rt-patches/0020-net-sched-Use-msleep-instead-of-yield.patch

   1 From 60758036b88c410553d86192794e49df643b788d Mon Sep 17 00:00:00 2001
   2 From: Marc Kleine-Budde <mkl@pengutronix.de>
   3 Date: Wed, 5 Mar 2014 00:49:47 +0100
   4 Subject: [PATCH 020/366] net: sched: Use msleep() instead of yield()
   5
   6 On PREEMPT_RT enabled systems the interrupt handler run as threads at prio 50
   7 (by default). If a high priority userspace process tries to shut down a busy
   8 network interface it might spin in a yield loop waiting for the device to
   9 become idle. With the interrupt thread having a lower priority than the
  10 looping process it might never be scheduled and so result in a deadlock on UP
  11 systems.
  12
  13 With Magic SysRq the following backtrace can be produced:
  14
  15 > test_app R running      0   174    168 0x00000000
  16 > [<c02c7070>] (__schedule+0x220/0x3fc) from [<c02c7870>] (preempt_schedule_irq+0x48/0x80)
  17 > [<c02c7870>] (preempt_schedule_irq+0x48/0x80) from [<c0008fa8>] (svc_preempt+0x8/0x20)
  18 > [<c0008fa8>] (svc_preempt+0x8/0x20) from [<c001a984>] (local_bh_enable+0x18/0x88)
  19 > [<c001a984>] (local_bh_enable+0x18/0x88) from [<c025316c>] (dev_deactivate_many+0x220/0x264)
  20 > [<c025316c>] (dev_deactivate_many+0x220/0x264) from [<c023be04>] (__dev_close_many+0x64/0xd4)
  21 > [<c023be04>] (__dev_close_many+0x64/0xd4) from [<c023be9c>] (__dev_close+0x28/0x3c)
  22 > [<c023be9c>] (__dev_close+0x28/0x3c) from [<c023f7f0>] (__dev_change_flags+0x88/0x130)
  23 > [<c023f7f0>] (__dev_change_flags+0x88/0x130) from [<c023f904>] (dev_change_flags+0x10/0x48)
  24 > [<c023f904>] (dev_change_flags+0x10/0x48) from [<c024c140>] (do_setlink+0x370/0x7ec)
  25 > [<c024c140>] (do_setlink+0x370/0x7ec) from [<c024d2f0>] (rtnl_newlink+0x2b4/0x450)
  26 > [<c024d2f0>] (rtnl_newlink+0x2b4/0x450) from [<c024cfa0>] (rtnetlink_rcv_msg+0x158/0x1f4)
  27 > [<c024cfa0>] (rtnetlink_rcv_msg+0x158/0x1f4) from [<c0256740>] (netlink_rcv_skb+0xac/0xc0)
  28 > [<c0256740>] (netlink_rcv_skb+0xac/0xc0) from [<c024bbd8>] (rtnetlink_rcv+0x18/0x24)
  29 > [<c024bbd8>] (rtnetlink_rcv+0x18/0x24) from [<c02561b8>] (netlink_unicast+0x13c/0x198)
  30 > [<c02561b8>] (netlink_unicast+0x13c/0x198) from [<c025651c>] (netlink_sendmsg+0x264/0x2e0)
  31 > [<c025651c>] (netlink_sendmsg+0x264/0x2e0) from [<c022af98>] (sock_sendmsg+0x78/0x98)
  32 > [<c022af98>] (sock_sendmsg+0x78/0x98) from [<c022bb50>] (___sys_sendmsg.part.25+0x268/0x278)
  33 > [<c022bb50>] (___sys_sendmsg.part.25+0x268/0x278) from [<c022cf08>] (__sys_sendmsg+0x48/0x78)
  34 > [<c022cf08>] (__sys_sendmsg+0x48/0x78) from [<c0009320>] (ret_fast_syscall+0x0/0x2c)
  35
  36 This patch works around the problem by replacing yield() by msleep(1), giving
  37 the interrupt thread time to finish, similar to other changes contained in the
  38 rt patch set. Using wait_for_completion() instead would probably be a better
  39 solution.
  40
  41 Signed-off-by: Marc Kleine-Budde <mkl@pengutronix.de>
  42 Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
  43 ---
  44  net/sched/sch_generic.c | 2 +-
  45  1 file changed, 1 insertion(+), 1 deletion(-)
  46
  47 diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
  48 index 2704a13..3f02edc 100644
  49 --- a/net/sched/sch_generic.c
  50 +++ b/net/sched/sch_generic.c
  51 @@ -900,7 +900,7 @@ void dev_deactivate_many(struct list_head *head)
  52         /* Wait for outstanding qdisc_run calls. */
  53         list_for_each_entry(dev, head, close_list)
  54                 while (some_qdisc_is_busy(dev))
  55 -                       yield();
  56 +                       msleep(1);
  57  }
  58
  59  void dev_deactivate(struct net_device *dev)
  60 --
  61 1.9.1
  62