On Mon, May 20, 2024 at 8:20 AM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
>
> On Wed, May 15, 2024 at 08:32:30AM -0600, Andrii Nakryiko wrote:
> > On Wed, May 15, 2024 at 3:30 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Wed, May 08, 2024 at 02:26:03PM -0700, Andrii Nakryiko wrote:
> > >
> > > > +static void fixup_uretprobe_trampoline_entries(struct perf_callchain_entry *entry,
> > > > + int start_entry_idx)
> > > > +{
> > > > +#ifdef CONFIG_UPROBES
> > > > + struct uprobe_task *utask = current->utask;
> > > > + struct return_instance *ri;
> > > > + __u64 *cur_ip, *last_ip, tramp_addr;
> > > > +
> > > > + if (likely(!utask || !utask->return_instances))
> > > > + return;
> > > > +
> > > > + cur_ip = &entry->ip[start_entry_idx];
> > > > + last_ip = &entry->ip[entry->nr - 1];
> > > > + ri = utask->return_instances;
> > > > + tramp_addr = uprobe_get_trampoline_vaddr();
> > > > +
> > > > + /* If there are pending uretprobes for current thread, they are
> > >
> > > Comment style fail. Also 'for *the* current thread'.
> > >
> >
> > ack, will fix
> >
> > > > + * recorded in a list inside utask->return_instances; each such
> > > > + * pending uretprobe replaces traced user function's return address on
> > > > + * the stack, so when stack trace is captured, instead of seeing
> > > > + * actual function's return address, we'll have one or many uretprobe
> > > > + * trampoline addresses in the stack trace, which are not helpful and
> > > > + * misleading to users.
> > >
> > > I would beg to differ, what if the uprobe is causing the performance
> > > issue?
> >
> > If uprobe/uretprobe code itself is causing performance issues, you'll
> > see that in other stack traces, where this code will be actively
> > running on CPU. I don't think we make anything worse here.
>
> I think we do similar thing in kernel unwind for rethook trampoline used
> in fprobe/kretprobe code, so seems ok to me to do it for uprobes as well
>
> >
> > Here we are talking about the case where the uprobe part is done and
> > it hijacked the return address on the stack, uretprobe is not yet
> > running (and so not causing any performance issues). The presence of
> > this "snooping" (pending) uretprobe is irrelevant to the user that is
> > capturing stack trace. Right now address in [uprobes] VMA section
> > installed by uretprobe infra code is directly replacing correct and
> > actual calling function address.
> >
> > Worst case, one can argue that both [uprobes] and original caller
> > address should be in the stack trace, but I think it still will be
> > confusing to users. And also will make implementation less efficient
> > because now we'll need to insert entries into the array and shift
> > everything around.
>
> agreed this would be confusing.. also as you noted above the return
> trampoline did not get executed yet at the time of the callstack,
> so it's bit misleading
>
> might be stupid idea.. but we do have the 'special' context entries
> that we store in the callstack to mark user/kernel/guest context ..
only when explicitly requested (add_mark argument to
get_perf_callchain), right? BPF doesn't ever set this to true and
generally speaking users don't care and shouldn't care about pending
uretprobe. I think we are conflating unrelated things here, uretprobe
is not running, so it's not really in the stack trace. I'd just do
nothing about it, it should stay transparent.
If uretprobe *handler* is causing issues, you'll see that in all the
other stack traces (according to relative CPU/resource usage of that
handler).
> maybe we could add some special entry (context does not fit too well)
> to point out there's uretprobe going on .. perf tool could print
> 'uretprobe' hint when displaying the original address
>
> jirka
>
> >
> > So as I mentioned above, if the concern is seeing uprobe/uretprobe
> > code using CPU, that doesn't change, we'll see that in the overall set
> > of captured stack traces (be it custom uprobe handler code or BPF
> > program).
> >
> > >
> > > While I do think it makes sense to fix the unwind in the sense that we
> > > should be able to continue the unwind, I don't think it makes sense to
> > > completely hide the presence of uprobes.
> >
> > Unwind isn't broken in this sense, we do unwind the entire stack trace
> > (see examples in the later patch). We just don't capture actual
> > callers if they have uretprobe pending.
> >
> > >
> > > > + * So here we go over the pending list of uretprobes, and each
> > > > + * encountered trampoline address is replaced with actual return
> > > > + * address.
> > > > + */
> > > > + while (ri && cur_ip <= last_ip) {
> > > > + if (*cur_ip == tramp_addr) {
> > > > + *cur_ip = ri->orig_ret_vaddr;
> > > > + ri = ri->next;
> > > > + }
> > > > + cur_ip++;
> > > > + }
> > > > +#endif
> > > > +}
> >
