diff mbox series

[v2,3/3] xen: perform XenDevice clean-up in XenBus watch handler

Message ID 20190912121646.29148-4-paul.durrant@citrix.com (mailing list archive)
State New, archived
Headers show
Series xen: fix a potential crash in xen-bus | expand

Commit Message

Paul Durrant Sept. 12, 2019, 12:16 p.m. UTC
Cleaning up offine XenDevice objects directly in
xen_device_backend_changed() is dangerous as xen_device_unrealize() will
modify the watch list that is being walked. Even the QLIST_FOREACH_SAFE()
used in notifier_list_notify() is insufficient as *two* notifiers (for
the frontend and backend watches) are removed, thus potentially rendering
the 'next' pointer unsafe.

The solution is to use the XenBus backend_watch handler to do the clean-up
instead, as it is invoked whilst walking a separate watch list.

This patch therefore adds a new 'offline_devices' list to XenBus, to which
offline devices are added by xen_device_backend_changed(). The XenBus
backend_watch registration is also changed to not only invoke
xen_bus_enumerate() but also a new xen_bus_cleanup() function, which will
walk 'offline_devices' and perform the necessary actions.
For safety a an extra 'online' check is also added to
xen_bus_type_enumerate() to make sure that no attempt is made to create a
new XenDevice object for a backend that is offline.

NOTE: This patch also include some cosmetic changes:
      - substitute the local variable name 'backend_state'
        in xen_bus_type_enumerate() with 'state', since there
        is no ambiguity with any other state in that context.
      - change xen_device_state_is_active() to
        xen_device_frontend_is_active() (and pass a XenDevice directly)
        since the state tests contained therein only apply to a frontend.
      - use 'state' rather then 'xendev->backend_state' in
        xen_device_backend_changed() to shorten the code.

Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
---
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Anthony Perard <anthony.perard@citrix.com>

v2:
 - Make sure we don't try to add a XenDevice to 'offline_devices' more than
   once
---
 hw/xen/trace-events      |  2 +
 hw/xen/xen-bus.c         | 92 +++++++++++++++++++++++++++++-----------
 include/hw/xen/xen-bus.h |  2 +
 3 files changed, 71 insertions(+), 25 deletions(-)

Comments

Anthony PERARD Sept. 12, 2019, 3:04 p.m. UTC | #1
On Thu, Sep 12, 2019 at 01:16:46PM +0100, Paul Durrant wrote:
> Cleaning up offine XenDevice objects directly in
              ^ offline

> xen_device_backend_changed() is dangerous as xen_device_unrealize() will
> modify the watch list that is being walked. Even the QLIST_FOREACH_SAFE()
> used in notifier_list_notify() is insufficient as *two* notifiers (for
> the frontend and backend watches) are removed, thus potentially rendering
> the 'next' pointer unsafe.
> 
> The solution is to use the XenBus backend_watch handler to do the clean-up
> instead, as it is invoked whilst walking a separate watch list.
> 
> This patch therefore adds a new 'offline_devices' list to XenBus, to which
> offline devices are added by xen_device_backend_changed(). The XenBus
> backend_watch registration is also changed to not only invoke
> xen_bus_enumerate() but also a new xen_bus_cleanup() function, which will
> walk 'offline_devices' and perform the necessary actions.
> For safety a an extra 'online' check is also added to
             ^ one 'a' too many?

> xen_bus_type_enumerate() to make sure that no attempt is made to create a
> new XenDevice object for a backend that is offline.
> 
> NOTE: This patch also include some cosmetic changes:
>       - substitute the local variable name 'backend_state'
>         in xen_bus_type_enumerate() with 'state', since there
>         is no ambiguity with any other state in that context.
>       - change xen_device_state_is_active() to
>         xen_device_frontend_is_active() (and pass a XenDevice directly)
>         since the state tests contained therein only apply to a frontend.
>       - use 'state' rather then 'xendev->backend_state' in
>         xen_device_backend_changed() to shorten the code.
> 
> Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
> ---
> 
> v2:
>  - Make sure we don't try to add a XenDevice to 'offline_devices' more than
>    once
> ---
>  
>      /*
>       * If a backend is still 'online' then we should leave it alone but,
> -     * if a backend is not 'online', then the device should be destroyed
> -     * once the state is Closed.
> +     * if a backend is not 'online', then the device is a candidate
> +     * for destruction. Hence add it to the 'offline' list to be cleaned
> +     * by xen_bus_cleanup().
>       */
> -    if (!xendev->backend_online &&
> -        (xendev->backend_state == XenbusStateClosed ||
> -         xendev->backend_state == XenbusStateInitialising ||
> -         xendev->backend_state == XenbusStateInitWait ||
> -         xendev->backend_state == XenbusStateUnknown)) {
> -        Error *local_err = NULL;
> +    if (!online &&
> +        (state == XenbusStateClosed ||  state == XenbusStateInitialising ||
> +         state == XenbusStateInitWait || state == XenbusStateUnknown) &&
> +        !QLIST_NEXT(xendev, list)) {

Could you add a note about this QLIST_NEXT? I don't think it's going to
be obvious enough why we check that there are no next item. I might only
understand it just because of your reply to the v1 of the patch.
(Well the changelog of the patch also point out what it's for.)

> +        XenBus *xenbus = XEN_BUS(qdev_get_parent_bus(DEVICE(xendev)));
>  
> -        if (!xen_backend_try_device_destroy(xendev, &local_err)) {
> -            object_unparent(OBJECT(xendev));
> -        }
> +        QLIST_INSERT_HEAD(&xenbus->offline_devices, xendev, list);
>  
> -        if (local_err) {
> -            error_report_err(local_err);
> -        }
> +        /*
> +         * Re-write the state to cause a XenBus backend_watch notification,
> +         * resulting in a call to xen_bus_cleanup().
> +         */
> +        xen_device_backend_printf(xendev, "state", "%u", state);

It kind of feels wrong to rely on xenstore to notify QEMU's xenbus
driver that a device needs cleanup. But that does the job.

With a note about the use of QLIST_NEXT and the few typo fixed, the
patch will be good to go.

Thanks,
Paul Durrant Sept. 12, 2019, 3:15 p.m. UTC | #2
> -----Original Message-----
> From: Anthony PERARD <anthony.perard@citrix.com>
> Sent: 12 September 2019 16:04
> To: Paul Durrant <Paul.Durrant@citrix.com>
> Cc: qemu-devel@nongnu.org; xen-devel@lists.xenproject.org; Stefano Stabellini <sstabellini@kernel.org>
> Subject: Re: [PATCH v2 3/3] xen: perform XenDevice clean-up in XenBus watch handler
> 
> On Thu, Sep 12, 2019 at 01:16:46PM +0100, Paul Durrant wrote:
> > Cleaning up offine XenDevice objects directly in
>               ^ offline
> 
> > xen_device_backend_changed() is dangerous as xen_device_unrealize() will
> > modify the watch list that is being walked. Even the QLIST_FOREACH_SAFE()
> > used in notifier_list_notify() is insufficient as *two* notifiers (for
> > the frontend and backend watches) are removed, thus potentially rendering
> > the 'next' pointer unsafe.
> >
> > The solution is to use the XenBus backend_watch handler to do the clean-up
> > instead, as it is invoked whilst walking a separate watch list.
> >
> > This patch therefore adds a new 'offline_devices' list to XenBus, to which
> > offline devices are added by xen_device_backend_changed(). The XenBus
> > backend_watch registration is also changed to not only invoke
> > xen_bus_enumerate() but also a new xen_bus_cleanup() function, which will
> > walk 'offline_devices' and perform the necessary actions.
> > For safety a an extra 'online' check is also added to
>              ^ one 'a' too many?
> 
> > xen_bus_type_enumerate() to make sure that no attempt is made to create a
> > new XenDevice object for a backend that is offline.
> >
> > NOTE: This patch also include some cosmetic changes:
> >       - substitute the local variable name 'backend_state'
> >         in xen_bus_type_enumerate() with 'state', since there
> >         is no ambiguity with any other state in that context.
> >       - change xen_device_state_is_active() to
> >         xen_device_frontend_is_active() (and pass a XenDevice directly)
> >         since the state tests contained therein only apply to a frontend.
> >       - use 'state' rather then 'xendev->backend_state' in
> >         xen_device_backend_changed() to shorten the code.
> >
> > Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
> > ---
> >
> > v2:
> >  - Make sure we don't try to add a XenDevice to 'offline_devices' more than
> >    once
> > ---
> >
> >      /*
> >       * If a backend is still 'online' then we should leave it alone but,
> > -     * if a backend is not 'online', then the device should be destroyed
> > -     * once the state is Closed.
> > +     * if a backend is not 'online', then the device is a candidate
> > +     * for destruction. Hence add it to the 'offline' list to be cleaned
> > +     * by xen_bus_cleanup().
> >       */
> > -    if (!xendev->backend_online &&
> > -        (xendev->backend_state == XenbusStateClosed ||
> > -         xendev->backend_state == XenbusStateInitialising ||
> > -         xendev->backend_state == XenbusStateInitWait ||
> > -         xendev->backend_state == XenbusStateUnknown)) {
> > -        Error *local_err = NULL;
> > +    if (!online &&
> > +        (state == XenbusStateClosed ||  state == XenbusStateInitialising ||
> > +         state == XenbusStateInitWait || state == XenbusStateUnknown) &&
> > +        !QLIST_NEXT(xendev, list)) {
> 
> Could you add a note about this QLIST_NEXT? I don't think it's going to
> be obvious enough why we check that there are no next item. I might only
> understand it just because of your reply to the v1 of the patch.
> (Well the changelog of the patch also point out what it's for.)
> 

Sure, it's worth a comment.

> > +        XenBus *xenbus = XEN_BUS(qdev_get_parent_bus(DEVICE(xendev)));
> >
> > -        if (!xen_backend_try_device_destroy(xendev, &local_err)) {
> > -            object_unparent(OBJECT(xendev));
> > -        }
> > +        QLIST_INSERT_HEAD(&xenbus->offline_devices, xendev, list);
> >
> > -        if (local_err) {
> > -            error_report_err(local_err);
> > -        }
> > +        /*
> > +         * Re-write the state to cause a XenBus backend_watch notification,
> > +         * resulting in a call to xen_bus_cleanup().
> > +         */
> > +        xen_device_backend_printf(xendev, "state", "%u", state);
> 
> It kind of feels wrong to rely on xenstore to notify QEMU's xenbus
> driver that a device needs cleanup. But that does the job.
> 

I had originally considered setting up an event notifier but that seemed like more fds than were strictly necessary ;-)

> With a note about the use of QLIST_NEXT and the few typo fixed, the
> patch will be good to go.
> 

Cool. I'll tidy up and send a v3.

  Paul

> Thanks,
> 
> --
> Anthony PERARD
Paul Durrant Sept. 13, 2019, 8:03 a.m. UTC | #3
> -----Original Message-----
> From: Xen-devel <xen-devel-bounces@lists.xenproject.org> On Behalf Of Paul Durrant
> Sent: 12 September 2019 16:16
> To: Anthony Perard <anthony.perard@citrix.com>
> Cc: xen-devel@lists.xenproject.org; Stefano Stabellini <sstabellini@kernel.org>; qemu-devel@nongnu.org
> Subject: Re: [Xen-devel] [PATCH v2 3/3] xen: perform XenDevice clean-up in XenBus watch handler
> 
> > -----Original Message-----
> > From: Anthony PERARD <anthony.perard@citrix.com>
> > Sent: 12 September 2019 16:04
> > To: Paul Durrant <Paul.Durrant@citrix.com>
> > Cc: qemu-devel@nongnu.org; xen-devel@lists.xenproject.org; Stefano Stabellini
> <sstabellini@kernel.org>
> > Subject: Re: [PATCH v2 3/3] xen: perform XenDevice clean-up in XenBus watch handler
> >
> > On Thu, Sep 12, 2019 at 01:16:46PM +0100, Paul Durrant wrote:
> > > Cleaning up offine XenDevice objects directly in
> >               ^ offline
> >
> > > xen_device_backend_changed() is dangerous as xen_device_unrealize() will
> > > modify the watch list that is being walked. Even the QLIST_FOREACH_SAFE()
> > > used in notifier_list_notify() is insufficient as *two* notifiers (for
> > > the frontend and backend watches) are removed, thus potentially rendering
> > > the 'next' pointer unsafe.
> > >
> > > The solution is to use the XenBus backend_watch handler to do the clean-up
> > > instead, as it is invoked whilst walking a separate watch list.
> > >
> > > This patch therefore adds a new 'offline_devices' list to XenBus, to which
> > > offline devices are added by xen_device_backend_changed(). The XenBus
> > > backend_watch registration is also changed to not only invoke
> > > xen_bus_enumerate() but also a new xen_bus_cleanup() function, which will
> > > walk 'offline_devices' and perform the necessary actions.
> > > For safety a an extra 'online' check is also added to
> >              ^ one 'a' too many?
> >
> > > xen_bus_type_enumerate() to make sure that no attempt is made to create a
> > > new XenDevice object for a backend that is offline.
> > >
> > > NOTE: This patch also include some cosmetic changes:
> > >       - substitute the local variable name 'backend_state'
> > >         in xen_bus_type_enumerate() with 'state', since there
> > >         is no ambiguity with any other state in that context.
> > >       - change xen_device_state_is_active() to
> > >         xen_device_frontend_is_active() (and pass a XenDevice directly)
> > >         since the state tests contained therein only apply to a frontend.
> > >       - use 'state' rather then 'xendev->backend_state' in
> > >         xen_device_backend_changed() to shorten the code.
> > >
> > > Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
> > > ---
> > >
> > > v2:
> > >  - Make sure we don't try to add a XenDevice to 'offline_devices' more than
> > >    once
> > > ---
> > >
> > >      /*
> > >       * If a backend is still 'online' then we should leave it alone but,
> > > -     * if a backend is not 'online', then the device should be destroyed
> > > -     * once the state is Closed.
> > > +     * if a backend is not 'online', then the device is a candidate
> > > +     * for destruction. Hence add it to the 'offline' list to be cleaned
> > > +     * by xen_bus_cleanup().
> > >       */
> > > -    if (!xendev->backend_online &&
> > > -        (xendev->backend_state == XenbusStateClosed ||
> > > -         xendev->backend_state == XenbusStateInitialising ||
> > > -         xendev->backend_state == XenbusStateInitWait ||
> > > -         xendev->backend_state == XenbusStateUnknown)) {
> > > -        Error *local_err = NULL;
> > > +    if (!online &&
> > > +        (state == XenbusStateClosed ||  state == XenbusStateInitialising ||
> > > +         state == XenbusStateInitWait || state == XenbusStateUnknown) &&
> > > +        !QLIST_NEXT(xendev, list)) {
> >
> > Could you add a note about this QLIST_NEXT? I don't think it's going to
> > be obvious enough why we check that there are no next item. I might only
> > understand it just because of your reply to the v1 of the patch.
> > (Well the changelog of the patch also point out what it's for.)
> >
> 
> Sure, it's worth a comment.

Actually, on closer inspection !QLIST_NEXT() is an insufficient checked. I had assumed that QLISTs were full doubly-linked lists but they are not; the last element on a list will still have a NULL next pointer. It will be sufficient, and also clearer to the reader, if I add a boolean to XenDevice which is set when it is added to the list.

  Paul
diff mbox series

Patch

diff --git a/hw/xen/trace-events b/hw/xen/trace-events
index 80ce3dafad..e6885bc751 100644
--- a/hw/xen/trace-events
+++ b/hw/xen/trace-events
@@ -17,8 +17,10 @@  xen_domid_restrict(int err) "err: %u"
 xen_bus_realize(void) ""
 xen_bus_unrealize(void) ""
 xen_bus_enumerate(void) ""
+xen_bus_cleanup(void) ""
 xen_bus_type_enumerate(const char *type) "type: %s"
 xen_bus_backend_create(const char *type, const char *path) "type: %s path: %s"
+xen_bus_device_cleanup(const char *type, char *name) "type: %s name: %s"
 xen_bus_add_watch(const char *node, const char *key) "node: %s key: %s"
 xen_bus_remove_watch(const char *node, const char *key) "node: %s key: %s"
 xen_device_realize(const char *type, char *name) "type: %s name: %s"
diff --git a/hw/xen/xen-bus.c b/hw/xen/xen-bus.c
index 810a4e2df3..43e29b8771 100644
--- a/hw/xen/xen-bus.c
+++ b/hw/xen/xen-bus.c
@@ -340,13 +340,18 @@  static void xen_bus_type_enumerate(XenBus *xenbus, const char *type)
     for (i = 0; i < n; i++) {
         char *backend_path = g_strdup_printf("%s/%s", domain_path,
                                              backend[i]);
-        enum xenbus_state backend_state;
+        enum xenbus_state state;
+        unsigned int online;
 
         if (xs_node_scanf(xenbus->xsh, XBT_NULL, backend_path, "state",
-                          NULL, "%u", &backend_state) != 1)
-            backend_state = XenbusStateUnknown;
+                          NULL, "%u", &state) != 1)
+            state = XenbusStateUnknown;
 
-        if (backend_state == XenbusStateInitialising) {
+        if (xs_node_scanf(xenbus->xsh, XBT_NULL, backend_path, "online",
+                          NULL, "%u", &online) != 1)
+            online = 0;
+
+        if (online && state == XenbusStateInitialising) {
             Error *local_err = NULL;
 
             xen_bus_backend_create(xenbus, type, backend[i], backend_path,
@@ -365,9 +370,8 @@  out:
     g_free(domain_path);
 }
 
-static void xen_bus_enumerate(void *opaque)
+static void xen_bus_enumerate(XenBus *xenbus)
 {
-    XenBus *xenbus = opaque;
     char **type;
     unsigned int i, n;
 
@@ -385,6 +389,44 @@  static void xen_bus_enumerate(void *opaque)
     free(type);
 }
 
+static void xen_bus_device_cleanup(XenDevice *xendev)
+{
+    const char *type = object_get_typename(OBJECT(xendev));
+    Error *local_err = NULL;
+
+    trace_xen_bus_device_cleanup(type, xendev->name);
+
+    g_assert(!xendev->backend_online);
+
+    if (!xen_backend_try_device_destroy(xendev, &local_err)) {
+        object_unparent(OBJECT(xendev));
+    }
+
+    if (local_err) {
+        error_report_err(local_err);
+    }
+}
+
+static void xen_bus_cleanup(XenBus *xenbus)
+{
+    XenDevice *xendev, *next;
+
+    trace_xen_bus_cleanup();
+
+    QLIST_FOREACH_SAFE(xendev, &xenbus->offline_devices, list, next) {
+        QLIST_REMOVE(xendev, list);
+        xen_bus_device_cleanup(xendev);
+    }
+}
+
+static void xen_bus_backend_changed(void *opaque)
+{
+    XenBus *xenbus = opaque;
+
+    xen_bus_enumerate(xenbus);
+    xen_bus_cleanup(xenbus);
+}
+
 static void xen_bus_unrealize(BusState *bus, Error **errp)
 {
     XenBus *xenbus = XEN_BUS(bus);
@@ -433,7 +475,7 @@  static void xen_bus_realize(BusState *bus, Error **errp)
 
     xenbus->backend_watch =
         xen_bus_add_watch(xenbus, "", /* domain root node */
-                          "backend", xen_bus_enumerate, &local_err);
+                          "backend", xen_bus_backend_changed, &local_err);
     if (local_err) {
         /* This need not be treated as a hard error so don't propagate */
         error_reportf_err(local_err,
@@ -555,9 +597,9 @@  static void xen_device_backend_set_online(XenDevice *xendev, bool online)
  * Tell from the state whether the frontend is likely alive,
  * i.e. it will react to a change of state of the backend.
  */
-static bool xen_device_state_is_active(enum xenbus_state state)
+static bool xen_device_frontend_is_active(XenDevice *xendev)
 {
-    switch (state) {
+    switch (xendev->frontend_state) {
     case XenbusStateInitWait:
     case XenbusStateInitialised:
     case XenbusStateConnected:
@@ -594,30 +636,30 @@  static void xen_device_backend_changed(void *opaque)
      * state to Closing, but there is no active frontend then set the
      * backend state to Closed.
      */
-    if (xendev->backend_state == XenbusStateClosing &&
-        !xen_device_state_is_active(xendev->frontend_state)) {
+    if (state == XenbusStateClosing &&
+        !xen_device_frontend_is_active(xendev)) {
         xen_device_backend_set_state(xendev, XenbusStateClosed);
     }
 
     /*
      * If a backend is still 'online' then we should leave it alone but,
-     * if a backend is not 'online', then the device should be destroyed
-     * once the state is Closed.
+     * if a backend is not 'online', then the device is a candidate
+     * for destruction. Hence add it to the 'offline' list to be cleaned
+     * by xen_bus_cleanup().
      */
-    if (!xendev->backend_online &&
-        (xendev->backend_state == XenbusStateClosed ||
-         xendev->backend_state == XenbusStateInitialising ||
-         xendev->backend_state == XenbusStateInitWait ||
-         xendev->backend_state == XenbusStateUnknown)) {
-        Error *local_err = NULL;
+    if (!online &&
+        (state == XenbusStateClosed ||  state == XenbusStateInitialising ||
+         state == XenbusStateInitWait || state == XenbusStateUnknown) &&
+        !QLIST_NEXT(xendev, list)) {
+        XenBus *xenbus = XEN_BUS(qdev_get_parent_bus(DEVICE(xendev)));
 
-        if (!xen_backend_try_device_destroy(xendev, &local_err)) {
-            object_unparent(OBJECT(xendev));
-        }
+        QLIST_INSERT_HEAD(&xenbus->offline_devices, xendev, list);
 
-        if (local_err) {
-            error_report_err(local_err);
-        }
+        /*
+         * Re-write the state to cause a XenBus backend_watch notification,
+         * resulting in a call to xen_bus_cleanup().
+         */
+        xen_device_backend_printf(xendev, "state", "%u", state);
     }
 }
 
diff --git a/include/hw/xen/xen-bus.h b/include/hw/xen/xen-bus.h
index 0d198148f6..15d71aff70 100644
--- a/include/hw/xen/xen-bus.h
+++ b/include/hw/xen/xen-bus.h
@@ -33,6 +33,7 @@  typedef struct XenDevice {
     xengnttab_handle *xgth;
     bool feature_grant_copy;
     QLIST_HEAD(, XenEventChannel) event_channels;
+    QLIST_ENTRY(XenDevice) list;
 } XenDevice;
 
 typedef char *(*XenDeviceGetName)(XenDevice *xendev, Error **errp);
@@ -68,6 +69,7 @@  typedef struct XenBus {
     struct xs_handle *xsh;
     XenWatchList *watch_list;
     XenWatch *backend_watch;
+    QLIST_HEAD(, XenDevice) offline_devices;
 } XenBus;
 
 typedef struct XenBusClass {