Message-ID: <20190402024714.GA14213@innovation.ch>
Date:   Mon, 1 Apr 2019 19:47:14 -0700
From:   "Life is hard, and then you die" <ronald@...ovation.ch>
To:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Cc:     Dmitry Torokhov <dmitry.torokhov@...il.com>,
        Henrik Rydberg <rydberg@...math.org>,
        Andy Shevchenko <andriy.shevchenko@...ux.intel.com>,
        Sergey Senozhatsky <sergey.senozhatsky@...il.com>,
        Steven Rostedt <rostedt@...dmis.org>,
        "Rafael J. Wysocki" <rafael@...nel.org>,
        Lukas Wunner <lukas@...ner.de>,
        Federico Lorenzi <federico@...velground.com>,
        linux-input@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v3 3/4] driver core: add dev_print_hex_dump() logging
 function.


On Thu, Mar 28, 2019 at 12:29:52PM +0100, Greg Kroah-Hartman wrote:
> On Thu, Mar 28, 2019 at 03:27:55AM -0700, Life is hard, and then you die wrote:
> > 
> > On Thu, Mar 28, 2019 at 06:29:17AM +0100, Greg Kroah-Hartman wrote:
> > > On Wed, Mar 27, 2019 at 05:28:17PM -0700, Life is hard, and then you die wrote:
> > > > 
> > > > On Wed, Mar 27, 2019 at 11:37:57AM +0900, Greg Kroah-Hartman wrote:
> > > > > On Tue, Mar 26, 2019 at 06:48:06PM -0700, Ronald Tschalär wrote:
> > > > > > This is the dev_xxx() analog to print_hex_dump(), using dev_printk()
> > > > > > instead of straight printk() to match other dev_xxx() logging functions.
> > > > > > ---
> > > > > >  drivers/base/core.c    | 43 ++++++++++++++++++++++++++++++++++++++++++
> > > > > >  include/linux/device.h | 15 +++++++++++++++
> > > > > >  2 files changed, 58 insertions(+)
[snip]
> > > > > Anyway, no, please do not do this.  Please do not dump large hex values
> > > > > like this to the kernel log, it does not help anyone.
> > > > > 
> > > > > You can do this while debugging, sure, but not for "real" kernel code.
> > > > 
> > > > As used by this driver, it is definitely called for debugging only and
> > > > must be explicitly enabled via a module param. But having the ability
> > > > for folks to easily generate and print out debugging info has proven
> > > > quite valuable.
> > > 
> > > We have dynamic debugging, no need for module parameters at all.  This
> > > isn't the 1990's anymore :)
> > 
> > I am aware of dynamic debugging, but there are several issues with it
> > from the perspective of the logging I'm doing here (I mentioned these
> > in response to an earlier review already):
> > 
> >   1. Dynamic debugging can't be enabled at a function or line level on
> >      the kernel command line, so this means that to debug issues
> >      during boot you have to enable everything, which can be way too
> >      verbose
> 
> You can, and should enable it at a function or line level by writing to
> the proper kernel file in debugfs.
> 
> You have read Documentation/admin-guide/dynamic-debug-howto.rst, right?

Yes, and I've played with the parameters quite a bit.

> No need to do anything on the command line, that's so old-school :)

Sorry if I'm being unduly dense, but then how to enable debugging
during early boot? The only other alternative I see is modifying the
initrd, and asking folks to do that is noticeably more complicated
than having them add something to the command line in grub. So from my
perspective I find kernel params far from old-school :-)

> >   2. The expressions to enable the individual logging statements are
> >      quite brittle (in particular the line-based ones) since they
> >      (may) change any time the code is changed - having stable
> >      commands to give to users and put in documentation (e.g.
> >      "echo 0x200 > /sys/module/applespi/parameters/debug") is
> >      quite valuable.
> > 
> >      One way to work around this would be to put every single logging
> >      statement in a function of its own, thereby ensuring a stable
> >      name is associated with each one.
> 
> Again, read the documentation, this works today as-is.

I have read it (we're talking about the dynamic debug docs here), but
I just don't see how it addresses this in any way.

> >   3. In at least two places we log different types of packets in the
> >      same lines of code (protected by a "if (log_mask & PKT_TYPE)") -
> >      dynamic-debug would only allow enabling or disabling logging of
> >      all packets. This could be worked around by creating a separate
> >      (but essentially duplicate) logging function for each packet type
> >      and some lookup table to call the appropriate one. Not very
> >      pretty IMO, though.
> 
> True, but you can use tracepoints as well, that would probably be much
> easier when you are logging data streams.  You can also use an ebpf
> program with the tracepoints to log just what you need/want to when you
> want to as well.

Thanks for the hint, I hadn't paid much attention to tracepoints till
now. They do solve most of these issues I've mentioned here, though.
So I've actually gone and implemented the tracing as tracepoints now.
Two issues I noticed though:

 1. since filters can't be enabled on the command line (and yes, we
    seem to disagree on the usefulness of the command line) it means
    I had to essentially create tracepoints for every trace+filter
    combo I may want to enable (in my case it's just one field, so I
    ended up with 8 tracepoints instead of 1). Not a big deal in my
    case, but could get ugly.

 2. in cases where there is relevant log output too, folks have to be
    told to provide both trace and dmesg output, the outputs have to
    merged back together again. Though I note that with the use of the
    tp_printk kernel param this inconvenience goes away again (but
    we're back to logging the traces in the kernel log :-) ).

> >   4. In a couple places we log both the sent command and the received
> >      response, with the log-mask determining for which types of
> >      packets to do this. With a log mask there is one bit to set to
> >      get both logged; with dynamic debugging you'd have to enable the
> >      writing and receiving parts separately (not a huge deal, but yet
> >      one place where things are a bit more complicated than
> >      necessary).
> > 
> > Except for the first, none of these are insurmountable, but together
> > they don't make dynamic debugging very attractive for this sort of
> > logging.
> 
> Look into it some more, we have all of this covered today, no need for
> just a single driver to do something fancy on its own :)

The tracepoints do indeed cover this too.


  Cheers,

  Ronald

Powered by blists - more mailing lists