race between cgroup_subsys->fork() and cgroup_migrate()
Subsystems (3)
Hello,
I am trying to backport cgroup_pids.c and I fail to understand pids_fork() which does
css = task_get_css(task, pids_cgrp_id);
pids = css_pids(css);
/*
* If the association has changed, we have to revert and reapply the
* charge/uncharge on the wrong hierarchy to the current one. Since
* the association can only change due to an organisation event, its
* okay for us to ignore the limit in this case.
*/
if (pids != old_pids) {
pids_uncharge(old_pids, 1);
pids_charge(pids, 1);
}But if the association has changed, pids_can_attach() which moved the child into another cgrp has called pids_uncharge(old_pids) too?
IOW. Suppose that the new child is moved right before cgroup_post_fork() does
for_each_subsys_which(...) ss->fork(child);
doesn't this mean that after ss->fork() we do the same sequence
pids_uncharge(old_pids, 1); pids_charge(pids, 1);
twice? Note that threadgroup_change_begin/end depends on CLONE_THREAD. So we can actually hit WARN_ON() in pids_cancel().
However, we can't simply remove this uncharge/charge afaics. We need this in case when the parent was moved to another cgroup before cgroup_post_fork(), and then css_set_move_task() moves the child.
I know almost nothing about cgroups, perhaps I missed something, please correct me.
If am right. How about the patch below? percpu_down_read() is cheap. And we can simplify cgroup_pids after this change.
And. We can probably unify cgroup_threadgroup_rwsem and dup_mmap_sem. Note that if we take cgroup_threadgroup_rwsem for reading if CLONE_THREAD, otherwise we take another percpu-rwsem in dup_mmap(), dup_mmap_sem.
Or I am totally confused?
Oleg.
kernel/fork.c +2 -4
@@ -1368,8 +1368,7 @@ static struct task_struct *copy_process( p->real_start_time =ktime_get_boot_ns (); p->io_context =NULL ; p->audit_context =NULL ; -if (clone_flags &CLONE_THREAD ) -threadgroup_change_begin (current); +threadgroup_change_begin (current);cgroup_fork (p);#ifdef CONFIG_NUMA p->mempolicy =mpol_dup (p->mempolicy);
@@ -1610,8 +1609,7 @@ static struct task_struct *copy_process(proc_fork_connector (p);cgroup_post_fork (p, cgrp_ss_priv); -if (clone_flags &CLONE_THREAD ) -threadgroup_change_end (current); +threadgroup_change_end (current);perf_event_fork (p);trace_task_newtask (p, clone_flags);
kernel/cgroup_pids.c +2 -19
@@ -243,27 +243,10 @@ static void pids_cancel_fork(struct taskstatic void pids_fork (struct task_struct *task,void *priv) { -struct cgroup_subsys_state *css; -struct cgroup_subsys_state *old_css = priv; -struct pids_cgroup *pids; -struct pids_cgroup *old_pids =css_pids (old_css); - - css =task_get_css (task, pids_cgrp_id); - pids =css_pids (css); - -/* -* If the association has changed, we have to revert and reapply the -* charge/uncharge on the wrong hierarchy to the current one. Since -* the association can only change due to an organisation event, its -* okay for us to ignore the limit in this case. -*/ -if (pids != old_pids) { -pids_uncharge (old_pids,1 ); -pids_charge (pids,1 ); - } +struct cgroup_subsys_state *css = priv; +WARN_ON (task_css (task, pids_cgrp_id) != css);css_put (css); -css_put (old_css); }static void pids_free (struct task_struct *task)
Hello, Oleg.
Quoted 9 lines · Oleg Nesterov
IOW. Suppose that the new child is moved right before
cgroup_post_fork()doesfor_each_subsys_which(...) ss->fork(child);doesn't this mean that after
ss->fork()we do the same sequencepids_uncharge(old_pids, 1); pids_charge(pids, 1);
You're absolutely right.
Quoted 9 lines
twice? Note that threadgroup_change_begin/end depends on
CLONE_THREAD. So we can actually hitWARN_ON()inpids_cancel().However, we can't simply remove this uncharge/charge afaics. We need this in case when the parent was moved to another cgroup before
cgroup_post_fork(), and thencss_set_move_task()moves the child.I know almost nothing about cgroups, perhaps I missed something, please correct me.
I can't think of anything better than what you're proposing. Thanks a lot for tracking it down and fixing it.
Quoted 6 lines
If am right. How about the patch below?
percpu_down_read()is cheap. And we can simplifycgroup_pidsafter this change.And. We can probably unify
cgroup_threadgroup_rwsemanddup_mmap_sem. Note that if we takecgroup_threadgroup_rwsemfor reading ifCLONE_THREAD, otherwise we take another percpu-rwsem indup_mmap(),dup_mmap_sem.
Sounds perfect. As this needs to go through -stable, can you please resend the patch with proper description and SOB? Please also update the now incorrect comment in can_attach.
Thanks a lot!
Signature · 1 line
tejun
Quoted 3 lines · Tejun Heo
Sounds perfect. As this needs to go through -stable, can you please resend the patch with proper description and SOB? Please also update the now incorrect comment in
can_attach.
Ooh, the patch triggers RCU warning from task_css(). It's spurious and I think the right thing to do at least for now is using task_css_check() and explain what's going on.
Thanks.
Signature · 1 line
tejun
Quoted 8 lines · Tejun Heo
Quoted 3 lines · Tejun Heo
Sounds perfect. As this needs to go through -stable, can you please resend the patch with proper description and SOB? Please also update the now incorrect comment in
can_attach.Ooh, the patch triggers RCU warning from
task_css(). It's spurious and I think the right thing to do at least for now is usingtask_css_check()and explain what's going on.
And I spotted a bug in cgroup core in controller enable path. While the race between fork and attach exists and needs to be fixed, I think the cgroup core bug is the main source of discrepancy. Will update once I learn more.
Thanks.
Signature · 1 line
tejun
Quoted 5 lines · Tejun Heo
Quoted 3 lines · Tejun Heo
Sounds perfect. As this needs to go through -stable, can you please resend the patch with proper description and SOB? Please also update the now incorrect comment in
can_attach.
OK, will do tomorrow (sorry, can't do today).
Quoted 3 lines
Ooh, the patch triggers RCU warning from
task_css(). It's spurious and I think the right thing to do at least for now is usingtask_css_check()and explain what's going on.
WARN_ON() in pids_fork() I guess. Thanks. I didn't expect you will actually apply this patch, I didn't even try to compile it ;)
Plus this patch forgets to unconditionalize another threadgroup_change_end() in the error path of copy_process().
Again, if we do this, we can make other cleanups/simplifications. For example, we can kill cgrp_ss_priv[CGROUP_CANFORK_COUNT] in copy_process().
But I see another email from you, will reply in a minute.
Oleg.
Quoted 5 lines · Tejun Heo
And I spotted a bug in cgroup core in controller enable path. While the race between fork and attach exists and needs to be fixed, I think the cgroup core bug is the main source of discrepancy. Will update once I learn more.
OK. I do not know what exactly do you mean, perhaps if you fix this problem the race between fork and attach goes away and in this case the fix I sent is not needed? I'll wait for more info.
Oleg.
Quoted 2 lines
Again, if we do this, we can make other cleanups/simplifications. For example, we can kill cgrp_ss_priv[CGROUP_CANFORK_COUNT] in
copy_process().
Thank god, I never liked that code. ;)
Signature · 2 lines
Aleksa Sarai (cyphar) www.cyphar.com