Re: [PATCH RFC 04/10] perf: Introduce deferred user callchains
From: Peter Zijlstra
Date: Mon Nov 13 2023 - 12:21:31 EST
On Mon, Nov 13, 2023 at 08:56:39AM -0800, Namhyung Kim wrote:
> Ok, then BPF gets no user call stacks even with sframes.
Well, you could obviously run another BPF program at return to user and
stitch there.
> Ok, but ...
> If the task is preempted, the call stack would be from another
> task (if it also has the pending call stacks) and we need to
> check which user call stack matches which kernel call stack.
> So there's no guarantee we can just use adjacent samples.
So upon requesting a user backtrace for the task it could request a
token (from eg a global atomic u64 in the most crude case) and place
this in the task_struct. The backtrace will emit this as forward
reference for the user trace.
Once the task_work generates the user stacktrace, it will again dump
this token along with the user unwind, after which it will reset the
token for this this task (to 0).