Re: [PATCH 2/3] bpf: Optimize get_modules_for_addrs()

[Jiri Olsa <olsajiri@gmail.com>]

On Mon, Jan 09, 2023 at 04:51:37PM +0800, Leizhen (ThunderTown) wrote:
> 
> 
> On 2023/1/6 17:45, Jiri Olsa wrote:
> > On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
> >> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
> >>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> >>>> Function __module_address() can quickly return the pointer of the module
> >>>> to which an address belongs. We do not need to traverse the symbols of all
> >>>> modules to check whether each address in addrs[] is the start address of
> >>>> the corresponding symbol, because register_fprobe_ips() will do this check
> >>>> later.
> >>
> >> hum, for some reason I can see only replies to this patch and
> >> not the actual patch.. I'll dig it out of the lore I guess
> >>
> >>>>
> >>>> Assuming that there are m modules, each module has n symbols on average,
> >>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> >>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> >>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> >>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> >>>> the ratio is still greater than 1. Therefore, the new method will
> >>>> generally have better performance.
> >>
> >> could you try to benchmark that? I tried something similar but was not
> >> able to get better performance
> > 
> > hm looks like I tried the smilar thing (below) like you did,
> 
> Yes. I just found out you're working on this improvement, too.
> 
> > but wasn't able to get better performace
> 
> Your implementation below is already the limit that can be optimized.
> If the performance is not improved, it indicates that this place is
> not the bottleneck.
> 
> > 
> > I guess your goal is to get rid of the module arg in
> > module_kallsyms_on_each_symbol callback that we use?
> 
> It's not a bad thing to keep argument 'mod' for function
> module_kallsyms_on_each_symbol(), but for kallsyms_on_each_symbol(),
> it's completely redundant. Now these two functions often use the
> same hook function. So I carefully analyzed get_modules_for_addrs(),
> which is the only place that involves the use of parameter 'mod'.
> Looks like there's a possibility of eliminating parameter 'mod'.
> 
> > I'm ok with the change if the performace is not worse
> 
> OK, thanks.
> 
> > 
> > jirka
> > 
> > 
> > ---
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 5b9008bc597b..3280c22009f1 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2692,23 +2692,16 @@ struct module_addr_args {
> >  	int mods_cap;
> >  };
> >  
> > -static int module_callback(void *data, const char *name,
> > -			   struct module *mod, unsigned long addr)
> > +static int add_module(struct module_addr_args *args, struct module *mod)
> >  {
> > -	struct module_addr_args *args = data;
> >  	struct module **mods;
> >  
> > -	/* We iterate all modules symbols and for each we:
> > -	 * - search for it in provided addresses array
> > -	 * - if found we check if we already have the module pointer stored
> > -	 *   (we iterate modules sequentially, so we can check just the last
> > -	 *   module pointer)
> > +	/* We iterate sorted addresses and for each within module we:
> > +	 * - check if we already have the module pointer stored for it
> > +	 *   (we iterate sorted addresses sequentially, so we can check
> > +	 *   just the last module pointer)
> >  	 * - take module reference and store it
> >  	 */
> > -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> > -		       bpf_kprobe_multi_addrs_cmp))
> > -		return 0;
> > -
> >  	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> >  		return 0;
> 
> There'll be problems Petr mentioned.
> 
> https://lkml.org/lkml/2023/1/5/191

ok, makes sense.. I guess we could just search args->mods in here?
are you going to send new version, or should I update my patch with that?

thanks,
jirka