On 13/05/2022 13:21, Markus Armbruster wrote:
Laurent Vivier <lviv...@redhat.com> writes:
As qemu_opts_parse_noisily() flattens the QAPI structures ("type" field
of Netdev structure can collides with "type" field of SocketAddress),
To remember how this works, I have to write a more verbose version of
the above. Why not post it then, so here goes.
qemu_init() passes the argument of -netdev, -nic, and -net to
net_client_parse().
net_client_parse() parses with qemu_opts_parse_noisily(), passing
QemuOptsList qemu_netdev_opts for -netdev, qemu_nic_opts for -nic, and
qemu_net_opts for -net. Their desc[] are all empty, which means any
keys are accepted. The result of the parse (a QemuOpts) is stored in
the QemuOptsList.
Note that QemuOpts is flat by design. In some places, we layer non-flat
on top using dotted keys convention, but not here.
net_init_clients() iterates over the stored QemuOpts, and passes them to
net_init_netdev(), net_param_nic(), or net_init_client(), respectively.
These functions pass the QemuOpts to net_client_init(). They also do
other things with the QemuOpts, which we can ignore here.
net_client_init() uses the opts visitor to convert the (flat) QemOpts to
a (non-flat) QAPI object Netdev. Netdev is also the argument of QMP
command netdev_add.
The opts visitor was an early attempt to support QAPI in
(QemuOpts-based) CLI. It restricts QAPI types to a certain shape; see
commit eb7ee2cbeb "qapi: introduce OptsVisitor".
A more modern way to support QAPI is qobject_input_visitor_new_str().
It uses keyval_parse() instead of QemuOpts for KEY=VALUE,... syntax, and
it also supports JSON syntax. The former isn't quite as expressive as
JSON, but it's a lot closer than QemuOpts + opts visitor.
we introduce a way to bypass qemu_opts_parse_noisily() and use directly
visit_type_Netdev() to parse the backend parameters.
This commit paves the way to use of the modern way instead.
I'm going to copy your analysis to the commit message of the patch.
Signed-off-by: Laurent Vivier <lviv...@redhat.com>
---
net/net.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 54 insertions(+)
diff --git a/net/net.c b/net/net.c
index 58c05c200622..2aab7167316c 100644
--- a/net/net.c
+++ b/net/net.c
@@ -54,6 +54,7 @@
#include "net/colo-compare.h"
#include "net/filter.h"
#include "qapi/string-output-visitor.h"
+#include "qapi/qobject-input-visitor.h"
/* Net bridge is currently not supported for W32. */
#if !defined(_WIN32)
@@ -63,6 +64,17 @@
static VMChangeStateEntry *net_change_state_entry;
static QTAILQ_HEAD(, NetClientState) net_clients;
+typedef struct NetdevQueueEntry {
+ bool is_netdev;
+ Netdev *nd;
+ Location loc;
+ QSIMPLEQ_ENTRY(NetdevQueueEntry) entry;
+} NetdevQueueEntry;
+
+typedef QSIMPLEQ_HEAD(, NetdevQueueEntry) NetdevQueue;
+
+static NetdevQueue nd_queue = QSIMPLEQ_HEAD_INITIALIZER(nd_queue);
+
/***********************************************************/
/* network device redirectors */
@@ -1559,6 +1571,19 @@ int net_init_clients(Error **errp)
QTAILQ_INIT(&net_clients);
+ while (!QSIMPLEQ_EMPTY(&nd_queue)) {
+ NetdevQueueEntry *nd = QSIMPLEQ_FIRST(&nd_queue);
+
+ QSIMPLEQ_REMOVE_HEAD(&nd_queue, entry);
+ loc_push_restore(&nd->loc);
+ if (net_client_init1(nd->nd, nd->is_netdev, errp) < 0) {
I think you need to loc_pop() here.
+ return -1;
+ }
Since the only caller passes &error_fatal, I'd be tempted to ditch the
@errp argument, and simply do
net_client_init1(nd->nd, nd->is_netdev, &error_fatal);
It's what we do for -blockdev, -device, and -object.
I've added a patch to remove the @errp from the net_init_clients() arguments.
+ loc_pop(&nd->loc);
+ qapi_free_Netdev(nd->nd);
+ g_free(nd);
+ }
+
if (qemu_opts_foreach(qemu_find_opts("netdev"),
net_init_netdev, NULL, errp)) {
return -1;
@@ -1575,8 +1600,37 @@ int net_init_clients(Error **errp)
return 0;
}
+/*
+ * netdev_is_modern() returns true when the backend needs to bypass
+ * qemu_opts_parse_noisily()
+ */
+static bool netdev_is_modern(const char *optarg)
+{
+ return false;
+}
+
int net_client_parse(QemuOptsList *opts_list, const char *optarg)
{
+ if (netdev_is_modern(optarg)) {
+ /*
+ * We need to bypass qemu_opts_parse_noisily() to accept
+ * new style object like addr.type=inet in SocketAddress
+ */
I'm not sure this will makes sense to future readers.
What about "Use modern, more expressive syntax"?
Done.
+ Visitor *v;
+ NetdevQueueEntry *nd;
+
+ v = qobject_input_visitor_new_str(optarg, "type",
+ &error_fatal);
+ nd = g_new(NetdevQueueEntry, 1);
+ visit_type_Netdev(v, NULL, &nd->nd, &error_fatal);
+ visit_free(v);
+ loc_save(&nd->loc);
+ nd->is_netdev = strcmp(opts_list->name, "netdev") == 0;
+
+ QSIMPLEQ_INSERT_TAIL(&nd_queue, nd, entry);
+ return 0;
+ }
Matches what we do for -blockdev, except we additionally have
nd->is_netdev. We need it for calling net_client_init1().
If netdev_is_modern(optarg), then the only use of parameter @opts_list
is opts_list->name in the initialization of nd->is_netdev.
There's a bit of code smell, I'm afraid.
I don't see what is the problem.
I believe @is_netdev needs to be true for -netdev, -nic, and netdev_add;
false for -net.
Will we ever use the modern syntax with -net?
Yes, I think we should support the same syntax with -netdev and -net.
My first iteration was to pass is_netdev=true to net_client_init1() and Stefano reported a
problem with "-net" with things like that:
-net dgram,id=socket0,local.type=inet,local.host=localhost,local.port=1234,\
remote.type=inet,remote.host=localhost,remote.port=1235
-net nic,model=virtio,macaddr=9a:2b:2c:2d:2e:2f
Any chance we can deprecate -net?
Who can decide of that?
+
if (!qemu_opts_parse_noisily(opts_list, optarg, true)) {
return -1;
}
Thanks,
Laurent
