kexec and kdump for Xen PVonHVM guests

Index: linux-3.0/drivers/xen/xenbus/xenbus_xs.c
===================================================================

drivers/xen/xenbus/xenbus_xs.c +3 -0

@@ -620,6 +620,20 @@ static struct xenbus_watch *find_watch(c
 	return NULL;
 }
 
+static int xs_introduce(const char *domid, const char *mfn, const char *port)
+{
+	struct kvec iov[3];

drivers/xen/xenbus/xenbus_xs.c +3 -1

+}
 /* Register callback to watch this node. */
 int register_xenbus_watch(struct xenbus_watch *watch)
 {

@@ -867,10 +881,11 @@ static int xenbus_thread(void *unused)
 	return 0;
 }
 
-int xs_init(void)
+int xs_init(unsigned long xen_store_mfn)
 {
 	int err;
 	struct task_struct *task;
+	char domid[12], mfn[24], port[24];
 
 	INIT_LIST_HEAD(&xs_state.reply_list);
 	spin_lock_init(&xs_state.reply_lock);

drivers/xen/xenbus/xenbus_xs.c +7 -0

+	if (xen_hvm_domain()) {
+		/* shutdown watches for kexec boot */
+		err = xs_introduce(domid, mfn, port);
+		if (err)
+			printk(KERN_WARNING "xs_introduce failed: %d\n", err);
+	}
+
 	return 0;
 }

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

drivers/xen/xenbus/xenbus_xs.c static struct xenbus_watch *find_watch(c · line 635

Quoted 4 lines

+++ linux-3.0/drivers/xen/xenbus/xenbus_xs.c

drivers/xen/xenbus/xenbus_xs.c +2 -0

+	return xs_error(xs_talkv(XBT_NIL, XS_INTRODUCE, iov,
+				 ARRAY_SIZE(iov), NULL));

What should we do if this fails?

drivers/xen/xenbus/xenbus_xs.c int xs_init(void) · line 917

Quoted 3 lines

drivers/xen/xenbus/xenbus_xs.c +3 -0

+	snprintf(domid, sizeof(domid), "%u", DOMID_SELF);
+	snprintf(mfn, sizeof(mfn), "%lu", xen_store_mfn);
+	snprintf(port, sizeof(port), "%d", xen_store_evtchn);

These can be within the if, or better within the xs_introduce function itself.

During a kexec boot some virqs such as timer and debugirq were already registered by the old kernel. The hypervisor will return -EEXISTS from the new EVTCHNOP_bind_virq request and the BUG in bind_virq_to_irq() triggers. Catch the -EEXISTS error and loop through all possible ports to find what port belongs to the virq/cpu combo.

Signed-off-by Olaf Hering <[email protected]>

---
 drivers/xen/events.c |   40 +++++++++++++++++++++++++++++++++++-----
 1 file changed, 35 insertions(+), 5 deletions(-)

Index: linux-3.0/drivers/xen/events.c
===================================================================

drivers/xen/events.c +35 -5

@@ -877,11 +877,35 @@ static int bind_interdomain_evtchn_to_ir
 	return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 }
 
+/* BITS_PER_LONG is used in Xen */
+#define MAX_EVTCHNS (64 * 64)
+
+static int find_virq(unsigned int virq, unsigned int cpu)
+{
+	struct evtchn_status status;
+	int port, rc = -ENOENT;
+
+	memset(&status, 0, sizeof(status));
+	for (port = 0; port <= MAX_EVTCHNS; port++) {
+		status.dom = DOMID_SELF;
+		status.port = port;
+		rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
+		if (rc < 0)
+			continue;
+		if (status.status != EVTCHNSTAT_virq)
+			continue;
+		if (status.u.virq == virq && status.vcpu == cpu) {
+			rc = port;
+			break;
+		}
+	}
+	return rc;
+}
 
 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 {
 	struct evtchn_bind_virq bind_virq;
-	int evtchn, irq;
+	int evtchn, irq, ret;
 
 	spin_lock(&irq_mapping_update_lock);
 

@@ -897,10 +921,16 @@ int bind_virq_to_irq(unsigned int virq,
 
 		bind_virq.virq = virq;
 		bind_virq.vcpu = cpu;
-		if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
-						&bind_virq) != 0)
-			BUG();
-		evtchn = bind_virq.port;
+		ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
+						&bind_virq);
+		if (ret == 0)
+			evtchn = bind_virq.port;
+		else {
+			if (ret == -EEXIST)
+				ret = find_virq(virq, cpu);
+			BUG_ON(ret < 0);
+			evtchn = ret;
+		}
 
 		xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Quoted 5 lines · Olaf Hering

During a kexec boot some virqs such as timer and debugirq were already registered by the old kernel. The hypervisor will return -EEXISTS from the new EVTCHNOP_bind_virq request and the BUG in bind_virq_to_irq() triggers. Catch the -EEXISTS error and loop through all possible ports to find what port belongs to the virq/cpu combo.

Would it be better to proactively just query the status of all event channels early on, like you do in find_virq, and setup the irq info structures as appropriate? Rather than waiting for an -EEXISTS I mean.

Quoted 7 lines · Ian Campbell

Quoted 5 lines · Olaf Hering

During a kexec boot some virqs such as timer and debugirq were already registered by the old kernel. The hypervisor will return -EEXISTS from the new EVTCHNOP_bind_virq request and the BUG in bind_virq_to_irq() triggers. Catch the -EEXISTS error and loop through all possible ports to find what port belongs to the virq/cpu combo.

Quoted 3 lines

Would it be better to proactively just query the status of all event channels early on, like you do in find_virq, and setup the irq info structures as appropriate? Rather than waiting for an -EEXISTS I mean.

Doing one hypercall in the common case is cheaper than doing a dozen in the kexec case.

If you prefer I will rearrange this part and query first.

Olaf

Quoted 5 lines · Olaf Hering

During a kexec boot some virqs such as timer and debugirq were already registered by the old kernel. The hypervisor will return -EEXISTS from the new EVTCHNOP_bind_virq request and the BUG in bind_virq_to_irq() triggers. Catch the -EEXISTS error and loop through all possible ports to find what port belongs to the virq/cpu combo.

Would it be better to proactively just query the status of all event channels early on, like you do in find_virq, and setup the irq info structures as appropriate? Rather than waiting for an -EEXISTS I mean.

Quoted 7 lines · Ian Campbell

Quoted 5 lines · Olaf Hering

During a kexec boot some virqs such as timer and debugirq were already registered by the old kernel. The hypervisor will return -EEXISTS from the new EVTCHNOP_bind_virq request and the BUG in bind_virq_to_irq() triggers. Catch the -EEXISTS error and loop through all possible ports to find what port belongs to the virq/cpu combo.

Quoted 3 lines

Would it be better to proactively just query the status of all event channels early on, like you do in find_virq, and setup the irq info structures as appropriate? Rather than waiting for an -EEXISTS I mean.

Now that I read that again more carefully:
No idea if thats possible, I will leave that answer to Jeremy/Konrad.

After triggering a crash dump in a HVM guest, the PV backend drivers will remain in Connected state. When the kdump kernel starts the PV drivers will skip such devices. As a result, no root device is found and the vmcore cant be saved.

A similar situation happens after a kexec boot, here the devices will be in the Closed state.

With this change all frontend devices with state XenbusStateConnected or XenbusStateClosed will be reset by changing the state file to Closing/Closed/Initializing. This will trigger a disconnect in the backend drivers. Now the frontend drivers will find the backend drivers in state Initwait and can connect.

Signed-off-by Olaf Hering <[email protected]>

---
 drivers/xen/xenbus/xenbus_comms.c          |    4 -
 drivers/xen/xenbus/xenbus_probe_frontend.c |  116 +++++++++++++++++++++++++++++
 2 files changed, 119 insertions(+), 1 deletion(-)

Index: linux-3.0/drivers/xen/xenbus/xenbus_comms.c
===================================================================

drivers/xen/xenbus/xenbus_comms.c +3 -1

@@ -212,7 +212,9 @@ int xb_init_comms(void)
 		printk(KERN_WARNING "XENBUS response ring is not quiescent "
 		       "(%08x:%08x): fixing up\n",
 		       intf->rsp_cons, intf->rsp_prod);
-		intf->rsp_cons = intf->rsp_prod;
+		/* breaks kdump */
+		if (!reset_devices)
+			intf->rsp_cons = intf->rsp_prod;
 	}
 
 	if (xenbus_irq) {

Index: linux-3.0/drivers/xen/xenbus/xenbus_probe_frontend.c
===================================================================

drivers/xen/xenbus/xenbus_probe_frontend.c +109 -0

@@ -252,10 +252,126 @@ int __xenbus_register_frontend(struct xe
 }
 EXPORT_SYMBOL_GPL(__xenbus_register_frontend);
 
+static DECLARE_WAIT_QUEUE_HEAD(backend_state_wq);
+static int backend_state;
+
+static void xenbus_reset_backend_state_changed(struct xenbus_watch *w,
+					const char **v, unsigned int l)
+{
+	xenbus_scanf(XBT_NIL, v[XS_WATCH_PATH], "", "%i", &backend_state);
+	printk(KERN_DEBUG "XENBUS: %s %s\n",
+			v[XS_WATCH_PATH], xenbus_strstate(backend_state));
+	wake_up(&backend_state_wq);
+}
+
+static void xenbus_reset_wait_for_backend(int expected)
+{
+	wait_event_interruptible(backend_state_wq, backend_state == expected);
+}
+
+/*
+ * Reset frontend if it is in Connected or Closed state.
+ * Wait for backend to catch up.
+ * State Connected happens during kdump, Closed after kexec.
+ */
+static void xenbus_reset_frontend(char *fe, char *be, int be_state)
+{
+	struct xenbus_watch be_watch;
+
+	printk(KERN_DEBUG "XENBUS: backend %s %s\n",
+			be, xenbus_strstate(be_state));
+
+	memset(&be_watch, 0, sizeof(be_watch));
+	be_watch.node = kasprintf(GFP_NOIO | __GFP_HIGH, "%s/state", be);
+	if (!be_watch.node)
+		return;
+
+	be_watch.callback = xenbus_reset_backend_state_changed;
+	backend_state = XenbusStateUnknown;
+
+	printk(KERN_INFO "XENBUS: triggering reconnect on %s\n", be);
+	register_xenbus_watch(&be_watch);
+
+	switch (be_state) {
+	case XenbusStateConnected:
+		xenbus_printf(XBT_NIL, fe, "state", "%d", XenbusStateClosing);
+		xenbus_reset_wait_for_backend(XenbusStateClosing);
+
+	case XenbusStateClosing:
+		xenbus_printf(XBT_NIL, fe, "state", "%d", XenbusStateClosed);
+		xenbus_reset_wait_for_backend(XenbusStateClosed);
+
+	case XenbusStateClosed:
+		xenbus_printf(XBT_NIL, fe, "state", "%d", XenbusStateInitialising);
+		xenbus_reset_wait_for_backend(XenbusStateInitWait);
+	}
+
+	unregister_xenbus_watch(&be_watch);
+	printk(KERN_INFO "XENBUS: reconnect done on %s\n", be);
+	kfree(be_watch.node);
+}
+
+static void xenbus_check_frontend(char *class, char *dev)
+{
+	int be_state, fe_state, err;
+	char *backend, *frontend;
+
+	frontend = kasprintf(GFP_NOIO | __GFP_HIGH, "device/%s/%s", class, dev);
+	if (!frontend)
+		return;
+
+	err = xenbus_scanf(XBT_NIL, frontend, "state", "%i", &fe_state);
+	if (err != 1)
+		goto out;
+
+	switch (fe_state) {
+	case XenbusStateConnected:
+	case XenbusStateClosed:
+		printk(KERN_DEBUG "XENBUS: frontend %s %s\n",
+				frontend, xenbus_strstate(fe_state));
+		backend = xenbus_read(XBT_NIL, frontend, "backend", NULL);
+		if (!backend || IS_ERR(backend))
+			goto out;
+		err = xenbus_scanf(XBT_NIL, backend, "state", "%i", &be_state);
+		if (err == 1)
+			xenbus_reset_frontend(frontend, backend, be_state);
+		kfree(backend);
+		break;
+	default:
+		break;
+	}
+out:
+	kfree(frontend);
+}
+
+static void xenbus_reset_state(void)
+{
+	char **devclass, **dev;
+	int devclass_n, dev_n;
+	int i, j;
+
+	devclass = xenbus_directory(XBT_NIL, "device", "", &devclass_n);
+	if (IS_ERR(devclass))
+		return;
+
+	for (i = 0; i < devclass_n; i++) {
+		dev = xenbus_directory(XBT_NIL, "device", devclass[i], &dev_n);
+		if (IS_ERR(dev))
+			continue;
+		for (j = 0; j < dev_n; j++)
+			xenbus_check_frontend(devclass[i], dev[j]);
+		kfree(dev);

drivers/xen/xenbus/xenbus_probe_frontend.c +1 -0

+		xenbus_reset_state();
 	/* Enumerate devices in xenstore and watch for changes. */
 	xenbus_probe_devices(&xenbus_frontend);
 	register_xenbus_watch(&fe_watch);

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

drivers/xen/xenbus/xenbus_probe_frontend.c int __xenbus_register_frontend(struct xe · line 373

Quoted 6 lines

static int frontend_probe_and_watch(struct notifier_block *notifier,

drivers/xen/xenbus/xenbus_probe_frontend.c +2 -0

 				   unsigned long event,
 				   void *data)
 {
+	/* reset devices in Connected or Closed state */
+	if (xen_hvm_domain())

                               && reset_devices ??

How long should we wait for the backend to respond? Should we add a timeout and countdown similar to wait_for_devices?

It's unfortunate that this code is effectively serialising on each device. It would be much preferable to kick off all the resets and then wait for them to occur. You could probably do this by incrementing a counter for each device you reset and decrementing it each time a watch triggers then wait for the counter to hit zero.

Quoted 7 lines · Ian Campbell

drivers/xen/xenbus/xenbus_probe_frontend.c int __xenbus_register_frontend(struct xe · line 373

Quoted 6 lines

static int frontend_probe_and_watch(struct notifier_block *notifier,

drivers/xen/xenbus/xenbus_probe_frontend.c +2 -0

 				   unsigned long event,
 				   void *data)
 {
+	/* reset devices in Connected or Closed state */
+	if (xen_hvm_domain())

                               && reset_devices ??

No, reset_devices is passed as kernel cmdline option to a kdump boot. But its not part of a kexec boot.

Quoted 2 lines

How long should we wait for the backend to respond? Should we add a timeout and countdown similar to wait_for_devices?

Adding a timeout to catch a confused backend is a good idea. That would give one at least a chance to poke around in a rescue shell.

Quoted 5 lines

It's unfortunate that this code is effectively serialising on each device. It would be much preferable to kick off all the resets and then wait for them to occur. You could probably do this by incrementing a counter for each device you reset and decrementing it each time a watch triggers then wait for the counter to hit zero.

That feature needs more thought. Since xenbus_reset_state() is only executed in the kexec/kdump case, the average use case is not slowed down.

Olaf

The following series implements kexec and kdump in a Xen PVonHVM guest.

It is (should be) available via git:

git://github.com/olafhering/linux.git xen-kexec-3.0

The kexec or kdump kernel has to take care of already allocated virqs, PV devices in Closed or Connected state, and of registered watches in the old kernel. With the three patches these conditions are checked during boot of the new kernel rather than in the reboot/crash path.

A fixed kexec-tools-2.0.2 package is required:
http://lists.infradead.org/pipermail/kexec/2011-May/005026.html
http://lists.infradead.org/pipermail/kexec/2011-August/005339.html

Another fix is for xenstored, it has to accept the XS_INTRODUCE from a guest:
http://lists.xensource.com/archives/html/xen-devel/2011-08/msg00007.html

One open issue is the balloon driver. It removes pages from the guest and gives them back to the hypervisor. The kexec kernel is not aware of the fact that some pages are unavailable, and hangs or crashes. The workaround for the time being is:

if test -f /sys/devices/system/xen_memory/xen_memory0/target -a \
-f /sys/devices/system/xen_memory/xen_memory0/target_kb
then

    cat /sys/devices/system/xen_memory/xen_memory0/target > \
        /sys/devices/system/xen_memory/xen_memory0/target_kb

fi
kexec -e

This has to be resolved with another series of changes.