| Message ID | 20200303173904.23492-1-jgross@suse.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v2] xen/sched: fix cpu offlining with core scheduling | expand |
On 03.03.2020 18:39, Juergen Gross wrote: > --- a/xen/common/sched/core.c > +++ b/xen/common/sched/core.c > @@ -2299,6 +2299,10 @@ void sched_context_switched(struct vcpu *vprev, struct vcpu *vnext) > rcu_read_unlock(&sched_res_rculock); > } > > +/* > + * Switch to a new context or keep the current one running. > + * On x86 it won't return, so it will drop the still held sched_res_rculock. > + */ > static void sched_context_switch(struct vcpu *vprev, struct vcpu *vnext, > bool reset_idle_unit, s_time_t now) > { I don't follow the comment: There's return continue_running(vprev); in the function which afaict can happen on all architectures. The lock gets dropped there too. I see no path through this function where the lock would not get dropped. > @@ -2408,6 +2412,9 @@ static struct vcpu *sched_force_context_switch(struct vcpu *vprev, > * zero do_schedule() is called and the rendezvous counter for leaving > * context_switch() is set. All other members will wait until the counter is > * becoming zero, dropping the schedule lock in between. > + * Either returns the new unit to run, or NULL if no context switch is > + * required or (on ARM) has already been performed. If NULL is returned > + * sched_res_rculock has been dropped. I guess official Arm folks would like Arm to not be spelled all upper case anymore. > @@ -2482,6 +2490,21 @@ static struct sched_unit *sched_wait_rendezvous_in(struct sched_unit *prev, > atomic_set(&prev->next_task->rendezvous_out_cnt, 0); > prev->rendezvous_in_cnt = 0; > } > + > + /* > + * Check for scheduling resourced switched. This happens when we are > + * moved away from our cpupool and cpus are subject of the idle > + * scheduler now. > + */ The 'd' on both "resourced" and "switched" are odd to read at least to me, and hence make me uncertain whether I actually correctly understand what is meant here. Jan
On 04.03.20 10:53, Jan Beulich wrote: > On 03.03.2020 18:39, Juergen Gross wrote: >> --- a/xen/common/sched/core.c >> +++ b/xen/common/sched/core.c >> @@ -2299,6 +2299,10 @@ void sched_context_switched(struct vcpu *vprev, struct vcpu *vnext) >> rcu_read_unlock(&sched_res_rculock); >> } >> >> +/* >> + * Switch to a new context or keep the current one running. >> + * On x86 it won't return, so it will drop the still held sched_res_rculock. >> + */ >> static void sched_context_switch(struct vcpu *vprev, struct vcpu *vnext, >> bool reset_idle_unit, s_time_t now) >> { > > I don't follow the comment: There's > > return continue_running(vprev); > > in the function which afaict can happen on all architectures. > The lock gets dropped there too. I see no path through this > function where the lock would not get dropped. It was meant as reasoning: due to the fact that sched_context_switch() won't return on x86, it is required that it will drop sched_res_rculock. > >> @@ -2408,6 +2412,9 @@ static struct vcpu *sched_force_context_switch(struct vcpu *vprev, >> * zero do_schedule() is called and the rendezvous counter for leaving >> * context_switch() is set. All other members will wait until the counter is >> * becoming zero, dropping the schedule lock in between. >> + * Either returns the new unit to run, or NULL if no context switch is >> + * required or (on ARM) has already been performed. If NULL is returned >> + * sched_res_rculock has been dropped. > > I guess official Arm folks would like Arm to not be spelled all > upper case anymore. Okay. > >> @@ -2482,6 +2490,21 @@ static struct sched_unit *sched_wait_rendezvous_in(struct sched_unit *prev, >> atomic_set(&prev->next_task->rendezvous_out_cnt, 0); >> prev->rendezvous_in_cnt = 0; >> } >> + >> + /* >> + * Check for scheduling resourced switched. This happens when we are >> + * moved away from our cpupool and cpus are subject of the idle >> + * scheduler now. >> + */ > > The 'd' on both "resourced" and "switched" are odd to read at > least to me, and hence make me uncertain whether I actually > correctly understand what is meant here. Should be "resources", of course. Juergen
diff --git a/xen/common/sched/core.c b/xen/common/sched/core.c index 7e8e7d2c39..5d8343b327 100644 --- a/xen/common/sched/core.c +++ b/xen/common/sched/core.c @@ -2299,6 +2299,10 @@ void sched_context_switched(struct vcpu *vprev, struct vcpu *vnext) rcu_read_unlock(&sched_res_rculock); } +/* + * Switch to a new context or keep the current one running. + * On x86 it won't return, so it will drop the still held sched_res_rculock. + */ static void sched_context_switch(struct vcpu *vprev, struct vcpu *vnext, bool reset_idle_unit, s_time_t now) { @@ -2408,6 +2412,9 @@ static struct vcpu *sched_force_context_switch(struct vcpu *vprev, * zero do_schedule() is called and the rendezvous counter for leaving * context_switch() is set. All other members will wait until the counter is * becoming zero, dropping the schedule lock in between. + * Either returns the new unit to run, or NULL if no context switch is + * required or (on ARM) has already been performed. If NULL is returned + * sched_res_rculock has been dropped. */ static struct sched_unit *sched_wait_rendezvous_in(struct sched_unit *prev, spinlock_t **lock, int cpu, @@ -2415,7 +2422,8 @@ static struct sched_unit *sched_wait_rendezvous_in(struct sched_unit *prev, { struct sched_unit *next; struct vcpu *v; - unsigned int gran = get_sched_res(cpu)->granularity; + struct sched_resource *sr = get_sched_res(cpu); + unsigned int gran = sr->granularity; if ( !--prev->rendezvous_in_cnt ) { @@ -2482,6 +2490,21 @@ static struct sched_unit *sched_wait_rendezvous_in(struct sched_unit *prev, atomic_set(&prev->next_task->rendezvous_out_cnt, 0); prev->rendezvous_in_cnt = 0; } + + /* + * Check for scheduling resourced switched. This happens when we are + * moved away from our cpupool and cpus are subject of the idle + * scheduler now. + */ + if ( unlikely(sr != get_sched_res(cpu)) ) + { + ASSERT(is_idle_unit(prev)); + atomic_set(&prev->next_task->rendezvous_out_cnt, 0); + prev->rendezvous_in_cnt = 0; + pcpu_schedule_unlock_irq(*lock, cpu); + rcu_read_unlock(&sched_res_rculock); + return NULL; + } } return prev->next_task; @@ -2567,11 +2590,11 @@ static void schedule(void) rcu_read_lock(&sched_res_rculock); + lock = pcpu_schedule_lock_irq(cpu); + sr = get_sched_res(cpu); gran = sr->granularity; - lock = pcpu_schedule_lock_irq(cpu); - if ( prev->rendezvous_in_cnt ) { /* @@ -3151,7 +3174,10 @@ int schedule_cpu_rm(unsigned int cpu) per_cpu(sched_res_idx, cpu_iter) = 0; if ( cpu_iter == cpu ) { - idle_vcpu[cpu_iter]->sched_unit->priv = NULL; + unit = idle_vcpu[cpu_iter]->sched_unit; + unit->priv = NULL; + atomic_set(&unit->next_task->rendezvous_out_cnt, 0); + unit->rendezvous_in_cnt = 0; } else { @@ -3182,6 +3208,8 @@ int schedule_cpu_rm(unsigned int cpu) } sr->scheduler = &sched_idle_ops; sr->sched_priv = NULL; + sr->granularity = 1; + sr->cpupool = NULL; smp_mb(); sr->schedule_lock = &sched_free_cpu_lock; @@ -3194,9 +3222,6 @@ int schedule_cpu_rm(unsigned int cpu) sched_free_udata(old_ops, vpriv_old); sched_free_pdata(old_ops, ppriv_old, cpu); - sr->granularity = 1; - sr->cpupool = NULL; - out: rcu_read_unlock(&sched_res_rculock); xfree(sr_new);
Offlining a cpu with core scheduling active can result in a hanging system. Reason is the scheduling resource and unit of the to be removed cpus needs to be split in order to remove the cpu from its cpupool and move it to the idle scheduler. In case one of the involved cpus happens to have received a sched slave event due to a vcpu former having been running on that cpu being woken up again, it can happen that this cpu will enter sched_wait_rendezvous_in() while its scheduling resource is just about to be split. It might wait for ever for the other sibling to join, which will never happen due to the resources already being modified. This can easily be avoided by: - resetting the rendezvous counters of the idle unit which is kept - checking for a new scheduling resource in sched_wait_rendezvous_in() after reacquiring the scheduling lock and resetting the counters in that case without scheduling another vcpu - moving schedule resource modifications (in schedule_cpu_rm()) and retrieving (schedule(), sched_slave() is fine already, others are not critical) into locked regions Reported-by: Igor Druzhinin <igor.druzhinin@citrix.com> Signed-off-by: Juergen Gross <jgross@suse.com> --- V2: - fix unlocking, add some related comments --- xen/common/sched/core.c | 39 ++++++++++++++++++++++++++++++++------- 1 file changed, 32 insertions(+), 7 deletions(-)