Merge branch 't/makefile_cleanups'

[paraslash.git] / vss.c

diff --git a/vss.c b/vss.c
index b9afc8ed637374c82918f3448695a1a2fdaa94a2..4a8aafa8d6e0c2ae239d09342ac10f3b08263828 100644 (file)
--- a/vss.c
+++ b/vss.c
@@ -132,6 +132,7 @@ struct fec_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 */
@@ -360,8 +361,17 @@ static void vss_get_chunk(int chunk_num, struct vss_task *vsst,
 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;
        /*
@@ -371,8 +381,6 @@ static void compute_group_size(struct vss_task *vsst, struct fec_group *g,
         * of exactly one chunk for these audio formats.
         */
        for (i = 0;; i++) {
-               char *buf;
-               size_t len;
                int chunk_num = g->first_chunk + i;
 
                if (g->bytes > 0 && i >= max_chunks) /* duration limit */
@@ -501,7 +509,7 @@ static int setup_next_fec_group(struct fec_client *fc, struct vss_task *vsst)
 {
        int ret, i, k, n, data_slices;
        size_t len;
-       char *buf;
+       char *buf, *p;
        struct fec_group *g = &fc->group;
 
        if (fc->state == FEC_STATE_NONE) {
@@ -561,16 +569,20 @@ static int setup_next_fec_group(struct fec_client *fc, struct vss_task *vsst)
                assert(i == g->num_header_slices - 1);
        }
 
-       /* setup data slices */
+       /*
+        * 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 (; i < g->num_header_slices + data_slices; i++) {
-               if (buf + g->slice_bytes > vsst->map + mmd->size) {
+       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.
+                        * We must make a copy for this slice since using p
+                        * directly would exceed the buffer.
                         */
-                       uint32_t payload_size = vsst->map + mmd->size - buf;
-                       memcpy(fc->extra_src_buf, buf, payload_size);
+                       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);
@@ -578,8 +590,8 @@ static int setup_next_fec_group(struct fec_client *fc, struct vss_task *vsst)
                        i++;
                        break;
                }
-               fc->src_data[i] = (const unsigned char *)buf;
-               buf += g->slice_bytes;
+               fc->src_data[i] = (const unsigned char *)p;
+               p += g->slice_bytes;
        }
        if (i < k) {
                /* use arbitrary data for all remaining slices */
@@ -952,6 +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;
@@ -986,8 +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);
@@ -1074,6 +1090,25 @@ static void vss_send(struct vss_task *vsst)
                        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++;
        }