Re: [PATCH v6 5/6] tracing: Show inode and device major:minor in deferred user space stacktrace

[Arnaldo Carvalho de Melo <arnaldo.melo@gmail.com>]

On August 29, 2025 1:59:21 PM GMT-03:00, Linus Torvalds <torvalds@linux-foundation.org> wrote:
>On Fri, 29 Aug 2025 at 09:49, Steven Rostedt <rostedt@goodmis.org> wrote:
>>
>> What do I use to make the hash?
>
>Literally just '(unsigned long)(vma->vm_file)'.
>
>Nothing else.
>
>> One thing this is trying to do is not have to look up the path name for
>> every line of a stack trace.
>
>That's the *opposite* of what I've been suggesting. I've literally
>been talking about just saving off the hash of the file pointer.
>
>(And I just suggested that what you actually save off isn't even the
>hash - just the value - and that you can hash it later at a less
>critical point in time)
>
>Don't do *any* work at all at trace collection time. All you need is
>to literally access three fields in the 'vma':
>
> - 'vm_start' and 'vm_pgoff' are needed to calculate the offset in the
>file using the user space address
>
> - save off the value of 'vm_file' for later hashing
>
>and I really think you're done.
>
>Then, for the actual trace, you need two things:
>
> - you need the mmap trace event that has the 'file' value, and you
>create a mmap event with that value hashed, and at that point you also
>output the pathname and/or things like the build ID
>
> - for the stack trace events, you output the offset in the file, and
>you hash and output the file value
>
>now, in user space, you have all you need. All you do is match the
>hashes. They are random numbers, and user space cannot know what they
>are. They are just cookies as a mapping ID.
>
>And look, now you have the pathname and the build ID - or whatever you
>saved off in that mmap event. And at stack trace time, you needed to
>do *nothing*.
>
>And mmap is rare enough - and heavy enough - that doing that pathname
>and build ID at *that* point is a non-issue.


Or using a preexisting one in the DSO used for the executable mmap.

As long as we don't lose those mmap events due to memory pressure/lost events and we have timestamps to order it all before lookups, yeah should work.

- Arnaldo 

>
>See what I'm trying to say?
>
>               Linus

- Arnaldo