Re: [PATCH bpf-next 2/2] libbpf: usdt aarch64 arg parsing support

[Alan Maguire <alan.maguire@oracle.com>]

On Mon, 11 Apr 2022, Andrii Nakryiko wrote:

> On Fri, Apr 8, 2022 at 3:53 PM Alan Maguire <alan.maguire@oracle.com> wrote:
> >
> > Parsing of USDT arguments is architecture-specific; on aarch64 it is
> > relatively easy since registers used are x[0-31], sp.  Format is
> > slightly different compared to x86_64; forms are
> >
> > - "size @ [ reg[,offset] ]" for dereferences, for example
> >   "-8 @ [ sp, 76 ]" ; " -4 @ [ sp ]"
> > - "size @ reg" for register values; for example
> >   "-4@x0"
> > - "size @ value" for raw values; for example
> >   "-8@1"
> >
> > Signed-off-by: Alan Maguire <alan.maguire@oracle.com>
> > ---
> >  tools/lib/bpf/usdt.c | 50 +++++++++++++++++++++++++++++++++++++++++++++++++-
> >  1 file changed, 49 insertions(+), 1 deletion(-)
> >
> > diff --git a/tools/lib/bpf/usdt.c b/tools/lib/bpf/usdt.c
> > index 0677bbd..6165d40 100644
> > --- a/tools/lib/bpf/usdt.c
> > +++ b/tools/lib/bpf/usdt.c
> > @@ -1170,7 +1170,7 @@ static int parse_usdt_spec(struct usdt_spec *spec, const struct usdt_note *note,
> >
> >  /* Architecture-specific logic for parsing USDT argument location specs */
> >
> > -#if defined(__x86_64__) || defined(__i386__) || defined(__s390x__)
> > +#if defined(__x86_64__) || defined(__i386__) || defined(__s390x__) || defined(__aarch64__)
> >
> >  static int init_usdt_arg_spec(struct usdt_arg_spec *arg, enum usdt_arg_type arg_type, int arg_sz,
> >                               __u64 val_off, int reg_off)
> > @@ -1316,6 +1316,54 @@ static int parse_usdt_arg(const char *arg_str, int arg_num, struct usdt_arg_spec
> >         return len;
> >  }
> >
> > +#elif defined(__aarch64__)
> > +
> > +static int calc_pt_regs_off(const char *reg_name)
> > +{
> > +       int reg_num;
> > +
> > +       if (sscanf(reg_name, "x%d", &reg_num) == 1) {
> > +               if (reg_num >= 0 && reg_num < 31)
> > +                       return offsetof(struct user_pt_regs, regs[reg_num]);
> > +       } else if (strcmp(reg_name, "sp") == 0) {
> > +               return offsetof(struct user_pt_regs, sp);
> > +       }
> > +       pr_warn("usdt: unrecognized register '%s'\n", reg_name);
> > +       return -ENOENT;
> > +}
> > +
> > +static int parse_usdt_arg(const char *arg_str, int arg_num, struct usdt_arg_spec *arg)
> > +{
> > +       char *reg_name = NULL;
> > +       int arg_sz, len, ret;
> > +       long off = 0;
> > +
> > +       if (sscanf(arg_str, " %d @ \[ %m[^,], %ld ] %n", &arg_sz, &reg_name, &off, &len) == 3 ||
> > +           sscanf(arg_str, " %d @ \[ %m[a-z0-9] ] %n", &arg_sz, &reg_name, &len) == 2) {
> 
> I'm not sure about the behavior here w.r.t. reg_name and memory
> allocation. What if first sscanf() matches reg_name but fails at %ld,
> will reg_name be allocated and then second sscanf() will reallocate
> (and thus we'll have a memory leak).
> 
> We might have similar problems in other implementations, actually...
> 
> Either way, came here to ask to split two sscanfs into two separate
> branches, so that we have a clear linear pattern. One sscanf, handle
> it if successful, otherwise move on to next case.
> 

good point; I'll separate the sscanfs into branches for v2.

> Also a question about [a-z0-9] for register in one case and [^,] in
> another. Should the first one be [a-z0-9] as well?
>

probably no harm, yep.

I'll drop the refactoring patch too; I was a bit worried I'd break
Ilya's s390 code anyhow. 

Thanks!

Alan