Re: [PATCH 0/5] trace: inroduce qmp: trace namespace
Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> writes:
Hi all!
We have handle_qmp_command and qmp_command_repond trace points to trace qmp commands. They are very useful to debug problems involving management tools like libvirt.
But tracing all qmp commands is too much.
Here I suggest a kind of tracing namespace. Formally this series adds a trace points called qmp:<some-command> for every command, which may be enabled in separate like
--trace qmp:drive-backup
or by pattern like
--trace qmp:block-job-*
or similarly with help of qmp command trace-event-set-state.
This also allows to enable tracing of some qmp commands permanently (by downstream patch or in libvirt xml). For example, I'm going to enable tracing of block job comamnds and blockdev-* commands in Virtuozzo. Qemu logs are often too empty (for example, in comparison with Libvirt), logging block jobs is not too much but will be very helpful.
What exactly is traced? Peeking at PATCH 5... looks like it's input that makes it to qmp_dispatch() and command responses, but not events. Fine print on "input that makes it to qmp_dispatch()": * You trace right before we execute the command, not when we receive, parse and enqueue input. * Corollary: input with certain errors is not traced. * You don't trace the input text, you trace the unparsed parse tree. All fine, I presume. Existing tracepoints in monitor/qmp.c, and what information they send (inessential bits omitted for clarity): * handle_qmp_command Handling a QMP command: unparsed parse tree Fine print, safe to ignore: - Out-of-band commands will be executed right away, in-band commands will be queued. Tracepoints monitor_qmp_in_band_enqueue and monitor_qmp_in_band_dequeue provide insight into that. - This also receives and queues parse errors, without tracing them. Tracepoint monitor_qmp_err_in_band traces them as they are dequeued. * monitor_qmp_cmd_in_band About to execute in-band command: command ID, if any * monitor_qmp_cmd_out_of_band About to execute out-of-band command: command ID, if any * monitor_qmp_respond About to send command response or event: QObject For input, --trace qmp:* is like --trace handle_qmp_command, except it traces late rather than early. For output, --trace qmp:* is like --trace monitor_qmp_respond less events. The main improvement over existing tracepoints seems to be the ability to filter on command names. To get that, you overload the @name argument of QMP command trace-event-set-state. In addition to the documented meaning "Event name pattern", it also has an alternate, undocumented meaning "QMP command name pattern". The "undocumented" part is easy enough to fix. However, QMP heavily frowns on argument values that need to be parsed. But before we discuss this in depth, we should decide whether we want the filtering feature. Management applications can enable and disable tracing as needed, but doing it all in QEMU might be more convenient or robust. Libvirt logs all QMP traffic. I doubt it'll make use of your filtering feature. Cc'ing libvir-list just in case. Another way to log all traffic is to route it through socat -x or similar. Opinions?
12.10.2021 14:49, Markus Armbruster wrote:
Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> writes:
Hi all!
We have handle_qmp_command and qmp_command_repond trace points to trace qmp commands. They are very useful to debug problems involving management tools like libvirt.
But tracing all qmp commands is too much.
Here I suggest a kind of tracing namespace. Formally this series adds a trace points called qmp:<some-command> for every command, which may be enabled in separate like
--trace qmp:drive-backup
or by pattern like
--trace qmp:block-job-*
or similarly with help of qmp command trace-event-set-state.
This also allows to enable tracing of some qmp commands permanently (by downstream patch or in libvirt xml). For example, I'm going to enable tracing of block job comamnds and blockdev-* commands in Virtuozzo. Qemu logs are often too empty (for example, in comparison with Libvirt), logging block jobs is not too much but will be very helpful.
What exactly is traced? Peeking at PATCH 5... looks like it's input that makes it to qmp_dispatch() and command responses, but not events.
Fine print on "input that makes it to qmp_dispatch()":
* You trace right before we execute the command, not when we receive, parse and enqueue input.
* Corollary: input with certain errors is not traced.
* You don't trace the input text, you trace the unparsed parse tree.
All fine, I presume.
Existing tracepoints in monitor/qmp.c, and what information they send (inessential bits omitted for clarity):
* handle_qmp_command
Handling a QMP command: unparsed parse tree
Fine print, safe to ignore:
- Out-of-band commands will be executed right away, in-band commands will be queued. Tracepoints monitor_qmp_in_band_enqueue and monitor_qmp_in_band_dequeue provide insight into that.
- This also receives and queues parse errors, without tracing them. Tracepoint monitor_qmp_err_in_band traces them as they are dequeued.
* monitor_qmp_cmd_in_band
About to execute in-band command: command ID, if any
* monitor_qmp_cmd_out_of_band
About to execute out-of-band command: command ID, if any
* monitor_qmp_respond
About to send command response or event: QObject
For input, --trace qmp:* is like --trace handle_qmp_command, except it traces late rather than early.
For output, --trace qmp:* is like --trace monitor_qmp_respond less events.
The main improvement over existing tracepoints seems to be the ability to filter on command names.
To get that, you overload the @name argument of QMP command trace-event-set-state. In addition to the documented meaning "Event name pattern", it also has an alternate, undocumented meaning "QMP command name pattern". The "undocumented" part is easy enough to fix. However, QMP heavily frowns on argument values that need to be parsed.
Still, pattern is parsed anyway, as pattern. But yes, this patch adds rather specific and tricky logic, which a lot more than just a pattern to search through the list. Another possible way is to update QAPI code generator to insert a personal trace point for each qmp command.. That seems more complicated to implement, but I can try.
But before we discuss this in depth, we should decide whether we want the filtering feature.
Management applications can enable and disable tracing as needed, but doing it all in QEMU might be more convenient or robust.
Libvirt logs all QMP traffic.
I am not sure.. At lest in Vz7 (based on RH7) it doesn't.. Is something changed in new versions? For example, I run a migration by virsh. In libvirtd.log I see migration events, but don't see the "migrate" command itself, neither migrate-set-parameters / migrate-set-capabilities.. So, I never could rely on libvirt logs in this area, and we usually use Qemu tracing to debug QMP traffic. Also, Libvirt is not the only source of QMP traffic.. In Vz7 backup is done separately. I do backup of VM. In vm log I see drive-backup trace point. In Libvirt I see only "new connection, client: /usr/libexec/vz_backup_client <cmdline follows>"..
I doubt it'll make use of your filtering feature. Cc'ing libvir-list just in case.
Another way to log all traffic is to route it through socat -x or similar.
Opinions?
The benefit of Qemu tracepoints, is that you can enable them independently of any environment/management tool. -- Best regards, Vladimir
On Thu, Oct 14, 2021 at 06:22:32PM +0300, Vladimir Sementsov-Ogievskiy wrote:
12.10.2021 14:49, Markus Armbruster wrote:
Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> writes:
Hi all!
We have handle_qmp_command and qmp_command_repond trace points to trace qmp commands. They are very useful to debug problems involving management tools like libvirt.
But tracing all qmp commands is too much.
Here I suggest a kind of tracing namespace. Formally this series adds a trace points called qmp:<some-command> for every command, which may be enabled in separate like
--trace qmp:drive-backup
or by pattern like
--trace qmp:block-job-*
or similarly with help of qmp command trace-event-set-state.
This also allows to enable tracing of some qmp commands permanently (by downstream patch or in libvirt xml). For example, I'm going to enable tracing of block job comamnds and blockdev-* commands in Virtuozzo. Qemu logs are often too empty (for example, in comparison with Libvirt), logging block jobs is not too much but will be very helpful.
What exactly is traced? Peeking at PATCH 5... looks like it's input that makes it to qmp_dispatch() and command responses, but not events.
Fine print on "input that makes it to qmp_dispatch()":
* You trace right before we execute the command, not when we receive, parse and enqueue input.
* Corollary: input with certain errors is not traced.
* You don't trace the input text, you trace the unparsed parse tree.
All fine, I presume.
Existing tracepoints in monitor/qmp.c, and what information they send (inessential bits omitted for clarity):
* handle_qmp_command
Handling a QMP command: unparsed parse tree
Fine print, safe to ignore:
- Out-of-band commands will be executed right away, in-band commands will be queued. Tracepoints monitor_qmp_in_band_enqueue and monitor_qmp_in_band_dequeue provide insight into that.
- This also receives and queues parse errors, without tracing them. Tracepoint monitor_qmp_err_in_band traces them as they are dequeued.
* monitor_qmp_cmd_in_band
About to execute in-band command: command ID, if any
* monitor_qmp_cmd_out_of_band
About to execute out-of-band command: command ID, if any
* monitor_qmp_respond
About to send command response or event: QObject
For input, --trace qmp:* is like --trace handle_qmp_command, except it traces late rather than early.
For output, --trace qmp:* is like --trace monitor_qmp_respond less events.
The main improvement over existing tracepoints seems to be the ability to filter on command names.
To get that, you overload the @name argument of QMP command trace-event-set-state. In addition to the documented meaning "Event name pattern", it also has an alternate, undocumented meaning "QMP command name pattern". The "undocumented" part is easy enough to fix. However, QMP heavily frowns on argument values that need to be parsed.
Still, pattern is parsed anyway, as pattern. But yes, this patch adds rather specific and tricky logic, which a lot more than just a pattern to search through the list.
Another possible way is to update QAPI code generator to insert a personal trace point for each qmp command.. That seems more complicated to implement, but I can try.
That's what came to mind when I saw this series too. The QAPI generator can create a trace event for each QMP command. That way each command has a dedicated trace event that can be enabled/disabled in the usual way (e.g. built-in "trace" monitor command, SystemTap scripts, etc) without introducing special syntax. Stefan
Stefan Hajnoczi <stefanha@redhat.com> writes:
On Thu, Oct 14, 2021 at 06:22:32PM +0300, Vladimir Sementsov-Ogievskiy wrote:
12.10.2021 14:49, Markus Armbruster wrote:
Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> writes:
[...]
Another possible way is to update QAPI code generator to insert a personal trace point for each qmp command.. That seems more complicated to implement, but I can try.
That's what came to mind when I saw this series too. The QAPI generator can create a trace event for each QMP command. That way each command has a dedicated trace event that can be enabled/disabled in the usual way (e.g. built-in "trace" monitor command, SystemTap scripts, etc) without introducing special syntax.
I like how this way stays entirely within the existing interface. I couldn't tell you how to implement it, though :)
participants (3)
-
Markus Armbruster -
Stefan Hajnoczi -
Vladimir Sementsov-Ogievskiy