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;
Date:   Mon, 18 Apr 2022 15:48:07 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Peter Gonda <pgonda@google.com>
Cc:     Mingwei Zhang <mizhang@google.com>, kvm <kvm@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] KVM: SEV: Add cond_resched() to loop in
 sev_clflush_pages()
Message-ID: <Yl2IN6CHQzkts4XE@google.com>
References: <20220330164306.2376085-1-pgonda@google.com>
 <CAL715W+S-SJwXBhYO=_T-9uAPLt6cQ-Hn+_+ehefAh6+kQ_zOA@mail.gmail.com>
 <YkYdlfYM/FWlMqMg@google.com>
 <CAL715WLhy7EkJCyO7vzak3O8iw8GDRHkPF8aRtDedPXO1vx_Qw@mail.gmail.com>
 <Yk3bSmQTspjZHUZf@google.com>
 <CAMkAt6obVDW_LFvQzUYw6v7okiNq1KAbUOMoM3bN6zeJUGg6Xw@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAMkAt6obVDW_LFvQzUYw6v7okiNq1KAbUOMoM3bN6zeJUGg6Xw@mail.gmail.com>
Precedence: bulk

On Wed, Apr 06, 2022, Peter Gonda wrote:
> On Wed, Apr 6, 2022 at 12:26 PM Sean Christopherson <seanjc@google.com> wrote:
> >
> > On Wed, Apr 06, 2022, Mingwei Zhang wrote:
> > > Hi Sean,
> > >
> > > > > > diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
> > > > > > index 75fa6dd268f0..c2fe89ecdb2d 100644
> > > > > > --- a/arch/x86/kvm/svm/sev.c
> > > > > > +++ b/arch/x86/kvm/svm/sev.c
> > > > > > @@ -465,6 +465,7 @@ static void sev_clflush_pages(struct page *pages[], unsigned long npages)
> > > > > >                 page_virtual = kmap_atomic(pages[i]);
> > > > > >                 clflush_cache_range(page_virtual, PAGE_SIZE);
> > > > > >                 kunmap_atomic(page_virtual);
> > > > > > +               cond_resched();
> > > > >
> > > > > If you add cond_resched() here, the frequency (once per 4K) might be
> > > > > too high. You may want to do it once per X pages, where X could be
> > > > > something like 1G/4K?
> > > >
> > > > No, every iteration is perfectly ok.  The "cond"itional part means that this will
> > > > reschedule if and only if it actually needs to be rescheduled, e.g. if the task's
> > > > timeslice as expired.  The check for a needed reschedule is cheap, using
> > > > cond_resched() in tight-ish loops is ok and intended, e.g. KVM does a reched
> > > > check prior to enterring the guest.
> > >
> > > Double check on the code again. I think the point is not about flag
> > > checking. Obviously branch prediction could really help. The point I
> > > think is the 'call' to cond_resched(). Depending on the kernel
> > > configuration, cond_resched() may not always be inlined, at least this
> > > is my understanding so far? So if that is true, then it still might
> > > not always be the best to call cond_resched() that often.
> >
> > Eh, compared to the cost of 64 back-to-back CLFLUSHOPTs, the cost of __cond_resched()
> > is peanuts.  Even accounting for the rcu_all_qs() work, it's still dwarfed by the
> > cost of flushing data from the cache.  E.g. based on Agner Fog's wonderful uop
> > latencies[*], the actual flush time for a single page is going to be upwards of
> > 10k cycles, whereas __cond_resched() is going to well under 100 cycles in the happy
> > case of no work.  Even if those throughput numbers are off by an order of magnitude,
> > e.g. CLFLUSHOPT can complete in 15 cycles, that's still ~1k cycles.
> >
> > Peter, don't we also theoretically need cond_resched() in the loops in
> > sev_launch_update_data()?  AFAICT, there's no articifical restriction on the size
> > of the payload, i.e. the kernel is effectively relying on userspace to not update
> > large swaths of memory.
> 
> Yea we probably do want to cond_resched() in the for loop inside of
> sev_launch_update_data(). Ithink in  sev_dbg_crypt() userspace could
> request a large number of pages to be decrypted/encrypted for
> debugging but se have a call to sev_pin_memory() in the loop so that
> will have a cond_resded() inside of __get_users_pages(). Or should we
> have a cond_resded() inside of the loop in sev_dbg_crypt() too?

I believe sev_dbg_crypt() needs a cond_resched() of its own, sev_pin_memory()
isn't guaranteed to get into the slow path of internal_get_user_pages_fast().