[Devel] [PATCH rh8 5/8] ve/proc/stat: Introduce /proc/stat virtualized handler for Containers
Konstantin Khorenko
khorenko at virtuozzo.com
Tue Nov 3 16:22:16 MSK 2020
has been merged into
[PATCH RHEL8 COMMIT] ve/sched/stat: Introduce functions to calculate vcpustat data
--
Best regards,
Konstantin Khorenko,
Virtuozzo Linux Kernel Team
On 10/28/2020 06:57 PM, Konstantin Khorenko wrote:
> vz8 rebase notes:
> * "swap 0 0" line has been dropped
> * extra empty line between "intr" and "ctxt" has been dropped
>
> Known issues:
> - it's known "procs_blocked" is shown incorrectly inside a CT
> TODO: to fix
>
> Signed-off-by: Konstantin Khorenko <khorenko at virtuozzo.com>
>
> Commit messages of related patches in vz7:
> ==========================================
> sched: use cpuacct->cpustat for showing cpu stats
>
> In contrast to RH6, where tg->cpustat was used, now cpu stats are
> accounted in the cpuacct cgroup. So zap tg->cpustat and use
> cpuacct->cpustat for showing cpu.proc.stat instead. Fortunately cpu and
> cpuacct cgroups are always mounted together (even by systemd by
> default), so this will work.
>
> Related to https://jira.sw.ru/browse/PSBM-33642
>
> Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
>
> +++
> Use ve init task's css instead of opening cgroup via vfs
>
> Currently, whenever we need to get cpu or devices cgroup corresponding
> to a ve, we open it using cgroup_kernel_open(). This is inflexible,
> because it relies on the fact that all container cgroups are located at
> a specific location which can never change (at the top level). Since we
> want to move container cgroups to machine.slice, we need to rework this.
>
> This patch does the trick. It makes each ve remember its init task at
> container start, and use css corresponding to init task whenever we need
> to get a corresponding cgroup. Note, that after this patch is applied,
> we don't need to mount cpu and devices cgroup in kernel.
>
> https://jira.sw.ru/browse/PSBM-48629
>
> Signed-off-by: Vladimir Davydov <vdavydov at virtuozzo.com>
>
> +++
> ve/cpustat: don't try to update vcpustats for root_task_group
>
> root_task_group doesn't have vcpu stats. Attempt to update them leads
> to NULL-ptr deref:
>
> BUG: unable to handle kernel NULL pointer dereference at (null)
> IP: [<ffffffff810b440c>] cpu_cgroup_update_vcpustat+0x13c/0x620
> ...
> Call Trace:
> [<ffffffff810bee3b>] cpu_cgroup_get_stat+0x7b/0x180
> [<ffffffff810f1ef7>] ve_get_cpu_stat+0x27/0x70
> [<ffffffffa01836a1>] fill_cpu_stat+0x91/0x1e0 [vzmon]
> [<ffffffffa0183c6b>] vzcalls_ioctl+0x2bb/0x430 [vzmon]
> [<ffffffffa018d0d5>] vzctl_ioctl+0x45/0x60 [vzdev]
> [<ffffffff8120cfb5>] do_vfs_ioctl+0x255/0x4f0
> [<ffffffff8120d2a4>] SyS_ioctl+0x54/0xa0
> [<ffffffff81642ac9>] system_call_fastpath+0x16/0x1b
>
> So, return -ENOENT if we asked for vcpu stats of root_task_group.
>
> https://jira.sw.ru/browse/PSBM-48721
>
> Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
> Reviewed-by: Vladimir Davydov <vdavydov at virtuozzo.com>
>
> +++
> sched: Port cpustat related patches
>
> This patch ports:
>
> diff-sched-rework-_proc_stat-output
> diff-sched-fix-output-of-vestat-idle
> diff-sched-make-allowance-for-vcpu-rate-in-_proc_stat
> diff-sched-hide-steal-time-from-inside-CT
> diff-sched-cpu.proc.stat-always-count-nr_running-and-co-on-all-cpus
>
> Author: Vladimir Davydov
> Email: vdavydov at parallels.com
> Subject: sched: rework /proc/stat output
> Date: Thu, 29 May 2014 11:40:50 +0400
>
> Initially we mapped usage pct on physical cpu i to vcpu (i % nr_vcpus).
> Obviously, if a CT is running on physical cpus equal mod nr_vcpus, we'll
> miss usage on one or more vcpus. F.e., if there is a 2-vcpus CT with
> several cpu-eaters running on physical cpus 0, 2, we'll get vcpu 0
> 200%-busy, and vcpu 1 idling.
>
> To fix that, we changed behavior so that vcpu i usage equals total cpu
> usage divided by nr_vcpus. That led to customers' dissatisfaction,
> because such an algorithm reveals the fake.
>
> So, now we're going to use the first algorithm, but if the usage of one
> of vcpus turns out to be greater than abs time delta, we'll "move" the
> usage excess to other vcpus, so that one vcpu will never consume more
> than one pcpu. F.e., in the situation described above, we'll move 100%
> of vcpu 0 time to vcpu 1, so that both vcpus will be 100%-busy.
>
> To achieve that, we serialize access to /proc/stat and making readers
> update stats basing on the pcpu usage delta, so that it can fix up per
> vcpu usage to be <= 100% and calculate idle time accordingly.
>
> https://jira.sw.ru/browse/PSBM-26714
>
> Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
>
> Acked-by: Kirill Tkhai <ktkhai at parallels.com>
> =============================================================================
>
> Author: Vladimir Davydov
> Email: vdavydov at parallels.com
> Subject: sched: fix output of vestat:idle
> Date: Tue, 22 Jul 2014 12:25:24 +0400
>
> /proc/vz/vestat must report virtualized idle time, but since commit
> diff-sched-rework-_proc_stat-output it shows total time CTs have been
> idling on all physical cpus. This is, because in cpu_cgroup_get_stat we
> use task_group->cpustat instead of vcpustat. Fix it.
>
> https://jira.sw.ru/browse/PSBM-28403
> https://bugzilla.openvz.org/show_bug.cgi?id=3035
>
> Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
>
> Acked-by: Kirill Tkhai <ktkhai at parallels.com>
> =============================================================================
>
> Author: Vladimir Davydov
> Email: vdavydov at parallels.com
> Subject: sched: make allowance for vcpu rate in /proc/stat
> Date: Wed, 13 Aug 2014 15:44:36 +0400
>
> Currently, if cpulimit < cpus * 100 for a CT we can't get 100% cpu usage
> as reported by the CT's /proc/stat. This is, because when reporting cpu
> usage statistics we consider the maximal cpu time a CT can get to be
> equal to cpus * 100, so that in the above-mentioned setup there will
> always be idle time. It confuses our customers, because before commit
> diff-sched-rework-_proc_stat-output it was possible to get 100% usage
> inside a CT irrespective of its cpulimit/cpus setup. So let's fix it by
> defining the maximal cpu usage a CT can get to be equal to cpulimit.
>
> https://jira.sw.ru/browse/PSBM-28500
>
> Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
>
> Acked-by: Kirill Tkhai <ktkhai at parallels.com>
> =============================================================================
>
> Author: Kirill Tkhai
> Email: ktkhai at parallels.com
> Subject: sched: Hide steal time from inside CT
> Date: Tue, 6 May 2014 18:27:40 +0400
>
> https://jira.sw.ru/browse/PSBM-26587
>
>>From the BUG by Khorenko Konstantin:
>
> "FastVPS complains on incorrect idle time calculation
> (PSBM-23431) and about high _steal_ time reported inside a CT.
> Steal time is a time when a CT was ready to run on a physical CPU,
> but the CPU was busy with processes which belong to another CT.
> => in case we have 10 CTs which eat cpu time as much as possible,
> then steal time in each CT will be 90% and execution time 10% only.
>
> i suggest to make steal time always shown as 0 inside Containers in
> order not to confuse end users. At the same time steal time for
> Containers should be visible from the HN (host), this is useful for
> development.
>
> No objections neither from Pasha Emelyanov nor from Maik".
>
> So, we do this for not to scare users.
>
> Signed-off-by: Kirill Tkhai <ktkhai at parallels.com>
>
> Acked-by: Vladimir Davydov <vdavydov at parallels.com>
> =============================================================================
>
> Author: Vladimir Davydov
> Email: vdavydov at parallels.com
> Subject: sched: cpu.proc.stat: always count nr_running and co on all cpus
> Date: Wed, 30 Jul 2014 13:32:20 +0400
>
> Currently we count them only on cpus 0..nr_vcpus, which is obviously
> wrong, because those numbers are kept in a per-pcpu structure.
>
> https://jira.sw.ru/browse/PSBM-28277
>
> Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
>
> =============================================================================
>
> Related to https://jira.sw.ru/browse/PSBM-33642
>
> Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
>
> +++
> ve/cpustat: don't try to update vcpustats for root_task_group
>
> root_task_group doesn't have vcpu stats. Attempt to update them leads
> to NULL-ptr deref:
>
> BUG: unable to handle kernel NULL pointer dereference at (null)
> IP: [<ffffffff810b440c>] cpu_cgroup_update_vcpustat+0x13c/0x620
> ...
> Call Trace:
> [<ffffffff810bee3b>] cpu_cgroup_get_stat+0x7b/0x180
> [<ffffffff810f1ef7>] ve_get_cpu_stat+0x27/0x70
> [<ffffffffa01836a1>] fill_cpu_stat+0x91/0x1e0 [vzmon]
> [<ffffffffa0183c6b>] vzcalls_ioctl+0x2bb/0x430 [vzmon]
> [<ffffffffa018d0d5>] vzctl_ioctl+0x45/0x60 [vzdev]
> [<ffffffff8120cfb5>] do_vfs_ioctl+0x255/0x4f0
> [<ffffffff8120d2a4>] SyS_ioctl+0x54/0xa0
> [<ffffffff81642ac9>] system_call_fastpath+0x16/0x1b
>
> So, return -ENOENT if we asked for vcpu stats of root_task_group.
>
> https://jira.sw.ru/browse/PSBM-48721
>
> Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
> Reviewed-by: Vladimir Davydov <vdavydov at virtuozzo.com>
>
> +++
> ve/sched: take nr_cpus and cpu_rate from ve root task group
>
> Patchset description:
>
> ve: properly handle nr_cpus and cpu_rate for nested cgroups
>
> https://jira.sw.ru/browse/PSBM-69678
>
> Pavel Tikhomirov (3):
> cgroup: remove rcu_read_lock from cgroup_get_ve_root
> cgroup: make cgroup_get_ve_root visible in kernel/sched/core.c
> sched: take nr_cpus and cpu_rate from ve root task group
>
> =============================================================
> This patch description:
> Cpu view in container should depend only from root cpu cgroup
> nr_cpus/rate configuration. So replace tg->xxx references by
> tg_xxx(tg) helpers to get xxx from root ve cgroup. We still
> allow set/read rate and nr_cpus directly in nested cgroups,
> but they are just converted to corresponding cfs_period and
> cfs_quota setup, and does _not_ influence in container view
> of cpus and their stats.
>
> Also remove excessive rcu_read_lock/unlock as we have no rcu
> dereference in between, looks like some leftover for task_group()
> which differs in VZ6 and VZ7.
>
> https://jira.sw.ru/browse/PSBM-69678
>
> Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
> Reviewed-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
>
> +++
> ve/sched: remove no ve root cgroup warning
>
> We can get to ve_root_tg() from host's cgroup so it is expected
> to have no ve root cgroup for it. Call stack on task wakeup:
>
> wake_up_process -> try_to_wake_up -> select_task_rq_fair
> -> select_runnable_cpu -> check_cpulimit_spread -> tg_cpu_rate
> -> ve_root_tg
>
> Fixes: e661261 ("ve/sched: take nr_cpus and cpu_rate from ve root task group")
>
> Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
>
> +++
> ve/sched: take nr_cpus and cpu_rate from ve root task group
>
> Patchset description:
>
> ve: properly handle nr_cpus and cpu_rate for nested cgroups
>
> https://jira.sw.ru/browse/PSBM-69678
>
> Pavel Tikhomirov (3):
> cgroup: remove rcu_read_lock from cgroup_get_ve_root
> cgroup: make cgroup_get_ve_root visible in kernel/sched/core.c
> sched: take nr_cpus and cpu_rate from ve root task group
>
> =============================================================
> This patch description:
>
> Cpu view in container should depend only from root cpu cgroup
> nr_cpus/rate configuration. So replace tg->xxx references by
> tg_xxx(tg) helpers to get xxx from root ve cgroup. We still
> allow set/read rate and nr_cpus directly in nested cgroups,
> but they are just converted to corresponding cfs_period and
> cfs_quota setup, and does _not_ influence in container view
> of cpus and their stats.
>
> Also remove excessive rcu_read_lock/unlock as we have no rcu
> dereference in between, looks like some leftover for task_group()
> which differs in VZ6 and VZ7.
>
> https://jira.sw.ru/browse/PSBM-69678
>
> Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
> Reviewed-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
>
> +++
> ve/sched: remove no ve root cgroup warning
>
> We can get to ve_root_tg() from host's cgroup so it is expected
> to have no ve root cgroup for it. Call stack on task wakeup:
>
> wake_up_process -> try_to_wake_up -> select_task_rq_fair
> -> select_runnable_cpu -> check_cpulimit_spread -> tg_cpu_rate
> -> ve_root_tg
>
> Fixes: e661261 ("ve/sched: take nr_cpus and cpu_rate from ve root task group")
>
> Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
>
> +++
> sched: Fallback to 0-order allocations in sched_create_group()
>
> On large machines ->cpustat_last/->vcpustat arrays are large
> causing possibly failing high-order allocation:
>
> fio: page allocation failure: order:4, mode:0xc0d0
>
> Call Trace:
> dump_stack+0x19/0x1b
> warn_alloc_failed+0x110/0x180
> __alloc_pages_nodemask+0x7bf/0xc60
> alloc_pages_current+0x98/0x110
> kmalloc_order+0x18/0x40
> kmalloc_order_trace+0x26/0xa0
> __kmalloc+0x279/0x290
> sched_create_group+0xba/0x150
> sched_autogroup_create_attach+0x3f/0x1a0
> sys_setsid+0x73/0xc0
> system_call_fastpath+0x16/0x1b
>
> Use kvzalloc to fallback to vmalloc() and avoid failure if
> high order page is not available.
>
> https://jira.sw.ru/browse/PSBM-79891
> Fixes: 85fd6b2ff490 ("sched: Port cpustat related patches")
>
> Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
> Acked-by: Kirill Tkhai <ktkhai at virtuozzo.com>
>
> =============================================================
> proc/cpu/cgroup: make boottime in CT reveal the real start time
>
> When we call 'ps axfw -o pid,stime' in container to show the start time
> of processes of CT it looks like all of them are time travelers from the
> future, these happens in case the CT was recently migrated before.
>
> These happens because ps takes the start_time from /proc/<pid>/stat
> which should be relative to the boottime, so ps adds it to the boottime
> taken from /proc/stat and prints the results. But before these patch
> the boottime in /proc/stat in CT is just the time of cgroup creation
> (after migration), and the start_time in CT is relative to the real
> boottime of CT (before the migration).
>
> So make boottime in /proc/stat in CT be a real boottime of CT to fix
> these mess. Collateral damage is that we would always see host boottime
> in cpu.proc.stat files from host, but I don't think that cgroup creation
> times made any sense there anyway.
>
> https://jira.sw.ru/browse/PSBM-94263
> Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
>
> =============================================================
> sched: Account task_group::cpustat,taskstats,avenrun
>
> Extracted from "Initial patch".
>
> Signed-off-by: Kirill Tkhai <ktkhai at virtuozzo.com>
>
> =============================================================
> sched: Export per task_group statistics_work
>
> loadavg, iowait etc.
>
> Extracted from "Initial patch".
>
> Signed-off-by: Kirill Tkhai <ktkhai at virtuozzo.com>
>
> =============================================================
> ---
> kernel/sched/cpuacct.c | 159 +++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 159 insertions(+)
>
> diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
> index 9f0ec721aec7..b1460db447e3 100644
> --- a/kernel/sched/cpuacct.c
> +++ b/kernel/sched/cpuacct.c
> @@ -419,6 +419,51 @@ struct kernel_cpustat *cpuacct_cpustat(struct cgroup_subsys_state *css, int cpu)
> return per_cpu_ptr(css_ca(css)->cpustat, cpu);
> }
>
> +static void cpu_cgroup_update_stat(struct cgroup_subsys_state *cpu_css,
> + struct cgroup_subsys_state *cpuacct_css,
> + int i)
> +{
> +#if defined(CONFIG_SCHEDSTATS) && defined(CONFIG_FAIR_GROUP_SCHED)
> + struct task_group *tg = css_tg(cpu_css);
> + struct sched_entity *se = tg->se[i];
> + u64 *cpustat = cpuacct_cpustat(cpuacct_css, i)->cpustat;
> + u64 now = cpu_clock(i);
> + u64 delta, idle, iowait, steal, used;
> +
> + /* root_task_group has not sched entities */
> + if (tg == &root_task_group)
> + return;
> +
> + iowait = se->statistics.iowait_sum;
> + idle = se->statistics.sum_sleep_runtime;
> + steal = se->statistics.wait_sum;
> + used = se->sum_exec_runtime;
> +
> + if (idle > iowait)
> + idle -= iowait;
> + else
> + idle = 0;
> +
> + if (se->statistics.sleep_start) {
> + delta = now - se->statistics.sleep_start;
> + if ((s64)delta > 0)
> + idle += delta;
> + } else if (se->statistics.block_start) {
> + delta = now - se->statistics.block_start;
> + if ((s64)delta > 0)
> + iowait += delta;
> + } else if (se->statistics.wait_start) {
> + delta = now - se->statistics.wait_start;
> + if ((s64)delta > 0)
> + steal += delta;
> + }
> +
> + cpustat[CPUTIME_IDLE] = max(cpustat[CPUTIME_IDLE], idle);
> + cpustat[CPUTIME_IOWAIT] = max(cpustat[CPUTIME_IOWAIT], iowait);
> + cpustat[CPUTIME_STEAL] = steal;
> +#endif
> +}
> +
> static void fixup_vcpustat_delta_usage(struct kernel_cpustat *cur,
> struct kernel_cpustat *rem, int ind,
> u64 cur_usage, u64 target_usage,
> @@ -588,3 +633,117 @@ static void cpu_cgroup_update_vcpustat(struct cgroup_subsys_state *cpu_css,
> out_unlock:
> spin_unlock(&tg->vcpustat_lock);
> }
> +
> +int cpu_cgroup_proc_stat(struct cgroup_subsys_state *cpu_css,
> + struct cgroup_subsys_state *cpuacct_css,
> + struct seq_file *p)
> +{
> + int i;
> + s64 boot_sec;
> + u64 user, nice, system, idle, iowait, steal;
> + struct timespec64 boottime;
> + struct task_group *tg = css_tg(cpu_css);
> + bool virt = !ve_is_super(get_exec_env()) && tg != &root_task_group;
> + int nr_vcpus = tg_nr_cpus(tg);
> + struct kernel_cpustat *kcpustat;
> + unsigned long tg_nr_running = 0;
> + unsigned long tg_nr_iowait = 0;
> +
> + getboottime64(&boottime);
> +
> + /*
> + * In VE0 we always show host's boottime and in VEX we show real CT
> + * start time, even across CT migrations, as we rely on userspace to
> + * set real_start_timespec for us on resume.
> + */
> + boot_sec = boottime.tv_sec +
> + get_exec_env()->real_start_time / NSEC_PER_SEC;
> +
> + for_each_possible_cpu(i) {
> + cpu_cgroup_update_stat(cpu_css, cpuacct_css, i);
> +
> + /* root task group has autogrouping, so this doesn't hold */
> +#ifdef CONFIG_FAIR_GROUP_SCHED
> + tg_nr_running += tg->cfs_rq[i]->h_nr_running;
> + tg_nr_iowait += tg->cfs_rq[i]->nr_iowait;
> +#endif
> +#ifdef CONFIG_RT_GROUP_SCHED
> + tg_nr_running += tg->rt_rq[i]->rt_nr_running;
> +#endif
> + }
> +
> + if (virt)
> + cpu_cgroup_update_vcpustat(cpu_css, cpuacct_css);
> +
> + user = nice = system = idle = iowait = steal = 0;
> +
> + for (i = 0; i < (virt ? nr_vcpus : nr_cpu_ids); i++) {
> + if (!virt && !cpu_possible(i))
> + continue;
> +
> + kcpustat = virt ? tg->vcpustat + i :
> + cpuacct_cpustat(cpuacct_css, i);
> +
> + user += kcpustat->cpustat[CPUTIME_USER];
> + nice += kcpustat->cpustat[CPUTIME_NICE];
> + system += kcpustat->cpustat[CPUTIME_SYSTEM];
> + idle += kcpustat->cpustat[CPUTIME_IDLE];
> + iowait += kcpustat->cpustat[CPUTIME_IOWAIT];
> + steal += kcpustat->cpustat[CPUTIME_STEAL];
> + }
> + /* Don't scare CT users with high steal time */
> + if (!ve_is_super(get_exec_env()))
> + steal = 0;
> +
> + seq_printf(p, "cpu %llu %llu %llu %llu %llu 0 0 %llu\n",
> + (unsigned long long)nsec_to_clock_t(user),
> + (unsigned long long)nsec_to_clock_t(nice),
> + (unsigned long long)nsec_to_clock_t(system),
> + (unsigned long long)nsec_to_clock_t(idle),
> + (unsigned long long)nsec_to_clock_t(iowait),
> + virt ? 0ULL :
> + (unsigned long long)nsec_to_clock_t(steal));
> +
> + for (i = 0; i < (virt ? nr_vcpus : nr_cpu_ids); i++) {
> + if (!virt && !cpu_online(i))
> + continue;
> + kcpustat = virt ? tg->vcpustat + i :
> + cpuacct_cpustat(cpuacct_css, i);
> +
> + user = kcpustat->cpustat[CPUTIME_USER];
> + nice = kcpustat->cpustat[CPUTIME_NICE];
> + system = kcpustat->cpustat[CPUTIME_SYSTEM];
> + idle = kcpustat->cpustat[CPUTIME_IDLE];
> + iowait = kcpustat->cpustat[CPUTIME_IOWAIT];
> + steal = kcpustat->cpustat[CPUTIME_STEAL];
> + /* Don't scare CT users with high steal time */
> + if (!ve_is_super(get_exec_env()))
> + steal = 0;
> +
> + seq_printf(p,
> + "cpu%d %llu %llu %llu %llu %llu 0 0 %llu\n",
> + i,
> + (unsigned long long)nsec_to_clock_t(user),
> + (unsigned long long)nsec_to_clock_t(nice),
> + (unsigned long long)nsec_to_clock_t(system),
> + (unsigned long long)nsec_to_clock_t(idle),
> + (unsigned long long)nsec_to_clock_t(iowait),
> + virt ? 0ULL :
> + (unsigned long long)nsec_to_clock_t(steal));
> + }
> + seq_printf(p, "intr 0");
> +
> + seq_printf(p,
> + "\nctxt %llu\n"
> + "btime %llu\n"
> + "processes %lu\n"
> + "procs_running %lu\n"
> + "procs_blocked %lu\n",
> + nr_context_switches(),
> + (unsigned long long)boot_sec,
> + total_forks,
> + tg_nr_running,
> + tg_nr_iowait);
> +
> + return 0;
> +}
>
More information about the Devel
mailing list