src: fix the bug that tasks enqueue between clear_sched_state() and rebuild_sched_state()

[zxpdemonio]

sched_free_extrapad() is called between clear_sched_state() and rebuild_sched_state() in function __sync_sched_restore(). In function sched_free_extrapad(), free_percpu() may be called, which will result in queue_work_on() called, and a kwork will enqueue, between clear_sched_state() and rebuild_sched_state(), and it will enqueue again in function clear_sched_state(). In this case, cfs->h_nr_running will be wrong.

To fix this bug, we call sched_free_extrapad() later, after rebuild_sched_state().

Reported-by: Cruz Zhao CruzZhao@linux.alibaba.com

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Cruz Zhao seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[anolis-bot]

@zxpdemonio , a new test job has been submitted. Please wait in patience. The test job url: https://tone.openanolis.cn/ws/2r1b4c7z/test_result/72128

[ampresent]

Does sched_alloc_extrapad have the same problem? Even if not, maybe it's a good idea to move it to the bottom of __sync_sched_install. Both to keep behavior consistent, and avoid unnecessary confusion.

[anolis-bot]

@zxpdemonio , The CI test is completed, please check result:

Test Case		Test Result
schedule_test	x86_64	:x: FAIL
schedule_test	aarch64	:white_check_mark: SUCCESS
public_var_test		:white_check_mark: SUCCESS
var_uniformity_test		:white_check_mark: SUCCESS
cpu_throttle_test		:white_check_mark: SUCCESS
domain_rebuild_test		:white_check_mark: SUCCESS
sched_syscall_test		:white_check_mark: SUCCESS
mem_pressure_test		:white_check_mark: SUCCESS

Sorry, your test job failed. Please get the details in the link.

[zxpdemonio]

Does sched_alloc_extrapad have the same problem? Even if not, maybe it's a good idea to move it to the bottom of __sync_sched_install. Both to keep behavior consistent, and avoid unnecessary confusion.

Yeah, it does. In funciont simple_percpu_mempool_create(), if percpu_ptr failed to allocate, free_percpu() will be called. But sched_alloc_extrapad cannot be moved to the bottom of __sync_sched_install(), as the memory may be used in rebuild_sched_state(). We should move it to the top.

[anolis-bot]

@zxpdemonio , the code has been updated, so a new test job has been submitted. Please wait in patience. The test job url: https://tone.openanolis.cn/ws/2r1b4c7z/test_result/72149

[anolis-bot]

@zxpdemonio , The CI test is completed, please check result:

Test Case		Test Result
schedule_test	x86_64	:white_check_mark: SUCCESS
public_var_test		:white_check_mark: SUCCESS
var_uniformity_test		:white_check_mark: SUCCESS
cpu_throttle_test		:white_check_mark: SUCCESS
domain_rebuild_test		:white_check_mark: SUCCESS
sched_syscall_test		:white_check_mark: SUCCESS
mem_pressure_test		:white_check_mark: SUCCESS
schedule_test	aarch64	:white_check_mark: SUCCESS
public_var_test		:white_check_mark: SUCCESS
var_uniformity_test		:white_check_mark: SUCCESS
cpu_throttle_test		:white_check_mark: SUCCESS
domain_rebuild_test		:white_check_mark: SUCCESS
sched_syscall_test		:white_check_mark: SUCCESS
mem_pressure_test		:white_check_mark: SUCCESS

Congratulations, your test job passed!