Re: [PATCH v17 15/16] accel/tcg: Add info [tb-list|tb] commands to HMP

[Alex Bennée]

Richard Henderson <richard.henderson@linaro.org> writes:

> From: "Vanderson M. do Rosario" <vandersonmr2@gmail.com>
>
> These commands allow the exploration of TBs generated by the TCG.
> Understand which one hotter, with more guest/host instructions,
> and examine the guest code.
>
> The goal of this command is to allow the dynamic exploration of
> TCG behavior and code quality. Therefore, for now, a corresponding
> QMP command is not worthwhile.
>
> Example of output:
>
> ------------------------------
>
> TB id:0 | phys:0xa21f562e virt:0x0000000000000000 flags:0x00028010 0 inv/1
>         | exec:6171503732/0 guest inst cov:94.77%
>         | trans:1 ints: g:8 op:28 op_opt:24 spills:0
>         | h/g (host bytes / guest insts): 37.000000
>
> 0xa21f562e:  00002797      auipc         a5,8192           # 0xa21f762e
> 0xa21f5632:  a2278793      addi          a5,a5,-1502
> 0xa21f5636:  639c          ld            a5,0(a5)
> 0xa21f5638:  00178713      addi          a4,a5,1
> 0xa21f563c:  00002797      auipc         a5,8192           # 0xa21f763c
> 0xa21f5640:  a1478793      addi          a5,a5,-1516
> 0xa21f5644:  e398          sd            a4,0(a5)
> 0xa21f5646:  b7e5          j             -24               # 0xa21f562e
>
> Signed-off-by: Vanderson M. do Rosario <vandersonmr2@gmail.com>
> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
> Signed-off-by: Fei Wu <fei2.wu@intel.com>
> [rth: Split out of a larger patch]
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  accel/tcg/tb-context.h |  2 +
>  accel/tcg/monitor.c    | 91 ++++++++++++++++++++++++++++++++++++++++++
>  accel/tcg/tb-stats.c   |  2 +
>  hmp-commands-info.hx   | 14 +++++++
>  4 files changed, 109 insertions(+)
>
> diff --git a/accel/tcg/tb-context.h b/accel/tcg/tb-context.h
> index 4b1abe392b..29d87200b6 100644
> --- a/accel/tcg/tb-context.h
> +++ b/accel/tcg/tb-context.h
> @@ -35,6 +35,8 @@ struct TBContext {
>      /* statistics */
>      unsigned tb_flush_count;
>      unsigned tb_phys_invalidate_count;
> +
> +    GPtrArray *last_search;
>  };
>  
>  extern TBContext tb_ctx;
> diff --git a/accel/tcg/monitor.c b/accel/tcg/monitor.c
> index 370fea883c..1be3218715 100644
> --- a/accel/tcg/monitor.c
> +++ b/accel/tcg/monitor.c
> @@ -15,12 +15,14 @@
>  #include "qapi/qmp/qdict.h"
>  #include "monitor/monitor.h"
>  #include "monitor/hmp.h"
> +#include "monitor/hmp-target.h"
>  #include "sysemu/cpus.h"
>  #include "sysemu/cpu-timers.h"
>  #include "sysemu/tcg.h"
>  #include "tcg/tcg.h"
>  #include "tcg/tb-stats.h"
>  #include "exec/tb-flush.h"
> +#include "disas/disas.h"
>  #include "internal-common.h"
>  #include "tb-context.h"
>  
> @@ -303,10 +305,99 @@ static void hmp_tbstats(Monitor *mon, const QDict 
> *qdict)
>                            RUN_ON_CPU_HOST_INT(flags));
>  }
>  
> +static void hmp_info_tblist(Monitor *mon, const QDict *qdict)
> +{
> +    int max;
> +    const char *sortedby_str;
> +    GCompareFunc sort;
> +    GPtrArray *array;
> +
> +    if (!tcg_enabled()) {
> +        monitor_printf(mon, "Only available with accel=tcg\n");
> +        return;
> +    }
> +    if (!tb_stats_enabled) {
> +        monitor_printf(mon, "TB statistics not being recorded\n");
> +        return;
> +    }
> +
> +    max = qdict_get_try_int(qdict, "number", 10);
> +    sortedby_str = qdict_get_try_str(qdict, "sortedby");
> +
> +    if (sortedby_str == NULL || g_str_equal(sortedby_str, "hotness")) {
> +        sort = tb_stats_sort_by_coverage;
> +    } else if (g_str_equal(sortedby_str, "hg")) {
> +        sort = tb_stats_sort_by_hg;
> +    } else if (g_str_equal(sortedby_str, "spills")) {
> +        sort = tb_stats_sort_by_spills;
> +    } else {
> +        monitor_printf(mon, "Sort options are: hotness, hg, spills\n");
> +        return;
> +    }
> +
> +    g_ptr_array_unref(tb_ctx.last_search);
> +    tb_ctx.last_search = NULL;
> +
> +    array = tb_stats_collect(max, sort);
> +    max = array->len;
> +    if (max == 0) {
> +        monitor_printf(mon, "No TB statistics collected\n");
> +        g_ptr_array_free(array, true);
> +        return;
> +    }
> +
> +    for (int i = 0; i < max; ++i) {
> +        TBStatistics *s = g_ptr_array_index(array, i);
> +        g_autoptr(GString) buf = tb_stats_dump(s, i);
> +        monitor_puts(mon, buf->str);
> +    }
> +
> +    /* Remember for the next "info tb" */
> +    tb_ctx.last_search = array;
> +}
> +
> +static void hmp_info_tb(Monitor *mon, const QDict *qdict)
> +{
> +    GPtrArray *array;
> +    int id;
> +
> +    if (!tcg_enabled()) {
> +        monitor_printf(mon, "Only available with accel=tcg\n");
> +        return;
> +    }
> +
> +    array = g_ptr_array_ref(tb_ctx.last_search);
> +    if (!array) {
> +        monitor_printf(mon, "No TB statistics collected\n");
> +        return;
> +    }
> +
> +    id = qdict_get_int(qdict, "id");
> +    if (id < array->len) {
> +        TBStatistics *s = g_ptr_array_index(array, id);
> +        g_autoptr(GString) buf = tb_stats_dump(s, id);
> +        monitor_puts(mon, buf->str);
> +
> +        for (int i = s->tbs->len - 1; i >= 0; --i) {
> +            TranslationBlock *tb = g_ptr_array_index(s->tbs, i);
> +            if (!(tb->cflags & CF_INVALID)) {
> +                monitor_disas(mon, mon_get_cpu(mon), s->phys_pc,
> +                              tb->icount, MON_DISAS_GRA);
> +            }

I'm confused by the state I've got to:

  (qemu) info tb 0
  TB id:0 | phys:0x1f59d5918 virt=0 flags:0x00000051 invalid:0/1
          | exec:57340981/0 coverage:20.73%
          | trans:1 inst: g:1 op:19 op_opt:18 spills:0
          | h/g (host bytes / guest insts): 136.000000
  0x1f59d5918:  35000354  cbnz     w20, #0x1f59d5980
  (qemu) xp/5i 0x1f59d5918
  0x1f59d5918:  00000000  .byte    0x00, 0x00, 0x00, 0x00
  0x1f59d591c:  00000000  .byte    0x00, 0x00, 0x00, 0x00
  0x1f59d5920:  00000000  .byte    0x00, 0x00, 0x00, 0x00
  0x1f59d5924:  00000000  .byte    0x00, 0x00, 0x00, 0x00
  0x1f59d5928:  00000000  .byte    0x00, 0x00, 0x00, 0x00

It seems this is the kernels busy loop (so I assume resident in memory)
but trying to dump the instructions directly it fails. I assume the
physical memory address in each case is the same right?

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro