Hi Waiman
On Tue, May 7, 2024 at 2:04 AM Waiman Long <longman@xxxxxxxxxx> wrote:
>
> On 5/6/24 04:04, Xuewen Yan wrote:
> > Hi Peter
> >
> > On Mon, Apr 29, 2024 at 8:10 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >> On Mon, Apr 29, 2024 at 04:46:33PM +0800, Xuewen Yan wrote:
> >>> The commit 851a723e45d1c("sched: Always clear user_cpus_ptr in do_set_cpus_allowed()")
> >>> would clear the user_cpus_ptr when call the do_set_cpus_allowed.
> >>>
> >>> In order to determine whether the user_cpus_ptr is taking effect,
> >>> it is better to print the task's user_cpus_ptr.
> >> This is an ABI change and would mandate we forever more have this
> >> distinction. I don't think your changes justifies things sufficiently
> >> for this.
> > I added this mainly because online/offline cpu will produce different
> > results for the !top-cpuset task.
> >
> > For example:
> >
> > If the task was running, then offline task's cpus, would lead to clear
> > its user-mask.
> >
> > unisoc:/ # while true; do sleep 600; done&
> > [1] 6786
> > unisoc:/ # echo 6786 > /dev/cpuset/top-app/tasks
> > unisoc:/ # cat /dev/cpuset/top-app/cpus
> > 0-7
> > unisoc:/ # cat /proc/6786/status | grep Cpus
> > Cpus_allowed: ff
> > Cpus_allowed_list: 0-7
> > Cpus_user_allowed: (null)
> > Cpus_user_allowed_list: (null)
> >
> > unisoc:/ # taskset -p c0 6786
> > pid 6786's current affinity mask: ff
> > pid 6786's new affinity mask: c0
> > unisoc:/ # cat /proc/6786/status | grep Cpus
> > Cpus_allowed: c0
> > Cpus_allowed_list: 6-7
> > Cpus_user_allowed: c0
> > Cpus_user_allowed_list: 6-7
> >
> > After offline the cpu6 and cpu7, the user-mask would be cleared:
> >
> > unisoc:/ # echo 0 > /sys/devices/system/cpu/cpu7/online
> > unisoc:/ # cat /proc/6786/status | grep Cpus
> > Cpus_allowed: 40
> > Cpus_allowed_list: 6
> > Cpus_user_allowed: c0
> > Cpus_user_allowed_list: 6-7
> > ums9621_1h10:/ # echo 0 > /sys/devices/system/cpu/cpu6/online
> > ums9621_1h10:/ # cat /proc/6786/status | grep Cpus
> > Cpus_allowed: 3f
> > Cpus_allowed_list: 0-5
> > Cpus_user_allowed: (null)
> > Cpus_user_allowed_list: (null)
> >
> > When online the cpu6/7, the user-mask can not bring back:
> >
> > unisoc:/ # echo 1 > /sys/devices/system/cpu/cpu6/online
> > unisoc:/ # cat /proc/6786/status | grep Cpus
> > Cpus_allowed: 7f
> > Cpus_allowed_list: 0-6
> > Cpus_user_allowed: (null)
> > Cpus_user_allowed_list: (null)
> > unisoc:/ # echo 1 > /sys/devices/system/cpu/cpu7/online
> > unisoc:/ # cat /proc/6786/status | grep Cpus
> > Cpus_allowed: ff
> > Cpus_allowed_list: 0-7
> > Cpus_user_allowed: (null)
> > Cpus_user_allowed_list: (null)
> >
> > However, if we offline the cpu when the task is sleeping, at this
> > time, because would not call the fallback_cpu(), its user-mask will
> > not be cleared.
> >
> > unisoc:/ # while true; do sleep 600; done&
> > [1] 5990
> > unisoc:/ # echo 5990 > /dev/cpuset/top-app/tasks
> > unisoc:/ # cat /proc/5990/status | grep Cpus
> > Cpus_allowed: ff
> > Cpus_allowed_list: 0-7
> > Cpus_user_allowed: (null)
> > Cpus_user_allowed_list: (null)
> >
> > unisoc:/ # taskset -p c0 5990
> > pid 5990's current affinity mask: ff
> > pid 5990's new affinity mask: c0
> > unisoc:/ # cat /proc/5990/status | grep Cpus
> > Cpus_allowed: c0
> > Cpus_allowed_list: 6-7
> > Cpus_user_allowed: c0
> > Cpus_user_allowed_list: 6-7
> >
> > unisoc:/ # echo 0 > /sys/devices/system/cpu/cpu6/online
> > unisoc:/ # cat /proc/5990/status | grep Cpus
> > Cpus_allowed: 80
> > Cpus_allowed_list: 7
> > Cpus_user_allowed: c0
> > Cpus_user_allowed_list: 6-7
> > unisoc:/ # echo 0 > /sys/devices/system/cpu/cpu7/online
> > unisoc:/ # cat /proc/5990/status | grep Cpus
> > Cpus_allowed: 3f
> > Cpus_allowed_list: 0-5
> > Cpus_user_allowed: c0
> > Cpus_user_allowed_list: 6-7
> >
> >
> > After 10 minutes, it was waked up, it can also keep its user-mask:
> > ums9621_1h10:/ # cat /proc/5990/status | grep Cpus
> > Cpus_allowed: 3f
> > Cpus_allowed_list: 0-5
> > Cpus_user_allowed: c0
> > Cpus_user_allowed_list: 6-7
> >
> > In order to solve the above problem, I modified the following patch.
> > At this time, for !top-cpuset, regardless of whether the task is in
> > the running state when offline cpu, its cpu-mask can be maintained.
> > However, this patch may not be perfect yet, so I send the "Print
> > user_cpus_ptr for task status" patch first to debug more conveniently.
> >
> > --->
> >
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index 68cfa656b9b1..00879b6de8d4 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -1870,7 +1870,7 @@ extern void dl_bw_free(int cpu, u64 dl_bw);
> > #ifdef CONFIG_SMP
> >
> > /* do_set_cpus_allowed() - consider using set_cpus_allowed_ptr() instead */
> > -extern void do_set_cpus_allowed(struct task_struct *p, const struct
> > cpumask *new_mask);
> > +extern void do_set_cpus_allowed(struct task_struct *p, const struct
> > cpumask *new_mask, bool keep_user);
> >
> > /**
> > * set_cpus_allowed_ptr - set CPU affinity mask of a task
> > @@ -1886,7 +1886,7 @@ extern int dl_task_check_affinity(struct
> > task_struct *p, const struct cpumask *m
> > extern void force_compatible_cpus_allowed_ptr(struct task_struct *p);
> > extern void relax_compatible_cpus_allowed_ptr(struct task_struct *p);
> > #else
> > -static inline void do_set_cpus_allowed(struct task_struct *p, const
> > struct cpumask *new_mask)
> > +static inline void do_set_cpus_allowed(struct task_struct *p, const
> > struct cpumask *new_mask, bool keep_user)
> > {
> > }
> > static inline int set_cpus_allowed_ptr(struct task_struct *p, const
> > struct cpumask *new_mask)
> > diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> > index 7ee9994aee40..0c448f8a3829 100644
> > --- a/kernel/cgroup/cpuset.c
> > +++ b/kernel/cgroup/cpuset.c
> > @@ -4005,9 +4005,14 @@ bool cpuset_cpus_allowed_fallback(struct
> > task_struct *tsk)
> >
> > rcu_read_lock();
> > cs_mask = task_cs(tsk)->cpus_allowed;
> > - if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
> > - do_set_cpus_allowed(tsk, cs_mask);
> > - changed = true;
> > + if (cpumask_subset(cs_mask, possible_mask)) {
> > + if (is_in_v2_mode()) {
> > + do_set_cpus_allowed(tsk, cs_mask, false);
> > + changed = true;
> > + } else if (task_cs(tsk) != &top_cpuset) {
> > + do_set_cpus_allowed(tsk, cs_mask, true);
> > + changed = true;
> > + }
> > }
> > rcu_read_unlock();
> >
> > diff --git a/kernel/kthread.c b/kernel/kthread.c
> > index 7a7aa5f93c0c..7ede27630088 100644
> > --- a/kernel/kthread.c
> > +++ b/kernel/kthread.c
> > @@ -527,7 +527,7 @@ static void __kthread_bind_mask(struct task_struct
> > *p, const struct cpumask *mas
> >
> > /* It's safe because the task is inactive. */
> > raw_spin_lock_irqsave(&p->pi_lock, flags);
> > - do_set_cpus_allowed(p, mask);
> > + do_set_cpus_allowed(p, mask, false);
> > p->flags |= PF_NO_SETAFFINITY;
> > raw_spin_unlock_irqrestore(&p->pi_lock, flags);
> > }
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 33cfd522fc7c..623f89e65e6c 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -2855,18 +2855,21 @@ __do_set_cpus_allowed(struct task_struct *p,
> > struct affinity_context *ctx)
> > * Used for kthread_bind() and select_fallback_rq(), in both cases the user
> > * affinity (if any) should be destroyed too.
> > */
> > -void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
> > +void do_set_cpus_allowed(struct task_struct *p, const struct cpumask
> > *new_mask, bool keep_user)
> > {
> > struct affinity_context ac = {
> > .new_mask = new_mask,
> > .user_mask = NULL,
> > - .flags = SCA_USER, /* clear the user requested mask */
> > + .flags = 0, /* clear the user requested mask */
> > };
> > union cpumask_rcuhead {
> > cpumask_t cpumask;
> > struct rcu_head rcu;
> > };
> >
> > + if (!keep_user)
> > + ac.flags = SCA_USER;
> > +
> > __do_set_cpus_allowed(p, &ac);
> >
> > /*
> > @@ -2874,7 +2877,8 @@ void do_set_cpus_allowed(struct task_struct *p,
> > const struct cpumask *new_mask)
> > * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
> > * kfree_rcu().
> > */
> > - kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
> > + if (!keep_user)
> > + kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
> > }
> >
> > static cpumask_t *alloc_user_cpus_ptr(int node)
> > @@ -3664,7 +3668,7 @@ int select_fallback_rq(int cpu, struct task_struct *p)
> > *
> > * More yuck to audit.
> > */
> > - do_set_cpus_allowed(p, task_cpu_possible_mask(p));
> > + do_set_cpus_allowed(p,
> > task_cpu_possible_mask(p), false);
> > state = fail;
> > break;
> > case fail:
> >
> These changes essentially reverts commit 851a723e45d1c("sched: Always
> clear user_cpus_ptr in do_set_cpus_allowed()") except the additional
> caller in the cpuset code.
>
> How about the following less invasive change?
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 7019a40457a6..646837eab70c 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2796,21 +2796,24 @@ __do_set_cpus_allowed(struct task_struct *p,
> struct affinity_context *ctx)
> }
>
> /*
> - * Used for kthread_bind() and select_fallback_rq(), in both cases the user
> - * affinity (if any) should be destroyed too.
> + * Used for kthread_bind() and select_fallback_rq(). Destroy user affinity
> + * if no intersection with the new mask.
> */
> void do_set_cpus_allowed(struct task_struct *p, const struct cpumask
> *new_mask)
> {
> struct affinity_context ac = {
> .new_mask = new_mask,
> .user_mask = NULL,
> - .flags = SCA_USER, /* clear the user requested mask */
> + .flags = 0,
> };
> union cpumask_rcuhead {
> cpumask_t cpumask;
> struct rcu_head rcu;
> };
>
> + if (current->user_cpus_ptr &&
> !cpumask_intersects(current->user_cpus_ptr, new_mask))
Thanks for your suggestion, and I try it and as for me, it works well,
but I change the "current" to p.
I think “current” is inappropriate because what is changed here is the
mask of p.
It is possible that “p” and “current” are not equal.
I would send the next patch later and add your Suggested-by. Thanks
again for your advice!
BR
---
xuewen
> + ac.flags = SCA_USER; /* clear the user requested mask */
> +
> __do_set_cpus_allowed(p, &ac);
>
> /*
>
> No compilation test done. Note that there is a null check inside
> kfree_rcu() with no need for additional check.
>
> Regards,
> Longman
>
>
