Re: [Qemu-devel] [PATCH] qapi: Reintroduce CommandDisabled error class

[Markus Armbruster]

Michal Privoznik <address@hidden> writes:

> On 8/29/19 3:12 PM, Eric Blake wrote:
>> On 8/29/19 8:04 AM, Michal Privoznik wrote:
>>
>>>>> A bit of background: up until very recently libvirt used qemu-ga
>>>>> in all or nothing way. It didn't care why a qemu-ga command
>>>>> failed. But very recently a new API was introduced which
>>>>> implements 'best effort' approach (in some cases) and thus
>>>>> libvirt must differentiate between: {CommandNotFound,
>>>>> CommandDisabled} and some generic error. While the former classes
>>>>> mean the API can issue some other commands the latter raises a
>>>>> red flag causing the API to fail.
>>>>
>>>> Why do you need to distinguish CommandNotFound from CommandDisabled?
>>>
>>> I don't. That's why I've put them both in curly braces. Perhaps this
>>> says its better:
>>>
>>> switch (klass) {
>>>    case CommandNotFound:
>>>    case CommandDisabled:
>>>          /* okay */
>>>          break;
>>>
>>
>> So the obvious counter-question - why not use class CommandNotFound for
>> a command that was disabled, rather than readding another class that has
>> no distinctive purpose?
>>
>>
>
> Because disabling a command is not the same as nonexistent
> command. While a command can be disabled by user/sysadmin, they are
> disabled at runtime by qemu-ga itself for a short period of time
> (e.g. on FS freeze some commands are disabled - typically those which
> require write disk access). And I guess reporting CommandNotFound for
> a command that does exist only is disabled temporarily doesn't reflect
> the reality, does it?
>
> On the other hand, CommandNotFound would fix the issue for libvirt, so
> if you don't want to invent a new error class, then that's the way to
> go.

I'm fine with changing the error to CommandNotFound.

I'm reluctant to add back CommandDisabled.  I doubt it's necessary.

To arrive at an informed opinion, I had to figure out how this command
disablement stuff works.  I can just as well send it out, so here goes.

Let's review our command disable feature.

Commands are enabled on registration, see qmp_register_command().

To disable, call qmp_disable_command().  Only qga/main.c does, in two
places:

* ga_disable_non_whitelisted(): disable all commands except for
  ga_freeze_whitelist[], which is documented as /* commands that are
  safe to issue while filesystems are frozen */

* initialize_agent(): disable blacklisted commands.  I figure these are
  the ones blacklisted with -b, plus commands blacklisted due to build
  configuration.  The latter feels inappropriate; we should use QAPI
  schema conditionals to compile them out instead (QAPI conditionals
  didn't exist when the blacklisting code was written).

Disabled commands can be re-enabled with qmp_enable_command().  Only
qga/main.c does, in ga_enable_non_blacklisted().  I figure it re-enables
the commands ga_disable_non_whitelisted() disables.  Gets called when
guest-fsfreeze-freeze freezes nothing[1], and when guest-fsfreeze-thaw
succeeds[2].

Command dispatch fails when the command is disabled, in
do_qmp_dispatch().  The proposed patch changes the error reply.

QGA's guest-info shows whether a command is disabled
(GuestAgentCommandInfo member @enabled, set in qmp_command_info()).

QMP's query-commands skips disabled commands, in query_commands_cb().
Dead, as nothing ever disables QMP commands.  Skipping feels like a bad
idea anyway.

Analysis:

There are three kinds of disabled commands: compile-time (should be
compiled out instead), permanently blacklisted with -b, temporarily
disabled while filesystems are frozen.

There are two states: thawed (first two kinds disabled) and frozen (all
three kinds disabled).

Command guest-fsfreeze-freeze[3] goes to state frozen or else fails.

Command guest-fsfreeze-thaw goes to state thawed or else fails.

guest-fsfreeze-status reports the state.

Note that the transition to frozen (and thus the temporary command
disablement) is under the control of the QGA client.  There is no
TOCTTOU between guest-info telling you which commands are disabled and
executing the next command.  My point is: the client can figure out
whether a command is disabled before executing it.

Of course, that doesn't mean we should make it figure it out.


[1] POSIX only, WTF?

[2] Except for execute_fsfreeze_hook(), which can still fail the command
    on POSIX, WTF?

[3] guest-fsfreeze-freeze's doc comment notes "The frozen state is
    limited for up to 10 seconds by VSS."  Sounds like some spontaneous
    transition back to thawed.  If this is actually true, GAState member
    @frozen is not updated to reflect the spontaneous thaw.  WTF?