This introduces a new way of linking Serval executables and dynamic
libraries from static libraries like libservald.a -- called
"feature-driven" linking.
The Makefile now links servald and serval-tests from libservald.a,
rather than from an explicit list of object (.o) files. Thanks to the
section-based method for registering functions such as HTTP handlers,
CLI commands and MDP handlers, these object files had become
"stand-alone" and hence were no longer included in the link because
there was no unresolved reference that required them to be linked in.
The new "feature.h" provides the DECLARE_FEATURE(name) macro that each
stand-alone source file uses to declare the named feature(s) it
provides. Each executable can call the USE_FEATURE(name) macro in any
of its explicitly-linked source files to cause the corresponding
object(s) to be included in the link, eg, servald_features.c.
The DEFINE_BINDING() macro has been extended so that every individual
MDP binding is given a feature name based on its port number macro, eg,
"mdp_binding_MDP_PORT_ECHO".
Some features have been factored into their own separate source files so
they can be omitted or included in a build independently of each other:
- the MDP bindings for MDP_PORT_DNALOOKUP, MDP_PORT_ECHO,
MDP_PORT_TRACE, MDP_PORT_KEYMAPREQUEST, MDP_PORT_RHIZOME_xxx,
MDP_PORT_PROBE, MDP_PORT_STUN, MDP_PORT_STUNREQ
- the CLI "log" and "echo" commands
- the CLI "rhizome direct" command
The JNI source files are only compiled if the <jni.h> header is present,
otherwise they are omitted from libservald.so.
This is required to support feature-driven linking, in which the exact
list of which functions to call in response to a configuration re-load
is not known until link time.
The "config_change_log" trigger is called before "config_change", so
that the logger can respond to the change before other subsystems.
Add the --branch option to specify which libsodium branch to checkout
and build.
Invoke libsodium's autogen.sh script if libsodium's 'configure' script
is not present (eg, on development branches).
The 'dst' parameter is now (char *) instead of (unsigned char *) because
in general it is used to produce human-readable (ASCII) strings, not
binary data.
The CLI and server main loop now have no conditional JNI code. All JNI
code has been moved into separate source files, which #include the new
"jni_common.h" instead of <jni.h>. The "cli.h" header no longer
includes <jni.h>, so the rest of the Serval source code is now
unaffected by JNI definitions.
The 'cf_limbo' global variable is now thread-local, so that each thread
has its own independent copy of the loaded configuration. The JNI
server entry point now calls cf_init() once. The new 'cf_initialised'
flag prevents clobbering the config state by redundant calls to
cf_init().
The CLI "stop" command now sends SIGHUP to the specific thread in which
the server is running. This is achieved by writing the PID and TID
(Linux Thread ID) into the pidfile, separated by a space, on systems
that support the Linux gettid() and tgkill() system calls. The server's
signal handler has been overhauled, and its logging improved.
Changed the built libraries and updated INSTALL.md accordingly:
- the 'servaldwrap' executable is now built
- libservald.[a,so] now contains the full daemon executable
- libservalclient.[a,so] contains the client library
- libserval.a is no longer built
Building the 'servaldwrap' executable is a step towards ensuring that
the libservald.so library can be linked and executed, without requiring
an Android build to reveal any failure.
Added the SOURCE_PREFIX variable that allows the make to be invoked from
within any current working directory, not just the serval-dna
repository's root directory.
Fixed the serval_version.o target to invoke the version_string.sh
script with the --repository=DIR option, so that the build will work
even if the current working directory is not within the serval-dna
repository.
Re-ordered and added some comments for clarity, and added some missing
.PHONY declarations.
The execute() and executeOk() functions now set $TFWEXECUTED instead
of $executed.
The execute() and executeOk() functions now make a hard link from the
--stdout-file=PATH and --stderr-file=PATH files to the default files
of 'stdout' and 'stderr' so that post-fork handling can still cat their
contents into the log file.
The fork() function now sets the variables $TFWFORKSTDOUT_<label>
$TFWFORKSTDERR_<label> $TFWFORKPID_<label>.
Ignores a variety of warnings when compiling sqlite3.c.
Uses the new sqlite3_trace_v2() function for logging SQL statements, now
that sqlite3_trace() and sqlite3_profile() are deprecated.
Fixes a rhizomeops test failure caused because SQLite 3.12 increased the
default block size from 1024 to 4096, causing the fixed overhead of an
empty Rhizome db to exceed 32KB. This caused the test to fail because
the first "file add" failed. Now the test fixture adapts to the fixed
overhead.
Deletes the SQLite 3.10.2 source amalgamation.
The recent randomising of keyring slots broke an assumption
about the order of the 'keyring list' output. Now the test's
assertions do not assume any order.
Add a new 'build-libsodium.sh' script and instructions in
doc/Development.md. Update the configure.in script to check for
libsodium headers and library, and print helpful messages if not
present.
The in-memory identity creation test was failing intermittently because
the server only creates the in-memory identity on the first tick, and
under load the server sometimes does not tick for a while, leading to a
race with the "servald id self" command.
The test case now waits for the server to emit a tick before running the
"id self" command. This necessitated a new DEBUG() statement and a new
debug flag 'overlaytick'.
Also disabled Rhizome for all the keyring tests, as unnecessary Rhizome
DB creation was slowing down the fixtures.
The quit-on-monitor-client-disconnect test was non-deterministic
depending on load (eg, other concurrently running tests). Under load,
it was likely that the server did not process the "monitor quit" command
before checking for disconnect, so the server did not terminate.
The fix was to make the monitor interface read and process all queued
input from the client before checking for HUP or ERR condition on the
socket. With this fix, the "sleep 1" kludges before and after the echo
"monitor quit" to the console command are no longer needed.
In the process the monitor interface code was modernised: eg, now it
calls read_nonblock() instead of read(2).
