[PATCH 0/2 v4] libxl: fix event races exposed by libvirt

[PATCH 0/2 v4] libxl: fix event races exposed by libvirt

[Ian Jackson]

This is the final version of these patches which are needed to make
Jim Fehlig's libvirt libxl driver work.

 1/2 libxl: fix stale fd event callback race
 2/2 libxl: fix stale timeout event callback race

They are identical to the previous posting apart from commit messages
and fixing the bug Jim mentioned in <50FD98DC.4020203@suse.com>.

[PATCH 1/2] libxl: fix stale fd event callback race

[Ian Jackson]

Because there is not necessarily any lock held at the point the
application (eg, libvirt) calls libxl_osevent_occurred_timeout and
..._fd, in a multithreaded program those calls may be arbitrarily
delayed in relation to other activities within the program.

libxl therefore needs to be prepared to receive very old event
callbacks.  Arrange for this to be the case for fd callbacks.

This requires a new layer of indirection through a "hook nexus" struct
which can outlive the libxl__ev_foo.  Allocation and deallocation of
these nexi is mostly handled in the OSEVENT macros which wrap up
the application's callbacks.

Document the problem and the solution in a comment in libxl_event.c
just before the definition of struct libxl__osevent_hook_nexus.

There is still a race relating to libxl__osevent_occurred_timeout;
this will be addressed in the following patch.

Reported-by: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

--
v2:
  - Prepare for fixing timeout race too
  - Break out osevent_release_nexus()
  - nexusop argument to OSEVENT* macros
  - Clarify OSEVENT* nexusop hooks
  - osevent_ev_from_hook_nexus takes a libxl__osevent_hook_nexus*
---
 tools/libxl/libxl_event.c    |  184 +++++++++++++++++++++++++++++++++++------
 tools/libxl/libxl_internal.h |    8 ++-
 2 files changed, 163 insertions(+), 29 deletions(-)

diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
index 72cb723..f1fe425 100644
--- a/tools/libxl/libxl_event.c
+++ b/tools/libxl/libxl_event.c
@@ -38,23 +38,131 @@
  * The application's registration hooks should be called ONLY via
  * these macros, with the ctx locked.  Likewise all the "occurred"
  * entrypoints from the application should assert(!in_hook);
+ *
+ * During the hook call - including while the arguments are being
+ * evaluated - ev->nexus is guaranteed to be valid and refer to the
+ * nexus which is being used for this event registration.  The
+ * arguments should specify ev->nexus for the for_libxl argument and
+ * ev->nexus->for_app_reg (or a pointer to it) for for_app_reg.
  */
-#define OSEVENT_HOOK_INTERN(retval, hookname, ...) do {                      \
-    if (CTX->osevent_hooks) {                                                \
-        CTX->osevent_in_hook++;                                              \
-        retval CTX->osevent_hooks->hookname(CTX->osevent_user, __VA_ARGS__); \
-        CTX->osevent_in_hook--;                                              \
-    }                                                                        \
+#define OSEVENT_HOOK_INTERN(retval, failedp, evkind, hookop, nexusop, ...) do { \
+    if (CTX->osevent_hooks) {                                           \
+        CTX->osevent_in_hook++;                                         \
+        libxl__osevent_hook_nexi *nexi = &CTX->hook_##evkind##_nexi_idle; \
+        osevent_hook_pre_##nexusop(gc, ev, nexi, &ev->nexus);            \
+        retval CTX->osevent_hooks->evkind##_##hookop                    \
+            (CTX->osevent_user, __VA_ARGS__);                           \
+        if ((failedp))                                                  \
+            osevent_hook_failed_##nexusop(gc, ev, nexi, &ev->nexus);     \
+        CTX->osevent_in_hook--;                                         \
+    }                                                                   \
 } while (0)
 
-#define OSEVENT_HOOK(hookname, ...) ({                                       \
-    int osevent_hook_rc = 0;                                                 \
-    OSEVENT_HOOK_INTERN(osevent_hook_rc = , hookname, __VA_ARGS__);          \
-    osevent_hook_rc;                                                         \
+#define OSEVENT_HOOK(evkind, hookop, nexusop, ...) ({                   \
+    int osevent_hook_rc = 0;                                    \
+    OSEVENT_HOOK_INTERN(osevent_hook_rc =, !!osevent_hook_rc,   \
+                        evkind, hookop, nexusop, __VA_ARGS__);          \
+    osevent_hook_rc;                                            \
 })
 
-#define OSEVENT_HOOK_VOID(hookname, ...) \
-    OSEVENT_HOOK_INTERN(/* void */, hookname, __VA_ARGS__)
+#define OSEVENT_HOOK_VOID(evkind, hookop, nexusop, ...)                         \
+    OSEVENT_HOOK_INTERN(/* void */, 0, evkind, hookop, nexusop, __VA_ARGS__)
+
+/*
+ * The application's calls to libxl_osevent_occurred_... may be
+ * indefinitely delayed with respect to the rest of the program (since
+ * they are not necessarily called with any lock held).  So the
+ * for_libxl value we receive may be (almost) arbitrarily old.  All we
+ * know is that it came from this ctx.
+ *
+ * Therefore we may not free the object referred to by any for_libxl
+ * value until we free the whole libxl_ctx.  And if we reuse it we
+ * must be able to tell when an old use turns up, and discard the
+ * stale event.
+ *
+ * Thus we cannot use the ev directly as the for_libxl value - we need
+ * a layer of indirection.
+ *
+ * We do this by keeping a pool of libxl__osevent_hook_nexus structs,
+ * and use pointers to them as for_libxl values.  In fact, there are
+ * two pools: one for fds and one for timeouts.  This ensures that we
+ * don't risk a type error when we upcast nexus->ev.  In each nexus
+ * the ev is either null or points to a valid libxl__ev_time or
+ * libxl__ev_fd, as applicable.
+ *
+ * We /do/ allow ourselves to reassociate an old nexus with a new ev
+ * as otherwise we would have to leak nexi.  (This reassociation
+ * might, of course, be an old ev being reused for a new purpose so
+ * simply comparing the ev pointer is not sufficient.)  Thus the
+ * libxl_osevent_occurred functions need to check that the condition
+ * allegedly signalled by this event actually exists.
+ *
+ * The nexi and the lists are all protected by the ctx lock.
+ */
+ 
+struct libxl__osevent_hook_nexus {
+    void *ev;
+    void *for_app_reg;
+    LIBXL_SLIST_ENTRY(libxl__osevent_hook_nexus) next;
+};
+
+static void *osevent_ev_from_hook_nexus(libxl_ctx *ctx,
+           libxl__osevent_hook_nexus *nexus /* pass  void *for_libxl */)
+{
+    return nexus->ev;
+}
+
+static void osevent_release_nexus(libxl__gc *gc,
+                                  libxl__osevent_hook_nexi *nexi_idle,
+                                  libxl__osevent_hook_nexus *nexus)
+{
+    nexus->ev = 0;
+    LIBXL_SLIST_INSERT_HEAD(nexi_idle, nexus, next);
+}
+
+/*----- OSEVENT* hook functions for nexusop "alloc" -----*/
+static void osevent_hook_pre_alloc(libxl__gc *gc, void *ev,
+                                   libxl__osevent_hook_nexi *nexi_idle,
+                                   libxl__osevent_hook_nexus **nexus_r)
+{
+    libxl__osevent_hook_nexus *nexus = LIBXL_SLIST_FIRST(nexi_idle);
+    if (nexus) {
+        LIBXL_SLIST_REMOVE_HEAD(nexi_idle, next);
+    } else {
+        nexus = libxl__zalloc(NOGC, sizeof(*nexus));
+    }
+    nexus->ev = ev;
+    *nexus_r = nexus;
+}
+static void osevent_hook_failed_alloc(libxl__gc *gc, void *ev,
+                                      libxl__osevent_hook_nexi *nexi_idle,
+                                      libxl__osevent_hook_nexus **nexus)
+{
+    osevent_release_nexus(gc, nexi_idle, *nexus);
+}
+
+/*----- OSEVENT* hook functions for nexusop "release" -----*/
+static void osevent_hook_pre_release(libxl__gc *gc, void *ev,
+                                     libxl__osevent_hook_nexi *nexi_idle,
+                                     libxl__osevent_hook_nexus **nexus)
+{
+    osevent_release_nexus(gc, nexi_idle, *nexus);
+}
+static void osevent_hook_failed_release(libxl__gc *gc, void *ev,
+                                        libxl__osevent_hook_nexi *nexi_idle,
+                                        libxl__osevent_hook_nexus **nexus)
+{
+    abort();
+}
+
+/*----- OSEVENT* hook functions for nexusop "noop" -----*/
+static void osevent_hook_pre_noop(libxl__gc *gc, void *ev,
+                                  libxl__osevent_hook_nexi *nexi_idle,
+                                  libxl__osevent_hook_nexus **nexus) { }
+static void osevent_hook_failed_noop(libxl__gc *gc, void *ev,
+                                     libxl__osevent_hook_nexi *nexi_idle,
+                                     libxl__osevent_hook_nexus **nexus) { }
+
 
 /*
  * fd events
@@ -72,7 +180,8 @@ int libxl__ev_fd_register(libxl__gc *gc, libxl__ev_fd *ev,
 
     DBG("ev_fd=%p register fd=%d events=%x", ev, fd, events);
 
-    rc = OSEVENT_HOOK(fd_register, fd, &ev->for_app_reg, events, ev);
+    rc = OSEVENT_HOOK(fd,register, alloc, fd, &ev->nexus->for_app_reg,
+                      events, ev->nexus);
     if (rc) goto out;
 
     ev->fd = fd;
@@ -97,7 +206,7 @@ int libxl__ev_fd_modify(libxl__gc *gc, libxl__ev_fd *ev, short events)
 
     DBG("ev_fd=%p modify fd=%d events=%x", ev, ev->fd, events);
 
-    rc = OSEVENT_HOOK(fd_modify, ev->fd, &ev->for_app_reg, events);
+    rc = OSEVENT_HOOK(fd,modify, noop, ev->fd, &ev->nexus->for_app_reg, events);
     if (rc) goto out;
 
     ev->events = events;
@@ -119,7 +228,7 @@ void libxl__ev_fd_deregister(libxl__gc *gc, libxl__ev_fd *ev)
 
     DBG("ev_fd=%p deregister fd=%d", ev, ev->fd);
 
-    OSEVENT_HOOK_VOID(fd_deregister, ev->fd, ev->for_app_reg);
+    OSEVENT_HOOK_VOID(fd,deregister, release, ev->fd, ev->nexus->for_app_reg);
     LIBXL_LIST_REMOVE(ev, entry);
     ev->fd = -1;
 
@@ -171,7 +280,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 {
     int rc;
 
-    rc = OSEVENT_HOOK(timeout_register, &ev->for_app_reg, absolute, ev);
+    rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
+                      absolute, ev->nexus);
     if (rc) return rc;
 
     ev->infinite = 0;
@@ -184,7 +294,7 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     if (!ev->infinite) {
-        OSEVENT_HOOK_VOID(timeout_deregister, ev->for_app_reg);
+        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     }
 }
@@ -270,7 +380,8 @@ int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
         rc = time_register_finite(gc, ev, absolute);
         if (rc) goto out;
     } else {
-        rc = OSEVENT_HOOK(timeout_modify, &ev->for_app_reg, absolute);
+        rc = OSEVENT_HOOK(timeout,modify, noop,
+                          &ev->nexus->for_app_reg, absolute);
         if (rc) goto out;
 
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
@@ -1010,35 +1121,54 @@ void libxl_osevent_register_hooks(libxl_ctx *ctx,
 
 
 void libxl_osevent_occurred_fd(libxl_ctx *ctx, void *for_libxl,
-                               int fd, short events, short revents)
+                               int fd, short events_ign, short revents_ign)
 {
-    libxl__ev_fd *ev = for_libxl;
-
     EGC_INIT(ctx);
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
-    assert(fd == ev->fd);
-    revents &= ev->events;
-    if (revents)
-        ev->func(egc, ev, fd, ev->events, revents);
+    libxl__ev_fd *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    if (!ev) goto out;
+    if (ev->fd != fd) goto out;
 
+    struct pollfd check;
+    for (;;) {
+        check.fd = fd;
+        check.events = ev->events;
+        int r = poll(&check, 1, 0);
+        if (!r)
+            goto out;
+        if (r==1)
+            break;
+        assert(r<0);
+        if (errno != EINTR) {
+            LIBXL__EVENT_DISASTER(egc, "failed poll to check for fd", errno, 0);
+            goto out;
+        }
+    }
+
+    if (check.revents)
+        ev->func(egc, ev, fd, ev->events, check.revents);
+
+ out:
     CTX_UNLOCK;
     EGC_FREE;
 }
 
 void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
 {
-    libxl__ev_time *ev = for_libxl;
-
     EGC_INIT(ctx);
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
+    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    if (!ev) goto out;
     assert(!ev->infinite);
+
     LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     ev->func(egc, ev, &ev->abs);
 
+ out:
     CTX_UNLOCK;
     EGC_FREE;
 }
diff --git a/tools/libxl/libxl_internal.h b/tools/libxl/libxl_internal.h
index cba3616..6484bcb 100644
--- a/tools/libxl/libxl_internal.h
+++ b/tools/libxl/libxl_internal.h
@@ -136,6 +136,8 @@ typedef struct libxl__gc libxl__gc;
 typedef struct libxl__egc libxl__egc;
 typedef struct libxl__ao libxl__ao;
 typedef struct libxl__aop_occurred libxl__aop_occurred;
+typedef struct libxl__osevent_hook_nexus libxl__osevent_hook_nexus;
+typedef struct libxl__osevent_hook_nexi libxl__osevent_hook_nexi;
 
 _hidden void libxl__alloc_failed(libxl_ctx *, const char *func,
                          size_t nmemb, size_t size) __attribute__((noreturn));
@@ -163,7 +165,7 @@ struct libxl__ev_fd {
     libxl__ev_fd_callback *func;
     /* remainder is private for libxl__ev_fd... */
     LIBXL_LIST_ENTRY(libxl__ev_fd) entry;
-    void *for_app_reg;
+    libxl__osevent_hook_nexus *nexus;
 };
 
 
@@ -178,7 +180,7 @@ struct libxl__ev_time {
     int infinite; /* not registered in list or with app if infinite */
     LIBXL_TAILQ_ENTRY(libxl__ev_time) entry;
     struct timeval abs;
-    void *for_app_reg;
+    libxl__osevent_hook_nexus *nexus;
 };
 
 typedef struct libxl__ev_xswatch libxl__ev_xswatch;
@@ -329,6 +331,8 @@ struct libxl__ctx {
     libxl__poller poller_app; /* libxl_osevent_beforepoll and _afterpoll */
     LIBXL_LIST_HEAD(, libxl__poller) pollers_event, pollers_idle;
 
+    LIBXL_SLIST_HEAD(libxl__osevent_hook_nexi, libxl__osevent_hook_nexus)
+        hook_fd_nexi_idle, hook_timeout_nexi_idle;
     LIBXL_LIST_HEAD(, libxl__ev_fd) efds;
     LIBXL_TAILQ_HEAD(, libxl__ev_time) etimes;
 
-- 
1.7.2.5

[PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Because there is not necessarily any lock held at the point the
application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
multithreaded program those calls may be arbitrarily delayed in
relation to other activities within the program.

Specifically this means when ->timeout_deregister returns, libxl does
not know whether it can safely dispose of the for_libxl value or
whether it needs to retain it in case of an in-progress call to
_occurred_timeout.

The interface could be fixed by requiring the application to make a
new call into libxl to say that the deregistration was complete.

However that new call would have to be threaded through the
application's event loop; this is complicated and some application
authors are likely not to implement it properly.  Furthermore the
easiest way to implement this facility in most event loops is to queue
up a time event for "now".

Shortcut all of this by having libxl always call timeout_modify
setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
cause the application to call _occurred_timeout.  When processing this
calldown we see that we were no longer actually interested and simply
throw it away.

Additionally, there is a race between _occurred_timeout and
->timeout_modify.  If libxl ever adjusts the deadline for a timeout
the application may already be in the process of calling _occurred, in
which case the situation with for_app's lifetime becomes very
complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
have no callers) and promise to the application only ever to call
->timeout_modify with abs=={0,0}.  The application still needs to cope
with ->timeout_modify racing with its internal function which calls
_occurred_timeout.  Document this.

This is a forwards-compatible change for applications using the libxl
API, and will hopefully eliminate these races in callback-supplying
applications (such as libvirt) without the need for corresponding
changes to the application.

For clarity, fold the body of time_register_finite into its one
remaining call site.  This makes the semantics of ev->infinite
slightly clearer.

Cc: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

--
v3:
  - Fix null pointer dereference in case when hooks not supplied.
---
 tools/libxl/libxl_event.c |   89 +++++++--------------------------------------
 tools/libxl/libxl_event.h |   17 ++++++++-
 2 files changed, 29 insertions(+), 77 deletions(-)

diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
index f1fe425..f86c528 100644
--- a/tools/libxl/libxl_event.c
+++ b/tools/libxl/libxl_event.c
@@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
     return 0;
 }
 
-static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
-{
-    libxl__ev_time *evsearch;
-    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
-                              timercmp(&ev->abs, &evsearch->abs, >));
-    ev->infinite = 0;
-}
-
 static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
                                 struct timeval absolute)
 {
     int rc;
+    libxl__ev_time *evsearch;
 
     rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
                       absolute, ev->nexus);
@@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 
     ev->infinite = 0;
     ev->abs = absolute;
-    time_insert_finite(gc, ev);
+    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
+                              timercmp(&ev->abs, &evsearch->abs, >));
 
     return 0;
 }
@@ -294,7 +288,12 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     if (!ev->infinite) {
-        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
+        struct timeval right_away = { 0, 0 };
+        if (ev->nexus) /* only set if app provided hooks */
+            ev->nexus->ev = 0;
+        OSEVENT_HOOK_VOID(timeout,modify,
+                          noop /* release nexus in _occurred_ */,
+                          ev->nexus->for_app_reg, right_away);
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     }
 }
@@ -364,70 +363,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
     return rc;
 }
 
-int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
-                              struct timeval absolute)
-{
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify abs==%lu.%06lu",
-        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (ev->infinite) {
-        rc = time_register_finite(gc, ev, absolute);
-        if (rc) goto out;
-    } else {
-        rc = OSEVENT_HOOK(timeout,modify, noop,
-                          &ev->nexus->for_app_reg, absolute);
-        if (rc) goto out;
-
-        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
-        ev->abs = absolute;
-        time_insert_finite(gc, ev);
-    }
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
-int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
-                              int milliseconds)
-{
-    struct timeval absolute;
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (milliseconds < 0) {
-        time_deregister(gc, ev);
-        ev->infinite = 1;
-        rc = 0;
-        goto out;
-    }
-
-    rc = time_rel_to_abs(gc, milliseconds, &absolute);
-    if (rc) goto out;
-
-    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
-    if (rc) goto out;
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
 void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     CTX_LOCK;
@@ -1161,7 +1096,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
-    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    libxl__osevent_hook_nexus *nexus = for_libxl;
+    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
+
+    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
+
     if (!ev) goto out;
     assert(!ev->infinite);
 
diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
index 3bcb6d3..51f2721 100644
--- a/tools/libxl/libxl_event.h
+++ b/tools/libxl/libxl_event.h
@@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
   int (*timeout_register)(void *user, void **for_app_registration_out,
                           struct timeval abs, void *for_libxl);
   int (*timeout_modify)(void *user, void **for_app_registration_update,
-                         struct timeval abs);
-  void (*timeout_deregister)(void *user, void *for_app_registration);
+                         struct timeval abs)
+      /* only ever called with abs={0,0}, meaning ASAP */;
+  void (*timeout_deregister)(void *user, void *for_app_registration)
+      /* will never be called */;
 } libxl_osevent_hooks;
 
 /* The application which calls register_fd_hooks promises to
@@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
  * register (or modify), and pass it to subsequent calls to modify
  * or deregister.
  *
+ * Note that the application must cope with a call from libxl to
+ * timeout_modify racing with its own call to
+ * libxl__osevent_occurred_timeout.  libxl guarantees that
+ * timeout_modify will only be called with abs={0,0} but the
+ * application must still ensure that libxl's attempt to cause the
+ * timeout to occur immediately is safely ignored even the timeout is
+ * actually already in the process of occurring.
+ *
+ * timeout_deregister is not used because it forms part of a
+ * deprecated unsafe mode of use of the API.
+ *
  * osevent_register_hooks may be called only once for each libxl_ctx.
  * libxl may make calls to register/modify/deregister from within
  * any libxl function (indeed, it will usually call register from
-- 
1.7.2.5

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Because there is not necessarily any lock held at the point the
> application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
> multithreaded program those calls may be arbitrarily delayed in
> relation to other activities within the program.
>
> Specifically this means when ->timeout_deregister returns, libxl does
> not know whether it can safely dispose of the for_libxl value or
> whether it needs to retain it in case of an in-progress call to
> _occurred_timeout.
>
> The interface could be fixed by requiring the application to make a
> new call into libxl to say that the deregistration was complete.
>
> However that new call would have to be threaded through the
> application's event loop; this is complicated and some application
> authors are likely not to implement it properly.  Furthermore the
> easiest way to implement this facility in most event loops is to queue
> up a time event for "now".
>
> Shortcut all of this by having libxl always call timeout_modify
> setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
> cause the application to call _occurred_timeout.  When processing this
> calldown we see that we were no longer actually interested and simply
> throw it away.
>
> Additionally, there is a race between _occurred_timeout and
> ->timeout_modify.  If libxl ever adjusts the deadline for a timeout
> the application may already be in the process of calling _occurred, in
> which case the situation with for_app's lifetime becomes very
> complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
> have no callers) and promise to the application only ever to call
> ->timeout_modify with abs=={0,0}.  The application still needs to cope
> with ->timeout_modify racing with its internal function which calls
> _occurred_timeout.  Document this.
>
> This is a forwards-compatible change for applications using the libxl
> API, and will hopefully eliminate these races in callback-supplying
> applications (such as libvirt) without the need for corresponding
> changes to the application.
>
> For clarity, fold the body of time_register_finite into its one
> remaining call site.  This makes the semantics of ev->infinite
> slightly clearer.
>
> Cc: Bamvor Jian Zhang <bjzhang@suse.com>
> Cc: Ian Campbell <Ian.Campbell@citrix.com>
> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
>
> --
> v3:
>   - Fix null pointer dereference in case when hooks not supplied.
> ---
>  tools/libxl/libxl_event.c |   89 +++++++--------------------------------------
>  tools/libxl/libxl_event.h |   17 ++++++++-
>  2 files changed, 29 insertions(+), 77 deletions(-)
>
> diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
> index f1fe425..f86c528 100644
> --- a/tools/libxl/libxl_event.c
> +++ b/tools/libxl/libxl_event.c
> @@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
>      return 0;
>  }
>  
> -static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
> -{
> -    libxl__ev_time *evsearch;
> -    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
> -                              timercmp(&ev->abs, &evsearch->abs, >));
> -    ev->infinite = 0;
> -}
> -
>  static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
>                                  struct timeval absolute)
>  {
>      int rc;
> +    libxl__ev_time *evsearch;
>  
>      rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
>                        absolute, ev->nexus);
> @@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
>  
>      ev->infinite = 0;
>      ev->abs = absolute;
> -    time_insert_finite(gc, ev);
> +    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
> +                              timercmp(&ev->abs, &evsearch->abs, >));
>  
>      return 0;
>  }
> @@ -294,7 +288,12 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
>  static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
>  {
>      if (!ev->infinite) {
> -        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
> +        struct timeval right_away = { 0, 0 };
> +        if (ev->nexus) /* only set if app provided hooks */
> +            ev->nexus->ev = 0;
> +        OSEVENT_HOOK_VOID(timeout,modify,
> +                          noop /* release nexus in _occurred_ */,
> +                          ev->nexus->for_app_reg, right_away);
>   

I still hit a segfault in the modify hook.  The modify hook prototype is

  int (*timeout_modify)(void *user, void **for_app_registration_update,
                         struct timeval abs)

Shouldn't for_app_registration_update contain the data provided by the
app in timeout_register, with the option to update it?  With the below
change, I don't see the segfault.

+        OSEVENT_HOOK_VOID(timeout,modify,
+                          noop /* release nexus in _occurred_ */,
+                          &ev->nexus->for_app_reg, right_away);


Thanks,
Jim


>          LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
>      }
>  }
> @@ -364,70 +363,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
>      return rc;
>  }
>  
> -int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
> -                              struct timeval absolute)
> -{
> -    int rc;
> -
> -    CTX_LOCK;
> -
> -    DBG("ev_time=%p modify abs==%lu.%06lu",
> -        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
> -
> -    assert(libxl__ev_time_isregistered(ev));
> -
> -    if (ev->infinite) {
> -        rc = time_register_finite(gc, ev, absolute);
> -        if (rc) goto out;
> -    } else {
> -        rc = OSEVENT_HOOK(timeout,modify, noop,
> -                          &ev->nexus->for_app_reg, absolute);
> -        if (rc) goto out;
> -
> -        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
> -        ev->abs = absolute;
> -        time_insert_finite(gc, ev);
> -    }
> -
> -    rc = 0;
> - out:
> -    time_done_debug(gc,__func__,ev,rc);
> -    CTX_UNLOCK;
> -    return rc;
> -}
> -
> -int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
> -                              int milliseconds)
> -{
> -    struct timeval absolute;
> -    int rc;
> -
> -    CTX_LOCK;
> -
> -    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
> -
> -    assert(libxl__ev_time_isregistered(ev));
> -
> -    if (milliseconds < 0) {
> -        time_deregister(gc, ev);
> -        ev->infinite = 1;
> -        rc = 0;
> -        goto out;
> -    }
> -
> -    rc = time_rel_to_abs(gc, milliseconds, &absolute);
> -    if (rc) goto out;
> -
> -    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
> -    if (rc) goto out;
> -
> -    rc = 0;
> - out:
> -    time_done_debug(gc,__func__,ev,rc);
> -    CTX_UNLOCK;
> -    return rc;
> -}
> -
>  void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
>  {
>      CTX_LOCK;
> @@ -1161,7 +1096,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
>      CTX_LOCK;
>      assert(!CTX->osevent_in_hook);
>  
> -    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
> +    libxl__osevent_hook_nexus *nexus = for_libxl;
> +    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
> +
> +    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
> +
>      if (!ev) goto out;
>      assert(!ev->infinite);
>  
> diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
> index 3bcb6d3..51f2721 100644
> --- a/tools/libxl/libxl_event.h
> +++ b/tools/libxl/libxl_event.h
> @@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
>    int (*timeout_register)(void *user, void **for_app_registration_out,
>                            struct timeval abs, void *for_libxl);
>    int (*timeout_modify)(void *user, void **for_app_registration_update,
> -                         struct timeval abs);
> -  void (*timeout_deregister)(void *user, void *for_app_registration);
> +                         struct timeval abs)
> +      /* only ever called with abs={0,0}, meaning ASAP */;
> +  void (*timeout_deregister)(void *user, void *for_app_registration)
> +      /* will never be called */;
>  } libxl_osevent_hooks;
>  
>  /* The application which calls register_fd_hooks promises to
> @@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
>   * register (or modify), and pass it to subsequent calls to modify
>   * or deregister.
>   *
> + * Note that the application must cope with a call from libxl to
> + * timeout_modify racing with its own call to
> + * libxl__osevent_occurred_timeout.  libxl guarantees that
> + * timeout_modify will only be called with abs={0,0} but the
> + * application must still ensure that libxl's attempt to cause the
> + * timeout to occur immediately is safely ignored even the timeout is
> + * actually already in the process of occurring.
> + *
> + * timeout_deregister is not used because it forms part of a
> + * deprecated unsafe mode of use of the API.
> + *
>   * osevent_register_hooks may be called only once for each libxl_ctx.
>   * libxl may make calls to register/modify/deregister from within
>   * any libxl function (indeed, it will usually call register from
>   

DO NOT APPLY Re: [PATCH 0/2 v4] libxl: fix event races exposed by libvirt

[Ian Jackson]

Ian Jackson writes ("[PATCH 0/2 v4] libxl: fix event races exposed by libvirt"):
> This is the final version of these patches which are needed to make
> Jim Fehlig's libvirt libxl driver work.
> 
>  1/2 libxl: fix stale fd event callback race
>  2/2 libxl: fix stale timeout event callback race
> 
> They are identical to the previous posting apart from commit messages
> and fixing the bug Jim mentioned in <50FD98DC.4020203@suse.com>.

Something has gone wrong with my patch toolchain and the version
posted is not the version in my tree.

Thanks to Jim for spotting this.

Ian.

[PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Because there is not necessarily any lock held at the point the
application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
multithreaded program those calls may be arbitrarily delayed in
relation to other activities within the program.

Specifically this means when ->timeout_deregister returns, libxl does
not know whether it can safely dispose of the for_libxl value or
whether it needs to retain it in case of an in-progress call to
_occurred_timeout.

The interface could be fixed by requiring the application to make a
new call into libxl to say that the deregistration was complete.

However that new call would have to be threaded through the
application's event loop; this is complicated and some application
authors are likely not to implement it properly.  Furthermore the
easiest way to implement this facility in most event loops is to queue
up a time event for "now".

Shortcut all of this by having libxl always call timeout_modify
setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
cause the application to call _occurred_timeout.  When processing this
calldown we see that we were no longer actually interested and simply
throw it away.

Additionally, there is a race between _occurred_timeout and
->timeout_modify.  If libxl ever adjusts the deadline for a timeout
the application may already be in the process of calling _occurred, in
which case the situation with for_app's lifetime becomes very
complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
have no callers) and promise to the application only ever to call
->timeout_modify with abs=={0,0}.  The application still needs to cope
with ->timeout_modify racing with its internal function which calls
_occurred_timeout.  Document this.

This is a forwards-compatible change for applications using the libxl
API, and will hopefully eliminate these races in callback-supplying
applications (such as libvirt) without the need for corresponding
changes to the application.  (It is possible that this might expose
bugs in applications, though, as previously libxl would never call
libxl_osevent_hooks->timeout_modify and now it never calls
->timeout_deregister).

For clarity, fold the body of time_register_finite into its one
remaining call site.  This makes the semantics of ev->infinite
slightly clearer.

Cc: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Tested-by: Jim Fehlig <jfehlig@suse.com>
Acked-by: Jim Fehlig <jfehlig@suse.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

--
v4:
  - Fix the wrong number of levels of indirection of for_app_reg
    in the new call to ->timeout_modify.  (Reported by Jim Fehlig.)
  - Add Jim Fehlig's Tested-by and Acked-by.
  - Improve the commit message somewhat.
---
 tools/libxl/libxl_event.c |   89 +++++++--------------------------------------
 tools/libxl/libxl_event.h |   17 ++++++++-
 2 files changed, 29 insertions(+), 77 deletions(-)

diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
index f1fe425..bfe317b 100644
--- a/tools/libxl/libxl_event.c
+++ b/tools/libxl/libxl_event.c
@@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
     return 0;
 }
 
-static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
-{
-    libxl__ev_time *evsearch;
-    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
-                              timercmp(&ev->abs, &evsearch->abs, >));
-    ev->infinite = 0;
-}
-
 static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
                                 struct timeval absolute)
 {
     int rc;
+    libxl__ev_time *evsearch;
 
     rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
                       absolute, ev->nexus);
@@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 
     ev->infinite = 0;
     ev->abs = absolute;
-    time_insert_finite(gc, ev);
+    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
+                              timercmp(&ev->abs, &evsearch->abs, >));
 
     return 0;
 }
@@ -294,7 +288,12 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     if (!ev->infinite) {
-        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
+        struct timeval right_away = { 0, 0 };
+        if (ev->nexus) /* only set if app provided hooks */
+            ev->nexus->ev = 0;
+        OSEVENT_HOOK_VOID(timeout,modify,
+                          noop /* release nexus in _occurred_ */,
+                          &ev->nexus->for_app_reg, right_away);
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     }
 }
@@ -364,70 +363,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
     return rc;
 }
 
-int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
-                              struct timeval absolute)
-{
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify abs==%lu.%06lu",
-        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (ev->infinite) {
-        rc = time_register_finite(gc, ev, absolute);
-        if (rc) goto out;
-    } else {
-        rc = OSEVENT_HOOK(timeout,modify, noop,
-                          &ev->nexus->for_app_reg, absolute);
-        if (rc) goto out;
-
-        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
-        ev->abs = absolute;
-        time_insert_finite(gc, ev);
-    }
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
-int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
-                              int milliseconds)
-{
-    struct timeval absolute;
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (milliseconds < 0) {
-        time_deregister(gc, ev);
-        ev->infinite = 1;
-        rc = 0;
-        goto out;
-    }
-
-    rc = time_rel_to_abs(gc, milliseconds, &absolute);
-    if (rc) goto out;
-
-    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
-    if (rc) goto out;
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
 void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     CTX_LOCK;
@@ -1161,7 +1096,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
-    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    libxl__osevent_hook_nexus *nexus = for_libxl;
+    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
+
+    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
+
     if (!ev) goto out;
     assert(!ev->infinite);
 
diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
index 3bcb6d3..51f2721 100644
--- a/tools/libxl/libxl_event.h
+++ b/tools/libxl/libxl_event.h
@@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
   int (*timeout_register)(void *user, void **for_app_registration_out,
                           struct timeval abs, void *for_libxl);
   int (*timeout_modify)(void *user, void **for_app_registration_update,
-                         struct timeval abs);
-  void (*timeout_deregister)(void *user, void *for_app_registration);
+                         struct timeval abs)
+      /* only ever called with abs={0,0}, meaning ASAP */;
+  void (*timeout_deregister)(void *user, void *for_app_registration)
+      /* will never be called */;
 } libxl_osevent_hooks;
 
 /* The application which calls register_fd_hooks promises to
@@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
  * register (or modify), and pass it to subsequent calls to modify
  * or deregister.
  *
+ * Note that the application must cope with a call from libxl to
+ * timeout_modify racing with its own call to
+ * libxl__osevent_occurred_timeout.  libxl guarantees that
+ * timeout_modify will only be called with abs={0,0} but the
+ * application must still ensure that libxl's attempt to cause the
+ * timeout to occur immediately is safely ignored even the timeout is
+ * actually already in the process of occurring.
+ *
+ * timeout_deregister is not used because it forms part of a
+ * deprecated unsafe mode of use of the API.
+ *
  * osevent_register_hooks may be called only once for each libxl_ctx.
  * libxl may make calls to register/modify/deregister from within
  * any libxl function (indeed, it will usually call register from
-- 
1.7.2.5

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
>   
>> Ian Campbell wrote:
>>     
>>> One option is to add new hooks which libxl can call to take/release the
>>> application's event loop lock + a LIBXL_HAVE_EVENT_LOOP_LOCK define so
>>> the application can conditionally provide them.
>>>       
>> libvirt's event loop lock is private to the event impl and not exposed
>> to its numerous users.
>>     
>
> Right.  I still think it might be useful to provide a way for a
> consenting application to allow libxl to use the application's event
> loop lock (perhaps, its single giant lock) as the ctx lock.  If it had
> been possible in this case it would have eliminated these particular
> races, so it's a benefit for those applications.  And the extra
> complexity doesn't seem likely to introduce other bugs.
>
> But I think we should fault that feature in when we have a potential
> user for it, and from what you say that's not libvirt.
>   

Correct.  That approach doesn't really fit with libvirt's generic event
loop used by the various drivers.  I suppose the libxl driver could have
a private event loop, but I'd prefer to keep the pattern used by other
drivers.

Regards,
Jim

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> Ian Campbell wrote:
> > One option is to add new hooks which libxl can call to take/release the
> > application's event loop lock + a LIBXL_HAVE_EVENT_LOOP_LOCK define so
> > the application can conditionally provide them.
> 
> libvirt's event loop lock is private to the event impl and not exposed
> to its numerous users.

Right.  I still think it might be useful to provide a way for a
consenting application to allow libxl to use the application's event
loop lock (perhaps, its single giant lock) as the ctx lock.  If it had
been possible in this case it would have eliminated these particular
races, so it's a benefit for those applications.  And the extra
complexity doesn't seem likely to introduce other bugs.

But I think we should fault that feature in when we have a potential
user for it, and from what you say that's not libvirt.

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
>   
>> Yes, after thinking about it some more, I agree.
>>
>> As for the modify callback, do you agree that it is fine to ignore the
>> timeval parameter and just update the timer in libvirt's event loop to
>> fire immediately?  I.e. always treat the timeval parameter as containing
>> {0,0} regardless of "old" or "new" libxl?
>>     
>
> Yes.  That is fine.
>
> The reason is that old libxl doesn't ever call this function and new
> libxl always calls it with {0,0}.  If you're worried about this, add
> an assertion :-).  But it's theoretical.
>
>   
>> I think my patch here is correct
>> http://lists.xen.org/archives/html/xen-devel/2012-12/msg00985.html
>>     
>
> Having thought about it, I agree, provided that you also fix the
> potential integer overflow in libxlTimeoutRegisterEventHook.
>   

Right.  I'll squash a fix to handle that into my current patch and send
it to the libvirt dev list, cc'ing xen-devel too.

Regards,
Jim

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Campbell wrote:
> On Wed, 2012-12-12 at 18:01 +0000, Jim Fehlig wrote:
>   
>> Ian Jackson wrote:
>>     
>
>   
>>>   
>>>       
>>>>  After again reading
>>>> libxl_event.h, I'm considering the below patch in the libvirt libxl
>>>> driver.  The change is primarily inspired by this comment for
>>>> libxl_osevent_occurred_timeout:
>>>>     
>>>>         
>>> ...
>>>   
>>>       
>>>> /* Implicitly, on entry to this function the timeout has been
>>>>  * deregistered.  If _occurred_timeout is called, libxl will not
>>>>  * call timeout_deregister; if it wants to requeue the timeout it
>>>>  * will call timeout_register again.
>>>>  */
>>>>     
>>>>         
>>> Well your patch is only correct when used with the new libxl, with my
>>> patches.
>>>   
>>>       
>> Hmm, it is not clear to me how to make the libxl driver work correctly
>> with libxl pre and post your patches :-/.
>>     
>
> Ideally we will find a way to make this work without changes on the
> application side.
>   

I think Ian J. is right about applications still working, *if* they have
the callbacks coded correctly.  There are some bugs on the libvirt side
as well :).

> But if that turns out to be impossible and applications are going to
> need patching anyway then I think we should consider just fixing the API
> rather than playing tricks like the "modify to 0" thing to try and keep
> it compatible.
>
> One option is to add new hooks which libxl can call to take/release the
> application's event loop lock + a LIBXL_HAVE_EVENT_LOOP_LOCK define so
> the application can conditionally provide them.

libvirt's event loop lock is private to the event impl and not exposed
to its numerous users.

Regards,
Jim

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> Yes, after thinking about it some more, I agree.
> 
> As for the modify callback, do you agree that it is fine to ignore the
> timeval parameter and just update the timer in libvirt's event loop to
> fire immediately?  I.e. always treat the timeval parameter as containing
> {0,0} regardless of "old" or "new" libxl?

Yes.  That is fine.

The reason is that old libxl doesn't ever call this function and new
libxl always calls it with {0,0}.  If you're worried about this, add
an assertion :-).  But it's theoretical.

> I think my patch here is correct
> http://lists.xen.org/archives/html/xen-devel/2012-12/msg00985.html

Having thought about it, I agree, provided that you also fix the
potential integer overflow in libxlTimeoutRegisterEventHook.

> > Right.  Well then yes you need to call virEventRemoveTimeout here.
> > But that applies to both old and new libxl.
> 
> Right.
> 
> Thanks for the help with getting this issue resolved!

Thank you and I'm sorry to have caused trouble with my inadequate
analysis.

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
>   
>> Ian Jackson wrote:
>>     
>>> Well your patch is only correct when used with the new libxl, with my
>>> patches.
>>>       
>> Hmm, it is not clear to me how to make the libxl driver work correctly
>> with libxl pre and post your patches :-/.
>>     
>
> This should be straightforward I think, except for the deregistration
> race bug which is unavoidable with the old libxl.  Simply correctly
> implementing both the deregistration callback and the modification
> callback ought to do it.
>   

Yes, after thinking about it some more, I agree.

As for the modify callback, do you agree that it is fine to ignore the
timeval parameter and just update the timer in libvirt's event loop to
fire immediately?  I.e. always treat the timeval parameter as containing
{0,0} regardless of "old" or "new" libxl?  I think my patch here is correct

http://lists.xen.org/archives/html/xen-devel/2012-12/msg00985.html

>   
>>>> @@ -184,6 +184,8 @@ static void libxlTimerCallback(int timer
>>>> ATTRIBUTE_UNUSED, void *timer_v)
>>>>  {
>>>>      struct libxlOSEventHookTimerInfo *timer_info = timer_v;
>>>>  
>>>> +    virEventRemoveTimeout(timer_info->id);
>>>> +    timer_info->id = -1;
>>>>     
>>>>         
>>> I don't understand why you need this.  Doesn't libvirt remove timers
>>> when they fire ?  If it doesn't, do they otherwise not keep reoccurring ?
>>>       
>> No, timers are not removed when they fire.  And they are continuous, so
>> will keep firing at each timeout.  They can be disabled by setting the
>> timeout to -1, and can be set to fire on each iteration of the event
>> loop by setting the timeout to 0.  But they must be explicitly removed
>> with virEventRemoveTimeout when no longer needed.
>>     
>
> Right.  Well then yes you need to call virEventRemoveTimeout here.
> But that applies to both old and new libxl.
>   

Right.

Thanks for the help with getting this issue resolved!

Jim

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Ian Campbell writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> On Wed, 2012-12-12 at 18:01 +0000, Jim Fehlig wrote:
> > Hmm, it is not clear to me how to make the libxl driver work correctly
> > with libxl pre and post your patches :-/.
> 
> Ideally we will find a way to make this work without changes on the
> application side.

Yes, but if the application has bugs which are exposed by the new
approach I think it is probably simplest to fix those bugs.

The way I'm proposing at the moment means that there are two sets of
relevant changes to libxl and libvirt:

 - libxl always use timeout_modify and never _deregister.  Officially
   speaking this is a backwards-compatible change: libxl now promises
   to use only a strict subset of the documented interface provided by
   the application.  Any correct application will still work.

 - libvirt implement both _modify and _deregister correctly.  With old
   libxl this leaves the timeout deregistration race.  With new libxl
   there is no problem at all.

> One option is to add new hooks which libxl can call to take/release the
> application's event loop lock + a LIBXL_HAVE_EVENT_LOOP_LOCK define so
> the application can conditionally provide them. The upside is that I
> would expect this to result in much simpler code in both libxl and
> libvirt. The downside is that doing this kind of sucks from an API
> stability point of view, but if the application has to change anyway
> then we might as well do it cleanly instead of bending over backwards to
> keep the API the same.

I don't know what other users there might be who would be hurt by an
API change.

But I think from a deployment/testing/etc. point of view two separate
patches to libxl and libvirt which each make matters strictly better,
and which can be applied independently, does seem like the best
approach.

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> Ian Jackson wrote:
> > Well your patch is only correct when used with the new libxl, with my
> > patches.
> 
> Hmm, it is not clear to me how to make the libxl driver work correctly
> with libxl pre and post your patches :-/.

This should be straightforward I think, except for the deregistration
race bug which is unavoidable with the old libxl.  Simply correctly
implementing both the deregistration callback and the modification
callback ought to do it.

> >> @@ -184,6 +184,8 @@ static void libxlTimerCallback(int timer
> >> ATTRIBUTE_UNUSED, void *timer_v)
> >>  {
> >>      struct libxlOSEventHookTimerInfo *timer_info = timer_v;
> >>  
> >> +    virEventRemoveTimeout(timer_info->id);
> >> +    timer_info->id = -1;
> >>     
> >
> > I don't understand why you need this.  Doesn't libvirt remove timers
> > when they fire ?  If it doesn't, do they otherwise not keep reoccurring ?
> 
> No, timers are not removed when they fire.  And they are continuous, so
> will keep firing at each timeout.  They can be disabled by setting the
> timeout to -1, and can be set to fire on each iteration of the event
> loop by setting the timeout to 0.  But they must be explicitly removed
> with virEventRemoveTimeout when no longer needed.

Right.  Well then yes you need to call virEventRemoveTimeout here.
But that applies to both old and new libxl.

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Campbell]

On Wed, 2012-12-12 at 18:01 +0000, Jim Fehlig wrote:
> Ian Jackson wrote:

> >   
> >>  After again reading
> >> libxl_event.h, I'm considering the below patch in the libvirt libxl
> >> driver.  The change is primarily inspired by this comment for
> >> libxl_osevent_occurred_timeout:
> >>     
> > ...
> >   
> >> /* Implicitly, on entry to this function the timeout has been
> >>  * deregistered.  If _occurred_timeout is called, libxl will not
> >>  * call timeout_deregister; if it wants to requeue the timeout it
> >>  * will call timeout_register again.
> >>  */
> >>     
> >
> > Well your patch is only correct when used with the new libxl, with my
> > patches.
> >   
> 
> Hmm, it is not clear to me how to make the libxl driver work correctly
> with libxl pre and post your patches :-/.

Ideally we will find a way to make this work without changes on the
application side.

But if that turns out to be impossible and applications are going to
need patching anyway then I think we should consider just fixing the API
rather than playing tricks like the "modify to 0" thing to try and keep
it compatible.

One option is to add new hooks which libxl can call to take/release the
application's event loop lock + a LIBXL_HAVE_EVENT_LOOP_LOCK define so
the application can conditionally provide them. The upside is that I
would expect this to result in much simpler code in both libxl and
libvirt. The downside is that doing this kind of sucks from an API
stability point of view, but if the application has to change anyway
then we might as well do it cleanly instead of bending over backwards to
keep the API the same.

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
>   
>> Ian Jackson wrote:
>>     
>>> Specifically, this code has an integer arithmetic overflow.
>>>       
>> Well, this patch is removing that buggy code :).
>>     
>
> I think you need to have some code somewhere which does the
> complicated arithmetic dance to avoid the integer overflow.  Does your
> timer registration function have the same bug ?
>   

Ah yes.  I'll take care of it following your suggestion around
beforepoll_internal.

>   
>>  After again reading
>> libxl_event.h, I'm considering the below patch in the libvirt libxl
>> driver.  The change is primarily inspired by this comment for
>> libxl_osevent_occurred_timeout:
>>     
> ...
>   
>> /* Implicitly, on entry to this function the timeout has been
>>  * deregistered.  If _occurred_timeout is called, libxl will not
>>  * call timeout_deregister; if it wants to requeue the timeout it
>>  * will call timeout_register again.
>>  */
>>     
>
> Well your patch is only correct when used with the new libxl, with my
> patches.
>   

Hmm, it is not clear to me how to make the libxl driver work correctly
with libxl pre and post your patches :-/.

>   
>> diff --git a/src/libxl/libxl_driver.c b/src/libxl/libxl_driver.c
>> index 302f81c..d4055be 100644
>> --- a/src/libxl/libxl_driver.c
>> +++ b/src/libxl/libxl_driver.c
>> @@ -184,6 +184,8 @@ static void libxlTimerCallback(int timer
>> ATTRIBUTE_UNUSED, void *timer_v)
>>  {
>>      struct libxlOSEventHookTimerInfo *timer_info = timer_v;
>>  
>> +    virEventRemoveTimeout(timer_info->id);
>> +    timer_info->id = -1;
>>     
>
> I don't understand why you need this.  Doesn't libvirt remove timers
> when they fire ?  If it doesn't, do they otherwise not keep reoccurring ?
>   

No, timers are not removed when they fire.  And they are continuous, so
will keep firing at each timeout.  They can be disabled by setting the
timeout to -1, and can be set to fire on each iteration of the event
loop by setting the timeout to 0.  But they must be explicitly removed
with virEventRemoveTimeout when no longer needed.

Regards,
Jim

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> Ian Jackson wrote:
> > Specifically, this code has an integer arithmetic overflow.
> 
> Well, this patch is removing that buggy code :).

I think you need to have some code somewhere which does the
complicated arithmetic dance to avoid the integer overflow.  Does your
timer registration function have the same bug ?

>  After again reading
> libxl_event.h, I'm considering the below patch in the libvirt libxl
> driver.  The change is primarily inspired by this comment for
> libxl_osevent_occurred_timeout:
...
> /* Implicitly, on entry to this function the timeout has been
>  * deregistered.  If _occurred_timeout is called, libxl will not
>  * call timeout_deregister; if it wants to requeue the timeout it
>  * will call timeout_register again.
>  */

Well your patch is only correct when used with the new libxl, with my
patches.

> diff --git a/src/libxl/libxl_driver.c b/src/libxl/libxl_driver.c
> index 302f81c..d4055be 100644
> --- a/src/libxl/libxl_driver.c
> +++ b/src/libxl/libxl_driver.c
> @@ -184,6 +184,8 @@ static void libxlTimerCallback(int timer
> ATTRIBUTE_UNUSED, void *timer_v)
>  {
>      struct libxlOSEventHookTimerInfo *timer_info = timer_v;
>  
> +    virEventRemoveTimeout(timer_info->id);
> +    timer_info->id = -1;

I don't understand why you need this.  Doesn't libvirt remove timers
when they fire ?  If it doesn't, do they otherwise not keep
reoccurring ?

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
>   
>> Ian Jackson wrote:
>>     
>>> This is a forwards-compatible change for applications using the libxl
>>> API, and will hopefully eliminate these races in callback-supplying
>>> applications (such as libvirt) without the need for corresponding
>>> changes to the application.
>>>       
>
> When I wrote this of course I forgot to mention that previously libxl
> would never call libxl_osevent_hooks->timeout_modify and now it never
> calls ->timeout_deregister.
>
> So this change can expose bugs in the application's implementation of
> ->timeout_modify.
>
>   
>> I'm not sure how to avoid changing the application (libvirt). Currently,
>> the libvirt libxl driver removes the timeout from the libvirt event loop
>> in the timeout_deregister() function. This function is never called now
>> and hence the timeout is never removed. I had to make the following
>> change in the libxl driver to use your patches
>>     
>
> I think this is because of a bug of the kind I describe above.
>
>   
>> - gettimeofday(&now, NULL);
>> - timeout = (abs_t.tv_usec - now.tv_usec) / 1000;
>> - timeout += (abs_t.tv_sec - now.tv_sec) * 1000;
>>     
>
> Specifically, this code has an integer arithmetic overflow.
>   

Well, this patch is removing that buggy code :).  After again reading
libxl_event.h, I'm considering the below patch in the libvirt libxl
driver.  The change is primarily inspired by this comment for
libxl_osevent_occurred_timeout:

/* Implicitly, on entry to this function the timeout has been
 * deregistered.  If _occurred_timeout is called, libxl will not
 * call timeout_deregister; if it wants to requeue the timeout it
 * will call timeout_register again.
 */

Regards,
Jim

diff --git a/src/libxl/libxl_driver.c b/src/libxl/libxl_driver.c
index 302f81c..d4055be 100644
--- a/src/libxl/libxl_driver.c
+++ b/src/libxl/libxl_driver.c
@@ -184,6 +184,8 @@ static void libxlTimerCallback(int timer
ATTRIBUTE_UNUSED, void *timer_v)
 {
     struct libxlOSEventHookTimerInfo *timer_info = timer_v;
 
+    virEventRemoveTimeout(timer_info->id);
+    timer_info->id = -1;
     libxl_osevent_occurred_timeout(timer_info->priv->ctx,
timer_info->xl_priv);
 }
 
@@ -225,16 +227,12 @@ static int libxlTimeoutRegisterEventHook(void *priv,
 
 static int libxlTimeoutModifyEventHook(void *priv ATTRIBUTE_UNUSED,
                                        void **hndp,
-                                       struct timeval abs_t)
+                                       struct timeval abs_t
ATTRIBUTE_UNUSED)
 {
-    struct timeval now;
-    int timeout;
     struct libxlOSEventHookTimerInfo *timer_info = *hndp;
 
-    gettimeofday(&now, NULL);
-    timeout = (abs_t.tv_usec - now.tv_usec) / 1000;
-    timeout += (abs_t.tv_sec - now.tv_sec) * 1000;
-    virEventUpdateTimeout(timer_info->id, timeout);
+    if (timer_info->id > 0)
+        virEventUpdateTimeout(timer_info->id, 0);
     return 0;
 }

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Ian Jackson writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> > - gettimeofday(&now, NULL);
> > - timeout = (abs_t.tv_usec - now.tv_usec) / 1000;
> > - timeout += (abs_t.tv_sec - now.tv_sec) * 1000;
> 
> Specifically, this code has an integer arithmetic overflow.

If you look at beforepoll_internal in libxl_event.c, near line 859,
you can see how I tackled this same problem there.

Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Jim Fehlig writes ("Re: [Xen-devel] [PATCH 2/2] libxl: fix stale timeout event callback race"):
> Ian Jackson wrote:
> > This is a forwards-compatible change for applications using the libxl
> > API, and will hopefully eliminate these races in callback-supplying
> > applications (such as libvirt) without the need for corresponding
> > changes to the application.

When I wrote this of course I forgot to mention that previously libxl
would never call libxl_osevent_hooks->timeout_modify and now it never
calls ->timeout_deregister.

So this change can expose bugs in the application's implementation of
->timeout_modify.

> I'm not sure how to avoid changing the application (libvirt). Currently,
> the libvirt libxl driver removes the timeout from the libvirt event loop
> in the timeout_deregister() function. This function is never called now
> and hence the timeout is never removed. I had to make the following
> change in the libxl driver to use your patches

I think this is because of a bug of the kind I describe above.

> - gettimeofday(&now, NULL);
> - timeout = (abs_t.tv_usec - now.tv_usec) / 1000;
> - timeout += (abs_t.tv_sec - now.tv_sec) * 1000;

Specifically, this code has an integer arithmetic overflow.

If libxl calls this function with a value of abs_t which is more than
about INT_MAX/1000 seconds (24 days on a 32-bit machine) in the past,
the calculation (abs_t.tv_sec - now.tv_sec) * 1000; overflows.

And the patch makes libxl always pass abs_t={0,0} which is more than
24 days ago unless the code is running during January 1970 :-).

> - virEventUpdateTimeout(timer_info->id, timeout);

Also, what does this do if timeout is negative ?

Thanks,
Ian.

Re: [PATCH 2/2] libxl: fix stale timeout event callback race

[Jim Fehlig]

Ian Jackson wrote:
> Because there is not necessarily any lock held at the point the
> application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
> multithreaded program those calls may be arbitrarily delayed in
> relation to other activities within the program.
>
> Specifically this means when ->timeout_deregister returns, libxl does
> not know whether it can safely dispose of the for_libxl value or
> whether it needs to retain it in case of an in-progress call to
> _occurred_timeout.
>
> The interface could be fixed by requiring the application to make a
> new call into libxl to say that the deregistration was complete.
>
> However that new call would have to be threaded through the
> application's event loop; this is complicated and some application
> authors are likely not to implement it properly.  Furthermore the
> easiest way to implement this facility in most event loops is to queue
> up a time event for "now".
>
> Shortcut all of this by having libxl always call timeout_modify
> setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
> cause the application to call _occurred_timeout.  When processing this
> calldown we see that we were no longer actually interested and simply
> throw it away.
>
> Additionally, there is a race between _occurred_timeout and
> ->timeout_modify.  If libxl ever adjusts the deadline for a timeout
> the application may already be in the process of calling _occurred, in
> which case the situation with for_app's lifetime becomes very
> complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
> have no callers) and promise to the application only ever to call
> ->timeout_modify with abs=={0,0}.  The application still needs to cope
> with ->timeout_modify racing with its internal function which calls
> _occurred_timeout.  Document this.
>
> This is a forwards-compatible change for applications using the libxl
> API, and will hopefully eliminate these races in callback-supplying
> applications (such as libvirt) without the need for corresponding
> changes to the application.
>   

I'm not sure how to avoid changing the application (libvirt). Currently,
the libvirt libxl driver removes the timeout from the libvirt event loop
in the timeout_deregister() function. This function is never called now
and hence the timeout is never removed. I had to make the following
change in the libxl driver to use your patches

diff --git a/src/libxl/libxl_driver.c b/src/libxl/libxl_driver.c
index 302f81c..d8f078e 100644
--- a/src/libxl/libxl_driver.c
+++ b/src/libxl/libxl_driver.c
@@ -225,16 +225,11 @@ static int libxlTimeoutRegisterEventHook(void *priv,

static int libxlTimeoutModifyEventHook(void *priv ATTRIBUTE_UNUSED,
void **hndp,
- struct timeval abs_t)
+ struct timeval abs_t ATTRIBUTE_UNUSED)
{
- struct timeval now;
- int timeout;
struct libxlOSEventHookTimerInfo *timer_info = *hndp;

- gettimeofday(&now, NULL);
- timeout = (abs_t.tv_usec - now.tv_usec) / 1000;
- timeout += (abs_t.tv_sec - now.tv_sec) * 1000;
- virEventUpdateTimeout(timer_info->id, timeout);
+ virEventRemoveTimeout(timer_info->id);
return 0;
}

I could also call virEventUpdateTimeout() with a timeout of 0 to make it
fire on the next iteration of the event loop, and then remove the
timeout in the callback before invoking
libxl_osevent_occurred_timeout(). But either way, changes need to be
made to the application.

> For clarity, fold the body of time_register_finite into its one
> remaining call site.  This makes the semantics of ev->infinite
> slightly clearer.
>
> Cc: Bamvor Jian Zhang <bjzhang@suse.com>
> Cc: Ian Campbell <Ian.Campbell@citrix.com>
> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
>
> --
> v3:
>   - Fix null pointer dereference in case when hooks not supplied.
> ---
>  tools/libxl/libxl_event.c |   89 +++++++--------------------------------------
>  tools/libxl/libxl_event.h |   17 ++++++++-
>  2 files changed, 29 insertions(+), 77 deletions(-)
>
> diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
> index f1fe425..f86c528 100644
> --- a/tools/libxl/libxl_event.c
> +++ b/tools/libxl/libxl_event.c
> @@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
>      return 0;
>  }
>  
> -static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
> -{
> -    libxl__ev_time *evsearch;
> -    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
> -                              timercmp(&ev->abs, &evsearch->abs, >));
> -    ev->infinite = 0;
> -}
> -
>  static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
>                                  struct timeval absolute)
>  {
>      int rc;
> +    libxl__ev_time *evsearch;
>  
>      rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
>                        absolute, ev->nexus);
> @@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
>  
>      ev->infinite = 0;
>      ev->abs = absolute;
> -    time_insert_finite(gc, ev);
> +    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
> +                              timercmp(&ev->abs, &evsearch->abs, >));
>  
>      return 0;
>  }
> @@ -294,7 +288,12 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
>  static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
>  {
>      if (!ev->infinite) {
> -        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
> +        struct timeval right_away = { 0, 0 };
> +        if (ev->nexus) /* only set if app provided hooks */
> +            ev->nexus->ev = 0;
> +        OSEVENT_HOOK_VOID(timeout,modify,
>   

Nit again about the space here.

> +                          noop /* release nexus in _occurred_ */,
> +                          ev->nexus->for_app_reg, right_away);
>   

I had to change this to &ev->nexus->for_app_reg to get it to work
properly. But with this change, and the aforementioned hack to the
libvirt libxl driver, I'm not seeing the race any longer.

Regards,
Jim

>          LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
>      }
>  }
> @@ -364,70 +363,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
>      return rc;
>  }
>  
> -int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
> -                              struct timeval absolute)
> -{
> -    int rc;
> -
> -    CTX_LOCK;
> -
> -    DBG("ev_time=%p modify abs==%lu.%06lu",
> -        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
> -
> -    assert(libxl__ev_time_isregistered(ev));
> -
> -    if (ev->infinite) {
> -        rc = time_register_finite(gc, ev, absolute);
> -        if (rc) goto out;
> -    } else {
> -        rc = OSEVENT_HOOK(timeout,modify, noop,
> -                          &ev->nexus->for_app_reg, absolute);
> -        if (rc) goto out;
> -
> -        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
> -        ev->abs = absolute;
> -        time_insert_finite(gc, ev);
> -    }
> -
> -    rc = 0;
> - out:
> -    time_done_debug(gc,__func__,ev,rc);
> -    CTX_UNLOCK;
> -    return rc;
> -}
> -
> -int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
> -                              int milliseconds)
> -{
> -    struct timeval absolute;
> -    int rc;
> -
> -    CTX_LOCK;
> -
> -    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
> -
> -    assert(libxl__ev_time_isregistered(ev));
> -
> -    if (milliseconds < 0) {
> -        time_deregister(gc, ev);
> -        ev->infinite = 1;
> -        rc = 0;
> -        goto out;
> -    }
> -
> -    rc = time_rel_to_abs(gc, milliseconds, &absolute);
> -    if (rc) goto out;
> -
> -    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
> -    if (rc) goto out;
> -
> -    rc = 0;
> - out:
> -    time_done_debug(gc,__func__,ev,rc);
> -    CTX_UNLOCK;
> -    return rc;
> -}
> -
>  void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
>  {
>      CTX_LOCK;
> @@ -1161,7 +1096,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
>      CTX_LOCK;
>      assert(!CTX->osevent_in_hook);
>  
> -    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
> +    libxl__osevent_hook_nexus *nexus = for_libxl;
> +    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
> +
> +    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
> +
>      if (!ev) goto out;
>      assert(!ev->infinite);
>  
> diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
> index 3bcb6d3..51f2721 100644
> --- a/tools/libxl/libxl_event.h
> +++ b/tools/libxl/libxl_event.h
> @@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
>    int (*timeout_register)(void *user, void **for_app_registration_out,
>                            struct timeval abs, void *for_libxl);
>    int (*timeout_modify)(void *user, void **for_app_registration_update,
> -                         struct timeval abs);
> -  void (*timeout_deregister)(void *user, void *for_app_registration);
> +                         struct timeval abs)
> +      /* only ever called with abs={0,0}, meaning ASAP */;
> +  void (*timeout_deregister)(void *user, void *for_app_registration)
> +      /* will never be called */;
>  } libxl_osevent_hooks;
>  
>  /* The application which calls register_fd_hooks promises to
> @@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
>   * register (or modify), and pass it to subsequent calls to modify
>   * or deregister.
>   *
> + * Note that the application must cope with a call from libxl to
> + * timeout_modify racing with its own call to
> + * libxl__osevent_occurred_timeout.  libxl guarantees that
> + * timeout_modify will only be called with abs={0,0} but the
> + * application must still ensure that libxl's attempt to cause the
> + * timeout to occur immediately is safely ignored even the timeout is
> + * actually already in the process of occurring.
> + *
> + * timeout_deregister is not used because it forms part of a
> + * deprecated unsafe mode of use of the API.
> + *
>   * osevent_register_hooks may be called only once for each libxl_ctx.
>   * libxl may make calls to register/modify/deregister from within
>   * any libxl function (indeed, it will usually call register from
>   

[PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Because there is not necessarily any lock held at the point the
application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
multithreaded program those calls may be arbitrarily delayed in
relation to other activities within the program.

Specifically this means when ->timeout_deregister returns, libxl does
not know whether it can safely dispose of the for_libxl value or
whether it needs to retain it in case of an in-progress call to
_occurred_timeout.

The interface could be fixed by requiring the application to make a
new call into libxl to say that the deregistration was complete.

However that new call would have to be threaded through the
application's event loop; this is complicated and some application
authors are likely not to implement it properly.  Furthermore the
easiest way to implement this facility in most event loops is to queue
up a time event for "now".

Shortcut all of this by having libxl always call timeout_modify
setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
cause the application to call _occurred_timeout.  When processing this
calldown we see that we were no longer actually interested and simply
throw it away.

Additionally, there is a race between _occurred_timeout and
->timeout_modify.  If libxl ever adjusts the deadline for a timeout
the application may already be in the process of calling _occurred, in
which case the situation with for_app's lifetime becomes very
complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
have no callers) and promise to the application only ever to call
->timeout_modify with abs=={0,0}.  The application still needs to cope
with ->timeout_modify racing with its internal function which calls
_occurred_timeout.  Document this.

This is a forwards-compatible change for applications using the libxl
API, and will hopefully eliminate these races in callback-supplying
applications (such as libvirt) without the need for corresponding
changes to the application.

For clarity, fold the body of time_register_finite into its one
remaining call site.  This makes the semantics of ev->infinite
slightly clearer.

Cc: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

--
v3:
  - Fix null pointer dereference in case when hooks not supplied.
---
 tools/libxl/libxl_event.c |   89 +++++++--------------------------------------
 tools/libxl/libxl_event.h |   17 ++++++++-
 2 files changed, 29 insertions(+), 77 deletions(-)

diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
index f1fe425..f86c528 100644
--- a/tools/libxl/libxl_event.c
+++ b/tools/libxl/libxl_event.c
@@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
     return 0;
 }
 
-static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
-{
-    libxl__ev_time *evsearch;
-    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
-                              timercmp(&ev->abs, &evsearch->abs, >));
-    ev->infinite = 0;
-}
-
 static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
                                 struct timeval absolute)
 {
     int rc;
+    libxl__ev_time *evsearch;
 
     rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
                       absolute, ev->nexus);
@@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 
     ev->infinite = 0;
     ev->abs = absolute;
-    time_insert_finite(gc, ev);
+    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
+                              timercmp(&ev->abs, &evsearch->abs, >));
 
     return 0;
 }
@@ -294,7 +288,12 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     if (!ev->infinite) {
-        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
+        struct timeval right_away = { 0, 0 };
+        if (ev->nexus) /* only set if app provided hooks */
+            ev->nexus->ev = 0;
+        OSEVENT_HOOK_VOID(timeout,modify,
+                          noop /* release nexus in _occurred_ */,
+                          ev->nexus->for_app_reg, right_away);
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     }
 }
@@ -364,70 +363,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
     return rc;
 }
 
-int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
-                              struct timeval absolute)
-{
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify abs==%lu.%06lu",
-        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (ev->infinite) {
-        rc = time_register_finite(gc, ev, absolute);
-        if (rc) goto out;
-    } else {
-        rc = OSEVENT_HOOK(timeout,modify, noop,
-                          &ev->nexus->for_app_reg, absolute);
-        if (rc) goto out;
-
-        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
-        ev->abs = absolute;
-        time_insert_finite(gc, ev);
-    }
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
-int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
-                              int milliseconds)
-{
-    struct timeval absolute;
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (milliseconds < 0) {
-        time_deregister(gc, ev);
-        ev->infinite = 1;
-        rc = 0;
-        goto out;
-    }
-
-    rc = time_rel_to_abs(gc, milliseconds, &absolute);
-    if (rc) goto out;
-
-    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
-    if (rc) goto out;
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
 void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     CTX_LOCK;
@@ -1161,7 +1096,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
-    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    libxl__osevent_hook_nexus *nexus = for_libxl;
+    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
+
+    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
+
     if (!ev) goto out;
     assert(!ev->infinite);
 
diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
index 3bcb6d3..51f2721 100644
--- a/tools/libxl/libxl_event.h
+++ b/tools/libxl/libxl_event.h
@@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
   int (*timeout_register)(void *user, void **for_app_registration_out,
                           struct timeval abs, void *for_libxl);
   int (*timeout_modify)(void *user, void **for_app_registration_update,
-                         struct timeval abs);
-  void (*timeout_deregister)(void *user, void *for_app_registration);
+                         struct timeval abs)
+      /* only ever called with abs={0,0}, meaning ASAP */;
+  void (*timeout_deregister)(void *user, void *for_app_registration)
+      /* will never be called */;
 } libxl_osevent_hooks;
 
 /* The application which calls register_fd_hooks promises to
@@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
  * register (or modify), and pass it to subsequent calls to modify
  * or deregister.
  *
+ * Note that the application must cope with a call from libxl to
+ * timeout_modify racing with its own call to
+ * libxl__osevent_occurred_timeout.  libxl guarantees that
+ * timeout_modify will only be called with abs={0,0} but the
+ * application must still ensure that libxl's attempt to cause the
+ * timeout to occur immediately is safely ignored even the timeout is
+ * actually already in the process of occurring.
+ *
+ * timeout_deregister is not used because it forms part of a
+ * deprecated unsafe mode of use of the API.
+ *
  * osevent_register_hooks may be called only once for each libxl_ctx.
  * libxl may make calls to register/modify/deregister from within
  * any libxl function (indeed, it will usually call register from
-- 
1.7.2.5

[PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Because there is not necessarily any lock held at the point the
application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
multithreaded program those calls may be arbitrarily delayed in
relation to other activities within the program.

Specifically this means when ->timeout_deregister returns, libxl does
not know whether it can safely dispose of the for_libxl value or
whether it needs to retain it in case of an in-progress call to
_occurred_timeout.

The interface could be fixed by requiring the application to make a
new call into libxl to say that the deregistration was complete.

However that new call would have to be threaded through the
application's event loop; this is complicated and some application
authors are likely not to implement it properly.  Furthermore the
easiest way to implement this facility in most event loops is to queue
up a time event for "now".

Shortcut all of this by having libxl always call timeout_modify
setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
cause the application to call _occurred_timeout.  When processing this
calldown we see that we were no longer actually interested and simply
throw it away.

Additionally, there is a race between _occurred_timeout and
->timeout_modify.  If libxl ever adjusts the deadline for a timeout
the application may already be in the process of calling _occurred, in
which case the situation with for_app's lifetime becomes very
complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
have no callers) and promise to the application only ever to call
->timeout_modify with abs=={0,0}.  The application still needs to cope
with ->timeout_modify racing with its internal function which calls
_occurred_timeout.  Document this.

This is a forwards-compatible change for applications using the libxl
API, and will hopefully eliminate these races in callback-supplying
applications (such as libvirt) without the need for corresponding
changes to the application.

For clarity, fold the body of time_register_finite into its one
remaining call site.  This makes the semantics of ev->infinite
slightly clearer.

Cc: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
---
 tools/libxl/libxl_event.c |   88 +++++++--------------------------------------
 tools/libxl/libxl_event.h |   17 ++++++++-
 2 files changed, 28 insertions(+), 77 deletions(-)

diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
index f1fe425..65c34da 100644
--- a/tools/libxl/libxl_event.c
+++ b/tools/libxl/libxl_event.c
@@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
     return 0;
 }
 
-static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
-{
-    libxl__ev_time *evsearch;
-    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
-                              timercmp(&ev->abs, &evsearch->abs, >));
-    ev->infinite = 0;
-}
-
 static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
                                 struct timeval absolute)
 {
     int rc;
+    libxl__ev_time *evsearch;
 
     rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
                       absolute, ev->nexus);
@@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 
     ev->infinite = 0;
     ev->abs = absolute;
-    time_insert_finite(gc, ev);
+    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
+                              timercmp(&ev->abs, &evsearch->abs, >));
 
     return 0;
 }
@@ -294,7 +288,11 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     if (!ev->infinite) {
-        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
+        struct timeval right_away = { 0, 0 };
+        ev->nexus->ev = 0;
+        OSEVENT_HOOK_VOID(timeout,modify,
+                          noop /* release nexus in _occurred_ */,
+                          ev->nexus->for_app_reg, right_away);
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     }
 }
@@ -364,70 +362,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
     return rc;
 }
 
-int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
-                              struct timeval absolute)
-{
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify abs==%lu.%06lu",
-        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (ev->infinite) {
-        rc = time_register_finite(gc, ev, absolute);
-        if (rc) goto out;
-    } else {
-        rc = OSEVENT_HOOK(timeout,modify, noop,
-                          &ev->nexus->for_app_reg, absolute);
-        if (rc) goto out;
-
-        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
-        ev->abs = absolute;
-        time_insert_finite(gc, ev);
-    }
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
-int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
-                              int milliseconds)
-{
-    struct timeval absolute;
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (milliseconds < 0) {
-        time_deregister(gc, ev);
-        ev->infinite = 1;
-        rc = 0;
-        goto out;
-    }
-
-    rc = time_rel_to_abs(gc, milliseconds, &absolute);
-    if (rc) goto out;
-
-    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
-    if (rc) goto out;
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
 void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     CTX_LOCK;
@@ -1161,7 +1095,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
-    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    libxl__osevent_hook_nexus *nexus = for_libxl;
+    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
+
+    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
+
     if (!ev) goto out;
     assert(!ev->infinite);
 
diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
index 3bcb6d3..51f2721 100644
--- a/tools/libxl/libxl_event.h
+++ b/tools/libxl/libxl_event.h
@@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
   int (*timeout_register)(void *user, void **for_app_registration_out,
                           struct timeval abs, void *for_libxl);
   int (*timeout_modify)(void *user, void **for_app_registration_update,
-                         struct timeval abs);
-  void (*timeout_deregister)(void *user, void *for_app_registration);
+                         struct timeval abs)
+      /* only ever called with abs={0,0}, meaning ASAP */;
+  void (*timeout_deregister)(void *user, void *for_app_registration)
+      /* will never be called */;
 } libxl_osevent_hooks;
 
 /* The application which calls register_fd_hooks promises to
@@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
  * register (or modify), and pass it to subsequent calls to modify
  * or deregister.
  *
+ * Note that the application must cope with a call from libxl to
+ * timeout_modify racing with its own call to
+ * libxl__osevent_occurred_timeout.  libxl guarantees that
+ * timeout_modify will only be called with abs={0,0} but the
+ * application must still ensure that libxl's attempt to cause the
+ * timeout to occur immediately is safely ignored even the timeout is
+ * actually already in the process of occurring.
+ *
+ * timeout_deregister is not used because it forms part of a
+ * deprecated unsafe mode of use of the API.
+ *
  * osevent_register_hooks may be called only once for each libxl_ctx.
  * libxl may make calls to register/modify/deregister from within
  * any libxl function (indeed, it will usually call register from
-- 
1.7.2.5

[PATCH 2/2] libxl: fix stale timeout event callback race

[Ian Jackson]

Because there is not necessarily any lock held at the point the
application (eg, libvirt) calls libxl_osevent_occurred_timeout, in a
multithreaded program those calls may be arbitrarily delayed in
relation to other activities within the program.

Specifically this means when ->timeout_deregister returns, libxl does
not know whether it can safely dispose of the for_libxl value or
whether it needs to retain it in case of an in-progress call to
_occurred_timeout.

The interface could be fixed by requiring the application to make a
new call into libxl to say that the deregistration was complete.

However that new call would have to be threaded through the
application's event loop; this is complicated and some application
authors are likely not to implement it properly.  Furthermore the
easiest way to implement this facility in most event loops is to queue
up a time event for "now".

Shortcut all of this by having libxl always call timeout_modify
setting abs={0,0} (ie, ASAP) instead of timeout_deregister.  This will
cause the application to call _occurred_timeout.  When processing this
calldown we see that we were no longer actually interested and simply
throw it away.

Additionally, there is a race between _occurred_timeout and
->timeout_modify.  If libxl ever adjusts the deadline for a timeout
the application may already be in the process of calling _occurred, in
which case the situation with for_app's lifetime becomes very
complicated.  Therefore abolish libxl__ev_time_modify_{abs,rel} (which
have no callers) and promise to the application only ever to call
->timeout_modify with abs=={0,0}.  The application still needs to cope
with ->timeout_modify racing with its internal function which calls
_occurred_timeout.  Document this.

This is a forwards-compatible change for applications using the libxl
API, and will hopefully eliminate these races in callback-supplying
applications (such as libvirt) without the need for corresponding
changes to the application.

For clarity, fold the body of time_register_finite into its one
remaining call site.  This makes the semantics of ev->infinite
slightly clearer.

Cc: Bamvor Jian Zhang <bjzhang@suse.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
---
 tools/libxl/libxl_event.c |   88 +++++++--------------------------------------
 tools/libxl/libxl_event.h |   17 ++++++++-
 2 files changed, 28 insertions(+), 77 deletions(-)

diff --git a/tools/libxl/libxl_event.c b/tools/libxl/libxl_event.c
index f1fe425..65c34da 100644
--- a/tools/libxl/libxl_event.c
+++ b/tools/libxl/libxl_event.c
@@ -267,18 +267,11 @@ static int time_rel_to_abs(libxl__gc *gc, int ms, struct timeval *abs_out)
     return 0;
 }
 
-static void time_insert_finite(libxl__gc *gc, libxl__ev_time *ev)
-{
-    libxl__ev_time *evsearch;
-    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
-                              timercmp(&ev->abs, &evsearch->abs, >));
-    ev->infinite = 0;
-}
-
 static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
                                 struct timeval absolute)
 {
     int rc;
+    libxl__ev_time *evsearch;
 
     rc = OSEVENT_HOOK(timeout,register, alloc, &ev->nexus->for_app_reg,
                       absolute, ev->nexus);
@@ -286,7 +279,8 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 
     ev->infinite = 0;
     ev->abs = absolute;
-    time_insert_finite(gc, ev);
+    LIBXL_TAILQ_INSERT_SORTED(&CTX->etimes, entry, ev, evsearch, /*empty*/,
+                              timercmp(&ev->abs, &evsearch->abs, >));
 
     return 0;
 }
@@ -294,7 +288,11 @@ static int time_register_finite(libxl__gc *gc, libxl__ev_time *ev,
 static void time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     if (!ev->infinite) {
-        OSEVENT_HOOK_VOID(timeout,deregister, release, ev->nexus->for_app_reg);
+        struct timeval right_away = { 0, 0 };
+        ev->nexus->ev = 0;
+        OSEVENT_HOOK_VOID(timeout,modify,
+                          noop /* release nexus in _occurred_ */,
+                          ev->nexus->for_app_reg, right_away);
         LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
     }
 }
@@ -364,70 +362,6 @@ int libxl__ev_time_register_rel(libxl__gc *gc, libxl__ev_time *ev,
     return rc;
 }
 
-int libxl__ev_time_modify_abs(libxl__gc *gc, libxl__ev_time *ev,
-                              struct timeval absolute)
-{
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify abs==%lu.%06lu",
-        ev, (unsigned long)absolute.tv_sec, (unsigned long)absolute.tv_usec);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (ev->infinite) {
-        rc = time_register_finite(gc, ev, absolute);
-        if (rc) goto out;
-    } else {
-        rc = OSEVENT_HOOK(timeout,modify, noop,
-                          &ev->nexus->for_app_reg, absolute);
-        if (rc) goto out;
-
-        LIBXL_TAILQ_REMOVE(&CTX->etimes, ev, entry);
-        ev->abs = absolute;
-        time_insert_finite(gc, ev);
-    }
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
-int libxl__ev_time_modify_rel(libxl__gc *gc, libxl__ev_time *ev,
-                              int milliseconds)
-{
-    struct timeval absolute;
-    int rc;
-
-    CTX_LOCK;
-
-    DBG("ev_time=%p modify ms=%d", ev, milliseconds);
-
-    assert(libxl__ev_time_isregistered(ev));
-
-    if (milliseconds < 0) {
-        time_deregister(gc, ev);
-        ev->infinite = 1;
-        rc = 0;
-        goto out;
-    }
-
-    rc = time_rel_to_abs(gc, milliseconds, &absolute);
-    if (rc) goto out;
-
-    rc = libxl__ev_time_modify_abs(gc, ev, absolute);
-    if (rc) goto out;
-
-    rc = 0;
- out:
-    time_done_debug(gc,__func__,ev,rc);
-    CTX_UNLOCK;
-    return rc;
-}
-
 void libxl__ev_time_deregister(libxl__gc *gc, libxl__ev_time *ev)
 {
     CTX_LOCK;
@@ -1161,7 +1095,11 @@ void libxl_osevent_occurred_timeout(libxl_ctx *ctx, void *for_libxl)
     CTX_LOCK;
     assert(!CTX->osevent_in_hook);
 
-    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, for_libxl);
+    libxl__osevent_hook_nexus *nexus = for_libxl;
+    libxl__ev_time *ev = osevent_ev_from_hook_nexus(ctx, nexus);
+
+    osevent_release_nexus(gc, &CTX->hook_timeout_nexi_idle, nexus);
+
     if (!ev) goto out;
     assert(!ev->infinite);
 
diff --git a/tools/libxl/libxl_event.h b/tools/libxl/libxl_event.h
index 3bcb6d3..51f2721 100644
--- a/tools/libxl/libxl_event.h
+++ b/tools/libxl/libxl_event.h
@@ -287,8 +287,10 @@ typedef struct libxl_osevent_hooks {
   int (*timeout_register)(void *user, void **for_app_registration_out,
                           struct timeval abs, void *for_libxl);
   int (*timeout_modify)(void *user, void **for_app_registration_update,
-                         struct timeval abs);
-  void (*timeout_deregister)(void *user, void *for_app_registration);
+                         struct timeval abs)
+      /* only ever called with abs={0,0}, meaning ASAP */;
+  void (*timeout_deregister)(void *user, void *for_app_registration)
+      /* will never be called */;
 } libxl_osevent_hooks;
 
 /* The application which calls register_fd_hooks promises to
@@ -337,6 +339,17 @@ typedef struct libxl_osevent_hooks {
  * register (or modify), and pass it to subsequent calls to modify
  * or deregister.
  *
+ * Note that the application must cope with a call from libxl to
+ * timeout_modify racing with its own call to
+ * libxl__osevent_occurred_timeout.  libxl guarantees that
+ * timeout_modify will only be called with abs={0,0} but the
+ * application must still ensure that libxl's attempt to cause the
+ * timeout to occur immediately is safely ignored even the timeout is
+ * actually already in the process of occurring.
+ *
+ * timeout_deregister is not used because it forms part of a
+ * deprecated unsafe mode of use of the API.
+ *
  * osevent_register_hooks may be called only once for each libxl_ctx.
  * libxl may make calls to register/modify/deregister from within
  * any libxl function (indeed, it will usually call register from
-- 
1.7.2.5