Serialize method names in message encoding #5392
Conversation
|
@FUDCo Oh, I forgot to mention: we need to update |
Changes message encoding to eliminate the `method` field and move its contents into the `args` capdata object (now renamed `methargs`) so that the method name will be serialized, allowing non-string values (notably symbols) to be used as method selectors (and also allow remote function call, i.e., no method selector at all, to be supported). refs #2481
| resultPolicy = 'ignore', | ||
| slots = [], | ||
| ) { | ||
| function replacer(_, arg) { |
There was a problem hiding this comment.
Why introduce a replacer that special cases bigints and undefined but nothing else? What is special about these two?
| const kref = kernel.getRootObject(vatID); | ||
| const kpid = kernel.queueToKref(kref, method, args, resultPolicy); | ||
| const methargs = { | ||
| body: JSON.stringify([method, args], replacer), |
There was a problem hiding this comment.
Is this an additional layer of JSON encoding, or does it replace a JSON.stringify removed elsewhere?
There was a problem hiding this comment.
Previously callers of queueToVatRoot had to provide a pre-cooked capdata object (and of course creating a capdata requires a JSON.stringify call to produce the body string). queueToVatRoot is used by tests to inject messages into the kernel from the outside. This particular PR required updating a lot of tests, so one of the things I did was move some of the necessary quasi-serialization work that previously each test had to do by hand into queueToVatRoot in order to make the API less delicate. Ideally we would just use marshal and avoid a lot of the mess here, but unfortunately, we can't. This entry point has to weirdly straddle the world as seen from inside a vat and world as seen from inside the kernel -- it's used to send messages into a vat as if the caller was another vat sending the message, but the caller isn't in a vat; it's a piece of unit test code that in effect exists in kernel space. In particular, the stuff that would otherwise be done by convertValToSlot must instead be done by providing the slots array (containing raw kref strings) directly, because there is no clist for the sender. Tests can get away with this because they construct their worlds to be just so and the worlds they construct tend to be very simple.
The replacer you asked about is migrated from some of the tests which sent messages with arguments containing (in this case) bigints and undefineds. and this freed the callers from having to construct @qclass structures manually. This is another piece of making the API less delicate. Conceivably other @qclass things might need to be added in the future if tests require them, but I don't expect that.
Changes message encoding to eliminate the
methodfield and move its contents into theargscapdata object (now renamedmethargs) so that the method name will be serialized, allowing non-string values (notably symbols) to be used as method selectors (and also allows remote function call, i.e., no method selector at all, to be supported).Fixes #2481