Re: [PATCH for-8.2 v2 1/2] qapi/migration: Deduplicate migration parameter field comments

[Peter Xu]

On Sat, Aug 05, 2023 at 10:12:00AM +0200, Markus Armbruster wrote:
> Peter Xu <peterx@redhat.com> writes:
> 
> > On Fri, Aug 04, 2023 at 05:48:49PM +0100, Daniel P. Berrangé wrote:
> >> On Fri, Aug 04, 2023 at 12:46:18PM -0400, Peter Xu wrote:
> >> > On Fri, Aug 04, 2023 at 05:29:19PM +0100, Daniel P. Berrangé wrote:
> >> > > On Fri, Aug 04, 2023 at 12:01:54PM -0400, Peter Xu wrote:
> >> > > > On Fri, Aug 04, 2023 at 02:59:07PM +0100, Daniel P. Berrangé wrote:
> >> > > > > On Fri, Aug 04, 2023 at 02:28:05PM +0200, Markus Armbruster wrote:
> >> > > > > > Peter Xu <peterx@redhat.com> writes:
> >> > > > > > 
> >> > > > > > > We used to have three objects that have always the same list 
> >> > > > > > > of parameters
> >> > > > > > 
> >> > > > > > We have!
> >> > > > > > 
> >> > > > > > > and comments are always duplicated:
> >> > > > > > >
> >> > > > > > >   - @MigrationParameter
> >> > > > > > >   - @MigrationParameters
> >> > > > > > >   - @MigrateSetParameters
> >> > > > > > >
> >> > > > > > > Before we can deduplicate the code, it's fairly 
> >> > > > > > > straightforward to
> >> > > > > > > deduplicate the comments first, so for each time we add a new 
> >> > > > > > > migration
> >> > > > > > > parameter we don't need to copy the same paragraphs three 
> >> > > > > > > times.
> >> > > > > > 
> >> > > > > > De-duplicating the code would be nice, but we haven't done so in 
> >> > > > > > years,
> >> > > > > > which suggests it's hard enough not to be worth the trouble.
> >> > > > > 
> >> > > > > The "MigrationParameter" enumeration isn't actually used in
> >> > > > > QMP at all.
> >> > > > > 
> >> > > > > It is only used in HMP for hmp_migrate_set_parameter and
> >> > > > > hmp_info_migrate_parameters. So it is questionable documenting
> >> > > > > that enum in the QMP reference docs at all.
> >> > > > > 
> >> > > > > 1c1
> >> > > > > < { 'struct': 'MigrationParameters',
> >> > > > > ---
> >> > > > > > { 'struct': 'MigrateSetParameters',
> >> > > > > 14,16c14,16
> >> > > > > <             '*tls-creds': 'str',
> >> > > > > <             '*tls-hostname': 'str',
> >> > > > > <             '*tls-authz': 'str',
> >> > > > > ---
> >> > > > > >             '*tls-creds': 'StrOrNull',
> >> > > > > >             '*tls-hostname': 'StrOrNull',
> >> > > > > >             '*tls-authz': 'StrOrNull',
> >> > > > > 
> >> > > > > Is it not valid to use StrOrNull in both cases and thus
> >> > > > > delete the duplication here ?
> >> > > > 
> >> > > > I tested removing MigrateSetParameters by replacing it with
> >> > > > MigrationParameters and it looks all fine here... I manually tested 
> >> > > > qmp/hmp
> >> > > > on set/query parameters, and qtests are all happy.
> >> > > 
> >> > > I meant the other way around, such we would be using 'StrOrNull'
> >> > > in all scenarios.
> >> > 
> >> > Yes, that should also work and even without worrying on nulls.  I just 
> >> > took
> >> > a random one replacing the other.
> >> > 
> >> > > 
> >> > > > 
> >> > > > The only thing I see that may affect it is we used to logically allow
> >> > > > taking things like '"tls-authz": null' in the json input, but now we 
> >> > > > won't
> >> > > > allow that because we'll be asking for a string type only.
> >> > > > 
> >> > > > Since we have query-qmp-schema I suppose we're all fine, because 
> >> > > > logically
> >> > > > the mgmt app (libvirt?) will still query that to understand the 
> >> > > > protocol,
> >> > > > so now we'll have (response of query-qmp-schema):
> >> > > > 
> >> > > >         {
> >> > > >             "arg-type": "144",
> >> > > >             "meta-type": "command",
> >> > > >             "name": "migrate-set-parameters",
> >> > > >             "ret-type": "0"
> >> > > >         },
> >> > > > 
> >> > > > Where 144 can start to point to MigrationParameters, rather than
> >> > > > MigrateSetParameters.
> >> > > > 
> >> > > > Ok, then what if the mgmt app doesn't care and just used "null" in 
> >> > > > tls-*
> >> > > > fields when setting?  Funnily I tried it and actually anything that 
> >> > > > does
> >> > > > migrate-set-parameters with a "null" passed over to tls-* fields will
> >> > > > already crash qemu...
> >> > > > 
> >> > > > ./migration/options.c:1333: migrate_params_apply: Assertion 
> >> > > > `params->tls_authz->type == QTYPE_QSTRING' failed.
> >> > > > 
> >> > > > #0  0x00007f72f4b2a844 in __pthread_kill_implementation () at 
> >> > > > /lib64/libc.so.6
> >> > > > #1  0x00007f72f4ad9abe in raise () at /lib64/libc.so.6
> >> > > > #2  0x00007f72f4ac287f in abort () at /lib64/libc.so.6
> >> > > > #3  0x00007f72f4ac279b in _nl_load_domain.cold () at /lib64/libc.so.6
> >> > > > #4  0x00007f72f4ad2147 in  () at /lib64/libc.so.6
> >> > > > #5  0x00005573308740e6 in migrate_params_apply 
> >> > > > (params=0x7ffc74fd09d0, errp=0x7ffc74fd0998) at 
> >> > > > ../migration/options.c:1333
> >> > > > #6  0x0000557330874591 in qmp_migrate_set_parameters 
> >> > > > (params=0x7ffc74fd09d0, errp=0x7ffc74fd0998) at 
> >> > > > ../migration/options.c:1433
> >> > > > #7  0x0000557330cb9132 in qmp_marshal_migrate_set_parameters 
> >> > > > (args=0x7f72e00036d0, ret=0x7f72f133cd98, errp=0x7f72f133cd90) at 
> >> > > > qapi/qapi-commands-migration.c:214
> >> > > > #8  0x0000557330d07fab in do_qmp_dispatch_bh (opaque=0x7f72f133ce30) 
> >> > > > at ../qapi/qmp-dispatch.c:128
> >> > > > #9  0x0000557330d33bbb in aio_bh_call (bh=0x5573337d7920) at 
> >> > > > ../util/async.c:169
> >> > > > #10 0x0000557330d33cd8 in aio_bh_poll (ctx=0x55733356e7d0) at 
> >> > > > ../util/async.c:216
> >> > > > #11 0x0000557330d17a19 in aio_dispatch (ctx=0x55733356e7d0) at 
> >> > > > ../util/aio-posix.c:423
> >> > > > #12 0x0000557330d34117 in aio_ctx_dispatch (source=0x55733356e7d0, 
> >> > > > callback=0x0, user_data=0x0) at ../util/async.c:358
> >> > > > #13 0x00007f72f5a8848c in g_main_context_dispatch () at 
> >> > > > /lib64/libglib-2.0.so.0
> >> > > > #14 0x0000557330d358d4 in glib_pollfds_poll () at 
> >> > > > ../util/main-loop.c:290
> >> > > > #15 0x0000557330d35951 in os_host_main_loop_wait (timeout=0) at 
> >> > > > ../util/main-loop.c:313
> >> > > > #16 0x0000557330d35a5f in main_loop_wait (nonblocking=0) at 
> >> > > > ../util/main-loop.c:592
> >> > > > #17 0x000055733083aee0 in qemu_main_loop () at 
> >> > > > ../softmmu/runstate.c:732
> >> > > > #18 0x0000557330b0921b in qemu_default_main () at 
> >> > > > ../softmmu/main.c:37
> >> > > > #19 0x0000557330b09251 in main (argc=35, argv=0x7ffc74fd0ec8) at 
> >> > > > ../softmmu/main.c:48
> >> > > > 
> >> > > > Then I suppose it means all mgmt apps are not using "null" anyway, 
> >> > > > and it
> >> > > > makes more sense to me to just remove MigrateSetParameters (by 
> >> > > > replacing it
> >> > > > with MigrationParameters).
> >> > > 
> >> > > It shouldn't be crashing,  because qmp_migrate_set_parameters()
> >> > > is turning 'null' into  "", which means the assert ought to
> >> > > never fire. Did you have a local modiification that caused
> >> > > this crash perhaps ?
> >> > 
> >> > I think it just got overlooked when introducing tls-authz to not have 
> >> > added
> >> > that special code in qmp_migrate_set_parameters(), the other two are 
> >> > fine.
> >> 
> >> Oh right yes, pre-existing bug.
> >
> > So do we really care about "null" in any form over "" (empty str) here for
> > tls-* parameters?
> 
> In my opinion, the use of "" was a design mistake.  Here's my argument:
> 
> commit 01fa55982692fb51a16049b63b571651a1053989
> Author: Markus Armbruster <armbru@redhat.com>
> Date:   Tue Jul 18 14:42:04 2017 +0200
> 
>     migration: Use JSON null instead of "" to reset parameter to default
>     
>     migrate-set-parameters sets migration parameters according to is
>     arguments like this:
>     
>     * Present means "set the parameter to this value"
>     
>     * Absent means "leave the parameter unchanged"
>     
>     * Except for parameters tls_creds and tls_hostname, "" means "reset
>       the parameter to its default value
>     
>     The first two are perfectly normal: presence of the parameter makes
>     the command do something.
>     
>     The third one overloads the parameter with a second meaning.  The
>     overloading is *implicit*, i.e. it's not visible in the types.  Works
>     here, because "" is neither a valid TLS credentials ID, nor a valid
>     host name.
>     
>     Pressing argument values the schema accepts, but are semantically
>     invalid, into service to mean "reset to default" is not general, as
>     suitable invalid values need not exist.  I also find it ugly.
>     
>     To clean this up, we could add a separate flag argument to ask for
>     "reset to default", or add a distinct value to @tls_creds and
>     @tls_hostname.  This commit implements the latter: add JSON null to
>     the values of @tls_creds and @tls_hostname, deprecate "".
>     
>     Because we're so close to the 2.10 freeze, implement it in the
>     stupidest way possible: have qmp_migrate_set_parameters() rewrite null
>     to "" before anything else can see the null.  The proper way to do it
>     would be rewriting "" to null, but that requires fixing up code to
>     work with null.  Add TODO comments for that.
>     
>     Signed-off-by: Markus Armbruster <armbru@redhat.com>
>     Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
>     Reviewed-by: Eric Blake <eblake@redhat.com>

I see.  Personally I think as long as the interface is 100% clear I'll be
all fine (say, no possible misuse of "").  But keeping StrOrNull may
definitely be cleaner I guess.

> 
> > To fix this tls-authz bug we can add one more QTYPE_QNULL to QTYPE_QSTRING
> > convertion, but I'd rather just use "str" for all tls* fields and remove
> > the other two instead, if "null" is not important to anyone.
> 
> "Important" sounds too much like absolutes :)
> 
> I think we have a tradeoff here.  If perpetuating the unclean and ugly
> use of "" is what it takes to de-triplicate migration parameters, we may
> decide to accept that.

I don't think it's a must.  As Dan raised, we can convert str -> StrOrNull
for MigrationParameters. I assume it won't affect query-migrate-parameters
anyway OTOH.

I assume it means there's nothing yet obvious that we overlooked on the
whole idea.  Let me propose the formal patchset early next week.  It'll be
mostly the patch I attached but just add those extra logics for StrOrNull,
so the diffstat might be less attractive but hopefully still good enough to
be accepted.

Thanks,

-- 
Peter Xu