A few minor issues around formatting collisions, nothing huge.
.gitignore fixes too
Conflicts:
source/common/base_dispatch_common.c
source/extensions/stdapi/server/stdapi.c
The work that was done a while back to fix up command dispatching allowed
inline commands to run so that the server could be told to shutdown. Those
commands that want the server to terminate (such as migrate and shutdown)
should have returned `FALSE` instead of `TRUE` to tell the server thread to
stop.
I have no idea why those values were incorrect, but it's my work so it's
definitely my fault. I will have to sick back and lick my wounds for a while.
I hate it when I'm stupid.
Thanks to Kevin Mitnick for the bug, and @todb-r7 for the investigating the
history.
Redmine: [FixRM #8696]
In previous work done during the command refactor the mechanism for overriding
commands changed such that it wasn't invoking commands if they were overriden
by an extension. This, it would appear, broke some stuff. Badly.
This commit fixes this issue by reinstating the way things were done before.
If a base command exists, it is always executed. If an extension also exists
which overrides this command then the base command result is ignored, the
extension command is executed, and the result of that command is returned.
POSIX was out of whack with Windows as a result of the changes made
around channels. The schedular in posix was very different, and this
commit brings it into line.
Other than the obvious issues, a non-obvious issue with the changes
was that the channel was being freed up on close prior to the thread
terminating. This doesn't appear to be an issue on Windows, but was
causing crashes on close in POSIX.
The changes go quite deep. This changeset requires a lot of testing.
This work contains a bunch of changes around command dispatching. The
goals for this bit of work were to:
* Provide the ability for commands to be executed on the same thread as
the server rather than always creating new threads and executing them on
those threads.
* Have the means for _special_ commands, such as `exit` and `migrate` to
shut down the server cleanly without having to rely on signalling across
threads or by doing brutal thread termination via shared global handles.
This should not only fix the dirty shutdown problem on Windows which
leaves tasks dangling (or based on the prior attempt at fixing, crashing
stuff as well), it should also help clean up the shutdown process in
POSIX.
These changes hit a very important part of Meterpreter and so should be
reviewed with intense scrutnity. I expect this PR to garner a log of
critique and most likely a number of changes before being included in the
main line.
The `PacketDispatcher` was modified to include a new function pointer
called an `inline_handler`. This new member indicates that there's a
handler which should be invoked inline. While this sits alongside the
existing `handler`, they are actually mutually exclusive. If an
`inline_handler` is specified then the `handler` is ignored and it is
assumed that the command is intended to be handled inline. The signature
of the inline handler is different to the standard handler, and this is
why a new function pointer was added rather than a simple flag. Addition of
this parameter meant that the basic command structure changed, and that
obviously affects all of the extensions and their respective commands.
This changeset therefore updates each of those command declarations so
that they use the new macros that hide this detail.
Other things to be mindful of:
* This version of the code reads the command's `method` prior to invoking
any other function, and after that the command itself is passed around to
the threaded or non-threaded routes for invocation. An extra thread
parameter was included as as result, and an overload for the
`thread_create` function was added which supported this new parameter.
This was named `thread_create3` because
`thread_create_with_another_paramter` sounded a bit crap.
* The migration code, which originally had a `thread_kill` and an event
wait once the new meterpreter session had been created, has been modified
to not do any waiting at all. Instead it finishes execution as fast as
possible and returns control to the server which should respond by
shutting down in a clean way.
* Originally the code always attempted to call a command handler in the
base command list and then, if found, would also call an "overload" in
the extension commands list. From the investigation that I did, it
appears that the overloaded methods did nothing in the base (they'd
early out during invocation). As a result, the new way of doing things
acts like a 'true' overload in that the extension commands are searched
first, and if one is found this is the command that is executed. Any
base commands with the same method name will not get executed. In the
case where there is no extension command found, the base command list is
then queried. If a command is found that command is instead invoked.
* The POSIX version still compiles cleanly, but I've never been able to
build a version that runs on my machines. I'm not sure if there's a
trick to getting POSIX builds to run, and if there is I don't know it.
Whoever scrutinises this build should make sure that the POSIX version
that they build can still run and (hopefully) exit cleanly.
I've added lots of documentation, but there's always room for improvement.
Hopefully this will fix the `*_tcp` side of Redmine 8438.
Bring on the feedback!
