]> code.delx.au - pulseaudio/commitdiff
port module-combine to new core
authorLennart Poettering <lennart@poettering.net>
Thu, 30 Aug 2007 22:58:12 +0000 (22:58 +0000)
committerLennart Poettering <lennart@poettering.net>
Thu, 30 Aug 2007 22:58:12 +0000 (22:58 +0000)
git-svn-id: file:///home/lennart/svn/public/pulseaudio/branches/lennart@1734 fefdeb5f-60dc-0310-8127-8f9354f1896f

src/Makefile.am
src/modules/module-combine.c

index 0bdc3f7926d04a7b1bb831255b32e208e6b4d55a..079610ab4b53121b476f96a39969902ff9be0a79 100644 (file)
@@ -930,8 +930,8 @@ modlibexec_LTLIBRARIES += \
                module-sine.la \
                module-native-protocol-tcp.la \
                module-native-protocol-fd.la \
-               module-esound-protocol-tcp.la
-#              module-combine.la \
+               module-esound-protocol-tcp.la \
+               module-combine.la
 #              module-tunnel-sink.la \
 #              module-tunnel-source.la \
 #              module-esound-sink.la
@@ -1192,9 +1192,9 @@ module_null_sink_la_LIBADD = $(AM_LIBADD) libpulsecore.la
 
 # Couplings
 
-#module_combine_la_SOURCES = modules/module-combine.c
-#module_combine_la_LDFLAGS = -module -avoid-version
-#module_combine_la_LIBADD = $(AM_LIBADD) libpulsecore.la
+module_combine_la_SOURCES = modules/module-combine.c
+module_combine_la_LDFLAGS = -module -avoid-version
+module_combine_la_LIBADD = $(AM_LIBADD) libpulsecore.la
 
 module_match_la_SOURCES = modules/module-match.c
 module_match_la_LDFLAGS = -module -avoid-version
index b6d718de2affe084c4929a2359d8404c07efbfac..c4fab457d7f6c501c768f635097c79707c1a8d0c 100644 (file)
 #include <config.h>
 #endif
 
-#include <assert.h>
 #include <stdio.h>
+#include <errno.h>
 
 #include <pulse/timeval.h>
 #include <pulse/xmalloc.h>
 
+#include <pulsecore/macro.h>
 #include <pulsecore/module.h>
 #include <pulsecore/llist.h>
 #include <pulsecore/sink.h>
 #include <pulsecore/core-util.h>
 #include <pulsecore/modargs.h>
 #include <pulsecore/namereg.h>
+#include <pulsecore/mutex.h>
+#include <pulsecore/thread.h>
+#include <pulsecore/thread-mq.h>
+#include <pulsecore/rtpoll.h>
+#include <pulsecore/rtclock.h>
+#include <pulsecore/core-error.h>
 
 #include "module-combine-symdef.h"
 
@@ -55,13 +62,12 @@ PA_MODULE_USAGE(
         "format=<sample format> "
         "channels=<number of channels> "
         "rate=<sample rate> "
-        "channel_map=<channel map> ")
+        "channel_map=<channel map>")
 
 #define DEFAULT_SINK_NAME "combined"
 #define MEMBLOCKQ_MAXLENGTH (1024*170)
-#define RENDER_SIZE (1024*10)
 
-#define DEFAULT_ADJUST_TIME 20
+#define DEFAULT_ADJUST_TIME 10
 
 static const char* const valid_modargs[] = {
     "sink_name",
@@ -78,95 +84,135 @@ static const char* const valid_modargs[] = {
 
 struct output {
     struct userdata *userdata;
+    pa_sink *sink;
     pa_sink_input *sink_input;
-    size_t counter;
+
+    pa_asyncmsgq *asyncmsgq;
+    pa_rtpoll_item *rtpoll_item;
+    
     pa_memblockq *memblockq;
+
     pa_usec_t total_latency;
+
     PA_LLIST_FIELDS(struct output);
 };
 
 struct userdata {
-    pa_module *module;
     pa_core *core;
+    pa_module *module;
     pa_sink *sink;
-    unsigned n_outputs;
+
+    pa_thread *thread;
+    pa_thread_mq thread_mq;
+    pa_rtpoll *rtpoll;
+
+    pa_mutex *mutex;
+    
     struct output *master;
-    pa_time_event *time_event;
-    uint32_t adjust_time;
 
-    PA_LLIST_HEAD(struct output, outputs);
+    pa_time_event *time_event; 
+    uint32_t adjust_time; 
+
+    int automatic;
+    size_t block_size;
+
+    struct timespec timestamp;
+
+    pa_hook_slot *sink_new_slot, *sink_unlink_slot, *sink_state_changed_slot;
+
+    pa_resample_method_t resample_method;
+
+    struct timespec adjust_timestamp;
+    
+    pa_idxset* outputs; /* managed in main context */
+
+    struct {
+        PA_LLIST_HEAD(struct output, outputs); /* managed in IO thread context */
+        struct output *master;
+    } thread_info;
 };
 
-static void output_free(struct output *o);
-static void clear_up(struct userdata *u);
+enum {
+    SINK_MESSAGE_DETACH = PA_SINK_MESSAGE_MAX,
+    SINK_MESSAGE_ATTACH,
+    SINK_MESSAGE_ADD_OUTPUT,
+    SINK_MESSAGE_REMOVE_OUTPUT
+};
 
-static void update_usage(struct userdata *u) {
-    pa_module_set_used(u->module, u->sink ? pa_sink_used_by(u->sink) : 0);
-}
+static void output_free(struct output *o);
+static int output_create_sink_input(struct userdata *u, struct output *o);
+static int update_master(struct userdata *u, struct output *o);
+static int pick_master(struct userdata *u);
 
 static void adjust_rates(struct userdata *u) {
     struct output *o;
     pa_usec_t max_sink_latency = 0, min_total_latency = (pa_usec_t) -1, target_latency;
     uint32_t base_rate;
-    assert(u && u->sink);
+    uint32_t idx;
 
-    for (o = u->outputs; o; o = o->next) {
-        uint32_t sink_latency = o->sink_input->sink ? pa_sink_get_latency(o->sink_input->sink) : 0;
+    pa_assert(u);
+    pa_sink_assert_ref(u->sink);
 
-        o->total_latency = sink_latency + pa_sink_input_get_latency(o->sink_input);
+    if (pa_idxset_size(u->outputs) <= 0)
+        return;
+
+    if (!PA_SINK_OPENED(pa_sink_get_state(u->sink)))
+        return;
+    
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx)) {
+        uint32_t sink_latency;
+
+        if (!o->sink_input || !PA_SINK_OPENED(pa_sink_get_state(o->sink)))
+            continue;
 
+        sink_latency = o->sink_input->sink ? pa_sink_get_latency(o->sink_input->sink) : 0;
+        o->total_latency = sink_latency + pa_sink_input_get_latency(o->sink_input);
+        
         if (sink_latency > max_sink_latency)
             max_sink_latency = sink_latency;
-
+        
         if (o->total_latency < min_total_latency)
             min_total_latency = o->total_latency;
     }
 
-    assert(min_total_latency != (pa_usec_t) -1);
-
+    if (min_total_latency == (pa_usec_t) -1)
+        return;
+        
     target_latency = max_sink_latency > min_total_latency ? max_sink_latency : min_total_latency;
-
+        
     pa_log_info("[%s] target latency is %0.0f usec.", u->sink->name, (float) target_latency);
-
+    pa_log_info("[%s] master is %s", u->sink->name, u->master->sink->description);
+        
     base_rate = u->sink->sample_spec.rate;
-
-    for (o = u->outputs; o; o = o->next) {
+        
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx)) {
         uint32_t r = base_rate;
 
+        if (!o->sink_input || !PA_SINK_OPENED(pa_sink_get_state(o->sink)))
+            continue;
+        
         if (o->total_latency < target_latency)
             r -= (uint32_t) (((((double) target_latency - o->total_latency))/u->adjust_time)*r/ 1000000);
         else if (o->total_latency > target_latency)
             r += (uint32_t) (((((double) o->total_latency - target_latency))/u->adjust_time)*r/ 1000000);
-
-        if (r < (uint32_t) (base_rate*0.9) || r > (uint32_t) (base_rate*1.1))
+        
+        if (r < (uint32_t) (base_rate*0.9) || r > (uint32_t) (base_rate*1.1)) {
             pa_log_warn("[%s] sample rates too different, not adjusting (%u vs. %u).", o->sink_input->name, base_rate, r);
-        else {
+            pa_sink_input_set_rate(o->sink_input, base_rate);
+        } else {
             pa_log_info("[%s] new rate is %u Hz; ratio is %0.3f; latency is %0.0f usec.", o->sink_input->name, r, (double) r / base_rate, (float) o->total_latency);
             pa_sink_input_set_rate(o->sink_input, r);
         }
     }
 }
 
-static void request_memblock(struct userdata *u) {
-    pa_memchunk chunk;
-    struct output *o;
-    assert(u && u->sink);
-
-    update_usage(u);
-
-    if (pa_sink_render(u->sink, RENDER_SIZE, &chunk) < 0)
-        return;
-
-    for (o = u->outputs; o; o = o->next)
-        pa_memblockq_push_align(o->memblockq, &chunk);
-
-    pa_memblock_unref(chunk.memblock);
-}
-
-static void time_callback(pa_mainloop_api*a, pa_time_event* e, PA_GCC_UNUSED const struct timeval *tv, void *userdata) {
+static void time_callback(pa_mainloop_api*a, pa_time_event* e, const struct timeval *tv, void *userdata) {
     struct userdata *u = userdata;
     struct timeval n;
-    assert(u && a && u->time_event == e);
+    
+    pa_assert(u);
+    pa_assert(a);
+    pa_assert(u->time_event == e);
 
     adjust_rates(u);
 
@@ -175,166 +221,735 @@ static void time_callback(pa_mainloop_api*a, pa_time_event* e, PA_GCC_UNUSED con
     u->sink->core->mainloop->time_restart(e, &n);
 }
 
+static void thread_func(void *userdata) {
+    struct userdata *u = userdata;
+
+    pa_assert(u);
+
+    pa_log_debug("Thread starting up");
+
+    pa_thread_mq_install(&u->thread_mq);
+    pa_rtpoll_install(u->rtpoll);
+
+    pa_rtclock_get(&u->timestamp);
+
+    /* This is only run when were are in NULL mode, to make sure that
+     * playback doesn't stop. In all other cases we hook our stuff
+     * into the master sink. */
+    
+    for (;;) {
+        int ret;
+
+        /* Render some data and drop it immediately */
+        if (u->sink->thread_info.state == PA_SINK_RUNNING) {
+            struct timespec now;
+            
+            pa_rtclock_get(&now);
+
+            if (pa_timespec_cmp(&u->timestamp, &now) <= 0) {
+                pa_sink_skip(u->sink, u->block_size);
+                pa_timespec_add(&u->timestamp, pa_bytes_to_usec(u->block_size, &u->sink->sample_spec));
+            }
+
+            pa_rtpoll_set_timer_absolute(u->rtpoll, &u->timestamp);
+        } else
+            pa_rtpoll_set_timer_disabled(u->rtpoll);
+
+        /* Now give the sink inputs some to time to process their data */
+        if ((ret = pa_sink_process_inputs(u->sink)) < 0)
+            goto fail;
+        if (ret > 0)
+            continue;
+
+        /* Check whether there is a message for us to process */
+        if ((ret = pa_thread_mq_process(&u->thread_mq) < 0))
+            goto finish;
+        if (ret > 0)
+            continue;
+
+        /* Hmm, nothing to do. Let's sleep */
+        if (pa_rtpoll_run(u->rtpoll) < 0) {
+            pa_log("poll() failed: %s", pa_cstrerror(errno));
+            goto fail;
+        }
+    }
+
+fail:
+    /* We have to continue processing messages until we receive the
+     * SHUTDOWN message */
+    pa_asyncmsgq_post(u->thread_mq.outq, PA_MSGOBJECT(u->core), PA_CORE_MESSAGE_UNLOAD_MODULE, u->module, 0, NULL, NULL);
+    pa_asyncmsgq_wait_for(u->thread_mq.inq, PA_MESSAGE_SHUTDOWN);
+
+finish:
+    pa_log_debug("Thread shutting down");
+}
+
+static void request_memblock(struct output *o) {
+    pa_memchunk chunk;
+    
+    pa_assert(o);
+    pa_sink_input_assert_ref(o->sink_input);
+    pa_sink_assert_ref(o->userdata->sink);
+    
+    /* If another thread already prepared some data we received
+     * the data over the asyncmsgq, hence let's first process
+     * it. */
+    while (pa_asyncmsgq_get(o->asyncmsgq, NULL, NULL, NULL, NULL, &chunk, 0) == 0) {
+        pa_memblockq_push_align(o->memblockq, &chunk);
+        pa_asyncmsgq_done(o->asyncmsgq, 0);
+    }
+    
+    /* Check whether we're now readable */
+    if (pa_memblockq_is_readable(o->memblockq))
+        return;
+    
+    /* OK, we need to prepare new data */
+    pa_mutex_lock(o->userdata->mutex);
+
+    if (PA_SINK_OPENED(o->userdata->sink->thread_info.state)) {
+    
+        /* Maybe there's some data now? */
+        while (pa_asyncmsgq_get(o->asyncmsgq, NULL, NULL, NULL, NULL, &chunk, 0) == 0) {
+            pa_memblockq_push_align(o->memblockq, &chunk);
+            pa_asyncmsgq_done(o->asyncmsgq, 0);
+        }
+        
+        /* Ok, now let's prepare some data if we really have to */
+        while (!pa_memblockq_is_readable(o->memblockq)) {
+            struct output *j;
+            
+            /* Do it! */
+            pa_sink_render(o->userdata->sink, o->userdata->block_size, &chunk);
+            
+            /* OK, let's send this data to the other threads */
+            for (j = o->userdata->thread_info.outputs; j; j = j->next)
+                if (j != o && j->sink_input)
+                    pa_asyncmsgq_post(j->asyncmsgq, NULL, 0, NULL, 0, &chunk, NULL);
+            
+            /* And push it into our own queue */
+            pa_memblockq_push_align(o->memblockq, &chunk);
+            pa_memblock_unref(chunk.memblock);
+        }
+    }
+    
+    pa_mutex_unlock(o->userdata->mutex);
+}
+
+/* Called from I/O trhead context */
 static int sink_input_peek_cb(pa_sink_input *i, pa_memchunk *chunk) {
-    struct output *o = i->userdata;
-    assert(i && o && o->sink_input && chunk);
+    struct output *o;
 
-    if (pa_memblockq_peek(o->memblockq, chunk) >= 0)
-        return 0;
+    pa_sink_input_assert_ref(i);
+    o = i->userdata;
+    pa_assert(o);
 
-    /* Try harder */
-    request_memblock(o->userdata);
+    /* If necessary, get some new data */
+    request_memblock(o);
 
-    return pa_memblockq_peek(o->memblockq, chunk);
+    return  pa_memblockq_peek(o->memblockq, chunk);
 }
 
-static void sink_input_drop_cb(pa_sink_input *i, const pa_memchunk *chunk, size_t length) {
-    struct output *o = i->userdata;
-    assert(i && o && o->sink_input && chunk && length);
+/* Called from I/O thread context */
+static void sink_input_drop_cb(pa_sink_input *i, size_t length) {
+    struct output *o;
+
+    pa_sink_input_assert_ref(i);
+    pa_assert(length > 0);
+    o = i->userdata;
+    pa_assert(o);
 
-    pa_memblockq_drop(o->memblockq, chunk, length);
-    o->counter += length;
+    pa_memblockq_drop(o->memblockq, length);
 }
 
+/* Called from I/O thread context */
+static int sink_input_process_cb(pa_sink_input *i) {
+    struct output *o;
+    pa_memchunk chunk;
+    int r = 0;
+    
+    pa_sink_input_assert_ref(i);
+    o = i->userdata;
+    pa_assert(o);
+
+    /* Move all data in the asyncmsgq into our memblockq */
+    
+    while (pa_asyncmsgq_get(o->asyncmsgq, NULL, NULL, NULL, NULL, &chunk, 0) == 0) {
+        if (PA_SINK_OPENED(i->sink->thread_info.state))
+            pa_memblockq_push_align(o->memblockq, &chunk);
+        pa_asyncmsgq_done(o->asyncmsgq, 0);
+    }
+
+    /* If the sink is suspended, flush our queue */
+    if (!PA_SINK_OPENED(i->sink->thread_info.state))
+        pa_memblockq_flush(o->memblockq);
+
+    if (o == o->userdata->thread_info.master) {
+        pa_mutex_lock(o->userdata->mutex);
+        r = pa_sink_process_inputs(o->userdata->sink);
+        pa_mutex_unlock(o->userdata->mutex);
+    }
+    
+    return r;
+}
+
+/* Called from I/O thread context */
+static void sink_input_attach_cb(pa_sink_input *i) {
+    struct output *o;
+
+    pa_sink_input_assert_ref(i);
+    o = i->userdata;
+    pa_assert(o);
+
+    pa_assert(!o->rtpoll_item);
+    o->rtpoll_item = pa_rtpoll_item_new_asyncmsgq(i->sink->rtpoll, o->asyncmsgq);
+}
+
+/* Called from I/O thread context */
+static void sink_input_detach_cb(pa_sink_input *i) {
+    struct output *o;
+
+    pa_sink_input_assert_ref(i);
+    o = i->userdata;
+    pa_assert(o);
+
+    pa_assert(o->rtpoll_item);
+    pa_rtpoll_item_free(o->rtpoll_item);
+    o->rtpoll_item = NULL;
+}
+
+/* Called from main context */
 static void sink_input_kill_cb(pa_sink_input *i) {
-    struct output *o = i->userdata;
-    assert(i && o && o->sink_input);
+    struct output *o;
+
+    pa_sink_input_assert_ref(i);
+    o = i->userdata;
+    pa_assert(o);
+
+    pa_sink_input_unlink(o->sink_input);
+    pa_sink_input_unref(o->sink_input);
+    o->sink_input = NULL;
+    
     pa_module_unload_request(o->userdata->module);
-    clear_up(o->userdata);
 }
 
-static pa_usec_t sink_input_get_latency_cb(pa_sink_input *i) {
-    struct output *o = i->userdata;
-    assert(i && o && o->sink_input);
+/* Called from thread context */
+static int sink_input_process_msg(pa_msgobject *obj, int code, void *data, int64_t offset, pa_memchunk *chunk) {
+    struct output *o = PA_SINK_INPUT(obj)->userdata;
+
+    switch (code) {
+        
+        case PA_SINK_INPUT_MESSAGE_GET_LATENCY: {
+             pa_usec_t *r = data;
+
+            *r = pa_bytes_to_usec(pa_memblockq_get_length(o->memblockq), &o->sink_input->sample_spec);
+
+            /* Fall through, the default handler will add in the extra
+             * latency added by the resampler */
+            break;
+        }
 
-    return pa_bytes_to_usec(pa_memblockq_get_length(o->memblockq), &i->sample_spec);
+    }
+    
+    return pa_sink_input_process_msg(obj, code, data, offset, chunk);
 }
 
-static pa_usec_t sink_get_latency_cb(pa_sink *s) {
-    struct userdata *u = s->userdata;
-    assert(s && u && u->sink && u->master);
+static int suspend(struct userdata *u) {
+    struct output *o;
+    uint32_t idx;
+    
+    pa_assert(u);
+
+    /* Let's suspend by unlinking all streams */
+    
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx)) {
+        pa_sink_input_unlink(o->sink_input);
+        pa_sink_input_unref(o->sink_input);
+        o->sink_input = NULL;
+    }
 
-    return
-        pa_sink_input_get_latency(u->master->sink_input) +
-        pa_sink_get_latency(u->master->sink_input->sink);
+    if (pick_master(u) < 0)
+        pa_module_unload_request(u->module);
+
+    pa_log_info("Device suspended...");
+    
+    return 0;
 }
 
-static void sink_notify(pa_sink *s) {
-    struct userdata *u;
+static int unsuspend(struct userdata *u) {
     struct output *o;
+    uint32_t idx;
+
+    pa_assert(u);
+    
+    /* Let's resume */
+    
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx)) {
+            
+        if (output_create_sink_input(u, o) < 0)
+            output_free(o);
+        else
+            pa_sink_input_put(o->sink_input);
+    }
+
+    if (pick_master(u) < 0)
+        pa_module_unload_request(u->module);
+    
+    pa_log_info("Resumed successfully...");
+    return 0;
+}
+
+static int sink_set_state(pa_sink *sink, pa_sink_state_t state) {
+    struct userdata *u;
+    
+    pa_sink_assert_ref(sink);
+    u = sink->userdata;
+    pa_assert(u);
+
+    /* Please note that in contrast to the ALSA modules we call
+     * suspend/unsuspend from main context here! */
+    
+    switch (state) {
+        case PA_SINK_SUSPENDED:
+            pa_assert(PA_SINK_OPENED(pa_sink_get_state(u->sink)));
+            
+            if (suspend(u) < 0)
+                return -1;
+
+            break;
+
+        case PA_SINK_IDLE:
+        case PA_SINK_RUNNING:
+
+            if (pa_sink_get_state(u->sink) == PA_SINK_SUSPENDED) {
+                if (unsuspend(u) < 0)
+                    return -1;
+            }
+                    
+            break;
+
+        case PA_SINK_UNLINKED:
+        case PA_SINK_INIT:
+            ;
+    }
+
+    return 0;
+}
+
+/* Called from thread context of the master */
+static int sink_process_msg(pa_msgobject *o, int code, void *data, int64_t offset, pa_memchunk *chunk) {
+    struct userdata *u = PA_SINK(o)->userdata;
+
+    switch (code) {
+        
+        case PA_SINK_MESSAGE_SET_STATE:
+
+            if ((pa_sink_state_t) PA_PTR_TO_UINT(data) == PA_SINK_RUNNING) {
+                /* Only useful when running in NULL mode, i.e. when no
+                 * master sink is attached */       
+                pa_rtclock_get(&u->timestamp);
+            }
+            
+            break;
+            
+        case PA_SINK_MESSAGE_GET_LATENCY: {
+            struct timespec now;
+
+            /* This code will only be called when running in NULL
+             * mode, i.e. when no master sink is attached. See
+             * sink_get_latency_cb() below */
+            pa_rtclock_get(&now);
+            
+            if (pa_timespec_cmp(&u->timestamp, &now) > 0)
+                *((pa_usec_t*) data) = 0;
+            else
+                *((pa_usec_t*) data) = pa_timespec_diff(&u->timestamp, &now);
+            break;
+        }
+
+        case SINK_MESSAGE_DETACH: {
+            pa_sink_input *i;
+            void *state = NULL;
+
+            /* We're detaching all our input streams artificially, so
+             * that we can driver our sink from a different sink */
+
+            while ((i = pa_hashmap_iterate(u->sink->thread_info.inputs, &state, NULL)))
+                if (i->detach)
+                    i->detach(i);
+
+            u->thread_info.master = NULL;
+            
+            break;
+        }
 
-    assert(s);
+        case SINK_MESSAGE_ATTACH: {
+            pa_sink_input *i;
+            void *state = NULL;
+
+            /* We're attached all our input streams artificially again */
+
+            while ((i = pa_hashmap_iterate(u->sink->thread_info.inputs, &state, NULL)))
+                if (i->attach)
+                    i->attach(i);
+
+            u->thread_info.master = data;
+            
+            break;
+        }
+
+        case SINK_MESSAGE_ADD_OUTPUT:
+            PA_LLIST_PREPEND(struct output, u->thread_info.outputs, (struct output*) data);
+            break;
+
+        case SINK_MESSAGE_REMOVE_OUTPUT:
+            PA_LLIST_REMOVE(struct output, u->thread_info.outputs, (struct output*) data);
+            break;
+    }
+    
+    return pa_sink_process_msg(o, code, data, offset, chunk);
+}
+
+/* Called from main context */
+static pa_usec_t sink_get_latency_cb(pa_sink *s) {
+    struct userdata *u;
+
+    pa_sink_assert_ref(s);
     u = s->userdata;
-    assert(u);
+    pa_assert(u);
+
+    if (u->master) {
+        /* If we have a master sink, we just return the latency of it
+         * and add our own buffering on top */
+
+        if (!u->master->sink_input)
+            return 0;
+        
+        return
+            pa_sink_input_get_latency(u->master->sink_input) +
+            pa_sink_get_latency(u->master->sink_input->sink);
+        
+    } else {
+        pa_usec_t usec;
+
+        /* We have no master, hence let's ask our own thread which
+         * implements the NULL sink */
+        
+        if (pa_asyncmsgq_send(s->asyncmsgq, PA_MSGOBJECT(s), PA_SINK_MESSAGE_GET_LATENCY, &usec, 0, NULL) < 0)
+            return 0;
+
+        return usec;
+    }
+}
+
+static void update_description(struct userdata *u) {
+    int first = 1;
+    char *t;
+    struct output *o;
+    uint32_t idx;
+    
+    pa_assert(u);
 
-    for (o = u->outputs; o; o = o->next)
-        pa_sink_notify(o->sink_input->sink);
+    if (pa_idxset_isempty(u->outputs)) {
+        pa_sink_set_description(u->sink, "Simultaneous output");
+        return;
+    }
+
+    t = pa_xstrdup("Simultaneous output to");
+    
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx)) {
+        char *e;
+        
+        if (first) {
+            e = pa_sprintf_malloc("%s %s", t, o->sink->description);
+            first = 0;
+        } else
+            e = pa_sprintf_malloc("%s, %s", t, o->sink->description);
+        
+        pa_xfree(t);
+        t = e;
+    }
+    
+    pa_sink_set_description(u->sink, t);
+    pa_xfree(t);
 }
 
-static struct output *output_new(struct userdata *u, pa_sink *sink, int resample_method) {
-    struct output *o = NULL;
-    char t[256];
-    pa_sink_input_new_data data;
+static int update_master(struct userdata *u, struct output *o) {
+    pa_assert(u);
 
-    assert(u && sink && u->sink);
+    /* Make sure everything is detached from the old thread before we move our stuff to a new thread */
+    if (u->sink && PA_SINK_LINKED(pa_sink_get_state(u->sink)))
+        pa_asyncmsgq_send(u->sink->asyncmsgq, PA_MSGOBJECT(u->sink), SINK_MESSAGE_DETACH, NULL, 0, NULL);
+    
+    if (o) {
+        /* If we have a master sink we run our own sink in its thread */
+
+        pa_assert(o->sink_input);
+        pa_assert(PA_SINK_OPENED(pa_sink_get_state(o->sink)));
+        
+        if (u->thread) {
+            /* If we previously were in NULL mode, let's kill the thread */
+            pa_asyncmsgq_send(u->thread_mq.inq, NULL, PA_MESSAGE_SHUTDOWN, NULL, 0, NULL);
+            pa_thread_free(u->thread);
+            u->thread = NULL;
+
+            pa_assert(u->rtpoll);
+            pa_rtpoll_free(u->rtpoll);
+            u->rtpoll = NULL;
+        }
 
-    o = pa_xmalloc(sizeof(struct output));
-    o->userdata = u;
+        pa_sink_set_asyncmsgq(u->sink, o->sink->asyncmsgq);
+        pa_sink_set_rtpoll(u->sink, o->sink->rtpoll);
+        u->master = o;
 
-    o->counter = 0;
-    o->memblockq = pa_memblockq_new(
-            0,
-            MEMBLOCKQ_MAXLENGTH,
-            MEMBLOCKQ_MAXLENGTH,
-            pa_frame_size(&u->sink->sample_spec),
-            1,
-            0,
-            NULL);
+        pa_log_info("Master sink is now '%s'", o->sink_input->sink->name);
 
-    pa_snprintf(t, sizeof(t), "Output stream #%u of sink %s", u->n_outputs+1, u->sink->name);
+    } else {
 
+        /* We have no master sink, let's create our own thread */
+        
+        pa_sink_set_asyncmsgq(u->sink, u->thread_mq.inq);
+        u->master = NULL;
+
+        if (!u->thread) {
+            pa_assert(!u->rtpoll);
+            
+            u->rtpoll = pa_rtpoll_new();
+            pa_rtpoll_item_new_asyncmsgq(u->rtpoll, u->thread_mq.inq);
+
+            pa_sink_set_rtpoll(u->sink, u->rtpoll);
+            
+            if (!(u->thread = pa_thread_new(thread_func, u))) {
+                pa_log("Failed to create thread.");
+                return -1;
+            }
+        }
+        
+        pa_log_info("No suitable master sink found, going to NULL mode\n");
+    }
+
+    /* Now attach everything again */
+    if (u->sink && PA_SINK_LINKED(pa_sink_get_state(u->sink)))
+        pa_asyncmsgq_send(u->sink->asyncmsgq, PA_MSGOBJECT(u->sink), SINK_MESSAGE_ATTACH, u->master, 0, NULL);
+
+    return 0;
+}
+
+static int pick_master(struct userdata *u) {
+    struct output *o;
+    uint32_t idx;
+    pa_assert(u);
+
+    if (u->master && u->master->sink_input && PA_SINK_OPENED(pa_sink_get_state(u->master->sink)))
+        return update_master(u, u->master);
+
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx))
+        if (o->sink_input && PA_SINK_OPENED(pa_sink_get_state(o->sink)))
+            return update_master(u, o);
+
+    return update_master(u, NULL);
+}
+
+static int output_create_sink_input(struct userdata *u, struct output *o) {
+    pa_sink_input_new_data data;
+    char *t;
+    
+    pa_assert(u);
+    pa_assert(!o->sink_input);
+
+    t = pa_sprintf_malloc("Simultaneous output on %s", o->sink->description);
+    
     pa_sink_input_new_data_init(&data);
-    data.sink = sink;
+    data.sink = o->sink;
     data.driver = __FILE__;
     data.name = t;
     pa_sink_input_new_data_set_sample_spec(&data, &u->sink->sample_spec);
     pa_sink_input_new_data_set_channel_map(&data, &u->sink->channel_map);
     data.module = u->module;
 
-    if (!(o->sink_input = pa_sink_input_new(u->core, &data, PA_SINK_INPUT_VARIABLE_RATE)))
-        goto fail;
+    o->sink_input = pa_sink_input_new(u->core, &data, PA_SINK_INPUT_VARIABLE_RATE|PA_SINK_INPUT_DONT_MOVE);
 
-    o->sink_input->get_latency = sink_input_get_latency_cb;
+    pa_xfree(t);
+
+    if (!o->sink_input)
+        return -1;
+    
+    o->sink_input->parent.process_msg = sink_input_process_msg;
     o->sink_input->peek = sink_input_peek_cb;
     o->sink_input->drop = sink_input_drop_cb;
+    o->sink_input->process = sink_input_process_cb;
+    o->sink_input->attach = sink_input_attach_cb;
+    o->sink_input->detach = sink_input_detach_cb;
     o->sink_input->kill = sink_input_kill_cb;
     o->sink_input->userdata = o;
+    
+    return 0;
+}
+
+static struct output *output_new(struct userdata *u, pa_sink *sink, int resample_method) {
+    struct output *o;
+
+    pa_assert(u);
+    pa_assert(sink);
+    pa_assert(u->sink);
+
+    o = pa_xnew(struct output, 1);
+    o->userdata = u;
+    o->asyncmsgq = pa_asyncmsgq_new(0);
+    o->rtpoll_item = NULL;
+    o->sink = sink;
+    o->sink_input = NULL;
+    o->memblockq = pa_memblockq_new(
+            0,
+            MEMBLOCKQ_MAXLENGTH,
+            MEMBLOCKQ_MAXLENGTH,
+            pa_frame_size(&u->sink->sample_spec),
+            1,
+            0,
+            NULL);
+
+
+    update_description(u);
+
+    pa_assert_se(pa_idxset_put(u->outputs, o, NULL) == 0);
+    if (u->sink && PA_SINK_LINKED(pa_sink_get_state(u->sink)))
+        pa_asyncmsgq_send(u->sink->asyncmsgq, PA_MSGOBJECT(u->sink), SINK_MESSAGE_ADD_OUTPUT, o, 0, NULL);
+    else
+        PA_LLIST_PREPEND(struct output, u->thread_info.outputs, o);
+
+    if (PA_SINK_OPENED(pa_sink_get_state(sink)))
+        if (output_create_sink_input(u, o) < 0)
+            goto fail;
 
-    PA_LLIST_PREPEND(struct output, u->outputs, o);
-    u->n_outputs++;
     return o;
 
 fail:
 
     if (o) {
         if (o->sink_input) {
-            pa_sink_input_disconnect(o->sink_input);
+            pa_sink_input_unlink(o->sink_input);
             pa_sink_input_unref(o->sink_input);
         }
 
         if (o->memblockq)
             pa_memblockq_free(o->memblockq);
 
+        if (o->asyncmsgq)
+            pa_asyncmsgq_unref(o->asyncmsgq);
+
         pa_xfree(o);
     }
 
     return NULL;
 }
 
-static void output_free(struct output *o) {
-    assert(o);
-    PA_LLIST_REMOVE(struct output, o->userdata->outputs, o);
-    o->userdata->n_outputs--;
-    pa_memblockq_free(o->memblockq);
-    pa_sink_input_disconnect(o->sink_input);
-    pa_sink_input_unref(o->sink_input);
-    pa_xfree(o);
+static pa_hook_result_t sink_new_hook_cb(pa_core *c, pa_sink *s, struct userdata* u) {
+    struct output *o;
+    
+    pa_core_assert_ref(c);
+    pa_sink_assert_ref(s);
+    pa_assert(u);
+    pa_assert(u->automatic);
+
+    if (!(s->flags & PA_SINK_HARDWARE) || s == u->sink)
+        return PA_HOOK_OK;
+
+    pa_log_info("Configuring new sink: %s", s->name);
+    
+    if (!(o = output_new(u, s, u->resample_method))) {
+        pa_log("Failed to create sink input on sink '%s'.", s->name);
+        return PA_HOOK_OK;
+    }
+
+    if (pick_master(u) < 0)
+        pa_module_unload_request(u->module);
+    
+    if (o->sink_input)
+        pa_sink_input_put(o->sink_input);
+        
+    return PA_HOOK_OK;
 }
 
-static void clear_up(struct userdata *u) {
+static pa_hook_result_t sink_unlink_hook_cb(pa_core *c, pa_sink *s, struct userdata* u) {
     struct output *o;
-    assert(u);
+    uint32_t idx;
+    
+    pa_assert(c);
+    pa_sink_assert_ref(s);
+    pa_assert(u);
+
+    if (s == u->sink)
+        return PA_HOOK_OK;
+
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx))
+        if (o->sink == s)
+            break;
+
+    if (!o)
+        return PA_HOOK_OK;
+
+    pa_log_info("Unconfiguring sink: %s", s->name);
+    
+    output_free(o);
+
+    if (pick_master(u) < 0)
+        pa_module_unload_request(u->module);
+    
+    return PA_HOOK_OK;
+}
 
-    if (u->time_event) {
-        u->core->mainloop->time_free(u->time_event);
-        u->time_event = NULL;
-    }
+static pa_hook_result_t sink_state_changed_hook_cb(pa_core *c, pa_sink *s, struct userdata* u) {
+    struct output *o;
+    uint32_t idx;
+    pa_sink_state_t state;
 
-    while ((o = u->outputs))
-        output_free(o);
+    if (s == u->sink)
+        return PA_HOOK_OK;
 
-    u->master = NULL;
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx))
+        if (o->sink == s)
+            break;
 
-    if (u->sink) {
-        pa_sink_disconnect(u->sink);
-        pa_sink_unref(u->sink);
-        u->sink = NULL;
+    if (!o)
+        return PA_HOOK_OK;
+
+    state = pa_sink_get_state(s);
+    
+    if (PA_SINK_OPENED(state) && !o->sink_input) {
+        output_create_sink_input(u, o);
+
+        if (pick_master(u) < 0)
+            pa_module_unload_request(u->module);
+
+        if (o->sink_input)
+            pa_sink_input_put(o->sink_input);
     }
+        
+    if (state == PA_SINK_SUSPENDED && o->sink_input) {
+        pa_sink_input_unlink(o->sink_input);
+        pa_sink_input_unref(o->sink_input);
+        o->sink_input = NULL;
+
+        pa_memblockq_flush(o->memblockq);
+
+        if (pick_master(u) < 0)
+            pa_module_unload_request(u->module);
+    }
+
+    return PA_HOOK_OK;
 }
 
-int pa__init(pa_core *c, pa_module*m) {
+int pa__init(pa_module*m) {
     struct userdata *u;
     pa_modargs *ma = NULL;
     const char *master_name, *slaves, *rm;
-    pa_sink *master_sink;
-    char *n = NULL;
-    const char*split_state;
-    struct timeval tv;
+    pa_sink *master_sink = NULL;
     int resample_method = -1;
     pa_sample_spec ss;
     pa_channel_map map;
+    struct output *o;
+    uint32_t idx;
 
-    assert(c && m);
+    pa_assert(m);
 
     if (!(ma = pa_modargs_new(m->argument, valid_modargs))) {
         pa_log("failed to parse module arguments");
@@ -349,115 +964,256 @@ int pa__init(pa_core *c, pa_module*m) {
     }
 
     u = pa_xnew(struct userdata, 1);
+    u->core = m->core;
+    u->module = m;
     m->userdata = u;
     u->sink = NULL;
-    u->n_outputs = 0;
-    u->master = NULL;
-    u->module = m;
-    u->core = c;
-    u->time_event = NULL;
-    u->adjust_time = DEFAULT_ADJUST_TIME;
-    PA_LLIST_HEAD_INIT(struct output, u->outputs);
-
+    u->thread_info.master = u->master = NULL;
+    u->time_event = NULL; 
+    u->adjust_time = DEFAULT_ADJUST_TIME; 
+    u->mutex = pa_mutex_new(0);
+    pa_thread_mq_init(&u->thread_mq, m->core->mainloop);
+    u->rtpoll = NULL;
+    u->thread = NULL;
+    PA_LLIST_HEAD_INIT(struct output, u->thread_info.outputs);
+    u->resample_method = resample_method;
+    u->outputs = pa_idxset_new(NULL, NULL);
+    pa_timespec_reset(&u->adjust_timestamp);
+    
     if (pa_modargs_get_value_u32(ma, "adjust_time", &u->adjust_time) < 0) {
-        pa_log("failed to parse adjust_time value");
+        pa_log("Failed to parse adjust_time value");
         goto fail;
     }
 
-    if (!(master_name = pa_modargs_get_value(ma, "master", NULL)) || !(slaves = pa_modargs_get_value(ma, "slaves", NULL))) {
-        pa_log("no master or slave sinks specified");
+    master_name = pa_modargs_get_value(ma, "master", NULL);
+    slaves = pa_modargs_get_value(ma, "slaves", NULL);
+    if (!master_name != !slaves) {
+        pa_log("No master or slave sinks specified");
         goto fail;
     }
 
-    if (!(master_sink = pa_namereg_get(c, master_name, PA_NAMEREG_SINK, 1))) {
-        pa_log("invalid master sink '%s'", master_name);
-        goto fail;
+    if (master_name) {
+        if (!(master_sink = pa_namereg_get(m->core, master_name, PA_NAMEREG_SINK, 1))) {
+            pa_log("Invalid master sink '%s'", master_name);
+            goto fail;
+        }
+        
+        ss = master_sink->sample_spec;
+        u->automatic = 0;
+    } else {
+        master_sink = NULL;
+        ss = m->core->default_sample_spec;
+        u->automatic = 1;
     }
 
-    ss = master_sink->sample_spec;
     if ((pa_modargs_get_sample_spec(ma, &ss) < 0)) {
-        pa_log("invalid sample specification.");
+        pa_log("Invalid sample specification.");
         goto fail;
     }
 
-    if (ss.channels == master_sink->sample_spec.channels)
+    if (master_sink && ss.channels == master_sink->sample_spec.channels)
         map = master_sink->channel_map;
     else
         pa_channel_map_init_auto(&map, ss.channels, PA_CHANNEL_MAP_DEFAULT);
 
     if ((pa_modargs_get_channel_map(ma, &map) < 0)) {
-        pa_log("invalid channel map.");
+        pa_log("Invalid channel map.");
         goto fail;
     }
-
+    
     if (ss.channels != map.channels) {
-        pa_log("channel map and sample specification don't match.");
+        pa_log("Channel map and sample specification don't match.");
         goto fail;
     }
 
-    if (!(u->sink = pa_sink_new(c, __FILE__, pa_modargs_get_value(ma, "sink_name", DEFAULT_SINK_NAME), 0, &ss, &map))) {
-        pa_log("failed to create sink");
+    if (!(u->sink = pa_sink_new(m->core, __FILE__, pa_modargs_get_value(ma, "sink_name", DEFAULT_SINK_NAME), 0, &ss, &map))) {
+        pa_log("Failed to create sink");
         goto fail;
     }
 
-    pa_sink_set_owner(u->sink, m);
-    pa_sink_set_description(u->sink, "Combined Sink");
+    u->sink->parent.process_msg = sink_process_msg;
     u->sink->get_latency = sink_get_latency_cb;
-    u->sink->notify = sink_notify;
+    u->sink->set_state = sink_set_state;
     u->sink->userdata = u;
 
-    if (!(u->master = output_new(u, master_sink, resample_method))) {
-        pa_log("failed to create master sink input on sink '%s'.", u->sink->name);
-        goto fail;
-    }
-
-    split_state = NULL;
-    while ((n = pa_split(slaves, ",", &split_state))) {
-        pa_sink *slave_sink;
-
-        if (!(slave_sink = pa_namereg_get(c, n, PA_NAMEREG_SINK, 1))) {
-            pa_log("invalid slave sink '%s'", n);
+    u->sink->flags = PA_SINK_CAN_SUSPEND|PA_SINK_LATENCY;
+    pa_sink_set_module(u->sink, m);
+    pa_sink_set_description(u->sink, "Simultaneous output");
+
+    u->block_size = pa_bytes_per_second(&ss) / 20; /* 50 ms */
+    if (u->block_size <= 0)
+        u->block_size = pa_frame_size(&ss);
+    
+    if (!u->automatic) {
+        const char*split_state;
+        char *n = NULL;
+        pa_assert(slaves);
+
+        /* The master and slaves have been specified manually */
+        
+        if (!(u->master = output_new(u, master_sink, resample_method))) {
+            pa_log("Failed to create master sink input on sink '%s'.", master_sink->name);
             goto fail;
         }
+    
+        split_state = NULL;
+        while ((n = pa_split(slaves, ",", &split_state))) {
+            pa_sink *slave_sink;
+            
+            if (!(slave_sink = pa_namereg_get(m->core, n, PA_NAMEREG_SINK, 1)) || slave_sink == u->sink) {
+                pa_log("Invalid slave sink '%s'", n);
+                pa_xfree(n);
+                goto fail;
+            }
+            
+            pa_xfree(n);
+            
+            if (!output_new(u, slave_sink, resample_method)) {
+                pa_log("Failed to create slave sink input on sink '%s'.", slave_sink->name);
+                goto fail;
+            }
+        }
 
-        pa_xfree(n);
+        if (pa_idxset_size(u->outputs) <= 1)
+            pa_log_warn("WARNING: No slave sinks specified.");
 
-        if (!output_new(u, slave_sink, resample_method)) {
-            pa_log("failed to create slave sink input on sink '%s'.", slave_sink->name);
-            goto fail;
+        u->sink_new_slot = NULL;
+        
+    } else {
+        pa_sink *s;
+
+        /* We're in automatic mode, we elect one hw sink to the master
+         * and attach all other hw sinks as slaves to it */
+
+        for (s = pa_idxset_first(m->core->sinks, &idx); s; s = pa_idxset_next(m->core->sinks, &idx)) {
+
+            if (!(s->flags & PA_SINK_HARDWARE) || s == u->sink)
+                continue;
+
+            if (!output_new(u, s, resample_method)) {
+                pa_log("Failed to create sink input on sink '%s'.", s->name);
+                goto fail;
+            }
         }
+
+        u->sink_new_slot = pa_hook_connect(&m->core->hooks[PA_CORE_HOOK_SINK_NEW_POST], (pa_hook_cb_t) sink_new_hook_cb, u);
     }
 
-    if (u->n_outputs <= 1)
-        pa_log_warn("WARNING: no slave sinks specified.");
+    u->sink_unlink_slot = pa_hook_connect(&m->core->hooks[PA_CORE_HOOK_SINK_UNLINK], (pa_hook_cb_t) sink_unlink_hook_cb, u);
+    u->sink_state_changed_slot = pa_hook_connect(&m->core->hooks[PA_CORE_HOOK_SINK_STATE_CHANGED], (pa_hook_cb_t) sink_state_changed_hook_cb, u);
+    
+    if (pick_master(u) < 0)
+        goto fail;
 
+    /* Activate the sink and the sink inputs */
+    pa_sink_put(u->sink);
+    for (o = pa_idxset_first(u->outputs, &idx); o; o = pa_idxset_next(u->outputs, &idx))
+        pa_sink_input_put(o->sink_input);
+    
     if (u->adjust_time > 0) {
+        struct timeval tv;
         pa_gettimeofday(&tv);
         tv.tv_sec += u->adjust_time;
-        u->time_event = c->mainloop->time_new(c->mainloop, &tv, time_callback, u);
+        u->time_event = m->core->mainloop->time_new(m->core->mainloop, &tv, time_callback, u);
     }
 
     pa_modargs_free(ma);
+    
     return 0;
 
 fail:
-    pa_xfree(n);
 
     if (ma)
         pa_modargs_free(ma);
 
-    pa__done(c, m);
+    pa__done(m);
+    
     return -1;
 }
 
-void pa__done(pa_core *c, pa_module*m) {
+static void output_free(struct output *o) {
+    pa_assert(o);
+
+    if (o->userdata) {
+        if (o->userdata->sink && PA_SINK_LINKED(pa_sink_get_state(o->userdata->sink)))
+            pa_asyncmsgq_send(o->userdata->sink->asyncmsgq, PA_MSGOBJECT(o->userdata->sink), SINK_MESSAGE_REMOVE_OUTPUT, o, 0, NULL);
+        else
+            PA_LLIST_REMOVE(struct output, o->userdata->thread_info.outputs, o);
+    }
+
+    pa_assert_se(pa_idxset_remove_by_data(o->userdata->outputs, o, NULL));
+
+    if (o->userdata->master == o) {
+        /* Make sure the master points to a different output */
+        o->userdata->master = NULL;
+        pick_master(o->userdata);
+    }
+    
+    update_description(o->userdata);
+    
+    if (o->sink_input) {
+        pa_sink_input_unlink(o->sink_input);
+        pa_sink_input_unref(o->sink_input);
+    }
+
+    if (o->rtpoll_item)
+        pa_rtpoll_item_free(o->rtpoll_item);
+
+    if (o->memblockq)
+        pa_memblockq_free(o->memblockq);
+
+    if (o->asyncmsgq)
+        pa_asyncmsgq_unref(o->asyncmsgq);
+    
+    pa_xfree(o);
+}
+
+void pa__done(pa_module*m) {
     struct userdata *u;
-    assert(c && m);
+    struct output *o;
+    
+    pa_assert(m);
 
     if (!(u = m->userdata))
         return;
 
-    clear_up(u);
+    if (u->sink_new_slot)
+        pa_hook_slot_free(u->sink_new_slot);
+
+    if (u->sink_unlink_slot)
+        pa_hook_slot_free(u->sink_unlink_slot);
+    
+    if (u->sink_state_changed_slot)
+        pa_hook_slot_free(u->sink_state_changed_slot);
+
+    if (u->sink)
+        pa_sink_unlink(u->sink);
+
+    if (u->outputs) {
+        while ((o = pa_idxset_first(u->outputs, NULL)))
+            output_free(o);
+        
+        pa_idxset_free(u->outputs, NULL, NULL);
+    }
+
+    if (u->thread) {
+        pa_asyncmsgq_send(u->thread_mq.inq, NULL, PA_MESSAGE_SHUTDOWN, NULL, 0, NULL);
+        pa_thread_free(u->thread);
+    }
+
+    pa_thread_mq_done(&u->thread_mq);
+    
+    if (u->sink)
+        pa_sink_unref(u->sink);
+
+    if (u->rtpoll)
+        pa_rtpoll_free(u->rtpoll);
+    
+    if (u->time_event)
+        u->core->mainloop->time_free(u->time_event);
+    
+    pa_mutex_free(u->mutex);
+        
     pa_xfree(u);
 }