linux-kernel - Re: [PATCH v6 1/4] rcu: Make call

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <8344B0AB-608E-44DA-8FEE-3FE56EDF9172@joelfernandes.org>
Date:   Mon, 26 Sep 2022 19:47:50 -0400
From:   Joel Fernandes <joel@...lfernandes.org>
To:     paulmck@...nel.org
Cc:     Uladzislau Rezki <urezki@...il.com>, rcu@...r.kernel.org,
        linux-kernel@...r.kernel.org, rushikesh.s.kadam@...el.com,
        neeraj.iitr10@...il.com, frederic@...nel.org, rostedt@...dmis.org
Subject: Re: [PATCH v6 1/4] rcu: Make call_rcu() lazy to save power



> On Sep 26, 2022, at 6:32 PM, Paul E. McKenney <paulmck@...nel.org> 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