[bug-report] possible performance problem in ret_to_user_from_irq

[Hui Tang]

On 2023/1/3 22:59, Jens Axboe wrote:
> On 1/3/23 7:34?AM, Russell King (Oracle) wrote:
>> On Tue, Jan 03, 2023 at 07:25:26AM -0700, Jens Axboe wrote:
>>> On 1/3/23 3:06?AM, Russell King (Oracle) wrote:
>>>> On Mon, Dec 26, 2022 at 04:45:20PM +0800, Hui Tang wrote:
>>>>> hi folks.
>>>>>
>>>>> I found a performance problem which is introduced by commit
>>>>> 32d59773da38 ("arm: add support for TIF_NOTIFY_SIGNAL").
>>>>> After the commit,  any bit in the range of 0..15 will cause
>>>>> do_work_pending() to be invoked. More frequent do_work_pending()
>>>>> invoked possible result in worse performance.
>>>>>
>>>>> Some of the tests I've done? as follows:
>>>>> lmbench test			base		with patch
>>>>> ./lat_ctx -P 1 -s 0  2		7.3167		11.04
>>>>> ./lat_ctx -P 1 -s 16 2          8.0467		14.5367
>>>>> ./lat_ctx -P 1 -s 64 2		7.8667		11.43
>>>>> ./lat_ctx -P 1 -s 16 16		16.47		18.3667
>>>>> ./lat_pipe -P 1			28.1671		44.7904
>>>>>
>>>>> libMicro-0.4.1 test		base		with patch
>>>>> ./cascade_cond -E -C 200\
>>>>>  -L -S -W -N "c_cond_1" -I 100	286.3333	358
>>>>>
>>>>> When I adjust test bit, the performance problem gone.
>>>>> -	movs	r1, r1, lsl #16
>>>>> +	ldr	r2, =#_TIF_WORK_MASK
>>>>> +	tst	r1, r2
>>>>>
>>>>> Does anyone have a good suggestion for this problem?
>>>>> should just test _TIF_WORK_MASK, as before?
>>>>
>>>> I think it should be fine - but I would suggest re-organising the
>>>> TIF definitions so that those TIF bits that shouldn't trigger
>>>> do_work_pending are not in the first 16 bits.
>>>>
>>>> Note that all four bits in _TIF_SYSCALL_WORK need to stay within
>>>> an 8-bit even-bit-aligned range, so the value is suitable for an
>>>> immediate assembly constant.
>>>>
>>>> I'd suggest moving the TIF definitions for 20 to 19, and 4..7 to
>>>> 20..23, and then 8 to 4.
>>>
>>> Like this?
>>>
>>> diff --git a/arch/arm/include/asm/thread_info.h b/arch/arm/include/asm/thread_info.h
>>> index aecc403b2880..7f092cb55a41 100644
>>> --- a/arch/arm/include/asm/thread_info.h
>>> +++ b/arch/arm/include/asm/thread_info.h
>>> @@ -128,15 +128,16 @@ extern int vfp_restore_user_hwstate(struct user_vfp *,
>>>  #define TIF_NEED_RESCHED	1	/* rescheduling necessary */
>>>  #define TIF_NOTIFY_RESUME	2	/* callback before returning to user */
>>>  #define TIF_UPROBE		3	/* breakpointed or singlestepping */
>>> -#define TIF_SYSCALL_TRACE	4	/* syscall trace active */
>>> -#define TIF_SYSCALL_AUDIT	5	/* syscall auditing active */
>>> -#define TIF_SYSCALL_TRACEPOINT	6	/* syscall tracepoint instrumentation */
>>> -#define TIF_SECCOMP		7	/* seccomp syscall filtering active */
>>> -#define TIF_NOTIFY_SIGNAL	8	/* signal notifications exist */
>>> +#define TIF_NOTIFY_SIGNAL	4	/* signal notifications exist */
>>>
>>>  #define TIF_USING_IWMMXT	17
>>>  #define TIF_MEMDIE		18	/* is terminating due to OOM killer */
>>> -#define TIF_RESTORE_SIGMASK	20
>>> +#define TIF_RESTORE_SIGMASK	19
>>> +#define TIF_SYSCALL_TRACE	20	/* syscall trace active */
>>> +#define TIF_SYSCALL_AUDIT	21	/* syscall auditing active */
>>> +#define TIF_SYSCALL_TRACEPOINT	22	/* syscall tracepoint instrumentation */
>>> +#define TIF_SECCOMP		23	/* seccomp syscall filtering active */
>>> +
>>>
>>>  #define _TIF_SIGPENDING		(1 << TIF_SIGPENDING)
>>>  #define _TIF_NEED_RESCHED	(1 << TIF_NEED_RESCHED)
>>
>> Yep, LGTM, thanks.
>
> Hui Tang, can you give it a whirl? Just checked and it applies to
> 5.10-stable as well, just with a slight offset.

With the latest patch, the testcase rusults shown in the 'new patch' column.
I also retested previous commit of 32d59773da38, the results shown in the 'base' column.

lmbench test                base        32d59773da38  	new patch
./lat_ctx -P 1 -s 0  2      8.04          11.04		8.25
./lat_ctx -P 1 -s 16 2      9.08          14.5367     	9.26
./lat_ctx -P 1 -s 64 2      8.78          11.43		8.71
./lat_ctx -P 1 -s 16 16     17.22         18.3667     	17.32
./lat_pipe -P 1             43.5021       44.7904     	41.3729

libMicro-0.4.1 test        	  base  32d59773da38    new patch
./cascade_cond -E -C 200\
  -L -S -W -N "c_cond_1" -I 100    281    358     	 281

The performance problem also seem to gone with the latest patch, thanks.