X-Git-Url: http://git.tuebingen.mpg.de/?p=paraslash.git;a=blobdiff_plain;f=vss.c;h=bfea37a191f8b90db60db720916eb37ebf900223;hp=25082834eb0160de013dd11639f3354160ee261b;hb=e5fbc490c2c16ecfa7bce58a18e11a0f7d382b91;hpb=e56d454fbd865f34cb5e043ff43585754631b079

diff --git a/vss.c b/vss.c
index 25082834..bfea37a1 100644
--- a/vss.c
+++ b/vss.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1997-2010 Andre Noll <maan@systemlinux.org>
+ * Copyright (C) 1997-2012 Andre Noll <maan@systemlinux.org>
  *
  * Licensed under the GPL v2. For licencing details see COPYING.
  */
@@ -12,7 +12,6 @@
  */
 
 #include <regex.h>
-#include <dirent.h>
 #include <osl.h>
 
 #include "para.h"
@@ -27,10 +26,10 @@
 #include "server.cmdline.h"
 #include "list.h"
 #include "send.h"
+#include "sched.h"
 #include "vss.h"
 #include "ipc.h"
 #include "fd.h"
-#include "sched.h"
 
 extern struct misc_meta_data *mmd;
 
@@ -86,7 +85,7 @@ struct vss_task {
 	/** Used by the scheduler. */
 	struct task task;
 	/** Pointer to the header of the mapped audio file. */
-	const char *header_buf;
+	char *header_buf;
 	/** Length of the audio file header. */
 	size_t header_len;
 	/** Time between audio file headers are sent. */
@@ -129,8 +128,11 @@ struct fec_group {
 	struct timeval slice_duration;
 	/** Group contains the audio file header that occupies that many slices. */
 	uint8_t num_header_slices;
+	/** Number of bytes per slice for this group. */
+	uint16_t slice_bytes;
 };
 
+/** A FEC client is always in one of these states. */
 enum fec_client_state {
 	FEC_STATE_NONE = 0,	/**< not initialized and not enabled */
 	FEC_STATE_DISABLED,	/**< temporarily disabled */
@@ -165,6 +167,8 @@ struct fec_client {
 	struct timeval next_header_time;
 	/** Used for the last source pointer of an audio file. */
 	unsigned char *extra_src_buf;
+	/** Needed for the last slice of the audio file header. */
+	unsigned char *extra_header_buf;
 	/** Extra slices needed to store largest chunk + header. */
 	int num_extra_slices;
 	/** Contains the FEC-encoded data. */
@@ -209,47 +213,68 @@ static void write_fec_header(struct fec_client *fc, struct vss_task *vsst)
 	write_u32(buf + 14, g->bytes);
 
 	write_u8(buf + 18, fc->current_slice_num);
-	write_u16(buf + 20, fc->mps - FEC_HEADER_SIZE);
+	write_u8(buf + 19, 0); /* unused */
+	write_u16(buf + 20, g->slice_bytes);
 	write_u8(buf + 22, g->first_chunk? 0 : 1);
 	write_u8(buf + 23, vsst->header_len? 1 : 0);
-	memset(buf + 24, 0, 7);
+	memset(buf + 24, 0, 8);
 }
 
-static int need_audio_header(struct fec_client *fc, struct vss_task *vsst)
+static bool need_audio_header(struct fec_client *fc, struct vss_task *vsst)
 {
 	if (!mmd->current_chunk) {
 		tv_add(now, &vsst->header_interval, &fc->next_header_time);
-		return 0;
+		return false;
 	}
 	if (!vsst->header_buf)
-		return 0;
-	if (!vsst->header_len)
-		return 0;
-	if (fc->group.num && tv_diff(&fc->next_header_time, now, NULL) > 0)
-		return 0;
+		return false;
+	if (vsst->header_len == 0)
+		return false;
+	if (fc->group.num > 0) {
+		if (!fc->fcp->need_periodic_header)
+			return false;
+		if (tv_diff(&fc->next_header_time, now, NULL) > 0)
+			return false;
+	}
 	tv_add(now, &vsst->header_interval, &fc->next_header_time);
-	return 1;
+	return true;
 }
 
-static int num_slices(long unsigned bytes, int mps, int rs)
+static bool need_data_slices(struct fec_client *fc, struct vss_task *vsst)
+{
+	if (fc->group.num > 0)
+		return true;
+	if (!vsst->header_buf)
+		return true;
+	if (vsst->header_len == 0)
+		return true;
+	if (fc->fcp->need_periodic_header)
+		return true;
+	return false;
+}
+
+static int num_slices(long unsigned bytes, int max_payload, int rs)
 {
-	int m = mps - FEC_HEADER_SIZE;
 	int ret;
 
-	assert(m > 0);
+	assert(max_payload > 0);
 	assert(rs > 0);
-	ret = DIV_ROUND_UP(bytes, m);
+	ret = DIV_ROUND_UP(bytes, max_payload);
 	if (ret + rs > 255)
 		return -E_BAD_CT;
 	return ret;
 }
 
 /* set group start and group duration */
-static void set_group_timing(struct fec_client *fc, struct fec_group *g)
+static void set_group_timing(struct fec_client *fc, struct vss_task *vsst)
 {
+	struct fec_group *g = &fc->group;
 	struct timeval *chunk_tv = vss_chunk_time();
 
-	tv_scale(g->num_chunks, chunk_tv, &g->duration);
+	if (!need_data_slices(fc, vsst))
+		ms2tv(200, &g->duration);
+	else
+		tv_scale(g->num_chunks, chunk_tv, &g->duration);
 	tv_divide(fc->fcp->slices_per_group + fc->num_extra_slices,
 		&g->duration, &g->slice_duration);
 	PARA_DEBUG_LOG("durations (group/chunk/slice): %lu/%lu/%lu\n",
@@ -258,7 +283,7 @@ static void set_group_timing(struct fec_client *fc, struct fec_group *g)
 
 static int initialize_fec_client(struct fec_client *fc, struct vss_task *vsst)
 {
-	int k, n, ret, mps;
+	int k, n, ret;
 	int hs, ds, rs; /* header/data/redundant slices */
 	struct fec_client_parms *fcp = fc->fcp;
 
@@ -272,59 +297,220 @@ static int initialize_fec_client(struct fec_client *fc, struct vss_task *vsst)
 		ret = fcp->init_fec(fc->sc);
 		if (ret < 0)
 			return ret;
-		mps = ret;
+		fc->mps = ret;
 	} else
-		mps = generic_max_transport_msg_size(fc->sc->fd);
-	if (mps <= FEC_HEADER_SIZE)
+		fc->mps = generic_max_transport_msg_size(fc->sc->fd);
+	if (fc->mps <= FEC_HEADER_SIZE)
 		return -ERRNO_TO_PARA_ERROR(EINVAL);
 
 	rs = fc->fcp->slices_per_group - fc->fcp->data_slices_per_group;
-	ret = num_slices(vsst->header_len, mps, rs);
+	ret = num_slices(vsst->header_len, fc->mps - FEC_HEADER_SIZE, rs);
 	if (ret < 0)
-		goto err;
+		return ret;
 	hs = ret;
-	ret = num_slices(afh_get_largest_chunk_size(&mmd->afd.afhi),
-		mps, rs);
+	ret = num_slices(mmd->afd.max_chunk_size, fc->mps - FEC_HEADER_SIZE, rs);
 	if (ret < 0)
-		goto err;
+		return ret;
 	ds = ret;
-	k = ret + ds;
+	if (fc->fcp->need_periodic_header)
+		k = hs + ds;
+	else
+		k = PARA_MAX(hs, ds);
 	if (k < fc->fcp->data_slices_per_group)
 		k = fc->fcp->data_slices_per_group;
+	fc->num_extra_slices = k - fc->fcp->data_slices_per_group;
 	n = k + rs;
-	PARA_CRIT_LOG("hs: %d, ds: %d, rs: %d, k: %d, n: %d\n", hs, ds, rs, k, n);
 	fec_free(fc->parms);
 	ret = fec_new(k, n, &fc->parms);
 	if (ret < 0)
 		return ret;
-	fc->num_extra_slices = k - fc->fcp->data_slices_per_group;
-	PARA_NOTICE_LOG("fec parms %d:%d:%d (%d extra slices)\n",
-		mps, k, n, fc->num_extra_slices);
+	PARA_INFO_LOG("mps: %d, k: %d, n: %d, extra slices: %d\n",
+		fc->mps, k, n, fc->num_extra_slices);
 	fc->src_data = para_realloc(fc->src_data, k * sizeof(char *));
-	fc->enc_buf = para_realloc(fc->enc_buf, mps);
-	memset(fc->enc_buf, 0, mps);
-	fc->extra_src_buf = para_realloc(fc->extra_src_buf, mps);
-	memset(fc->extra_src_buf, 0, mps);
+	fc->enc_buf = para_realloc(fc->enc_buf, fc->mps);
+	fc->extra_src_buf = para_realloc(fc->extra_src_buf, fc->mps);
+	fc->extra_header_buf = para_realloc(fc->extra_header_buf, fc->mps);
 
-	fc->mps = mps;
 	fc->state = FEC_STATE_READY_TO_RUN;
 	fc->next_header_time.tv_sec = 0;
 	fc->stream_start = *now;
 	fc->first_stream_chunk = mmd->current_chunk;
 	return 1;
-err:
-	fec_free(fc->parms);
-	return ret;
+}
+
+static void vss_get_chunk(int chunk_num, struct vss_task *vsst,
+		char **buf, size_t *sz)
+{
+	/*
+	 * Chunk zero is special for header streams: It is the first portion of
+	 * the audio file which consists of the audio file header. It may be
+	 * arbitrary large due to embedded meta data. Audio format handlers may
+	 * replace the header by a stripped one with meta data omitted which is
+	 * of bounded size. We always use the stripped header for streaming
+	 * rather than the unmodified header (chunk zero).
+	 */
+	if (chunk_num == 0 && vsst->header_len > 0) {
+		*buf = vsst->header_buf; /* stripped header */
+		*sz = vsst->header_len;
+		return;
+	}
+	afh_get_chunk(chunk_num, &mmd->afd.afhi, vsst->map, (const char **)buf,
+		sz);
+}
+
+static void compute_group_size(struct vss_task *vsst, struct fec_group *g,
+		int max_bytes)
+{
+	char *buf;
+	size_t len;
+	int i, max_chunks = PARA_MAX(1LU, 150 / tv2ms(vss_chunk_time()));
+
+	if (g->first_chunk == 0) {
+		g->num_chunks = 1;
+		vss_get_chunk(0, vsst, &buf, &len);
+		g->bytes = len;
+		return;
+	}
+
+	g->num_chunks = 0;
+	g->bytes = 0;
+	/*
+	 * Include chunks into the group until the group duration is at least
+	 * 150ms.  For ogg and wma, a single chunk's duration (ogg page/wma
+	 * super frame) is already larger than 150ms, so a FEC group consists
+	 * of exactly one chunk for these audio formats.
+	 */
+	for (i = 0;; i++) {
+		int chunk_num = g->first_chunk + i;
+
+		if (g->bytes > 0 && i >= max_chunks) /* duration limit */
+			break;
+		if (chunk_num >= mmd->afd.afhi.chunks_total) /* eof */
+			break;
+		vss_get_chunk(chunk_num, vsst, &buf, &len);
+		if (g->bytes + len > max_bytes)
+			break;
+		/* Include this chunk */
+		g->bytes += len;
+		g->num_chunks++;
+	}
+	assert(g->num_chunks);
+}
+
+/*
+ * Compute the slice size of the next group.
+ *
+ * The FEC parameters n and k are fixed but the slice size varies per
+ * FEC group.  We'd like to choose slices as small as possible to avoid
+ * unnecessary FEC calculations but large enough to guarantee that the
+ * k data slices suffice to encode the header (if needed) and the data
+ * chunk(s).
+ *
+ * Once we know the payload of the next group, we define the number s
+ * of bytes per slice for this group by
+ *
+ * 	s = ceil(payload / k)
+ *
+ * However, for header streams, computing s is more complicated since no
+ * overlapping of header and data slices is possible. Hence we have k >=
+ * 2 and s must satisfy
+ *
+ * (*)	ceil(h / s) + ceil(d / s) <= k
+ *
+ * where h and d are payload of the header and the data chunk(s)
+ * respectively. In general there is no value for s such that (*)
+ * becomes an equality, for example if h = 4000, d = 5000 and k = 10.
+ *
+ * We use the following approach for computing a suitable value for s:
+ *
+ * Let
+ * 	k1 := ceil(k * min(h, d) / (h + d)),
+ * 	k2 := k - k1.
+ *
+ * Note that k >= 2 implies k1 > 0 and k2 > 0, so
+ *
+ * 	s := max(ceil(min(h, d) / k1), ceil(max(h, d) / k2))
+ *
+ * is well-defined. Inequality (*) holds for this value of s since k1
+ * slices suffice to store min(h, d) while k2 slices suffice to store
+ * max(h, d), i.e. the first addent of (*) is bounded by k1 and the
+ * second by k2.
+ *
+ * For the above example we obtain
+ *
+ * 	k1 = ceil(10 * 4000 / 9000) = 5, k2 = 5,
+ * 	s = max(4000 / 5, 5000 / 5) = 1000,
+ *
+ * which is optimal since a slice size of 999 bytes would already require
+ * 11 slices.
+ */
+static int compute_slice_size(struct fec_client *fc, struct vss_task *vsst)
+{
+	struct fec_group *g = &fc->group;
+	int k = fc->fcp->data_slices_per_group + fc->num_extra_slices;
+	int n = fc->fcp->slices_per_group + fc->num_extra_slices;
+	int ret, k1, k2, h, d, min, max, sum;
+	int max_slice_bytes = fc->mps - FEC_HEADER_SIZE;
+	int max_group_bytes;
+
+	if (!need_audio_header(fc, vsst)) {
+		max_group_bytes = k * max_slice_bytes;
+		g->num_header_slices = 0;
+		compute_group_size(vsst, g, max_group_bytes);
+		g->slice_bytes = DIV_ROUND_UP(g->bytes, k);
+		if (g->slice_bytes == 0)
+			g->slice_bytes = 1;
+		return 1;
+	}
+	if (!need_data_slices(fc, vsst)) {
+		g->bytes = 0;
+		g->num_chunks = 0;
+		g->slice_bytes = DIV_ROUND_UP(vsst->header_len, k);
+		g->num_header_slices = k;
+		return 1;
+	}
+	h = vsst->header_len;
+	max_group_bytes = (k - num_slices(h, max_slice_bytes, n - k))
+		* max_slice_bytes;
+	compute_group_size(vsst, g, max_group_bytes);
+	d = g->bytes;
+	if (d == 0) {
+		g->slice_bytes = DIV_ROUND_UP(h, k);
+		ret = num_slices(vsst->header_len, g->slice_bytes, n - k);
+		if (ret < 0)
+			return ret;
+		g->num_header_slices = ret;
+		return 1;
+	}
+	min = PARA_MIN(h, d);
+	max = PARA_MAX(h, d);
+	sum = h + d;
+	k1 = DIV_ROUND_UP(k * min, sum);
+	k2 = k - k1;
+	assert(k1 > 0);
+	assert(k2 > 0);
+
+	g->slice_bytes = PARA_MAX(DIV_ROUND_UP(min, k1), DIV_ROUND_UP(max, k2));
+	/*
+	 * This value of s := g->slice_bytes satisfies inequality (*) above,
+	 * but it might be larger than max_slice_bytes. However, we know that
+	 * max_slice_bytes are sufficient to store header and data, so:
+	 */
+	g->slice_bytes = PARA_MIN((int)g->slice_bytes, max_slice_bytes);
+
+	ret = num_slices(vsst->header_len, g->slice_bytes, n - k);
+	if (ret < 0)
+		return ret;
+	g->num_header_slices = ret;
+	return 1;
 }
 
 static int setup_next_fec_group(struct fec_client *fc, struct vss_task *vsst)
 {
 	int ret, i, k, n, data_slices;
 	size_t len;
-	const char *buf, *start_buf;
+	char *buf, *p;
 	struct fec_group *g = &fc->group;
-	unsigned slice_bytes;
-	uint32_t max_data_size;
 
 	if (fc->state == FEC_STATE_NONE) {
 		ret = initialize_fec_client(fc, vsst);
@@ -333,7 +519,6 @@ static int setup_next_fec_group(struct fec_client *fc, struct vss_task *vsst)
 		g->first_chunk = mmd->current_chunk;
 		g->num = 0;
 		g->start = *now;
-		
 	} else {
 		struct timeval tmp;
 		if (g->first_chunk + g->num_chunks >= mmd->afd.afhi.chunks_total)
@@ -344,75 +529,82 @@ static int setup_next_fec_group(struct fec_client *fc, struct vss_task *vsst)
 		 */
 		tmp = g->start;
 		tv_add(&tmp, &g->duration, &g->start);
-		set_group_timing(fc, g);
+		set_group_timing(fc, vsst);
 		g->first_chunk += g->num_chunks;
 		g->num++;
 	}
-	slice_bytes = fc->mps - FEC_HEADER_SIZE;
-	PARA_CRIT_LOG("slice_bytes: %d\n", slice_bytes);
 	k = fc->fcp->data_slices_per_group + fc->num_extra_slices;
 	n = fc->fcp->slices_per_group + fc->num_extra_slices;
-	PARA_CRIT_LOG("k: %d, n: %d\n", k, n);
-	if (need_audio_header(fc, vsst)) {
-		ret = num_slices(vsst->header_len, slice_bytes, n - k);
-		if (ret < 0)
-			return ret;
-		g->num_header_slices = ret;
-	} else
-		g->num_header_slices = 0;
-	afh_get_chunk(g->first_chunk, &mmd->afd.afhi, vsst->map, &start_buf,
-		&len);
-	data_slices = k - g->num_header_slices;
-	assert(data_slices);
-	max_data_size = slice_bytes * data_slices;
-	g->bytes = 0;
-	for (i = g->first_chunk; i < mmd->afd.afhi.chunks_total; i++) {
-		afh_get_chunk(i, &mmd->afd.afhi, vsst->map, &buf, &len);
-		if (g->bytes + len > max_data_size)
-			break;
-		g->bytes += len;
-	}
-	g->num_chunks = i - g->first_chunk;
-	assert(g->num_chunks);
+
+	compute_slice_size(fc, vsst);
+	assert(g->slice_bytes > 0);
+	ret = num_slices(g->bytes, g->slice_bytes, n - k);
+	if (ret < 0)
+		return ret;
+	data_slices = ret;
+	assert(g->num_header_slices + data_slices <= k);
 	fc->current_slice_num = 0;
 	if (g->num == 0)
-		set_group_timing(fc, g);
-
+		set_group_timing(fc, vsst);
 	/* setup header slices */
 	buf = vsst->header_buf;
 	for (i = 0; i < g->num_header_slices; i++) {
-		fc->src_data[i] = (const unsigned char *)buf;
-		buf += slice_bytes;
+		uint32_t payload_size;
+		if (buf + g->slice_bytes <= vsst->header_buf + vsst->header_len) {
+			fc->src_data[i] = (const unsigned char *)buf;
+			buf += g->slice_bytes;
+			continue;
+		}
+		/*
+		 * Can not use vss->header_buf for this slice as it
+		 * goes beyond the buffer. This slice will not be fully
+		 * used.
+		 */
+		payload_size = vsst->header_buf + vsst->header_len - buf;
+		memcpy(fc->extra_header_buf, buf, payload_size);
+		if (payload_size < g->slice_bytes)
+			memset(fc->extra_header_buf + payload_size, 0,
+				g->slice_bytes - payload_size);
+		fc->src_data[i] = fc->extra_header_buf;
+		assert(i == g->num_header_slices - 1);
 	}
 
-	/* setup data slices */
-	buf = start_buf;
-	for (i = g->num_header_slices; i < k; i++) {
-		if (buf + slice_bytes > vsst->map + mmd->size)
+	/*
+	 * Setup data slices. Note that for ogg streams chunk 0 points to a
+	 * buffer on the heap rather than to the mapped audio file.
+	 */
+	vss_get_chunk(g->first_chunk, vsst, &buf, &len);
+	for (p = buf; i < g->num_header_slices + data_slices; i++) {
+		if (p + g->slice_bytes > buf + g->bytes) {
 			/*
-			 * Can not use the memory mapped audio file for this
-			 * slice as it goes beyond the map. This slice will not
-			 * be fully used.
+			 * We must make a copy for this slice since using p
+			 * directly would exceed the buffer.
 			 */
+			uint32_t payload_size = buf + g->bytes - p;
+			assert(payload_size + FEC_HEADER_SIZE <= fc->mps);
+			memcpy(fc->extra_src_buf, p, payload_size);
+			if (payload_size < g->slice_bytes)
+				memset(fc->extra_src_buf + payload_size, 0,
+					g->slice_bytes - payload_size);
+			fc->src_data[i] = fc->extra_src_buf;
+			i++;
 			break;
-		fc->src_data[i] = (const unsigned char *)buf;
-		buf += slice_bytes;
+		}
+		fc->src_data[i] = (const unsigned char *)p;
+		p += g->slice_bytes;
 	}
 	if (i < k) {
-		uint32_t payload_size = vsst->map + mmd->size - buf;
-		memcpy(fc->extra_src_buf, buf, payload_size);
-		fc->src_data[i] = fc->extra_src_buf;
-		i++;
 		/* use arbitrary data for all remaining slices */
 		buf = vsst->map;
 		for (; i < k; i++)
 			fc->src_data[i] = (const unsigned char *)buf;
 	}
-	PARA_DEBUG_LOG("FEC group %d: %d chunks (%d - %d), "
-		"%d header slices, %d data slices\n",
+	PARA_DEBUG_LOG("FEC group %d: %d chunks (%d - %d), %d bytes\n",
 		g->num, g->num_chunks, g->first_chunk,
-		g->first_chunk + g->num_chunks - 1,
-		g->num_header_slices, data_slices
+		g->first_chunk + g->num_chunks - 1, g->bytes
+	);
+	PARA_DEBUG_LOG("slice_bytes: %d, %d header slices, %d data slices\n",
+		g->slice_bytes, g->num_header_slices, data_slices
 	);
 	return 1;
 }
@@ -433,7 +625,7 @@ static int compute_next_fec_slice(struct fec_client *fc, struct vss_task *vsst)
 	}
 	write_fec_header(fc, vsst);
 	fec_encode(fc->parms, fc->src_data, fc->enc_buf + FEC_HEADER_SIZE,
-		fc->current_slice_num, fc->mps - FEC_HEADER_SIZE);
+		fc->current_slice_num, fc->group.slice_bytes);
 	return 1;
 }
 
@@ -483,6 +675,7 @@ void vss_del_fec_client(struct fec_client *fc)
 	free(fc->src_data);
 	free(fc->enc_buf);
 	free(fc->extra_src_buf);
+	free(fc->extra_header_buf);
 	fec_free(fc->parms);
 	free(fc);
 }
@@ -507,26 +700,6 @@ static int next_slice_is_due(struct fec_client *fc, struct timeval *diff)
 	return ret < 0? 1 : 0;
 }
 
-static void compute_slice_timeout(struct timeval *timeout)
-{
-	struct fec_client *fc;
-
-	list_for_each_entry(fc, &fec_client_list, node) {
-		struct timeval diff;
-
-		if (fc->state != FEC_STATE_READY_TO_RUN)
-			continue;
-		if (next_slice_is_due(fc, &diff)) {
-			timeout->tv_sec = 0;
-			timeout->tv_usec = 0;
-			return;
-		}
-		/* timeout = min(timeout, diff) */
-		if (tv_diff(&diff, timeout, NULL) < 0)
-			*timeout = diff;
-	}
-}
-
 static void set_eof_barrier(struct vss_task *vsst)
 {
 	struct fec_client *fc;
@@ -617,42 +790,38 @@ static int chk_barrier(const char *bname, const struct timeval *barrier,
 	return -1;
 }
 
-/*
- * != NULL: timeout for next chunk
- * NULL: nothing to do
- */
-static struct timeval *vss_compute_timeout(struct vss_task *vsst)
+static void vss_compute_timeout(struct sched *s, struct vss_task *vsst)
 {
-	static struct timeval the_timeout;
-	struct timeval next_chunk;
+	struct timeval tv;
+	struct fec_client *fc;
 
-	if (vss_next() && vsst->map) {
-		/* only sleep a bit, nec*/
-		the_timeout.tv_sec = 0;
-		the_timeout.tv_usec = 100;
-		return &the_timeout;
-	}
-	if (chk_barrier("autoplay_delay", &vsst->autoplay_barrier,
-			&the_timeout, 1) < 0)
-		return &the_timeout;
-	if (chk_barrier("eof", &vsst->eof_barrier, &the_timeout, 1) < 0)
-		return &the_timeout;
-	if (chk_barrier("data send", &vsst->data_send_barrier,
-			&the_timeout, 1) < 0)
-		return &the_timeout;
 	if (!vss_playing() || !vsst->map)
-		return NULL;
+		return;
+	if (vss_next() && vsst->map) /* only sleep a bit, nec*/
+		return sched_request_timeout_ms(100, s);
+
+	/* Each of these barriers must have passed until we may proceed */
+	if (sched_request_barrier(&vsst->autoplay_barrier, s) == 1)
+		return;
+	if (sched_request_barrier(&vsst->eof_barrier, s) == 1)
+		return;
+	if (sched_request_barrier(&vsst->data_send_barrier, s) == 1)
+		return;
+	/*
+	 * Compute the select timeout as the minimal time until the next
+	 * chunk/slice is due for any client.
+	 */
 	compute_chunk_time(mmd->chunks_sent, &mmd->afd.afhi.chunk_tv,
-		&mmd->stream_start, &next_chunk);
-	if (chk_barrier("chunk", &next_chunk, &the_timeout, 0) >= 0) {
-		/* chunk is due or bof */
-		the_timeout.tv_sec = 0;
-		the_timeout.tv_usec = 0;
-		return &the_timeout;
+		&mmd->stream_start, &tv);
+	if (sched_request_barrier_or_min_delay(&tv, s) == 0)
+		return;
+	list_for_each_entry(fc, &fec_client_list, node) {
+		if (fc->state != FEC_STATE_READY_TO_RUN)
+			continue;
+		if (next_slice_is_due(fc, &tv))
+			return sched_min_delay(s);
+		sched_request_timeout(&tv, s);
 	}
-	/* compute min of current timeout and next slice time */
-	compute_slice_timeout(&the_timeout);
-	return &the_timeout;
 }
 
 static void vss_eof(struct vss_task *vsst)
@@ -663,6 +832,8 @@ static void vss_eof(struct vss_task *vsst)
 	if (mmd->new_vss_status_flags & VSS_NOMORE)
 		mmd->new_vss_status_flags = VSS_NEXT;
 	set_eof_barrier(vsst);
+	afh_free_header(vsst->header_buf, mmd->afd.audio_format_id);
+	vsst->header_buf = NULL;
 	para_munmap(vsst->map, mmd->size);
 	vsst->map = NULL;
 	mmd->chunks_sent = 0;
@@ -719,7 +890,6 @@ static void set_mmd_offset(void)
 static void vss_pre_select(struct sched *s, struct task *t)
 {
 	int i;
-	struct timeval *tv;
 	struct vss_task *vsst = container_of(t, struct vss_task, task);
 
 	if (!vsst->map || vss_next() || vss_paused() || vss_repos()) {
@@ -757,9 +927,7 @@ static void vss_pre_select(struct sched *s, struct task *t)
 			continue;
 		senders[i].pre_select(&s->max_fileno, &s->rfds, &s->wfds);
 	}
-	tv = vss_compute_timeout(vsst);
-	if (tv)
-		sched_request_timeout(tv, s);
+	vss_compute_timeout(s, vsst);
 }
 
 static int recv_afs_msg(int afs_socket, int *fd, uint32_t *code, uint32_t *data)
@@ -796,6 +964,10 @@ static int recv_afs_msg(int afs_socket, int *fd, uint32_t *code, uint32_t *data)
 	return 1;
 }
 
+#ifndef MAP_POPULATE
+#define MAP_POPULATE 0
+#endif
+
 static void recv_afs_result(struct vss_task *vsst, fd_set *rfds)
 {
 	int ret, passed_fd, shmid;
@@ -830,8 +1002,8 @@ static void recv_afs_result(struct vss_task *vsst, fd_set *rfds)
 	}
 	mmd->size = statbuf.st_size;
 	mmd->mtime = statbuf.st_mtime;
-	ret = para_mmap(mmd->size, PROT_READ, MAP_PRIVATE, passed_fd,
-		0, &vsst->map);
+	ret = para_mmap(mmd->size, PROT_READ, MAP_PRIVATE | MAP_POPULATE,
+		passed_fd, 0, &vsst->map);
 	if (ret < 0)
 		goto err;
 	close(passed_fd);
@@ -841,8 +1013,8 @@ static void recv_afs_result(struct vss_task *vsst, fd_set *rfds)
 	mmd->events++;
 	mmd->num_played++;
 	mmd->new_vss_status_flags &= (~VSS_NEXT);
-	afh_get_header(&mmd->afd.afhi, vsst->map, &vsst->header_buf,
-		&vsst->header_len);
+	afh_get_header(&mmd->afd.afhi, mmd->afd.audio_format_id,
+		vsst->map, mmd->size, &vsst->header_buf, &vsst->header_len);
 	return;
 err:
 	free(mmd->afd.afhi.chunk_table);
@@ -884,8 +1056,9 @@ static void vss_send(struct vss_task *vsst)
 		if (compute_next_fec_slice(fc, vsst) <= 0)
 			continue;
 		PARA_DEBUG_LOG("sending %d:%d (%u bytes)\n", fc->group.num,
-			fc->current_slice_num, fc->mps);
-		fc->fcp->send_fec(fc->sc, (char *)fc->enc_buf, fc->mps);
+			fc->current_slice_num, fc->group.slice_bytes);
+		fc->fcp->send_fec(fc->sc, (char *)fc->enc_buf,
+			fc->group.slice_bytes + FEC_HEADER_SIZE);
 		fc->current_slice_num++;
 		fec_active = 1;
 	}
@@ -897,7 +1070,7 @@ static void vss_send(struct vss_task *vsst)
 	compute_chunk_time(mmd->chunks_sent, &mmd->afd.afhi.chunk_tv,
 		&mmd->stream_start, &due);
 	if (tv_diff(&due, now, NULL) <= 0) {
-		const char *buf;
+		char *buf;
 		size_t len;
 
 		if (!mmd->chunks_sent) {
@@ -910,14 +1083,32 @@ static void vss_send(struct vss_task *vsst)
 		 * they might have still some data queued which can be sent in
 		 * this case.
 		 */
-		afh_get_chunk(mmd->current_chunk, &mmd->afd.afhi, vsst->map,
-			&buf, &len);
+		vss_get_chunk(mmd->current_chunk, vsst, &buf, &len);
 		for (i = 0; senders[i].name; i++) {
 			if (!senders[i].send)
 				continue;
 			senders[i].send(mmd->current_chunk, mmd->chunks_sent,
 				buf, len, vsst->header_buf, vsst->header_len);
 		}
+		/*
+		 * Prefault next chunk(s)
+		 *
+		 * If the backing device of the memory-mapped audio file is
+		 * slow and read-ahead is turned off or prevented for some
+		 * reason, e.g. due to memory pressure, it may take much longer
+		 * than the chunk interval to get the next chunk on the wire,
+		 * causing buffer underruns on the client side. Mapping the
+		 * file with MAP_POPULATE seems to help a bit, but it does not
+		 * eliminate the delays completely. Moreover, it is supported
+		 * only on Linux. So we do our own read-ahead here.
+		 */
+		if (mmd->current_chunk > 0) { /* chunk 0 might be on the heap */
+			buf += len;
+			for (i = 0; i < 5 && buf < vsst->map + mmd->size; i++) {
+				__a_unused volatile char x = *buf;
+				buf += 4096;
+			}
+		}
 		mmd->chunks_sent++;
 		mmd->current_chunk++;
 	}
@@ -966,11 +1157,12 @@ static void vss_post_select(struct sched *s, struct task *t)
  * Initialize the virtual streaming system task.
  *
  * \param afs_socket The fd for communication with afs.
+ * \param s The scheduler to register the vss task to.
  *
  * This also initializes all supported senders and starts streaming
  * if the --autoplay command line flag was given.
  */
-void init_vss_task(int afs_socket)
+void init_vss_task(int afs_socket, struct sched *s)
 {
 	static struct vss_task vss_task_struct, *vsst = &vss_task_struct;
 	int i;
@@ -1002,5 +1194,6 @@ void init_vss_task(int afs_socket)
 		tv_add(&vsst->autoplay_barrier, &vsst->announce_tv,
 			&vsst->data_send_barrier);
 	}
-	register_task(&vsst->task);
+	sprintf(vsst->task.status, "vss task");
+	register_task(s, &vsst->task);
 }