Re: [BUG] Cpustall with kcore

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Jan 01, 2009 at 09:19:14PM +0100, Eric Sesterhenn wrote:
> hi,
> 
> when i issue a "cat /proc/kcore > /dev/null", the kernel issues  the
> following warning and i cant kill the cat process anymore.
> 
> [  356.948019] INFO: RCU detected CPU 0 stall (t=2500 jiffies)
> [  356.948019] Pid: 4156, comm: cat Not tainted 2.6.28 #77
> [  356.948019] Call Trace:
> [  356.948019]  [<c05a99f6>] ? printk+0x18/0x1a
> [  356.948019]  [<c016d682>] __rcu_pending+0x62/0x250
> [  356.948019]  [<c014becb>] ? trace_hardirqs_off+0xb/0x10
> [  356.948019]  [<c016d882>] rcu_pending+0x12/0x40
> [  356.948019]  [<c0132c5a>] update_process_times+0x2a/0x60
> [  356.948019]  [<c0149a65>] tick_sched_timer+0x55/0xc0
> [  356.948019]  [<c0149a10>] ? tick_sched_timer+0x0/0xc0
> [  356.948019]  [<c0140ab0>] __run_hrtimer+0x70/0xe0
> [  356.948019]  [<c0149a10>] ? tick_sched_timer+0x0/0xc0
> [  356.948019]  [<c014128d>] hrtimer_interrupt+0xed/0x190
> [  356.948019]  [<c01059cb>] timer_interrupt+0x3b/0x50
> [  356.948019]  [<c016a779>] handle_IRQ_event+0x29/0x60
> [  356.948019]  [<c016c505>] handle_level_irq+0x65/0xe0
> [  356.948019]  [<c016c4a0>] ? handle_level_irq+0x0/0xe0
> [  356.948019]  <IRQ>  [<c014e321>] ?
> trace_hardirqs_on_caller+0x151/0x1c0
> [  356.948019]  [<c0103bac>] ? common_interrupt+0x2c/0x34
> [  356.948019]  [<c018007b>] ? setup_per_zone_pages_min+0x11b/0x270
> [  356.948019]  [<c029fdc2>] ? copy_to_user+0x112/0x120
> [  356.948019]  [<c01ec705>] ? read_kcore+0x385/0x440
> [  356.948019]  [<c01ec380>] ? read_kcore+0x0/0x440
> [  356.948019]  [<c01e38dd>] ? proc_reg_read+0x5d/0x90
> [  356.948019]  [<c01a6129>] ? vfs_read+0x99/0x160
> [  356.948019]  [<c010357f>] ? sysenter_exit+0xf/0x16
> [  356.948019]  [<c01e3880>] ? proc_reg_read+0x0/0x90
> [  356.948019]  [<c01a62ad>] ? sys_read+0x3d/0x70
> [  356.948019]  [<c0103551>] ? sysenter_do_call+0x12/0x31

This message indicates that RCU is no longer getting any response from
CPU 0.  This is most likely because CPU 0 is spinning in an infinite loop
with interrupts disabled -- otherwise CPU 0 would have reported itself
and given a stack trace.

So, do you have any way to collect CPU 0's stack trace?  Kernel debugger
or some such?

> RCU options are:
> # CONFIG_CLASSIC_RCU is not set
> # CONFIG_TREE_RCU is not set
> CONFIG_PREEMPT_RCU=y
> CONFIG_RCU_TRACE=y
> # CONFIG_TREE_RCU_TRACE is not set
> CONFIG_PREEMPT_RCU_TRACE=y
> CONFIG_RCU_TORTURE_TEST=m
> 
> I tried this on another machine, which completely hangs after the cat,
> i cant log in anymore via ssh. This one has
> 
> # CONFIG_CLASSIC_RCU is not set
> CONFIG_TREE_RCU=y
> # CONFIG_PREEMPT_RCU is not set
> CONFIG_RCU_TRACE=y
> CONFIG_RCU_FANOUT=32
> # CONFIG_RCU_FANOUT_EXACT is not set
> CONFIG_TREE_RCU_TRACE=y
> # CONFIG_PREEMPT_RCU_TRACE is not set
> CONFIG_RCU_TORTURE_TEST=m
> CONFIG_RCU_CPU_STALL_DETECTOR=y

If I am correct about the CPU going out to lunch with interrupts
disabled, the type of RCU will be irrelevant.  It would nevertheless
be interesting to try Classic RCU with CPU stall detection enabled.

							Thanx, Paul