Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]


Groups > linux.kernel > #1331574

[PATCH 2/3] random: Make input to output pool balancing per cpu

From Andi Kleen <andi@firstfloor.org>
Newsgroups linux.kernel
Subject [PATCH 2/3] random: Make input to output pool balancing per cpu
Date 2016-02-11 00:10 +0100
Message-ID <r0M3x-pI-33@gated-at.bofh.it> (permalink)
References <r0M3w-pI-7@gated-at.bofh.it>
Organization linux.* mail to news gateway

Show all headers | View raw


From: Andi Kleen <ak@linux.intel.com>

The load balancing from input pool to output pools was
essentially unlocked. Before it didn't matter much because
there were only two choices (blocking and non blocking).

But now with the distributed non blocking pools we have
a lot more pools, and unlocked access of the counters
may systematically deprive some nodes from their deserved
entropy.

Turn the round-robin state into per CPU variables
to avoid any possibility of races. This code already
runs with preemption disabled.

v2: Check for non initialized pools.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 drivers/char/random.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

diff --git a/drivers/char/random.c b/drivers/char/random.c
index e7e02c0..a395f783 100644
--- a/drivers/char/random.c
+++ b/drivers/char/random.c
@@ -774,15 +774,20 @@ retry:
 		if (entropy_bits > random_write_wakeup_bits &&
 		    r->initialized &&
 		    r->entropy_total >= 2*random_read_wakeup_bits) {
-			static struct entropy_store *last = &blocking_pool;
-			static int next_pool = -1;
-			struct entropy_store *other = &blocking_pool;
+			static DEFINE_PER_CPU(struct entropy_store *, lastp) =
+				&blocking_pool;
+			static DEFINE_PER_CPU(int, next_pool);
+			struct entropy_store *other = &blocking_pool, *last;
+			int np;
 
 			/* -1: use blocking pool, 0<=max_node: node nb pool */
-			if (next_pool > -1)
-				other = nonblocking_node_pool[next_pool];
-			if (++next_pool >= num_possible_nodes())
-				next_pool = -1;
+			np = __this_cpu_read(next_pool);
+			if (np > -1 && nonblocking_node_pool)
+				other = nonblocking_node_pool[np];
+			if (++np >= num_possible_nodes())
+				np = -1;
+			__this_cpu_write(next_pool, np);
+			last = __this_cpu_read(lastp);
 			if (other->entropy_count <=
 			    3 * other->poolinfo->poolfracbits / 4)
 				last = other;
@@ -791,6 +796,7 @@ retry:
 				schedule_work(&last->push_work);
 				r->entropy_total = 0;
 			}
+			__this_cpu_write(lastp, last);
 		}
 	}
 }
-- 
2.4.3

Back to linux.kernel | Previous | NextPrevious in thread | Find similar | Unroll thread


Thread

Scalable random patchkit revisited Andi Kleen <andi@firstfloor.org> - 2016-02-11 00:10 +0100
  [PATCH 1/3] Make /dev/urandom scalable Andi Kleen <andi@firstfloor.org> - 2016-02-11 00:10 +0100
  [PATCH 3/3] random: Add pool name to urandom_read trace point Andi Kleen <andi@firstfloor.org> - 2016-02-11 00:10 +0100
  [PATCH 2/3] random: Make input to output pool balancing per cpu Andi Kleen <andi@firstfloor.org> - 2016-02-11 00:10 +0100

csiph-web