On 7/28/23 15:03, Tomasz Figa wrote:
CAUTION: Email originated externally, do not click links or open attachments unless you recognize the sender and know the content is safe.
On Fri, Jul 28, 2023 at 3:55 PM Hsia-Jun Li <Randy.Li@xxxxxxxxxxxxx> wrote:
On 7/28/23 14:46, Tomasz Figa wrote:
CAUTION: Email originated externally, do not click links or open attachments unless you recognize the sender and know the content is safe.
On Mon, Jul 17, 2023 at 4:44 PM Hsia-Jun Li <Randy.Li@xxxxxxxxxxxxx> wrote:
On 7/12/23 18:48, Tomasz Figa wrote:
CAUTION: Email originated externally, do not click links or open attachments unless you recognize the sender and know the content is safe.
On Mon, Jul 03, 2023 at 04:35:30PM +0800, Hsia-Jun Li wrote:
On 7/3/23 16:09, Benjamin Gaignard wrote:
CAUTION: Email originated externally, do not click links or open
attachments unless you recognize the sender and know the content is
safe.
Le 30/06/2023 à 11:51, Hsia-Jun Li a écrit :
On 6/22/23 21:13, Benjamin Gaignard wrote:
CAUTION: Email originated externally, do not click links or open
attachments unless you recognize the sender and know the content is
safe.
After changing bufs arrays to a dynamic allocated array
VB2_MAX_FRAME doesn't mean anything for videobuf2 core.
I think make it 64 which is the VB2_MAX_FRAME in Android GKI kernel is
more reasonable.
It would be hard to iterate the whole array, it would go worse with a
filter. Such iterate may need to go twice because you mix
post-processing buffer and decoding buffer(with MV) in the same array.
Here I don't want to change drivers behavior so I keep the same value.
If it happens that they need more buffers, like for dynamic resolution
change
feature for Verisilicon VP9 decoder, case by case patches will be needed.
I just don't like the idea that using a variant length array here.
"I don't like" is not an argument. We had a number of arguments for
using a generic helper (originally idr, but later decided to go with
XArray, because the former is now deprecated) that we pointed out in
our review comments for previous revisions. It wasn't really about the
size being variable, but rather avoiding open coding things in vb2 and
duplicating what's already implemented in generic code.
I just want to say I don't think we need a variable length array to
store the buffer here.
And the below is the reason that such a case could be avoided in the
first place.
And I could explain why you won't need so many buffers for the performance
of decoding.
VP9 could support 10 reference frames in dpb.
Even for those frequent resolution changing test set, it would only happen
to two resolutions,
32 would be enough for 20 buffers of two resolution plus golden frames. It
also leaves enough slots for re-order latency.
If your case had more two resolutions, likes low->medium->high.
I would suggest just skip the medium resolutions, just allocate the lower
one first for fast playback then the highest for all the possible
medium cases. Reallocation happens frequently would only cause memory
fragment, nothing benefits your performance.
We have mechanisms in the kernel to deal with memory fragmentation
(migration/compaction) and it would still only matters for the
pathologic cases of hardware that require physically contiguous memory.
Modern hardware with proper DMA capabilities can either scatter-gather
or are equipped with an IOMMU, so the allocation always happens in page
granularity and fragmentation is avoided.
Unfortunately, there are more devices that didn't have a IOMMU attached
to it, supporting scatter gather is more odd.
It would be more likely that IOMMU would be disabled for the performance
reason.
These days IOMMU is totally mandatory if you want to think about
having any level of security in your system. Sure, there could be some
systems that are completely isolated from any external environment,
like some offline industry automation machines, but then arguably
their running conditions would also be quite static and require very
little memory re-allocation.
Vendor just decided not to included such hardware.
That is why From ION to DMA-heap, people like to allocate from a cavout
out memory.
I also don't buy the performance reason. CPUs have been behind MMUs
for ages and nobody is running them with paging disabled for
performance reasons. Similarly, most of the modern consumer systems
Page lookup would increase the delay. Besides a few upstream devices
prove them only use a level 1 page table without TBL.
That's just an excuse for a bad hardware design/implementation. As I
said, there are good IOMMU implementations out there that don't suffer
from performance issues.
I could do nothing about that.
Besides, even with TLB, cache missing could happen frequently,
especially we need to access many (5~16, 10 usually) buffers and more
11MBytes each in a hardware processing.
You can't have a very large TLB.
(mobile phones, PCs) run with IOMMUs enabled for pretty much anything
because of the security reason and they don't seem to be having any
If the page is secure, you can't operate it in a insecure IOMMU or MMU.
The most security way here, we should use a dedicated memory(or a zone
in unified memory).
You still need something to enforce that the hardware is not accessing
memory that it's not supposed to access. How do you do that without an
IOMMU?
If you know the arm security pipeline and security controller, you could
found we could reserved a range of memory for a security id(devices in
secure world may be a different security domain).
Besides, a MMU or security MPU could mark some pages for the secure
world access only, it doesn't mean the device need an IOMMU to access
them. The MPU could filter the access through the AXI id.
I believe there are more users in mobile for DMA-heap than kernel's dma
allocation API.
Yes, but that's completely separate from whether there is an IOMMU or
not. It's just a different allocation API.
The memory heap would mean a dedicated memory usually(we don't talk
about system heap or why there are many vendor heaps). Dedicated memory
means contiguous memory in the most of cases.
performance issues. In fact, it can improve the performance, because
memory allocation is much easier and without contiguous careouts (as
we used to have long ago on Android devices) the extra memory can be
used for buffers and caches to improve system performance.
Best regards,
Tomasz
Best regards,
Tomasz
Remove it from the core definitions but keep it for drivers internal
needs.
Signed-off-by: Benjamin Gaignard <benjamin.gaignard@xxxxxxxxxxxxx>
---
drivers/media/common/videobuf2/videobuf2-core.c | 2 ++
drivers/media/platform/amphion/vdec.c | 1 +
.../media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c | 2 ++
drivers/media/platform/qcom/venus/hfi.h | 2 ++
drivers/media/platform/verisilicon/hantro_hw.h | 2 ++
drivers/staging/media/ipu3/ipu3-v4l2.c | 2 ++
include/media/videobuf2-core.h | 1 -
include/media/videobuf2-v4l2.h | 4 ----
8 files changed, 11 insertions(+), 5 deletions(-)
diff --git a/drivers/media/common/videobuf2/videobuf2-core.c
b/drivers/media/common/videobuf2/videobuf2-core.c
index 86e1e926fa45..899783f67580 100644
--- a/drivers/media/common/videobuf2/videobuf2-core.c
+++ b/drivers/media/common/videobuf2/videobuf2-core.c
@@ -31,6 +31,8 @@
#include <trace/events/vb2.h>
+#define VB2_MAX_FRAME 32
+
static int debug;
module_param(debug, int, 0644);
diff --git a/drivers/media/platform/amphion/vdec.c
b/drivers/media/platform/amphion/vdec.c
index 3fa1a74a2e20..b3219f6d17fa 100644
--- a/drivers/media/platform/amphion/vdec.c
+++ b/drivers/media/platform/amphion/vdec.c
@@ -28,6 +28,7 @@
#define VDEC_MIN_BUFFER_CAP 8
#define VDEC_MIN_BUFFER_OUT 8
+#define VB2_MAX_FRAME 32
struct vdec_fs_info {
char name[8];
diff --git
a/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
b/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
index 6532a69f1fa8..a1e0f24bb91c 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
@@ -16,6 +16,8 @@
#include "../vdec_drv_if.h"
#include "../vdec_vpu_if.h"
+#define VB2_MAX_FRAME 32
+
/* reset_frame_context defined in VP9 spec */
#define VP9_RESET_FRAME_CONTEXT_NONE0 0
#define VP9_RESET_FRAME_CONTEXT_NONE1 1
diff --git a/drivers/media/platform/qcom/venus/hfi.h
b/drivers/media/platform/qcom/venus/hfi.h
index f25d412d6553..bd5ca5a8b945 100644
--- a/drivers/media/platform/qcom/venus/hfi.h
+++ b/drivers/media/platform/qcom/venus/hfi.h
@@ -10,6 +10,8 @@
#include "hfi_helper.h"
+#define VB2_MAX_FRAME 32
+
#define VIDC_SESSION_TYPE_VPE 0
#define VIDC_SESSION_TYPE_ENC 1
#define VIDC_SESSION_TYPE_DEC 2
diff --git a/drivers/media/platform/verisilicon/hantro_hw.h
b/drivers/media/platform/verisilicon/hantro_hw.h
index e83f0c523a30..9e8faf7ba6fb 100644
--- a/drivers/media/platform/verisilicon/hantro_hw.h
+++ b/drivers/media/platform/verisilicon/hantro_hw.h
@@ -15,6 +15,8 @@
#include <media/v4l2-vp9.h>
#include <media/videobuf2-core.h>
+#define VB2_MAX_FRAME 32
+
#define DEC_8190_ALIGN_MASK 0x07U
#define MB_DIM 16
diff --git a/drivers/staging/media/ipu3/ipu3-v4l2.c
b/drivers/staging/media/ipu3/ipu3-v4l2.c
index e530767e80a5..6627b5c2d4d6 100644
--- a/drivers/staging/media/ipu3/ipu3-v4l2.c
+++ b/drivers/staging/media/ipu3/ipu3-v4l2.c
@@ -10,6 +10,8 @@
#include "ipu3.h"
#include "ipu3-dmamap.h"
+#define VB2_MAX_FRAME 32
+
/******************** v4l2_subdev_ops ********************/
#define IPU3_RUNNING_MODE_VIDEO 0
diff --git a/include/media/videobuf2-core.h
b/include/media/videobuf2-core.h
index 77921cf894ef..080b783d608d 100644
--- a/include/media/videobuf2-core.h
+++ b/include/media/videobuf2-core.h
@@ -20,7 +20,6 @@
#include <media/media-request.h>
#include <media/frame_vector.h>
-#define VB2_MAX_FRAME (32)
#define VB2_MAX_PLANES (8)
/**
diff --git a/include/media/videobuf2-v4l2.h
b/include/media/videobuf2-v4l2.h
index 5a845887850b..88a7a565170e 100644
--- a/include/media/videobuf2-v4l2.h
+++ b/include/media/videobuf2-v4l2.h
@@ -15,10 +15,6 @@
#include <linux/videodev2.h>
#include <media/videobuf2-core.h>
-#if VB2_MAX_FRAME != VIDEO_MAX_FRAME
-#error VB2_MAX_FRAME != VIDEO_MAX_FRAME
-#endif
-
#if VB2_MAX_PLANES != VIDEO_MAX_PLANES
#error VB2_MAX_PLANES != VIDEO_MAX_PLANES
#endif
--
2.39.2
--
Hsia-Jun(Randy) Li
--
Hsia-Jun(Randy) Li
--
Hsia-Jun(Randy) Li
--
Hsia-Jun(Randy) Li
