OSDN Git Service

migration: Unify and trace vmstate field_exists() checks
authorPeter Xu <peterx@redhat.com>
Wed, 6 Sep 2023 20:47:22 +0000 (16:47 -0400)
committerJuan Quintela <quintela@redhat.com>
Wed, 4 Oct 2023 11:19:47 +0000 (13:19 +0200)
For both save/load we actually share the logic on deciding whether a field
should exist.  Merge the checks into a helper and use it for both save and
load.  When doing so, add documentations and reformat the code to make it
much easier to read.

The real benefit here (besides code cleanups) is we add a trace-point for
this; this is a known spot where we can easily break migration
compatibilities between binaries, and this trace point will be critical for
us to identify such issues.

For example, this will be handy when debugging things like:

https://gitlab.com/qemu-project/qemu/-/issues/932

Reviewed-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
Message-ID: <20230906204722.514474-1-peterx@redhat.com>

migration/trace-events
migration/vmstate.c

index 3e9649a..002abe3 100644 (file)
@@ -66,6 +66,7 @@ vmstate_save_state_loop(const char *name, const char *field, int n_elems) "%s/%s
 vmstate_save_state_top(const char *idstr) "%s"
 vmstate_subsection_save_loop(const char *name, const char *sub) "%s/%s"
 vmstate_subsection_save_top(const char *idstr) "%s"
+vmstate_field_exists(const char *vmsd, const char *name, int field_version, int version, int result) "%s:%s field_version %d version %d result %d"
 
 # vmstate-types.c
 get_qtailq(const char *name, int version_id) "%s v%d"
index 4cde30b..1cf9e45 100644 (file)
@@ -26,6 +26,30 @@ static int vmstate_subsection_save(QEMUFile *f, const VMStateDescription *vmsd,
 static int vmstate_subsection_load(QEMUFile *f, const VMStateDescription *vmsd,
                                    void *opaque);
 
+/* Whether this field should exist for either save or load the VM? */
+static bool
+vmstate_field_exists(const VMStateDescription *vmsd, const VMStateField *field,
+                     void *opaque, int version_id)
+{
+    bool result;
+
+    if (field->field_exists) {
+        /* If there's the function checker, that's the solo truth */
+        result = field->field_exists(opaque, version_id);
+        trace_vmstate_field_exists(vmsd->name, field->name, field->version_id,
+                                   version_id, result);
+    } else {
+        /*
+         * Otherwise, we only save/load if field version is same or older.
+         * For example, when loading from an old binary with old version,
+         * we ignore new fields with newer version_ids.
+         */
+        result = field->version_id <= version_id;
+    }
+
+    return result;
+}
+
 static int vmstate_n_elems(void *opaque, const VMStateField *field)
 {
     int n_elems = 1;
@@ -105,10 +129,7 @@ int vmstate_load_state(QEMUFile *f, const VMStateDescription *vmsd,
     }
     while (field->name) {
         trace_vmstate_load_state_field(vmsd->name, field->name);
-        if ((field->field_exists &&
-             field->field_exists(opaque, version_id)) ||
-            (!field->field_exists &&
-             field->version_id <= version_id)) {
+        if (vmstate_field_exists(vmsd, field, opaque, version_id)) {
             void *first_elem = opaque + field->offset;
             int i, n_elems = vmstate_n_elems(opaque, field);
             int size = vmstate_size(opaque, field);
@@ -349,10 +370,7 @@ int vmstate_save_state_v(QEMUFile *f, const VMStateDescription *vmsd,
     }
 
     while (field->name) {
-        if ((field->field_exists &&
-             field->field_exists(opaque, version_id)) ||
-            (!field->field_exists &&
-             field->version_id <= version_id)) {
+        if (vmstate_field_exists(vmsd, field, opaque, version_id)) {
             void *first_elem = opaque + field->offset;
             int i, n_elems = vmstate_n_elems(opaque, field);
             int size = vmstate_size(opaque, field);