Re: [PATCH v2 0/5] media: rkvdec: Enable multi-core support

From: Lukas Splavec

Date: Fri Nov 13 2026 - 10:11:57 EST


On Mon, Aug 10, 2026 at 11:22:10AM +0100, Detlev Casanova wrote:
> Since the driver is used for decoding on rk3588 and that the SoC has
> 2 identical decoding cores, enable support for it.

Hi Detlev,

I tested this series together with Sven's RGA3 multicore v1 on an RK3588
board (Giada D77, both vdec cores enabled), with 4 to 8 concurrent
GStreamer 1.28 pipelines (v4l2slh264dec / v4l2slh265dec, 1080p to 4K60,
H.264 and HEVC 8/10-bit).

Multicore works fine: both cores are used and throughput scales roughly
linearly up to 8 streams; 4x HEVC 4K60 runs at about 65% busy per core.

I encountered a problem on v7.1-rc1. I hit an rk_iommu page fault storm
followed by a decoder timeout and device freeze, when pipelines were
started and stopped in overlapping manner. I was able to replicate the
same behaviour without these and Sven's patches on plain v7.1-rc1, but
not on v7.0.

The cause seems to be the inverted error check in iommu_map_nosync() from
d6c65b0fd6 ("iommupt: Avoid rewalking during map"), already fixed by
6fc7e8a3 ("iommu: Fix loss of errno on map failure for classic ops") in
v7.1-rc4. With that fix applied on top of both series, 150 rounds of my
stress test pass with no faults. So it is not caused by this series, but
a rebase onto >= v7.1-rc4 would avoid confusing other testers.

The underlying map failure is likely a separate issue: the RCB SRAM
buffers are mapped at IOVA == kernel VA, which the dma-iommu allocator
does not know about, so capture buffers can collide with them. That is
present since v7.0 and independent of this series, so it is probably a
matter for a separate thread.

Tested-by: Lukas Splavec <lukas.splavec@xxxxxxxxxx>

Thanks,
Lukas