Re: [Qemu-devel] [PATCH v4 6/7] monitor: adding tb_stats hmp command

["Alex Bennée" <alex.bennee@linaro.org>]

vandersonmr <vandersonmr2@gmail.com> writes:

> Adding tb_stats [start|pause|stop|filter] command to hmp.
> This allows controlling the collection of statistics.
> It is also possible to set the level of collection:
> all, jit, or exec.
>
> The goal of this command is to allow the dynamic exploration
> of the TCG behavior and quality. Therefore, for now, a
> corresponding QMP command is not worthwhile.
>
> Signed-off-by: Vanderson M. do Rosario <vandersonmr2@gmail.com>
> ---
>  accel/tcg/tb-stats.c    | 107 ++++++++++++++++++++++++++++++++++++++++
>  hmp-commands.hx         |  17 +++++++
>  include/exec/tb-stats.h |  13 +++++
>  include/qemu/log.h      |   1 +
>  monitor/misc.c          |  40 +++++++++++++++
>  5 files changed, 178 insertions(+)
>
> diff --git a/accel/tcg/tb-stats.c b/accel/tcg/tb-stats.c
> index 44497d4f9b..6c330e1b02 100644
> --- a/accel/tcg/tb-stats.c
> +++ b/accel/tcg/tb-stats.c
> @@ -6,6 +6,9 @@
>
>  #include "qemu/qemu-print.h"
>
> +/* only accessed in safe work */
> +static GList *last_search;
> +
>  struct jit_profile_info {
>      uint64_t translations;
>      uint64_t aborted;
> @@ -104,4 +107,108 @@ void dump_jit_profile_info(TCGProfile *s)
>      }
>  }
>
> +static void dessaloc_tbstats(void *p, uint32_t hash, void *userp)
> +{
> +    g_free(p);
> +}

surely free_tbstats would be a better name?

> +
> +void clean_tbstats(void)
> +{
> +    /* remove all tb_stats */
> +    qht_iter(&tb_ctx.tb_stats, dessaloc_tbstats, NULL);
> +    qht_destroy(&tb_ctx.tb_stats);
> +}
> +
> +static void reset_tbstats_flag(void *p, uint32_t hash, void *userp)
> +{
> +    uint32_t flag = *((int *)userp);
> +    TBStatistics *tbs = p;
> +    tbs->stats_enabled = flag;
> +}
> +
> +void set_tbstats_flags(uint32_t flag)
> +{
> +    /* iterate over tbstats setting their flag as TB_NOTHING */
> +    qht_iter(&tb_ctx.tb_stats, reset_tbstats_flag, &flag);
> +}
> +
> +void do_hmp_tbstats_safe(CPUState *cpu, run_on_cpu_data icmd)
> +{
> +    struct TbstatsCommand *cmdinfo = icmd.host_ptr;
> +    int cmd = cmdinfo->cmd;
> +    uint32_t level = cmdinfo->level;
> +
> +    /* for safe, always pause TB generation while running this commands */
> +    mmap_lock();

This doesn't serialise in all cases - in fact it's running as safe work
which is our serial case. The mmap_lock is really only needed for code
generation in linux-user. I think the lock/unlock can be dropped from
these operations.

> +
> +    switch (cmd) {
> +    case START:
> +        if (tb_stats_collection_paused()) {
> +            set_tbstats_flags(level);
> +        } else {
> +            if (tb_stats_collection_enabled()) {
> +                qemu_printf("TB information already being recorded");
> +                return;
> +            }
> +            qht_init(&tb_ctx.tb_stats, tb_stats_cmp, CODE_GEN_HTABLE_SIZE,
> +                        QHT_MODE_AUTO_RESIZE);

I'm wary of initialising this in multiple places. Maybe move this to
enable_collect_tb_stats and drop the initialisation in tb_htable_init.

> +        }
> +
> +        set_default_tbstats_flag(level);
> +        enable_collect_tb_stats();
> +        tb_flush(cpu);
> +        break;
> +    case PAUSE:
> +        if (!tb_stats_collection_enabled()) {
> +            qemu_printf("TB information not being recorded");
> +            return;
> +        }
> +
> +        /* Continue to create TBStatistic structures but stop collecting statistics */
> +        pause_collect_tb_stats();
> +        tb_flush(cpu);
> +        set_default_tbstats_flag(TB_NOTHING);
> +        set_tbstats_flags(TB_PAUSED);

Minor style suggestion - make the tb_flush the last thing you do once
you have moved things around. You are in safe work so nothing should be
happening anyway and it reads better.

> +        break;
> +    case STOP:
> +        if (!tb_stats_collection_enabled()) {
> +            qemu_printf("TB information not being recorded");
> +            return;
> +        }
> +
> +        /* Dissalloc all TBStatistics structures and stop creating new ones */
> +        disable_collect_tb_stats();
> +        tb_flush(cpu);
> +        clean_tbstats();

As above.

> +        break;
> +    case FILTER:
> +        if (!tb_stats_collection_enabled()) {
> +            qemu_printf("TB information not being recorded");
> +            return;
> +        }
> +        if (!last_search) {
> +            qemu_printf("no search on record! execute info tbs before filtering!");
> +            return;
> +        }
> +
> +        tb_flush(cpu);
> +        set_default_tbstats_flag(TB_NOTHING);
> +
> +        /* Set all tbstats as paused, then return only the ones from last_search */
> +        pause_collect_tb_stats();
> +        set_tbstats_flags(TB_PAUSED);
> +
> +        for (GList *iter = last_search; iter; iter = g_list_next(iter)) {
> +            TBStatistics *tbs = iter->data;
> +            tbs->stats_enabled = level;
> +        }

If we are only interested in tracking the N hotestest at this point we
want to prevent allocation of new TBStats and free all the unused ones?

> +
> +        break;
> +    default: /* INVALID */

g_assert_not_reached()? This would be an internal bug if we ever sent
the wrong command?

> +        break;
> +    }
> +
> +    mmap_unlock();
> +    g_free(cmdinfo);
> +}
>
> diff --git a/hmp-commands.hx b/hmp-commands.hx
> index bfa5681dd2..419898751e 100644
> --- a/hmp-commands.hx
> +++ b/hmp-commands.hx
> @@ -1885,6 +1885,23 @@ STEXI
>  @findex qemu-io
>  Executes a qemu-io command on the given block device.
>
> +ETEXI
> +#if defined(CONFIG_TCG)
> +    {
> +        .name       = "tb_stats",
> +        .args_type  = "command:s,level:s?",
> +        .params     = "command [stats_level]",
> +        .help       = "Control tb statistics collection:"
> +                        "tb_stats (start|pause|stop|filter)
> [all|jit_stats|exec_stats]",

What does filter do?

> +        .cmd        = hmp_tbstats,
> +    },
> +#endif
> +
> +STEXI
> +@item tb_stats
> +@findex
> +Control recording tb statistics
> +
>  ETEXI
>
>      {
> diff --git a/include/exec/tb-stats.h b/include/exec/tb-stats.h
> index 7d775f2c0d..d1debd3262 100644
> --- a/include/exec/tb-stats.h
> +++ b/include/exec/tb-stats.h
> @@ -6,6 +6,9 @@
>  #include "exec/tb-context.h"
>  #include "tcg.h"
>
> +enum SortBy { SORT_BY_HOTNESS, SORT_BY_HG /* Host/Guest */, SORT_BY_SPILLS };
> +enum TbstatsCmd { START, PAUSE, STOP, FILTER };
> +
>  #define tb_stats_enabled(tb, JIT_STATS) \
>      (tb && tb->tb_stats && (tb->tb_stats->stats_enabled & JIT_STATS))
>
> @@ -60,4 +63,14 @@ bool tb_stats_cmp(const void *ap, const void *bp);
>
>  void dump_jit_profile_info(TCGProfile *s);
>
> +void set_tbstats_flags(uint32_t flags);
> +void clean_tbstats(void);
> +
> +struct TbstatsCommand {
> +    enum TbstatsCmd cmd;
> +    uint32_t level;
> +};
> +
> +void do_hmp_tbstats_safe(CPUState *cpu, run_on_cpu_data icmd);
> +
>  #endif
> diff --git a/include/qemu/log.h b/include/qemu/log.h
> index 93642603e5..47071d72c7 100644
> --- a/include/qemu/log.h
> +++ b/include/qemu/log.h
> @@ -133,6 +133,7 @@ void qemu_log_close(void);
>  #define TB_NOTHING    0
>  #define TB_EXEC_STATS (1 << 1)
>  #define TB_JIT_STATS  (1 << 2)
> +#define TB_PAUSED     (1 << 3)
>
>  void enable_collect_tb_stats(void);
>  void disable_collect_tb_stats(void);
> diff --git a/monitor/misc.c b/monitor/misc.c
> index 00338c002a..aea9b0db4f 100644
> --- a/monitor/misc.c
> +++ b/monitor/misc.c
> @@ -459,6 +459,46 @@ static void hmp_info_registers(Monitor *mon, const QDict *qdict)
>  }
>
>  #ifdef CONFIG_TCG
> +static void hmp_tbstats(Monitor *mon, const QDict *qdict)
> +{
> +    if (!tcg_enabled()) {
> +        error_report("TB information is only available with accel=tcg");
> +        return;
> +    }
> +
> +    char *cmd = (char *) qdict_get_try_str(qdict, "command");
> +    enum TbstatsCmd icmd = -1;
> +
> +    if (strcmp(cmd, "start") == 0) {
> +        icmd = START;
> +    } else if (strcmp(cmd, "pause") == 0) {
> +        icmd = PAUSE;
> +    } else if (strcmp(cmd, "stop") == 0) {
> +        icmd = STOP;
> +    } else if (strcmp(cmd, "filter") == 0) {
> +        icmd = FILTER;
> +    } else {
> +        error_report("invalid command!");
> +    }
> +
> +    char *slevel = (char *) qdict_get_try_str(qdict, "level");
> +    uint32_t level = TB_EXEC_STATS | TB_JIT_STATS;
> +    if (slevel) {
> +        if (strcmp(slevel, "jit_stats") == 0) {
> +            level = TB_JIT_STATS;
> +        } else if (strcmp(slevel, "exec_stats") == 0) {
> +            level = TB_EXEC_STATS;
> +        }
> +    }
> +
> +    struct TbstatsCommand *tbscommand = g_new0(struct TbstatsCommand, 1);
> +    tbscommand->cmd = icmd;
> +    tbscommand->level = level;

It might be worth considering squashing these into a single pointer to
avoid the allocation and free?

> +    async_safe_run_on_cpu(first_cpu, do_hmp_tbstats_safe,
> +                          RUN_ON_CPU_HOST_PTR(tbscommand));
> +
> +}
> +
>  static void hmp_info_jit(Monitor *mon, const QDict *qdict)
>  {
>      if (!tcg_enabled()) {


--
Alex Bennée