core scheduling: add PR_SCHED_CORE_SHARE
Subsystems (11)
cover core scheduling: add PR_SCHED_CORE_SHARE
Hey everyone,
This adds the new PR_CORE_SCHED prctl() command PR_SCHED_CORE_SHARE to allow a third process to pull a core scheduling domain from one task and push it to another task.
The core scheduling uapi is exposed via the PR_SCHED_CORE option of the prctl() system call. Two commands can be used to alter the core scheduling domain of a task:
1. PR_SCHED_CORE_SHARE_TO
This command takes the cookie for the caller's core scheduling domain and applies it to a target task identified by passing a pid.
2. PR_SCHED_CORE_SHARE_FROM
This command takes the cookie for a task's core scheduling domain and applies it to the calling task.
While these options cover nearly all use-cases they are rather inconvient for some common use-cases. A vm/container manager often supervises a large number of vms/containers:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
Where none of the vms/container are its immediate children.
For container managers each container often has a separate supervising process and the workload is the parent of the container. In the example below the supervising process is "[lxc monitor]" and the workload is "/sbin/init" and all descendant processes:
├─[lxc monitor] /var/lib/lxd/containers imp1
│ └─systemd
│ ├─agetty -o -p -- \\u --noclear --keep-baud console 115200,38400,9600 linux
│ ├─cron -f -P
│ ├─dbus-daemon --system --address=systemd: --nofork --nopidfile --systemd-activation --syslog-only
│ ├─networkd-dispat /usr/bin/networkd-dispatcher --run-startup-triggers
│ ├─rsyslogd -n -iNONE
│ │ ├─{rsyslogd}
│ │ └─{rsyslogd}
│ ├─systemd-journal
│ ├─systemd-logind
│ ├─systemd-network
│ ├─systemd-resolve
│ └─systemd-udevdSimiliar in spirit but different in layout a vm often has a supervising process and multiple threads for each vcpu:
├─qemu-system-x86 -S -name f2-vm [...]
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ └─{qemu-system-x86}So ultimately an approximation of that layout would be:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
| | | |
vcpus workload vcpus workload
(/sbin/init) (/sbin/init)For containers a new core scheduling domain is allocated for the init process. Any descendant processes and threads init spawns will automatically inherit the correct core scheduling domain.
For vms a new core scheduling domain is allocated and each vcpu thread will be made to join the new core scheduling domain.
Whenever the tool or library that we use to run containers or vms exposes an option to automatically create a new core scheduling domain we will make use of it. However that is not always the case. In such cases the vm/container manager will need to allocate and set the core scheduling domain for the relevant processes or threads.
Neither the vm/container mananger nor the indivial vm/container supervisors are supposed to run in any or the same core scheduling domain as the respective vcpus/workloads.
So in order to create to create a new core scheduling domain we need to fork() off a new helper process which allocates a core scheduling domain and then pushes the cookie for the core scheduling domain to the relevant vcpus/workloads.
This works but things get rather tricky, especially for containers, when a new process is supposed to be spawned into a running container. An important step in creating a new process inside a running container involves:
- getting a handle on the container's init process (pid or nowadays often a pidfd)
- getting a handle on the container's namespaces (namespace file descriptors reachable via /proc/<init-pid>/ns/<ns-typ> or nowadays often a pidfd)
- calling setns() either on each namespace file descriptor individually or on the pidfd of the init process
An important sub-step here is to attach to the container's pid namespace via setns(). After attaching to the container's pid namespace any process created via a fork()-like system calls will be a full member of the container's pid namespace.
So attaching often involves two child processes. The first child simply attaches to the namespaces of the container including the container's pid namespace. The second child fork()s and ultimately exec()s thereby guaranteeing that the newly created process is a full member of the container's pid namespace:
first_child = fork();
if (first_child == 0) {
setns(CLONE_NEWPID);
second_child = fork();
if (second_child == 0) {
execlp();
}
}As part of this we also need to make sure that the second child - the one ultimately exec()ing the relevant programm in an already running container - joins the core scheduling domain of the container. When the container runs in a new pid namespace this can usually be done by calling:
first_child = fork();
if (first_child == 0) {
setns(CLONE_NEWPID);
second_child = fork();
if (second_child == 0) {
prctl(PR_SCHED_CORE, PR_SCHED_CORE_SHARE_FROM,
1, PR_SCHED_CORE_SCOPE_THREAD, 0);
execlp();
}
}from the second child since we know that pid 1 in a container running inside of a separate pid namespace is the correct process to get the core scheduling domain from.
However, this doesn't work when the container does not run in a separate pid namespace or when it shares the pid namespace with another container. In these scenarios we can't simply call PR_SCHED_CORE_SHARE_FROM from the second child since we don't know the correct pid number to call it on in the pid namespace.
(Note it is of course possible to learn the pid of the process in the relevant pid namespace but it is rather complex involving three separate processes and an AF_UNIX domain socket over which to send a message including struct ucred from which to learn the relevant pid. But that doesn't work in all cases since it requires privileges to translate arbitrary pids. In any case, this is not an option for performance reasons alone. However, I do also have a separate patchset in [1] allowing translation of pids between pid namespaces which will help with that in the future - something which I had discussed with Joel a while back but haven't pushed for yet since implementing it early 2020. Both patches are useful independent of one another.)
Additionally, we ideally always want to manage the core scheduling domain from the first child since the first child knows the pids for the relevant processes in its current pid namespace. The first child knows the pid of the init process in the current pid namespace from which to pull the core scheduling domain and it knows the pid of the second child it created to which to apply the core scheduling domain.
The core scheduling domain of the first child needs to be unaffected as it might run sensitive codepaths that should not be exposed in smt attacks.
The new PR_CORE_SCHED_SHARE command for the PR_SCHED_CORE prctl() option allows to support this and other use-cases by making it possible to pull the core scheduling domain from a task identified via its pid and push it to another task identified via its pid from a third managing task:
prctl(PR_SCHED_CORE, PR_SCHED_CORE_SHARE,
<pid-to-which-to-apply-coresched-domain>, PR_SCHED_CORE_SCOPE_{THREAD,THREAD_GROUP,PROCESS_GROUP}, <pid-from-which-to-take-coresched-domain>)
In order to use PR_SCHED_CORE_SHARE the caller must have ptrace_may_access() rights to both the task from which to take the core scheduling domain and to the task to which to apply the core scheduling domain. If the caller passes zero as the 5th argument then its own core scheduling domain is applied to the target making the option adhere to regular prctl() semantics.
Thanks!
Christian
Christian Brauner (3):
pid: introduce task_by_pid()
sched/prctl: add PR_SCHED_CORE_SHARE command
tests: add new PR_SCHED_CORE_SHARE test
arch/mips/kernel/mips-mt-fpaff.c | 14 +----- arch/x86/kernel/cpu/resctrl/rdtgroup.c | 19 +++----- block/ioprio.c | 10 +---- include/linux/sched.h | 9 +++- include/uapi/linux/prctl.h | 3 +- kernel/cgroup/cgroup.c | 12 ++--- kernel/events/core.c | 5 +-- kernel/futex/syscalls.c | 20 +++------ kernel/pid.c | 5 +++ kernel/sched/core.c | 27 ++++-------- kernel/sched/core_sched.c | 44 ++++++++++++++----- kernel/sys.c | 12 ++--- mm/mempolicy.c | 2 +- tools/testing/selftests/sched/cs_prctl_test.c | 23 ++++++++++ 14 files changed, 105 insertions(+), 100 deletions(-)
base-commit: e783362eb54cd99b2cac8b3a9aeac942e6f6ac07
Signature · 1 line
2.32.0
RFC 1/3 pid: introduce task_by_pid()
We have a lot of places that open code
if (who)
p = find_task_by_vpid(who);
else
p = current;Introduce a simpler helper which can be used instead.
--- arch/mips/kernel/mips-mt-fpaff.c | 14 ++----------- arch/x86/kernel/cpu/resctrl/rdtgroup.c | 19 +++++++----------- block/ioprio.c | 10 ++-------- include/linux/sched.h | 7 +++++++ kernel/cgroup/cgroup.c | 12 ++++-------- kernel/events/core.c | 5 +---- kernel/futex/syscalls.c | 20 ++++++------------- kernel/pid.c | 5 +++++ kernel/sched/core.c | 27 ++++++++------------------ kernel/sched/core_sched.c | 12 ++++-------- kernel/sys.c | 12 +++--------- mm/mempolicy.c | 2 +- 12 files changed, 50 insertions(+), 95 deletions(-)
arch/mips/kernel/mips-mt-fpaff.c +2 -12
@@ -33,16 +33,6 @@ unsigned long mt_fpemul_threshold;* updated when kernel/sched/core.c changes. */ -/* -* find_process_by_pid - find a process with a matching PID value. -* used in sys_sched_set/getaffinity() in kernel/sched/core.c, so -* cloned here. -*/ -static inline struct task_struct *find_process_by_pid (pid_t pid) -{ -return pid ?find_task_by_vpid (pid) : current; -} -/* * check the target process has a UID that matches the current process's */
@@ -79,7 +69,7 @@ asmlinkage long mipsmt_sys_sched_setaffinity(pid_t pid, unsigned int len,cpus_read_lock ();rcu_read_lock (); - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (!p) {rcu_read_unlock ();cpus_read_unlock ();
@@ -170,7 +160,7 @@ asmlinkage long mipsmt_sys_sched_getaffinity(pid_t pid, unsigned int len,rcu_read_lock (); retval = -ESRCH ; - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (!p) goto out_unlock; retval =security_task_getscheduler (p);
arch/x86/kernel/cpu/resctrl/rdtgroup.c +7 -12
@@ -660,19 +660,14 @@ static int rdtgroup_move_task(pid_t pid, struct rdtgroup *rdtgrp,int ret;rcu_read_lock (); -if (pid) { - tsk =find_task_by_vpid (pid); -if (!tsk) { -rcu_read_unlock (); -rdt_last_cmd_printf ("No task %d\n" , pid); -return -ESRCH ; - } - }else { - tsk = current; - } - -get_task_struct (tsk); + tsk =task_by_pid (pid); +if (tsk) +get_task_struct (tsk);rcu_read_unlock (); +if (!tsk) { +rdt_last_cmd_printf ("No task %d\n" , pid); +return -ESRCH ; + } ret =rdtgroup_task_write_permission (tsk, of);if (!ret)
block/ioprio.c +2 -8
@@ -81,10 +81,7 @@ SYSCALL_DEFINE3(ioprio_set, int, which, int, who, int, ioprio)rcu_read_lock ();switch (which) {case IOPRIO_WHO_PROCESS : -if (!who) - p = current; -else - p =find_task_by_vpid (who); + p =task_by_pid (who);if (p) ret =set_task_ioprio (p, ioprio);break ;
@@ -176,10 +173,7 @@ SYSCALL_DEFINE2(ioprio_get, int, which, int, who)rcu_read_lock ();switch (which) {case IOPRIO_WHO_PROCESS : -if (!who) - p = current; -else - p =find_task_by_vpid (who); + p =task_by_pid (who);if (p) ret =get_task_ioprio (p);break ;
include/linux/sched.h +7 -0
@@ -1912,6 +1912,13 @@ extern unsigned long init_stack[THREAD_SIZE / sizeof(unsigned long)];*/ extern struct task_struct *find_task_by_vpid (pid_t nr); +/** +* task_by_pid - find a process with a matching PID value. +* @pid: the pid in question. +* +* The task of @pid, if found. %NULL otherwise. +*/ +extern struct task_struct *task_by_pid (pid_t nr);extern struct task_struct *find_task_by_pid_ns (pid_t nr,struct pid_namespace *ns);/*
kernel/cgroup/cgroup.c +4 -8
@@ -2839,14 +2839,10 @@ struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup, }rcu_read_lock (); -if (pid) { - tsk =find_task_by_vpid (pid); -if (!tsk) { - tsk =ERR_PTR (-ESRCH ); - goto out_unlock_threadgroup; - } - }else { - tsk = current; + tsk =task_by_pid (pid); +if (!tsk) { + tsk =ERR_PTR (-ESRCH ); + goto out_unlock_threadgroup; }if (threadgroup)
kernel/events/core.c +1 -4
@@ -4608,10 +4608,7 @@ find_lively_task_by_vpid(pid_t vpid)struct task_struct *task;rcu_read_lock (); -if (!vpid) - task = current; -else - task =find_task_by_vpid (vpid); + task =task_by_pid (vpid);if (task)get_task_struct (task);rcu_read_unlock ();
kernel/futex/syscalls.c +6 -14
@@ -57,13 +57,9 @@ SYSCALL_DEFINE3(get_robust_list, int, pid,rcu_read_lock (); ret = -ESRCH ; -if (!pid) - p = current; -else { - p =find_task_by_vpid (pid); -if (!p) - goto err_unlock; - } + p =task_by_pid (pid); +if (!p) + goto err_unlock; ret = -EPERM ;if (!ptrace_may_access (p,PTRACE_MODE_READ_REALCREDS ))
@@ -326,13 +322,9 @@ COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,rcu_read_lock (); ret = -ESRCH ; -if (!pid) - p = current; -else { - p =find_task_by_vpid (pid); -if (!p) - goto err_unlock; - } + p =task_by_pid (pid); +if (!p) + goto err_unlock; ret = -EPERM ;if (!ptrace_may_access (p,PTRACE_MODE_READ_REALCREDS ))
kernel/pid.c +5 -0
@@ -422,6 +422,11 @@ struct task_struct *find_task_by_vpid(pid_t vnr)return find_task_by_pid_ns (vnr,task_active_pid_ns (current)); } +struct task_struct *task_by_pid (pid_t nr) +{ +return nr ?find_task_by_vpid (nr) : current; +} +struct task_struct *find_get_task_by_vpid (pid_t nr) {struct task_struct *task;
kernel/sched/core.c +8 -19
@@ -7174,17 +7174,6 @@ unsigned long sched_cpu_util(int cpu, unsigned long max) }#endif /* CONFIG_SMP */ -/** -* find_process_by_pid - find a process with a matching PID value. -* @pid: the pid in question. -* -* The task of @pid, if found. %NULL otherwise. -*/ -static struct task_struct *find_process_by_pid (pid_t pid) -{ -return pid ?find_task_by_vpid (pid) : current; -} -/* * sched_setparam() passes in -1 for its policy, to let the functions * it calls know not to change it.
@@ -7627,7 +7616,7 @@ do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)rcu_read_lock (); retval = -ESRCH ; - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (likely (p))get_task_struct (p);rcu_read_unlock ();
@@ -7750,7 +7739,7 @@ SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,rcu_read_lock (); retval = -ESRCH ; - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (likely (p))get_task_struct (p);rcu_read_unlock ();
@@ -7782,7 +7771,7 @@ SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) retval = -ESRCH ;rcu_read_lock (); - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (p) { retval =security_task_getscheduler (p);if (!retval)
@@ -7811,7 +7800,7 @@ SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)return -EINVAL ;rcu_read_lock (); - p =find_process_by_pid (pid); + p =task_by_pid (pid); retval = -ESRCH ;if (!p) goto out_unlock;
@@ -7894,7 +7883,7 @@ SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,return -EINVAL ;rcu_read_lock (); - p =find_process_by_pid (pid); + p =task_by_pid (pid); retval = -ESRCH ;if (!p) goto out_unlock;
@@ -8003,7 +7992,7 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)rcu_read_lock (); - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (!p) {rcu_read_unlock ();return -ESRCH ;
@@ -8082,7 +8071,7 @@ long sched_getaffinity(pid_t pid, struct cpumask *mask)rcu_read_lock (); retval = -ESRCH ; - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (!p) goto out_unlock;
@@ -8482,7 +8471,7 @@ static int sched_rr_get_interval(pid_t pid, struct timespec64 *t) retval = -ESRCH ;rcu_read_lock (); - p =find_process_by_pid (pid); + p =task_by_pid (pid);if (!p) goto out_unlock;
kernel/sched/core_sched.c +4 -8
@@ -148,14 +148,10 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,return -EINVAL ;rcu_read_lock (); -if (pid ==0 ) { - task = current; - }else { - task =find_task_by_vpid (pid); -if (!task) { -rcu_read_unlock (); -return -ESRCH ; - } + task =task_by_pid (pid); +if (!task) { +rcu_read_unlock (); +return -ESRCH ; }get_task_struct (task);rcu_read_unlock ();
kernel/sys.c +3 -9
@@ -222,10 +222,7 @@ SYSCALL_DEFINE3(setpriority, int, which, int, who, int, niceval)rcu_read_lock ();switch (which) {case PRIO_PROCESS : -if (who) - p =find_task_by_vpid (who); -else - p = current; + p =task_by_pid (who);if (p) error =set_one_prio (p, niceval, error);break ;
@@ -285,10 +282,7 @@ SYSCALL_DEFINE2(getpriority, int, which, int, who)rcu_read_lock ();switch (which) {case PRIO_PROCESS : -if (who) - p =find_task_by_vpid (who); -else - p = current; + p =task_by_pid (who);if (p) { niceval =nice_to_rlimit (task_nice (p));if (niceval > retval)
@@ -1659,7 +1653,7 @@ SYSCALL_DEFINE4(prlimit64, pid_t, pid, unsigned int, resource, }rcu_read_lock (); - tsk = pid ?find_task_by_vpid (pid) : current; + tsk =task_by_pid (pid);if (!tsk) {rcu_read_unlock ();return -ESRCH ;
mm/mempolicy.c +1 -1
@@ -1613,7 +1613,7 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,/* Find the mm_struct */ rcu_read_lock (); - task = pid ?find_task_by_vpid (pid) : current; + task =task_by_pid (pid);if (!task) {rcu_read_unlock (); err = -ESRCH ;
Signature · 1 line
2.32.0
Re: [resend RFC 1/3] pid: introduce task_by_pid()
Quoted 9 lines · Christian Brauner
We have a lot of places that open code
if (who)
p = find_task_by_vpid(who); else p = current;Introduce a simpler helper which can be used instead.
Quoted 12 lines
Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc Jens Axboe <[email protected]>Cc Peter Zijlstra <[email protected]>Cc Juri Lelli <[email protected]>Cc Vincent Guittot <[email protected]>Cc Steven Rostedt <[email protected]>Cc Daniel Bristot de Oliveira <[email protected]>Cc Tejun Heo <[email protected]>Cc Sebastian Andrzej Siewior <[email protected]>Signed-off-by Christian Brauner <[email protected]>
For cgroup part:
Thanks.
Signature · 1 line
tejun
RFC 2/3 sched/prctl: add PR_SCHED_CORE_SHARE command
In reply to Christian BraunerThis adds the new PR_CORE_SCHED prctl() command PR_SCHED_CORE_SHARE to allow a third process to pull a core scheduling domain from one task and push it to another task.
The core scheduling uapi is exposed via the PR_SCHED_CORE option of the prctl() system call. Two commands can be used to alter the core scheduling domain of a task:
1. PR_SCHED_CORE_SHARE_TO
This command takes the cookie for the caller's core scheduling domain and applies it to a target task identified by passing a pid.
2. PR_SCHED_CORE_SHARE_FROM
This command takes the cookie for a task's core scheduling domain and applies it to the calling task.
While these options cover nearly all use-cases they are rather inconvient for some common use-cases. A vm/container manager often supervises a large number of vms/containers:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
Where none of the vms/container are its immediate children.
For container managers each container often has a separate supervising process and the workload is the parent of the container. In the example below the supervising process is "[lxc monitor]" and the workload is "/sbin/init" and all descendant processes:
├─[lxc monitor] /var/lib/lxd/containers imp1
│ └─systemd
│ ├─agetty -o -p -- \\u --noclear --keep-baud console 115200,38400,9600 linux
│ ├─cron -f -P
│ ├─dbus-daemon --system --address=systemd: --nofork --nopidfile --systemd-activation --syslog-only
│ ├─networkd-dispat /usr/bin/networkd-dispatcher --run-startup-triggers
│ ├─rsyslogd -n -iNONE
│ │ ├─{rsyslogd}
│ │ └─{rsyslogd}
│ ├─systemd-journal
│ ├─systemd-logind
│ ├─systemd-network
│ ├─systemd-resolve
│ └─systemd-udevdSimiliar in spirit but different in layout a vm often has a supervising process and multiple threads for each vcpu:
├─qemu-system-x86 -S -name f2-vm [...]
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ ├─{qemu-system-x86}
│ └─{qemu-system-x86}So ultimately an approximation of that layout would be:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
| | | |
vcpus workload vcpus workload
(/sbin/init) (/sbin/init)For containers a new core scheduling domain is allocated for the init process. Any descendant processes and threads init spawns will automatically inherit the correct core scheduling domain.
For vms a new core scheduling domain is allocated and each vcpu thread will be made to join the new core scheduling domain.
Whenever the tool or library that we use to run containers or vms exposes an option to automatically create a new core scheduling domain we will make use of it. However that is not always the case. In such cases the vm/container manager will need to allocate and set the core scheduling domain for the relevant processes or threads.
Neither the vm/container mananger nor the indivial vm/container supervisors are supposed to run in any or the same core scheduling domain as the respective vcpus/workloads.
So in order to create to create a new core scheduling domain we need to fork() off a new helper process which allocates a core scheduling domain and then pushes the cookie for the core scheduling domain to the relevant vcpus/workloads.
This works but things get rather tricky, especially for containers, when a new process is supposed to be spawned into a running container. An important step in creating a new process inside a running container involves:
- getting a handle on the container's init process (pid or nowadays often a pidfd)
- getting a handle on the container's namespaces (namespace file descriptors reachable via /proc/<init-pid>/ns/<ns-typ> or nowadays often a pidfd)
- calling setns() either on each namespace file descriptor individually or on the pidfd of the init process
An important sub-step here is to attach to the container's pid namespace via setns(). After attaching to the container's pid namespace any process created via a fork()-like system calls will be a full member of the container's pid namespace.
So attaching often involves two child processes. The first child simply attaches to the namespaces of the container including the container's pid namespace. The second child fork()s and ultimately exec()s thereby guaranteeing that the newly created process is a full member of the container's pid namespace:
first_child = fork();
if (first_child == 0) {
setns(CLONE_NEWPID);
second_child = fork();
if (second_child == 0) {
execlp();
}
}As part of this we also need to make sure that the second child - the one ultimately exec()ing the relevant programm in an already running container - joins the core scheduling domain of the container. When the container runs in a new pid namespace this can usually be done by calling:
first_child = fork();
if (first_child == 0) {
setns(CLONE_NEWPID);
second_child = fork();
if (second_child == 0) {
prctl(PR_SCHED_CORE, PR_SCHED_CORE_SHARE_FROM,
1, PR_SCHED_CORE_SCOPE_THREAD, 0);
execlp();
}
}from the second child since we know that pid 1 in a container running inside of a separate pid namespace is the correct process to get the core scheduling domain from.
However, this doesn't work when the container does not run in a separate pid namespace or when it shares the pid namespace with another container. In these scenarios we can't simply call PR_SCHED_CORE_SHARE_FROM from the second child since we don't know the correct pid number to call it on in the pid namespace.
(Note it is of course possible to learn the pid of the process in the relevant pid namespace but it is rather complex involving three separate processes and an AF_UNIX domain socket over which to send a message including struct ucred from which to learn the relevant pid. But that doesn't work in all cases since it requires privileges to translate arbitrary pids. In any case, this is not an option for performance reasons alone. However, I do also have a separate patchset in [1] allowing translation of pids between pid namespaces which will help with that in the future - something which I had discussed with Joel a while back but haven't pushed for yet since implementing it early 2020. Both patches are useful independent of one another.)
Additionally, we ideally always want to manage the core scheduling domain from the first child since the first child knows the pids for the relevant processes in its current pid namespace. The first child knows the pid of the init process in the current pid namespace from which to pull the core scheduling domain and it knows the pid of the second child it created to which to apply the core scheduling domain.
The core scheduling domain of the first child needs to be unaffected as it might run sensitive codepaths that should not be exposed in smt attacks.
The new PR_CORE_SCHED_SHARE command for the PR_SCHED_CORE prctl() option allows to support this and other use-cases by making it possible to pull the core scheduling domain from a task identified via its pid and push it to another task identified via its pid from a third managing task:
prctl(PR_SCHED_CORE, PR_SCHED_CORE_SHARE,
<pid-to-which-to-apply-coresched-domain>, PR_SCHED_CORE_SCOPE_{THREAD,THREAD_GROUP,PROCESS_GROUP}, <pid-from-which-to-take-coresched-domain>)
In order to use PR_SCHED_CORE_SHARE the caller must have ptrace_may_access() rights to both the task from which to take the core scheduling domain and to the task to which to apply the core scheduling domain. If the caller passes zero as the 5th argument then its own core scheduling domain is applied to the target making the option adhere to regular prctl() semantics.
[1]: https://git.kernel.org/brauner/h/ioctl_ns_get_init_pid
https://git.kernel.org/brauner/c/1ad81fd698dd7e6511c3db422eba42dec3ce1b08
--- include/linux/sched.h | 2 +- include/uapi/linux/prctl.h | 3 ++- kernel/sched/core_sched.c | 32 +++++++++++++++++++++++++++++--- 3 files changed, 32 insertions(+), 5 deletions(-)
include/linux/sched.h +1 -1
@@ -2341,7 +2341,7 @@ const struct cpumask *sched_trace_rd_span(struct root_domain *rd);extern void sched_core_free (struct task_struct *tsk);extern void sched_core_fork (struct task_struct *p);extern int sched_core_share_pid (unsigned int cmd,pid_t pid,enum pid_type type, -unsigned long uaddr); +unsigned long arg);#else static inline void sched_core_free (struct task_struct *tsk) { }static inline void sched_core_fork (struct task_struct *p) { }
include/uapi/linux/prctl.h +2 -1
@@ -267,7 +267,8 @@ struct prctl_mm_map {# define PR_SCHED_CORE_CREATE 1/* create unique core_sched cookie */ # define PR_SCHED_CORE_SHARE_TO 2/* push core_sched cookie to pid */ # define PR_SCHED_CORE_SHARE_FROM 3/* pull core_sched cookie to pid */ -# define PR_SCHED_CORE_MAX 4 +# define PR_SCHED_CORE_SHARE 4 +# define PR_SCHED_CORE_MAX 5# define PR_SCHED_CORE_SCOPE_THREAD 0# define PR_SCHED_CORE_SCOPE_THREAD_GROUP 1# define PR_SCHED_CORE_SCOPE_PROCESS_GROUP 2
kernel/sched/core_sched.c +29 -3
@@ -129,9 +129,10 @@ static void __sched_core_set(struct task_struct *p, unsigned long cookie)/* Called from prctl interface: PR_SCHED_CORE */ int sched_core_share_pid (unsigned int cmd,pid_t pid,enum pid_type type, -unsigned long uaddr) +unsigned long arg) { -unsigned long cookie =0 , id =0 ; +unsigned long cookie =0 , id =0 , uaddr =0 ; +pid_t pid_share =-1 ;struct task_struct *task, *p;struct pid *grp;int err =0 ;
@@ -144,9 +145,20 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,BUILD_BUG_ON (PR_SCHED_CORE_SCOPE_PROCESS_GROUP !=PIDTYPE_PGID );if (type >PIDTYPE_PGID || cmd >=PR_SCHED_CORE_MAX || pid <0 || - (cmd !=PR_SCHED_CORE_GET && uaddr)) + (cmd !=PR_SCHED_CORE_GET && cmd !=PR_SCHED_CORE_SHARE && arg))return -EINVAL ; +switch (cmd) { +case PR_SCHED_CORE_GET : + uaddr = arg; +break ; +case PR_SCHED_CORE_SHARE : + pid_share = arg; +if (pid_share <0 ) +return -EINVAL ; +break ; + } +rcu_read_lock (); task =task_by_pid (pid);if (!task) {
@@ -200,6 +212,20 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,__sched_core_set (current, cookie); goto out; +case PR_SCHED_CORE_SHARE : +rcu_read_lock (); + p =task_by_pid (pid_share); +if (!p) + err = -ESRCH ; +else if (!ptrace_may_access (p,PTRACE_MODE_READ_REALCREDS )) + err = -EPERM ; +if (!err) + cookie =sched_core_clone_cookie (p); +rcu_read_unlock (); +if (err) + goto out; +break ; +default : err = -EINVAL ; goto out;
Signature · 1 line
2.32.0
Re: [resend RFC 2/3] sched/prctl: add PR_SCHED_CORE_SHARE command
+Josh Don Who was involved with upstream development of the interface.
Quoted 311 lines · Christian Brauner
This adds the new
PR_CORE_SCHEDprctl()commandPR_SCHED_CORE_SHAREto allow a third process to pull a core scheduling domain from one task and push it to another task.The core scheduling uapi is exposed via the
PR_SCHED_COREoption of theprctl()system call. Two commands can be used to alter the core scheduling domain of a task:1.
PR_SCHED_CORE_SHARE_TO
This command takes the cookie for the caller's core scheduling domain and applies it to a target task identified by passing a pid.2.
PR_SCHED_CORE_SHARE_FROM
This command takes the cookie for a task's core scheduling domain and applies it to the calling task.While these options cover nearly all use-cases they are rather inconvient for some common use-cases. A vm/container manager often supervises a large number of vms/containers:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
Where none of the vms/container are its immediate children.
For container managers each container often has a separate supervising process and the workload is the parent of the container. In the example below the supervising process is "[lxc monitor]" and the workload is "
/sbin/init" and all descendant processes:├─[lxc monitor] /var/lib/lxd/containers imp1 │ └─systemd │ ├─agetty -o -p -- \\u --noclear --keep-baud console 115200,38400,9600 linux │ ├─cron -f -P │ ├─dbus-daemon --system --address=systemd: --nofork --nopidfile --systemd-activation --syslog-only │ ├─networkd-dispat /usr/bin/networkd-dispatcher --run-startup-triggers │ ├─rsyslogd -n -iNONE │ │ ├─{rsyslogd} │ │ └─{rsyslogd} │ ├─systemd-journal │ ├─systemd-logind │ ├─systemd-network │ ├─systemd-resolve │ └─systemd-udevdSimiliar in spirit but different in layout a vm often has a supervising process and multiple threads for each vcpu:
├─qemu-system-x86 -S -name f2-vm [...] │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ └─{qemu-system-x86}So ultimately an approximation of that layout would be:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
| | | | vcpus workload vcpus workload (/sbin/init) (/sbin/init)For containers a new core scheduling domain is allocated for the init process. Any descendant processes and threads init spawns will automatically inherit the correct core scheduling domain.
For vms a new core scheduling domain is allocated and each vcpu thread will be made to join the new core scheduling domain.
Whenever the tool or library that we use to run containers or vms exposes an option to automatically create a new core scheduling domain we will make use of it. However that is not always the case. In such cases the vm/container manager will need to allocate and set the core scheduling domain for the relevant processes or threads.
Neither the vm/container mananger nor the indivial vm/container supervisors are supposed to run in any or the same core scheduling domain as the respective vcpus/workloads.
So in order to create to create a new core scheduling domain we need to
fork()off a new helper process which allocates a core scheduling domain and then pushes the cookie for the core scheduling domain to the relevant vcpus/workloads.This works but things get rather tricky, especially for containers, when a new process is supposed to be spawned into a running container. An important step in creating a new process inside a running container involves:
- getting a handle on the container's init process (pid or nowadays often a pidfd)
- getting a handle on the container's namespaces (namespace file descriptors reachable via /proc/<init-pid>/ns/<ns-typ> or nowadays often a pidfd)
- callingsetns()either on each namespace file descriptor individually or on the pidfd of the init processAn important sub-step here is to attach to the container's pid namespace via
setns(). After attaching to the container's pid namespace any process created via a fork()-like system calls will be a full member of the container's pid namespace.So attaching often involves two child processes. The first child simply attaches to the namespaces of the container including the container's pid namespace. The second child fork()s and ultimately exec()s thereby guaranteeing that the newly created process is a full member of the container's pid namespace:
first_child=fork();if (first_child == 0) { setns(CLONE_NEWPID); second_child = fork(); if (second_child == 0) { execlp(); } }As part of this we also need to make sure that the second child - the one ultimately exec()ing the relevant programm in an already running container - joins the core scheduling domain of the container. When the container runs in a new pid namespace this can usually be done by calling:
first_child=fork();if (first_child == 0) { setns(CLONE_NEWPID); second_child = fork(); if (second_child == 0) { prctl(PR_SCHED_CORE, PR_SCHED_CORE_SHARE_FROM, 1, PR_SCHED_CORE_SCOPE_THREAD, 0); execlp(); } }from the second child since we know that pid 1 in a container running inside of a separate pid namespace is the correct process to get the core scheduling domain from.
However, this doesn't work when the container does not run in a separate pid namespace or when it shares the pid namespace with another container. In these scenarios we can't simply call
PR_SCHED_CORE_SHARE_FROMfrom the second child since we don't know the correct pid number to call it on in the pid namespace.(Note it is of course possible to learn the pid of the process in the relevant pid namespace but it is rather complex involving three separate processes and an
AF_UNIXdomain socket over which to send a message including struct ucred from which to learn the relevant pid. But that doesn't work in all cases since it requires privileges to translate arbitrary pids. In any case, this is not an option for performance reasons alone. However, I do also have a separate patchset in [1] allowing translation of pids between pid namespaces which will help with that in the future - something which I had discussed with Joel a while back but haven't pushed for yet since implementing it early 2020. Both patches are useful independent of one another.)Additionally, we ideally always want to manage the core scheduling domain from the first child since the first child knows the pids for the relevant processes in its current pid namespace. The first child knows the pid of the init process in the current pid namespace from which to pull the core scheduling domain and it knows the pid of the second child it created to which to apply the core scheduling domain.
The core scheduling domain of the first child needs to be unaffected as it might run sensitive codepaths that should not be exposed in smt attacks.
The new
PR_CORE_SCHED_SHAREcommand for thePR_SCHED_COREprctl()option allows to support this and other use-cases by making it possible to pull the core scheduling domain from a task identified via its pid and push it to another task identified via its pid from a third managing task:prctl(PR_SCHED_CORE,
PR_SCHED_CORE_SHARE,
<pid-to-which-to-apply-coresched-domain>, PR_SCHED_CORE_SCOPE_{THREAD,THREAD_GROUP,PROCESS_GROUP}, <pid-from-which-to-take-coresched-domain>)In order to use
PR_SCHED_CORE_SHAREthe caller must haveptrace_may_access()rights to both the task from which to take the core scheduling domain and to the task to which to apply the core scheduling domain. If the caller passes zero as the 5th argument then its own core scheduling domain is applied to the target making the option adhere to regularprctl()semantics.[1]: https://git.kernel.org/brauner/h/ioctl_ns_get_init_pid
https://git.kernel.org/brauner/c/1ad81fd698dd7e6511c3db422eba42dec3ce1b08Cc Peter Collingbourne <[email protected]>Cc Dietmar Eggemann <[email protected]>Cc Joel Fernandes <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Mel Gorman <[email protected]>Cc Vincent Guittot <[email protected]>Cc Chris Hyser <[email protected]>Cc Juri Lelli <[email protected]>Cc Catalin Marinas <[email protected]>Cc Ingo Molnar <[email protected]>Cc Daniel Bristot de Oliveira <[email protected]>Cc Steven Rostedt <[email protected]>Cc Ben Segall <[email protected]>Cc Balbir Singh <[email protected]>Cc Peter Zijlstra <[email protected]>Signed-off-by Christian Brauner <[email protected]>--- include/linux/sched.h | 2 +- include/uapi/linux/prctl.h | 3 ++- kernel/sched/core_sched.c | 32 +++++++++++++++++++++++++++++--- 3 files changed, 32 insertions(+), 5 deletions(-)include/linux/sched.h +1
-1@@ -2341,7 +2341,7 @@ const struct cpumask *sched_trace_rd_span(struct root_domain *rd);extern void sched_core_free (struct task_struct *tsk);extern void sched_core_fork (struct task_struct *p);extern int sched_core_share_pid (unsigned int cmd,pid_t pid,enum pid_type type, -unsigned long uaddr); +unsigned long arg);#else static inline void sched_core_free (struct task_struct *tsk) { }static inline void sched_core_fork (struct task_struct *p) { }include/uapi/linux/prctl.h +2
-1@@ -267,7 +267,8 @@ struct prctl_mm_map {# define PR_SCHED_CORE_CREATE 1/* create unique core_sched cookie */ # define PR_SCHED_CORE_SHARE_TO 2/* push core_sched cookie to pid */ # define PR_SCHED_CORE_SHARE_FROM 3/* pull core_sched cookie to pid */ -# define PR_SCHED_CORE_MAX 4 +# define PR_SCHED_CORE_SHARE 4 +# define PR_SCHED_CORE_MAX 5# define PR_SCHED_CORE_SCOPE_THREAD 0# define PR_SCHED_CORE_SCOPE_THREAD_GROUP 1# define PR_SCHED_CORE_SCOPE_PROCESS_GROUP 2kernel/sched/core_sched.c +29
-4@@ -129,9 +129,10 @@ static void __sched_core_set(struct task_struct *p, unsigned long cookie)/* Called from prctl interface: PR_SCHED_CORE */ int sched_core_share_pid (unsigned int cmd,pid_t pid,enum pid_type type, -unsigned long uaddr) +unsigned long arg) { -unsigned long cookie =0 , id =0 ; +unsigned long cookie =0 , id =0 , uaddr =0 ; +pid_t pid_share =-1 ;struct task_struct *task, *p;struct pid *grp;int err =0 ;@@ -144,9 +145,20 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,BUILD_BUG_ON (PR_SCHED_CORE_SCOPE_PROCESS_GROUP !=PIDTYPE_PGID );if (type >PIDTYPE_PGID || cmd >=PR_SCHED_CORE_MAX || pid <0 || - (cmd !=PR_SCHED_CORE_GET && uaddr)) + (cmd !=PR_SCHED_CORE_GET && cmd !=PR_SCHED_CORE_SHARE && arg))return -EINVAL ; +switch (cmd) { +case PR_SCHED_CORE_GET : + uaddr = arg; +break ; +case PR_SCHED_CORE_SHARE : + pid_share = arg; +if (pid_share <0 ) +return -EINVAL ; +break ; + } +rcu_read_lock (); task =task_by_pid (pid);if (!task) {@@ -200,6 +212,20 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,__sched_core_set (current, cookie); goto out; +case PR_SCHED_CORE_SHARE : +rcu_read_lock (); + p =task_by_pid (pid_share); +if (!p) + err = -ESRCH ; +else if (!ptrace_may_access (p,PTRACE_MODE_READ_REALCREDS )) + err = -EPERM ; +if (!err) + cookie =sched_core_clone_cookie (p); +rcu_read_unlock (); +if (err) + goto out; +break ; +default : err = -EINVAL ; goto out; --2.32.0
Re: [resend RFC 2/3] sched/prctl: add PR_SCHED_CORE_SHARE command
In reply to Christian BraunerHey Christian,
This seems like a reasonable extension of the interface to me.
Quoted 8 lines
kernel/sched/core_sched.c +4
-0@@ -200,6 +212,20 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,__sched_core_set (current, cookie); goto out; +case PR_SCHED_CORE_SHARE : +rcu_read_lock (); + p =task_by_pid (pid_share); +if (!p)
kernel/sched/core_sched.c int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type, · line 228
Quoted 10 lines
kernel/sched/core_sched.c +10
-0+ err = -ESRCH ; + else if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS)) + err = -EPERM ; +if (!err) + cookie =sched_core_clone_cookie (p); +rcu_read_unlock (); +if (err) + goto out; +break ; +
Did you consider folding this into SCHED_CORE_SHARE_TO? SHARE_TO isn't using the last arg right now; it could use it as an override for the task we copy the cookie from instead of always choosing 'current'. Since the code currently rejects any SCHED_CORE prctl calls with a non-zero last arg for commands other than SCHED_CORE_GET, this would be a safe change for userspace.
Re: [resend RFC 2/3] sched/prctl: add PR_SCHED_CORE_SHARE command
Quoted 21 lines · Josh Don
Hey Christian,
This seems like a reasonable extension of the interface to me.
Quoted 17 lines
kernel/sched/core_sched.c +13
-0@@ -200,6 +212,20 @@ int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,__sched_core_set (current, cookie); goto out; +case PR_SCHED_CORE_SHARE : +rcu_read_lock (); + p =task_by_pid (pid_share); +if (!p) + err = -ESRCH ; +else if (!ptrace_may_access (p,PTRACE_MODE_READ_REALCREDS )) + err = -EPERM ; +if (!err) + cookie =sched_core_clone_cookie (p); +rcu_read_unlock (); +if (err) + goto out; +break ;
Quoted 8 lines
Review int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type, · line 228
Quoted 1 line
(quoted diff) +1
-0+Did you consider folding this into
SCHED_CORE_SHARE_TO?SHARE_TOisn't using the last arg right now; it could use it as an override for the task we copy the cookie from instead of always choosing 'current'. Since the code currently rejects anySCHED_COREprctl calls with a non-zero last arg for commands other thanSCHED_CORE_GET, this would be a safe change for userspace.
Yeah, that sounds good to me too. I can certainly rework the patch to do that!
RFC 3/3 tests: add new PR_SCHED_CORE_SHARE test
In reply to Christian BraunerAdd tests for the new PR_SCHED_CORE_SHARE command.
--- tools/testing/selftests/sched/cs_prctl_test.c | 23 +++++++++++++++++++ 1 file changed, 23 insertions(+)
tools/testing/selftests/sched/cs_prctl_test.c +23 -0
@@ -229,6 +229,7 @@ int main(int argc, char *argv[])int pidx;int pid;int opt; +int i;while ((opt =getopt (argc, argv,":hkT:P:d:" )) !=-1 ) {switch (opt) {
@@ -325,6 +326,28 @@ int main(int argc, char *argv[])validate (get_cs_cookie (pid) !=0 );validate (get_cs_cookie (pid) ==get_cs_cookie (procs[pidx].thr_tids[0 ])); +printf ("\n## Set a new cookie on a single thread/PR_SCHED_CORE_SCOPE_THREAD [%d]\n" , pid); +if (_prctl (PR_SCHED_CORE ,PR_SCHED_CORE_CREATE , pid,PR_SCHED_CORE_SCOPE_THREAD ,0 ) <0 ) +handle_error ("core_sched create failed -- PR_SCHED_CORE_SCOPE_THREAD" ); +disp_processes (num_processes, procs); + +validate (get_cs_cookie (pid) !=get_cs_cookie (procs[pidx].thr_tids[0 ])); + +printf ("\n## Copy cookie from a thread [%d] to [%d] as PR_SCHED_CORE_SCOPE_THREAD\n" , pid, procs[pidx].thr_tids[0 ]); +if (_prctl (PR_SCHED_CORE ,PR_SCHED_CORE_SHARE , procs[pidx].thr_tids[0 ],PR_SCHED_CORE_SCOPE_THREAD , pid) <0 ) +handle_error ("core_sched share cookie from and to thread failed -- PR_SCHED_CORE_SCOPE_THREAD" ); +disp_processes (num_processes, procs); + +validate (get_cs_cookie (pid) ==get_cs_cookie (procs[pidx].thr_tids[0 ])); + +printf ("\n## Copy cookie from a thread [%d] to [%d] as PR_SCHED_CORE_SCOPE_THREAD_GROUP\n" , pid, pid); +if (_prctl (PR_SCHED_CORE ,PR_SCHED_CORE_SHARE , pid,PR_SCHED_CORE_SCOPE_THREAD_GROUP , pid) <0 ) +handle_error ("core_sched share cookie from and to thread-group failed -- PR_SCHED_CORE_SCOPE_THREAD_GROUP" ); +disp_processes (num_processes, procs); + +for (i =0 ; i < procs[pidx].num_threads; ++i) +validate (get_cs_cookie (pid) ==get_cs_cookie (procs[pidx].thr_tids[i])); +if (errors) {printf ("TESTS FAILED. errors: %d\n" , errors); res =10 ;
Signature · 1 line
2.32.0
+Josh Don Who was involved with upstream development of the interface.
Quoted 227 lines · Christian Brauner
Hey everyone,
This adds the new
PR_CORE_SCHEDprctl()commandPR_SCHED_CORE_SHAREto allow a third process to pull a core scheduling domain from one task and push it to another task.The core scheduling uapi is exposed via the
PR_SCHED_COREoption of theprctl()system call. Two commands can be used to alter the core scheduling domain of a task:1.
PR_SCHED_CORE_SHARE_TO
This command takes the cookie for the caller's core scheduling domain and applies it to a target task identified by passing a pid.2.
PR_SCHED_CORE_SHARE_FROM
This command takes the cookie for a task's core scheduling domain and applies it to the calling task.While these options cover nearly all use-cases they are rather inconvient for some common use-cases. A vm/container manager often supervises a large number of vms/containers:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
Where none of the vms/container are its immediate children.
For container managers each container often has a separate supervising process and the workload is the parent of the container. In the example below the supervising process is "[lxc monitor]" and the workload is "
/sbin/init" and all descendant processes:├─[lxc monitor] /var/lib/lxd/containers imp1 │ └─systemd │ ├─agetty -o -p -- \\u --noclear --keep-baud console 115200,38400,9600 linux │ ├─cron -f -P │ ├─dbus-daemon --system --address=systemd: --nofork --nopidfile --systemd-activation --syslog-only │ ├─networkd-dispat /usr/bin/networkd-dispatcher --run-startup-triggers │ ├─rsyslogd -n -iNONE │ │ ├─{rsyslogd} │ │ └─{rsyslogd} │ ├─systemd-journal │ ├─systemd-logind │ ├─systemd-network │ ├─systemd-resolve │ └─systemd-udevdSimiliar in spirit but different in layout a vm often has a supervising process and multiple threads for each vcpu:
├─qemu-system-x86 -S -name f2-vm [...] │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ ├─{qemu-system-x86} │ └─{qemu-system-x86}So ultimately an approximation of that layout would be:
vm/container manager
vm-supervisor-1 container-supervisor-1 vm-supervisor-2 container-supervisor-2
| | | | vcpus workload vcpus workload (/sbin/init) (/sbin/init)For containers a new core scheduling domain is allocated for the init process. Any descendant processes and threads init spawns will automatically inherit the correct core scheduling domain.
For vms a new core scheduling domain is allocated and each vcpu thread will be made to join the new core scheduling domain.
Whenever the tool or library that we use to run containers or vms exposes an option to automatically create a new core scheduling domain we will make use of it. However that is not always the case. In such cases the vm/container manager will need to allocate and set the core scheduling domain for the relevant processes or threads.
Neither the vm/container mananger nor the indivial vm/container supervisors are supposed to run in any or the same core scheduling domain as the respective vcpus/workloads.
So in order to create to create a new core scheduling domain we need to
fork()off a new helper process which allocates a core scheduling domain and then pushes the cookie for the core scheduling domain to the relevant vcpus/workloads.This works but things get rather tricky, especially for containers, when a new process is supposed to be spawned into a running container. An important step in creating a new process inside a running container involves:
- getting a handle on the container's init process (pid or nowadays often a pidfd)
- getting a handle on the container's namespaces (namespace file descriptors reachable via /proc/<init-pid>/ns/<ns-typ> or nowadays often a pidfd)
- callingsetns()either on each namespace file descriptor individually or on the pidfd of the init processAn important sub-step here is to attach to the container's pid namespace via
setns(). After attaching to the container's pid namespace any process created via a fork()-like system calls will be a full member of the container's pid namespace.So attaching often involves two child processes. The first child simply attaches to the namespaces of the container including the container's pid namespace. The second child fork()s and ultimately exec()s thereby guaranteeing that the newly created process is a full member of the container's pid namespace:
first_child=fork();if (first_child == 0) { setns(CLONE_NEWPID); second_child = fork(); if (second_child == 0) { execlp(); } }As part of this we also need to make sure that the second child - the one ultimately exec()ing the relevant programm in an already running container - joins the core scheduling domain of the container. When the container runs in a new pid namespace this can usually be done by calling:
first_child=fork();if (first_child == 0) { setns(CLONE_NEWPID); second_child = fork(); if (second_child == 0) { prctl(PR_SCHED_CORE, PR_SCHED_CORE_SHARE_FROM, 1, PR_SCHED_CORE_SCOPE_THREAD, 0); execlp(); } }from the second child since we know that pid 1 in a container running inside of a separate pid namespace is the correct process to get the core scheduling domain from.
However, this doesn't work when the container does not run in a separate pid namespace or when it shares the pid namespace with another container. In these scenarios we can't simply call
PR_SCHED_CORE_SHARE_FROMfrom the second child since we don't know the correct pid number to call it on in the pid namespace.(Note it is of course possible to learn the pid of the process in the relevant pid namespace but it is rather complex involving three separate processes and an
AF_UNIXdomain socket over which to send a message including struct ucred from which to learn the relevant pid. But that doesn't work in all cases since it requires privileges to translate arbitrary pids. In any case, this is not an option for performance reasons alone. However, I do also have a separate patchset in [1] allowing translation of pids between pid namespaces which will help with that in the future - something which I had discussed with Joel a while back but haven't pushed for yet since implementing it early 2020. Both patches are useful independent of one another.)Additionally, we ideally always want to manage the core scheduling domain from the first child since the first child knows the pids for the relevant processes in its current pid namespace. The first child knows the pid of the init process in the current pid namespace from which to pull the core scheduling domain and it knows the pid of the second child it created to which to apply the core scheduling domain.
The core scheduling domain of the first child needs to be unaffected as it might run sensitive codepaths that should not be exposed in smt attacks.
The new
PR_CORE_SCHED_SHAREcommand for thePR_SCHED_COREprctl()option allows to support this and other use-cases by making it possible to pull the core scheduling domain from a task identified via its pid and push it to another task identified via its pid from a third managing task:prctl(PR_SCHED_CORE,
PR_SCHED_CORE_SHARE,
<pid-to-which-to-apply-coresched-domain>, PR_SCHED_CORE_SCOPE_{THREAD,THREAD_GROUP,PROCESS_GROUP}, <pid-from-which-to-take-coresched-domain>)In order to use
PR_SCHED_CORE_SHAREthe caller must haveptrace_may_access()rights to both the task from which to take the core scheduling domain and to the task to which to apply the core scheduling domain. If the caller passes zero as the 5th argument then its own core scheduling domain is applied to the target making the option adhere to regularprctl()semantics.Thanks!
ChristianChristian Brauner (3):
pid: introducetask_by_pid()
sched/prctl: addPR_SCHED_CORE_SHAREcommand
tests: add newPR_SCHED_CORE_SHAREtestarch/mips/kernel/mips-mt-fpaff.c | 14 +----- arch/x86/kernel/cpu/resctrl/rdtgroup.c | 19 +++----- block/ioprio.c | 10 +---- include/linux/sched.h | 9 +++- include/uapi/linux/prctl.h | 3 +- kernel/cgroup/cgroup.c | 12 ++--- kernel/events/core.c | 5 +-- kernel/futex/syscalls.c | 20 +++------ kernel/pid.c | 5 +++ kernel/sched/core.c | 27 ++++-------- kernel/sched/core_sched.c | 44 ++++++++++++++----- kernel/sys.c | 12 ++--- mm/mempolicy.c | 2 +- tools/testing/selftests/sched/cs_prctl_test.c | 23 ++++++++++ 14 files changed, 105 insertions(+), 100 deletions(-)base-commit:
e783362eb54cd99b2cac8b3a9aeac942e6f6ac07--
2.32.0