ff-decode 0.14.2

use super::{
    AvFrameGuard, DecodeError, Duration, KEYFRAME_SEEK_TOLERANCE_SECS, VideoDecoderInner,
    VideoFrame,
};

impl VideoDecoderInner {
    /// Returns the current playback position.
    pub(crate) fn position(&self) -> Duration {
        self.position
    }

    /// Returns whether end of file has been reached.
    pub(crate) fn is_eof(&self) -> bool {
        self.eof
    }

    /// Returns whether the source is a live or streaming input.
    ///
    /// Live sources have the `AVFMT_TS_DISCONT` flag set on their `AVInputFormat`.
    /// Seeking is not meaningful on live sources.
    pub(crate) fn is_live(&self) -> bool {
        self.is_live
    }

    /// Converts a `Duration` to a presentation timestamp (PTS) in stream time_base units.
    ///
    /// # Arguments
    ///
    /// * `duration` - The duration to convert.
    ///
    /// # Returns
    ///
    /// The timestamp in stream time_base units.
    ///
    /// # Note
    ///
    /// av_seek_frame expects timestamps in stream time_base units when using a specific stream_index.
    fn duration_to_pts(&self, duration: Duration) -> i64 {
        // Convert duration to stream time_base units for seeking
        // SAFETY:
        // - format_ctx is valid: owned by VideoDecoderInner, initialized in constructor via avformat_open_input
        // - stream_index is valid: validated during decoder creation (find_stream_info + codec opening)
        // - streams array access is valid: guaranteed by FFmpeg after successful avformat_open_input
        let time_base = unsafe {
            let stream = (*self.format_ctx).streams.add(self.stream_index as usize);
            (*(*stream)).time_base
        };

        // Convert: duration (seconds) * (time_base.den / time_base.num) = PTS
        let time_base_f64 = time_base.den as f64 / time_base.num as f64;
        (duration.as_secs_f64() * time_base_f64) as i64
    }

    /// Converts a presentation timestamp (PTS) to a `Duration`.
    ///
    /// # Arguments
    ///
    /// * `pts` - The presentation timestamp in stream time base units.
    ///
    /// # Returns
    ///
    /// The duration corresponding to the PTS.
    ///
    /// # Safety
    ///
    /// Caller must ensure that `format_ctx` and `stream_index` are valid.
    /// Seeks to a specified position in the video stream.
    ///
    /// This method performs efficient seeking without reopening the file.
    /// It uses `av_seek_frame` internally and flushes the decoder buffers.
    ///
    /// # Performance Characteristics
    ///
    /// - **Keyframe seek**: 5-10ms for typical GOP sizes (1-2 seconds)
    /// - **Exact seek**: Proportional to distance from nearest keyframe
    /// - **Large GOP videos**: May require sequential decoding from distant keyframe
    ///
    /// For videos with sparse keyframes (GOP > 2 seconds), the method will
    /// decode frames sequentially from the nearest keyframe to reach the target.
    /// This ensures correct frame data but may take longer (10-50ms for very large GOPs).
    ///
    /// # Arguments
    ///
    /// * `position` - Target position to seek to.
    /// * `mode` - Seek mode (Keyframe, Exact, or Backward).
    ///
    /// # Errors
    ///
    /// Returns [`DecodeError::SeekFailed`] if the seek operation fails.
    pub(crate) fn seek(
        &mut self,
        position: Duration,
        mode: crate::SeekMode,
    ) -> Result<(), DecodeError> {
        use crate::SeekMode;

        let timestamp = self.duration_to_pts(position);

        // All seek modes use BACKWARD flag to find the nearest keyframe at or before target.
        // The difference between modes is in the post-seek processing below.
        let flags = ff_sys::avformat::seek_flags::BACKWARD;

        // 1. Clear any pending packet and frame to avoid reading stale data after seek
        // SAFETY:
        // - packet is valid: allocated in constructor, owned by VideoDecoderInner
        // - frame is valid: allocated in constructor, owned by VideoDecoderInner
        unsafe {
            ff_sys::av_packet_unref(self.packet);
            ff_sys::av_frame_unref(self.frame);
        }

        // 2. Seek in the format context (file is NOT reopened)
        // Use av_seek_frame with the stream index and timestamp in stream time_base units
        // SAFETY:
        // - format_ctx is valid: owned by VideoDecoderInner, initialized via avformat_open_input
        // - stream_index is valid: validated during decoder creation
        // - timestamp is valid: converted from Duration using stream's time_base
        unsafe {
            ff_sys::avformat::seek_frame(
                self.format_ctx,
                self.stream_index as i32,
                timestamp,
                flags,
            )
            .map_err(|e| DecodeError::SeekFailed {
                target: position,
                reason: ff_sys::av_error_string(e),
            })?;
        }

        // 3. Flush decoder buffers to clear any cached frames
        // SAFETY: codec_ctx is valid: owned by VideoDecoderInner, initialized via avcodec_open2
        unsafe {
            ff_sys::avcodec::flush_buffers(self.codec_ctx);
        }

        // 4. Drain any remaining frames from the decoder after flush
        // This ensures no stale frames are returned after the seek
        // SAFETY:
        // - codec_ctx is valid: owned by VideoDecoderInner, initialized via avcodec_open2
        // - frame is valid: allocated in constructor, owned by VideoDecoderInner
        unsafe {
            loop {
                let ret = ff_sys::avcodec_receive_frame(self.codec_ctx, self.frame);
                if ret == ff_sys::error_codes::EAGAIN || ret == ff_sys::error_codes::EOF {
                    // No more frames in the decoder buffer
                    break;
                } else if ret == 0 {
                    // Got a frame, unref it and continue draining
                    ff_sys::av_frame_unref(self.frame);
                } else {
                    // Other error, break out
                    break;
                }
            }
        }

        // 5. Reset internal state
        self.eof = false;
        // Note: We don't update self.position here because it will be updated
        // when the next frame is decoded. This ensures position reflects actual decoded position.

        // 6. Skip forward to the target position
        //
        // Context: av_seek_frame with BACKWARD flag seeks to the nearest keyframe *at or before*
        // the target timestamp. For videos with sparse keyframes (large GOP size), this may
        // land far from the target (e.g., at the first keyframe for GOP=entire video).
        //
        // Solution: Decode frames sequentially from the keyframe until reaching the target.
        // This is necessary because H.264/H.265 P-frames and B-frames depend on previous
        // frames for reconstruction, so we must decode all intermediate frames.
        //
        // Performance Impact:
        // - Typical GOP (1-2s): 30-60 frames to skip, ~5-10ms overhead
        // - Large GOP (5-10s): 150-300 frames to skip, ~20-50ms overhead
        // - Worst case (single keyframe): May decode entire video, ~100ms-1s
        if mode == SeekMode::Exact {
            // For exact mode, decode until we reach or pass the exact target
            self.skip_to_exact(position)?;
        } else {
            // For keyframe/backward modes, decode until we're reasonably close to the target
            // Rationale: Balances accuracy with performance for common use cases
            let tolerance = Duration::from_secs(KEYFRAME_SEEK_TOLERANCE_SECS);
            let min_position = position.saturating_sub(tolerance);

            while let Some(frame) = self.decode_one()? {
                let frame_time = frame.timestamp().as_duration();
                if frame_time >= min_position {
                    // We're close enough to the target
                    break;
                }
                // Continue decoding to get closer (frames are automatically dropped)
            }
        }

        Ok(())
    }

    /// Skips frames until reaching the exact target position.
    ///
    /// This is used by [`Self::seek`] when `SeekMode::Exact` is specified.
    /// It decodes and discards frames from the nearest keyframe until
    /// reaching the target position.
    ///
    /// # Performance
    ///
    /// Time complexity is O(n) where n is the number of frames between the
    /// keyframe and target. For a 30fps video with 2-second GOP:
    /// - Worst case: ~60 frames to decode, ~10-20ms
    /// - Average case: ~30 frames to decode, ~5-10ms
    ///
    /// # Arguments
    ///
    /// * `target` - The exact target position.
    ///
    /// # Errors
    ///
    /// Returns [`DecodeError::SeekFailed`] if EOF is reached before the target position.
    fn skip_to_exact(&mut self, target: Duration) -> Result<(), DecodeError> {
        loop {
            match self.decode_one()? {
                Some(frame) => {
                    let frame_time = frame.timestamp().as_duration();
                    if frame_time >= target {
                        // Reached or passed the target frame
                        // Position will be updated by decode_one() which was just called
                        break;
                    }
                    // Continue decoding (frame is automatically dropped)
                }
                None => {
                    // Reached EOF before finding target frame
                    return Err(DecodeError::SeekFailed {
                        target,
                        reason: "Reached end of stream before target position".to_string(),
                    });
                }
            }
        }
        Ok(())
    }

    /// Flushes the decoder's internal buffers.
    ///
    /// This clears any cached frames and resets the decoder state.
    /// The decoder is ready to receive new packets after flushing.
    pub(crate) fn flush(&mut self) {
        // SAFETY: codec_ctx is valid and owned by this instance
        unsafe {
            ff_sys::avcodec::flush_buffers(self.codec_ctx);
        }
        self.eof = false;
    }

    /// Scales a video frame to the specified dimensions while preserving aspect ratio.
    ///
    /// This method uses SwScale to resize frames efficiently using a "fit-within"
    /// strategy that preserves the original aspect ratio.
    ///
    /// # Aspect Ratio Preservation
    ///
    /// The frame is scaled to fit within `(target_width, target_height)` while
    /// maintaining its original aspect ratio. The output dimensions will be at most
    /// the target size, with at least one dimension matching the target. No letterboxing
    /// or pillarboxing is applied - the frame is simply scaled down to fit.
    ///
    /// # Arguments
    ///
    /// * `frame` - The source frame to scale.
    /// * `target_width` - Desired width in pixels.
    /// * `target_height` - Desired height in pixels.
    ///
    /// # Returns
    ///
    /// A new `VideoFrame` scaled to fit within the target dimensions.
    ///
    /// # Errors
    ///
    /// Returns [`DecodeError`] if SwScale context creation or scaling fails.
    ///
    /// # Performance
    ///
    /// - Caches SwScale context for repeated calls with same dimensions
    /// - Context creation: ~0.1-0.5ms (only on first call or dimension change)
    /// - Typical scaling time: 1-3ms for 1080p → 320x180
    /// - Uses bilinear interpolation for quality/performance balance
    ///
    /// # Cache Behavior
    ///
    /// The SwScale context is cached based on source/target dimensions and format.
    /// When generating multiple thumbnails with the same size (e.g., via `thumbnails()`),
    /// the context is reused, eliminating the ~0.1-0.5ms creation overhead per thumbnail.
    pub(crate) fn scale_frame(
        &mut self,
        frame: &VideoFrame,
        target_width: u32,
        target_height: u32,
    ) -> Result<VideoFrame, DecodeError> {
        let src_width = frame.width();
        let src_height = frame.height();
        let src_format = frame.format();

        // Calculate scaled dimensions to preserve aspect ratio (fit within target)
        let src_aspect = src_width as f64 / src_height as f64;
        let target_aspect = target_width as f64 / target_height as f64;

        let (scaled_width, scaled_height) = if src_aspect > target_aspect {
            // Source is wider - fit to width
            let height = (target_width as f64 / src_aspect).round() as u32;
            (target_width, height)
        } else {
            // Source is taller or equal - fit to height
            let width = (target_height as f64 * src_aspect).round() as u32;
            (width, target_height)
        };

        // Convert pixel format to FFmpeg format
        let av_format = Self::pixel_format_to_av(src_format);

        // Cache key: (src_width, src_height, scaled_width, scaled_height, format)
        let cache_key = (
            src_width,
            src_height,
            scaled_width,
            scaled_height,
            av_format,
        );

        // SAFETY: We're creating temporary FFmpeg objects for scaling
        unsafe {
            // Check if we can reuse the cached SwScale context
            let (sws_ctx, is_cached) = if let (Some(cached_ctx), Some(cached_key)) =
                (self.thumbnail_sws_ctx, self.thumbnail_cache_key)
            {
                if cached_key == cache_key {
                    // Cache hit - reuse existing context
                    (cached_ctx, true)
                } else {
                    // Cache miss - free old context and create new one
                    ff_sys::swscale::free_context(cached_ctx);
                    // Clear cache immediately to prevent dangling pointer
                    self.thumbnail_sws_ctx = None;
                    self.thumbnail_cache_key = None;

                    let new_ctx = ff_sys::swscale::get_context(
                        src_width as i32,
                        src_height as i32,
                        av_format,
                        scaled_width as i32,
                        scaled_height as i32,
                        av_format,
                        ff_sys::swscale::scale_flags::BILINEAR,
                    )
                    .map_err(|e| DecodeError::Ffmpeg {
                        code: 0,
                        message: format!("Failed to create scaling context: {e}"),
                    })?;

                    // Don't cache yet - will cache after successful scaling
                    (new_ctx, false)
                }
            } else {
                // No cache - create new context
                let new_ctx = ff_sys::swscale::get_context(
                    src_width as i32,
                    src_height as i32,
                    av_format,
                    scaled_width as i32,
                    scaled_height as i32,
                    av_format,
                    ff_sys::swscale::scale_flags::BILINEAR,
                )
                .map_err(|e| DecodeError::Ffmpeg {
                    code: 0,
                    message: format!("Failed to create scaling context: {e}"),
                })?;

                // Don't cache yet - will cache after successful scaling
                (new_ctx, false)
            };

            // Set up source frame with VideoFrame data
            let src_frame_guard = AvFrameGuard::new()?;
            let src_frame = src_frame_guard.as_ptr();

            (*src_frame).width = src_width as i32;
            (*src_frame).height = src_height as i32;
            (*src_frame).format = av_format;

            // Set up source frame data pointers directly from VideoFrame (no copy)
            let planes = frame.planes();
            let strides = frame.strides();

            for (i, plane_data) in planes.iter().enumerate() {
                if i >= ff_sys::AV_NUM_DATA_POINTERS as usize {
                    break;
                }
                (*src_frame).data[i] = plane_data.as_ref().as_ptr().cast_mut();
                (*src_frame).linesize[i] = strides[i] as i32;
            }

            // Allocate destination frame
            let dst_frame_guard = AvFrameGuard::new()?;
            let dst_frame = dst_frame_guard.as_ptr();

            (*dst_frame).width = scaled_width as i32;
            (*dst_frame).height = scaled_height as i32;
            (*dst_frame).format = av_format;

            // Allocate buffer for destination frame
            let buffer_ret = ff_sys::av_frame_get_buffer(dst_frame, 0);
            if buffer_ret < 0 {
                // Clean up context if not cached
                if !is_cached {
                    ff_sys::swscale::free_context(sws_ctx);
                }
                return Err(DecodeError::Ffmpeg {
                    code: buffer_ret,
                    message: format!(
                        "Failed to allocate destination frame buffer: {}",
                        ff_sys::av_error_string(buffer_ret)
                    ),
                });
            }

            // Perform scaling
            let scale_result = ff_sys::swscale::scale(
                sws_ctx,
                (*src_frame).data.as_ptr() as *const *const u8,
                (*src_frame).linesize.as_ptr(),
                0,
                src_height as i32,
                (*dst_frame).data.as_ptr() as *const *mut u8,
                (*dst_frame).linesize.as_ptr(),
            );

            if let Err(e) = scale_result {
                // Clean up context if not cached
                if !is_cached {
                    ff_sys::swscale::free_context(sws_ctx);
                }
                return Err(DecodeError::Ffmpeg {
                    code: 0,
                    message: format!("Failed to scale frame: {e}"),
                });
            }

            // Scaling successful - cache the context if it's new
            if !is_cached {
                self.thumbnail_sws_ctx = Some(sws_ctx);
                self.thumbnail_cache_key = Some(cache_key);
            }

            // Copy timestamp
            (*dst_frame).pts = frame.timestamp().pts();

            // Convert destination frame to VideoFrame
            let video_frame = self.av_frame_to_video_frame(dst_frame)?;

            Ok(video_frame)
        }
    }

    // ── Reconnect helpers ─────────────────────────────────────────────────────

    /// Attempts to reconnect to the stream URL using exponential backoff.
    ///
    /// Called from `decode_one()` when `StreamInterrupted` is received and
    /// `NetworkOptions::reconnect_on_error` is `true`. After all attempts fail,
    /// returns a `StreamInterrupted` error.
    pub(super) fn attempt_reconnect(&mut self) -> Result<(), DecodeError> {
        let url = match self.url.as_deref() {
            Some(u) => u.to_owned(),
            None => return Ok(()), // file-path source: no reconnect
        };
        let max = self.network_opts.max_reconnect_attempts;

        for attempt in 1..=max {
            let backoff_ms = 100u64 * (1u64 << (attempt - 1).min(10));
            log::warn!(
                "reconnecting attempt={attempt} url={} backoff_ms={backoff_ms}",
                crate::network::sanitize_url(&url)
            );
            std::thread::sleep(Duration::from_millis(backoff_ms));
            match self.reopen(&url) {
                Ok(()) => {
                    self.reconnect_count += 1;
                    log::info!(
                        "reconnected attempt={attempt} url={} total_reconnects={}",
                        crate::network::sanitize_url(&url),
                        self.reconnect_count
                    );
                    return Ok(());
                }
                Err(e) => log::warn!("reconnect attempt={attempt} failed err={e}"),
            }
        }

        Err(DecodeError::StreamInterrupted {
            code: 0,
            endpoint: crate::network::sanitize_url(&url),
            message: format!("stream did not recover after {max} attempts"),
        })
    }

    /// Closes the current `AVFormatContext`, re-opens the URL, re-reads stream info,
    /// re-finds the video stream, and flushes the codec.
    fn reopen(&mut self, url: &str) -> Result<(), DecodeError> {
        // Close the current format context. `avformat_close_input` sets the pointer
        // to null — this matches the null check in Drop so no double-free occurs.
        // SAFETY: self.format_ctx is valid and owned exclusively by self.
        unsafe {
            ff_sys::avformat::close_input(std::ptr::addr_of_mut!(self.format_ctx));
        }

        // Re-open the URL with the stored network timeouts.
        // SAFETY: url is a valid UTF-8 network URL string.
        self.format_ctx = unsafe {
            ff_sys::avformat::open_input_url(
                url,
                self.network_opts.connect_timeout,
                self.network_opts.read_timeout,
            )
            .map_err(|e| crate::network::map_network_error(e, crate::network::sanitize_url(url)))?
        };

        // Re-read stream information.
        // SAFETY: self.format_ctx is valid and freshly opened.
        unsafe {
            ff_sys::avformat::find_stream_info(self.format_ctx).map_err(|e| {
                DecodeError::Ffmpeg {
                    code: e,
                    message: format!(
                        "reconnect find_stream_info failed: {}",
                        ff_sys::av_error_string(e)
                    ),
                }
            })?;
        }

        // Re-find the video stream (index may differ in theory after reconnect).
        // SAFETY: self.format_ctx is valid.
        let (stream_index, _) = unsafe { Self::find_video_stream(self.format_ctx) }
            .ok_or_else(|| DecodeError::NoVideoStream { path: url.into() })?;
        self.stream_index = stream_index as i32;

        // Flush codec buffers to discard stale decoded state from before the drop.
        // SAFETY: self.codec_ctx is valid and has not been freed.
        unsafe {
            ff_sys::avcodec::flush_buffers(self.codec_ctx);
        }

        self.eof = false;
        Ok(())
    }
}