| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Fri, 1 Mar 2019 02:04:48 +0800
From: Yu Chen <yu.c.chen@...el.com>
To: "Rafael J. Wysocki" <rafael@...nel.org>
Cc: "Rafael J. Wysocki" <rjw@...ysocki.net>,
Len Brown <lenb@...nel.org>,
Viresh Kumar <viresh.kumar@...aro.org>,
Srinivas Pandruvada <srinivas.pandruvada@...ux.intel.com>,
ACPI Devel Maling List <linux-acpi@...r.kernel.org>,
Linux PM <linux-pm@...r.kernel.org>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH][RFC] ACPI: add "processor.broadcast_ppc" hook to
broadcast _PPC to all online CPUs
On Thu, Feb 14, 2019 at 11:21:13AM +0100, Rafael J. Wysocki wrote:
> On Wed, Feb 13, 2019 at 5:44 PM Yu Chen <yu.c.chen@...el.com> wrote:
> >
> > Hi Rafael,
> > On Mon, Feb 11, 2019 at 11:41:26AM +0100, Rafael J. Wysocki wrote:
> > > On Sat, Feb 9, 2019 at 12:54 PM Chen Yu <yu.c.chen@...el.com> wrote:
> > > >
> > > > On Dell Inc. XPS13 9333, the BIOS changes the value of
> > > > MSR_IA32_MISC_ENABLE_TURBO_DISABLE at runtime (e.g., when
> > > > the power source changes), the maximum frequency of the
> > > > CPU is not updated accordingly. This is due to the policy's
> > > > cpuinfo.max is not updated when _PPC notifier fires.
> > > >
> > > > Fix this problem by updating the policy's cpuinfo.max
> > > > and broadcast the _PPC notifier to all online CPUs.
> > > >
> > > > Link: https://bugzilla.kernel.org/show_bug.cgi?id=200759
> > > > Reported-and-tested-by: Gabriele Mazzotta <gabriele.mzt@...il.com>
> > > > Originally-by: Srinivas Pandruvada <srinivas.pandruvada@...ux.intel.com>
> > > > Signed-off-by: Chen Yu <yu.c.chen@...el.com>
> > > > ---
> > > > drivers/acpi/processor_perflib.c | 16 ++++++++++++++--
> > > > drivers/cpufreq/cpufreq.c | 2 ++
> > > > drivers/cpufreq/intel_pstate.c | 15 ++++++++++++++-
> > > > 3 files changed, 30 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/drivers/acpi/processor_perflib.c b/drivers/acpi/processor_perflib.c
> > > > index a303fd0e108c..737dbf5aa7f7 100644
> > > > --- a/drivers/acpi/processor_perflib.c
> > > > +++ b/drivers/acpi/processor_perflib.c
> > > > @@ -63,6 +63,10 @@ module_param(ignore_ppc, int, 0644);
> > > > MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
> > > > "limited by BIOS, this should help");
> > > >
> > > > +static int broadcast_ppc;
> > > > +module_param(broadcast_ppc, int, 0644);
> > > > +MODULE_PARM_DESC(broadcast_ppc, "Broadcast the ppc to all online CPUs");
> > > > +
> > > > #define PPC_REGISTERED 1
> > > > #define PPC_IN_USE 2
> > > >
> > > > @@ -180,8 +184,16 @@ void acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
> > > > else
> > > > acpi_processor_ppc_ost(pr->handle, 0);
> > > > }
> > > > - if (ret >= 0)
> > > > - cpufreq_update_policy(pr->id);
> > > > + if (ret >= 0) {
> > > > + if (broadcast_ppc) {
> > > > + int cpu;
> > > > +
> > > > + for_each_possible_cpu(cpu)
> > > > + cpufreq_update_policy(cpu);
> > > > + } else {
> > > > + cpufreq_update_policy(pr->id);
> > > > + }
> > > > + }
> > > > }
> > > >
> > > > int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
> > > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > > > index e35a886e00bc..95e08816b512 100644
> > > > --- a/drivers/cpufreq/cpufreq.c
> > > > +++ b/drivers/cpufreq/cpufreq.c
> > > > @@ -2237,6 +2237,8 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy,
> > > >
> > > > policy->min = new_policy->min;
> > > > policy->max = new_policy->max;
> > > > + policy->cpuinfo.max_freq = new_policy->cpuinfo.max_freq;
> > > > + policy->cpuinfo.min_freq = new_policy->cpuinfo.min_freq;
> > > > trace_cpu_frequency_limits(policy);
> > > >
> > > > policy->cached_target_freq = UINT_MAX;
> > > > diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
> > > > index dd66decf2087..e1881313c396 100644
> > > > --- a/drivers/cpufreq/intel_pstate.c
> > > > +++ b/drivers/cpufreq/intel_pstate.c
> > > > @@ -2081,11 +2081,24 @@ static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
> > > >
> > > > static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
> > > > {
> > > > + int max_freq;
> > > > struct cpudata *cpu = all_cpu_data[policy->cpu];
> > > >
> > > > update_turbo_state();
> > >
> > > Well, update_turbo_state() should handle the case at hand already.
> > >
> > > That's what it's for actually: It checks if
> > > MSR_IA32_MISC_ENABLE_TURBO_DISABLE is set and sets
> > > global.turbo_disabled is that's the case.
> > >
> > > Why isn't that sufficient?
> > >
> > update_turbo_state() changes the flag of global.turbo_diabled but we
> > need to also leverage it to adjust the policy.max accordingly. This is why
> > we add intel_pstate_get_max_freq() to get the updated max freq in
> > intel_pstate_verify_policy().
>
> Yes, that's why intel_pstate_verify_policy() passes the return value
> of intel_pstate_get_max_freq() as the second arg
> cpufreq_verify_within_limits(), so really my question was about why
> cpuinfo.max_freq needed to be updated (below).
>
Ok.
> > > > + max_freq = intel_pstate_get_max_freq(cpu);
> > > > +
> > > > + if (acpi_ppc && policy->max == policy->cpuinfo.max_freq &&
> > > > + max_freq != policy->cpuinfo.max_freq) {
> > > > + /*
> > > > + * System was not running under any constraints, but the
> > > > + * current max possible frequency is changed. So reset
> > > > + * policy limits.
> > > > + */
> > > > + policy->cpuinfo.max_freq = policy->max = max_freq;
> > > > + }
> > >
> > > Why does policy->cpuinfo.max_freq need to be updated?
> > >
> > This is my understanding:
> > There's a corner case that, if the system boots with battery,
> > the max cpu frequency will not scale up if we plug the AC later.
>
> I see. The *initial* cpuinfo.max_freq may be too low. This part is
> missing from your patch changelog.
>
> The driver is not expected to update cpuinfo.max_freq after init.
> That may not actually break anything, even though it is racy in
> principle, but if it is done, it needs to be done in the "passive"
> mode too and that may be more problematic.
>
Do you mean updating it for "passive" mode might not be suitable?
> Anyway, this is more fundamental than you seem to be thinking.
>
> > According to the log provided by Gabriele Mazzotta, if the system
> > boot up with battery, then plug the AC after boot up, the max perf ratio
> > and policy->cpuinfo.max will remain 17 rather than increasing to
> > 30(when AC plugged thus turbo enabled):
> >
> > [ 52.158810] CPU 0: _PPC is 6 - frequency limited
> > [ 52.158822] intel_pstate: set_policy cpuinfo.max 1700000 policy->max 1700000
> > [ 52.158825] intel_pstate: cpu:0 max_state 30 min_policy_perf:8 max_policy_perf:17
> > [ 52.158827] intel_pstate: cpu:0 global_min:8 global_max:30
> > [ 52.158829] intel_pstate: cpu:0 max_perf_ratio:17 min_perf_ratio:8
> >
> > This is caused by:
> > In current intel_pstate, there's only one chance for policy.cpuinfo.max to get updated
> > which is during boot up in __intel_pstate_cpu_init(). If the turbo status changes,
> > we might need to also update the policy->cpuinfo.max to tell user that the hardware
> > status has changed.
> >
> > So we give it a chance to adjust the policy.cpuinfo.max and policy.max in
> > cpufreq_driver->verify() according to turbo status, this is what this patch mainly
> > aims to do.
> >
> > Besides, since on this platform there's only one _PPC notification for one CPU, it is
> > necessary to broadcast the notification to all CPUs on this package. And this patch
> > broadcast it to all online CPUs to make the change simpler.
>
> You're trying to make two substantial changes in one go, broadcasting
> _PPC and updating cpuinfo.max_freq. Don't do that, they need to be
> separate changes.
Ok, I'll send version 2 out which is composed of two modifications to
address different problems.
>
> Moreover, we may want to address the initial cpuinfo.max_freq issue in
> a different way.
Adjust the cpuinfo.max_freq in .verify() seems to be a proper place since
this callback is invoked at a very early stage for both active and passive
mode, and I did not see race condition for them. (Could not quite catch
what we talked last time on this, if you have proposed another suggestion
on how to update the cpuinfo.max_freq) :)
Thanks,
Ryan(Yu)
>
> Thanks,
> Rafael
Powered by blists - more mailing lists