Change stream results to indicate cancellation #498
Conversation
| otherwise, completed successfully. In all cases, any number of elements (from | ||
| `0` to `n`) may have *first* been copied into or out of the buffer passed to | ||
| the `read` or `write` and so this number is packed into the `i32` result. |
There was a problem hiding this comment.
Would it be possible to provide a guarantee that if CANCELLED is returned then the number of elements transferred is always 0?
There was a problem hiding this comment.
I was thinking that it was indeed possible to receive CANCELLED and >0 elements. Hypothetically this could occur even with component-to-component streaming if a few partial writes succeed but then, when the reader tries to cancel, the writer is (in a hypothetical future with threads) on a different thread and the runtime is holding a lock while doing a large copy between buffers, and maybe we allow in this case the cancellation to return BLOCKED, and then it seems like we'd expect CANCELLED with >0 elements (as opposed to COMPLETED; the idea being that COMPLETED means "on its own", which is perhaps a distinction without a difference; I dunno?)
There was a problem hiding this comment.
Ok makes sense to me, and I don't think this should be too too hard to bind in various languages. There's going to be some subtelty to interpreting the return values but I don't think that'll be too bad.
I was sort of having second thoughts about this too so I think lining things up seems reasonable. As you say getting multi-value means that we can lift all these restrictions anyway |
Updating to WebAssembly/component-model#498, will require bytecodealliance/wit-bindgen#1273
As first discussed in #444 and resolving #490, this PR explicitly propagates whether an async stream/future read/write completed or was cancelled. The
Buffer.MAX_LENGTHis lowered to 228-1 to matchTable.MAX_LENGTHand leave a nice 4-bit space for any future additions.(Originally, I was going to put the 2-bit "result" field in the MSBs such that
Buffer.MAX_LENGTHcould be increased if we wanted in the future, but I noticed this is the reverse of thecanon lowerresult, which seemed gratuitously inconsistent. Also, I realized that once we change the CABI to take advantage of multi-return, this issue will go away, since we'll just return twoi32s. I'm not wed to this detail, though.)