All of lore.kernel.org
 help / color / mirror / Atom feed
From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
To: David Howells <dhowells@redhat.com>
Cc: linux-kernel@vger.kernel.org, mingo@kernel.org,
	torvalds@linux-foundation.org, mark.rutland@arm.com,
	linux-arch@vger.kernel.org, peterz@infradead.org,
	will.deacon@arm.com, Jonathan Corbet <corbet@lwn.net>,
	Alexander Kuleshov <kuleshovmail@gmail.com>
Subject: Re: [PATCH RFC tip/core/rcu 12/15] lib/assoc_array: Remove smp_read_barrier_depends()
Date: Tue, 10 Oct 2017 08:50:42 -0700	[thread overview]
Message-ID: <20171010155042.GD3521@linux.vnet.ibm.com> (raw)
In-Reply-To: <8079.1507628146@warthog.procyon.org.uk>

On Tue, Oct 10, 2017 at 10:35:46AM +0100, David Howells wrote:
> Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> 
> >  static inline void *assoc_array_ptr_to_leaf(const struct assoc_array_ptr *x)
> >  {
> > -	return (void *)((unsigned long)x & ~ASSOC_ARRAY_PTR_TYPE_MASK);
> > +	return (void *)((unsigned long)READ_ONCE(x) & /* Address dependency. */
> > +		~ASSOC_ARRAY_PTR_TYPE_MASK);
> >  }
> 
> This is the wrong place to do this.  assoc_array_ptr_to_leaf() is effectively
> no more than a special cast; it removes a metadata bit from a pointer.  x is
> the value we're modifying, not *x, and x was read by the caller.

OK, so in at least one of these cases, the barrier is already provided
by the ACCESS_ONCE(array->root) in assoc_array_walk().  I took a quick
look at the other instances of ->root and convinced myself that they
were update-side accesses, but please let me know if I am missing
something.

I marked the ACCESS_ONCE() instances that seemed to me to be heading
address-dependency chains, please let me know how I did.  ;-)

> >  static inline
> >  unsigned long __assoc_array_ptr_to_meta(const struct assoc_array_ptr *x)
> >  {
> > -	return (unsigned long)x &
> > +	return (unsigned long)READ_ONCE(x) & /* Address dependency. */
> >  		~(ASSOC_ARRAY_PTR_SUBTYPE_MASK | ASSOC_ARRAY_PTR_TYPE_MASK);
> >  }
> 
> Ditto.
> 
> > -		ptr = READ_ONCE(node->slots[slot]);
> > +		ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
> >  		has_meta |= (unsigned long)ptr;
> >  		if (ptr && assoc_array_ptr_is_leaf(ptr)) {
> > -			/* We need a barrier between the read of the pointer
> > -			 * and dereferencing the pointer - but only if we are
> > -			 * actually going to dereference it.
> > -			 */
> > -			smp_read_barrier_depends();
> > -
> 
> For example, you can see the READ_ONCE() here; that is already done.
> 
> Can you also not just delete the comment, but rephrase it to indicate that a
> barrier is necessary and it's done by READ_ONCE(), please?

So, is the patch below at least somewhat less confused?  ;-)

							Thanx, Paul

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

commit a25170d6998c587b296c75be09b49b5e31bc515b
Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Date:   Mon Oct 9 11:39:57 2017 -0700

    lib/assoc_array: Remove smp_read_barrier_depends()
    
    Now that smp_read_barrier_depends() is implied by READ_ONCE(), adding
    READ_ONCE() to assoc_array_ptr_to_leaf() and __assoc_array_ptr_to_meta()
    allows the several smp_read_barrier_depends() calls to be removed from
    lib/assoc_array.c.  This commit makes this change.
    
    Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
    Cc: Jonathan Corbet <corbet@lwn.net>
    Cc: Mark Rutland <mark.rutland@arm.com>
    Cc: Alexander Kuleshov <kuleshovmail@gmail.com>

diff --git a/lib/assoc_array.c b/lib/assoc_array.c
index fe7953aead82..7e483294f7d7 100644
--- a/lib/assoc_array.c
+++ b/lib/assoc_array.c
@@ -38,12 +38,10 @@ static int assoc_array_subtree_iterate(const struct assoc_array_ptr *root,
 	if (assoc_array_ptr_is_shortcut(cursor)) {
 		/* Descend through a shortcut */
 		shortcut = assoc_array_ptr_to_shortcut(cursor);
-		smp_read_barrier_depends();
-		cursor = READ_ONCE(shortcut->next_node);
+		cursor = READ_ONCE(shortcut->next_node); /* Address dependency. */
 	}
 
 	node = assoc_array_ptr_to_node(cursor);
-	smp_read_barrier_depends();
 	slot = 0;
 
 	/* We perform two passes of each node.
@@ -55,15 +53,12 @@ static int assoc_array_subtree_iterate(const struct assoc_array_ptr *root,
 	 */
 	has_meta = 0;
 	for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
-		ptr = READ_ONCE(node->slots[slot]);
+		ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
 		has_meta |= (unsigned long)ptr;
 		if (ptr && assoc_array_ptr_is_leaf(ptr)) {
-			/* We need a barrier between the read of the pointer
-			 * and dereferencing the pointer - but only if we are
-			 * actually going to dereference it.
+			/* We need a barrier between the read of the pointer,
+			 * which is supplied by the above READ_ONCE().
 			 */
-			smp_read_barrier_depends();
-
 			/* Invoke the callback */
 			ret = iterator(assoc_array_ptr_to_leaf(ptr),
 				       iterator_data);
@@ -86,10 +81,8 @@ static int assoc_array_subtree_iterate(const struct assoc_array_ptr *root,
 
 continue_node:
 	node = assoc_array_ptr_to_node(cursor);
-	smp_read_barrier_depends();
-
 	for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
-		ptr = READ_ONCE(node->slots[slot]);
+		ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
 		if (assoc_array_ptr_is_meta(ptr)) {
 			cursor = ptr;
 			goto begin_node;
@@ -98,16 +91,15 @@ static int assoc_array_subtree_iterate(const struct assoc_array_ptr *root,
 
 finished_node:
 	/* Move up to the parent (may need to skip back over a shortcut) */
-	parent = READ_ONCE(node->back_pointer);
+	parent = READ_ONCE(node->back_pointer); /* Address dependency. */
 	slot = node->parent_slot;
 	if (parent == stop)
 		return 0;
 
 	if (assoc_array_ptr_is_shortcut(parent)) {
 		shortcut = assoc_array_ptr_to_shortcut(parent);
-		smp_read_barrier_depends();
 		cursor = parent;
-		parent = READ_ONCE(shortcut->back_pointer);
+		parent = READ_ONCE(shortcut->back_pointer); /* Address dependency. */
 		slot = shortcut->parent_slot;
 		if (parent == stop)
 			return 0;
@@ -147,7 +139,7 @@ int assoc_array_iterate(const struct assoc_array *array,
 					void *iterator_data),
 			void *iterator_data)
 {
-	struct assoc_array_ptr *root = READ_ONCE(array->root);
+	struct assoc_array_ptr *root = READ_ONCE(array->root); /* Address dependency. */
 
 	if (!root)
 		return 0;
@@ -194,7 +186,7 @@ assoc_array_walk(const struct assoc_array *array,
 
 	pr_devel("-->%s()\n", __func__);
 
-	cursor = READ_ONCE(array->root);
+	cursor = READ_ONCE(array->root);  /* Address dependency. */
 	if (!cursor)
 		return assoc_array_walk_tree_empty;
 
@@ -216,11 +208,9 @@ assoc_array_walk(const struct assoc_array *array,
 
 consider_node:
 	node = assoc_array_ptr_to_node(cursor);
-	smp_read_barrier_depends();
-
 	slot = segments >> (level & ASSOC_ARRAY_KEY_CHUNK_MASK);
 	slot &= ASSOC_ARRAY_FAN_MASK;
-	ptr = READ_ONCE(node->slots[slot]);
+	ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
 
 	pr_devel("consider slot %x [ix=%d type=%lu]\n",
 		 slot, level, (unsigned long)ptr & 3);
@@ -254,7 +244,6 @@ assoc_array_walk(const struct assoc_array *array,
 	cursor = ptr;
 follow_shortcut:
 	shortcut = assoc_array_ptr_to_shortcut(cursor);
-	smp_read_barrier_depends();
 	pr_devel("shortcut to %d\n", shortcut->skip_to_level);
 	sc_level = level + ASSOC_ARRAY_LEVEL_STEP;
 	BUG_ON(sc_level > shortcut->skip_to_level);
@@ -294,7 +283,7 @@ assoc_array_walk(const struct assoc_array *array,
 	} while (sc_level < shortcut->skip_to_level);
 
 	/* The shortcut matches the leaf's index to this point. */
-	cursor = READ_ONCE(shortcut->next_node);
+	cursor = READ_ONCE(shortcut->next_node); /* Address dependency. */
 	if (((level ^ sc_level) & ~ASSOC_ARRAY_KEY_CHUNK_MASK) != 0) {
 		level = sc_level;
 		goto jumped;
@@ -330,21 +319,19 @@ void *assoc_array_find(const struct assoc_array *array,
 	    assoc_array_walk_found_terminal_node)
 		return NULL;
 
-	node = result.terminal_node.node;
-	smp_read_barrier_depends();
+	node = READ_ONCE(result.terminal_node.node); /* Address dependency. */
 
 	/* If the target key is available to us, it's has to be pointed to by
 	 * the terminal node.
 	 */
 	for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
-		ptr = READ_ONCE(node->slots[slot]);
+		ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
 		if (ptr && assoc_array_ptr_is_leaf(ptr)) {
 			/* We need a barrier between the read of the pointer
 			 * and dereferencing the pointer - but only if we are
 			 * actually going to dereference it.
 			 */
 			leaf = assoc_array_ptr_to_leaf(ptr);
-			smp_read_barrier_depends();
 			if (ops->compare_object(leaf, index_key))
 				return (void *)leaf;
 		}

  reply	other threads:[~2017-10-10 15:50 UTC|newest]

Thread overview: 70+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2017-10-10  0:19 [PATCH RFC tip/core/rcu 0/15] Remove to-be-unneeded smp_read_barrier_depends() Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 01/15] doc: READ_ONCE() now implies smp_barrier_depends() Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 02/15] mn10300: READ_ONCE() now implies smp_read_barrier_depends() Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 03/15] drivers/net/ethernet/qlogic/qed: Fix __qed_spq_block() ordering Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 04/15] fs/dcache: Use release-acquire for name/length update Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 05/15] percpu: READ_ONCE() now implies smp_read_barrier_depends() Paul E. McKenney
2017-10-10 14:08   ` Tejun Heo
2017-10-10 15:30     ` Paul E. McKenney
2017-10-10 15:49       ` Tejun Heo
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 06/15] rcu: Adjust read-side accessor comments for READ_ONCE() Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 07/15] rtnetlink: Update now-misleading smp_read_barrier_depends() comment Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 08/15] seqlock: Remove now-redundant smp_read_barrier_depends() Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 09/15] uprobes: " Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 10/15] locking: Remove smp_read_barrier_depends() from queued_spin_lock_slowpath() Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 11/15] tracepoint: Remove smp_read_barrier_depends() from comment Paul E. McKenney
2017-10-10  0:31   ` Steven Rostedt
2017-10-10  1:12     ` Mathieu Desnoyers
2017-10-10 15:32       ` Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 12/15] lib/assoc_array: Remove smp_read_barrier_depends() Paul E. McKenney
2017-10-10  8:39   ` Peter Zijlstra
2017-10-10  9:36   ` David Howells
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 13/15] mm/ksm: Remove now-redundant smp_read_barrier_depends() Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 14/15] netfilter: " Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  8:43   ` Peter Zijlstra
2017-10-10 15:56     ` Paul E. McKenney
2017-10-10  0:22 ` [PATCH RFC tip/core/rcu 15/15] keyring: " Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  0:22   ` Paul E. McKenney
2017-10-10  9:35 ` [PATCH RFC tip/core/rcu 12/15] lib/assoc_array: Remove smp_read_barrier_depends() David Howells
2017-10-10 15:50   ` Paul E. McKenney [this message]
2017-10-10 15:54     ` Peter Zijlstra
2017-10-10 16:05       ` Paul E. McKenney
2017-10-11 12:19   ` David Howells
2017-10-11 12:22     ` Will Deacon
2017-10-11 12:54       ` Paul E. McKenney
2017-10-11 14:18         ` Will Deacon
2017-10-11 14:50           ` Paul E. McKenney
2017-10-11 12:58     ` Paul E. McKenney
2017-10-11 15:17     ` David Howells
2017-10-11 15:59       ` Paul E. McKenney
2017-10-11 16:12         ` Peter Zijlstra
2017-10-11 16:24           ` Peter Zijlstra
2017-10-11 16:47             ` Paul E. McKenney
2017-10-11 16:54               ` Peter Zijlstra
2017-10-11 17:06                 ` Paul E. McKenney
2017-10-11 17:11                   ` Peter Zijlstra
2017-10-11 17:34                     ` Paul E. McKenney
2017-10-11 18:43                       ` Dmitry Vyukov
2017-10-11 18:56                         ` Linus Torvalds
2017-10-11 19:28                           ` Paul E. McKenney
2017-10-11 19:59                           ` Dmitry Vyukov
2017-10-11 17:14                   ` Paul E. McKenney
2017-10-11 17:19             ` Mark Rutland
2017-10-11 16:50           ` Paul E. McKenney
2017-10-11 16:07       ` David Howells
2017-10-11 16:17         ` Peter Zijlstra
2017-10-11 16:19         ` Paul E. McKenney
2017-10-11 15:28     ` David Howells
2017-10-11 16:02       ` Paul E. McKenney
2017-10-10  9:59 ` David Howells
2017-10-10 15:52   ` Paul E. McKenney
2017-10-11 12:21 ` [PATCH RFC tip/core/rcu 0/15] Remove to-be-unneeded smp_read_barrier_depends() David Howells
2017-10-11 12:56   ` Paul E. McKenney

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20171010155042.GD3521@linux.vnet.ibm.com \
    --to=paulmck@linux.vnet.ibm.com \
    --cc=corbet@lwn.net \
    --cc=dhowells@redhat.com \
    --cc=kuleshovmail@gmail.com \
    --cc=linux-arch@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mark.rutland@arm.com \
    --cc=mingo@kernel.org \
    --cc=peterz@infradead.org \
    --cc=torvalds@linux-foundation.org \
    --cc=will.deacon@arm.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.