Re: [PATCH v2] x86/resctrl: Fix WARN in get_domain_from_cpu()

[Tony Luck <tony.luck@intel.com>]

On Wed, Feb 21, 2024 at 02:59:43PM -0800, Reinette Chatre wrote:
> Hi Tony,
> 
> On 2/21/2024 11:31 AM, Tony Luck wrote:
> > reset_all_ctrls() and resctrl_arch_update_domains() use on_each_cpu_mask()
> > to call rdt_ctrl_update() on potentially one CPU from each domain.
> > 
> > But this means rdt_ctrl_update() needs to figure out which domain to
> > apply changes to. Doing so requires a search of all domains in a resource,
> > which can only be done safely if cpus_lock is held. Both callers do hold
> > this lock, but there isn't a way for a function called on another CPU
> > via IPI to verify this.
> > 
> > Fix by adding the target domain to the msr_param structure and passing an
> > array with CDP_NUM_TYPES entries. Then calling for each domain separately
> > using smp_call_function_single()
> 
> This work contains no changes to get_domain_from_cpu(). I expect the WARN
> within it to be removed as intended with [1] and then this work can build
> on that without urgency. As I understand, to support the stated goal of this
> work, I expect get_domain_from_cpu() in the end to not have any WARN or
> IS_ENABLED checks, but just a lockdep_assert_cpus_held(). 
> 
> Do you have different expectations?

Same expectations. Boris should apply the simple fix (delete the WARN
that is giving a false positive) for this current cycle.

If there is support for my patch (with changes/fixes you point out
below), then it could be added in the future and get_domain_from_cpu()
can use lockdep_assert_cpus_held().

> > Change the low level cat_wrmsr(), mba_wrmsr_intel(), and mba_wrmsr_amd()
> > functions to just take a msr_param structure since it contains the
> > rdt_resource and rdt_domain information.
> 
> Could moving the rdt_domain into msr_param be done in a separate patch?

I can break it into more pieces if there is enthusiam to apply it.

> ...
> 
> > diff --git a/arch/x86/kernel/cpu/resctrl/ctrlmondata.c b/arch/x86/kernel/cpu/resctrl/ctrlmondata.c
> > index 7997b47743a2..09f6e624f1bb 100644
> > --- a/arch/x86/kernel/cpu/resctrl/ctrlmondata.c
> > +++ b/arch/x86/kernel/cpu/resctrl/ctrlmondata.c
> > @@ -272,22 +272,6 @@ static u32 get_config_index(u32 closid, enum resctrl_conf_type type)
> >  	}
> >  }
> >  
> > -static bool apply_config(struct rdt_hw_domain *hw_dom,
> > -			 struct resctrl_staged_config *cfg, u32 idx,
> > -			 cpumask_var_t cpu_mask)
> > -{
> > -	struct rdt_domain *dom = &hw_dom->d_resctrl;
> > -
> > -	if (cfg->new_ctrl != hw_dom->ctrl_val[idx]) {
> > -		cpumask_set_cpu(cpumask_any(&dom->cpu_mask), cpu_mask);
> > -		hw_dom->ctrl_val[idx] = cfg->new_ctrl;
> > -
> > -		return true;
> > -	}
> > -
> > -	return false;
> > -}
> > -
> >  int resctrl_arch_update_one(struct rdt_resource *r, struct rdt_domain *d,
> >  			    u32 closid, enum resctrl_conf_type t, u32 cfg_val)
> >  {
> > @@ -304,59 +288,50 @@ int resctrl_arch_update_one(struct rdt_resource *r, struct rdt_domain *d,
> >  	msr_param.res = r;
> >  	msr_param.low = idx;
> >  	msr_param.high = idx + 1;
> > -	hw_res->msr_update(d, &msr_param, r);
> > +	hw_res->msr_update(&msr_param);
> >  
> 
> Is this missing setting the domain in msr_param?

Indeed yes.

> >  	return 0;
> >  }
> >  
> >  int resctrl_arch_update_domains(struct rdt_resource *r, u32 closid)
> >  {
> > +	struct msr_param msr_param[CDP_NUM_TYPES];
> >  	struct resctrl_staged_config *cfg;
> >  	struct rdt_hw_domain *hw_dom;
> > -	struct msr_param msr_param;
> >  	enum resctrl_conf_type t;
> > -	cpumask_var_t cpu_mask;
> >  	struct rdt_domain *d;
> > +	bool need_update;
> > +	int cpu;
> >  	u32 idx;
> >  
> >  	/* Walking r->domains, ensure it can't race with cpuhp */
> >  	lockdep_assert_cpus_held();
> >  
> > -	if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
> > -		return -ENOMEM;
> > -
> > -	msr_param.res = NULL;
> > +	memset(msr_param, 0, sizeof(msr_param));
> >  	list_for_each_entry(d, &r->domains, list) {
> >  		hw_dom = resctrl_to_arch_dom(d);
> > +		need_update = false;
> >  		for (t = 0; t < CDP_NUM_TYPES; t++) {
> >  			cfg = &hw_dom->d_resctrl.staged_config[t];
> >  			if (!cfg->have_new_ctrl)
> >  				continue;
> >  
> >  			idx = get_config_index(closid, t);
> > -			if (!apply_config(hw_dom, cfg, idx, cpu_mask))
> > +			if (cfg->new_ctrl == hw_dom->ctrl_val[idx])
> >  				continue;
> > -
> > -			if (!msr_param.res) {
> > -				msr_param.low = idx;
> > -				msr_param.high = msr_param.low + 1;
> > -				msr_param.res = r;
> > -			} else {
> > -				msr_param.low = min(msr_param.low, idx);
> > -				msr_param.high = max(msr_param.high, idx + 1);
> > -			}
> > +			hw_dom->ctrl_val[idx] = cfg->new_ctrl;
> > +			cpu = cpumask_any(&d->cpu_mask);
> > +
> > +			msr_param[t].low = idx;
> > +			msr_param[t].high = msr_param[t].low + 1;
> > +			msr_param[t].res = r;
> > +			msr_param[t].dom = d;
> > +			need_update = true;
> >  		}
> > +		if (need_update)
> > +			smp_call_function_single(cpu, rdt_ctrl_update, &msr_param, 1);
> 
> It is not clear to me why it is needed to pass this additional data. Why not
> just use the original mechanism of letting the low and high of msr_param span the
> multiple indices that need updating? There can still be a "need_update" but it
> can be set when msr_param gets its first data. Any other index needing updating
> can just update low/high and a single msr_param can be used.

For some reason this morning I thought that the domain needed to be
different. It isn't, so keeping the code that just adjusts the range
of MSRs will work just fine.

The "need_update" variable isn't required. I will move the
	msr_param.res = NULL;
inside the for each domain loop, and can use non-NULL value to
decide whether to IPI a CPU.

-Tony