};
+
+/**
+ * enum um_timetravel_ops - Operation codes
+ */
+enum um_timetravel_ops {
+	/**
+	 * @UM_TIMETRAVEL_ACK: response (ACK) to any previous message,
+	 *	this usually doesn't carry any data in the 'time' field
+	 *	unless otherwise specified below
+	 */
+	UM_TIMETRAVEL_ACK		= 0,
+
+	/**
+	 * @UM_TIMETRAVEL_START: initialize the connection, the time
+	 *	field contains an (arbitrary) ID to possibly be able
+	 *	to distinguish the connections.
+	 */
+	UM_TIMETRAVEL_START		= 1,
+
+	/**
+	 * @UM_TIMETRAVEL_REQUEST: request to run at the given time
+	 *	(host -> calendar)
+	 */
+	UM_TIMETRAVEL_REQUEST		= 2,
+
+	/**
+	 * @UM_TIMETRAVEL_WAIT: Indicate waiting for the previously requested
+	 *	runtime, new requests may be made while waiting (e.g. due to
+	 *	interrupts); the time field is ignored. The calendar must process
+	 *	this message and later	send a %UM_TIMETRAVEL_RUN message when
+	 *	the host can run again.
+	 *	(host -> calendar)
+	 */
+	UM_TIMETRAVEL_WAIT		= 3,
+
+	/**
+	 * @UM_TIMETRAVEL_GET: return the current time from the calendar in the
+	 *	ACK message, the time in the request message is ignored
+	 *	(host -> calendar)
+	 */
+	UM_TIMETRAVEL_GET		= 4,
+
+	/**
+	 * @UM_TIMETRAVEL_UPDATE: time update to the calendar, must be sent e.g.
+	 *	before kicking an interrupt to another calendar
+	 *	(host -> calendar)
+	 */
+	UM_TIMETRAVEL_UPDATE		= 5,
+
+	/**
+	 * @UM_TIMETRAVEL_RUN: run time request granted, current time is in
+	 *	the time field
+	 *	(calendar -> host)
+	 */
+	UM_TIMETRAVEL_RUN		= 6,
+
+	/**
+	 * @UM_TIMETRAVEL_FREE_UNTIL: Enable free-running until the given time,
+	 *	this is a message from the calendar telling the host that it can
+	 *	freely do its own scheduling for anything before the indicated
+	 *	time.
+	 *	Note that if a calendar sends this message once, the host may
+	 *	assume that it will also do so in the future, if it implements
+	 *	wraparound semantics for the time field.
+	 *	(calendar -> host)
+	 */
+	UM_TIMETRAVEL_FREE_UNTIL	= 7,
+
+	/**
+	 * @UM_TIMETRAVEL_GET_TOD: Return time of day, typically used once at
+	 *	boot by the virtual machines to get a synchronized time from
+	 *	the simulation.
+	 */
+	UM_TIMETRAVEL_GET_TOD		= 8,
+};
+
+#endif /* _UAPI_LINUX_UM_TIMETRAVEL_H */