Re: [PATCH v3] Allow user probes on versioned symbols

[Paul Clarke]

On 04/04/2017 09:26 AM, Arnaldo Carvalho de Melo wrote:
> Em Tue, Apr 04, 2017 at 11:18:02PM +0900, Masami Hiramatsu escreveu:
> > On Mon, 3 Apr 2017 11:46:58 -0300
> > Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> > > > > But apart from those problems, I think that one should be able to ask
> > > > > for a versioned symbol, to probe just apps using that specific version,
> > >
> > > > I agree, but wasn't trying to tackle that at the moment.  I can look into it, though.
> > >
> > > > > for instance, we should consider the whole name as two functions, which
> > > > > in fact, they are, no?
> > >
> > > > I'm not sure I understand what you mean here.  Do you mean we should set a probe at every version of a given symbol name?  For example, if there are symbols:
> > > > a@@V2
> > > > a@xxxx
> > > > a@V1
> > >
> > > > ...for a request to set a probe at "a", we'd actually set a probe at all 3?
> > >
> > > I think that we should just probe the default for that symbol and have a
> > > way to probe all of them, perhaps using the wildcard, i.e.:
> > >
> > > [root@jouet linux]# nm /lib64/libpthread-2.24.so  | grep ' pthread_cond_timedwait'
> > > 000000000000dd90 T pthread_cond_timedwait@xxxxxxxxxxx
> > > 000000000000d6e0 T pthread_cond_timedwait@@GLIBC_2.3.2
> > > [root@jouet linux]#
> > >
> > >   # perf probe -x /lib64/libpthread-2.24.so pthread_cond_timedwait
> > >
> > > should be equivalent to:
> > >
> > >   # perf probe -x /lib64/libpthread-2.24.so pthread_cond_timedwait@@GLIBC_2.3.2
> > >
> > > Which matches how these versioned symbols are resolved by the linker,
> > > no?
> > >
> > > I.e. when 'pthread_cond_timedwait' is specified and the symbol table
> > > lookup fails, I think we should re-lookup for
> > > 'pthread_cond_timedwait@@*', i.e. we should have a
> > > symbol__find_default_by_name(), which will take the
> > > "pthread_cond_timedwait" and use a symbol comparison using
> > > strncmp(strlen(key)), matching, should then look at right after the
> > > common part looking for the double @@.
>
> > Hm, this 'fallback'process sounds good idea to me.

I just sent a patch that does what you suggest, above.  To avoid duplicating the code in symbols_find_by_name, I added a parameter to tell it whether to ignore default symbol tags or not.

> This is just trying to keep the semantics used by the original user of
> this syntax, i.e. the linker.
>
> > BTW, how would we support other SYMBOL@VERSION, since we already
> > use '@' for specifying source code?
> > One possible way is to support it directly in perf-probe. If it
> > failed to find probe point from dwarf, try to find from symbol
> > map by using '@VERSION' suffix.
>
> Right, we would be overloading that @ symbol, since version numbers
> usually are very different of file source names :-)

There's not a lot of syntactic difference between "file" and "tag".  I don't think there is any standard for what either can be.  One might expect a "file" to be name.extension, where extension is a finite set (possibly fairly large).  A tag can be almost anything, I think.  One might expect it to end with a number.  I don't think there's a guarantee of either case, though.

It would seem one way to determine whether it's a file or a tag is to try to find it in the symbol tables.  If it's not there, assume it's a file.  (Or vice-versa.)

PC