Previously, working with alternates required two lookup arrays
and some indirection: for type Foo, we created Foo_qtypes[]
which maps each qtype to a value of the generated FooKind enum,
then look up that value in FooKind_lookup[] like we do for other
union types.
This has a couple of subtle bugs. First, the generator was
creating a call with a parameter '(int *) &(*obj)->type' where
type is an enum type; this is unsafe if the compiler chooses
to store the enum type in a different size than int, where
assigning through the wrong size pointer can corrupt data or
cause a SIGBUS.
Related bug, not not fixed in this patch: qapi-visit.py's
gen_visit_enum() generates a cast of its enum * argument to
int *. Marked FIXME.
Second, since the values of the FooKind enum start at zero, all
entries of the Foo_qtypes[] array that were not explicitly
initialized will map to the same branch of the union as the
first member of the alternate, rather than triggering a desired
failure in visit_get_next_type(). Fortunately, the bug seldom
bites; the very next thing the input visitor does is try to
parse the incoming JSON with the wrong parser, which normally
fails; the output visitor is not used with a C struct in that
state, and the dealloc visitor has nothing to clean up (so
there is no leak).
However, the second bug IS observable in one case: parsing an
integer causes unusual behavior in an alternate that contains
at least a 'number' member but no 'int' member, because the
'number' parser accepts QTYPE_QINT in addition to the expected
QTYPE_QFLOAT (that is, since 'int' is not a member, the type
QTYPE_QINT accidentally maps to FooKind 0; if this enum value
is the 'number' branch the integer parses successfully, but if
the 'number' branch is not first, some other branch tries to
parse the integer and rejects it). A later patch will worry
about fixing alternates to always parse all inputs that a
non-alternate 'number' would accept, for now this is still
marked FIXME in the updated test-qmp-input-visitor.c, to
merely point out that new undesired behavior of 'ans' matches
the existing undesired behavior of 'asn'.
This patch fixes the default-initialization bug by deleting the
indirection, and modifying get_next_type() to directly assign a
QTypeCode parameter. This in turn fixes the type-casting bug,
as we are no longer casting a pointer to enum to a questionable
size. There is no longer a need to generate an implicit FooKind
enum associated with the alternate type (since the QMP wire
format never uses the stringized counterparts of the C union
member names). Since the updated visit_get_next_type() does not
know which qtypes are expected, the generated visitor is
modified to generate an error statement if an unexpected type is
encountered.
Callers now have to know the QTYPE_* mapping when looking at the
discriminator; but so far, only the testsuite was even using the
C struct of an alternate types. I considered the possibility of
keeping the internal enum FooKind, but initialized differently
than most generated arrays, as in:
typedef enum FooKind {
FOO_KIND_A = QTYPE_QDICT,
FOO_KIND_B = QTYPE_QINT,
} FooKind;
to create nicer aliases for knowing when to use foo->a or foo->b
when inspecting foo->type; but it turned out to add too much
complexity, especially without a client.
There is a user-visible side effect to this change, but I
consider it to be an improvement. Previously,
the invalid QMP command:
{"execute":"blockdev-add", "arguments":{"options":
{"driver":"raw", "id":"a", "file":true}}}
failed with:
{"error": {"class": "GenericError",
"desc": "Invalid parameter type for 'file', expected: QDict"}}
(visit_get_next_type() succeeded, and the error comes from the
visit_type_BlockdevOptions() expecting {}; there is no mention of
the fact that a string would also work). Now it fails with:
{"error": {"class": "GenericError",
"desc": "Invalid parameter type for 'file', expected: BlockdevRef"}}
(the error when the next type doesn't match any expected types for
the overall alternate).
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1449033659-25497-5-git-send-email-eblake@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
What's more meta than using qapi to define qapi? :)
Convert QType into a full-fledged[*] builtin qapi enum type, so
that a subsequent patch can then use it as the discriminator
type of qapi alternate types. Fortunately, the judicious use of
'prefix' in the qapi definition avoids churn to the spelling of
the enum constants.
To avoid circular definitions, we have to flip the order of
inclusion between "qobject.h" vs. "qapi-types.h". Back in commit
28770e0, we had the latter include the former, so that we could
use 'QObject *' for our implementation of 'any'. But that usage
also works with only a forward declaration, whereas the
definition of QObject requires QType to be a complete type.
[*] The type has to be builtin, rather than declared in
qapi/common.json, because we want to use it for alternates even
when common.json is not included. But since it is the first
builtin enum type, we have to add special cases to qapi-types
and qapi-visit to only emit definitions once, even when two
qapi files are being compiled into the same binary (the way we
already handled builtin list types like 'intList'). We may
need to revisit how multiple qapi files share common types,
but that's a project for another day.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1449033659-25497-4-git-send-email-eblake@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
The documentation claims that alternates are useful for
allowing two or more types, although nothing enforces this.
Meanwhile, it is silent on whether empty unions are allowed.
In practice, the generated code will compile, in part because
we have a 'void *data' branch; but attempting to visit such a
type will cause an abort(). While there's no technical reason
that a degenerate union could not be made to work, it's harder
to justify the time spent in chasing known (the current
abort() during visit) and unknown corner cases, than it would
be to just outlaw them. A future patch will probably take the
approach of forbidding them; in the meantime, we can at least
add testsuite coverage to make it obvious where things stand.
In addition to adding tests to expose the problems, we also
need to adjust existing tests that are meant to test something
else, but which could fail for the wrong reason if we reject
degenerate alternates/unions.
Note that empty structs are explicitly supported (for example,
right now they are the only way to specify that one branch of a
flat union adds no additional members), and empty enums are
covered by the testsuite as working (even if they do not seem
to have much use).
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <1443565276-4535-8-git-send-email-eblake@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Eric Blake