Re: [LSF/MM/BPF TOPIC] tracing the source of errors

[Dave Chinner]

On Thu, Feb 08, 2024 at 03:57:27PM +0000, Matthew Wilcox wrote:
> On Thu, Feb 08, 2024 at 08:37:29AM +1100, Dave Chinner wrote:
> > ftrace using the function_graph tracer will emit the return values
> > of the functions if you use it with the 'funcgraph-retval' option.
> 
> OK, but that may not be fine grained enough.  Why is mmap() returning
> -ENOMEM?
> 
> unsigned long do_mmap(struct file *file, unsigned long addr,
> ...
>        /* Careful about overflows.. */
>         len = PAGE_ALIGN(len);
>         if (!len)
>                 return -ENOMEM;
> ...
>         /* Too many mappings? */
>         if (mm->map_count > sysctl_max_map_count)
>                 return -ENOMEM;
> 
> So it can distinguish between mmap() returning ENOMEM because
> get_unmapped_area() returned ENOMEM and do_mmap() returning ENOMEM of
> its own accord (right?),

The call stack trace should tell you which function the error
originated from, yes?

> but it can't tell you which of the above two
> cases you hit.  Or can it?

Never used it, but it might be able to - the "sym-offset" option
will display exact offsets of the function in the trace. If you then
add "funcgraph-tail" it will emit the function being returned from.
If the return location is generated with an offset indicating the
actual return, then it might tell you the exact location.

If it doesn't, then this would seem like a reasonable thing to add
to ftrace - function return tracing with a filter to grab the return
location when the return value is less than 0 seems exactly the sort
of thing ftrace was intended to be used for...

-Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx