dragonflight/services/framecache/client/fc_client.c

/**
 * fc_client.c — Consumer-side framecache client implementation.
 */
#include "fc_client.h"
#include "../src/slot.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <fcntl.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <semaphore.h>
#include <time.h>
#include <unistd.h>

#define SHM_DIR    "/dev/shm/framecache"
#define SEM_PREFIX "/framecache-"
#define SEM_SUFFIX "-write"

struct fc_consumer {
    int      shm_fd;
    void    *base;
    size_t   shm_size;
    sem_t   *sem;
    uint64_t read_cursor;   /* consumer's own position in the ring */
    uint64_t local_dropped; /* frames skipped by this consumer */
    uint8_t *copy_buf;      /* consumer-owned copy buffer: [video frame_size][audio audio_max] */
    uint32_t frame_size;    /* cached from header (video bytes) */
    uint32_t audio_max;     /* cached from header (audio region capacity) */
    char     slot_id[FC_MAX_SLOT_ID];
};

static uint64_t now_us(void)
{
    struct timespec ts;
    clock_gettime(CLOCK_REALTIME, &ts);
    return (uint64_t)ts.tv_sec * 1000000ULL + ts.tv_nsec / 1000ULL;
}

fc_consumer_t *fc_consumer_open(const char *slot_id, uint64_t wait_ms)
{
    char shm_path[128], sem_name[128];
    snprintf(shm_path, sizeof shm_path, "%s/%s", SHM_DIR, slot_id);
    snprintf(sem_name, sizeof sem_name, "%s%s%s", SEM_PREFIX, slot_id, SEM_SUFFIX);

    uint64_t deadline = now_us() + wait_ms * 1000ULL;
    int fd = -1;
    while (1) {
        fd = open(shm_path, O_RDONLY);
        if (fd >= 0) break;
        if (now_us() >= deadline) return NULL;
        struct timespec ts = { .tv_nsec = 100000000 }; /* 100ms */
        nanosleep(&ts, NULL);
    }

    /* Read header to get frame_size */
    fc_header_t hdr;
    if (pread(fd, &hdr, sizeof hdr, 0) != sizeof hdr || hdr.magic != FC_MAGIC) {
        close(fd); return NULL;
    }
    /* Version gate: an old reader against a new writer (or vice-versa) computes
     * the wrong per-entry stride and would misparse. Fail safe. */
    if (hdr.version != FC_VERSION) {
        fprintf(stderr, "[fc_client] slot %s version %u != expected %u — refusing\n",
                slot_id, hdr.version, FC_VERSION);
        close(fd); return NULL;
    }
    size_t total = fc_slot_shm_size(hdr.frame_size);

    void *base = mmap(NULL, total, PROT_READ, MAP_SHARED, fd, 0);
    if (base == MAP_FAILED) { close(fd); return NULL; }

    sem_t *sem = sem_open(sem_name, 0);
    if (sem == SEM_FAILED) { munmap(base, total); close(fd); return NULL; }

    fc_consumer_t *c = calloc(1, sizeof *c);
    if (!c) { sem_close(sem); munmap(base, total); close(fd); return NULL; }

    /* Consumer-owned copy buffer — fc_consumer_read copies the frame here and
     * re-validates the cursor afterward, so a writer lapping a slow consumer
     * cannot corrupt the frame the caller is using. Sized for video + audio so
     * the frame-coupled audio is copied atomically with its video. */
    c->copy_buf = malloc((size_t)hdr.frame_size + hdr.audio_max_bytes);
    if (!c->copy_buf) {
        free(c); sem_close(sem); munmap(base, total); close(fd); return NULL;
    }

    c->shm_fd     = fd;
    c->base       = base;
    c->shm_size   = total;
    c->sem        = sem;
    c->frame_size = hdr.frame_size;
    c->audio_max  = hdr.audio_max_bytes;
    /* Start reading from the current write position so we don't replay old frames */
    c->read_cursor  = atomic_load_explicit(
        &((fc_header_t *)base)->write_cursor, memory_order_acquire);
    c->local_dropped = 0;
    strncpy(c->slot_id, slot_id, FC_MAX_SLOT_ID - 1);
    return c;
}

int fc_consumer_read(fc_consumer_t *c, fc_frame_ref_t *ref, uint64_t timeout_ms)
{
    fc_header_t *hdr = (fc_header_t *)c->base;
    int dropped = 0;   /* set when this call skipped one or more frames */

    /* ── Wait for new data ──────────────────────────────────────────────
     * The semaphore is used ONLY as an edge-wakeup hint, never as a frame
     * counter. The writer posts once per frame, but a consumer that skips
     * frames (lap) or reads less often than the writer posts would otherwise
     * leave the count climbing unbounded — causing sem_timedwait to never
     * block (100% CPU busy-spin) and eventually EOVERFLOW. So:
     *   - cursor-diff (write_cursor - read_cursor) is the SOURCE OF TRUTH for
     *     whether a frame is available.
     *   - we drain the semaphore to zero (sem_trywait loop) so the count never
     *     accumulates.
     *   - if no frame is available we block on ONE sem_timedwait for wakeup. */
    for (;;) {
        uint64_t write_cur = atomic_load_explicit(&hdr->write_cursor,
                                                   memory_order_acquire);

        /* Lap detection: if the writer is more than ring_depth ahead, the
         * oldest unread frames have been overwritten — skip to the oldest
         * still-valid frame. */
        if (write_cur > c->read_cursor + hdr->ring_depth) {
            uint64_t skipped = write_cur - c->read_cursor - hdr->ring_depth;
            c->read_cursor   = write_cur - hdr->ring_depth;
            c->local_dropped += skipped;
            /* NOTE: do NOT write hdr->dropped_frames here — the consumer maps
             * the shm PROT_READ (read-only), so an atomic write would SIGSEGV.
             * Per-consumer drops are tracked in c->local_dropped and exposed
             * via fc_consumer_dropped(). The writer owns hdr->dropped_frames. */
            dropped = 1;
        }

        if (c->read_cursor < write_cur) {
            /* A frame is available — drain the semaphore so its count never
             * accumulates, then read+copy below. */
            while (sem_trywait(c->sem) == 0) { /* drain */ }
            break;
        }

        /* No frame yet — drain stale posts, then block for a wakeup. */
        while (sem_trywait(c->sem) == 0) { /* drain */ }

        struct timespec abs_ts;
        clock_gettime(CLOCK_REALTIME, &abs_ts);
        abs_ts.tv_sec  += (time_t)(timeout_ms / 1000);
        abs_ts.tv_nsec += (long)((timeout_ms % 1000) * 1000000L);
        if (abs_ts.tv_nsec >= 1000000000L) { abs_ts.tv_sec++; abs_ts.tv_nsec -= 1000000000L; }

        int w = sem_timedwait(c->sem, &abs_ts);
        if (w != 0) {
            if (errno == ETIMEDOUT) {
                /* Re-check the cursor once more before giving up — the writer
                 * may have advanced between our check and the wait. */
                uint64_t wc2 = atomic_load_explicit(&hdr->write_cursor,
                                                    memory_order_acquire);
                if (c->read_cursor < wc2) continue;
                return FC_TIMEOUT;
            }
            if (errno == EINTR) continue;
            return FC_ERROR;
        }
        /* Woken — loop to re-evaluate cursor-diff. */
    }

    /* ── Copy the entry (video + this frame's audio) into the owned buffer ─
     * Both are copied from the SAME ring entry in the SAME iteration, so the
     * audio handed to the caller is exactly this video frame's embedded audio —
     * frame-coupled, no second buffer to drift. copy_buf layout mirrors the
     * shm entry: [video frame_size][audio audio_max]. */
    fc_frame_t *frame = fc_frame_at(c->base, hdr->frame_size, c->read_cursor);
    uint32_t fsz = frame->size;
    if (fsz > hdr->frame_size) fsz = hdr->frame_size;
    uint32_t asz = frame->audio_size;
    if (asz > c->audio_max) asz = c->audio_max;
    uint64_t pts  = frame->pts_us;
    uint64_t wall = frame->wall_us;
    memcpy(c->copy_buf, frame->data, fsz);
    if (asz)
        memcpy(c->copy_buf + c->frame_size,
               fc_frame_audio(frame, hdr->frame_size), asz);

    /* ── Re-validate AFTER the copy ─────────────────────────────────────
     * If the writer lapped us during the copy (overwrote this slot), the copy
     * may be torn — discard it and signal DROPPED so the caller reads again. */
    uint64_t write_after = atomic_load_explicit(&hdr->write_cursor,
                                                memory_order_acquire);
    if (write_after > c->read_cursor + hdr->ring_depth) {
        uint64_t skipped = write_after - c->read_cursor - hdr->ring_depth;
        c->read_cursor   = write_after - hdr->ring_depth;
        c->local_dropped += skipped;
        return FC_LAPPED;   /* copy torn — ref not valid, caller reads again */
    }

    /* Copy is valid. */
    ref->data       = c->copy_buf;
    ref->size       = fsz;
    ref->audio      = asz ? (c->copy_buf + c->frame_size) : NULL;
    ref->audio_size = asz;
    ref->pts_us     = pts;
    ref->wall_us    = wall;
    ref->seq        = c->read_cursor;

    c->read_cursor++;
    return dropped ? FC_DROPPED : FC_OK;
}

void fc_consumer_close(fc_consumer_t *c)
{
    if (!c) return;
    if (c->copy_buf) free(c->copy_buf);
    sem_close(c->sem);
    munmap(c->base, c->shm_size);
    close(c->shm_fd);
    free(c);
}

uint64_t fc_consumer_write_cursor(fc_consumer_t *c)
{
    fc_header_t *hdr = (fc_header_t *)c->base;
    return atomic_load(&hdr->write_cursor);
}

uint64_t fc_consumer_dropped(fc_consumer_t *c)
{
    return c->local_dropped;
}