diff mbox series

kernel: Introduce a write lock/unlock wrapper for tasklist_lock

Message ID 20231225081932.17752-1-quic_aiquny@quicinc.com
State New
Headers show
Series kernel: Introduce a write lock/unlock wrapper for tasklist_lock | expand

Commit Message

Aiqun Yu (Maria) Dec. 25, 2023, 8:19 a.m. UTC
As a rwlock for tasklist_lock, there are multiple scenarios to acquire
read lock which write lock needed to be waiting for.
In freeze_process/thaw_processes it can take about 200+ms for holding read
lock of tasklist_lock by walking and freezing/thawing tasks in commercial
devices. And write_lock_irq will have preempt disabled and local irq
disabled to spin until the tasklist_lock can be acquired. This leading to
a bad responsive performance of current system.
Take an example:
1. cpu0 is holding read lock of tasklist_lock to thaw_processes.
2. cpu1 is waiting write lock of tasklist_lock to exec a new thread with
   preempt_disabled and local irq disabled.
3. cpu2 is waiting write lock of tasklist_lock to do_exit with
   preempt_disabled and local irq disabled.
4. cpu3 is waiting write lock of tasklist_lock to do_exit with
   preempt_disabled and local irq disabled.
So introduce a write lock/unlock wrapper for tasklist_lock specificly.
The current taskslist_lock writers all have write_lock_irq to hold
tasklist_lock, and write_unlock_irq to release tasklist_lock, that means
the writers are not suitable or workable to wait on tasklist_lock in irq
disabled scenarios. So the write lock/unlock wrapper here only follow the
current design of directly use local_irq_disable and local_irq_enable,
and not take already irq disabled writer callers into account.
Use write_trylock in the loop and enabled irq for cpu to repsond if lock
cannot be taken.

Signed-off-by: Maria Yu <quic_aiquny@quicinc.com>
---
 fs/exec.c                  | 10 +++++-----
 include/linux/sched/task.h | 29 +++++++++++++++++++++++++++++
 kernel/exit.c              | 16 ++++++++--------
 kernel/fork.c              |  6 +++---
 kernel/ptrace.c            | 12 ++++++------
 kernel/sys.c               |  8 ++++----
 security/keys/keyctl.c     |  4 ++--
 7 files changed, 57 insertions(+), 28 deletions(-)


base-commit: 88035e5694a86a7167d490bb95e9df97a9bb162b

Comments

Jarkko Sakkinen Jan. 3, 2024, 2:04 p.m. UTC | #1
On Mon Dec 25, 2023 at 10:19 AM EET, Maria Yu wrote:
> As a rwlock for tasklist_lock, there are multiple scenarios to acquire
> read lock which write lock needed to be waiting for.
> In freeze_process/thaw_processes it can take about 200+ms for holding read
> lock of tasklist_lock by walking and freezing/thawing tasks in commercial
> devices. And write_lock_irq will have preempt disabled and local irq
> disabled to spin until the tasklist_lock can be acquired. This leading to
> a bad responsive performance of current system.
> Take an example:
> 1. cpu0 is holding read lock of tasklist_lock to thaw_processes.
> 2. cpu1 is waiting write lock of tasklist_lock to exec a new thread with
>    preempt_disabled and local irq disabled.
> 3. cpu2 is waiting write lock of tasklist_lock to do_exit with
>    preempt_disabled and local irq disabled.
> 4. cpu3 is waiting write lock of tasklist_lock to do_exit with
>    preempt_disabled and local irq disabled.
> So introduce a write lock/unlock wrapper for tasklist_lock specificly.
> The current taskslist_lock writers all have write_lock_irq to hold
> tasklist_lock, and write_unlock_irq to release tasklist_lock, that means
> the writers are not suitable or workable to wait on tasklist_lock in irq
> disabled scenarios. So the write lock/unlock wrapper here only follow the
> current design of directly use local_irq_disable and local_irq_enable,
> and not take already irq disabled writer callers into account.
> Use write_trylock in the loop and enabled irq for cpu to repsond if lock
> cannot be taken.
>
> Signed-off-by: Maria Yu <quic_aiquny@quicinc.com>
> ---
>  fs/exec.c                  | 10 +++++-----
>  include/linux/sched/task.h | 29 +++++++++++++++++++++++++++++
>  kernel/exit.c              | 16 ++++++++--------
>  kernel/fork.c              |  6 +++---
>  kernel/ptrace.c            | 12 ++++++------
>  kernel/sys.c               |  8 ++++----
>  security/keys/keyctl.c     |  4 ++--
>  7 files changed, 57 insertions(+), 28 deletions(-)
>
> diff --git a/fs/exec.c b/fs/exec.c
> index 4aa19b24f281..030eef6852eb 100644
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -1086,7 +1086,7 @@ static int de_thread(struct task_struct *tsk)
>  
>  		for (;;) {
>  			cgroup_threadgroup_change_begin(tsk);
> -			write_lock_irq(&tasklist_lock);
> +			write_lock_tasklist_lock();
>  			/*
>  			 * Do this under tasklist_lock to ensure that
>  			 * exit_notify() can't miss ->group_exec_task
> @@ -1095,7 +1095,7 @@ static int de_thread(struct task_struct *tsk)
>  			if (likely(leader->exit_state))
>  				break;
>  			__set_current_state(TASK_KILLABLE);
> -			write_unlock_irq(&tasklist_lock);
> +			write_unlock_tasklist_lock();
>  			cgroup_threadgroup_change_end(tsk);
>  			schedule();
>  			if (__fatal_signal_pending(tsk))
> @@ -1150,7 +1150,7 @@ static int de_thread(struct task_struct *tsk)
>  		 */
>  		if (unlikely(leader->ptrace))
>  			__wake_up_parent(leader, leader->parent);
> -		write_unlock_irq(&tasklist_lock);
> +		write_unlock_tasklist_lock();
>  		cgroup_threadgroup_change_end(tsk);
>  
>  		release_task(leader);
> @@ -1198,13 +1198,13 @@ static int unshare_sighand(struct task_struct *me)
>  
>  		refcount_set(&newsighand->count, 1);
>  
> -		write_lock_irq(&tasklist_lock);
> +		write_lock_tasklist_lock();
>  		spin_lock(&oldsighand->siglock);
>  		memcpy(newsighand->action, oldsighand->action,
>  		       sizeof(newsighand->action));
>  		rcu_assign_pointer(me->sighand, newsighand);
>  		spin_unlock(&oldsighand->siglock);
> -		write_unlock_irq(&tasklist_lock);
> +		write_unlock_tasklist_lock();
>  
>  		__cleanup_sighand(oldsighand);
>  	}
> diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> index a23af225c898..6f69d9a3c868 100644
> --- a/include/linux/sched/task.h
> +++ b/include/linux/sched/task.h
> @@ -50,6 +50,35 @@ struct kernel_clone_args {
>   * a separate lock).
>   */
>  extern rwlock_t tasklist_lock;
> +
> +/*
> + * Tasklist_lock is a special lock, it takes a good amount of time of
> + * taskslist_lock readers to finish, and the pure write_irq_lock api
> + * will do local_irq_disable at the very first, and put the current cpu
> + * waiting for the lock while is non-responsive for interrupts.
> + *
> + * The current taskslist_lock writers all have write_lock_irq to hold
> + * tasklist_lock, and write_unlock_irq to release tasklist_lock, that
> + * means the writers are not suitable or workable to wait on
> + * tasklist_lock in irq disabled scenarios. So the write lock/unlock
> + * wrapper here only follow the current design of directly use
> + * local_irq_disable and local_irq_enable.
> + */
> +static inline void write_lock_tasklist_lock(void)
> +{
> +	while (1) {
> +		local_irq_disable();
> +		if (write_trylock(&tasklist_lock))
> +			break;
> +		local_irq_enable();
> +		cpu_relax();
> +	}

Maybe:

	local_irq_disable();
	while (!write_trylock(&tasklist_lock)) {
		local_irq_enable();
		cpu_relax();
		local_irq_disable();
	}

BR, Jarkko
diff mbox series

Patch

diff --git a/fs/exec.c b/fs/exec.c
index 4aa19b24f281..030eef6852eb 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1086,7 +1086,7 @@  static int de_thread(struct task_struct *tsk)
 
 		for (;;) {
 			cgroup_threadgroup_change_begin(tsk);
-			write_lock_irq(&tasklist_lock);
+			write_lock_tasklist_lock();
 			/*
 			 * Do this under tasklist_lock to ensure that
 			 * exit_notify() can't miss ->group_exec_task
@@ -1095,7 +1095,7 @@  static int de_thread(struct task_struct *tsk)
 			if (likely(leader->exit_state))
 				break;
 			__set_current_state(TASK_KILLABLE);
-			write_unlock_irq(&tasklist_lock);
+			write_unlock_tasklist_lock();
 			cgroup_threadgroup_change_end(tsk);
 			schedule();
 			if (__fatal_signal_pending(tsk))
@@ -1150,7 +1150,7 @@  static int de_thread(struct task_struct *tsk)
 		 */
 		if (unlikely(leader->ptrace))
 			__wake_up_parent(leader, leader->parent);
-		write_unlock_irq(&tasklist_lock);
+		write_unlock_tasklist_lock();
 		cgroup_threadgroup_change_end(tsk);
 
 		release_task(leader);
@@ -1198,13 +1198,13 @@  static int unshare_sighand(struct task_struct *me)
 
 		refcount_set(&newsighand->count, 1);
 
-		write_lock_irq(&tasklist_lock);
+		write_lock_tasklist_lock();
 		spin_lock(&oldsighand->siglock);
 		memcpy(newsighand->action, oldsighand->action,
 		       sizeof(newsighand->action));
 		rcu_assign_pointer(me->sighand, newsighand);
 		spin_unlock(&oldsighand->siglock);
-		write_unlock_irq(&tasklist_lock);
+		write_unlock_tasklist_lock();
 
 		__cleanup_sighand(oldsighand);
 	}
diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
index a23af225c898..6f69d9a3c868 100644
--- a/include/linux/sched/task.h
+++ b/include/linux/sched/task.h
@@ -50,6 +50,35 @@  struct kernel_clone_args {
  * a separate lock).
  */
 extern rwlock_t tasklist_lock;
+
+/*
+ * Tasklist_lock is a special lock, it takes a good amount of time of
+ * taskslist_lock readers to finish, and the pure write_irq_lock api
+ * will do local_irq_disable at the very first, and put the current cpu
+ * waiting for the lock while is non-responsive for interrupts.
+ *
+ * The current taskslist_lock writers all have write_lock_irq to hold
+ * tasklist_lock, and write_unlock_irq to release tasklist_lock, that
+ * means the writers are not suitable or workable to wait on
+ * tasklist_lock in irq disabled scenarios. So the write lock/unlock
+ * wrapper here only follow the current design of directly use
+ * local_irq_disable and local_irq_enable.
+ */
+static inline void write_lock_tasklist_lock(void)
+{
+	while (1) {
+		local_irq_disable();
+		if (write_trylock(&tasklist_lock))
+			break;
+		local_irq_enable();
+		cpu_relax();
+	}
+}
+static inline void write_unlock_tasklist_lock(void)
+{
+	write_unlock_irq(&tasklist_lock);
+}
+
 extern spinlock_t mmlist_lock;
 
 extern union thread_union init_thread_union;
diff --git a/kernel/exit.c b/kernel/exit.c
index ee9f43bed49a..18b00f477079 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -251,7 +251,7 @@  void release_task(struct task_struct *p)
 
 	cgroup_release(p);
 
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 	ptrace_release_task(p);
 	thread_pid = get_pid(p->thread_pid);
 	__exit_signal(p);
@@ -275,7 +275,7 @@  void release_task(struct task_struct *p)
 			leader->exit_state = EXIT_DEAD;
 	}
 
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 	seccomp_filter_release(p);
 	proc_flush_pid(thread_pid);
 	put_pid(thread_pid);
@@ -598,7 +598,7 @@  static struct task_struct *find_child_reaper(struct task_struct *father,
 		return reaper;
 	}
 
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 
 	list_for_each_entry_safe(p, n, dead, ptrace_entry) {
 		list_del_init(&p->ptrace_entry);
@@ -606,7 +606,7 @@  static struct task_struct *find_child_reaper(struct task_struct *father,
 	}
 
 	zap_pid_ns_processes(pid_ns);
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 
 	return father;
 }
@@ -730,7 +730,7 @@  static void exit_notify(struct task_struct *tsk, int group_dead)
 	struct task_struct *p, *n;
 	LIST_HEAD(dead);
 
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 	forget_original_parent(tsk, &dead);
 
 	if (group_dead)
@@ -758,7 +758,7 @@  static void exit_notify(struct task_struct *tsk, int group_dead)
 	/* mt-exec, de_thread() is waiting for group leader */
 	if (unlikely(tsk->signal->notify_count < 0))
 		wake_up_process(tsk->signal->group_exec_task);
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 
 	list_for_each_entry_safe(p, n, &dead, ptrace_entry) {
 		list_del_init(&p->ptrace_entry);
@@ -1172,7 +1172,7 @@  static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
 	wo->wo_stat = status;
 
 	if (state == EXIT_TRACE) {
-		write_lock_irq(&tasklist_lock);
+		write_lock_tasklist_lock();
 		/* We dropped tasklist, ptracer could die and untrace */
 		ptrace_unlink(p);
 
@@ -1181,7 +1181,7 @@  static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
 		if (do_notify_parent(p, p->exit_signal))
 			state = EXIT_DEAD;
 		p->exit_state = state;
-		write_unlock_irq(&tasklist_lock);
+		write_unlock_tasklist_lock();
 	}
 	if (state == EXIT_DEAD)
 		release_task(p);
diff --git a/kernel/fork.c b/kernel/fork.c
index 10917c3e1f03..06c4b4ab9102 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -2623,7 +2623,7 @@  __latent_entropy struct task_struct *copy_process(
 	 * Make it visible to the rest of the system, but dont wake it up yet.
 	 * Need tasklist lock for parent etc handling!
 	 */
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 
 	/* CLONE_PARENT re-uses the old parent */
 	if (clone_flags & (CLONE_PARENT|CLONE_THREAD)) {
@@ -2714,7 +2714,7 @@  __latent_entropy struct task_struct *copy_process(
 	hlist_del_init(&delayed.node);
 	spin_unlock(&current->sighand->siglock);
 	syscall_tracepoint_update(p);
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 
 	if (pidfile)
 		fd_install(pidfd, pidfile);
@@ -2735,7 +2735,7 @@  __latent_entropy struct task_struct *copy_process(
 bad_fork_cancel_cgroup:
 	sched_core_free(p);
 	spin_unlock(&current->sighand->siglock);
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 	cgroup_cancel_fork(p, args);
 bad_fork_put_pidfd:
 	if (clone_flags & CLONE_PIDFD) {
diff --git a/kernel/ptrace.c b/kernel/ptrace.c
index d8b5e13a2229..a8d7e2d06f3e 100644
--- a/kernel/ptrace.c
+++ b/kernel/ptrace.c
@@ -435,7 +435,7 @@  static int ptrace_attach(struct task_struct *task, long request,
 	if (retval)
 		goto unlock_creds;
 
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 	retval = -EPERM;
 	if (unlikely(task->exit_state))
 		goto unlock_tasklist;
@@ -479,7 +479,7 @@  static int ptrace_attach(struct task_struct *task, long request,
 
 	retval = 0;
 unlock_tasklist:
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 unlock_creds:
 	mutex_unlock(&task->signal->cred_guard_mutex);
 out:
@@ -508,7 +508,7 @@  static int ptrace_traceme(void)
 {
 	int ret = -EPERM;
 
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 	/* Are we already being traced? */
 	if (!current->ptrace) {
 		ret = security_ptrace_traceme(current->parent);
@@ -522,7 +522,7 @@  static int ptrace_traceme(void)
 			ptrace_link(current, current->real_parent);
 		}
 	}
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 
 	return ret;
 }
@@ -588,7 +588,7 @@  static int ptrace_detach(struct task_struct *child, unsigned int data)
 	/* Architecture-specific hardware disable .. */
 	ptrace_disable(child);
 
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 	/*
 	 * We rely on ptrace_freeze_traced(). It can't be killed and
 	 * untraced by another thread, it can't be a zombie.
@@ -600,7 +600,7 @@  static int ptrace_detach(struct task_struct *child, unsigned int data)
 	 */
 	child->exit_code = data;
 	__ptrace_detach(current, child);
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 
 	proc_ptrace_connector(child, PTRACE_DETACH);
 
diff --git a/kernel/sys.c b/kernel/sys.c
index e219fcfa112d..0b1647d3ed32 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -1088,7 +1088,7 @@  SYSCALL_DEFINE2(setpgid, pid_t, pid, pid_t, pgid)
 	/* From this point forward we keep holding onto the tasklist lock
 	 * so that our parent does not change from under us. -DaveM
 	 */
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 
 	err = -ESRCH;
 	p = find_task_by_vpid(pid);
@@ -1136,7 +1136,7 @@  SYSCALL_DEFINE2(setpgid, pid_t, pid, pid_t, pgid)
 	err = 0;
 out:
 	/* All paths lead to here, thus we are safe. -DaveM */
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 	rcu_read_unlock();
 	return err;
 }
@@ -1229,7 +1229,7 @@  int ksys_setsid(void)
 	pid_t session = pid_vnr(sid);
 	int err = -EPERM;
 
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 	/* Fail if I am already a session leader */
 	if (group_leader->signal->leader)
 		goto out;
@@ -1247,7 +1247,7 @@  int ksys_setsid(void)
 
 	err = session;
 out:
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 	if (err > 0) {
 		proc_sid_connector(group_leader);
 		sched_autogroup_create_attach(group_leader);
diff --git a/security/keys/keyctl.c b/security/keys/keyctl.c
index 19be69fa4d05..dd8aed20486a 100644
--- a/security/keys/keyctl.c
+++ b/security/keys/keyctl.c
@@ -1652,7 +1652,7 @@  long keyctl_session_to_parent(void)
 
 	me = current;
 	rcu_read_lock();
-	write_lock_irq(&tasklist_lock);
+	write_lock_tasklist_lock();
 
 	ret = -EPERM;
 	oldwork = NULL;
@@ -1702,7 +1702,7 @@  long keyctl_session_to_parent(void)
 	if (!ret)
 		newwork = NULL;
 unlock:
-	write_unlock_irq(&tasklist_lock);
+	write_unlock_tasklist_lock();
 	rcu_read_unlock();
 	if (oldwork)
 		put_cred(container_of(oldwork, struct cred, rcu));