Re: [Qemu-devel] [PATCH] atomic failures on qemu-system-riscv64

[Alistair Francis]

On Tue, Sep 24, 2019 at 1:04 PM Palmer Dabbelt <address@hidden> wrote:
>
> On Tue, 24 Sep 2019 11:29:25 PDT (-0700), address@hidden wrote:
> > On Mon, Jun 24, 2019 at 11:21 AM Joel Sing <address@hidden> wrote:
> >>
> >> On 19-06-17 16:52:44, Richard Henderson wrote:
> >> > On 6/16/19 12:19 PM, Joel Sing wrote:
> >> > > +    /*
> >> > > +     * Clear the load reservation, since an SC must fail if there is
> >> > > +     * an SC to any address, in between an LR and SC pair.
> >> > > +     */
> >> > > +    tcg_gen_movi_tl(load_res, 0);
> >> > > +
> >> > >      gen_set_label(l2);
> >> >
> >> > This clear needs to be moved down below label l2.
> >> > Otherwise, with lr / sc / sc, the second sc could succeed in error.
> >>
> >> Indeed, thanks.
> >>
> >> > FWIW, other targets have used -1 as the "invalid" load reservation, 
> >> > since the
> >> > architecture does not require address 0 to be unmapped.  This should be 
> >> > quite
> >> > visible in M-mode with paging disabled and ram at offset 0.  Often, other
> >> > targets require alignment for the lr/sc address, though I don't see that 
> >> > for riscv.
> >>
> >> I've switched to -1 as suggested. Regarding the alignment for 
> >> reservations, the
> >> specification does require this, although I do not recall seeing any 
> >> enforcement
> >> of this by qemu itself.
> >>
> >> New diff follows.
> >>
> >> From 8ef31a2ce8ef1cbeee92995a0b2994f480e9bb6d Mon Sep 17 00:00:00 2001
> >> From: Joel Sing <address@hidden>
> >> Date: Tue, 25 Jun 2019 02:44:24 +1000
> >> Subject: [PATCH] Clear load reservations on qemu riscv target
> >>
> >> This prevents a load reservation from being placed in one context/process,
> >> then being used in another, resulting in an SC succeeding incorrectly and
> >> breaking atomics.
> >>
> >> Signed-off-by: Joel Sing <address@hidden>
> >> ---
> >>  target/riscv/cpu.c                      | 1 +
> >>  target/riscv/cpu_helper.c               | 9 +++++++++
> >>  target/riscv/insn_trans/trans_rva.inc.c | 8 +++++++-
> >>  3 files changed, 17 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/target/riscv/cpu.c b/target/riscv/cpu.c
> >> index d61bce6d55..e7c8bf48fc 100644
> >> --- a/target/riscv/cpu.c
> >> +++ b/target/riscv/cpu.c
> >> @@ -281,6 +281,7 @@ static void riscv_cpu_reset(CPUState *cs)
> >>      env->pc = env->resetvec;
> >>  #endif
> >>      cs->exception_index = EXCP_NONE;
> >> +    env->load_res = -1;
> >>      set_default_nan_mode(1, &env->fp_status);
> >>  }
> >>
> >> diff --git a/target/riscv/cpu_helper.c b/target/riscv/cpu_helper.c
> >> index b17f169681..6a07b12e65 100644
> >> --- a/target/riscv/cpu_helper.c
> >> +++ b/target/riscv/cpu_helper.c
> >> @@ -113,6 +113,15 @@ void riscv_cpu_set_mode(CPURISCVState *env, 
> >> target_ulong newpriv)
> >>      }
> >>      /* tlb_flush is unnecessary as mode is contained in mmu_idx */
> >>      env->priv = newpriv;
> >> +
> >> +    /* Clear the load reservation - otherwise a reservation placed in one
> >> +     * context/process can be used by another, resulting in an SC 
> >> succeeding
> >> +     * incorrectly. Version 2.2 of the ISA specification explicitly 
> >> requires
> >> +     * this behaviour, while later revisions say that the kernel "should" 
> >> use
> >> +     * an SC instruction to force the yielding of a load reservation on a
> >> +     * preemptive context switch. As a result, do both.
> >> +     */
> >> +    env->load_res = -1;
> >>  }
> >>
> >>  /* get_physical_address - get the physical address for this virtual 
> >> address
> >> diff --git a/target/riscv/insn_trans/trans_rva.inc.c 
> >> b/target/riscv/insn_trans/trans_rva.inc.c
> >> index f6dbbc065e..fadd88849e 100644
> >> --- a/target/riscv/insn_trans/trans_rva.inc.c
> >> +++ b/target/riscv/insn_trans/trans_rva.inc.c
> >> @@ -61,7 +61,7 @@ static inline bool gen_sc(DisasContext *ctx, arg_atomic 
> >> *a, TCGMemOp mop)
> >>
> >>      gen_set_label(l1);
> >>      /*
> >> -     * Address comparion failure.  However, we still need to
> >> +     * Address comparison failure.  However, we still need to
> >>       * provide the memory barrier implied by AQ/RL.
> >>       */
> >>      tcg_gen_mb(TCG_MO_ALL + a->aq * TCG_BAR_LDAQ + a->rl * TCG_BAR_STRL);
> >> @@ -69,6 +69,12 @@ static inline bool gen_sc(DisasContext *ctx, arg_atomic 
> >> *a, TCGMemOp mop)
> >>      gen_set_gpr(a->rd, dat);
> >>
> >>      gen_set_label(l2);
> >> +    /*
> >> +     * Clear the load reservation, since an SC must fail if there is
> >> +     * an SC to any address, in between an LR and SC pair.
> >> +     */
> >> +    tcg_gen_movi_tl(load_res, -1);
> >> +
> >>      tcg_temp_free(dat);
> >>      tcg_temp_free(src1);
> >>      tcg_temp_free(src2);
> >> --
> >
> > This patch causes boot failures when booting systemd built with musl on 
> > RV64.
> >
> > It could possibly be a musl bug, but I wanted to throw that out here
> > first to see what people think.
>
> Looking at the musl port, I see at least one bug in their atomics jumping out
> at me:
>
> diff --git a/arch/riscv64/atomic_arch.h b/arch/riscv64/atomic_arch.h
> index c9765342..41ad4d04 100644
> --- a/arch/riscv64/atomic_arch.h
> +++ b/arch/riscv64/atomic_arch.h
> @@ -14,7 +14,7 @@ static inline int a_cas(volatile int *p, int t, int s)
>                 "       sc.w.aqrl %1, %4, (%2)\n"
>                 "       bnez %1, 1b\n"
>                 "1:"
> -               : "=&r"(old), "=r"(tmp)
> +               : "=&r"(old), "=&r"(tmp)
>                 : "r"(p), "r"(t), "r"(s)
>                 : "memory");
>         return old;
> @@ -31,7 +31,7 @@ static inline void *a_cas_p(volatile void *p, void *t, void 
> *s)
>                 "       sc.d.aqrl %1, %4, (%2)\n"
>                 "       bnez %1, 1b\n"
>                 "1:"
> -               : "=&r"(old), "=r"(tmp)
> +               : "=&r"(old), "=&r"(tmp)
>                 : "r"(p), "r"(t), "r"(s)
>                 : "memory");
>         return old;
>
> It's a shot in the dark as to whether that'll fix your bug, but I could at
> least see a mechanism for it: before we yielded load reservations on context
> switches then that backwards branch would never be taken, so we wouldn't 
> notice
> if tmp was allocated into one of the same registers as the inputs.  Even if it
> doesn't fix your issue it's still a bug so I'll send the patch out, just LMK 
> so
> I can indicate how important the issue is.

I haven't had a chance to test this fix yet. The bug was reported by
Khem (and other OE people) as it's break musl for RISC-V.

>
> This should manifest on hardware, but it looks like we managed to drop that SC
> patch.  I'll go send the patch out now...

Thanks, do you mind CCing me?

Alistair

>
> >
> > Alistair
> >
> >> 2.21.0
> >>
> >>