[DTrace-devel] [PATCH v3 14/19] Ignore clauses in USDT trampoline if we know they are impossible

[Kris Van Hees]

On Wed, Oct 16, 2024 at 12:01:39PM -0400, eugene.loh at oracle.com wrote:
> From: Eugene Loh <eugene.loh at oracle.com>
> 
> An underlying probe might have to call all sorts of clauses if new
> USDT processes start up.  Currently, the underlying probe trampoline
> simply loops over all clauses, deciding whether to call a clause
> based on a run-time bit mask.
> 
> While this approach works, it can mean, e.g., that clauses are being
> loaded unnecessarily into BPF programs, bloating code.  It also means
> that the trampoline is looping over unnecessarily many clauses.
> 
> Meanwhile, it is possible to know in certain cases that a clause
> could never be called for a particular underlying probe.  The
> heuristics can be tricky, but that challenge can be faced incrementally.
> 
> Introduce an ignore_clause() function that, for an underlying probe,
> determines whether some clause (denoted by index) can safely be
> ignored.
> 
> The same ignore_clause() function should be called both during code
> generation of the trampoline as well as during the construction of
> the bit mask.  Further, for a given clause n and underlying probe upp,
> the function should always give the same output.  The set of ignored
> clauses should remain fixed over the lifetime of an underlying probe.
> 
> For now, conservatively, ignore_clause() ignores nothing.  Later
> patches will introduce heuristics for ignoring clauses.
> 
> Signed-off-by: Eugene Loh <eugene.loh at oracle.com>

Reviewed-by: Kris Van Hees <kris.van.hees at oracle.com>

> ---
>  libdtrace/dt_prov_uprobe.c | 20 +++++++++++++++++++-
>  1 file changed, 19 insertions(+), 1 deletion(-)
> 
> diff --git a/libdtrace/dt_prov_uprobe.c b/libdtrace/dt_prov_uprobe.c
> index 0289c9223..a3c04478c 100644
> --- a/libdtrace/dt_prov_uprobe.c
> +++ b/libdtrace/dt_prov_uprobe.c
> @@ -263,6 +263,17 @@ clean_usdt_probes(dtrace_hdl_t *dtp)
>  	return 0;
>  }
>  
> +/*
> + * Judge whether clause "n" could ever be called as a USDT probe
> + * for this underlying probe.
> + */
> +static int
> +ignore_clause(dtrace_hdl_t *dtp, int n, const dt_probe_t *uprp)
> +{
> +	/* To be safe, ignore nothing. */
> +	return 0;
> +}
> +
>  static int add_probe_uprobe(dtrace_hdl_t *dtp, dt_probe_t *prp)
>  {
>  	dtrace_difo_t   *dp;
> @@ -388,6 +399,9 @@ static int add_probe_usdt(dtrace_hdl_t *dtp, dt_probe_t *prp)
>  				if (stp == NULL)
>  					continue;
>  
> +				if (ignore_clause(dtp, n, uprp))
> +					continue;
> +
>  				if (dt_gmatch(prp->desc->prv, stp->dtsd_ecbdesc->dted_probe.prv) &&
>  				    dt_gmatch(prp->desc->mod, stp->dtsd_ecbdesc->dted_probe.mod) &&
>  				    dt_gmatch(prp->desc->fun, stp->dtsd_ecbdesc->dted_probe.fun) &&
> @@ -862,13 +876,17 @@ static int trampoline(dt_pcb_t *pcb, uint_t exitlbl)
>  	for (n = 0; n < dtp->dt_stmt_nextid; n++) {
>  		dtrace_stmtdesc_t *stp;
>  		dt_ident_t	*idp;
> -		uint_t		lbl_next = dt_irlist_label(dlp);
> +		uint_t		lbl_next;
>  
>  		stp = dtp->dt_stmts[n];
>  		if (stp == NULL)
>  			continue;
>  
> +		if (ignore_clause(dtp, n, uprp))
> +			continue;
> +
>  		idp = stp->dtsd_clause;
> +		lbl_next = dt_irlist_label(dlp);
>  
>  		/* If the lowest %r6 bit is 0, skip over this clause. */
>  		emit(dlp,  BPF_MOV_REG(BPF_REG_1, BPF_REG_6));
> -- 
> 2.43.5
>