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

[Joel Sing <joel@sing.id.au>]

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 <joel@sing.id.au>
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 <joel@sing.id.au>
---
 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);
-- 
2.21.0