[BUG]: when printk too more through serial, cpu up is failed.

[BUG]: when printk too more through serial, cpu up is failed.

[Shuge]

Hi all,
     When the kernel printk too many log, the cpu is failed to come online.
The problem is this:
For example, cpu0 bring up cpu1:

       a. cpu0 call cpu_up:
          cpu_up()
          ->_cpu_up()
             ->__cpu_notify(CPU_UP_PREPARE)
             ->__cpu_up()
                ->boot_secondary()
#       ->wait_for_completion_timeout(&cpu_running, msecs_to_jiffires(1000))
          -> if (!cpu_online(cpu)) {
                   pr_crit("CPU%u: failed to come online\n", cpu);
                   ret = -EIO;
               }
          ->cpu_notify(CPU_ONLINE)

       b. cpu1 enter kernel:
       secondary_start_kernel()
@   ->printk("CPU%u: Booted secondary processor\n", cpu)
*   ->calibrate_delay()
       ->set_cpu_online()
       ->complete(cpu_running)
         ->cpumask_set_cpu()

    While cpu0 run to mark #,  which wait that cpu1 complete 
cpu_running, and set online.
Generally, cpu0 can get it. But if the __log_buf is too large or other 
threads write
it unceasing, then cpu1 come to mark @ or * in this moment. Cpu1 is busy 
outputing
buffer, which cost time more than 1s, and cpu1 have not join in sched, 
so cpu0 wait it timeout.
    By reading printk.c, I found that can_use_console() always return 
true, which be called by
console_trylock_for_printk(). Because, have_callable_console() return 
ture always, if the console
driver set CON_ANYTIME flag. I think that cpu should not output the 
__log_buf in coming online,
even though have_callable_console() is true.

/*
  * Can we actually use the console at this time on this cpu?
  *
  * Console drivers may assume that per-cpu resources have
  * been allocated. So unless they're explicitly marked as
  * being able to cope (CON_ANYTIME) don't call them until
  * this CPU is officially up.
  */
static inline int can_use_console(unsigned int cpu)
{
     return cpu_online(cpu) || have_callable_console();
}

In can_use_console, why not is &&, but ||?

Kernel Version: 3.3.0
Arch: arm

This problem trace:
[ 1722.150286] CPU1: Booted secondary processor
[ 1722.154613] Calibrating delay loop (skipped) already calibrated this CPU
      .........
[ 1725.134687] CPU1: failed to come online
      .........
[ 1728.055547] ------------[ cut here ]------------
[ 1728.060219] kernel BUG at kernel/workqueue.c:3332!
[ 1728.065010] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP
[ 1728.070503] Modules linked in: cdc_ether rtl8150 mcs7830 qf9700 asix 
usbnet mma7660 sw_keyboard gt9xx g2d_33 sunxi_csi1 gc2035 ov5640 
videobuf_dma_contig videobuf_core dc_sunxi(O) pvrsrvkm(O) nand(O)
[ 1728.088474] CPU: 0    Tainted: G           O  (3.3.0 #9)
[ 1728.093796] PC is at trustee_thread+0x24/0x514
[ 1728.098239] LR is at trustee_thread+0x1c/0x514
[ 1728.102683] pc : [<c059c750>]    lr : [<c059c748>]    psr: 20000013
[ 1728.102690] sp : ec05bf68  ip : ec05bf68  fp : ec05bfb4
[ 1728.114161] r10: 00000000  r9 : 00000000  r8 : 00000000
[ 1728.119383] r7 : 00000013  r6 : c059c72c  r5 : 00000001  r4 : e5e162c0
[ 1728.125906] r3 : 00000000  r2 : 00000001  r1 : ec05bf40  r0 : 00000000
[ 1728.132432] Flags: nzCv  IRQs on  FIQs on  Mode SVC_32  ISA ARM  
Segment kernel
[ 1728.139736] Control: 10c53c7d  Table: 6e35c06a  DAC: 00000015
      .........
[ 1728.709566] Process workqueue_trust (pid: 22196, stack limit = 
0xec05a2f8)
[ 1728.716436] Stack: (0xec05bf68 to 0xec05c000)
[ 1728.720796] bf60:                   c05a6c3c c007f98c c07ed100 
c07ed100 c05a788c ed089de0
[ 1728.728973] bf80: e5e162c0 c059c72c 00000013 ed089de0 e5e162c0 
c059c72c 00000013 00000000
[ 1728.737148] bfa0: 00000000 00000000 ec05bff4 ec05bfb8 c00732fc 
c059c738 00000000 00000000
[ 1728.745324] bfc0: e5e162c0 00000000 00000000 00000000 ec05bfd0 
ec05bfd0 00000000 ed089de0
[ 1728.753501] bfe0: c0073260 c0057d50 00000000 ec05bff8 c0057d50 
c007326c 00d851b1 da280100
[ 1728.761669] Backtrace:
[ 1728.764140] [<c059c72c>] (trustee_thread+0x0/0x514) from [<c00732fc>] 
(kthread+0x9c/0xa8)
[ 1728.772321] [<c0073260>] (kthread+0x0/0xa8) from [<c0057d50>] 
(do_exit+0x0/0x764)
[ 1728.779797]  r6:c0057d50 r5:c0073260 r4:ed089de0
[ 1728.784457] Code: e1a04000 ebf36dbd e1550000 0a000000 (e7f001f2)
[ 1729.130494] ---[ end trace 3deba094f67f9e11 ]---
[ 1729.140018] Kernel panic - not syncing: Fatal exception
[ 1730.355119] SMP: failed to stop secondary CPUs

Re: [BUG]: when printk too more through serial, cpu up is failed.

[Greg KH]

On Thu, Mar 14, 2013 at 09:51:34PM +0800, Shuge wrote:
> Hi all,
>     When the kernel printk too many log, the cpu is failed to come online.
> The problem is this:
> For example, cpu0 bring up cpu1:
> 
>       a. cpu0 call cpu_up:
>          cpu_up()
>          ->_cpu_up()
>             ->__cpu_notify(CPU_UP_PREPARE)
>             ->__cpu_up()
>                ->boot_secondary()
> #       ->wait_for_completion_timeout(&cpu_running, msecs_to_jiffires(1000))
>          -> if (!cpu_online(cpu)) {
>                   pr_crit("CPU%u: failed to come online\n", cpu);
>                   ret = -EIO;
>               }
>          ->cpu_notify(CPU_ONLINE)
> 
>       b. cpu1 enter kernel:
>       secondary_start_kernel()
> @   ->printk("CPU%u: Booted secondary processor\n", cpu)
> *   ->calibrate_delay()
>       ->set_cpu_online()
>       ->complete(cpu_running)
>         ->cpumask_set_cpu()
> 
>    While cpu0 run to mark #,  which wait that cpu1 complete
> cpu_running, and set online.
> Generally, cpu0 can get it. But if the __log_buf is too large or
> other threads write
> it unceasing, then cpu1 come to mark @ or * in this moment. Cpu1 is
> busy outputing
> buffer, which cost time more than 1s, and cpu1 have not join in
> sched, so cpu0 wait it timeout.
>    By reading printk.c, I found that can_use_console() always return
> true, which be called by
> console_trylock_for_printk(). Because, have_callable_console()
> return ture always, if the console
> driver set CON_ANYTIME flag. I think that cpu should not output the
> __log_buf in coming online,
> even though have_callable_console() is true.
> 
> /*
>  * Can we actually use the console at this time on this cpu?
>  *
>  * Console drivers may assume that per-cpu resources have
>  * been allocated. So unless they're explicitly marked as
>  * being able to cope (CON_ANYTIME) don't call them until
>  * this CPU is officially up.
>  */
> static inline int can_use_console(unsigned int cpu)
> {
>     return cpu_online(cpu) || have_callable_console();
> }
> 
> In can_use_console, why not is &&, but ||?
> 
> Kernel Version: 3.3.0

Why such an old and obsolete kernel version?  Please try this on 3.8,
lots of work have gone into the printk area that should have solved this
issue.

greg k-h

Re: [BUG]: when printk too more through serial, cpu up is failed.

[Shuge]

于 2013年03月14日 22:05, Greg KH 写道:
> On Thu, Mar 14, 2013 at 09:51:34PM +0800, Shuge wrote:
>> Hi all,
>>      When the kernel printk too many log, the cpu is failed to come online.
>> The problem is this:
>> For example, cpu0 bring up cpu1:
>>
>>        a. cpu0 call cpu_up:
>>           cpu_up()
>>           ->_cpu_up()
>>              ->__cpu_notify(CPU_UP_PREPARE)
>>              ->__cpu_up()
>>                 ->boot_secondary()
>> #       ->wait_for_completion_timeout(&cpu_running, msecs_to_jiffires(1000))
>>           -> if (!cpu_online(cpu)) {
>>                    pr_crit("CPU%u: failed to come online\n", cpu);
>>                    ret = -EIO;
>>                }
>>           ->cpu_notify(CPU_ONLINE)
>>
>>        b. cpu1 enter kernel:
>>        secondary_start_kernel()
>> @   ->printk("CPU%u: Booted secondary processor\n", cpu)
>> *   ->calibrate_delay()
>>        ->set_cpu_online()
>>        ->complete(cpu_running)
>>          ->cpumask_set_cpu()
>>
>>     While cpu0 run to mark #,  which wait that cpu1 complete
>> cpu_running, and set online.
>> Generally, cpu0 can get it. But if the __log_buf is too large or
>> other threads write
>> it unceasing, then cpu1 come to mark @ or * in this moment. Cpu1 is
>> busy outputing
>> buffer, which cost time more than 1s, and cpu1 have not join in
>> sched, so cpu0 wait it timeout.
>>     By reading printk.c, I found that can_use_console() always return
>> true, which be called by
>> console_trylock_for_printk(). Because, have_callable_console()
>> return ture always, if the console
>> driver set CON_ANYTIME flag. I think that cpu should not output the
>> __log_buf in coming online,
>> even though have_callable_console() is true.
>>
>> /*
>>   * Can we actually use the console at this time on this cpu?
>>   *
>>   * Console drivers may assume that per-cpu resources have
>>   * been allocated. So unless they're explicitly marked as
>>   * being able to cope (CON_ANYTIME) don't call them until
>>   * this CPU is officially up.
>>   */
>> static inline int can_use_console(unsigned int cpu)
>> {
>>      return cpu_online(cpu) || have_callable_console();
>> }
>>
>> In can_use_console, why not is &&, but ||?
>>
>> Kernel Version: 3.3.0
> Why such an old and obsolete kernel version?  Please try this on 3.8,
> lots of work have gone into the printk area that should have solved this
> issue.
>
> greg k-h

I saw the printk.c in version 3.9, it still check 
console_trylock_for_printk() to decide to call console_unlock. In 
vprintk_emit(), cpu1 also have the opportunity to execute 
console_unlock() at coming online time.
Once cpu which is coming online can output buffer, nothing can interrupt 
it until buffer is empty.But we can't ensure that none always write the 
__log_buf. It is danger! I think, the solution is that we should prevent 
to use console at coming online.

Re: [BUG]: when printk too more through serial, cpu up is failed.

[Greg KH]

On Fri, Mar 15, 2013 at 09:26:26AM +0800, Shuge wrote:
> 于 2013年03月14日 22:05, Greg KH 写道:
> >On Thu, Mar 14, 2013 at 09:51:34PM +0800, Shuge wrote:
> >>Hi all,
> >>     When the kernel printk too many log, the cpu is failed to come online.
> >>The problem is this:
> >>For example, cpu0 bring up cpu1:
> >>
> >>       a. cpu0 call cpu_up:
> >>          cpu_up()
> >>          ->_cpu_up()
> >>             ->__cpu_notify(CPU_UP_PREPARE)
> >>             ->__cpu_up()
> >>                ->boot_secondary()
> >>#       ->wait_for_completion_timeout(&cpu_running, msecs_to_jiffires(1000))
> >>          -> if (!cpu_online(cpu)) {
> >>                   pr_crit("CPU%u: failed to come online\n", cpu);
> >>                   ret = -EIO;
> >>               }
> >>          ->cpu_notify(CPU_ONLINE)
> >>
> >>       b. cpu1 enter kernel:
> >>       secondary_start_kernel()
> >>@   ->printk("CPU%u: Booted secondary processor\n", cpu)
> >>*   ->calibrate_delay()
> >>       ->set_cpu_online()
> >>       ->complete(cpu_running)
> >>         ->cpumask_set_cpu()
> >>
> >>    While cpu0 run to mark #,  which wait that cpu1 complete
> >>cpu_running, and set online.
> >>Generally, cpu0 can get it. But if the __log_buf is too large or
> >>other threads write
> >>it unceasing, then cpu1 come to mark @ or * in this moment. Cpu1 is
> >>busy outputing
> >>buffer, which cost time more than 1s, and cpu1 have not join in
> >>sched, so cpu0 wait it timeout.
> >>    By reading printk.c, I found that can_use_console() always return
> >>true, which be called by
> >>console_trylock_for_printk(). Because, have_callable_console()
> >>return ture always, if the console
> >>driver set CON_ANYTIME flag. I think that cpu should not output the
> >>__log_buf in coming online,
> >>even though have_callable_console() is true.
> >>
> >>/*
> >>  * Can we actually use the console at this time on this cpu?
> >>  *
> >>  * Console drivers may assume that per-cpu resources have
> >>  * been allocated. So unless they're explicitly marked as
> >>  * being able to cope (CON_ANYTIME) don't call them until
> >>  * this CPU is officially up.
> >>  */
> >>static inline int can_use_console(unsigned int cpu)
> >>{
> >>     return cpu_online(cpu) || have_callable_console();
> >>}
> >>
> >>In can_use_console, why not is &&, but ||?
> >>
> >>Kernel Version: 3.3.0
> >Why such an old and obsolete kernel version?  Please try this on 3.8,
> >lots of work have gone into the printk area that should have solved this
> >issue.
> >
> >greg k-h
> 
> I saw the printk.c in version 3.9, it still check
> console_trylock_for_printk() to decide to call console_unlock. In
> vprintk_emit(), cpu1 also have the opportunity to execute
> console_unlock() at coming online time.
> Once cpu which is coming online can output buffer, nothing can
> interrupt it until buffer is empty.But we can't ensure that none
> always write the __log_buf. It is danger!

Do you really hit this with a real system?  Is your cpu just really slow
in initializing?  What is the actual time it takes?

> I think, the solution is that we should prevent to use console at
> coming online.

Ok, what would be your proposed patch to solve this?

greg k-h