Also re-word some error messages:
- "Key does not exist: %s"
- "Invalid channel: %<PRIu64>"
- "Request array size must be 4 (request) or 3 (notification)"
- "String cannot contain newlines"
References #6150
Fixes problem introduced by “api: Allow kObjectTypeNil to be zero without
breaking compatibility”: apparently there are clients which use metadata and
there are which aren’t. For the first that commit would not be needed, for the
second that commit misses this critical piece.
During testing found the following bugs:
1. msgpack-gen.lua script is completely unprepared for Float values either in
return type or in arguments. Specifically:
1. At the time of writing relevant code FLOAT_OBJ did not exist as well as
FLOATING_OBJ, but it would be used by msgpack-gen.lua should return type
be Float. I added FLOATING_OBJ macros later because did not know that
msgpack-gen.lua uses these _OBJ macros, otherwise it would be FLOAT_OBJ.
2. msgpack-gen.lua should use .data.floating in place of .data.float. But it
did not expect that .data subattribute may have name different from
lowercased type name.
2. vim_replace_termcodes returned its argument as-is if it receives an empty
string (as well as _vim_id*() functions did). But if something in returned
argument lives in an allocated memory such action will cause double free:
once when freeing arguments, then when freeing return value. It did not cause
problems yet because msgpack bindings return empty string as {NULL, 0} and
nothing was actually allocated.
3. New code in msgpack-gen.lua popped arguments in reversed order, making lua
bindings’ signatures be different from API ones.
Reasonings:
1. It is not used for anything, but scope dictionaries currenly. So there is no
need to generalize and split it into dict_set_var (which will contain some
scope-dictionary-specific checks) and dict_set_value (which will work for any
dictionary).
2. Check for key size is no longer valid for non-scope dictionaries: you *can*
use empty keys there. In scope dictionaries also, but you actually are not
supposed to store there anything, but variables.
Note that actually one may still do
let b:[''] = 1
and “bypass” check for variable name. It won’t change what `echo b:` will show,
but it may affect code which iterates over scope dictionary keys and sets them
to something (if there is such code).
v7.4.2024 changed a few function signatures of functions that we use in
Neovim-specific code, e.g. the API.
Due to that the commit for 7.4.2024 doesn't build on its own, only together with
this commit.
Also fixed dumping of partials by encode_vim_to_object and added code which is
able to work with partials and dictionaries to test/unit/eval/helpers.lua
(mostly copied from #5119, except for partials handling).
This makes gdb backtraces much more meaningful: specifically I now know at which
line it crashes in place of seeing that it crashes at
TYPVAL_ENCODE_DEFINE_CONV_FUNCTIONS macros invocation.
Problem: Leaking memory when there is a cycle involving a job and a
partial.
Solution: Add a copyID to job and channel. Set references in items referred
by them. Go through all jobs and channels to find unreferenced
items. Also, decrement reference counts when garbage collecting.
107e1eef1d
Since data.integer is a different (larger) integer type than
data.{buffer,window,tabpage}, we cannot abuse the union by using
data.integer to access the value for all 4 types. Instead, remove the
{buffer,window,tabpage} members and always use the integer member.
In order to accomodate this, perform distinct validation and coercion
between the Integer type and Buffer/Window/Tabpage types in
object_to_vim, msgpack_rpc helpers, and gendispatch.lua.
API level is disconnected from NVIM version. The API metadata holds the
current API level, and the lowest backwards-compatible level supported
by this instance.
Release 0.1.6 is the first release that reports the Nvim version and API
level.
metadata['version'] = {
major: 0,
minor: 1,
patch: 6,
api_level: 1,
api_compatible: 0,
api_prerelease: false,
}
The API level may remain unchanged across Nvim releases if the API has
not changed.
When changing the API,
- set NVIM_API_PRERELEASE to true
- increment NVIM_API_LEVEL (at most once per Nvim version)
- adjust NVIM_API_LEVEL_COMPAT if backwards-compatibility was broken
api_level_0.mpack was generated from Nvim 0.1.5 with:
nvim --api-info
The API level is disconnected from the NVIM version. The API metadata
holds the current API level, and the lowest backwards-compatible level
supported by this instance.
Release 0.1.6 will be the first release reporting the Nvim version and
API level.
metadata['version'] = {
major: 0,
minor: 1,
patch: 6,
prerelease: true,
api_level: 1,
api_compatible: 0,
}
The API level may remain unchanged across Neovim releases if the API has
not changed.
When changing the API the CMake variable NVIM_API_PRERELEASE is set to
true, and NVIM_API_CURRENT/NVIM_API_COMPATIBILITY are incremented
accordingly.
The functional tests check the API table against fixtures of past
versions of Neovim. It compares all the functions in the old table with
the new one, it does ignore some metadata attributes that do not alter
the function signature or were removed since 0.1.5. Currently the only
fixture is 0.mpack, generated from Neovim 0.1.5 with nvim --api-info.
If a conversion for a container fails in object_to_vim(), the memory for
the container in the returned/converted value is freed, but the returned
value keeps a pointer to the freed memory. Calling later clear_tv() on
this value leads to an invalid memory access.
Set v_type to VAR_UNKNOWN in the converted value on failure, so that
clear_tv() has no effect.
Also adds one exception to linter rules:
typedef struct {
kvec_t(Object) stack;
} EncodedData;
is completely valid (from the style guide point of view) code.
This ought to prevent stack overflow, but I do not see this actually working:
*lua* code crashes with stack overflow when trying to deserialize msgpack from
Neovim, Neovim is fine even if nesting level is increased 100x (though test
becomes very slow); not sure how recursive function may survive this. So it
looks like there are currently only two positive effects:
1. NULL lists are returned as empty (#4596).
2. Functional tests are slightly more fast. Very slightly. Checked for Release
build for test/functional/eval tests because benchmarking of debug mode is
not very useful.
