Re: [mm] 4e2c82a409: ltp.overcommit_memory01.fail

[Feng Tang <feng.tang@intel.com>]

On Thu, Jul 09, 2020 at 10:15:19PM +0800, Feng Tang wrote:
> Hi Qian Cai,
> 
> On Thu, Jul 09, 2020 at 09:40:40AM -0400, Qian Cai wrote:
> > > > > Can we change the batch firstly, then sync the global counter, finally
> > > > > change the overcommit policy?
> > > > 
> > > > These reorderings are really head scratching :)
> > > > 
> > > > I've thought about this before when Qian Cai first reported the warning
> > > > message, as kernel had a check: 
> > > > 
> > > > 	VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
> > > > 			-(s64)vm_committed_as_batch * num_online_cpus(),
> > > > 			"memory commitment underflow");
> > > > 
> > > > If the batch is decreased first, the warning will be easier/earlier to be
> > > > triggered, so I didn't brought this up when handling the warning message.
> > > > 
> > > > But it might work now, as the warning has been removed.
> > > 
> > > I tested the reorder way, and the test could pass in 100 times run. The
> > > new order when changing policy to OVERCOMMIT_NEVER:
> > >   1. re-compute the batch ( to the smaller one)
> > >   2. do the on_each_cpu sync
> > >   3. really change the policy to NEVER.
> > > 
> > > It solves one of previous concern, that after the sync is done on cpuX,
> > > but before the whole sync on all cpus are done, there is a window that
> > > the percpu-counter could be enlarged again.
> > > 
> > > IIRC Andi had concern about read side cost when doing the sync, my
> > > understanding is most of the readers (malloc/free/map/unmap) are using
> > > percpu_counter_read_positive, which is a fast path without involving lock.
> > > 
> > > As for the problem itself, I agree with Michal's point, that usually there
> > > is no normal case that will change the overcommit_policy too frequently.
> > > 
> > > The code logic is mainly in overcommit_policy_handler(), based on the
> > > previous sync fix. please help to review, thanks!
> > > 
> > > int overcommit_policy_handler(struct ctl_table *table, int write, void *buffer,
> > > 		size_t *lenp, loff_t *ppos)
> > > {
> > > 	int ret;
> > > 
> > > 	if (write) {
> > > 		int new_policy;
> > > 		struct ctl_table t;
> > > 
> > > 		t = *table;
> > > 		t.data = &new_policy;
> > > 		ret = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
> > > 		if (ret)
> > > 			return ret;
> > > 
> > > 		mm_compute_batch(new_policy);
> > > 		if (new_policy == OVERCOMMIT_NEVER)
> > > 			schedule_on_each_cpu(sync_overcommit_as);
> > > 		sysctl_overcommit_memory = new_policy;
> > > 	} else {
> > > 		ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
> > > 	}
> > > 
> > > 	return ret;
> > > }
> > 
> > Rather than having to indent those many lines, how about this?
> 
> Thanks for the cleanup suggestion.
> 
> > t = *table;
> > t.data = &new_policy;
> 
> The input table->data is actually &sysctl_overcommit_memory, so
> there is a problem for "read" case, it will return the 'new_policy'
> value instead of real sysctl_overcommit_memory.
> 
> It should work after adding a check
> 	if (write)
> 		t.data = &new_policy;
> 
> > ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
> 			    --> &t
 
Give it a second thought, my previous way has more indents and lines,
but it is easier to be understood that we have special handling for
'write' case. So I would prefer using it. 

Thoughts?

Thanks,
Feng

> Thanks,
> Feng
> 	
> > if (ret || !write)
> > 	return ret;
> > mm_compute_batch(new_policy);
> > if (new_policy == OVERCOMMIT_NEVER)
> > 	schedule_on_each_cpu(sync_overcommit_as);
> > 
> > sysctl_overcommit_memory = new_policy;
> > return ret;