Mr. Rostedt,

I sent you trace.data last week and i looked into it in kernelshark meanwhile.I
found it took more than 20us in _raw_spin_unlock_irq.

I looked into the code, the call stack is like that:

hrtimer_try_to_cancel -->
unlock_hrtimer_base -->
raw_spin_unlock_irqrestore -->
do_raw_spin_unlock -->
arch_spin_unlock

or

finish_task_switch -->
finish_lock_switch -->
raw_spin_unlock_irq -->
do_raw_spin_unlock -->
arch_spin_unlock

Normally, it's working fine, however it will be reproduced when a app is
being opened like firefox.

Have you ever had the same problem like latency from a spinlock?Could
you please share me some of
your experience in this case? many thanks.

PS:arch_spin_unlock in arch/arm64/include/asm/spinlock.h

#define arch_spin_lock_flags(lock, flags) arch_spin_lock(lock)

static inline void arch_spin_lock(arch_spinlock_t *lock)
{
        unsigned int tmp;
        arch_spinlock_t lockval, newval;

        asm volatile(
        /* Atomically increment the next ticket. */
        ARM64_LSE_ATOMIC_INSN(
        /* LL/SC */
"       prfm    pstl1strm, %3\n"
"1:     ldaxr   %w0, %3\n"
"       add     %w1, %w0, %w5\n"
"       stxr    %w2, %w1, %3\n"
"       cbnz    %w2, 1b\n",
        /* LSE atomics */
"       mov     %w2, %w5\n"
"       ldadda  %w2, %w0, %3\n"
        __nops(3)
        )

        /* Did we get the lock? */
"       eor     %w1, %w0, %w0, ror #16\n"
"       cbz     %w1, 3f\n"
        /*
         * No: spin on the owner. Send a local event to avoid missing an
         * unlock before the exclusive load.
         */
"       sevl\n"
"2:     wfe\n"
"       ldaxrh  %w2, %4\n"
"       eor     %w1, %w2, %w0, lsr #16\n"
"       cbnz    %w1, 2b\n"
        /* We got the lock. Critical section starts here. */
"3:"
        : "=&r" (lockval), "=&r" (newval), "=&r" (tmp), "+Q" (*lock)
        : "Q" (lock->owner), "I" (1 << TICKET_SHIFT)
        : "memory");
}

static inline int arch_spin_trylock(arch_spinlock_t *lock)
{
        unsigned int tmp;
        arch_spinlock_t lockval;

        asm volatile(ARM64_LSE_ATOMIC_INSN(
        /* LL/SC */
        "       prfm    pstl1strm, %2\n"
        "1:     ldaxr   %w0, %2\n"
        "       eor     %w1, %w0, %w0, ror #16\n"
        "       cbnz    %w1, 2f\n"
        "       add     %w0, %w0, %3\n"
        "       stxr    %w1, %w0, %2\n"
        "       cbnz    %w1, 1b\n"
        "2:",
        /* LSE atomics */
        "       ldr     %w0, %2\n"
        "       eor     %w1, %w0, %w0, ror #16\n"
        "       cbnz    %w1, 1f\n"
        "       add     %w1, %w0, %3\n"
        "       casa    %w0, %w1, %2\n"
        "       and     %w1, %w1, #0xffff\n"
        "       eor     %w1, %w1, %w0, lsr #16\n"
        "1:")
        : "=&r" (lockval), "=&r" (tmp), "+Q" (*lock)
        : "I" (1 << TICKET_SHIFT)
        : "memory");

        return !tmp;
}

static inline void arch_spin_unlock(arch_spinlock_t *lock)
{
        unsigned long tmp;

        asm volatile(ARM64_LSE_ATOMIC_INSN(
        /* LL/SC */
        "       ldrh    %w1, %0\n"
        "       add     %w1, %w1, #1\n"
        "       stlrh   %w1, %0",
        /* LSE atomics */
        "       mov     %w1, #1\n"
        "       staddlh %w1, %0\n"
        __nops(1))
        : "=Q" (lock->owner), "=&r" (tmp)
        :
        : "memory");
}

BR

chensong

On Wed, May 13, 2020 at 9:27 AM chen song [off-list ref] wrote:

Mr. Rostedt,

The compressed trace.dat is in the attachment with trace-cmd report
followed with "-l". Thanks again for helping me with this. Please let
me know if there is anything i can do for further investigation.

BR

/Song

On Tue, May 12, 2020 at 9:14 PM Steven Rostedt [off-list ref] wrote:

quoted

On Tue, 12 May 2020 18:56:10 +0800
chen song [off-list ref] wrote:

quoted

Dear Rostedt,

This is chensong who was asking question in #linux-rt channel. I
upgrated my rt-tests to the lastest version (cyclictest V 1.80) and
trace-cmd as well(trace-cmd version 2.9.dev).

I ran commands like these:
trace-cmd start -e irq -e sched -e timer -p function -l '*spin*';
cyclictest -p 90 -m -c 0 -i 100 -h 100 -q -l 1000000 -b 80 --tracemark
trace-cmd extract
trace-cmd report

BTW, trace-cmd report -l, will show the preemption numbers.

quoted

The log of last cycle before cyclictest reached the threshold and quit
is like that:

Your email client line wrapped everything, and it's unreadable.

Can you send my the trace.dat file. Compress it first.

Thanks,

-- Steve

quoted

<idle>-0     [000]   130.330659: irq_handler_entry:    irq=4 name=arch_timer
<idle>-0     [000]   130.330665: function:
_raw_spin_lock_irqsave
<idle>-0     [000]   130.330666: hrtimer_cancel:
hrtimer=0xffffff801234bdc8
<idle>-0     [000]   130.330667: function:
_raw_spin_unlock_irqrestore
<idle>-0     [000]   130.330668: hrtimer_expire_entry:
hrtimer=0xffffff801234bdc8 now=130207149982 function=hrtimer_wakeup/0x0
<idle>-0     [000]   130.330669: function:
_raw_spin_lock_irqsave
<idle>-0     [000]   130.330682: sched_waking:         comm=cyclictest
pid=1720 prio=9 target_cpu=000
<idle>-0     [000]   130.330695: function:             _raw_spin_lock
<idle>-0     [000]   130.330697: function:             _raw_spin_lock
<idle>-0     [000]   130.330699: sched_wakeup:         cyclictest:1720
[9] success=1 CPU:000
<idle>-0     [000]   130.330699: function:
_raw_spin_unlock_irqrestore
<idle>-0     [000]   130.330700: hrtimer_expire_exit:
hrtimer=0xffffff801234bdc8
<idle>-0     [000]   130.330701: function:             _raw_spin_lock_irq
<idle>-0     [000]   130.330702: function:
_raw_spin_unlock_irqrestore
<idle>-0     [000]   130.330703: irq_handler_exit:     irq=4 ret=handled

It took 44us from irq_handler_entry to irq_handler_exit, compare to
normal case which generally took only around 15us, it produced around
30us extra latency. I also found something wrong like
"_raw_spin_lock_irqsave from 130.330669 to 130.330682" and "sched_waking
from 130.330682 to 130.330695".

PS: cyclictest was running on CPU0 and a firefox was running on CPU1
meanwhile.

I tried to dig deeper, is there any way to find more information, please
let me know, many thanks.

Here is more log for reference:

<idle>-0     [000]   130.330659: irq_handler_entry:    irq=4 name=arch_timer
           <idle>-0     [000]   130.330665: function:
_raw_spin_lock_irqsave
           <idle>-0     [000]   130.330666: hrtimer_cancel:
hrtimer=0xffffff801234bdc8
           <idle>-0     [000]   130.330667: function:
_raw_spin_unlock_irqrestore
           <idle>-0     [000]   130.330668: hrtimer_expire_entry:
hrtimer=0xffffff801234bdc8 now=130207149982 function=hrtimer_wakeup/0x0
           <idle>-0     [000]   130.330669: function:
_raw_spin_lock_irqsave
           <idle>-0     [000]   130.330682: sched_waking:
comm=cyclictest pid=1720 prio=9 target_cpu=000
           <idle>-0     [000]   130.330695: function:
_raw_spin_lock
           <idle>-0     [000]   130.330697: function:
_raw_spin_lock
           <idle>-0     [000]   130.330699: sched_wakeup:
cyclictest:1720 [9] success=1 CPU:000


           <idle>-0     [000]   130.330699: function:
_raw_spin_unlock_irqrestore
           <idle>-0     [000]   130.330700: hrtimer_expire_exit:
hrtimer=0xffffff801234bdc8
           <idle>-0     [000]   130.330701: function:
_raw_spin_lock_irq
           <idle>-0     [000]   130.330702: function:
_raw_spin_unlock_irqrestore
           <idle>-0     [000]   130.330703: irq_handler_exit:     irq=4
ret=handled

          firefox-1722  [001]   130.330705: function:
_raw_spin_lock_irqsave
           <idle>-0     [000]   130.330706: function:
_raw_spin_lock_irqsave
           <idle>-0     [000]   130.330706: hrtimer_cancel:
hrtimer=0xffffffc07ffbe960
          firefox-1722  [001]   130.330706: function:
_raw_spin_lock
          firefox-1722  [001]   130.330707: function:
_raw_spin_unlock_irqrestore
           <idle>-0     [000]   130.330708: function:
_raw_spin_unlock_irqrestore
          firefox-1722  [001]   130.330708: function:
rt_spin_lock
           <idle>-0     [000]   130.330709: function:
_raw_spin_lock_irqsave
           <idle>-0     [000]   130.330709: hrtimer_start:
hrtimer=0xffffffc07ffbe960 function=tick_sched_timer/0x0
expires=130210000000 softexpires=130210000000
           <idle>-0     [000]   130.330711: function:
_raw_spin_unlock_irqrestore
          firefox-1722  [001]   130.330711: function:
rt_spin_unlock
          firefox-1722  [001]   130.330712: function:
_raw_spin_lock_irqsave
           <idle>-0     [000]   130.330712: function:
_raw_spin_lock
          firefox-1722  [001]   130.330712: function:
_raw_spin_lock
          firefox-1722  [001]   130.330713: function:
_raw_spin_unlock_irqrestore
           <idle>-0     [000]   130.330713: sched_switch:
swapper/0:0 [120] R ==> cyclictest:1720 [9]
          firefox-1722  [001]   130.330714: function:
_raw_spin_lock_irqsave

       cyclictest-1720  [000]   130.330715: function:
_raw_spin_unlock_irq
          firefox-1722  [001]   130.330715: function:
_raw_spin_lock
          firefox-1722  [001]   130.330716: function:
_raw_spin_unlock_irqrestore
          firefox-1722  [001]   130.330718: function:
_raw_spin_lock_irqsave
          firefox-1722  [001]   130.330719: function:
_raw_spin_lock
          firefox-1722  [001]   130.330719: function:
_raw_spin_unlock_irqrestore
          firefox-1722  [001]   130.330722: function:
rt_spin_lock
          firefox-1722  [001]   130.330722: function:
_raw_spin_lock_irqsave
          firefox-1722  [001]   130.330723: function:
_raw_spin_lock
          firefox-1722  [001]   130.330724: function:
_raw_spin_unlock_irqrestore
          firefox-1722  [001]   130.330726: function:
rt_spin_lock
          firefox-1722  [001]   130.330727: function:
rt_spin_unlock
          firefox-1722  [001]   130.330728: function:
rt_spin_lock
          firefox-1722  [001]   130.330729: function:
rt_spin_unlock
          firefox-1722  [001]   130.330730: function:
rt_spin_unlock
          firefox-1722  [001]   130.330731: function:
_raw_spin_lock_irqsave
          firefox-1722  [001]   130.330732: function:
_raw_spin_lock
          firefox-1722  [001]   130.330733: function:
_raw_spin_unlock_irqrestore
       cyclictest-1720  [000]   130.330749: print:
tracing_mark_write: hit latency threshold (84 > 80)

BR

/Song

On Thu, May 7, 2020 at 4:56 PM chen song [off-list ref] wrote:

quoted

Dear experts,

I had a 4.14.4 kernel with its rtpatch running on a dual-core arm64
device, the desktop is ubuntu.

In general, the max latency of cyclictest is around 50us in empty load
which is acceptable. However, if i open a firefox duing cyclictest is
running, the max latency becomes higher up to 200us.

I tried to use ftrace to debug this issue,"cyclictest -p 90 -m -c 0 -i
10 -n -h 100 -q -l 10000000 -b 80", "-b 80" means enable ftrace event
and terminate cyclictest when latency is higher than 80us.

Here is a piece of log:

6034916919us+: cpu_idle: state=1 cpu_id=1
6034916981us+: hrtimer_cancel: hrtimer=ffffff8012663dc8
6034916994us+: sched_waking: comm=cyclictest pid=2321 prio=9 target_cpu=001
6034917032us+: cpu_idle: state=4294967295 cpu_id=1
6034917044us+: hrtimer_start: hrtimer=ffffffc07ffd5960
function=tick_sched_timer expires=6034920000000
softexpires=6034920000000 mode=ABS|PINNED
6034917066us+: sys_clock_nanosleep -> 0x0

In this case, I can only see it's stalled in some points but i could'n
find what happens, in other words, the root cause of the latency.

Is there any better way to debug this issue? I guess some of you might
meet the same issue, how did you approach to the root cause? I would
appreciate it so much if you could give me some advice, many thanks.

Best regards,

/Song


============================================================
Here is the log for reference:
   <idle>-0       1d..h1.. 6034916878us : irq_handler_entry: irq=3
name=arch_timer    -- this is a normal routine
   <idle>-0       1d..h2.. 6034916879us : hrtimer_cancel:
hrtimer=ffffff8012663dc8
   <idle>-0       1d..h1.. 6034916879us : hrtimer_expire_entry:
hrtimer=ffffff8012663dc8 function=hrtimer_wakeup now=6034916525493
   <idle>-0       1d..h2.. 6034916880us : sched_waking: comm=cyclictest
pid=2321 prio=9 target_cpu=001
   <idle>-0       1dn.h3.. 6034916882us : sched_wakeup: comm=cyclictest
pid=2321 prio=9 target_cpu=001
   <idle>-0       1dn.h1.. 6034916883us : hrtimer_expire_exit:
hrtimer=ffffff8012663dc8
   <idle>-0       1dn.h1.. 6034916884us : irq_handler_exit: irq=3
ret=handled
   <idle>-0       1.n..1.. 6034916886us : cpu_idle: state=4294967295
cpu_id=1
   <idle>-0       1dn..2.. 6034916888us : hrtimer_cancel:
hrtimer=ffffffc07ffd5960
   <idle>-0       1dn..2.. 6034916890us : hrtimer_start:
hrtimer=ffffffc07ffd5960 function=tick_sched_timer expires=6034920000000
softexpires=6034920000000 mode=ABS|PINNED
   <idle>-0       1dn..1.. 6034916891us : rcu_utilization: Start context
switch
   <idle>-0       1dn..1.. 6034916892us : rcu_utilization: End context
switch
   <idle>-0       1d...2.. 6034916894us : sched_switch:
prev_comm=swapper/1 prev_pid=0 prev_prio=120 prev_state=R ==>
next_comm=cyclictest next_pid=2321 next_prio=9
cyclicte-2321    1....... 6034916897us : kfree:
call_site=ffffff80081501f0 ptr=          (null)
cyclicte-2321    1....... 6034916898us : sys_exit: NR 115 = 0
cyclicte-2321    1....1.. 6034916898us : sys_clock_nanosleep -> 0x0
cyclicte-2321    1....... 6034916902us : sys_enter: NR 115 (1, 1,
7fbd20b448, 0, 0, 7fbd20be00)
cyclicte-2321    1....1.. 6034916903us :
sys_clock_nanosleep(which_clock: 1, flags: 1, rqtp: 7fbd20b448, rmtp: 0)
cyclicte-2321    1....... 6034916904us : hrtimer_init:
hrtimer=ffffff8012663dc8 clockid=CLOCK_MONOTONIC mode=ABS
cyclicte-2321    1d...1.. 6034916906us : hrtimer_start:
hrtimer=ffffff8012663dc8 function=hrtimer_wakeup expires=6034916622448
softexpires=6034916622448 mode=ABS
cyclicte-2321    1d...1.. 6034916907us : rcu_utilization: Start context
switch
cyclicte-2321    1d...1.. 6034916908us : rcu_utilization: End context switch
cyclicte-2321    1d...2.. 6034916911us : sched_switch:
prev_comm=cyclictest prev_pid=2321 prev_prio=9 prev_state=S ==>
next_comm=swapper/1 next_pid=0 next_prio=120
   <idle>-0       1d...1.. 6034916915us : tick_stop: success=1
dependency=NONE
   <idle>-0       1d...2.. 6034916916us : hrtimer_cancel:
hrtimer=ffffffc07ffd5960
   <idle>-0       1d...2.. 6034916917us : hrtimer_start:
hrtimer=ffffffc07ffd5960 function=tick_sched_timer expires=6035010000000
softexpires=6035010000000 mode=ABS|PINNED
   <idle>-0       1d...1.. 6034916919us+: cpu_idle: state=1 cpu_id=1
                                     --(1)
   <idle>-0       1d..h1.. 6034916977us : irq_handler_entry: irq=3
name=arch_timer
   <idle>-0       1d..h2.. 6034916981us+: hrtimer_cancel:
hrtimer=ffffff8012663dc8                          --(2)
   <idle>-0       1d..h1.. 6034916993us : hrtimer_expire_entry:
hrtimer=ffffff8012663dc8 function=hrtimer_wakeup now=6034916628193
   <idle>-0       1d..h2.. 6034916994us+: sched_waking: comm=cyclictest
pid=2321 prio=9 target_cpu=001      --(3)
   <idle>-0       1dn.h3.. 6034917016us : sched_wakeup: comm=cyclictest
pid=2321 prio=9 target_cpu=001
   <idle>-0       1dn.h1.. 6034917016us : hrtimer_expire_exit:
hrtimer=ffffff8012663dc8
   <idle>-0       1dn.h1.. 6034917026us : irq_handler_exit: irq=3
ret=handled
   <idle>-0       1.n..1.. 6034917032us+: cpu_idle: state=4294967295
cpu_id=1                               --(4)
   <idle>-0       1dn..2.. 6034917043us : hrtimer_cancel:
hrtimer=ffffffc07ffd5960
   <idle>-0       1dn..2.. 6034917044us+: hrtimer_start:
hrtimer=ffffffc07ffd5960 function=tick_sched_timer expires=6034920000000
softexpires=6034920000000 mode=ABS|PINNED  --(5)
   <idle>-0       1dn..1.. 6034917056us : rcu_utilization: Start context
switch
   <idle>-0       1dn..1.. 6034917057us : rcu_utilization: End context
switch
   <idle>-0       1d...2.. 6034917062us : sched_switch:
prev_comm=swapper/1 prev_pid=0 prev_prio=120 prev_state=R ==>
next_comm=cyclictest next_pid=2321 next_prio=9
cyclicte-2321    1....... 6034917065us : kfree:
call_site=ffffff80081501f0 ptr=          (null)
cyclicte-2321    1....... 6034917066us : sys_exit: NR 115 = 0
cyclicte-2321    1....1.. 6034917066us+: sys_clock_nanosleep -> 0x0
                                                     --(5)
cyclicte-2321    1....... 6034917142us : sys_enter: NR 64 (5,
7fbd20c500, 1f, 7fbd20bf08, 2, 7fbd20be00)
cyclicte-2321    1....1.. 6034917143us : sys_write(fd: 5, buf:
7fbd20c500, count: 1f)
cyclicte-2321    1....... 6034917151us : tracing_mark_write: hit latency
threshold (91 > 80)
cyclicte-2321    1....... 6034917155us : kfree:
call_site=ffffff80081501f0 ptr=          (null)
cyclicte-2321    1....... 6034917156us : sys_exit: NR 64 = 31
cyclicte-2321    1....1.. 6034917157us : sys_write -> 0x1f
cyclicte-2321    1....... 6034917159us : sys_enter: NR 64 (4, 408650, 1,
7fbd20bf08, 2, 7fbd20be00)
cyclicte-2321    1....1.. 6034917160us : sys_write(fd: 4, buf: 408650,
count: 1)