[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20221215195854.GL4001@paulmck-ThinkPad-P17-Gen-1>
Date: Thu, 15 Dec 2022 11:58:54 -0800
From: "Paul E. McKenney" <paulmck@...nel.org>
To: Joel Fernandes <joel@...lfernandes.org>
Cc: Frederic Weisbecker <frederic@...nel.org>, boqun.feng@...il.com,
neeraj.iitr10@...il.com, urezki@...il.com, rcu@...r.kernel.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH RFC] srcu: Yet more detail for
srcu_readers_active_idx_check() comments
On Thu, Dec 15, 2022 at 05:48:46PM +0000, Joel Fernandes wrote:
> On Thu, Dec 15, 2022 at 5:08 PM Paul E. McKenney <paulmck@...nel.org> wrote:
>
> > > Scenario for the reader to increment the old idx once:
> > >
> > > _ Assume ssp->srcu_idx is initially 0.
> > > _ The READER reads idx that is 0
> > > _ The updater runs and flips the idx that is now 1
> > > _ The reader resumes with 0 as an index but on the next srcu_read_lock()
> > > it will see the new idx which is 1
> > >
> > > What could be the scenario for it to increment the old idx twice?
> >
> > Unless I am missing something, the reader must reference the
> > srcu_unlock_count[old_idx] and then do smp_mb() before it will be
> > absolutely guaranteed of seeing the new value of ->srcu_idx.
>
> I think both of you are right depending on how the flip raced with the
> first reader's unlock in that specific task.
There are indeed a variety of scenarios and also a variety of failure
cases.
> If the first read section's srcu_read_unlock() and its corresponding
> smp_mb() happened before the flip, then the increment of old idx
> would happen only once. The next srcu_read_lock() will read the new
> index. If the srcu_read_unlock() and it's corresponding smp_mb()
> happened after the flip, the old_idx will be sampled again and can be
> incremented twice. So it depends on how the flip races with
> srcu_read_unlock().
I do understand that a number of people like reasoning about
memory-barrier ordering, courtesy of the sequentially consistent portions
of the C and C++ memory models, but thinking in terms of the accesses
surrounding the memory barriers has been far less error-prone.
> Also, since this is all hard to reason about I started making some
> diagrams, LOL. For your amusement, here is why need to scan both idx
> during grace period detection: https://i.imgur.com/jz4bNKd.png
Nice!
I suggest placing a gap between GP 2 and GP 3. That way, you can make it
very clear that Reader 1's critical section starts after the end of GP 2
(thus clearly never blocking GP 2) and before GP 3 (thus possibly having
a reference to some data that is going to be freed at the end of GP 3).
I also suggest coloring Reader 1 red and Reader 2 green, given that the
color red generally indicates danger.
Thanx, Paul
Powered by blists - more mailing lists