reduce tasklist_lock hold time on exit and do some pid cleanup
v6 cover reduce tasklist_lock hold time on exit and do some pid cleanup
I fixed fat-fingering and touchedup some commit messages. git diff between the old and new branch does not show any code changes.
hopefully we will be done here after this iteration 8->
old cover letter:
The clone side contends against exit side in a way which avoidably exacerbates the problem by the latter waiting on locks held by the former while holding the tasklist_lock.
Whacking this for both add_device_randomness and pids allocation gives me a 15% speed up for thread creation/destruction in a 24-core vm.
The random patch is worth about 4%.
The new bottleneck is pidmap_lock itself, with the biggest problem being the allocation itself taking the lock *twice*.
Bench (plop into will-it-scale):
$ cat tests/threadspawn1.c
char *testcase_description = "Thread creation and teardown";
static void *worker(void *arg)
{
return (NULL);
}void testcase(unsigned long long *iterations, unsigned long nr)
{
pthread_t thread;
int error;
while (1) {
error = pthread_create(&thread, NULL, worker, NULL);
assert(error == 0);
error = pthread_join(thread, NULL);
assert(error == 0);
(*iterations)++;
}
}v6:
- expand on the commit message in 3/5 pid: sprinkle tasklist_lock asserts
- move fat-fingered free_pids call to the right patch
v5:
- whack scripts/selinux/genheaders/genheaders which accidentally got in
- rebased on next-20250205
v4:
- justify moving get_pid in the commit message with a one-liner
- drop the tty unref patch -- it is completely optional and Oleg has his own variant
- add the ACK by Oleg
v3:
- keep procfs flush where it was, instead hoist get_pid outside of the lock
- make detach_pid et al accept an array argument of pids to populate
- sprinkle asserts
- drop irq trips around pidmap_lock
- move tty unref outside of tasklist_lock
Mateusz Guzik (5):
exit: perform add_device_randomness() without tasklist_lock
exit: hoist get_pid() in release_task() outside of tasklist_lock
pid: sprinkle tasklist_lock asserts
pid: perform free_pid() calls outside of tasklist_lock
pid: drop irq disablement around pidmap_lock
include/linux/pid.h | 7 ++-- kernel/exit.c | 36 +++++++++++++------- kernel/pid.c | 82 +++++++++++++++++++++++++-------------------- kernel/sys.c | 14 +++++--- 4 files changed, 82 insertions(+), 57 deletions(-)
Signature · 1 line
2.43.0
PATCH v6 1/5 exit: perform add_device_randomness() without tasklist_lock
Parallel calls to add_device_randomness() contend on their own.
The clone side aleady runs outside of tasklist_lock, which in turn means any caller on the exit side extends the tasklist_lock hold time while contending on the random-private lock.
--- kernel/exit.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
kernel/exit.c +2 -3
@@ -174,9 +174,6 @@ static void __exit_signal(struct task_struct *tsk) sig->curr_target =next_thread (tsk); } -add_device_randomness ((const void *) &tsk->se.sum_exec_runtime, -sizeof (unsigned long long )); -/* * Accumulate here the counters for all threads as they die. We could * skip the group leader because it is the last user of signal_struct,
@@ -278,6 +275,8 @@ void release_task(struct task_struct *p)write_unlock_irq (&tasklist_lock);proc_flush_pid (thread_pid);put_pid (thread_pid); +add_device_randomness (&p->se.sum_exec_runtime, +sizeof (p->se.sum_exec_runtime));release_thread (p);put_task_struct_rcu_user (p);
Signature · 1 line
2.43.0
PATCH v6 2/5 exit: hoist get_pid() in release_task() outside of tasklist_lock
In reply to Mateusz GuzikReduces hold time as get_pid() contains an atomic.
--- kernel/exit.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
kernel/exit.c +2 -1
@@ -248,9 +248,10 @@ void release_task(struct task_struct *p)cgroup_release (p); + thread_pid =get_pid (p->thread_pid); +write_lock_irq (&tasklist_lock);ptrace_release_task (p); - thread_pid =get_pid (p->thread_pid);__exit_signal (p);/*
Signature · 1 line
2.43.0
PATCH v6 3/5 pid: sprinkle tasklist_lock asserts
In reply to Mateusz GuzikThey cost nothing on production kernels and document the requirement of holding the tasklist_lock lock in respective routines.
--- kernel/pid.c | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-)
kernel/pid.c +12 -3
@@ -339,17 +339,23 @@ static struct pid **task_pid_ptr(struct task_struct *task, enum pid_type type)*/ void attach_pid (struct task_struct *task,enum pid_type type) { -struct pid *pid = *task_pid_ptr (task, type); +struct pid *pid; + +lockdep_assert_held_write (&tasklist_lock); + + pid = *task_pid_ptr (task, type);hlist_add_head_rcu (&task->pid_links[type], &pid->tasks[type]); }static void __change_pid (struct task_struct *task,enum pid_type type,struct pid *new) { -struct pid **pid_ptr =task_pid_ptr (task, type); -struct pid *pid; +struct pid **pid_ptr, *pid;int tmp; +lockdep_assert_held_write (&tasklist_lock); + + pid_ptr =task_pid_ptr (task, type); pid = *pid_ptr;hlist_del_rcu (&task->pid_links[type]);
@@ -386,6 +392,8 @@ void exchange_tids(struct task_struct *left, struct task_struct *right)struct hlist_head *head1 = &pid1->tasks[PIDTYPE_PID ];struct hlist_head *head2 = &pid2->tasks[PIDTYPE_PID ]; +lockdep_assert_held_write (&tasklist_lock); +/* Swap the single entry tid lists */ hlists_swap_heads_rcu (head1, head2);
@@ -403,6 +411,7 @@ void transfer_pid(struct task_struct *old, struct task_struct *new, enum pid_type type) {WARN_ON_ONCE (type ==PIDTYPE_PID ); +lockdep_assert_held_write (&tasklist_lock);hlist_replace_rcu (&old->pid_links[type], &new->pid_links[type]); }
Signature · 1 line
2.43.0
PATCH v6 4/5 pid: perform free_pid() calls outside of tasklist_lock
In reply to Mateusz GuzikAs the clone side already executes pid allocation with only pidmap_lock held, issuing free_pid() while still holding tasklist_lock exacerbates total hold time of the latter.
More things may show up later which require initial clean up with the lock held and allow finishing without it. For that reason a struct to collect such work is added instead of merely passing the pid array.
--- include/linux/pid.h | 7 ++++--- kernel/exit.c | 28 ++++++++++++++++++++-------- kernel/pid.c | 44 ++++++++++++++++++++++---------------------- kernel/sys.c | 14 +++++++++----- 4 files changed, 55 insertions(+), 38 deletions(-)
include/linux/pid.h +4 -3
@@ -101,9 +101,9 @@ extern struct pid *get_task_pid(struct task_struct *task, enum pid_type type);* these helpers must be called with the tasklist_lock write-held. */ extern void attach_pid (struct task_struct *task,enum pid_type ); -extern void detach_pid (struct task_struct *task,enum pid_type ); -extern void change_pid (struct task_struct *task,enum pid_type , -struct pid *pid); +void detach_pid (struct pid **pids,struct task_struct *task,enum pid_type ); +void change_pid (struct pid **pids,struct task_struct *task,enum pid_type , +struct pid *pid);extern void exchange_tids (struct task_struct *task,struct task_struct *old);extern void transfer_pid (struct task_struct *old,struct task_struct *new,enum pid_type );
@@ -129,6 +129,7 @@ extern struct pid *find_ge_pid(int nr, struct pid_namespace *);extern struct pid *alloc_pid (struct pid_namespace *ns,pid_t *set_tid,size_t set_tid_size);extern void free_pid (struct pid *pid); +void free_pids (struct pid **pids);extern void disable_pid_allocation (struct pid_namespace *ns);/*
kernel/exit.c +20 -8
@@ -122,14 +122,22 @@ static __init int kernel_exit_sysfs_init(void)late_initcall (kernel_exit_sysfs_init);#endif -static void __unhash_process (struct task_struct *p,bool group_dead) +/* +* For things release_task() would like to do *after* tasklist_lock is released. +*/ +struct release_task_post { +struct pid *pids[PIDTYPE_MAX ]; +}; + +static void __unhash_process (struct release_task_post *post,struct task_struct *p, +bool group_dead) { nr_threads--; -detach_pid (p,PIDTYPE_PID ); +detach_pid (post->pids, p,PIDTYPE_PID );if (group_dead) { -detach_pid (p,PIDTYPE_TGID ); -detach_pid (p,PIDTYPE_PGID ); -detach_pid (p,PIDTYPE_SID ); +detach_pid (post->pids, p,PIDTYPE_TGID ); +detach_pid (post->pids, p,PIDTYPE_PGID ); +detach_pid (post->pids, p,PIDTYPE_SID );list_del_rcu (&p->tasks);list_del_init (&p->sibling);
@@ -141,7 +149,7 @@ static void __unhash_process(struct task_struct *p, bool group_dead)/* * This function expects the tasklist_lock write-locked. */ -static void __exit_signal (struct task_struct *tsk) +static void __exit_signal (struct release_task_post *post,struct task_struct *tsk) {struct signal_struct *sig = tsk->signal;bool group_dead =thread_group_leader (tsk);
@@ -194,7 +202,7 @@ static void __exit_signal(struct task_struct *tsk)task_io_accounting_add (&sig->ioac, &tsk->ioac); sig->sum_sched_runtime += tsk->se.sum_exec_runtime; sig->nr_threads--; -__unhash_process (tsk, group_dead); +__unhash_process (post, tsk, group_dead);write_sequnlock (&sig->stats_lock);/*
@@ -236,10 +244,13 @@ void __weak release_thread(struct task_struct *dead_task)void release_task (struct task_struct *p) { +struct release_task_post post;struct task_struct *leader;struct pid *thread_pid;int zap_leader; repeat: +memset (&post,0 ,sizeof (post)); +/* don't need to get the RCU readlock here - the process is dead and * can't be modifying its own credentials. But shut RCU-lockdep up */ rcu_read_lock ();
@@ -252,7 +263,7 @@ void release_task(struct task_struct *p)write_lock_irq (&tasklist_lock);ptrace_release_task (p); -__exit_signal (p); +__exit_signal (&post, p);/* * If we are the last non-leader member of the thread
@@ -278,6 +289,7 @@ void release_task(struct task_struct *p)put_pid (thread_pid);add_device_randomness (&p->se.sum_exec_runtime,sizeof (p->se.sum_exec_runtime)); +free_pids (post.pids);release_thread (p);put_task_struct_rcu_user (p);
kernel/pid.c +22 -22
84 lines not on this page · raw message
kernel/sys.c +9 -5
@@ -1085,6 +1085,7 @@ SYSCALL_DEFINE2(setpgid, pid_t, pid, pid_t, pgid) {struct task_struct *p;struct task_struct *group_leader = current->group_leader; +struct pid *pids[PIDTYPE_MAX ] = {0 };struct pid *pgrp;int err;
@@ -1142,13 +1143,14 @@ SYSCALL_DEFINE2(setpgid, pid_t, pid, pid_t, pgid) goto out;if (task_pgrp (p) != pgrp) -change_pid (p,PIDTYPE_PGID , pgrp); +change_pid (pids, p,PIDTYPE_PGID , pgrp); err =0 ; out:/* All paths lead to here, thus we are safe. -DaveM */ write_unlock_irq (&tasklist_lock);rcu_read_unlock (); +free_pids (pids);return err; }
@@ -1222,21 +1224,22 @@ SYSCALL_DEFINE1(getsid, pid_t, pid)return retval; } -static void set_special_pids (struct pid *pid) +static void set_special_pids (struct pid **pids,struct pid *pid) {struct task_struct *curr = current->group_leader;if (task_session (curr) != pid) -change_pid (curr,PIDTYPE_SID , pid); +change_pid (pids, curr,PIDTYPE_SID , pid);if (task_pgrp (curr) != pid) -change_pid (curr,PIDTYPE_PGID , pid); +change_pid (pids, curr,PIDTYPE_PGID , pid); }int ksys_setsid (void ) {struct task_struct *group_leader = current->group_leader;struct pid *sid =task_pid (group_leader); +struct pid *pids[PIDTYPE_MAX ] = {0 };pid_t session =pid_vnr (sid);int err = -EPERM ;
@@ -1252,13 +1255,14 @@ int ksys_setsid(void) goto out; group_leader->signal->leader =1 ; -set_special_pids (sid); +set_special_pids (pids, sid);proc_clear_tty (group_leader); err = session; out:write_unlock_irq (&tasklist_lock); +free_pids (pids);if (err >0 ) {proc_sid_connector (group_leader);sched_autogroup_create_attach (group_leader);
Signature · 1 line
2.43.0
PATCH v6 5/5 pid: drop irq disablement around pidmap_lock
In reply to Mateusz GuzikIt no longer serves any purpose now that the tasklist_lock -> pidmap_lock ordering got eliminated.
--- kernel/pid.c | 23 +++++++++++------------ 1 file changed, 11 insertions(+), 12 deletions(-)
kernel/pid.c +11 -12
@@ -115,11 +115,10 @@ static void delayed_put_pid(struct rcu_head *rhp)void free_pid (struct pid *pid) {int i; -unsigned long flags;lockdep_assert_not_held (&tasklist_lock); -spin_lock_irqsave (&pidmap_lock, flags); +spin_lock (&pidmap_lock);for (i =0 ; i <= pid->level; i++) {struct upid *upid = pid->numbers + i;struct pid_namespace *ns = upid->ns;
@@ -142,7 +141,7 @@ void free_pid(struct pid *pid)idr_remove (&ns->idr, upid->nr); }pidfs_remove_pid (pid); -spin_unlock_irqrestore (&pidmap_lock, flags); +spin_unlock (&pidmap_lock);call_rcu (&pid->rcu, delayed_put_pid); }
@@ -210,7 +209,7 @@ struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid, }idr_preload (GFP_KERNEL ); -spin_lock_irq (&pidmap_lock); +spin_lock (&pidmap_lock);if (tid) { nr = idr_alloc(&tmp->idr, NULL, tid,
@@ -237,7 +236,7 @@ struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid, nr =idr_alloc_cyclic (&tmp->idr,NULL , pid_min, pid_max,GFP_ATOMIC ); } -spin_unlock_irq (&pidmap_lock); +spin_unlock (&pidmap_lock);idr_preload_end ();if (nr <0 ) {
@@ -271,7 +270,7 @@ struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid, upid = pid->numbers + ns->level;idr_preload (GFP_KERNEL ); -spin_lock_irq (&pidmap_lock); +spin_lock (&pidmap_lock);if (!(ns->pid_allocated &PIDNS_ADDING )) goto out_unlock;pidfs_add_pid (pid);
@@ -280,18 +279,18 @@ struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid,idr_replace (&upid->ns->idr, pid, upid->nr); upid->ns->pid_allocated++; } -spin_unlock_irq (&pidmap_lock); +spin_unlock (&pidmap_lock);idr_preload_end ();return pid; out_unlock: -spin_unlock_irq (&pidmap_lock); +spin_unlock (&pidmap_lock);idr_preload_end ();put_pid_ns (ns); out_free: -spin_lock_irq (&pidmap_lock); +spin_lock (&pidmap_lock);while (++i <= ns->level) { upid = pid->numbers + i;idr_remove (&upid->ns->idr, upid->nr);
@@ -301,7 +300,7 @@ struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid,if (ns->pid_allocated ==PIDNS_ADDING )idr_set_cursor (&ns->idr,0 ); -spin_unlock_irq (&pidmap_lock); +spin_unlock (&pidmap_lock);kmem_cache_free (ns->pid_cachep, pid);return ERR_PTR (retval);
@@ -309,9 +308,9 @@ struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid,void disable_pid_allocation (struct pid_namespace *ns) { -spin_lock_irq (&pidmap_lock); +spin_lock (&pidmap_lock); ns->pid_allocated &= ~PIDNS_ADDING ; -spin_unlock_irq (&pidmap_lock); +spin_unlock (&pidmap_lock); }struct pid *find_pid_ns (int nr,struct pid_namespace *ns)
Signature · 1 line
2.43.0
Quoted 2 lines · Mateusz Guzik
I fixed fat-fingering and touchedup some commit messages. git diff between the old and new branch does not show any code changes.
Thanks for doing these changes.
Quoted 81 lines
hopefully we will be done here after this iteration 8->
old cover letter:
The clone side contends against exit side in a way which avoidably exacerbates the problem by the latter waiting on locks held by the former while holding the
tasklist_lock.Whacking this for both
add_device_randomnessand pids allocation gives me a 15% speed up for thread creation/destruction in a 24-core vm.The random patch is worth about 4%.
The new bottleneck is
pidmap_lockitself, with the biggest problem being the allocation itself taking the lock *twice*.Bench (plop into will-it-scale):
$ cat tests/threadspawn1.cchar
*testcase_description= "Thread creation and teardown";static void *worker(void *arg)
{ return (NULL); }void
testcase(unsigned long long *iterations, unsigned long nr){ pthread_t thread; int error; while (1) { error = pthread_create(&thread, NULL, worker, NULL); assert(error == 0); error = pthread_join(thread, NULL); assert(error == 0); (*iterations)++; } }v6:
- expand on the commit message in 3/5 pid: sprinkletasklist_lockasserts
- move fat-fingeredfree_pidscall to the right patchv5:
- whackscripts/selinux/genheaders/genheaderswhich accidentally got in
- rebased on next-20250205v4:
- justify movingget_pidin the commit message with a one-liner
- drop the tty unref patch -- it is completely optional and Oleg has his own variant
- add the ACK by Olegv3:
- keep procfs flush where it was, instead hoistget_pidoutside of the lock
- makedetach_pidet al accept an array argument of pids to populate
- sprinkle asserts
- drop irq trips aroundpidmap_lock
- move tty unref outside oftasklist_lockMateusz Guzik (5):
exit: performadd_device_randomness()withouttasklist_lock
exit: hoistget_pid()inrelease_task()outside oftasklist_lock
pid: sprinkletasklist_lockasserts
pid: performfree_pid()calls outside oftasklist_lock
pid: drop irq disablement aroundpidmap_lockinclude/linux/pid.h | 7 ++-- kernel/exit.c | 36 +++++++++++++------- kernel/pid.c | 82 +++++++++++++++++++++++++-------------------- kernel/sys.c | 14 +++++--- 4 files changed, 82 insertions(+), 57 deletions(-)Signature · 1 line
2.43.0
Christian Brauner
AppliedQuoted 8 lines · Mateusz Guzik
I fixed fat-fingering and touchedup some commit messages. git diff between the old and new branch does not show any code changes.
hopefully we will be done here after this iteration 8->
old cover letter:
[...]
This is great!
---
Applied to the kernel-6.15.tasklist_lock branch of the vfs/vfs.git tree. Patches in the kernel-6.15.tasklist_lock branch should appear in linux-next soon.
Please report any outstanding bugs that were missed during review in a new review to the original patch series allowing us to drop it.
It's encouraged to provide Acked-bys and Reviewed-bys even though the patch has now been applied. If possible patch trailers will be updated.
Note that commit hashes shown below are subject to change due to rebase, trailer updates or similar. If in doubt, please check the listed branch.
tree: https://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs.git branch: kernel-6.15.tasklist_lock
[1/5] exit: perform add_device_randomness() without tasklist_lock
https://git.kernel.org/vfs/vfs/c/c8be3764feff [2/5] exit: hoist get_pid() in release_task() outside of tasklist_lock https://git.kernel.org/vfs/vfs/c/9ee2997cbf14 [3/5] pid: sprinkle tasklist_lock asserts https://git.kernel.org/vfs/vfs/c/7d0030e1660c [4/5] pid: perform free_pid() calls outside of tasklist_lock https://git.kernel.org/vfs/vfs/c/7b15589a575f [5/5] pid: drop irq disablement around pidmap_lock https://git.kernel.org/vfs/vfs/c/07e518c464e8
Quoted 2 lines · Christian Brauner
Applied to the kernel-6.15.tasklist_lock branch of the vfs/vfs.git tree.
Great.
Then perhaps you can also take the related
[PATCH v2 0/2] exit: change the release_task() paths to call flush_sigqueue() lockless https://lore.kernel.org/all/[email protected]/
changes?
Oleg.