Andre Noll [Wed, 8 Jan 2014 05:16:15 +0000 (05:16 +0000)]
task_register() conversion: stdin task
This renames stdin_set_defaults() to stdin_task_register() and changes
the function to register the stdin task. Before this patch, the task
was registered in the callers.
Andre Noll [Sat, 25 Jan 2014 18:41:45 +0000 (19:41 +0100)]
task_register() conversion: receivers
This adds a new public function, task_reap(), to sched.c. It is
called by para_audiod and para_play to free the memory occupied by
the receiver node after EOF. sched_shutdown() can not be used for
this purpose since the scheduler stays active during the life time of
these programs (i.e. schedule() never returns) while receiver nodes
come and go.
The new task_reap() has to face the problem that it is called
from another task's ->post_select() method, so removing the task
being reaped from the scheduler task list is not possible in
task_reap(). Hence this patch adds the new flag "dead" to struct
task. It is initially unset and is turned on in task_reap() to indicate
that (a) the task has exited (i.e. ->post_select() returned negative)
and (b) task_reap() has been called to fetch the exit status. Only
if this flag is set, the scheduler removes the task from the task list.
Andre Noll [Mon, 30 Dec 2013 19:27:04 +0000 (19:27 +0000)]
sched: Introduce alternative task API.
In the current implementation struct task is public so users of this
structure can mess with internal scheduler details as they please. This
has led to many bugs and questionable code in the past. This commit
is the first step to overcome this design mistake. At the end of this
patch series struct task can be made private to sched.c.
This commit introduces the following new public functions:
It also adds the new public task_info structure which carries the
information passed to the scheduler when a new task is registered. This
structure will stay public while struct task will become private once
all users have been converted.
task_register() is supposed to eventually replace register_task(). The
main difference of the two is that the new function returns a _pointer_
to a task structure which is allocated dynamically. Users are not
supposed to look at the fields of this pointer directly.
task_context() is a temporary helper which can be removed again at the
end of the series. Its sole purpose is to return the context pointer
which was passed at task register time as part of struct task_info.
The final new function, sched_shutdown(), deallocates the task
structures allocated during task_register() to cleanly shut down the
scheduler after all tasks have terminated.
All users need to be converted to the new API. This patch only converts
the stdout task though. The other tasks will be converted in subsequent
patches. The scheduler can tell if a task was registered using the
new API by means of the new ->owned_by_sched bit of struct task.
This boolean variable can also be removed after all tasks have been
converted.
Some users will need to query the exit status of a terminated
task. Hence we keep all tasks on the task list after ->post_select()
returned negative but call neither ->pre_select() nor ->post_select()
any more for such tasks. This leads to the concept of zombie tasks.
Andre Noll [Mon, 30 Dec 2013 17:37:41 +0000 (17:37 +0000)]
Simplify sched: Use only a single task list.
Currently the scheduler maintains two linked task lists, the pre_select
and the post_select list. This is completely unnecessary and bloats
the code for no good reason. This patch makes it use a single list
only and updates the documentation accordingly.
Andre Noll [Sat, 3 May 2014 17:56:01 +0000 (19:56 +0200)]
Merge branch 't/alsa_improvements'
Was cooking for almost two months.
* t/alsa_improvements:
alsa: Remove pointless initialization.
alsa_mix: Allow non-positive mixer values.
alsa: New writer option: --buffer-time.
alsa: Work around poll fd problems.
alsa: Set period time.
alsa: Also dump hardware params.
alsa: Improve help text of --device.
alsa: Improve documentation of struct private_alsa_write_data.
alsa: Don't set t->error in ->pre_select().
Andre Noll [Tue, 31 Dec 2013 15:44:24 +0000 (15:44 +0000)]
Remove client_disconnect().
Everything this function does can be done as well when the task
terminates by returning negative in ->post_select().
This also avoids to leak the stream cipher structure of the previous
connection when running in interactive mode.
1,032 bytes in 1 blocks are definitely lost in loss record 61 of 85
at 0x402994A: malloc (vg_replace_malloc.c:263)
by 0x804C689: para_malloc (string.c:71)
by 0x804FF87: sc_new (crypt.c:268)
by 0x804E204: client_post_select (client_common.c:370)
by 0x804D203: schedule (sched.c:59)
by 0x804A2E7: main (client.c:500)
1,032 bytes in 1 blocks are definitely lost in loss record 62 of 85
at 0x402994A: malloc (vg_replace_malloc.c:263)
by 0x804C689: para_malloc (string.c:71)
by 0x804FF87: sc_new (crypt.c:268)
by 0x804E217: client_post_select (client_common.c:371)
by 0x804D203: schedule (sched.c:59)
by 0x804A2E7: main (client.c:500)
Andre Noll [Tue, 31 Dec 2013 13:29:16 +0000 (13:29 +0000)]
Fix memory leak in btr_splice_out_node().
Commit 072391fc (Improve btr_splice_out_node(), 2013-03) modified this
function to take a pointer to a buffer tree node pointer so that the
node pointer can be invalidated after it is spliced out. However, this
also means the caller can no longer free its resources. Hence we must
free the btrn in btr_splice_out_node() in order to avoid memory leaks.
Andre Noll [Sun, 9 Mar 2014 14:21:48 +0000 (15:21 +0100)]
ao_write: Call ao_initialize() only once.
It is possible that another instance of the ao writer is active
when aow_close() is called. Calling ao_shutdown() in ao_close()
might disturb the other instance and lead to a segmentation fault.
Fix this by calling ao_initialize() only once during aow_init()
and never call ao_shutdown() any more.
Andre Noll [Sun, 9 Mar 2014 12:23:56 +0000 (13:23 +0100)]
ao_write: Simplify locking.
Currently we lock the mutex at the beginnint of each iteration
of the main loop of aow_play() and drop the lock before calling
ao_play(). On errors we leave the loop either with the mutex locked
locked or unlocked, depending on the error condition that caused us
to break out of the loop.
This is unnecessarily complex. This patch simplifies the locking
by always leave the loop with the mutex locked.
Andre Noll [Mon, 3 Feb 2014 20:06:00 +0000 (21:06 +0100)]
ao_write: Avoid segfault on exit.
If para_audiod is terminated by a signal it calls the ->close() method
of each active writer. In this method the ao writer calls ao_close()
and frees all resources but leaves the play thread active. So the
thread continues to run, prints
ao_alsa ERROR: write error: File descriptor in bad state
and probably crashes afterwards due to accessing freed memory. Fix this
by cancelling the thread in ->close() if it exists. Ironically enough,
this re-introduces a call to pthread_join(), but it's OK here since
we the cancelled thread should exit immediately (so we don't block),
and we are about to exit anyway.
Andre Noll [Mon, 3 Feb 2014 19:30:11 +0000 (20:30 +0100)]
ao_write: Avoid pthread_join().
The call to pthread_join() may block for hundreds of milliseconds,
depending of the amount of buffered data. This is not acceptable in
a post_select hook. Among other problems this leads to an incorrect time
display and fake "time jump" warnings.
This patch drops the call to pthread_join() and makes ->post_select()
return non-negative until the play thread has terminated. In order to
have a way to tell if the thread is still running we let the thread
(i.e. aow_play()) remove its own buffer tree node on exit instead
of doing this in ->post_select() after pthread_join(). Hence if
->thread_btrn is NULL, we know the thread has terminated. It would
be easier to use pthread_tryjoin_np() but this would not be portable,
as this function is a non-standard GNU extension.
Andre Noll [Fri, 4 Apr 2014 12:23:44 +0000 (14:23 +0200)]
audiod: Avoid segfault on exit.
When para_audiod terminates, it calls clean_exit() which closes
all slots unconditionally, regardless of whether they are in use.
This may lead to a segfault due to an invalid read:
==25253== Invalid read of size 4
==25253== at 0x40EBEA: close_slot (audiod.c:443)
==25253== by 0x40FCE8: clean_exit (audiod.c:1163)
==25253== by 0x40FD7D: signal_post_select (audiod.c:1014)
==25253== by 0x41490A: schedule (sched.c:65)
==25253== by 0x406BC3: main (audiod.c:1436)
==25253== Address 0x5c19d8 is not stack'd, malloc'd or (recently) free'd
This bug was introduced 9 month ago in commit eec23a69 (audiod:
Fix memory leak on exit: close slots). Fix it by calling
close_unused_slots(). This also simplifies clean_exit() a bit.
Andre Noll [Sun, 30 Mar 2014 16:31:57 +0000 (18:31 +0200)]
Merge branch 't/misc'
Various fixes, improvements, cleanups. Cooking since 2014-02-22.
* t/misc: (29 commits)
build: Don't link with -lreadline if readline was not found.
audiod: Skip NULL pointer check in compute_time_diff().
audiod: Make compute_time_diff() return void.
com_stat(): Remove pointless uptime variable.
gcrypt: Fix gcc warning on Ubuntu Lucid.
flac: Try to link also without -logg.
version.c: Fix comment of version_single_line().
doxygen: Expand all macros, in particular config.h.
recv_common.c: Improve documentation of check_receiver_arg().
audiod: get_time_string() comment fix.
configure: Really print opus audio file handler if opus lib was found.
Overhaul doxygen main page.
afs.h: Don't try to list all supported audio formats.
Change copyright year to 2014.
Add link to sideband.h in doxygen main page.
Doxify error2.c and add GPL header.
Add -Wdeclaration-after-statement.
Add some missing includes.
Makefile.real: Add clean2 to the list of phony targets.
mood.c: Fix a trivial whitespace issue.
...
Andre Noll [Sun, 23 Mar 2014 18:21:37 +0000 (19:21 +0100)]
Merge branch 't/fade_improvements'
Cooking since 2014-02-08.
* t/fade_improvements:
fade: Improve error diagnostics.
fade: Switch to the fade-in mood *before* sleeping.
fade: Allow to set more than one channel in sleep mode.
fade: Implement new mode "set".
fade: Log to stderr.
Andre Noll [Mon, 9 Sep 2013 03:53:32 +0000 (03:53 +0000)]
alsa: New writer option: --buffer-time.
This removes the old way of setting of private_alsa_write_data->buffer_time
Currently we set the buffer time to the minimum of the maximum
buffer time and 500ms. This results in different values for different
hardware. In view of the introduction of the sync filter it seems to
be desirable to have more control over this value.
This patch adds the --buffer-time config option to the alsa
writer which allows to specify a suitable value, the default being
170ms. The unit of private_alsa_write_data->buffer_time is changed
from microseconds to milliseconds as this avoids to divide by 1000
in ->pre_select().
Andre Noll [Thu, 13 Jun 2013 08:53:55 +0000 (10:53 +0200)]
alsa: Work around poll fd problems.
In its ->pre_select method the alsa writer adds the alsa poll file
descriptor to the set of fds to be monitored during the next select
call. When this fd is ready for I/O, select() returns and ->post_select
is called. Currently ->post_select checks if the alsa poll fd is set
and returns early if it is not.
The problem is that on some hardware/kernel/alsa-lib combinations,
the poll fd is never marked as ready for I/O. This happens at least
on the raspberry pi with linux-3.10.10 and alsa-lib-1.0.27. The result
is no audio at all.
We address this problem in two ways:
* In alsa_init() we remember the buffer duration for later use in
->pre_select where we ask the scheduler to impose half of the duration
as an upper bound on the select timeout.
* In ->post_select, if there is input data available, we don't
check any more whether the poll fd is ready but try to write
unconditionally. On EAGAIN, we read from the poll fd to discard any
pending events.
Andre Noll [Thu, 27 Feb 2014 18:10:55 +0000 (19:10 +0100)]
Merge branch 't/cmd_improvements'
Cooking since 2014-01-16.
* t/cmd_improvements:
afs.cmd: Strip redundant "sort" from ls help output.
audiod.cmd: Improve grab client help text.
afs.cmd: Improve second level indenting.
afd.cmd: Fix a trivial whitespace issue.
Andre Noll [Wed, 26 Feb 2014 18:16:03 +0000 (19:16 +0100)]
build: Check whether rl_free_keymap is declared.
Readline versions up to and including 6.2 miss to declare the public
symbol rl_free_keymap(). We currently work around this issue by
manually declaring the function in interactive.c.
The recently released readline-6.3, however, does declare that symbol.
Since we compile with -Wredundant-decls this leads to the following
warning:
interactive.c:239:6: warning: redundant redeclaration of 'rl_free_keymap' [-Wredundant-decls]
void rl_free_keymap(Keymap keymap);
This patch gets rid of the warning by adding a check to configure.ac
which detects whether the symbol is declared in the readline header
file. Only if it is uncdelared we declare it manually as before.
Andre Noll [Sat, 22 Feb 2014 16:29:10 +0000 (17:29 +0100)]
Merge branch 't/sync'
Cooking since 2014-01-14.
* t/sync:
The sync filter.
net: Let maksock() callers perform flowopt cleanup.
net: makesock_addrinfo(): Make socketfd local to the loop.
net: Let makesock() continue on setsockopt() failure.
net: Reduce indentation level in makesock_addrinfo().
net: Further simplify makesock_addrinfo().
net: Replace the double loop of lookup_address() by a single loop.
net: makesock(): Combine code for passive sockets.
net: Change makesock_addrinfo() to set given flowopts before SO_REUSEADDR.
net: Clarify code flow of makesock_addrinfo().
net: Remove unnecessary condition in makesock_addrinfo().
net: Remove pointless initialization in makesock_addrinfo().
net: Kill dead code in makesock_addrinfo().
net: Remove networking headers from para.h.
net: Improve error handling of makesock_addrinfo().
net: Split makesock(), part 2: Introduce makesock_addrinfo().
net: Split makesock(), part 1: Introduce lookup_address().
net: Simplify makesock().
net: Fix parse_url().
net: Correct \return text of parse_url().
Andre Noll [Mon, 27 Jan 2014 12:54:49 +0000 (13:54 +0100)]
flac: Try to link also without -logg.
If libflac was compiled without ogg support, the build system should
include flac even if libogg is not installed. Currently this does
not work because the autoconf test tries to link with -logg, so it
always fails if libogg is not available.
Andre Noll [Fri, 3 Jan 2014 05:59:22 +0000 (05:59 +0000)]
Overhaul doxygen main page.
It got way too large, and contained quite a few links to .c files
whose doxygen page is not very interesting. For example, the doxygen
page of the .c files of filters, receivers and writers contain only
the init function since all other functions are static.
This commit limits the number of links on the main page considerably.
The new page tries to focus on important data structures and APIs
and does not strive for completeness.
Andre Noll [Wed, 1 Jan 2014 00:00:34 +0000 (00:00 +0000)]
Change copyright year to 2014.
This year, we're really on time. The changes in this patch were
created by the following silly script:
files=$(git grep -l 'Copyright (C) [0-9]\{4\}\(-2013\)* Andre Noll')
sed --in-place= -e 's/Copyright (C) \([0-9]\{4\}\)-2013 Andre Noll/Copyright (C) \1-2014 Andre Noll/1' $files
sed --in-place= -e 's/Copyright (C) 2013 Andre Noll/Copyright (C) 2013-2014 Andre Noll/1' $files
Andre Noll [Sun, 29 Dec 2013 01:45:49 +0000 (01:45 +0000)]
Remove old changelog files.
This simplifies Makefile.real a bit, and of course it's still in the
git history in case somebody is really that interested in the ancient
history of this project.
Andre Noll [Fri, 17 Jan 2014 01:27:14 +0000 (02:27 +0100)]
test-lib: Make colors work on FreeBSD.
Some implementations of tput don't accept capnames (bold, setaf,
sgr0) but only the older termcap capability names (md, AF, me). This
patch teaches test-lib.sh to try both variants and succeed if either
variant works.