7:56 p.m.
# HG changeset patch
# User Hollis Blanchard <hollisb(a)us.ibm.com>
# Date 1250181138 25200
# Node ID 36aa26ca62cabda72e027f075889c7a0973a4e6b
# Parent c345ed88f03432da53efee161d505bd64c02c32e
[RFC] Implement libvirt event callback management.
Libvirt requires that users implement their own event monitoring and
callback-management infrastructure. Actually, two similar lists are needed: one
for file descriptors to monitor, and one for pending timers.
This patch starts exactly one event-monitoring thread per libvirt-cim instance.
Multiple provider threads in the same process will share this thread.
Signed-off-by: Hollis Blanchard <hollisb(a)us.ibm.com>
---
This builds but isn't tested, and may very well contain embarrassing list
manipulation or locking bugs. It really needs a consumer to test, i.e. for
someone to register a domain event callback.
Comments welcome.
Please CC me on replies.
diff --git a/libxkutil/Makefile.am b/libxkutil/Makefile.am
--- a/libxkutil/Makefile.am
+++ b/libxkutil/Makefile.am
@@ -5,14 +5,14 @@ SUBDIRS = tests
CFLAGS += $(CFLAGS_STRICT)
noinst_HEADERS = cs_util.h misc_util.h device_parsing.h xmlgen.h infostore.h \
- pool_parsing.h
+ pool_parsing.h event.h
lib_LTLIBRARIES = libxkutil.la
AM_LDFLAGS = -lvirt -luuid
libxkutil_la_SOURCES = cs_util_instance.c misc_util.c device_parsing.c \
- xmlgen.c infostore.c pool_parsing.c
+ xmlgen.c infostore.c pool_parsing.c event.c
noinst_PROGRAMS = xml_parse_test
diff --git a/libxkutil/event.c b/libxkutil/event.c
new file mode 100644
--- /dev/null
+++ b/libxkutil/event.c
@@ -0,0 +1,367 @@
+/*
+ * Copyright IBM Corp. 2009
+ *
+ * Authors:
+ * Hollis Blanchard <hollisb(a)us.ibm.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ */
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <stdarg.h>
+#include <unistd.h>
+#include <pthread.h>
+#include <errno.h>
+#include <poll.h>
+#include <values.h>
+
+#include <libvirt/libvirt.h>
+#include <libvirt/virterror.h>
+
+#include "event.h"
+
+struct timer {
+ int id;
+ int timeout;
+ void *opaque;
+ virFreeCallback ff;
+ struct timer *next;
+ bool deleted;
+};
+
+struct watch {
+ int id;
+ int fd;
+ int events;
+ virEventHandleCallback cb;
+ void *opaque;
+ virFreeCallback ff;
+ struct watch *next;
+ bool deleted;
+};
+
+static pthread_t watch_thread_id;
+
+static int next_watch_id;
+static struct watch *watch_list;
+static int watch_count;
+static pthread_mutex_t watch_list_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+static int next_timer_id;
+static struct timer *timer_list;
+static pthread_mutex_t timer_list_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+
+static int eventAddHandle(int fd, int events, virEventHandleCallback cb,
+ void *opaque, virFreeCallback ff)
+{
+ struct watch *watch;
+
+ CU_DEBUG("%s", __func__);
+
+ watch = malloc(sizeof(struct watch));
+ if (!watch)
+ return -ENOMEM;
+
+ watch->id = next_watch_id++;
+ watch->fd = fd;
+ watch->events = events;
+ watch->cb = cb;
+ watch->opaque = opaque;
+ watch->ff = ff;
+
+ pthread_mutex_lock(&watch_list_mutex);
+
+ watch->next = watch_list;
+ watch_list = watch;
+ watch_count++;
+
+ pthread_mutex_unlock(&watch_list_mutex);
+
+ return watch->id;
+}
+
+static void eventUpdateHandle(int id, int events)
+{
+ struct watch *cur;
+
+ CU_DEBUG("%s %d", __func__, id);
+
+ for (cur = watch_list; cur != NULL; cur = cur->next) {
+ if (cur->id == id) {
+ cur->events = events;
+ break;
+ }
+ }
+}
+
+/* To avoid locking problems, watches are just flagged here, and the memory is
+ * freed later. */
+static int eventRemoveHandle(int id)
+{
+ struct watch *cur;
+
+ CU_DEBUG("%s %d", __func__, id);
+
+ for (cur = watch_list; cur != NULL; cur = cur->next) {
+ if (cur->id == id) {
+ cur->deleted = 1;
+ break;
+ }
+ }
+
+ return 0;
+}
+
+/* Delete all watches marked for deletion. */
+static void event_watch_free_deleted(void)
+{
+ struct watch *cur;
+ struct watch **link;
+
+ CU_DEBUG("%s", __func__);
+
+ pthread_mutex_lock(&watch_list_mutex);
+
+ cur = watch_list;
+ link = &watch_list;
+ while (cur != NULL) {
+ struct watch *next = cur->next;
+
+ if (cur->deleted) {
+ *link = next;
+
+ cur->ff(cur->opaque);
+ free(cur);
+ watch_count--;
+ } else
+ link = &cur->next;
+
+ cur = next;
+ }
+
+ pthread_mutex_unlock(&watch_list_mutex);
+}
+
+static int eventAddTimeout(int timeout, virEventTimeoutCallback cb,
+ void *opaque, virFreeCallback ff)
+{
+ struct timer *timer;
+
+ CU_DEBUG("%s", __func__);
+
+ timer = malloc(sizeof(struct timer));
+ if (!timer)
+ return -ENOMEM;
+
+ timer->id = next_timer_id++;
+ timer->timeout = timeout;
+ timer->opaque = opaque;
+ timer->ff = ff;
+
+ pthread_mutex_lock(&timer_list_mutex);
+
+ timer->next = timer_list;
+ timer_list = timer;
+
+ pthread_mutex_unlock(&timer_list_mutex);
+
+ return timer->id;
+}
+
+static void eventUpdateTimeout(int id, int timeout)
+{
+ struct timer *cur;
+
+ CU_DEBUG("%s %d", __func__, id);
+
+ for (cur = timer_list; cur != NULL; cur = cur->next) {
+ if (cur->id == id) {
+ cur->timeout = timeout;
+ break;
+ }
+ }
+}
+
+static int eventRemoveTimeout(int id)
+{
+ struct timer *cur;
+
+ CU_DEBUG("%s %d", __func__, id);
+
+ for (cur = timer_list; cur != NULL; cur = cur->next) {
+ if (cur->id == id) {
+ cur->deleted = 1;
+ break;
+ }
+ }
+
+ return 0;
+}
+
+/* Delete all timers marked for deletion. */
+static void event_timer_free_deleted(void)
+{
+ struct timer *cur;
+ struct timer **link;
+
+ CU_DEBUG("%s", __func__);
+
+ pthread_mutex_lock(&timer_list_mutex);
+
+ cur = timer_list;
+ link = &timer_list;
+ while (cur != NULL) {
+ struct timer *next = cur->next;
+
+ if (cur->deleted) {
+ *link = next;
+
+ cur->ff(cur->opaque);
+ free(cur);
+ } else
+ link = &cur->next;
+
+ cur = next;
+ }
+
+ pthread_mutex_unlock(&timer_list_mutex);
+}
+
+
+
+static int poll_to_libvirt_events(int pevents)
+{
+ int vevents = 0;
+
+ if (pevents & POLLIN)
+ vevents |= VIR_EVENT_HANDLE_READABLE;
+
+ if (pevents & POLLOUT)
+ vevents |= VIR_EVENT_HANDLE_WRITABLE;
+
+ if (pevents & POLLERR)
+ vevents |= VIR_EVENT_HANDLE_ERROR;
+
+ if (pevents & POLLHUP)
+ vevents |= VIR_EVENT_HANDLE_HANGUP;
+
+ return vevents;
+}
+
+static int libvirt_to_poll_events(int vevents)
+{
+ int pevents = 0;
+
+ if (vevents & VIR_EVENT_HANDLE_READABLE)
+ pevents |= POLLIN;
+
+ if (vevents & VIR_EVENT_HANDLE_WRITABLE)
+ pevents |= POLLOUT;
+
+ if (vevents & VIR_EVENT_HANDLE_ERROR)
+ pevents |= POLLERR;
+
+ if (vevents & VIR_EVENT_HANDLE_HANGUP)
+ pevents |= POLLHUP;
+
+ return pevents;
+}
+
+static void invoke_callback(struct watch *watch, struct pollfd *pollfd)
+{
+ int vevents = poll_to_libvirt_events(watch->events);
+
+ watch->cb(watch->id, watch->fd, vevents, watch->opaque);
+}
+
+static int event_next_timeout(void)
+{
+ struct timer *cur;
+ int closest = MAXINT;
+
+ for (cur = timer_list; cur != NULL; cur = cur->next)
+ if (cur->timeout < closest)
+ closest = cur->timeout;
+
+ return closest;
+}
+
+/* One thread to watch all fds for all events for all libvirt threads. */
+static void *event_thread(void *ptr)
+{
+ while (1) {
+ struct watch *cur;
+ struct pollfd *pollfds;
+ struct pollfd *pollfd;
+ int timeout;
+ int i;
+
+ pollfds = malloc(sizeof(struct pollfd) * watch_count);
+
+ /* fill in pollfds array from our watch list */
+ for (pollfd = &pollfds[0], cur = watch_list;
+ cur != NULL;
+ pollfd++, cur = cur->next) {
+ pollfd->fd = cur->fd;
+ pollfd->events = libvirt_to_poll_events(cur->events);
+ }
+
+ timeout = event_next_timeout();
+
+ poll(pollfds, watch_count, timeout);
+
+ /* invoke callbacks */
+ for (i = 0; i < watch_count; i++)
+ for (cur = watch_list; cur != NULL; cur = cur->next)
+ if (cur->fd == pollfds[i].fd
+ && !cur->deleted) {
+ invoke_callback(cur, &pollfds[i]);
+ break;
+ }
+
+ free(pollfds);
+
+ event_watch_free_deleted();
+ event_timer_free_deleted();
+ }
+
+ return NULL;
+}
+
+void init_events(void)
+{
+ static pthread_mutex_t thread_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+ CU_DEBUG("%s", __func__);
+
+ pthread_mutex_lock(&thread_mutex);
+
+ if (!watch_thread_id) {
+ virEventRegisterImpl(eventAddHandle,
+ eventUpdateHandle,
+ eventRemoveHandle,
+ eventAddTimeout,
+ eventUpdateTimeout,
+ eventRemoveTimeout);
+
+ pthread_create(&watch_thread_id, NULL, event_thread, NULL);
+ }
+
+ pthread_mutex_unlock(&thread_mutex);
+}
diff --git a/libxkutil/event.h b/libxkutil/event.h
new file mode 100644
--- /dev/null
+++ b/libxkutil/event.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright IBM Corp. 2009
+ *
+ * Authors:
+ * Hollis Blanchard <hollisb(a)us.ibm.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ */
+
+#ifndef __EVENT_H
+#define __EVENT_H
+
+#include <libcmpiutil.h>
+
+void init_events(void);
+
+#endif /* __EVENT_H */
diff --git a/libxkutil/misc_util.c b/libxkutil/misc_util.c
--- a/libxkutil/misc_util.c
+++ b/libxkutil/misc_util.c
@@ -38,6 +38,7 @@
#include "misc_util.h"
#include "cs_util.h"
+#include "event.h"
#include <config.h>
@@ -490,6 +491,7 @@ bool parse_instanceid(const CMPIObjectPa
bool libvirt_cim_init(void)
{
+ init_events();
return virInitialize() == 0;
}
10:05 p.m.
Hollis Blanchard wrote:
# HG changeset patch
# User Hollis Blanchard <hollisb(a)us.ibm.com>
# Date 1250181138 25200
# Node ID 36aa26ca62cabda72e027f075889c7a0973a4e6b
# Parent c345ed88f03432da53efee161d505bd64c02c32e
[RFC] Implement libvirt event callback management.
Libvirt requires that users implement their own event monitoring and
callback-management infrastructure. Actually, two similar lists are needed: one
for file descriptors to monitor, and one for pending timers.
This patch starts exactly one event-monitoring thread per libvirt-cim instance.
Multiple provider threads in the same process will share this thread.
Signed-off-by: Hollis Blanchard <hollisb(a)us.ibm.com>
---
This builds but isn't tested, and may very well contain embarrassing list
manipulation or locking bugs. It really needs a consumer to test, i.e. for
someone to register a domain event callback.
Comments welcome.
This looks great Hollis. Thanks!
I haven't had a chance to test it, but Richard is working on some code
that will utilize this.
static int eventAddHandle(int fd, int events, virEventHandleCallback
cb,
+ void *opaque, virFreeCallback ff)
+{
+ struct watch *watch;
+
+ CU_DEBUG("%s", __func__);
+
We really should include the function name in our CU_DEBUG() statement.
I have some other ideas as to how to improve CU_DEBUG, so I will
include the function when I work on those enhancements.
--
Kaitlin Rupert
IBM Linux Technology Center
kaitlin(a)linux.vnet.ibm.com
1:24 a.m.
Hi Hollis,
I applied your patch and the patches Richard submitted to do some
testing. I noticed a few issues commented below...
+/* Delete all watches marked for deletion. */
+static void event_watch_free_deleted(void)
+{
+ struct watch *cur;
+ struct watch **link;
+
+ CU_DEBUG("%s", __func__);
+
+ pthread_mutex_lock(&watch_list_mutex);
+
+ cur = watch_list;
+ link = &watch_list;
+ while (cur != NULL) {
+ struct watch *next = cur->next;
+
+ if (cur->deleted) {
+ *link = next;
+
+ cur->ff(cur->opaque);
I'm seeing a seg fault here because once we reach this point, cur is NULL.
+ free(cur);
+ watch_count--;
+ } else
+ link = &cur->next;
+
+ cur = next;
+ }
+
+ pthread_mutex_unlock(&watch_list_mutex);
+}
+
+/* One thread to watch all fds for all events for all libvirt
threads. */
+static void *event_thread(void *ptr)
+{
+ while (1) {
+ struct watch *cur;
+ struct pollfd *pollfds;
+ struct pollfd *pollfd;
+ int timeout;
+ int i;
+
+ pollfds = malloc(sizeof(struct pollfd) * watch_count);
+
+ /* fill in pollfds array from our watch list */
+ for (pollfd = &pollfds[0], cur = watch_list;
+ cur != NULL;
+ pollfd++, cur = cur->next) {
+ pollfd->fd = cur->fd;
+ pollfd->events = libvirt_to_poll_events(cur->events);
+ }
+
+ timeout = event_next_timeout();
+
+ poll(pollfds, watch_count, timeout);
+
+ /* invoke callbacks */
+ for (i = 0; i < watch_count; i++)
+ for (cur = watch_list; cur != NULL; cur = cur->next)
+ if (cur->fd == pollfds[i].fd
When I generate a event, poll never seems to catch it. I tried forcing
this by changing the timeout value to a minute. And then generating the
event. I can see from the libvirt debug that the event has been generated:
15:00:34.232: debug : virEventRunOnce:567 : Poll got 1 event
15:00:34.239: debug : virEventDispatchHandles:450 : Dispatch n=2 f=8 w=3
e=1 0x7f2a57d6e6b0
In the eventAddHandle() call, we have:
event.c(75): eventAddHandle
event.c(82): ++++++++++++++watch->id is 0
event.c(84): ++++++++++++++watch->fd is 11
event.c(86): ++++++++++++++watch->events is 1
event.c(88): ++++++++++++++watch->cb is 0x6b8f4c0, cb is 0x6b8f4c0
event.c(90): ++++++++++++++watch->opaque is 0x7fffdc014900
I haven't tracked down what is happening here.
+ && !cur->deleted) {
+ invoke_callback(cur, &pollfds[i]);
+ break;
+ }
+
+ free(pollfds);
+
+ event_watch_free_deleted();
+ event_timer_free_deleted();
+ }
+
+ return NULL;
+}
+
+void init_events(void)
+{
+ static pthread_mutex_t thread_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+ CU_DEBUG("%s", __func__);
+
+ pthread_mutex_lock(&thread_mutex);
+
+ if (!watch_thread_id) {
+ virEventRegisterImpl(eventAddHandle,
+ eventUpdateHandle,
+ eventRemoveHandle,
+ eventAddTimeout,
+ eventUpdateTimeout,
+ eventRemoveTimeout);
+
+ pthread_create(&watch_thread_id, NULL, event_thread, NULL);
+ }
+
+ pthread_mutex_unlock(&thread_mutex);
+}
--
Kaitlin Rupert
IBM Linux Technology Center
kaitlin(a)linux.vnet.ibm.com
1:45 a.m.
On Tue, 2009-08-18 at 16:24 -0700, Kaitlin Rupert wrote:
Hi Hollis,
I applied your patch and the patches Richard submitted to do some
testing. I noticed a few issues commented below...
> +/* Delete all watches marked for deletion. */
> +static void event_watch_free_deleted(void)
> +{
> + struct watch *cur;
> + struct watch **link;
> +
> + CU_DEBUG("%s", __func__);
> +
> + pthread_mutex_lock(&watch_list_mutex);
> +
> + cur = watch_list;
> + link = &watch_list;
> + while (cur != NULL) {
> + struct watch *next = cur->next;
> +
> + if (cur->deleted) {
> + *link = next;
> +
> + cur->ff(cur->opaque);
I'm seeing a seg fault here because once we reach this point, cur is NULL.
> + free(cur);
> + watch_count--;
> + } else
> + link = &cur->next;
> +
> + cur = next;
> + }
> +
> + pthread_mutex_unlock(&watch_list_mutex);
> +}
> +
Am I missing something obvious? The loop only runs while cur is not
NULL.
Moreover, the list is locked, so concurrent list modifications wouldn't
even explain it... unless the other user didn't take the lock (which we
do on purpose sometimes). From the CU_DEBUG log, can you tell if there's
any concurrency going on?
> +/* One thread to watch all fds for all events for all libvirt
threads. */
> +static void *event_thread(void *ptr)
> +{
> + while (1) {
> + struct watch *cur;
> + struct pollfd *pollfds;
> + struct pollfd *pollfd;
> + int timeout;
> + int i;
> +
> + pollfds = malloc(sizeof(struct pollfd) * watch_count);
> +
> + /* fill in pollfds array from our watch list */
> + for (pollfd = &pollfds[0], cur = watch_list;
> + cur != NULL;
> + pollfd++, cur = cur->next) {
> + pollfd->fd = cur->fd;
> + pollfd->events =
libvirt_to_poll_events(cur->events);
> + }
> +
> + timeout = event_next_timeout();
> +
> + poll(pollfds, watch_count, timeout);
> +
> + /* invoke callbacks */
> + for (i = 0; i < watch_count; i++)
> + for (cur = watch_list; cur != NULL; cur = cur->next)
> + if (cur->fd == pollfds[i].fd
When I generate a event, poll never seems to catch it. I tried forcing
this by changing the timeout value to a minute. And then generating the
event. I can see from the libvirt debug that the event has been generated:
15:00:34.232: debug : virEventRunOnce:567 : Poll got 1 event
15:00:34.239: debug : virEventDispatchHandles:450 : Dispatch n=2 f=8 w=3
e=1 0x7f2a57d6e6b0
These messages are from qemud? So the event is generated on the other
side of the "remote" pipe?
In the eventAddHandle() call, we have:
event.c(75): eventAddHandle
event.c(82): ++++++++++++++watch->id is 0
event.c(84): ++++++++++++++watch->fd is 11
event.c(86): ++++++++++++++watch->events is 1
event.c(88): ++++++++++++++watch->cb is 0x6b8f4c0, cb is 0x6b8f4c0
event.c(90): ++++++++++++++watch->opaque is 0x7fffdc014900
These seem reasonable, so we should have a list with one entry.
I haven't tracked down what is happening here.
> + && !cur->deleted) {
> + invoke_callback(cur, &pollfds[i]);
> + break;
> + }
> +
> + free(pollfds);
> +
> + event_watch_free_deleted();
> + event_timer_free_deleted();
> + }
> +
> + return NULL;
> +}
Hmm, actually I'd expect the opposite: this code seems to *always*
invoke the callback, even when no event is pending (oops!).
Since that's not happening, it may be that the pollfds structure isn't
being created properly, in particular the fd member.
Another possibility is that watch_count is somehow 0, which would also
hose pollfds allocation...
I guess strace output (for poll()), CU_DEBUG logs, and some gdb work
would help.
--
Hollis Blanchard
IBM Linux Technology Center
2:33 a.m.
Hollis Blanchard wrote:
On Tue, 2009-08-18 at 16:24 -0700, Kaitlin Rupert wrote:
> Hi Hollis,
>
> I applied your patch and the patches Richard submitted to do some
> testing. I noticed a few issues commented below...
>
>
>> +/* Delete all watches marked for deletion. */
>> +static void event_watch_free_deleted(void)
>> +{
>> + struct watch *cur;
>> + struct watch **link;
>> +
>> + CU_DEBUG("%s", __func__);
>> +
>> + pthread_mutex_lock(&watch_list_mutex);
>> +
>> + cur = watch_list;
>> + link = &watch_list;
>> + while (cur != NULL) {
>> + struct watch *next = cur->next;
>> +
>> + if (cur->deleted) {
>> + *link = next;
>> +
>> + cur->ff(cur->opaque);
> I'm seeing a seg fault here because once we reach this point, cur is NULL.
>
>> + free(cur);
>> + watch_count--;
>> + } else
>> + link = &cur->next;
>> +
>> + cur = next;
>> + }
>> +
>> + pthread_mutex_unlock(&watch_list_mutex);
>> +}
>> +
Am I missing something obvious? The loop only runs while cur is not
NULL.
Agreed. I meant to say in my previous email that I am confused as well.
Actually, I put a check for (cur != NULL) right about the call to
cur->ff - and cur is definitely NULL at that point.
I haven't spent enough time debugging it, here's what little I have so far:
event.c(142): event_watch_free_deleted
event.c(146): +++++++++watch_list addr is 0x7fffdc014b30
Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7ffff61c6710 (LWP 22188)]
0x0000000000000000 in ?? ()
Missing separate debuginfos, use: debuginfo-install
tog-pegasus-2.9.0-3.fc12.x86_64
(gdb) bt
#0 0x0000000000000000 in ?? ()
#1 0x00000000069811f1 in event_watch_free_deleted () at event.c:156
#2 event_thread () at event.c:375
#3 0x0000000003c5e91a in start_thread () from /lib64/libpthread.so.0
#4 0x0000000004936b2d in clone () from /lib64/libc.so.6
#5 0x0000000000000000 in ?? ()
(gdb) up
#1 0x00000000069811f1 in event_watch_free_deleted () at event.c:156
156 cur->ff(cur->opaque);
(gdb) print cur
$1 = (struct watch *) 0x0
(gdb)
Moreover, the list is locked, so concurrent list modifications wouldn't
even explain it... unless the other user didn't take the lock (which we
do on purpose sometimes). From the CU_DEBUG log, can you tell if there's
any concurrency going on?
I'll investigate this more tomorrow. Richard - or maybe you can?
>> +/* One thread to watch all fds for all events for all libvirt threads. */
>> +static void *event_thread(void *ptr)
>> +{
>> + while (1) {
>> + struct watch *cur;
>> + struct pollfd *pollfds;
>> + struct pollfd *pollfd;
>> + int timeout;
>> + int i;
>> +
>> + pollfds = malloc(sizeof(struct pollfd) * watch_count);
>> +
>> + /* fill in pollfds array from our watch list */
>> + for (pollfd = &pollfds[0], cur = watch_list;
>> + cur != NULL;
>> + pollfd++, cur = cur->next) {
>> + pollfd->fd = cur->fd;
>> + pollfd->events =
libvirt_to_poll_events(cur->events);
>> + }
>> +
>> + timeout = event_next_timeout();
>> +
>> + poll(pollfds, watch_count, timeout);
>> +
>> + /* invoke callbacks */
>> + for (i = 0; i < watch_count; i++)
>> + for (cur = watch_list; cur != NULL; cur = cur->next)
>> + if (cur->fd == pollfds[i].fd
> When I generate a event, poll never seems to catch it. I tried forcing
> this by changing the timeout value to a minute. And then generating the
> event. I can see from the libvirt debug that the event has been generated:
>
> 15:00:34.232: debug : virEventRunOnce:567 : Poll got 1 event
> 15:00:34.239: debug : virEventDispatchHandles:450 : Dispatch n=2 f=8 w=3
> e=1 0x7f2a57d6e6b0
These messages are from qemud? So the event is generated on the other
side of the "remote" pipe?
> In the eventAddHandle() call, we have:
>
> event.c(75): eventAddHandle
> event.c(82): ++++++++++++++watch->id is 0
> event.c(84): ++++++++++++++watch->fd is 11
> event.c(86): ++++++++++++++watch->events is 1
> event.c(88): ++++++++++++++watch->cb is 0x6b8f4c0, cb is 0x6b8f4c0
> event.c(90): ++++++++++++++watch->opaque is 0x7fffdc014900
These seem reasonable, so we should have a list with one entry.
> I haven't tracked down what is happening here.
>
>
>> + && !cur->deleted) {
>> + invoke_callback(cur, &pollfds[i]);
>> + break;
>> + }
>> +
>> + free(pollfds);
>> +
>> + event_watch_free_deleted();
>> + event_timer_free_deleted();
>> + }
>> +
>> + return NULL;
>> +}
Hmm, actually I'd expect the opposite: this code seems to *always*
invoke the callback, even when no event is pending (oops!).
Since that's not happening, it may be that the pollfds structure isn't
being created properly, in particular the fd member.
I should have said in my previous email - the value of pollfds[i].fd is
always 0. In looking at it more, I think the following is happening:
1) The CIMOM loads Richard's provider and init_events() is called.
2) This starts the event_thread. At this point, we don't have a handle,
so the following for loop fails:
for (pollfd = &pollfds[0], cur = watch_list;
cur != NULL;
pollfd++, cur = cur->next) {}
3) We call poll()
4) In the meantime, the provider's ActivateFilter() function has been
called and the callback is registered
5) eventUpdateHandle() is called, but it doesn't update our pollfds list
5) We generate a domain event and libvirt dispatches it
5) Our fd in pollfds[i] is still 0, so we don't have a match.
event.c(388): init_events
[New Thread 0x7ffff61c6710 (LWP 22188)]
event.c(320): ^^^^^^^^^^in event_thread
event.c(330): ^^^^^^^^^^got pollfds
event.c(335): ^^^^^^^^^^cur is NULL!!
event.c(352): ^^^^^^^^^timeout is 30000
std_indication.c(231): Calling handler->activate_fn
Virt_GuestCrashAlertIndication.c(334): ActivateFilter for
KVM_GuestCrashAlertIndication
misc_util.c(76): Connecting to libvirt with uri `qemu:///system'
event.c(75): eventAddHandle
event.c(82): ++++++++++++++watch->id is 0
event.c(84): ++++++++++++++watch->fd is 11
event.c(86): ++++++++++++++watch->events is 1
event.c(88): ++++++++++++++watch->cb is 0x6de44c0, cb is 0x6de44c0
event.c(90): ++++++++++++++watch->opaque is 0x7fffdc014900
event.c(174): eventAddTimeout
Virt_GuestCrashAlertIndication.c(338): Registering callback function
event.c(108): eventUpdateHandle 0
event.c(108): eventUpdateHandle 0
Virt_GuestCrashAlertIndication.c(343): ret val: 0
std_indication.c(270): IndicationVirt_GuestCrashAlertIndicationProvider:
indications enabled
Virt_GuestCrashAlertIndication.c(365): EnableIndications
event.c(358): ^^^^^^^^^^loop through watch_count
event.c(360): ^^^^^^^^^^loop through watch_list
event.c(361): ==================cur->fd is 11, pollfds[i].fd is 0
event.c(362): ==================cur->deleted is 7
Another possibility is that watch_count is somehow 0, which would also
hose pollfds allocation...
I guess strace output (for poll()), CU_DEBUG logs, and some gdb work
would help.
--
Kaitlin Rupert
IBM Linux Technology Center
kaitlin(a)linux.vnet.ibm.com
4:27 a.m.
>
> Moreover, the list is locked, so concurrent list modifications wouldn't
> even explain it... unless the other user didn't take the lock (which we
> do on purpose sometimes). From the CU_DEBUG log, can you tell if there's
> any concurrency going on?
I'll investigate this more tomorrow. Richard - or maybe you can?
Yes, I can check it tomorrow
--
Richard Maciel, MSc
IBM Linux Technology Center
rmaciel(a)linux.vnet.ibm.com
5510
days inactive
5516
days old
5 comments
3 participants
participants (3)
-
Hollis Blanchard -
Kaitlin Rupert -
Richard Maciel