Re: [PATCH v3 1/3] target/riscv: Fix tb->flags FS status

From: Alistair Francis <alistair23@gmail.com>
To: Richard Henderson <richard.henderson@linaro.org>
Cc: "open list:RISC-V" <qemu-riscv@nongnu.org>,
	Sagar Karandikar <sagark@eecs.berkeley.edu>,
	Bastian Koppelmann <kbastian@mail.uni-paderborn.de>,
	"qemu-devel@nongnu.org Developers" <qemu-devel@nongnu.org>,
	Alistair Francis <Alistair.Francis@wdc.com>,
	Palmer Dabbelt <palmer@dabbelt.com>,
	shihpo.hung@sifive.com
Subject: Re: [PATCH v3 1/3] target/riscv: Fix tb->flags FS status
Date: Thu, 16 Jan 2020 09:06:46 +1000	[thread overview]
Message-ID: <CAKmqyKPQQPJsTiBmRtQ7hZOk0Ly62qKSmmFY_m8G2S9Pw20k5Q@mail.gmail.com> (raw)
In-Reply-To: <8d63e90b-0779-6432-3a35-1b759f5ae279@linaro.org>

On Thu, Jan 16, 2020 at 7:46 AM Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 1/14/20 8:28 PM, Alistair Francis wrote:
> > On Wed, Jan 15, 2020 at 4:18 PM <shihpo.hung@sifive.com> wrote:
> >>
> >> It was found that running libquantum on riscv-linux qemu produced an
> >> incorrect result. After investigation, FP registers are not saved
> >> during context switch due to incorrect mstatus.FS.
> >>
> >> In current implementation tb->flags merges all non-disabled state to
> >> dirty. This means the code in mark_fs_dirty in translate.c that
> >> handles initial and clean states is unreachable.
> >>
> >> This patch fixes it and is successfully tested with:
> >>   libquantum
> >>
> >> Thanks to Richard for pointing out the actual bug.
> >>
> >> v3: remove the redundant condition
> >> v2: root cause FS problem
> >>
> >> Suggested-by: Richard Henderson <richard.henderson@linaro.org>
> >> Signed-off-by: ShihPo Hung <shihpo.hung@sifive.com>
> >> Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
> >> ---
> >>  target/riscv/cpu.h | 5 +----
> >>  1 file changed, 1 insertion(+), 4 deletions(-)
> >>
> >> diff --git a/target/riscv/cpu.h b/target/riscv/cpu.h
> >> index e59343e..de0a8d8 100644
> >> --- a/target/riscv/cpu.h
> >> +++ b/target/riscv/cpu.h
> >> @@ -293,10 +293,7 @@ static inline void cpu_get_tb_cpu_state(CPURISCVState *env, target_ulong *pc,
> >>  #ifdef CONFIG_USER_ONLY
> >>      *flags = TB_FLAGS_MSTATUS_FS;
> >>  #else
> >> -    *flags = cpu_mmu_index(env, 0);
> >> -    if (riscv_cpu_fp_enabled(env)) {
> >> -        *flags |= TB_FLAGS_MSTATUS_FS;
> >> -    }
> >> +    *flags = cpu_mmu_index(env, 0) | (env->mstatus & MSTATUS_FS);
> >
> > I don't think this is right, you should use the riscv_cpu_fp_enabled() function.
> >
> > Right now it's the same as env->mstatus & MSTATUS_FS but when the
> > Hypervisor extension goes in riscv_cpu_fp_enabled() will be more
> > complex.
>
> Hmm.  Are you sure something like
>
>   flags |= riscv_cpu_effective_mstatus(env) & MSTATUS_FS;
>
> wouldn't be more appropriate for the hypervisor extension?

I was more thinking:

    if (riscv_cpu_fp_enabled(env)) {
        *flags |= env->mstatus & MSTATUS_FS;
    }

as floating point can be disabled from multiple places when we have
the H extension.

>
> I guess I should have another browse through your hv patchset, but I worry now
> about bare uses of env->mstatus, if they no longer mean what they appear to mean.

That was why this was all refacted in the first place as we now need
to check against env->vsstatus as well (depending on virt status).

Alistair

>
>
> r~