GithubFollow/One-KVM
1
0
Fork 0
mirror of https://github.com/mofeng-git/One-KVM.git synced 2026-09-13 02:54:26 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

perf(video): 新增 RKMPP DMA 采集编码通路并完善恢复逻辑

Browse Source 
支持原生 HDMI 和 UVC 缓冲区导出，增加同步 RKMPP 编码及可选 MJPEG 硬件转码。
校验帧布局和缓冲区租约，在 DMA 不可用或编码失败时回退到复制通路。
保留自定义码率和 GOP 策略，重开采集时同步 HDMI 源帧率，并区分 UVC 超时状态。

验证：88 个视频测试通过（含 4 个新增回归测试）；ARM64 cargo check --tests 通过。
This commit is contained in:
mofeng-git
2026-09-05 20:55:30 +08:00
parent 620fe0be54
commit db9d79554a
11 changed files with 1680 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
libs/hwcodec/build.rs
View File

@@ -488,6 +488,7 @@ mod ffmpeg {
}
}
builder.file(ffmpeg_hw_dir.join("ffmpeg_hw_mjpeg_h26x.cpp"));
builder.file(ffmpeg_hw_dir.join("rkmpp_dmabuf.cpp"));
} else {
println!(
"cargo:info=Skipping ffmpeg_hw_mjpeg_h26x.cpp (RKMPP) for arch {}",

38
libs/hwcodec/cpp/ffmpeg_hw/rkmpp_dma_jpeg.h Normal file
View File

@@ -0,0 +1,38 @@
#pragma once
#include <stddef.h>
#include <stdint.h>
// Validate bounded JPEG headers before giving hardware a fixed-size output
// buffer. Only baseline 8-bit JPEG is accepted; other streams use copy fallback.
// No scan-data traversal or full-packet copy is needed.
inline bool rkmpp_dma_jpeg_header(const uint8_t *data, size_t size, int width, int height) {
if (!data || size < 4 || data[0] != 0xff || data[1] != 0xd8) return false;
size_t pos = 2;
bool sof = false;
while (pos < size) {
if (data[pos++] != 0xff) return false;
while (pos < size && data[pos] == 0xff) ++pos;
if (pos == size) return false;
const unsigned marker = data[pos++];
if (!marker || marker == 0xd8 || marker == 0xd9 || marker == 1 ||
(marker >= 0xd0 && marker <= 0xd7)) return false;
if (size - pos < 2) return false;
const size_t length = (size_t(data[pos]) << 8) | data[pos + 1];
if (length < 2 || length > size - pos) return false;
if (marker == 0xc0) {
if (sof || length < 8 || data[pos + 2] != 8) return false;
const unsigned h = (unsigned(data[pos + 3]) << 8) | data[pos + 4];
const unsigned w = (unsigned(data[pos + 5]) << 8) | data[pos + 6];
const unsigned components = data[pos + 7];
if (w != unsigned(width) || h != unsigned(height) ||
(components != 1 && components != 3) || length != 8 + 3 * components) return false;
sof = true;
} else if (marker >= 0xc0 && marker <= 0xcf && marker != 0xc4 && marker != 0xcc) {
return false; // Progressive, lossless, extended or differential SOF.
} else if (marker == 0xda) {
return sof && length >= 6 && size - pos > length;
}
pos += length;
}
return false;
}

329
libs/hwcodec/cpp/ffmpeg_hw/rkmpp_dmabuf.cpp Normal file
View File

@@ -0,0 +1,329 @@
#include "rkmpp_dmabuf_ffi.h"
#include <array>
#include <cstdio>
#include <limits>
#include <new>
#include "rkmpp_dma_jpeg.h"
// Native MPP is already linked by the ARM FFmpeg/RKMPP build. Keep this
// optional for toolchains which only supply the FFmpeg headers.
#if defined(__linux__) && __has_include(<rockchip/rk_mpi.h>)
#define HAVE_MPP_DMA 1
#include <cerrno>
#include <linux/dma-buf.h>
#include <sys/ioctl.h>
extern "C" {
#include <rockchip/rk_mpi.h>
#include <rockchip/mpp_buffer.h>
#include <rockchip/mpp_frame.h>
#include <rockchip/mpp_packet.h>
#include <rockchip/mpp_task.h>
#include <rockchip/rk_venc_cfg.h>
#include <rockchip/rk_venc_rc.h>
}
#endif
static thread_local char dma_error[192] = {};
static int fail(const char *operation, int code) {
std::snprintf(dma_error, sizeof(dma_error), "%s (ret=%d)", operation, code);
return -1;
}
#ifdef HAVE_MPP_DMA
struct RkmppDmaEncoder {
MppCtx ctx = nullptr;
MppApi *api = nullptr;
MppEncCfg cfg = nullptr;
MppPacket packet = nullptr;
std::array<MppBuffer, 16> buffers{};
std::array<size_t, 16> capacities{};
size_t count = 0;
size_t minimum = 0;
int width = 0, height = 0, stride = 0;
MppFrameFormat format = MPP_FMT_YUV420SP;
bool jpeg = false;
MppCtx decoder = nullptr;
MppApi *dec_api = nullptr;
MppBufferGroup decoded_group = nullptr;
MppBuffer decoded_buffer = nullptr;
MppFrame decoded_frame = nullptr;
MppPacket input_packet = nullptr;
bool decoded_layout_set = false;
int decoded_stride = 0, decoded_vstride = 0;
void close() {
if (packet) mpp_packet_deinit(&packet);
if (ctx) {
// A timeout must not expose still-in-use input to V4L2 QBUF.
api->reset(ctx);
mpp_destroy(ctx);
ctx = nullptr;
}
// On any decoder failure, end hardware access before releasing the
// input packet, exported buffers, or allowing the caller's QBUF.
if (decoder) {
dec_api->reset(decoder);
mpp_destroy(decoder);
decoder = nullptr;
}
if (input_packet) mpp_packet_deinit(&input_packet);
if (decoded_frame) mpp_frame_deinit(&decoded_frame);
if (decoded_buffer) { mpp_buffer_put(decoded_buffer); decoded_buffer = nullptr; }
if (decoded_group) { mpp_buffer_group_put(decoded_group); decoded_group = nullptr; }
for (auto &buffer : buffers) {
if (buffer) { mpp_buffer_put(buffer); buffer = nullptr; }
}
if (cfg) { mpp_enc_cfg_deinit(cfg); cfg = nullptr; }
}
~RkmppDmaEncoder() { close(); }
};
static bool set_cfg(RkmppDmaEncoder *e, const char *key, int value) {
int ret = mpp_enc_cfg_set_s32(e->cfg, key, value);
if (ret) fail(key, ret);
return ret == 0;
}
extern "C" int rkmpp_dma_reconfigure(RkmppDmaEncoder *e, int kbps, int gop) {
if (!e || !e->ctx || kbps <= 0 || kbps > 1000000 || gop <= 0)
return fail("invalid DMA encoder configuration", -1);
const int bps = kbps * 1000;
if (!set_cfg(e, "rc:bps_target", bps) ||
!set_cfg(e, "rc:bps_max", bps + bps / 16) ||
!set_cfg(e, "rc:bps_min", bps - bps / 16) ||
!set_cfg(e, "rc:gop", gop)) return -1;
int ret = e->api->control(e->ctx, MPP_ENC_SET_CFG, e->cfg);
return ret ? fail("MPP_ENC_SET_CFG", ret) : 0;
}
extern "C" RkmppDmaEncoder *rkmpp_dma_new(
int width, int height, int stride, int format, int codec, int fps,
int kbps, int gop, const int *fds, const size_t *sizes, size_t count) {
if (width <= 0 || height <= 0 || width > 8192 || height > 8192 ||
(width & 1) || (height & 1) || format < 0 || format > 4 ||
codec < 0 || codec > 1 || fps <= 0 || fps > 240 ||
(format != 4 && stride < width * (format == 1 || format == 3 ? 3 : format == 2 ? 2 : 1)) ||
(format == 2 && stride % 16 != 0) || !fds || !sizes || !count || count > 16) {
fail("invalid DMA frame layout", -1); return nullptr;
}
auto *e = new (std::nothrow) RkmppDmaEncoder;
if (!e) { fail("allocate DMA encoder", -1); return nullptr; }
e->width = width; e->height = height; e->stride = stride; e->count = count;
e->jpeg = format == 4;
if (e->jpeg) e->stride = stride = (width + 15) & ~15;
const int vstride = e->jpeg ? (height + 15) & ~15 : height;
e->format = format == 3 ? MPP_FMT_RGB888 : format == 2 ? MPP_FMT_YUV422_YUYV : format == 1 ? MPP_FMT_BGR888 : MPP_FMT_YUV420SP;
auto abort_init = [e](const char *op, int ret) -> RkmppDmaEncoder * {
fail(op, ret); delete e; return nullptr;
};
int ret = mpp_create(&e->ctx, &e->api);
if (ret) return abort_init("mpp_create", ret);
RK_S64 timeout = 2000;
ret = e->api->control(e->ctx, MPP_SET_OUTPUT_TIMEOUT, &timeout);
if (ret) return abort_init("MPP_SET_OUTPUT_TIMEOUT", ret);
ret = e->api->control(e->ctx, MPP_SET_INPUT_TIMEOUT, &timeout);
if (ret) return abort_init("MPP_SET_INPUT_TIMEOUT", ret);
ret = mpp_init(e->ctx, MPP_CTX_ENC, codec ? MPP_VIDEO_CodingHEVC : MPP_VIDEO_CodingAVC);
if (ret) return abort_init("mpp_init", ret);
ret = mpp_enc_cfg_init(&e->cfg);
if (ret) return abort_init("mpp_enc_cfg_init", ret);
ret = e->api->control(e->ctx, MPP_ENC_GET_CFG, e->cfg);
if (ret) return abort_init("MPP_ENC_GET_CFG", ret);
if (!set_cfg(e, "prep:width", width) || !set_cfg(e, "prep:height", height) ||
!set_cfg(e, "prep:hor_stride", stride) || !set_cfg(e, "prep:ver_stride", vstride) ||
!set_cfg(e, "prep:format", e->format) || !set_cfg(e, "rc:mode", MPP_ENC_RC_MODE_CBR) ||
!set_cfg(e, "rc:fps_in_flex", 0) || !set_cfg(e, "rc:fps_in_num", fps) ||
!set_cfg(e, "rc:fps_in_denorm", 1) || !set_cfg(e, "rc:fps_out_flex", 0) ||
!set_cfg(e, "rc:fps_out_num", fps) || !set_cfg(e, "rc:fps_out_denorm", 1) ||
!set_cfg(e, "codec:type", codec ? MPP_VIDEO_CodingHEVC : MPP_VIDEO_CodingAVC)) {
delete e; return nullptr;
}
// Match the browser-friendly baseline profile used by the existing RKMPP
// byte encoder, rather than inheriting MPP's High-profile default.
const int level = int64_t(width) * height * fps <= int64_t(1920) * 1080 * 60 ? 42 : 52;
if (!codec && (!set_cfg(e, "h264:profile", 66) || !set_cfg(e, "h264:level", level) ||
!set_cfg(e, "h264:cabac_en", 0) || !set_cfg(e, "h264:trans8x8", 0))) {
delete e; return nullptr;
}
if (rkmpp_dma_reconfigure(e, kbps, gop)) { delete e; return nullptr; }
MppEncHeaderMode mode = MPP_ENC_HEADER_MODE_EACH_IDR;
ret = e->api->control(e->ctx, MPP_ENC_SET_HEADER_MODE, &mode);
if (ret) return abort_init("MPP_ENC_SET_HEADER_MODE", ret);
// Reject arithmetic overflow even on 32-bit ARM; stride is supplied by a driver.
if (size_t(stride) > std::numeric_limits<size_t>::max() / size_t(height))
return abort_init("DMA buffer size overflow", -1);
size_t minimum = size_t(stride) * height;
if (format == 0) {
if (minimum > std::numeric_limits<size_t>::max() / 3)
return abort_init("DMA buffer size overflow", -1);
minimum = minimum * 3 / 2;
}
if (e->jpeg) minimum = 68; // SOI/payload plus bounded hardware read headroom.
e->minimum = minimum;
for (size_t i = 0; i < count; ++i) {
if (fds[i] < 0 || sizes[i] < minimum) return abort_init("short DMA buffer", -1);
MppBufferInfo info{};
info.type = MPP_BUFFER_TYPE_EXT_DMA; info.fd = fds[i];
info.size = sizes[i]; info.index = static_cast<int>(i);
ret = mpp_buffer_import(&e->buffers[i], &info);
if (ret) return abort_init("mpp_buffer_import", ret);
e->capacities[i] = sizes[i];
}
if (e->jpeg) {
ret = mpp_create(&e->decoder, &e->dec_api);
if (ret) return abort_init("mpp_create JPEG decoder", ret);
ret = mpp_init(e->decoder, MPP_CTX_DEC, MPP_VIDEO_CodingMJPEG);
if (ret) return abort_init("mpp_init JPEG decoder", ret);
MppFrameFormat output = MPP_FMT_YUV420SP;
ret = e->dec_api->control(e->decoder, MPP_DEC_SET_OUTPUT_FORMAT, &output);
if (ret) return abort_init("JPEG NV12 output", ret);
ret = mpp_buffer_group_get_internal(&e->decoded_group, MPP_BUFFER_TYPE_DRM);
if (ret) return abort_init("JPEG output buffer group", ret);
// MPP JPEG requires aligned storage; reserve the conservative size used
// by its advanced-task decoder demo. One output reused after encode.
ret = mpp_buffer_get(e->decoded_group, &e->decoded_buffer, size_t(stride) * vstride * 4);
if (ret) return abort_init("JPEG output buffer", ret);
ret = mpp_frame_init(&e->decoded_frame);
if (ret) return abort_init("JPEG output frame", ret);
mpp_frame_set_buffer(e->decoded_frame, e->decoded_buffer);
}
return e;
}
static int dma_read_sync(MppBuffer buffer, bool start) {
dma_buf_sync sync{};
sync.flags = DMA_BUF_SYNC_READ | (start ? DMA_BUF_SYNC_START : DMA_BUF_SYNC_END);
int ret;
do { ret = ioctl(mpp_buffer_get_fd(buffer), DMA_BUF_IOCTL_SYNC, &sync); }
while (ret < 0 && errno == EINTR);
return ret;
}
static int decode_jpeg(RkmppDmaEncoder *e, size_t index, size_t bytes_used) {
MppBuffer input = e->buffers[index];
// The parser reads only header bytes with explicit DMA CPU-read ownership.
if (dma_read_sync(input, true)) return fail("JPEG DMA read sync start", errno);
const auto *data = static_cast<const uint8_t *>(mpp_buffer_get_ptr(input));
const bool valid = rkmpp_dma_jpeg_header(data, bytes_used, e->width, e->height);
if (dma_read_sync(input, false)) return fail("JPEG DMA read sync end", errno);
if (!valid) return fail("unsupported/mismatched JPEG header", -1);
int ret = mpp_packet_init_with_buffer(&e->input_packet, input);
if (ret) return fail("JPEG input packet", ret);
mpp_packet_set_length(e->input_packet, bytes_used);
MppTask task = nullptr;
ret = e->dec_api->poll(e->decoder, MPP_PORT_INPUT, static_cast<MppPollType>(2000));
if (ret) return fail("JPEG input poll", ret);
ret = e->dec_api->dequeue(e->decoder, MPP_PORT_INPUT, &task);
if (ret || !task) return fail("JPEG input task", ret);
ret = mpp_task_meta_set_packet(task, KEY_INPUT_PACKET, e->input_packet);
if (ret) return fail("JPEG input metadata", ret);
ret = mpp_task_meta_set_frame(task, KEY_OUTPUT_FRAME, e->decoded_frame);
if (ret) return fail("JPEG output metadata", ret);
ret = e->dec_api->enqueue(e->decoder, MPP_PORT_INPUT, task);
if (ret) return fail("JPEG submit", ret);
task = nullptr;
ret = e->dec_api->poll(e->decoder, MPP_PORT_OUTPUT, static_cast<MppPollType>(2000));
if (ret) return fail("JPEG output poll", ret);
ret = e->dec_api->dequeue(e->decoder, MPP_PORT_OUTPUT, &task);
if (ret || !task) return fail("JPEG output task", ret);
MppFrame result = nullptr;
ret = mpp_task_meta_get_frame(task, KEY_OUTPUT_FRAME, &result);
if (ret || result != e->decoded_frame) return fail("JPEG output frame mismatch", ret);
ret = e->dec_api->enqueue(e->decoder, MPP_PORT_OUTPUT, task);
if (ret) return fail("JPEG return output task", ret);
ret = e->dec_api->poll(e->decoder, MPP_PORT_INPUT, static_cast<MppPollType>(2000));
if (ret) return fail("JPEG input completion", ret);
mpp_packet_deinit(&e->input_packet);
if (mpp_frame_get_errinfo(result) || mpp_frame_get_discard(result) ||
mpp_frame_get_info_change(result) ||
mpp_frame_get_width(result) != unsigned(e->width) ||
mpp_frame_get_height(result) != unsigned(e->height) ||
mpp_frame_get_fmt(result) != MPP_FMT_YUV420SP ||
mpp_frame_get_buffer(result) != e->decoded_buffer) {
std::snprintf(dma_error, sizeof(dma_error),
"invalid JPEG decoded frame: err=%u discard=%u info_change=%u size=%ux%u fmt=%x buffer_match=%d",
mpp_frame_get_errinfo(result), mpp_frame_get_discard(result), mpp_frame_get_info_change(result),
mpp_frame_get_width(result), mpp_frame_get_height(result), unsigned(mpp_frame_get_fmt(result)),
int(mpp_frame_get_buffer(result) == e->decoded_buffer));
return -1;
}
const int hs = mpp_frame_get_hor_stride(result), vs = mpp_frame_get_ver_stride(result);
if (hs < e->width || vs < e->height || hs > 8192 || vs > 8192 || (hs & 15) || (vs & 15) ||
size_t(hs) * vs * 3 / 2 > mpp_buffer_get_size(e->decoded_buffer))
return fail("invalid JPEG decoded stride", -1);
if (!e->decoded_layout_set) {
if (!set_cfg(e, "prep:hor_stride", hs) || !set_cfg(e, "prep:ver_stride", vs)) return -1;
ret = e->api->control(e->ctx, MPP_ENC_SET_CFG, e->cfg);
if (ret) return fail("JPEG encoder layout", ret);
e->decoded_stride = hs; e->decoded_vstride = vs; e->decoded_layout_set = true;
} else if (hs != e->decoded_stride || vs != e->decoded_vstride) {
return fail("JPEG decoded layout changed", -1);
}
return 0;
}
extern "C" int rkmpp_dma_encode(RkmppDmaEncoder *e, size_t index, size_t bytes_used, int fresh_fd, int64_t pts_us,
int force_idr, const uint8_t **data, size_t *size) {
if (!e || !e->ctx || index >= e->count || !data || !size)
return fail("invalid DMA encode call", -1);
*data = nullptr; *size = 0;
if (e->packet) mpp_packet_deinit(&e->packet);
auto abort_encode = [e](const char *op, int ret) {
fail(op, ret); e->close(); return -1;
};
if (bytes_used > e->capacities[index] ||
(e->jpeg ? (bytes_used < 4 || e->capacities[index] - bytes_used < 64) : bytes_used != e->minimum))
return abort_encode("invalid DMA payload length", -1);
if (fresh_fd >= 0) {
MppBuffer replacement = nullptr;
MppBufferInfo info{};
info.type = MPP_BUFFER_TYPE_EXT_DMA; info.fd = fresh_fd;
info.size = e->capacities[index]; info.index = static_cast<int>(index);
const int ret = mpp_buffer_import(&replacement, &info);
if (ret) return abort_encode("refresh USB DMA import", ret);
if (e->buffers[index]) mpp_buffer_put(e->buffers[index]);
e->buffers[index] = replacement;
}
if (e->jpeg && decode_jpeg(e, index, bytes_used)) {
// Preserve the detailed decoder failure while ending BOTH engines.
e->close(); return -1;
}
if (force_idr) {
int ret = e->api->control(e->ctx, MPP_ENC_SET_IDR_FRAME, nullptr);
if (ret) return abort_encode("MPP_ENC_SET_IDR_FRAME", ret);
}
MppFrame frame = e->jpeg ? e->decoded_frame : nullptr;
int ret = 0;
if (!e->jpeg) {
ret = mpp_frame_init(&frame);
if (ret) return abort_encode("mpp_frame_init", ret);
mpp_frame_set_width(frame, e->width); mpp_frame_set_height(frame, e->height);
mpp_frame_set_hor_stride(frame, e->stride); mpp_frame_set_ver_stride(frame, e->height);
mpp_frame_set_fmt(frame, e->format);
mpp_frame_set_buffer(frame, e->buffers[index]);
}
mpp_frame_set_pts(frame, pts_us);
ret = e->api->encode_put_frame(e->ctx, frame);
if (!e->jpeg) mpp_frame_deinit(&frame);
if (ret) return abort_encode("encode_put_frame", ret);
ret = e->api->encode_get_packet(e->ctx, &e->packet);
if (ret || !e->packet) return abort_encode("encode_get_packet", ret);
if (mpp_packet_is_partition(e->packet) || !mpp_packet_get_length(e->packet))
return abort_encode("incomplete DMA encoder output", -1);
// One synchronous input, no temporal scalability/reordering/split output.
// A completed packet is the input-consumption barrier for this mode.
*data = static_cast<const uint8_t *>(mpp_packet_get_pos(e->packet));
*size = mpp_packet_get_length(e->packet);
return 0;
}
extern "C" void rkmpp_dma_free(RkmppDmaEncoder *e) { delete e; }
#else
extern "C" RkmppDmaEncoder *rkmpp_dma_new(int,int,int,int,int,int,int,int,const int*,const size_t*,size_t) {
fail("RKMPP DMA support not built", -1); return nullptr;
}
extern "C" int rkmpp_dma_encode(RkmppDmaEncoder*,size_t,size_t,int,int64_t,int,const uint8_t**,size_t*) { return -1; }
extern "C" int rkmpp_dma_reconfigure(RkmppDmaEncoder*,int,int) { return -1; }
extern "C" void rkmpp_dma_free(RkmppDmaEncoder*) {}
#endif
extern "C" const char *rkmpp_dma_error(void) { return dma_error; }

31
libs/hwcodec/cpp/ffmpeg_hw/rkmpp_dmabuf_ffi.h Normal file
View File

@@ -0,0 +1,31 @@
#pragma once
#include <stddef.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
typedef struct RkmppDmaEncoder RkmppDmaEncoder;
// format: 0=NV12, 1=BGR24, 2=YUYV, 3=RGB24 (byte stride), 4=MJPEG (stride ignored).
// codec: 0=H264, 1=HEVC. MJPEG is decoded to hardware NV12, then encoded.
RkmppDmaEncoder *rkmpp_dma_new(int width, int height, int stride, int format,
int codec, int fps, int kbps, int gop,
const int *fds, const size_t *sizes, size_t count);
// Synchronous input-completion boundary. On failure the encoder is destroyed
// internally BEFORE returning, so it cannot keep reading the capture buffer.
// Output is borrowed until the next call or destruction; copy before reusing it.
// bytes_used must be the actual captured payload. MJPEG needs 64 bytes of
// readable allocation headroom; never pass sizeimage as the compressed length.
// fresh_fd=-1 reuses the original import (native HDMI). UVC supplies a new
// export of this dequeued slot. Keep it open until replacement/free; the old
// export can be closed AFTER this call, including on failure.
int rkmpp_dma_encode(RkmppDmaEncoder *encoder, size_t index, size_t bytes_used, int fresh_fd, int64_t pts_us,
int force_idr, const uint8_t **data, size_t *size);
int rkmpp_dma_reconfigure(RkmppDmaEncoder *encoder, int kbps, int gop);
void rkmpp_dma_free(RkmppDmaEncoder *encoder);
const char *rkmpp_dma_error(void);
#ifdef __cplusplus
}
#endif

5
libs/hwcodec/src/lib.rs
View File

@@ -5,6 +5,11 @@ pub mod ffmpeg;
#[cfg(any(target_arch = "aarch64", target_arch = "arm", feature = "rkmpp"))]
pub mod ffmpeg_hw;
pub mod ffmpeg_ram;
#[cfg(all(
target_os = "linux",
any(target_arch = "aarch64", target_arch = "arm", feature = "rkmpp")
))]
pub mod rkmpp_dmabuf;
#[no_mangle]
pub extern "C" fn hwcodec_log(level: i32, message: *const std::os::raw::c_char) {

193
libs/hwcodec/src/rkmpp_dmabuf.rs Normal file
View File

@@ -0,0 +1,193 @@
//! Synchronous RKMPP encoder for pre-exported V4L2 DMA buffers.
//! Unlike the byte-slice encoder this never reads raw pixels on the CPU.
use std::ffi::{c_char, c_int, c_void, CStr};
use std::os::fd::{AsRawFd, OwnedFd};
use std::ptr::NonNull;
unsafe extern "C" {
fn rkmpp_dma_new(
width: c_int,
height: c_int,
stride: c_int,
format: c_int,
codec: c_int,
fps: c_int,
kbps: c_int,
gop: c_int,
fds: *const c_int,
sizes: *const usize,
count: usize,
) -> *mut c_void;
fn rkmpp_dma_encode(
encoder: *mut c_void,
index: usize,
bytes_used: usize,
fresh_fd: c_int,
pts_us: i64,
force_idr: c_int,
data: *mut *const u8,
size: *mut usize,
) -> c_int;
fn rkmpp_dma_reconfigure(encoder: *mut c_void, kbps: c_int, gop: c_int) -> c_int;
fn rkmpp_dma_free(encoder: *mut c_void);
fn rkmpp_dma_error() -> *const c_char;
}
#[derive(Debug, Clone, Copy)]
#[repr(i32)]
pub enum DmaFormat {
Nv12 = 0,
Bgr24 = 1,
Yuyv = 2,
Rgb24 = 3,
Mjpeg = 4,
}
pub struct DmaEncoderConfig {
pub width: u32,
pub height: u32,
pub stride: u32,
pub format: DmaFormat,
pub hevc: bool,
pub fps: u32,
pub bitrate_kbps: u32,
pub gop: u32,
}
pub struct DmaEncoder {
ctx: NonNull<c_void>,
// Export FDs remain open until AFTER mpp_destroy and imported buffer release.
_buffers: Vec<(OwnedFd, usize)>,
// Keep refreshed exports alive until native replacement/release has ended
// all references to the previous import. At most one FD per capture slot.
fresh_buffers: Vec<Option<OwnedFd>>,
}
// Exclusive ownership: the context can move between threads but all calls are sequential.
unsafe impl Send for DmaEncoder {}
fn last_error() -> String {
unsafe {
CStr::from_ptr(rkmpp_dma_error())
.to_string_lossy()
.into_owned()
}
}
impl DmaEncoder {
pub fn new(config: DmaEncoderConfig, buffers: Vec<(OwnedFd, usize)>) -> Result<Self, String> {
let fds: Vec<_> = buffers.iter().map(|(fd, _)| fd.as_raw_fd()).collect();
let sizes: Vec<_> = buffers.iter().map(|(_, size)| *size).collect();
for value in [
config.width,
config.height,
config.stride,
config.fps,
config.bitrate_kbps,
config.gop,
] {
if value > c_int::MAX as u32 {
return Err("DMA encoder parameter overflow".into());
}
}
let ptr = unsafe {
rkmpp_dma_new(
config.width as _,
config.height as _,
config.stride as _,
config.format as c_int,
config.hevc as _,
config.fps as _,
config.bitrate_kbps as _,
config.gop as _,
fds.as_ptr(),
sizes.as_ptr(),
buffers.len(),
)
};
Ok(Self {
ctx: NonNull::new(ptr).ok_or_else(last_error)?,
fresh_buffers: (0..buffers.len()).map(|_| None).collect(),
_buffers: buffers,
})
}
/// # Safety
/// The indexed buffer must be dequeued and exclusively leased to this call.
/// Do not requeue/write it until this function returns. On failure native MPP
/// is synchronously destroyed before returning, ending all input access.
/// `bytes_used` is the actual DQBUF payload length, not the buffer capacity.
/// A refreshed FD, if supplied, must refer to the same leased capture slot
/// with the capacity registered at construction. Ownership is retained here.
pub unsafe fn encode(
&mut self,
index: usize,
bytes_used: usize,
fresh_fd: Option<OwnedFd>,
pts_ms: i64,
force_idr: bool,
) -> Result<Vec<u8>, String> {
let mut data = std::ptr::null();
let mut size = 0;
if index >= self.fresh_buffers.len() {
return Err("Invalid DMA capture index".into());
}
let ret = unsafe {
rkmpp_dma_encode(
self.ctx.as_ptr(),
index,
bytes_used,
fresh_fd.as_ref().map_or(-1, AsRawFd::as_raw_fd),
pts_ms.saturating_mul(1000),
force_idr as _,
&mut data,
&mut size,
)
};
if fresh_fd.is_some() {
// Native has now released the previous import, or destroyed both
// hardware contexts on error. Only now may its old FD be closed.
self.fresh_buffers[index] = fresh_fd;
}
if ret != 0 {
return Err(last_error());
}
if data.is_null() || size == 0 {
return Err("Empty RKMPP DMA packet".into());
}
// Copy only the compressed output, releasing the driver's packet promptly
// regardless of how long a network subscriber retains its Bytes.
Ok(unsafe { std::slice::from_raw_parts(data, size) }.to_vec())
}
pub fn reconfigure(&mut self, kbps: u32, gop: u32) -> Result<(), String> {
if kbps > c_int::MAX as u32 || gop > c_int::MAX as u32 {
return Err("DMA encoder parameter overflow".into());
}
if unsafe { rkmpp_dma_reconfigure(self.ctx.as_ptr(), kbps as _, gop as _) } != 0 {
return Err(last_error());
}
Ok(())
}
}
impl Drop for DmaEncoder {
fn drop(&mut self) {
unsafe { rkmpp_dma_free(self.ctx.as_ptr()) };
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn format_discriminants_match_native_abi() {
assert_eq!(DmaFormat::Nv12 as c_int, 0);
assert_eq!(DmaFormat::Bgr24 as c_int, 1);
assert_eq!(DmaFormat::Yuyv as c_int, 2);
assert_eq!(DmaFormat::Rgb24 as c_int, 3);
assert_eq!(DmaFormat::Mjpeg as c_int, 4);
}
}

201
src/video/capture/dmabuf_layout.rs Normal file
View File

@@ -0,0 +1,201 @@
//! Conservative, dependency-free eligibility checks for linear RKMPP input.
pub struct DmaCaptureLayout<'a> {
pub native_hdmi: bool,
pub driver: &'a str,
pub bus_info: &'a str,
pub configurable_usb: bool,
pub single_planar: bool,
pub fourcc: [u8; 4],
pub width: u32,
pub height: u32,
pub stride: u32,
}
impl DmaCaptureLayout<'_> {
/// Minimum readable bytes, not the driver's page-aligned allocation size.
pub fn minimum_bytes(&self) -> Option<usize> {
if self.width == 0
|| self.height == 0
|| self.width > 8192
|| self.height > 8192
|| self.width % 2 != 0
|| self.height % 2 != 0
{
return None;
}
let bytes_per_row = if self.native_hdmi {
match &self.fourcc {
b"NV12" => self.width,
b"BGR3" => self.width.checked_mul(3)?,
_ => return None,
}
} else if self.configurable_usb
&& self.single_planar
&& self.driver == "uvcvideo"
&& self.bus_info.starts_with("usb-")
{
match &self.fourcc {
// Compressed frames have variable bytesused and no byte stride.
b"MJPG" => return Some(4),
b"YUYV" if self.stride % 16 == 0 => self.width.checked_mul(2)?,
b"NV12" if self.stride % 16 == 0 => self.width,
b"RGB3" if self.stride % 16 == 0 => self.width.checked_mul(3)?,
_ => return None,
}
} else {
return None;
};
if self.stride < bytes_per_row {
return None;
}
let size = (self.stride as usize).checked_mul(self.height as usize)?;
if self.fourcc == *b"NV12" {
size.checked_mul(3)?.checked_div(2)
} else {
Some(size)
}
}
}
/// An MJPEG DMA packet has a bounded payload, not stride * height bytes.
/// Reserve readable headroom for the MPP bitstream reader without modifying
/// capture memory. The decoder receives only `used`, never the allocation size.
pub fn valid_payload(
compressed: bool,
used: usize,
capacity: usize,
expected: Option<usize>,
) -> bool {
if compressed {
used >= 4 && used.checked_add(64).is_some_and(|end| end <= capacity)
} else {
used > 0 && Some(used) == expected && used <= capacity
}
}
#[cfg(test)]
mod tests {
use super::*;
fn usb() -> DmaCaptureLayout<'static> {
DmaCaptureLayout {
native_hdmi: false,
driver: "uvcvideo",
bus_info: "usb-fc880000.usb-1.1",
configurable_usb: true,
single_planar: true,
fourcc: *b"YUYV",
width: 1920,
height: 1080,
stride: 3840,
}
}
#[test]
fn usb_yuyv_uses_byte_stride_and_supports_padding() {
let mut layout = usb();
assert_eq!(layout.minimum_bytes(), Some(4_147_200));
layout.width = 640;
layout.height = 480;
layout.stride = 1280;
assert_eq!(layout.minimum_bytes(), Some(614_400));
layout.stride = 1296;
assert_eq!(layout.minimum_bytes(), Some(622_080));
}
#[test]
fn usb_requires_correct_driver_bus_queue_and_control_mode() {
let mut layout = usb();
layout.driver = "rkcif";
assert_eq!(layout.minimum_bytes(), None);
layout = usb();
layout.bus_info = "platform:hdmi";
assert_eq!(layout.minimum_bytes(), None);
layout = usb();
layout.single_planar = false;
assert_eq!(layout.minimum_bytes(), None);
layout = usb();
layout.configurable_usb = false;
assert_eq!(layout.minimum_bytes(), None);
}
#[test]
fn unverified_usb_formats_stay_on_copy_path() {
for fourcc in [
*b"H264", *b"NV21", *b"NV16", *b"NV24", *b"BGR3", *b"YU12", *b"UYVY", *b"YVYU",
*b"BAD!",
] {
let mut layout = usb();
layout.fourcc = fourcc;
assert_eq!(layout.minimum_bytes(), None, "{fourcc:?}");
}
}
#[test]
fn usb_nv12_rgb_and_mjpeg_layouts() {
let mut layout = usb();
layout.fourcc = *b"NV12";
layout.stride = 1920;
assert_eq!(layout.minimum_bytes(), Some(3_110_400));
layout.fourcc = *b"RGB3";
layout.stride = 5760;
assert_eq!(layout.minimum_bytes(), Some(6_220_800));
layout.stride = 1920;
assert_eq!(layout.minimum_bytes(), None);
layout.fourcc = *b"MJPG";
layout.stride = 0;
assert_eq!(layout.minimum_bytes(), Some(4));
}
#[test]
fn compressed_payload_is_bounded_and_not_allocation_size() {
assert!(valid_payload(true, 63163, 4147200, None));
for used in [0, 3, 4147200, usize::MAX] {
assert!(!valid_payload(true, used, 4147200, None));
}
assert!(valid_payload(true, 4, 68, None));
assert!(!valid_payload(true, 4, 67, None));
assert!(valid_payload(false, 614400, 614400, Some(614400)));
assert!(!valid_payload(false, 614399, 614400, Some(614400)));
assert!(!valid_payload(false, 614400, 614399, Some(614400)));
}
#[test]
fn malformed_geometry_or_stride_is_rejected() {
for (w, h, stride) in [
(0, 1080, 3840),
(1920, 0, 3840),
(1919, 1080, 3840),
(1920, 1079, 3840),
(8194, 1080, 16384),
(1920, 8194, 3840),
(1920, 1080, 0),
(1920, 1080, 1920),
(1920, 1080, 3841),
(u32::MAX, u32::MAX, u32::MAX),
] {
let mut layout = usb();
layout.width = w;
layout.height = h;
layout.stride = stride;
assert_eq!(layout.minimum_bytes(), None);
}
}
#[test]
fn native_hdmi_formats_are_preserved_but_not_expanded() {
let mut layout = usb();
layout.native_hdmi = true;
layout.single_planar = false;
layout.fourcc = *b"BGR3";
layout.stride = 5760;
assert_eq!(layout.minimum_bytes(), Some(6_220_800));
layout.fourcc = *b"NV12";
layout.stride = 1920;
assert_eq!(layout.minimum_bytes(), Some(3_110_400));
layout.fourcc = *b"YUYV";
layout.stride = 3840;
assert_eq!(layout.minimum_bytes(), None);
}
}

230
src/video/capture/linux.rs
View File

@@ -3,6 +3,8 @@
use std::fs::File;
use std::io;
use std::os::fd::AsFd;
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
use std::os::fd::OwnedFd;
use std::os::unix::fs::OpenOptionsExt;
use std::path::{Path, PathBuf};
use std::time::{Duration, Instant};
@@ -29,6 +31,10 @@ use crate::video::device::VideoControlMode;
use crate::video::format::{PixelFormat, Resolution};
use crate::video::signal::SignalStatus;
#[cfg(any(test, target_arch = "aarch64", target_arch = "arm"))]
#[path = "dmabuf_layout.rs"]
mod dmabuf_layout;
/// Metadata for a captured frame.
#[derive(Debug, Clone, Copy)]
pub struct CaptureMeta {
@@ -67,6 +73,8 @@ pub struct CaptureStream {
bridge_kind: Option<CsiBridgeKind>,
native_hdmirx_state: Option<NativeHdmirxState>,
native_hdmirx_next_state_check: Option<Instant>,
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
dma_layout_bytes: Option<usize>,
}
fn open_capture_device(path: &Path) -> io::Result<File> {
@@ -319,6 +327,24 @@ impl CaptureStream {
mappings.push(plane_maps);
}
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
let dma_layout_bytes = PixelFormat::from_v4l2r(actual_fmt.pixelformat).and_then(|format| {
dmabuf_layout::DmaCaptureLayout {
native_hdmi: is_native_hdmirx,
driver: &caps.driver,
bus_info: &caps.bus_info,
configurable_usb: !is_source_following
&& bridge.kind.is_none()
&& !bridge.has_subdev(),
single_planar: queue == QueueType::VideoCapture,
fourcc: format.to_fourcc(),
width: actual_resolution.width,
height: actual_resolution.height,
stride,
}
.minimum_bytes()
});
let mut stream = Self {
fd,
queue,
@@ -332,6 +358,8 @@ impl CaptureStream {
bridge_kind: bridge.kind,
native_hdmirx_state,
native_hdmirx_next_state_check,
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
dma_layout_bytes,
};
stream.queue_all_buffers()?;
@@ -421,10 +449,7 @@ impl CaptureStream {
}
}
pub fn next_into(
&mut self,
dst: &mut Vec<u8>,
) -> std::result::Result<CaptureMeta, CaptureReadError> {
fn dequeue_buffer(&mut self) -> std::result::Result<V4l2Buffer, CaptureReadError> {
self.wait_ready()?;
// Several vendor BSPs update G_FMT/DV timings without making the
@@ -455,6 +480,143 @@ impl CaptureStream {
};
CaptureReadError::Io(error)
})?;
Ok(dqbuf)
}
/// Native HDMI NV12/BGR24 and single-planar USB UVC YUYV/NV12/RGB24/MJPEG.
/// Actual EXPBUF/import support is probed separately; failure retains copy.
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
pub(crate) fn supports_rkmpp_dmabuf(&self) -> bool {
self.dma_layout_bytes.is_some_and(|minimum| {
(2..=16).contains(&self.mappings.len())
&& self
.mappings
.iter()
.all(|planes| planes.len() == 1 && planes[0].len() >= minimum)
})
}
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
pub(crate) fn export_dmabufs(&self) -> io::Result<Vec<(OwnedFd, usize)>> {
if !self.supports_rkmpp_dmabuf() {
return Err(io::Error::new(
io::ErrorKind::Unsupported,
"Unsupported RKMPP DMA capture layout",
));
}
self.mappings
.iter()
.enumerate()
.map(|(index, planes)| {
let fd = ioctl::expbuf(&self.fd, self.queue, index, 0, ioctl::ExpbufFlags::CLOEXEC)
.map_err(|error| io::Error::other(error.to_string()))?;
Ok((fd, planes[0].len()))
})
.collect()
}
/// Run a synchronous consumer while a buffer is dequeued. QBUF occurs only
/// after the callback returns, including its error path. Consumers must end
/// hardware access before returning; see hwcodec::rkmpp_dmabuf::encode.
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
pub(crate) fn with_next_dmabuf<T>(
&mut self,
consume: impl FnOnce(usize, usize, Option<OwnedFd>) -> T,
) -> std::result::Result<(CaptureMeta, T), CaptureReadError> {
let buffer = self.dequeue_buffer()?;
let index = buffer.as_v4l2_buffer().index as usize;
let sequence = buffer.as_v4l2_buffer().sequence as u64;
if index >= self.mappings.len() {
return Err(
io::Error::new(io::ErrorKind::InvalidData, "Invalid capture buffer index").into(),
);
}
let expected = self.expected_capture_bytes();
let mapped_size = self.mappings[index][0].len();
let native_hdmi = self.native_hdmirx_state.is_some();
let compressed = self.format == PixelFormat::Mjpeg;
let lease = BufferReturn(Some(|| {
self.queue_buffer(index as u32)
.map_err(|e| io::Error::other(e.to_string()))
}));
if buffer.as_v4l2_buffer().flags & v4l2r::bindings::V4L2_BUF_FLAG_ERROR != 0 {
// A corrupt UVC frame is not a source change or a DMA failure.
// Return it without ever letting the encoder read its payload.
lease.finish()?;
return Err(io::Error::from(io::ErrorKind::WouldBlock).into());
}
if !native_hdmi
&& buffer.as_v4l2_buffer().field != v4l2r::bindings::v4l2_field_V4L2_FIELD_NONE
{
return Err(io::Error::new(
io::ErrorKind::InvalidData,
"Interlaced USB DMA frames are not supported",
)
.into());
}
let mut planes = buffer.planes_iter();
let plane = planes
.next()
.ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "Missing DMA plane"))?;
if planes.next().is_some() || plane.data_offset.copied().unwrap_or(0) != 0 {
return Err(io::Error::new(
io::ErrorKind::InvalidData,
"Unsupported DMA plane offset/layout",
)
.into());
}
let bytes_used = *plane.bytesused as usize;
if !dmabuf_layout::valid_payload(compressed, bytes_used, mapped_size, expected) {
if !native_hdmi {
// An unexpected UVC payload is not evidence of a source mode
// change. Disable DMA instead of reopening it indefinitely.
return Err(io::Error::new(
io::ErrorKind::InvalidData,
"Unexpected USB DMA payload length",
)
.into());
}
return Err(CaptureReadError::SourceChanged);
}
// UVC commonly fills vmalloc memory on the CPU. Older BSP exporters
// cache DMA attachments without usable per-frame CPU-access sync hooks.
// A fresh export object forces a fresh device mapping of this completed
// frame. Reuse the actual capture allocation, not a stale attachment.
let fresh_fd = if !native_hdmi {
Some(
ioctl::expbuf(
&self.fd,
self.queue,
index,
0,
ioctl::ExpbufFlags::CLOEXEC | ioctl::ExpbufFlags::RDWR,
)
.map_err(|error| {
io::Error::new(
io::ErrorKind::InvalidData,
format!("USB DMA re-export failed: {error}"),
)
})?,
)
} else {
None
};
let output = consume(index, bytes_used, fresh_fd);
lease.finish()?;
Ok((
CaptureMeta {
bytes_used,
sequence,
},
output,
))
}
pub fn next_into(
&mut self,
dst: &mut Vec<u8>,
) -> std::result::Result<CaptureMeta, CaptureReadError> {
let dqbuf = self.dequeue_buffer()?;
let index = dqbuf.as_v4l2_buffer().index as usize;
let sequence = dqbuf.as_v4l2_buffer().sequence as u64;
@@ -664,7 +826,7 @@ impl CaptureStream {
Ok(())
}
fn queue_buffer(&mut self, index: u32) -> Result<()> {
fn queue_buffer(&self, index: u32) -> Result<()> {
let handle = MmapHandle;
let planes = self.mappings[index as usize]
.iter()
@@ -682,6 +844,64 @@ impl CaptureStream {
}
}
#[cfg(any(test, target_arch = "aarch64", target_arch = "arm"))]
struct BufferReturn<F: FnOnce() -> io::Result<()>>(Option<F>);
#[cfg(any(test, target_arch = "aarch64", target_arch = "arm"))]
impl<F: FnOnce() -> io::Result<()>> BufferReturn<F> {
fn finish(mut self) -> io::Result<()> {
self.0.take().expect("capture lease already returned")()
}
}
#[cfg(any(test, target_arch = "aarch64", target_arch = "arm"))]
impl<F: FnOnce() -> io::Result<()>> Drop for BufferReturn<F> {
fn drop(&mut self) {
if let Some(return_buffer) = self.0.take() {
if let Err(error) = return_buffer() {
warn!("Failed to return leased capture buffer: {}", error);
}
}
}
}
#[cfg(test)]
mod dma_lease_tests {
use super::*;
use std::cell::RefCell;
#[test]
fn returns_buffer_once_after_consumer_and_does_not_retry_failed_qbuf() {
let operations = RefCell::new(Vec::new());
let lease = BufferReturn(Some(|| {
operations.borrow_mut().push("qbuf");
Err(io::Error::other("device lost"))
}));
operations.borrow_mut().push("encode completed");
assert!(lease.finish().is_err());
assert_eq!(*operations.borrow(), ["encode completed", "qbuf"]);
}
#[test]
fn returns_buffer_on_validation_error_or_unwind() {
let returns = std::cell::Cell::new(0);
{
let _lease = BufferReturn(Some(|| {
returns.set(returns.get() + 1);
Ok(())
}));
}
let _ = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
let _lease = BufferReturn(Some(|| {
returns.set(returns.get() + 1);
Ok(())
}));
panic!("consumer panic");
}));
assert_eq!(returns.get(), 2);
}
}
impl Drop for CaptureStream {
fn drop(&mut self) {
// Release ordering matters on rkcif: a subsequent open()/S_FMT from a

51
src/video/capture/status.rs
View File

@@ -2,8 +2,32 @@
use std::io;
#[cfg(any(
test,
all(target_os = "linux", any(target_arch = "aarch64", target_arch = "arm"))
))]
use crate::video::device::VideoControlMode;
use crate::video::signal::SignalStatus;
#[cfg(any(
test,
all(target_os = "linux", any(target_arch = "aarch64", target_arch = "arm"))
))]
pub(crate) fn capture_recovery_status(
control_mode: VideoControlMode,
error: &io::Error,
) -> SignalStatus {
if control_mode == VideoControlMode::Configurable && error.kind() == io::ErrorKind::TimedOut {
return SignalStatus::UvcCaptureStall;
}
match classify_capture_io_error(error) {
CaptureIoErrorKind::TransientSignal {
status: Some(status),
} => status,
_ => SignalStatus::NoSignal,
}
}
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum CaptureIoErrorKind {
DeviceLost,
@@ -52,6 +76,33 @@ pub fn capture_error_log_key(err: &io::Error) -> String {
mod tests {
use super::*;
#[test]
fn recovery_distinguishes_uvc_stalls_from_hdmi_signal_loss() {
let timeout = io::Error::from(io::ErrorKind::TimedOut);
assert_eq!(
capture_recovery_status(VideoControlMode::Configurable, &timeout),
SignalStatus::UvcCaptureStall
);
assert_eq!(
capture_recovery_status(VideoControlMode::SourceFollowing, &timeout),
SignalStatus::NoSignal
);
assert_eq!(
capture_recovery_status(
VideoControlMode::Configurable,
&io::Error::from_raw_os_error(71)
),
SignalStatus::UvcUsbError
);
assert_eq!(
capture_recovery_status(
VideoControlMode::SourceFollowing,
&io::Error::from_raw_os_error(5)
),
SignalStatus::NoSignal
);
}
#[test]
fn maps_known_signal_status_strings() {
assert_eq!(

392
src/video/pipeline/dmabuf.rs Normal file
View File

@@ -0,0 +1,392 @@
//! RKMPP-only capture/encode worker. Raw buffer ownership never crosses into a
//! latest-frame slot or a network subscriber. Other encoders use shared.rs.
use super::*;
use crate::video::capture::status::capture_recovery_status;
use crate::video::codec::registry::EncoderRegistry;
use hwcodec::rkmpp_dmabuf::{DmaEncoder, DmaEncoderConfig, DmaFormat};
pub(super) fn eligible(config: &SharedVideoPipelineConfig) -> bool {
if std::env::var("ONE_KVM_RKMPP_DMABUF").as_deref() == Ok("0") {
return false;
}
// The UVC per-frame mapping needed by older BSPs costs more than copying
// compressed packets in our current tests. Keep JPEG DMA opt-in; raw DMA
// remains automatic. The existing JPEG hardware transcode is the default.
if config.input_format == PixelFormat::Mjpeg
&& std::env::var("ONE_KVM_RKMPP_MJPEG_DMABUF").as_deref() != Ok("1")
{
return false;
}
let registry = EncoderRegistry::global();
let selected = match config.encoder_backend {
Some(backend) => registry.encoder_with_backend(config.output_codec, backend),
None => registry.best_available_encoder(config.output_codec),
};
rkmpp_dma_eligible(
selected.map(|e| e.backend),
config.output_codec,
config.input_format,
)
}
pub(super) fn prepare(
stream: &CaptureStream,
config: &SharedVideoPipelineConfig,
) -> Result<DmaEncoder> {
let buffers = stream
.export_dmabufs()
.map_err(|e| AppError::VideoError(e.to_string()))?;
DmaEncoder::new(
DmaEncoderConfig {
width: config.resolution.width,
height: config.resolution.height,
stride: stream.stride(),
format: match stream.format() {
PixelFormat::Nv12 => DmaFormat::Nv12,
PixelFormat::Bgr24 => DmaFormat::Bgr24,
PixelFormat::Yuyv => DmaFormat::Yuyv,
PixelFormat::Rgb24 => DmaFormat::Rgb24,
PixelFormat::Mjpeg => DmaFormat::Mjpeg,
_ => return Err(AppError::VideoError("Unsupported DMA format".into())),
},
hevc: config.output_codec == VideoEncoderType::H265,
fps: config.fps,
bitrate_kbps: config.bitrate_kbps(),
gop: config.gop_size().max(1),
},
buffers,
)
.map_err(AppError::VideoError)
}
enum CaptureEncoder {
Dma(DmaEncoder),
Copy(Box<EncoderThreadState>),
}
// Field order is intentional, including during unwinding: destroy the encoder
// and its imported FDs before STREAMOFF/unmap/REQBUFS(0).
struct ActiveCapture {
encoder: Option<CaptureEncoder>,
stream: CaptureStream,
}
impl ActiveCapture {
fn fallback(&mut self, config: &SharedVideoPipelineConfig) -> Result<()> {
drop(self.encoder.take());
self.encoder = Some(CaptureEncoder::Copy(Box::new(build_encoder_state(config)?)));
Ok(())
}
}
struct Completion(Arc<SharedVideoPipeline>);
impl Drop for Completion {
fn drop(&mut self) {
self.0.running_flag.store(false, Ordering::Release);
self.0.clear_cmd_tx();
let _ = self.0.encoder_done.send(true);
let _ = self.0.running.send(false);
info!("RKMPP capture/encode worker stopped and device resources released");
}
}
#[allow(clippy::too_many_arguments)]
pub(super) fn start(
pipeline: Arc<SharedVideoPipeline>,
stream: CaptureStream,
encoder: DmaEncoder,
config: SharedVideoPipelineConfig,
device: std::path::PathBuf,
buffer_count: u32,
bridge: BridgeContext,
) -> Result<()> {
let (tx, rx) = mpsc::unbounded_channel();
*pipeline.cmd_tx.write() = Some(tx);
pipeline.running_flag.store(true, Ordering::Release);
let _ = pipeline.encoder_done.send(false);
let _ = pipeline.running.send(true);
let worker = pipeline.clone();
info!(
"RKMPP DMA candidate: device={} format={:?} resolution={:?} stride={}",
device.display(),
stream.format(),
stream.resolution(),
stream.stride()
);
let active = ActiveCapture {
encoder: Some(CaptureEncoder::Dma(encoder)),
stream,
};
let result = std::thread::Builder::new()
.name("rkmpp-dmabuf".into())
.spawn(move || {
let _completion = Completion(worker.clone());
if let Err(error) = run(&worker, active, config, device, buffer_count, bridge, rx) {
error!("RKMPP DMA worker failed: {}", error);
}
});
if let Err(error) = result {
drop(Completion(pipeline));
return Err(AppError::VideoError(format!(
"Failed to start RKMPP DMA worker: {error}"
)));
}
info!("RKMPP DMA capture path active: no CPU raw-frame copies (ONE_KVM_RKMPP_DMABUF=0 disables it)");
Ok(())
}
#[allow(clippy::too_many_arguments)]
fn run(
pipeline: &Arc<SharedVideoPipeline>,
initial: ActiveCapture,
mut config: SharedVideoPipelineConfig,
device: std::path::PathBuf,
buffer_count: u32,
bridge: BridgeContext,
mut commands: mpsc::UnboundedReceiver<PipelineCmd>,
) -> Result<()> {
let policy = CaptureRecoveryPolicy::new(config.control_mode);
let mut active = Some(initial);
let mut allow_dma = true;
let mut failures = 0u32;
let mut idle_since: Option<Instant> = None;
let buffer_pool = Arc::new(FrameBufferPool::new(2)); // allocated only on fallback
let mut fps_frames = 0u32;
let mut fps_start = Instant::now();
let errors = LogThrottler::with_secs(5);
while pipeline.running_flag.load(Ordering::Acquire) {
if pipeline.subscriber_count() == 0 {
if idle_since.get_or_insert_with(Instant::now).elapsed()
>= Duration::from_secs(AUTO_STOP_GRACE_PERIOD_SECS)
{
break;
}
std::thread::sleep(Duration::from_millis(50));
continue;
}
idle_since = None;
while let Ok(command) = commands.try_recv() {
let PipelineCmd::SetBitrate { preset } = command;
// Preserve Custom values and preset-specific GOPs across fallback/reopen.
config.bitrate_preset = preset;
if let Some(capture) = active.as_mut() {
match capture.encoder.as_mut().expect("active encoder") {
CaptureEncoder::Dma(encoder) => {
if let Err(error) =
encoder.reconfigure(config.bitrate_kbps(), config.gop_size().max(1))
{
warn!(
"RKMPP DMA reconfigure failed, using copy encoder: {}",
error
);
capture.fallback(&config)?;
allow_dma = false;
pipeline.keyframe_requested.store(true, Ordering::Release);
}
}
CaptureEncoder::Copy(encoder) => {
pipeline.apply_cmd(encoder, PipelineCmd::SetBitrate { preset })?
}
}
}
}
if active.is_none() {
match open_capture_stream_for_retry(
&device,
config.resolution,
config.input_format,
config.fps,
buffer_count.max(2),
Duration::from_secs(2),
bridge.clone(),
config.control_mode,
is_device_lost_message,
) {
CaptureOpenResult::Opened(stream) => {
if stream.resolution() != config.resolution
|| stream.format() != config.input_format
{
*pipeline.pending_sync_geometry.lock() =
Some((stream.resolution(), stream.format()));
break;
}
config.align_source_fps(stream.source_fps());
// Update only timing: a concurrently queued bitrate command
// must retain the user's latest preset in the shared config.
pipeline.config.blocking_write().fps = config.fps;
let encoder = if allow_dma && stream.supports_rkmpp_dmabuf() {
match prepare(&stream, &config) {
Ok(encoder) => CaptureEncoder::Dma(encoder),
Err(error) => {
warn!("RKMPP DMA reopen failed, using copy encoder: {}", error);
allow_dma = false;
CaptureEncoder::Copy(Box::new(build_encoder_state(&config)?))
}
}
} else {
CaptureEncoder::Copy(Box::new(build_encoder_state(&config)?))
};
active = Some(ActiveCapture {
encoder: Some(encoder),
stream,
});
pipeline.keyframe_requested.store(true, Ordering::Release);
}
CaptureOpenResult::NoSignal(status) => {
failures = failures.saturating_add(1);
let delay = policy.retry_delay(failures);
pipeline.notify_state(PipelineStateNotification::no_signal(
status,
Some(delay.as_millis() as u64),
));
wait_for_source_change(&bridge, delay, || {
pipeline.running_flag.load(Ordering::Acquire)
});
continue;
}
CaptureOpenResult::DeviceLost(reason) => {
pipeline.mark_device_lost(reason);
break;
}
CaptureOpenResult::Fatal => break,
}
}
let capture = active.as_mut().expect("opened capture");
let result = match capture.encoder.as_mut().expect("active encoder") {
CaptureEncoder::Dma(encoder) => {
let pts = pipeline.pts_ms();
capture
.stream
.with_next_dmabuf(|index, bytes_used, fresh_fd| {
let keyframe = pipeline.keyframe_requested.swap(false, Ordering::AcqRel);
// The callback holds the dequeue lease until native encode
// completes (or destroys MPP on error), before QBUF.
unsafe { encoder.encode(index, bytes_used, fresh_fd, pts, keyframe) }
})
.map(|(_, packet)| {
packet
.map(|packet| {
let data = Bytes::from(packet);
let is_keyframe = match config.output_codec {
VideoEncoderType::H264 => h264_bitstream::is_keyframe(&data),
VideoEncoderType::H265 => h265_bitstream::is_keyframe(&data),
_ => false,
};
let (data, is_keyframe) = pipeline.inspect_and_parameterize_packet(
config.output_codec,
data,
is_keyframe,
);
if config.output_codec == VideoEncoderType::H264 {
pipeline.update_h264_profile_level_id(&data);
}
vec![EncodedVideoFrame {
data,
pts_ms: pts,
is_keyframe,
sequence: pipeline.sequence.fetch_add(1, Ordering::Relaxed) + 1,
duration: Duration::from_micros(
1_000_000 / config.fps.max(1) as u64,
),
codec: config.output_codec,
}]
})
.map_err(AppError::VideoError)
})
}
CaptureEncoder::Copy(encoder) => {
let mut raw = buffer_pool.take(0);
capture.stream.next_into(&mut raw).map(|meta| {
let frame = VideoFrame::from_pooled(
Arc::new(FrameBuffer::new(raw, Some(buffer_pool.clone()))),
config.resolution,
config.input_format,
capture.stream.stride(),
meta.sequence,
);
pipeline.encode_frame_sync(encoder, &frame)
})
}
};
match result {
Ok(Ok(frames)) => {
failures = 0;
pipeline.notify_state(PipelineStateNotification::streaming(
config.resolution,
config.input_format,
config.fps,
));
for frame in frames {
pipeline.broadcast_encoded(Arc::new(frame));
fps_frames += 1;
}
}
Ok(Err(error)) => {
if matches!(capture.encoder, Some(CaptureEncoder::Dma(_))) {
warn!("RKMPP DMA encode failed; disabling DMA for this pipeline and using copy encoder: {}", error);
capture.fallback(&config)?;
allow_dma = false;
pipeline.keyframe_requested.store(true, Ordering::Release);
} else if errors.should_log("copy_encode") {
error!("RKMPP copy encode failed: {}", error);
}
}
Err(CaptureReadError::Io(error)) if error.kind() == std::io::ErrorKind::WouldBlock => {
continue
}
Err(CaptureReadError::Io(error))
if error.kind() == std::io::ErrorKind::InvalidData && allow_dma =>
{
warn!(
"Unsupported RKMPP DMA frame layout, using copy encoder: {}",
error
);
capture.fallback(&config)?;
allow_dma = false;
pipeline.keyframe_requested.store(true, Ordering::Release);
}
Err(error) => {
let mut status = SignalStatus::NoSignal;
if let CaptureReadError::Io(ref io) = error {
if classify_capture_io_error(io) == CaptureIoErrorKind::DeviceLost
|| is_device_lost_message(&io.to_string())
{
pipeline.mark_device_lost(io.to_string());
break;
}
if errors.should_log("capture") {
warn!("RKMPP DMA capture recovery: {}", io);
}
status = capture_recovery_status(config.control_mode, io);
}
// ActiveCapture drops encoder/imports before the V4L2 stream.
drop(active.take());
failures = failures.saturating_add(1);
let delay = policy.retry_delay(failures);
pipeline.notify_state(PipelineStateNotification::no_signal(
status,
Some(delay.as_millis() as u64),
));
if !matches!(error, CaptureReadError::SourceChanged) {
wait_for_source_change(&bridge, delay, || {
pipeline.running_flag.load(Ordering::Acquire)
});
}
}
}
if fps_start.elapsed() >= Duration::from_secs(1) {
pipeline.stats.blocking_lock().current_fps =
fps_frames as f32 / fps_start.elapsed().as_secs_f32();
fps_frames = 0;
fps_start = Instant::now();
}
}
// Explicitly release in the worker before Completion publishes stopped.
drop(active);
Ok(())
}

247
src/video/pipeline/shared.rs
View File

@@ -29,6 +29,96 @@ use tracing::{debug, error, info, trace, warn};
use super::encoder_state::{build_encoder_state, should_parallel_decode_mjpeg, EncoderThreadState};
#[cfg(all(target_os = "linux", any(target_arch = "aarch64", target_arch = "arm")))]
#[path = "dmabuf.rs"]
mod dmabuf;
#[cfg(any(
test,
all(target_os = "linux", any(target_arch = "aarch64", target_arch = "arm"))
))]
fn rkmpp_dma_eligible(
backend: Option<EncoderBackend>,
codec: VideoEncoderType,
format: PixelFormat,
) -> bool {
backend == Some(EncoderBackend::Rkmpp)
&& matches!(codec, VideoEncoderType::H264 | VideoEncoderType::H265)
&& matches!(
format,
PixelFormat::Bgr24
| PixelFormat::Nv12
| PixelFormat::Yuyv
| PixelFormat::Rgb24
| PixelFormat::Mjpeg
)
}
#[cfg(test)]
mod dma_selection_tests {
use super::*;
#[test]
fn only_selected_rkmpp_uses_dma() {
for backend in [
EncoderBackend::Software,
EncoderBackend::Vaapi,
EncoderBackend::Nvenc,
EncoderBackend::Qsv,
EncoderBackend::Amf,
EncoderBackend::V4l2m2m,
] {
for codec in [VideoEncoderType::H264, VideoEncoderType::H265] {
for format in [
PixelFormat::Bgr24,
PixelFormat::Nv12,
PixelFormat::Yuyv,
PixelFormat::Rgb24,
PixelFormat::Mjpeg,
] {
assert!(!rkmpp_dma_eligible(Some(backend), codec, format));
}
}
}
assert!(!rkmpp_dma_eligible(
None,
VideoEncoderType::H264,
PixelFormat::Nv12
));
for codec in [VideoEncoderType::H264, VideoEncoderType::H265] {
for format in [
PixelFormat::Bgr24,
PixelFormat::Nv12,
PixelFormat::Yuyv,
PixelFormat::Rgb24,
PixelFormat::Mjpeg,
] {
assert!(rkmpp_dma_eligible(
Some(EncoderBackend::Rkmpp),
codec,
format
));
}
}
for format in [
PixelFormat::Nv16,
PixelFormat::Nv21,
PixelFormat::Nv24,
PixelFormat::Yuv420,
] {
assert!(!rkmpp_dma_eligible(
Some(EncoderBackend::Rkmpp),
VideoEncoderType::H264,
format
));
}
assert!(!rkmpp_dma_eligible(
Some(EncoderBackend::Rkmpp),
VideoEncoderType::VP9,
PixelFormat::Nv12
));
}
}
/// Grace period before auto-stopping pipeline when no subscribers (in seconds)
const AUTO_STOP_GRACE_PERIOD_SECS: u64 = 3;
/// After this many consecutive timeouts, log a prominent warning.
@@ -172,7 +262,9 @@ pub struct EncodedVideoFrame {
}
enum PipelineCmd {
SetBitrate { bitrate_kbps: u32, gop: u32 },
SetBitrate {
preset: crate::video::codec::BitratePreset,
},
}
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -255,6 +347,15 @@ impl Default for SharedVideoPipelineConfig {
}
impl SharedVideoPipelineConfig {
/// Keep encoder timing aligned with the negotiated HDMI source on every open.
fn align_source_fps(&mut self, source_fps: Option<f64>) {
if self.control_mode == VideoControlMode::SourceFollowing {
if let Some(fps) = source_fps {
self.fps = fps.round().clamp(1.0, 120.0) as u32;
}
}
}
/// Get effective bitrate in kbps
pub fn bitrate_kbps(&self) -> u32 {
self.bitrate_preset.bitrate_kbps()
@@ -538,14 +639,13 @@ impl SharedVideoPipeline {
fn apply_cmd(&self, state: &mut EncoderThreadState, cmd: PipelineCmd) -> Result<()> {
match cmd {
PipelineCmd::SetBitrate { bitrate_kbps, gop } => {
#[cfg(not(any(target_arch = "aarch64", target_arch = "arm")))]
let _ = gop;
PipelineCmd::SetBitrate { preset } => {
let bitrate_kbps = preset.bitrate_kbps();
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
if state.ffmpeg_hw_enabled {
if let Some(ref mut pipeline) = state.ffmpeg_hw_pipeline {
pipeline
.reconfigure(bitrate_kbps as i32, gop as i32)
.reconfigure(bitrate_kbps as i32, preset.gop_size(state.fps) as i32)
.map_err(|e| {
let detail = if e.is_empty() {
ffmpeg_hw_last_error()
@@ -767,7 +867,8 @@ impl SharedVideoPipeline {
subdev_path.clone(),
parse_bridge_kind(bridge_kind.as_deref()),
);
let preopened: Option<CaptureStream> = match open_capture_stream(
#[allow(unused_mut)]
let mut preopened: Option<CaptureStream> = match open_capture_stream(
&device_path,
config.resolution,
config.input_format,
@@ -781,11 +882,7 @@ impl SharedVideoPipeline {
let negotiated_res = s.resolution();
let negotiated_fmt = s.format();
let previous = (config.resolution, config.input_format, config.fps);
if config.control_mode == VideoControlMode::SourceFollowing {
if let Some(source_fps) = s.source_fps() {
config.fps = source_fps.round().clamp(1.0, 120.0) as u32;
}
}
config.align_source_fps(s.source_fps());
config.resolution = negotiated_res;
config.input_format = negotiated_fmt;
if previous != (config.resolution, config.input_format, config.fps) {
@@ -822,6 +919,32 @@ impl SharedVideoPipeline {
Err(e) => return Err(e),
};
#[cfg(all(target_os = "linux", any(target_arch = "aarch64", target_arch = "arm")))]
if dmabuf::eligible(&config) {
if let Some(stream) = preopened.as_ref().filter(|s| s.supports_rkmpp_dmabuf()) {
match dmabuf::prepare(stream, &config) {
Ok(encoder) => {
return dmabuf::start(
self.clone(),
preopened.take().expect("preopened DMA capture"),
encoder,
config,
device_path,
buffer_count,
BridgeContext::from_parts(
subdev_path,
parse_bridge_kind(bridge_kind.as_deref()),
),
);
}
Err(error) => warn!(
"RKMPP DMA unavailable; using existing copy pipeline: {}",
error
),
}
}
}
let mut encoder_config = config.clone();
if parallel_mjpeg_decode {
encoder_config.input_format = PixelFormat::Nv12;
@@ -1400,23 +1523,7 @@ impl SharedVideoPipeline {
let input_format = state.input_format;
let raw_frame = frame.data();
let process_start = PROCESS_START.get_or_init(Instant::now);
let current_ts_us = process_start.elapsed().as_micros() as i64;
let start_ts_us = self.pipeline_start_time_us.load(Ordering::Acquire);
let pts_ms = if start_ts_us == 0 {
let start_ts_us = match self.pipeline_start_time_us.compare_exchange(
0,
current_ts_us,
Ordering::AcqRel,
Ordering::Acquire,
) {
Ok(_) => current_ts_us,
Err(existing) => existing,
};
current_ts_us.saturating_sub(start_ts_us) / 1000
} else {
current_ts_us.saturating_sub(start_ts_us) / 1000
};
let pts_ms = self.pts_ms();
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
if state.ffmpeg_hw_enabled {
@@ -1551,6 +1658,28 @@ impl SharedVideoPipeline {
}
}
fn pts_ms(&self) -> i64 {
let current_ts_us = PROCESS_START
.get_or_init(Instant::now)
.elapsed()
.as_micros() as i64;
let start_ts_us = self.pipeline_start_time_us.load(Ordering::Acquire);
let start_ts_us = if start_ts_us == 0 {
match self.pipeline_start_time_us.compare_exchange(
0,
current_ts_us,
Ordering::AcqRel,
Ordering::Acquire,
) {
Ok(_) => current_ts_us,
Err(existing) => existing,
}
} else {
start_ts_us
};
current_ts_us.saturating_sub(start_ts_us) / 1000
}
/// Stop the pipeline (non-blocking, does not wait for capture thread to exit)
pub fn stop(&self) {
if self.running_flag.swap(false, Ordering::AcqRel) {
@@ -1630,13 +1759,11 @@ impl SharedVideoPipeline {
&self,
preset: crate::video::codec::BitratePreset,
) -> Result<()> {
let bitrate_kbps = preset.bitrate_kbps();
let gop = {
{
let mut config = self.config.write().await;
config.bitrate_preset = preset;
config.gop_size()
};
self.send_cmd(PipelineCmd::SetBitrate { bitrate_kbps, gop });
}
self.send_cmd(PipelineCmd::SetBitrate { preset });
Ok(())
}
@@ -1831,6 +1958,60 @@ mod tests {
use super::*;
use crate::video::codec::BitratePreset;
#[tokio::test]
async fn bitrate_commands_preserve_custom_values_and_gop_policy() {
let pipeline = SharedVideoPipeline::new(SharedVideoPipelineConfig::default()).unwrap();
let (tx, mut rx) = mpsc::unbounded_channel();
*pipeline.cmd_tx.write() = Some(tx);
for preset in [
BitratePreset::Custom(2500),
BitratePreset::Custom(1000),
BitratePreset::Speed,
BitratePreset::Quality,
] {
pipeline.set_bitrate_preset(preset).await.unwrap();
let PipelineCmd::SetBitrate { preset: received } = rx.try_recv().unwrap();
assert_eq!(received, preset);
assert_eq!(pipeline.config().await.bitrate_preset, preset);
// Rebuilt encoders must retain the preset's policy at the new FPS.
let restored = SharedVideoPipelineConfig {
bitrate_preset: received,
fps: 60,
..Default::default()
};
assert_eq!(restored.bitrate_kbps(), preset.bitrate_kbps());
assert_eq!(restored.gop_size(), preset.gop_size(60));
}
}
#[test]
fn source_reopen_updates_fps_and_gop_without_changing_geometry_or_bitrate() {
let mut config = SharedVideoPipelineConfig {
control_mode: VideoControlMode::SourceFollowing,
resolution: Resolution::HD1080,
fps: 60,
bitrate_preset: BitratePreset::Quality,
..Default::default()
};
config.align_source_fps(Some(29.97));
assert_eq!(config.fps, 30);
assert_eq!(config.gop_size(), 60);
assert_eq!(config.resolution, Resolution::HD1080);
assert_eq!(config.bitrate_kbps(), 8000);
config.align_source_fps(None);
assert_eq!(config.fps, 30);
config.align_source_fps(Some(59.94));
assert_eq!(config.fps, 60);
assert_eq!(config.gop_size(), 120);
}
#[test]
fn configurable_capture_keeps_requested_fps() {
let mut config = SharedVideoPipelineConfig::default();
config.align_source_fps(Some(60.0));
assert_eq!(config.fps, 30);
}
#[test]
fn test_pipeline_config() {
let h264 = SharedVideoPipelineConfig::h264(Resolution::HD1080, BitratePreset::Balanced);