[VP] Fix VEBOX surface config crash in 60i->30p deinterlacing

[tsukumijima]

Summary

Fix a segfault (SIGSEGV / exit code 139) in VpResourceManager::AssignVeboxResource() that occurs when performing non-bob deinterlacing (normal/advanced mode, 60i->30p) on pure softlet platforms (Meteor Lake and later).

Background

I originally encountered this crash while using QSVEncC (a hardware-accelerated video encoder leveraging Intel Quick Sync Video via libvpl) on an Intel Core Ultra 7 265K (ArrowLake-S) system running Ubuntu 22.04.5 LTS.
After discussions with @rigaya (the author of QSVEncC), we confirmed that the crash occurs in the Intel Media Driver, not in QSVEncC itself.

This is a significant regression for real-world use cases. In Japan, terrestrial TV broadcasts use MPEG-2 TS (1440x1080i, interlaced), and QSVEncC is widely used to transcode these recordings to H.264/H.265 with hardware-accelerated deinterlacing.
This workflow was working fine on older platforms (e.g., i5-9400 / Gen9, i5-12400 / ADL, Arc A310 / DG2), but is now completely broken on Meteor Lake and later platforms where the pure softlet VP adapter is used.

The bug also reproduces with FFmpeg's QSV VPP deinterlace filter (vpp_qsv), confirming this is not application-specific but a driver-level issue affecting any VA-API consumer that uses VEBOX-based 60i->30p deinterlacing.

We would appreciate prompt merging and backporting to distribution packages (e.g., intel-media-va-driver-non-free on Ubuntu), as this currently renders non-bob deinterlacing unusable on all MTL+ platforms.

Root Cause

In AssignVeboxResource(), the VEBOX_SURFACES_CONFIG is constructed with m_sameSamples passed directly as the sameSample field. However, InitSurfaceConfigMap() only registers entries with sameSample=true for 60fps DI (b64DI=true) configurations. The 60i->30p section (lines 695-705, b64DI=false) has no entries with sameSample=true:

note: the source code comments use "30i -> 30p" counting interlaced frames rather than fields, which is equivalent to "60i -> 30p" in broadcast terminology.

//30i -> 30p sfc Enable
AddSurfaceConfig(false, true,   false, ...);  // sameSample=false
AddSurfaceConfig(false, true,   false, ...);  // sameSample=false
...
//30i -> 30p sfc disable
AddSurfaceConfig(false, false,  false, ...);  // sameSample=false
...

When b60fpsDi=false (60i->30p) and m_sameSamples=true, the constructed config value has no matching entry in the map, causing MOS_STATUS_INVALID_PARAMETER to be returned and ultimately a segfault.

Fix

Gate m_sameSamples by b60fpsDi so that sameSample is only set for 60fps DI (60i->60p) paths, where the surface config map does have matching entries:

const bool sameSamplesForConfig = b60fpsDi ? m_sameSamples : false;
VEBOX_SURFACES_CONFIG cfg(b60fpsDi, caps.bSFC, sameSamplesForConfig, ...);

Affected Platforms

This bug affects all pure softlet VP adapter platforms that go through InitSurfaceConfigMap():

Platform	Adapter	Affected
TGL / ADL / RPL (Gen12)	VpPipelineG12Adapter (legacy hybrid)	No
DG2 / Arc (Xe HPM)	VpPipelineAdapterXe_Hpm (legacy hybrid)	No
MTL (Xe LPM+)	VpPipelineAdapterXe_Lpm_Plus (pure softlet)	Yes
ARL (Xe LPM+)	VpPipelineAdapterXe_Lpm_Plus (pure softlet)	Yes
LNL (Xe2 LPM)	VpPipelineAdapterXe2_Lpm (pure softlet)	Yes
BMG (Xe2 HPM)	VpPipelineAdapterXe2_Hpm (pure softlet)	Yes
PTL (Xe3 LPM)	VpPipelineAdapterXe3_Lpm (pure softlet)	Yes
NVL (Xe3P LPM)	VpPipelineAdapterXe3P_Lpm (pure softlet)	Yes

Legacy hybrid adapters (TGL/ADL/RPL, DG2/Arc) inherit from both VphalState* and VpPipelineAdapterLegacy, using a different VEBOX resource management path that does not go through InitSurfaceConfigMap(). This is why the bug was not noticed on older platforms.

Bug Origin

Introduced in commit f38f67c ([VP] 30pRefdi case Linux, 2022-11-09), which added 60i->30p config entries to InitSurfaceConfigMap() but only with sameSample=false variants. The bug remained latent until the pure softlet VP adapter transition starting with Meteor Lake (late 2023).

The bug is codec-agnostic: m_sameSamples is computed from frame IDs and buffer object tracking in the VA-API layer, with no codec-specific logic. Any interlaced source (MPEG-2, H.264 interlaced, VC-1, etc.) can trigger it.

Reproduction

Hardware: Intel Core Ultra 7 265K (ArrowLake-S, iGPU) — or any MTL+ platform
OS: Ubuntu 22.04.5 LTS, kernel 6.8.0-100-generic
Driver: intel-media-va-driver-non-free 25.2.4 (iHD 25.2.4)

Input: Any interlaced video file (e.g., MPEG-2 TS, 1440x1080i)

QSVEncC 8.04

# Crashes (SIGSEGV / exit 139) — non-bob DI (60i->30p, b60fpsDi=false)
QSVEncC --avhw -i input.ts --vpp-deinterlace normal -o output.mp4

# Works fine — bob mode (60i->60p, b60fpsDi=true)
QSVEncC --avhw -i input.ts --vpp-deinterlace bob -o output.mp4

FFmpeg 7.1

# Crashes (SIGSEGV / exit 139) — advanced DI at frame rate (60i->30p)
ffmpeg -hwaccel qsv -c:v mpeg2_qsv -i input.ts \
  -vf 'vpp_qsv=deinterlace=2:rate=frame' -c:v h264_qsv -an output.mp4

# Works fine — bob mode (60i->60p)
ffmpeg -hwaccel qsv -c:v mpeg2_qsv -i input.ts \
  -vf 'vpp_qsv=deinterlace=1' -c:v h264_qsv -an output.mp4

# Works fine — advanced DI at field rate (60i->60p, b60fpsDi=true)
ffmpeg -hwaccel qsv -c:v mpeg2_qsv -i input.ts \
  -vf 'deinterlace_qsv=mode=advanced' -c:v h264_qsv -an output.mp4

Test Results

The "After Fix" column was tested by building iHD_drv_video.so from this branch and loading it via LIBVA_DRIVERS_PATH at runtime, on the same system listed above.

Tool	DI Mode	Output	Before Fix (25.2.4)	After Fix (this branch)
QSVEncC	normal	30p	SIGSEGV (exit 139)	OK
QSVEncC	normal (avsw decode)	30p	SIGSEGV (exit 139)	OK
QSVEncC	bob	60p	OK	OK
FFmpeg	vpp_qsv advanced, rate=frame	30p	SIGSEGV (exit 139)	OK
FFmpeg	vpp_qsv bob	60p	OK	OK
FFmpeg	deinterlace_qsv advanced	60p	OK	OK

PS: This investigation and fix were performed using Codex (GPT-5.3-Codex) and Claude Code (Claude Opus 4.6).
However, it has been confirmed on an actual Core Ultra 7 265K machine that applying this patch to iHD_drv_video.so successfully avoids this error and enables deinterlacing.
Therefore, it is undeniable that this patch is effective, at the very least.

[umbrageodotus]

Updates?