netdev - [PATCH net-next] gro_cells: avoid using synchronize_rcu() in gro_cells

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Date:   Sat, 19 Feb 2022 19:57:39 -0800
From:   Eric Dumazet <eric.dumazet@...il.com>
To:     "David S . Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>
Cc:     netdev <netdev@...r.kernel.org>,
        Eric Dumazet <edumazet@...gle.com>,
        Eric Dumazet <eric.dumazet@...il.com>
Subject: [PATCH net-next] gro_cells: avoid using synchronize_rcu() in gro_cells_destroy()

From: Eric Dumazet <edumazet@...gle.com>

Another thing making netns dismantles potentially very slow is located
in gro_cells_destroy(),
whenever cleanup_net() has to remove a device using gro_cells framework.

RTNL is not held at this stage, so synchronize_net()
is calling synchronize_rcu():

netdev_run_todo()
 ip_tunnel_dev_free()
  gro_cells_destroy()
   synchronize_net()
    synchronize_rcu() // Ouch.

This patch uses call_rcu(), and gave me a 25x performance improvement
in my tests.

cleanup_net() is no longer blocked ~10 ms per synchronize_rcu()
call.

In the case we could not allocate the memory needed to queue the
deferred free, use synchronize_rcu_expedited()

Signed-off-by: Eric Dumazet <edumazet@...gle.com>
---
 net/core/gro_cells.c | 36 +++++++++++++++++++++++++++++++-----
 1 file changed, 31 insertions(+), 5 deletions(-)

diff --git a/net/core/gro_cells.c b/net/core/gro_cells.c
index 6eb2e5ec2c5068e1d798557e55d084b785187a9b..46fa7d93fd9696755efd56b72731f08e821042e1 100644
--- a/net/core/gro_cells.c
+++ b/net/core/gro_cells.c
@@ -89,8 +89,23 @@ int gro_cells_init(struct gro_cells *gcells, struct net_device *dev)
 }
 EXPORT_SYMBOL(gro_cells_init);
 
+struct percpu_free_defer {
+	struct rcu_head rcu;
+	void __percpu	*ptr;
+};
+
+void percpu_free_defer_callback(struct rcu_head *head)
+{
+	struct percpu_free_defer *defer;
+
+	defer = container_of(head, struct percpu_free_defer, rcu);
+	free_percpu(defer->ptr);
+	kfree(defer);
+}
+
 void gro_cells_destroy(struct gro_cells *gcells)
 {
+	struct percpu_free_defer *defer;
 	int i;
 
 	if (!gcells->cells)
@@ -102,12 +117,23 @@ void gro_cells_destroy(struct gro_cells *gcells)
 		__netif_napi_del(&cell->napi);
 		__skb_queue_purge(&cell->napi_skbs);
 	}
-	/* This barrier is needed because netpoll could access dev->napi_list
-	 * under rcu protection.
+	/* We need to observe an rcu grace period before freeing ->cells,
+	 * because netpoll could access dev->napi_list under rcu protection.
+	 * Try hard using call_rcu() instead of synchronize_rcu(),
+	 * because we might be called from cleanup_net(), and we
+	 * definitely do not want to block this critical task.
 	 */
-	synchronize_net();
-
-	free_percpu(gcells->cells);
+	defer = kmalloc(sizeof(*defer), GFP_KERNEL | __GFP_NOWARN);
+	if (likely(defer)) {
+		defer->ptr = gcells->cells;
+		call_rcu(&defer->rcu, percpu_free_defer_callback);
+	} else {
+		/* We do not hold RTNL at this point, synchronize_net()
+		 * would not be able to expedite this sync.
+		*/
+		synchronize_rcu_expedited();
+		free_percpu(gcells->cells);
+	}
 	gcells->cells = NULL;
 }
 EXPORT_SYMBOL(gro_cells_destroy);
-- 
2.35.1.473.g83b2b277ed-goog