PR #25906 opened by Anton Kesy (antonkesy) URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25906 Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25906.patch
## Problem When a QSV decoder outputs frames to system memory (no -hwaccel qsv), it keeps a buffer pool for the decoded frames. When the stream changes resolution or format, the decoder reinitializes and creates a new pool, but never frees the old one. Every resolution change leaks the old pool plus all frame buffers in it (several MiB per change at 1080p). ## Fix Free the old pool before creating the new one. `av_buffer_pool_uninit()` should be safe here since the frames are still held by the caller, and the old pool is released once they are returned. ### How to reproduce Use any H.264 clip that changes resolution (e.g. 1080p -> 720p -> 1080p) and loop it so the decoder reinitializes many times: ``` ffmpeg -benchmark -stream_loop 0 -c:v h264_qsv -i size_change.mp4 -an -f null - ffmpeg -benchmark -stream_loop 50 -c:v h264_qsv -i size_change.mp4 -an -f null - ``` Compare the `bench: maxrss=` line printed by each command. ### Results: Before: ``` passes=1 peak memory=219 MiB passes=51 peak memory=2634 MiB ``` After: ``` passes=1 peak memory=194 MiB passes=51 peak memory=193 MiB ``` >From 9a45d2422447742842f0d063a4c98d7cc4e1600d Mon Sep 17 00:00:00 2001 From: Anton Kesy <[email protected]> Date: Mon, 9 Nov 2026 09:11:40 +0100 Subject: [PATCH] avcodec/qsvdec: free the system memory pool on reinit qsv_decode_init_context() runs again after a resolution or format change and overwrote q->pool without releasing the old one. When decoding to system memory (no hw_frames_ctx), every reinit leaked the old pool and all buffers allocated from it. Signed-off-by: Anton Kesy <[email protected]> --- libavcodec/qsvdec.c | 1 + 1 file changed, 1 insertion(+) diff --git a/libavcodec/qsvdec.c b/libavcodec/qsvdec.c index 990e46ea1d7..8e493146519 100644 --- a/libavcodec/qsvdec.c +++ b/libavcodec/qsvdec.c @@ -417,6 +417,7 @@ static int qsv_decode_init_context(AVCodecContext *avctx, QSVContext *q, mfxVide ret = av_image_get_buffer_size(avctx->pix_fmt, FFALIGN(avctx->coded_width, 128), FFALIGN(avctx->coded_height, 64), 1); if (ret < 0) return ret; + av_buffer_pool_uninit(&q->pool); q->pool = av_buffer_pool_init(ret, av_buffer_allocz); } return 0; -- 2.54.0 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
