We've had all the required pieces for doing a type-safe representation
of netdev_add as a flat union for quite some time now (since
0e55c381f6 in v2.7.0, released in 2016), but did not make the final
switch to using it because of concern about whether a command-line
regression in accepting "1" in place of 1 for integer arguments would
be problematic. Back then, we did not have the deprecation cycle to
allow us to make progress. But now that we have waited so long, other
problems have crept in: for example, our desire to add
qemu-storage-daemon is hampered by the inability to express net
objects, and we are unable to introspect what we actually accept.
Additionally, our round-trip through QemuOpts silently eats any
argument that expands to an array, rendering dnssearch, hostfwd, and
guestfwd useless through QMP:
{"execute": "netdev_add", "arguments": { "id": "netdev0",
"type": "user", "dnssearch": [
{ "str": "8.8.8.8" }, { "str": "8.8.4.4" }
]}}
So without further ado, let's turn on proper QAPI. netdev_add() was a
trivial wrapper around net_client_init(), which did a few steps prior
to calling net_client_init1(); with this patch, we now skip directly
to net_client_init1(). In addition to fixing array parameters, the
following additional differences occur:
- {"execute": "netdev_add", "arguments": {"type": "help"}}
no longer attempts to print help to stdout and exit. Bug fix, broken
in 547203ead4 'net: List available netdevs with "-netdev help"',
v2.12.0.
- {"execute": "netdev_add", "arguments': {... "ipv6-net": "..." }}
no longer attempts to desugar the undocumented ipv6-net magic string
into the proper "ipv6-prefix" and "ipv6-prefixlen". Undocumented
misfeature, introduced in commit 7aac531ef2 "qapi-schema, qemu-options
& slirp: Adding Qemu options for IPv6 addresses", v2.6.0.
- {'execute':'netdev_add',
'arguments':{'id':'net2', 'type':'hubport', 'hubid':"2"}}
{"error": {"class": "GenericError", "desc": "Invalid parameter type for 'hubid', expected: integer"}}
Used to succeed: since our command line treats everything as strings,
our not-so-round-trip conversion from QAPI -> QemuOpts -> QAPI lost
the original typing and turned everything into a string; now that we
skip the QemuOpts, the JSON input has to match the exact QAPI type.
But this stricter QMP is desirable, and introspection is sufficient
for any affected applications to make sure they use it correctly.
In qmp_netdev_add(), we still have to create a QemuOpts object so that
qmp_netdev_del() will be able to remove a hotplugged network device;
but the opts->head remains empty since we now manage all parsing
through the QAPI object rather than QemuOpts; a separate patch will
address the abuse of QemuOpts as a witness for whether a
NetClientState is a netdev. In the meantime, our argument that we are
okay requires auditing all uses of option group "netdev":
- qemu_netdev_opts: option group definition, empty .desc[]
- CLI (CLI netdev parsing ends before monitors start, so while
monitors can mess with CLI netdevs, CLI cannot mess with
monitor netdevs):
- main() case QEMU_OPTION_netdev: store CLI definition
- main() case QEMU_OPTION_readconfig, case QEMU_OPTION_writeconfig:
similar, dealing only with CLI
- net_init_clients(): Pass CLI to net_client_init()
- Monitor:
- hmp_netdev_add(): straightforward parse into net_client_init()
- qmp_netdev_add(): subject of this patch, used to add full
object to option group, now just adds bare-bones id
- qmp_netdev_del(), netdev_del_completion(): check the option group
solely for id, as a 'is this a netdev' predicate
Reported-by: Alex Kirillov <lekiravi@yandex-team.ru>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20200317201711.322764-2-eblake@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
[Commit message typo fixed]
Signed-off-by: Markus Armbruster <armbru@redhat.com>