Message-ID: <20171205155537.w6hohhbvjest3pbm@pathway.suse.cz>
Date:   Tue, 5 Dec 2017 16:55:37 +0100
From:   Petr Mladek <pmladek@...e.com>
To:     Fengguang Wu <fengguang.wu@...el.com>
Cc:     Kevin Hilman <khilman@...libre.com>,
        Mark Brown <broonie@...nel.org>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Sergey Senozhatsky <sergey.senozhatsky@...il.com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Linus Torvalds <torvalds@...ux-foundation.org>
Subject: Re: kernel CI: printk loglevels in kernel boot logs?

On Wed 2017-11-22 22:10:58, Fengguang Wu wrote:
> On Wed, Nov 22, 2017 at 12:34:48PM +0100, Petr Mladek wrote:
> > On Wed 2017-11-22 11:27:02, Fengguang Wu wrote:
> > > [CC LKML for possible printk improvements]
> > > 
> > > On Wed, Nov 22, 2017 at 09:56:10AM +0800, Fengguang Wu wrote:
> > > > Hi Kevin,
> > > >
> > > > On Tue, Nov 21, 2017 at 12:27:48PM -0800, Kevin Hilman wrote:
> > > > > Hi Fenguang,
> > > > >
> > > > > In automated testing, for ease of parsing kernel boot logs (especially
> > > > > separating warnings and errors from debug, info etc.)
> > 
> > Would be enough to use console_loglevel to filter out the non-interesting
> > messages?
> 
> Nah we do prefer and actually require keeping the full dmesg. For one
> thing, it's good to have the context information. The 0day test farm
> will dump logs and debug info to the serial console. And we'll parse
> the normal dmesg lines to detect when the kernel boots up and
> shutdown (to split the continuous serial console log into dmesg files
> for each kernel boot), when the initramfs is loading (to judge whether
> the kernel oops is related to the initrd), when the testing begins,
> etc. For embedded systems people may even control machine boot and
> collect artifacts via the serial console.

OK, I understand why you would need this feature.

> > > > > Right now we can get this info from a "dmesg --raw" after bootup, but
> > > > > it would be really nice in certain automation frameworks to have a
> > > > > kernel command-line option to enable printing of loglevels in default
> > > > > boot log.
> > > >
> > > > Agreed.
> > > >
> > > > > This is espeically useful when ingesting kernel logs into advanced
> > > > > search/analytics frameworks (I'm playing with and ELK stack: Elastic
> > > > > Search, Logstash, Kibana).
> > 
> > This suggests that people are already filtering the kernel log using
> > external tools. Therefore there is a danger that the format might
> > have already became the kernel ABI and we could not change it easily.
> > 
> > A solution would be to printk the loglevel after the timestamp.
> > But this would be ugly because dmesg --raw prints it at the beginning.
> > 
> > To make it clear. I understand that "show_loglevel" command line argument
> > would be useful for you. But I am afraid that it is not worth changing
> > the format. There would need to be wide interest into the change.
> > Also there would need to be evidence that the existing solutions
> > (dmesg --raw, console_loglevel) are not enough in many real life
> > scenarios.
> 
> As Sergey said in the other thread, it should be safe because the
> syslog format will be explicitly enabled by those who want and ready
> to parse it.

I would have agreed with this argument few weeks ago but I do not get
it any longer. If there are tools depending on the message format
and people enable the extended format, the tools might stop working.

You might argue that these people are shooting into their legs.
But other people might argue that kernel changed ABI and broke
userspace tools. Note that these tools do not have any chance
to force kernel using the old format, they are consumers.

It might be especially annoying if these tools are part of some
complex infrastructure that is used to access the boot logs
and admins get blind. Then it might be hard for them to debug
the problem, especially if they got this feature as part of
a bigger system update.

Am I too paranoid? Steven? Linus?

> > > [...]
> > > 
> > > > As you may know I'm currently reporting kernel oops in mainline
> > > > kernel, hoping to clear noisy oops there -- they obviously hurt
> > > > catching and bisecting new oops.
> > > >
> > > > I have to say the warnings are much more noisy than kernel oops
> > > > in 2 ways:
> > > >
> > > > 1) It's common for a *normal* kernel boot to have a dozen of old
> > > > warnings.
> > 
> > I guess that you solve this by a "clever" diff between old and new
> > log.
> 
> That does help filtering out the majority old warnings. For example,
> if it's an inherent HW issue, it may remain there forever and never
> show up after diff.
> 
> However some noises will still go through and the number is large
> enough to make "tester side blacklist" a big maintenance headache.
> As you may see from the listing of warning messages waiting for bisect
> in 0day now.
> 
> > The extra info suggested below might help to filter out
> > some messages. But the problems will still be there.
> > I am not sure that the categories are worth it, see below.
> > 
> > 
> > > > 2) Many warnings do not necessarily mean something should or could be
> > > > fixed -- they may well be mentioning some HW problem, or an alert
> > > > message to the user.
> > > >
> > > > So there is a much bigger and messy problem than catching the warnings:
> > > > ways to effectively mark or filter real warnings that automated testing
> > > > should catch and report.
> > > >
> > > > For filtering, we currently accumulated the below blacklist:
> > > >
> > > > https://github.com/fengguang/lkp-tests/blob/master/etc/kmsg-blacklist
> > > >
> > > [...]
> > > 
> > > For the marking part, I wonder if there can be a clear rule that
> > > allows developers to distinguish 2 kind of information for users and
> > > testers:
> > > 
> > > - "bug" message: indicating a regression that should be reported and fixed
> > > 
> > > - "fact" message: well there's a problem, but we kernel developers
> > >  probably can do nothing about it. It's just to let the user know
> > >  about the fact. The fix might be something like replacing a broken
> > >  disk drive.
> > > 
> > > Those message types are orthogonal to severity of the problem (log
> > > levels), so the current log levels are actually not sufficient for
> > > distinguishing these kind of situations.
> > 
> > It is interesting idea but I am afraid that it would not work
> > in reality from several reasons:
> > 
> >  + It is yet another complexity. I am not sure that developers would
> >    want to think about it when creating new messages.
> 
> [moved ahead the below 2 points]
> >  + The proposal it rather strigthforward. But does all messages
> >    clearly fit into one of the two categories? What about
> >    messages that are not about problems? For example, most
> >    of pr_info(), pr_notice() messages?
> > 
> > 
> >  + Are the two categories worth the effort? Especially when the
> >    second one might still eventually be a kernel bug?
> > 
> >    If we add more categories, it might get too complicated.
> >    For example, like setting severities and priorities
> >    in the classic bugzilla.
> > 
> 
> Let me answer the 3 points together. Sorry if I was not clear -- I
> merely give the 2 categories as example. However the key point is not
> about creating categories. My proposal would be something like this:
> 
>        #define KERNEL_BUG "[kernel bug]"
> 
> Then use that tag wherever maintainers want to track regressions:
> 
>        pr_err(KERNEL_BUG "alg: Unexpected test result for %s: %d\n", name, err);

We already have something like this, see include/linux/printk.h

#define FW_BUG		"[Firmware Bug]: "
#define FW_WARN		"[Firmware Warn]: "
#define FW_INFO		"[Firmware Info]: "
#define HW_ERR		"[Hardware Error]: "

IMHO, the last one is a good example. It was added 7 years ago
by the commit c6de9f08912311ddc1 ("x86, mce: Add HW_ERR printk prefix
for hardware error logging" and it is still used only in two
mce-related files:

arch/x86/kernel/cpu/mcheck/mce.c
drivers/edac/mce_amd.c

I guess that this prefix would make sense in many more messages
but people do not use it. IMHO, they are either not aware of it
or they simply do not care.

IMHO, this is a good example that this approach does not work.

It seems that Andrew shared my skepticism when HW_ERR was introduced.
See the sentence "There have been various proposals to improve
the usefulness and to rationally categorise things in way which are
more useful to operators, but nothing seems to ever get over the
line." at
https://lkml.kernel.org/r/20100610164607.b2908334.akpm@linux-foundation.org

IMHO, we would need to come up with another approach that would
be easy to understand and could be "more" enforced.

Best Regards,
Petr

Powered by blists - more mailing lists