diff mbox series

target/ppc: Implement ISA v3.1 wait variants

Message ID 20210517024651.2200837-1-npiggin@gmail.com
State New
Headers show
Series target/ppc: Implement ISA v3.1 wait variants | expand

Commit Message

Nicholas Piggin May 17, 2021, 2:46 a.m. UTC
ISA v3.1 adds new variations of wait, specified by the WC field. These
are not compatible with the wait 0 implementation, because they add
additional conditions that cause the processor to resume, which can
cause software to hang or run very slowly.

Add the new wait variants with a trivial no-op implementation, which is
allowed, as explained in comments: software must not depend on any
particular architected WC condition having caused resumption of
execution, therefore a no-op implementation is architecturally correct.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
Implementing cpu_relax() in Linux with wait 2,0 (pause_short) causes a
hang on boot without this patch.

 target/ppc/translate.c | 39 +++++++++++++++++++++++++++++++++------
 1 file changed, 33 insertions(+), 6 deletions(-)

Comments

David Gibson May 17, 2021, 5:39 a.m. UTC | #1
On Mon, May 17, 2021 at 12:46:51PM +1000, Nicholas Piggin wrote:
> ISA v3.1 adds new variations of wait, specified by the WC field. These
> are not compatible with the wait 0 implementation, because they add
> additional conditions that cause the processor to resume, which can
> cause software to hang or run very slowly.
> 
> Add the new wait variants with a trivial no-op implementation, which is
> allowed, as explained in comments: software must not depend on any
> particular architected WC condition having caused resumption of
> execution, therefore a no-op implementation is architecturally correct.
> 
> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>

Logic looks fine.  There is no test on the CPU's features or model
here, though, so this will change behaviour for pre-3.1 CPUs as well.

What would invoking these wait variants (presumably reserved) on
earlier CPUs do?

> ---
> Implementing cpu_relax() in Linux with wait 2,0 (pause_short) causes a
> hang on boot without this patch.
> 
>  target/ppc/translate.c | 39 +++++++++++++++++++++++++++++++++------
>  1 file changed, 33 insertions(+), 6 deletions(-)
> 
> diff --git a/target/ppc/translate.c b/target/ppc/translate.c
> index a6381208a5..80db450cab 100644
> --- a/target/ppc/translate.c
> +++ b/target/ppc/translate.c
> @@ -3619,12 +3619,39 @@ static void gen_sync(DisasContext *ctx)
>  /* wait */
>  static void gen_wait(DisasContext *ctx)
>  {
> -    TCGv_i32 t0 = tcg_const_i32(1);
> -    tcg_gen_st_i32(t0, cpu_env,
> -                   -offsetof(PowerPCCPU, env) + offsetof(CPUState, halted));
> -    tcg_temp_free_i32(t0);
> -    /* Stop translation, as the CPU is supposed to sleep from now */
> -    gen_exception_nip(ctx, EXCP_HLT, ctx->base.pc_next);
> +    uint32_t wc = (ctx->opcode >> 21) & 3;
> +
> +    /*
> +     * wait 0 waits for an exception to occur.
> +     */
> +    if (wc == 0) {
> +        TCGv_i32 t0 = tcg_const_i32(1);
> +        tcg_gen_st_i32(t0, cpu_env,
> +                       -offsetof(PowerPCCPU, env) + offsetof(CPUState, halted));
> +        tcg_temp_free_i32(t0);
> +        /* Stop translation, as the CPU is supposed to sleep from now */
> +        gen_exception_nip(ctx, EXCP_HLT, ctx->base.pc_next);
> +    }
> +
> +    /*
> +     * Other wait types must not wait until an exception occurs because
> +     * ignoring their other wake-up conditions could cause a hang.
> +     *
> +     * wait 1 (waitrsv) waits for an exception or a reservation to be lost.
> +     * This can happen for implementation specific reasons, so it can be
> +     * implemented as a no-op.
> +     *
> +     * wait 2 waits for an exception or an amount of time to pass. This is
> +     * implementation specific so it can be implemented as a no-op.
> +     *
> +     * wait 3 is reserved, so it may be implemented as a no-op.
> +     *
> +     * ISA v3.1 does allow for execution to resume "in the rare case of
> +     * an implementation-dependent event", so in any case software must
> +     * not depend on the architected resumption condition to become
> +     * true, so no-op implementations are architecturally correct (if
> +     * suboptimal).
> +     */
>  }
>  
>  #if defined(TARGET_PPC64)
Nicholas Piggin May 17, 2021, 7:19 a.m. UTC | #2
Excerpts from David Gibson's message of May 17, 2021 3:39 pm:
> On Mon, May 17, 2021 at 12:46:51PM +1000, Nicholas Piggin wrote:
>> ISA v3.1 adds new variations of wait, specified by the WC field. These
>> are not compatible with the wait 0 implementation, because they add
>> additional conditions that cause the processor to resume, which can
>> cause software to hang or run very slowly.
>> 
>> Add the new wait variants with a trivial no-op implementation, which is
>> allowed, as explained in comments: software must not depend on any
>> particular architected WC condition having caused resumption of
>> execution, therefore a no-op implementation is architecturally correct.
>> 
>> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> 
> Logic looks fine.  There is no test on the CPU's features or model
> here, though, so this will change behaviour for pre-3.1 CPUs as well.

Huh. 2.06-2.07 has very similar WC bits as 3.1, but 3.0 removed them
and made them reserved. I should have looked back but I'd assumed
they weren't there either.

Existing code treats WC != 0 as invalid on pre-3.0 processors AFAIKS,
so that's not quite right for 2.06-7 (they should look more like 3.1).

But before that it looks like it was just wait with no WC field.

> What would invoking these wait variants (presumably reserved) on
> earlier CPUs do?

Prior to 2.06, it looks like there is no WC field, and so they should 
generate a program check. So that just leaves the incorrect program
checks for 2.06-7, something like this should do it:

-GEN_HANDLER_E(wait, 0x1F, 0x1E, 0x00, 0x039FF801, PPC_NONE, PPC2_ISA300),
+GEN_HANDLER_E(wait, 0x1F, 0x1E, 0x00, 0x039FF801, PPC_NONE, PPC2_ISA206),

2.06-3.1 should all be fine with this patch, AFAIKS they all have words 
to the effect that WC != 0 is subject to implementation defined 
behaviour and may be treated as a no-op or not implemented.

Thanks,
Nick
David Gibson May 24, 2021, 4:49 a.m. UTC | #3
On Mon, May 17, 2021 at 05:19:06PM +1000, Nicholas Piggin wrote:
> Excerpts from David Gibson's message of May 17, 2021 3:39 pm:
> > On Mon, May 17, 2021 at 12:46:51PM +1000, Nicholas Piggin wrote:
> >> ISA v3.1 adds new variations of wait, specified by the WC field. These
> >> are not compatible with the wait 0 implementation, because they add
> >> additional conditions that cause the processor to resume, which can
> >> cause software to hang or run very slowly.
> >> 
> >> Add the new wait variants with a trivial no-op implementation, which is
> >> allowed, as explained in comments: software must not depend on any
> >> particular architected WC condition having caused resumption of
> >> execution, therefore a no-op implementation is architecturally correct.
> >> 
> >> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> > 
> > Logic looks fine.  There is no test on the CPU's features or model
> > here, though, so this will change behaviour for pre-3.1 CPUs as well.
> 
> Huh. 2.06-2.07 has very similar WC bits as 3.1, but 3.0 removed them
> and made them reserved. I should have looked back but I'd assumed
> they weren't there either.
> 
> Existing code treats WC != 0 as invalid on pre-3.0 processors AFAIKS,
> so that's not quite right for 2.06-7 (they should look more like 3.1).
> 
> But before that it looks like it was just wait with no WC field.
> 
> > What would invoking these wait variants (presumably reserved) on
> > earlier CPUs do?
> 
> Prior to 2.06, it looks like there is no WC field, and so they should 
> generate a program check. So that just leaves the incorrect program
> checks for 2.06-7, something like this should do it:
> 
> -GEN_HANDLER_E(wait, 0x1F, 0x1E, 0x00, 0x039FF801, PPC_NONE, PPC2_ISA300),
> +GEN_HANDLER_E(wait, 0x1F, 0x1E, 0x00, 0x039FF801, PPC_NONE, PPC2_ISA206),

Ok, can you update with such a change, and put some of this
explanation of the history in a comment.

> 2.06-3.1 should all be fine with this patch, AFAIKS they all have words 
> to the effect that WC != 0 is subject to implementation defined 
> behaviour and may be treated as a no-op or not implemented.

Ok.  Note that we do try to match specific CPU behaviour, not just the
architecture, although the architecture is obviously more important.
diff mbox series

Patch

diff --git a/target/ppc/translate.c b/target/ppc/translate.c
index a6381208a5..80db450cab 100644
--- a/target/ppc/translate.c
+++ b/target/ppc/translate.c
@@ -3619,12 +3619,39 @@  static void gen_sync(DisasContext *ctx)
 /* wait */
 static void gen_wait(DisasContext *ctx)
 {
-    TCGv_i32 t0 = tcg_const_i32(1);
-    tcg_gen_st_i32(t0, cpu_env,
-                   -offsetof(PowerPCCPU, env) + offsetof(CPUState, halted));
-    tcg_temp_free_i32(t0);
-    /* Stop translation, as the CPU is supposed to sleep from now */
-    gen_exception_nip(ctx, EXCP_HLT, ctx->base.pc_next);
+    uint32_t wc = (ctx->opcode >> 21) & 3;
+
+    /*
+     * wait 0 waits for an exception to occur.
+     */
+    if (wc == 0) {
+        TCGv_i32 t0 = tcg_const_i32(1);
+        tcg_gen_st_i32(t0, cpu_env,
+                       -offsetof(PowerPCCPU, env) + offsetof(CPUState, halted));
+        tcg_temp_free_i32(t0);
+        /* Stop translation, as the CPU is supposed to sleep from now */
+        gen_exception_nip(ctx, EXCP_HLT, ctx->base.pc_next);
+    }
+
+    /*
+     * Other wait types must not wait until an exception occurs because
+     * ignoring their other wake-up conditions could cause a hang.
+     *
+     * wait 1 (waitrsv) waits for an exception or a reservation to be lost.
+     * This can happen for implementation specific reasons, so it can be
+     * implemented as a no-op.
+     *
+     * wait 2 waits for an exception or an amount of time to pass. This is
+     * implementation specific so it can be implemented as a no-op.
+     *
+     * wait 3 is reserved, so it may be implemented as a no-op.
+     *
+     * ISA v3.1 does allow for execution to resume "in the rare case of
+     * an implementation-dependent event", so in any case software must
+     * not depend on the architected resumption condition to become
+     * true, so no-op implementations are architecturally correct (if
+     * suboptimal).
+     */
 }
 
 #if defined(TARGET_PPC64)