if a pending inbound session did not complete a handshake after an unclean close from a previous session the
remote udp endpoint would remain stuck mapped as authed and thus any further attempts from the remote would
be silently dropped as it entered a stuck state in the state machine. this was happening as a small part
of the state machine was hidden in the implementation details of iwp, but instead should be in the super type
as it is logic exclusively outside the details which every dialect would have regardless of its details.
this commit will unmap the udp endpoint every time it needs to in the link layer state machine, independat of
the implementation details of the diact.
We trigger a pump immediately, but this is racey because we add to our
plaintext data in a worker thread; if the worker thread runs after the
pump then it ends up leaving plaintext to be handled, but there's no
wakeup until the next one.
This was the cause of seeing a random +1s and bunching added to ping
responses sometimes: it wasn't until the *next* ping goes through the
network that the plaintext queue gets processed, at which point it
flushes the old one and often the new one together.
The fix here gets rid of the map of sessions needing wakeups and instead
adds an atomic flag to all of them to let us figure out which ones
need to be flushed.
currently creating an outbound session will cancel if we have any session
at all with the relay. instead, only cancel if we have an outbound session
to that relay. this is useful for reachability testing.
All #ifndef guards on headers have been removed, I think,
in favor of #pragma once
Headers are now included as `#include "filename"` if the included file
resides in the same directory as the file including it, or any
subdirectory therein. Otherwise they are included as
`#include <project/top/dir/relative/path/filename>`
The above does not include system/os headers.
loop->call(...) is similar to the old logic->Call(...), but is smart
about the current thread: if called from within the event loop it simply
runs the argument directly, otherwise it queues it.
Similarly most of the other event loop calls are also now thread-aware:
for example, `call_later(...)` can queue the job directly when called if
in the event loop rather than having to double-queue through the even
loop (once to call, then inside the call to initiate the time).
- removes all the llarp_ev_* functions, replacing with methods/classes/functions in the llarp
namespace.
- banish ev/ev.h to the void
- Passes various things by const lvalue ref, especially shared_ptr's that don't need to be copied
(to avoid an atomic refcount increment/decrement).
- Add a llarp::UDPHandle abstract class for UDP handling
- Removes the UDP tick handler; code that needs tick can just do a separate handler on the event
loop outside the UDP socket.
- Adds an "OwnedBuffer" which owns its own memory but is implicitly convertible to a llarp_buffer_t.
This is mostly needed to take over ownership of buffers from uvw without copying them as,
currently, uvw does its own allocation (pending some open upstream issues/PRs).
- Logic:
- add `make_caller`/`call_forever`/`call_every` utility functions to abstract Call wrapping and
dependent timed tasks.
- Add inLogicThread() so that code can tell its inside the logic thread (typically for
debugging assertions).
- get rid of janky integer returns and dealing with cancellations on call_later: the other methods
added here and the event loop code remove the need for them.
- Event loop:
- redo everything with uvw instead of libuv
- rename EventLoopWakeup::Wakeup to EventLoopWakeup::Trigger to better reflect what it does.
- add EventLoopRepeater for repeated events, and replace the code that reschedules itself every
time it is called with a repeater.
- Split up `EventLoop::run()` into a non-virtual base method and abstract `run_loop()` methods;
the base method does a couple extra setup/teardown things that don't need to be in the derived class.
- udp_listen is replaced with ev->udp(...) which returns a new UDPHandle object rather that
needing gross C-style-but-not-actually-C-compatible structs.
- Remove unused register_poll_fd_(un)readable
- Use shared_ptr for EventLoopWakeup rather than returning a raw pointer; uvw lets us not have to
worry about having the event loop class maintain ownership of it.
- Add factory EventLoop::create() function to create a default (uvw-based) event loop (previously
this was one of the llarp_ev_blahblah unnamespaced functions).
- ev_libuv: this is mostly rewritten; all of the glue code/structs, in particular, are gone as
they are no longer needed with uvw.
- DNS:
- Rename DnsHandler to DnsInterceptor to better describe what it does (this is the code that
intercepts all DNS to the tun IP range for Android).
- endpoint:
- remove unused "isolated network" code
- remove distinct (but actually always the same) variables for router/endpoint logic objects
- llarp_buffer_t
- make constructors type-safe against being called with points to non-size-1 values
- tun packet reading:
- read all available packets off the device/file descriptor; previously we were reading one packet
at a time then returning to the event loop to poll again.
- ReadNextPacket() now returns a 0-size packet if the read would block (so that we can implement
the previous point).
- ReadNextPacket() now throws on I/O error
- Miscellaneous code cleanups/simplifications
- Previous android java and jni code updated to work, but with much love
still needed to make it work nicely, e.g. handling when the VPN is
turned off.
- DNS handling refactored to allow android to intercept and handle DNS
requests as we can't set the system DNS to use a high port
(and apparently Chrome ignores system DNS settings anyway)
- add packet router structure to allow separate handling of specific
intercepted traffic, e.g. UDP traffic to port 53 gets handled by our
DNS handler rather than being naively forwarded as exit traffic.
- For now, android lokinet is exit-only and hard-coded to use exit.loki
as its exit. The exit will be configurable before release, but
allowing to not use exit-only mode is more of a challenge.
- some old gitignore remnants which were matching to things we don't
want them to (and are no longer relevant) removed
- some minor changes to CI configuration
* partial tun code refactor
* take out the trash
* move vpn platform code into llarp/vpn/platform.cpp
* fix hive build
* fix win32
* fix memory leak on win32
* reduce cpu use
* make macos compile
* win32 patches:
* use wepoll for zmq
* use all cores on windows iocp read loop
* fix zmq patch for windows
* clean up cmake for win32
* add uninstall before reinstall option to win32 installer
* more ipv6 stuff
* make it compile
* fix up route poker
* remove an unneeded code block in macos wtf
* always use call to system
* fix route poker behavior on macos
* disable ipv6 on windows for now
* cpu perf improvement:
* colease calls to Router::PumpLL to 1 per event loop wakeup
* set up THEN add addresses
* emulate proactor event loop on win32
* remove excessively verbose error message
* fix issue #1499
* exclude uv_poll from win32 so that it can start up
* update logtag to include directory
* create minidump on windows if there was a crash
* make windows happy
* use dmp suffix on minidump files
* typo fix
* address feedback from jason
* use PROJECT_SOURCE_DIR instead of CMAKE_SOURCE_DIR
* quote $@ in apply-patches in case path has spaces in it
* address feedback from tom
* remove llarp/ev/pipe
* add comments for clairification
* make event loop queue size constant named
* Config file API/comment improvements
API improvements:
=================
Make the config API use position-independent tag parameters (Required,
Default{123}, MultiValue) rather than a sequence of bools with
overloads. For example, instead of:
conf.defineOption<int>("a", "b", false, true, 123, [] { ... });
you now write:
conf.defineOption<int>("a", "b", MultiValue, Default{123}, [] { ... });
The tags are:
- Required
- MultiValue
- Default{value}
plus new abilities (see below):
- Hidden
- RelayOnly
- ClientOnly
- Comment{"line1", "line2", "line3"}
Made option definition more powerful:
=====================================
- `Hidden` allows you to define an option that won't show up in the
generated config file if it isn't set.
- `RelayOnly`/`ClientOnly` sets up an option that is only accepted and
only shows up for relay or client configs. (If neither is specified
the option shows up in both modes).
- `Comment{...}` lets the option comments be specified as part of the
defineOption.
Comment improvements
====================
- Rewrote comments for various options to expand on details.
- Inlined all the comments with the option definitions.
- Several options that were missing comments got comments added.
- Made various options for deprecated and or internal options hidden by
default so that they don't show up in a default config file.
- show the section comment (but not option comments) *after* the
[section] tag instead of before it as it makes more sense that way
(particularly for the [bind] section which has a new long comment to
describe how it works).
Disable profiling by default
============================
We had this weird state where we use and store profiling by default but
never *load* it when starting up. This commit makes us just not use
profiling at all unless explicitly enabled.
Other misc changes:
===================
- change default worker threads to 0 (= num cpus) instead of 1, and fix
it to allow 0.
- Actually apply worker-threads option
- fixed default data-dir value erroneously having quotes around it
- reordered ifname/ifaddr/mapaddr (was previously mapaddr/ifaddr/ifname)
as mapaddr is a sort of specialization of ifaddr and so makes more
sense to come after it (particularly because it now references ifaddr
in its help message).
- removed peer-stats option (since we always require it for relays and
never use it for clients)
- removed router profiles filename option (this doesn't need to be
configurable)
- removed defunct `service-node-seed` option
- Change default logging output file to "" (which means stdout), and
also made "-" work for stdout.
* Router hive compilation fixes
* Comments for SNApp SRV settings in ini file
* Add extra blank line after section comments
* Better deprecated option handling
Allow {client,relay}-only options in {relay,client} configs to be
specified as implicitly deprecated options: they warn, and don't set
anything.
Add an explicit `Deprecated` tag and move deprecated option handling
into definition.cpp.
* Move backwards compat options into section definitions
Keep the "addBackwardsCompatibleConfigOptions" only for options in
sections that no longer exist.
* Fix INI parsing issues & C++17-ify
- don't allow inline comments because it seems they aren't allowed in
ini formats in general, and is going to cause problems if there is a
comment character in a value (e.g. an exit auth string). Additionally
it was breaking on a line such as:
# some comment; see?
because it was treating only `; see?` as the comment and then producing
an error message about the rest of the line being invalid.
- make section parsing stricter: the `[` and `]` have to be at the
beginning at end of the line now (after stripping whitespace).
- Move whitespace stripping to the top since everything in here does it.
- chop off string_view suffix/prefix rather than maintaining position
values
- fix potential infinite loop/segfault when given a line such as `]foo[`
* Make config parsing failure fatal
Load() LogError's and returns false on failure, so we weren't aborting
on config file errors.
* Formatting: allow `{}` for empty functions/structs
Instead of using two lines when empty:
{
}
* Make default dns bind 127.0.0.1 on non-Linux
* Don't show empty section; fix tests
We can conceivably have sections that only make sense for clients or
relays, and so want to completely omit that section if we have no
options for the type of config being generated.
Also fixes missing empty lines between tests.
Co-authored-by: Thomas Winget <tewinget@gmail.com>
* route poking:
* remove popen() call, replace with reading /proc/net/route for getting default route
* dynamically poke and unpoke routes on runtime
* swap intros and fix rpc endpoint for version to return what the ui expects
* use std::string::find_first_not_of instead of using a lambda
This replaces all use of std::optional's `opt.value()` with `*opt`
because macOS is great and the ghost of Steve Jobs says that actually
supporting std::optional's value() method is not for chumps before macOS
10.14. So don't use it because Apple is great.
Pretty much all of our use of it actually is done better with operator*
anyway (since operator* doesn't do a check that the optional has a
value).
Also replaced *most* of the `has_value()` calls with direct bool
context, except for one in the config section which looked really
confusing at a glance without a has_value().
This commit reflects changes to clang-format rules. Unfortunately,
these rule changes create a massive change to the codebase, which
causes an apparent rewrite of git history.
Git blame's --ignore-rev flag can be used to ignore this commit when
attempting to `git blame` some code.
These aren't needed: CMake already knows how to follow #includes and
rebuild when headers change as long as the headers are included
*somewhere*. The extra .cpp files here just require building a bunch of
.cpp files with just header content that we just end up throw away
during linking (since the same things will also be compiled in whatever
other compilation units include the same headers).
- util::Mutex is now a std::shared_timed_mutex, which is capable of
exclusive and shared locks.
- util::Lock is still present as a std::lock_guard<util::Mutex>.
- the locking annotations are preserved, but updated to the latest
supported by clang rather than using abseil's older/deprecated ones.
- ACQUIRE_LOCK macro is gone since we don't pass mutexes by pointer into
locks anymore (WTF abseil).
- ReleasableLock is gone. Instead there are now some llarp::util helper
methods to obtain unique and/or shared locks:
- `auto lock = util::unique_lock(mutex);` gets an RAII-but-also
unlockable object (std::unique_lock<T>, with T inferred from
`mutex`).
- `auto lock = util::shared_lock(mutex);` gets an RAII shared (i.e.
"reader") lock of the mutex.
- `auto lock = util::unique_locks(mutex1, mutex2, mutex3);` can be
used to atomically lock multiple mutexes at once (returning a
tuple of the locks).
This are templated on the mutex which makes them a bit more flexible
than using a concrete type: they can be used for any type of lockable
mutex, not only util::Mutex. (Some of the code here uses them for
getting locks around a std::mutex). Until C++17, using the RAII types
is painfully verbose:
```C++
// pre-C++17 - needing to figure out the mutex type here is annoying:
std::unique_lock<util::Mutex> lock(mutex);
// pre-C++17 and even more verbose (but at least the type isn't needed):
std::unique_lock<decltype(mutex)> lock(mutex);
// our compromise:
auto lock = util::unique_lock(mutex);
// C++17:
std::unique_lock lock(mutex);
```
All of these functions will also warn (under gcc or clang) if you
discard the return value. You can also do fancy things like
`auto l = util::unique_lock(mutex, std::adopt_lock)` (which lets a
lock take over an already-locked mutex).
- metrics code is gone, which also removes a big pile of code that was
only used by metrics:
- llarp::util::Scheduler
- llarp:🧵:TimerQueue
- llarp::util::Stopwatch
