Re: [PATCH v5 2/9] fprobe: Add ftrace based probe APIs

[Masami Hiramatsu]

On Tue, 25 Jan 2022 19:11:52 +0100
Jiri Olsa <jolsa@xxxxxxxxxx> wrote:

> On Tue, Jan 25, 2022 at 05:41:24PM +0100, Jiri Olsa wrote:
> > On Tue, Jan 25, 2022 at 09:11:57PM +0900, Masami Hiramatsu wrote:
> > 
> > SNIP
> > 
> > > +
> > > +/* Convert ftrace location address from symbols */
> > > +static int convert_func_addresses(struct fprobe *fp)
> > > +{
> > > +	unsigned long addr, size;
> > > +	unsigned int i;
> > > +
> > > +	/* Convert symbols to symbol address */
> > > +	if (fp->syms) {
> > > +		fp->addrs = kcalloc(fp->nentry, sizeof(*fp->addrs), GFP_KERNEL);
> > > +		if (!fp->addrs)
> > > +			return -ENOMEM;
> > > +
> > > +		for (i = 0; i < fp->nentry; i++) {
> > > +			fp->addrs[i] = kallsyms_lookup_name(fp->syms[i]);
> > > +			if (!fp->addrs[i])	/* Maybe wrong symbol */
> > > +				goto error;
> > > +		}
> > > +	}
> > > +
> > > +	/* Convert symbol address to ftrace location. */
> > > +	for (i = 0; i < fp->nentry; i++) {
> > > +		if (!kallsyms_lookup_size_offset(fp->addrs[i], &size, NULL))
> > > +			size = MCOUNT_INSN_SIZE;
> > > +		addr = ftrace_location_range(fp->addrs[i], fp->addrs[i] + size);
> > 
> > you need to substract 1 from 'end' in here, as explained in
> > __within_notrace_func comment:
> > 
> >         /*
> >          * Since ftrace_location_range() does inclusive range check, we need
> >          * to subtract 1 byte from the end address.
> >          */
> > 
> > like in the patch below

Good catch, I missed that point.

> > also this convert is for archs where address from kallsyms does not match
> > the real attach addresss, like for arm you mentioned earlier, right?

Yes, that's right.

> > 
> > could we have that arch specific, so we don't have extra heavy search
> > loop for archs that do not need it?

Hmm, is that so heavy? Even though, this kind of convertion is better
to be implemented in the arch-specific ftrace. They may know the
best way to do, because the offset is fixed for each arch.

> 
> one more question..
> 
> I'm adding support for user to pass function symbols to bpf fprobe link
> and I thought I'd pass symbols array to register_fprobe, but I'd need to
> copy the whole array of strings from user space first, which could take
> lot of memory considering attachment of 10k+ functions
> 
> so I'm thinking better way is to resolve symbols already in bpf fprobe
> link code and pass just addresses to register_fprobe

That is OK. Fprobe accepts either ::syms or ::addrs.

> 
> I assume you want to keep symbol interface, right? could we have some
> flag ensuring the conversion code is skipped, so we don't go through
> it twice?

Yeah, we still have many unused bits in fprobe::flags. :)

Thank you,

> in any case I need addresses before I call register_fprobe, because
> of the bpf cookies setup
> 
> thanks,
> jirka
> 


-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>