Re: [PATCH 0/8] Generic IRQ entry/exit support for powerpc

[Samir M <samir@linux.ibm.com>]

On 11/11/25 10:09 am, Samir M wrote:
> On 02/11/25 5:23 pm, Mukesh Kumar Chaurasiya wrote:
>> Adding support for the generic irq entry/exit handling for PowerPC. The
>> goal is to bring PowerPC in line with other architectures that already
>> use the common irq entry infrastructure, reducing duplicated code and
>> making it easier to share future changes in entry/exit paths.
>>
>> This is slightly tested of ppc64le and ppc32.
>>
>> The performance benchmarks from perf bench basic syscall are below:
>>
>> | Metric     | W/O Generic Framework | With Generic Framework | Change |
>> | ---------- | --------------------- | ---------------------- | ------ |
>> | Total time | 0.939 [sec]           | 0.938 [sec]            | ~0%    |
>> | usecs/op   | 0.093900              | 0.093882               | ~0%    |
>> | ops/sec    | 1,06,49,615           | 1,06,51,725            | ~0%    |
>>
>> Thats very close to performance earlier with arch specific handling.
>>
>> Tests done:
>>   - Build and boot on ppc64le pseries.
>>   - Build and boot on ppc64le powernv8 powernv9 powernv10.
>>   - Build and boot on ppc32.
>>   - Performance benchmark done with perf syscall basic on pseries.
>>
>> Changelog:
>>
>> RFC -> PATCH
>>   - Fix for ppc32 spitting out kuap lock warnings.
>>   - ppc64le powernv8 crash fix.
>>   - Review comments incorporated from previous RFC.
>> RFC 
>> https://lore.kernel.org/all/20250908210235.137300-2-mchauras@linux.ibm.com/
>>
>> Mukesh Kumar Chaurasiya (8):
>>    powerpc: rename arch_irq_disabled_regs
>>    powerpc: Prepare to build with generic entry/exit framework
>>    powerpc: introduce arch_enter_from_user_mode
>>    powerpc: Introduce syscall exit arch functions
>>    powerpc: add exit_flags field in pt_regs
>>    powerpc: Prepare for IRQ entry exit
>>    powerpc: Enable IRQ generic entry/exit path.
>>    powerpc: Enable Generic Entry/Exit for syscalls.
>>
>>   arch/powerpc/Kconfig                    |   2 +
>>   arch/powerpc/include/asm/entry-common.h | 539 ++++++++++++++++++++++++
>>   arch/powerpc/include/asm/hw_irq.h       |   4 +-
>>   arch/powerpc/include/asm/interrupt.h    | 401 +++---------------
>>   arch/powerpc/include/asm/ptrace.h       |   3 +
>>   arch/powerpc/include/asm/stacktrace.h   |   6 +
>>   arch/powerpc/include/asm/syscall.h      |   5 +
>>   arch/powerpc/include/asm/thread_info.h  |   1 +
>>   arch/powerpc/include/uapi/asm/ptrace.h  |  14 +-
>>   arch/powerpc/kernel/asm-offsets.c       |   1 +
>>   arch/powerpc/kernel/interrupt.c         | 258 +++---------
>>   arch/powerpc/kernel/ptrace/ptrace.c     | 142 +------
>>   arch/powerpc/kernel/signal.c            |   8 +
>>   arch/powerpc/kernel/syscall.c           | 119 +-----
>>   arch/powerpc/kernel/traps.c             |   2 +-
>>   arch/powerpc/kernel/watchdog.c          |   2 +-
>>   arch/powerpc/perf/core-book3s.c         |   2 +-
>>   17 files changed, 693 insertions(+), 816 deletions(-)
>>   create mode 100644 arch/powerpc/include/asm/entry-common.h
>>
> Hi,
>
> I have reviewed and tested the generic IRQ entry/exist patch series. 
> Below are my observations:
>
> 
Test Coverage

> • Successfully ran LTP (specially syscall) and entire LTP test suite, 
> without observing any regressions or issues related to the 
> implementation.
>
> 
System Configuration

> • CPUs: 160

> • Kernel: v6.18.0-rc1+

> • Processor mode: Shared (uncapped)
>
> 
Performance Evaluation

> • Conducted benchmarking using perf bench syscall basic -l and 
> hackbench.

> • No functional regressions observed, and results were consistent with 
> expectations.
>
>     •    Results for perf bench syscall**Loops = 100,000**
> **Loops = 100,000**
> | Metric       | W/O Generic Framework      | With Generic Framework  
>   | Improvement |
> |----------|-----------------------:|-----------------------:|------------:| 
>
> | usecs/op   |              0.125328              | 0.128839         
> |     ~-2.80% |
> | ops/sec     |             7,979,645              |  7,762,047       
>     |     ~-2.73% |
>
> **Loops = 1,000,000**
> | Metric        | W/O Generic Framework         | With Generic 
> Framework             | Improvement |
> |----------|-----------------------:|-----------------------:|------------:| 
>
> | usecs/op   |              0.125015              | 0.127885         
> |     ~-2.30% |
> | ops/sec     |             7,999,051              |  7,819,546       
>     |     ~-2.24% |
>
> **Loops = 10,000,000**
> | Metric        | W/O Generic Framework    | With Generic Framework   
>  | Improvement |
> |----------|-----------------------:|-----------------------:|------------:| 
>
> | usecs/op   |              0.124613              | 0.127426         
> |     ~-2.26% |
> | ops/sec     |             8,024,827              |  7,847,735       
>     |     ~-2.21% |
>
> **Overall (aggregated across all runs)**
> | Metric         | W/O Generic Framework   | With Generic Framework   
>  | Improvement |
> | ---------- | 
> ---------------------:|-----------------------:|------------:|
> | Total time    |           1.384 [sec]            |  1.415 [sec]      
>          |     ~-2.27% |
> | usecs/op     |              0.124656            | 0.127480         
> |     ~-2.27% |
> | ops/sec       |             8,022,098            |  7,844,423       
>     |     ~-2.21% |
>
> A 2% performance degradation was observed with the perf bench syscall.
>
>     •    Results for hackbench
>
> | Metric        | W/O Generic Framework    | With Generic Framework   
> | Improvement |
> |----------|---------------------- 
> :|-----------------------:|------------:|
> | Min Time   | 142.055 (sec).                   | 141.699 (sec)       
>        | 0.25%
> | Max Time  | 143.791 (sec).                   | 143.206 (sec)         
>    | 0.41%
> | Avg Time   | 142.925 (sec)                    | 142.472 (sec)       
>        | 0.32%
>
> So overall 0.3 % improvement is observed across 10 runs.
>
> Please add below tag for the patch set.
> 
Tested-by: Samir M <samir@linux.ibm.com>
> Thank You !!
>
>
> Regards,
> Samir.
>
Hi,

Apologies for the earlier email. The benchmark results table was not 
properly formatted in that version, so I am re-sending the results below 
for clarity.

I have reviewed and tested the generic IRQ entry/exist patch series. 
Below are my observations:


Test Coverage
• Successfully ran LTP (specially syscall) and entire LTP test suite, 
without observing any regressions or issues related to the implementation.


System Configuration
• CPUs: 160
• Kernel: v6.18.0-rc1+
• Processor mode: Shared (uncapped)


Performance Evaluation
• Conducted benchmarking using perf bench syscall basic -l and hackbench.
• No functional regressions observed, and results were consistent with 
expectations.

     •    Results for perf bench syscall

Loops = 100,000
+-----------+------------------------+------------------------+------------+
| Metric      | W/O Generic Framework     | With Generic Framework    | 
Improvement |
+-----------+------------------------+------------------------+------------+
| usecs/op  |           0.125328                 |  0.128839            
        | ~-2.80%     |
| ops/sec    |            7,979,645               |  7,762,047          
         | ~-2.73%     |
+-----------+------------------------+------------------------+------------+

Loops = 1,000,000
+-----------+------------------------+------------------------+------------+
| Metric      | W/O Generic Framework  | With Generic Framework | 
Improvement |
+-----------+------------------------+------------------------+------------+
| usecs/op  |          0.125015               |        0.127885         
         | ~-2.30%     |
| ops/sec    |          7,999,051              |        7,819,546       
            | ~-2.24%     |
+-----------+------------------------+------------------------+------------+

Loops = 10,000,000
+-----------+------------------------+------------------------+------------+
| Metric      | W/O Generic Framework  | With Generic Framework  | 
Improvement |
+-----------+------------------------+------------------------+------------+
| usecs/op  |         0.124613                |        0.127426         
         | ~-2.26%     |
| ops/sec    |         8,024,827               |        7,847,735       
            | ~-2.21%     |
+-----------+------------------------+------------------------+------------+

Overall (aggregated across all runs)
+-------------+------------------------+------------------------+----------+
| Metric         | W/O Generic Framework  | With Generic Framework | 
Improvement |
+-------------+------------------------+------------------------+----------+
| Total time   |        1.384 [sec]               |         1.415 [sec]  
               | ~-2.27%     |
| usecs/op     |        0.124656                 | 0.127480              
      | ~-2.27%     |
| ops/sec       |        8,022,098                | 7,844,423            
       | ~-2.21%     |
+-------------+------------------------+------------------------+----------+
A 2% performance degradation was observed with the perf bench syscall.

     •    Results for hackbench

+-----------+---------------------------+---------------------------+------+
| Metric        | W/O Generic Framework     | With Generic Framework    
| Improvement |
+-----------+---------------------------+---------------------------+------+
| Min Time  |    142.055 (sec)                     |       141.699 
(sec)             |  +0.25%      |
| Max Time  |   143.791 (sec)                     |       143.206 (sec)  
             |  +0.41%      |
| Avg Time  |   142.925 (sec)                      |      142.472 (sec)  
             |  +0.32%      |
+-----------+---------------------------+---------------------------+------+

So overall 0.3 % improvement is observed across 10 runs.

Please add below tag for the patch set.


Tested-by: Samir M <samir@linux.ibm.com>
Thank You !!


Regards,
Samir.