[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH v2 7/8] target-arm: A64: Create Instruction Synd
From: |
Edgar E. Iglesias |
Subject: |
Re: [Qemu-devel] [PATCH v2 7/8] target-arm: A64: Create Instruction Syndromes for Data Aborts |
Date: |
Fri, 29 Apr 2016 13:58:42 +0200 |
User-agent: |
Mutt/1.5.23 (2014-03-12) |
On Thu, Feb 25, 2016 at 06:26:45PM +0000, Peter Maydell wrote:
> On 19 February 2016 at 20:04, Edgar E. Iglesias
> <address@hidden> wrote:
> > From: "Edgar E. Iglesias" <address@hidden>
> >
> > Add support for generating the instruction syndrome for Data Aborts.
> > These syndromes are used by hypervisors for example to trap and emulate
> > memory accesses.
> >
> > We save the decoded data out-of-band with the TBs at translation time.
> > When exceptions hit, the extra data attached to the TB is used to
> > recreate the state needed to encode instruction syndromes.
> > This avoids the need to emit moves with every load/store.
> >
> > Based on a suggestion from Peter Maydell.
> >
> > Suggested-by: Peter Maydell <address@hidden>
> > Signed-off-by: Edgar E. Iglesias <address@hidden>
> > ---
> > target-arm/cpu.h | 2 +-
> > target-arm/op_helper.c | 47 +++++++++++++++++++++++-------
> > target-arm/translate-a64.c | 72
> > +++++++++++++++++++++++++++++++++++++++++++++-
> > target-arm/translate.c | 5 +++-
> > target-arm/translate.h | 14 +++++++++
> > 5 files changed, 127 insertions(+), 13 deletions(-)
> >
> > diff --git a/target-arm/cpu.h b/target-arm/cpu.h
> > index 1623821..c3ade8d 100644
> > --- a/target-arm/cpu.h
> > +++ b/target-arm/cpu.h
> > @@ -95,7 +95,7 @@
> > struct arm_boot_info;
> >
> > #define NB_MMU_MODES 7
> > -#define TARGET_INSN_START_EXTRA_WORDS 1
> > +#define TARGET_INSN_START_EXTRA_WORDS 2
>
> A comment saying what our two extra words are would probably be helpful.
Yes, I've added a comment.
>
> >
> > /* We currently assume float and double are IEEE single and double
> > precision respectively.
>
> > +static inline uint32_t merge_syn_data_abort(uint32_t template_syn,
> > + unsigned int target_el,
> > + bool same_el,
> > + bool s1ptw, int is_write,
> > + int fsc)
> > +{
> > + uint32_t syn;
> > +
> > + if (target_el != 2 || s1ptw) {
> > + /* ISV is only set for data aborts routed to EL2 and
> > + * never for stage-1 page table walks faulting on stage 2.
> > + *
> > + * See ARMv8 specs, D7-1974:
> > + * ISS encoding for an exception from a Data Abort, the
> > + * ISV field.
> > + */
> > + template_syn = 0;
>
> IL should be set if ISV is clear, but this template_syn value doesn't
> have that. It might be simpler to just
> return syn_data_abort_no_isv(stuff);
> here rather than using the OR-fields-together code path.
>
> + }
> + /* Fields: IL, ISV, SAS, SSE, SRT, SF and AR come from the template
> + * syndrome created at translation time.
> + * Now we create the runtime syndrome with the remaining fields.
> + */
> + syn = syn_data_abort(same_el,
> + 0, 0, 0, 0, 0, 0,
> + 0, 0, s1ptw, is_write == 1, fsc,
> + 0);
>
> Since you pass in 0 for isv here, the IL bit will always be set
> in syn and thus in the returned value, even if the template_syn value
> from translate time said it should be clear.
Thanks, I've reworked this code. Hopefully fixing this issues.
>
> + /* Merge runtime syndrome with the template syndrome. */
> + syn |= template_syn;
> + return syn;
> +}
>
> > diff --git a/target-arm/translate-a64.c b/target-arm/translate-a64.c
> > index 9e26d5e..5250c08 100644
> > --- a/target-arm/translate-a64.c
> > +++ b/target-arm/translate-a64.c
> > @@ -1803,6 +1803,24 @@ static void gen_store_exclusive(DisasContext *s, int
> > rd, int rt, int rt2,
> > }
> > #endif
> >
> > +static void disas_ldr_update_isyn_sse_sf(DisasContext *s, int size,
> > + bool is_signed, int opc)
> > +{
> > + int opc0 = extract32(opc, 0, 1);
> > + int regsize;
> > +
> > + s->isyn.dabt.sse = is_signed;
> > + /* Update the Sixty-Four bit (SF) registersize. This logic is derived
> > + * from the ARMv8 specs for LDR (Shared decode for all encodings).
> > + */
> > + if (is_signed) {
> > + regsize = opc0 ? 32 : 64;
> > + } else {
> > + regsize = size == 3 ? 64 : 32;
> > + }
> > + s->isyn.dabt.sf = regsize == 64;
> > +}
> > +
> > /* C3.3.6 Load/store exclusive
> > *
> > * 31 30 29 24 23 22 21 20 16 15 14 10 9 5 4 0
> > @@ -1859,6 +1877,12 @@ static void disas_ldst_excl(DisasContext *s,
> > uint32_t insn)
> > } else {
> > do_gpr_ld(s, tcg_rt, tcg_addr, size, false, false);
> > }
> > +
> > + /* Generate ISV for non-exclusive accesses including LASR. */
>
> The comment says "for non-exclusive accesses" but aren't we going to
> end up generating one for all accesses given where we are in the code ?
My understanding is that it's only for the case where is_excl gets set and
we end up calling gen_load/store_exclusive that we shouldn't create the ISS.
For other loads/stores, including load acquires, we should create one I think.
>
> > + s->isyn.dabt.valid = true;
> > + s->isyn.dabt.sse = false;
> > + s->isyn.dabt.sf = size == 3 ? 64 : 32;
> > + s->isyn.dabt.ar = is_lasr;
> > }
> > }
> >
> > @@ -1901,6 +1925,11 @@ static void disas_ld_lit(DisasContext *s, uint32_t
> > insn)
> > }
> > size = 2 + extract32(opc, 0, 1);
> > is_signed = extract32(opc, 1, 1);
> > +
> > + /* Enable ISV generation. */
> > + s->isyn.dabt.valid = true;
> > + s->isyn.dabt.sse = is_signed;
> > + s->isyn.dabt.sf = size == 3 ? 64 : 32;
>
> This code to update syndrome values ought to work, but it feels
> a little fragile in the face of future changes which might introduce
> code paths which add new guest load/store ops but don't set the syndrome
> info right.
>
> How would it look if we instead set the syndrome information in
> the lower level functions which directly call the tcg_gen_qemu_ld/st
> functions? You'd need to pass more arguments into those functions,
> of course, but I think that would make it easier to review and be
> confident that we set the syndrome correctly for all the cases that
> we need to.
I've massaged the code to be more explicit about ISS generation, moving
the setup closer to tcg_gen_qemu_ld/st as you suggest. It looks OK, I'll
include the new version in v3.
>
>
> > @@ -2685,6 +2737,23 @@ static void disas_ldst(DisasContext *s, uint32_t
> > insn)
> > unallocated_encoding(s);
> > break;
> > }
> > +
> > + /* The saved insn syndrome state in insn_start is already zero
> > (invalid).
> > + * We only update it if we have a valid insn syndrome to save.
> > + */
> > + if (s->isyn.dabt.valid) {
> > + isyn32 = syn_data_abort(0,
> > + s->isyn.dabt.valid,
> > + s->isyn.dabt.sas,
> > + s->isyn.dabt.sse,
> > + s->isyn.dabt.srt,
> > + s->isyn.dabt.sf,
> > + s->isyn.dabt.ar,
> > + 0, 0, 0, 0, 0, false);
> > + /* We don't use the lower 14 bits. */
> > + isyn32 >>= 14;
> > + tcg_set_insn_param(s->insn_start_idx, 2, isyn32);
> > + }
> > }
> >
> > /* C3.4.6 PC-rel. addressing
> > @@ -11087,7 +11156,8 @@ void gen_intermediate_code_a64(ARMCPU *cpu,
> > TranslationBlock *tb)
> > tcg_clear_temp_count();
> >
> > do {
> > - tcg_gen_insn_start(dc->pc, 0);
> > + dc->insn_start_idx = tcg_op_buf_count();
> > + tcg_gen_insn_start(dc->pc, 0, 0);
> > num_insns++;
> >
> > if (unlikely(!QTAILQ_EMPTY(&cs->breakpoints))) {
> > diff --git a/target-arm/translate.c b/target-arm/translate.c
> > index e69145d..7d83c94 100644
> > --- a/target-arm/translate.c
> > +++ b/target-arm/translate.c
> > @@ -11406,7 +11406,8 @@ void gen_intermediate_code(CPUARMState *env,
> > TranslationBlock *tb)
> > }
> > do {
> > tcg_gen_insn_start(dc->pc,
> > - (dc->condexec_cond << 4) | (dc->condexec_mask
> > >> 1));
> > + (dc->condexec_cond << 4) | (dc->condexec_mask
> > >> 1),
> > + 0);
> > num_insns++;
> >
> > #ifdef CONFIG_USER_ONLY
> > @@ -11722,8 +11723,10 @@ void restore_state_to_opc(CPUARMState *env,
> > TranslationBlock *tb,
> > if (is_a64(env)) {
> > env->pc = data[0];
> > env->condexec_bits = 0;
> > + env->exception.syndrome = data[2] << 14;
> > } else {
> > env->regs[15] = data[0];
> > env->condexec_bits = data[1];
> > + env->exception.syndrome = data[2] << 14;
>
> We should have a symbolic constant rather than this hardcoded 14
> (here and where we put the word into the insn stream).
Added one.
Cheers,
Edgar
>
> > }
> > }
> > diff --git a/target-arm/translate.h b/target-arm/translate.h
> > index 53ef971..e002c90 100644
> > --- a/target-arm/translate.h
> > +++ b/target-arm/translate.h
> > @@ -58,6 +58,20 @@ typedef struct DisasContext {
> > bool ss_same_el;
> > /* Bottom two bits of XScale c15_cpar coprocessor access control reg */
> > int c15_cpar;
> > + /* State needed to create the Data Abort template syndrome. */
> > + struct {
> > + /* Data Abort section. */
> > + struct {
> > + bool valid;
> > + unsigned int sas;
> > + bool sse;
> > + unsigned int srt;
> > + bool sf;
> > + bool ar;
> > + } dabt;
> > + } isyn;
> > + /* TCG op index of the current insn_start. */
> > + int insn_start_idx;
> > #define TMP_A64_MAX 16
> > int tmp_a64_count;
> > TCGv_i64 tmp_a64[TMP_A64_MAX];
> > --
> > 1.9.1
>
> thanks
> -- PMM
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- Re: [Qemu-devel] [PATCH v2 7/8] target-arm: A64: Create Instruction Syndromes for Data Aborts,
Edgar E. Iglesias <=