Re: [PATCH v1 2/8] tracing/ftrace: guard syscall probe with preempt_notrace

[Mathieu Desnoyers]

On 2024-10-04 03:04, Steven Rostedt wrote:
> On Thu, 3 Oct 2024 20:26:29 -0400
> Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
>
> > static void ftrace_syscall_enter(void *data, struct pt_regs *regs, long id)
> > {
> >           struct trace_array *tr = data;
> >           struct trace_event_file *trace_file;
> >           struct syscall_trace_enter *entry;
> >           struct syscall_metadata *sys_data;
> >           struct trace_event_buffer fbuffer;
> >           unsigned long args[6];
> >           int syscall_nr;
> >           int size;
> >
> >           syscall_nr = trace_get_syscall_nr(current, regs);
> >           if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
> >                   return;
> >
> >           /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE) */
> >           trace_file = rcu_dereference_sched(tr->enter_syscall_files[syscall_nr]);
> >
> > ^^^^ this function explicitly states that preempt needs to be disabled by
> > tracepoints.
>
> Ah, I should have known it was the syscall portion. I don't care for this
> hidden dependency. I rather add a preempt disable here and not expect it to
> be disabled when called.

Which is exactly what this patch is doing.

> >           if (!trace_file)
> >                   return;
> >
> >           if (trace_trigger_soft_disabled(trace_file))
> >                   return;
> >
> >           sys_data = syscall_nr_to_meta(syscall_nr);
> >           if (!sys_data)
> >                   return;
> >
> >           size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
> >
> >           entry = trace_event_buffer_reserve(&fbuffer, trace_file, size);
> > ^^^^ it reserves space in the ring buffer without disabling preemption explicitly.
> >
> > And also:
> >
> > void *trace_event_buffer_reserve(struct trace_event_buffer *fbuffer,
> >                                    struct trace_event_file *trace_file,
> >                                    unsigned long len)
> > {
> >           struct trace_event_call *event_call = trace_file->event_call;
> >
> >           if ((trace_file->flags & EVENT_FILE_FL_PID_FILTER) &&
> >               trace_event_ignore_this_pid(trace_file))
> >                   return NULL;
> >
> >           /*
> >            * If CONFIG_PREEMPTION is enabled, then the tracepoint itself disables
> >            * preemption (adding one to the preempt_count). Since we are
> >            * interested in the preempt_count at the time the tracepoint was
> >            * hit, we need to subtract one to offset the increment.
> >            */
> > ^^^ This function also explicitly expects preemption to be disabled.
> >
> > So I rest my case. The change I'm introducing for tracepoints
> > don't make any assumptions about whether or not each tracer require
> > preempt off or not: it keeps the behavior the _same_ as it was before.
> >
> > Then it's up to each tracer's developer to change the behavior of their
> > own callbacks as they see fit. But I'm not introducing regressions in
> > tracers with the "big switch" change of making syscall tracepoints
> > faultable. This will belong to changes that are specific to each tracer.
>
>
> I rather remove these dependencies at the source. So, IMHO, these places
> should be "fixed" first.
>
> At least for the ftrace users. But I think the same can be done for the
> other users as well. BPF already stated it just needs "migrate_disable()".
> Let's see what perf has.
>
> We can then audit all the tracepoint users to make sure they do not need
> preemption disabled.

Why does it need to be a broad refactoring of the entire world ? What is
wrong with the simple approach of introducing this tracepoint faultable
syscall support as a no-op from the tracer's perspective ?

Then we can build on top and figure out if we want to relax things
on a tracer-per-tracer basis.

Thanks,

Mathieu



--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com