GitLab

This replaces problematic data pass in Lua with files saved on disk.
Original problem is that Lua in some cases like closures duplicates
environment and that means also all data. If this happens few times
(from measurements it seems like two or three times in code segment we
are dropping in this) then we can have few times bigger memory
requirements.

This also temporally disables test test-sys-pkgupdate-plan-unapproved as
it does not expect additional files. This is going to be solved in
following commit.

This was originaly introduced to allow local packages to be installed.
It was later replaced with localrepo and was marked as obsoleted for
long time.

Morpher is obsoleted for more than year and should not be used any more.
We are doing now huge changes with approaching Turris OS 4.0 and that is
best time to drop this.

This is huge change but it makes sense. Original name is now pretty
confusing and it should be clear that this tool is part of updater-ng.
This new name should signal that.
This new name also contains whole word "transaction" so it should
possibly be more clearer what it really does. This also adds comment
that this tool is dangerous to play with.

I originally was thinking that Lua is doing something nasty like
duplicating original file descriptor for stdout but it turns out that
exec was a culprit. For some reason a call to exec truncates (or at
least removes unfinished line) from stdout. I have no idea why it
behaves that way but adding flush just after callback (before exec)
solves that problem and allows us to correctly use print in callback.
The weird part is that the same thing does not happen with stderr but
non the less I also added flush for it.

This also removes possibly hack to run C/Lua code in subprocess without
providing command. It is overhead but it is more systematic to use
'true' as a command instead of possibly in code unexpected value NULL.

This changes timeout from seconds to milliseconds. This is consistent
with previous events design.
Previously used seconds had to be multiplied and milliseconds divided
respectively to hook to current code. This just makes more sense.

We want to do sometimes more than just edit environment for subprocess.
Defining full fledge callback that is called in subprocess is better
alternative. It allows even more changes to be done before execution of
command.

These functions are intended as a replacement for events set of
functions. We no longer need parallel script execution and also we should
get rid of embedded busybox.

This replaces way how we define state in pkgupdate. Now we have
predefined enum that we use to identify sate instead of just using
string.

We do change directory and it makes sense to use absolute path either
way so let's go for it.

I forgot to update lua_coverage2info script when I renamed modules. Also
one file name was renamed to its historical variant.

This file was left in tree from internal:// uri.

opkg-trans is backend tool that is usable for direct packages
manipulation. It has no use of model and board specification (it does
not care). This makes it that way.

This uri is no longer used anywhere now.

These suppressions are not required as we are either using openssl with
PURITY enabled or valgrind suppressions are system defined (Debian).

These new functions should be used for correct cleanup. Idea is that you
will register your cleanup function that would be called if fatal error
is detected and if no such error happens then you will call cleanup_run
to call and unregister given cleanup function anyway.

Since we are using new version of luacheck we can drop some of the old
definitions. Modules are now automatically detected by luacheck and
listing them out is no longer necessary.

This also moves some configuration variables from backend to separate
syscnf module.
Original design where hard coded in values were used as defaults is not
replaced with design where set function is required to be called before
anything is done. Commonly that is suppose to be done in C before we
start using other Lua modules.
Thanks to new options --model and --board we can now set target even
outside of the router environment without requiring /tmp tainting. It's
now also not possible to run updater's executables on non-openwrt
platform without specifying those two options.

There is also new upstream version of luacheck that does more stuff.
Because of that there are some fixes in code.

There is no reason why we should have our own version of check bundled
in. We can use system wide version instead.

This tests that even if won't do any changes after replan that we will
rin postupdate hooks.

Restricting accessible uris for script at first look looks like an very
good feature but in reality it cripples scripts ability and adds no
additional security because if we use https and signatures then there
should be no worry about script badness. And if it's then this is just
small garden wall anyone exploiting https and signatures could go
around just by using same https exploit again and again so this really
doesn't make any sense.

Yes this allows us to pass sub-script a invalid uri but it doesn't
matter if it won't use if and if it will then it will fail on its own.
So just do that check when we need it not always.

It makes little to no sense to validate extra arguments when they are
pushed to sub-script context as if they would be invalid the target
script would already failed. This is just useless check and would add
complexity as it makes table from any argument passed to it so this way
it discards uri.system_cas and uri.no_crl and creates bugged system.