[Qemu-devel] [PATCH] Improve cmpxchg emulation.

[Qemu-devel] [PATCH] Improve cmpxchg emulation.

[Jonathan A. Kollasch]

Change the accumulator only after performing the redundant write during
cmpxchg.  This fixes pthreaded programs using fork() in NetBSD/i386
guests.

>From Andreas Gustafsson in https://bugs.launchpad.net/qemu/+bug/569760.

Signed-off-by: Jonathan A. Kollasch <jakllsch@kollasch.net>
---
 target-i386/translate.c |   12 ++++++++----
 1 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/target-i386/translate.c b/target-i386/translate.c
index 7b6e3c2..391830f 100644
--- a/target-i386/translate.c
+++ b/target-i386/translate.c
@@ -4879,20 +4879,24 @@ static target_ulong disas_insn(DisasContext *s, target_ulong pc_start)
             tcg_gen_sub_tl(t2, cpu_regs[R_EAX], t0);
             gen_extu(ot, t2);
             tcg_gen_brcondi_tl(TCG_COND_EQ, t2, 0, label1);
+	    label2 = gen_new_label();
             if (mod == 3) {
-                label2 = gen_new_label();
                 gen_op_mov_reg_v(ot, R_EAX, t0);
                 tcg_gen_br(label2);
                 gen_set_label(label1);
                 gen_op_mov_reg_v(ot, rm, t1);
-                gen_set_label(label2);
             } else {
-                tcg_gen_mov_tl(t1, t0);
+                /* perform no-op store cycle like physical cpu; must be
+		   before changing accumulator to ensure idempotency if
+		   the store faults and the instruction is restarted
+		*/
+                gen_op_st_v(ot + s->mem_index, t0, a0);
                 gen_op_mov_reg_v(ot, R_EAX, t0);
+                tcg_gen_br(label2);
                 gen_set_label(label1);
-                /* always store */
                 gen_op_st_v(ot + s->mem_index, t1, a0);
             }
+	    gen_set_label(label2);
             tcg_gen_mov_tl(cpu_cc_src, t0);
             tcg_gen_mov_tl(cpu_cc_dst, t2);
             s->cc_op = CC_OP_SUBB + ot;
-- 
1.6.6.2

Re: [Qemu-devel] [PATCH] Improve cmpxchg emulation.

[malc]

On Wed, 8 Sep 2010, Jonathan A. Kollasch wrote:

> Change the accumulator only after performing the redundant write during
> cmpxchg.  This fixes pthreaded programs using fork() in NetBSD/i386
> guests.
> 
> >From Andreas Gustafsson in https://bugs.launchpad.net/qemu/+bug/569760.

This is tab damaged.

Secondly it looks as if this addresses only a small part of a general
problem [1], also in a very naive and inefficient way, while also opening
a hole can of worms (should real parallel execution for TCG be ever
implemented)

[1] http://www.mail-archive.com/qemu-devel@nongnu.org/msg40000.html

> 
> Signed-off-by: Jonathan A. Kollasch <jakllsch@kollasch.net>
> ---
>  target-i386/translate.c |   12 ++++++++----
>  1 files changed, 8 insertions(+), 4 deletions(-)
> 
> diff --git a/target-i386/translate.c b/target-i386/translate.c
> index 7b6e3c2..391830f 100644
> --- a/target-i386/translate.c
> +++ b/target-i386/translate.c
> @@ -4879,20 +4879,24 @@ static target_ulong disas_insn(DisasContext *s, target_ulong pc_start)
>              tcg_gen_sub_tl(t2, cpu_regs[R_EAX], t0);
>              gen_extu(ot, t2);
>              tcg_gen_brcondi_tl(TCG_COND_EQ, t2, 0, label1);
> +	    label2 = gen_new_label();
>              if (mod == 3) {
> -                label2 = gen_new_label();
>                  gen_op_mov_reg_v(ot, R_EAX, t0);
>                  tcg_gen_br(label2);
>                  gen_set_label(label1);
>                  gen_op_mov_reg_v(ot, rm, t1);
> -                gen_set_label(label2);
>              } else {
> -                tcg_gen_mov_tl(t1, t0);
> +                /* perform no-op store cycle like physical cpu; must be
> +		   before changing accumulator to ensure idempotency if
> +		   the store faults and the instruction is restarted
> +		*/
> +                gen_op_st_v(ot + s->mem_index, t0, a0);
>                  gen_op_mov_reg_v(ot, R_EAX, t0);
> +                tcg_gen_br(label2);
>                  gen_set_label(label1);
> -                /* always store */
>                  gen_op_st_v(ot + s->mem_index, t1, a0);
>              }
> +	    gen_set_label(label2);
>              tcg_gen_mov_tl(cpu_cc_src, t0);
>              tcg_gen_mov_tl(cpu_cc_dst, t2);
>              s->cc_op = CC_OP_SUBB + ot;
> 

-- 
mailto:av1474@comtv.ru