IRQ Tracing Problem in Linux 2.6.28 Kernel

IRQ Tracing Problem in Linux 2.6.28 Kernel

[Sol Kavy]

The following back trace represents a deadlock in Ubicom's SMP port of 2.6.28 kernel.   I am sure that we are doing something unexpected.  I would appreciate the community's help in understanding what is going wrong.

Thanks in advance for any pointers,

Sol Kavy

Problem:
Ubicom's initial port does not use GENERIC_CLOCKEVENTS.  Instead it uses a periodic timer based on HZ.  The periodic timer calls do_timer() on each tick.

From the arch directory perspective, we are required to hold the xtime_lock before calling do_timer().  The lock is indeed help by cpu 3 as evidenced in the output below.

The call to get_jiffies_64() at the top of the backtrace is attempting to read the jiffies in a reliable fashion.  The caller is required to wait for the xtime_lock not to be held.  Clearly, since we are in  a path that is holding the xtime_lock, this will never make forward progress.

What is unclear to me is why other ports are not seeing the same problem?  

Perhaps it is because most ports now set GENERIC_CLOCKEVENTS which uses an entirely different mechanism for doing things.  I am in the middle of switching the port to use GENERIC_CLOCKEVENTS but would like to understand this failure in more detail.

Any feedback is greatly appreciated,

Sol

Config Flags:
# CONFIG_GENERIC_CLOCKEVENTS is not set
CONFIG_GENERIC_HARDIRQS=y
CONFIG_IRQ_PER_CPU=y
CONFIG_TRACE_IRQFLAGS_SUPPORT=y
# CONFIG_DEBUG_SHIRQ is not set
CONFIG_TRACE_IRQFLAGS=y
CONFIG_IRQSOFF_TRACER=y

State of the lock:
(gdb) p xtime_lock
$5 = {sequence = 47089, lock = {raw_lock = {lock = 1}, magic = 3735899821, 
    owner_cpu = 3, owner = 0x42b01160}}

This is a backtrace from CPU 3:

 (gdb) bt
#0  get_jiffies_64 () at include/linux/seqlock.h:94
#1  0x4044f558 in sched_clock () at kernel/sched_clock.c:40
#2  0x4045f108 in ring_buffer_time_stamp (cpu=3) at kernel/trace/ring_buffer.c:58
#3  0x40464c50 in ftrace_now (cpu=3) at kernel/trace/trace.c:77
#4  0x404656ec in trace_hardirqs_off () at kernel/trace/trace_irqsoff.c:207
#5  0x40413020 in _spin_lock_irqsave (lock=0x3) at kernel/spinlock.c:82
#6  0x40451b2c in clocksource_get_next () at kernel/time/clocksource.c:254
#7  0x3ffd08ac in update_wall_time () at kernel/time/timekeeping.c:182
#8  0x4043bcd8 in do_timer (ticks=0) at kernel/timer.c:1125
#9  0x404169f8 in timer_tick (irq=<value optimized out>, 
    dummy=<value optimized out>) at arch/ubicom32/kernel/time.c:126
#10 0x3ffcefb0 in handle_IRQ_event (irq=<value optimized out>, 
    action=0x406fe6f4) at kernel/irq/handle.c:142
#11 0x3ffcee20 in __do_IRQ (irq=<value optimized out>)
    at kernel/irq/handle.c:239
#12 0x3ffcfcc8 in do_IRQ (irq=47089, regs=<value optimized out>)
    at arch/ubicom32/kernel/irq.c:250
#13 0x4041018c in sys_call_table ()
#14 0x4044f558 in sched_clock () at kernel/sched_clock.c:40
#15 0x00000008 in ?? ()




--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs

Re: IRQ Tracing Problem in Linux 2.6.28 Kernel

[Gedare Bloom]

On Mon, Apr 13, 2009 at 6:00 PM, Sol Kavy <skavy@ubicom.com> wrote:
> The following back trace represents a deadlock in Ubicom's SMP port of 2.6.28 kernel.   I am sure that we are doing something unexpected.  I would appreciate the community's help in understanding what is going wrong.
>
> Thanks in advance for any pointers,
>
> Sol Kavy
>
> Problem:
> Ubicom's initial port does not use GENERIC_CLOCKEVENTS.  Instead it uses a periodic timer based on HZ.  The periodic timer calls do_timer() on each tick.
>
> From the arch directory perspective, we are required to hold the xtime_lock before calling do_timer().  The lock is indeed help by cpu 3 as evidenced in the output below.
>
> The call to get_jiffies_64() at the top of the backtrace is attempting to read the jiffies in a reliable fashion.  The caller is required to wait for the xtime_lock not to be held.  Clearly, since we are in  a path that is holding the xtime_lock, this will never make forward progress.
>
> What is unclear to me is why other ports are not seeing the same problem?
>
> Perhaps it is because most ports now set GENERIC_CLOCKEVENTS which uses an entirely different mechanism for doing things.  I am in the middle of switching the port to use GENERIC_CLOCKEVENTS but would like to understand this failure in more detail.
>
> Any feedback is greatly appreciated,
>
> Sol
>
> Config Flags:
> # CONFIG_GENERIC_CLOCKEVENTS is not set
> CONFIG_GENERIC_HARDIRQS=y
> CONFIG_IRQ_PER_CPU=y
> CONFIG_TRACE_IRQFLAGS_SUPPORT=y
> # CONFIG_DEBUG_SHIRQ is not set
> CONFIG_TRACE_IRQFLAGS=y
> CONFIG_IRQSOFF_TRACER=y
>
> State of the lock:
> (gdb) p xtime_lock
> $5 = {sequence = 47089, lock = {raw_lock = {lock = 1}, magic = 3735899821,
>    owner_cpu = 3, owner = 0x42b01160}}
>
> This is a backtrace from CPU 3:
>
>  (gdb) bt
> #0  get_jiffies_64 () at include/linux/seqlock.h:94
> #1  0x4044f558 in sched_clock () at kernel/sched_clock.c:40
> #2  0x4045f108 in ring_buffer_time_stamp (cpu=3) at kernel/trace/ring_buffer.c:58
> #3  0x40464c50 in ftrace_now (cpu=3) at kernel/trace/trace.c:77
> #4  0x404656ec in trace_hardirqs_off () at kernel/trace/trace_irqsoff.c:207
> #5  0x40413020 in _spin_lock_irqsave (lock=0x3) at kernel/spinlock.c:82
> #6  0x40451b2c in clocksource_get_next () at kernel/time/clocksource.c:254
> #7  0x3ffd08ac in update_wall_time () at kernel/time/timekeeping.c:182
> #8  0x4043bcd8 in do_timer (ticks=0) at kernel/timer.c:1125
> #9  0x404169f8 in timer_tick (irq=<value optimized out>,
>     dummy=<value optimized out>) at arch/ubicom32/kernel/time.c:126
> #10 0x3ffcefb0 in handle_IRQ_event (irq=<value optimized out>,
>     action=0x406fe6f4) at kernel/irq/handle.c:142
> #11 0x3ffcee20 in __do_IRQ (irq=<value optimized out>)
>     at kernel/irq/handle.c:239
> #12 0x3ffcfcc8 in do_IRQ (irq=47089, regs=<value optimized out>)
>     at arch/ubicom32/kernel/irq.c:250
> #13 0x4041018c in sys_call_table ()
> #14 0x4044f558 in sched_clock () at kernel/sched_clock.c:40
> #15 0x00000008 in ?? ()
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.linux-learn.org/faqs
>

Hi Sol,

Caveat: I'm not familiar with porting Linux :)

get_jiffies_64 does attempt to acquire xtime_lock, but since you know
the lock is already held in do_timer, can you simply read the value
directly?  You'll just have to make sure every code path through
sched_clock() has xtime_lock...

Here is an example from the lxr that might be relevant:
http://lxr.linux.no/linux+v2.6.29/arch/frv/kernel/time.c#L151

Gedare Bloom
--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs

RE: IRQ Tracing Problem in Linux 2.6.28 Kernel

[Sol Kavy]

Every arch's timer interrupt does the following:

	write_seqlock(&xtime_lock);
	do_timer(1);
	write_sequnlock(&xtime_lock);

This is required because the first thing do_timer() does is increment jiffies_64.

	void do_timer(unsigned long ticks)
	{
		jiffies_64 += ticks;
		update_times(ticks);
	}

The frames 8-0 in the backtrace below are in Linux main code (which as a port, we don't want to change).   The arch calls do_timer() to advance time (and must do so with the lock held).  However, if the Linux code is then going to eventually call get_jiffies_64(), this leads to a deadlock.  The only time that get_jiffies_64() is called is when you have selected CONFIG_IRQSOFF_TRACER and CONFIG_TRACE_IRQFLAGS.   

My next step is to build this on a x86 and try to understand why that port does not run into the same problem.

Sol


-----Original Message-----
From: Gedare Bloom [mailto:gedare@gwmail.gwu.edu] 
Sent: Tuesday, April 14, 2009 5:52 AM
To: Sol Kavy
Cc: linux-newbie@vger.kernel.org
Subject: Re: IRQ Tracing Problem in Linux 2.6.28 Kernel

On Mon, Apr 13, 2009 at 6:00 PM, Sol Kavy <skavy@ubicom.com> wrote:
> The following back trace represents a deadlock in Ubicom's SMP port of 2.6.28 kernel.   I am sure that we are doing something unexpected.  I would appreciate the community's help in understanding what is going wrong.
>
> Thanks in advance for any pointers,
>
> Sol Kavy
>
> Problem:
> Ubicom's initial port does not use GENERIC_CLOCKEVENTS.  Instead it uses a periodic timer based on HZ.  The periodic timer calls do_timer() on each tick.
>
> From the arch directory perspective, we are required to hold the xtime_lock before calling do_timer().  The lock is indeed help by cpu 3 as evidenced in the output below.
>
> The call to get_jiffies_64() at the top of the backtrace is attempting to read the jiffies in a reliable fashion.  The caller is required to wait for the xtime_lock not to be held.  Clearly, since we are in  a path that is holding the xtime_lock, this will never make forward progress.
>
> What is unclear to me is why other ports are not seeing the same problem?
>
> Perhaps it is because most ports now set GENERIC_CLOCKEVENTS which uses an entirely different mechanism for doing things.  I am in the middle of switching the port to use GENERIC_CLOCKEVENTS but would like to understand this failure in more detail.
>
> Any feedback is greatly appreciated,
>
> Sol
>
> Config Flags:
> # CONFIG_GENERIC_CLOCKEVENTS is not set
> CONFIG_GENERIC_HARDIRQS=y
> CONFIG_IRQ_PER_CPU=y
> CONFIG_TRACE_IRQFLAGS_SUPPORT=y
> # CONFIG_DEBUG_SHIRQ is not set
> CONFIG_TRACE_IRQFLAGS=y
> CONFIG_IRQSOFF_TRACER=y
>
> State of the lock:
> (gdb) p xtime_lock
> $5 = {sequence = 47089, lock = {raw_lock = {lock = 1}, magic = 3735899821,
>    owner_cpu = 3, owner = 0x42b01160}}
>
> This is a backtrace from CPU 3:
>
>  (gdb) bt
> #0  get_jiffies_64 () at include/linux/seqlock.h:94
> #1  0x4044f558 in sched_clock () at kernel/sched_clock.c:40
> #2  0x4045f108 in ring_buffer_time_stamp (cpu=3) at kernel/trace/ring_buffer.c:58
> #3  0x40464c50 in ftrace_now (cpu=3) at kernel/trace/trace.c:77
> #4  0x404656ec in trace_hardirqs_off () at kernel/trace/trace_irqsoff.c:207
> #5  0x40413020 in _spin_lock_irqsave (lock=0x3) at kernel/spinlock.c:82
> #6  0x40451b2c in clocksource_get_next () at kernel/time/clocksource.c:254
> #7  0x3ffd08ac in update_wall_time () at kernel/time/timekeeping.c:182
> #8  0x4043bcd8 in do_timer (ticks=0) at kernel/timer.c:1125
> #9  0x404169f8 in timer_tick (irq=<value optimized out>,
>     dummy=<value optimized out>) at arch/ubicom32/kernel/time.c:126
> #10 0x3ffcefb0 in handle_IRQ_event (irq=<value optimized out>,
>     action=0x406fe6f4) at kernel/irq/handle.c:142
> #11 0x3ffcee20 in __do_IRQ (irq=<value optimized out>)
>     at kernel/irq/handle.c:239
> #12 0x3ffcfcc8 in do_IRQ (irq=47089, regs=<value optimized out>)
>     at arch/ubicom32/kernel/irq.c:250
> #13 0x4041018c in sys_call_table ()
> #14 0x4044f558 in sched_clock () at kernel/sched_clock.c:40
> #15 0x00000008 in ?? ()
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.linux-learn.org/faqs
>

Hi Sol,

Caveat: I'm not familiar with porting Linux :)

get_jiffies_64 does attempt to acquire xtime_lock, but since you know
the lock is already held in do_timer, can you simply read the value
directly?  You'll just have to make sure every code path through
sched_clock() has xtime_lock...

Here is an example from the lxr that might be relevant:
http://lxr.linux.no/linux+v2.6.29/arch/frv/kernel/time.c#L151

Gedare Bloom


--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs

RE: IRQ Tracing Problem in Linux 2.6.28 Kernel

[Sol Kavy]

I have discovered why other architectures do not have the same problem.   The back trace indicates a real defect (i.e. deadlock) in the generic code.  

Most architectures override sched_clock() with their own version.  Kernel/sched_clock.c:38 is a weak alias that will be overridden if an arch directory supplies its own.

Most of the arch directories provide an implementation that directly access the jiffies_64 variable "without" acquiring the xtime_lock.  

Some of the implementations provide a "poor" implementation in that the value of the jiffies_64 during a 32 rollover is not taken into account.  If sched_clock() is to be called while holding xtime_lock, the core implementation should be modified not to call get_jiffies_64() (which requires the xlock) but to use something like the following:

unsigned long long sched_clock(void)
{
	unsigned long long my_jiffies;
	unsigned long jiffies_top;
	unsigned long jiffies_bottom;

	do {
		jiffies_top = jiffies_64 >> 32;
		jiffies_bottom = jiffies_64 & 0xffffffff;
	} while(unlikely(jiffies_top != (unsigned long)(jiffies_64 >> 32))); 

	my_jiffies = ((unsigned long long)jiffies_top << 32) | (jiffies_bottom);
	return (my_jiffies - INITIAL_JIFFIES) * (NSEC_PER_SEC / HZ);
}

-----Original Message-----
From: linux-newbie-owner@vger.kernel.org [mailto:linux-newbie-owner@vger.kernel.org] On Behalf Of Sol Kavy
Sent: Monday, April 13, 2009 3:01 PM
To: linux-newbie@vger.kernel.org
Subject: IRQ Tracing Problem in Linux 2.6.28 Kernel

The following back trace represents a deadlock in Ubicom's SMP port of 2.6.28 kernel.   I am sure that we are doing something unexpected.  I would appreciate the community's help in understanding what is going wrong.

Thanks in advance for any pointers,

Sol Kavy

Problem:
Ubicom's initial port does not use GENERIC_CLOCKEVENTS.  Instead it uses a periodic timer based on HZ.  The periodic timer calls do_timer() on each tick.

From the arch directory perspective, we are required to hold the xtime_lock before calling do_timer().  The lock is indeed help by cpu 3 as evidenced in the output below.

The call to get_jiffies_64() at the top of the backtrace is attempting to read the jiffies in a reliable fashion.  The caller is required to wait for the xtime_lock not to be held.  Clearly, since we are in  a path that is holding the xtime_lock, this will never make forward progress.

What is unclear to me is why other ports are not seeing the same problem?  

Perhaps it is because most ports now set GENERIC_CLOCKEVENTS which uses an entirely different mechanism for doing things.  I am in the middle of switching the port to use GENERIC_CLOCKEVENTS but would like to understand this failure in more detail.

Any feedback is greatly appreciated,

Sol

Config Flags:
# CONFIG_GENERIC_CLOCKEVENTS is not set
CONFIG_GENERIC_HARDIRQS=y
CONFIG_IRQ_PER_CPU=y
CONFIG_TRACE_IRQFLAGS_SUPPORT=y
# CONFIG_DEBUG_SHIRQ is not set
CONFIG_TRACE_IRQFLAGS=y
CONFIG_IRQSOFF_TRACER=y

State of the lock:
(gdb) p xtime_lock
$5 = {sequence = 47089, lock = {raw_lock = {lock = 1}, magic = 3735899821, 
    owner_cpu = 3, owner = 0x42b01160}}

This is a backtrace from CPU 3:

 (gdb) bt
#0  get_jiffies_64 () at include/linux/seqlock.h:94
#1  0x4044f558 in sched_clock () at kernel/sched_clock.c:40
#2  0x4045f108 in ring_buffer_time_stamp (cpu=3) at kernel/trace/ring_buffer.c:58
#3  0x40464c50 in ftrace_now (cpu=3) at kernel/trace/trace.c:77
#4  0x404656ec in trace_hardirqs_off () at kernel/trace/trace_irqsoff.c:207
#5  0x40413020 in _spin_lock_irqsave (lock=0x3) at kernel/spinlock.c:82
#6  0x40451b2c in clocksource_get_next () at kernel/time/clocksource.c:254
#7  0x3ffd08ac in update_wall_time () at kernel/time/timekeeping.c:182
#8  0x4043bcd8 in do_timer (ticks=0) at kernel/timer.c:1125
#9  0x404169f8 in timer_tick (irq=<value optimized out>, 
    dummy=<value optimized out>) at arch/ubicom32/kernel/time.c:126
#10 0x3ffcefb0 in handle_IRQ_event (irq=<value optimized out>, 
    action=0x406fe6f4) at kernel/irq/handle.c:142
#11 0x3ffcee20 in __do_IRQ (irq=<value optimized out>)
    at kernel/irq/handle.c:239
#12 0x3ffcfcc8 in do_IRQ (irq=47089, regs=<value optimized out>)
    at arch/ubicom32/kernel/irq.c:250
#13 0x4041018c in sys_call_table ()
#14 0x4044f558 in sched_clock () at kernel/sched_clock.c:40
#15 0x00000008 in ?? ()




--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs


--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs

Re: IRQ Tracing Problem in Linux 2.6.28 Kernel

[Peter Teoh]

Just my guess......

On Tue, Apr 14, 2009 at 10:58 PM, Sol Kavy <skavy@ubicom.com> wrote:
>
> I have discovered why other architectures do not have the same problem.   The back trace indicates a real defect (i.e. deadlock) in the generic code.
>
> Most architectures override sched_clock() with their own version.  Kernel/sched_clock.c:38 is a weak alias that will be overridden if an arch directory supplies its own.
>
> Most of the arch directories provide an implementation that directly access the jiffies_64 variable "without" acquiring the xtime_lock.
>
> Some of the implementations provide a "poor" implementation in that the value of the jiffies_64 during a 32 rollover is not taken into account.  If sched_clock() is to be called while holding xtime_lock, the core implementation should be modified not to call get_jiffies_64() (which requires the xlock) but to use something like the following:
>
> unsigned long long sched_clock(void)
> {
>        unsigned long long my_jiffies;
>        unsigned long jiffies_top;
>        unsigned long jiffies_bottom;
>
>        do {
>                jiffies_top = jiffies_64 >> 32;
>                jiffies_bottom = jiffies_64 & 0xffffffff;

in general this type of operation is only done when u are in 32bit
mode.   In 64bit mode, u can do it in ONE atomic assembly
instruction.....so no locking needed.

>        } while(unlikely(jiffies_top != (unsigned long)(jiffies_64 >> 32)));
>
>        my_jiffies = ((unsigned long long)jiffies_top << 32) | (jiffies_bottom);
>        return (my_jiffies - INITIAL_JIFFIES) * (NSEC_PER_SEC / HZ);
> }

-- 
Regards,
Peter Teoh
--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs

Re: IRQ Tracing Problem in Linux 2.6.28 Kernel

[Jeffrey Cao]

On 2009-04-13, Sol Kavy <skavy@ubicom.com> wrote:
> The following back trace represents a deadlock in Ubicom's SMP port of 2.6.28 kernel.   I am sure that we are doing something unexpected.  I would appreciate the community's help in understanding what is going wrong.
>
> Thanks in advance for any pointers,
>
> Sol Kavy
>
> Problem:
> Ubicom's initial port does not use GENERIC_CLOCKEVENTS.  Instead it uses a periodic timer based on HZ.  The periodic timer calls do_timer() on each tick.
>
> From the arch directory perspective, we are required to hold the xtime_lock before calling do_timer().  The lock is indeed help by cpu 3 as evidenced in the output below.
>
> The call to get_jiffies_64() at the top of the backtrace is attempting to read the jiffies in a reliable fashion.  The caller is required to wait for the xtime_lock not to be held.  Clearly, since we are in  a path that is holding the xtime_lock, this will never make forward progress.
For x86 arch, function get_jiffies_64() seems not to wait the xtime_lock,
but to do something related to CPU ordering:
get_jiffies_64()
|->read_seqbegin()
   |->smp_rmb()
      |->alternative("lock; addl $0,0(%%esp)", "lfence", X86_FEATURE_XMM2)
I'm not sure if this is the same as to accquire xtime_lock spinlock. Maybe this
is a point you need check.


Jeffrey

--
To unsubscribe from this list: send the line "unsubscribe linux-newbie" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.linux-learn.org/faqs