Andre Noll [Sun, 13 Dec 2015 14:16:27 +0000 (15:16 +0100)]
Convert manual and NEWS from grutatxt to markdown.
Grutatext still works but the project seems to be dead, so we have
to switch to something else eventually.
This patch converts the two grutatext documents in the tree to the
markdown language, getting rid of the dependency on grutatext at
the expense of a new dependency on the markdown executable. Only the
discount markdown converter has been tested and is known to work.
The commit also changes the "Tools" section of the manual to list
discount instead of grutatext.
Andre Noll [Tue, 6 Oct 2015 02:32:31 +0000 (02:32 +0000)]
Constify struct filter access and introduce filter_get().
This modifies all users of the filter API to not access the filter
array directly. Instead all callers now obtain a const pointer to
the filter structure through the new filter_get() accessor function.
However, the filter array can not be made constant yet because the
->init methods modify the filter structure. This requires some casts
in filter_common.c unfortunately.
Andre Noll [Wed, 30 Sep 2015 01:09:24 +0000 (01:09 +0000)]
Fix documentation of check_receiver_arg().
Receiver name and options are NOT separated by a colon. Also, all three
receivers (http, udp, dccp) are always supported, so it's pointless
to talk about supported paraslash receivers. This patch simplifies
the comment accordingly.
Andre Noll [Wed, 9 Sep 2015 18:30:24 +0000 (20:30 +0200)]
afh_recv: Improve documentation of --begin-chunk.
The help text referred to --info, which was removed four years ago
in commit 06f33ca1. This patch also inlucdes a couple of minor other
improvements to the same paragraph.
Andre Noll [Sat, 4 Oct 2014 23:52:50 +0000 (23:52 +0000)]
afh: Make ->chunks_total and ->seconds_total fixed-size.
These members of struct afh_info are stored as 4-byte quantities in
the serialized afhi blob created by save_afhi(), so the structure
should declare them as uint32_t rather than unsigned long.
Fortunately, this bug is benign since save_afhi() uses the write_u32()
helper from portable_io.h which does the right thing, regardless of
the type of the variable passed.
Andre Noll [Sat, 5 Sep 2015 19:49:16 +0000 (21:49 +0200)]
aft.c: Prefer localtime() over localtime_r().
While localtime(3) is required to behave as though tzset(3) was
called, localtime_r(3) does not have this requirement, and portable
code should thus call tzset(3) before localtime_r(3). This is not
the case for get_local_time() of aft.c.
As we are single-threaded here, we don't need the thread-safe version,
so it is easiest to switch to localtime(3).
Andre Noll [Sun, 4 Oct 2015 20:25:06 +0000 (20:25 +0000)]
build: Fix audiod_command.c dependency.
The effect of these typos is that the dependencies are effectively
ignored. That is, changes to audiod_command.c do not cause a rebuild
of the command_list files.
Andre Noll [Sat, 3 Oct 2015 22:48:38 +0000 (22:48 +0000)]
play: Handle empty arguments to --key-map gracefully.
Executing para_play with an empty argument to --key-map results in
a read which starts one byte past the allocated buffer:
==24163== Invalid read of size 1
==24163== at 0x402A1DA: index (mc_replace_strmem.c:223)
==24163== by 0x804DA22: main (play.c:187)
==24163== Address 0x498e331 is 0 bytes after a block of size 1 alloc'd
This patch fixes the issue by rejecting empty arguments as invalid.
Andre Noll [Sat, 3 Oct 2015 17:51:38 +0000 (17:51 +0000)]
i9e: Avoid key binding macros.
The key binding code of interactive.c is quite an ugly hack which is
marked with a FIXME comment since forever. We bind each key sequence
given in the ->bound_keyseqs array of struct i9e_client_info to a
key code which is then mapped to a command handler provoded by the
application. The bindings started at key 64 with an arbitrary limit
of 32 possible mappings.
Besides being ugly, the problem with this mapping scheme is that
upper case keys are also in this range and are hence also mapped to
the specified commands.
This commit takes another approach. We now bind key sequences to a
function instead of a macro, passing ISFUNC as the first parameter to
libreadline's rl_generic_bind(). All key sequences are bound to the
same function, dispatch_key(), which calls the application-defined
key handler.
This is still not optimal because we need to look up the key sequence
again in dispatch_key(). But since this is not a fast path anyway,
it should be OK.
* refs/heads/t/command_handler_cleanups: (39 commits)
com_addatt(): Return negative on errors
com_rm(): In force mode, don't complain if no paths matched.
aft: Unify handling of hash and path duplicates.
afs: Provide pbout para_buffer for each callback.
afs: Make afs callbacks more flexible.
afs: Rename callback_function to afs_callback.
com_check(): Add attribute checking.
Let afs_event() return int.
playlist_check_callback(): Return negative on errors
mood_check_callback(): Return negative on errors
com_mvblob(): Return negative on errors
com_addblob(): Return negative on errors
com_rmblob(): Return negative on errors, cleanup
com_catblob(): Return negative on errors
com_lsblob(): Return negative on errors
rmatt: Cleanup callback.
com_rmatt(): Return negative on errors
com_mvatt(): Return negative on errors
com_lsatt(): Return negative on errors
com_init(): Return negative on errors
...
Andre Noll [Sat, 17 Oct 2015 14:47:01 +0000 (16:47 +0200)]
build: Fix m4 dependencies for BSD m4.
FreeBSD and NetBSD ship an m4 version which does not understand
the --debug=i option which we use to generate the .m4d dependency
files. This results in empty dependency files on those systems.
This was broken since m4 dependencies were introduced two years ago
in commit e249d105 (Generate ggo dependencies automatically). The
bug went unnoticed for so long because the problematic m4 command
is the first command of a pipeline (hence the exit code is lost),
and because the command redirects stderr, so the error message is
not printed to the terminal.
This commit fixes the bug by using the (portable) -s option of
m4 rather than --debug. The -s option outputs line synchronization
directives which include the file name, so it is equally well suited
for generating the dependency files.
Andre Noll [Tue, 9 Jun 2015 17:42:31 +0000 (19:42 +0200)]
web: Clear DOT_FONTNAME in Doxyfile.
Doxygen-1.8.6. warns about DOT_FONTNAME being set:
Warning: doxygen no longer ships with the FreeSans font.
You may want to clear or change DOT_FONTNAME.
Otherwise you run the risk that the wrong font is being used for dot generated graphs.
The warning is harmless because the DOT_FONTNAME option has no effect
in our setup because HAVE_DOT is set to "NO". Setting DOT_FONTNAME
to the empty string, as recommended in the explanatory text after
the warning, silences the warning.
Andre Noll [Tue, 9 Jun 2015 17:23:13 +0000 (19:23 +0200)]
web: Update Doxyfile to doxygen-1.8.6.
The system on which the public web pages are generated is now running
on Ubuntu-14.04 which ships doxygen-1.8.6. This version prints a
couple of warnings about obsolete tags in Doxyfile:
Warning: Tag `SYMBOL_CACHE_SIZE' at line 289 of file - has become obsolete.
To avoid this warning please remove this line from your configuration file or upgrade it using "doxygen -u"
Warning: Tag `SHOW_DIRECTORIES' at line 483 of file - has become obsolete.
To avoid this warning please remove this line from your configuration file or upgrade it using "doxygen -u"
Warning: Tag `HTML_ALIGN_MEMBERS' at line 820 of file - has become obsolete.
To avoid this warning please remove this line from your configuration file or upgrade it using "doxygen -u"
Warning: Tag `USE_INLINE_TREES' at line 989 of file - has become obsolete.
To avoid this warning please remove this line from your configuration file or upgrade it using "doxygen -u"
This patch is the result of running doxygen -u. With the patch applied,
the above warnings go away.
There is no need to be backwards compatible here because the web pages
are not created during a normal build.
Andre Noll [Wed, 12 Aug 2015 17:24:50 +0000 (19:24 +0200)]
error.h: Never call (para_)strerror() on osl errors.
In the (inlined) function para_strerror() of error.h we compile in
the call to osl_strerror() if and only if osl.h has been included
(in which case the preprocessor macro _OSL_H is defined). This is
wrong because an osl error code might well be passed to a function
defined in a compilation unit which does not include osl.h. If this
function calls para_strerror(), we will segfault or worse.
We need to check at link time whether osl_strerror is a defined
symbol rather than checking a preprocessor macro at compile time.
Fortunately, a little linker fu helps us out here. This patch
introduces a weak reference to osl_strerror to tell whether
osl_strerror is defined.
Andre Noll [Mon, 10 Aug 2015 18:00:49 +0000 (20:00 +0200)]
Clarify para_strerror().
para_strerror() needs to distinguish three kinds of errors: paraslash
errors, errors from the osl library, and system (libc) errors. This
is achieved through dedicated bits in the error code which are
set for errors from osl and libc function calls. These bits tell
para_strerror() which function to call in order to obtain the text
that corresponds to the error code.
If such a dedicated bit is set, para_strerror() first clears the bit,
then calls the library strerror() function that corresponds to the
bit. The code to clear the bit is
num & ((1 << OSL_ERROR_BIT) - 1))
and similar for libc errors. However, this expression clears *all*
high bits, not only bit number OSL_ERROR_BIT. This is not a problem
since the higher bits are not set under normal circumstances, but it
is better to fix this anyway. The new code is
Andre Noll [Sun, 27 Sep 2015 12:35:05 +0000 (12:35 +0000)]
Merge branch 'refs/heads/t/taggers'
Cooking in next since 2015-04-26.
* refs/heads/t/taggers:
aac: Fix compilation without libmp4v2.
The mp4 tagger.
The mp3 tagger.
The flac tagger.
The ogg/speex tagger.
The ogg/vorbis tagger.
The ogg/opus tagger.
The wma tagger.
This directory was used by a private script, but these days the script
creates its temporary files elsewhere. Hence it is no longer neccessary
to remove the directory in the receipe of the maintainer-clean target.
Andre Noll [Sat, 5 Sep 2015 11:01:18 +0000 (13:01 +0200)]
string.c: Kill E_STRTOLL.
This error code is unnecessary because para_atoi64() returns it only
if strtoll(3) did not perform any conversion and we have already a
more descriptive error code for this case: E_ATOI_NO_DIGITS.
Two new comments to para_atoi64() explain this in detail.
Andre Noll [Thu, 17 Sep 2015 18:07:02 +0000 (20:07 +0200)]
Merge branch 'refs/heads/t/com_ls_compat'
A single commit which adds support for the new syntax for the -l and -s options
of the ls command. These options should now be specified as -l=v rather than
-lv, for example. The old syntax still works, but support will be dropped in
v0.6.0.
Andre Noll [Mon, 14 Sep 2015 18:08:19 +0000 (20:08 +0200)]
build: Include flac afh only if libogg is available.
In configure.ac we have
AC_DEFUN([NEED_FLAC_OBJECTS], [{
test "$HAVE_OGG" = 'yes' -a "$HAVE_FLAC" = 'yes'
}])
In particular, we do not include the flac audio format handler on
systems where libflac is installed but libogg is not. In afh_common.c,
however, we only check for HAVE_FLAC. This results in a build failure
on those systems.
Fix this by modifying the check in afh_common.c to match the check
in configure.ac.
Andre Noll [Sun, 6 Sep 2015 13:05:30 +0000 (15:05 +0200)]
com_add(): Allow paths and afhi to exceed 64K.
The add command has the unnecessary limitation that for each audio
file to be added, the length of the path plus the size of serialized
audio format handler info structure (afhi) must not exceed 64K.
An afhi bigger than that is quite possible, given that it contains
the meta information of the file, for example an ID3v2 tag which can
contain almost unlimited amounts of data. If the 64K limit is exceeded,
all kinds of bad things may happen. So this issue definitely needs
to be addressed.
Specifically, save_add_callback_buffer() must be fixed. In this
function we serialize the data to be passed from the add command
handler to its callback. This includes path, afhi and the chunk table,
which are all variable in size. To cope with variable sizes, the
start of the buffer contains fields in which the offsets are stored
at which afhi and chunk table are located, relative to the start of
the serialized buffer. The problem is that the offset fields are only
16 bit wide, allowing for a maximal size of 64K.
This commit makes the offsets 32 bit wide and reorders them a
bit. Also, the CAB_AUDIO_FORMAT_OFFSET identifier is renamed to
CAB_AUDIO_FORMAT_ID_OFFSET.
Andre Noll [Wed, 1 Apr 2015 00:25:04 +0000 (00:25 +0000)]
server.h: Improve documentation of mmd struct.
This adds a sentence to the description of the structure which explains
how the misc_meta_data structure is used for communication betweeen
the server process and the command handlers. The patch also contains
some minor language improvements.
Andre Noll [Fri, 4 Sep 2015 07:24:12 +0000 (09:24 +0200)]
rmblob: Generate proper BLOB_REMOVE_EVENT.
com_rmblob_callback() of blob.c is responsible for removing blobs from
any of the four blob tables (images, lyrics, moods, playlists). After
a blob has been removed, the function generates an AFS event to tell
the other tables about the removal.
However, the call to afs_event() passes the event identifier
BLOB_RENAME rather than BLOB_REMOVE to the event handlers, which is
clearly incorrect. Fortunately, this does not matter because the only
event handler which cares about blob events is the one in mood.c,
and this handler treats the two events identically. Nevertheless,
it's is a bug that should be fixed.
This bug was introduced eight years ago when event handling was
introduced: commit 02d818d9 (Complete afs event handling).
Andre Noll [Wed, 26 Aug 2015 20:24:02 +0000 (22:24 +0200)]
Introduce new syntax for com_ls().
The ls command is the only command that takes one-letter arguments
which immediately follow the option. For example, -lv activates
verbose listing mode. If we ever want to support short option combining
(like in rm -rf) for paraslash commands, the one-letter arguments are
problematic because, in the above example, -lv can also be interpreted
as -l -v.
This commit encourages a different syntax for the -l and -s options
of the ls command while keeping the one-letter options working
for backwards compatibility. The new official way to enable verbose
listing mode is -l=v. This is analogous to how the "stat" and "touch"
commands expect their arguments.
The ls completer and the documentation are also updated according to
the new syntax.
Andre Noll [Mon, 31 Aug 2015 20:00:17 +0000 (22:00 +0200)]
com_rm(): In force mode, don't complain if no paths matched.
The documentation says that the rm command stays silent and exits
successfully if none of the given patterns matched any path of the
audio file table.
This was true until commit b02b7155 (com_rm(): Return negative on
errors), which effectively made -f a no-op by mistake. Since then
the rm command prints an error message and fails if there is no
match.
Andre Noll [Sat, 29 Aug 2015 13:10:25 +0000 (15:10 +0200)]
setatt: Do not abort if no pattern is given.
It is an error if no pattern is given to the setatt command.
For example,
setatt foo+ bar-
should result in a syntax error because no file name pattern is given.
The code in aft.c contains an assertion to detect this type of error,
but the afs proccess terminates if the condition of the assert
statement is not fulfilled:
Andre Noll [Fri, 28 Aug 2015 20:33:57 +0000 (22:33 +0200)]
lsatt: Fix sort order.
The -i option had the opposite effect of what the documentation
says. That is, lsatt -i sorted the attribute list by name while the
default was to sort by id. This patch reverts the logic in com_lsatt()
to let the implementation match the documentation.
Andre Noll [Sat, 22 Aug 2015 14:21:53 +0000 (16:21 +0200)]
server: Avoid segfault in com_sender().
If exactly one argument is given to the sender command, and this
argument is the name of an existing sender, the sender command
segfaults due to the NULL pointer dereference. The problem is an
off-by-one bug in the check for the number of arguments.
This patch makes sure we never dereference argv[2] if it is NULL.
Andre Noll [Wed, 1 Apr 2015 02:18:29 +0000 (02:18 +0000)]
oggdec: Add documentation of output size constants.
The OGGDEC_MAX_OUTPUT_SIZE and OGGDEC_OUTPUT_CHUNK_SIZE defines in
oggdec_filter.c are undocumented. This commit adds comments for both which
explain the meaning of the two sizes.
Andre Noll [Mon, 1 Jun 2015 07:36:28 +0000 (09:36 +0200)]
aft: Unify handling of hash and path duplicates.
These are symmetric but were not treated as such. Specifically:
* We ignored fatal errors for finding hash sisters while path
brothers were handled strictly.
* For hash sisters, we had a helper which implicitly special
cased the non-fatal E_OSL_RB_KEY_NOT_FOUND case while the
same check was open-coded for path brothers.
This adds the find_path_brother() helper which is analogous to
find_hash_sister() and fixes up the latter function to perform strict
error checking rather than relying on the implementation detail that
osl_get_row() sets the result pointer to NULL if no entry was found.
Andre Noll [Wed, 8 Apr 2015 03:35:21 +0000 (03:35 +0000)]
afs: Provide pbout para_buffer for each callback.
Most afs callbacks define a para_buffer to pass output from the
callback to the command handler. Hence the code which defines and
initializes the para_buffer is duplicated many times.
This commit gets rid of the duplication by moving the initialization
to the common call_callback(). The para_buffer becomes part of
struct afs_callback_arg, a pointer to which is passed to every
callback. The buffer is also flushed and freed in call_callback()
so that the callbacks don't need to care about it any more. This also
allows to make flush_and_free_pb() static since only a single caller
in afs.c remains.
This change simplifies the callbacks considerably. The callbacks of the
rm, setatt, lsatt, lsblob and touch commands don't even need their own
"action_data" structure any more since it was only necessary to pass
the para_buffer to the ->action method.
Andre Noll [Tue, 12 May 2015 07:44:42 +0000 (09:44 +0200)]
afs: Make afs callbacks more flexible.
Currently we pass the information for callbacks (an int and a pointer
to an osl_object) as separate arguments. If additional information
must be passed to some callbacks, every callback must be modified to
match the new prototype, even those which won't use the new argument.
This commit introduces struct afs_callback_arg which contains the two
callback arguments and changes all callbacks to receive a pointer
to such a structure. This is an equivalent transformation with no
visible change in semantics.
With this commit in place it is easy to provide additional information
by simply extending struct afs_callback as appropriate.
Andre Noll [Tue, 7 Apr 2015 21:18:34 +0000 (21:18 +0000)]
com_check(): Add attribute checking.
The afs info stored in the audio file table contains the attribute
bit mask of each audio file. If there is a bit set which does
not correspond to an attribute defined in the attribute table, we
have an inconsistency. This commit adds a check that reports such
inconsistencies.
com_check(), which is part of afs.c, calls the attribute check
callback of attribute.c via the callback mechanism. The callback
computes the logical or of all defined bits and passes this bit mask
to aft_check_attributes() of aft.c to check each audio file against
the mask. Hence two new public functions are required.
Andre Noll [Thu, 9 Apr 2015 12:51:49 +0000 (12:51 +0000)]
Let afs_event() return int.
It is usually a critical error if an afs event handler returns an
error. Currently we only print a log message an continue in this
case. The callers of afs_event() which trigger the event have no way
to tell that something went wrong, since this function returns void.
By returning int instead of void the callers can abort in the error
case. Most of the callers are afs callbacks. These can propagate the
error code to the command handler process, which will translate the
error code into a string and send it to the client. All callbacks
are changed in this way.
Andre Noll [Sun, 5 Apr 2015 22:22:04 +0000 (22:22 +0000)]
rmatt: Cleanup callback.
Get rid of struct remove_attribute_action_data, since only the para
buffer is necessary. This changes the code to pass a pointer to the
para buffer itself as the data pointer for ->action() of the pattern
matching loop.
Andre Noll [Tue, 13 Jan 2015 23:52:02 +0000 (00:52 +0100)]
command_util.bash: Kill make_proto().
It is always a bad idea to parse C code with a regex in a script. In
addition, it is completely unnecessary in this case.
This commit changes command_util.bash to define, in addition to the
old XXX_CMD_ARRAY, another preprocessor macro XXX__COMMAND_HANDLERS
containing the comma separated list of command handlers without
any type information instead of grepping the source files. A simple
typedef is used to declare all command handlers.
Avoiding all the grep/sed calls reduces the (warm cache) make dep
time by ~10%.
Andre Noll [Tue, 13 Jan 2015 22:53:28 +0000 (23:53 +0100)]
audiod: Clean up fd closing logic in command handlers.
audiod_command.c contains this comment:
/*
* command handlers don't close their fd on errors (ret < 0) so that
* its caller can send an error message. Otherwise (ret >= 0) it's up
* to each individual command to close the fd if necessary.
*/
This is a somewhat weird rule and this commit gets rid of it. Instead,
from now on the command handlers must not close their file descriptor
and handle_connect() closes it unconditionally.
The grab and stat commands need special treatment, which was the reason
for imposing the above rule. They need to keep the file descriptor open
to send the status items or the grabbed stream to the client. This
patch makes these two handlers create a copy of the descriptor with
dup(2). The new approach is simpler and less error-prone.
Andre Noll [Tue, 13 Jan 2015 23:00:41 +0000 (00:00 +0100)]
server: Remove command pointer from struct command_context.
Command handlers should not know about this implementation detail.
This commit also changes parse_sb_command() to not only parse the
command line but to actually run the command in case the caller has
sufficient permissions. The function is renamed to run_command()
to reflect this change.
We use the opportunity to clean up the allocation and freeing of
the command buffer and the argument vector. Both are now freed in
the same function they were allocated, which is considered good
programming practice.