cbc6c440e3 doc: add comments and release-notes for JSON-RPC 2.0 (Matthew Zipkin)
e7ee80dcf2 rpc: JSON-RPC 2.0 should not respond to "notifications" (Matthew Zipkin)
bf1a1f1662 rpc: Avoid returning HTTP errors for JSON-RPC 2.0 requests (Matthew Zipkin)
466b90562f rpc: Add "jsonrpc" field and drop null "result"/"error" fields (Matthew Zipkin)
2ca1460ae3 rpc: identify JSON-RPC 2.0 requests (Matthew Zipkin)
a64a2b77e0 rpc: refactor single/batch requests (Matthew Zipkin)
df6e3756d6 rpc: Avoid copies in JSONRPCReplyObj() (Matthew Zipkin)
09416f9ec4 test: cover JSONRPC 2.0 requests, batches, and notifications (Matthew Zipkin)
4202c170da test: refactor interface_rpc.py (Matthew Zipkin)
Pull request description:
Closes https://github.com/bitcoin/bitcoin/issues/2960
Bitcoin Core's JSONRPC server behaves with a special blend of 1.0, 1.1 and 2.0 behaviors. This introduces compliance issues with more strict clients. There are the major misbehaviors that I found:
- returning non-200 HTTP codes for RPC errors like "Method not found" (this is not a server error or an HTTP error)
- returning both `"error"` and `"result"` fields together in a response object.
- different error-handling behavior for single and batched RPC requests (batches contain errors in the response but single requests will actually throw HTTP errors)
https://github.com/bitcoin/bitcoin/pull/15495 added regression tests after a discussion in https://github.com/bitcoin/bitcoin/pull/15381 to kinda lock in our RPC behavior to preserve backwards compatibility.
https://github.com/bitcoin/bitcoin/pull/12435 was an attempt to allow strict 2.0 compliance behind a flag, but was abandoned.
The approach in this PR is not strict and preserves backwards compatibility in a familiar bitcoin-y way: all old behavior is preserved, but new rules are applied to clients that opt in. One of the rules in the [JSON RPC 2.0 spec](https://www.jsonrpc.org/specification#request_object) is that the kv pair `"jsonrpc": "2.0"` must be present in the request. Well, let's just use that to trigger strict 2.0 behavior! When that kv pair is included in a request object, the [response will adhere to strict JSON-RPC 2.0 rules](https://www.jsonrpc.org/specification#response_object), essentially:
- always return HTTP 200 "OK" unless there really is a server error or malformed request
- either return `"error"` OR `"result"` but never both
- same behavior for single and batch requests
If this is merged next steps can be:
- Refactor bitcoin-cli to always use strict 2.0
- Refactor the python test framework to always use strict 2.0 for everything
- Begin deprecation process for 1.0/1.1 behavior (?)
If we can one day remove the old 1.0/1.1 behavior we can clean up the rpc code quite a bit.
ACKs for top commit:
cbergqvist:
re ACK cbc6c440e3
ryanofsky:
Code review ACK cbc6c440e3. Just suggested changes since the last review: changing uncaught exception error code from PARSE_ERROR to MISC_ERROR, renaming a few things, and adding comments.
tdb3:
re ACK for cbc6c440e3
Tree-SHA512: 0b702ed32368b34b29ad570d090951a7aeb56e3b0f2baf745bd32fdc58ef68fee6b0b8fad901f1ca42573ed714b150303829cddad4a34ca7ad847350feeedb36
992c714451 common: Don't terminate on null character in UrlDecode (Fabian Jahr)
099fa57151 scripted-diff: Modernize name of urlDecode function and param (Fabian Jahr)
8f39aaae41 refactor: Remove hooking code for urlDecode (Fabian Jahr)
650d43ec15 refactor: Replace libevent use in urlDecode with our own code (Fabian Jahr)
46bc6c2aaa test: Add unit tests for urlDecode (Fabian Jahr)
Pull request description:
Fixes#29654 (as a side-effect)
Removing dependencies is a general goal of the project and the xz backdoor has been an additional wake up call recently. Libevent shows many of the same symptoms, few maintainers and slow releases. While libevent can not be removed completely over night we should start removing it’s usage where it's possible, ideally with the end goal to removing it completely.
This is a pretty easy win in that direction. The [`evhttp_uridecode` function from libevent](e0a4574ba2/http.c (L3542)) we were using in `urlDecode` could be easily emulated in fewer LOC. This also ports the [applicable test vectors over from libevent](https://github.com/libevent/libevent/blob/master/test/regress_http.c#L3430).
ACKs for top commit:
achow101:
ACK 992c714451
theStack:
Code-review ACK 992c714451
maflcko:
ACK 992c714451👈
stickies-v:
ACK 992c714451
Tree-SHA512: 78f76ae7ab3b6710eab2aaac20f55eb0da7803e057eaa6220e865f328666a5399ef1a479702aaf630b2f974ad3aa15e2b6adac9c11bc8c3d4be21e8af1667fea
The point of this was to be able to build bitcoin-tx and bitcoin-wallet without libevent, see #18504.
Now that we use our own implementation of urlDecode this is not needed anymore.
Co-authored-by: stickies-v <stickies-v@protonmail.com>
c3e632b441 Bugfix: bitcoin-cli: Check length of peer.transport_protocol_type (Luke Dashjr)
Pull request description:
"v" would dereference beyond the string length, and "v10" would show as '1'
Turn both of these cases into a blank, like anything else unexpected currently is.
ACKs for top commit:
sipa:
utACK c3e632b441.
hernanmarino:
utACK c3e632b441
alfonsoromanz:
ACK c3e632b441
achow101:
ACK c3e632b441
Tree-SHA512: f641e4412521adae7c8c8e1f268bdaaa223d9048d8286e3df4b13905faaa0d601155ce581cd649f760cab2acc4122356fa94a44714f1f190845552100105eda0
CLI -netinfo will currently break when calling it on a node that is running
pre-v26 bitcoind, as `getpeerinfo` doesn't yet return a transport_protocol_type
field.
Fix this by adding an `IsNull()` check as already done for other fields, and also:
- avoid checking for the full string "detecting", and instead do the cheaper
check for the most frequent case of the string starting with "v"
- drop displaying the "v" prefix in all the rows, as it doesn't add useful
information, and instead use "v" for the column header
- display nothing during peer setup, like for the -netinfo mping and ping columns
