Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210809152651.2297337-1-maz@kernel.org> <20210809152651.2297337-14-maz@kernel.org>
 <CAOQ_QshjRWh=onT-j3dWgmVfnAXsMFJoz0i7OEezQxkW4O9KZA@mail.gmail.com> <87im0ebi9m.wl-maz@kernel.org>
In-Reply-To: <87im0ebi9m.wl-maz@kernel.org>
From:   Oliver Upton <oupton@google.com>
Date:   Mon, 9 Aug 2021 11:17:38 -0700
Message-ID: <CAOQ_Qsh6h_afu_-WjJ2JTP=gYBYW4hbC92qmxzyJ8dgRCCgLKw@mail.gmail.com>
Subject: Re: [PATCH 13/13] arm64: Add CNT{P,V}CTSS_EL0 alternatives to cnt{p,v}ct_el0
To:     Marc Zyngier <maz@kernel.org>
Cc:     linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        Mark Rutland <mark.rutland@arm.com>,
        Daniel Lezcano <daniel.lezcano@linaro.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Peter Shier <pshier@google.com>,
        Raghavendra Rao Ananta <rananta@google.com>,
        Ricardo Koller <ricarkol@google.com>,
        Will Deacon <will@kernel.org>,
        Catalin Marinas <catalin.marinas@arm.com>,
        Linus Walleij <linus.walleij@linaro.org>,
        kernel-team@android.com
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Aug 9, 2021 at 11:11 AM Marc Zyngier <maz@kernel.org> wrote:
>
> On Mon, 09 Aug 2021 17:42:00 +0100,
> Oliver Upton <oupton@google.com> wrote:
> >
> > On Mon, Aug 9, 2021 at 8:48 AM Marc Zyngier <maz@kernel.org> wrote:
> > >
> > > CNTPCTSS_EL0 and CNTVCTSS_EL0 are alternatives to the usual
> > > CNTPCT_EL0 and CNTVCT_EL0 that do not require a previous ISB
> > > to be synchronised (SS stands for Self-Synchronising).
> > >
> > > Use the ARM64_HAS_ECV capability to control alternative sequences
> > > that switch to these low(er)-cost primitives. Note that the
> > > counter access in the VDSO is for now left alone until we decide
> > > whether we want to allow this.
> >
> > What remains to be figured out before we add this to the vDSO (and
> > presumably advertise to userspace through some standard convention)?
>
> We need to understand what breaks if we runtime-patch the VDSO just
> like we do with the rest of the kernel. To start with, the debug
> version of the shared object is not the same as the object presented
> to the process. Maybe that's not a problem, but I would tend to err on
> the side of caution.

I would too, but there sadly are instances of Linux patching *user*
memory already (go look at how KVM/x86 handles the VMCALL/VMMCALL
instruction). But yes, I would much prefer the debug vDSO correspond
to the actual instructions.

> An alternative suggested by Ard was to have a separate function
> altogether for the counter access and an ifunc mapping to pick the
> right one.
>

Hmm, this does sound promising.

> > It would be nice to skip the trap handler altogether, unless there's a
> > can of worms lurking that I'm not aware of.
>
> The trap handlers are only there to work around errata. If you look at
> the arch timer code, you will notice that there is a bunch of SoCs and
> CPUs that do not have a reliable counter, and for which we have to
> trap the virtual counter accesses from userspace (as well as the
> VDSO).
>
> On sane platforms, userspace is free to use the virtual counter
> without any trap.

/facepalm I was about 2 cups of coffee short when writing this :) Thanks!

--
Oliver