轉自:https://blog.csdn.net/dog250/article/details/105756168
大約10年前,我寫過兩篇關於Linux內核CFS調度器的文章:
https://blog.csdn.net/dog250/article/details/5302865
https://blog.csdn.net/dog250/article/details/5302864
我覺得這兩篇文章是垃圾,但我又不刪,留着給自己噴吧!
不就是一個內核參數 kernel.sched_child_runs_first 嗎?在今天看來,驗證它是否起作用實在太簡單了。
首先解釋一下 為什么要子進程先運行 。
因為fork的行為造成了后續的COW(copy on write),一般而言子進程會調用exec而替換掉需要COW的地址空間,子進程先運行可以避免不必要的COW開銷。
那么對於CFS調度器而言,kernel.sched_child_runs_first是否有作用呢?我們試一下便知道,依然使用那兩篇垃圾文章中的例子:
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <unistd.h>
int main(int argc,char *argv[])
{
int v = atoi(argv[1]);
printf("%d\n", getpid());
nice(v);
int i = 90000;
while (i-->0) {
v++;
}
if(fork() == 0) {
printf("sub\n");
}
printf("main,%d\n",v);
}
我們設置好內核參數后,看看到底哪個先打印出來:
[root@localhost test]# sysctl -w kernel.sched_child_runs_first=1
kernel.sched_child_runs_first = 1
[root@localhost test]#
[root@localhost test]# ./a.out 10
5101
main,90010
[root@localhost test]# sub
[root@localhost test]# ./a.out -10
5105
sub
main,89990
[root@localhost test]# ./a.out -10
5108
main,89990
[root@localhost test]# sub
[root@localhost test]# ./a.out -10
5112
main,89990
[root@localhost test]# sub
[root@localhost test]# ./a.out 10
5117
main,90010
[root@localhost test]# sub
不用試了,它不起作用,不管你有沒有設置START_DEBIT這個feature!它和START_DEBIT根本沒有關系,dog250在2010年寫的那些東西故弄玄虛,把簡單問題復雜化!還扯什么START_DEBIT,還扯什么統計概覽,真是無中生有,垃圾啊垃圾。
正確的排查問題的方法完全就不是這個思路!
現在,我來展示正確的做法。在實驗之前,澄清一個事實, 不要用printf來確認到底誰先運行! 因為printf太復雜了,執行它的周期太久,有可能雖然子進程先運行但卻是父進程先打印出來。
所以,我用exit:
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <unistd.h>
int main(int argc,char *argv[])
{
int v = atoi(argv[1]);
printf("%d\n", getpid());
nice(v);
if(fork() == 0) {
exit(0);
}
exit(0);
}
我們用操作系統的方式去觀測,而不是用printf,這次,我們用stap:
#!/usr/bin/stap -g
global g_se;
global g_cfs_rq;
probe begin {
g_cfs_rq = 0;
g_se = 0;
}
probe kernel.function("__schedule")
{
t_curr = task_current();
if (task_execname(t_curr) == "a.out")
printf("[_schedule] current task: %s[%d]\n", task_execname(t_curr), task_pid(t_curr));
}
probe kernel.function("do_exit")
{
t_curr = task_current();
if (task_execname(t_curr) == "a.out")
printf("Exit task: %s[%d]\n", task_execname(t_curr), task_pid(t_curr));
}
probe kernel.function("pick_next_task_fair")
{
g_cfs_rq = &$rq->cfs;
}
function container_of_entity:long(se:long)
{
offset = &@cast(0, "struct task_struct")->se;
return se - offset;
}
probe kernel.function("pick_next_task_fair").return
{
if($return != 0) {
se = &$return->se;
t_se = container_of_entity(se);
t_curr = task_current();
if (task_execname(t_se) == "a.out" || task_execname(t_curr) == "a.out") {
printf("[pick_next_task_fair] Return task: %s[%d] From current: %s[%d]\n", task_execname(t_se), task_pid(t_se), task_execname(t_curr), task_pid(t_curr));
}
}
}
probe kernel.function("wake_up_new_task")
{
g_se = &$p->se;
g_cfs_rq = @cast(g_se, "struct sched_entity")->cfs_rq;
}
probe kernel.function("wake_up_new_task").return
{
t_se = container_of_entity(g_se);
tname = task_execname(t_se);
vruntime = @cast(g_se, "struct sched_entity")->vruntime;
if (tname == "a.out") {
curr = @cast(g_cfs_rq, "struct cfs_rq")->curr;
t_curr = container_of_entity(curr);
curr_vruntime = @cast(curr, "struct sched_entity")->vruntime;
printf("[wake_up_new_task] current:[%s][%d] curr:%d new:%d del:%d\n",
task_execname(t_curr), task_pid(t_curr), curr_vruntime, vruntime,
curr_vruntime - vruntime);
}
g_se = 0;
g_cfs_rq = 0;
}
probe kernel.function("place_entity")
{
t_initial = $initial;
if (t_initial == 1) {
g_cfs_rq = $cfs_rq;
g_se = $se;
}
}
probe kernel.function("place_entity").return
{
if (g_se) {
t_se = container_of_entity(g_se);
tname = task_execname(t_se);
vruntime = @cast(g_se, "struct sched_entity")->vruntime;
if (tname == "a.out") {
curr = @cast(g_cfs_rq, "struct cfs_rq")->curr;
t_curr = container_of_entity(curr);
curr_vruntime = @cast(curr, "struct sched_entity")->vruntime;
printf("[place_entity] name:[%s][%d] curr:%d new:%d delta:%d\n",
task_execname(t_curr), task_pid(t_curr), curr_vruntime, vruntime,
curr_vruntime - vruntime);
}
g_se = 0;
g_cfs_rq = 0;
}
}
執行它,然后運行多次a.out,到底發生了什么,你就徹底知道了,下面是一個結果:
[root@localhost test]# ./a.out 10
5653
main,90010
[root@localhost test]# sub
# 另一個終端上打印的stap信息
[_schedule] current task: a.out[5653]
[pick_next_task_fair] Return task: a.out[5653] From current: a.out[5653]
# 父進程fork子進程,並設置了它的初始vruntime。
# 后續的child runs first檢查會resched current
[place_entity] name:[a.out][5653] curr:74161009564 new:74192039854 delta:-31030290
# 注意,這里在fork中發生了切換,why??因為在fork中spin_unlock的時候會check resched!
# 這就發生了task_fork_fair最后釋放rq lock時!
[_schedule] current task: a.out[5653]
[pick_next_task_fair] Return task: rcu_sched[10] From current: a.out[5653]
[pick_next_task_fair] Return task: a.out[5653] From current: sshd[1392]
# 父進程返回運行,wakeup子進程,然而vruntime的delta卻不足一個granularity!
# 不足一個granularity,子進程無法搶占父進程!
# 換句話說,之前由於child runs first進行的resched已經失效!
[wake_up_new_task] current:[a.out][5653] curr:74192443179 new:74192132619 del:310560
# 依然是父進程先運行。
Exit task: a.out[5653]
[_schedule] current task: a.out[5653]
[pick_next_task_fair] Return task: rcu_sched[10] From current: a.out[5653]
[pick_next_task_fair] Return task: a.out[5654] From current: sshd[1392]
# 子進程被調度運行
Exit task: a.out[5654]
很尷尬的事發生了,為了child runs first而執行resched_task,需要lock住rq,之所以要resched_task可能是交換父子的vruntime之后,希望子進程繼續運行下去,替換父進程。
然而unlock rq時的check preempt卻白白消耗了這次resched的機會!為什么說白白消耗呢?因為調用sched_fork的時候,子進程尚未准備好,也就是說,它尚不足以被wakeup!
只要在rq unlock時check preempt時候,父進程被其它進程搶占(在父進程優先級低時更容易發生!!),那么子進程大概率不會runs first了,因為后面還要check granularity!
如果我們用更小的nice值運行a.out,那么子進程還是有機會runs first的,因為在rq unlock的時候,父進程不容易被其它進程搶占進而消費掉resched的機會,留到后面wakeup child的時候,還可以使用,此時子進程已經擁有了執行的條件,進而搶占掉父進程!
或者說,即便這次搶占父進程失敗,那么它的vruntime已經低於父進程,它在紅黑樹中的位置是比父進程更加leftmost的,終究還是要比父進程runs first。
好了,情景分析完畢,該解題了!如何讓kernel.sched_child_runs_first如其意之所表達,真正做到child runs first呢?
2010年dog250寫的那個patch是錯的,沒有意義。真正的解法是:
在wakeup child的時候再check sched_child_runs_first,進而resched。
我們來驗證一下,由於我懶得為這個重新編譯一遍內核,所以我采用stap guru hook的方式來玩玩。
首先我們廢除原始的sched_child_runs_first判斷,這很容易,關掉這個開關即可:
[root@localhost test]# sysctl -w kernel.sched_child_runs_first=0
kernel.sched_child_runs_first = 0
1
2
然后,我們以guru模式運行下面的stap腳本:
#!/usr/bin/stap -g
global g_p;
probe begin {
g_p = 0;
}
%{
static void *(*_resched_task)(struct task_struct *p);
%}
function resched(tsk:long, tskp:long)
%{
struct task_struct *task = NULL, *parent = NULL;
struct sched_entity *pse = NULL, *cse = NULL;
task = (struct task_struct *)STAP_ARG_tsk;
parent = (struct task_struct *)STAP_ARG_tskp;
cse = &task->se;
pse = &parent->se;
if (_resched_task == NULL)
_resched_task = (void *)kallsyms_lookup_name("resched_task");
if (_resched_task && pse->vruntime < cse->vruntime) {
swap(pse->vruntime, cse->vruntime);
STAP_PRINTF("---[%lu]------[%lu]-------\n", pse->vruntime, cse->vruntime);
_resched_task(current);
}
%}
probe kernel.function("check_preempt_wakeup")
{
g_p = $p;
}
// 這里的trick在於,由於我們的父子a.out都是純CPU型的,只在創建時被wakeup一次,所以hook該點。
probe kernel.function("check_preempt_wakeup").return
{
parent = @cast(g_p, "struct task_struct")->parent;
// 這里過濾掉了除了我們的fork場景之外的所有其它的wakeup場景。
if (task_execname(g_p) == "a.out" || task_execname(parent) == "a.out") {
resched(g_p, parent);
}
g_p = 0;
}
來吧,執行之!為了觀測效果,我們可以再次同時執行之前的腳本(hook不要沖突即可):
[root@localhost test]# ./a.out 10
6988
sub
main,90010
# 以下是輸出
[_schedule] current task: a.out[6988]
[pick_next_task_fair] Return task: kworker/0:0[5380] From current: a.out[6988]
[pick_next_task_fair] Return task: a.out[6988] From current: sshd[1392]
[_schedule] current task: a.out[6988]
[pick_next_task_fair] Return task: rcu_sched[10] From current: a.out[6988]
[pick_next_task_fair] Return task: a.out[6988] From current: sshd[1392]
[place_entity] name:[a.out][6988] curr:78347075684 new:78440166555 delta:-93090871
# 在place_entity和wake_up_new_task之間沒有被打斷!
# 因為把resched從place移動到了wakeup的時候。
[wake_up_new_task] current:[a.out][6988] curr:78440166555 new:78347480815 del:92685740
[_schedule] current task: a.out[6988]
[pick_next_task_fair] Return task: a.out[6989] From current: a.out[6988]
# 子進程runs first,優先退出!
Exit task: a.out[6989]
[_schedule] current task: a.out[6989]
[pick_next_task_fair] Return task: kworker/0:0[5380] From current: a.out[6989]
[pick_next_task_fair] Return task: a.out[6988] From current: sshd[1392]
[_schedule] current task: a.out[6988]
[pick_next_task_fair] Return task: kworker/0:0[5380] From current: a.out[6988]
[pick_next_task_fair] Return task: a.out[6988] From current: sshd[1392]
# 父進程在后
Exit task: a.out[6988]
[_schedule] current task: a.out[6988]
[pick_next_task_fair] Return task: systemd[1] From current: a.out[6988]
多試幾次,還是這樣的結果。
現在,是時候回到printf了,雖然它可能並不准,但是肉眼觀測,讓經理信服,也只能靠它了:
[root@localhost test]# ./a.out 18
sub
main,90018
[root@localhost test]# ./a.out -18
sub
main,89982
[root@localhost test]# ./a.out -10
sub
main,89990
[root@localhost test]# ./a.out 10
sub
main,90010
[root@localhost test]# ./a.out 1
sub
main,90001
[root@localhost test]# ./a.out -1
sub
main,89999
[root@localhost test]# ./a.out 0
sub
main,90000
咋試咋舒服!
好了,現在該出patch了。這個patch才是真的有效的:
Date: Thu, 23 Apr 2020 22:42:07 +0800
Subject: [PATCH] fix child runs first
---
kernel/sched/fair.c | 20 +++++++++++---------
1 file changed, 11 insertions(+), 9 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 6a33137..f7f83a3 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4564,6 +4564,17 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_
int scale = cfs_rq->nr_running >= sched_nr_latency;
int next_buddy_marked = 0;
+ if ((wake_flags&WF_FORK) && sysctl_sched_child_runs_first && se &&
+ entity_before(se, pse)) {
+ /*
+ * Upon rescheduling, sched_class::put_prev_task() will place
+ * 'current' within the tree based on its new key value.
+ */
+ swap(se->vruntime, pse->vruntime);
+ resched_task(curr);
+ }
+
+
if (unlikely(se == pse))
return;
@@ -7086,15 +7097,6 @@ static void task_fork_fair(struct task_struct *p)
se->vruntime = curr->vruntime;
place_entity(cfs_rq, se, 1);
- if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
- /*
- * Upon rescheduling, sched_class::put_prev_task() will place
- * 'current' within the tree based on its new key value.
- */
- swap(curr->vruntime, se->vruntime);
- resched_task(rq->curr);
- }
-
se->vruntime -= cfs_rq->min_vruntime;
raw_spin_unlock_irqrestore(&rq->lock, flags);
--
1.8.3.1
對了,為了讓這個patch不是真正可以打入內核的,我特意在老舊的內核上制作了這個patch。真正手藝人的玩法永遠不是制作真正的patch,二進制hook不好嗎?哈哈!
浙江溫州皮鞋濕,下雨進水不會胖。
————————————————
版權聲明:本文為CSDN博主「dog250」的原創文章,遵循CC 4.0 BY-SA版權協議,轉載請附上原文出處鏈接及本聲明。
原文鏈接:https://blog.csdn.net/dog250/article/details/105756168