[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <b70994c4-04dd-4120-8598-fe1d604a07d8@paulmck-laptop>
Date: Mon, 21 Oct 2024 06:57:26 -0700
From: "Paul E. McKenney" <paulmck@...nel.org>
To: Peter Zijlstra <peterz@...radead.org>
Cc: Andrii Nakryiko <andrii.nakryiko@...il.com>,
Andrii Nakryiko <andrii@...nel.org>,
linux-trace-kernel@...r.kernel.org, oleg@...hat.com,
rostedt@...dmis.org, mhiramat@...nel.org, mingo@...nel.org,
bpf@...r.kernel.org, linux-kernel@...r.kernel.org, jolsa@...nel.org
Subject: Re: [PATCH v2 tip/perf/core 2/2] uprobes: SRCU-protect uretprobe
lifetime (with timeout)
On Mon, Oct 21, 2024 at 12:48:15PM +0200, Peter Zijlstra wrote:
> On Fri, Oct 18, 2024 at 11:22:09AM -0700, Andrii Nakryiko wrote:
>
> > > So... after a few readings I think I'm mostly okay with this. But I got
> > > annoyed by the whole HPROBE_STABLE with uprobe=NULL weirdness. Also,
> > > that data_race() usage is weird, what is that about?
> >
> > People keep saying that evil KCSAN will come after me if I don't add
> > data_race() for values that can change under me, so I add it to make
> > it explicit that it's fine. But I can of course just drop data_race(),
> > as it has no bearing on correctness.
>
> AFAICT this was READ_ONCE() vs xchg(), and that should work. Otherwise I
> have to yell at KCSAN people again :-)
>
> > > And then there's the case where we end up doing:
> > >
> > > try_get_uprobe()
> > > put_uprobe()
> > > try_get_uprobe()
> > >
> > > in the dup path. Yes, it's unlikely, but gah.
> > >
> > >
> > > So how about something like this?
> >
> > Yep, it makes sense to start with HPROBE_GONE if it's already NULL, no
> > problem. I'll roll those changes in.
> >
> > I'm fine with the `bool get` flag as well. Will incorporate all that
> > into the next revision, thanks!
> >
> > The only problem I can see is in the assumption that `srcu_idx < 0` is
> > never going to be returned by srcu_read_lock(). Paul says that it can
> > only be 0 or 1, but it's not codified as part of a contract.
>
> Yeah, [0,1] is the current range. Fundamentally that thing is an array
> index, so negative values are out and generally safe to use as 'error'
> codes. Paul can't we simply document that the SRCU cookie is always a
> positive integer (or zero) and the negative space shall not be used?
We are looking at a few approaches, but they all guarantee that the
return value will be non-negative. My current guess is that we will
just document this non-negative return value, but in all cases, you
should feel free to assume non-negative starting now.
Thanx, Paul
> > So until we change that, probably safer to pass an extra bool
> > specifying whether srcu_idx is valid or not, is that OK?
>
> I think Changeing the SRCU documentation to provide us this guarantee
> should be an achievable goal.
>
> > (and I assume you want me to drop verbose comments for various states, right?)
>
> I axed the comments because I made them invalid and didn't care enough
> to fix them up. If you like them feel free to amend them to reflect the
> new state of things.
Powered by blists - more mailing lists