Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230523085045.29391-1-kweifat@gmail.com> <CAJZ5v0ifp1088wY7o=7pnBVBm=_3H0M4sfq6=gmyChZD6R9g1g@mail.gmail.com>
 <b5e35f904174905d8f90df3f49944b22389126c7.camel@linux.intel.com>
 <CAJZ5v0ioQm95ZQ5LCCoDtVNX1TVQN_=sgzB_RRe5SAOOucpWJg@mail.gmail.com> <96d81ba4c501134c77869a08b843b39fc1bf7526.camel@linux.intel.com>
In-Reply-To: <96d81ba4c501134c77869a08b843b39fc1bf7526.camel@linux.intel.com>
From:   "Rafael J. Wysocki" <rafael@kernel.org>
Date:   Tue, 23 May 2023 16:34:40 +0200
Message-ID: <CAJZ5v0gc7x6x+aAic+z=usJY-1nLMOFheEj4-JfPmB5-tWRRbA@mail.gmail.com>
Subject: Re: [PATCH v2] cpufreq: intel_pstate: Avoid initializing variables prematurely
To:     srinivas pandruvada <srinivas.pandruvada@linux.intel.com>
Cc:     "Rafael J. Wysocki" <rafael@kernel.org>,
        Fieah Lim <kweifat@gmail.com>, lenb@kernel.org,
        linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8BIT
Precedence: bulk

On Tue, May 23, 2023 at 4:02 PM srinivas pandruvada
<srinivas.pandruvada@linux.intel.com> wrote:
>
> On Tue, 2023-05-23 at 15:37 +0200, Rafael J. Wysocki wrote:
> > On Tue, May 23, 2023 at 2:20 PM srinivas pandruvada
> > <srinivas.pandruvada@linux.intel.com> wrote:
> > >
> > > On Tue, 2023-05-23 at 13:08 +0200, Rafael J. Wysocki wrote:
> > > > On Tue, May 23, 2023 at 10:51 AM Fieah Lim <kweifat@gmail.com>
> > > > wrote:
> > > > >
> > > > > We should avoid initializing some rather expensive data
> > > > > when the function has a chance to bail out earlier
> > > > > before actually using it.
> > > > > This applies to the following initializations:
> > > > >
> > > > >  - cpudata *cpu = all_cpu_data; (in everywhere)
> > > > >  - this_cpu = smp_processor_id(); (in notify_hwp_interrupt)
> > > > >  - hwp_cap = READ_ONCE(cpu->hwp_cap_cached); (in
> > > > > intel_pstate_hwp_boost_up)
> > > > >
> > > > > These initializations are premature because there is a chance
> > > > > that the function will bail out before actually using the data.
> > > > > I think this qualifies as a micro-optimization,
> > > > > especially in such a hot path.
> > > > >
> > > > > While at it, tidy up how and when we initialize
> > > > > all of the cpu_data pointers, for the sake of consistency.
> > > > >
> > > > > A side note on the intel_pstate_cpu_online change:
> > > > > we simply don't have to initialize cpudata just
> > > > > for the pr_debug, while policy->cpu is being there.
> > > > >
> > > > > Signed-off-by: Fieah Lim <kweifat@gmail.com>
> > > > > ---
> > > > > V1 -> V2: Rewrite changelog for better explanation.
> > > > >
> > >
> > > [...]
> > >
> > > > >  void notify_hwp_interrupt(void)
> > > > >  {
> > > > > -       unsigned int this_cpu = smp_processor_id();
> > > > > +       unsigned int this_cpu;
> > > > >         struct cpudata *cpudata;
> > > > >         unsigned long flags;
> > > > >         u64 value;
> > > > > @@ -1591,6 +1593,8 @@ void notify_hwp_interrupt(void)
> > > > >         if (!(value & 0x01))
> > > > >                 return;
> > > > >
> > > > > +       this_cpu = smp_processor_id();
> > > > > +
> > > > >         spin_lock_irqsave(&hwp_notify_lock, flags);
> > > > >
> > > > >         if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
> > > >
> > > > This is a place where it may really matter for performance, but
> > > > how
> > > > much?  Can you actually estimate this?
> > >
> > > If DEBUG_PREEMPT is defined
> > > ~12 instructions (most of them with latency = 1 in dependency
> > > chain)
> >
> > I really meant "estimate the effect of this change on performance",
> > because I'm not sure if it is going to be visible in any test.
> >
> > But yes, skipping it if not needed at least makes some sense.
> It will have neglible effect.

Well, I agree.

> I can measure it, but may be next week.

No need really.