]> git.sesse.net Git - ffmpeg/blobdiff - libavformat/hls.c
Merge commit '6f243b17c537646b894857d43dfdac65f85ab377'
[ffmpeg] / libavformat / hls.c
index 9af3356a3874e0401d3e9ea24766b88e48fc77db..3897723b7defac135e7a4e3ab49b51acfdedf78a 100644 (file)
@@ -27,6 +27,7 @@
  */
 
 #include "libavutil/avstring.h"
+#include "libavutil/avassert.h"
 #include "libavutil/intreadwrite.h"
 #include "libavutil/mathematics.h"
 #include "libavutil/opt.h"
 #include "internal.h"
 #include "avio_internal.h"
 #include "url.h"
+#include "id3v2.h"
 
 #define INITIAL_BUFFER_SIZE 32768
 
 #define MAX_FIELD_LEN 64
 #define MAX_CHARACTERISTICS_LEN 512
 
+#define MPEG_TIME_BASE 90000
+#define MPEG_TIME_BASE_Q (AVRational){1, MPEG_TIME_BASE}
+
 /*
  * An apple http stream consists of a playlist with media segment files,
  * played sequentially. There may be several playlists with the same
@@ -63,14 +68,20 @@ struct segment {
     int64_t duration;
     int64_t url_offset;
     int64_t size;
-    char url[MAX_URL_SIZE];
-    char key[MAX_URL_SIZE];
+    char *url;
+    char *key;
     enum KeyType key_type;
     uint8_t iv[16];
 };
 
 struct rendition;
 
+enum PlaylistType {
+    PLS_TYPE_UNSPECIFIED,
+    PLS_TYPE_EVENT,
+    PLS_TYPE_VOD
+};
+
 /*
  * Each playlist has its own demuxer. If it currently is active,
  * it has an open AVIOContext too, and potentially an AVPacket
@@ -88,6 +99,7 @@ struct playlist {
     int stream_offset;
 
     int finished;
+    enum PlaylistType type;
     int64_t target_duration;
     int start_seq_no;
     int n_segments;
@@ -100,6 +112,22 @@ struct playlist {
     char key_url[MAX_URL_SIZE];
     uint8_t key[16];
 
+    /* ID3 timestamp handling (elementary audio streams have ID3 timestamps
+     * (and possibly other ID3 tags) in the beginning of each segment) */
+    int is_id3_timestamped; /* -1: not yet known */
+    int64_t id3_mpegts_timestamp; /* in mpegts tb */
+    int64_t id3_offset; /* in stream original tb */
+    uint8_t* id3_buf; /* temp buffer for id3 parsing */
+    unsigned int id3_buf_size;
+    AVDictionary *id3_initial; /* data from first id3 tag */
+    int id3_found; /* ID3 tag found at some point */
+    int id3_changed; /* ID3 tag data has changed at some point */
+    ID3v2ExtraMeta *id3_deferred_extra; /* stored here until subdemuxer is opened */
+
+    int64_t seek_timestamp;
+    int seek_flags;
+    int seek_stream_index; /* into subdemuxer stream array */
+
     /* Renditions associated with this playlist, if any.
      * Alternative rendition playlists have a single rendition associated
      * with them, and variant main Media Playlists may have
@@ -144,11 +172,9 @@ typedef struct HLSContext {
     struct rendition **renditions;
 
     int cur_seq_no;
-    int end_of_segment;
     int first_packet;
     int64_t first_timestamp;
-    int64_t seek_timestamp;
-    int seek_flags;
+    int64_t cur_timestamp;
     AVIOInterruptCB *interrupt_callback;
     char *user_agent;                    ///< holds HTTP user agent set as an AVOption to the HTTP protocol context
     char *cookies;                       ///< holds HTTP cookie values set in either the initial response or as an AVOption to the HTTP protocol context
@@ -166,8 +192,11 @@ static int read_chomp_line(AVIOContext *s, char *buf, int maxlen)
 static void free_segment_list(struct playlist *pls)
 {
     int i;
-    for (i = 0; i < pls->n_segments; i++)
+    for (i = 0; i < pls->n_segments; i++) {
+        av_free(pls->segments[i]->key);
+        av_free(pls->segments[i]->url);
         av_free(pls->segments[i]);
+    }
     av_freep(&pls->segments);
     pls->n_segments = 0;
 }
@@ -179,6 +208,9 @@ static void free_playlist_list(HLSContext *c)
         struct playlist *pls = c->playlists[i];
         free_segment_list(pls);
         av_freep(&pls->renditions);
+        av_freep(&pls->id3_buf);
+        av_dict_free(&pls->id3_initial);
+        ff_id3v2_free_extra_meta(&pls->id3_deferred_extra);
         av_free_packet(&pls->pkt);
         av_free(pls->pb.buffer);
         if (pls->input)
@@ -234,6 +266,11 @@ static struct playlist *new_playlist(HLSContext *c, const char *url,
         return NULL;
     reset_packet(&pls->pkt);
     ff_make_absolute_url(pls->url, sizeof(pls->url), base, url);
+    pls->seek_timestamp = AV_NOPTS_VALUE;
+
+    pls->is_id3_timestamped = -1;
+    pls->id3_mpegts_timestamp = AV_NOPTS_VALUE;
+
     dynarray_add(&c->playlists, &c->n_playlists, pls);
     return pls;
 }
@@ -438,6 +475,22 @@ static void handle_rendition_args(struct rendition_info *info, const char *key,
      */
 }
 
+/* used by parse_playlist to allocate a new variant+playlist when the
+ * playlist is detected to be a Media Playlist (not Master Playlist)
+ * and we have no parent Master Playlist (parsing of which would have
+ * allocated the variant and playlist already) */
+static int ensure_playlist(HLSContext *c, struct playlist **pls, const char *url)
+{
+    if (*pls)
+        return 0;
+    if (!new_variant(c, NULL, url, NULL))
+        return AVERROR(ENOMEM);
+    *pls = c->playlists[c->n_playlists - 1];
+    return 0;
+}
+
+/* pls = NULL  => Master Playlist or parentless Media Playlist
+ * pls = !NULL => parented Media Playlist, playlist+variant allocated */
 static int parse_playlist(HLSContext *c, const char *url,
                           struct playlist *pls, AVIOContext *in)
 {
@@ -454,6 +507,7 @@ static int parse_playlist(HLSContext *c, const char *url,
     int64_t seg_size = -1;
     uint8_t *new_url = NULL;
     struct variant_info variant_info;
+    char tmp_str[MAX_URL_SIZE];
 
     if (!in) {
         AVDictionary *opts = NULL;
@@ -485,6 +539,7 @@ static int parse_playlist(HLSContext *c, const char *url,
     if (pls) {
         free_segment_list(pls);
         pls->finished = 0;
+        pls->type = PLS_TYPE_UNSPECIFIED;
     }
     while (!url_feof(in)) {
         read_chomp_line(in, line, sizeof(line));
@@ -512,23 +567,23 @@ static int parse_playlist(HLSContext *c, const char *url,
                                &info);
             new_rendition(c, &info, url);
         } else if (av_strstart(line, "#EXT-X-TARGETDURATION:", &ptr)) {
-            if (!pls) {
-                if (!new_variant(c, NULL, url, NULL)) {
-                    ret = AVERROR(ENOMEM);
-                    goto fail;
-                }
-                pls = c->playlists[c->n_playlists - 1];
-            }
+            ret = ensure_playlist(c, &pls, url);
+            if (ret < 0)
+                goto fail;
             pls->target_duration = atoi(ptr) * AV_TIME_BASE;
         } else if (av_strstart(line, "#EXT-X-MEDIA-SEQUENCE:", &ptr)) {
-            if (!pls) {
-                if (!new_variant(c, NULL, url, NULL)) {
-                    ret = AVERROR(ENOMEM);
-                    goto fail;
-                }
-                pls = c->playlists[c->n_playlists - 1];
-            }
+            ret = ensure_playlist(c, &pls, url);
+            if (ret < 0)
+                goto fail;
             pls->start_seq_no = atoi(ptr);
+        } else if (av_strstart(line, "#EXT-X-PLAYLIST-TYPE:", &ptr)) {
+            ret = ensure_playlist(c, &pls, url);
+            if (ret < 0)
+                goto fail;
+            if (!strcmp(ptr, "EVENT"))
+                pls->type = PLS_TYPE_EVENT;
+            else if (!strcmp(ptr, "VOD"))
+                pls->type = PLS_TYPE_VOD;
         } else if (av_strstart(line, "#EXT-X-ENDLIST", &ptr)) {
             if (pls)
                 pls->finished = 1;
@@ -573,8 +628,28 @@ static int parse_playlist(HLSContext *c, const char *url,
                     memset(seg->iv, 0, sizeof(seg->iv));
                     AV_WB32(seg->iv + 12, seq);
                 }
-                ff_make_absolute_url(seg->key, sizeof(seg->key), url, key);
-                ff_make_absolute_url(seg->url, sizeof(seg->url), url, line);
+
+                if (key_type != KEY_NONE) {
+                    ff_make_absolute_url(tmp_str, sizeof(tmp_str), url, key);
+                    seg->key = av_strdup(tmp_str);
+                    if (!seg->key) {
+                        av_free(seg);
+                        ret = AVERROR(ENOMEM);
+                        goto fail;
+                    }
+                } else {
+                    seg->key = NULL;
+                }
+
+                ff_make_absolute_url(tmp_str, sizeof(tmp_str), url, line);
+                seg->url = av_strdup(tmp_str);
+                if (!seg->url) {
+                    av_free(seg->key);
+                    av_free(seg);
+                    ret = AVERROR(ENOMEM);
+                    goto fail;
+                }
+
                 dynarray_add(&pls->segments, &pls->n_segments, seg);
                 is_segment = 0;
 
@@ -600,6 +675,231 @@ fail:
     return ret;
 }
 
+enum ReadFromURLMode {
+    READ_NORMAL,
+    READ_COMPLETE,
+};
+
+/* read from URLContext, limiting read to current segment */
+static int read_from_url(struct playlist *pls, uint8_t *buf, int buf_size,
+                         enum ReadFromURLMode mode)
+{
+    int ret;
+    struct segment *seg = pls->segments[pls->cur_seq_no - pls->start_seq_no];
+
+     /* limit read if the segment was only a part of a file */
+    if (seg->size >= 0)
+        buf_size = FFMIN(buf_size, seg->size - pls->cur_seg_offset);
+
+    if (mode == READ_COMPLETE)
+        ret = ffurl_read_complete(pls->input, buf, buf_size);
+    else
+        ret = ffurl_read(pls->input, buf, buf_size);
+
+    if (ret > 0)
+        pls->cur_seg_offset += ret;
+
+    return ret;
+}
+
+/* Parse the raw ID3 data and pass contents to caller */
+static void parse_id3(AVFormatContext *s, AVIOContext *pb,
+                      AVDictionary **metadata, int64_t *dts,
+                      ID3v2ExtraMetaAPIC **apic, ID3v2ExtraMeta **extra_meta)
+{
+    static const char id3_priv_owner_ts[] = "com.apple.streaming.transportStreamTimestamp";
+    ID3v2ExtraMeta *meta;
+
+    ff_id3v2_read_dict(pb, metadata, ID3v2_DEFAULT_MAGIC, extra_meta);
+    for (meta = *extra_meta; meta; meta = meta->next) {
+        if (!strcmp(meta->tag, "PRIV")) {
+            ID3v2ExtraMetaPRIV *priv = meta->data;
+            if (priv->datasize == 8 && !strcmp(priv->owner, id3_priv_owner_ts)) {
+                /* 33-bit MPEG timestamp */
+                int64_t ts = AV_RB64(priv->data);
+                av_log(s, AV_LOG_DEBUG, "HLS ID3 audio timestamp %"PRId64"\n", ts);
+                if ((ts & ~((1ULL << 33) - 1)) == 0)
+                    *dts = ts;
+                else
+                    av_log(s, AV_LOG_ERROR, "Invalid HLS ID3 audio timestamp %"PRId64"\n", ts);
+            }
+        } else if (!strcmp(meta->tag, "APIC") && apic)
+            *apic = meta->data;
+    }
+}
+
+/* Check if the ID3 metadata contents have changed */
+static int id3_has_changed_values(struct playlist *pls, AVDictionary *metadata,
+                                  ID3v2ExtraMetaAPIC *apic)
+{
+    AVDictionaryEntry *entry = NULL;
+    AVDictionaryEntry *oldentry;
+    /* check that no keys have changed values */
+    while ((entry = av_dict_get(metadata, "", entry, AV_DICT_IGNORE_SUFFIX))) {
+        oldentry = av_dict_get(pls->id3_initial, entry->key, NULL, AV_DICT_MATCH_CASE);
+        if (!oldentry || strcmp(oldentry->value, entry->value) != 0)
+            return 1;
+    }
+
+    /* check if apic appeared */
+    if (apic && (pls->ctx->nb_streams != 2 || !pls->ctx->streams[1]->attached_pic.data))
+        return 1;
+
+    if (apic) {
+        int size = pls->ctx->streams[1]->attached_pic.size;
+        if (size != apic->buf->size - FF_INPUT_BUFFER_PADDING_SIZE)
+            return 1;
+
+        if (memcmp(apic->buf->data, pls->ctx->streams[1]->attached_pic.data, size) != 0)
+            return 1;
+    }
+
+    return 0;
+}
+
+/* Parse ID3 data and handle the found data */
+static void handle_id3(AVIOContext *pb, struct playlist *pls)
+{
+    AVDictionary *metadata = NULL;
+    ID3v2ExtraMetaAPIC *apic = NULL;
+    ID3v2ExtraMeta *extra_meta = NULL;
+    int64_t timestamp = AV_NOPTS_VALUE;
+
+    parse_id3(pls->ctx, pb, &metadata, &timestamp, &apic, &extra_meta);
+
+    if (timestamp != AV_NOPTS_VALUE) {
+        pls->id3_mpegts_timestamp = timestamp;
+        pls->id3_offset = 0;
+    }
+
+    if (!pls->id3_found) {
+        /* initial ID3 tags */
+        av_assert0(!pls->id3_deferred_extra);
+        pls->id3_found = 1;
+
+        /* get picture attachment and set text metadata */
+        if (pls->ctx->nb_streams)
+            ff_id3v2_parse_apic(pls->ctx, &extra_meta);
+        else
+            /* demuxer not yet opened, defer picture attachment */
+            pls->id3_deferred_extra = extra_meta;
+
+        av_dict_copy(&pls->ctx->metadata, metadata, 0);
+        pls->id3_initial = metadata;
+
+    } else {
+        if (!pls->id3_changed && id3_has_changed_values(pls, metadata, apic)) {
+            avpriv_report_missing_feature(pls->ctx, "Changing ID3 metadata in HLS audio elementary stream");
+            pls->id3_changed = 1;
+        }
+        av_dict_free(&metadata);
+    }
+
+    if (!pls->id3_deferred_extra)
+        ff_id3v2_free_extra_meta(&extra_meta);
+}
+
+/* Intercept and handle ID3 tags between URLContext and AVIOContext */
+static void intercept_id3(struct playlist *pls, uint8_t *buf,
+                         int buf_size, int *len)
+{
+    /* intercept id3 tags, we do not want to pass them to the raw
+     * demuxer on all segment switches */
+    int bytes;
+    int id3_buf_pos = 0;
+    int fill_buf = 0;
+
+    /* gather all the id3 tags */
+    while (1) {
+        /* see if we can retrieve enough data for ID3 header */
+        if (*len < ID3v2_HEADER_SIZE && buf_size >= ID3v2_HEADER_SIZE) {
+            bytes = read_from_url(pls, buf + *len, ID3v2_HEADER_SIZE - *len, READ_COMPLETE);
+            if (bytes > 0) {
+
+                if (bytes == ID3v2_HEADER_SIZE - *len)
+                    /* no EOF yet, so fill the caller buffer again after
+                     * we have stripped the ID3 tags */
+                    fill_buf = 1;
+
+                *len += bytes;
+
+            } else if (*len <= 0) {
+                /* error/EOF */
+                *len = bytes;
+                fill_buf = 0;
+            }
+        }
+
+        if (*len < ID3v2_HEADER_SIZE)
+            break;
+
+        if (ff_id3v2_match(buf, ID3v2_DEFAULT_MAGIC)) {
+            struct segment *seg = pls->segments[pls->cur_seq_no - pls->start_seq_no];
+            int64_t maxsize = seg->size >= 0 ? seg->size : 1024*1024;
+            int taglen = ff_id3v2_tag_len(buf);
+            int tag_got_bytes = FFMIN(taglen, *len);
+            int remaining = taglen - tag_got_bytes;
+
+            if (taglen > maxsize) {
+                av_log(pls->ctx, AV_LOG_ERROR, "Too large HLS ID3 tag (%d > %"PRId64" bytes)\n",
+                       taglen, maxsize);
+                break;
+            }
+
+            /*
+             * Copy the id3 tag to our temporary id3 buffer.
+             * We could read a small id3 tag directly without memcpy, but
+             * we would still need to copy the large tags, and handling
+             * both of those cases together with the possibility for multiple
+             * tags would make the handling a bit complex.
+             */
+            pls->id3_buf = av_fast_realloc(pls->id3_buf, &pls->id3_buf_size, id3_buf_pos + taglen);
+            if (!pls->id3_buf)
+                break;
+            memcpy(pls->id3_buf + id3_buf_pos, buf, tag_got_bytes);
+            id3_buf_pos += tag_got_bytes;
+
+            /* strip the intercepted bytes */
+            *len -= tag_got_bytes;
+            memmove(buf, buf + tag_got_bytes, *len);
+            av_log(pls->ctx, AV_LOG_DEBUG, "Stripped %d HLS ID3 bytes\n", tag_got_bytes);
+
+            if (remaining > 0) {
+                /* read the rest of the tag in */
+                if (read_from_url(pls, pls->id3_buf + id3_buf_pos, remaining, READ_COMPLETE) != remaining)
+                    break;
+                id3_buf_pos += remaining;
+                av_log(pls->ctx, AV_LOG_DEBUG, "Stripped additional %d HLS ID3 bytes\n", remaining);
+            }
+
+        } else {
+            /* no more ID3 tags */
+            break;
+        }
+    }
+
+    /* re-fill buffer for the caller unless EOF */
+    if (*len >= 0 && (fill_buf || *len == 0)) {
+        bytes = read_from_url(pls, buf + *len, buf_size - *len, READ_NORMAL);
+
+        /* ignore error if we already had some data */
+        if (bytes >= 0)
+            *len += bytes;
+        else if (*len == 0)
+            *len = bytes;
+    }
+
+    if (pls->id3_buf) {
+        /* Now parse all the ID3 tags */
+        AVIOContext id3ioctx;
+        ffio_init_context(&id3ioctx, pls->id3_buf, id3_buf_pos, 0, NULL, NULL, NULL, NULL);
+        handle_id3(&id3ioctx, pls);
+    }
+
+    if (pls->is_id3_timestamped == -1)
+        pls->is_id3_timestamped = (pls->id3_mpegts_timestamp != AV_NOPTS_VALUE);
+}
+
 static int open_input(HLSContext *c, struct playlist *pls)
 {
     AVDictionary *opts = NULL;
@@ -680,7 +980,7 @@ static int open_input(HLSContext *c, struct playlist *pls)
     /* Seek to the requested position. If this was a HTTP request, the offset
      * should already be where want it to, but this allows e.g. local testing
      * without a HTTP server. */
-    if (ret == 0) {
+    if (ret == 0 && seg->key_type == KEY_NONE) {
         int seekret = ffurl_seek(pls->input, seg->url_offset, SEEK_SET);
         if (seekret < 0) {
             av_log(pls->parent, AV_LOG_ERROR, "Unable to seek to offset %"PRId64" of HLS segment '%s'\n", seg->url_offset, seg->url);
@@ -697,24 +997,47 @@ cleanup:
     return ret;
 }
 
+static int64_t default_reload_interval(struct playlist *pls)
+{
+    return pls->n_segments > 0 ?
+                          pls->segments[pls->n_segments - 1]->duration :
+                          pls->target_duration;
+}
+
 static int read_data(void *opaque, uint8_t *buf, int buf_size)
 {
     struct playlist *v = opaque;
     HLSContext *c = v->parent->priv_data;
     int ret, i;
-    int actual_read_size;
-    struct segment *seg;
+    int just_opened = 0;
 
+restart:
     if (!v->needed)
         return AVERROR_EOF;
 
-restart:
     if (!v->input) {
+        int64_t reload_interval;
+
+        /* Check that the playlist is still needed before opening a new
+         * segment. */
+        if (v->ctx && v->ctx->nb_streams &&
+            v->parent->nb_streams >= v->stream_offset + v->ctx->nb_streams) {
+            v->needed = 0;
+            for (i = v->stream_offset; i < v->stream_offset + v->ctx->nb_streams;
+                i++) {
+                if (v->parent->streams[i]->discard < AVDISCARD_ALL)
+                    v->needed = 1;
+            }
+        }
+        if (!v->needed) {
+            av_log(v->parent, AV_LOG_INFO, "No longer receiving playlist %d\n",
+                v->index);
+            return AVERROR_EOF;
+        }
+
         /* If this is a live stream and the reload interval has elapsed since
          * the last playlist reload, reload the playlists now. */
-        int64_t reload_interval = v->n_segments > 0 ?
-                                  v->segments[v->n_segments - 1]->duration :
-                                  v->target_duration;
+        reload_interval = default_reload_interval(v);
 
 reload:
         if (!v->finished &&
@@ -753,40 +1076,25 @@ reload:
                    v->index);
             return ret;
         }
+        just_opened = 1;
     }
-    /* limit read if the segment was only a part of a file */
-    seg = v->segments[v->cur_seq_no - v->start_seq_no];
-    if (seg->size >= 0)
-        actual_read_size = FFMIN(buf_size, seg->size - v->cur_seg_offset);
-    else
-        actual_read_size = buf_size;
 
-    ret = ffurl_read(v->input, buf, actual_read_size);
+    ret = read_from_url(v, buf, buf_size, READ_NORMAL);
     if (ret > 0) {
-        v->cur_seg_offset += ret;
+        if (just_opened && v->is_id3_timestamped != 0) {
+            /* Intercept ID3 tags here, elementary audio streams are required
+             * to convey timestamps using them in the beginning of each segment. */
+            intercept_id3(v, buf, buf_size, &ret);
+        }
+
         return ret;
     }
     ffurl_close(v->input);
     v->input = NULL;
     v->cur_seq_no++;
 
-    c->end_of_segment = 1;
     c->cur_seq_no = v->cur_seq_no;
 
-    if (v->ctx && v->ctx->nb_streams &&
-        v->parent->nb_streams >= v->stream_offset + v->ctx->nb_streams) {
-        v->needed = 0;
-        for (i = v->stream_offset; i < v->stream_offset + v->ctx->nb_streams;
-             i++) {
-            if (v->parent->streams[i]->discard < AVDISCARD_ALL)
-                v->needed = 1;
-        }
-    }
-    if (!v->needed) {
-        av_log(v->parent, AV_LOG_INFO, "No longer receiving playlist %d\n",
-               v->index);
-        return AVERROR_EOF;
-    }
     goto restart;
 }
 
@@ -863,6 +1171,71 @@ static void add_metadata_from_renditions(AVFormatContext *s, struct playlist *pl
     }
 }
 
+/* if timestamp was in valid range: returns 1 and sets seq_no
+ * if not: returns 0 and sets seq_no to closest segment */
+static int find_timestamp_in_playlist(HLSContext *c, struct playlist *pls,
+                                      int64_t timestamp, int *seq_no)
+{
+    int i;
+    int64_t pos = c->first_timestamp == AV_NOPTS_VALUE ?
+                  0 : c->first_timestamp;
+
+    if (timestamp < pos) {
+        *seq_no = pls->start_seq_no;
+        return 0;
+    }
+
+    for (i = 0; i < pls->n_segments; i++) {
+        int64_t diff = pos + pls->segments[i]->duration - timestamp;
+        if (diff > 0) {
+            *seq_no = pls->start_seq_no + i;
+            return 1;
+        }
+        pos += pls->segments[i]->duration;
+    }
+
+    *seq_no = pls->start_seq_no + pls->n_segments - 1;
+
+    return 0;
+}
+
+static int select_cur_seq_no(HLSContext *c, struct playlist *pls)
+{
+    int seq_no;
+
+    if (!pls->finished && !c->first_packet &&
+        av_gettime() - pls->last_load_time >= default_reload_interval(pls))
+        /* reload the playlist since it was suspended */
+        parse_playlist(c, pls->url, pls, NULL);
+
+    /* If playback is already in progress (we are just selecting a new
+     * playlist) and this is a complete file, find the matching segment
+     * by counting durations. */
+    if (pls->finished && c->cur_timestamp != AV_NOPTS_VALUE) {
+        find_timestamp_in_playlist(c, pls, c->cur_timestamp, &seq_no);
+        return seq_no;
+    }
+
+    if (!pls->finished) {
+        if (!c->first_packet && /* we are doing a segment selection during playback */
+            c->cur_seq_no >= pls->start_seq_no &&
+            c->cur_seq_no < pls->start_seq_no + pls->n_segments)
+            /* While spec 3.4.3 says that we cannot assume anything about the
+             * content at the same sequence number on different playlists,
+             * in practice this seems to work and doing it otherwise would
+             * require us to download a segment to inspect its timestamps. */
+            return c->cur_seq_no;
+
+        /* If this is a live stream with more than 3 segments, start at the
+         * third last segment. */
+        if (pls->n_segments > 3)
+            return pls->start_seq_no + pls->n_segments - 3;
+    }
+
+    /* Otherwise just start on the first segment. */
+    return pls->start_seq_no;
+}
+
 static int hls_read_header(AVFormatContext *s)
 {
     URLContext *u = (s->flags & AVFMT_FLAG_CUSTOM_IO) ? NULL : s->pb->opaque;
@@ -871,6 +1244,10 @@ static int hls_read_header(AVFormatContext *s)
 
     c->interrupt_callback = &s->interrupt_callback;
 
+    c->first_packet = 1;
+    c->first_timestamp = AV_NOPTS_VALUE;
+    c->cur_timestamp = AV_NOPTS_VALUE;
+
     // if the URL context is good, read important options we must broker later
     if (u && u->prot->priv_data_class) {
         // get the previous user agent & set back to null if string size is zero
@@ -953,12 +1330,7 @@ static int hls_read_header(AVFormatContext *s)
         pls->index  = i;
         pls->needed = 1;
         pls->parent = s;
-
-        /* If this is a live stream with more than 3 segments, start at the
-         * third last segment. */
-        pls->cur_seq_no = pls->start_seq_no;
-        if (!pls->finished && pls->n_segments > 3)
-            pls->cur_seq_no = pls->start_seq_no + pls->n_segments - 3;
+        pls->cur_seq_no = select_cur_seq_no(c, pls);
 
         pls->read_buffer = av_malloc(INITIAL_BUFFER_SIZE);
         ffio_init_context(&pls->pb, pls->read_buffer, INITIAL_BUFFER_SIZE, 0, pls,
@@ -982,11 +1354,21 @@ static int hls_read_header(AVFormatContext *s)
         if (ret < 0)
             goto fail;
 
+        if (pls->id3_deferred_extra && pls->ctx->nb_streams == 1) {
+            ff_id3v2_parse_apic(pls->ctx, &pls->id3_deferred_extra);
+            avformat_queue_attached_pictures(pls->ctx);
+            ff_id3v2_free_extra_meta(&pls->id3_deferred_extra);
+            pls->id3_deferred_extra = NULL;
+        }
+
         pls->ctx->ctx_flags &= ~AVFMTCTX_NOHEADER;
         ret = avformat_find_stream_info(pls->ctx, NULL);
         if (ret < 0)
             goto fail;
 
+        if (pls->is_id3_timestamped == -1)
+            av_log(s, AV_LOG_WARNING, "No expected HTTP requests have been made\n");
+
         /* Create new AVStreams for each stream in this playlist */
         for (j = 0; j < pls->ctx->nb_streams; j++) {
             AVStream *st = avformat_new_stream(s, NULL);
@@ -996,8 +1378,13 @@ static int hls_read_header(AVFormatContext *s)
                 goto fail;
             }
             st->id = i;
-            avpriv_set_pts_info(st, ist->pts_wrap_bits, ist->time_base.num, ist->time_base.den);
+
             avcodec_copy_context(st->codec, pls->ctx->streams[j]->codec);
+
+            if (pls->is_id3_timestamped) /* custom timestamps via id3 */
+                avpriv_set_pts_info(st, 33, 1, MPEG_TIME_BASE);
+            else
+                avpriv_set_pts_info(st, ist->pts_wrap_bits, ist->time_base.num, ist->time_base.den);
         }
 
         add_metadata_from_renditions(s, pls, AVMEDIA_TYPE_AUDIO);
@@ -1037,10 +1424,6 @@ static int hls_read_header(AVFormatContext *s)
         }
     }
 
-    c->first_packet = 1;
-    c->first_timestamp = AV_NOPTS_VALUE;
-    c->seek_timestamp  = AV_NOPTS_VALUE;
-
     return 0;
 fail:
     free_playlist_list(c);
@@ -1069,9 +1452,15 @@ static int recheck_discard_flags(AVFormatContext *s, int first)
         if (pls->cur_needed && !pls->needed) {
             pls->needed = 1;
             changed = 1;
-            pls->cur_seq_no = c->cur_seq_no;
+            pls->cur_seq_no = select_cur_seq_no(c, pls);
             pls->pb.eof_reached = 0;
-            av_log(s, AV_LOG_INFO, "Now receiving playlist %d\n", i);
+            if (c->cur_timestamp != AV_NOPTS_VALUE) {
+                /* catch up */
+                pls->seek_timestamp = c->cur_timestamp;
+                pls->seek_flags = AVSEEK_FLAG_ANY;
+                pls->seek_stream_index = -1;
+            }
+            av_log(s, AV_LOG_INFO, "Now receiving playlist %d, segment %d\n", i, pls->cur_seq_no);
         } else if (first && !pls->cur_needed && pls->needed) {
             if (pls->input)
                 ffurl_close(pls->input);
@@ -1084,18 +1473,55 @@ static int recheck_discard_flags(AVFormatContext *s, int first)
     return changed;
 }
 
+static void fill_timing_for_id3_timestamped_stream(struct playlist *pls)
+{
+    if (pls->id3_offset >= 0) {
+        pls->pkt.dts = pls->id3_mpegts_timestamp +
+                                 av_rescale_q(pls->id3_offset,
+                                              pls->ctx->streams[pls->pkt.stream_index]->time_base,
+                                              MPEG_TIME_BASE_Q);
+        if (pls->pkt.duration)
+            pls->id3_offset += pls->pkt.duration;
+        else
+            pls->id3_offset = -1;
+    } else {
+        /* there have been packets with unknown duration
+         * since the last id3 tag, should not normally happen */
+        pls->pkt.dts = AV_NOPTS_VALUE;
+    }
+
+    if (pls->pkt.duration)
+        pls->pkt.duration = av_rescale_q(pls->pkt.duration,
+                                         pls->ctx->streams[pls->pkt.stream_index]->time_base,
+                                         MPEG_TIME_BASE_Q);
+
+    pls->pkt.pts = AV_NOPTS_VALUE;
+}
+
+static AVRational get_timebase(struct playlist *pls)
+{
+    if (pls->is_id3_timestamped)
+        return MPEG_TIME_BASE_Q;
+
+    return pls->ctx->streams[pls->pkt.stream_index]->time_base;
+}
+
+static int compare_ts_with_wrapdetect(int64_t ts_a, struct playlist *pls_a,
+                                      int64_t ts_b, struct playlist *pls_b)
+{
+    int64_t scaled_ts_a = av_rescale_q(ts_a, get_timebase(pls_a), MPEG_TIME_BASE_Q);
+    int64_t scaled_ts_b = av_rescale_q(ts_b, get_timebase(pls_b), MPEG_TIME_BASE_Q);
+
+    return av_compare_mod(scaled_ts_a, scaled_ts_b, 1LL << 33);
+}
+
 static int hls_read_packet(AVFormatContext *s, AVPacket *pkt)
 {
     HLSContext *c = s->priv_data;
     int ret, i, minplaylist = -1;
 
-    if (c->first_packet) {
-        recheck_discard_flags(s, 1);
-        c->first_packet = 0;
-    }
+    recheck_discard_flags(s, c->first_packet);
 
-start:
-    c->end_of_segment = 0;
     for (i = 0; i < c->n_playlists; i++) {
         struct playlist *pls = c->playlists[i];
         /* Make sure we've got one buffered packet from each open playlist
@@ -1103,7 +1529,7 @@ start:
         if (pls->needed && !pls->pkt.data) {
             while (1) {
                 int64_t ts_diff;
-                AVStream *st;
+                AVRational tb;
                 ret = av_read_frame(pls->ctx, &pls->pkt);
                 if (ret < 0) {
                     if (!url_feof(&pls->pb) && ret != AVERROR_EOF)
@@ -1111,71 +1537,72 @@ start:
                     reset_packet(&pls->pkt);
                     break;
                 } else {
+                    /* stream_index check prevents matching picture attachments etc. */
+                    if (pls->is_id3_timestamped && pls->pkt.stream_index == 0) {
+                        /* audio elementary streams are id3 timestamped */
+                        fill_timing_for_id3_timestamped_stream(pls);
+                    }
+
                     if (c->first_timestamp == AV_NOPTS_VALUE &&
                         pls->pkt.dts       != AV_NOPTS_VALUE)
                         c->first_timestamp = av_rescale_q(pls->pkt.dts,
-                            pls->ctx->streams[pls->pkt.stream_index]->time_base,
-                            AV_TIME_BASE_Q);
+                            get_timebase(pls), AV_TIME_BASE_Q);
                 }
 
-                if (c->seek_timestamp == AV_NOPTS_VALUE)
+                if (pls->seek_timestamp == AV_NOPTS_VALUE)
                     break;
 
-                if (pls->pkt.dts == AV_NOPTS_VALUE) {
-                    c->seek_timestamp = AV_NOPTS_VALUE;
-                    break;
-                }
+                if (pls->seek_stream_index < 0 ||
+                    pls->seek_stream_index == pls->pkt.stream_index) {
 
-                st = pls->ctx->streams[pls->pkt.stream_index];
-                ts_diff = av_rescale_rnd(pls->pkt.dts, AV_TIME_BASE,
-                                         st->time_base.den, AV_ROUND_DOWN) -
-                          c->seek_timestamp;
-                if (ts_diff >= 0 && (c->seek_flags  & AVSEEK_FLAG_ANY ||
-                                     pls->pkt.flags & AV_PKT_FLAG_KEY)) {
-                    c->seek_timestamp = AV_NOPTS_VALUE;
-                    break;
+                    if (pls->pkt.dts == AV_NOPTS_VALUE) {
+                        pls->seek_timestamp = AV_NOPTS_VALUE;
+                        break;
+                    }
+
+                    tb = get_timebase(pls);
+                    ts_diff = av_rescale_rnd(pls->pkt.dts, AV_TIME_BASE,
+                                            tb.den, AV_ROUND_DOWN) -
+                            pls->seek_timestamp;
+                    if (ts_diff >= 0 && (pls->seek_flags  & AVSEEK_FLAG_ANY ||
+                                        pls->pkt.flags & AV_PKT_FLAG_KEY)) {
+                        pls->seek_timestamp = AV_NOPTS_VALUE;
+                        break;
+                    }
                 }
                 av_free_packet(&pls->pkt);
                 reset_packet(&pls->pkt);
             }
         }
-        /* Check if this stream still is on an earlier segment number, or
-         * has the packet with the lowest dts */
+        /* Check if this stream has the packet with the lowest dts */
         if (pls->pkt.data) {
             struct playlist *minpls = minplaylist < 0 ?
                                      NULL : c->playlists[minplaylist];
-            if (minplaylist < 0 || pls->cur_seq_no < minpls->cur_seq_no) {
+            if (minplaylist < 0) {
                 minplaylist = i;
-            } else if (pls->cur_seq_no == minpls->cur_seq_no) {
+            } else {
                 int64_t dts     =    pls->pkt.dts;
                 int64_t mindts  = minpls->pkt.dts;
-                AVStream *st    =    pls->ctx->streams[pls->pkt.stream_index];
-                AVStream *minst = minpls->ctx->streams[minpls->pkt.stream_index];
 
-                if (dts == AV_NOPTS_VALUE) {
+                if (dts == AV_NOPTS_VALUE ||
+                    (mindts != AV_NOPTS_VALUE && compare_ts_with_wrapdetect(dts, pls, mindts, minpls) < 0))
                     minplaylist = i;
-                } else if (mindts != AV_NOPTS_VALUE) {
-                    if (st->start_time    != AV_NOPTS_VALUE)
-                        dts    -= st->start_time;
-                    if (minst->start_time != AV_NOPTS_VALUE)
-                        mindts -= minst->start_time;
-
-                    if (av_compare_ts(dts, st->time_base,
-                                      mindts, minst->time_base) < 0)
-                        minplaylist = i;
-                }
             }
         }
     }
-    if (c->end_of_segment) {
-        if (recheck_discard_flags(s, 0))
-            goto start;
-    }
+
     /* If we got a packet, return it */
     if (minplaylist >= 0) {
-        *pkt = c->playlists[minplaylist]->pkt;
-        pkt->stream_index += c->playlists[minplaylist]->stream_offset;
+        struct playlist *pls = c->playlists[minplaylist];
+        *pkt = pls->pkt;
+        pkt->stream_index += pls->stream_offset;
         reset_packet(&c->playlists[minplaylist]->pkt);
+
+        if (pkt->dts != AV_NOPTS_VALUE)
+            c->cur_timestamp = av_rescale_q(pkt->dts,
+                                            pls->ctx->streams[pls->pkt.stream_index]->time_base,
+                                            AV_TIME_BASE_Q);
+
         return 0;
     }
     return AVERROR_EOF;
@@ -1195,32 +1622,49 @@ static int hls_read_seek(AVFormatContext *s, int stream_index,
                                int64_t timestamp, int flags)
 {
     HLSContext *c = s->priv_data;
-    int i, j, ret;
+    struct playlist *seek_pls = NULL;
+    int i, seq_no;
+    int64_t first_timestamp, seek_timestamp, duration;
 
-    if ((flags & AVSEEK_FLAG_BYTE) || !c->variants[0]->playlists[0]->finished)
+    if ((flags & AVSEEK_FLAG_BYTE) ||
+        !(c->variants[0]->playlists[0]->finished || c->variants[0]->playlists[0]->type == PLS_TYPE_EVENT))
         return AVERROR(ENOSYS);
 
-    c->seek_flags     = flags;
-    c->seek_timestamp = stream_index < 0 ? timestamp :
-                        av_rescale_rnd(timestamp, AV_TIME_BASE,
-                                       s->streams[stream_index]->time_base.den,
-                                       flags & AVSEEK_FLAG_BACKWARD ?
-                                       AV_ROUND_DOWN : AV_ROUND_UP);
-    timestamp = av_rescale_rnd(timestamp, AV_TIME_BASE, stream_index >= 0 ?
-                               s->streams[stream_index]->time_base.den :
-                               AV_TIME_BASE, flags & AVSEEK_FLAG_BACKWARD ?
-                               AV_ROUND_DOWN : AV_ROUND_UP);
-    if (s->duration < c->seek_timestamp) {
-        c->seek_timestamp = AV_NOPTS_VALUE;
+    first_timestamp = c->first_timestamp == AV_NOPTS_VALUE ?
+                      0 : c->first_timestamp;
+
+    seek_timestamp = av_rescale_rnd(timestamp, AV_TIME_BASE,
+                                    s->streams[stream_index]->time_base.den,
+                                    flags & AVSEEK_FLAG_BACKWARD ?
+                                    AV_ROUND_DOWN : AV_ROUND_UP);
+
+    duration = s->duration == AV_NOPTS_VALUE ?
+               0 : s->duration;
+
+    if (0 < duration && duration < seek_timestamp - first_timestamp)
         return AVERROR(EIO);
+
+    /* find the playlist with the specified stream */
+    for (i = 0; i < c->n_playlists; i++) {
+        struct playlist *pls = c->playlists[i];
+        if (stream_index >= pls->stream_offset &&
+            stream_index - pls->stream_offset < pls->ctx->nb_streams) {
+            seek_pls = pls;
+            break;
+        }
     }
+    /* check if the timestamp is valid for the playlist with the
+     * specified stream index */
+    if (!seek_pls || !find_timestamp_in_playlist(c, seek_pls, seek_timestamp, &seq_no))
+        return AVERROR(EIO);
+
+    /* set segment now so we do not need to search again below */
+    seek_pls->cur_seq_no = seq_no;
+    seek_pls->seek_stream_index = stream_index - seek_pls->stream_offset;
 
-    ret = AVERROR(EIO);
     for (i = 0; i < c->n_playlists; i++) {
         /* Reset reading */
         struct playlist *pls = c->playlists[i];
-        int64_t pos = c->first_timestamp == AV_NOPTS_VALUE ?
-                      0 : c->first_timestamp;
         if (pls->input) {
             ffurl_close(pls->input);
             pls->input = NULL;
@@ -1232,21 +1676,26 @@ static int hls_read_seek(AVFormatContext *s, int stream_index,
         pls->pb.buf_end = pls->pb.buf_ptr = pls->pb.buffer;
         /* Reset the pos, to let the mpegts demuxer know we've seeked. */
         pls->pb.pos = 0;
-
-        /* Locate the segment that contains the target timestamp */
-        for (j = 0; j < pls->n_segments; j++) {
-            if (timestamp >= pos &&
-                timestamp < pos + pls->segments[j]->duration) {
-                pls->cur_seq_no = pls->start_seq_no + j;
-                ret = 0;
-                break;
-            }
-            pos += pls->segments[j]->duration;
+        /* Flush the packet queue of the subdemuxer. */
+        ff_read_frame_flush(pls->ctx);
+
+        pls->seek_timestamp = seek_timestamp;
+        pls->seek_flags = flags;
+
+        if (pls != seek_pls) {
+            /* set closest segment seq_no for playlists not handled above */
+            find_timestamp_in_playlist(c, pls, seek_timestamp, &pls->cur_seq_no);
+            /* seek the playlist to the given position without taking
+             * keyframes into account since this playlist does not have the
+             * specified stream where we should look for the keyframes */
+            pls->seek_stream_index = -1;
+            pls->seek_flags |= AVSEEK_FLAG_ANY;
         }
-        if (ret)
-            c->seek_timestamp = AV_NOPTS_VALUE;
     }
-    return ret;
+
+    c->cur_timestamp = seek_timestamp;
+
+    return 0;
 }
 
 static int hls_probe(AVProbeData *p)