Skip to contentSkip to conversation

media: iris: Fix FPS calculation and VPP FW overhead

PATCH In review Bryan O'Donoghue 4 participants 5 messages Jun 2 – Jun 3

Subsystems (4)
lore
Share & tools
lore

Bryan O'Donoghue

author maintainer
#0 lore raw

PATCH 1/1 media: iris: Fix FPS calculation and VPP FW overhead

Use div_u64() instead of mult_fract as u64 operator division fails on 32 bit systems which don't link against libgcc.

Fixes 5c66647a5c3e ("media: iris: add FPS calculation and VPP FW overhead in frequency formula")

---
Quick fix for u32 bit division error in -next.

---
 drivers/media/platform/qcom/iris/iris_vpu_common.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/media/platform/qcom/iris/iris_vpu_common.c +1 -1
index 5a85568c5ee1c..37dbfe433a085 100644
@@ -444,7 +444,7 @@ u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz
 
 	/* 1.05 is VPP FW overhead */
 	if (inst->fw_caps[STAGE].value == STAGE_2)
-		vpp_cycles += mult_frac(vpp_cycles, 5, 100);
+		vpp_cycles += div_u64(vpp_cycles * 5, 100);
 
 	vsp_cycles = fps * data_size * 8;
 	vsp_cycles = div_u64(vsp_cycles, 2);

---
base-commit: 6a75e3d4f6428b90f398354212e3a2e0172851d6
change-id: 20260602-fixbranch-b5e1adc00e92

Best regards,

Signature · 1 line
Bryan O'Donoghue <[email protected]>
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] 32-bit integer overflow in mult_frac frequency calculation affects both 32-bit and 64-bit platforms.
- [High] Critical 32-bit integer overflow when calculating vsp_cycles leads to severe VPU under-clocking on 32-bit platforms. --

commit 77f6f57aadf558885bf4697f0e3df01315999c7b
Author: Bryan O'Donoghue <[email protected]>

media: iris: Fix FPS calculation and VPP FW overhead

This commit replaces the mult_frac() macro with div_u64() for calculating vpp_cycles. This fixes a compilation issue on 32-bit systems lacking libgcc, where 64-bit operator division would otherwise fail.

Fixes 5c66647a5c3e ("media: iris: add FPS calculation and VPP FW overhead in frequency formula")

drivers/media/platform/qcom/iris/iris_vpu_common.c u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz

Quoted 5 lines
drivers/media/platform/qcom/iris/iris_vpu_common.c +0 -0
index 5a85568c5ee1c..37dbfe433a085 100644
@@ -444,7 +444,7 @@ u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz

[Severity: High]
This is a pre-existing issue, but can the call to mult_frac() for vpp_cycles earlier in this function overflow on both 32-bit and 64-bit platforms?

In iris_vpu3x_vpu4x_calculate_frequency():

    vpp_cycles = mult_frac(mbs_per_second, caps->mb_cycles_vpp, ...);

Because mbs_per_second is a u32, the mult_frac() macro performs the internal multiplication in 32-bit arithmetic.

For high resolutions and framerates natively supported by the driver, mbs_per_second exceeds 21.4M. Multiplying this by caps->mb_cycles_vpp (e.g., 200) produces a value greater than 4.29 billion, wrapping around the maximum 32-bit unsigned limit.

Could this regression result in severely under-provisioning the hardware VPU clock and causing encoding or decoding failures?

drivers/media/platform/qcom/iris/iris_vpu_common.c u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz · line 449

Quoted 7 lines
drivers/media/platform/qcom/iris/iris_vpu_common.c +1 -1
 	/* 1.05 is VPP FW overhead */
 	if (inst->fw_caps[STAGE].value == STAGE_2)
-		vpp_cycles += mult_frac(vpp_cycles, 5, 100);
+		vpp_cycles += div_u64(vpp_cycles * 5, 100);
 
 	vsp_cycles = fps * data_size * 8;

[Severity: High]
This is a pre-existing issue, but does this calculation overflow on 32-bit platforms?

The variable fps is a u32, data_size is a size_t, and 8 is an int. On 32-bit architectures, this multiplication evaluates entirely in 32-bit arithmetic before being zero-extended and assigned to the 64-bit vsp_cycles variable.

For standard 4K 60fps encoding, an uncompressed data_size of around 12.4MB yields 60 * 12441600 * 8 = 5,971,968,000, which wraps around the maximum 32-bit limit.

Since this patch explicitly addresses 32-bit systems, will this regression cause the VSP to be severely under-clocked during common use cases on those platforms?

Quoted 1 line
drivers/media/platform/qcom/iris/iris_vpu_common.c +0 -0
 	vsp_cycles = div_u64(vsp_cycles, 2);
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author maintainer
#2 lore raw
Quoted 64 lines · On 02/06/2026 22:15, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] 32-bit integer overflow in mult_frac frequency calculation affects both 32-bit and 64-bit platforms.
- [High] Critical 32-bit integer overflow when calculating vsp_cycles leads to severe VPU under-clocking on 32-bit platforms. --

commit 77f6f57aadf558885bf4697f0e3df01315999c7b
Author: Bryan O'Donoghue <[email protected]>

media: iris: Fix FPS calculation and VPP FW overhead

This commit replaces the mult_frac() macro with div_u64() for calculating vpp_cycles. This fixes a compilation issue on 32-bit systems lacking libgcc, where 64-bit operator division would otherwise fail.

Fixes 5c66647a5c3e ("media: iris: add FPS calculation and VPP FW overhead in frequency formula")

drivers/media/platform/qcom/iris/iris_vpu_common.c u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz

Quoted 5 lines
drivers/media/platform/qcom/iris/iris_vpu_common.c +0 -0
index 5a85568c5ee1c..37dbfe433a085 100644
@@ -444,7 +444,7 @@ u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz

[Severity: High]
This is a pre-existing issue, but can the call to mult_frac() for vpp_cycles earlier in this function overflow on both 32-bit and 64-bit platforms?

In iris_vpu3x_vpu4x_calculate_frequency():

     vpp_cycles = mult_frac(mbs_per_second, caps->mb_cycles_vpp, ...);

Because mbs_per_second is a u32, the mult_frac() macro performs the internal multiplication in 32-bit arithmetic.

For high resolutions and framerates natively supported by the driver, mbs_per_second exceeds 21.4M. Multiplying this by caps->mb_cycles_vpp (e.g., 200) produces a value greater than 4.29 billion, wrapping around the maximum 32-bit unsigned limit.

Could this regression result in severely under-provisioning the hardware VPU clock and causing encoding or decoding failures?

drivers/media/platform/qcom/iris/iris_vpu_common.c u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz · line 449

Quoted 7 lines
drivers/media/platform/qcom/iris/iris_vpu_common.c +1 -1
  	/* 1.05 is VPP FW overhead */
  	if (inst->fw_caps[STAGE].value == STAGE_2)
-		vpp_cycles += mult_frac(vpp_cycles, 5, 100);
+		vpp_cycles += div_u64(vpp_cycles * 5, 100);
 
  	vsp_cycles = fps * data_size * 8;

[Severity: High]
This is a pre-existing issue, but does this calculation overflow on 32-bit platforms?

The variable fps is a u32, data_size is a size_t, and 8 is an int. On 32-bit architectures, this multiplication evaluates entirely in 32-bit arithmetic before being zero-extended and assigned to the 64-bit vsp_cycles variable.

For standard 4K 60fps encoding, an uncompressed data_size of around 12.4MB yields 60 * 12441600 * 8 = 5,971,968,000, which wraps around the maximum 32-bit limit.

Quoted 8 lines

Since this patch explicitly addresses 32-bit systems, will this regression cause the VSP to be severely under-clocked during common use cases on those platforms?

drivers/media/platform/qcom/iris/iris_vpu_common.c u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz · line 450

Quoted 1 line
drivers/media/platform/qcom/iris/iris_vpu_common.c +0 -0
  	vsp_cycles = div_u64(vsp_cycles, 2);

--
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1

Yeah but no but yeah.

This driver only really runs on 64 bit systems so making it resilient for 32 bit is another day's pile of proverbial to shovel.

---
bod

maintainer Reviewed-by on v1 1/1
#3 lore raw
In reply to Bryan O'Donoghue
Quoted 7 lines · Bryan O'Donoghue

Use div_u64() instead of mult_fract as u64 operator division fails on 32 bit systems which don't link against libgcc.

Fixes 5c66647a5c3e ("media: iris: add FPS calculation and VPP FW overhead in frequency formula")
Quoted 5 lines

---
Quick fix for u32 bit division error in -next.

---
  drivers/media/platform/qcom/iris/iris_vpu_common.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
Reviewed-by on v1 1/1
#4 lore raw
In reply to Bryan O'Donoghue
Quoted 18 lines · Bryan O'Donoghue

Use div_u64() instead of mult_fract as u64 operator division fails on 32 bit systems which don't link against libgcc.

Fixes 5c66647a5c3e ("media: iris: add FPS calculation and VPP FW overhead in frequency formula")

---
Quick fix for u32 bit division error in -next.

---
 drivers/media/platform/qcom/iris/iris_vpu_common.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/media/platform/qcom/iris/iris_vpu_common.c +0 -0
index 5a85568c5ee1c..37dbfe433a085 100644

drivers/media/platform/qcom/iris/iris_vpu_common.c u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz · line 447

Quoted 5 lines
drivers/media/platform/qcom/iris/iris_vpu_common.c +1 -1
@@ -444,7 +444,7 @@ u64 iris_vpu3x_vpu4x_calculate_frequency(struct iris_inst *inst, size_t data_siz
 
 	/* 1.05 is VPP FW overhead */
 	if (inst->fw_caps[STAGE].value == STAGE_2)
-		vpp_cycles += mult_frac(vpp_cycles, 5, 100);
+		vpp_cycles += div_u64(vpp_cycles * 5, 100);
Quoted 1 line
Signature · 2 lines
With best wishes
Dmitry