> On Sep 26, 2022, at 6:32 PM, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Mon, Sep 26, 2022 at 09:02:21PM +0000, Joel Fernandes wrote:
>> On Mon, Sep 26, 2022 at 09:32:44PM +0200, Uladzislau Rezki wrote:
>> [...]
>>>>>> On my KVM machine the boot time is affected:
>>>>>>
>>>>>> <snip>
>>>>>> [ 2.273406] e1000 0000:00:03.0 eth0: Intel(R) PRO/1000 Network Connection
>>>>>> [ 11.945283] e1000 0000:00:03.0 ens3: renamed from eth0
>>>>>> [ 22.165198] sr 1:0:0:0: [sr0] scsi3-mmc drive: 4x/4x cd/rw xa/form2 tray
>>>>>> [ 22.165206] cdrom: Uniform CD-ROM driver Revision: 3.20
>>>>>> [ 32.406981] sr 1:0:0:0: Attached scsi CD-ROM sr0
>>>>>> [ 104.115418] process '/usr/bin/fstype' started with executable stack
>>>>>> [ 104.170142] EXT4-fs (sda1): mounted filesystem with ordered data mode. Quota mode: none.
>>>>>> [ 104.340125] systemd[1]: systemd 241 running in system mode. (+PAM +AUDIT +SELINUX +IMA +APPARMOR +SMACK +SYSVINIT +UTMP +LIBCRYPTSETUP +GCRYPT +GNUTLS +ACL +XZ +LZ4 +SECCOMP +BLKID +ELFUTILS +KMOD -IDN2 +IDN -PCRE2 default-hierarchy=hybrid)
>>>>>> [ 104.340193] systemd[1]: Detected virtualization kvm.
>>>>>> [ 104.340196] systemd[1]: Detected architecture x86-64.
>>>>>> [ 104.359032] systemd[1]: Set hostname to <pc638>.
>>>>>> [ 105.740109] random: crng init done
>>>>>> [ 105.741267] systemd[1]: Reached target Remote File Systems.
>>>>>> <snip>
>>>>>>
>>>>>> 2 - 11 and second delay is between 32 - 104. So there are still users which must
>>>>>> be waiting for "RCU" in a sync way.
>>>>>
>>>>> I was wondering if you can compare boot logs and see which timestamp does the
>>>>> slow down start from. That way, we can narrow down the callback. Also another
>>>>> idea is, add "trace_event=rcu:rcu_callback,rcu:rcu_invoke_callback
>>>>> ftrace_dump_on_oops" to the boot params, and then manually call
>>>>> "tracing_off(); panic();" from the code at the first printk that seems off in
>>>>> your comparison of good vs bad. For example, if "crng init done" timestamp is
>>>>> off, put the "tracing_off(); panic();" there. Then grab the serial console
>>>>> output to see what were the last callbacks that was queued/invoked.
>>>>
>>>> We do seem to be in need of some way to quickly and easily locate the
>>>> callback that needed to be _flush() due to a wakeup.
>>>>
>>> <snip>
>>> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
>>> index aeea9731ef80..fe1146d97f1a 100644
>>> --- a/kernel/workqueue.c
>>> +++ b/kernel/workqueue.c
>>> @@ -1771,7 +1771,7 @@ bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
>>>
>>> if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
>>> rwork->wq = wq;
>>> - call_rcu(&rwork->rcu, rcu_work_rcufn);
>>> + call_rcu_flush(&rwork->rcu, rcu_work_rcufn);
>>> return true;
>>> }
>>>
>>> <snip>
>>>
>>> ?
>>>
>>> But it does not fully solve my boot-up issue. Will debug tomorrow further.
>>
>> Ah, but at least its progress, thanks. Could you send me a patch to include
>> in the next revision with details of this?
>>
>>>> Might one more proactive approach be to use Coccinelle to locate such
>>>> callback functions? We might not want -all- callbacks that do wakeups
>>>> to use call_rcu_flush(), but knowing which are which should speed up
>>>> slow-boot debugging by quite a bit.
>>>>
>>>> Or is there a better way to do this?
>>>>
>>> I am not sure what Coccinelle is. If we had something automated that measures
>>> a boot time and if needed does some profiling it would be good. Otherwise it
>>> is a manual debugging mainly, IMHO.
>>
>> Paul, What about using a default-off kernel CONFIG that splats on all lazy
>> call_rcu() callbacks that do a wake up. We could use the trace hooks to do it
>> in kernel I think. I can talk to Steve to get ideas on how to do that but I
>> think it can be done purely from trace events (we might need a new
>> trace_end_invoke_callback to fire after the callback is invoked). Thoughts?
>
> Could you look for wakeups invoked between trace_rcu_batch_start() and
> trace_rcu_batch_end() that are not from interrupt context? This would
> of course need to be associated with a task rather than a CPU.
Yes this sounds good, but we also need to know if the callbacks are lazy or not since wake-up is ok from a non lazy one. I think I’ll need a table to track that at queuing time.
> Note that you would need to check for wakeups from interrupt handlers
> even with the extra trace_end_invoke_callback(). The window where an
> interrupt handler could do a wakeup would be reduced, but not eliminated.
True! Since this is a debugging option, can we not just disable interrupts across callback invocation?
Thanks,
- Joel
>
> Thanx, Paul
