[Qemu-devel] Re: QError conversion problems: putting errors in context

[Markus Armbruster]

Anthony Liguori <address@hidden> writes:

> Hi Markus,
>
> On 02/12/2010 11:48 AM, Markus Armbruster wrote:
>> Our QError conversions were pretty straightforward so far.  For example,
>> when we found
>>
>>                  monitor_printf(mon, "device is not removable\n");
>>
>> in eject_device(), we created the obvious QError class for it:
>>
>> #define QERR_DEVICE_NOT_REMOVABLE \
>>      "{ 'class': 'DeviceNotRemovable', 'data': { 'device': %s } }"
>>
>> with the obvious pretty-print template:
>>
>>      {
>>          .error_fmt = QERR_DEVICE_NOT_REMOVABLE,
>>          .desc      = "Device %(device) is not removable",
>>      },
>>
>> and replaced the print with
>>
>>                  qemu_error_new(QERR_DEVICE_NOT_REMOVABLE,
>>                                 bdrv_get_device_name(bs));
>>
>> This even improved the message from "device is not removable" to "device
>> hda is not removable".  Commit 2c2a6bb8.
>>
>>
>> We also ran into cases like this one, in qdev_device_add():
>>
>> -        qemu_error("Device \"%s\" not found.  Try -device '?' for a 
>> list.\n",
>> -                   driver);
>> +        qemu_error_new(QERR_DEVICE_NOT_FOUND, driver);
>>
>> Here, the pretty-print template became
>>
>>          .desc      = "The %(device) device has not been found",
>>
>> Note the loss of the helpful "Try -device '?' for a list." part.  That's
>> because the same error is used in other places, where that piece of
>> advice doesn't apply.  Commit 3ced9f7a.
>>
>>
>> For similar reasons, "Invalid CPU index" changed to the more generic
>> "Invalid parameter index" in monitor command cpu (commit cc0c4185).
>> More examples in my private tree: set_link's "invalid link status 'off';
>> only 'up' or 'down' valid" becomes "Invalid parameter up_or_down", and
>> migrate's "migration already in progress" becomes "Already in progress".
>>
>>
>> Most of the conversions touched only monitor code.  Things get much more
>> complicated for code shared between the monitor and other places.  For
>> instance, there's this one in qdev_device_add():
>>
>>          qemu_error("-device: no driver specified\n");
>>
>> Of course, this message is pretty pathetic already, in several ways:
>>
>> * A command line can contain several -device, and the error message
>>    leaves finding the offending one to the user.
>>
>> * It's even worse with configuration files.  FILE:LINENR is par for that
>>    course.
>>
>> * It talks about -device even when we're in the monitor or a
>>    configuration file.  To trigger it in the monitor, try "device_add
>>    a=b".
>>
>>  From QMP's point of view, the appropriate error to use is this one:
>>
>> #define QERR_MISSING_PARAMETER \
>>      "{ 'class': 'MissingParameter', 'data': { 'name': %s } }"
>>
>> It's pretty-print template is
>>
>>          .desc      = "Parameter %(name) is missing",
>>
>> This changes the message "-device: no driver specified" to "Parameter
>> driver is missing".  We go from bad to worse.
>>
>>
>> I think we have a few related but nevertheless distinct issues here:
>>
>> * We need to identify the error message's object.  Proper identification
>>    depends on the context in which the code is executing:
>>
>>    - Command line: the option, with its argument, if any.
>>
>>    - Config file: filename, line number.
>>
>>    - Monitor: command (this is trivial)
>>
>>    This is what's lacking in the "no driver specified" example, even
>>    before conversion to QError.
>>
>>    Dragging context information along all over the place so we can use it
>>    to make better error messages would be cumbersome and invasive.  There
>>    must be a better way.
>>
>>    Note that qemu_error_new() already has a primitive notion of context:
>>    it distinguishes between human monitor, QMP monitor, and anything
>>    else.  This could be refined, so it can add suitable context
>>    information to the error without encumbering its callers.  For
>>    instance, a hypothetical error message "pants on fire", emitted with
>>
>>        qemu_error_new(QERR_PANTS_ON_FILE)
>>
>>    should become
>>
>>    - "qemu-system-x86_64: -device pants,color=black: pants on fire" in
>>      the context of command line option "-device pants,color=black",
>>
>>    - "vm.cfg:123: pants on fire" in the context of configuration file
>>      vm.cfg, line 123,
>>
>>    - "pants on fire" in the human monitor, and
>>
>>    - { "error": { "class": "PantsOnFire", "data": {},
>>                   "desc": "pants on fire" } }
>>      in QMP.
>>
>> * Occasionally, we'd like to add help, but not for all uses of the same
>>    error, and not in all contexts.
>>
>>    This is the "Try -device '?' for a list" example.  We'd like to print
>>    that if we're working for -device (user qdev_device_add(), context
>>    command line).  And maybe "Try device_add ? for a list" if we're
>>    working for device_add (same user, context human monitor).
>>
>>    I believe this is rare enough to permit an ad hoc solution, like this:
>>
>>          qemu_error_new(QERR_DEVICE_NOT_FOUND, driver);
>>          printf_if_cmdline("Try -device '?' for a list.\n");
>>          printf_if_human_monitor("Try device_add ? for a list.\n");
>>
>>    Wouldn't work for the monitor right now, because the actual printing
>>    of the error gets delayed until the handler returns, but that's
>>    fixable.
>>
>> * Nice human-readable messages sometimes need information that should
>>    not go into QMP.
>>
>>    This is the "Migration already in progress" example.  Since different
>>    things can be in progress, it is tempting to do
>>
>>      #define QERR_ALREADY_IN_PROGRESS \
>>          "{ 'class': 'AlreadyInProgress', 'data': { 'activity': %s } }"
>>
>>    with pretty-print template
>>
>>      .desc      = "%(activity) already in progress"
>>
>>    and use it like this:
>>
>>      qemu_error_new(QERR_FAILED, "migration");
>>
>>    But that puts a rather pointless 'activity' member into the data
>>    object.
>>
>>    An obvious way to avoid that is to adopt a convention like "if the key
>>    starts with '_', it's internal, and not to be passed over QMP".
>>
>>    Another way (suggested by Anthony) is to permit overriding the
>>    pretty-print template somehow.  More flexible, but also more clumsy to
>>    use.
>>
>> Comments?
>>    
>
> Thanks for the write-up.  Here's my view of how error reporting ought
> to work:
>
> 0) A user tries to do something via some piece of code that serves as
> the UI.  This may be the -device option handling, the config file
> parser, QMP, or the human monitor.
> 1) Something, somewhere, generates an error.  This is may be as
> primitive as an errno or as sophisticated as a QError object.  In both
> cases, the error consists of structured data will well defined
> semantics.
> 2) That error is propagated up the call chain back to the UI code.
> 3) The UI code then decides how to present the error to the user.

This is how QError actually works.  The trouble is that our error
presentation is too primitive.

> In an ideal world, qdev_device_add would either return a QError object
> or just an errno.  The caller of qdev_device_add would then decide how
> to display that error to the user.
>
> The whole notion of having a description table baked into QError is
> wrong IMHO.  You want as much contextual information as possible when
> generating an error for a user and the place you have that is at the
> point you're interacting with the user.  It makes no sense to generate
> error messages deep in the bowels of QEMU.

Fine in theory, but there are practical problems with lifting all error
message generation up into UI code.

Lifting errors means we duplicate them in every UI where they can occur.
This is maximally flexible: we can do completely different things in
every single case.  But I doubt we need that flexibility, because in
practice, a useful error message can be created from a part that depends
on the error, but not the context (this is the message's subject), and
another part that depends on context but not the error (this is it's
object).  This is how my "pants on fire" example works.

Another practical consideration is what you have to do to report a new
error.  Before QError, you just print it where it gets diagnosed.  One
place to change.  With QError, you define a QERR_<FOO> in qerror.h, you
define its pretty-printing in qerror.c, and then you create the QError
where the error gets diagnosed.  Three places.  With error lifting, you
get to define the pretty-printing once for each UI.  N+2 places.  I
count N==3 right now.

We may wish to vary the pretty-printing even within a single UI.  For
instance, in my "Migration already in progress" example,
QERR_ALREADY_IN_PROGRESS gets pretty-printed specially for command
"migrate".  Yet a few more places to change.

> I'm a fan of decoupling the pretty printing from qerror.c and moving
> it into the monitor/command line parser.

I'm no friend of qerror.c either, but duplicating it several times over
doesn't strike me as an improvement.