As many of you know MusicDNS is an acoustic fingerprinting service and a software development kit provided by MusicIP. The fingerprinting client library that looks up and identifies audio files based on existing fingerprints is called libofa. MusicBrainz has a great audio tagger called Picard which can tag audio files by querying this MusicDNS service.
There is, however, a simple problem. Picard is a GUI and thus doesn’t allow batch tagging of audio files from command line.
Hence I decided to write my own tool for generating acoustic fingerprints and for querying MusicDNS service. I’ve chosen to use libsndfile to do the decoding as libofa expects raw audio data. libsndfile is a C library for reading and writing files containing sampled sound through one standard library interface. It’s pretty easy to use and its API hides most of the low-level details from the programmer.
The tool is named afprint, released under GPLv2. Following the UNIX philosophy it just does one thing, calculation of acoustic fingerprint and duration of the given audio file.
Usage is simple:
alip@harikalardiyari> afprint -h
afprint-0.1.0-7b17577 audio fingerprinting tool
Usage: afprint [-hVv0] <infile>
Options:
-h, --help Display usage and exit
-V, --version Display version and exit
-v, --verbose Be verbose
-0, --print0 Delimit path and fingerprint by null character instead of space
If <infile> is '-' afprint reads from standard input.
alip@harikalardiyari> afprint -v sample.ogg
[dump_print.294] Format: OGG (OGG Container format)
[dump_print.295] Frames: 2188368
[dump_print.296] Channels: 1
[dump_print.297] Samplerate: 44100Hz
[dump_print.298] Duration: 49735ms
[dump_print.302] essential frames: 5953500 > frames: 2188368, adjusting
sample.ogg 49735 ARaJDAgL...
afprint decodes the audio data using libsndfile and feeds it to libofa. It also calculates the duration of the audio file and prints them in format: FILENAME DURATION FINGERPRINT
Reading from standard input is tricky because pipes aren’t seekable thus it’s not possible to calculate the duration of the audio file. For this reason, when the audio data is fed via standard input, when <infile> is -, afprint saves this data into a temporary file and reads from it. This makes it possible to calculate acoustic fingerprints of Mp3 files, which libsndfile doesn’t support, easily.
alip@harikalardiyari> mpg123 -q --au - 01_san_francisco.mp3|afprint -v -
[wav.c:388] warning: Cannot rewind AU file. File-format isn't fully conform now.
[wav.c:388] warning: Cannot rewind AU file. File-format isn't fully conform now.
[dump_print.294] Format: AU (Sun/NeXT)
[dump_print.295] Frames: 8000111
[dump_print.296] Channels: 2
[dump_print.297] Samplerate: 44100Hz
[dump_print.298] Duration: 181820ms
/dev/stdin.au 181820 AQMZN...
Note the --au option passed to mpg123 as --wav doesn’t work.
So far so good, now we need a tool to query the MusicDNS server to find out the PUID of the audio file and query MusicBrainz to get the audio tags.
I’ve written a simple Perl script to do the job. The script, which has the name puidlookup, reads audio fingerprints from standard input and queries the MusicDNS server. Optionally it can query MusicBrainz as well to receive the tags.
Here are the requirements:
Usage is simple, just pipe afprint’s output to puidlookup.
alip@harikalardiyari> puidlookup -h
Usage: puidlookup [-hVv0]
-h, --help Display usage and exit
-V, --version Display version and exit
-v, --verbose Be verbose
-0, --null Expect input is null delimited
-m, --musicbrainz Look up PUIDs from MusicBrainz
(requires WebService-MusicBrainz)
-l, --limit Limit results to the given number
puidlookup reads filename, duration and audio fingerprint from standard input
The --null option responds to afprint’s --print0 option. These options are useful if filenames have spaces or other weird characters in it.
By default it only queries MusicDNS:
alip@harikalardiyari> afprint 04sheep.ogg | puidlookup
ARTIST='Pink Floyd'
TITLE='Sheep'
PUID=930806c1-e1e0-588a-b7de-2dacb1b8b11e
The --musicbrainz option can be used to query MusicBrainz:
alip@harikalardiyari> afprint 04sheep.ogg | puidlookup --musicbrainz
PUID=930806c1-e1e0-588a-b7de-2dacb1b8b11e
TRACKID=431a85dd-e22b-4626-91c9-c0abb8058d3f
ARTISTID=83d91898-7763-47d7-b03b-b92132375c47
ARTIST='Pink Floyd'
TITLE='Sheep'
TRACK=4
ALBUM='Animals'
The output is quoted so it’s safe to pass to eval, making it easy to integrate with shell scripts.
Last step is writing a tagger script to tag audio files. I’ve written a shell script called ofatag which uses envtag. It recognizes Mp3 files using the file command and decodes using mpg123, other formats are directly fed to afprint.
Now, to tag your files using MusicBrainz web services just do
ofatag /path/to/music/*.mp3 /path/to/music/*.ogg
etc.
I haven’t released a version yet because it’s all pretty new and needs testing. So please test it and report back! Any comments, thoughts, patches are appreciated.
sydbox-0.6.3 is released. ( tarball, sign, sha1sum )
- Resolve path of non-abstract UNIX sockets
- Intercept dup family calls and fcntl calls to see if a socket descriptor we care about has been duplicated
ptrace is a system call which is used for process tracing and debugging. This system call is available on many operating systems. However each operating system has different versions.
I want to explain about my efforts to port sydbox to FreeBSD. The ptrace implementation of FreeBSD is similar to Linux’. The request PT_SYSCALL is available to stop the traced process at every system call and exit similar to PTRACE_SYSCALL of Linux. In addition to that FreeBSD has the requests PT_TO_SCE and PT_TO_SCX which stops the traced process only at the beginning of system call entry or exit. This is a feature I really miss on Linux.
There is, however, a big difference, I’m inclined to call it a bug, about ptrace on FreeBSD. When a traced process is stopped at the entry of a system call, there’s no way to prevent the execution of this system call. On Linux this is done by changing the system call number to either something invalid like 0xbadca11 or something harmless like getpid.
Here is an example:
/* denying system calls using ptrace on Linux
*/
#include <assert.h>
#include <fcntl.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/reg.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <sys/ptrace.h>
#include <linux/ptrace.h>
#if defined(__x86__)
#define ORIG_ACCUM (4 * ORIG_EAX)
#elif defined(__x86_64__)
#define ORIG_ACCUM (8 * ORIG_RAX)
#else
#error unsupported architecture
#endif
int main(void)
{
int status;
pid_t pid;
if ((pid = fork()) < 0) {
perror("fork");
abort();
}
else if (pid == 0) {
ptrace(PTRACE_TRACEME, 0, NULL, NULL);
kill(getpid(), SIGSTOP);
open("foo.bar", O_WRONLY | O_CREAT);
_exit(0);
}
if (waitpid(pid, &status, 0) < 0) {
perror("waitpid");
abort();
}
assert(WIFSTOPPED(status));
assert(WSTOPSIG(status) == SIGSTOP);
if (ptrace(PTRACE_SYSCALL, pid, NULL, NULL) < 0) {
perror("ptrace(PTRACE_SYSCALL, ...)");
ptrace(PTRACE_KILL, pid, NULL, NULL);
abort();
}
if (waitpid(pid, &status, 0) < 0) {
perror("waitpid");
ptrace(PTRACE_KILL, pid, NULL, NULL);
abort();
}
assert(WIFSTOPPED(status));
assert(WSTOPSIG(status) == SIGTRAP);
/* Change the system call to something invalid, so it will be denied.
*/
if (ptrace(PTRACE_POKEUSER, pid, ORIG_ACCUM, 0xbadca11) < 0) {
perror("ptrace(PTRACE_POKEUSER, ...)");
ptrace(PTRACE_KILL, pid, NULL, NULL);
abort();
}
/* Let the process continue */
ptrace(PTRACE_CONT, pid, NULL, NULL);
waitpid(pid, &status, 0);
assert(WIFEXITED(status));
exit(WEXITSTATUS(status));
}
Now although the traced process calls open("foo.bar", O_WRONLY | O_CREAT) the file foo.bar won’t be created because the tracer process denies the system call.
Here is the same example for FreeBSD:
/* denying system calls using ptrace on FreeBSD
*/
#include <assert.h>
#include <fcntl.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <sys/ptrace.h>
#include <machine/reg.h>
int main(void)
{
int status;
pid_t pid;
struct reg r;
if ((pid = fork()) < 0) {
perror("fork");
abort();
}
else if (pid == 0) {
ptrace(PT_TRACE_ME, 0, NULL, 0);
kill(getpid(), SIGSTOP);
open("foo.bar", O_WRONLY | O_CREAT);
_exit(0);
}
if (waitpid(pid, &status, 0) < 0) {
perror("waitpid");
abort();
}
assert(WIFSTOPPED(status));
assert(WSTOPSIG(status) == SIGSTOP);
if (ptrace(PT_SYSCALL, pid, (caddr_t)1, 0) < 0) {
perror("ptrace(PT_SYSCALL, ...)");
ptrace(PT_KILL, pid, (caddr_t)1, 0);
abort();
}
if (waitpid(pid, &status, 0) < 0) {
perror("waitpid");
ptrace(PT_KILL, pid, (caddr_t)1, 0);
abort();
}
assert(WIFSTOPPED(status));
assert(WSTOPSIG(status) == SIGTRAP);
/* Change the system call to something invalid, so it will be denied.
*/
if (ptrace(PT_GETREGS, pid, (caddr_t)&r, 0) < 0) {
perror("ptrace(PT_GETREGS, ...)");
ptrace(PT_KILL, pid, (caddr_t)1, 0);
abort();
}
r.r_eax = 0xbadca11;
if (ptrace(PT_SETREGS, pid, (caddr_t)&r, 0) < 0) {
perror("ptrace(PT_SETREGS, ...)");
ptrace(PT_KILL, pid, (caddr_t)1, 0);
abort();
}
/* Let the process continue */
ptrace(PT_CONTINUE, pid, (caddr_t)1, 0);
exit(0);
}
We expect the same to happen here, the file foo.bar shouldn’t be created. But it’s created. Replace the PT_GETREGS and PT_SETREGS calls with a PT_KILL to terminate process with signal SIGKILL. The file will still be created! So there’s no way to deny a system call using ptrace which makes it impossible to port sydbox to FreeBSD without patching the kernel.
None of the other BSD’s, neither NetBSD nor DragonFlyBSD nor OpenBSD, has the ptrace request PT_SYSCALL so I haven’t checked if the behaviour is the same on these systems.
As many of you know sydbox can do network sandboxing but for some reasons we didn’t have it on by default on Exherbo.
For those who don’t know much about sydbox and network sandboxing let me explain it briefly. Network sandboxing has three modes:
- allow: All network connections are allowed.
- local: Only local network connections are allowed.
- deny: No network connections are allowed.
In addition to that there’s a restrict_connect option which disallows connects to all addresses except addresses that one of the parents has bind()‘ed to.
There’s also a network white list which specifies the additional network addresses that are allowed in local and deny modes.
On Exherbo we use the mode local with restrict_connect option enabled.
One limitation of sydbox was it couldn’t white list bind() addresses whose port were zero. The reason is obvious. The only place we can look up the actual port is /proc/net/tcp, or /proc/net/tcp6 for ipv6, and we need to do this before the bind() call has completed. The problem arises here. The /proc/net/tcp entry is only created after the bind() call has succeeded.
The solution isn’t entirely trivial. We have to note the file descriptor argument of bind() along with the socket family and socket address and intercept the subsequent listen() call. Only then we can look up the port argument from /proc/net/tcp.
The sydbox master has a simple implementation to solve this problem. If the port argument of a bind() call is zero, we save the file descriptor and the corresponding socket family and address to a GHashTable. After that the subsequent listen() call is intercepted and if the file descriptor of the listen() call matches a file descriptor in the hash table, sydbox looks up the port from /proc/net/tcp, fills it in and white lists the address.
With sydbox-0.4, which I’ll release after some testing, network sandboxing will be on by default again for the Paludis profile.
Just to be on the secure side ;)
alip@jaim.at
