[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-kernel
Subject:    Re: [PATCH v3] tty: tty_io: remove hung_up_tty_fops
From:       "Paul E. McKenney" <paulmck () kernel ! org>
Date:       2024-05-03 23:59:12
Message-ID: 2beaba9f-6f83-4a7c-8835-fe5fe88a006c () paulmck-laptop
[Download RAW message or body]

On Thu, May 02, 2024 at 09:37:17AM -0700, Boqun Feng wrote:
> On Wed, May 01, 2024 at 03:32:34PM -0700, Paul E. McKenney wrote:
> > On Wed, May 01, 2024 at 02:49:17PM -0700, Paul E. McKenney wrote:
> > > On Wed, May 01, 2024 at 02:20:35PM -0700, Linus Torvalds wrote:
> > > > On Wed, 1 May 2024 at 14:06, Linus Torvalds
> > > > <torvalds@linux-foundation.org> wrote:
> > 
> > [ . . . ]
> > 
> > > > I'd love to see an extension where "const volatile" basically means
> > > > exactly that: the volatile tells the compiler that it can't
> > > > rematerialize by doing the load multiple times, but the "const" would
> > > > say that if the compiler sees two or more accesses, it can still CSE
> > > > them.
> > 
> > Except that "const volatile" already means "you cannot write to it,
> > and reads will not be fused".  :-/
> > 
> > > No promises, other than that if we don't ask, they won't say "yes".
> > > 
> > > Let me see what can be done.
> > 
> > >From a semantics viewpoint __atomic_load_n(&x, __ATOMIC_RELAXED) would
> > work for loading from x.  The compilers that I tried currently do not
> > fuse loads, but they are allowed to do so.
> 
> Yeah, I wonder the same, from what I read, "const volatile" seems to
> be just a (non-volatile) relaxed atomic load.

Hmmm...  Maybe something like this very lightly tested patch?

(I did not immediately see a use case for WRITE_ONCE_MERGEABLE(),
but that is likely a failure of imagination on my part.)

------------------------------------------------------------------------

diff --git a/include/asm-generic/rwonce.h b/include/asm-generic/rwonce.h
index 8d0a6280e9824..55e87a8aec56f 100644
--- a/include/asm-generic/rwonce.h
+++ b/include/asm-generic/rwonce.h
@@ -79,6 +79,15 @@ unsigned long __read_once_word_nocheck(const void *addr)
 	(typeof(x))__read_once_word_nocheck(&(x));			\
 })
 
+/*
+ * Use READ_ONCE_MERGEABLE() and WRITE_ONCE_MERGEABLE() when you need to
+ * avoid duplicating or tearing a load or store, respectively, but when
+ * it is OK to merge nearby loads and stores.  It must also be OK for a
+ * later nearby load to take its value directly from a prior store.
+ */
+#define READ_ONCE_MERGEABLE(x) __atomic_load_n(&x, __ATOMIC_RELAXED)
+#define WRITE_ONCE_MERGEABLE(x, val) __atomic_store_n(&x, val, __ATOMIC_RELAXED)
+
 static __no_kasan_or_inline
 unsigned long read_word_at_a_time(const void *addr)
 {
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index d5507ac1bbf19..b37c0dbde8cde 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -459,7 +459,7 @@ static void adjust_jiffies_till_sched_qs(void)
 		return;
 	}
 	/* Otherwise, set to third fqs scan, but bound below on large system. */
-	j = READ_ONCE(jiffies_till_first_fqs) +
+	j = READ_ONCE_MERGEABLE(jiffies_till_first_fqs) +
 		      2 * READ_ONCE(jiffies_till_next_fqs);
 	if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
 		j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;

[prev in list] [next in list] [prev in thread] [next in thread]